initial commit, 4.5 stable

thirdparty/libtheora/COPYING

@@ -0,0 +1,28 @@
+Copyright (C) 2002-2009 Xiph.org Foundation
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+
+- Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+
+- Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+
+- Neither the name of the Xiph.org Foundation nor the names of its
+contributors may be used to endorse or promote products derived from
+this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION
+OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

thirdparty/libtheora/LICENSE

@@ -0,0 +1,18 @@
+Please see the file COPYING for the copyright license for this software.
+
+In addition to and irrespective of the copyright license associated
+with this software, On2 Technologies, Inc. makes the following statement
+regarding technology used in this software:
+
+  On2 represents and warrants that it shall not assert any rights
+  relating to infringement of On2's registered patents, nor initiate
+  any litigation asserting such rights, against any person who, or
+  entity which utilizes the On2 VP3 Codec Software, including any
+  use, distribution, and sale of said Software; which make changes,
+  modifications, and improvements in said Software; and to use,
+  distribute, and sell said changes as well as applications for other
+  fields of use.
+
+This reference implementation is originally derived from the On2 VP3
+Codec Software, and the Theora video format is essentially compatible
+with the VP3 video format, consisting of a backward-compatible superset.

thirdparty/libtheora/apiwrapper.c

@@ -0,0 +1,166 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+#include <stdlib.h>
+#include <string.h>
+#include <limits.h>
+#include "apiwrapper.h"
+
+
+
+const char *theora_version_string(void){
+  return th_version_string();
+}
+
+ogg_uint32_t theora_version_number(void){
+  return th_version_number();
+}
+
+void theora_info_init(theora_info *_ci){
+  memset(_ci,0,sizeof(*_ci));
+}
+
+void theora_info_clear(theora_info *_ci){
+  th_api_wrapper *api;
+  api=(th_api_wrapper *)_ci->codec_setup;
+  memset(_ci,0,sizeof(*_ci));
+  if(api!=NULL){
+    if(api->clear!=NULL)(*api->clear)(api);
+    _ogg_free(api);
+  }
+}
+
+void theora_clear(theora_state *_th){
+  /*Provide compatibility with mixed encoder and decoder shared lib versions.*/
+  if(_th->internal_decode!=NULL){
+    (*((oc_state_dispatch_vtable *)_th->internal_decode)->clear)(_th);
+  }
+  if(_th->internal_encode!=NULL){
+    (*((oc_state_dispatch_vtable *)_th->internal_encode)->clear)(_th);
+  }
+  if(_th->i!=NULL)theora_info_clear(_th->i);
+  memset(_th,0,sizeof(*_th));
+}
+
+int theora_control(theora_state *_th,int _req,void *_buf,size_t _buf_sz){
+  /*Provide compatibility with mixed encoder and decoder shared lib versions.*/
+  if(_th->internal_decode!=NULL){
+    return (*((oc_state_dispatch_vtable *)_th->internal_decode)->control)(_th,
+     _req,_buf,_buf_sz);
+  }
+  else if(_th->internal_encode!=NULL){
+    return (*((oc_state_dispatch_vtable *)_th->internal_encode)->control)(_th,
+     _req,_buf,_buf_sz);
+  }
+  else return TH_EINVAL;
+}
+
+ogg_int64_t theora_granule_frame(theora_state *_th,ogg_int64_t _gp){
+  /*Provide compatibility with mixed encoder and decoder shared lib versions.*/
+  if(_th->internal_decode!=NULL){
+    return (*((oc_state_dispatch_vtable *)_th->internal_decode)->granule_frame)(
+     _th,_gp);
+  }
+  else if(_th->internal_encode!=NULL){
+    return (*((oc_state_dispatch_vtable *)_th->internal_encode)->granule_frame)(
+     _th,_gp);
+  }
+  else return -1;
+}
+
+double theora_granule_time(theora_state *_th, ogg_int64_t _gp){
+  /*Provide compatibility with mixed encoder and decoder shared lib versions.*/
+  if(_th->internal_decode!=NULL){
+    return (*((oc_state_dispatch_vtable *)_th->internal_decode)->granule_time)(
+     _th,_gp);
+  }
+  else if(_th->internal_encode!=NULL){
+    return (*((oc_state_dispatch_vtable *)_th->internal_encode)->granule_time)(
+     _th,_gp);
+  }
+  else return -1;
+}
+
+void oc_theora_info2th_info(th_info *_info,const theora_info *_ci){
+  _info->version_major=_ci->version_major;
+  _info->version_minor=_ci->version_minor;
+  _info->version_subminor=_ci->version_subminor;
+  _info->frame_width=_ci->width;
+  _info->frame_height=_ci->height;
+  _info->pic_width=_ci->frame_width;
+  _info->pic_height=_ci->frame_height;
+  _info->pic_x=_ci->offset_x;
+  _info->pic_y=_ci->offset_y;
+  _info->fps_numerator=_ci->fps_numerator;
+  _info->fps_denominator=_ci->fps_denominator;
+  _info->aspect_numerator=_ci->aspect_numerator;
+  _info->aspect_denominator=_ci->aspect_denominator;
+  switch(_ci->colorspace){
+    case OC_CS_ITU_REC_470M:_info->colorspace=TH_CS_ITU_REC_470M;break;
+    case OC_CS_ITU_REC_470BG:_info->colorspace=TH_CS_ITU_REC_470BG;break;
+    default:_info->colorspace=TH_CS_UNSPECIFIED;break;
+  }
+  switch(_ci->pixelformat){
+    case OC_PF_420:_info->pixel_fmt=TH_PF_420;break;
+    case OC_PF_422:_info->pixel_fmt=TH_PF_422;break;
+    case OC_PF_444:_info->pixel_fmt=TH_PF_444;break;
+    default:_info->pixel_fmt=TH_PF_RSVD;
+  }
+  _info->target_bitrate=_ci->target_bitrate;
+  _info->quality=_ci->quality;
+  _info->keyframe_granule_shift=_ci->keyframe_frequency_force>0?
+   OC_MINI(31,oc_ilog(_ci->keyframe_frequency_force-1)):0;
+}
+
+int theora_packet_isheader(ogg_packet *_op){
+  return th_packet_isheader(_op);
+}
+
+int theora_packet_iskeyframe(ogg_packet *_op){
+  return th_packet_iskeyframe(_op);
+}
+
+int theora_granule_shift(theora_info *_ci){
+  /*This breaks when keyframe_frequency_force is not positive or is larger than
+     2**31 (if your int is more than 32 bits), but that's what the original
+     function does.*/
+  return oc_ilog(_ci->keyframe_frequency_force-1);
+}
+
+void theora_comment_init(theora_comment *_tc){
+  th_comment_init((th_comment *)_tc);
+}
+
+char *theora_comment_query(theora_comment *_tc,char *_tag,int _count){
+  return th_comment_query((th_comment *)_tc,_tag,_count);
+}
+
+int theora_comment_query_count(theora_comment *_tc,char *_tag){
+  return th_comment_query_count((th_comment *)_tc,_tag);
+}
+
+void theora_comment_clear(theora_comment *_tc){
+  th_comment_clear((th_comment *)_tc);
+}
+
+void theora_comment_add(theora_comment *_tc,char *_comment){
+  th_comment_add((th_comment *)_tc,_comment);
+}
+
+void theora_comment_add_tag(theora_comment *_tc, char *_tag, char *_value){
+  th_comment_add_tag((th_comment *)_tc,_tag,_value);
+}

thirdparty/libtheora/apiwrapper.h

@@ -0,0 +1,54 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+#if !defined(_apiwrapper_H)
+# define _apiwrapper_H (1)
+# include <ogg/ogg.h>
+# include <theora/theora.h>
+# include "theora/theoradec.h"
+# include "theora/theoraenc.h"
+# include "state.h"
+
+typedef struct th_api_wrapper th_api_wrapper;
+typedef struct th_api_info    th_api_info;
+
+/*Provide an entry point for the codec setup to clear itself in case we ever
+   want to break pieces off into a common base library shared by encoder and
+   decoder.
+  In addition, this makes several other pieces of the API wrapper cleaner.*/
+typedef void (*oc_setup_clear_func)(void *_ts);
+
+/*Generally only one of these pointers will be non-NULL in any given instance.
+  Technically we do not even really need this struct, since we should be able
+   to figure out which one from "context", but doing it this way makes sure we
+   don't flub it up.*/
+struct th_api_wrapper{
+  oc_setup_clear_func  clear;
+  th_setup_info       *setup;
+  th_dec_ctx          *decode;
+  th_enc_ctx          *encode;
+};
+
+struct th_api_info{
+  th_api_wrapper api;
+  theora_info    info;
+};
+
+
+void oc_theora_info2th_info(th_info *_info,const theora_info *_ci);
+
+#endif

thirdparty/libtheora/bitpack.c

@@ -0,0 +1,114 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE OggTheora SOURCE CODE IS (C) COPYRIGHT 1994-2009             *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function: packing variable sized words into an octet stream
+
+ ********************************************************************/
+#include <string.h>
+#include <stdlib.h>
+#include "bitpack.h"
+
+/*We're 'MSb' endian; if we write a word but read individual bits,
+   then we'll read the MSb first.*/
+
+void oc_pack_readinit(oc_pack_buf *_b,unsigned char *_buf,long _bytes){
+  memset(_b,0,sizeof(*_b));
+  _b->ptr=_buf;
+  _b->stop=_buf+_bytes;
+}
+
+static oc_pb_window oc_pack_refill(oc_pack_buf *_b,int _bits){
+  const unsigned char *ptr;
+  const unsigned char *stop;
+  oc_pb_window         window;
+  int                  available;
+  unsigned             shift;
+  stop=_b->stop;
+  ptr=_b->ptr;
+  window=_b->window;
+  available=_b->bits;
+  shift=OC_PB_WINDOW_SIZE-available;
+  while(7<shift&&ptr<stop){
+    shift-=8;
+    window|=(oc_pb_window)*ptr++<<shift;
+  }
+  _b->ptr=ptr;
+  available=OC_PB_WINDOW_SIZE-shift;
+  if(_bits>available){
+    if(ptr>=stop){
+      _b->eof=1;
+      available=OC_LOTS_OF_BITS;
+    }
+    else window|=*ptr>>(available&7);
+  }
+  _b->bits=available;
+  return window;
+}
+
+int oc_pack_look1(oc_pack_buf *_b){
+  oc_pb_window window;
+  int          available;
+  window=_b->window;
+  available=_b->bits;
+  if(available<1)_b->window=window=oc_pack_refill(_b,1);
+  return window>>OC_PB_WINDOW_SIZE-1;
+}
+
+void oc_pack_adv1(oc_pack_buf *_b){
+  _b->window<<=1;
+  _b->bits--;
+}
+
+/*Here we assume that 0<=_bits&&_bits<=32.*/
+long oc_pack_read_c(oc_pack_buf *_b,int _bits){
+  oc_pb_window window;
+  int          available;
+  long         result;
+  window=_b->window;
+  available=_b->bits;
+  if(_bits==0)return 0;
+  if(available<_bits){
+    window=oc_pack_refill(_b,_bits);
+    available=_b->bits;
+  }
+  result=window>>OC_PB_WINDOW_SIZE-_bits;
+  available-=_bits;
+  window<<=1;
+  window<<=_bits-1;
+  _b->window=window;
+  _b->bits=available;
+  return result;
+}
+
+int oc_pack_read1_c(oc_pack_buf *_b){
+  oc_pb_window window;
+  int          available;
+  int          result;
+  window=_b->window;
+  available=_b->bits;
+  if(available<1){
+    window=oc_pack_refill(_b,1);
+    available=_b->bits;
+  }
+  result=window>>OC_PB_WINDOW_SIZE-1;
+  available--;
+  window<<=1;
+  _b->window=window;
+  _b->bits=available;
+  return result;
+}
+
+long oc_pack_bytes_left(oc_pack_buf *_b){
+  if(_b->eof)return -1;
+  return _b->stop-_b->ptr+(_b->bits>>3);
+}

thirdparty/libtheora/bitpack.h

@@ -0,0 +1,76 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE OggTheora SOURCE CODE IS (C) COPYRIGHT 1994-2009             *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function: packing variable sized words into an octet stream
+
+ ********************************************************************/
+#if !defined(_bitpack_H)
+# define _bitpack_H (1)
+# include <stddef.h>
+# include <limits.h>
+# include "internal.h"
+
+
+
+typedef size_t             oc_pb_window;
+typedef struct oc_pack_buf oc_pack_buf;
+
+
+
+/*Custom bitpacker implementations.*/
+# if defined(OC_ARM_ASM)
+#  include "arm/armbits.h"
+# endif
+
+# if !defined(oc_pack_read)
+#  define oc_pack_read oc_pack_read_c
+# endif
+# if !defined(oc_pack_read1)
+#  define oc_pack_read1 oc_pack_read1_c
+# endif
+# if !defined(oc_huff_token_decode)
+#  define oc_huff_token_decode oc_huff_token_decode_c
+# endif
+
+# define OC_PB_WINDOW_SIZE ((int)sizeof(oc_pb_window)*CHAR_BIT)
+/*This is meant to be a large, positive constant that can still be efficiently
+   loaded as an immediate (on platforms like ARM, for example).
+  Even relatively modest values like 100 would work fine.*/
+# define OC_LOTS_OF_BITS (0x40000000)
+
+
+
+struct oc_pack_buf{
+  const unsigned char *stop;
+  const unsigned char *ptr;
+  oc_pb_window         window;
+  int                  bits;
+  int                  eof;
+};
+
+void oc_pack_readinit(oc_pack_buf *_b,unsigned char *_buf,long _bytes);
+int oc_pack_look1(oc_pack_buf *_b);
+void oc_pack_adv1(oc_pack_buf *_b);
+/*Here we assume 0<=_bits&&_bits<=32.*/
+long oc_pack_read_c(oc_pack_buf *_b,int _bits);
+int oc_pack_read1_c(oc_pack_buf *_b);
+/* returns -1 for read beyond EOF, or the number of whole bytes available */
+long oc_pack_bytes_left(oc_pack_buf *_b);
+
+/*These two functions are implemented locally in huffdec.c*/
+/*Read in bits without advancing the bitptr.
+  Here we assume 0<=_bits&&_bits<=32.*/
+/*static int oc_pack_look(oc_pack_buf *_b,int _bits);*/
+/*static void oc_pack_adv(oc_pack_buf *_b,int _bits);*/
+
+#endif

thirdparty/libtheora/collect.c

@@ -0,0 +1,974 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2011                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function: mode selection code
+
+ ********************************************************************/
+#include <stdio.h>
+#include <limits.h>
+#include <math.h>
+#include <string.h>
+#include "collect.h"
+
+#if defined(OC_COLLECT_METRICS)
+
+int              OC_HAS_MODE_METRICS;
+double           OC_MODE_RD_WEIGHT_SATD[OC_LOGQ_BINS][3][2][OC_COMP_BINS];
+double           OC_MODE_RD_WEIGHT_SAD[OC_LOGQ_BINS][3][2][OC_COMP_BINS];
+oc_mode_metrics  OC_MODE_METRICS_SATD[OC_LOGQ_BINS-1][3][2][OC_COMP_BINS];
+oc_mode_metrics  OC_MODE_METRICS_SAD[OC_LOGQ_BINS-1][3][2][OC_COMP_BINS];
+const char      *OC_MODE_METRICS_FILENAME="modedec.stats";
+
+void oc_mode_metrics_add(oc_mode_metrics *_metrics,
+ double _w,int _s,int _q,int _r,double _d){
+  if(_metrics->w>0){
+    double ds;
+    double dq;
+    double dr;
+    double dd;
+    double ds2;
+    double dq2;
+    double s2;
+    double sq;
+    double q2;
+    double sr;
+    double qr;
+    double sd;
+    double qd;
+    double s2q;
+    double sq2;
+    double w;
+    double wa;
+    double rwa;
+    double rwa2;
+    double rwb;
+    double rwb2;
+    double rw2;
+    double rw3;
+    double rw4;
+    wa=_metrics->w;
+    ds=_s-_metrics->s/wa;
+    dq=_q-_metrics->q/wa;
+    dr=_r-_metrics->r/wa;
+    dd=_d-_metrics->d/wa;
+    ds2=ds*ds;
+    dq2=dq*dq;
+    s2=_metrics->s2;
+    sq=_metrics->sq;
+    q2=_metrics->q2;
+    sr=_metrics->sr;
+    qr=_metrics->qr;
+    sd=_metrics->sd;
+    qd=_metrics->qd;
+    s2q=_metrics->s2q;
+    sq2=_metrics->sq2;
+    w=wa+_w;
+    rwa=wa/w;
+    rwb=_w/w;
+    rwa2=rwa*rwa;
+    rwb2=rwb*rwb;
+    rw2=wa*rwb;
+    rw3=rw2*(rwa2-rwb2);
+    rw4=_w*rwa2*rwa2+wa*rwb2*rwb2;
+    _metrics->s2q2+=-2*(ds*sq2+dq*s2q)*rwb
+     +(ds2*q2+4*ds*dq*sq+dq2*s2)*rwb2+ds2*dq2*rw4;
+    _metrics->s2q+=(-2*ds*sq-dq*s2)*rwb+ds2*dq*rw3;
+    _metrics->sq2+=(-ds*q2-2*dq*sq)*rwb+ds*dq2*rw3;
+    _metrics->sqr+=(-ds*qr-dq*sr-dr*sq)*rwb+ds*dq*dr*rw3;
+    _metrics->sqd+=(-ds*qd-dq*sd-dd*sq)*rwb+ds*dq*dd*rw3;
+    _metrics->s2+=ds2*rw2;
+    _metrics->sq+=ds*dq*rw2;
+    _metrics->q2+=dq2*rw2;
+    _metrics->sr+=ds*dr*rw2;
+    _metrics->qr+=dq*dr*rw2;
+    _metrics->r2+=dr*dr*rw2;
+    _metrics->sd+=ds*dd*rw2;
+    _metrics->qd+=dq*dd*rw2;
+    _metrics->d2+=dd*dd*rw2;
+  }
+  _metrics->w+=_w;
+  _metrics->s+=_s*_w;
+  _metrics->q+=_q*_w;
+  _metrics->r+=_r*_w;
+  _metrics->d+=_d*_w;
+}
+
+void oc_mode_metrics_merge(oc_mode_metrics *_dst,
+ const oc_mode_metrics *_src,int _n){
+  int i;
+  /*Find a non-empty set of metrics.*/
+  for(i=0;i<_n&&_src[i].w==0;i++);
+  if(i>=_n){
+    memset(_dst,0,sizeof(*_dst));
+    return;
+  }
+  memcpy(_dst,_src+i,sizeof(*_dst));
+  /*And iterate over the remaining non-empty sets of metrics.*/
+  for(i++;i<_n;i++)if(_src[i].w!=0){
+    double ds;
+    double dq;
+    double dr;
+    double dd;
+    double ds2;
+    double dq2;
+    double s2a;
+    double s2b;
+    double sqa;
+    double sqb;
+    double q2a;
+    double q2b;
+    double sra;
+    double srb;
+    double qra;
+    double qrb;
+    double sda;
+    double sdb;
+    double qda;
+    double qdb;
+    double s2qa;
+    double s2qb;
+    double sq2a;
+    double sq2b;
+    double w;
+    double wa;
+    double wb;
+    double rwa;
+    double rwb;
+    double rwa2;
+    double rwb2;
+    double rw2;
+    double rw3;
+    double rw4;
+    wa=_dst->w;
+    wb=_src[i].w;
+    ds=_src[i].s/wb-_dst->s/wa;
+    dq=_src[i].q/wb-_dst->q/wa;
+    dr=_src[i].r/wb-_dst->r/wa;
+    dd=_src[i].d/wb-_dst->d/wa;
+    ds2=ds*ds;
+    dq2=dq*dq;
+    s2a=_dst->s2;
+    sqa=_dst->sq;
+    q2a=_dst->q2;
+    sra=_dst->sr;
+    qra=_dst->qr;
+    sda=_dst->sd;
+    qda=_dst->qd;
+    s2qa=_dst->s2q;
+    sq2a=_dst->sq2;
+    s2b=_src[i].s2;
+    sqb=_src[i].sq;
+    q2b=_src[i].q2;
+    srb=_src[i].sr;
+    qrb=_src[i].qr;
+    sdb=_src[i].sd;
+    qdb=_src[i].qd;
+    s2qb=_src[i].s2q;
+    sq2b=_src[i].sq2;
+    w=wa+wb;
+    if(w==0)rwa=rwb=0;
+    else{
+      rwa=wa/w;
+      rwb=wb/w;
+    }
+    rwa2=rwa*rwa;
+    rwb2=rwb*rwb;
+    rw2=wa*rwb;
+    rw3=rw2*(rwa2-rwb2);
+    rw4=wb*rwa2*rwa2+wa*rwb2*rwb2;
+    /*
+    (1,1,1) ->
+     (0,0,0)#
+     (1,0,0) C(1,1)*C(1,0)*C(1,0)->  d^{1,0,0}*(rwa*B_{0,1,1}-rwb*A_{0,1,1})
+     (0,1,0) C(1,0)*C(1,1)*C(1,0)->  d^{0,1,0}*(rwa*B_{1,0,1}-rwb*A_{1,0,1})
+     (0,0,1) C(1,0)*C(1,0)*C(1,1)->  d^{0,0,1}*(rwa*B_{1,1,0}-rwb*A_{1,1,0})
+     (1,1,0)*
+     (1,0,1)*
+     (0,1,1)*
+     (1,1,1) C(1,1)*C(1,1)*C(1,1)->  d^{1,1,1}*(rwa^3*wb-rwb^3*wa)
+    (2,1) ->
+     (0,0)#
+     (1,0) C(2,1)*C(1,1)->2*d^{1,0}*(rwa*B_{1,1}-rwb*A_{1,1})
+     (0,1) C(2,0)*C(1,1)->  d^{0,1}*(rwa*B_{2,0}-rwb*A_{2,0})
+     (2,0)*
+     (1,1)*
+     (2,1) C(2,2)*C(1,1)->  d^{2,1}*(rwa^3*wb-rwb^3*wa)
+    (2,2) ->
+     (0,0)#
+     (1,0) C(2,1)*C(2,0)->2*d^{1,0}*(rwa*B_{1,2}-rwb*A_{1,2})
+     (0,1) C(2,0)*C(2,1)->2*d^{0,1}*(rwa*B_{2,1}-rwb*A_{2,1})
+     (2,0) C(2,2)*C(2,0)->  d^{2,0}*(rwa^2*B_{0,2}+rwb^2*A_{0,2})
+     (1,1) C(2,1)*C(2,1)->4*d^{1,1}*(rwa^2*B_{1,1}+rwb^2*A_{1,1})
+     (0,2) C(2,0)*C(2,2)->  d^{0,2}*(rwa^2*B_{2,0}+rwb^2*A_{2,0})
+     (1,2)*
+     (2,1)*
+     (2,2) C(2,2)*C(2,2)*d^{2,2}*(rwa^4*wb+rwb^4*wa)
+    */
+    _dst->s2q2+=_src[i].s2q2+2*(ds*(rwa*sq2b-rwb*sq2a)+dq*(rwa*s2qb-rwb*s2qa))
+     +ds2*(rwa2*q2b+rwb2*q2a)+4*ds*dq*(rwa2*sqb+rwb2*sqa)
+     +dq2*(rwa2*s2b+rwb2*s2a)+ds2*dq2*rw4;
+    _dst->s2q+=_src[i].s2q+2*ds*(rwa*sqb-rwb*sqa)
+     +dq*(rwa*s2b-rwb*s2a)+ds2*dq*rw3;
+    _dst->sq2+=_src[i].sq2+ds*(rwa*q2b-rwb*q2a)
+     +2*dq*(rwa*sqb-rwb*sqa)+ds*dq2*rw3;
+    _dst->sqr+=_src[i].sqr+ds*(rwa*qrb-rwb*qra)+dq*(rwa*srb-rwb*sra)
+     +dr*(rwa*sqb-rwb*sqa)+ds*dq*dr*rw3;
+    _dst->sqd+=_src[i].sqd+ds*(rwa*qdb-rwb*qda)+dq*(rwa*sdb-rwb*sda)
+     +dd*(rwa*sqb-rwb*sqa)+ds*dq*dd*rw3;
+    _dst->s2+=_src[i].s2+ds2*rw2;
+    _dst->sq+=_src[i].sq+ds*dq*rw2;
+    _dst->q2+=_src[i].q2+dq2*rw2;
+    _dst->sr+=_src[i].sr+ds*dr*rw2;
+    _dst->qr+=_src[i].qr+dq*dr*rw2;
+    _dst->r2+=_src[i].r2+dr*dr*rw2;
+    _dst->sd+=_src[i].sd+ds*dd*rw2;
+    _dst->qd+=_src[i].qd+dq*dd*rw2;
+    _dst->d2+=_src[i].d2+dd*dd*rw2;
+    _dst->w+=_src[i].w;
+    _dst->s+=_src[i].s;
+    _dst->q+=_src[i].q;
+    _dst->r+=_src[i].r;
+    _dst->d+=_src[i].d;
+  }
+}
+
+/*Adjust a single corner of a set of metric bins to minimize the squared
+   prediction error of R and D.
+  Each bin is assumed to cover a quad like so:
+    (s0,q0)    (s1,q0)
+       A----------B
+       |          |
+       |          |
+       |          |
+       |          |
+       C----------Z
+    (s0,q1)    (s1,q1)
+  The values A, B, and C are fixed, and Z is the free parameter.
+  Then, for example, R_i is predicted via bilinear interpolation as
+    x_i=(s_i-s0)/(s1-s0)
+    y_i=(q_i-q0)/(q1-q0)
+    dRds1_i=A+(B-A)*x_i
+    dRds2_i=C+(Z-C)*x_i
+    R_i=dRds1_i+(dRds2_i-dRds1_i)*y_i
+  To find the Z that minimizes the squared prediction error over i, this can
+   be rewritten as
+    R_i-(A+(B-A)*x_i+(C-A)*y_i+(A-B-C)*x_i*y_i)=x_i*y_i*Z
+  Letting X={...,x_i*y_i,...}^T and
+   Y={...,R_i-(A+(B-A)*x_i+(C-A)*y_i+(A-B-C)*x_i*y_i),...}^T,
+   the optimal Z is given by Z=(X^T.Y)/(X^T.X).
+  Now, we need to compute these dot products without actually storing data for
+   each sample.
+  Starting with X^T.X, we have
+   X^T.X = sum(x_i^2*y_i^2) = sum((s_i-s0)^2*(q_i-q0)^2)/((s1-s0)^2*(q1-q0)^2).
+  Expanding the interior of the sum in a monomial basis of s_i and q_i gives
+    s0^2*q0^2  *(1)
+     -2*s0*q0^2*(s_i)
+     -2*s0^2*q0*(q_i)
+     +q0^2     *(s_i^2)
+     +4*s0*q0  *(s_i*q_i)
+     +s0^2     *(q_i^2)
+     -2*q0     *(s_i^2*q_i)
+     -2*s0     *(s_i*q_i^2)
+     +1        *(s_i^2*q_i^2).
+  However, computing things directly in this basis leads to gross numerical
+   errors, as most of the terms will have similar size and destructive
+   cancellation results.
+  A much better basis is the central (co-)moment basis:
+    {1,s_i-sbar,q_i-qbar,(s_i-sbar)^2,(s_i-sbar)*(q_i-qbar),(q_i-qbar)^2,
+     (s_i-sbar)^2*(q_i-qbar),(s_i-sbar)*(q_i-qbar)^2,(s_i-sbar)^2*(q_i-qbar)^2},
+   where sbar and qbar are the average s and q values over the bin,
+   respectively.
+  In that basis, letting ds=sbar-s0 and dq=qbar-q0, (s_i-s0)^2*(q_i-q0)^2 is
+    ds^2*dq^2*(1)
+     +dq^2   *((s_i-sbar)^2)
+     +4*ds*dq*((s_i-sbar)*(q_i-qbar))
+     +ds^2   *((q_i-qbar)^2)
+     +2*dq   *((s_i-sbar)^2*(q_i-qbar))
+     +2*ds   *((s_i-sbar)*(q_i-qbar)^2)
+     +1      *((s_i-sbar)^2*(q_i-qbar)^2).
+  With these expressions in the central (co-)moment bases, all we need to do
+   is compute sums over the (co-)moment terms, which can be done
+   incrementally (see oc_mode_metrics_add() and oc_mode_metrics_merge()),
+   with no need to store the individual samples.
+  Now, for X^T.Y, we have
+    X^T.Y = sum((R_i-A-((B-A)/(s1-s0))*(s_i-s0)-((C-A)/(q1-q0))*(q_i-q0)
+     -((A-B-C)/((s1-s0)*(q1-q0)))*(s_i-s0)*(q_i-q0))*(s_i-s0)*(q_i-q0))/
+     ((s1-s0)*(q1-q0)),
+   or, rewriting the constants to simplify notation,
+    X^T.Y = sum((C0+C1*(s_i-s0)+C2*(q_i-q0)
+     +C3*(s_i-s0)*(q_i-q0)+R_i)*(s_i-s0)*(q_i-q0))/((s1-s0)*(q1-q0)).
+  Again, converting to the central (co-)moment basis, the interior of the
+   above sum is
+    ds*dq*(rbar+C0+C1*ds+C2*dq+C3*ds*dq)  *(1)
+     +(C1*dq+C3*dq^2)                     *((s_i-sbar)^2)
+     +(rbar+C0+2*C1*ds+2*C2*dq+4*C3*ds*dq)*((s_i-sbar)*(q_i-qbar))
+     +(C2*ds+C3*ds^2)                     *((q_i-qbar)^2)
+     +dq                                  *((s_i-sbar)*(r_i-rbar))
+     +ds                                  *((q_i-qbar)*(r_i-rbar))
+     +(C1+2*C3*dq)                        *((s_i-sbar)^2*(q_i-qbar))
+     +(C2+2*C3*ds)                        *((s_i-sbar)*(q_i-qbar)^2)
+     +1                                   *((s_i-sbar)*(q_i-qbar)*(r_i-rbar))
+     +C3                                  *((s_i-sbar)^2*(q_i-qbar)^2).
+  You might think it would be easier (if perhaps slightly less robust) to
+   accumulate terms directly around s0 and q0.
+  However, we update each corner of the bins in turn, so we would have to
+   change basis to move the sums from corner to corner anyway.*/
+double oc_mode_metrics_solve(double *_r,double *_d,
+ const oc_mode_metrics *_metrics,const int *_s0,const int *_s1,
+ const int *_q0,const int *_q1,
+ const double *_ra,const double *_rb,const double *_rc,
+ const double *_da,const double *_db,const double *_dc,int _n){
+  double xx;
+  double rxy;
+  double dxy;
+  double wt;
+  int i;
+  xx=rxy=dxy=wt=0;
+  for(i=0;i<_n;i++)if(_metrics[i].w>0){
+    double s10;
+    double q10;
+    double sq10;
+    double ds;
+    double dq;
+    double ds2;
+    double dq2;
+    double r;
+    double d;
+    double s2;
+    double sq;
+    double q2;
+    double sr;
+    double qr;
+    double sd;
+    double qd;
+    double s2q;
+    double sq2;
+    double sqr;
+    double sqd;
+    double s2q2;
+    double c0;
+    double c1;
+    double c2;
+    double c3;
+    double w;
+    w=_metrics[i].w;
+    wt+=w;
+    s10=_s1[i]-_s0[i];
+    q10=_q1[i]-_q0[i];
+    sq10=s10*q10;
+    ds=_metrics[i].s/w-_s0[i];
+    dq=_metrics[i].q/w-_q0[i];
+    ds2=ds*ds;
+    dq2=dq*dq;
+    s2=_metrics[i].s2;
+    sq=_metrics[i].sq;
+    q2=_metrics[i].q2;
+    s2q=_metrics[i].s2q;
+    sq2=_metrics[i].sq2;
+    s2q2=_metrics[i].s2q2;
+    xx+=(dq2*(ds2*w+s2)+4*ds*dq*sq+ds2*q2+2*(dq*s2q+ds*sq2)+s2q2)/(sq10*sq10);
+    r=_metrics[i].r/w;
+    sr=_metrics[i].sr;
+    qr=_metrics[i].qr;
+    sqr=_metrics[i].sqr;
+    c0=-_ra[i];
+    c1=-(_rb[i]-_ra[i])/s10;
+    c2=-(_rc[i]-_ra[i])/q10;
+    c3=-(_ra[i]-_rb[i]-_rc[i])/sq10;
+    rxy+=(ds*dq*(r+c0+c1*ds+c2*dq+c3*ds*dq)*w+(c1*dq+c3*dq2)*s2
+     +(r+c0+2*(c1*ds+(c2+2*c3*ds)*dq))*sq+(c2*ds+c3*ds2)*q2+dq*sr+ds*qr
+     +(c1+2*c3*dq)*s2q+(c2+2*c3*ds)*sq2+sqr+c3*s2q2)/sq10;
+    d=_metrics[i].d/w;
+    sd=_metrics[i].sd;
+    qd=_metrics[i].qd;
+    sqd=_metrics[i].sqd;
+    c0=-_da[i];
+    c1=-(_db[i]-_da[i])/s10;
+    c2=-(_dc[i]-_da[i])/q10;
+    c3=-(_da[i]-_db[i]-_dc[i])/sq10;
+    dxy+=(ds*dq*(d+c0+c1*ds+c2*dq+c3*ds*dq)*w+(c1*dq+c3*dq2)*s2
+     +(d+c0+2*(c1*ds+(c2+2*c3*ds)*dq))*sq+(c2*ds+c3*ds2)*q2+dq*sd+ds*qd
+     +(c1+2*c3*dq)*s2q+(c2+2*c3*ds)*sq2+sqd+c3*s2q2)/sq10;
+  }
+  if(xx>1E-3){
+    *_r=rxy/xx;
+    *_d=dxy/xx;
+  }
+  else{
+    *_r=0;
+    *_d=0;
+  }
+  return wt;
+}
+
+/*Compile collected SATD/logq/rate/RMSE metrics into a form that's immediately
+   useful for mode decision.*/
+void oc_mode_metrics_update(oc_mode_metrics (*_metrics)[3][2][OC_COMP_BINS],
+ int _niters_min,int _reweight,oc_mode_rd (*_table)[3][2][OC_COMP_BINS],
+ int _shift,double (*_weight)[3][2][OC_COMP_BINS]){
+  int niters;
+  int prevdr;
+  int prevdd;
+  int dr;
+  int dd;
+  int pli;
+  int qti;
+  int qi;
+  int si;
+  dd=dr=INT_MAX;
+  niters=0;
+  /*The encoder interpolates rate and RMSE terms bilinearly from an
+     OC_LOGQ_BINS by OC_COMP_BINS grid of sample points in _table.
+    To find the sample values at the grid points that minimize the total
+     squared prediction error actually requires solving a relatively sparse
+     linear system with a number of variables equal to the number of grid
+     points.
+    Instead of writing a general sparse linear system solver, we just use
+     Gauss-Seidel iteration, i.e., we update one grid point at time until
+     they stop changing.*/
+  do{
+    prevdr=dr;
+    prevdd=dd;
+    dd=dr=0;
+    for(pli=0;pli<3;pli++){
+      for(qti=0;qti<2;qti++){
+        for(qi=0;qi<OC_LOGQ_BINS;qi++){
+          for(si=0;si<OC_COMP_BINS;si++){
+            oc_mode_metrics m[4];
+            int             s0[4];
+            int             s1[4];
+            int             q0[4];
+            int             q1[4];
+            double          ra[4];
+            double          rb[4];
+            double          rc[4];
+            double          da[4];
+            double          db[4];
+            double          dc[4];
+            double          r;
+            double          d;
+            int             rate;
+            int             rmse;
+            int             ds;
+            int             n;
+            n=0;
+            /*Collect the statistics for the (up to) four bins grid point
+               (si,qi) touches.*/
+            if(qi>0&&si>0){
+              q0[n]=OC_MODE_LOGQ[qi-1][pli][qti];
+              q1[n]=OC_MODE_LOGQ[qi][pli][qti];
+              s0[n]=si-1<<_shift;
+              s1[n]=si<<_shift;
+              ra[n]=ldexp(_table[qi-1][pli][qti][si-1].rate,-OC_BIT_SCALE);
+              da[n]=ldexp(_table[qi-1][pli][qti][si-1].rmse,-OC_RMSE_SCALE);
+              rb[n]=ldexp(_table[qi-1][pli][qti][si].rate,-OC_BIT_SCALE);
+              db[n]=ldexp(_table[qi-1][pli][qti][si].rmse,-OC_RMSE_SCALE);
+              rc[n]=ldexp(_table[qi][pli][qti][si-1].rate,-OC_BIT_SCALE);
+              dc[n]=ldexp(_table[qi][pli][qti][si-1].rmse,-OC_RMSE_SCALE);
+              *(m+n++)=*(_metrics[qi-1][pli][qti]+si-1);
+            }
+            if(qi>0){
+              ds=si+1<OC_COMP_BINS?1:-1;
+              q0[n]=OC_MODE_LOGQ[qi-1][pli][qti];
+              q1[n]=OC_MODE_LOGQ[qi][pli][qti];
+              s0[n]=si+ds<<_shift;
+              s1[n]=si<<_shift;
+              ra[n]=ldexp(_table[qi-1][pli][qti][si+ds].rate,-OC_BIT_SCALE);
+              da[n]=
+               ldexp(_table[qi-1][pli][qti][si+ds].rmse,-OC_RMSE_SCALE);
+              rb[n]=ldexp(_table[qi-1][pli][qti][si].rate,-OC_BIT_SCALE);
+              db[n]=ldexp(_table[qi-1][pli][qti][si].rmse,-OC_RMSE_SCALE);
+              rc[n]=ldexp(_table[qi][pli][qti][si+ds].rate,-OC_BIT_SCALE);
+              dc[n]=ldexp(_table[qi][pli][qti][si+ds].rmse,-OC_RMSE_SCALE);
+              *(m+n++)=*(_metrics[qi-1][pli][qti]+si);
+            }
+            if(qi+1<OC_LOGQ_BINS&&si>0){
+              q0[n]=OC_MODE_LOGQ[qi+1][pli][qti];
+              q1[n]=OC_MODE_LOGQ[qi][pli][qti];
+              s0[n]=si-1<<_shift;
+              s1[n]=si<<_shift;
+              ra[n]=ldexp(_table[qi+1][pli][qti][si-1].rate,-OC_BIT_SCALE);
+              da[n]=ldexp(_table[qi+1][pli][qti][si-1].rmse,-OC_RMSE_SCALE);
+              rb[n]=ldexp(_table[qi+1][pli][qti][si].rate,-OC_BIT_SCALE);
+              db[n]=ldexp(_table[qi+1][pli][qti][si].rmse,-OC_RMSE_SCALE);
+              rc[n]=ldexp(_table[qi][pli][qti][si-1].rate,-OC_BIT_SCALE);
+              dc[n]=ldexp(_table[qi][pli][qti][si-1].rmse,-OC_RMSE_SCALE);
+              *(m+n++)=*(_metrics[qi][pli][qti]+si-1);
+            }
+            if(qi+1<OC_LOGQ_BINS){
+              ds=si+1<OC_COMP_BINS?1:-1;
+              q0[n]=OC_MODE_LOGQ[qi+1][pli][qti];
+              q1[n]=OC_MODE_LOGQ[qi][pli][qti];
+              s0[n]=si+ds<<_shift;
+              s1[n]=si<<_shift;
+              ra[n]=ldexp(_table[qi+1][pli][qti][si+ds].rate,-OC_BIT_SCALE);
+              da[n]=
+               ldexp(_table[qi+1][pli][qti][si+ds].rmse,-OC_RMSE_SCALE);
+              rb[n]=ldexp(_table[qi+1][pli][qti][si].rate,-OC_BIT_SCALE);
+              db[n]=ldexp(_table[qi+1][pli][qti][si].rmse,-OC_RMSE_SCALE);
+              rc[n]=ldexp(_table[qi][pli][qti][si+ds].rate,-OC_BIT_SCALE);
+              dc[n]=ldexp(_table[qi][pli][qti][si+ds].rmse,-OC_RMSE_SCALE);
+              *(m+n++)=*(_metrics[qi][pli][qti]+si);
+            }
+            /*On the first pass, initialize with a simple weighted average of
+               the neighboring bins.*/
+            if(!OC_HAS_MODE_METRICS&&niters==0){
+              double w;
+              w=r=d=0;
+              while(n-->0){
+                w+=m[n].w;
+                r+=m[n].r;
+                d+=m[n].d;
+              }
+              r=w>1E-3?r/w:0;
+              d=w>1E-3?d/w:0;
+              _weight[qi][pli][qti][si]=w;
+            }
+            else{
+              /*Update the grid point and save the weight for later.*/
+              _weight[qi][pli][qti][si]=
+               oc_mode_metrics_solve(&r,&d,m,s0,s1,q0,q1,ra,rb,rc,da,db,dc,n);
+            }
+            rate=OC_CLAMPI(-32768,(int)(ldexp(r,OC_BIT_SCALE)+0.5),32767);
+            rmse=OC_CLAMPI(-32768,(int)(ldexp(d,OC_RMSE_SCALE)+0.5),32767);
+            dr+=abs(rate-_table[qi][pli][qti][si].rate);
+            dd+=abs(rmse-_table[qi][pli][qti][si].rmse);
+            _table[qi][pli][qti][si].rate=(ogg_int16_t)rate;
+            _table[qi][pli][qti][si].rmse=(ogg_int16_t)rmse;
+          }
+        }
+      }
+    }
+  }
+  /*After a fixed number of initial iterations, only iterate so long as the
+     total change is decreasing.
+    This ensures we don't oscillate forever, which is a danger, as all of our
+     results are rounded fairly coarsely.*/
+  while((dr>0||dd>0)&&(niters++<_niters_min||(dr<prevdr&&dd<prevdd)));
+  if(_reweight){
+    /*Now, reduce the values of the optimal solution until we get enough
+       samples in each bin to overcome the constant OC_ZWEIGHT factor.
+      This encourages sampling under-populated bins and prevents a single large
+       sample early on from discouraging coding in that bin ever again.*/
+    for(pli=0;pli<3;pli++){
+      for(qti=0;qti<2;qti++){
+        for(qi=0;qi<OC_LOGQ_BINS;qi++){
+          for(si=0;si<OC_COMP_BINS;si++){
+            double wt;
+            wt=_weight[qi][pli][qti][si];
+            wt/=OC_ZWEIGHT+wt;
+            _table[qi][pli][qti][si].rate=(ogg_int16_t)
+             (_table[qi][pli][qti][si].rate*wt+0.5);
+            _table[qi][pli][qti][si].rmse=(ogg_int16_t)
+             (_table[qi][pli][qti][si].rmse*wt+0.5);
+          }
+        }
+      }
+    }
+  }
+}
+
+/*Dump the in memory mode metrics to a file.
+  Note this data format isn't portable between different platforms.*/
+void oc_mode_metrics_dump(void){
+  FILE *fmetrics;
+  fmetrics=fopen(OC_MODE_METRICS_FILENAME,"wb");
+  if(fmetrics!=NULL){
+    (void)fwrite(OC_MODE_LOGQ,sizeof(OC_MODE_LOGQ),1,fmetrics);
+    (void)fwrite(OC_MODE_METRICS_SATD,sizeof(OC_MODE_METRICS_SATD),1,fmetrics);
+    (void)fwrite(OC_MODE_METRICS_SAD,sizeof(OC_MODE_METRICS_SAD),1,fmetrics);
+    fclose(fmetrics);
+  }
+}
+
+void oc_mode_metrics_print_rd(FILE *_fout,const char *_table_name,
+#if !defined(OC_COLLECT_METRICS)
+ const oc_mode_rd (*_mode_rd_table)[3][2][OC_COMP_BINS]){
+#else
+ oc_mode_rd (*_mode_rd_table)[3][2][OC_COMP_BINS]){
+#endif
+  int qii;
+  fprintf(_fout,
+   "# if !defined(OC_COLLECT_METRICS)\n"
+   "static const\n"
+   "# endif\n"
+   "oc_mode_rd %s[OC_LOGQ_BINS][3][2][OC_COMP_BINS]={\n",_table_name);
+  for(qii=0;qii<OC_LOGQ_BINS;qii++){
+    int pli;
+    fprintf(_fout,"  {\n");
+    for(pli=0;pli<3;pli++){
+      int qti;
+      fprintf(_fout,"    {\n");
+      for(qti=0;qti<2;qti++){
+        int bin;
+        int qi;
+        static const char *pl_names[3]={"Y'","Cb","Cr"};
+        static const char *qti_names[2]={"INTRA","INTER"};
+        qi=(63*qii+(OC_LOGQ_BINS-1>>1))/(OC_LOGQ_BINS-1);
+        fprintf(_fout,"      /*%s  qi=%i  %s*/\n",
+         pl_names[pli],qi,qti_names[qti]);
+        fprintf(_fout,"      {\n");
+        fprintf(_fout,"        ");
+        for(bin=0;bin<OC_COMP_BINS;bin++){
+          if(bin&&!(bin&0x3))fprintf(_fout,"\n        ");
+          fprintf(_fout,"{%5i,%5i}",
+           _mode_rd_table[qii][pli][qti][bin].rate,
+           _mode_rd_table[qii][pli][qti][bin].rmse);
+          if(bin+1<OC_COMP_BINS)fprintf(_fout,",");
+        }
+        fprintf(_fout,"\n      }");
+        if(qti<1)fprintf(_fout,",");
+        fprintf(_fout,"\n");
+      }
+      fprintf(_fout,"    }");
+      if(pli<2)fprintf(_fout,",");
+      fprintf(_fout,"\n");
+    }
+    fprintf(_fout,"  }");
+    if(qii+1<OC_LOGQ_BINS)fprintf(_fout,",");
+    fprintf(_fout,"\n");
+  }
+  fprintf(_fout,
+   "};\n"
+   "\n");
+}
+
+void oc_mode_metrics_print(FILE *_fout){
+  int qii;
+  fprintf(_fout,
+   "/*File generated by libtheora with OC_COLLECT_METRICS"
+   " defined at compile time.*/\n"
+   "#if !defined(_modedec_H)\n"
+   "# define _modedec_H (1)\n"
+   "# include \"encint.h\"\n"
+   "\n"
+   "\n"
+   "\n"
+   "/*The log of the average quantizer for each of the OC_MODE_RD table rows\n"
+   "   (e.g., for the represented qi's, and each pli and qti), in Q10 format.\n"
+   "  The actual statistics used by the encoder will be interpolated from\n"
+   "   that table based on log_plq for the actual quantization matrix used.*/\n"
+   "# if !defined(OC_COLLECT_METRICS)\n"
+   "static const\n"
+   "# endif\n"
+   "ogg_int16_t OC_MODE_LOGQ[OC_LOGQ_BINS][3][2]={\n");
+  for(qii=0;qii<OC_LOGQ_BINS;qii++){
+    fprintf(_fout,"  { {0x%04X,0x%04X},{0x%04X,0x%04X},{0x%04X,0x%04X} }%s\n",
+     OC_MODE_LOGQ[qii][0][0],OC_MODE_LOGQ[qii][0][1],OC_MODE_LOGQ[qii][1][0],
+     OC_MODE_LOGQ[qii][1][1],OC_MODE_LOGQ[qii][2][0],OC_MODE_LOGQ[qii][2][1],
+     qii+1<OC_LOGQ_BINS?",":"");
+  }
+  fprintf(_fout,
+   "};\n"
+   "\n");
+  oc_mode_metrics_print_rd(_fout,"OC_MODE_RD_SATD",OC_MODE_RD_SATD);
+  oc_mode_metrics_print_rd(_fout,"OC_MODE_RD_SAD",OC_MODE_RD_SAD);
+  fprintf(_fout,
+   "#endif\n");
+}
+
+
+# if !defined(OC_COLLECT_NO_ENC_FUNCS)
+void oc_enc_mode_metrics_load(oc_enc_ctx *_enc){
+  oc_restore_fpu(&_enc->state);
+  /*Load any existing mode metrics if we haven't already.*/
+  if(!OC_HAS_MODE_METRICS){
+    FILE *fmetrics;
+    memset(OC_MODE_METRICS_SATD,0,sizeof(OC_MODE_METRICS_SATD));
+    memset(OC_MODE_METRICS_SAD,0,sizeof(OC_MODE_METRICS_SAD));
+    fmetrics=fopen(OC_MODE_METRICS_FILENAME,"rb");
+    if(fmetrics!=NULL){
+      /*Read in the binary structures as written my oc_mode_metrics_dump().
+        Note this format isn't portable between different platforms.*/
+      (void)fread(OC_MODE_LOGQ,sizeof(OC_MODE_LOGQ),1,fmetrics);
+      (void)fread(OC_MODE_METRICS_SATD,sizeof(OC_MODE_METRICS_SATD),1,fmetrics);
+      (void)fread(OC_MODE_METRICS_SAD,sizeof(OC_MODE_METRICS_SAD),1,fmetrics);
+      fclose(fmetrics);
+    }
+    else{
+      int qii;
+      int qi;
+      int pli;
+      int qti;
+      for(qii=0;qii<OC_LOGQ_BINS;qii++){
+        qi=(63*qii+(OC_LOGQ_BINS-1>>1))/(OC_LOGQ_BINS-1);
+        for(pli=0;pli<3;pli++)for(qti=0;qti<2;qti++){
+          OC_MODE_LOGQ[qii][pli][qti]=_enc->log_plq[qi][pli][qti];
+        }
+      }
+    }
+    oc_mode_metrics_update(OC_MODE_METRICS_SATD,100,1,
+     OC_MODE_RD_SATD,OC_SATD_SHIFT,OC_MODE_RD_WEIGHT_SATD);
+    oc_mode_metrics_update(OC_MODE_METRICS_SAD,100,1,
+     OC_MODE_RD_SAD,OC_SAD_SHIFT,OC_MODE_RD_WEIGHT_SAD);
+    OC_HAS_MODE_METRICS=1;
+  }
+}
+
+/*The following token skipping code used to also be used in the decoder (and
+   even at one point other places in the encoder).
+  However, it was obsoleted by other optimizations, and is now only used here.
+  It has been moved here to avoid generating the code when it's not needed.*/
+
+/*Determines the number of blocks or coefficients to be skipped for a given
+   token value.
+  _token:      The token value to skip.
+  _extra_bits: The extra bits attached to this token.
+  Return: A positive value indicates that number of coefficients are to be
+           skipped in the current block.
+          Otherwise, the negative of the return value indicates that number of
+           blocks are to be ended.*/
+typedef ptrdiff_t (*oc_token_skip_func)(int _token,int _extra_bits);
+
+/*Handles the simple end of block tokens.*/
+static ptrdiff_t oc_token_skip_eob(int _token,int _extra_bits){
+  int nblocks_adjust;
+  nblocks_adjust=OC_UNIBBLE_TABLE32(0,1,2,3,7,15,0,0,_token)+1;
+  return -_extra_bits-nblocks_adjust;
+}
+
+/*The last EOB token has a special case, where an EOB run of size zero ends all
+   the remaining blocks in the frame.*/
+static ptrdiff_t oc_token_skip_eob6(int _token,int _extra_bits){
+  /*Note: We want to return -PTRDIFF_MAX, but that requires C99, which is not
+     yet available everywhere; this should be equivalent.*/
+  if(!_extra_bits)return -(~(size_t)0>>1);
+  return -_extra_bits;
+}
+
+/*Handles the pure zero run tokens.*/
+static ptrdiff_t oc_token_skip_zrl(int _token,int _extra_bits){
+  return _extra_bits+1;
+}
+
+/*Handles a normal coefficient value token.*/
+static ptrdiff_t oc_token_skip_val(void){
+  return 1;
+}
+
+/*Handles a category 1A zero run/coefficient value combo token.*/
+static ptrdiff_t oc_token_skip_run_cat1a(int _token){
+  return _token-OC_DCT_RUN_CAT1A+2;
+}
+
+/*Handles category 1b, 1c, 2a, and 2b zero run/coefficient value combo tokens.*/
+static ptrdiff_t oc_token_skip_run(int _token,int _extra_bits){
+  int run_cati;
+  int ncoeffs_mask;
+  int ncoeffs_adjust;
+  run_cati=_token-OC_DCT_RUN_CAT1B;
+  ncoeffs_mask=OC_BYTE_TABLE32(3,7,0,1,run_cati);
+  ncoeffs_adjust=OC_BYTE_TABLE32(7,11,2,3,run_cati);
+  return (_extra_bits&ncoeffs_mask)+ncoeffs_adjust;
+}
+
+/*A jump table for computing the number of coefficients or blocks to skip for
+   a given token value.
+  This reduces all the conditional branches, etc., needed to parse these token
+   values down to one indirect jump.*/
+static const oc_token_skip_func OC_TOKEN_SKIP_TABLE[TH_NDCT_TOKENS]={
+  oc_token_skip_eob,
+  oc_token_skip_eob,
+  oc_token_skip_eob,
+  oc_token_skip_eob,
+  oc_token_skip_eob,
+  oc_token_skip_eob,
+  oc_token_skip_eob6,
+  oc_token_skip_zrl,
+  oc_token_skip_zrl,
+  (oc_token_skip_func)oc_token_skip_val,
+  (oc_token_skip_func)oc_token_skip_val,
+  (oc_token_skip_func)oc_token_skip_val,
+  (oc_token_skip_func)oc_token_skip_val,
+  (oc_token_skip_func)oc_token_skip_val,
+  (oc_token_skip_func)oc_token_skip_val,
+  (oc_token_skip_func)oc_token_skip_val,
+  (oc_token_skip_func)oc_token_skip_val,
+  (oc_token_skip_func)oc_token_skip_val,
+  (oc_token_skip_func)oc_token_skip_val,
+  (oc_token_skip_func)oc_token_skip_val,
+  (oc_token_skip_func)oc_token_skip_val,
+  (oc_token_skip_func)oc_token_skip_val,
+  (oc_token_skip_func)oc_token_skip_val,
+  (oc_token_skip_func)oc_token_skip_run_cat1a,
+  (oc_token_skip_func)oc_token_skip_run_cat1a,
+  (oc_token_skip_func)oc_token_skip_run_cat1a,
+  (oc_token_skip_func)oc_token_skip_run_cat1a,
+  (oc_token_skip_func)oc_token_skip_run_cat1a,
+  oc_token_skip_run,
+  oc_token_skip_run,
+  oc_token_skip_run,
+  oc_token_skip_run
+};
+
+/*Determines the number of blocks or coefficients to be skipped for a given
+   token value.
+  _token:      The token value to skip.
+  _extra_bits: The extra bits attached to this token.
+  Return: A positive value indicates that number of coefficients are to be
+           skipped in the current block.
+          Otherwise, the negative of the return value indicates that number of
+           blocks are to be ended.
+          0 will never be returned, so that at least one coefficient in one
+           block will always be decoded for every token.*/
+static ptrdiff_t oc_dct_token_skip(int _token,int _extra_bits){
+  return (*OC_TOKEN_SKIP_TABLE[_token])(_token,_extra_bits);
+}
+
+
+void oc_enc_mode_metrics_collect(oc_enc_ctx *_enc){
+  static const unsigned char OC_ZZI_HUFF_OFFSET[64]={
+     0,16,16,16,16,16,32,32,
+    32,32,32,32,32,32,32,48,
+    48,48,48,48,48,48,48,48,
+    48,48,48,48,64,64,64,64,
+    64,64,64,64,64,64,64,64,
+    64,64,64,64,64,64,64,64,
+    64,64,64,64,64,64,64,64
+  };
+  const oc_fragment *frags;
+  const unsigned    *frag_sad;
+  const unsigned    *frag_satd;
+  const unsigned    *frag_ssd;
+  const ptrdiff_t   *coded_fragis;
+  ptrdiff_t          ncoded_fragis;
+  ptrdiff_t          fragii;
+  double             fragw;
+  int                modelines[3][3][2];
+  int                qti;
+  int                qii;
+  int                qi;
+  int                pli;
+  int                zzi;
+  int                token;
+  int                eb;
+  oc_restore_fpu(&_enc->state);
+  /*Figure out which metric bins to use for this frame's quantizers.*/
+  for(qii=0;qii<_enc->state.nqis;qii++){
+    for(pli=0;pli<3;pli++){
+      for(qti=0;qti<2;qti++){
+        int log_plq;
+        int modeline;
+        log_plq=_enc->log_plq[_enc->state.qis[qii]][pli][qti];
+        for(modeline=0;modeline<OC_LOGQ_BINS-1&&
+         OC_MODE_LOGQ[modeline+1][pli][qti]>log_plq;modeline++);
+        modelines[qii][pli][qti]=modeline;
+      }
+    }
+  }
+  qti=_enc->state.frame_type;
+  frags=_enc->state.frags;
+  frag_sad=_enc->frag_sad;
+  frag_satd=_enc->frag_satd;
+  frag_ssd=_enc->frag_ssd;
+  coded_fragis=_enc->state.coded_fragis;
+  ncoded_fragis=fragii=0;
+  /*Weight the fragments by the inverse frame size; this prevents HD content
+     from dominating the statistics.*/
+  fragw=1.0/_enc->state.nfrags;
+  for(pli=0;pli<3;pli++){
+    ptrdiff_t ti[64];
+    int       eob_token[64];
+    int       eob_run[64];
+    /*Set up token indices and eob run counts.
+      We don't bother trying to figure out the real cost of the runs that span
+       coefficients; instead we use the costs that were available when R-D
+       token optimization was done.*/
+    for(zzi=0;zzi<64;zzi++){
+      ti[zzi]=_enc->dct_token_offs[pli][zzi];
+      if(ti[zzi]>0){
+        token=_enc->dct_tokens[pli][zzi][0];
+        eb=_enc->extra_bits[pli][zzi][0];
+        eob_token[zzi]=token;
+        eob_run[zzi]=-oc_dct_token_skip(token,eb);
+      }
+      else{
+        eob_token[zzi]=OC_NDCT_EOB_TOKEN_MAX;
+        eob_run[zzi]=0;
+      }
+    }
+    /*Scan the list of coded fragments for this plane.*/
+    ncoded_fragis+=_enc->state.ncoded_fragis[pli];
+    for(;fragii<ncoded_fragis;fragii++){
+      ptrdiff_t fragi;
+      int       frag_bits;
+      int       huffi;
+      int       skip;
+      int       mb_mode;
+      unsigned  sad;
+      unsigned  satd;
+      double    sqrt_ssd;
+      int       bin;
+      int       qtj;
+      fragi=coded_fragis[fragii];
+      frag_bits=0;
+      for(zzi=0;zzi<64;){
+        if(eob_run[zzi]>0){
+          /*We've reached the end of the block.*/
+          eob_run[zzi]--;
+          break;
+        }
+        huffi=_enc->huff_idxs[qti][zzi>0][pli+1>>1]
+         +OC_ZZI_HUFF_OFFSET[zzi];
+        if(eob_token[zzi]<OC_NDCT_EOB_TOKEN_MAX){
+          /*This token caused an EOB run to be flushed.
+            Therefore it gets the bits associated with it.*/
+          frag_bits+=_enc->huff_codes[huffi][eob_token[zzi]].nbits
+           +OC_DCT_TOKEN_EXTRA_BITS[eob_token[zzi]];
+          eob_token[zzi]=OC_NDCT_EOB_TOKEN_MAX;
+        }
+        token=_enc->dct_tokens[pli][zzi][ti[zzi]];
+        eb=_enc->extra_bits[pli][zzi][ti[zzi]];
+        ti[zzi]++;
+        skip=oc_dct_token_skip(token,eb);
+        if(skip<0){
+          eob_token[zzi]=token;
+          eob_run[zzi]=-skip;
+        }
+        else{
+          /*A regular DCT value token; accumulate the bits for it.*/
+          frag_bits+=_enc->huff_codes[huffi][token].nbits
+           +OC_DCT_TOKEN_EXTRA_BITS[token];
+          zzi+=skip;
+        }
+      }
+      mb_mode=frags[fragi].mb_mode;
+      qii=frags[fragi].qii;
+      qi=_enc->state.qis[qii];
+      sad=frag_sad[fragi]<<(pli+1&2);
+      satd=frag_satd[fragi]<<(pli+1&2);
+      sqrt_ssd=sqrt(frag_ssd[fragi]);
+      qtj=mb_mode!=OC_MODE_INTRA;
+      /*Accumulate statistics.
+        The rate (frag_bits) and RMSE (sqrt(frag_ssd)) are not scaled by
+         OC_BIT_SCALE and OC_RMSE_SCALE; this lets us change the scale factor
+         yet still use old data.*/
+      bin=OC_MINI(satd>>OC_SATD_SHIFT,OC_COMP_BINS-1);
+      oc_mode_metrics_add(
+       OC_MODE_METRICS_SATD[modelines[qii][pli][qtj]][pli][qtj]+bin,
+       fragw,satd,_enc->log_plq[qi][pli][qtj],frag_bits,sqrt_ssd);
+      bin=OC_MINI(sad>>OC_SAD_SHIFT,OC_COMP_BINS-1);
+      oc_mode_metrics_add(
+       OC_MODE_METRICS_SAD[modelines[qii][pli][qtj]][pli][qtj]+bin,
+       fragw,sad,_enc->log_plq[qi][pli][qtj],frag_bits,sqrt_ssd);
+    }
+  }
+  /*Update global SA(T)D/logq/rate/RMSE estimation matrix.*/
+  oc_mode_metrics_update(OC_MODE_METRICS_SATD,4,1,
+   OC_MODE_RD_SATD,OC_SATD_SHIFT,OC_MODE_RD_WEIGHT_SATD);
+  oc_mode_metrics_update(OC_MODE_METRICS_SAD,4,1,
+   OC_MODE_RD_SAD,OC_SAD_SHIFT,OC_MODE_RD_WEIGHT_SAD);
+}
+# endif
+
+#endif

thirdparty/libtheora/collect.h

@@ -0,0 +1,108 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation https://www.xiph.org/                 *
+ *                                                                  *
+ ********************************************************************
+
+  function: mode selection code
+
+ ********************************************************************/
+#if !defined(_collect_H)
+# define _collect_H (1)
+# include "encint.h"
+# if defined(OC_COLLECT_METRICS)
+#  include <stdio.h>
+
+
+
+typedef struct oc_mode_metrics oc_mode_metrics;
+
+
+
+/**Sets the file name to load/store mode metrics from/to.
+ * The file name string is stored by reference, and so must be valid for the
+ *  lifetime of the encoder.
+ * Mode metric collection uses global tables; do not attempt to perform
+ *  multiple collections at once.
+ * \param[in] _buf <tt>char[]</tt> The file name.
+ * \retval TH_EIMPL   Not supported by this implementation.*/
+#define TH_ENCCTL_SET_METRICS_FILE (0x8000)
+
+
+
+/*Accumulates various weighted sums of the measurements.
+  w -> weight
+  s -> SATD
+  q -> log quantizer
+  r -> rate (in bits)
+  d -> RMSE
+  All of the single letters correspond to direct, weighted sums, e.g.,
+   w=sum(w_i), s=sum(s_i*w_i), etc.
+  The others correspond to central moments (or co-moments) of the given order,
+   e.g., sq=sum((s_i-s/w)*(q_i-q/w)*w_i).
+  Because we need some moments up to fourth order, we use central moments to
+   minimize the dynamic range and prevent rounding error from dominating the
+   calculations.*/
+struct oc_mode_metrics{
+  double w;
+  double s;
+  double q;
+  double r;
+  double d;
+  double s2;
+  double sq;
+  double q2;
+  double sr;
+  double qr;
+  double r2;
+  double sd;
+  double qd;
+  double d2;
+  double s2q;
+  double sq2;
+  double sqr;
+  double sqd;
+  double s2q2;
+};
+
+
+# define OC_ZWEIGHT   (0.25)
+
+/*TODO: It may be helpful (for block-level quantizers especially) to separate
+   out the contributions from AC and DC into separate tables.*/
+
+extern ogg_int16_t OC_MODE_LOGQ[OC_LOGQ_BINS][3][2];
+extern oc_mode_rd  OC_MODE_RD_SATD[OC_LOGQ_BINS][3][2][OC_COMP_BINS];
+extern oc_mode_rd  OC_MODE_RD_SAD[OC_LOGQ_BINS][3][2][OC_COMP_BINS];
+
+extern int              OC_HAS_MODE_METRICS;
+extern oc_mode_metrics  OC_MODE_METRICS_SATD[OC_LOGQ_BINS-1][3][2][OC_COMP_BINS];
+extern oc_mode_metrics  OC_MODE_METRICS_SAD[OC_LOGQ_BINS-1][3][2][OC_COMP_BINS];
+extern const char      *OC_MODE_METRICS_FILENAME;
+
+void oc_mode_metrics_dump();
+void oc_mode_metrics_print(FILE *_fout);
+
+void oc_mode_metrics_add(oc_mode_metrics *_metrics,
+ double _w,int _s,int _q,int _r,double _d);
+void oc_mode_metrics_merge(oc_mode_metrics *_dst,
+ const oc_mode_metrics *_src,int _n);
+double oc_mode_metrics_solve(double *_r,double *_d,
+ const oc_mode_metrics *_metrics,const int *_s0,const int *_s1,
+ const int *_q0,const int *_q1,
+ const double *_ra,const double *_rb,const double *_rc,
+ const double *_da,const double *_db,const double *_dc,int _n);
+void oc_mode_metrics_update(oc_mode_metrics (*_metrics)[3][2][OC_COMP_BINS],
+ int _niters_min,int _reweight,oc_mode_rd (*_table)[3][2][OC_COMP_BINS],
+ int shift,double (*_weight)[3][2][OC_COMP_BINS]);
+void oc_enc_mode_metrics_load(oc_enc_ctx *_enc);
+void oc_enc_mode_metrics_collect(oc_enc_ctx *_enc);
+
+# endif
+#endif

thirdparty/libtheora/dct.h

@@ -0,0 +1,31 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+/*Definitions shared by the forward and inverse DCT transforms.*/
+#if !defined(_dct_H)
+# define _dct_H (1)
+
+/*cos(n*pi/16) (resp. sin(m*pi/16)) scaled by 65536.*/
+#define OC_C1S7 ((ogg_int32_t)64277)
+#define OC_C2S6 ((ogg_int32_t)60547)
+#define OC_C3S5 ((ogg_int32_t)54491)
+#define OC_C4S4 ((ogg_int32_t)46341)
+#define OC_C5S3 ((ogg_int32_t)36410)
+#define OC_C6S2 ((ogg_int32_t)25080)
+#define OC_C7S1 ((ogg_int32_t)12785)
+
+#endif

thirdparty/libtheora/decapiwrapper.c

@@ -0,0 +1,193 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+#include <stdlib.h>
+#include <string.h>
+#include <limits.h>
+#include "apiwrapper.h"
+#include "decint.h"
+#include "theora/theoradec.h"
+
+static void th_dec_api_clear(th_api_wrapper *_api){
+  if(_api->setup)th_setup_free(_api->setup);
+  if(_api->decode)th_decode_free(_api->decode);
+  memset(_api,0,sizeof(*_api));
+}
+
+static void theora_decode_clear(theora_state *_td){
+  if(_td->i!=NULL)theora_info_clear(_td->i);
+  memset(_td,0,sizeof(*_td));
+}
+
+static int theora_decode_control(theora_state *_td,int _req,
+ void *_buf,size_t _buf_sz){
+  return th_decode_ctl(((th_api_wrapper *)_td->i->codec_setup)->decode,
+   _req,_buf,_buf_sz);
+}
+
+static ogg_int64_t theora_decode_granule_frame(theora_state *_td,
+ ogg_int64_t _gp){
+  return th_granule_frame(((th_api_wrapper *)_td->i->codec_setup)->decode,_gp);
+}
+
+static double theora_decode_granule_time(theora_state *_td,ogg_int64_t _gp){
+  return th_granule_time(((th_api_wrapper *)_td->i->codec_setup)->decode,_gp);
+}
+
+static const oc_state_dispatch_vtable OC_DEC_DISPATCH_VTBL={
+  (oc_state_clear_func)theora_decode_clear,
+  (oc_state_control_func)theora_decode_control,
+  (oc_state_granule_frame_func)theora_decode_granule_frame,
+  (oc_state_granule_time_func)theora_decode_granule_time,
+};
+
+static void th_info2theora_info(theora_info *_ci,const th_info *_info){
+  _ci->version_major=_info->version_major;
+  _ci->version_minor=_info->version_minor;
+  _ci->version_subminor=_info->version_subminor;
+  _ci->width=_info->frame_width;
+  _ci->height=_info->frame_height;
+  _ci->frame_width=_info->pic_width;
+  _ci->frame_height=_info->pic_height;
+  _ci->offset_x=_info->pic_x;
+  _ci->offset_y=_info->pic_y;
+  _ci->fps_numerator=_info->fps_numerator;
+  _ci->fps_denominator=_info->fps_denominator;
+  _ci->aspect_numerator=_info->aspect_numerator;
+  _ci->aspect_denominator=_info->aspect_denominator;
+  switch(_info->colorspace){
+    case TH_CS_ITU_REC_470M:_ci->colorspace=OC_CS_ITU_REC_470M;break;
+    case TH_CS_ITU_REC_470BG:_ci->colorspace=OC_CS_ITU_REC_470BG;break;
+    default:_ci->colorspace=OC_CS_UNSPECIFIED;break;
+  }
+  switch(_info->pixel_fmt){
+    case TH_PF_420:_ci->pixelformat=OC_PF_420;break;
+    case TH_PF_422:_ci->pixelformat=OC_PF_422;break;
+    case TH_PF_444:_ci->pixelformat=OC_PF_444;break;
+    default:_ci->pixelformat=OC_PF_RSVD;
+  }
+  _ci->target_bitrate=_info->target_bitrate;
+  _ci->quality=_info->quality;
+  _ci->keyframe_frequency_force=1<<_info->keyframe_granule_shift;
+}
+
+int theora_decode_init(theora_state *_td,theora_info *_ci){
+  th_api_info    *apiinfo;
+  th_api_wrapper *api;
+  th_info         info;
+  api=(th_api_wrapper *)_ci->codec_setup;
+  /*Allocate our own combined API wrapper/theora_info struct.
+    We put them both in one malloc'd block so that when the API wrapper is
+     freed, the info struct goes with it.
+    This avoids having to figure out whether or not we need to free the info
+     struct in either theora_info_clear() or theora_clear().*/
+  apiinfo=(th_api_info *)_ogg_calloc(1,sizeof(*apiinfo));
+  if(apiinfo==NULL)return OC_FAULT;
+  /*Make our own copy of the info struct, since its lifetime should be
+     independent of the one we were passed in.*/
+  *&apiinfo->info=*_ci;
+  /*Convert the info struct now instead of saving the the one we decoded with
+     theora_decode_header(), since the user might have modified values (i.e.,
+     color space, aspect ratio, etc. can be specified from a higher level).
+    The user also might be doing something "clever" with the header packets if
+     they are not using an Ogg encapsulation.*/
+  oc_theora_info2th_info(&info,_ci);
+  /*Don't bother to copy the setup info; th_decode_alloc() makes its own copy
+     of the stuff it needs.*/
+  apiinfo->api.decode=th_decode_alloc(&info,api->setup);
+  if(apiinfo->api.decode==NULL){
+    _ogg_free(apiinfo);
+    return OC_EINVAL;
+  }
+  apiinfo->api.clear=(oc_setup_clear_func)th_dec_api_clear;
+  _td->internal_encode=NULL;
+  /*Provide entry points for ABI compatibility with old decoder shared libs.*/
+  _td->internal_decode=(void *)&OC_DEC_DISPATCH_VTBL;
+  _td->granulepos=0;
+  _td->i=&apiinfo->info;
+  _td->i->codec_setup=&apiinfo->api;
+  return 0;
+}
+
+int theora_decode_header(theora_info *_ci,theora_comment *_cc,ogg_packet *_op){
+  th_api_wrapper *api;
+  th_info         info;
+  int             ret;
+  api=(th_api_wrapper *)_ci->codec_setup;
+  /*Allocate an API wrapper struct on demand, since it will not also include a
+     theora_info struct like the ones that are used in a theora_state struct.*/
+  if(api==NULL){
+    _ci->codec_setup=_ogg_calloc(1,sizeof(*api));
+    if(_ci->codec_setup==NULL)return OC_FAULT;
+    api=(th_api_wrapper *)_ci->codec_setup;
+    api->clear=(oc_setup_clear_func)th_dec_api_clear;
+  }
+  /*Convert from the theora_info struct instead of saving our own th_info
+     struct between calls.
+    The user might be doing something "clever" with the header packets if they
+     are not using an Ogg encapsulation, and we don't want to break this.*/
+  oc_theora_info2th_info(&info,_ci);
+  /*We rely on the fact that theora_comment and th_comment structures are
+     actually identical.
+    Take care not to change this fact unless you change the code here as
+     well!*/
+  ret=th_decode_headerin(&info,(th_comment *)_cc,&api->setup,_op);
+  /*We also rely on the fact that the error return code values are the same,
+    and that the implementations of these two functions return the same set of
+    them.
+   Note that theora_decode_header() really can return OC_NOTFORMAT, even
+    though it is not currently documented to do so.*/
+  if(ret<0)return ret;
+  th_info2theora_info(_ci,&info);
+  return 0;
+}
+
+int theora_decode_packetin(theora_state *_td,ogg_packet *_op){
+  th_api_wrapper *api;
+  ogg_int64_t     gp;
+  int             ret;
+  if(!_td||!_td->i||!_td->i->codec_setup)return OC_FAULT;
+  api=(th_api_wrapper *)_td->i->codec_setup;
+  ret=th_decode_packetin(api->decode,_op,&gp);
+  if(ret<0)return OC_BADPACKET;
+  _td->granulepos=gp;
+  return 0;
+}
+
+int theora_decode_YUVout(theora_state *_td,yuv_buffer *_yuv){
+  th_api_wrapper  *api;
+  th_dec_ctx      *decode;
+  th_ycbcr_buffer  buf;
+  int              ret;
+  if(!_td||!_td->i||!_td->i->codec_setup)return OC_FAULT;
+  api=(th_api_wrapper *)_td->i->codec_setup;
+  decode=(th_dec_ctx *)api->decode;
+  if(!decode)return OC_FAULT;
+  ret=th_decode_ycbcr_out(decode,buf);
+  if(ret>=0){
+    _yuv->y_width=buf[0].width;
+    _yuv->y_height=buf[0].height;
+    _yuv->y_stride=buf[0].stride;
+    _yuv->uv_width=buf[1].width;
+    _yuv->uv_height=buf[1].height;
+    _yuv->uv_stride=buf[1].stride;
+    _yuv->y=buf[0].data;
+    _yuv->u=buf[1].data;
+    _yuv->v=buf[2].data;
+  }
+  return ret;
+}

thirdparty/libtheora/decinfo.c

@@ -0,0 +1,279 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+#include <stdlib.h>
+#include <string.h>
+#include <limits.h>
+#include "decint.h"
+
+/*Only used for fuzzing.*/
+#if defined(HAVE_MEMORY_CONSTRAINT)
+static const int MAX_FUZZING_WIDTH = 16384;
+static const int MAX_FUZZING_HEIGHT = 16384;
+#endif
+
+
+/*Unpacks a series of octets from a given byte array into the pack buffer.
+  No checking is done to ensure the buffer contains enough data.
+  _opb: The pack buffer to read the octets from.
+  _buf: The byte array to store the unpacked bytes in.
+  _len: The number of octets to unpack.*/
+static void oc_unpack_octets(oc_pack_buf *_opb,char *_buf,size_t _len){
+  while(_len-->0){
+    long val;
+    val=oc_pack_read(_opb,8);
+    *_buf++=(char)val;
+  }
+}
+
+/*Unpacks a 32-bit integer encoded by octets in little-endian form.*/
+static long oc_unpack_length(oc_pack_buf *_opb){
+  long ret[4];
+  int  i;
+  for(i=0;i<4;i++)ret[i]=oc_pack_read(_opb,8);
+  return ret[0]|ret[1]<<8|ret[2]<<16|ret[3]<<24;
+}
+
+static int oc_info_unpack(oc_pack_buf *_opb,th_info *_info){
+  long val;
+  /*Check the codec bitstream version.*/
+  val=oc_pack_read(_opb,8);
+  _info->version_major=(unsigned char)val;
+  val=oc_pack_read(_opb,8);
+  _info->version_minor=(unsigned char)val;
+  val=oc_pack_read(_opb,8);
+  _info->version_subminor=(unsigned char)val;
+  /*verify we can parse this bitstream version.
+     We accept earlier minors and all subminors, by spec*/
+  if(_info->version_major>TH_VERSION_MAJOR||
+   (_info->version_major==TH_VERSION_MAJOR&&
+   _info->version_minor>TH_VERSION_MINOR)){
+    return TH_EVERSION;
+  }
+  /*Read the encoded frame description.*/
+  val=oc_pack_read(_opb,16);
+  _info->frame_width=(ogg_uint32_t)val<<4;
+  val=oc_pack_read(_opb,16);
+  _info->frame_height=(ogg_uint32_t)val<<4;
+  val=oc_pack_read(_opb,24);
+  _info->pic_width=(ogg_uint32_t)val;
+  val=oc_pack_read(_opb,24);
+  _info->pic_height=(ogg_uint32_t)val;
+  val=oc_pack_read(_opb,8);
+  _info->pic_x=(ogg_uint32_t)val;
+  val=oc_pack_read(_opb,8);
+  _info->pic_y=(ogg_uint32_t)val;
+  val=oc_pack_read(_opb,32);
+  _info->fps_numerator=(ogg_uint32_t)val;
+  val=oc_pack_read(_opb,32);
+  _info->fps_denominator=(ogg_uint32_t)val;
+  if(_info->frame_width==0||_info->frame_height==0||
+   _info->pic_width+_info->pic_x>_info->frame_width||
+   _info->pic_height+_info->pic_y>_info->frame_height||
+   _info->fps_numerator==0||_info->fps_denominator==0){
+    return TH_EBADHEADER;
+  }
+#if defined(HAVE_MEMORY_CONSTRAINT)
+  if(_info->frame_width>=MAX_FUZZING_WIDTH&&_info->frame_height>=MAX_FUZZING_HEIGHT){
+    return TH_EBADHEADER;
+  }
+#endif
+  /*Note: The sense of pic_y is inverted in what we pass back to the
+     application compared to how it is stored in the bitstream.
+    This is because the bitstream uses a right-handed coordinate system, while
+     applications expect a left-handed one.*/
+  _info->pic_y=_info->frame_height-_info->pic_height-_info->pic_y;
+  val=oc_pack_read(_opb,24);
+  _info->aspect_numerator=(ogg_uint32_t)val;
+  val=oc_pack_read(_opb,24);
+  _info->aspect_denominator=(ogg_uint32_t)val;
+  val=oc_pack_read(_opb,8);
+  _info->colorspace=(th_colorspace)val;
+  val=oc_pack_read(_opb,24);
+  _info->target_bitrate=(int)val;
+  val=oc_pack_read(_opb,6);
+  _info->quality=(int)val;
+  val=oc_pack_read(_opb,5);
+  _info->keyframe_granule_shift=(int)val;
+  val=oc_pack_read(_opb,2);
+  _info->pixel_fmt=(th_pixel_fmt)val;
+  if(_info->pixel_fmt==TH_PF_RSVD)return TH_EBADHEADER;
+  val=oc_pack_read(_opb,3);
+  if(val!=0||oc_pack_bytes_left(_opb)<0)return TH_EBADHEADER;
+  return 0;
+}
+
+static int oc_comment_unpack(oc_pack_buf *_opb,th_comment *_tc){
+  long len;
+  int  i;
+  /*Read the vendor string.*/
+  len=oc_unpack_length(_opb);
+  if(len<0||len>oc_pack_bytes_left(_opb))return TH_EBADHEADER;
+  _tc->vendor=_ogg_malloc((size_t)len+1);
+  if(_tc->vendor==NULL)return TH_EFAULT;
+  oc_unpack_octets(_opb,_tc->vendor,len);
+  _tc->vendor[len]='\0';
+  /*Read the user comments.*/
+  _tc->comments=(int)oc_unpack_length(_opb);
+  len=_tc->comments;
+  if(len<0||len>(LONG_MAX>>2)||len<<2>oc_pack_bytes_left(_opb)){
+    _tc->comments=0;
+    return TH_EBADHEADER;
+  }
+  if(0<_tc->comments){
+    _tc->comment_lengths=(int *)_ogg_malloc(
+     _tc->comments*sizeof(_tc->comment_lengths[0]));
+    _tc->user_comments=(char **)_ogg_malloc(
+     _tc->comments*sizeof(_tc->user_comments[0]));
+    if(_tc->comment_lengths==NULL||_tc->user_comments==NULL){
+      _tc->comments=0;
+      return TH_EFAULT;
+    }
+    for(i=0;i<_tc->comments;i++){
+      len=oc_unpack_length(_opb);
+      if(len<0||len>oc_pack_bytes_left(_opb)){
+        _tc->comments=i;
+        return TH_EBADHEADER;
+      }
+      _tc->comment_lengths[i]=len;
+      _tc->user_comments[i]=_ogg_malloc((size_t)len+1);
+      if(_tc->user_comments[i]==NULL){
+        _tc->comments=i;
+        return TH_EFAULT;
+      }
+      oc_unpack_octets(_opb,_tc->user_comments[i],len);
+      _tc->user_comments[i][len]='\0';
+    }
+  } else {
+    _tc->comment_lengths=NULL;
+    _tc->user_comments=NULL;
+  }
+  return oc_pack_bytes_left(_opb)<0?TH_EBADHEADER:0;
+}
+
+static int oc_setup_unpack(oc_pack_buf *_opb,th_setup_info *_setup){
+  int ret;
+  /*Read the quantizer tables.*/
+  ret=oc_quant_params_unpack(_opb,&_setup->qinfo);
+  if(ret<0)return ret;
+  /*Read the Huffman trees.*/
+  return oc_huff_trees_unpack(_opb,_setup->huff_tables);
+}
+
+static void oc_setup_clear(th_setup_info *_setup){
+  oc_quant_params_clear(&_setup->qinfo);
+  oc_huff_trees_clear(_setup->huff_tables);
+}
+
+static int oc_dec_headerin(oc_pack_buf *_opb,th_info *_info,
+ th_comment *_tc,th_setup_info **_setup,ogg_packet *_op){
+  char buffer[6];
+  long val;
+  int  packtype;
+  int  ret;
+  val=oc_pack_read(_opb,8);
+  packtype=(int)val;
+  /*If we're at a data packet...*/
+  if(!(packtype&0x80)){
+    /*Check to make sure we received all three headers...
+      If we haven't seen any valid headers, assume this is not actually
+       Theora.*/
+    if(_info->frame_width<=0)return TH_ENOTFORMAT;
+    /*Follow our documentation, which says we'll return TH_EFAULT if this
+       are NULL (_info was checked by our caller).*/
+    if(_tc==NULL)return TH_EFAULT;
+    /*And if any other headers were missing, declare this packet "out of
+       sequence" instead.*/
+    if(_tc->vendor==NULL)return TH_EBADHEADER;
+    /*Don't check this until it's needed, since we allow passing NULL for the
+       arguments that we're not expecting the next header to fill in yet.*/
+    if(_setup==NULL)return TH_EFAULT;
+    if(*_setup==NULL)return TH_EBADHEADER;
+    /*If we got everything, we're done.*/
+    return 0;
+  }
+  /*Check the codec string.*/
+  oc_unpack_octets(_opb,buffer,6);
+  if(memcmp(buffer,"theora",6)!=0)return TH_ENOTFORMAT;
+  switch(packtype){
+    /*Codec info header.*/
+    case 0x80:{
+      /*This should be the first packet, and we should not already be
+         initialized.*/
+      if(!_op->b_o_s||_info->frame_width>0)return TH_EBADHEADER;
+      ret=oc_info_unpack(_opb,_info);
+      if(ret<0)th_info_clear(_info);
+      else ret=3;
+    }break;
+    /*Comment header.*/
+    case 0x81:{
+      if(_tc==NULL)return TH_EFAULT;
+      /*We should have already decoded the info header, and should not yet have
+         decoded the comment header.*/
+      if(_info->frame_width==0||_tc->vendor!=NULL)return TH_EBADHEADER;
+      ret=oc_comment_unpack(_opb,_tc);
+      if(ret<0)th_comment_clear(_tc);
+      else ret=2;
+    }break;
+    /*Codec setup header.*/
+    case 0x82:{
+      oc_setup_info *setup;
+      if(_tc==NULL||_setup==NULL)return TH_EFAULT;
+      /*We should have already decoded the info header and the comment header,
+         and should not yet have decoded the setup header.*/
+      if(_info->frame_width==0||_tc->vendor==NULL||*_setup!=NULL){
+        return TH_EBADHEADER;
+      }
+      setup=(oc_setup_info *)_ogg_calloc(1,sizeof(*setup));
+      if(setup==NULL)return TH_EFAULT;
+      ret=oc_setup_unpack(_opb,setup);
+      if(ret<0){
+        oc_setup_clear(setup);
+        _ogg_free(setup);
+      }
+      else{
+        *_setup=setup;
+        ret=1;
+      }
+    }break;
+    default:{
+      /*We don't know what this header is.*/
+      return TH_EBADHEADER;
+    }break;
+  }
+  return ret;
+}
+
+
+/*Decodes one header packet.
+  This should be called repeatedly with the packets at the beginning of the
+   stream until it returns 0.*/
+int th_decode_headerin(th_info *_info,th_comment *_tc,
+ th_setup_info **_setup,ogg_packet *_op){
+  oc_pack_buf opb;
+  if(_op==NULL)return TH_EBADHEADER;
+  if(_info==NULL)return TH_EFAULT;
+  oc_pack_readinit(&opb,_op->packet,_op->bytes);
+  return oc_dec_headerin(&opb,_info,_tc,_setup,_op);
+}
+
+void th_setup_free(th_setup_info *_setup){
+  if(_setup!=NULL){
+    oc_setup_clear(_setup);
+    _ogg_free(_setup);
+  }
+}

thirdparty/libtheora/decint.h

@@ -0,0 +1,185 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+#include <limits.h>
+#if !defined(_decint_H)
+# define _decint_H (1)
+# include "theora/theoradec.h"
+# include "state.h"
+# include "bitpack.h"
+# include "huffdec.h"
+# include "dequant.h"
+
+typedef struct th_setup_info         oc_setup_info;
+typedef struct oc_dec_opt_vtable     oc_dec_opt_vtable;
+typedef struct oc_dec_pipeline_state oc_dec_pipeline_state;
+typedef struct th_dec_ctx            oc_dec_ctx;
+
+
+
+/*Decoder-specific accelerated functions.*/
+# if defined(OC_C64X_ASM)
+#  include "c64x/c64xdec.h"
+# endif
+
+# if !defined(oc_dec_accel_init)
+#  define oc_dec_accel_init oc_dec_accel_init_c
+# endif
+# if defined(OC_DEC_USE_VTABLE)
+#  if !defined(oc_dec_dc_unpredict_mcu_plane)
+#   define oc_dec_dc_unpredict_mcu_plane(_dec,_pipe,_pli) \
+ ((*(_dec)->opt_vtable.dc_unpredict_mcu_plane)(_dec,_pipe,_pli))
+#  endif
+# else
+#  if !defined(oc_dec_dc_unpredict_mcu_plane)
+#   define oc_dec_dc_unpredict_mcu_plane oc_dec_dc_unpredict_mcu_plane_c
+#  endif
+# endif
+
+
+
+/*Constants for the packet-in state machine specific to the decoder.*/
+
+/*Next packet to read: Data packet.*/
+#define OC_PACKET_DATA (0)
+
+
+
+struct th_setup_info{
+  /*The Huffman codes.*/
+  ogg_int16_t   *huff_tables[TH_NHUFFMAN_TABLES];
+  /*The quantization parameters.*/
+  th_quant_info  qinfo;
+};
+
+
+
+/*Decoder specific functions with accelerated variants.*/
+struct oc_dec_opt_vtable{
+  void (*dc_unpredict_mcu_plane)(oc_dec_ctx *_dec,
+   oc_dec_pipeline_state *_pipe,int _pli);
+};
+
+
+
+struct oc_dec_pipeline_state{
+  /*Decoded DCT coefficients.
+    These are placed here instead of on the stack so that they can persist
+     between blocks, which makes clearing them back to zero much faster when
+     only a few non-zero coefficients were decoded.
+    It requires at least 65 elements because the zig-zag index array uses the
+     65th element as a dumping ground for out-of-range indices to protect us
+     from buffer overflow.
+    We make it fully twice as large so that the second half can serve as the
+     reconstruction buffer, which saves passing another parameter to all the
+     acceleration functions.
+    It also solves problems with 16-byte alignment for NEON on ARM.
+    gcc (as of 4.2.1) only seems to be able to give stack variables 8-byte
+     alignment, and silently produces incorrect results if you ask for 16.
+    Finally, keeping it off the stack means there's less likely to be a data
+     hazard between the NEON co-processor and the regular ARM core, which avoids
+     unnecessary stalls.*/
+  OC_ALIGN16(ogg_int16_t dct_coeffs[128]);
+  OC_ALIGN16(signed char bounding_values[256]);
+  ptrdiff_t           ti[3][64];
+  ptrdiff_t           ebi[3][64];
+  ptrdiff_t           eob_runs[3][64];
+  const ptrdiff_t    *coded_fragis[3];
+  const ptrdiff_t    *uncoded_fragis[3];
+  ptrdiff_t           ncoded_fragis[3];
+  ptrdiff_t           nuncoded_fragis[3];
+  const ogg_uint16_t *dequant[3][3][2];
+  int                 fragy0[3];
+  int                 fragy_end[3];
+  int                 pred_last[3][4];
+  int                 mcu_nvfrags;
+  int                 loop_filter;
+  int                 pp_level;
+};
+
+
+struct th_dec_ctx{
+  /*Shared encoder/decoder state.*/
+  oc_theora_state        state;
+  /*Whether or not packets are ready to be emitted.
+    This takes on negative values while there are remaining header packets to
+     be emitted, reaches 0 when the codec is ready for input, and goes to 1
+     when a frame has been processed and a data packet is ready.*/
+  int                    packet_state;
+  /*Buffer in which to assemble packets.*/
+  oc_pack_buf            opb;
+  /*Huffman decode trees.*/
+  ogg_int16_t           *huff_tables[TH_NHUFFMAN_TABLES];
+  /*The index of the first token in each plane for each coefficient.*/
+  ptrdiff_t              ti0[3][64];
+  /*The number of outstanding EOB runs at the start of each coefficient in each
+     plane.*/
+  ptrdiff_t              eob_runs[3][64];
+  /*The DCT token lists.*/
+  unsigned char         *dct_tokens;
+  /*The extra bits associated with DCT tokens.*/
+  unsigned char         *extra_bits;
+  /*The number of dct tokens unpacked so far.*/
+  int                    dct_tokens_count;
+  /*The out-of-loop post-processing level.*/
+  int                    pp_level;
+  /*The DC scale used for out-of-loop deblocking.*/
+  int                    pp_dc_scale[64];
+  /*The sharpen modifier used for out-of-loop deringing.*/
+  int                    pp_sharp_mod[64];
+  /*The DC quantization index of each block.*/
+  unsigned char         *dc_qis;
+  /*The variance of each block.*/
+  int                   *variances;
+  /*The storage for the post-processed frame buffer.*/
+  unsigned char         *pp_frame_data;
+  /*Whether or not the post-processsed frame buffer has space for chroma.*/
+  int                    pp_frame_state;
+  /*The buffer used for the post-processed frame.
+    Note that this is _not_ guaranteed to have the same strides and offsets as
+     the reference frame buffers.*/
+  th_ycbcr_buffer        pp_frame_buf;
+  /*The striped decode callback function.*/
+  th_stripe_callback     stripe_cb;
+  oc_dec_pipeline_state  pipe;
+# if defined(OC_DEC_USE_VTABLE)
+  /*Table for decoder acceleration functions.*/
+  oc_dec_opt_vtable      opt_vtable;
+# endif
+# if defined(HAVE_CAIRO)
+  /*Output metrics for debugging.*/
+  int                    telemetry_mbmode;
+  int                    telemetry_mv;
+  int                    telemetry_qi;
+  int                    telemetry_bits;
+  int                    telemetry_frame_bytes;
+  int                    telemetry_coding_bytes;
+  int                    telemetry_mode_bytes;
+  int                    telemetry_mv_bytes;
+  int                    telemetry_qi_bytes;
+  int                    telemetry_dc_bytes;
+  unsigned char         *telemetry_frame_data;
+# endif
+};
+
+/*Default pure-C implementations of decoder-specific accelerated functions.*/
+void oc_dec_accel_init_c(oc_dec_ctx *_dec);
+
+void oc_dec_dc_unpredict_mcu_plane_c(oc_dec_ctx *_dec,
+ oc_dec_pipeline_state *_pipe,int _pli);
+
+#endif

thirdparty/libtheora/dequant.c

@@ -0,0 +1,182 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+#include <stdlib.h>
+#include <string.h>
+#include <ogg/ogg.h>
+#include "dequant.h"
+#include "decint.h"
+
+int oc_quant_params_unpack(oc_pack_buf *_opb,th_quant_info *_qinfo){
+  th_quant_base *base_mats;
+  long           val;
+  int            nbase_mats;
+  int            sizes[64];
+  int            indices[64];
+  int            nbits;
+  int            bmi;
+  int            ci;
+  int            qti;
+  int            pli;
+  int            qri;
+  int            qi;
+  int            i;
+  val=oc_pack_read(_opb,3);
+  nbits=(int)val;
+  for(qi=0;qi<64;qi++){
+    val=oc_pack_read(_opb,nbits);
+    _qinfo->loop_filter_limits[qi]=(unsigned char)val;
+  }
+  val=oc_pack_read(_opb,4);
+  nbits=(int)val+1;
+  for(qi=0;qi<64;qi++){
+    val=oc_pack_read(_opb,nbits);
+    _qinfo->ac_scale[qi]=(ogg_uint16_t)val;
+  }
+  val=oc_pack_read(_opb,4);
+  nbits=(int)val+1;
+  for(qi=0;qi<64;qi++){
+    val=oc_pack_read(_opb,nbits);
+    _qinfo->dc_scale[qi]=(ogg_uint16_t)val;
+  }
+  val=oc_pack_read(_opb,9);
+  nbase_mats=(int)val+1;
+  base_mats=_ogg_malloc(nbase_mats*sizeof(base_mats[0]));
+  if(base_mats==NULL)return TH_EFAULT;
+  for(bmi=0;bmi<nbase_mats;bmi++){
+    for(ci=0;ci<64;ci++){
+      val=oc_pack_read(_opb,8);
+      base_mats[bmi][ci]=(unsigned char)val;
+    }
+  }
+  nbits=oc_ilog(nbase_mats-1);
+  for(i=0;i<6;i++){
+    th_quant_ranges *qranges;
+    th_quant_base   *qrbms;
+    int             *qrsizes;
+    qti=i/3;
+    pli=i%3;
+    qranges=_qinfo->qi_ranges[qti]+pli;
+    if(i>0){
+      val=oc_pack_read1(_opb);
+      if(!val){
+        int qtj;
+        int plj;
+        if(qti>0){
+          val=oc_pack_read1(_opb);
+          if(val){
+            qtj=qti-1;
+            plj=pli;
+          }
+          else{
+            qtj=(i-1)/3;
+            plj=(i-1)%3;
+          }
+        }
+        else{
+          qtj=(i-1)/3;
+          plj=(i-1)%3;
+        }
+        *qranges=*(_qinfo->qi_ranges[qtj]+plj);
+        continue;
+      }
+    }
+    val=oc_pack_read(_opb,nbits);
+    indices[0]=(int)val;
+    for(qi=qri=0;qi<63;){
+      val=oc_pack_read(_opb,oc_ilog(62-qi));
+      sizes[qri]=(int)val+1;
+      qi+=(int)val+1;
+      val=oc_pack_read(_opb,nbits);
+      indices[++qri]=(int)val;
+    }
+    /*Note: The caller is responsible for cleaning up any partially
+       constructed qinfo.*/
+    if(qi>63){
+      _ogg_free(base_mats);
+      return TH_EBADHEADER;
+    }
+    qranges->nranges=qri;
+    qranges->sizes=qrsizes=(int *)_ogg_malloc(qri*sizeof(qrsizes[0]));
+    if(qranges->sizes==NULL){
+      /*Note: The caller is responsible for cleaning up any partially
+         constructed qinfo.*/
+      _ogg_free(base_mats);
+      return TH_EFAULT;
+    }
+    memcpy(qrsizes,sizes,qri*sizeof(qrsizes[0]));
+    qrbms=(th_quant_base *)_ogg_malloc((qri+1)*sizeof(qrbms[0]));
+    if(qrbms==NULL){
+      /*Note: The caller is responsible for cleaning up any partially
+         constructed qinfo.*/
+      _ogg_free(base_mats);
+      return TH_EFAULT;
+    }
+    qranges->base_matrices=(const th_quant_base *)qrbms;
+    do{
+      bmi=indices[qri];
+      /*Note: The caller is responsible for cleaning up any partially
+         constructed qinfo.*/
+      if(bmi>=nbase_mats){
+        _ogg_free(base_mats);
+        return TH_EBADHEADER;
+      }
+      memcpy(qrbms[qri],base_mats[bmi],sizeof(qrbms[qri]));
+    }
+    while(qri-->0);
+  }
+  _ogg_free(base_mats);
+  return 0;
+}
+
+void oc_quant_params_clear(th_quant_info *_qinfo){
+  int i;
+  for(i=6;i-->0;){
+    int qti;
+    int pli;
+    qti=i/3;
+    pli=i%3;
+    /*Clear any duplicate pointer references.*/
+    if(i>0){
+      int qtj;
+      int plj;
+      qtj=(i-1)/3;
+      plj=(i-1)%3;
+      if(_qinfo->qi_ranges[qti][pli].sizes==
+       _qinfo->qi_ranges[qtj][plj].sizes){
+        _qinfo->qi_ranges[qti][pli].sizes=NULL;
+      }
+      if(_qinfo->qi_ranges[qti][pli].base_matrices==
+       _qinfo->qi_ranges[qtj][plj].base_matrices){
+        _qinfo->qi_ranges[qti][pli].base_matrices=NULL;
+      }
+    }
+    if(qti>0){
+      if(_qinfo->qi_ranges[1][pli].sizes==
+       _qinfo->qi_ranges[0][pli].sizes){
+        _qinfo->qi_ranges[1][pli].sizes=NULL;
+      }
+      if(_qinfo->qi_ranges[1][pli].base_matrices==
+       _qinfo->qi_ranges[0][pli].base_matrices){
+        _qinfo->qi_ranges[1][pli].base_matrices=NULL;
+      }
+    }
+    /*Now free all the non-duplicate storage.*/
+    _ogg_free((void *)_qinfo->qi_ranges[qti][pli].sizes);
+    _ogg_free((void *)_qinfo->qi_ranges[qti][pli].base_matrices);
+  }
+}

thirdparty/libtheora/dequant.h

@@ -0,0 +1,27 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+#if !defined(_dequant_H)
+# define _dequant_H (1)
+# include "quant.h"
+# include "bitpack.h"
+
+int oc_quant_params_unpack(oc_pack_buf *_opb,
+ th_quant_info *_qinfo);
+void oc_quant_params_clear(th_quant_info *_qinfo);
+
+#endif

thirdparty/libtheora/encapiwrapper.c

@@ -0,0 +1,168 @@
+#include <stdlib.h>
+#include <string.h>
+#include <limits.h>
+#include "apiwrapper.h"
+#include "encint.h"
+#include "theora/theoraenc.h"
+
+
+
+static void th_enc_api_clear(th_api_wrapper *_api){
+  if(_api->encode)th_encode_free(_api->encode);
+  memset(_api,0,sizeof(*_api));
+}
+
+static void theora_encode_clear(theora_state *_te){
+  if(_te->i!=NULL)theora_info_clear(_te->i);
+  memset(_te,0,sizeof(*_te));
+}
+
+static int theora_encode_control(theora_state *_te,int _req,
+ void *_buf,size_t _buf_sz){
+  return th_encode_ctl(((th_api_wrapper *)_te->i->codec_setup)->encode,
+   _req,_buf,_buf_sz);
+}
+
+static ogg_int64_t theora_encode_granule_frame(theora_state *_te,
+ ogg_int64_t _gp){
+  return th_granule_frame(((th_api_wrapper *)_te->i->codec_setup)->encode,_gp);
+}
+
+static double theora_encode_granule_time(theora_state *_te,ogg_int64_t _gp){
+  return th_granule_time(((th_api_wrapper *)_te->i->codec_setup)->encode,_gp);
+}
+
+static const oc_state_dispatch_vtable OC_ENC_DISPATCH_VTBL={
+  (oc_state_clear_func)theora_encode_clear,
+  (oc_state_control_func)theora_encode_control,
+  (oc_state_granule_frame_func)theora_encode_granule_frame,
+  (oc_state_granule_time_func)theora_encode_granule_time,
+};
+
+int theora_encode_init(theora_state *_te,theora_info *_ci){
+  th_api_info *apiinfo;
+  th_info      info;
+  ogg_uint32_t keyframe_frequency_force;
+  /*Allocate our own combined API wrapper/theora_info struct.
+    We put them both in one malloc'd block so that when the API wrapper is
+     freed, the info struct goes with it.
+    This avoids having to figure out whether or not we need to free the info
+     struct in either theora_info_clear() or theora_clear().*/
+  apiinfo=(th_api_info *)_ogg_malloc(sizeof(*apiinfo));
+  if(apiinfo==NULL)return TH_EFAULT;
+  /*Make our own copy of the info struct, since its lifetime should be
+     independent of the one we were passed in.*/
+  *&apiinfo->info=*_ci;
+  oc_theora_info2th_info(&info,_ci);
+  apiinfo->api.encode=th_encode_alloc(&info);
+  if(apiinfo->api.encode==NULL){
+    _ogg_free(apiinfo);
+    return OC_EINVAL;
+  }
+  apiinfo->api.clear=(oc_setup_clear_func)th_enc_api_clear;
+  /*Provide entry points for ABI compatibility with old decoder shared libs.*/
+  _te->internal_encode=(void *)&OC_ENC_DISPATCH_VTBL;
+  _te->internal_decode=NULL;
+  _te->granulepos=0;
+  _te->i=&apiinfo->info;
+  _te->i->codec_setup=&apiinfo->api;
+  /*Set the precise requested keyframe frequency.*/
+  keyframe_frequency_force=_ci->keyframe_auto_p?
+   _ci->keyframe_frequency_force:_ci->keyframe_frequency;
+  th_encode_ctl(apiinfo->api.encode,
+   TH_ENCCTL_SET_KEYFRAME_FREQUENCY_FORCE,
+   &keyframe_frequency_force,sizeof(keyframe_frequency_force));
+  /*TODO: Additional codec setup using the extra fields in theora_info.*/
+  return 0;
+}
+
+int theora_encode_YUVin(theora_state *_te,yuv_buffer *_yuv){
+  th_api_wrapper  *api;
+  th_ycbcr_buffer  buf;
+  int              ret;
+  api=(th_api_wrapper *)_te->i->codec_setup;
+  buf[0].width=_yuv->y_width;
+  buf[0].height=_yuv->y_height;
+  buf[0].stride=_yuv->y_stride;
+  buf[0].data=_yuv->y;
+  buf[1].width=_yuv->uv_width;
+  buf[1].height=_yuv->uv_height;
+  buf[1].stride=_yuv->uv_stride;
+  buf[1].data=_yuv->u;
+  buf[2].width=_yuv->uv_width;
+  buf[2].height=_yuv->uv_height;
+  buf[2].stride=_yuv->uv_stride;
+  buf[2].data=_yuv->v;
+  ret=th_encode_ycbcr_in(api->encode,buf);
+  if(ret<0)return ret;
+  _te->granulepos=api->encode->state.granpos;
+  return ret;
+}
+
+int theora_encode_packetout(theora_state *_te,int _last_p,ogg_packet *_op){
+  th_api_wrapper *api;
+  api=(th_api_wrapper *)_te->i->codec_setup;
+  return th_encode_packetout(api->encode,_last_p,_op);
+}
+
+int theora_encode_header(theora_state *_te,ogg_packet *_op){
+  oc_enc_ctx     *enc;
+  th_api_wrapper *api;
+  int             ret;
+  api=(th_api_wrapper *)_te->i->codec_setup;
+  enc=api->encode;
+  /*If we've already started encoding, fail.*/
+  if(enc->packet_state>OC_PACKET_EMPTY||enc->state.granpos!=0){
+    return TH_EINVAL;
+  }
+  /*Reset the state to make sure we output an info packet.*/
+  enc->packet_state=OC_PACKET_INFO_HDR;
+  ret=th_encode_flushheader(api->encode,NULL,_op);
+  return ret>=0?0:ret;
+}
+
+int theora_encode_comment(theora_comment *_tc,ogg_packet *_op){
+  oggpack_buffer  opb;
+  void           *buf;
+  int             packet_state;
+  int             ret;
+  packet_state=OC_PACKET_COMMENT_HDR;
+  oggpackB_writeinit(&opb);
+  ret=oc_state_flushheader(NULL,&packet_state,&opb,NULL,NULL,
+   th_version_string(),(th_comment *)_tc,_op);
+  if(ret>=0){
+    /*The oggpack_buffer's lifetime ends with this function, so we have to
+       copy out the packet contents.
+      Presumably the application knows it is supposed to free this.
+      This part works nothing like the Vorbis API, and the documentation on it
+       has been wrong for some time, claiming libtheora owned the memory.*/
+    buf=_ogg_malloc(_op->bytes);
+    if(buf==NULL){
+      _op->packet=NULL;
+      ret=TH_EFAULT;
+    }
+    else{
+      memcpy(buf,_op->packet,_op->bytes);
+      _op->packet=buf;
+      ret=0;
+    }
+  }
+  oggpack_writeclear(&opb);
+  return ret;
+}
+
+int theora_encode_tables(theora_state *_te,ogg_packet *_op){
+  oc_enc_ctx     *enc;
+  th_api_wrapper *api;
+  int             ret;
+  api=(th_api_wrapper *)_te->i->codec_setup;
+  enc=api->encode;
+  /*If we've already started encoding, fail.*/
+  if(enc->packet_state>OC_PACKET_EMPTY||enc->state.granpos!=0){
+    return TH_EINVAL;
+  }
+  /*Reset the state to make sure we output a setup packet.*/
+  enc->packet_state=OC_PACKET_SETUP_HDR;
+  ret=th_encode_flushheader(api->encode,NULL,_op);
+  return ret>=0?0:ret;
+}

thirdparty/libtheora/encfrag.c

@@ -0,0 +1,378 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009,2025           *
+ * by the Xiph.Org Foundation https://www.xiph.org/                 *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+#include <stdlib.h>
+#include <string.h>
+#include "encint.h"
+
+
+void oc_enc_frag_sub_c(ogg_int16_t _diff[64],const unsigned char *_src,
+ const unsigned char *_ref,int _ystride){
+  int i;
+  for(i=0;i<8;i++){
+    int j;
+    for(j=0;j<8;j++)_diff[i*8+j]=(ogg_int16_t)(_src[j]-_ref[j]);
+    _src+=_ystride;
+    _ref+=_ystride;
+  }
+}
+
+void oc_enc_frag_sub_128_c(ogg_int16_t _diff[64],
+ const unsigned char *_src,int _ystride){
+  int i;
+  for(i=0;i<8;i++){
+    int j;
+    for(j=0;j<8;j++)_diff[i*8+j]=(ogg_int16_t)(_src[j]-128);
+    _src+=_ystride;
+  }
+}
+
+unsigned oc_enc_frag_sad_c(const unsigned char *_src,
+ const unsigned char *_ref,int _ystride){
+  unsigned sad;
+  int      i;
+  sad=0;
+  for(i=8;i-->0;){
+    int j;
+    for(j=0;j<8;j++)sad+=abs(_src[j]-_ref[j]);
+    _src+=_ystride;
+    _ref+=_ystride;
+  }
+  return sad;
+}
+
+unsigned oc_enc_frag_sad_thresh_c(const unsigned char *_src,
+ const unsigned char *_ref,int _ystride,unsigned _thresh){
+  unsigned sad;
+  int      i;
+  sad=0;
+  for(i=8;i-->0;){
+    int j;
+    for(j=0;j<8;j++)sad+=abs(_src[j]-_ref[j]);
+    if(sad>_thresh)break;
+    _src+=_ystride;
+    _ref+=_ystride;
+  }
+  return sad;
+}
+
+unsigned oc_enc_frag_sad2_thresh_c(const unsigned char *_src,
+ const unsigned char *_ref1,const unsigned char *_ref2,int _ystride,
+ unsigned _thresh){
+  unsigned sad;
+  int      i;
+  sad=0;
+  for(i=8;i-->0;){
+    int j;
+    for(j=0;j<8;j++)sad+=abs(_src[j]-(_ref1[j]+_ref2[j]>>1));
+    if(sad>_thresh)break;
+    _src+=_ystride;
+    _ref1+=_ystride;
+    _ref2+=_ystride;
+  }
+  return sad;
+}
+
+unsigned oc_enc_frag_intra_sad_c(const unsigned char *_src, int _ystride){
+  const unsigned char *src = _src;
+  int dc;
+  unsigned sad;
+  int      i;
+  dc=0;
+  for(i=8;i-->0;){
+    int j;
+    for(j=0;j<8;j++)dc+=src[j];
+    src+=_ystride;
+  }
+  dc=dc+32>>6;
+  sad=0;
+  for(i=8;i-->0;){
+    int j;
+    for(j=0;j<8;j++)sad+=abs(_src[j]-dc);
+    _src+=_ystride;
+  }
+  return sad;
+}
+
+static void oc_diff_hadamard(ogg_int16_t _buf[64],const unsigned char *_src,
+ const unsigned char *_ref,int _ystride){
+  int i;
+  for(i=0;i<8;i++){
+    int t0;
+    int t1;
+    int t2;
+    int t3;
+    int t4;
+    int t5;
+    int t6;
+    int t7;
+    int r;
+    /*Hadamard stage 1:*/
+    t0=_src[0]-_ref[0]+_src[4]-_ref[4];
+    t4=_src[0]-_ref[0]-_src[4]+_ref[4];
+    t1=_src[1]-_ref[1]+_src[5]-_ref[5];
+    t5=_src[1]-_ref[1]-_src[5]+_ref[5];
+    t2=_src[2]-_ref[2]+_src[6]-_ref[6];
+    t6=_src[2]-_ref[2]-_src[6]+_ref[6];
+    t3=_src[3]-_ref[3]+_src[7]-_ref[7];
+    t7=_src[3]-_ref[3]-_src[7]+_ref[7];
+    /*Hadamard stage 2:*/
+    r=t0;
+    t0+=t2;
+    t2=r-t2;
+    r=t1;
+    t1+=t3;
+    t3=r-t3;
+    r=t4;
+    t4+=t6;
+    t6=r-t6;
+    r=t5;
+    t5+=t7;
+    t7=r-t7;
+    /*Hadamard stage 3:*/
+    _buf[0*8+i]=(ogg_int16_t)(t0+t1);
+    _buf[1*8+i]=(ogg_int16_t)(t0-t1);
+    _buf[2*8+i]=(ogg_int16_t)(t2+t3);
+    _buf[3*8+i]=(ogg_int16_t)(t2-t3);
+    _buf[4*8+i]=(ogg_int16_t)(t4+t5);
+    _buf[5*8+i]=(ogg_int16_t)(t4-t5);
+    _buf[6*8+i]=(ogg_int16_t)(t6+t7);
+    _buf[7*8+i]=(ogg_int16_t)(t6-t7);
+    _src+=_ystride;
+    _ref+=_ystride;
+  }
+}
+
+static void oc_diff_hadamard2(ogg_int16_t _buf[64],const unsigned char *_src,
+ const unsigned char *_ref1,const unsigned char *_ref2,int _ystride){
+  int i;
+  for(i=0;i<8;i++){
+    int t0;
+    int t1;
+    int t2;
+    int t3;
+    int t4;
+    int t5;
+    int t6;
+    int t7;
+    int r;
+    /*Hadamard stage 1:*/
+    r=_ref1[0]+_ref2[0]>>1;
+    t4=_ref1[4]+_ref2[4]>>1;
+    t0=_src[0]-r+_src[4]-t4;
+    t4=_src[0]-r-_src[4]+t4;
+    r=_ref1[1]+_ref2[1]>>1;
+    t5=_ref1[5]+_ref2[5]>>1;
+    t1=_src[1]-r+_src[5]-t5;
+    t5=_src[1]-r-_src[5]+t5;
+    r=_ref1[2]+_ref2[2]>>1;
+    t6=_ref1[6]+_ref2[6]>>1;
+    t2=_src[2]-r+_src[6]-t6;
+    t6=_src[2]-r-_src[6]+t6;
+    r=_ref1[3]+_ref2[3]>>1;
+    t7=_ref1[7]+_ref2[7]>>1;
+    t3=_src[3]-r+_src[7]-t7;
+    t7=_src[3]-r-_src[7]+t7;
+    /*Hadamard stage 2:*/
+    r=t0;
+    t0+=t2;
+    t2=r-t2;
+    r=t1;
+    t1+=t3;
+    t3=r-t3;
+    r=t4;
+    t4+=t6;
+    t6=r-t6;
+    r=t5;
+    t5+=t7;
+    t7=r-t7;
+    /*Hadamard stage 3:*/
+    _buf[0*8+i]=(ogg_int16_t)(t0+t1);
+    _buf[1*8+i]=(ogg_int16_t)(t0-t1);
+    _buf[2*8+i]=(ogg_int16_t)(t2+t3);
+    _buf[3*8+i]=(ogg_int16_t)(t2-t3);
+    _buf[4*8+i]=(ogg_int16_t)(t4+t5);
+    _buf[5*8+i]=(ogg_int16_t)(t4-t5);
+    _buf[6*8+i]=(ogg_int16_t)(t6+t7);
+    _buf[7*8+i]=(ogg_int16_t)(t6-t7);
+    _src+=_ystride;
+    _ref1+=_ystride;
+    _ref2+=_ystride;
+  }
+}
+
+static void oc_intra_hadamard(ogg_int16_t _buf[64],const unsigned char *_src,
+ int _ystride){
+  int i;
+  for(i=0;i<8;i++){
+    int t0;
+    int t1;
+    int t2;
+    int t3;
+    int t4;
+    int t5;
+    int t6;
+    int t7;
+    int r;
+    /*Hadamard stage 1:*/
+    t0=_src[0]+_src[4];
+    t4=_src[0]-_src[4];
+    t1=_src[1]+_src[5];
+    t5=_src[1]-_src[5];
+    t2=_src[2]+_src[6];
+    t6=_src[2]-_src[6];
+    t3=_src[3]+_src[7];
+    t7=_src[3]-_src[7];
+    /*Hadamard stage 2:*/
+    r=t0;
+    t0+=t2;
+    t2=r-t2;
+    r=t1;
+    t1+=t3;
+    t3=r-t3;
+    r=t4;
+    t4+=t6;
+    t6=r-t6;
+    r=t5;
+    t5+=t7;
+    t7=r-t7;
+    /*Hadamard stage 3:*/
+    _buf[0*8+i]=(ogg_int16_t)(t0+t1);
+    _buf[1*8+i]=(ogg_int16_t)(t0-t1);
+    _buf[2*8+i]=(ogg_int16_t)(t2+t3);
+    _buf[3*8+i]=(ogg_int16_t)(t2-t3);
+    _buf[4*8+i]=(ogg_int16_t)(t4+t5);
+    _buf[5*8+i]=(ogg_int16_t)(t4-t5);
+    _buf[6*8+i]=(ogg_int16_t)(t6+t7);
+    _buf[7*8+i]=(ogg_int16_t)(t6-t7);
+    _src+=_ystride;
+  }
+}
+
+unsigned oc_hadamard_sad(int *_dc,const ogg_int16_t _buf[64]){
+  unsigned sad;
+  int      dc;
+  int      t0;
+  int      t1;
+  int      t2;
+  int      t3;
+  int      t4;
+  int      t5;
+  int      t6;
+  int      t7;
+  int      r;
+  int      i;
+  sad=dc=0;
+  for(i=0;i<8;i++){
+    /*Hadamard stage 1:*/
+    t0=_buf[i*8+0]+_buf[i*8+4];
+    t4=_buf[i*8+0]-_buf[i*8+4];
+    t1=_buf[i*8+1]+_buf[i*8+5];
+    t5=_buf[i*8+1]-_buf[i*8+5];
+    t2=_buf[i*8+2]+_buf[i*8+6];
+    t6=_buf[i*8+2]-_buf[i*8+6];
+    t3=_buf[i*8+3]+_buf[i*8+7];
+    t7=_buf[i*8+3]-_buf[i*8+7];
+    /*Hadamard stage 2:*/
+    r=t0;
+    t0+=t2;
+    t2=r-t2;
+    r=t1;
+    t1+=t3;
+    t3=r-t3;
+    r=t4;
+    t4+=t6;
+    t6=r-t6;
+    r=t5;
+    t5+=t7;
+    t7=r-t7;
+    /*Hadamard stage 3:*/
+    r=abs(t0+t1)&-(i>0);
+    r+=abs(t0-t1);
+    r+=abs(t2+t3);
+    r+=abs(t2-t3);
+    r+=abs(t4+t5);
+    r+=abs(t4-t5);
+    r+=abs(t6+t7);
+    r+=abs(t6-t7);
+    sad+=r;
+  }
+  dc=_buf[0]+_buf[1]+_buf[2]+_buf[3]+_buf[4]+_buf[5]+_buf[6]+_buf[7];
+  *_dc=dc;
+  return sad;
+}
+
+unsigned oc_enc_frag_satd_c(int *_dc,const unsigned char *_src,
+ const unsigned char *_ref,int _ystride){
+  ogg_int16_t buf[64];
+  oc_diff_hadamard(buf,_src,_ref,_ystride);
+  return oc_hadamard_sad(_dc,buf);
+}
+
+unsigned oc_enc_frag_satd2_c(int *_dc,const unsigned char *_src,
+ const unsigned char *_ref1,const unsigned char *_ref2,int _ystride){
+  ogg_int16_t buf[64];
+  oc_diff_hadamard2(buf,_src,_ref1,_ref2,_ystride);
+  return oc_hadamard_sad(_dc,buf);
+}
+
+unsigned oc_enc_frag_intra_satd_c(int *_dc,
+ const unsigned char *_src,int _ystride){
+  ogg_int16_t buf[64];
+  oc_intra_hadamard(buf,_src,_ystride);
+  return oc_hadamard_sad(_dc,buf);
+}
+
+unsigned oc_enc_frag_ssd_c(const unsigned char *_src,
+ const unsigned char *_ref,int _ystride){
+  unsigned ret;
+  int      y;
+  int      x;
+  ret=0;
+  for(y=0;y<8;y++){
+    for(x=0;x<8;x++)ret+=(_src[x]-_ref[x])*(_src[x]-_ref[x]);
+    _src+=_ystride;
+    _ref+=_ystride;
+  }
+  return ret;
+}
+
+unsigned oc_enc_frag_border_ssd_c(const unsigned char *_src,
+ const unsigned char *_ref,int _ystride,ogg_int64_t _mask){
+  unsigned ret;
+  int      y;
+  int      x;
+  ret=0;
+  for(y=0;y<8;y++){
+    for(x=0;x<8;x++,_mask>>=1){
+      if(_mask&1)ret+=(_src[x]-_ref[x])*(_src[x]-_ref[x]);
+    }
+    _src+=_ystride;
+    _ref+=_ystride;
+  }
+  return ret;
+}
+
+void oc_enc_frag_copy2_c(unsigned char *_dst,
+ const unsigned char *_src1,const unsigned char *_src2,int _ystride){
+  int i;
+  int j;
+  for(i=8;i-->0;){
+    for(j=0;j<8;j++)_dst[j]=_src1[j]+_src2[j]>>1;
+    _dst+=_ystride;
+    _src1+=_ystride;
+    _src2+=_ystride;
+  }
+}

thirdparty/libtheora/encinfo.c

@@ -0,0 +1,121 @@
+#include <stdlib.h>
+#include <string.h>
+#include "state.h"
+#include "enquant.h"
+#include "huffenc.h"
+
+
+
+/*Packs a series of octets from a given byte array into the pack buffer.
+  _opb: The pack buffer to store the octets in.
+  _buf: The byte array containing the bytes to pack.
+  _len: The number of octets to pack.*/
+static void oc_pack_octets(oggpack_buffer *_opb,const char *_buf,int _len){
+  int i;
+  for(i=0;i<_len;i++)oggpackB_write(_opb,_buf[i],8);
+}
+
+
+
+int oc_state_flushheader(oc_theora_state *_state,int *_packet_state,
+ oggpack_buffer *_opb,const th_quant_info *_qinfo,
+ const th_huff_code _codes[TH_NHUFFMAN_TABLES][TH_NDCT_TOKENS],
+ const char *_vendor,th_comment *_tc,ogg_packet *_op){
+  unsigned char *packet;
+  int            b_o_s;
+  if(_op==NULL)return TH_EFAULT;
+  switch(*_packet_state){
+    /*Codec info header.*/
+    case OC_PACKET_INFO_HDR:{
+      if(_state==NULL)return TH_EFAULT;
+      oggpackB_reset(_opb);
+      /*Mark this packet as the info header.*/
+      oggpackB_write(_opb,0x80,8);
+      /*Write the codec string.*/
+      oc_pack_octets(_opb,"theora",6);
+      /*Write the codec bitstream version.*/
+      oggpackB_write(_opb,TH_VERSION_MAJOR,8);
+      oggpackB_write(_opb,TH_VERSION_MINOR,8);
+      oggpackB_write(_opb,TH_VERSION_SUB,8);
+      /*Describe the encoded frame.*/
+      oggpackB_write(_opb,_state->info.frame_width>>4,16);
+      oggpackB_write(_opb,_state->info.frame_height>>4,16);
+      oggpackB_write(_opb,_state->info.pic_width,24);
+      oggpackB_write(_opb,_state->info.pic_height,24);
+      oggpackB_write(_opb,_state->info.pic_x,8);
+      oggpackB_write(_opb,_state->info.pic_y,8);
+      oggpackB_write(_opb,_state->info.fps_numerator,32);
+      oggpackB_write(_opb,_state->info.fps_denominator,32);
+      oggpackB_write(_opb,_state->info.aspect_numerator,24);
+      oggpackB_write(_opb,_state->info.aspect_denominator,24);
+      oggpackB_write(_opb,_state->info.colorspace,8);
+      oggpackB_write(_opb,_state->info.target_bitrate,24);
+      oggpackB_write(_opb,_state->info.quality,6);
+      oggpackB_write(_opb,_state->info.keyframe_granule_shift,5);
+      oggpackB_write(_opb,_state->info.pixel_fmt,2);
+      /*Spare configuration bits.*/
+      oggpackB_write(_opb,0,3);
+      b_o_s=1;
+    }break;
+    /*Comment header.*/
+    case OC_PACKET_COMMENT_HDR:{
+      int vendor_len;
+      int i;
+      if(_tc==NULL)return TH_EFAULT;
+      vendor_len=strlen(_vendor);
+      oggpackB_reset(_opb);
+      /*Mark this packet as the comment header.*/
+      oggpackB_write(_opb,0x81,8);
+      /*Write the codec string.*/
+      oc_pack_octets(_opb,"theora",6);
+      /*Write the vendor string.*/
+      oggpack_write(_opb,vendor_len,32);
+      oc_pack_octets(_opb,_vendor,vendor_len);
+      oggpack_write(_opb,_tc->comments,32);
+      for(i=0;i<_tc->comments;i++){
+        if(_tc->user_comments[i]!=NULL){
+          oggpack_write(_opb,_tc->comment_lengths[i],32);
+          oc_pack_octets(_opb,_tc->user_comments[i],_tc->comment_lengths[i]);
+        }
+        else oggpack_write(_opb,0,32);
+      }
+      b_o_s=0;
+    }break;
+    /*Codec setup header.*/
+    case OC_PACKET_SETUP_HDR:{
+      int ret;
+      oggpackB_reset(_opb);
+      /*Mark this packet as the setup header.*/
+      oggpackB_write(_opb,0x82,8);
+      /*Write the codec string.*/
+      oc_pack_octets(_opb,"theora",6);
+      /*Write the quantizer tables.*/
+      oc_quant_params_pack(_opb,_qinfo);
+      /*Write the huffman codes.*/
+      ret=oc_huff_codes_pack(_opb,_codes);
+      /*This should never happen, because we validate the tables when they
+         are set.
+        If you see, it's a good chance memory is being corrupted.*/
+      if(ret<0)return ret;
+      b_o_s=0;
+    }break;
+    /*No more headers to emit.*/
+    default:return 0;
+  }
+  /*This is kind of fugly: we hand the user a buffer which they do not own.
+    We will overwrite it when the next packet is output, so the user better be
+     done with it by then.
+    Vorbis is little better: it hands back buffers that it will free the next
+     time the headers are requested, or when the encoder is cleared.
+    Hopefully libogg2 will make this much cleaner.*/
+  packet=oggpackB_get_buffer(_opb);
+  /*If there's no packet, malloc failed while writing.*/
+  if(packet==NULL)return TH_EFAULT;
+  _op->packet=packet;
+  _op->bytes=oggpackB_bytes(_opb);
+  _op->b_o_s=b_o_s;
+  _op->e_o_s=0;
+  _op->granulepos=0;
+  _op->packetno=*_packet_state+3;
+  return ++(*_packet_state)+3;
+}

thirdparty/libtheora/encint.h

@@ -0,0 +1,844 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation https://www.xiph.org/                 *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+#if !defined(_encint_H)
+# define _encint_H (1)
+# include "theora/theoraenc.h"
+# include "state.h"
+# include "mathops.h"
+# include "enquant.h"
+# include "huffenc.h"
+/*# define OC_COLLECT_METRICS*/
+
+
+
+typedef oc_mv                         oc_mv2[2];
+
+typedef struct oc_enc_opt_vtable      oc_enc_opt_vtable;
+typedef struct oc_enc_opt_data        oc_enc_opt_data;
+typedef struct oc_mb_enc_info         oc_mb_enc_info;
+typedef struct oc_mode_scheme_chooser oc_mode_scheme_chooser;
+typedef struct oc_fr_state            oc_fr_state;
+typedef struct oc_qii_state           oc_qii_state;
+typedef struct oc_enc_pipeline_state  oc_enc_pipeline_state;
+typedef struct oc_mode_rd             oc_mode_rd;
+typedef struct oc_iir_filter          oc_iir_filter;
+typedef struct oc_frame_metrics       oc_frame_metrics;
+typedef struct oc_rc_state            oc_rc_state;
+typedef struct th_enc_ctx             oc_enc_ctx;
+typedef struct oc_token_checkpoint    oc_token_checkpoint;
+
+
+
+/*Encoder-specific accelerated functions.*/
+# if defined(OC_X86_ASM)
+#  if defined(_MSC_VER)
+#   include "x86_vc/x86enc.h"
+#  else
+#   include "x86/x86enc.h"
+#  endif
+# endif
+# if defined(OC_ARM_ASM)
+#  include "arm/armenc.h"
+# endif
+
+# if !defined(oc_enc_accel_init)
+#  define oc_enc_accel_init oc_enc_accel_init_c
+# endif
+# if defined(OC_ENC_USE_VTABLE)
+#  if !defined(oc_enc_frag_sub)
+#   define oc_enc_frag_sub(_enc,_diff,_src,_ref,_ystride) \
+  ((*(_enc)->opt_vtable.frag_sub)(_diff,_src,_ref,_ystride))
+#  endif
+#  if !defined(oc_enc_frag_sub_128)
+#   define oc_enc_frag_sub_128(_enc,_diff,_src,_ystride) \
+  ((*(_enc)->opt_vtable.frag_sub_128)(_diff,_src,_ystride))
+#  endif
+#  if !defined(oc_enc_frag_sad)
+#   define oc_enc_frag_sad(_enc,_src,_ref,_ystride) \
+  ((*(_enc)->opt_vtable.frag_sad)(_src,_ref,_ystride))
+#  endif
+#  if !defined(oc_enc_frag_sad_thresh)
+#   define oc_enc_frag_sad_thresh(_enc,_src,_ref,_ystride,_thresh) \
+  ((*(_enc)->opt_vtable.frag_sad_thresh)(_src,_ref,_ystride,_thresh))
+#  endif
+#  if !defined(oc_enc_frag_sad2_thresh)
+#   define oc_enc_frag_sad2_thresh(_enc,_src,_ref1,_ref2,_ystride,_thresh) \
+  ((*(_enc)->opt_vtable.frag_sad2_thresh)(_src,_ref1,_ref2,_ystride,_thresh))
+#  endif
+#  if !defined(oc_enc_frag_intra_sad)
+#   define oc_enc_frag_intra_sad(_enc,_src,_ystride) \
+  ((*(_enc)->opt_vtable.frag_intra_sad)(_src,_ystride))
+#  endif
+#  if !defined(oc_enc_frag_satd)
+#   define oc_enc_frag_satd(_enc,_dc,_src,_ref,_ystride) \
+  ((*(_enc)->opt_vtable.frag_satd)(_dc,_src,_ref,_ystride))
+#  endif
+#  if !defined(oc_enc_frag_satd2)
+#   define oc_enc_frag_satd2(_enc,_dc,_src,_ref1,_ref2,_ystride) \
+  ((*(_enc)->opt_vtable.frag_satd2)(_dc,_src,_ref1,_ref2,_ystride))
+#  endif
+#  if !defined(oc_enc_frag_intra_satd)
+#   define oc_enc_frag_intra_satd(_enc,_dc,_src,_ystride) \
+  ((*(_enc)->opt_vtable.frag_intra_satd)(_dc,_src,_ystride))
+#  endif
+#  if !defined(oc_enc_frag_ssd)
+#   define oc_enc_frag_ssd(_enc,_src,_ref,_ystride) \
+  ((*(_enc)->opt_vtable.frag_ssd)(_src,_ref,_ystride))
+#  endif
+#  if !defined(oc_enc_frag_border_ssd)
+#   define oc_enc_frag_border_ssd(_enc,_src,_ref,_ystride,_mask) \
+  ((*(_enc)->opt_vtable.frag_border_ssd)(_src,_ref,_ystride,_mask))
+#  endif
+#  if !defined(oc_enc_frag_copy2)
+#   define oc_enc_frag_copy2(_enc,_dst,_src1,_src2,_ystride) \
+  ((*(_enc)->opt_vtable.frag_copy2)(_dst,_src1,_src2,_ystride))
+#  endif
+#  if !defined(oc_enc_enquant_table_init)
+#   define oc_enc_enquant_table_init(_enc,_enquant,_dequant) \
+  ((*(_enc)->opt_vtable.enquant_table_init)(_enquant,_dequant))
+#  endif
+#  if !defined(oc_enc_enquant_table_fixup)
+#   define oc_enc_enquant_table_fixup(_enc,_enquant,_nqis) \
+  ((*(_enc)->opt_vtable.enquant_table_fixup)(_enquant,_nqis))
+#  endif
+#  if !defined(oc_enc_quantize)
+#   define oc_enc_quantize(_enc,_qdct,_dct,_dequant,_enquant) \
+  ((*(_enc)->opt_vtable.quantize)(_qdct,_dct,_dequant,_enquant))
+#  endif
+#  if !defined(oc_enc_frag_recon_intra)
+#   define oc_enc_frag_recon_intra(_enc,_dst,_ystride,_residue) \
+  ((*(_enc)->opt_vtable.frag_recon_intra)(_dst,_ystride,_residue))
+#  endif
+#  if !defined(oc_enc_frag_recon_inter)
+#   define oc_enc_frag_recon_inter(_enc,_dst,_src,_ystride,_residue) \
+  ((*(_enc)->opt_vtable.frag_recon_inter)(_dst,_src,_ystride,_residue))
+#  endif
+#  if !defined(oc_enc_fdct8x8)
+#   define oc_enc_fdct8x8(_enc,_y,_x) \
+  ((*(_enc)->opt_vtable.fdct8x8)(_y,_x))
+#  endif
+# else
+#  if !defined(oc_enc_frag_sub)
+#   define oc_enc_frag_sub(_enc,_diff,_src,_ref,_ystride) \
+  oc_enc_frag_sub_c(_diff,_src,_ref,_ystride)
+#  endif
+#  if !defined(oc_enc_frag_sub_128)
+#   define oc_enc_frag_sub_128(_enc,_diff,_src,_ystride) \
+  oc_enc_frag_sub_128_c(_diff,_src,_ystride)
+#  endif
+#  if !defined(oc_enc_frag_sad)
+#   define oc_enc_frag_sad(_enc,_src,_ref,_ystride) \
+  oc_enc_frag_sad_c(_src,_ref,_ystride)
+#  endif
+#  if !defined(oc_enc_frag_sad_thresh)
+#   define oc_enc_frag_sad_thresh(_enc,_src,_ref,_ystride,_thresh) \
+  oc_enc_frag_sad_thresh_c(_src,_ref,_ystride,_thresh)
+#  endif
+#  if !defined(oc_enc_frag_sad2_thresh)
+#   define oc_enc_frag_sad2_thresh(_enc,_src,_ref1,_ref2,_ystride,_thresh) \
+  oc_enc_frag_sad2_thresh_c(_src,_ref1,_ref2,_ystride,_thresh)
+#  endif
+#  if !defined(oc_enc_frag_intra_sad)
+#   define oc_enc_frag_intra_sad(_enc,_src,_ystride) \
+  oc_enc_frag_intra_sad_c(_src,_ystride)
+#  endif
+#  if !defined(oc_enc_frag_satd)
+#   define oc_enc_frag_satd(_enc,_dc,_src,_ref,_ystride) \
+  oc_enc_frag_satd_c(_dc,_src,_ref,_ystride)
+#  endif
+#  if !defined(oc_enc_frag_satd2)
+#   define oc_enc_frag_satd2(_enc,_dc,_src,_ref1,_ref2,_ystride) \
+  oc_enc_frag_satd2_c(_dc,_src,_ref1,_ref2,_ystride)
+#  endif
+#  if !defined(oc_enc_frag_intra_satd)
+#   define oc_enc_frag_intra_satd(_enc,_dc,_src,_ystride) \
+  oc_enc_frag_intra_satd_c(_dc,_src,_ystride)
+#  endif
+#  if !defined(oc_enc_frag_ssd)
+#   define oc_enc_frag_ssd(_enc,_src,_ref,_ystride) \
+  oc_enc_frag_ssd_c(_src,_ref,_ystride)
+#  endif
+#  if !defined(oc_enc_frag_border_ssd)
+#   define oc_enc_frag_border_ssd(_enc,_src,_ref,_ystride,_mask) \
+  oc_enc_frag_border_ssd_c(_src,_ref,_ystride,_mask)
+#  endif
+#  if !defined(oc_enc_frag_copy2)
+#   define oc_enc_frag_copy2(_enc,_dst,_src1,_src2,_ystride) \
+  oc_enc_frag_copy2_c(_dst,_src1,_src2,_ystride)
+#  endif
+#  if !defined(oc_enc_enquant_table_init)
+#   define oc_enc_enquant_table_init(_enc,_enquant,_dequant) \
+  oc_enc_enquant_table_init_c(_enquant,_dequant)
+#  endif
+#  if !defined(oc_enc_enquant_table_fixup)
+#   define oc_enc_enquant_table_fixup(_enc,_enquant,_nqis) \
+  oc_enc_enquant_table_fixup_c(_enquant,_nqis)
+#  endif
+#  if !defined(oc_enc_quantize)
+#   define oc_enc_quantize(_enc,_qdct,_dct,_dequant,_enquant) \
+  oc_enc_quantize_c(_qdct,_dct,_dequant,_enquant)
+#  endif
+#  if !defined(oc_enc_frag_recon_intra)
+#   define oc_enc_frag_recon_intra(_enc,_dst,_ystride,_residue) \
+  oc_frag_recon_intra_c(_dst,_ystride,_residue)
+#  endif
+#  if !defined(oc_enc_frag_recon_inter)
+#   define oc_enc_frag_recon_inter(_enc,_dst,_src,_ystride,_residue) \
+  oc_frag_recon_inter_c(_dst,_src,_ystride,_residue)
+#  endif
+#  if !defined(oc_enc_fdct8x8)
+#   define oc_enc_fdct8x8(_enc,_y,_x) oc_enc_fdct8x8_c(_y,_x)
+#  endif
+# endif
+
+
+
+/*Constants for the packet-out state machine specific to the encoder.*/
+
+/*Next packet to emit: Data packet, but none are ready yet.*/
+#define OC_PACKET_EMPTY (0)
+/*Next packet to emit: Data packet, and one is ready.*/
+#define OC_PACKET_READY (1)
+
+/*All features enabled.*/
+#define OC_SP_LEVEL_SLOW          (0)
+/*Enable early skip.*/
+#define OC_SP_LEVEL_EARLY_SKIP    (1)
+/*Use analysis shortcuts, single quantizer, and faster tokenization.*/
+#define OC_SP_LEVEL_FAST_ANALYSIS (2)
+/*Use SAD instead of SATD*/
+#define OC_SP_LEVEL_NOSATD        (3)
+/*Disable motion compensation.*/
+#define OC_SP_LEVEL_NOMC          (4)
+/*Maximum valid speed level.*/
+#define OC_SP_LEVEL_MAX           (4)
+
+
+/*The number of extra bits of precision at which to store rate metrics.*/
+# define OC_BIT_SCALE  (6)
+/*The number of extra bits of precision at which to store RMSE metrics.
+  This must be at least half OC_BIT_SCALE (rounded up).*/
+# define OC_RMSE_SCALE (5)
+/*The number of quantizer bins to partition statistics into.*/
+# define OC_LOGQ_BINS  (8)
+/*The number of SAD/SATD bins to partition statistics into.*/
+# define OC_COMP_BINS   (24)
+/*The number of bits of precision to drop from SAD and SATD scores
+   to assign them to a bin.*/
+# define OC_SAD_SHIFT  (6)
+# define OC_SATD_SHIFT (9)
+
+/*Masking is applied by scaling the D used in R-D optimization (via rd_scale)
+   or the lambda parameter (via rd_iscale).
+  These are only equivalent within a single block; when more than one block is
+   being considered, the former is the interpretation used.*/
+
+/*This must be at least 4 for OC_RD_SKIP_SCALE() to work below.*/
+# define OC_RD_SCALE_BITS (12-OC_BIT_SCALE)
+# define OC_RD_ISCALE_BITS (11)
+
+/*This macro is applied to _ssd values with just 4 bits of headroom
+   ((15-OC_RMSE_SCALE)*2+OC_BIT_SCALE+2); since we want to allow rd_scales as
+   large as 16, and need additional fractional bits, our only recourse that
+   doesn't lose precision on blocks with very small SSDs is to use a wider
+   multiply.*/
+# if LONG_MAX>2147483647
+#  define OC_RD_SCALE(_ssd,_rd_scale) \
+ ((unsigned)((unsigned long)(_ssd)*(_rd_scale) \
+ +((1<<OC_RD_SCALE_BITS)>>1)>>OC_RD_SCALE_BITS))
+# else
+#  define OC_RD_SCALE(_ssd,_rd_scale) \
+ (((_ssd)>>OC_RD_SCALE_BITS)*(_rd_scale) \
+ +(((_ssd)&(1<<OC_RD_SCALE_BITS)-1)*(_rd_scale) \
+ +((1<<OC_RD_SCALE_BITS)>>1)>>OC_RD_SCALE_BITS))
+# endif
+# define OC_RD_SKIP_SCALE(_ssd,_rd_scale) \
+ ((_ssd)*(_rd_scale)+((1<<OC_RD_SCALE_BITS-4)>>1)>>OC_RD_SCALE_BITS-4)
+# define OC_RD_ISCALE(_lambda,_rd_iscale) \
+ ((_lambda)*(_rd_iscale)+((1<<OC_RD_ISCALE_BITS)>>1)>>OC_RD_ISCALE_BITS)
+
+
+/*The bits used for each of the MB mode codebooks.*/
+extern const unsigned char OC_MODE_BITS[2][OC_NMODES];
+
+/*The bits used for each of the MV codebooks.*/
+extern const unsigned char OC_MV_BITS[2][64];
+
+/*The minimum value that can be stored in a SB run for each codeword.
+  The last entry is the upper bound on the length of a single SB run.*/
+extern const ogg_uint16_t  OC_SB_RUN_VAL_MIN[8];
+/*The bits used for each SB run codeword.*/
+extern const unsigned char OC_SB_RUN_CODE_NBITS[7];
+
+/*The bits used for each block run length (starting with 1).*/
+extern const unsigned char OC_BLOCK_RUN_CODE_NBITS[30];
+
+
+
+/*Encoder specific functions with accelerated variants.*/
+struct oc_enc_opt_vtable{
+  void     (*frag_sub)(ogg_int16_t _diff[64],const unsigned char *_src,
+   const unsigned char *_ref,int _ystride);
+  void     (*frag_sub_128)(ogg_int16_t _diff[64],
+   const unsigned char *_src,int _ystride);
+  unsigned (*frag_sad)(const unsigned char *_src,
+   const unsigned char *_ref,int _ystride);
+  unsigned (*frag_sad_thresh)(const unsigned char *_src,
+   const unsigned char *_ref,int _ystride,unsigned _thresh);
+  unsigned (*frag_sad2_thresh)(const unsigned char *_src,
+   const unsigned char *_ref1,const unsigned char *_ref2,int _ystride,
+   unsigned _thresh);
+  unsigned (*frag_intra_sad)(const unsigned char *_src,int _ystride);
+  unsigned (*frag_satd)(int *_dc,const unsigned char *_src,
+   const unsigned char *_ref,int _ystride);
+  unsigned (*frag_satd2)(int *_dc,const unsigned char *_src,
+   const unsigned char *_ref1,const unsigned char *_ref2,int _ystride);
+  unsigned (*frag_intra_satd)(int *_dc,const unsigned char *_src,int _ystride);
+  unsigned (*frag_ssd)(const unsigned char *_src,
+   const unsigned char *_ref,int _ystride);
+  unsigned (*frag_border_ssd)(const unsigned char *_src,
+   const unsigned char *_ref,int _ystride,ogg_int64_t _mask);
+  void     (*frag_copy2)(unsigned char *_dst,
+   const unsigned char *_src1,const unsigned char *_src2,int _ystride);
+  void     (*enquant_table_init)(void *_enquant,
+   const ogg_uint16_t _dequant[64]);
+  void     (*enquant_table_fixup)(void *_enquant[3][3][2],int _nqis);
+  int      (*quantize)(ogg_int16_t _qdct[64],const ogg_int16_t _dct[64],
+   const ogg_uint16_t _dequant[64],const void *_enquant);
+  void     (*frag_recon_intra)(unsigned char *_dst,int _ystride,
+   const ogg_int16_t _residue[64]);
+  void     (*frag_recon_inter)(unsigned char *_dst,
+   const unsigned char *_src,int _ystride,const ogg_int16_t _residue[64]);
+  void     (*fdct8x8)(ogg_int16_t _y[64],const ogg_int16_t _x[64]);
+};
+
+
+/*Encoder specific data that varies according to which variants of the above
+   functions are used.*/
+struct oc_enc_opt_data{
+  /*The size of a single quantizer table.
+    This must be a multiple of enquant_table_alignment.*/
+  size_t               enquant_table_size;
+  /*The alignment required for the quantizer tables.
+    This must be a positive power of two.*/
+  int                  enquant_table_alignment;
+};
+
+
+void oc_enc_accel_init(oc_enc_ctx *_enc);
+
+
+
+/*Encoder-specific macroblock information.*/
+struct oc_mb_enc_info{
+  /*Neighboring macro blocks that have MVs available from the current frame.*/
+  unsigned      cneighbors[4];
+  /*Neighboring macro blocks to use for MVs from the previous frame.*/
+  unsigned      pneighbors[4];
+  /*The number of current-frame neighbors.*/
+  unsigned char ncneighbors;
+  /*The number of previous-frame neighbors.*/
+  unsigned char npneighbors;
+  /*Flags indicating which MB modes have been refined.*/
+  unsigned char refined;
+  /*Motion vectors for a macro block for the current frame and the
+     previous two frames.
+    Each is a set of 2 vectors against OC_FRAME_GOLD and OC_FRAME_PREV, which
+     can be used to estimate constant velocity and constant acceleration
+     predictors.
+    Uninitialized MVs are (0,0).*/
+  oc_mv2        analysis_mv[3];
+  /*Current unrefined analysis MVs.*/
+  oc_mv         unref_mv[2];
+  /*Unrefined block MVs.*/
+  oc_mv         block_mv[4];
+  /*Refined block MVs.*/
+  oc_mv         ref_mv[4];
+  /*Minimum motion estimation error from the analysis stage.*/
+  ogg_uint16_t  error[2];
+  /*MB error for half-pel refinement for each frame type.*/
+  unsigned      satd[2];
+  /*Block error for half-pel refinement.*/
+  unsigned      block_satd[4];
+};
+
+
+
+/*State machine to estimate the opportunity cost of coding a MB mode.*/
+struct oc_mode_scheme_chooser{
+  /*Pointers to the a list containing the index of each mode in the mode
+     alphabet used by each scheme.
+    The first entry points to the dynamic scheme0_ranks, while the remaining 7
+     point to the constant entries stored in OC_MODE_SCHEMES.*/
+  const unsigned char *mode_ranks[8];
+  /*The ranks for each mode when coded with scheme 0.
+    These are optimized so that the more frequent modes have lower ranks.*/
+  unsigned char        scheme0_ranks[OC_NMODES];
+  /*The list of modes, sorted in descending order of frequency, that
+    corresponds to the ranks above.*/
+  unsigned char        scheme0_list[OC_NMODES];
+  /*The number of times each mode has been chosen so far.*/
+  unsigned             mode_counts[OC_NMODES];
+  /*The list of mode coding schemes, sorted in ascending order of bit cost.*/
+  unsigned char        scheme_list[8];
+  /*The number of bits used by each mode coding scheme.*/
+  ptrdiff_t            scheme_bits[8];
+};
+
+
+void oc_mode_scheme_chooser_init(oc_mode_scheme_chooser *_chooser);
+
+
+
+/*State to track coded block flags and their bit cost.
+  We use opportunity cost to measure the bits required to code or skip the next
+   block, using the cheaper of the cost to code it fully or partially, so long
+   as both are possible.*/
+struct oc_fr_state{
+  /*The number of bits required for the coded block flags so far this frame.*/
+  ptrdiff_t  bits;
+  /*The length of the current run for the partial super block flag, not
+     including the current super block.*/
+  unsigned   sb_partial_count:16;
+  /*The length of the current run for the full super block flag, not
+     including the current super block.*/
+  unsigned   sb_full_count:16;
+  /*The length of the coded block flag run when the current super block
+     started.*/
+  unsigned   b_coded_count_prev:6;
+  /*The coded block flag when the current super block started.*/
+  signed int b_coded_prev:2;
+  /*The length of the current coded block flag run.*/
+  unsigned   b_coded_count:6;
+  /*The current coded block flag.*/
+  signed int b_coded:2;
+  /*The number of blocks processed in the current super block.*/
+  unsigned   b_count:5;
+  /*Whether or not it is cheaper to code the current super block partially,
+     even if it could still be coded fully.*/
+  unsigned   sb_prefer_partial:1;
+  /*Whether the last super block was coded partially.*/
+  signed int sb_partial:2;
+  /*The number of bits required for the flags for the current super block.*/
+  unsigned   sb_bits:6;
+  /*Whether the last non-partial super block was coded fully.*/
+  signed int sb_full:2;
+};
+
+
+
+struct oc_qii_state{
+  ptrdiff_t  bits;
+  unsigned   qi01_count:14;
+  signed int qi01:2;
+  unsigned   qi12_count:14;
+  signed int qi12:2;
+};
+
+
+
+/*Temporary encoder state for the analysis pipeline.*/
+struct oc_enc_pipeline_state{
+  /*DCT coefficient storage.
+    This is kept off the stack because a) gcc can't align things on the stack
+     reliably on ARM, and b) it avoids (unintentional) data hazards between
+     ARM and NEON code.*/
+  OC_ALIGN16(ogg_int16_t dct_data[64*3]);
+  OC_ALIGN16(signed char bounding_values[256]);
+  oc_fr_state         fr[3];
+  oc_qii_state        qs[3];
+  /*Skip SSD storage for the current MCU in each plane.*/
+  unsigned           *skip_ssd[3];
+  /*Coded/uncoded fragment lists for each plane for the current MCU.*/
+  ptrdiff_t          *coded_fragis[3];
+  ptrdiff_t          *uncoded_fragis[3];
+  ptrdiff_t           ncoded_fragis[3];
+  ptrdiff_t           nuncoded_fragis[3];
+  /*The starting fragment for the current MCU in each plane.*/
+  ptrdiff_t           froffset[3];
+  /*The starting row for the current MCU in each plane.*/
+  int                 fragy0[3];
+  /*The ending row for the current MCU in each plane.*/
+  int                 fragy_end[3];
+  /*The starting superblock for the current MCU in each plane.*/
+  unsigned            sbi0[3];
+  /*The ending superblock for the current MCU in each plane.*/
+  unsigned            sbi_end[3];
+  /*The number of tokens for zzi=1 for each color plane.*/
+  int                 ndct_tokens1[3];
+  /*The outstanding eob_run count for zzi=1 for each color plane.*/
+  int                 eob_run1[3];
+  /*Whether or not the loop filter is enabled.*/
+  int                 loop_filter;
+};
+
+
+
+/*Statistics used to estimate R-D cost of a block in a given coding mode.
+  See modedec.h for more details.*/
+struct oc_mode_rd{
+  /*The expected bits used by the DCT tokens, shifted by OC_BIT_SCALE.*/
+  ogg_int16_t rate;
+  /*The expected square root of the sum of squared errors, shifted by
+     OC_RMSE_SCALE.*/
+  ogg_int16_t rmse;
+};
+
+# if defined(OC_COLLECT_METRICS)
+#  include "collect.h"
+# endif
+
+
+
+/*A 2nd order low-pass Bessel follower.
+  We use this for rate control because it has fast reaction time, but is
+   critically damped.*/
+struct oc_iir_filter{
+  ogg_int32_t c[2];
+  ogg_int64_t g;
+  ogg_int32_t x[2];
+  ogg_int32_t y[2];
+};
+
+
+
+/*The 2-pass metrics associated with a single frame.*/
+struct oc_frame_metrics{
+  /*The log base 2 of the scale factor for this frame in Q24 format.*/
+  ogg_int32_t   log_scale;
+  /*The number of application-requested duplicates of this frame.*/
+  unsigned      dup_count:31;
+  /*The frame type from pass 1.*/
+  unsigned      frame_type:1;
+  /*The frame activity average from pass 1.*/
+  unsigned      activity_avg;
+};
+
+
+
+/*Rate control state information.*/
+struct oc_rc_state{
+  /*The target average bits per frame.*/
+  ogg_int64_t        bits_per_frame;
+  /*The current buffer fullness (bits available to be used).*/
+  ogg_int64_t        fullness;
+  /*The target buffer fullness.
+    This is where we'd like to be by the last keyframe the appears in the next
+     buf_delay frames.*/
+  ogg_int64_t        target;
+  /*The maximum buffer fullness (total size of the buffer).*/
+  ogg_int64_t        max;
+  /*The log of the number of pixels in a frame in Q57 format.*/
+  ogg_int64_t        log_npixels;
+  /*The exponent used in the rate model in Q8 format.*/
+  unsigned           exp[2];
+  /*The number of frames to distribute the buffer usage over.*/
+  int                buf_delay;
+  /*The total drop count from the previous frame.
+    This includes duplicates explicitly requested via the
+     TH_ENCCTL_SET_DUP_COUNT API as well as frames we chose to drop ourselves.*/
+  ogg_uint32_t       prev_drop_count;
+  /*The log of an estimated scale factor used to obtain the real framerate, for
+     VFR sources or, e.g., 12 fps content doubled to 24 fps, etc.*/
+  ogg_int64_t        log_drop_scale;
+  /*The log of estimated scale factor for the rate model in Q57 format.*/
+  ogg_int64_t        log_scale[2];
+  /*The log of the target quantizer level in Q57 format.*/
+  ogg_int64_t        log_qtarget;
+  /*Will we drop frames to meet bitrate target?*/
+  unsigned char      drop_frames;
+  /*Do we respect the maximum buffer fullness?*/
+  unsigned char      cap_overflow;
+  /*Can the reservoir go negative?*/
+  unsigned char      cap_underflow;
+  /*Second-order lowpass filters to track scale and VFR.*/
+  oc_iir_filter      scalefilter[2];
+  int                inter_count;
+  int                inter_delay;
+  int                inter_delay_target;
+  oc_iir_filter      vfrfilter;
+  /*Two-pass mode state.
+    0 => 1-pass encoding.
+    1 => 1st pass of 2-pass encoding.
+    2 => 2nd pass of 2-pass encoding.*/
+  int                twopass;
+  /*Buffer for current frame metrics.*/
+  unsigned char      twopass_buffer[48];
+  /*The number of bytes in the frame metrics buffer.
+    When 2-pass encoding is enabled, this is set to 0 after each frame is
+     submitted, and must be non-zero before the next frame will be accepted.*/
+  int                twopass_buffer_bytes;
+  int                twopass_buffer_fill;
+  /*Whether or not to force the next frame to be a keyframe.*/
+  unsigned char      twopass_force_kf;
+  /*The metrics for the previous frame.*/
+  oc_frame_metrics   prev_metrics;
+  /*The metrics for the current frame.*/
+  oc_frame_metrics   cur_metrics;
+  /*The buffered metrics for future frames.*/
+  oc_frame_metrics  *frame_metrics;
+  int                nframe_metrics;
+  int                cframe_metrics;
+  /*The index of the current frame in the circular metric buffer.*/
+  int                frame_metrics_head;
+  /*The frame count of each type (keyframes, delta frames, and dup frames);
+     32 bits limits us to 2.268 years at 60 fps.*/
+  ogg_uint32_t       frames_total[3];
+  /*The number of frames of each type yet to be processed.*/
+  ogg_uint32_t       frames_left[3];
+  /*The sum of the scale values for each frame type.*/
+  ogg_int64_t        scale_sum[2];
+  /*The start of the window over which the current scale sums are taken.*/
+  int                scale_window0;
+  /*The end of the window over which the current scale sums are taken.*/
+  int                scale_window_end;
+  /*The frame count of each type in the current 2-pass window; this does not
+     include dup frames.*/
+  int                nframes[3];
+  /*The total accumulated estimation bias.*/
+  ogg_int64_t        rate_bias;
+};
+
+
+void oc_rc_state_init(oc_rc_state *_rc,oc_enc_ctx *_enc);
+void oc_rc_state_clear(oc_rc_state *_rc);
+
+void oc_enc_rc_resize(oc_enc_ctx *_enc);
+int oc_enc_select_qi(oc_enc_ctx *_enc,int _qti,int _clamp);
+void oc_enc_calc_lambda(oc_enc_ctx *_enc,int _frame_type);
+int oc_enc_update_rc_state(oc_enc_ctx *_enc,
+ long _bits,int _qti,int _qi,int _trial,int _droppable);
+int oc_enc_rc_2pass_out(oc_enc_ctx *_enc,unsigned char **_buf);
+int oc_enc_rc_2pass_in(oc_enc_ctx *_enc,unsigned char *_buf,size_t _bytes);
+
+
+
+/*The internal encoder state.*/
+struct th_enc_ctx{
+  /*Shared encoder/decoder state.*/
+  oc_theora_state          state;
+  /*Buffer in which to assemble packets.*/
+  oggpack_buffer           opb;
+  /*Encoder-specific macroblock information.*/
+  oc_mb_enc_info          *mb_info;
+  /*DC coefficients after prediction.*/
+  ogg_int16_t             *frag_dc;
+  /*The list of coded macro blocks, in coded order.*/
+  unsigned                *coded_mbis;
+  /*The number of coded macro blocks.*/
+  size_t                   ncoded_mbis;
+  /*Whether or not packets are ready to be emitted.
+    This takes on negative values while there are remaining header packets to
+     be emitted, reaches 0 when the codec is ready for input, and becomes
+     positive when a frame has been processed and data packets are ready.*/
+  int                      packet_state;
+  /*The maximum distance between keyframes.*/
+  ogg_uint32_t             keyframe_frequency_force;
+  /*The number of duplicates to produce for the next frame.*/
+  ogg_uint32_t             dup_count;
+  /*The number of duplicates remaining to be emitted for the current frame.*/
+  ogg_uint32_t             nqueued_dups;
+  /*The number of duplicates emitted for the last frame.*/
+  ogg_uint32_t             prev_dup_count;
+  /*The current speed level.*/
+  int                      sp_level;
+  /*Whether or not VP3 compatibility mode has been enabled.*/
+  unsigned char            vp3_compatible;
+  /*Whether or not any INTER frames have been coded.*/
+  unsigned char            coded_inter_frame;
+  /*Whether or not previous frame was dropped.*/
+  unsigned char            prevframe_dropped;
+  /*Stores most recently chosen Huffman tables for each frame type, DC and AC
+     coefficients, and luma and chroma tokens.
+    The actual Huffman table used for a given coefficient depends not only on
+     the choice made here, but also its index in the zig-zag ordering.*/
+  unsigned char            huff_idxs[2][2][2];
+  /*Current count of bits used by each MV coding mode.*/
+  size_t                   mv_bits[2];
+  /*The mode scheme chooser for estimating mode coding costs.*/
+  oc_mode_scheme_chooser   chooser;
+  /*Temporary encoder state for the analysis pipeline.*/
+  oc_enc_pipeline_state    pipe;
+  /*The number of vertical super blocks in an MCU.*/
+  int                      mcu_nvsbs;
+  /*The SSD error for skipping each fragment in the current MCU.*/
+  unsigned                *mcu_skip_ssd;
+  /*The masking scale factors for chroma blocks in the current MCU.*/
+  ogg_uint16_t            *mcu_rd_scale;
+  ogg_uint16_t            *mcu_rd_iscale;
+  /*The DCT token lists for each coefficient and each plane.*/
+  unsigned char          **dct_tokens[3];
+  /*The extra bits associated with each DCT token.*/
+  ogg_uint16_t           **extra_bits[3];
+  /*The number of DCT tokens for each coefficient for each plane.*/
+  ptrdiff_t                ndct_tokens[3][64];
+  /*Pending EOB runs for each coefficient for each plane.*/
+  ogg_uint16_t             eob_run[3][64];
+  /*The offset of the first DCT token for each coefficient for each plane.*/
+  unsigned char            dct_token_offs[3][64];
+  /*The last DC coefficient for each plane and reference frame.*/
+  int                      dc_pred_last[3][4];
+#if defined(OC_COLLECT_METRICS)
+  /*Fragment SAD statistics for MB mode estimation metrics.*/
+  unsigned                *frag_sad;
+  /*Fragment SATD statistics for MB mode estimation metrics.*/
+  unsigned                *frag_satd;
+  /*Fragment SSD statistics for MB mode estimation metrics.*/
+  unsigned                *frag_ssd;
+#endif
+  /*The R-D optimization parameter.*/
+  int                      lambda;
+  /*The average block "activity" of the previous frame.*/
+  unsigned                 activity_avg;
+  /*The average MB luma of the previous frame.*/
+  unsigned                 luma_avg;
+  /*The huffman tables in use.*/
+  th_huff_code             huff_codes[TH_NHUFFMAN_TABLES][TH_NDCT_TOKENS];
+  /*The quantization parameters in use.*/
+  th_quant_info            qinfo;
+  /*The original DC coefficients saved off from the dequatization tables.*/
+  ogg_uint16_t             dequant_dc[64][3][2];
+  /*Condensed dequantization tables.*/
+  const ogg_uint16_t      *dequant[3][3][2];
+  /*Condensed quantization tables.*/
+  void                    *enquant[3][3][2];
+  /*The full set of quantization tables.*/
+  void                    *enquant_tables[64][3][2];
+  /*Storage for the quantization tables.*/
+  unsigned char           *enquant_table_data;
+  /*An "average" quantizer for each frame type (INTRA or INTER) and qi value.
+    This is used to parameterize the rate control decisions.
+    They are kept in the log domain to simplify later processing.
+    These are DCT domain quantizers, and so are scaled by an additional factor
+     of 4 from the pixel domain.*/
+  ogg_int64_t              log_qavg[2][64];
+  /*The "average" quantizer further partitioned by color plane.
+    This is used to parameterize mode decision.
+    These are DCT domain quantizers, and so are scaled by an additional factor
+     of 4 from the pixel domain.*/
+  ogg_int16_t              log_plq[64][3][2];
+  /*The R-D scale factors to apply to chroma blocks for a given frame type
+     (INTRA or INTER) and qi value.
+    The first is the "D" modifier (rd_scale), while the second is the "lambda"
+     modifier (rd_iscale).*/
+  ogg_uint16_t             chroma_rd_scale[2][64][2];
+  /*The interpolated mode decision R-D lookup tables for the current
+     quantizers, color plane, and quantization type.*/
+  oc_mode_rd               mode_rd[3][3][2][OC_COMP_BINS];
+  /*The buffer state used to drive rate control.*/
+  oc_rc_state              rc;
+# if defined(OC_ENC_USE_VTABLE)
+  /*Table for encoder acceleration functions.*/
+  oc_enc_opt_vtable        opt_vtable;
+# endif
+  /*Table for encoder data used by accelerated functions.*/
+  oc_enc_opt_data          opt_data;
+};
+
+
+void oc_enc_analyze_intra(oc_enc_ctx *_enc,int _recode);
+int oc_enc_analyze_inter(oc_enc_ctx *_enc,int _allow_keyframe,int _recode);
+
+
+
+/*Perform fullpel motion search for a single MB against both reference frames.*/
+void oc_mcenc_search(oc_enc_ctx *_enc,int _mbi);
+/*Refine a MB MV for one frame.*/
+void oc_mcenc_refine1mv(oc_enc_ctx *_enc,int _mbi,int _frame);
+/*Refine the block MVs.*/
+void oc_mcenc_refine4mv(oc_enc_ctx *_enc,int _mbi);
+
+
+
+/*Used to rollback a tokenlog transaction when we retroactively decide to skip
+   a fragment.
+  A checkpoint is taken right before each token is added.*/
+struct oc_token_checkpoint{
+  /*The color plane the token was added to.*/
+  unsigned char pli;
+  /*The zig-zag index the token was added to.*/
+  unsigned char zzi;
+  /*The outstanding EOB run count before the token was added.*/
+  ogg_uint16_t  eob_run;
+  /*The token count before the token was added.*/
+  ptrdiff_t     ndct_tokens;
+};
+
+
+
+void oc_enc_tokenize_start(oc_enc_ctx *_enc);
+int oc_enc_tokenize_ac(oc_enc_ctx *_enc,int _pli,ptrdiff_t _fragi,
+ ogg_int16_t *_qdct_out,const ogg_int16_t *_qdct_in,
+ const ogg_uint16_t *_dequant,const ogg_int16_t *_dct,
+ int _zzi,oc_token_checkpoint **_stack,int _lambda,int _acmin);
+int oc_enc_tokenize_ac_fast(oc_enc_ctx *_enc,int _pli,ptrdiff_t _fragi,
+ ogg_int16_t *_qdct_out,const ogg_int16_t *_qdct_in,
+ const ogg_uint16_t *_dequant,const ogg_int16_t *_dct,
+ int _zzi,oc_token_checkpoint **_stack,int _lambda,int _acmin);
+void oc_enc_tokenlog_rollback(oc_enc_ctx *_enc,
+ const oc_token_checkpoint *_stack,int _n);
+void oc_enc_pred_dc_frag_rows(oc_enc_ctx *_enc,
+ int _pli,int _fragy0,int _frag_yend);
+void oc_enc_tokenize_dc_frag_list(oc_enc_ctx *_enc,int _pli,
+ const ptrdiff_t *_coded_fragis,ptrdiff_t _ncoded_fragis,
+ int _prev_ndct_tokens1,int _prev_eob_run1);
+void oc_enc_tokenize_finish(oc_enc_ctx *_enc);
+
+
+
+/*Utility routine to encode one of the header packets.*/
+int oc_state_flushheader(oc_theora_state *_state,int *_packet_state,
+ oggpack_buffer *_opb,const th_quant_info *_qinfo,
+ const th_huff_code _codes[TH_NHUFFMAN_TABLES][TH_NDCT_TOKENS],
+ const char *_vendor,th_comment *_tc,ogg_packet *_op);
+
+
+
+/*Default pure-C implementations of encoder-specific accelerated functions.*/
+void oc_enc_accel_init_c(oc_enc_ctx *_enc);
+
+void oc_enc_frag_sub_c(ogg_int16_t _diff[64],
+ const unsigned char *_src,const unsigned char *_ref,int _ystride);
+void oc_enc_frag_sub_128_c(ogg_int16_t _diff[64],
+ const unsigned char *_src,int _ystride);
+unsigned oc_enc_frag_sad_c(const unsigned char *_src,
+ const unsigned char *_ref,int _ystride);
+unsigned oc_enc_frag_sad_thresh_c(const unsigned char *_src,
+ const unsigned char *_ref,int _ystride,unsigned _thresh);
+unsigned oc_enc_frag_sad2_thresh_c(const unsigned char *_src,
+ const unsigned char *_ref1,const unsigned char *_ref2,int _ystride,
+ unsigned _thresh);
+unsigned oc_enc_frag_intra_sad_c(const unsigned char *_src, int _ystride);
+unsigned oc_enc_frag_satd_c(int *_dc,const unsigned char *_src,
+ const unsigned char *_ref,int _ystride);
+unsigned oc_enc_frag_satd2_c(int *_dc,const unsigned char *_src,
+ const unsigned char *_ref1,const unsigned char *_ref2,int _ystride);
+unsigned oc_enc_frag_intra_satd_c(int *_dc,
+ const unsigned char *_src,int _ystride);
+unsigned oc_enc_frag_ssd_c(const unsigned char *_src,
+ const unsigned char *_ref,int _ystride);
+unsigned oc_enc_frag_border_ssd_c(const unsigned char *_src,
+ const unsigned char *_ref,int _ystride,ogg_int64_t _mask);
+void oc_enc_frag_copy2_c(unsigned char *_dst,
+ const unsigned char *_src1,const unsigned char *_src2,int _ystride);
+void oc_enc_enquant_table_init_c(void *_enquant,
+ const ogg_uint16_t _dequant[64]);
+void oc_enc_enquant_table_fixup_c(void *_enquant[3][3][2],int _nqis);
+int oc_enc_quantize_c(ogg_int16_t _qdct[64],const ogg_int16_t _dct[64],
+ const ogg_uint16_t _dequant[64],const void *_enquant);
+void oc_enc_fdct8x8_c(ogg_int16_t _y[64],const ogg_int16_t _x[64]);
+
+#endif

thirdparty/libtheora/encoder_disabled.c

@@ -0,0 +1,69 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation https://www.xiph.org/                 *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+#include "apiwrapper.h"
+#include "encint.h"
+
+const th_quant_info TH_VP31_QUANT_INFO = {};
+const th_huff_code TH_VP31_HUFF_CODES[TH_NHUFFMAN_TABLES][TH_NDCT_TOKENS];
+
+th_enc_ctx *th_encode_alloc(const th_info *_info){
+  return NULL;
+}
+
+void th_encode_free(th_enc_ctx *_enc){}
+
+
+int th_encode_ctl(th_enc_ctx *_enc,int _req,void *_buf,size_t _buf_sz){
+  return OC_DISABLED;
+}
+
+int th_encode_flushheader(th_enc_ctx *_enc,th_comment *_tc,ogg_packet *_op){
+  return OC_DISABLED;
+}
+
+int th_encode_ycbcr_in(th_enc_ctx *_enc,th_ycbcr_buffer _img){
+  return OC_DISABLED;
+}
+
+int th_encode_packetout(th_enc_ctx *_enc,int _last_p,ogg_packet *_op){
+  return OC_DISABLED;
+}
+
+
+
+int theora_encode_init(theora_state *_te,theora_info *_ci){
+  return OC_DISABLED;
+}
+
+int theora_encode_YUVin(theora_state *_te,yuv_buffer *_yuv){
+  return OC_DISABLED;
+}
+
+int theora_encode_packetout(theora_state *_te,int _last_p,ogg_packet *_op){
+  return OC_DISABLED;
+}
+
+int theora_encode_header(theora_state *_te,ogg_packet *_op){
+  return OC_DISABLED;
+}
+
+int theora_encode_comment(theora_comment *_tc,ogg_packet *_op){
+  return OC_DISABLED;
+}
+
+int theora_encode_tables(theora_state *_te,ogg_packet *_op){
+  return OC_DISABLED;
+}

thirdparty/libtheora/enquant.c

@@ -0,0 +1,369 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation https://www.xiph.org/                 *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+#include <stdlib.h>
+#include <string.h>
+#include "encint.h"
+
+
+
+int oc_quant_params_clone(th_quant_info *_dst,const th_quant_info *_src){
+  int i;
+  memcpy(_dst,_src,sizeof(*_dst));
+  memset(_dst->qi_ranges,0,sizeof(_dst->qi_ranges));
+  for(i=0;i<6;i++){
+    int nranges;
+    int qti;
+    int pli;
+    int qtj;
+    int plj;
+    int pdup;
+    int qdup;
+    qti=i/3;
+    pli=i%3;
+    qtj=(i-1)/3;
+    plj=(i-1)%3;
+    nranges=_src->qi_ranges[qti][pli].nranges;
+    /*Check for those duplicates that can be cleanly handled by
+       oc_quant_params_clear().*/
+    pdup=i>0&&nranges<=_src->qi_ranges[qtj][plj].nranges;
+    qdup=qti>0&&nranges<=_src->qi_ranges[0][pli].nranges;
+    _dst->qi_ranges[qti][pli].nranges=nranges;
+    if(pdup&&_src->qi_ranges[qti][pli].sizes==_src->qi_ranges[qtj][plj].sizes){
+      _dst->qi_ranges[qti][pli].sizes=_dst->qi_ranges[qtj][plj].sizes;
+    }
+    else if(qdup&&_src->qi_ranges[1][pli].sizes==_src->qi_ranges[0][pli].sizes){
+      _dst->qi_ranges[1][pli].sizes=_dst->qi_ranges[0][pli].sizes;
+    }
+    else{
+      int *sizes;
+      sizes=(int *)_ogg_malloc(nranges*sizeof(*sizes));
+      /*Note: The caller is responsible for cleaning up any partially
+         constructed qinfo.*/
+      if(sizes==NULL)return TH_EFAULT;
+      memcpy(sizes,_src->qi_ranges[qti][pli].sizes,nranges*sizeof(*sizes));
+      _dst->qi_ranges[qti][pli].sizes=sizes;
+    }
+    if(pdup&&_src->qi_ranges[qti][pli].base_matrices==
+     _src->qi_ranges[qtj][plj].base_matrices){
+      _dst->qi_ranges[qti][pli].base_matrices=
+       _dst->qi_ranges[qtj][plj].base_matrices;
+    }
+    else if(qdup&&_src->qi_ranges[1][pli].base_matrices==
+     _src->qi_ranges[0][pli].base_matrices){
+      _dst->qi_ranges[1][pli].base_matrices=
+       _dst->qi_ranges[0][pli].base_matrices;
+    }
+    else{
+      th_quant_base *base_matrices;
+      base_matrices=(th_quant_base *)_ogg_malloc(
+       (nranges+1)*sizeof(*base_matrices));
+      /*Note: The caller is responsible for cleaning up any partially
+         constructed qinfo.*/
+      if(base_matrices==NULL)return TH_EFAULT;
+      memcpy(base_matrices,_src->qi_ranges[qti][pli].base_matrices,
+       (nranges+1)*sizeof(*base_matrices));
+      _dst->qi_ranges[qti][pli].base_matrices=
+       (const th_quant_base *)base_matrices;
+    }
+  }
+  return 0;
+}
+
+void oc_quant_params_pack(oggpack_buffer *_opb,const th_quant_info *_qinfo){
+  const th_quant_ranges *qranges;
+  const th_quant_base   *base_mats[2*3*64];
+  int                    indices[2][3][64];
+  int                    nbase_mats;
+  int                    nbits;
+  int                    ci;
+  int                    qi;
+  int                    qri;
+  int                    qti;
+  int                    pli;
+  int                    qtj;
+  int                    plj;
+  int                    bmi;
+  int                    i;
+  i=_qinfo->loop_filter_limits[0];
+  for(qi=1;qi<64;qi++)i=OC_MAXI(i,_qinfo->loop_filter_limits[qi]);
+  nbits=OC_ILOG_32(i);
+  oggpackB_write(_opb,nbits,3);
+  for(qi=0;qi<64;qi++){
+    oggpackB_write(_opb,_qinfo->loop_filter_limits[qi],nbits);
+  }
+  /*580 bits for VP3.*/
+  i=1;
+  for(qi=0;qi<64;qi++)i=OC_MAXI(_qinfo->ac_scale[qi],i);
+  nbits=OC_ILOGNZ_32(i);
+  oggpackB_write(_opb,nbits-1,4);
+  for(qi=0;qi<64;qi++)oggpackB_write(_opb,_qinfo->ac_scale[qi],nbits);
+  /*516 bits for VP3.*/
+  i=1;
+  for(qi=0;qi<64;qi++)i=OC_MAXI(_qinfo->dc_scale[qi],i);
+  nbits=OC_ILOGNZ_32(i);
+  oggpackB_write(_opb,nbits-1,4);
+  for(qi=0;qi<64;qi++)oggpackB_write(_opb,_qinfo->dc_scale[qi],nbits);
+  /*Consolidate any duplicate base matrices.*/
+  nbase_mats=0;
+  for(qti=0;qti<2;qti++)for(pli=0;pli<3;pli++){
+    qranges=_qinfo->qi_ranges[qti]+pli;
+    for(qri=0;qri<=qranges->nranges;qri++){
+      for(bmi=0;;bmi++){
+        if(bmi>=nbase_mats){
+          base_mats[bmi]=qranges->base_matrices+qri;
+          indices[qti][pli][qri]=nbase_mats++;
+          break;
+        }
+        else if(memcmp(base_mats[bmi][0],qranges->base_matrices[qri],
+         sizeof(base_mats[bmi][0]))==0){
+          indices[qti][pli][qri]=bmi;
+          break;
+        }
+      }
+    }
+  }
+  /*Write out the list of unique base matrices.
+    1545 bits for VP3 matrices.*/
+  oggpackB_write(_opb,nbase_mats-1,9);
+  for(bmi=0;bmi<nbase_mats;bmi++){
+    for(ci=0;ci<64;ci++)oggpackB_write(_opb,base_mats[bmi][0][ci],8);
+  }
+  /*Now store quant ranges and their associated indices into the base matrix
+     list.
+    46 bits for VP3 matrices.*/
+  nbits=OC_ILOG_32(nbase_mats-1);
+  for(i=0;i<6;i++){
+    qti=i/3;
+    pli=i%3;
+    qranges=_qinfo->qi_ranges[qti]+pli;
+    if(i>0){
+      if(qti>0){
+        if(qranges->nranges==_qinfo->qi_ranges[qti-1][pli].nranges&&
+         memcmp(qranges->sizes,_qinfo->qi_ranges[qti-1][pli].sizes,
+         qranges->nranges*sizeof(qranges->sizes[0]))==0&&
+         memcmp(indices[qti][pli],indices[qti-1][pli],
+         (qranges->nranges+1)*sizeof(indices[qti][pli][0]))==0){
+          oggpackB_write(_opb,1,2);
+          continue;
+        }
+      }
+      qtj=(i-1)/3;
+      plj=(i-1)%3;
+      if(qranges->nranges==_qinfo->qi_ranges[qtj][plj].nranges&&
+       memcmp(qranges->sizes,_qinfo->qi_ranges[qtj][plj].sizes,
+       qranges->nranges*sizeof(qranges->sizes[0]))==0&&
+       memcmp(indices[qti][pli],indices[qtj][plj],
+       (qranges->nranges+1)*sizeof(indices[qti][pli][0]))==0){
+        oggpackB_write(_opb,0,1+(qti>0));
+        continue;
+      }
+      oggpackB_write(_opb,1,1);
+    }
+    oggpackB_write(_opb,indices[qti][pli][0],nbits);
+    for(qi=qri=0;qi<63;qri++){
+      oggpackB_write(_opb,qranges->sizes[qri]-1,OC_ILOG_32(62-qi));
+      qi+=qranges->sizes[qri];
+      oggpackB_write(_opb,indices[qti][pli][qri+1],nbits);
+    }
+  }
+}
+
+void oc_iquant_init(oc_iquant *_this,ogg_uint16_t _d){
+  ogg_uint32_t t;
+  int          l;
+  _d<<=1;
+  l=OC_ILOGNZ_32(_d)-1;
+  t=1+((ogg_uint32_t)1<<16+l)/_d;
+  _this->m=(ogg_int16_t)(t-0x10000);
+  _this->l=l;
+}
+
+void oc_enc_enquant_table_init_c(void *_enquant,
+ const ogg_uint16_t _dequant[64]){
+  oc_iquant *enquant;
+  int        zzi;
+  /*In the original VP3.2 code, the rounding offset and the size of the
+     dead zone around 0 were controlled by a "sharpness" parameter.
+    We now R-D optimize the tokens for each block after quantization,
+     so the rounding offset should always be 1/2, and an explicit dead
+     zone is unnecessary.
+    Hence, all of that VP3.2 code is gone from here, and the remaining
+     floating point code has been implemented as equivalent integer
+     code with exact precision.*/
+  enquant=(oc_iquant *)_enquant;
+  for(zzi=0;zzi<64;zzi++)oc_iquant_init(enquant+zzi,_dequant[zzi]);
+}
+
+void oc_enc_enquant_table_fixup_c(void *_enquant[3][3][2],int _nqis){
+  int pli;
+  int qii;
+  int qti;
+  for(pli=0;pli<3;pli++)for(qii=1;qii<_nqis;qii++)for(qti=0;qti<2;qti++){
+    *((oc_iquant *)_enquant[pli][qii][qti])=
+     *((oc_iquant *)_enquant[pli][0][qti]);
+  }
+}
+
+int oc_enc_quantize_c(ogg_int16_t _qdct[64],const ogg_int16_t _dct[64],
+ const ogg_uint16_t _dequant[64],const void *_enquant){
+  const oc_iquant *enquant;
+  int              nonzero;
+  int              zzi;
+  int              val;
+  int              d;
+  int              s;
+  enquant=(const oc_iquant *)_enquant;
+  nonzero=0;
+  for(zzi=0;zzi<64;zzi++){
+    val=_dct[zzi];
+    d=_dequant[zzi];
+    val=val<<1;
+    if(abs(val)>=d){
+      s=OC_SIGNMASK(val);
+      /*The bias added here rounds ties away from zero, since token
+         optimization can only decrease the magnitude of the quantized
+         value.*/
+      val+=d+s^s;
+      /*Note the arithmetic right shift is not guaranteed by ANSI C.
+        Hopefully no one still uses ones-complement architectures.*/
+      val=((enquant[zzi].m*(ogg_int32_t)val>>16)+val>>enquant[zzi].l)-s;
+      _qdct[zzi]=(ogg_int16_t)val;
+      nonzero=zzi;
+    }
+    else _qdct[zzi]=0;
+  }
+  return nonzero;
+}
+
+
+
+/*This table gives the square root of the fraction of the squared magnitude of
+   each DCT coefficient relative to the total, scaled by 2**16, for both INTRA
+   and INTER modes.
+  These values were measured after motion-compensated prediction, before
+   quantization, over a large set of test video (from QCIF to 1080p) encoded at
+   all possible rates.
+  The DC coefficient takes into account the DPCM prediction (using the
+   quantized values from neighboring blocks, as the encoder does, but still
+   before quantization of the coefficient in the current block).
+  The results differ significantly from the expected variance (e.g., using an
+   AR(1) model of the signal with rho=0.95, as is frequently done to compute
+   the coding gain of the DCT).
+  We use them to estimate an "average" quantizer for a given quantizer matrix,
+   as this is used to parameterize a number of the rate control decisions.
+  These values are themselves probably quantizer-matrix dependent, since the
+   shape of the matrix affects the noise distribution in the reference frames,
+   but they should at least give us _some_ amount of adaptivity to different
+   matrices, as opposed to hard-coding a table of average Q values for the
+   current set.
+  The main features they capture are that a) only a few of the quantizers in
+   the upper-left corner contribute anything significant at all (though INTER
+   mode is significantly flatter) and b) the DPCM prediction of the DC
+   coefficient gives a very minor improvement in the INTRA case and a quite
+   significant one in the INTER case (over the expected variance).*/
+static const ogg_uint16_t OC_RPSD[2][64]={
+  {
+    52725,17370,10399, 6867, 5115, 3798, 2942, 2076,
+    17370, 9900, 6948, 4994, 3836, 2869, 2229, 1619,
+    10399, 6948, 5516, 4202, 3376, 2573, 2015, 1461,
+     6867, 4994, 4202, 3377, 2800, 2164, 1718, 1243,
+     5115, 3836, 3376, 2800, 2391, 1884, 1530, 1091,
+     3798, 2869, 2573, 2164, 1884, 1495, 1212,  873,
+     2942, 2229, 2015, 1718, 1530, 1212, 1001,  704,
+     2076, 1619, 1461, 1243, 1091,  873,  704,  474
+  },
+  {
+    23411,15604,13529,11601,10683, 8958, 7840, 6142,
+    15604,11901,10718, 9108, 8290, 6961, 6023, 4487,
+    13529,10718, 9961, 8527, 7945, 6689, 5742, 4333,
+    11601, 9108, 8527, 7414, 7084, 5923, 5175, 3743,
+    10683, 8290, 7945, 7084, 6771, 5754, 4793, 3504,
+     8958, 6961, 6689, 5923, 5754, 4679, 3936, 2989,
+     7840, 6023, 5742, 5175, 4793, 3936, 3522, 2558,
+     6142, 4487, 4333, 3743, 3504, 2989, 2558, 1829
+  }
+};
+
+/*The fraction of the squared magnitude of the residuals in each color channel
+   relative to the total, scaled by 2**16, for each pixel format.
+  These values were measured after motion-compensated prediction, before
+   quantization, over a large set of test video encoded at all possible rates.
+  TODO: These values are only from INTER frames; they should be re-measured for
+   INTRA frames.*/
+static const ogg_uint16_t OC_PCD[4][3]={
+  {59926, 3038, 2572},
+  {55201, 5597, 4738},
+  {55201, 5597, 4738},
+  {47682, 9669, 8185}
+};
+
+
+/*Compute "average" quantizers for each qi level to use for rate control.
+  We do one for each color channel, as well as an average across color
+   channels, separately for INTER and INTRA, since their behavior is very
+   different.
+  The basic approach is to compute a harmonic average of the squared quantizer,
+   weighted by the expected squared magnitude of the DCT coefficients.
+  Under the (not quite true) assumption that DCT coefficients are
+   Laplacian-distributed, this preserves the product Q*lambda, where
+   lambda=sqrt(2/sigma**2) is the Laplacian distribution parameter (not to be
+   confused with the lambda used in R-D optimization throughout most of the
+   rest of the code), when the distributions from multiple coefficients are
+   pooled.
+  The value Q*lambda completely determines the entropy of coefficients drawn
+   from a Laplacian distribution, and thus the expected bitrate.*/
+void oc_enquant_qavg_init(ogg_int64_t _log_qavg[2][64],
+ ogg_int16_t _log_plq[64][3][2],ogg_uint16_t _chroma_rd_scale[2][64][2],
+ ogg_uint16_t *_dequant[64][3][2],int _pixel_fmt){
+  int qi;
+  int pli;
+  int qti;
+  int ci;
+  for(qti=0;qti<2;qti++)for(qi=0;qi<64;qi++){
+    ogg_int64_t  q2;
+    ogg_uint32_t qp[3];
+    ogg_uint32_t cqp;
+    ogg_uint32_t d;
+    q2=0;
+    for(pli=0;pli<3;pli++){
+      qp[pli]=0;
+      for(ci=0;ci<64;ci++){
+        unsigned rq;
+        unsigned qd;
+        qd=_dequant[qi][pli][qti][OC_IZIG_ZAG[ci]];
+        rq=(OC_RPSD[qti][ci]+(qd>>1))/qd;
+        qp[pli]+=rq*(ogg_uint32_t)rq;
+      }
+      q2+=OC_PCD[_pixel_fmt][pli]*(ogg_int64_t)qp[pli];
+      /*plq=1.0/sqrt(qp)*/
+      _log_plq[qi][pli][qti]=
+       (ogg_int16_t)(OC_Q10(32)-oc_blog32_q10(qp[pli])>>1);
+    }
+    d=OC_PCD[_pixel_fmt][1]+OC_PCD[_pixel_fmt][2];
+    cqp=(ogg_uint32_t)((OC_PCD[_pixel_fmt][1]*(ogg_int64_t)qp[1]+
+     OC_PCD[_pixel_fmt][2]*(ogg_int64_t)qp[2]+(d>>1))/d);
+    /*chroma_rd_scale=clamp(0.25,cqp/qp[0],4)*/
+    d=OC_MAXI(qp[0]+(1<<OC_RD_SCALE_BITS-1)>>OC_RD_SCALE_BITS,1);
+    d=OC_CLAMPI(1<<OC_RD_SCALE_BITS-2,(cqp+(d>>1))/d,4<<OC_RD_SCALE_BITS);
+    _chroma_rd_scale[qti][qi][0]=(ogg_int16_t)d;
+    /*chroma_rd_iscale=clamp(0.25,qp[0]/cqp,4)*/
+    d=OC_MAXI(OC_RD_ISCALE(cqp,1),1);
+    d=OC_CLAMPI(1<<OC_RD_ISCALE_BITS-2,(qp[0]+(d>>1))/d,4<<OC_RD_ISCALE_BITS);
+    _chroma_rd_scale[qti][qi][1]=(ogg_int16_t)d;
+    /*qavg=1.0/sqrt(q2).*/
+    _log_qavg[qti][qi]=OC_Q57(48)-oc_blog64(q2)>>1;
+  }
+}

thirdparty/libtheora/enquant.h

@@ -0,0 +1,26 @@
+#if !defined(_enquant_H)
+# define _enquant_H (1)
+# include "quant.h"
+
+typedef struct oc_iquant oc_iquant;
+
+#define OC_QUANT_MAX_LOG (OC_Q57(OC_STATIC_ILOG_32(OC_QUANT_MAX)-1))
+
+/*Used to compute x/d via ((x*m>>16)+x>>l)+(x<0))
+   (i.e., one 16x16->16 mul, 2 shifts, and 2 adds).
+  This is not an approximation; for 16-bit x and d, it is exact.*/
+struct oc_iquant{
+  ogg_int16_t m;
+  ogg_int16_t l;
+};
+
+
+
+int oc_quant_params_clone(th_quant_info *_dst,const th_quant_info *_src);
+void oc_quant_params_pack(oggpack_buffer *_opb,const th_quant_info *_qinfo);
+void oc_iquant_init(oc_iquant *_this,ogg_uint16_t _d);
+void oc_enquant_qavg_init(ogg_int64_t _log_qavg[2][64],
+ ogg_int16_t _log_plq[64][3][2],ogg_uint16_t _pl_rd_scale[2][64][2],
+ ogg_uint16_t *_dequant[64][3][2],int _pixel_fmt);
+
+#endif

thirdparty/libtheora/fdct.c

@@ -0,0 +1,416 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation https://www.xiph.org/                 *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+#include "encint.h"
+#include "dct.h"
+
+
+
+/*Performs a forward 8 point Type-II DCT transform.
+  The output is scaled by a factor of 2 from the orthonormal version of the
+   transform.
+  _y: The buffer to store the result in.
+      Data will be placed the first 8 entries (e.g., in a row of an 8x8 block).
+  _x: The input coefficients.
+      Every 8th entry is used (e.g., from a column of an 8x8 block).*/
+static void oc_fdct8(ogg_int16_t _y[8],const ogg_int16_t *_x){
+  int t0;
+  int t1;
+  int t2;
+  int t3;
+  int t4;
+  int t5;
+  int t6;
+  int t7;
+  int r;
+  int s;
+  int u;
+  int v;
+  /*Stage 1:*/
+  /*0-7 butterfly.*/
+  t0=_x[0<<3]+(int)_x[7<<3];
+  t7=_x[0<<3]-(int)_x[7<<3];
+  /*1-6 butterfly.*/
+  t1=_x[1<<3]+(int)_x[6<<3];
+  t6=_x[1<<3]-(int)_x[6<<3];
+  /*2-5 butterfly.*/
+  t2=_x[2<<3]+(int)_x[5<<3];
+  t5=_x[2<<3]-(int)_x[5<<3];
+  /*3-4 butterfly.*/
+  t3=_x[3<<3]+(int)_x[4<<3];
+  t4=_x[3<<3]-(int)_x[4<<3];
+  /*Stage 2:*/
+  /*0-3 butterfly.*/
+  r=t0+t3;
+  t3=t0-t3;
+  t0=r;
+  /*1-2 butterfly.*/
+  r=t1+t2;
+  t2=t1-t2;
+  t1=r;
+  /*6-5 butterfly.*/
+  r=t6+t5;
+  t5=t6-t5;
+  t6=r;
+  /*Stages 3 and 4 are where all the approximation occurs.
+    These are chosen to be as close to an exact inverse of the approximations
+     made in the iDCT as possible, while still using mostly 16-bit arithmetic.
+    We use some 16x16->32 signed MACs, but those still commonly execute in 1
+     cycle on a 16-bit DSP.
+    For example, s=(27146*t5+0x4000>>16)+t5+(t5!=0) is an exact inverse of
+     t5=(OC_C4S4*s>>16).
+    That is, applying the latter to the output of the former will recover t5
+     exactly (over the valid input range of t5, -23171...23169).
+    We increase the rounding bias to 0xB500 in this particular case so that
+     errors inverting the subsequent butterfly are not one-sided (e.g., the
+     mean error is very close to zero).
+    The (t5!=0) term could be replaced simply by 1, but we want to send 0 to 0.
+    The fDCT of an all-zeros block will still not be zero, because of the
+     biases we added at the very beginning of the process, but it will be close
+     enough that it is guaranteed to round to zero.*/
+  /*Stage 3:*/
+  /*4-5 butterfly.*/
+  s=(27146*t5+0xB500>>16)+t5+(t5!=0)>>1;
+  r=t4+s;
+  t5=t4-s;
+  t4=r;
+  /*7-6 butterfly.*/
+  s=(27146*t6+0xB500>>16)+t6+(t6!=0)>>1;
+  r=t7+s;
+  t6=t7-s;
+  t7=r;
+  /*Stage 4:*/
+  /*0-1 butterfly.*/
+  r=(27146*t0+0x4000>>16)+t0+(t0!=0);
+  s=(27146*t1+0xB500>>16)+t1+(t1!=0);
+  u=r+s>>1;
+  v=r-u;
+  _y[0]=u;
+  _y[4]=v;
+  /*3-2 rotation by 6pi/16*/
+  u=(OC_C6S2*t2+OC_C2S6*t3+0x6CB7>>16)+(t3!=0);
+  s=(OC_C6S2*u>>16)-t2;
+  v=(s*21600+0x2800>>18)+s+(s!=0);
+  _y[2]=u;
+  _y[6]=v;
+  /*6-5 rotation by 3pi/16*/
+  u=(OC_C5S3*t6+OC_C3S5*t5+0x0E3D>>16)+(t5!=0);
+  s=t6-(OC_C5S3*u>>16);
+  v=(s*26568+0x3400>>17)+s+(s!=0);
+  _y[5]=u;
+  _y[3]=v;
+  /*7-4 rotation by 7pi/16*/
+  u=(OC_C7S1*t4+OC_C1S7*t7+0x7B1B>>16)+(t7!=0);
+  s=(OC_C7S1*u>>16)-t4;
+  v=(s*20539+0x3000>>20)+s+(s!=0);
+  _y[1]=u;
+  _y[7]=v;
+}
+
+/*Performs a forward 8x8 Type-II DCT transform.
+  The output is scaled by a factor of 4 relative to the orthonormal version
+   of the transform.
+  _y: The buffer to store the result in.
+      This may be the same as _x.
+  _x: The input coefficients. */
+void oc_enc_fdct8x8_c(ogg_int16_t _y[64],const ogg_int16_t _x[64]){
+  const ogg_int16_t *in;
+  ogg_int16_t       *end;
+  ogg_int16_t       *out;
+  ogg_int16_t        w[64];
+  int                i;
+  /*Add two extra bits of working precision to improve accuracy; any more and
+     we could overflow.*/
+  for(i=0;i<64;i++)w[i]=_x[i]<<2;
+  /*These biases correct for some systematic error that remains in the full
+     fDCT->iDCT round trip.*/
+  w[0]+=(w[0]!=0)+1;
+  w[1]++;
+  w[8]--;
+  /*Transform columns of w into rows of _y.*/
+  for(in=w,out=_y,end=out+64;out<end;in++,out+=8)oc_fdct8(out,in);
+  /*Transform columns of _y into rows of w.*/
+  for(in=_y,out=w,end=out+64;out<end;in++,out+=8)oc_fdct8(out,in);
+  /*Round the result back to the external working precision (which is still
+     scaled by four relative to the orthogonal result).
+    TODO: We should just update the external working precision.*/
+  for(i=0;i<64;i++)_y[i]=w[OC_FZIG_ZAG[i]]+2>>2;
+}
+
+
+
+/*This does not seem to outperform simple LFE border padding before MC.
+  It yields higher PSNR, but much higher bitrate usage.*/
+#if 0
+typedef struct oc_extension_info oc_extension_info;
+
+
+
+/*Information needed to pad boundary blocks.
+  We multiply each row/column by an extension matrix that fills in the padding
+   values as a linear combination of the active values, so that an equivalent
+   number of coefficients are forced to zero.
+  This costs at most 16 multiplies, the same as a 1-D fDCT itself, and as
+   little as 7 multiplies.
+  We compute the extension matrices for every possible shape in advance, as
+   there are only 35.
+  The coefficients for all matrices are stored in a single array to take
+   advantage of the overlap and repetitiveness of many of the shapes.
+  A similar technique is applied to the offsets into this array.
+  This reduces the required table storage by about 48%.
+  See tools/extgen.c for details.
+  We could conceivably do the same for all 256 possible shapes.*/
+struct oc_extension_info{
+  /*The mask of the active pixels in the shape.*/
+  short                     mask;
+  /*The number of active pixels in the shape.*/
+  short                     na;
+  /*The extension matrix.
+    This is (8-na)xna*/
+  const ogg_int16_t *const *ext;
+  /*The pixel indices: na active pixels followed by 8-na padding pixels.*/
+  unsigned char             pi[8];
+  /*The coefficient indices: na unconstrained coefficients followed by 8-na
+     coefficients to be forced to zero.*/
+  unsigned char             ci[8];
+};
+
+
+/*The number of shapes we need.*/
+#define OC_NSHAPES   (35)
+
+static const ogg_int16_t OC_EXT_COEFFS[229]={
+  0x7FFF,0xE1F8,0x6903,0xAA79,0x5587,0x7FFF,0x1E08,0x7FFF,
+  0x5587,0xAA79,0x6903,0xE1F8,0x7FFF,0x0000,0x0000,0x0000,
+  0x7FFF,0x0000,0x0000,0x7FFF,0x8000,0x7FFF,0x0000,0x0000,
+  0x7FFF,0xE1F8,0x1E08,0xB0A7,0xAA1D,0x337C,0x7FFF,0x4345,
+  0x2267,0x4345,0x7FFF,0x337C,0xAA1D,0xB0A7,0x8A8C,0x4F59,
+  0x03B4,0xE2D6,0x7FFF,0x2CF3,0x7FFF,0xE2D6,0x03B4,0x4F59,
+  0x8A8C,0x1103,0x7AEF,0x5225,0xDF60,0xC288,0xDF60,0x5225,
+  0x7AEF,0x1103,0x668A,0xD6EE,0x3A16,0x0E6C,0xFA07,0x0E6C,
+  0x3A16,0xD6EE,0x668A,0x2A79,0x2402,0x980F,0x50F5,0x4882,
+  0x50F5,0x980F,0x2402,0x2A79,0xF976,0x2768,0x5F22,0x2768,
+  0xF976,0x1F91,0x76C1,0xE9AE,0x76C1,0x1F91,0x7FFF,0xD185,
+  0x0FC8,0xD185,0x7FFF,0x4F59,0x4345,0xED62,0x4345,0x4F59,
+  0xF574,0x5D99,0x2CF3,0x5D99,0xF574,0x5587,0x3505,0x30FC,
+  0xF482,0x953C,0xEAC4,0x7FFF,0x4F04,0x7FFF,0xEAC4,0x953C,
+  0xF482,0x30FC,0x4F04,0x273D,0xD8C3,0x273D,0x1E09,0x61F7,
+  0x1E09,0x273D,0xD8C3,0x273D,0x4F04,0x30FC,0xA57E,0x153C,
+  0x6AC4,0x3C7A,0x1E08,0x3C7A,0x6AC4,0x153C,0xA57E,0x7FFF,
+  0xA57E,0x5A82,0x6AC4,0x153C,0xC386,0xE1F8,0xC386,0x153C,
+  0x6AC4,0x5A82,0xD8C3,0x273D,0x7FFF,0xE1F7,0x7FFF,0x273D,
+  0xD8C3,0x4F04,0x30FC,0xD8C3,0x273D,0xD8C3,0x30FC,0x4F04,
+  0x1FC8,0x67AD,0x1853,0xE038,0x1853,0x67AD,0x1FC8,0x4546,
+  0xE038,0x1FC8,0x3ABA,0x1FC8,0xE038,0x4546,0x3505,0x5587,
+  0xF574,0xBC11,0x78F4,0x4AFB,0xE6F3,0x4E12,0x3C11,0xF8F4,
+  0x4AFB,0x3C7A,0xF88B,0x3C11,0x78F4,0xCAFB,0x7FFF,0x08CC,
+  0x070C,0x236D,0x5587,0x236D,0x070C,0xF88B,0x3C7A,0x4AFB,
+  0xF8F4,0x3C11,0x7FFF,0x153C,0xCAFB,0x153C,0x7FFF,0x1E08,
+  0xE1F8,0x7FFF,0x08CC,0x7FFF,0xCAFB,0x78F4,0x3C11,0x4E12,
+  0xE6F3,0x4AFB,0x78F4,0xBC11,0xFE3D,0x7FFF,0xFE3D,0x2F3A,
+  0x7FFF,0x2F3A,0x89BC,0x7FFF,0x89BC
+};
+
+static const ogg_int16_t *const OC_EXT_ROWS[96]={
+  OC_EXT_COEFFS+   0,OC_EXT_COEFFS+   0,OC_EXT_COEFFS+   0,OC_EXT_COEFFS+   0,
+  OC_EXT_COEFFS+   0,OC_EXT_COEFFS+   0,OC_EXT_COEFFS+   0,OC_EXT_COEFFS+   6,
+  OC_EXT_COEFFS+  27,OC_EXT_COEFFS+  38,OC_EXT_COEFFS+  43,OC_EXT_COEFFS+  32,
+  OC_EXT_COEFFS+  49,OC_EXT_COEFFS+  58,OC_EXT_COEFFS+  67,OC_EXT_COEFFS+  71,
+  OC_EXT_COEFFS+  62,OC_EXT_COEFFS+  53,OC_EXT_COEFFS+  12,OC_EXT_COEFFS+  15,
+  OC_EXT_COEFFS+  14,OC_EXT_COEFFS+  13,OC_EXT_COEFFS+  76,OC_EXT_COEFFS+  81,
+  OC_EXT_COEFFS+  86,OC_EXT_COEFFS+  91,OC_EXT_COEFFS+  96,OC_EXT_COEFFS+  98,
+  OC_EXT_COEFFS+  93,OC_EXT_COEFFS+  88,OC_EXT_COEFFS+  83,OC_EXT_COEFFS+  78,
+  OC_EXT_COEFFS+  12,OC_EXT_COEFFS+  15,OC_EXT_COEFFS+  15,OC_EXT_COEFFS+  12,
+  OC_EXT_COEFFS+  12,OC_EXT_COEFFS+  15,OC_EXT_COEFFS+  12,OC_EXT_COEFFS+  15,
+  OC_EXT_COEFFS+  15,OC_EXT_COEFFS+  12,OC_EXT_COEFFS+ 103,OC_EXT_COEFFS+ 108,
+  OC_EXT_COEFFS+ 126,OC_EXT_COEFFS+  16,OC_EXT_COEFFS+ 137,OC_EXT_COEFFS+ 141,
+  OC_EXT_COEFFS+  20,OC_EXT_COEFFS+ 130,OC_EXT_COEFFS+ 113,OC_EXT_COEFFS+ 116,
+  OC_EXT_COEFFS+ 146,OC_EXT_COEFFS+ 153,OC_EXT_COEFFS+ 160,OC_EXT_COEFFS+ 167,
+  OC_EXT_COEFFS+ 170,OC_EXT_COEFFS+ 163,OC_EXT_COEFFS+ 156,OC_EXT_COEFFS+ 149,
+  OC_EXT_COEFFS+ 119,OC_EXT_COEFFS+ 122,OC_EXT_COEFFS+ 174,OC_EXT_COEFFS+ 177,
+  OC_EXT_COEFFS+ 182,OC_EXT_COEFFS+ 187,OC_EXT_COEFFS+ 192,OC_EXT_COEFFS+ 197,
+  OC_EXT_COEFFS+ 202,OC_EXT_COEFFS+ 207,OC_EXT_COEFFS+ 210,OC_EXT_COEFFS+ 215,
+  OC_EXT_COEFFS+ 179,OC_EXT_COEFFS+ 189,OC_EXT_COEFFS+  24,OC_EXT_COEFFS+ 204,
+  OC_EXT_COEFFS+ 184,OC_EXT_COEFFS+ 194,OC_EXT_COEFFS+ 212,OC_EXT_COEFFS+ 199,
+  OC_EXT_COEFFS+ 217,OC_EXT_COEFFS+ 100,OC_EXT_COEFFS+ 134,OC_EXT_COEFFS+ 135,
+  OC_EXT_COEFFS+ 135,OC_EXT_COEFFS+  12,OC_EXT_COEFFS+  15,OC_EXT_COEFFS+ 134,
+  OC_EXT_COEFFS+ 134,OC_EXT_COEFFS+ 135,OC_EXT_COEFFS+ 220,OC_EXT_COEFFS+ 223,
+  OC_EXT_COEFFS+ 226,OC_EXT_COEFFS+ 227,OC_EXT_COEFFS+ 224,OC_EXT_COEFFS+ 221
+};
+
+static const oc_extension_info OC_EXTENSION_INFO[OC_NSHAPES]={
+  {0x7F,7,OC_EXT_ROWS+  0,{0,1,2,3,4,5,6,7},{0,1,2,4,5,6,7,3}},
+  {0xFE,7,OC_EXT_ROWS+  7,{1,2,3,4,5,6,7,0},{0,1,2,4,5,6,7,3}},
+  {0x3F,6,OC_EXT_ROWS+  8,{0,1,2,3,4,5,7,6},{0,1,3,4,6,7,5,2}},
+  {0xFC,6,OC_EXT_ROWS+ 10,{2,3,4,5,6,7,1,0},{0,1,3,4,6,7,5,2}},
+  {0x1F,5,OC_EXT_ROWS+ 12,{0,1,2,3,4,7,6,5},{0,2,3,5,7,6,4,1}},
+  {0xF8,5,OC_EXT_ROWS+ 15,{3,4,5,6,7,2,1,0},{0,2,3,5,7,6,4,1}},
+  {0x0F,4,OC_EXT_ROWS+ 18,{0,1,2,3,7,6,5,4},{0,2,4,6,7,5,3,1}},
+  {0xF0,4,OC_EXT_ROWS+ 18,{4,5,6,7,3,2,1,0},{0,2,4,6,7,5,3,1}},
+  {0x07,3,OC_EXT_ROWS+ 22,{0,1,2,7,6,5,4,3},{0,3,6,7,5,4,2,1}},
+  {0xE0,3,OC_EXT_ROWS+ 27,{5,6,7,4,3,2,1,0},{0,3,6,7,5,4,2,1}},
+  {0x03,2,OC_EXT_ROWS+ 32,{0,1,7,6,5,4,3,2},{0,4,7,6,5,3,2,1}},
+  {0xC0,2,OC_EXT_ROWS+ 32,{6,7,5,4,3,2,1,0},{0,4,7,6,5,3,2,1}},
+  {0x01,1,OC_EXT_ROWS+  0,{0,7,6,5,4,3,2,1},{0,7,6,5,4,3,2,1}},
+  {0x80,1,OC_EXT_ROWS+  0,{7,6,5,4,3,2,1,0},{0,7,6,5,4,3,2,1}},
+  {0x7E,6,OC_EXT_ROWS+ 42,{1,2,3,4,5,6,7,0},{0,1,2,5,6,7,4,3}},
+  {0x7C,5,OC_EXT_ROWS+ 44,{2,3,4,5,6,7,1,0},{0,1,4,5,7,6,3,2}},
+  {0x3E,5,OC_EXT_ROWS+ 47,{1,2,3,4,5,7,6,0},{0,1,4,5,7,6,3,2}},
+  {0x78,4,OC_EXT_ROWS+ 50,{3,4,5,6,7,2,1,0},{0,4,5,7,6,3,2,1}},
+  {0x3C,4,OC_EXT_ROWS+ 54,{2,3,4,5,7,6,1,0},{0,3,4,7,6,5,2,1}},
+  {0x1E,4,OC_EXT_ROWS+ 58,{1,2,3,4,7,6,5,0},{0,4,5,7,6,3,2,1}},
+  {0x70,3,OC_EXT_ROWS+ 62,{4,5,6,7,3,2,1,0},{0,5,7,6,4,3,2,1}},
+  {0x38,3,OC_EXT_ROWS+ 67,{3,4,5,7,6,2,1,0},{0,5,6,7,4,3,2,1}},
+  {0x1C,3,OC_EXT_ROWS+ 72,{2,3,4,7,6,5,1,0},{0,5,6,7,4,3,2,1}},
+  {0x0E,3,OC_EXT_ROWS+ 77,{1,2,3,7,6,5,4,0},{0,5,7,6,4,3,2,1}},
+  {0x60,2,OC_EXT_ROWS+ 82,{5,6,7,4,3,2,1,0},{0,2,7,6,5,4,3,1}},
+  {0x30,2,OC_EXT_ROWS+ 36,{4,5,7,6,3,2,1,0},{0,4,7,6,5,3,2,1}},
+  {0x18,2,OC_EXT_ROWS+ 90,{3,4,7,6,5,2,1,0},{0,1,7,6,5,4,3,2}},
+  {0x0C,2,OC_EXT_ROWS+ 34,{2,3,7,6,5,4,1,0},{0,4,7,6,5,3,2,1}},
+  {0x06,2,OC_EXT_ROWS+ 84,{1,2,7,6,5,4,3,0},{0,2,7,6,5,4,3,1}},
+  {0x40,1,OC_EXT_ROWS+  0,{6,7,5,4,3,2,1,0},{0,7,6,5,4,3,2,1}},
+  {0x20,1,OC_EXT_ROWS+  0,{5,7,6,4,3,2,1,0},{0,7,6,5,4,3,2,1}},
+  {0x10,1,OC_EXT_ROWS+  0,{4,7,6,5,3,2,1,0},{0,7,6,5,4,3,2,1}},
+  {0x08,1,OC_EXT_ROWS+  0,{3,7,6,5,4,2,1,0},{0,7,6,5,4,3,2,1}},
+  {0x04,1,OC_EXT_ROWS+  0,{2,7,6,5,4,3,1,0},{0,7,6,5,4,3,2,1}},
+  {0x02,1,OC_EXT_ROWS+  0,{1,7,6,5,4,3,2,0},{0,7,6,5,4,3,2,1}}
+};
+
+
+
+/*Pads a single column of a partial block and then performs a forward Type-II
+   DCT on the result.
+  The input is scaled by a factor of 4 and biased appropriately for the current
+   fDCT implementation.
+  The output is scaled by an additional factor of 2 from the orthonormal
+   version of the transform.
+  _y: The buffer to store the result in.
+      Data will be placed the first 8 entries (e.g., in a row of an 8x8 block).
+  _x: The input coefficients.
+      Every 8th entry is used (e.g., from a column of an 8x8 block).
+  _e: The extension information for the shape.*/
+static void oc_fdct8_ext(ogg_int16_t _y[8],ogg_int16_t *_x,
+ const oc_extension_info *_e){
+  const unsigned char *pi;
+  int                  na;
+  na=_e->na;
+  pi=_e->pi;
+  if(na==1){
+    int ci;
+    /*While the branch below is still correct for shapes with na==1, we can
+       perform the entire transform with just 1 multiply in this case instead
+       of 23.*/
+    _y[0]=(ogg_int16_t)(OC_DIV2_16(OC_C4S4*(_x[pi[0]])));
+    for(ci=1;ci<8;ci++)_y[ci]=0;
+  }
+  else{
+    const ogg_int16_t *const *ext;
+    int                       zpi;
+    int                       api;
+    int                       nz;
+    /*First multiply by the extension matrix to compute the padding values.*/
+    nz=8-na;
+    ext=_e->ext;
+    for(zpi=0;zpi<nz;zpi++){
+      ogg_int32_t v;
+      v=0;
+      for(api=0;api<na;api++){
+        v+=ext[zpi][api]*(ogg_int32_t)(_x[pi[api]<<3]<<1);
+      }
+      _x[pi[na+zpi]<<3]=(ogg_int16_t)(v+0x8000>>16)+1>>1;
+    }
+    oc_fdct8(_y,_x);
+  }
+}
+
+/*Performs a forward 8x8 Type-II DCT transform on blocks which overlap the
+   border of the picture region.
+  This method ONLY works with rectangular regions.
+  _border: A description of which pixels are inside the border.
+  _y:      The buffer to store the result in.
+           This may be the same as _x.
+  _x:      The input pixel values.
+           Pixel values outside the border will be ignored.*/
+void oc_fdct8x8_border(const oc_border_info *_border,
+ ogg_int16_t _y[64],const ogg_int16_t _x[64]){
+  ogg_int16_t             *in;
+  ogg_int16_t             *out;
+  ogg_int16_t              w[64];
+  ogg_int64_t              mask;
+  const oc_extension_info *cext;
+  const oc_extension_info *rext;
+  int                      cmask;
+  int                      rmask;
+  int                      ri;
+  int                      ci;
+  /*Identify the shapes of the non-zero rows and columns.*/
+  rmask=cmask=0;
+  mask=_border->mask;
+  for(ri=0;ri<8;ri++){
+    /*This aggregation is _only_ correct for rectangular masks.*/
+    cmask|=((mask&0xFF)!=0)<<ri;
+    rmask|=mask&0xFF;
+    mask>>=8;
+  }
+  /*Find the associated extension info for these shapes.*/
+  if(cmask==0xFF)cext=NULL;
+  else for(cext=OC_EXTENSION_INFO;cext->mask!=cmask;){
+    /*If we somehow can't find the shape, then just do an unpadded fDCT.
+      It won't be efficient, but it should still be correct.*/
+    if(++cext>=OC_EXTENSION_INFO+OC_NSHAPES){
+      oc_enc_fdct8x8_c(_y,_x);
+      return;
+    }
+  }
+  if(rmask==0xFF)rext=NULL;
+  else for(rext=OC_EXTENSION_INFO;rext->mask!=rmask;){
+    /*If we somehow can't find the shape, then just do an unpadded fDCT.
+      It won't be efficient, but it should still be correct.*/
+    if(++rext>=OC_EXTENSION_INFO+OC_NSHAPES){
+      oc_enc_fdct8x8_c(_y,_x);
+      return;
+    }
+  }
+  /*Add two extra bits of working precision to improve accuracy; any more and
+     we could overflow.*/
+  for(ci=0;ci<64;ci++)w[ci]=_x[ci]<<2;
+  /*These biases correct for some systematic error that remains in the full
+     fDCT->iDCT round trip.
+    We can safely add them before padding, since if these pixel values are
+     overwritten, we didn't care what they were anyway (and the unbiased values
+     will usually yield smaller DCT coefficient magnitudes).*/
+  w[0]+=(w[0]!=0)+1;
+  w[1]++;
+  w[8]--;
+  /*Transform the columns.
+    We can ignore zero columns without a problem.*/
+  in=w;
+  out=_y;
+  if(cext==NULL)for(ci=0;ci<8;ci++)oc_fdct8(out+(ci<<3),in+ci);
+  else for(ci=0;ci<8;ci++)if(rmask&(1<<ci))oc_fdct8_ext(out+(ci<<3),in+ci,cext);
+  /*Transform the rows.
+    We transform even rows that are supposedly zero, because rounding errors
+     may make them slightly non-zero, and this will give a more precise
+     reconstruction with very small quantizers.*/
+  in=_y;
+  out=w;
+  if(rext==NULL)for(ri=0;ri<8;ri++)oc_fdct8(out+(ri<<3),in+ri);
+  else for(ri=0;ri<8;ri++)oc_fdct8_ext(out+(ri<<3),in+ri,rext);
+  /*Round the result back to the external working precision (which is still
+     scaled by four relative to the orthogonal result).
+    TODO: We should just update the external working precision.*/
+  for(ci=0;ci<64;ci++)_y[ci]=w[ci]+2>>2;
+}
+#endif

thirdparty/libtheora/fragment.c

@@ -0,0 +1,82 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+#include <string.h>
+#include "internal.h"
+
+void oc_frag_copy_c(unsigned char *_dst,const unsigned char *_src,int _ystride){
+  int i;
+  for(i=8;i-->0;){
+    memcpy(_dst,_src,8*sizeof(*_dst));
+    _dst+=_ystride;
+    _src+=_ystride;
+  }
+}
+
+/*Copies the fragments specified by the lists of fragment indices from one
+   frame to another.
+  _dst_frame:     The reference frame to copy to.
+  _src_frame:     The reference frame to copy from.
+  _ystride:       The row stride of the reference frames.
+  _fragis:        A pointer to a list of fragment indices.
+  _nfragis:       The number of fragment indices to copy.
+  _frag_buf_offs: The offsets of fragments in the reference frames.*/
+void oc_frag_copy_list_c(unsigned char *_dst_frame,
+ const unsigned char *_src_frame,int _ystride,
+ const ptrdiff_t *_fragis,ptrdiff_t _nfragis,const ptrdiff_t *_frag_buf_offs){
+  ptrdiff_t fragii;
+  for(fragii=0;fragii<_nfragis;fragii++){
+    ptrdiff_t frag_buf_off;
+    frag_buf_off=_frag_buf_offs[_fragis[fragii]];
+    oc_frag_copy_c(_dst_frame+frag_buf_off,
+     _src_frame+frag_buf_off,_ystride);
+  }
+}
+
+void oc_frag_recon_intra_c(unsigned char *_dst,int _ystride,
+ const ogg_int16_t _residue[64]){
+  int i;
+  for(i=0;i<8;i++){
+    int j;
+    for(j=0;j<8;j++)_dst[j]=OC_CLAMP255(_residue[i*8+j]+128);
+    _dst+=_ystride;
+  }
+}
+
+void oc_frag_recon_inter_c(unsigned char *_dst,
+ const unsigned char *_src,int _ystride,const ogg_int16_t _residue[64]){
+  int i;
+  for(i=0;i<8;i++){
+    int j;
+    for(j=0;j<8;j++)_dst[j]=OC_CLAMP255(_residue[i*8+j]+_src[j]);
+    _dst+=_ystride;
+    _src+=_ystride;
+  }
+}
+
+void oc_frag_recon_inter2_c(unsigned char *_dst,const unsigned char *_src1,
+ const unsigned char *_src2,int _ystride,const ogg_int16_t _residue[64]){
+  int i;
+  for(i=0;i<8;i++){
+    int j;
+    for(j=0;j<8;j++)_dst[j]=OC_CLAMP255(_residue[i*8+j]+(_src1[j]+_src2[j]>>1));
+    _dst+=_ystride;
+    _src1+=_ystride;
+    _src2+=_ystride;
+  }
+}
+
+void oc_restore_fpu_c(void){}

thirdparty/libtheora/huffdec.c

@@ -0,0 +1,513 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009,2025           *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+#include <stdlib.h>
+#include <string.h>
+#include <ogg/ogg.h>
+#include "huffdec.h"
+#include "decint.h"
+
+
+
+/*Instead of storing every branching in the tree, subtrees can be collapsed
+   into one node, with a table of size 1<<nbits pointing directly to its
+   descedents nbits levels down.
+  This allows more than one bit to be read at a time, and avoids following all
+   the intermediate branches with next to no increased code complexity once
+   the collapsed tree has been built.
+  We do _not_ require that a subtree be complete to be collapsed, but instead
+   store duplicate pointers in the table, and record the actual depth of the
+   node below its parent.
+  This tells us the number of bits to advance the stream after reaching it.
+
+  This turns out to be equivalent to the method described in \cite{Hash95},
+   without the requirement that codewords be sorted by length.
+  If the codewords were sorted by length (so-called ``canonical-codes''), they
+   could be decoded much faster via either Lindell and Moffat's approach or
+   Hashemian's Condensed Huffman Code approach, the latter of which has an
+   extremely small memory footprint.
+  We can't use Choueka et al.'s finite state machine approach, which is
+   extremely fast, because we can't allow multiple symbols to be output at a
+   time; the codebook can and does change between symbols.
+  It also has very large memory requirements, which impairs cache coherency.
+
+  We store the tree packed in an array of 16-bit integers (words).
+  Each node consists of a single word, followed consecutively by two or more
+   indices of its children.
+  Let n be the value of this first word.
+  This is the number of bits that need to be read to traverse the node, and
+   must be positive.
+  1<<n entries follow in the array, each an index to a child node.
+  If the child is positive, then it is the index of another internal node in
+   the table.
+  If the child is negative or zero, then it is a leaf node.
+  These are stored directly in the child pointer to save space, since they only
+   require a single word.
+  If a leaf node would have been encountered before reading n bits, then it is
+   duplicated the necessary number of times in this table.
+  Leaf nodes pack both a token value and their actual depth in the tree.
+  The token in the leaf node is (-leaf&255).
+  The number of bits that need to be consumed to reach the leaf, starting from
+   the current node, is (-leaf>>8).
+
+  @ARTICLE{Hash95,
+    author="Reza Hashemian",
+    title="Memory Efficient and High-Speed Search {Huffman} Coding",
+    journal="{IEEE} Transactions on Communications",
+    volume=43,
+    number=10,
+    pages="2576--2581",
+    month=Oct,
+    year=1995
+  }*/
+
+
+
+/*The map from external spec-defined tokens to internal tokens.
+  This is constructed so that any extra bits read with the original token value
+   can be masked off the least significant bits of its internal token index.
+  In addition, all of the tokens which require additional extra bits are placed
+   at the start of the list, and grouped by type.
+  OC_DCT_REPEAT_RUN3_TOKEN is placed first, as it is an extra-special case, so
+   giving it index 0 may simplify comparisons on some architectures.
+  These requirements require some substantial reordering.*/
+static const unsigned char OC_DCT_TOKEN_MAP[TH_NDCT_TOKENS]={
+  /*OC_DCT_EOB1_TOKEN (0 extra bits)*/
+  15,
+  /*OC_DCT_EOB2_TOKEN (0 extra bits)*/
+  16,
+  /*OC_DCT_EOB3_TOKEN (0 extra bits)*/
+  17,
+  /*OC_DCT_REPEAT_RUN0_TOKEN (2 extra bits)*/
+  88,
+  /*OC_DCT_REPEAT_RUN1_TOKEN (3 extra bits)*/
+  80,
+  /*OC_DCT_REPEAT_RUN2_TOKEN (4 extra bits)*/
+   1,
+  /*OC_DCT_REPEAT_RUN3_TOKEN (12 extra bits)*/
+   0,
+  /*OC_DCT_SHORT_ZRL_TOKEN (3 extra bits)*/
+  48,
+  /*OC_DCT_ZRL_TOKEN (6 extra bits)*/
+  14,
+  /*OC_ONE_TOKEN (0 extra bits)*/
+  56,
+  /*OC_MINUS_ONE_TOKEN (0 extra bits)*/
+  57,
+  /*OC_TWO_TOKEN (0 extra bits)*/
+  58,
+  /*OC_MINUS_TWO_TOKEN (0 extra bits)*/
+  59,
+  /*OC_DCT_VAL_CAT2 (1 extra bit)*/
+  60,
+  62,
+  64,
+  66,
+  /*OC_DCT_VAL_CAT3 (2 extra bits)*/
+  68,
+  /*OC_DCT_VAL_CAT4 (3 extra bits)*/
+  72,
+  /*OC_DCT_VAL_CAT5 (4 extra bits)*/
+   2,
+  /*OC_DCT_VAL_CAT6 (5 extra bits)*/
+   4,
+  /*OC_DCT_VAL_CAT7 (6 extra bits)*/
+   6,
+  /*OC_DCT_VAL_CAT8 (10 extra bits)*/
+   8,
+  /*OC_DCT_RUN_CAT1A (1 extra bit)*/
+  18,
+  20,
+  22,
+  24,
+  26,
+  /*OC_DCT_RUN_CAT1B (3 extra bits)*/
+  32,
+  /*OC_DCT_RUN_CAT1C (4 extra bits)*/
+  12,
+  /*OC_DCT_RUN_CAT2A (2 extra bits)*/
+  28,
+  /*OC_DCT_RUN_CAT2B (3 extra bits)*/
+  40
+};
+
+/*The log base 2 of number of internal tokens associated with each of the spec
+   tokens (i.e., how many of the extra bits are folded into the token value).
+  Increasing the maximum value beyond 3 will enlarge the amount of stack
+   required for tree construction.*/
+static const unsigned char OC_DCT_TOKEN_MAP_LOG_NENTRIES[TH_NDCT_TOKENS]={
+  0,0,0,2,3,0,0,3,0,0,0,0,0,1,1,1,1,2,3,1,1,1,2,1,1,1,1,1,3,1,2,3
+};
+
+
+/*The size a lookup table is allowed to grow to relative to the number of
+   unique nodes it contains.
+  E.g., if OC_HUFF_SLUSH is 4, then at most 75% of the space in the tree is
+   wasted (1/4 of the space must be used).
+  Larger numbers can decode tokens with fewer read operations, while smaller
+   numbers may save more space.
+  With a sample file:
+  32233473 read calls are required when no tree collapsing is done (100.0%).
+  19269269 read calls are required when OC_HUFF_SLUSH is 1 (59.8%).
+  11144969 read calls are required when OC_HUFF_SLUSH is 2 (34.6%).
+  10538563 read calls are required when OC_HUFF_SLUSH is 4 (32.7%).
+  10192578 read calls are required when OC_HUFF_SLUSH is 8 (31.6%).
+  Since a value of 2 gets us the vast majority of the speed-up with only a
+   small amount of wasted memory, this is what we use.
+  This value must be less than 128, or you could create a tree with more than
+   32767 entries, which would overflow the 16-bit words used to index it.*/
+#define OC_HUFF_SLUSH (2)
+/*The root of the tree is on the fast path, and a larger value here is more
+   beneficial than elsewhere in the tree.
+  7 appears to give the best performance, trading off between increased use of
+   the single-read fast path and cache footprint for the tables, though
+   obviously this will depend on your cache size.
+  Using 7 here, the VP3 tables are about twice as large compared to using 2.*/
+#define OC_ROOT_HUFF_SLUSH (7)
+
+
+
+/*Unpacks a Huffman codebook.
+  _opb:    The buffer to unpack from.
+  _tokens: Stores a list of internal tokens, in the order they were found in
+            the codebook, and the lengths of their corresponding codewords.
+           This is enough to completely define the codebook, while minimizing
+            stack usage and avoiding temporary allocations (for platforms
+            where free() is a no-op).
+  Return: The number of internal tokens in the codebook, or a negative value
+   on error.*/
+int oc_huff_tree_unpack(oc_pack_buf *_opb,unsigned char _tokens[256][2]){
+  ogg_uint32_t code;
+  int          len;
+  int          ntokens;
+  int          nleaves;
+  code=0;
+  len=ntokens=nleaves=0;
+  for(;;){
+    long bits;
+    bits=oc_pack_read1(_opb);
+    /*Only process nodes so long as there's more bits in the buffer.*/
+    if(oc_pack_bytes_left(_opb)<0)return TH_EBADHEADER;
+    /*Read an internal node:*/
+    if(!bits){
+      len++;
+      /*Don't allow codewords longer than 32 bits.*/
+      if(len>32)return TH_EBADHEADER;
+    }
+    /*Read a leaf node:*/
+    else{
+      ogg_uint32_t code_bit;
+      int          neb;
+      int          nentries;
+      int          token;
+      /*Don't allow more than 32 spec-tokens per codebook.*/
+      if(++nleaves>32)return TH_EBADHEADER;
+      bits=oc_pack_read(_opb,OC_NDCT_TOKEN_BITS);
+      neb=OC_DCT_TOKEN_MAP_LOG_NENTRIES[bits];
+      token=OC_DCT_TOKEN_MAP[bits];
+      nentries=1<<neb;
+      while(nentries-->0){
+        _tokens[ntokens][0]=(unsigned char)token++;
+        _tokens[ntokens][1]=(unsigned char)(len+neb);
+        ntokens++;
+      }
+      if(len<=0)break;
+      code_bit=0x80000000U>>len-1;
+      while(len>0&&(code&code_bit)){
+        code^=code_bit;
+        code_bit<<=1;
+        len--;
+      }
+      if(len<=0)break;
+      code|=code_bit;
+    }
+  }
+  return ntokens;
+}
+
+/*Count how many tokens would be required to fill a subtree at depth _depth.
+  _tokens: A list of internal tokens, in the order they are found in the
+            codebook, and the lengths of their corresponding codewords.
+  _depth:  The depth of the desired node in the corresponding tree structure.
+  Return: The number of tokens that belong to that subtree.*/
+static int oc_huff_subtree_tokens(unsigned char _tokens[][2],int _depth){
+  ogg_uint32_t code;
+  int          ti;
+  code=0;
+  ti=0;
+  do{
+    if(_tokens[ti][1]-_depth<32)code+=0x80000000U>>_tokens[ti++][1]-_depth;
+    else{
+      /*Because of the expanded internal tokens, we can have codewords as long
+         as 35 bits.
+        A single recursion here is enough to advance past them.*/
+      code++;
+      ti+=oc_huff_subtree_tokens(_tokens+ti,_depth+31);
+    }
+  }
+  while(code<0x80000000U);
+  return ti;
+}
+
+/*Compute the number of bits to use for a collapsed tree node at the given
+   depth.
+  _tokens:  A list of internal tokens, in the order they are found in the
+             codebook, and the lengths of their corresponding codewords.
+  _ntokens: The number of tokens corresponding to this tree node.
+  _depth:   The depth of this tree node.
+  Return: The number of bits to use for a collapsed tree node rooted here.
+          This is always at least one, even if this was a leaf node.*/
+static int oc_huff_tree_collapse_depth(unsigned char _tokens[][2],
+ int _ntokens,int _depth){
+  int got_leaves;
+  int loccupancy;
+  int occupancy;
+  int slush;
+  int nbits;
+  int best_nbits;
+  slush=_depth>0?OC_HUFF_SLUSH:OC_ROOT_HUFF_SLUSH;
+  /*It's legal to have a tree with just a single node, which requires no bits
+     to decode and always returns the same token.
+    However, no encoder actually does this (yet).
+    To avoid a special case in oc_huff_token_decode(), we force the number of
+     lookahead bits to be at least one.
+    This will produce a tree that looks ahead one bit and then advances the
+     stream zero bits.*/
+  nbits=1;
+  occupancy=2;
+  got_leaves=1;
+  do{
+    int ti;
+    if(got_leaves)best_nbits=nbits;
+    nbits++;
+    got_leaves=0;
+    loccupancy=occupancy;
+    for(occupancy=ti=0;ti<_ntokens;occupancy++){
+      if(_tokens[ti][1]<_depth+nbits)ti++;
+      else if(_tokens[ti][1]==_depth+nbits){
+        got_leaves=1;
+        ti++;
+      }
+      else ti+=oc_huff_subtree_tokens(_tokens+ti,_depth+nbits);
+    }
+  }
+  while(occupancy>loccupancy&&occupancy*slush>=1<<nbits);
+  return best_nbits;
+}
+
+/*Determines the size in words of a Huffman tree node that represents a
+   subtree of depth _nbits.
+  _nbits: The depth of the subtree.
+          This must be greater than zero.
+  Return: The number of words required to store the node.*/
+static size_t oc_huff_node_size(int _nbits){
+  return 1+(1<<_nbits);
+}
+
+/*Produces a collapsed-tree representation of the given token list.
+  _tree: The storage for the collapsed Huffman tree.
+         This may be NULL to compute the required storage size instead of
+          constructing the tree.
+  _tokens:  A list of internal tokens, in the order they are found in the
+             codebook, and the lengths of their corresponding codewords.
+  _ntokens: The number of tokens corresponding to this tree node.
+  Return: The number of words required to store the tree.*/
+static size_t oc_huff_tree_collapse(ogg_int16_t *_tree,
+ unsigned char _tokens[][2],int _ntokens){
+  ogg_int16_t   node[34];
+  unsigned char depth[34];
+  unsigned char last[34];
+  size_t        ntree;
+  int           ti;
+  int           l;
+  depth[0]=0;
+  last[0]=(unsigned char)(_ntokens-1);
+  ntree=0;
+  ti=0;
+  l=0;
+  do{
+    int nbits;
+    nbits=oc_huff_tree_collapse_depth(_tokens+ti,last[l]+1-ti,depth[l]);
+    node[l]=(ogg_int16_t)ntree;
+    ntree+=oc_huff_node_size(nbits);
+    if(_tree!=NULL)_tree[node[l]++]=(ogg_int16_t)nbits;
+    do{
+      while(ti<=last[l]&&_tokens[ti][1]<=depth[l]+nbits){
+        if(_tree!=NULL){
+          ogg_int16_t leaf;
+          int         nentries;
+          nentries=1<<depth[l]+nbits-_tokens[ti][1];
+          leaf=(ogg_int16_t)-(_tokens[ti][1]-depth[l]<<8|_tokens[ti][0]);
+          while(nentries-->0)_tree[node[l]++]=leaf;
+        }
+        ti++;
+      }
+      if(ti<=last[l]){
+        /*We need to recurse*/
+        depth[l+1]=(unsigned char)(depth[l]+nbits);
+        if(_tree!=NULL)_tree[node[l]++]=(ogg_int16_t)ntree;
+        l++;
+        last[l]=
+         (unsigned char)(ti+oc_huff_subtree_tokens(_tokens+ti,depth[l])-1);
+        break;
+      }
+      /*Pop back up a level of recursion.*/
+      else if(l-->0)nbits=depth[l+1]-depth[l];
+    }
+    while(l>=0);
+  }
+  while(l>=0);
+  return ntree;
+}
+
+/*Unpacks a set of Huffman trees, and reduces them to a collapsed
+   representation.
+  _opb:   The buffer to unpack the trees from.
+  _nodes: The table to fill with the Huffman trees.
+  Return: 0 on success, or a negative value on error.
+          The caller is responsible for cleaning up any partially initialized
+           _nodes on failure.*/
+int oc_huff_trees_unpack(oc_pack_buf *_opb,
+ ogg_int16_t *_nodes[TH_NHUFFMAN_TABLES]){
+  int i;
+  for(i=0;i<TH_NHUFFMAN_TABLES;i++){
+    unsigned char  tokens[256][2];
+    int            ntokens;
+    ogg_int16_t   *tree;
+    size_t         size;
+    /*Unpack the full tree into a temporary buffer.*/
+    ntokens=oc_huff_tree_unpack(_opb,tokens);
+    if(ntokens<0)return ntokens;
+    /*Figure out how big the collapsed tree will be and allocate space for it.*/
+    size=oc_huff_tree_collapse(NULL,tokens,ntokens);
+    /*This should never happen; if it does it means you set OC_HUFF_SLUSH or
+       OC_ROOT_HUFF_SLUSH too large.*/
+    if(size>32767)return TH_EIMPL;
+    tree=(ogg_int16_t *)_ogg_malloc(size*sizeof(*tree));
+    if(tree==NULL)return TH_EFAULT;
+    /*Construct the collapsed the tree.*/
+    oc_huff_tree_collapse(tree,tokens,ntokens);
+    _nodes[i]=tree;
+  }
+  return 0;
+}
+
+/*Determines the size in words of a Huffman subtree.
+  _tree: The complete Huffman tree.
+  _node: The index of the root of the desired subtree.
+  Return: The number of words required to store the tree.*/
+static size_t oc_huff_tree_size(const ogg_int16_t *_tree,int _node){
+  size_t size;
+  int    nchildren;
+  int    n;
+  int    i;
+  n=_tree[_node];
+  size=oc_huff_node_size(n);
+  nchildren=1<<n;
+  i=0;
+  do{
+    int child;
+    child=_tree[_node+i+1];
+    if(child<=0)i+=1<<n-(-child>>8);
+    else{
+      size+=oc_huff_tree_size(_tree,child);
+      i++;
+    }
+  }
+  while(i<nchildren);
+  return size;
+}
+
+/*Makes a copy of the given set of Huffman trees.
+  _dst: The array to store the copy in.
+  _src: The array of trees to copy.*/
+int oc_huff_trees_copy(ogg_int16_t *_dst[TH_NHUFFMAN_TABLES],
+ const ogg_int16_t *const _src[TH_NHUFFMAN_TABLES]){
+  int i;
+  for(i=0;i<TH_NHUFFMAN_TABLES;i++){
+    size_t size;
+    size=oc_huff_tree_size(_src[i],0);
+    _dst[i]=(ogg_int16_t *)_ogg_malloc(size*sizeof(*_dst[i]));
+    if(_dst[i]==NULL){
+      while(i-->0)_ogg_free(_dst[i]);
+      return TH_EFAULT;
+    }
+    memcpy(_dst[i],_src[i],size*sizeof(*_dst[i]));
+  }
+  return 0;
+}
+
+/*Frees the memory used by a set of Huffman trees.
+  _nodes: The array of trees to free.*/
+void oc_huff_trees_clear(ogg_int16_t *_nodes[TH_NHUFFMAN_TABLES]){
+  int i;
+  for(i=0;i<TH_NHUFFMAN_TABLES;i++)_ogg_free(_nodes[i]);
+}
+
+
+/*Unpacks a single token using the given Huffman tree.
+  _opb:  The buffer to unpack the token from.
+  _node: The tree to unpack the token with.
+  Return: The token value.*/
+int oc_huff_token_decode_c(oc_pack_buf *_opb,const ogg_int16_t *_tree){
+  const unsigned char *ptr;
+  const unsigned char *stop;
+  oc_pb_window         window;
+  int                  available;
+  long                 bits;
+  int                  node;
+  int                  n;
+  ptr=_opb->ptr;
+  window=_opb->window;
+  stop=_opb->stop;
+  available=_opb->bits;
+  node=0;
+  for(;;){
+    n=_tree[node];
+    if(n>available){
+      unsigned shift;
+      shift=OC_PB_WINDOW_SIZE-available;
+      do{
+        /*We don't bother setting eof because we won't check for it after we've
+           started decoding DCT tokens.*/
+        if(ptr>=stop){
+          shift=(unsigned)-OC_LOTS_OF_BITS;
+          break;
+        }
+        shift-=8;
+        window|=(oc_pb_window)*ptr++<<shift;
+      }
+      while(shift>=8);
+      /*Note: We never request more than 24 bits, so there's no need to fill in
+         the last partial byte here.*/
+      available=OC_PB_WINDOW_SIZE-shift;
+    }
+    bits=window>>OC_PB_WINDOW_SIZE-n;
+    node=_tree[node+1+bits];
+    if(node<=0)break;
+    window<<=n;
+    available-=n;
+  }
+  node=-node;
+  n=node>>8;
+  window<<=n;
+  available-=n;
+  _opb->ptr=ptr;
+  _opb->window=window;
+  _opb->bits=available;
+  return node&255;
+}

thirdparty/libtheora/huffdec.h

@@ -0,0 +1,32 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+#if !defined(_huffdec_H)
+# define _huffdec_H (1)
+# include "huffman.h"
+# include "bitpack.h"
+
+
+
+int oc_huff_trees_unpack(oc_pack_buf *_opb,
+ ogg_int16_t *_nodes[TH_NHUFFMAN_TABLES]);
+int oc_huff_trees_copy(ogg_int16_t *_dst[TH_NHUFFMAN_TABLES],
+ const ogg_int16_t *const _src[TH_NHUFFMAN_TABLES]);
+void oc_huff_trees_clear(ogg_int16_t *_nodes[TH_NHUFFMAN_TABLES]);
+int oc_huff_token_decode_c(oc_pack_buf *_opb,const ogg_int16_t *_node);
+
+#endif

thirdparty/libtheora/huffenc.c

@@ -0,0 +1,966 @@
+#include <stdlib.h>
+#include <string.h>
+#include <ogg/ogg.h>
+#include "huffenc.h"
+
+
+
+/*The default Huffman codes used for VP3.1.*/
+const th_huff_code TH_VP31_HUFF_CODES[TH_NHUFFMAN_TABLES][TH_NDCT_TOKENS]={
+  {
+    {0x002D, 6},{0x0026, 7},{0x0166, 9},{0x004E, 8},
+    {0x02CE,10},{0x059E,11},{0x027D,11},{0x0008, 5},
+    {0x04F9,12},{0x000F, 4},{0x000E, 4},{0x001B, 5},
+    {0x0006, 4},{0x0008, 4},{0x0005, 4},{0x001A, 5},
+    {0x0015, 5},{0x0007, 4},{0x000C, 4},{0x0001, 3},
+    {0x0000, 3},{0x0009, 4},{0x0017, 5},{0x0029, 6},
+    {0x0028, 6},{0x00B2, 8},{0x04F8,12},{0x059F,11},
+    {0x009E, 9},{0x013F,10},{0x0012, 6},{0x0058, 7}
+  },
+  {
+    {0x0010, 5},{0x0047, 7},{0x01FF, 9},{0x008C, 8},
+    {0x03FC,10},{0x046A,11},{0x0469,11},{0x0022, 6},
+    {0x11A1,13},{0x000E, 4},{0x000D, 4},{0x0004, 4},
+    {0x0005, 4},{0x0009, 4},{0x0006, 4},{0x001E, 5},
+    {0x0016, 5},{0x0007, 4},{0x000C, 4},{0x0001, 3},
+    {0x0000, 3},{0x000A, 4},{0x0017, 5},{0x007D, 7},
+    {0x007E, 7},{0x011B, 9},{0x08D1,12},{0x03FD,10},
+    {0x046B,11},{0x11A0,13},{0x007C, 7},{0x00FE, 8}
+  },
+  {
+    {0x0016, 5},{0x0020, 6},{0x0086, 8},{0x0087, 8},
+    {0x0367,10},{0x06CC,11},{0x06CB,11},{0x006E, 7},
+    {0x366D,14},{0x000F, 4},{0x000E, 4},{0x0004, 4},
+    {0x0005, 4},{0x000A, 4},{0x0006, 4},{0x001A, 5},
+    {0x0011, 5},{0x0007, 4},{0x000C, 4},{0x0001, 3},
+    {0x0000, 3},{0x0009, 4},{0x0017, 5},{0x006F, 7},
+    {0x006D, 7},{0x0364,10},{0x0D9A,12},{0x06CA,11},
+    {0x1B37,13},{0x366C,14},{0x0042, 7},{0x00D8, 8}
+  },
+  {
+    {0x0000, 4},{0x002D, 6},{0x00F7, 8},{0x0058, 7},
+    {0x0167, 9},{0x02CB,10},{0x02CA,10},{0x000E, 6},
+    {0x1661,13},{0x0003, 3},{0x0002, 3},{0x0008, 4},
+    {0x0009, 4},{0x000D, 4},{0x0002, 4},{0x001F, 5},
+    {0x0017, 5},{0x0001, 4},{0x000C, 4},{0x000E, 4},
+    {0x000A, 4},{0x0006, 5},{0x0078, 7},{0x000F, 6},
+    {0x007A, 7},{0x0164, 9},{0x0599,11},{0x02CD,10},
+    {0x0B31,12},{0x1660,13},{0x0079, 7},{0x00F6, 8}
+  },
+  {
+    {0x0003, 4},{0x003C, 6},{0x000F, 7},{0x007A, 7},
+    {0x001D, 8},{0x0020, 9},{0x0072,10},{0x0006, 6},
+    {0x0399,13},{0x0004, 3},{0x0005, 3},{0x0005, 4},
+    {0x0006, 4},{0x000E, 4},{0x0004, 4},{0x0000, 4},
+    {0x0019, 5},{0x0002, 4},{0x000D, 4},{0x0007, 4},
+    {0x001F, 5},{0x0030, 6},{0x0011, 8},{0x0031, 6},
+    {0x0005, 6},{0x0021, 9},{0x00E7,11},{0x0038, 9},
+    {0x01CD,12},{0x0398,13},{0x007B, 7},{0x0009, 7}
+  },
+  {
+    {0x0009, 4},{0x0002, 5},{0x0074, 7},{0x0007, 6},
+    {0x00EC, 8},{0x00D1, 9},{0x01A6,10},{0x0006, 6},
+    {0x0D21,13},{0x0005, 3},{0x0006, 3},{0x0008, 4},
+    {0x0007, 4},{0x000F, 4},{0x0004, 4},{0x0000, 4},
+    {0x001C, 5},{0x0002, 4},{0x0005, 4},{0x0003, 4},
+    {0x000C, 5},{0x0035, 7},{0x01A7,10},{0x001B, 6},
+    {0x0077, 7},{0x01A5,10},{0x0349,11},{0x00D0, 9},
+    {0x0691,12},{0x0D20,13},{0x0075, 7},{0x00ED, 8}
+  },
+  {
+    {0x000A, 4},{0x000C, 5},{0x0012, 6},{0x001B, 6},
+    {0x00B7, 8},{0x016C, 9},{0x0099, 9},{0x005A, 7},
+    {0x16D8,13},{0x0007, 3},{0x0006, 3},{0x0009, 4},
+    {0x0008, 4},{0x0000, 3},{0x0005, 4},{0x0017, 5},
+    {0x000E, 5},{0x0002, 4},{0x0003, 4},{0x000F, 5},
+    {0x001A, 6},{0x004D, 8},{0x2DB3,14},{0x002C, 6},
+    {0x0011, 6},{0x02DA,10},{0x05B7,11},{0x0098, 9},
+    {0x0B6D,12},{0x2DB2,14},{0x0010, 6},{0x0027, 7}
+  },
+  {
+    {0x000D, 4},{0x000F, 5},{0x001D, 6},{0x0008, 5},
+    {0x0051, 7},{0x0056, 8},{0x00AF, 9},{0x002A, 7},
+    {0x148A,13},{0x0007, 3},{0x0000, 2},{0x0008, 4},
+    {0x0009, 4},{0x000C, 4},{0x0006, 4},{0x0017, 5},
+    {0x000B, 5},{0x0016, 5},{0x0015, 5},{0x0009, 5},
+    {0x0050, 7},{0x00AE, 9},{0x2917,14},{0x001C, 6},
+    {0x0014, 6},{0x0290,10},{0x0523,11},{0x0149, 9},
+    {0x0A44,12},{0x2916,14},{0x0053, 7},{0x00A5, 8}
+  },
+  {
+    {0x0001, 4},{0x001D, 6},{0x00F5, 8},{0x00F4, 8},
+    {0x024D,10},{0x0499,11},{0x0498,11},{0x0001, 5},
+    {0x0021, 6},{0x0006, 3},{0x0005, 3},{0x0006, 4},
+    {0x0005, 4},{0x0002, 4},{0x0007, 5},{0x0025, 6},
+    {0x007B, 7},{0x001C, 6},{0x0020, 6},{0x000D, 6},
+    {0x0048, 7},{0x0092, 8},{0x0127, 9},{0x000E, 4},
+    {0x0004, 4},{0x0011, 5},{0x000C, 6},{0x003C, 6},
+    {0x000F, 5},{0x0000, 5},{0x001F, 5},{0x0013, 5}
+  },
+  {
+    {0x0005, 4},{0x003C, 6},{0x0040, 7},{0x000D, 7},
+    {0x0031, 9},{0x0061,10},{0x0060,10},{0x0002, 5},
+    {0x00F5, 8},{0x0006, 3},{0x0005, 3},{0x0007, 4},
+    {0x0006, 4},{0x0002, 4},{0x0009, 5},{0x0025, 6},
+    {0x0007, 6},{0x0021, 6},{0x0024, 6},{0x0010, 6},
+    {0x0041, 7},{0x00F4, 8},{0x0019, 8},{0x000E, 4},
+    {0x0003, 4},{0x0011, 5},{0x0011, 6},{0x003F, 6},
+    {0x003E, 6},{0x007B, 7},{0x0000, 4},{0x0013, 5}
+  },
+  {
+    {0x000A, 4},{0x0007, 5},{0x0001, 6},{0x0009, 6},
+    {0x0131, 9},{0x0261,10},{0x0260,10},{0x0015, 6},
+    {0x0001, 7},{0x0007, 3},{0x0006, 3},{0x0008, 4},
+    {0x0007, 4},{0x0006, 4},{0x0012, 5},{0x002F, 6},
+    {0x0014, 6},{0x0027, 6},{0x002D, 6},{0x0016, 6},
+    {0x004D, 7},{0x0099, 8},{0x0000, 7},{0x0004, 4},
+    {0x0001, 4},{0x0005, 5},{0x0017, 6},{0x002E, 6},
+    {0x002C, 6},{0x0008, 6},{0x0006, 5},{0x0001, 5}
+  },
+  {
+    {0x0000, 3},{0x000E, 5},{0x0017, 6},{0x002A, 6},
+    {0x0010, 7},{0x00F9,10},{0x00F8,10},{0x001E, 7},
+    {0x003F, 8},{0x0007, 3},{0x0006, 3},{0x0009, 4},
+    {0x0008, 4},{0x0006, 4},{0x000F, 5},{0x0005, 5},
+    {0x0016, 6},{0x0029, 6},{0x002B, 6},{0x0015, 6},
+    {0x0050, 7},{0x0011, 7},{0x007D, 9},{0x0004, 4},
+    {0x0017, 5},{0x0006, 5},{0x0014, 6},{0x002C, 6},
+    {0x002D, 6},{0x000E, 6},{0x0009, 6},{0x0051, 7}
+  },
+  {
+    {0x0002, 3},{0x0018, 5},{0x002F, 6},{0x000D, 5},
+    {0x0053, 7},{0x0295,10},{0x0294,10},{0x00A4, 8},
+    {0x007C, 8},{0x0000, 2},{0x0007, 3},{0x0009, 4},
+    {0x0008, 4},{0x001B, 5},{0x000C, 5},{0x0028, 6},
+    {0x006A, 7},{0x001E, 6},{0x001D, 6},{0x0069, 7},
+    {0x00D7, 8},{0x007D, 8},{0x014B, 9},{0x0019, 5},
+    {0x0016, 5},{0x002E, 6},{0x001C, 6},{0x002B, 6},
+    {0x002A, 6},{0x0068, 7},{0x003F, 7},{0x00D6, 8}
+  },
+  {
+    {0x0002, 3},{0x001B, 5},{0x000C, 5},{0x0018, 5},
+    {0x0029, 6},{0x007F, 8},{0x02F0,10},{0x0198, 9},
+    {0x0179, 9},{0x0000, 2},{0x0007, 3},{0x0009, 4},
+    {0x0008, 4},{0x001A, 5},{0x000D, 5},{0x002A, 6},
+    {0x0064, 7},{0x001E, 6},{0x0067, 7},{0x005F, 7},
+    {0x00CD, 8},{0x007E, 8},{0x02F1,10},{0x0016, 5},
+    {0x000E, 5},{0x002E, 6},{0x0065, 7},{0x002B, 6},
+    {0x0028, 6},{0x003E, 7},{0x00BD, 8},{0x0199, 9}
+  },
+  {
+    {0x0002, 3},{0x0007, 4},{0x0016, 5},{0x0006, 4},
+    {0x0036, 6},{0x005C, 7},{0x015D, 9},{0x015C, 9},
+    {0x02BF,10},{0x0000, 2},{0x0007, 3},{0x0009, 4},
+    {0x0008, 4},{0x0018, 5},{0x0034, 6},{0x002A, 6},
+    {0x005E, 7},{0x006A, 7},{0x0064, 7},{0x005D, 7},
+    {0x00CB, 8},{0x00AD, 8},{0x02BE,10},{0x0014, 5},
+    {0x0033, 6},{0x006E, 7},{0x005F, 7},{0x006F, 7},
+    {0x006B, 7},{0x00CA, 8},{0x00AC, 8},{0x015E, 9}
+  },
+  {
+    {0x000F, 4},{0x001D, 5},{0x0018, 5},{0x000B, 4},
+    {0x0019, 5},{0x0029, 6},{0x00D6, 8},{0x0551,11},
+    {0x0AA1,12},{0x0001, 2},{0x0000, 2},{0x0009, 4},
+    {0x0008, 4},{0x001B, 5},{0x0038, 6},{0x0028, 6},
+    {0x0057, 7},{0x006A, 7},{0x0068, 7},{0x0056, 7},
+    {0x00E5, 8},{0x0155, 9},{0x0AA0,12},{0x0073, 7},
+    {0x0069, 7},{0x00D7, 8},{0x00AB, 8},{0x00E4, 8},
+    {0x00A9, 8},{0x0151, 9},{0x0150, 9},{0x02A9,10}
+  },
+  {
+    {0x0008, 5},{0x0025, 7},{0x017A, 9},{0x02F7,10},
+    {0x0BDB,12},{0x17B4,13},{0x2F6B,14},{0x001D, 5},
+    {0x2F6A,14},{0x0008, 4},{0x0007, 4},{0x0001, 4},
+    {0x0002, 4},{0x000A, 4},{0x0006, 4},{0x0000, 4},
+    {0x001C, 5},{0x0009, 4},{0x000D, 4},{0x000F, 4},
+    {0x000C, 4},{0x0003, 4},{0x000A, 5},{0x0016, 5},
+    {0x0013, 6},{0x005D, 7},{0x0024, 7},{0x00BC, 8},
+    {0x005C, 7},{0x05EC,11},{0x000B, 5},{0x005F, 7}
+  },
+  {
+    {0x000F, 5},{0x0010, 6},{0x004B, 8},{0x00C6, 8},
+    {0x031D,10},{0x0C71,12},{0x0C70,12},{0x0001, 4},
+    {0x0C73,12},{0x0008, 4},{0x0009, 4},{0x0002, 4},
+    {0x0003, 4},{0x000B, 4},{0x0006, 4},{0x0000, 4},
+    {0x001C, 5},{0x0005, 4},{0x000D, 4},{0x000F, 4},
+    {0x000A, 4},{0x0019, 5},{0x0013, 6},{0x001D, 5},
+    {0x0030, 6},{0x0062, 7},{0x0024, 7},{0x004A, 8},
+    {0x018F, 9},{0x0C72,12},{0x000E, 5},{0x0011, 6}
+  },
+  {
+    {0x001B, 5},{0x0003, 6},{0x008D, 8},{0x0040, 7},
+    {0x0239,10},{0x0471,11},{0x08E0,12},{0x0003, 4},
+    {0x11C3,13},{0x000A, 4},{0x0009, 4},{0x0004, 4},
+    {0x0005, 4},{0x000E, 4},{0x0007, 4},{0x0001, 4},
+    {0x001E, 5},{0x0006, 4},{0x000C, 4},{0x000B, 4},
+    {0x0002, 4},{0x0000, 5},{0x0041, 7},{0x001F, 5},
+    {0x0022, 6},{0x0002, 6},{0x008F, 8},{0x008C, 8},
+    {0x011D, 9},{0x11C2,13},{0x001A, 5},{0x0021, 6}
+  },
+  {
+    {0x001F, 5},{0x0003, 6},{0x0003, 7},{0x0043, 7},
+    {0x000B, 9},{0x0015,10},{0x0051,12},{0x0003, 4},
+    {0x0050,12},{0x000D, 4},{0x000C, 4},{0x0004, 4},
+    {0x0006, 4},{0x000E, 4},{0x000A, 4},{0x0001, 4},
+    {0x001E, 5},{0x0005, 4},{0x0009, 4},{0x0007, 4},
+    {0x0011, 5},{0x0002, 6},{0x0004, 8},{0x0002, 4},
+    {0x002D, 6},{0x0020, 6},{0x0042, 7},{0x0001, 7},
+    {0x0000, 7},{0x0029,11},{0x0017, 5},{0x002C, 6}
+  },
+  {
+    {0x0003, 4},{0x001F, 6},{0x003A, 7},{0x005D, 7},
+    {0x0173, 9},{0x02E4,10},{0x172D,13},{0x0004, 4},
+    {0x172C,13},{0x000F, 4},{0x000E, 4},{0x0009, 4},
+    {0x0008, 4},{0x000C, 4},{0x000A, 4},{0x0001, 4},
+    {0x0016, 5},{0x0002, 4},{0x0005, 4},{0x001A, 5},
+    {0x002F, 6},{0x0038, 7},{0x05CA,11},{0x0006, 4},
+    {0x0037, 6},{0x001E, 6},{0x003B, 7},{0x0039, 7},
+    {0x00B8, 8},{0x0B97,12},{0x0000, 4},{0x0036, 6}
+  },
+  {
+    {0x0006, 4},{0x0037, 6},{0x005D, 7},{0x000C, 6},
+    {0x00B9, 8},{0x02E3,10},{0x05C4,11},{0x0004, 4},
+    {0x1715,13},{0x0000, 3},{0x000F, 4},{0x0008, 4},
+    {0x0007, 4},{0x000C, 4},{0x0009, 4},{0x001D, 5},
+    {0x0016, 5},{0x001C, 5},{0x001A, 5},{0x000B, 5},
+    {0x005E, 7},{0x0170, 9},{0x1714,13},{0x000A, 4},
+    {0x000A, 5},{0x0036, 6},{0x005F, 7},{0x001B, 7},
+    {0x001A, 7},{0x0B8B,12},{0x0002, 4},{0x0007, 5}
+  },
+  {
+    {0x000C, 4},{0x000B, 5},{0x0079, 7},{0x0022, 6},
+    {0x00F0, 8},{0x0119, 9},{0x0230,10},{0x001D, 5},
+    {0x08C4,12},{0x0001, 3},{0x0000, 3},{0x000A, 4},
+    {0x0009, 4},{0x000B, 4},{0x0007, 4},{0x001C, 5},
+    {0x003D, 6},{0x000D, 5},{0x0008, 5},{0x0015, 6},
+    {0x008D, 8},{0x118B,13},{0x118A,13},{0x000D, 4},
+    {0x0010, 5},{0x0009, 5},{0x0014, 6},{0x0047, 7},
+    {0x00F1, 8},{0x0463,11},{0x001F, 5},{0x000C, 5}
+  },
+  {
+    {0x0000, 3},{0x001A, 5},{0x0033, 6},{0x000C, 5},
+    {0x0046, 7},{0x01E3, 9},{0x03C5,10},{0x0017, 5},
+    {0x1E21,13},{0x0002, 3},{0x0001, 3},{0x0009, 4},
+    {0x000A, 4},{0x0007, 4},{0x001B, 5},{0x003D, 6},
+    {0x001B, 6},{0x0022, 6},{0x0079, 7},{0x00F0, 8},
+    {0x1E20,13},{0x1E23,13},{0x1E22,13},{0x000E, 4},
+    {0x0016, 5},{0x0018, 5},{0x0032, 6},{0x001A, 6},
+    {0x0047, 7},{0x0789,11},{0x001F, 5},{0x0010, 5}
+  },
+  {
+    {0x001D, 5},{0x0061, 7},{0x004E, 8},{0x009E, 9},
+    {0x027C,11},{0x09F5,13},{0x09F4,13},{0x0003, 4},
+    {0x0060, 7},{0x0000, 3},{0x000F, 4},{0x000B, 4},
+    {0x000A, 4},{0x0009, 4},{0x0005, 4},{0x000D, 5},
+    {0x0031, 6},{0x0008, 5},{0x0038, 6},{0x0012, 6},
+    {0x0026, 7},{0x013F,10},{0x04FB,12},{0x000D, 4},
+    {0x0002, 4},{0x000C, 5},{0x0039, 6},{0x001C, 6},
+    {0x000F, 5},{0x001D, 6},{0x0008, 4},{0x0019, 5}
+  },
+  {
+    {0x0007, 4},{0x0019, 6},{0x00AB, 8},{0x00AA, 8},
+    {0x0119,10},{0x0461,12},{0x0460,12},{0x001B, 5},
+    {0x0047, 8},{0x0001, 3},{0x0000, 3},{0x000C, 4},
+    {0x000B, 4},{0x0009, 4},{0x0005, 4},{0x000D, 5},
+    {0x0035, 6},{0x003D, 6},{0x003C, 6},{0x0018, 6},
+    {0x0022, 7},{0x008D, 9},{0x0231,11},{0x000E, 4},
+    {0x001F, 5},{0x0009, 5},{0x002B, 6},{0x0010, 6},
+    {0x0034, 6},{0x0054, 7},{0x0008, 4},{0x0014, 5}
+  },
+  {
+    {0x000C, 4},{0x0005, 5},{0x0008, 6},{0x005B, 7},
+    {0x004D, 9},{0x0131,11},{0x0261,12},{0x001A, 5},
+    {0x0012, 7},{0x0000, 3},{0x000F, 4},{0x000A, 4},
+    {0x0009, 4},{0x0006, 4},{0x001B, 5},{0x0006, 5},
+    {0x001C, 6},{0x002C, 6},{0x0015, 6},{0x005A, 7},
+    {0x0027, 8},{0x0099,10},{0x0260,12},{0x000E, 4},
+    {0x0004, 4},{0x000F, 5},{0x0007, 5},{0x001D, 6},
+    {0x000B, 5},{0x0014, 6},{0x0008, 4},{0x0017, 5}
+  },
+  {
+    {0x000F, 4},{0x0013, 5},{0x0075, 7},{0x0024, 6},
+    {0x0095, 8},{0x0251,10},{0x04A0,11},{0x0010, 5},
+    {0x00C8, 8},{0x0002, 3},{0x0001, 3},{0x0001, 4},
+    {0x0000, 4},{0x001A, 5},{0x0011, 5},{0x002C, 6},
+    {0x0065, 7},{0x0074, 7},{0x004B, 7},{0x00C9, 8},
+    {0x0129, 9},{0x0943,12},{0x0942,12},{0x0003, 3},
+    {0x000A, 4},{0x001C, 5},{0x0018, 5},{0x0033, 6},
+    {0x0017, 5},{0x002D, 6},{0x001B, 5},{0x003B, 6}
+  },
+  {
+    {0x0003, 3},{0x001A, 5},{0x002D, 6},{0x0038, 6},
+    {0x0028, 7},{0x0395,10},{0x0E51,12},{0x0037, 6},
+    {0x00E4, 8},{0x0001, 3},{0x0000, 3},{0x001F, 5},
+    {0x001E, 5},{0x0017, 5},{0x003A, 6},{0x0073, 7},
+    {0x002A, 7},{0x002B, 7},{0x0029, 7},{0x01CB, 9},
+    {0x0729,11},{0x1CA1,13},{0x1CA0,13},{0x0004, 3},
+    {0x000A, 4},{0x0004, 4},{0x0018, 5},{0x0036, 6},
+    {0x000B, 5},{0x002C, 6},{0x0019, 5},{0x003B, 6}
+  },
+  {
+    {0x0004, 3},{0x0004, 4},{0x003F, 6},{0x0017, 5},
+    {0x0075, 7},{0x01F5, 9},{0x07D1,11},{0x0017, 6},
+    {0x01F6, 9},{0x0001, 3},{0x0000, 3},{0x001B, 5},
+    {0x001A, 5},{0x000A, 5},{0x0032, 6},{0x0074, 7},
+    {0x00F8, 8},{0x00F9, 8},{0x01F7, 9},{0x03E9,10},
+    {0x0FA0,12},{0x1F43,13},{0x1F42,13},{0x0003, 3},
+    {0x000A, 4},{0x001E, 5},{0x001C, 5},{0x003B, 6},
+    {0x0018, 5},{0x0016, 6},{0x0016, 5},{0x0033, 6}
+  },
+  {
+    {0x0004, 3},{0x0007, 4},{0x0018, 5},{0x001E, 5},
+    {0x0036, 6},{0x0031, 7},{0x0177, 9},{0x0077, 7},
+    {0x0176, 9},{0x0001, 3},{0x0000, 3},{0x001A, 5},
+    {0x0019, 5},{0x003A, 6},{0x0019, 6},{0x005C, 7},
+    {0x00BA, 8},{0x0061, 8},{0x00C1, 9},{0x0180,10},
+    {0x0302,11},{0x0607,12},{0x0606,12},{0x0002, 3},
+    {0x000A, 4},{0x001F, 5},{0x001C, 5},{0x0037, 6},
+    {0x0016, 5},{0x0076, 7},{0x000D, 5},{0x002F, 6}
+  },
+  {
+    {0x0000, 3},{0x000A, 4},{0x001A, 5},{0x000C, 4},
+    {0x001D, 5},{0x0039, 6},{0x0078, 7},{0x005E, 7},
+    {0x0393,11},{0x0002, 3},{0x0001, 3},{0x0016, 5},
+    {0x000F, 5},{0x002E, 6},{0x005F, 7},{0x0073, 8},
+    {0x00E5, 9},{0x01C8,10},{0x0E4A,13},{0x1C97,14},
+    {0x1C96,14},{0x0E49,13},{0x0E48,13},{0x0004, 3},
+    {0x0006, 4},{0x001F, 5},{0x001B, 5},{0x001D, 6},
+    {0x0038, 6},{0x0038, 7},{0x003D, 6},{0x0079, 7}
+  },
+  {
+    {0x000B, 5},{0x002B, 7},{0x0054, 8},{0x01B7, 9},
+    {0x06D9,11},{0x0DB1,12},{0x0DB0,12},{0x0002, 4},
+    {0x00AB, 9},{0x0009, 4},{0x000A, 4},{0x0007, 4},
+    {0x0008, 4},{0x000F, 4},{0x000C, 4},{0x0003, 4},
+    {0x001D, 5},{0x0004, 4},{0x000B, 4},{0x0006, 4},
+    {0x001A, 5},{0x0003, 6},{0x00AA, 9},{0x0001, 4},
+    {0x0000, 5},{0x0014, 6},{0x006C, 7},{0x00DA, 8},
+    {0x0002, 6},{0x036D,10},{0x001C, 5},{0x0037, 6}
+  },
+  {
+    {0x001D, 5},{0x0004, 6},{0x00B6, 8},{0x006A, 8},
+    {0x05B9,11},{0x16E1,13},{0x16E0,13},{0x0007, 4},
+    {0x016F, 9},{0x000C, 4},{0x000D, 4},{0x0009, 4},
+    {0x0008, 4},{0x000F, 4},{0x000A, 4},{0x0003, 4},
+    {0x0017, 5},{0x0002, 4},{0x0004, 4},{0x001C, 5},
+    {0x002C, 6},{0x006B, 8},{0x0B71,12},{0x0005, 4},
+    {0x0003, 5},{0x001B, 6},{0x005A, 7},{0x0034, 7},
+    {0x0005, 6},{0x02DD,10},{0x0000, 4},{0x000C, 5}
+  },
+  {
+    {0x0003, 4},{0x007F, 7},{0x00A1, 8},{0x00A0, 8},
+    {0x020C,10},{0x0834,12},{0x106B,13},{0x0007, 4},
+    {0x0082, 8},{0x000E, 4},{0x000D, 4},{0x000B, 4},
+    {0x000C, 4},{0x0000, 3},{0x0009, 4},{0x0002, 4},
+    {0x0011, 5},{0x001E, 5},{0x0015, 5},{0x003E, 6},
+    {0x0040, 7},{0x041B,11},{0x106A,13},{0x0006, 4},
+    {0x000A, 5},{0x0029, 6},{0x007E, 7},{0x0051, 7},
+    {0x0021, 6},{0x0107, 9},{0x0004, 4},{0x000B, 5}
+  },
+  {
+    {0x0007, 4},{0x001B, 6},{0x00F6, 8},{0x00E9, 8},
+    {0x03A1,10},{0x0740,11},{0x0E82,12},{0x001F, 5},
+    {0x01EF, 9},{0x0001, 3},{0x0002, 3},{0x000B, 4},
+    {0x000C, 4},{0x000D, 4},{0x0008, 4},{0x001C, 5},
+    {0x0003, 5},{0x0012, 5},{0x0002, 5},{0x0075, 7},
+    {0x01D1, 9},{0x1D07,13},{0x1D06,13},{0x000A, 4},
+    {0x0013, 5},{0x003B, 6},{0x001A, 6},{0x007A, 7},
+    {0x003C, 6},{0x01EE, 9},{0x0000, 4},{0x000C, 5}
+  },
+  {
+    {0x000D, 4},{0x003D, 6},{0x0042, 7},{0x0037, 7},
+    {0x00D9, 9},{0x0362,11},{0x06C6,12},{0x001F, 5},
+    {0x0086, 8},{0x0001, 3},{0x0002, 3},{0x000C, 4},
+    {0x000B, 4},{0x000A, 4},{0x0001, 4},{0x000F, 5},
+    {0x0025, 6},{0x003C, 6},{0x001A, 6},{0x0087, 8},
+    {0x01B0,10},{0x0D8F,13},{0x0D8E,13},{0x000E, 4},
+    {0x0013, 5},{0x000C, 5},{0x0024, 6},{0x0020, 6},
+    {0x0011, 5},{0x006D, 8},{0x0000, 4},{0x000E, 5}
+  },
+  {
+    {0x0000, 3},{0x0012, 5},{0x0076, 7},{0x0077, 7},
+    {0x014D, 9},{0x0533,11},{0x14C9,13},{0x0013, 5},
+    {0x00A5, 8},{0x0002, 3},{0x0003, 3},{0x000B, 4},
+    {0x000C, 4},{0x0008, 4},{0x001A, 5},{0x002B, 6},
+    {0x0075, 7},{0x0074, 7},{0x00A7, 8},{0x0298,10},
+    {0x14C8,13},{0x14CB,13},{0x14CA,13},{0x000F, 4},
+    {0x001C, 5},{0x0007, 5},{0x002A, 6},{0x0028, 6},
+    {0x001B, 5},{0x00A4, 8},{0x0002, 4},{0x0006, 5}
+  },
+  {
+    {0x0002, 3},{0x001A, 5},{0x002B, 6},{0x003A, 6},
+    {0x00ED, 8},{0x0283,10},{0x0A0A,12},{0x0004, 5},
+    {0x00A1, 8},{0x0004, 3},{0x0003, 3},{0x000B, 4},
+    {0x000C, 4},{0x001F, 5},{0x0006, 5},{0x0077, 7},
+    {0x00A3, 8},{0x00A2, 8},{0x0140, 9},{0x1417,13},
+    {0x1416,13},{0x0A09,12},{0x0A08,12},{0x0000, 3},
+    {0x001E, 5},{0x0007, 5},{0x002A, 6},{0x0029, 6},
+    {0x001C, 5},{0x00EC, 8},{0x001B, 5},{0x0005, 5}
+  },
+  {
+    {0x0002, 3},{0x0002, 4},{0x0018, 5},{0x001D, 5},
+    {0x0035, 6},{0x00E4, 8},{0x01CF,11},{0x001D, 7},
+    {0x0072, 9},{0x0004, 3},{0x0005, 3},{0x0006, 4},
+    {0x0007, 4},{0x0006, 5},{0x0073, 7},{0x0038, 8},
+    {0x01CE,11},{0x039B,12},{0x0398,12},{0x0733,13},
+    {0x0732,13},{0x0735,13},{0x0734,13},{0x0000, 3},
+    {0x001F, 5},{0x001B, 5},{0x0034, 6},{0x000F, 6},
+    {0x001E, 5},{0x00E5, 8},{0x0019, 5},{0x0038, 6}
+  },
+  {
+    {0x0016, 5},{0x0050, 7},{0x0172, 9},{0x02E7,10},
+    {0x1732,13},{0x2E67,14},{0x2E66,14},{0x0006, 4},
+    {0x0051, 7},{0x0001, 3},{0x0000, 3},{0x000D, 4},
+    {0x000C, 4},{0x0009, 4},{0x001C, 5},{0x0009, 5},
+    {0x001C, 6},{0x001D, 6},{0x005D, 7},{0x00B8, 8},
+    {0x05CD,11},{0x1731,13},{0x1730,13},{0x000F, 4},
+    {0x0005, 4},{0x000F, 5},{0x0008, 5},{0x0029, 6},
+    {0x001D, 5},{0x002F, 6},{0x0008, 4},{0x0015, 5}
+  },
+  {
+    {0x0009, 4},{0x0021, 6},{0x0040, 7},{0x00AD, 8},
+    {0x02B0,10},{0x1589,13},{0x1588,13},{0x001C, 5},
+    {0x005F, 7},{0x0000, 3},{0x000F, 4},{0x000D, 4},
+    {0x000C, 4},{0x0006, 4},{0x0011, 5},{0x002A, 6},
+    {0x0057, 7},{0x005E, 7},{0x0041, 7},{0x0159, 9},
+    {0x0563,11},{0x158B,13},{0x158A,13},{0x0001, 3},
+    {0x0005, 4},{0x0014, 5},{0x003B, 6},{0x002E, 6},
+    {0x0004, 4},{0x003A, 6},{0x0007, 4},{0x0016, 5}
+  },
+  {
+    {0x000E, 4},{0x0007, 5},{0x0046, 7},{0x0045, 7},
+    {0x0064, 9},{0x032A,12},{0x0657,13},{0x0018, 5},
+    {0x000D, 6},{0x0000, 3},{0x000F, 4},{0x000A, 4},
+    {0x000B, 4},{0x001A, 5},{0x0036, 6},{0x0047, 7},
+    {0x0044, 7},{0x0018, 7},{0x0033, 8},{0x00CB,10},
+    {0x0656,13},{0x0329,12},{0x0328,12},{0x0002, 3},
+    {0x0006, 4},{0x0019, 5},{0x000E, 5},{0x0037, 6},
+    {0x0009, 4},{0x000F, 5},{0x0002, 4},{0x0010, 5}
+  },
+  {
+    {0x0003, 3},{0x0018, 5},{0x0023, 6},{0x0077, 7},
+    {0x0194, 9},{0x1956,13},{0x32AF,14},{0x003A, 6},
+    {0x0076, 7},{0x0002, 3},{0x0001, 3},{0x001F, 5},
+    {0x001E, 5},{0x0014, 5},{0x0022, 6},{0x0064, 7},
+    {0x0197, 9},{0x0196, 9},{0x032B,10},{0x0654,11},
+    {0x32AE,14},{0x1955,13},{0x1954,13},{0x0000, 3},
+    {0x0009, 4},{0x001C, 5},{0x0015, 5},{0x0010, 5},
+    {0x000D, 4},{0x0017, 5},{0x0016, 5},{0x0033, 6}
+  },
+  {
+    {0x0005, 3},{0x0006, 4},{0x003E, 6},{0x0010, 5},
+    {0x0048, 7},{0x093F,12},{0x24FA,14},{0x0032, 6},
+    {0x0067, 7},{0x0002, 3},{0x0001, 3},{0x001B, 5},
+    {0x001E, 5},{0x0034, 6},{0x0066, 7},{0x0092, 8},
+    {0x0126, 9},{0x024E,10},{0x049E,11},{0x49F7,15},
+    {0x49F6,15},{0x24F9,14},{0x24F8,14},{0x0000, 3},
+    {0x0007, 4},{0x0018, 5},{0x0011, 5},{0x003F, 6},
+    {0x000E, 4},{0x0013, 5},{0x0035, 6},{0x0025, 6}
+  },
+  {
+    {0x0005, 3},{0x0008, 4},{0x0012, 5},{0x001C, 5},
+    {0x001C, 6},{0x00EA, 9},{0x1D75,14},{0x001E, 6},
+    {0x0066, 7},{0x0001, 3},{0x0002, 3},{0x001B, 5},
+    {0x001A, 5},{0x001F, 6},{0x003B, 7},{0x0074, 8},
+    {0x01D6,10},{0x03AF,11},{0x1D74,14},{0x1D77,14},
+    {0x1D76,14},{0x0EB9,13},{0x0EB8,13},{0x000F, 4},
+    {0x0006, 4},{0x0013, 5},{0x003B, 6},{0x003A, 6},
+    {0x0000, 3},{0x0018, 5},{0x0032, 6},{0x0067, 7}
+  },
+  {
+    {0x0004, 3},{0x000A, 4},{0x001B, 5},{0x000C, 4},
+    {0x000D, 5},{0x00E6, 8},{0x0684,11},{0x0072, 7},
+    {0x00E7, 8},{0x0002, 3},{0x0001, 3},{0x0017, 5},
+    {0x0016, 5},{0x0018, 6},{0x00D1, 8},{0x01A0, 9},
+    {0x0686,11},{0x0D0F,12},{0x0D0A,12},{0x1A17,13},
+    {0x1A16,13},{0x1A1D,13},{0x1A1C,13},{0x000F, 4},
+    {0x001D, 5},{0x000E, 5},{0x0035, 6},{0x0038, 6},
+    {0x0000, 3},{0x000F, 5},{0x0019, 6},{0x0069, 7}
+  },
+  {
+    {0x0003, 3},{0x000C, 4},{0x001B, 5},{0x0000, 3},
+    {0x0003, 4},{0x002E, 6},{0x0051, 9},{0x00BC, 8},
+    {0x0053, 9},{0x0004, 3},{0x0002, 3},{0x0016, 5},
+    {0x0015, 5},{0x0015, 7},{0x0050, 9},{0x00A4,10},
+    {0x0294,12},{0x052B,13},{0x052A,13},{0x052D,13},
+    {0x052C,13},{0x052F,13},{0x052E,13},{0x000E, 4},
+    {0x001A, 5},{0x0004, 5},{0x0028, 6},{0x0029, 6},
+    {0x000F, 4},{0x000B, 6},{0x005F, 7},{0x00BD, 8}
+  },
+  {
+    {0x0003, 4},{0x0009, 6},{0x00D0, 8},{0x01A3, 9},
+    {0x0344,10},{0x0D14,12},{0x1A2B,13},{0x0004, 4},
+    {0x0015, 7},{0x0000, 3},{0x000F, 4},{0x000B, 4},
+    {0x000C, 4},{0x000E, 4},{0x0009, 4},{0x001B, 5},
+    {0x000A, 5},{0x0014, 5},{0x000D, 5},{0x002A, 6},
+    {0x0014, 7},{0x068B,11},{0x1A2A,13},{0x0008, 4},
+    {0x000B, 5},{0x002B, 6},{0x000B, 6},{0x0069, 7},
+    {0x0035, 6},{0x0008, 6},{0x0007, 4},{0x000C, 5}
+  },
+  {
+    {0x000A, 4},{0x003C, 6},{0x0032, 7},{0x0030, 7},
+    {0x00C5, 9},{0x0621,12},{0x0620,12},{0x001F, 5},
+    {0x0033, 7},{0x0001, 3},{0x0000, 3},{0x000E, 4},
+    {0x000D, 4},{0x000C, 4},{0x0004, 4},{0x000D, 5},
+    {0x0026, 6},{0x0027, 6},{0x0014, 6},{0x0063, 8},
+    {0x0189,10},{0x0623,12},{0x0622,12},{0x000B, 4},
+    {0x0012, 5},{0x003D, 6},{0x0022, 6},{0x0015, 6},
+    {0x000B, 5},{0x0023, 6},{0x0007, 4},{0x0010, 5}
+  },
+  {
+    {0x000F, 4},{0x000C, 5},{0x0043, 7},{0x0010, 6},
+    {0x0044, 8},{0x0114,10},{0x0455,12},{0x0018, 5},
+    {0x0023, 7},{0x0001, 3},{0x0000, 3},{0x000E, 4},
+    {0x000D, 4},{0x0009, 4},{0x0019, 5},{0x0009, 5},
+    {0x0017, 6},{0x0016, 6},{0x0042, 7},{0x008B, 9},
+    {0x0454,12},{0x0457,12},{0x0456,12},{0x000B, 4},
+    {0x0015, 5},{0x000A, 5},{0x0029, 6},{0x0020, 6},
+    {0x000D, 5},{0x0028, 6},{0x0007, 4},{0x0011, 5}
+  },
+  {
+    {0x0001, 3},{0x001A, 5},{0x0029, 6},{0x002A, 6},
+    {0x00A0, 8},{0x0285,10},{0x1425,13},{0x0002, 5},
+    {0x0000, 7},{0x0002, 3},{0x0003, 3},{0x000C, 4},
+    {0x000B, 4},{0x0008, 4},{0x0012, 5},{0x0001, 6},
+    {0x0051, 7},{0x0001, 7},{0x0143, 9},{0x0508,11},
+    {0x1424,13},{0x1427,13},{0x1426,13},{0x000F, 4},
+    {0x001C, 5},{0x0003, 5},{0x0037, 6},{0x002B, 6},
+    {0x0013, 5},{0x0036, 6},{0x001D, 5},{0x0001, 5}
+  },
+  {
+    {0x0004, 3},{0x001F, 5},{0x003D, 6},{0x0006, 5},
+    {0x0016, 7},{0x0053, 9},{0x014A,11},{0x0034, 6},
+    {0x002A, 8},{0x0002, 3},{0x0003, 3},{0x000B, 4},
+    {0x000C, 4},{0x001C, 5},{0x0037, 6},{0x0017, 7},
+    {0x002B, 8},{0x0028, 8},{0x00A4,10},{0x052D,13},
+    {0x052C,13},{0x052F,13},{0x052E,13},{0x0000, 3},
+    {0x001D, 5},{0x0007, 5},{0x0004, 5},{0x0035, 6},
+    {0x0014, 5},{0x0036, 6},{0x0015, 5},{0x003C, 6}
+  },
+  {
+    {0x0004, 3},{0x000A, 4},{0x0007, 5},{0x001D, 5},
+    {0x0009, 6},{0x01F3, 9},{0x07C7,11},{0x0008, 6},
+    {0x01F0, 9},{0x0003, 3},{0x0002, 3},{0x000D, 4},
+    {0x000C, 4},{0x0017, 5},{0x007D, 7},{0x01F2, 9},
+    {0x07C6,11},{0x07C5,11},{0x1F12,13},{0x3E27,14},
+    {0x3E26,14},{0x1F11,13},{0x1F10,13},{0x0000, 3},
+    {0x001E, 5},{0x0006, 5},{0x0039, 6},{0x0038, 6},
+    {0x003F, 6},{0x002C, 6},{0x0005, 5},{0x002D, 6}
+  },
+  {
+    {0x0002, 3},{0x0007, 4},{0x0018, 5},{0x0003, 4},
+    {0x0005, 5},{0x0035, 7},{0x004F, 9},{0x0012, 7},
+    {0x04E5,13},{0x0005, 3},{0x0004, 3},{0x000D, 4},
+    {0x000E, 4},{0x0033, 6},{0x0026, 8},{0x009D,10},
+    {0x04E4,13},{0x04E7,13},{0x04E6,13},{0x04E1,13},
+    {0x04E0,13},{0x04E3,13},{0x04E2,13},{0x0000, 3},
+    {0x001F, 5},{0x000C, 5},{0x003D, 6},{0x003C, 6},
+    {0x0032, 6},{0x0034, 7},{0x001B, 6},{0x0008, 6}
+  },
+  {
+    {0x0000, 3},{0x0004, 4},{0x001C, 5},{0x000F, 4},
+    {0x0002, 4},{0x0007, 5},{0x0075, 7},{0x00E8, 8},
+    {0x1D2A,13},{0x0005, 3},{0x0004, 3},{0x000D, 4},
+    {0x000C, 4},{0x0077, 7},{0x0E96,12},{0x3A57,14},
+    {0x3A56,14},{0x3A5D,14},{0x3A5C,14},{0x3A5F,14},
+    {0x3A5E,14},{0x1D29,13},{0x1D28,13},{0x0003, 3},
+    {0x0006, 5},{0x000A, 5},{0x002C, 7},{0x0017, 6},
+    {0x0076, 7},{0x01D3, 9},{0x03A4,10},{0x002D, 7}
+  },
+  {
+    {0x000A, 4},{0x0024, 6},{0x00BF, 8},{0x0085, 8},
+    {0x0211,10},{0x0842,12},{0x1087,13},{0x0018, 5},
+    {0x0020, 6},{0x0001, 3},{0x0002, 3},{0x000E, 4},
+    {0x000D, 4},{0x0007, 4},{0x0013, 5},{0x0025, 6},
+    {0x005E, 7},{0x0043, 7},{0x00BE, 8},{0x0109, 9},
+    {0x1086,13},{0x0841,12},{0x0840,12},{0x000F, 4},
+    {0x0001, 4},{0x0011, 5},{0x0000, 5},{0x002E, 6},
+    {0x0019, 5},{0x0001, 5},{0x0006, 4},{0x0016, 5}
+  },
+  {
+    {0x0002, 3},{0x000F, 5},{0x006F, 7},{0x0061, 7},
+    {0x0374,10},{0x1BA8,13},{0x3753,14},{0x0012, 5},
+    {0x0036, 6},{0x0000, 3},{0x0001, 3},{0x000A, 4},
+    {0x000B, 4},{0x001A, 5},{0x0031, 6},{0x0060, 7},
+    {0x00DC, 8},{0x01BB, 9},{0x06EB,11},{0x1BAB,13},
+    {0x3752,14},{0x3755,14},{0x3754,14},{0x000E, 4},
+    {0x0006, 4},{0x0013, 5},{0x000E, 5},{0x003E, 6},
+    {0x0008, 4},{0x001E, 5},{0x0019, 5},{0x003F, 6}
+  },
+  {
+    {0x0003, 3},{0x001C, 5},{0x0025, 6},{0x0024, 6},
+    {0x01DA, 9},{0x1DBD,13},{0x3B7C,14},{0x003C, 6},
+    {0x003D, 6},{0x0000, 3},{0x0001, 3},{0x000B, 4},
+    {0x000A, 4},{0x000B, 5},{0x0077, 7},{0x00EC, 8},
+    {0x03B6,10},{0x076E,11},{0x1DBF,13},{0x76FB,15},
+    {0x76FA,15},{0x3B79,14},{0x3B78,14},{0x000D, 4},
+    {0x001F, 5},{0x0013, 5},{0x000A, 5},{0x0008, 5},
+    {0x000C, 4},{0x0008, 4},{0x0009, 5},{0x003A, 6}
+  },
+  {
+    {0x0005, 3},{0x0003, 4},{0x0004, 5},{0x0010, 5},
+    {0x008F, 8},{0x0475,11},{0x11D1,13},{0x0079, 7},
+    {0x0027, 6},{0x0002, 3},{0x0003, 3},{0x0001, 4},
+    {0x0000, 4},{0x0026, 6},{0x0046, 7},{0x011C, 9},
+    {0x0477,11},{0x08ED,12},{0x11D0,13},{0x11D3,13},
+    {0x11D2,13},{0x11D9,13},{0x11D8,13},{0x000D, 4},
+    {0x001F, 5},{0x0012, 5},{0x0005, 5},{0x003D, 6},
+    {0x000C, 4},{0x000E, 4},{0x0022, 6},{0x0078, 7}
+  },
+  {
+    {0x0005, 3},{0x000C, 4},{0x001B, 5},{0x0000, 4},
+    {0x0006, 6},{0x03E2,10},{0x3E3D,14},{0x000F, 7},
+    {0x0034, 6},{0x0003, 3},{0x0002, 3},{0x001E, 5},
+    {0x001D, 5},{0x007D, 7},{0x01F0, 9},{0x07C6,11},
+    {0x3E3C,14},{0x3E3F,14},{0x3E3E,14},{0x3E39,14},
+    {0x3E38,14},{0x3E3B,14},{0x3E3A,14},{0x0008, 4},
+    {0x001C, 5},{0x0002, 5},{0x003F, 6},{0x0035, 6},
+    {0x0009, 4},{0x0001, 3},{0x000E, 7},{0x00F9, 8}
+  },
+  {
+    {0x0004, 3},{0x000B, 4},{0x0001, 4},{0x000A, 4},
+    {0x001E, 6},{0x00E0, 9},{0x0E1E,13},{0x0071, 8},
+    {0x0039, 7},{0x0007, 3},{0x0006, 3},{0x000D, 5},
+    {0x000C, 5},{0x0020, 7},{0x01C2,10},{0x1C3F,14},
+    {0x1C3E,14},{0x0E19,13},{0x0E18,13},{0x0E1B,13},
+    {0x0E1A,13},{0x0E1D,13},{0x0E1C,13},{0x0000, 4},
+    {0x0009, 5},{0x001D, 6},{0x001F, 6},{0x0011, 6},
+    {0x0005, 4},{0x0001, 3},{0x0043, 8},{0x0042, 8}
+  },
+  {
+    {0x0004, 3},{0x000D, 4},{0x0007, 4},{0x0002, 3},
+    {0x0014, 5},{0x016C, 9},{0x16D1,13},{0x02DF,10},
+    {0x016E, 9},{0x0000, 2},{0x0007, 3},{0x002C, 6},
+    {0x002B, 6},{0x02DE,10},{0x16D0,13},{0x16D3,13},
+    {0x16D2,13},{0x2DB5,14},{0x2DB4,14},{0x2DB7,14},
+    {0x2DB6,14},{0x16D9,13},{0x16D8,13},{0x000C, 5},
+    {0x002A, 6},{0x005A, 7},{0x001B, 6},{0x001A, 6},
+    {0x0017, 5},{0x000C, 4},{0x05B7,11},{0x05B5,11}
+  },
+  {
+    {0x0002, 2},{0x000F, 4},{0x001C, 5},{0x000C, 4},
+    {0x003B, 6},{0x01AC, 9},{0x1AD8,13},{0x35B3,14},
+    {0x35B2,14},{0x0001, 2},{0x0000, 2},{0x0069, 7},
+    {0x0068, 7},{0x35BD,14},{0x35BC,14},{0x35BF,14},
+    {0x35BE,14},{0x35B9,14},{0x35B8,14},{0x35BB,14},
+    {0x35BA,14},{0x35B5,14},{0x35B4,14},{0x01A9, 9},
+    {0x01A8, 9},{0x035A,10},{0x00D7, 8},{0x00D5, 8},
+    {0x003A, 6},{0x001B, 5},{0x35B7,14},{0x35B6,14}
+  },
+  {
+    {0x0000, 3},{0x0010, 5},{0x0072, 7},{0x0071, 7},
+    {0x0154, 9},{0x0AAB,12},{0x0AA8,12},{0x0014, 5},
+    {0x0070, 7},{0x0002, 3},{0x0003, 3},{0x000C, 4},
+    {0x000B, 4},{0x0003, 4},{0x0011, 5},{0x0073, 7},
+    {0x0054, 7},{0x00AB, 8},{0x02AB,10},{0x1553,13},
+    {0x1552,13},{0x1555,13},{0x1554,13},{0x000D, 4},
+    {0x001E, 5},{0x0012, 5},{0x003E, 6},{0x002B, 6},
+    {0x0002, 4},{0x003F, 6},{0x001D, 5},{0x0013, 5}
+  },
+  {
+    {0x0003, 3},{0x001F, 5},{0x0029, 6},{0x003D, 6},
+    {0x000C, 7},{0x0069,10},{0x0345,13},{0x0002, 5},
+    {0x0028, 6},{0x0002, 3},{0x0001, 3},{0x000E, 4},
+    {0x000C, 4},{0x0015, 5},{0x0007, 6},{0x001B, 8},
+    {0x006B,10},{0x006A,10},{0x0344,13},{0x0347,13},
+    {0x0346,13},{0x01A1,12},{0x01A0,12},{0x000B, 4},
+    {0x001A, 5},{0x0012, 5},{0x0000, 5},{0x003C, 6},
+    {0x0008, 4},{0x001B, 5},{0x0013, 5},{0x0001, 5}
+  },
+  {
+    {0x0004, 3},{0x0004, 4},{0x003F, 6},{0x0014, 5},
+    {0x0056, 7},{0x015C, 9},{0x15D5,13},{0x003C, 6},
+    {0x002A, 6},{0x0000, 3},{0x0001, 3},{0x000E, 4},
+    {0x000D, 4},{0x000C, 5},{0x00AF, 8},{0x02BB,10},
+    {0x15D4,13},{0x15D7,13},{0x15D6,13},{0x15D1,13},
+    {0x15D0,13},{0x15D3,13},{0x15D2,13},{0x000B, 4},
+    {0x0019, 5},{0x000D, 5},{0x003E, 6},{0x0031, 6},
+    {0x0007, 4},{0x0005, 4},{0x003D, 6},{0x0030, 6}
+  },
+  {
+    {0x0005, 3},{0x0008, 4},{0x001A, 5},{0x0000, 4},
+    {0x0036, 6},{0x0011, 8},{0x0106,12},{0x000A, 7},
+    {0x006E, 7},{0x0002, 3},{0x0003, 3},{0x0003, 4},
+    {0x0002, 4},{0x006F, 7},{0x0021, 9},{0x020F,13},
+    {0x020E,13},{0x0101,12},{0x0100,12},{0x0103,12},
+    {0x0102,12},{0x0105,12},{0x0104,12},{0x000C, 4},
+    {0x001E, 5},{0x0003, 5},{0x003E, 6},{0x003F, 6},
+    {0x0009, 4},{0x000E, 4},{0x000B, 7},{0x0009, 7}
+  },
+  {
+    {0x0002, 3},{0x000E, 4},{0x001E, 5},{0x000C, 4},
+    {0x001F, 5},{0x006E, 7},{0x00AD,10},{0x00AF,10},
+    {0x0014, 7},{0x0004, 3},{0x0003, 3},{0x001A, 5},
+    {0x0017, 5},{0x002A, 8},{0x0576,13},{0x0AEF,14},
+    {0x0AEE,14},{0x0571,13},{0x0570,13},{0x0573,13},
+    {0x0572,13},{0x0575,13},{0x0574,13},{0x0003, 4},
+    {0x0016, 5},{0x0004, 5},{0x0036, 6},{0x000B, 6},
+    {0x000A, 4},{0x0000, 3},{0x006F, 7},{0x00AC,10}
+  },
+  {
+    {0x0004, 3},{0x0005, 4},{0x0003, 3},{0x0001, 3},
+    {0x0004, 4},{0x002F, 6},{0x0526,11},{0x1495,13},
+    {0x00A6, 8},{0x0007, 3},{0x0006, 3},{0x002D, 6},
+    {0x002C, 6},{0x1494,13},{0x1497,13},{0x1496,13},
+    {0x1491,13},{0x1490,13},{0x1493,13},{0x1492,13},
+    {0x293D,14},{0x293C,14},{0x293F,14},{0x0000, 3},
+    {0x0028, 6},{0x00A5, 8},{0x0148, 9},{0x00A7, 8},
+    {0x002E, 6},{0x0015, 5},{0x0A4E,12},{0x293E,14}
+  },
+  {
+    {0x0004, 3},{0x0005, 4},{0x0003, 3},{0x0001, 3},
+    {0x0004, 4},{0x002F, 6},{0x0526,11},{0x1495,13},
+    {0x00A6, 8},{0x0007, 3},{0x0006, 3},{0x002D, 6},
+    {0x002C, 6},{0x1494,13},{0x1497,13},{0x1496,13},
+    {0x1491,13},{0x1490,13},{0x1493,13},{0x1492,13},
+    {0x293D,14},{0x293C,14},{0x293F,14},{0x0000, 3},
+    {0x0028, 6},{0x00A5, 8},{0x0148, 9},{0x00A7, 8},
+    {0x002E, 6},{0x0015, 5},{0x0A4E,12},{0x293E,14}
+  },
+  {
+    {0x0004, 3},{0x0005, 4},{0x0003, 3},{0x0001, 3},
+    {0x0004, 4},{0x002F, 6},{0x0526,11},{0x1495,13},
+    {0x00A6, 8},{0x0007, 3},{0x0006, 3},{0x002D, 6},
+    {0x002C, 6},{0x1494,13},{0x1497,13},{0x1496,13},
+    {0x1491,13},{0x1490,13},{0x1493,13},{0x1492,13},
+    {0x293D,14},{0x293C,14},{0x293F,14},{0x0000, 3},
+    {0x0028, 6},{0x00A5, 8},{0x0148, 9},{0x00A7, 8},
+    {0x002E, 6},{0x0015, 5},{0x0A4E,12},{0x293E,14}
+  },
+  {
+    {0x0003, 3},{0x0011, 5},{0x0020, 6},{0x0074, 7},
+    {0x010D, 9},{0x0863,12},{0x0860,12},{0x000A, 5},
+    {0x0075, 7},{0x0001, 3},{0x0000, 3},{0x000B, 4},
+    {0x000A, 4},{0x0018, 5},{0x0038, 6},{0x0042, 7},
+    {0x010F, 9},{0x010E, 9},{0x0219,10},{0x10C3,13},
+    {0x10C2,13},{0x10C5,13},{0x10C4,13},{0x000F, 4},
+    {0x0004, 4},{0x0019, 5},{0x000B, 5},{0x0039, 6},
+    {0x0009, 4},{0x001B, 5},{0x001A, 5},{0x003B, 6}
+  },
+  {
+    {0x0005, 3},{0x0001, 4},{0x003E, 6},{0x0001, 5},
+    {0x00E2, 8},{0x1C6F,13},{0x38D9,14},{0x0039, 6},
+    {0x001F, 6},{0x0002, 3},{0x0001, 3},{0x0009, 4},
+    {0x0008, 4},{0x0000, 5},{0x0070, 7},{0x01C7, 9},
+    {0x038C,10},{0x071A,11},{0x38D8,14},{0x38DB,14},
+    {0x38DA,14},{0x38DD,14},{0x38DC,14},{0x000D, 4},
+    {0x001D, 5},{0x000E, 5},{0x003F, 6},{0x003C, 6},
+    {0x000C, 4},{0x0006, 4},{0x003D, 6},{0x001E, 6}
+  },
+  {
+    {0x0006, 3},{0x000B, 4},{0x0011, 5},{0x001E, 5},
+    {0x0074, 7},{0x03AA,10},{0x1D5C,13},{0x0001, 6},
+    {0x0021, 6},{0x0001, 3},{0x0002, 3},{0x0007, 4},
+    {0x0006, 4},{0x003E, 6},{0x00EB, 8},{0x01D4, 9},
+    {0x0EAF,12},{0x3ABB,14},{0x3ABA,14},{0x1D59,13},
+    {0x1D58,13},{0x1D5B,13},{0x1D5A,13},{0x000A, 4},
+    {0x001C, 5},{0x0001, 5},{0x003F, 6},{0x003B, 6},
+    {0x0001, 4},{0x0009, 4},{0x0020, 6},{0x0000, 6}
+  },
+  {
+    {0x0004, 3},{0x000A, 4},{0x0017, 5},{0x0004, 4},
+    {0x0016, 6},{0x016A, 9},{0x16B1,13},{0x0017, 7},
+    {0x005B, 7},{0x0006, 3},{0x0007, 3},{0x0001, 4},
+    {0x0000, 4},{0x000A, 6},{0x02D7,10},{0x0B5A,12},
+    {0x16B0,13},{0x16B3,13},{0x16B2,13},{0x2D6D,14},
+    {0x2D6C,14},{0x2D6F,14},{0x2D6E,14},{0x0006, 4},
+    {0x000A, 5},{0x0004, 5},{0x002C, 6},{0x0017, 6},
+    {0x0003, 4},{0x0007, 4},{0x0016, 7},{0x00B4, 8}
+  },
+  {
+    {0x0005, 3},{0x000D, 4},{0x0005, 4},{0x0009, 4},
+    {0x0033, 6},{0x0193, 9},{0x192C,13},{0x0061, 8},
+    {0x0031, 7},{0x0000, 2},{0x0007, 3},{0x0010, 5},
+    {0x0011, 5},{0x00C8, 8},{0x192F,13},{0x325B,14},
+    {0x325A,14},{0x1929,13},{0x1928,13},{0x192B,13},
+    {0x192A,13},{0x325D,14},{0x325C,14},{0x0018, 5},
+    {0x001A, 6},{0x001B, 6},{0x0065, 7},{0x0019, 6},
+    {0x0004, 4},{0x0007, 4},{0x0060, 8},{0x0324,10}
+  },
+  {
+    {0x0006, 3},{0x0000, 3},{0x0002, 4},{0x000F, 4},
+    {0x0039, 6},{0x01D9, 9},{0x1D82,13},{0x0761,11},
+    {0x03BE,10},{0x0001, 2},{0x0002, 2},{0x000F, 6},
+    {0x000E, 6},{0x0762,11},{0x3B07,14},{0x3B06,14},
+    {0x3B1D,14},{0x3B1C,14},{0x3B1F,14},{0x3B1E,14},
+    {0x3B19,14},{0x3B18,14},{0x3B1B,14},{0x0038, 6},
+    {0x01DE, 9},{0x00ED, 8},{0x03BF,10},{0x00EE, 8},
+    {0x003A, 6},{0x0006, 5},{0x0EC0,12},{0x3B1A,14}
+  },
+  {
+    {0x0000, 2},{0x0002, 3},{0x000F, 5},{0x0006, 4},
+    {0x001C, 6},{0x01D0,10},{0x0E8C,13},{0x1D1B,14},
+    {0x1D1A,14},{0x0003, 2},{0x0002, 2},{0x00EA, 9},
+    {0x00E9, 9},{0x0E89,13},{0x0E88,13},{0x0E8B,13},
+    {0x0E8A,13},{0x1D65,14},{0x1D64,14},{0x1D67,14},
+    {0x1D66,14},{0x1D61,14},{0x1D60,14},{0x03AD,11},
+    {0x1D63,14},{0x1D62,14},{0x1D1D,14},{0x1D1C,14},
+    {0x003B, 7},{0x01D7,10},{0x1D1F,14},{0x1D1E,14}
+  },
+  {
+    {0x0002, 2},{0x000F, 4},{0x001C, 5},{0x000C, 4},
+    {0x003B, 6},{0x01AC, 9},{0x1AD8,13},{0x35B3,14},
+    {0x35B2,14},{0x0001, 2},{0x0000, 2},{0x0069, 7},
+    {0x0068, 7},{0x35BD,14},{0x35BC,14},{0x35BF,14},
+    {0x35BE,14},{0x35B9,14},{0x35B8,14},{0x35BB,14},
+    {0x35BA,14},{0x35B5,14},{0x35B4,14},{0x01A9, 9},
+    {0x01A8, 9},{0x035A,10},{0x00D7, 8},{0x00D5, 8},
+    {0x003A, 6},{0x001B, 5},{0x35B7,14},{0x35B6,14}
+  }
+};
+
+
+
+/*A description of a Huffman code value used when encoding the tree.*/
+typedef struct{
+  /*The bit pattern, left-shifted so that the MSB of all patterns is
+     aligned.*/
+  ogg_uint32_t pattern;
+  /*The amount the bit pattern was shifted.*/
+  int          shift;
+  /*The token this bit pattern represents.*/
+  int          token;
+}oc_huff_entry;
+
+
+
+/*Compares two oc_huff_entry structures by their bit patterns.
+  _c1: The first entry to compare.
+  _c2: The second entry to compare.
+  Return: <0 if _c1<_c2, >0 if _c1>_c2.*/
+static int huff_entry_cmp(const void *_c1,const void *_c2){
+  ogg_uint32_t b1;
+  ogg_uint32_t b2;
+  b1=((const oc_huff_entry *)_c1)->pattern;
+  b2=((const oc_huff_entry *)_c2)->pattern;
+  return b1<b2?-1:b1>b2?1:0;
+}
+
+/*Encodes a description of the given Huffman tables.
+  Although the codes are stored in the encoder as flat arrays, in the bit
+   stream and in the decoder they are structured as a tree.
+  This function recovers the tree structure from the flat array and then
+   writes it out.
+  Note that the codes MUST form a Huffman code, and not merely a prefix-free
+   code, since the binary tree is assumed to be full.
+  _opb:   The buffer to store the tree in.
+  _codes: The Huffman tables to pack.
+  Return: 0 on success, or a negative value if one of the given Huffman tables
+   does not form a full, prefix-free code.*/
+int oc_huff_codes_pack(oggpack_buffer *_opb,
+ const th_huff_code _codes[TH_NHUFFMAN_TABLES][TH_NDCT_TOKENS]){
+  int i;
+  for(i=0;i<TH_NHUFFMAN_TABLES;i++){
+    oc_huff_entry entries[TH_NDCT_TOKENS];
+    int           bpos;
+    int           maxlen;
+    int           mask;
+    int           j;
+    /*First, find the maximum code length so we can align all the bit
+       patterns.*/
+    maxlen=_codes[i][0].nbits;
+    for(j=1;j<TH_NDCT_TOKENS;j++)maxlen=OC_MAXI(_codes[i][j].nbits,maxlen);
+    /*It's improbable that a code with more than 32 bits could pass the
+       validation below, but abort early in any case.*/
+    if(maxlen>32)return TH_EINVAL;
+    mask=(1<<(maxlen>>1)<<(maxlen+1>>1))-1;
+    /*Copy over the codes into our temporary workspace.
+      The bit patterns are aligned, and the original entry each code is from
+       is stored as well.*/
+    for(j=0;j<TH_NDCT_TOKENS;j++){
+      entries[j].shift=maxlen-_codes[i][j].nbits;
+      entries[j].pattern=_codes[i][j].pattern<<entries[j].shift&mask;
+      entries[j].token=j;
+    }
+    /*Sort the codes into ascending order.
+      This is the order the leaves of the tree will be traversed.*/
+    qsort(entries,TH_NDCT_TOKENS,sizeof(entries[0]),huff_entry_cmp);
+    /*For each leaf of the tree:*/
+    bpos=maxlen;
+    for(j=0;j<TH_NDCT_TOKENS;j++){
+      ogg_uint32_t bit;
+      /*Fail if this code has no bits at all.
+        Technically a codebook with a single 0-bit entry is legal, but the
+         encoder currently does not support codebooks which do not contain all
+         the tokens.*/
+      if(entries[j].shift>=maxlen)return TH_EINVAL;
+      /*Descend into the tree, writing a bit for each branch.*/
+      for(;bpos>entries[j].shift;bpos--)oggpackB_write(_opb,0,1);
+      /*Mark this as a leaf node, and write its value.*/
+      oggpackB_write(_opb,1,1);
+      oggpackB_write(_opb,entries[j].token,5);
+      /*For each 1 branch we've descended, back up the tree until we reach a
+         0 branch.*/
+      bit=(ogg_uint32_t)1<<bpos;
+      for(;entries[j].pattern&bit;bpos++)bit<<=1;
+      /*Validate the code.*/
+      if(j+1<TH_NDCT_TOKENS){
+        mask=~(bit-1)<<1;
+        /*The next entry should have a 1 bit where we had a 0, and should
+           match our code above that bit.
+          This verifies both fullness and prefix-freeness simultaneously.*/
+        if(!(entries[j+1].pattern&bit)||
+         (entries[j].pattern&mask)!=(entries[j+1].pattern&mask)){
+          return TH_EINVAL;
+        }
+      }
+      /*If there are no more codes, we should have ascended back to the top
+         of the tree.*/
+      else if(bpos<maxlen)return TH_EINVAL;
+    }
+  }
+  return 0;
+}
+
+/*This is used to copy the configuration of an existing setup header for use by
+   the encoder.
+  The decoder uses a completely different data structure for the Huffman
+   codebooks.*/
+int oc_huff_codes_unpack(oc_pack_buf *_opb,
+ th_huff_code _codes[TH_NHUFFMAN_TABLES][TH_NDCT_TOKENS]){
+  int i;
+  for(i=0;i<TH_NHUFFMAN_TABLES;i++){
+    ogg_uint32_t code;
+    int          len;
+    int          nleaves;
+    code=0;
+    len=nleaves=0;
+    memset(_codes[i],0,TH_NDCT_TOKENS*sizeof(*_codes[i]));
+    for(;;){
+      long bits;
+      bits=oc_pack_read1(_opb);
+      /*Only process nodes so long as there's more bits in the buffer.*/
+      if(oc_pack_bytes_left(_opb)<0)return TH_EBADHEADER;
+      /*Read an internal node:*/
+      if(!bits){
+        len++;
+        /*Don't allow codewords longer than 32 bits.*/
+        if(len>32)return TH_EBADHEADER;
+      }
+      /*Read a leaf node:*/
+      else{
+        ogg_uint32_t code_bit;
+        /*Don't allow more than 32 tokens per codebook.*/
+        if(++nleaves>32)return TH_EBADHEADER;
+        bits=oc_pack_read(_opb,OC_NDCT_TOKEN_BITS);
+        /*The current encoder does not support codebooks that do not contain
+           all of the tokens.*/
+        if(_codes[i][bits].nbits>0)return TH_EINVAL;
+        _codes[i][bits].pattern=code>>32-len;
+        _codes[i][bits].nbits=len;
+        code_bit=0x80000000U>>len-1;
+        while(len>0&&(code&code_bit)){
+          code^=code_bit;
+          code_bit<<=1;
+          len--;
+        }
+        if(len<=0)break;
+        code|=code_bit;
+      }
+    }
+    /*The current encoder does not support codebooks that do not contain all of
+       the tokens.*/
+    if(nleaves<32)return TH_EINVAL;
+  }
+  return 0;
+}

thirdparty/libtheora/huffenc.h

@@ -0,0 +1,22 @@
+#if !defined(_huffenc_H)
+# define _huffenc_H (1)
+# include "huffman.h"
+# include "bitpack.h"
+
+
+
+typedef th_huff_code                  th_huff_table[TH_NDCT_TOKENS];
+
+
+
+extern const th_huff_code
+ TH_VP31_HUFF_CODES[TH_NHUFFMAN_TABLES][TH_NDCT_TOKENS];
+
+
+
+int oc_huff_codes_pack(oggpack_buffer *_opb,
+ const th_huff_code _codes[TH_NHUFFMAN_TABLES][TH_NDCT_TOKENS]);
+int oc_huff_codes_unpack(oc_pack_buf *_opb,
+ th_huff_code _codes[TH_NHUFFMAN_TABLES][TH_NDCT_TOKENS]);
+
+#endif

thirdparty/libtheora/huffman.h

@@ -0,0 +1,70 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+#if !defined(_huffman_H)
+# define _huffman_H (1)
+# include "theora/codec.h"
+# include "ocintrin.h"
+
+/*The range of valid quantized DCT coefficient values.
+  VP3 used 511 in the encoder, but the bitstream is capable of 580.*/
+#define OC_DCT_VAL_RANGE         (580)
+
+#define OC_NDCT_TOKEN_BITS       (5)
+
+#define OC_DCT_EOB1_TOKEN        (0)
+#define OC_DCT_EOB2_TOKEN        (1)
+#define OC_DCT_EOB3_TOKEN        (2)
+#define OC_DCT_REPEAT_RUN0_TOKEN (3)
+#define OC_DCT_REPEAT_RUN1_TOKEN (4)
+#define OC_DCT_REPEAT_RUN2_TOKEN (5)
+#define OC_DCT_REPEAT_RUN3_TOKEN (6)
+
+#define OC_DCT_SHORT_ZRL_TOKEN   (7)
+#define OC_DCT_ZRL_TOKEN         (8)
+
+#define OC_ONE_TOKEN             (9)
+#define OC_MINUS_ONE_TOKEN       (10)
+#define OC_TWO_TOKEN             (11)
+#define OC_MINUS_TWO_TOKEN       (12)
+
+#define OC_DCT_VAL_CAT2          (13)
+#define OC_DCT_VAL_CAT3          (17)
+#define OC_DCT_VAL_CAT4          (18)
+#define OC_DCT_VAL_CAT5          (19)
+#define OC_DCT_VAL_CAT6          (20)
+#define OC_DCT_VAL_CAT7          (21)
+#define OC_DCT_VAL_CAT8          (22)
+
+#define OC_DCT_RUN_CAT1A         (23)
+#define OC_DCT_RUN_CAT1B         (28)
+#define OC_DCT_RUN_CAT1C         (29)
+#define OC_DCT_RUN_CAT2A         (30)
+#define OC_DCT_RUN_CAT2B         (31)
+
+#define OC_NDCT_EOB_TOKEN_MAX    (7)
+#define OC_NDCT_ZRL_TOKEN_MAX    (9)
+#define OC_NDCT_VAL_MAX          (23)
+#define OC_NDCT_VAL_CAT1_MAX     (13)
+#define OC_NDCT_VAL_CAT2_MAX     (17)
+#define OC_NDCT_VAL_CAT2_SIZE    (OC_NDCT_VAL_CAT2_MAX-OC_DCT_VAL_CAT2)
+#define OC_NDCT_RUN_MAX          (32)
+#define OC_NDCT_RUN_CAT1A_MAX    (28)
+
+extern const unsigned char OC_DCT_TOKEN_EXTRA_BITS[TH_NDCT_TOKENS];
+
+#endif

thirdparty/libtheora/idct.c

@@ -0,0 +1,330 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+#include <string.h>
+#include "internal.h"
+#include "dct.h"
+
+/*Performs an inverse 8 point Type-II DCT transform.
+  The output is scaled by a factor of 2 relative to the orthonormal version of
+   the transform.
+  _y: The buffer to store the result in.
+      Data will be placed in every 8th entry (e.g., in a column of an 8x8
+       block).
+  _x: The input coefficients.
+      The first 8 entries are used (e.g., from a row of an 8x8 block).*/
+static void idct8(ogg_int16_t *_y,const ogg_int16_t _x[8]){
+  ogg_int32_t t[8];
+  ogg_int32_t r;
+  /*Stage 1:*/
+  /*0-1 butterfly.*/
+  t[0]=OC_C4S4*(ogg_int16_t)(_x[0]+_x[4])>>16;
+  t[1]=OC_C4S4*(ogg_int16_t)(_x[0]-_x[4])>>16;
+  /*2-3 rotation by 6pi/16.*/
+  t[2]=(OC_C6S2*_x[2]>>16)-(OC_C2S6*_x[6]>>16);
+  t[3]=(OC_C2S6*_x[2]>>16)+(OC_C6S2*_x[6]>>16);
+  /*4-7 rotation by 7pi/16.*/
+  t[4]=(OC_C7S1*_x[1]>>16)-(OC_C1S7*_x[7]>>16);
+  /*5-6 rotation by 3pi/16.*/
+  t[5]=(OC_C3S5*_x[5]>>16)-(OC_C5S3*_x[3]>>16);
+  t[6]=(OC_C5S3*_x[5]>>16)+(OC_C3S5*_x[3]>>16);
+  t[7]=(OC_C1S7*_x[1]>>16)+(OC_C7S1*_x[7]>>16);
+  /*Stage 2:*/
+  /*4-5 butterfly.*/
+  r=t[4]+t[5];
+  t[5]=OC_C4S4*(ogg_int16_t)(t[4]-t[5])>>16;
+  t[4]=r;
+  /*7-6 butterfly.*/
+  r=t[7]+t[6];
+  t[6]=OC_C4S4*(ogg_int16_t)(t[7]-t[6])>>16;
+  t[7]=r;
+  /*Stage 3:*/
+  /*0-3 butterfly.*/
+  r=t[0]+t[3];
+  t[3]=t[0]-t[3];
+  t[0]=r;
+  /*1-2 butterfly.*/
+  r=t[1]+t[2];
+  t[2]=t[1]-t[2];
+  t[1]=r;
+  /*6-5 butterfly.*/
+  r=t[6]+t[5];
+  t[5]=t[6]-t[5];
+  t[6]=r;
+  /*Stage 4:*/
+  /*0-7 butterfly.*/
+  _y[0<<3]=(ogg_int16_t)(t[0]+t[7]);
+  /*1-6 butterfly.*/
+  _y[1<<3]=(ogg_int16_t)(t[1]+t[6]);
+  /*2-5 butterfly.*/
+  _y[2<<3]=(ogg_int16_t)(t[2]+t[5]);
+  /*3-4 butterfly.*/
+  _y[3<<3]=(ogg_int16_t)(t[3]+t[4]);
+  _y[4<<3]=(ogg_int16_t)(t[3]-t[4]);
+  _y[5<<3]=(ogg_int16_t)(t[2]-t[5]);
+  _y[6<<3]=(ogg_int16_t)(t[1]-t[6]);
+  _y[7<<3]=(ogg_int16_t)(t[0]-t[7]);
+}
+
+/*Performs an inverse 8 point Type-II DCT transform.
+  The output is scaled by a factor of 2 relative to the orthonormal version of
+   the transform.
+  _y: The buffer to store the result in.
+      Data will be placed in every 8th entry (e.g., in a column of an 8x8
+       block).
+  _x: The input coefficients.
+      Only the first 4 entries are used.
+      The other 4 are assumed to be 0.*/
+static void idct8_4(ogg_int16_t *_y,const ogg_int16_t _x[8]){
+  ogg_int32_t t[8];
+  ogg_int32_t r;
+  /*Stage 1:*/
+  t[0]=OC_C4S4*_x[0]>>16;
+  t[2]=OC_C6S2*_x[2]>>16;
+  t[3]=OC_C2S6*_x[2]>>16;
+  t[4]=OC_C7S1*_x[1]>>16;
+  t[5]=-(OC_C5S3*_x[3]>>16);
+  t[6]=OC_C3S5*_x[3]>>16;
+  t[7]=OC_C1S7*_x[1]>>16;
+  /*Stage 2:*/
+  r=t[4]+t[5];
+  t[5]=OC_C4S4*(ogg_int16_t)(t[4]-t[5])>>16;
+  t[4]=r;
+  r=t[7]+t[6];
+  t[6]=OC_C4S4*(ogg_int16_t)(t[7]-t[6])>>16;
+  t[7]=r;
+  /*Stage 3:*/
+  t[1]=t[0]+t[2];
+  t[2]=t[0]-t[2];
+  r=t[0]+t[3];
+  t[3]=t[0]-t[3];
+  t[0]=r;
+  r=t[6]+t[5];
+  t[5]=t[6]-t[5];
+  t[6]=r;
+  /*Stage 4:*/
+  _y[0<<3]=(ogg_int16_t)(t[0]+t[7]);
+  _y[1<<3]=(ogg_int16_t)(t[1]+t[6]);
+  _y[2<<3]=(ogg_int16_t)(t[2]+t[5]);
+  _y[3<<3]=(ogg_int16_t)(t[3]+t[4]);
+  _y[4<<3]=(ogg_int16_t)(t[3]-t[4]);
+  _y[5<<3]=(ogg_int16_t)(t[2]-t[5]);
+  _y[6<<3]=(ogg_int16_t)(t[1]-t[6]);
+  _y[7<<3]=(ogg_int16_t)(t[0]-t[7]);
+}
+
+/*Performs an inverse 8 point Type-II DCT transform.
+  The output is scaled by a factor of 2 relative to the orthonormal version of
+   the transform.
+  _y: The buffer to store the result in.
+      Data will be placed in every 8th entry (e.g., in a column of an 8x8
+       block).
+  _x: The input coefficients.
+      Only the first 3 entries are used.
+      The other 5 are assumed to be 0.*/
+static void idct8_3(ogg_int16_t *_y,const ogg_int16_t _x[8]){
+  ogg_int32_t t[8];
+  ogg_int32_t r;
+  /*Stage 1:*/
+  t[0]=OC_C4S4*_x[0]>>16;
+  t[2]=OC_C6S2*_x[2]>>16;
+  t[3]=OC_C2S6*_x[2]>>16;
+  t[4]=OC_C7S1*_x[1]>>16;
+  t[7]=OC_C1S7*_x[1]>>16;
+  /*Stage 2:*/
+  t[5]=OC_C4S4*t[4]>>16;
+  t[6]=OC_C4S4*t[7]>>16;
+  /*Stage 3:*/
+  t[1]=t[0]+t[2];
+  t[2]=t[0]-t[2];
+  r=t[0]+t[3];
+  t[3]=t[0]-t[3];
+  t[0]=r;
+  r=t[6]+t[5];
+  t[5]=t[6]-t[5];
+  t[6]=r;
+  /*Stage 4:*/
+  _y[0<<3]=(ogg_int16_t)(t[0]+t[7]);
+  _y[1<<3]=(ogg_int16_t)(t[1]+t[6]);
+  _y[2<<3]=(ogg_int16_t)(t[2]+t[5]);
+  _y[3<<3]=(ogg_int16_t)(t[3]+t[4]);
+  _y[4<<3]=(ogg_int16_t)(t[3]-t[4]);
+  _y[5<<3]=(ogg_int16_t)(t[2]-t[5]);
+  _y[6<<3]=(ogg_int16_t)(t[1]-t[6]);
+  _y[7<<3]=(ogg_int16_t)(t[0]-t[7]);
+}
+
+/*Performs an inverse 8 point Type-II DCT transform.
+  The output is scaled by a factor of 2 relative to the orthonormal version of
+   the transform.
+  _y: The buffer to store the result in.
+      Data will be placed in every 8th entry (e.g., in a column of an 8x8
+       block).
+  _x: The input coefficients.
+      Only the first 2 entries are used.
+      The other 6 are assumed to be 0.*/
+static void idct8_2(ogg_int16_t *_y,const ogg_int16_t _x[8]){
+  ogg_int32_t t[8];
+  ogg_int32_t r;
+  /*Stage 1:*/
+  t[0]=OC_C4S4*_x[0]>>16;
+  t[4]=OC_C7S1*_x[1]>>16;
+  t[7]=OC_C1S7*_x[1]>>16;
+  /*Stage 2:*/
+  t[5]=OC_C4S4*t[4]>>16;
+  t[6]=OC_C4S4*t[7]>>16;
+  /*Stage 3:*/
+  r=t[6]+t[5];
+  t[5]=t[6]-t[5];
+  t[6]=r;
+  /*Stage 4:*/
+  _y[0<<3]=(ogg_int16_t)(t[0]+t[7]);
+  _y[1<<3]=(ogg_int16_t)(t[0]+t[6]);
+  _y[2<<3]=(ogg_int16_t)(t[0]+t[5]);
+  _y[3<<3]=(ogg_int16_t)(t[0]+t[4]);
+  _y[4<<3]=(ogg_int16_t)(t[0]-t[4]);
+  _y[5<<3]=(ogg_int16_t)(t[0]-t[5]);
+  _y[6<<3]=(ogg_int16_t)(t[0]-t[6]);
+  _y[7<<3]=(ogg_int16_t)(t[0]-t[7]);
+}
+
+/*Performs an inverse 8 point Type-II DCT transform.
+  The output is scaled by a factor of 2 relative to the orthonormal version of
+   the transform.
+  _y: The buffer to store the result in.
+      Data will be placed in every 8th entry (e.g., in a column of an 8x8
+       block).
+  _x: The input coefficients.
+      Only the first entry is used.
+      The other 7 are assumed to be 0.*/
+static void idct8_1(ogg_int16_t *_y,const ogg_int16_t _x[1]){
+  _y[0<<3]=_y[1<<3]=_y[2<<3]=_y[3<<3]=
+   _y[4<<3]=_y[5<<3]=_y[6<<3]=_y[7<<3]=(ogg_int16_t)(OC_C4S4*_x[0]>>16);
+}
+
+/*Performs an inverse 8x8 Type-II DCT transform.
+  The input is assumed to be scaled by a factor of 4 relative to orthonormal
+   version of the transform.
+  All coefficients but the first 3 in zig-zag scan order are assumed to be 0:
+   x  x  0  0  0  0  0  0
+   x  0  0  0  0  0  0  0
+   0  0  0  0  0  0  0  0
+   0  0  0  0  0  0  0  0
+   0  0  0  0  0  0  0  0
+   0  0  0  0  0  0  0  0
+   0  0  0  0  0  0  0  0
+   0  0  0  0  0  0  0  0
+  _y: The buffer to store the result in.
+      This may be the same as _x.
+  _x: The input coefficients.*/
+static void oc_idct8x8_3(ogg_int16_t _y[64],ogg_int16_t _x[64]){
+  ogg_int16_t w[64];
+  int         i;
+  /*Transform rows of x into columns of w.*/
+  idct8_2(w,_x);
+  idct8_1(w+1,_x+8);
+  /*Transform rows of w into columns of y.*/
+  for(i=0;i<8;i++)idct8_2(_y+i,w+i*8);
+  /*Adjust for the scale factor.*/
+  for(i=0;i<64;i++)_y[i]=(ogg_int16_t)(_y[i]+8>>4);
+  /*Clear input data for next block.*/
+  _x[0]=_x[1]=_x[8]=0;
+}
+
+/*Performs an inverse 8x8 Type-II DCT transform.
+  The input is assumed to be scaled by a factor of 4 relative to orthonormal
+   version of the transform.
+  All coefficients but the first 10 in zig-zag scan order are assumed to be 0:
+   x  x  x  x  0  0  0  0
+   x  x  x  0  0  0  0  0
+   x  x  0  0  0  0  0  0
+   x  0  0  0  0  0  0  0
+   0  0  0  0  0  0  0  0
+   0  0  0  0  0  0  0  0
+   0  0  0  0  0  0  0  0
+   0  0  0  0  0  0  0  0
+  _y: The buffer to store the result in.
+      This may be the same as _x.
+  _x: The input coefficients.*/
+static void oc_idct8x8_10(ogg_int16_t _y[64],ogg_int16_t _x[64]){
+  ogg_int16_t w[64];
+  int         i;
+  /*Transform rows of x into columns of w.*/
+  idct8_4(w,_x);
+  idct8_3(w+1,_x+8);
+  idct8_2(w+2,_x+16);
+  idct8_1(w+3,_x+24);
+  /*Transform rows of w into columns of y.*/
+  for(i=0;i<8;i++)idct8_4(_y+i,w+i*8);
+  /*Adjust for the scale factor.*/
+  for(i=0;i<64;i++)_y[i]=(ogg_int16_t)(_y[i]+8>>4);
+  /*Clear input data for next block.*/
+  _x[0]=_x[1]=_x[2]=_x[3]=_x[8]=_x[9]=_x[10]=_x[16]=_x[17]=_x[24]=0;
+}
+
+/*Performs an inverse 8x8 Type-II DCT transform.
+  The input is assumed to be scaled by a factor of 4 relative to orthonormal
+   version of the transform.
+  _y: The buffer to store the result in.
+      This may be the same as _x.
+  _x: The input coefficients.*/
+static void oc_idct8x8_slow(ogg_int16_t _y[64],ogg_int16_t _x[64]){
+  ogg_int16_t w[64];
+  int         i;
+  /*Transform rows of x into columns of w.*/
+  for(i=0;i<8;i++)idct8(w+i,_x+i*8);
+  /*Transform rows of w into columns of y.*/
+  for(i=0;i<8;i++)idct8(_y+i,w+i*8);
+  /*Adjust for the scale factor.*/
+  for(i=0;i<64;i++)_y[i]=(ogg_int16_t)(_y[i]+8>>4);
+  /*Clear input data for next block.*/
+  for(i=0;i<64;i++)_x[i]=0;
+}
+
+/*Performs an inverse 8x8 Type-II DCT transform.
+  The input is assumed to be scaled by a factor of 4 relative to orthonormal
+   version of the transform.*/
+void oc_idct8x8_c(ogg_int16_t _y[64],ogg_int16_t _x[64],int _last_zzi){
+  /*_last_zzi is subtly different from an actual count of the number of
+     coefficients we decoded for this block.
+    It contains the value of zzi BEFORE the final token in the block was
+     decoded.
+    In most cases this is an EOB token (the continuation of an EOB run from a
+     previous block counts), and so this is the same as the coefficient count.
+    However, in the case that the last token was NOT an EOB token, but filled
+     the block up with exactly 64 coefficients, _last_zzi will be less than 64.
+    Provided the last token was not a pure zero run, the minimum value it can
+     be is 46, and so that doesn't affect any of the cases in this routine.
+    However, if the last token WAS a pure zero run of length 63, then _last_zzi
+     will be 1 while the number of coefficients decoded is 64.
+    Thus, we will trigger the following special case, where the real
+     coefficient count would not.
+    Note also that a zero run of length 64 will give _last_zzi a value of 0,
+     but we still process the DC coefficient, which might have a non-zero value
+     due to DC prediction.
+    Although convoluted, this is arguably the correct behavior: it allows us to
+     use a smaller transform when the block ends with a long zero run instead
+     of a normal EOB token.
+    It could be smarter... multiple separate zero runs at the end of a block
+     will fool it, but an encoder that generates these really deserves what it
+     gets.
+    Needless to say we inherited this approach from VP3.*/
+  /*Then perform the iDCT.*/
+  if(_last_zzi<=3)oc_idct8x8_3(_y,_x);
+  else if(_last_zzi<=10)oc_idct8x8_10(_y,_x);
+  else oc_idct8x8_slow(_y,_x);
+}

thirdparty/libtheora/info.c

@@ -0,0 +1,131 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+#include <stdlib.h>
+#include <ctype.h>
+#include <string.h>
+#include "internal.h"
+
+
+
+/*This is more or less the same as strncasecmp, but that doesn't exist
+   everywhere, and this is a fairly trivial function, so we include it.
+  Note: We take advantage of the fact that we know _n is less than or equal to
+   the length of at least one of the strings.*/
+static int oc_tagcompare(const char *_s1,const char *_s2,int _n){
+  int c;
+  for(c=0;c<_n;c++){
+    if(toupper(_s1[c])!=toupper(_s2[c]))return !0;
+  }
+  return _s1[c]!='=';
+}
+
+
+
+void th_info_init(th_info *_info){
+  memset(_info,0,sizeof(*_info));
+  _info->version_major=TH_VERSION_MAJOR;
+  _info->version_minor=TH_VERSION_MINOR;
+  _info->version_subminor=TH_VERSION_SUB;
+  _info->keyframe_granule_shift=6;
+}
+
+void th_info_clear(th_info *_info){
+  memset(_info,0,sizeof(*_info));
+}
+
+
+
+void th_comment_init(th_comment *_tc){
+  memset(_tc,0,sizeof(*_tc));
+}
+
+void th_comment_add(th_comment *_tc,const char *_comment){
+  char **user_comments;
+  int   *comment_lengths;
+  int    comment_len;
+  user_comments=_ogg_realloc(_tc->user_comments,
+   (_tc->comments+2)*sizeof(*_tc->user_comments));
+  if(user_comments==NULL)return;
+  _tc->user_comments=user_comments;
+  comment_lengths=_ogg_realloc(_tc->comment_lengths,
+   (_tc->comments+2)*sizeof(*_tc->comment_lengths));
+  if(comment_lengths==NULL)return;
+  _tc->comment_lengths=comment_lengths;
+  comment_len=strlen(_comment);
+  comment_lengths[_tc->comments]=comment_len;
+  user_comments[_tc->comments]=_ogg_malloc(comment_len+1);
+  if(user_comments[_tc->comments]==NULL)return;
+  memcpy(_tc->user_comments[_tc->comments],_comment,comment_len+1);
+  _tc->comments++;
+  _tc->user_comments[_tc->comments]=NULL;
+}
+
+void th_comment_add_tag(th_comment *_tc,const char *_tag,const char *_val){
+  char *comment;
+  int   tag_len;
+  int   val_len;
+  tag_len=strlen(_tag);
+  val_len=strlen(_val);
+  /*+2 for '=' and '\0'.*/
+  comment=_ogg_malloc(tag_len+val_len+2);
+  if(comment==NULL)return;
+  memcpy(comment,_tag,tag_len);
+  comment[tag_len]='=';
+  memcpy(comment+tag_len+1,_val,val_len+1);
+  th_comment_add(_tc,comment);
+  _ogg_free(comment);
+}
+
+char *th_comment_query(th_comment *_tc,const char *_tag,int _count){
+  long i;
+  int  found;
+  int  tag_len;
+  tag_len=strlen(_tag);
+  found=0;
+  for(i=0;i<_tc->comments;i++){
+    if(!oc_tagcompare(_tc->user_comments[i],_tag,tag_len)){
+      /*We return a pointer to the data, not a copy.*/
+      if(_count==found++)return _tc->user_comments[i]+tag_len+1;
+    }
+  }
+  /*Didn't find anything.*/
+  return NULL;
+}
+
+int th_comment_query_count(th_comment *_tc,const char *_tag){
+  long i;
+  int  tag_len;
+  int  count;
+  tag_len=strlen(_tag);
+  count=0;
+  for(i=0;i<_tc->comments;i++){
+    if(!oc_tagcompare(_tc->user_comments[i],_tag,tag_len))count++;
+  }
+  return count;
+}
+
+void th_comment_clear(th_comment *_tc){
+  if(_tc!=NULL){
+    long i;
+    for(i=0;i<_tc->comments;i++)_ogg_free(_tc->user_comments[i]);
+    _ogg_free(_tc->user_comments);
+    _ogg_free(_tc->comment_lengths);
+    _ogg_free(_tc->vendor);
+    memset(_tc,0,sizeof(*_tc));
+  }
+}

thirdparty/libtheora/internal.c

@@ -0,0 +1,210 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+#include <stdlib.h>
+#include <limits.h>
+#include <string.h>
+#include "internal.h"
+
+
+
+/*A map from the index in the zig zag scan to the coefficient number in a
+   block.
+  All zig zag indices beyond 63 are sent to coefficient 64, so that zero runs
+   past the end of a block in bogus streams get mapped to a known location.*/
+const unsigned char OC_FZIG_ZAG[128]={
+   0, 1, 8,16, 9, 2, 3,10,
+  17,24,32,25,18,11, 4, 5,
+  12,19,26,33,40,48,41,34,
+  27,20,13, 6, 7,14,21,28,
+  35,42,49,56,57,50,43,36,
+  29,22,15,23,30,37,44,51,
+  58,59,52,45,38,31,39,46,
+  53,60,61,54,47,55,62,63,
+  64,64,64,64,64,64,64,64,
+  64,64,64,64,64,64,64,64,
+  64,64,64,64,64,64,64,64,
+  64,64,64,64,64,64,64,64,
+  64,64,64,64,64,64,64,64,
+  64,64,64,64,64,64,64,64,
+  64,64,64,64,64,64,64,64,
+  64,64,64,64,64,64,64,64
+};
+
+/*A map from the coefficient number in a block to its index in the zig zag
+   scan.*/
+const unsigned char OC_IZIG_ZAG[64]={
+   0, 1, 5, 6,14,15,27,28,
+   2, 4, 7,13,16,26,29,42,
+   3, 8,12,17,25,30,41,43,
+   9,11,18,24,31,40,44,53,
+  10,19,23,32,39,45,52,54,
+  20,22,33,38,46,51,55,60,
+  21,34,37,47,50,56,59,61,
+  35,36,48,49,57,58,62,63
+};
+
+/*A map from physical macro block ordering to bitstream macro block
+   ordering within a super block.*/
+const unsigned char OC_MB_MAP[2][2]={{0,3},{1,2}};
+
+/*A list of the indices in the oc_mb.map array that can be valid for each of
+   the various chroma decimation types.*/
+const unsigned char OC_MB_MAP_IDXS[TH_PF_NFORMATS][12]={
+  {0,1,2,3,4,8},
+  {0,1,2,3,4,5,8,9},
+  {0,1,2,3,4,6,8,10},
+  {0,1,2,3,4,5,6,7,8,9,10,11}
+};
+
+/*The number of indices in the oc_mb.map array that can be valid for each of
+   the various chroma decimation types.*/
+const unsigned char OC_MB_MAP_NIDXS[TH_PF_NFORMATS]={6,8,8,12};
+
+/*The number of extra bits that are coded with each of the DCT tokens.
+  Each DCT token has some fixed number of additional bits (possibly 0) stored
+   after the token itself, containing, for example, coefficient magnitude,
+   sign bits, etc.*/
+const unsigned char OC_DCT_TOKEN_EXTRA_BITS[TH_NDCT_TOKENS]={
+  0,0,0,2,3,4,12,3,6,
+  0,0,0,0,
+  1,1,1,1,2,3,4,5,6,10,
+  1,1,1,1,1,3,4,
+  2,3
+};
+
+
+
+int oc_ilog(unsigned _v){
+  int ret;
+  for(ret=0;_v;ret++)_v>>=1;
+  return ret;
+}
+
+
+
+void *oc_aligned_malloc(size_t _sz,size_t _align){
+  unsigned char *p;
+  if(_align-1>UCHAR_MAX||(_align&_align-1)||_sz>~(size_t)0-_align)return NULL;
+  p=(unsigned char *)_ogg_malloc(_sz+_align);
+  if(p!=NULL){
+    int offs;
+    offs=((p-(unsigned char *)0)-1&_align-1);
+    p[offs]=offs;
+    p+=offs+1;
+  }
+  return p;
+}
+
+void oc_aligned_free(void *_ptr){
+  unsigned char *p;
+  p=(unsigned char *)_ptr;
+  if(p!=NULL){
+    int offs;
+    offs=*--p;
+    _ogg_free(p-offs);
+  }
+}
+
+
+void **oc_malloc_2d(size_t _height,size_t _width,size_t _sz){
+  size_t  rowsz;
+  size_t  colsz;
+  size_t  datsz;
+  char   *ret;
+  colsz=_height*sizeof(void *);
+  rowsz=_sz*_width;
+  datsz=rowsz*_height;
+  /*Alloc array and row pointers.*/
+  ret=(char *)_ogg_malloc(datsz+colsz);
+  /*Initialize the array.*/
+  if(ret!=NULL){
+    size_t   i;
+    void   **p;
+    char    *datptr;
+    p=(void **)ret;
+    i=_height;
+    for(datptr=ret+colsz;i-->0;p++,datptr+=rowsz)*p=(void *)datptr;
+  }
+  return (void **)ret;
+}
+
+void **oc_calloc_2d(size_t _height,size_t _width,size_t _sz){
+  size_t  colsz;
+  size_t  rowsz;
+  size_t  datsz;
+  char   *ret;
+  colsz=_height*sizeof(void *);
+  rowsz=_sz*_width;
+  datsz=rowsz*_height;
+  /*Alloc array and row pointers.*/
+  ret=(char *)_ogg_calloc(datsz+colsz,1);
+  /*Initialize the array.*/
+  if(ret!=NULL){
+    size_t   i;
+    void   **p;
+    char    *datptr;
+    p=(void **)ret;
+    i=_height;
+    for(datptr=ret+colsz;i-->0;p++,datptr+=rowsz)*p=(void *)datptr;
+  }
+  return (void **)ret;
+}
+
+void oc_free_2d(void *_ptr){
+  _ogg_free(_ptr);
+}
+
+/*Fills in a Y'CbCr buffer with a pointer to the image data in the first
+   buffer, but with the opposite vertical orientation.
+  _dst: The destination buffer.
+        This can be the same as _src.
+  _src: The source buffer.*/
+void oc_ycbcr_buffer_flip(th_ycbcr_buffer _dst,
+ const th_ycbcr_buffer _src){
+  int pli;
+  for(pli=0;pli<3;pli++){
+    _dst[pli].width=_src[pli].width;
+    _dst[pli].height=_src[pli].height;
+    _dst[pli].stride=-_src[pli].stride;
+    _dst[pli].data=_src[pli].data
+     +(1-_dst[pli].height)*(ptrdiff_t)_dst[pli].stride;
+  }
+}
+
+const char *th_version_string(void){
+  return OC_VENDOR_STRING;
+}
+
+ogg_uint32_t th_version_number(void){
+  return (TH_VERSION_MAJOR<<16)+(TH_VERSION_MINOR<<8)+TH_VERSION_SUB;
+}
+
+/*Determines the packet type.
+  Note that this correctly interprets a 0-byte packet as a video data packet.
+  Return: 1 for a header packet, 0 for a data packet.*/
+int th_packet_isheader(ogg_packet *_op){
+  return _op->bytes>0?_op->packet[0]>>7:0;
+}
+
+/*Determines the frame type of a video data packet.
+  Note that this correctly interprets a 0-byte packet as a delta frame.
+  Return: 1 for a key frame, 0 for a delta frame, and -1 for a header
+           packet.*/
+int th_packet_iskeyframe(ogg_packet *_op){
+  return _op->bytes<=0?0:_op->packet[0]&0x80?-1:!(_op->packet[0]&0x40);
+}

thirdparty/libtheora/internal.h

@@ -0,0 +1,120 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+#if !defined(_internal_H)
+# define _internal_H (1)
+# include <stdlib.h>
+# include <limits.h>
+# if defined(HAVE_CONFIG_H)
+#  include "config.h"
+# endif
+# include "theora/codec.h"
+# include "theora/theora.h"
+# include "ocintrin.h"
+
+# if !defined(__GNUC_PREREQ)
+#  if defined(__GNUC__)&&defined(__GNUC_MINOR__)
+#   define __GNUC_PREREQ(_maj,_min) \
+ ((__GNUC__<<16)+__GNUC_MINOR__>=((_maj)<<16)+(_min))
+#  else
+#   define __GNUC_PREREQ(_maj,_min) 0
+#  endif
+# endif
+
+# if defined(_MSC_VER)
+/*Disable missing EMMS warnings.*/
+#  pragma warning(disable:4799)
+/*Thank you Microsoft, I know the order of operations.*/
+#  pragma warning(disable:4554)
+# endif
+/*You, too, gcc.*/
+# if __GNUC_PREREQ(4,2)
+#  pragma GCC diagnostic ignored "-Wparentheses"
+# endif
+/*Et tu, clang.*/
+# if defined(__clang__)
+#  pragma clang diagnostic ignored "-Wparentheses"
+# endif
+
+/*Some assembly constructs require aligned operands.
+  The following macros are _only_ intended for structure member declarations.
+  Although they will sometimes work on stack variables, gcc will often silently
+   ignore them.
+  A separate set of macros could be made for manual stack alignment, but we
+   don't actually require it anywhere.*/
+# if defined(OC_X86_ASM)||defined(OC_ARM_ASM)
+#  if defined(__GNUC__)
+#   define OC_ALIGN8(expr) expr __attribute__((aligned(8)))
+#   define OC_ALIGN16(expr) expr __attribute__((aligned(16)))
+#  elif defined(_MSC_VER)
+#   define OC_ALIGN8(expr) __declspec (align(8)) expr
+#   define OC_ALIGN16(expr) __declspec (align(16)) expr
+#  else
+#   error "Alignment macros required for this platform."
+#  endif
+# endif
+# if !defined(OC_ALIGN8)
+#  define OC_ALIGN8(expr) expr
+# endif
+# if !defined(OC_ALIGN16)
+#  define OC_ALIGN16(expr) expr
+# endif
+
+
+
+/*This library's version.*/
+# define OC_VENDOR_STRING "Xiph.Org libtheora 1.2.0 20250329 (Ptalarbvorm)"
+
+/*Theora bitstream version.*/
+# define TH_VERSION_MAJOR (3)
+# define TH_VERSION_MINOR (2)
+# define TH_VERSION_SUB   (1)
+# define TH_VERSION_CHECK(_info,_maj,_min,_sub) \
+ ((_info)->version_major>(_maj)||(_info)->version_major==(_maj)&& \
+ ((_info)->version_minor>(_min)||(_info)->version_minor==(_min)&& \
+ (_info)->version_subminor>=(_sub)))
+
+
+
+/*A map from the index in the zig zag scan to the coefficient number in a
+   block.*/
+extern const unsigned char OC_FZIG_ZAG[128];
+/*A map from the coefficient number in a block to its index in the zig zag
+   scan.*/
+extern const unsigned char OC_IZIG_ZAG[64];
+/*A map from physical macro block ordering to bitstream macro block
+   ordering within a super block.*/
+extern const unsigned char OC_MB_MAP[2][2];
+/*A list of the indices in the oc_mb_map array that can be valid for each of
+   the various chroma decimation types.*/
+extern const unsigned char OC_MB_MAP_IDXS[TH_PF_NFORMATS][12];
+/*The number of indices in the oc_mb_map array that can be valid for each of
+   the various chroma decimation types.*/
+extern const unsigned char OC_MB_MAP_NIDXS[TH_PF_NFORMATS];
+
+
+
+int oc_ilog(unsigned _v);
+void *oc_aligned_malloc(size_t _sz,size_t _align);
+void oc_aligned_free(void *_ptr);
+void **oc_malloc_2d(size_t _height,size_t _width,size_t _sz);
+void **oc_calloc_2d(size_t _height,size_t _width,size_t _sz);
+void oc_free_2d(void *_ptr);
+
+void oc_ycbcr_buffer_flip(th_ycbcr_buffer _dst,
+ const th_ycbcr_buffer _src);
+
+#endif

thirdparty/libtheora/mathops.c

@@ -0,0 +1,314 @@
+#include "internal.h"
+#include "mathops.h"
+
+/*The fastest fallback strategy for platforms with fast multiplication appears
+   to be based on de Bruijn sequences~\cite{LP98}.
+  Define OC_ILOG_NODEBRUIJN to use a simpler fallback on platforms where
+   multiplication or table lookups are too expensive.
+
+  @UNPUBLISHED{LP98,
+    author="Charles E. Leiserson and Harald Prokop",
+    title="Using de {Bruijn} Sequences to Index a 1 in a Computer Word",
+    month=Jun,
+    year=1998,
+    note="\url{http://supertech.csail.mit.edu/papers/debruijn.pdf}"
+  }*/
+#if !defined(OC_ILOG_NODEBRUIJN)&&!defined(OC_CLZ32)
+static const unsigned char OC_DEBRUIJN_IDX32[32]={
+   0, 1,28, 2,29,14,24, 3,30,22,20,15,25,17, 4, 8,
+  31,27,13,23,21,19,16, 7,26,12,18, 6,11, 5,10, 9
+};
+#endif
+
+int oc_ilog32(ogg_uint32_t _v){
+#if defined(OC_CLZ32)
+  return _v ? (OC_CLZ32_OFFS-OC_CLZ32(_v)) : 0;
+#else
+/*On a Pentium M, this branchless version tested as the fastest version without
+   multiplications on 1,000,000,000 random 32-bit integers, edging out a
+   similar version with branches, and a 256-entry LUT version.*/
+# if defined(OC_ILOG_NODEBRUIJN)
+  int ret;
+  int m;
+  ret=_v>0;
+  m=(_v>0xFFFFU)<<4;
+  _v>>=m;
+  ret|=m;
+  m=(_v>0xFFU)<<3;
+  _v>>=m;
+  ret|=m;
+  m=(_v>0xFU)<<2;
+  _v>>=m;
+  ret|=m;
+  m=(_v>3)<<1;
+  _v>>=m;
+  ret|=m;
+  ret+=_v>1;
+  return ret;
+/*This de Bruijn sequence version is faster if you have a fast multiplier.*/
+# else
+  int ret;
+  _v|=_v>>1;
+  _v|=_v>>2;
+  _v|=_v>>4;
+  _v|=_v>>8;
+  _v|=_v>>16;
+  ret=_v&1;
+  _v=(_v>>1)+1;
+  ret+=OC_DEBRUIJN_IDX32[_v*0x77CB531U>>27&0x1F];
+  return ret;
+# endif
+#endif
+}
+
+int oc_ilog64(ogg_int64_t _v){
+#if defined(CLZ64)
+  return _v ? CLZ64_OFFS-CLZ64(_v) : 0;
+#else
+/*If we don't have a fast 64-bit word implementation, split it into two 32-bit
+   halves.*/
+# if defined(OC_ILOG_NODEBRUIJN)|| \
+ defined(OC_CLZ32)||LONG_MAX<9223372036854775807LL
+  ogg_uint32_t v;
+  int          ret;
+  int          m;
+  m=(_v>0xFFFFFFFFU)<<5;
+  v=(ogg_uint32_t)(_v>>m);
+#  if defined(OC_CLZ32)
+  ret=m+OC_CLZ32_OFFS-OC_CLZ32(v)&-!!v;
+#  elif defined(OC_ILOG_NODEBRUIJN)
+  ret=v>0|m;
+  m=(v>0xFFFFU)<<4;
+  v>>=m;
+  ret|=m;
+  m=(v>0xFFU)<<3;
+  v>>=m;
+  ret|=m;
+  m=(v>0xFU)<<2;
+  v>>=m;
+  ret|=m;
+  m=(v>3)<<1;
+  v>>=m;
+  ret|=m;
+  ret+=v>1;
+  return ret;
+#  else
+  v|=v>>1;
+  v|=v>>2;
+  v|=v>>4;
+  v|=v>>8;
+  v|=v>>16;
+  ret=v&1|m;
+  v=(v>>1)+1;
+  ret+=OC_DEBRUIJN_IDX32[v*0x77CB531U>>27&0x1F];
+#  endif
+  return ret;
+/*Otherwise do it in one 64-bit multiply.*/
+# else
+  static const unsigned char OC_DEBRUIJN_IDX64[64]={
+     0, 1, 2, 7, 3,13, 8,19, 4,25,14,28, 9,34,20,40,
+     5,17,26,38,15,46,29,48,10,31,35,54,21,50,41,57,
+    63, 6,12,18,24,27,33,39,16,37,45,47,30,53,49,56,
+    62,11,23,32,36,44,52,55,61,22,43,51,60,42,59,58
+  };
+  int ret;
+  _v|=_v>>1;
+  _v|=_v>>2;
+  _v|=_v>>4;
+  _v|=_v>>8;
+  _v|=_v>>16;
+  _v|=_v>>32;
+  ret=(int)_v&1;
+  _v=(_v>>1)+1;
+  ret+=OC_DEBRUIJN_IDX64[_v*0x218A392CD3D5DBF>>58&0x3F];
+  return ret;
+# endif
+#endif
+}
+
+/*round(2**(62+i)*atanh(2**(-(i+1)))/log(2))*/
+static const ogg_int64_t OC_ATANH_LOG2[32]={
+  0x32B803473F7AD0F4LL,0x2F2A71BD4E25E916LL,0x2E68B244BB93BA06LL,
+  0x2E39FB9198CE62E4LL,0x2E2E683F68565C8FLL,0x2E2B850BE2077FC1LL,
+  0x2E2ACC58FE7B78DBLL,0x2E2A9E2DE52FD5F2LL,0x2E2A92A338D53EECLL,
+  0x2E2A8FC08F5E19B6LL,0x2E2A8F07E51A485ELL,0x2E2A8ED9BA8AF388LL,
+  0x2E2A8ECE2FE7384ALL,0x2E2A8ECB4D3E4B1ALL,0x2E2A8ECA94940FE8LL,
+  0x2E2A8ECA6669811DLL,0x2E2A8ECA5ADEDD6ALL,0x2E2A8ECA57FC347ELL,
+  0x2E2A8ECA57438A43LL,0x2E2A8ECA57155FB4LL,0x2E2A8ECA5709D510LL,
+  0x2E2A8ECA5706F267LL,0x2E2A8ECA570639BDLL,0x2E2A8ECA57060B92LL,
+  0x2E2A8ECA57060008LL,0x2E2A8ECA5705FD25LL,0x2E2A8ECA5705FC6CLL,
+  0x2E2A8ECA5705FC3ELL,0x2E2A8ECA5705FC33LL,0x2E2A8ECA5705FC30LL,
+  0x2E2A8ECA5705FC2FLL,0x2E2A8ECA5705FC2FLL
+};
+
+/*Computes the binary exponential of _z, a log base 2 in Q57 format.*/
+ogg_int64_t oc_bexp64(ogg_int64_t _z){
+  ogg_int64_t w;
+  ogg_int64_t z;
+  int         ipart;
+  ipart=(int)(_z>>57);
+  if(ipart<0)return 0;
+  if(ipart>=63)return 0x7FFFFFFFFFFFFFFFLL;
+  z=_z-OC_Q57(ipart);
+  if(z){
+    ogg_int64_t mask;
+    long        wlo;
+    int         i;
+    /*C doesn't give us 64x64->128 muls, so we use CORDIC.
+      This is not particularly fast, but it's not being used in time-critical
+       code; it is very accurate.*/
+    /*z is the fractional part of the log in Q62 format.
+      We need 1 bit of headroom since the magnitude can get larger than 1
+       during the iteration, and a sign bit.*/
+    z*=32;
+    /*w is the exponential in Q61 format (since it also needs headroom and can
+       get as large as 2.0); we could get another bit if we dropped the sign,
+       but we'll recover that bit later anyway.
+      Ideally this should start out as
+        \lim_{n->\infty} 2^{61}/\product_{i=1}^n \sqrt{1-2^{-2i}}
+       but in order to guarantee convergence we have to repeat iterations 4,
+        13 (=3*4+1), and 40 (=3*13+1, etc.), so it winds up somewhat larger.*/
+    w=0x26A3D0E401DD846DLL;
+    for(i=0;;i++){
+      mask=-(z<0);
+      w+=(w>>i+1)+mask^mask;
+      z-=OC_ATANH_LOG2[i]+mask^mask;
+      /*Repeat iteration 4.*/
+      if(i>=3)break;
+      z*=2;
+    }
+    for(;;i++){
+      mask=-(z<0);
+      w+=(w>>i+1)+mask^mask;
+      z-=OC_ATANH_LOG2[i]+mask^mask;
+      /*Repeat iteration 13.*/
+      if(i>=12)break;
+      z*=2;
+    }
+    for(;i<32;i++){
+      mask=-(z<0);
+      w+=(w>>i+1)+mask^mask;
+      z=(z-(OC_ATANH_LOG2[i]+mask^mask))*2;
+    }
+    wlo=0;
+    /*Skip the remaining iterations unless we really require that much
+       precision.
+      We could have bailed out earlier for smaller iparts, but that would
+       require initializing w from a table, as the limit doesn't converge to
+       61-bit precision until n=30.*/
+    if(ipart>30){
+      /*For these iterations, we just update the low bits, as the high bits
+         can't possibly be affected.
+        OC_ATANH_LOG2 has also converged (it actually did so one iteration
+         earlier, but that's no reason for an extra special case).*/
+      for(;;i++){
+        mask=-(z<0);
+        wlo+=(w>>i)+mask^mask;
+        z-=OC_ATANH_LOG2[31]+mask^mask;
+        /*Repeat iteration 40.*/
+        if(i>=39)break;
+        z*=2;
+      }
+      for(;i<61;i++){
+        mask=-(z<0);
+        wlo+=(w>>i)+mask^mask;
+        z=(z-(OC_ATANH_LOG2[31]+mask^mask))*2;
+      }
+    }
+    w=(w<<1)+wlo;
+  }
+  else w=(ogg_int64_t)1<<62;
+  if(ipart<62)w=(w>>61-ipart)+1>>1;
+  return w;
+}
+
+/*Computes the binary logarithm of _w, returned in Q57 format.*/
+ogg_int64_t oc_blog64(ogg_int64_t _w){
+  ogg_int64_t z;
+  int         ipart;
+  if(_w<=0)return -1;
+  ipart=OC_ILOGNZ_64(_w)-1;
+  if(ipart>61)_w>>=ipart-61;
+  else _w<<=61-ipart;
+  z=0;
+  if(_w&_w-1){
+    ogg_int64_t x;
+    ogg_int64_t y;
+    ogg_int64_t u;
+    ogg_int64_t mask;
+    int         i;
+    /*C doesn't give us 64x64->128 muls, so we use CORDIC.
+      This is not particularly fast, but it's not being used in time-critical
+       code; it is very accurate.*/
+    /*z is the fractional part of the log in Q61 format.*/
+    /*x and y are the cosh() and sinh(), respectively, in Q61 format.
+      We are computing z=2*atanh(y/x)=2*atanh((_w-1)/(_w+1)).*/
+    x=_w+((ogg_int64_t)1<<61);
+    y=_w-((ogg_int64_t)1<<61);
+    for(i=0;i<4;i++){
+      mask=-(y<0);
+      z+=(OC_ATANH_LOG2[i]>>i)+mask^mask;
+      u=x>>i+1;
+      x-=(y>>i+1)+mask^mask;
+      y-=u+mask^mask;
+    }
+    /*Repeat iteration 4.*/
+    for(i--;i<13;i++){
+      mask=-(y<0);
+      z+=(OC_ATANH_LOG2[i]>>i)+mask^mask;
+      u=x>>i+1;
+      x-=(y>>i+1)+mask^mask;
+      y-=u+mask^mask;
+    }
+    /*Repeat iteration 13.*/
+    for(i--;i<32;i++){
+      mask=-(y<0);
+      z+=(OC_ATANH_LOG2[i]>>i)+mask^mask;
+      u=x>>i+1;
+      x-=(y>>i+1)+mask^mask;
+      y-=u+mask^mask;
+    }
+    /*OC_ATANH_LOG2 has converged.*/
+    for(;i<40;i++){
+      mask=-(y<0);
+      z+=(OC_ATANH_LOG2[31]>>i)+mask^mask;
+      u=x>>i+1;
+      x-=(y>>i+1)+mask^mask;
+      y-=u+mask^mask;
+    }
+    /*Repeat iteration 40.*/
+    for(i--;i<62;i++){
+      mask=-(y<0);
+      z+=(OC_ATANH_LOG2[31]>>i)+mask^mask;
+      u=x>>i+1;
+      x-=(y>>i+1)+mask^mask;
+      y-=u+mask^mask;
+    }
+    z=z+8>>4;
+  }
+  return OC_Q57(ipart)+z;
+}
+
+/*Polynomial approximation of a binary exponential.
+  Q10 input, Q0 output.*/
+ogg_uint32_t oc_bexp32_q10(int _z){
+  unsigned n;
+  int      ipart;
+  ipart=_z>>10;
+  n=(_z&(1<<10)-1)<<4;
+  n=(n*((n*((n*((n*3548>>15)+6817)>>15)+15823)>>15)+22708)>>15)+16384;
+  return 14-ipart>0?n+(1<<13-ipart)>>14-ipart:n<<ipart-14;
+}
+
+/*Polynomial approximation of a binary logarithm.
+  Q0 input, Q10 output.*/
+int oc_blog32_q10(ogg_uint32_t _w){
+  int n;
+  int ipart;
+  int fpart;
+  if(_w<=0)return -1;
+  ipart=OC_ILOGNZ_32(_w);
+  n=(ipart-16>0?_w>>ipart-16:_w<<16-ipart)-32768-16384;
+  fpart=(n*((n*((n*((n*-1402>>15)+2546)>>15)-5216)>>15)+15745)>>15)-6793;
+  return (ipart<<10)+(fpart>>4);
+}

thirdparty/libtheora/mathops.h

@@ -0,0 +1,143 @@
+#if !defined(_mathops_H)
+# define _mathops_H (1)
+# include <ogg/ogg.h>
+
+# if __GNUC_PREREQ(3,4)
+#  include <limits.h>
+/*Note the casts to (int) below: this prevents OC_CLZ{32|64}_OFFS from
+   "upgrading" the type of an entire expression to an (unsigned) size_t.*/
+#  if INT_MAX>=2147483647
+#   define OC_CLZ32_OFFS ((int)sizeof(unsigned)*CHAR_BIT)
+#   define OC_CLZ32(_x) (__builtin_clz(_x))
+#  elif LONG_MAX>=2147483647L
+#   define OC_CLZ32_OFFS ((int)sizeof(unsigned long)*CHAR_BIT)
+#   define OC_CLZ32(_x) (__builtin_clzl(_x))
+#  endif
+#  if INT_MAX>=9223372036854775807LL
+#   define OC_CLZ64_OFFS ((int)sizeof(unsigned)*CHAR_BIT)
+#   define OC_CLZ64(_x) (__builtin_clz(_x))
+#  elif LONG_MAX>=9223372036854775807LL
+#   define OC_CLZ64_OFFS ((int)sizeof(unsigned long)*CHAR_BIT)
+#   define OC_CLZ64(_x) (__builtin_clzl(_x))
+#  elif LLONG_MAX>=9223372036854775807LL|| \
+    __LONG_LONG_MAX__>=9223372036854775807LL
+#   define OC_CLZ64_OFFS ((int)sizeof(unsigned long long)*CHAR_BIT)
+#   define OC_CLZ64(_x) (__builtin_clzll(_x))
+#  endif
+# endif
+
+
+
+/**
+ * oc_ilog32 - Integer binary logarithm of a 32-bit value.
+ * @_v: A 32-bit value.
+ * Returns floor(log2(_v))+1, or 0 if _v==0.
+ * This is the number of bits that would be required to represent _v in two's
+ *  complement notation with all of the leading zeros stripped.
+ * The OC_ILOG_32() or OC_ILOGNZ_32() macros may be able to use a builtin
+ *  function instead, which should be faster.
+ */
+int oc_ilog32(ogg_uint32_t _v);
+/**
+ * oc_ilog64 - Integer binary logarithm of a 64-bit value.
+ * @_v: A 64-bit value.
+ * Returns floor(log2(_v))+1, or 0 if _v==0.
+ * This is the number of bits that would be required to represent _v in two's
+ *  complement notation with all of the leading zeros stripped.
+ * The OC_ILOG_64() or OC_ILOGNZ_64() macros may be able to use a builtin
+ *  function instead, which should be faster.
+ */
+int oc_ilog64(ogg_int64_t _v);
+
+
+# if defined(OC_CLZ32)
+/**
+ * OC_ILOGNZ_32 - Integer binary logarithm of a non-zero 32-bit value.
+ * @_v: A non-zero 32-bit value.
+ * Returns floor(log2(_v))+1.
+ * This is the number of bits that would be required to represent _v in two's
+ *  complement notation with all of the leading zeros stripped.
+ * If _v is zero, the return value is undefined; use OC_ILOG_32() instead.
+ */
+#  define OC_ILOGNZ_32(_v) (OC_CLZ32_OFFS-OC_CLZ32(_v))
+/**
+ * OC_ILOG_32 - Integer binary logarithm of a 32-bit value.
+ * @_v: A 32-bit value.
+ * Returns floor(log2(_v))+1, or 0 if _v==0.
+ * This is the number of bits that would be required to represent _v in two's
+ *  complement notation with all of the leading zeros stripped.
+ */
+#  define OC_ILOG_32(_v)   ((_v)?OC_ILOGNZ_32(_v):0)
+# else
+#  define OC_ILOGNZ_32(_v) (oc_ilog32(_v))
+#  define OC_ILOG_32(_v)   (oc_ilog32(_v))
+# endif
+
+# if defined(CLZ64)
+/**
+ * OC_ILOGNZ_64 - Integer binary logarithm of a non-zero 64-bit value.
+ * @_v: A non-zero 64-bit value.
+ * Returns floor(log2(_v))+1.
+ * This is the number of bits that would be required to represent _v in two's
+ *  complement notation with all of the leading zeros stripped.
+ * If _v is zero, the return value is undefined; use OC_ILOG_64() instead.
+ */
+#  define OC_ILOGNZ_64(_v) (CLZ64_OFFS-CLZ64(_v))
+/**
+ * OC_ILOG_64 - Integer binary logarithm of a 64-bit value.
+ * @_v: A 64-bit value.
+ * Returns floor(log2(_v))+1, or 0 if _v==0.
+ * This is the number of bits that would be required to represent _v in two's
+ *  complement notation with all of the leading zeros stripped.
+ */
+#  define OC_ILOG_64(_v)   ((_v)?OC_ILOGNZ_64(_v):0)
+# else
+#  define OC_ILOGNZ_64(_v) (oc_ilog64(_v))
+#  define OC_ILOG_64(_v)   (oc_ilog64(_v))
+# endif
+
+# define OC_STATIC_ILOG0(_v) (!!(_v))
+# define OC_STATIC_ILOG1(_v) (((_v)&0x2)?2:OC_STATIC_ILOG0(_v))
+# define OC_STATIC_ILOG2(_v) \
+ (((_v)&0xC)?2+OC_STATIC_ILOG1((_v)>>2):OC_STATIC_ILOG1(_v))
+# define OC_STATIC_ILOG3(_v) \
+ (((_v)&0xF0)?4+OC_STATIC_ILOG2((_v)>>4):OC_STATIC_ILOG2(_v))
+# define OC_STATIC_ILOG4(_v) \
+ (((_v)&0xFF00)?8+OC_STATIC_ILOG3((_v)>>8):OC_STATIC_ILOG3(_v))
+# define OC_STATIC_ILOG5(_v) \
+ (((_v)&0xFFFF0000)?16+OC_STATIC_ILOG4((_v)>>16):OC_STATIC_ILOG4(_v))
+# define OC_STATIC_ILOG6(_v) \
+ (((_v)&0xFFFFFFFF00000000ULL)?32+OC_STATIC_ILOG5((_v)>>32):OC_STATIC_ILOG5(_v))
+/**
+ * OC_STATIC_ILOG_32 - The integer logarithm of an (unsigned, 32-bit) constant.
+ * @_v: A non-negative 32-bit constant.
+ * Returns floor(log2(_v))+1, or 0 if _v==0.
+ * This is the number of bits that would be required to represent _v in two's
+ *  complement notation with all of the leading zeros stripped.
+ * This macro is suitable for evaluation at compile time, but it should not be
+ *  used on values that can change at runtime, as it operates via exhaustive
+ *  search.
+ */
+# define OC_STATIC_ILOG_32(_v) (OC_STATIC_ILOG5((ogg_uint32_t)(_v)))
+/**
+ * OC_STATIC_ILOG_64 - The integer logarithm of an (unsigned, 64-bit) constant.
+ * @_v: A non-negative 64-bit constant.
+ * Returns floor(log2(_v))+1, or 0 if _v==0.
+ * This is the number of bits that would be required to represent _v in two's
+ *  complement notation with all of the leading zeros stripped.
+ * This macro is suitable for evaluation at compile time, but it should not be
+ *  used on values that can change at runtime, as it operates via exhaustive
+ *  search.
+ */
+# define OC_STATIC_ILOG_64(_v) (OC_STATIC_ILOG6((ogg_int64_t)(_v)))
+
+#define OC_Q57(_v) ((_v)*((ogg_int64_t)1<<57))
+#define OC_Q10(_v) ((_v)<<10)
+
+ogg_int64_t oc_bexp64(ogg_int64_t _z);
+ogg_int64_t oc_blog64(ogg_int64_t _w);
+
+ogg_uint32_t oc_bexp32_q10(int _z);
+int oc_blog32_q10(ogg_uint32_t _w);
+
+#endif

thirdparty/libtheora/mcenc.c

@@ -0,0 +1,791 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation https://www.xiph.org/                 *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+#include <stdlib.h>
+#include <limits.h>
+#include <string.h>
+#include "encint.h"
+
+
+
+typedef struct oc_mcenc_ctx           oc_mcenc_ctx;
+
+
+
+/*Temporary state used for motion estimation.*/
+struct oc_mcenc_ctx{
+  /*The candidate motion vectors.*/
+  int                candidates[13][2];
+  /*The start of the Set B candidates.*/
+  int                setb0;
+  /*The total number of candidates.*/
+  int                ncandidates;
+};
+
+
+
+/*The maximum Y plane SAD value for accepting the median predictor.*/
+#define OC_YSAD_THRESH1            (256)
+/*The amount to right shift the minimum error by when inflating it for
+   computing the second maximum Y plane SAD threshold.*/
+#define OC_YSAD_THRESH2_SCALE_BITS (4)
+/*The amount to add to the second maximum Y plane threshold when inflating
+   it.*/
+#define OC_YSAD_THRESH2_OFFSET     (64)
+
+/*The vector offsets in the X direction for each search site in the square
+   pattern.*/
+static const int OC_SQUARE_DX[9]={-1,0,1,-1,0,1,-1,0,1};
+/*The vector offsets in the Y direction for each search site in the square
+   pattern.*/
+static const int OC_SQUARE_DY[9]={-1,-1,-1,0,0,0,1,1,1};
+/*The number of sites to search for each boundary condition in the square
+   pattern.
+  Bit flags for the boundary conditions are as follows:
+  1: -16==dx
+  2:      dx==15(.5)
+  4: -16==dy
+  8:      dy==15(.5)*/
+static const int OC_SQUARE_NSITES[11]={8,5,5,0,5,3,3,0,5,3,3};
+/*The list of sites to search for each boundary condition in the square
+   pattern.*/
+static const int OC_SQUARE_SITES[11][8]={
+  /* -15.5<dx<31,       -15.5<dy<15(.5)*/
+  {0,1,2,3,5,6,7,8},
+  /*-15.5==dx,          -15.5<dy<15(.5)*/
+  {1,2,5,7,8},
+  /*     dx==15(.5),    -15.5<dy<15(.5)*/
+  {0,1,3,6,7},
+  /*-15.5==dx==15(.5),  -15.5<dy<15(.5)*/
+  {-1},
+  /* -15.5<dx<15(.5),  -15.5==dy*/
+  {3,5,6,7,8},
+  /*-15.5==dx,         -15.5==dy*/
+  {5,7,8},
+  /*     dx==15(.5),   -15.5==dy*/
+  {3,6,7},
+  /*-15.5==dx==15(.5), -15.5==dy*/
+  {-1},
+  /*-15.5dx<15(.5),           dy==15(.5)*/
+  {0,1,2,3,5},
+  /*-15.5==dx,                dy==15(.5)*/
+  {1,2,5},
+  /*       dx==15(.5),        dy==15(.5)*/
+  {0,1,3}
+};
+
+
+static void oc_mcenc_find_candidates_a(oc_enc_ctx *_enc,oc_mcenc_ctx *_mcenc,
+ oc_mv _accum,int _mbi,int _frame){
+  oc_mb_enc_info *embs;
+  int             accum_x;
+  int             accum_y;
+  int             a[3][2];
+  int             ncandidates;
+  unsigned        nmbi;
+  int             i;
+  embs=_enc->mb_info;
+  /*Skip a position to store the median predictor in.*/
+  ncandidates=1;
+  if(embs[_mbi].ncneighbors>0){
+    /*Fill in the first part of set A: the vectors from adjacent blocks.*/
+    for(i=0;i<embs[_mbi].ncneighbors;i++){
+      nmbi=embs[_mbi].cneighbors[i];
+      _mcenc->candidates[ncandidates][0]=
+       OC_MV_X(embs[nmbi].analysis_mv[0][_frame]);
+      _mcenc->candidates[ncandidates][1]=
+       OC_MV_Y(embs[nmbi].analysis_mv[0][_frame]);
+      ncandidates++;
+    }
+  }
+  accum_x=OC_MV_X(_accum);
+  accum_y=OC_MV_Y(_accum);
+  /*Add a few additional vectors to set A: the vectors used in the previous
+     frames and the (0,0) vector.*/
+  _mcenc->candidates[ncandidates][0]=accum_x;
+  _mcenc->candidates[ncandidates][1]=accum_y;
+  ncandidates++;
+  _mcenc->candidates[ncandidates][0]=OC_CLAMPI(-31,
+   OC_MV_X(embs[_mbi].analysis_mv[1][_frame])+accum_x,31);
+  _mcenc->candidates[ncandidates][1]=OC_CLAMPI(-31,
+   OC_MV_Y(embs[_mbi].analysis_mv[1][_frame])+accum_y,31);
+  ncandidates++;
+  _mcenc->candidates[ncandidates][0]=0;
+  _mcenc->candidates[ncandidates][1]=0;
+  ncandidates++;
+  /*Use the first three vectors of set A to find our best predictor: their
+     median.*/
+  memcpy(a,_mcenc->candidates+1,sizeof(a));
+  OC_SORT2I(a[0][0],a[1][0]);
+  OC_SORT2I(a[0][1],a[1][1]);
+  OC_SORT2I(a[1][0],a[2][0]);
+  OC_SORT2I(a[1][1],a[2][1]);
+  OC_SORT2I(a[0][0],a[1][0]);
+  OC_SORT2I(a[0][1],a[1][1]);
+  _mcenc->candidates[0][0]=a[1][0];
+  _mcenc->candidates[0][1]=a[1][1];
+  _mcenc->setb0=ncandidates;
+}
+
+static void oc_mcenc_find_candidates_b(oc_enc_ctx *_enc,oc_mcenc_ctx *_mcenc,
+ oc_mv _accum,int _mbi,int _frame){
+  oc_mb_enc_info *embs;
+  int             accum_x;
+  int             accum_y;
+  int             ncandidates;
+  embs=_enc->mb_info;
+  accum_x=OC_MV_X(_accum);
+  accum_y=OC_MV_Y(_accum);
+  /*Fill in set B: accelerated predictors for this and adjacent macro blocks.*/
+  ncandidates=_mcenc->setb0;
+  /*Use only the current block. Using more did not appear to be helpful
+    with the current selection logic due to escaping the local search too
+    quickly.*/
+  _mcenc->candidates[ncandidates][0]=OC_CLAMPI(-31,
+   2*OC_MV_X(embs[_mbi].analysis_mv[1][_frame])
+   -OC_MV_X(embs[_mbi].analysis_mv[2][_frame])+accum_x,31);
+  _mcenc->candidates[ncandidates][1]=OC_CLAMPI(-31,
+   2*OC_MV_Y(embs[_mbi].analysis_mv[1][_frame])
+   -OC_MV_Y(embs[_mbi].analysis_mv[2][_frame])+accum_y,31);
+  ncandidates++;
+  _mcenc->ncandidates=ncandidates;
+}
+
+static unsigned oc_sad16_halfpel(const oc_enc_ctx *_enc,
+ const ptrdiff_t *_frag_buf_offs,const ptrdiff_t _fragis[4],
+ int _mvoffset0,int _mvoffset1,const unsigned char *_src,
+ const unsigned char *_ref,int _ystride,unsigned _best_err){
+  unsigned err;
+  int      bi;
+  err=0;
+  for(bi=0;bi<4;bi++){
+    ptrdiff_t frag_offs;
+    frag_offs=_frag_buf_offs[_fragis[bi]];
+    err+=oc_enc_frag_sad2_thresh(_enc,_src+frag_offs,_ref+frag_offs+_mvoffset0,
+     _ref+frag_offs+_mvoffset1,_ystride,_best_err-err);
+  }
+  return err;
+}
+
+static unsigned oc_satd16_halfpel(const oc_enc_ctx *_enc,
+ const ptrdiff_t *_frag_buf_offs,const ptrdiff_t _fragis[4],
+ int _mvoffset0,int _mvoffset1,const unsigned char *_src,
+ const unsigned char *_ref,int _ystride,unsigned _best_err){
+  unsigned err;
+  int      dc;
+  int      bi;
+  err=0;
+  for(bi=0;bi<4;bi++){
+    ptrdiff_t frag_offs;
+    frag_offs=_frag_buf_offs[_fragis[bi]];
+    err+=oc_enc_frag_satd2(_enc,&dc,_src+frag_offs,
+     _ref+frag_offs+_mvoffset0,_ref+frag_offs+_mvoffset1,_ystride);
+    err+=abs(dc);
+  }
+  return err;
+}
+
+static unsigned oc_mcenc_ysad_check_mbcandidate_fullpel(const oc_enc_ctx *_enc,
+ const ptrdiff_t *_frag_buf_offs,const ptrdiff_t _fragis[4],int _dx,int _dy,
+ const unsigned char *_src,const unsigned char *_ref,int _ystride,
+ unsigned _block_err[4]){
+  unsigned err;
+  int      mvoffset;
+  int      bi;
+  mvoffset=_dx+_dy*_ystride;
+  err=0;
+  for(bi=0;bi<4;bi++){
+    ptrdiff_t frag_offs;
+    unsigned  block_err;
+    frag_offs=_frag_buf_offs[_fragis[bi]];
+    block_err=oc_enc_frag_sad(_enc,
+     _src+frag_offs,_ref+frag_offs+mvoffset,_ystride);
+    _block_err[bi]=block_err;
+    err+=block_err;
+  }
+  return err;
+}
+
+static int oc_mcenc_ysatd_check_mbcandidate_fullpel(const oc_enc_ctx *_enc,
+ const ptrdiff_t *_frag_buf_offs,const ptrdiff_t _fragis[4],int _dx,int _dy,
+ const unsigned char *_src,const unsigned char *_ref,int _ystride){
+  int mvoffset;
+  int err;
+  int bi;
+  mvoffset=_dx+_dy*_ystride;
+  err=0;
+  for(bi=0;bi<4;bi++){
+    ptrdiff_t frag_offs;
+    int       dc;
+    frag_offs=_frag_buf_offs[_fragis[bi]];
+    if(_enc->sp_level<OC_SP_LEVEL_NOSATD){
+      err+=oc_enc_frag_satd(_enc,&dc,
+       _src+frag_offs,_ref+frag_offs+mvoffset,_ystride);
+      err+=abs(dc);
+    }
+    else{
+      err+=oc_enc_frag_sad(_enc,
+       _src+frag_offs,_ref+frag_offs+mvoffset,_ystride);
+    }
+  }
+  return err;
+}
+
+static unsigned oc_mcenc_ysatd_check_bcandidate_fullpel(const oc_enc_ctx *_enc,
+ ptrdiff_t _frag_offs,int _dx,int _dy,
+ const unsigned char *_src,const unsigned char *_ref,int _ystride){
+  unsigned err;
+  int      dc;
+  err=oc_enc_frag_satd(_enc,&dc,
+   _src+_frag_offs,_ref+_frag_offs+_dx+_dy*_ystride,_ystride);
+  return err+abs(dc);
+}
+
+/*Perform a motion vector search for this macro block against a single
+   reference frame.
+  As a bonus, individual block motion vectors are computed as well, as much of
+   the work can be shared.
+  The actual motion vector is stored in the appropriate place in the
+   oc_mb_enc_info structure.
+  _accum:      Drop frame/golden MV accumulators.
+  _mbi:        The macro block index.
+  _frame:      The frame to use for SATD calculations and refinement,
+                either OC_FRAME_PREV or OC_FRAME_GOLD.
+  _frame_full: The frame to perform the 1px search on, one of OC_FRAME_PREV,
+                OC_FRAME_GOLD, OC_FRAME_PREV_ORIG, or OC_FRAME_GOLD_ORIG.*/
+void oc_mcenc_search_frame(oc_enc_ctx *_enc,oc_mv _accum,int _mbi,int _frame,
+ int _frame_full){
+  /*Note: Traditionally this search is done using a rate-distortion objective
+     function of the form D+lambda*R.
+    However, xiphmont tested this and found it produced a small degradation,
+     while requiring extra computation.
+    This is most likely due to Theora's peculiar MV encoding scheme: MVs are
+     not coded relative to a predictor, and the only truly cheap way to use a
+     MV is in the LAST or LAST2 MB modes, which are not being considered here.
+    Therefore if we use the MV found here, it's only because both LAST and
+     LAST2 performed poorly, and therefore the MB is not likely to be uniform
+     or suffer from the aperture problem.
+    Furthermore we would like to reuse the MV found here for as many MBs as
+     possible, so picking a slightly sub-optimal vector to save a bit or two
+     may cause increased degradation in many blocks to come.
+    We could artificially reduce lambda to compensate, but it's faster to just
+     disable it entirely, and use D (the distortion) as the sole criterion.*/
+  oc_mcenc_ctx         mcenc;
+  const ptrdiff_t     *frag_buf_offs;
+  const ptrdiff_t     *fragis;
+  const unsigned char *src;
+  const unsigned char *ref;
+  const unsigned char *satd_ref;
+  int                  ystride;
+  oc_mb_enc_info      *embs;
+  ogg_int32_t          hit_cache[31];
+  ogg_int32_t          hitbit;
+  unsigned             best_block_err[4];
+  unsigned             block_err[4];
+  unsigned             best_err;
+  int                  best_vec[2];
+  int                  best_block_vec[4][2];
+  int                  candx;
+  int                  candy;
+  int                  bi;
+  embs=_enc->mb_info;
+  /*Find some candidate motion vectors.*/
+  oc_mcenc_find_candidates_a(_enc,&mcenc,_accum,_mbi,_frame);
+  /*Clear the cache of locations we've examined.*/
+  memset(hit_cache,0,sizeof(hit_cache));
+  /*Start with the median predictor.*/
+  candx=OC_DIV2(mcenc.candidates[0][0]);
+  candy=OC_DIV2(mcenc.candidates[0][1]);
+  hit_cache[candy+15]|=(ogg_int32_t)1<<candx+15;
+  frag_buf_offs=_enc->state.frag_buf_offs;
+  fragis=_enc->state.mb_maps[_mbi][0];
+  src=_enc->state.ref_frame_data[OC_FRAME_IO];
+  ref=_enc->state.ref_frame_data[_frame_full];
+  satd_ref=_enc->state.ref_frame_data[_frame];
+  ystride=_enc->state.ref_ystride[0];
+  /*TODO: customize error function for speed/(quality+size) tradeoff.*/
+  best_err=oc_mcenc_ysad_check_mbcandidate_fullpel(_enc,
+   frag_buf_offs,fragis,candx,candy,src,ref,ystride,block_err);
+  best_vec[0]=candx;
+  best_vec[1]=candy;
+  if(_frame==OC_FRAME_PREV){
+    for(bi=0;bi<4;bi++){
+      best_block_err[bi]=block_err[bi];
+      best_block_vec[bi][0]=candx;
+      best_block_vec[bi][1]=candy;
+    }
+  }
+  /*If this predictor fails, move on to set A.*/
+  if(best_err>OC_YSAD_THRESH1){
+    unsigned err;
+    unsigned t2;
+    int      ncs;
+    int      ci;
+    /*Compute the early termination threshold for set A.*/
+    t2=embs[_mbi].error[_frame];
+    ncs=OC_MINI(3,embs[_mbi].ncneighbors);
+    for(ci=0;ci<ncs;ci++){
+      t2=OC_MAXI(t2,embs[embs[_mbi].cneighbors[ci]].error[_frame]);
+    }
+    t2+=(t2>>OC_YSAD_THRESH2_SCALE_BITS)+OC_YSAD_THRESH2_OFFSET;
+    /*Examine the candidates in set A.*/
+    for(ci=1;ci<mcenc.setb0;ci++){
+      candx=OC_DIV2(mcenc.candidates[ci][0]);
+      candy=OC_DIV2(mcenc.candidates[ci][1]);
+      /*If we've already examined this vector, then we would be using it if it
+         was better than what we are using.*/
+      hitbit=(ogg_int32_t)1<<candx+15;
+      if(hit_cache[candy+15]&hitbit)continue;
+      hit_cache[candy+15]|=hitbit;
+      err=oc_mcenc_ysad_check_mbcandidate_fullpel(_enc,
+       frag_buf_offs,fragis,candx,candy,src,ref,ystride,block_err);
+      if(err<best_err){
+        best_err=err;
+        best_vec[0]=candx;
+        best_vec[1]=candy;
+      }
+      if(_frame==OC_FRAME_PREV){
+        for(bi=0;bi<4;bi++)if(block_err[bi]<best_block_err[bi]){
+          best_block_err[bi]=block_err[bi];
+          best_block_vec[bi][0]=candx;
+          best_block_vec[bi][1]=candy;
+        }
+      }
+    }
+    if(best_err>t2){
+      oc_mcenc_find_candidates_b(_enc,&mcenc,_accum,_mbi,_frame);
+      /*Examine the candidates in set B.*/
+      for(;ci<mcenc.ncandidates;ci++){
+        candx=OC_DIV2(mcenc.candidates[ci][0]);
+        candy=OC_DIV2(mcenc.candidates[ci][1]);
+        hitbit=(ogg_int32_t)1<<candx+15;
+        if(hit_cache[candy+15]&hitbit)continue;
+        hit_cache[candy+15]|=hitbit;
+        err=oc_mcenc_ysad_check_mbcandidate_fullpel(_enc,
+         frag_buf_offs,fragis,candx,candy,src,ref,ystride,block_err);
+        if(err<best_err){
+          best_err=err;
+          best_vec[0]=candx;
+          best_vec[1]=candy;
+        }
+        if(_frame==OC_FRAME_PREV){
+          for(bi=0;bi<4;bi++)if(block_err[bi]<best_block_err[bi]){
+            best_block_err[bi]=block_err[bi];
+            best_block_vec[bi][0]=candx;
+            best_block_vec[bi][1]=candy;
+          }
+        }
+      }
+      /*Use the same threshold for set B as in set A.*/
+      if(best_err>t2){
+        int best_site;
+        int nsites;
+        int sitei;
+        int site;
+        int b;
+        /*Square pattern search.*/
+        for(;;){
+          best_site=4;
+          /*Compose the bit flags for boundary conditions.*/
+          b=OC_DIV16(-best_vec[0]+1)|OC_DIV16(best_vec[0]+1)<<1|
+           OC_DIV16(-best_vec[1]+1)<<2|OC_DIV16(best_vec[1]+1)<<3;
+          nsites=OC_SQUARE_NSITES[b];
+          for(sitei=0;sitei<nsites;sitei++){
+            site=OC_SQUARE_SITES[b][sitei];
+            candx=best_vec[0]+OC_SQUARE_DX[site];
+            candy=best_vec[1]+OC_SQUARE_DY[site];
+            hitbit=(ogg_int32_t)1<<candx+15;
+            if(hit_cache[candy+15]&hitbit)continue;
+            hit_cache[candy+15]|=hitbit;
+            err=oc_mcenc_ysad_check_mbcandidate_fullpel(_enc,
+             frag_buf_offs,fragis,candx,candy,src,ref,ystride,block_err);
+            if(err<best_err){
+              best_err=err;
+              best_site=site;
+            }
+            if(_frame==OC_FRAME_PREV){
+              for(bi=0;bi<4;bi++)if(block_err[bi]<best_block_err[bi]){
+                best_block_err[bi]=block_err[bi];
+                best_block_vec[bi][0]=candx;
+                best_block_vec[bi][1]=candy;
+              }
+            }
+          }
+          if(best_site==4)break;
+          best_vec[0]+=OC_SQUARE_DX[best_site];
+          best_vec[1]+=OC_SQUARE_DY[best_site];
+        }
+        /*Final 4-MV search.*/
+        /*Simply use 1/4 of the macro block set A and B threshold as the
+           individual block threshold.*/
+        if(_frame==OC_FRAME_PREV){
+          t2>>=2;
+          for(bi=0;bi<4;bi++){
+            if(best_block_err[bi]>t2){
+              /*Square pattern search.
+                We do this in a slightly interesting manner.
+                We continue to check the SAD of all four blocks in the
+                 macro block.
+                This gives us two things:
+                 1) We can continue to use the hit_cache to avoid duplicate
+                     checks.
+                    Otherwise we could continue to read it, but not write to it
+                     without saving and restoring it for each block.
+                    Note that we could still eliminate a large number of
+                     duplicate checks by taking into account the site we came
+                     from when choosing the site list.
+                    We can still do that to avoid extra hit_cache queries, and
+                     it might even be a speed win.
+                 2) It gives us a slightly better chance of escaping local
+                     minima.
+                    We would not be here if we weren't doing a fairly bad job
+                     in finding a good vector, and checking these vectors can
+                     save us from 100 to several thousand points off our SAD 1
+                     in 15 times.
+                TODO: Is this a good idea?
+                Who knows.
+                It needs more testing.*/
+              for(;;){
+                int bestx;
+                int besty;
+                int bj;
+                bestx=best_block_vec[bi][0];
+                besty=best_block_vec[bi][1];
+                /*Compose the bit flags for boundary conditions.*/
+                b=OC_DIV16(-bestx+1)|OC_DIV16(bestx+1)<<1|
+                 OC_DIV16(-besty+1)<<2|OC_DIV16(besty+1)<<3;
+                nsites=OC_SQUARE_NSITES[b];
+                for(sitei=0;sitei<nsites;sitei++){
+                  site=OC_SQUARE_SITES[b][sitei];
+                  candx=bestx+OC_SQUARE_DX[site];
+                  candy=besty+OC_SQUARE_DY[site];
+                  hitbit=(ogg_int32_t)1<<candx+15;
+                  if(hit_cache[candy+15]&hitbit)continue;
+                  hit_cache[candy+15]|=hitbit;
+                  err=oc_mcenc_ysad_check_mbcandidate_fullpel(_enc,
+                   frag_buf_offs,fragis,candx,candy,src,ref,ystride,block_err);
+                  if(err<best_err){
+                    best_err=err;
+                    best_vec[0]=candx;
+                    best_vec[1]=candy;
+                  }
+                  for(bj=0;bj<4;bj++)if(block_err[bj]<best_block_err[bj]){
+                    best_block_err[bj]=block_err[bj];
+                    best_block_vec[bj][0]=candx;
+                    best_block_vec[bj][1]=candy;
+                  }
+                }
+                if(best_block_vec[bi][0]==bestx&&best_block_vec[bi][1]==besty){
+                  break;
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+  embs[_mbi].error[_frame]=(ogg_uint16_t)best_err;
+  candx=best_vec[0];
+  candy=best_vec[1];
+  embs[_mbi].satd[_frame]=oc_mcenc_ysatd_check_mbcandidate_fullpel(_enc,
+   frag_buf_offs,fragis,candx,candy,src,satd_ref,ystride);
+  embs[_mbi].analysis_mv[0][_frame]=OC_MV(candx<<1,candy<<1);
+  if(_frame==OC_FRAME_PREV&&_enc->sp_level<OC_SP_LEVEL_FAST_ANALYSIS){
+    for(bi=0;bi<4;bi++){
+      candx=best_block_vec[bi][0];
+      candy=best_block_vec[bi][1];
+      embs[_mbi].block_satd[bi]=oc_mcenc_ysatd_check_bcandidate_fullpel(_enc,
+       frag_buf_offs[fragis[bi]],candx,candy,src,satd_ref,ystride);
+      embs[_mbi].block_mv[bi]=OC_MV(candx<<1,candy<<1);
+    }
+  }
+}
+
+void oc_mcenc_search(oc_enc_ctx *_enc,int _mbi){
+  oc_mv2 *mvs;
+  oc_mv   accum_p;
+  oc_mv   accum_g;
+  oc_mv   mv2_p;
+  mvs=_enc->mb_info[_mbi].analysis_mv;
+  if(_enc->prevframe_dropped)accum_p=mvs[0][OC_FRAME_PREV];
+  else accum_p=0;
+  accum_g=mvs[2][OC_FRAME_GOLD];
+  /*Move the motion vector predictors back a frame.*/
+  mv2_p=mvs[2][OC_FRAME_PREV];
+  mvs[2][OC_FRAME_GOLD]=mvs[1][OC_FRAME_GOLD];
+  mvs[2][OC_FRAME_PREV]=mvs[1][OC_FRAME_PREV];
+  mvs[1][OC_FRAME_GOLD]=mvs[0][OC_FRAME_GOLD];
+  mvs[1][OC_FRAME_PREV]=OC_MV_SUB(mvs[0][OC_FRAME_PREV],mv2_p);
+  /*Search the last frame.*/
+  oc_mcenc_search_frame(_enc,accum_p,_mbi,OC_FRAME_PREV,OC_FRAME_PREV_ORIG);
+  mvs[2][OC_FRAME_PREV]=accum_p;
+  /*GOLDEN MVs are different from PREV MVs in that they're each absolute
+     offsets from some frame in the past rather than relative offsets from the
+     frame before.
+    For predictor calculation to make sense, we need them to be in the same
+     form as PREV MVs.*/
+  mvs[1][OC_FRAME_GOLD]=OC_MV_SUB(mvs[1][OC_FRAME_GOLD],mvs[2][OC_FRAME_GOLD]);
+  mvs[2][OC_FRAME_GOLD]=OC_MV_SUB(mvs[2][OC_FRAME_GOLD],accum_g);
+  /*Search the golden frame.*/
+  oc_mcenc_search_frame(_enc,accum_g,_mbi,OC_FRAME_GOLD,OC_FRAME_GOLD_ORIG);
+  /*Put GOLDEN MVs back into absolute offset form.
+    The newest MV is already an absolute offset.*/
+  mvs[2][OC_FRAME_GOLD]=OC_MV_ADD(mvs[2][OC_FRAME_GOLD],accum_g);
+  mvs[1][OC_FRAME_GOLD]=OC_MV_ADD(mvs[1][OC_FRAME_GOLD],mvs[2][OC_FRAME_GOLD]);
+}
+
+#if 0
+static int oc_mcenc_ysad_halfpel_mbrefine(const oc_enc_ctx *_enc,int _mbi,
+ int _vec[2],int _best_err,int _frame){
+  const unsigned char *src;
+  const unsigned char *ref;
+  const ptrdiff_t     *frag_buf_offs;
+  const ptrdiff_t     *fragis;
+  int                  offset_y[9];
+  int                  ystride;
+  int                  mvoffset_base;
+  int                  best_site;
+  int                  sitei;
+  int                  err;
+  src=_enc->state.ref_frame_data[OC_FRAME_IO];
+  ref=_enc->state.ref_frame_data[_framei];
+  frag_buf_offs=_enc->state.frag_buf_offs;
+  fragis=_enc->state.mb_maps[_mbi][0];
+  ystride=_enc->state.ref_ystride[0];
+  mvoffset_base=_vec[0]+_vec[1]*ystride;
+  offset_y[0]=offset_y[1]=offset_y[2]=-ystride;
+  offset_y[3]=offset_y[5]=0;
+  offset_y[6]=offset_y[7]=offset_y[8]=ystride;
+  best_site=4;
+  for(sitei=0;sitei<8;sitei++){
+    int site;
+    int xmask;
+    int ymask;
+    int dx;
+    int dy;
+    int mvoffset0;
+    int mvoffset1;
+    site=OC_SQUARE_SITES[0][sitei];
+    dx=OC_SQUARE_DX[site];
+    dy=OC_SQUARE_DY[site];
+    /*The following code SHOULD be equivalent to
+        oc_state_get_mv_offsets(&_mcenc->enc.state,&mvoffset0,&mvoffset1,
+         (_vec[0]<<1)+dx,(_vec[1]<<1)+dy,ref_ystride,0);
+      However, it should also be much faster, as it involves no multiplies and
+       doesn't have to handle chroma vectors.*/
+    xmask=OC_SIGNMASK(((_vec[0]<<1)+dx)^dx);
+    ymask=OC_SIGNMASK(((_vec[1]<<1)+dy)^dy);
+    mvoffset0=mvoffset_base+(dx&xmask)+(offset_y[site]&ymask);
+    mvoffset1=mvoffset_base+(dx&~xmask)+(offset_y[site]&~ymask);
+    err=oc_sad16_halfpel(_enc,frag_buf_offs,fragis,
+     mvoffset0,mvoffset1,src,ref,ystride,_best_err);
+    if(err<_best_err){
+      _best_err=err;
+      best_site=site;
+    }
+  }
+  _vec[0]=(_vec[0]<<1)+OC_SQUARE_DX[best_site];
+  _vec[1]=(_vec[1]<<1)+OC_SQUARE_DY[best_site];
+  return _best_err;
+}
+#endif
+
+static unsigned oc_mcenc_ysatd_halfpel_mbrefine(const oc_enc_ctx *_enc,
+ int _mbi,int _vec[2],unsigned _best_err,int _frame){
+  const unsigned char *src;
+  const unsigned char *ref;
+  const ptrdiff_t     *frag_buf_offs;
+  const ptrdiff_t     *fragis;
+  int                  offset_y[9];
+  int                  ystride;
+  int                  mvoffset_base;
+  int                  best_site;
+  int                  sitei;
+  int                  err;
+  src=_enc->state.ref_frame_data[OC_FRAME_IO];
+  ref=_enc->state.ref_frame_data[_frame];
+  frag_buf_offs=_enc->state.frag_buf_offs;
+  fragis=_enc->state.mb_maps[_mbi][0];
+  ystride=_enc->state.ref_ystride[0];
+  mvoffset_base=_vec[0]+_vec[1]*ystride;
+  offset_y[0]=offset_y[1]=offset_y[2]=-ystride;
+  offset_y[3]=offset_y[5]=0;
+  offset_y[6]=offset_y[7]=offset_y[8]=ystride;
+  best_site=4;
+  for(sitei=0;sitei<8;sitei++){
+    int site;
+    int xmask;
+    int ymask;
+    int dx;
+    int dy;
+    int mvoffset0;
+    int mvoffset1;
+    site=OC_SQUARE_SITES[0][sitei];
+    dx=OC_SQUARE_DX[site];
+    dy=OC_SQUARE_DY[site];
+    /*The following code SHOULD be equivalent to
+        oc_state_get_mv_offsets(&_mcenc->enc.state,&mvoffset0,&mvoffset1,
+         (_vec[0]<<1)+dx,(_vec[1]<<1)+dy,ref_ystride,0);
+      However, it should also be much faster, as it involves no multiplies and
+       doesn't have to handle chroma vectors.*/
+    xmask=OC_SIGNMASK(((_vec[0]<<1)+dx)^dx);
+    ymask=OC_SIGNMASK(((_vec[1]<<1)+dy)^dy);
+    mvoffset0=mvoffset_base+(dx&xmask)+(offset_y[site]&ymask);
+    mvoffset1=mvoffset_base+(dx&~xmask)+(offset_y[site]&~ymask);
+    if(_enc->sp_level<OC_SP_LEVEL_NOSATD){
+      err=oc_satd16_halfpel(_enc,frag_buf_offs,fragis,
+       mvoffset0,mvoffset1,src,ref,ystride,_best_err);
+    }
+    else{
+      err=oc_sad16_halfpel(_enc,frag_buf_offs,fragis,
+           mvoffset0,mvoffset1,src,ref,ystride,_best_err);
+    }
+    if(err<_best_err){
+      _best_err=err;
+      best_site=site;
+    }
+  }
+  _vec[0]=(_vec[0]<<1)+OC_SQUARE_DX[best_site];
+  _vec[1]=(_vec[1]<<1)+OC_SQUARE_DY[best_site];
+  return _best_err;
+}
+
+void oc_mcenc_refine1mv(oc_enc_ctx *_enc,int _mbi,int _frame){
+  oc_mb_enc_info *embs;
+  int             vec[2];
+  embs=_enc->mb_info;
+  vec[0]=OC_DIV2(OC_MV_X(embs[_mbi].analysis_mv[0][_frame]));
+  vec[1]=OC_DIV2(OC_MV_Y(embs[_mbi].analysis_mv[0][_frame]));
+  embs[_mbi].satd[_frame]=oc_mcenc_ysatd_halfpel_mbrefine(_enc,
+   _mbi,vec,embs[_mbi].satd[_frame],_frame);
+  embs[_mbi].analysis_mv[0][_frame]=OC_MV(vec[0],vec[1]);
+}
+
+#if 0
+static int oc_mcenc_ysad_halfpel_brefine(const oc_enc_ctx *_enc,
+ int _vec[2],const unsigned char *_src,const unsigned char *_ref,int _ystride,
+ int _offset_y[9],unsigned _best_err){
+  int mvoffset_base;
+  int best_site;
+  int sitei;
+  mvoffset_base=_vec[0]+_vec[1]*_ystride;
+  best_site=4;
+  for(sitei=0;sitei<8;sitei++){
+    unsigned err;
+    int      site;
+    int      xmask;
+    int      ymask;
+    int      dx;
+    int      dy;
+    int      mvoffset0;
+    int      mvoffset1;
+    site=OC_SQUARE_SITES[0][sitei];
+    dx=OC_SQUARE_DX[site];
+    dy=OC_SQUARE_DY[site];
+    /*The following code SHOULD be equivalent to
+        oc_state_get_mv_offsets(&_mcenc->enc.state,&mvoffset0,&mvoffset1,
+         (_vec[0]<<1)+dx,(_vec[1]<<1)+dy,ref_ystride,0);
+      However, it should also be much faster, as it involves no multiplies and
+       doesn't have to handle chroma vectors.*/
+    xmask=OC_SIGNMASK(((_vec[0]<<1)+dx)^dx);
+    ymask=OC_SIGNMASK(((_vec[1]<<1)+dy)^dy);
+    mvoffset0=mvoffset_base+(dx&xmask)+(_offset_y[site]&ymask);
+    mvoffset1=mvoffset_base+(dx&~xmask)+(_offset_y[site]&~ymask);
+    err=oc_enc_frag_sad2_thresh(_enc,_src,
+     _ref+mvoffset0,_ref+mvoffset1,ystride,_best_err);
+    if(err<_best_err){
+      _best_err=err;
+      best_site=site;
+    }
+  }
+  _vec[0]=(_vec[0]<<1)+OC_SQUARE_DX[best_site];
+  _vec[1]=(_vec[1]<<1)+OC_SQUARE_DY[best_site];
+  return _best_err;
+}
+#endif
+
+static unsigned oc_mcenc_ysatd_halfpel_brefine(const oc_enc_ctx *_enc,
+ int _vec[2],const unsigned char *_src,const unsigned char *_ref,int _ystride,
+ int _offset_y[9],unsigned _best_err){
+  int mvoffset_base;
+  int best_site;
+  int sitei;
+  mvoffset_base=_vec[0]+_vec[1]*_ystride;
+  best_site=4;
+  for(sitei=0;sitei<8;sitei++){
+    unsigned err;
+    int      dc;
+    int      site;
+    int      xmask;
+    int      ymask;
+    int      dx;
+    int      dy;
+    int      mvoffset0;
+    int      mvoffset1;
+    site=OC_SQUARE_SITES[0][sitei];
+    dx=OC_SQUARE_DX[site];
+    dy=OC_SQUARE_DY[site];
+    /*The following code SHOULD be equivalent to
+        oc_state_get_mv_offsets(&_enc->state,&mvoffsets,0,
+         (_vec[0]<<1)+dx,(_vec[1]<<1)+dy);
+      However, it should also be much faster, as it involves no multiplies and
+       doesn't have to handle chroma vectors.*/
+    xmask=OC_SIGNMASK(((_vec[0]<<1)+dx)^dx);
+    ymask=OC_SIGNMASK(((_vec[1]<<1)+dy)^dy);
+    mvoffset0=mvoffset_base+(dx&xmask)+(_offset_y[site]&ymask);
+    mvoffset1=mvoffset_base+(dx&~xmask)+(_offset_y[site]&~ymask);
+    err=oc_enc_frag_satd2(_enc,&dc,_src,
+     _ref+mvoffset0,_ref+mvoffset1,_ystride);
+    err+=abs(dc);
+    if(err<_best_err){
+      _best_err=err;
+      best_site=site;
+    }
+  }
+  _vec[0]=(_vec[0]<<1)+OC_SQUARE_DX[best_site];
+  _vec[1]=(_vec[1]<<1)+OC_SQUARE_DY[best_site];
+  return _best_err;
+}
+
+void oc_mcenc_refine4mv(oc_enc_ctx *_enc,int _mbi){
+  oc_mb_enc_info      *embs;
+  const ptrdiff_t     *frag_buf_offs;
+  const ptrdiff_t     *fragis;
+  const unsigned char *src;
+  const unsigned char *ref;
+  int                  offset_y[9];
+  int                  ystride;
+  int                  bi;
+  ystride=_enc->state.ref_ystride[0];
+  frag_buf_offs=_enc->state.frag_buf_offs;
+  fragis=_enc->state.mb_maps[_mbi][0];
+  src=_enc->state.ref_frame_data[OC_FRAME_IO];
+  ref=_enc->state.ref_frame_data[OC_FRAME_PREV];
+  offset_y[0]=offset_y[1]=offset_y[2]=-ystride;
+  offset_y[3]=offset_y[5]=0;
+  offset_y[6]=offset_y[7]=offset_y[8]=ystride;
+  embs=_enc->mb_info;
+  for(bi=0;bi<4;bi++){
+    ptrdiff_t frag_offs;
+    int       vec[2];
+    frag_offs=frag_buf_offs[fragis[bi]];
+    vec[0]=OC_DIV2(OC_MV_X(embs[_mbi].block_mv[bi]));
+    vec[1]=OC_DIV2(OC_MV_Y(embs[_mbi].block_mv[bi]));
+    embs[_mbi].block_satd[bi]=oc_mcenc_ysatd_halfpel_brefine(_enc,vec,
+     src+frag_offs,ref+frag_offs,ystride,offset_y,embs[_mbi].block_satd[bi]);
+    embs[_mbi].ref_mv[bi]=OC_MV(vec[0],vec[1]);
+  }
+}

thirdparty/libtheora/ocintrin.h

@@ -0,0 +1,128 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+/*Some common macros for potential platform-specific optimization.*/
+#include <math.h>
+#if !defined(_ocintrin_H)
+# define _ocintrin_H (1)
+
+/*Some specific platforms may have optimized intrinsic or inline assembly
+   versions of these functions which can substantially improve performance.
+  We define macros for them to allow easy incorporation of these non-ANSI
+   features.*/
+
+/*Note that we do not provide a macro for abs(), because it is provided as a
+   library function, which we assume is translated into an intrinsic to avoid
+   the function call overhead and then implemented in the smartest way for the
+   target platform.
+  With modern gcc (4.x), this is true: it uses cmov instructions if the
+   architecture supports it and branchless bit-twiddling if it does not (the
+   speed difference between the two approaches is not measurable).
+  Interestingly, the bit-twiddling method was patented in 2000 (US 6,073,150)
+   by Sun Microsystems, despite prior art dating back to at least 1996:
+   https://web.archive.org/web/19961201174141/www.x86.org/ftp/articles/pentopt/PENTOPT.TXT
+  On gcc 3.x, however, our assumption is not true, as abs() is translated to a
+   conditional jump, which is horrible on deeply piplined architectures (e.g.,
+   all consumer architectures for the past decade or more).
+  Also be warned that -C*abs(x) where C is a constant is mis-optimized as
+   abs(C*x) on every gcc release before 4.2.3.
+  See bug https://gcc.gnu.org/bugzilla/show_bug.cgi?id=34130 */
+
+/*Modern gcc (4.x) can compile the naive versions of min and max with cmov if
+   given an appropriate architecture, but the branchless bit-twiddling versions
+   are just as fast, and do not require any special target architecture.
+  Earlier gcc versions (3.x) compiled both code to the same assembly
+   instructions, because of the way they represented ((_b)>(_a)) internally.*/
+#define OC_MAXI(_a,_b)      ((_a)-((_a)-(_b)&-((_b)>(_a))))
+#define OC_MINI(_a,_b)      ((_a)+((_b)-(_a)&-((_b)<(_a))))
+/*Clamps an integer into the given range.
+  If _a>_c, then the lower bound _a is respected over the upper bound _c (this
+   behavior is required to meet our documented API behavior).
+  _a: The lower bound.
+  _b: The value to clamp.
+  _c: The upper boud.*/
+#define OC_CLAMPI(_a,_b,_c) (OC_MAXI(_a,OC_MINI(_b,_c)))
+#define OC_CLAMP255(_x)     ((unsigned char)((((_x)<0)-1)&((_x)|-((_x)>255))))
+/*This has a chance of compiling branchless, and is just as fast as the
+   bit-twiddling method, which is slightly less portable, since it relies on a
+   sign-extended rightshift, which is not guaranteed by ANSI (but present on
+   every relevant platform).*/
+#define OC_SIGNI(_a)        (((_a)>0)-((_a)<0))
+/*Slightly more portable than relying on a sign-extended right-shift (which is
+   not guaranteed by ANSI), and just as fast, since gcc (3.x and 4.x both)
+   compile it into the right-shift anyway.*/
+#define OC_SIGNMASK(_a)     (-((_a)<0))
+/*Divides an integer by a power of two, truncating towards 0.
+  _dividend: The integer to divide.
+  _shift:    The non-negative power of two to divide by.
+  _rmask:    (1<<_shift)-1*/
+#define OC_DIV_POW2(_dividend,_shift,_rmask)\
+  ((_dividend)+(OC_SIGNMASK(_dividend)&(_rmask))>>(_shift))
+/*Divides _x by 65536, truncating towards 0.*/
+#define OC_DIV2_16(_x) OC_DIV_POW2(_x,16,0xFFFF)
+/*Divides _x by 2, truncating towards 0.*/
+#define OC_DIV2(_x) OC_DIV_POW2(_x,1,0x1)
+/*Divides _x by 8, truncating towards 0.*/
+#define OC_DIV8(_x) OC_DIV_POW2(_x,3,0x7)
+/*Divides _x by 16, truncating towards 0.*/
+#define OC_DIV16(_x) OC_DIV_POW2(_x,4,0xF)
+/*Right shifts _dividend by _shift, adding _rval, and subtracting one for
+   negative dividends first.
+  When _rval is (1<<_shift-1), this is equivalent to division with rounding
+   ties away from zero.*/
+#define OC_DIV_ROUND_POW2(_dividend,_shift,_rval)\
+  ((_dividend)+OC_SIGNMASK(_dividend)+(_rval)>>(_shift))
+/*Divides a _x by 2, rounding towards even numbers.*/
+#define OC_DIV2_RE(_x) ((_x)+((_x)>>1&1)>>1)
+/*Divides a _x by (1<<(_shift)), rounding towards even numbers.*/
+#define OC_DIV_POW2_RE(_x,_shift) \
+  ((_x)+((_x)>>(_shift)&1)+((1<<(_shift))-1>>1)>>(_shift))
+/*Swaps two integers _a and _b if _a>_b.*/
+#define OC_SORT2I(_a,_b) \
+  do{ \
+    int t__; \
+    t__=((_a)^(_b))&-((_b)<(_a)); \
+    (_a)^=t__; \
+    (_b)^=t__; \
+  } \
+  while(0)
+
+/*Accesses one of four (signed) bytes given an index.
+  This can be used to avoid small lookup tables.*/
+#define OC_BYTE_TABLE32(_a,_b,_c,_d,_i) \
+  ((signed char) \
+   (((_a)&0xFF|((_b)&0xFF)<<8|((_c)&0xFF)<<16|((_d)&0xFF)<<24)>>(_i)*8))
+/*Accesses one of eight (unsigned) nibbles given an index.
+  This can be used to avoid small lookup tables.*/
+#define OC_UNIBBLE_TABLE32(_a,_b,_c,_d,_e,_f,_g,_h,_i) \
+  ((((_a)&0xF|((_b)&0xF)<<4|((_c)&0xF)<<8|((_d)&0xF)<<12| \
+   ((_e)&0xF)<<16|((_f)&0xF)<<20|((_g)&0xF)<<24|((_h)&0xF)<<28)>>(_i)*4)&0xF)
+
+
+
+/*All of these macros should expect floats as arguments.*/
+#define OC_MAXF(_a,_b)      ((_a)<(_b)?(_b):(_a))
+#define OC_MINF(_a,_b)      ((_a)>(_b)?(_b):(_a))
+#define OC_CLAMPF(_a,_b,_c) (OC_MINF(_a,OC_MAXF(_b,_c)))
+#define OC_FABSF(_f)        ((float)fabs(_f))
+#define OC_SQRTF(_f)        ((float)sqrt(_f))
+#define OC_POWF(_b,_e)      ((float)pow(_b,_e))
+#define OC_LOGF(_f)         ((float)log(_f))
+#define OC_IFLOORF(_f)      ((int)floor(_f))
+#define OC_ICEILF(_f)       ((int)ceil(_f))
+
+#endif

thirdparty/libtheora/quant.c

@@ -0,0 +1,127 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+#include <stdlib.h>
+#include <string.h>
+#include <ogg/ogg.h>
+#include "quant.h"
+#include "decint.h"
+
+/*The maximum output of the DCT with +/- 255 inputs is +/- 8157.
+  These minimum quantizers ensure the result after quantization (and after
+   prediction for DC) will be no more than +/- 510.
+  The tokenization system can handle values up to +/- 580, so there is no need
+   to do any coefficient clamping.
+  I would rather have allowed smaller quantizers and had to clamp, but these
+   minimums were required when constructing the original VP3 matrices and have
+   been formalized in the spec.*/
+static const unsigned OC_DC_QUANT_MIN[2]={4<<2,8<<2};
+static const unsigned OC_AC_QUANT_MIN[2]={2<<2,4<<2};
+
+/*Initializes the dequantization tables from a set of quantizer info.
+  Currently the dequantizer (and elsewhere enquantizer) tables are expected to
+   be initialized as pointing to the storage reserved for them in the
+   oc_theora_state (resp. oc_enc_ctx) structure.
+  If some tables are duplicates of others, the pointers will be adjusted to
+   point to a single copy of the tables, but the storage for them will not be
+   freed.
+  If you're concerned about the memory footprint, the obvious thing to do is
+   to move the storage out of its fixed place in the structures and allocate
+   it on demand.
+  However, a much, much better option is to only store the quantization
+   matrices being used for the current frame, and to recalculate these as the
+   qi values change between frames (this is what VP3 did).*/
+void oc_dequant_tables_init(ogg_uint16_t *_dequant[64][3][2],
+ int _pp_dc_scale[64],const th_quant_info *_qinfo){
+  /*Coding mode: intra or inter.*/
+  int          qti;
+  /*Y', C_b, C_r*/
+  int          pli;
+  for(qti=0;qti<2;qti++)for(pli=0;pli<3;pli++){
+    /*Quality index.*/
+    int qi;
+    /*Range iterator.*/
+    int qri;
+    for(qi=0,qri=0;qri<=_qinfo->qi_ranges[qti][pli].nranges;qri++){
+      th_quant_base base;
+      ogg_uint32_t  q;
+      int           qi_start;
+      int           qi_end;
+      memcpy(base,_qinfo->qi_ranges[qti][pli].base_matrices[qri],
+       sizeof(base));
+      qi_start=qi;
+      if(qri==_qinfo->qi_ranges[qti][pli].nranges)qi_end=qi+1;
+      else qi_end=qi+_qinfo->qi_ranges[qti][pli].sizes[qri];
+      /*Iterate over quality indices in this range.*/
+      for(;;){
+        ogg_uint32_t qfac;
+        int          zzi;
+        int          ci;
+        /*In the original VP3.2 code, the rounding offset and the size of the
+           dead zone around 0 were controlled by a "sharpness" parameter.
+          The size of our dead zone is now controlled by the per-coefficient
+           quality thresholds returned by our HVS module.
+          We round down from a more accurate value when the quality of the
+           reconstruction does not fall below our threshold and it saves bits.
+          Hence, all of that VP3.2 code is gone from here, and the remaining
+           floating point code has been implemented as equivalent integer code
+           with exact precision.*/
+        qfac=(ogg_uint32_t)_qinfo->dc_scale[qi]*base[0];
+        /*For postprocessing, not dequantization.*/
+        if(_pp_dc_scale!=NULL)_pp_dc_scale[qi]=(int)(qfac/160);
+        /*Scale DC the coefficient from the proper table.*/
+        q=(qfac/100)<<2;
+        q=OC_CLAMPI(OC_DC_QUANT_MIN[qti],q,OC_QUANT_MAX);
+        _dequant[qi][pli][qti][0]=(ogg_uint16_t)q;
+        /*Now scale AC coefficients from the proper table.*/
+        for(zzi=1;zzi<64;zzi++){
+          q=((ogg_uint32_t)_qinfo->ac_scale[qi]*base[OC_FZIG_ZAG[zzi]]/100)<<2;
+          q=OC_CLAMPI(OC_AC_QUANT_MIN[qti],q,OC_QUANT_MAX);
+          _dequant[qi][pli][qti][zzi]=(ogg_uint16_t)q;
+        }
+        /*If this is a duplicate of a previous matrix, use that instead.
+          This simple check helps us improve cache coherency later.*/
+        {
+          int dupe;
+          int qtj;
+          int plj;
+          dupe=0;
+          for(qtj=0;qtj<=qti;qtj++){
+            for(plj=0;plj<(qtj<qti?3:pli);plj++){
+              if(!memcmp(_dequant[qi][pli][qti],_dequant[qi][plj][qtj],
+               sizeof(oc_quant_table))){
+                dupe=1;
+                break;
+              }
+            }
+            if(dupe)break;
+          }
+          if(dupe)_dequant[qi][pli][qti]=_dequant[qi][plj][qtj];
+        }
+        if(++qi>=qi_end)break;
+        /*Interpolate the next base matrix.*/
+        for(ci=0;ci<64;ci++){
+          base[ci]=(unsigned char)(
+           (2*((qi_end-qi)*_qinfo->qi_ranges[qti][pli].base_matrices[qri][ci]+
+           (qi-qi_start)*_qinfo->qi_ranges[qti][pli].base_matrices[qri+1][ci])
+           +_qinfo->qi_ranges[qti][pli].sizes[qri])/
+           (2*_qinfo->qi_ranges[qti][pli].sizes[qri]));
+        }
+      }
+    }
+  }
+}

thirdparty/libtheora/quant.h

@@ -0,0 +1,33 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+#if !defined(_quant_H)
+# define _quant_H (1)
+# include "theora/codec.h"
+# include "ocintrin.h"
+
+typedef ogg_uint16_t   oc_quant_table[64];
+
+
+/*Maximum scaled quantizer value.*/
+#define OC_QUANT_MAX          (1024<<2)
+
+
+void oc_dequant_tables_init(ogg_uint16_t *_dequant[64][3][2],
+ int _pp_dc_scale[64],const th_quant_info *_qinfo);
+
+#endif

thirdparty/libtheora/state.h

@@ -0,0 +1,552 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009,2025           *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+#if !defined(_state_H)
+# define _state_H (1)
+# include "internal.h"
+# include "huffman.h"
+# include "quant.h"
+
+
+
+/*A single quadrant of the map from a super block to fragment numbers.*/
+typedef ptrdiff_t       oc_sb_map_quad[4];
+/*A map from a super block to fragment numbers.*/
+typedef oc_sb_map_quad  oc_sb_map[4];
+/*A single plane of the map from a macro block to fragment numbers.*/
+typedef ptrdiff_t       oc_mb_map_plane[4];
+/*A map from a macro block to fragment numbers.*/
+typedef oc_mb_map_plane oc_mb_map[3];
+/*A motion vector.*/
+typedef ogg_int16_t     oc_mv;
+
+typedef struct oc_sb_flags              oc_sb_flags;
+typedef struct oc_border_info           oc_border_info;
+typedef struct oc_fragment              oc_fragment;
+typedef struct oc_fragment_plane        oc_fragment_plane;
+typedef struct oc_base_opt_vtable       oc_base_opt_vtable;
+typedef struct oc_base_opt_data         oc_base_opt_data;
+typedef struct oc_state_dispatch_vtable oc_state_dispatch_vtable;
+typedef struct oc_theora_state          oc_theora_state;
+
+
+
+/*Shared accelerated functions.*/
+# if defined(OC_X86_ASM)
+#  if defined(_MSC_VER)
+#   include "x86_vc/x86int.h"
+#  else
+#   include "x86/x86int.h"
+#  endif
+# endif
+# if defined(OC_ARM_ASM)
+#  include "arm/armint.h"
+# endif
+# if defined(OC_C64X_ASM)
+#  include "c64x/c64xint.h"
+# endif
+
+# if !defined(oc_state_accel_init)
+#  define oc_state_accel_init oc_state_accel_init_c
+# endif
+# if defined(OC_STATE_USE_VTABLE)
+#  if !defined(oc_frag_copy)
+#   define oc_frag_copy(_state,_dst,_src,_ystride) \
+  ((*(_state)->opt_vtable.frag_copy)(_dst,_src,_ystride))
+#  endif
+#  if !defined(oc_frag_copy_list)
+#   define oc_frag_copy_list(_state,_dst_frame,_src_frame,_ystride, \
+ _fragis,_nfragis,_frag_buf_offs) \
+ ((*(_state)->opt_vtable.frag_copy_list)(_dst_frame,_src_frame,_ystride, \
+  _fragis,_nfragis,_frag_buf_offs))
+#  endif
+#  if !defined(oc_frag_recon_intra)
+#   define oc_frag_recon_intra(_state,_dst,_dst_ystride,_residue) \
+  ((*(_state)->opt_vtable.frag_recon_intra)(_dst,_dst_ystride,_residue))
+#  endif
+#  if !defined(oc_frag_recon_inter)
+#   define oc_frag_recon_inter(_state,_dst,_src,_ystride,_residue) \
+  ((*(_state)->opt_vtable.frag_recon_inter)(_dst,_src,_ystride,_residue))
+#  endif
+#  if !defined(oc_frag_recon_inter2)
+#   define oc_frag_recon_inter2(_state,_dst,_src1,_src2,_ystride,_residue) \
+  ((*(_state)->opt_vtable.frag_recon_inter2)(_dst, \
+   _src1,_src2,_ystride,_residue))
+#  endif
+# if !defined(oc_idct8x8)
+#   define oc_idct8x8(_state,_y,_x,_last_zzi) \
+  ((*(_state)->opt_vtable.idct8x8)(_y,_x,_last_zzi))
+#  endif
+#  if !defined(oc_state_frag_recon)
+#   define oc_state_frag_recon(_state,_fragi, \
+ _pli,_dct_coeffs,_last_zzi,_dc_quant) \
+  ((*(_state)->opt_vtable.state_frag_recon)(_state,_fragi, \
+   _pli,_dct_coeffs,_last_zzi,_dc_quant))
+#  endif
+#  if !defined(oc_loop_filter_init)
+#   define oc_loop_filter_init(_state,_bv,_flimit) \
+  ((*(_state)->opt_vtable.loop_filter_init)(_bv,_flimit))
+#  endif
+#  if !defined(oc_state_loop_filter_frag_rows)
+#   define oc_state_loop_filter_frag_rows(_state, \
+ _bv,_refi,_pli,_fragy0,_fragy_end) \
+  ((*(_state)->opt_vtable.state_loop_filter_frag_rows)(_state, \
+   _bv,_refi,_pli,_fragy0,_fragy_end))
+#  endif
+#  if !defined(oc_restore_fpu)
+#   define oc_restore_fpu(_state) \
+  ((*(_state)->opt_vtable.restore_fpu)())
+#  endif
+# else
+#  if !defined(oc_frag_copy)
+#   define oc_frag_copy(_state,_dst,_src,_ystride) \
+  oc_frag_copy_c(_dst,_src,_ystride)
+#  endif
+#  if !defined(oc_frag_copy_list)
+#   define oc_frag_copy_list(_state,_dst_frame,_src_frame,_ystride, \
+ _fragis,_nfragis,_frag_buf_offs) \
+  oc_frag_copy_list_c(_dst_frame,_src_frame,_ystride, \
+  _fragis,_nfragis,_frag_buf_offs)
+#  endif
+#  if !defined(oc_frag_recon_intra)
+#   define oc_frag_recon_intra(_state,_dst,_dst_ystride,_residue) \
+  oc_frag_recon_intra_c(_dst,_dst_ystride,_residue)
+#  endif
+#  if !defined(oc_frag_recon_inter)
+#   define oc_frag_recon_inter(_state,_dst,_src,_ystride,_residue) \
+  oc_frag_recon_inter_c(_dst,_src,_ystride,_residue)
+#  endif
+#  if !defined(oc_frag_recon_inter2)
+#   define oc_frag_recon_inter2(_state,_dst,_src1,_src2,_ystride,_residue) \
+  oc_frag_recon_inter2_c(_dst,_src1,_src2,_ystride,_residue)
+#  endif
+#  if !defined(oc_idct8x8)
+#   define oc_idct8x8(_state,_y,_x,_last_zzi) oc_idct8x8_c(_y,_x,_last_zzi)
+#  endif
+#  if !defined(oc_state_frag_recon)
+#   define oc_state_frag_recon oc_state_frag_recon_c
+#  endif
+#  if !defined(oc_loop_filter_init)
+#   define oc_loop_filter_init(_state,_bv,_flimit) \
+  oc_loop_filter_init_c(_bv,_flimit)
+#  endif
+#  if !defined(oc_state_loop_filter_frag_rows)
+#   define oc_state_loop_filter_frag_rows oc_state_loop_filter_frag_rows_c
+#  endif
+#  if !defined(oc_restore_fpu)
+#   define oc_restore_fpu(_state) do{}while(0)
+#  endif
+# endif
+
+
+
+/*A keyframe.*/
+# define OC_INTRA_FRAME (0)
+/*A predicted frame.*/
+# define OC_INTER_FRAME (1)
+/*A frame of unknown type (frame type decision has not yet been made).*/
+# define OC_UNKWN_FRAME (-1)
+
+/*The amount of padding to add to the reconstructed frame buffers on all
+   sides.
+  This is used to allow unrestricted motion vectors without special casing.
+  This must be a multiple of 2.*/
+# define OC_UMV_PADDING (16)
+
+/*Frame classification indices.*/
+/*The previous golden frame.*/
+# define OC_FRAME_GOLD      (0)
+/*The previous frame.*/
+# define OC_FRAME_PREV      (1)
+/*The current frame.*/
+# define OC_FRAME_SELF      (2)
+/*Used to mark uncoded fragments (for DC prediction).*/
+# define OC_FRAME_NONE      (3)
+
+/*The input or output buffer.*/
+# define OC_FRAME_IO        (3)
+/*Uncompressed prev golden frame.*/
+# define OC_FRAME_GOLD_ORIG (4)
+/*Uncompressed previous frame. */
+# define OC_FRAME_PREV_ORIG (5)
+
+/*Macroblock modes.*/
+/*Macro block is invalid: It is never coded.*/
+# define OC_MODE_INVALID        (-1)
+/*Encoded difference from the same macro block in the previous frame.*/
+# define OC_MODE_INTER_NOMV     (0)
+/*Encoded with no motion compensated prediction.*/
+# define OC_MODE_INTRA          (1)
+/*Encoded difference from the previous frame offset by the given motion
+   vector.*/
+# define OC_MODE_INTER_MV       (2)
+/*Encoded difference from the previous frame offset by the last coded motion
+   vector.*/
+# define OC_MODE_INTER_MV_LAST  (3)
+/*Encoded difference from the previous frame offset by the second to last
+   coded motion vector.*/
+# define OC_MODE_INTER_MV_LAST2 (4)
+/*Encoded difference from the same macro block in the previous golden
+   frame.*/
+# define OC_MODE_GOLDEN_NOMV    (5)
+/*Encoded difference from the previous golden frame offset by the given motion
+   vector.*/
+# define OC_MODE_GOLDEN_MV      (6)
+/*Encoded difference from the previous frame offset by the individual motion
+   vectors given for each block.*/
+# define OC_MODE_INTER_MV_FOUR  (7)
+/*The number of (coded) modes.*/
+# define OC_NMODES              (8)
+
+/*Determines the reference frame used for a given MB mode.*/
+# define OC_FRAME_FOR_MODE(_x) \
+ OC_UNIBBLE_TABLE32(OC_FRAME_PREV,OC_FRAME_SELF,OC_FRAME_PREV,OC_FRAME_PREV, \
+  OC_FRAME_PREV,OC_FRAME_GOLD,OC_FRAME_GOLD,OC_FRAME_PREV,(_x))
+
+/*Constants for the packet state machine common between encoder and decoder.*/
+
+/*Next packet to emit/read: Codec info header.*/
+# define OC_PACKET_INFO_HDR    (-3)
+/*Next packet to emit/read: Comment header.*/
+# define OC_PACKET_COMMENT_HDR (-2)
+/*Next packet to emit/read: Codec setup header.*/
+# define OC_PACKET_SETUP_HDR   (-1)
+/*No more packets to emit/read.*/
+# define OC_PACKET_DONE        (INT_MAX)
+
+
+
+#define OC_MV(_x,_y)         ((oc_mv)((_x)&0xFF|(_y)*256))
+#define OC_MV_X(_mv)         ((signed char)(_mv))
+#define OC_MV_Y(_mv)         ((_mv)>>8)
+#define OC_MV_ADD(_mv1,_mv2) \
+  OC_MV(OC_MV_X(_mv1)+OC_MV_X(_mv2), \
+   OC_MV_Y(_mv1)+OC_MV_Y(_mv2))
+#define OC_MV_SUB(_mv1,_mv2) \
+  OC_MV(OC_MV_X(_mv1)-OC_MV_X(_mv2), \
+   OC_MV_Y(_mv1)-OC_MV_Y(_mv2))
+
+
+
+/*Super blocks are 32x32 segments of pixels in a single color plane indexed
+   in image order.
+  Internally, super blocks are broken up into four quadrants, each of which
+   contains a 2x2 pattern of blocks, each of which is an 8x8 block of pixels.
+  Quadrants, and the blocks within them, are indexed in a special order called
+   a "Hilbert curve" within the super block.
+
+  In order to differentiate between the Hilbert-curve indexing strategy and
+   the regular image order indexing strategy, blocks indexed in image order
+   are called "fragments".
+  Fragments are indexed in image order, left to right, then bottom to top,
+   from Y' plane to Cb plane to Cr plane.
+
+  The co-located fragments in all image planes corresponding to the location
+   of a single quadrant of a luma plane super block form a macro block.
+  Thus there is only a single set of macro blocks for all planes, each of which
+   contains between 6 and 12 fragments, depending on the pixel format.
+  Therefore macro block information is kept in a separate set of arrays from
+   super blocks to avoid unused space in the other planes.
+  The lists are indexed in super block order.
+  That is, the macro block corresponding to the macro block mbi in (luma plane)
+   super block sbi is at index (sbi<<2|mbi).
+  Thus the number of macro blocks in each dimension is always twice the number
+   of super blocks, even when only an odd number fall inside the coded frame.
+  These "extra" macro blocks are just an artifact of our internal data layout,
+   and not part of the coded stream; they are flagged with a negative MB mode.*/
+
+
+
+/*Super block information.*/
+struct oc_sb_flags{
+  unsigned char coded_fully:1;
+  unsigned char coded_partially:1;
+  unsigned char quad_valid:4;
+};
+
+
+
+/*Information about a fragment which intersects the border of the displayable
+   region.
+  This marks which pixels belong to the displayable region.*/
+struct oc_border_info{
+  /*A bit mask marking which pixels are in the displayable region.
+    Pixel (x,y) corresponds to bit (y<<3|x).*/
+  ogg_int64_t mask;
+  /*The number of pixels in the displayable region.
+    This is always positive, and always less than 64.*/
+  int         npixels;
+};
+
+
+
+/*Fragment information.*/
+struct oc_fragment{
+  /*A flag indicating whether or not this fragment is coded.*/
+  unsigned   coded:1;
+  /*A flag indicating that this entire fragment lies outside the displayable
+     region of the frame.
+    Note the contrast with an invalid macro block, which is outside the coded
+     frame, not just the displayable one.
+    There are no fragments outside the coded frame by construction.*/
+  unsigned   invalid:1;
+  /*The index of the quality index used for this fragment's AC coefficients.*/
+  unsigned   qii:4;
+  /*The index of the reference frame this fragment is predicted from.*/
+  unsigned   refi:2;
+  /*The mode of the macroblock this fragment belongs to.*/
+  unsigned   mb_mode:3;
+  /*The index of the associated border information for fragments which lie
+     partially outside the displayable region.
+    For fragments completely inside or outside this region, this is -1.
+    Note that the C standard requires an explicit signed keyword for bitfield
+     types, since some compilers may treat them as unsigned without it.*/
+  signed int borderi:5;
+  /*The prediction-corrected DC component.
+    Note that the C standard requires an explicit signed keyword for bitfield
+     types, since some compilers may treat them as unsigned without it.*/
+  signed int dc:16;
+};
+
+
+
+/*A description of each fragment plane.*/
+struct oc_fragment_plane{
+  /*The number of fragments in the horizontal direction.*/
+  int       nhfrags;
+  /*The number of fragments in the vertical direction.*/
+  int       nvfrags;
+  /*The offset of the first fragment in the plane.*/
+  ptrdiff_t froffset;
+  /*The total number of fragments in the plane.*/
+  ptrdiff_t nfrags;
+  /*The number of super blocks in the horizontal direction.*/
+  unsigned  nhsbs;
+  /*The number of super blocks in the vertical direction.*/
+  unsigned  nvsbs;
+  /*The offset of the first super block in the plane.*/
+  unsigned  sboffset;
+  /*The total number of super blocks in the plane.*/
+  unsigned  nsbs;
+};
+
+
+typedef void (*oc_state_loop_filter_frag_rows_func)(
+ const oc_theora_state *_state,signed char _bv[256],int _refi,int _pli,
+ int _fragy0,int _fragy_end);
+
+/*The shared (encoder and decoder) functions that have accelerated variants.*/
+struct oc_base_opt_vtable{
+  void (*frag_copy)(unsigned char *_dst,
+   const unsigned char *_src,int _ystride);
+  void (*frag_copy_list)(unsigned char *_dst_frame,
+   const unsigned char *_src_frame,int _ystride,
+   const ptrdiff_t *_fragis,ptrdiff_t _nfragis,const ptrdiff_t *_frag_buf_offs);
+  void (*frag_recon_intra)(unsigned char *_dst,int _ystride,
+   const ogg_int16_t _residue[64]);
+  void (*frag_recon_inter)(unsigned char *_dst,
+   const unsigned char *_src,int _ystride,const ogg_int16_t _residue[64]);
+  void (*frag_recon_inter2)(unsigned char *_dst,const unsigned char *_src1,
+   const unsigned char *_src2,int _ystride,const ogg_int16_t _residue[64]);
+  void (*idct8x8)(ogg_int16_t _y[64],ogg_int16_t _x[64],int _last_zzi);
+  void (*state_frag_recon)(const oc_theora_state *_state,ptrdiff_t _fragi,
+   int _pli,ogg_int16_t _dct_coeffs[128],int _last_zzi,ogg_uint16_t _dc_quant);
+  void (*loop_filter_init)(signed char _bv[256],int _flimit);
+  oc_state_loop_filter_frag_rows_func state_loop_filter_frag_rows;
+  void (*restore_fpu)(void);
+};
+
+/*The shared (encoder and decoder) tables that vary according to which variants
+   of the above functions are used.*/
+struct oc_base_opt_data{
+  const unsigned char *dct_fzig_zag;
+};
+
+
+/*State information common to both the encoder and decoder.*/
+struct oc_theora_state{
+  /*The stream information.*/
+  th_info             info;
+# if defined(OC_STATE_USE_VTABLE)
+  /*Table for shared accelerated functions.*/
+  oc_base_opt_vtable  opt_vtable;
+# endif
+  /*Table for shared data used by accelerated functions.*/
+  oc_base_opt_data    opt_data;
+  /*CPU flags to detect the presence of extended instruction sets.*/
+  ogg_uint32_t        cpu_flags;
+  /*The fragment plane descriptions.*/
+  oc_fragment_plane   fplanes[3];
+  /*The list of fragments, indexed in image order.*/
+  oc_fragment        *frags;
+  /*The the offset into the reference frame buffer to the upper-left pixel of
+     each fragment.*/
+  ptrdiff_t          *frag_buf_offs;
+  /*The motion vector for each fragment.*/
+  oc_mv              *frag_mvs;
+  /*The total number of fragments in a single frame.*/
+  ptrdiff_t           nfrags;
+  /*The list of super block maps, indexed in image order.*/
+  oc_sb_map          *sb_maps;
+  /*The list of super block flags, indexed in image order.*/
+  oc_sb_flags        *sb_flags;
+  /*The total number of super blocks in a single frame.*/
+  unsigned            nsbs;
+  /*The fragments from each color plane that belong to each macro block.
+    Fragments are stored in image order (left to right then top to bottom).
+    When chroma components are decimated, the extra fragments have an index of
+     -1.*/
+  oc_mb_map          *mb_maps;
+  /*The list of macro block modes.
+    A negative number indicates the macro block lies entirely outside the
+     coded frame.*/
+  signed char        *mb_modes;
+  /*The number of macro blocks in the X direction.*/
+  unsigned            nhmbs;
+  /*The number of macro blocks in the Y direction.*/
+  unsigned            nvmbs;
+  /*The total number of macro blocks.*/
+  size_t              nmbs;
+  /*The list of coded fragments, in coded order.
+    Uncoded fragments are stored in reverse order from the end of the list.*/
+  ptrdiff_t          *coded_fragis;
+  /*The number of coded fragments in each plane.*/
+  ptrdiff_t           ncoded_fragis[3];
+  /*The total number of coded fragments.*/
+  ptrdiff_t           ntotal_coded_fragis;
+  /*The actual buffers used for the reference frames.*/
+  th_ycbcr_buffer     ref_frame_bufs[6];
+  /*The index of the buffers being used for each OC_FRAME_* reference frame.*/
+  int                 ref_frame_idx[6];
+  /*The storage for the reference frame buffers.
+    This is just ref_frame_bufs[ref_frame_idx[i]][0].data, but is cached here
+     for faster look-up.*/
+  unsigned char      *ref_frame_data[6];
+  /*The handle used to allocate the reference frame buffers.*/
+  unsigned char      *ref_frame_handle;
+  /*The strides for each plane in the reference frames.*/
+  int                 ref_ystride[3];
+  /*The number of unique border patterns.*/
+  int                 nborders;
+  /*The unique border patterns for all border fragments.
+    The borderi field of fragments which straddle the border indexes this
+     list.*/
+  oc_border_info      borders[16];
+  /*The frame number of the last keyframe.*/
+  ogg_int64_t         keyframe_num;
+  /*The frame number of the current frame.*/
+  ogg_int64_t         curframe_num;
+  /*The granpos of the current frame.*/
+  ogg_int64_t         granpos;
+  /*The type of the current frame.*/
+  signed char         frame_type;
+  /*The bias to add to the frame count when computing granule positions.*/
+  unsigned char       granpos_bias;
+  /*The number of quality indices used in the current frame.*/
+  unsigned char       nqis;
+  /*The quality indices of the current frame.*/
+  unsigned char       qis[3];
+  /*The dequantization tables, stored in zig-zag order, and indexed by
+     qi, pli, qti, and zzi.*/
+  ogg_uint16_t       *dequant_tables[64][3][2];
+  OC_ALIGN16(oc_quant_table      dequant_table_data[64][3][2]);
+  /*Loop filter strength parameters.*/
+  unsigned char       loop_filter_limits[64];
+};
+
+
+
+/*The function type used to fill in the chroma plane motion vectors for a
+   macro block when 4 different motion vectors are specified in the luma
+   plane.
+  _cbmvs: The chroma block-level motion vectors to fill in.
+  _lmbmv: The luma macro-block level motion vector to fill in for use in
+           prediction.
+  _lbmvs: The luma block-level motion vectors.*/
+typedef void (*oc_set_chroma_mvs_func)(oc_mv _cbmvs[4],const oc_mv _lbmvs[4]);
+
+
+
+/*A table of functions used to fill in the Cb,Cr plane motion vectors for a
+   macro block when 4 different motion vectors are specified in the luma
+   plane.*/
+extern const oc_set_chroma_mvs_func OC_SET_CHROMA_MVS_TABLE[TH_PF_NFORMATS];
+
+
+
+int oc_state_init(oc_theora_state *_state,const th_info *_info,int _nrefs);
+void oc_state_clear(oc_theora_state *_state);
+void oc_state_accel_init_c(oc_theora_state *_state);
+void oc_state_borders_fill_rows(oc_theora_state *_state,int _refi,int _pli,
+ int _y0,int _yend);
+void oc_state_borders_fill_caps(oc_theora_state *_state,int _refi,int _pli);
+void oc_state_borders_fill(oc_theora_state *_state,int _refi);
+void oc_state_fill_buffer_ptrs(oc_theora_state *_state,int _buf_idx,
+ th_ycbcr_buffer _img);
+int oc_state_mbi_for_pos(oc_theora_state *_state,int _mbx,int _mby);
+int oc_state_get_mv_offsets(const oc_theora_state *_state,int _offsets[2],
+ int _pli,oc_mv _mv);
+
+void oc_loop_filter_init_c(signed char _bv[256],int _flimit);
+void oc_state_loop_filter(oc_theora_state *_state,int _frame);
+# if defined(OC_DUMP_IMAGES)
+int oc_state_dump_frame(const oc_theora_state *_state,int _frame,
+ const char *_suf);
+# endif
+
+/*Default pure-C implementations of shared accelerated functions.*/
+void oc_frag_copy_c(unsigned char *_dst,
+ const unsigned char *_src,int _src_ystride);
+void oc_frag_copy_list_c(unsigned char *_dst_frame,
+ const unsigned char *_src_frame,int _ystride,
+ const ptrdiff_t *_fragis,ptrdiff_t _nfragis,const ptrdiff_t *_frag_buf_offs);
+void oc_frag_recon_intra_c(unsigned char *_dst,int _dst_ystride,
+ const ogg_int16_t _residue[64]);
+void oc_frag_recon_inter_c(unsigned char *_dst,
+ const unsigned char *_src,int _ystride,const ogg_int16_t _residue[64]);
+void oc_frag_recon_inter2_c(unsigned char *_dst,const unsigned char *_src1,
+ const unsigned char *_src2,int _ystride,const ogg_int16_t _residue[64]);
+void oc_idct8x8_c(ogg_int16_t _y[64],ogg_int16_t _x[64],int _last_zzi);
+void oc_state_frag_recon_c(const oc_theora_state *_state,ptrdiff_t _fragi,
+ int _pli,ogg_int16_t _dct_coeffs[128],int _last_zzi,ogg_uint16_t _dc_quant);
+void oc_state_loop_filter_frag_rows_c(const oc_theora_state *_state,
+ signed char _bvarray[256],int _refi,int _pli,int _fragy0,int _fragy_end);
+void oc_restore_fpu_c(void);
+
+/*We need a way to call a few encoder functions without introducing a link-time
+   dependency into the decoder, while still allowing the old alpha API which
+   does not distinguish between encoder and decoder objects to be used.
+  We do this by placing a function table at the start of the encoder object
+   which can dispatch into the encoder library.
+  We do a similar thing for the decoder in case we ever decide to split off a
+   common base library.*/
+typedef void (*oc_state_clear_func)(theora_state *_th);
+typedef int (*oc_state_control_func)(theora_state *th,int _req,
+ void *_buf,size_t _buf_sz);
+typedef ogg_int64_t (*oc_state_granule_frame_func)(theora_state *_th,
+ ogg_int64_t _granulepos);
+typedef double (*oc_state_granule_time_func)(theora_state *_th,
+ ogg_int64_t _granulepos);
+
+
+struct oc_state_dispatch_vtable{
+  oc_state_clear_func         clear;
+  oc_state_control_func       control;
+  oc_state_granule_frame_func granule_frame;
+  oc_state_granule_time_func  granule_time;
+};
+
+#endif

thirdparty/libtheora/theora/codec.h

@@ -0,0 +1,605 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation https://www.xiph.org/                 *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+/**\mainpage
+ *
+ * \section intro Introduction
+ *
+ * This is the documentation for the <tt>libtheora</tt> C API.
+ *
+ * The \c libtheora package is the current reference
+ * implementation for <a href="https://www.theora.org/">Theora</a>, a free,
+ * patent-unencumbered video codec.
+ * Theora is derived from On2's VP3 codec with additional features and
+ *  integration with Ogg multimedia formats by
+ *  <a href="https://www.xiph.org/">the Xiph.Org Foundation</a>.
+ * Complete documentation of the format itself is available in
+ * <a href="https://www.theora.org/doc/Theora.pdf">the Theora
+ *  specification</a>.
+ *
+ * \section Organization
+ *
+ * The functions documented here are divided between two
+ * separate libraries:
+ * - \c libtheoraenc contains the encoder interface,
+ *   described in \ref encfuncs.
+ * - \c libtheoradec contains the decoder interface,
+ *   described in \ref decfuncs, \n
+ *   and additional \ref basefuncs.
+ *
+ * New code should link to \c libtheoradec. If using encoder
+ * features, it must also link to \c libtheoraenc.
+ *
+ * During initial development, prior to the 1.0 release,
+ * \c libtheora exported a different \ref oldfuncs which
+ * combined both encode and decode functions.
+ * In general, legacy API symbols can be identified
+ * by their \c theora_ or \c OC_ namespace prefixes.
+ * The current API uses \c th_ or \c TH_ instead.
+ *
+ * While deprecated, \c libtheoraenc and \c libtheoradec
+ * together export the legacy api as well at the one documented above.
+ * Likewise, the legacy \c libtheora included with this package
+ * exports the new 1.x API. Older code and build scripts can therefore
+ * but updated independently to the current scheme.
+ */
+
+/**\file
+ * The shared <tt>libtheoradec</tt> and <tt>libtheoraenc</tt> C API.
+ * You don't need to include this directly.*/
+
+#if !defined(OGG_THEORA_CODEC_HEADER)
+# define OGG_THEORA_CODEC_HEADER (1)
+# include <ogg/ogg.h>
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+
+
+/**\name Return codes*/
+/*@{*/
+/**An invalid pointer was provided.*/
+#define TH_EFAULT     (-1)
+/**An invalid argument was provided.*/
+#define TH_EINVAL     (-10)
+/**The contents of the header were incomplete, invalid, or unexpected.*/
+#define TH_EBADHEADER (-20)
+/**The header does not belong to a Theora stream.*/
+#define TH_ENOTFORMAT (-21)
+/**The bitstream version is too high.*/
+#define TH_EVERSION   (-22)
+/**The specified function is not implemented.*/
+#define TH_EIMPL      (-23)
+/**There were errors in the video data packet.*/
+#define TH_EBADPACKET (-24)
+/**The decoded packet represented a dropped frame.
+   The player can continue to display the current frame, as the contents of the
+    decoded frame buffer have not changed.*/
+#define TH_DUPFRAME   (1)
+/*@}*/
+
+/**The currently defined color space tags.
+ * See <a href="https://www.theora.org/doc/Theora.pdf">the Theora
+ *  specification</a>, Chapter 4, for exact details on the meaning
+ *  of each of these color spaces.*/
+typedef enum{
+  /**The color space was not specified at the encoder.
+      It may be conveyed by an external means.*/
+  TH_CS_UNSPECIFIED,
+  /**A color space designed for NTSC content.*/
+  TH_CS_ITU_REC_470M,
+  /**A color space designed for PAL/SECAM content.*/
+  TH_CS_ITU_REC_470BG,
+  /**The total number of currently defined color spaces.*/
+  TH_CS_NSPACES
+}th_colorspace;
+
+/**The currently defined pixel format tags.
+ * See <a href="https://www.theora.org/doc/Theora.pdf">the Theora
+ *  specification</a>, Section 4.4, for details on the precise sample
+ *  locations.*/
+typedef enum{
+  /**Chroma decimation by 2 in both the X and Y directions (4:2:0).
+     The Cb and Cr chroma planes are half the width and half the
+      height of the luma plane.*/
+  TH_PF_420,
+  /**Currently reserved.*/
+  TH_PF_RSVD,
+  /**Chroma decimation by 2 in the X direction (4:2:2).
+     The Cb and Cr chroma planes are half the width of the luma plane, but full
+      height.*/
+  TH_PF_422,
+  /**No chroma decimation (4:4:4).
+     The Cb and Cr chroma planes are full width and full height.*/
+  TH_PF_444,
+  /**The total number of currently defined pixel formats.*/
+  TH_PF_NFORMATS
+}th_pixel_fmt;
+
+
+
+/**A buffer for a single color plane in an uncompressed image.
+ * This contains the image data in a left-to-right, top-down format.
+ * Each row of pixels is stored contiguously in memory, but successive
+ *  rows need not be.
+ * Use \a stride to compute the offset of the next row.
+ * The encoder accepts both positive \a stride values (top-down in memory)
+ *  and negative (bottom-up in memory).
+ * The decoder currently always generates images with positive strides.*/
+typedef struct{
+  /**The width of this plane.*/
+  int            width;
+  /**The height of this plane.*/
+  int            height;
+  /**The offset in bytes between successive rows.*/
+  int            stride;
+  /**A pointer to the beginning of the first row.*/
+  unsigned char *data;
+}th_img_plane;
+
+/**A complete image buffer for an uncompressed frame.
+ * The chroma planes may be decimated by a factor of two in either
+ *  direction, as indicated by th_info#pixel_fmt.
+ * The width and height of the Y' plane must be multiples of 16.
+ * They may need to be cropped for display, using the rectangle
+ *  specified by th_info#pic_x, th_info#pic_y, th_info#pic_width,
+ *  and th_info#pic_height.
+ * All samples are 8 bits.
+ * \note The term YUV often used to describe a colorspace is ambiguous.
+ * The exact parameters of the RGB to YUV conversion process aside, in
+ *  many contexts the U and V channels actually have opposite meanings.
+ * To avoid this confusion, we are explicit: the name of the color
+ *  channels are Y'CbCr, and they appear in that order, always.
+ * The prime symbol denotes that the Y channel is non-linear.
+ * Cb and Cr stand for "Chroma blue" and "Chroma red", respectively.*/
+typedef th_img_plane th_ycbcr_buffer[3];
+
+/**Theora bitstream information.
+ * This contains the basic playback parameters for a stream, and corresponds to
+ *  the initial 'info' header packet.
+ * To initialize an encoder, the application fills in this structure and
+ *  passes it to th_encode_alloc().
+ * A default encoding mode is chosen based on the values of the #quality and
+ *  #target_bitrate fields.
+ * On decode, it is filled in by th_decode_headerin(), and then passed to
+ *  th_decode_alloc().
+ *
+ * Encoded Theora frames must be a multiple of 16 in size;
+ *  this is what the #frame_width and #frame_height members represent.
+ * To handle arbitrary picture sizes, a crop rectangle is specified in the
+ *  #pic_x, #pic_y, #pic_width and #pic_height members.
+ *
+ * All frame buffers contain pointers to the full, padded frame.
+ * However, the current encoder <em>will not</em> reference pixels outside of
+ *  the cropped picture region, and the application does not need to fill them
+ *  in.
+ * The decoder <em>will</em> allocate storage for a full frame, but the
+ *  application <em>should not</em> rely on the padding containing sensible
+ *  data.
+ *
+ * It is also generally recommended that the offsets and sizes should still be
+ *  multiples of 2 to avoid chroma sampling shifts when chroma is sub-sampled.
+ * See <a href="https://www.theora.org/doc/Theora.pdf">the Theora
+ *  specification</a>, Section 4.4, for more details.
+ *
+ * Frame rate, in frames per second, is stored as a rational fraction, as is
+ *  the pixel aspect ratio.
+ * Note that this refers to the aspect ratio of the individual pixels, not of
+ *  the overall frame itself.
+ * The frame aspect ratio can be computed from pixel aspect ratio using the
+ *  image dimensions.*/
+typedef struct{
+  /**\name Theora version
+   * Bitstream version information.*/
+  /*@{*/
+  unsigned char version_major;
+  unsigned char version_minor;
+  unsigned char version_subminor;
+  /*@}*/
+  /**The encoded frame width.
+   * This must be a multiple of 16, and less than 1048576.*/
+  ogg_uint32_t  frame_width;
+  /**The encoded frame height.
+   * This must be a multiple of 16, and less than 1048576.*/
+  ogg_uint32_t  frame_height;
+  /**The displayed picture width.
+   * This must be no larger than width.*/
+  ogg_uint32_t  pic_width;
+  /**The displayed picture height.
+   * This must be no larger than height.*/
+  ogg_uint32_t  pic_height;
+  /**The X offset of the displayed picture.
+   * This must be no larger than #frame_width-#pic_width or 255, whichever is
+   *  smaller.*/
+  ogg_uint32_t  pic_x;
+  /**The Y offset of the displayed picture.
+   * This must be no larger than #frame_height-#pic_height, and
+   *  #frame_height-#pic_height-#pic_y must be no larger than 255.
+   * This slightly funny restriction is due to the fact that the offset is
+   *  specified from the top of the image for consistency with the standard
+   *  graphics left-handed coordinate system used throughout this API, while
+   *  it is stored in the encoded stream as an offset from the bottom.*/
+  ogg_uint32_t  pic_y;
+  /**\name Frame rate
+   * The frame rate, as a fraction.
+   * If either is 0, the frame rate is undefined.*/
+  /*@{*/
+  ogg_uint32_t  fps_numerator;
+  ogg_uint32_t  fps_denominator;
+  /*@}*/
+  /**\name Aspect ratio
+   * The aspect ratio of the pixels.
+   * If either value is zero, the aspect ratio is undefined.
+   * If not specified by any external means, 1:1 should be assumed.
+   * The aspect ratio of the full picture can be computed as
+   * \code
+   *  aspect_numerator*pic_width/(aspect_denominator*pic_height).
+   * \endcode */
+  /*@{*/
+  ogg_uint32_t  aspect_numerator;
+  ogg_uint32_t  aspect_denominator;
+  /*@}*/
+  /**The color space.*/
+  th_colorspace colorspace;
+  /**The pixel format.*/
+  th_pixel_fmt  pixel_fmt;
+  /**The target bit-rate in bits per second.
+     If initializing an encoder with this struct, set this field to a non-zero
+      value to activate CBR encoding by default.*/
+  int           target_bitrate;
+  /**The target quality level.
+     Valid values range from 0 to 63, inclusive, with higher values giving
+      higher quality.
+     If initializing an encoder with this struct, and #target_bitrate is set
+      to zero, VBR encoding at this quality will be activated by default.*/
+  /*Currently this is set so that a qi of 0 corresponds to distortions of 24
+     times the JND, and each increase by 16 halves that value.
+    This gives us fine discrimination at low qualities, yet effective rate
+     control at high qualities.
+    The qi value 63 is special, however.
+    For this, the highest quality, we use one half of a JND for our threshold.
+    Due to the lower bounds placed on allowable quantizers in Theora, we will
+     not actually be able to achieve quality this good, but this should
+     provide as close to visually lossless quality as Theora is capable of.
+    We could lift the quantizer restrictions without breaking VP3.1
+     compatibility, but this would result in quantized coefficients that are
+     too large for the current bitstream to be able to store.
+    We'd have to redesign the token syntax to store these large coefficients,
+     which would make transcoding complex.*/
+  int           quality;
+  /**The amount to shift to extract the last keyframe number from the granule
+   *  position.
+   * This can be at most 31.
+   * th_info_init() will set this to a default value (currently <tt>6</tt>,
+   *  which is good for streaming applications), but you can set it to 0 to
+   *  make every frame a keyframe.
+   * The maximum distance between key frames is
+   *  <tt>1<<#keyframe_granule_shift</tt>.
+   * The keyframe frequency can be more finely controlled with
+   *  #TH_ENCCTL_SET_KEYFRAME_FREQUENCY_FORCE, which can also be adjusted
+   *  during encoding (for example, to force the next frame to be a keyframe),
+   *  but it cannot be set larger than the amount permitted by this field after
+   *  the headers have been output.*/
+  int           keyframe_granule_shift;
+}th_info;
+
+/**The comment information.
+ *
+ * This structure holds the in-stream metadata corresponding to
+ *  the 'comment' header packet.
+ * The comment header is meant to be used much like someone jotting a quick
+ *  note on the label of a video.
+ * It should be a short, to the point text note that can be more than a couple
+ *  words, but not more than a short paragraph.
+ *
+ * The metadata is stored as a series of (tag, value) pairs, in
+ *  length-encoded string vectors.
+ * The first occurrence of the '=' character delimits the tag and value.
+ * A particular tag may occur more than once, and order is significant.
+ * The character set encoding for the strings is always UTF-8, but the tag
+ *  names are limited to ASCII, and treated as case-insensitive.
+ * See <a href="https://www.theora.org/doc/Theora.pdf">the Theora
+ *  specification</a>, Section 6.3.3 for details.
+ *
+ * In filling in this structure, th_decode_headerin() will null-terminate
+ *  the user_comment strings for safety.
+ * However, the bitstream format itself treats them as 8-bit clean vectors,
+ *  possibly containing null characters, so the length array should be
+ *  treated as their authoritative length.
+ */
+typedef struct th_comment{
+  /**The array of comment string vectors.*/
+  char **user_comments;
+  /**An array of the corresponding length of each vector, in bytes.*/
+  int   *comment_lengths;
+  /**The total number of comment strings.*/
+  int    comments;
+  /**The null-terminated vendor string.
+     This identifies the software used to encode the stream.*/
+  char  *vendor;
+}th_comment;
+
+
+
+/**A single base matrix.*/
+typedef unsigned char th_quant_base[64];
+
+/**A set of \a qi ranges.*/
+typedef struct{
+  /**The number of ranges in the set.*/
+  int                  nranges;
+  /**The size of each of the #nranges ranges.
+     These must sum to 63.*/
+  const int           *sizes;
+  /**#nranges <tt>+1</tt> base matrices.
+     Matrices \a i and <tt>i+1</tt> form the endpoints of range \a i.*/
+  const th_quant_base *base_matrices;
+}th_quant_ranges;
+
+/**A complete set of quantization parameters.
+   The quantizer for each coefficient is calculated as:
+   \code
+    Q=MAX(MIN(qmin[qti][ci!=0],scale[ci!=0][qi]*base[qti][pli][qi][ci]/100),
+     1024).
+   \endcode
+
+   \a qti is the quantization type index: 0 for intra, 1 for inter.
+   <tt>ci!=0</tt> is 0 for the DC coefficient and 1 for AC coefficients.
+   \a qi is the quality index, ranging between 0 (low quality) and 63 (high
+    quality).
+   \a pli is the color plane index: 0 for Y', 1 for Cb, 2 for Cr.
+   \a ci is the DCT coefficient index.
+   Coefficient indices correspond to the normal 2D DCT block
+    ordering--row-major with low frequencies first--\em not zig-zag order.
+
+   Minimum quantizers are constant, and are given by:
+   \code
+   qmin[2][2]={{4,2},{8,4}}.
+   \endcode
+
+   Parameters that can be stored in the bitstream are as follows:
+    - The two scale matrices ac_scale and dc_scale.
+      \code
+      scale[2][64]={dc_scale,ac_scale}.
+      \endcode
+    - The base matrices for each \a qi, \a qti and \a pli (up to 384 in all).
+      In order to avoid storing a full 384 base matrices, only a sparse set of
+       matrices are stored, and the rest are linearly interpolated.
+      This is done as follows.
+      For each \a qti and \a pli, a series of \a n \a qi ranges is defined.
+      The size of each \a qi range can vary arbitrarily, but they must sum to
+       63.
+      Then, <tt>n+1</tt> matrices are specified, one for each endpoint of the
+       ranges.
+      For interpolation purposes, each range's endpoints are the first \a qi
+       value it contains and one past the last \a qi value it contains.
+      Fractional values are rounded to the nearest integer, with ties rounded
+       away from zero.
+
+      Base matrices are stored by reference, so if the same matrices are used
+       multiple times, they will only appear once in the bitstream.
+      The bitstream is also capable of omitting an entire set of ranges and
+       its associated matrices if they are the same as either the previous
+       set (indexed in row-major order) or if the inter set is the same as the
+       intra set.
+
+    - Loop filter limit values.
+      The same limits are used for the loop filter in all color planes, despite
+       potentially differing levels of quantization in each.
+
+   For the current encoder, <tt>scale[ci!=0][qi]</tt> must be no greater
+    than <tt>scale[ci!=0][qi-1]</tt> and <tt>base[qti][pli][qi][ci]</tt> must
+    be no greater than <tt>base[qti][pli][qi-1][ci]</tt>.
+   These two conditions ensure that the actual quantizer for a given \a qti,
+    \a pli, and \a ci does not increase as \a qi increases.
+   This is not required by the decoder.*/
+typedef struct{
+  /**The DC scaling factors.*/
+  ogg_uint16_t    dc_scale[64];
+  /**The AC scaling factors.*/
+  ogg_uint16_t    ac_scale[64];
+  /**The loop filter limit values.*/
+  unsigned char   loop_filter_limits[64];
+  /**The \a qi ranges for each \a ci and \a pli.*/
+  th_quant_ranges qi_ranges[2][3];
+}th_quant_info;
+
+
+
+/**The number of Huffman tables used by Theora.*/
+#define TH_NHUFFMAN_TABLES (80)
+/**The number of DCT token values in each table.*/
+#define TH_NDCT_TOKENS     (32)
+
+/**A Huffman code for a Theora DCT token.
+ * Each set of Huffman codes in a given table must form a complete, prefix-free
+ *  code.
+ * There is no requirement that all the tokens in a table have a valid code,
+ *  but the current encoder is not optimized to take advantage of this.
+ * If each of the five grouops of 16 tables does not contain at least one table
+ *  with a code for every token, then the encoder may fail to encode certain
+ *  frames.
+ * The complete table in the first group of 16 does not have to be in the same
+ *  place as the complete table in the other groups, but the complete tables in
+ *  the remaining four groups must all be in the same place.*/
+typedef struct{
+  /**The bit pattern for the code, with the LSbit of the pattern aligned in
+   *   the LSbit of the word.*/
+  ogg_uint32_t pattern;
+  /**The number of bits in the code.
+   * This must be between 0 and 32, inclusive.*/
+  int          nbits;
+}th_huff_code;
+
+
+
+/**\defgroup basefuncs Functions Shared by Encode and Decode*/
+/*@{*/
+/**\name Basic shared functions
+ * These functions return information about the library itself,
+ * or provide high-level information about codec state
+ * and packet type.
+ *
+ * You must link to \c libtheoradec if you use any of the
+ * functions in this section.*/
+/*@{*/
+/**Retrieves a human-readable string to identify the library vendor and
+ *  version.
+ * \return the version string.*/
+extern const char *th_version_string(void);
+/**Retrieves the library version number.
+ * This is the highest bitstream version that the encoder library will produce,
+ *  or that the decoder library can decode.
+ * This number is composed of a 16-bit major version, 8-bit minor version
+ * and 8 bit sub-version, composed as follows:
+ * \code
+ * (VERSION_MAJOR<<16)+(VERSION_MINOR<<8)+(VERSION_SUBMINOR)
+ * \endcode
+ * \return the version number.*/
+extern ogg_uint32_t th_version_number(void);
+/**Converts a granule position to an absolute frame index, starting at
+ *  <tt>0</tt>.
+ * The granule position is interpreted in the context of a given
+ *  #th_enc_ctx or #th_dec_ctx handle (either will suffice).
+ * \param _encdec  A previously allocated #th_enc_ctx or #th_dec_ctx
+ *                  handle.
+ * \param _granpos The granule position to convert.
+ * \returns The absolute frame index corresponding to \a _granpos.
+ * \retval -1 The given granule position was invalid (i.e. negative).*/
+extern ogg_int64_t th_granule_frame(void *_encdec,ogg_int64_t _granpos);
+/**Converts a granule position to an absolute time in seconds.
+ * The granule position is interpreted in the context of a given
+ *  #th_enc_ctx or #th_dec_ctx handle (either will suffice).
+ * \param _encdec  A previously allocated #th_enc_ctx or #th_dec_ctx
+ *                  handle.
+ * \param _granpos The granule position to convert.
+ * \return The absolute time in seconds corresponding to \a _granpos.
+ *         This is the "end time" for the frame, or the latest time it should
+ *          be displayed.
+ *         It is not the presentation time.
+ * \retval -1 The given granule position was invalid (i.e. negative).*/
+extern double th_granule_time(void *_encdec,ogg_int64_t _granpos);
+/**Determines whether a Theora packet is a header or not.
+ * This function does no verification beyond checking the packet type bit, so
+ *  it should not be used for bitstream identification; use
+ *  th_decode_headerin() for that.
+ * As per the Theora specification, an empty (0-byte) packet is treated as a
+ *  data packet (a delta frame with no coded blocks).
+ * \param _op An <tt>ogg_packet</tt> containing encoded Theora data.
+ * \retval 1 The packet is a header packet
+ * \retval 0 The packet is a video data packet.*/
+extern int th_packet_isheader(ogg_packet *_op);
+/**Determines whether a theora packet is a key frame or not.
+ * This function does no verification beyond checking the packet type and
+ *  key frame bits, so it should not be used for bitstream identification; use
+ *  th_decode_headerin() for that.
+ * As per the Theora specification, an empty (0-byte) packet is treated as a
+ *  delta frame (with no coded blocks).
+ * \param _op An <tt>ogg_packet</tt> containing encoded Theora data.
+ * \retval 1  The packet contains a key frame.
+ * \retval 0  The packet contains a delta frame.
+ * \retval -1 The packet is not a video data packet.*/
+extern int th_packet_iskeyframe(ogg_packet *_op);
+/*@}*/
+
+
+/**\name Functions for manipulating header data
+ * These functions manipulate the #th_info and #th_comment structures
+ * which describe video parameters and key-value metadata, respectively.
+ *
+ * You must link to \c libtheoradec if you use any of the
+ * functions in this section.*/
+/*@{*/
+/**Initializes a th_info structure.
+ * This should be called on a freshly allocated #th_info structure before
+ *  attempting to use it.
+ * \param _info The #th_info struct to initialize.*/
+extern void th_info_init(th_info *_info);
+/**Clears a #th_info structure.
+ * This should be called on a #th_info structure after it is no longer
+ *  needed.
+ * \param _info The #th_info struct to clear.*/
+extern void th_info_clear(th_info *_info);
+
+/**Initialize a #th_comment structure.
+ * This should be called on a freshly allocated #th_comment structure
+ *  before attempting to use it.
+ * \param _tc The #th_comment struct to initialize.*/
+extern void th_comment_init(th_comment *_tc);
+/**Add a comment to an initialized #th_comment structure.
+ * \note Neither th_comment_add() nor th_comment_add_tag() support
+ *  comments containing null values, although the bitstream format does
+ *  support them.
+ * To add such comments you will need to manipulate the #th_comment
+ *  structure directly.
+ * \param _tc      The #th_comment struct to add the comment to.
+ * \param _comment Must be a null-terminated UTF-8 string containing the
+ *                  comment in "TAG=the value" form.*/
+extern void th_comment_add(th_comment *_tc,const char *_comment);
+/**Add a comment to an initialized #th_comment structure.
+ * \note Neither th_comment_add() nor th_comment_add_tag() support
+ *  comments containing null values, although the bitstream format does
+ *  support them.
+ * To add such comments you will need to manipulate the #th_comment
+ *  structure directly.
+ * \param _tc  The #th_comment struct to add the comment to.
+ * \param _tag A null-terminated string containing the tag associated with
+ *              the comment.
+ * \param _val The corresponding value as a null-terminated string.*/
+extern void th_comment_add_tag(th_comment *_tc,const char *_tag,
+ const char *_val);
+/**Look up a comment value by its tag.
+ * \param _tc    An initialized #th_comment structure.
+ * \param _tag   The tag to look up.
+ * \param _count The instance of the tag.
+ *               The same tag can appear multiple times, each with a distinct
+ *                value, so an index is required to retrieve them all.
+ *               The order in which these values appear is significant and
+ *                should be preserved.
+ *               Use th_comment_query_count() to get the legal range for
+ *                the \a _count parameter.
+ * \return A pointer to the queried tag's value.
+ *         This points directly to data in the #th_comment structure.
+ *         It should not be modified or freed by the application, and
+ *          modifications to the structure may invalidate the pointer.
+ * \retval NULL If no matching tag is found.*/
+extern char *th_comment_query(th_comment *_tc,const char *_tag,int _count);
+/**Look up the number of instances of a tag.
+ * Call this first when querying for a specific tag and then iterate over the
+ *  number of instances with separate calls to th_comment_query() to
+ *  retrieve all the values for that tag in order.
+ * \param _tc    An initialized #th_comment structure.
+ * \param _tag   The tag to look up.
+ * \return The number of instances of this particular tag.*/
+extern int th_comment_query_count(th_comment *_tc,const char *_tag);
+/**Clears a #th_comment structure.
+ * This should be called on a #th_comment structure after it is no longer
+ *  needed.
+ * It will free all memory used by the structure members.
+ * \param _tc The #th_comment struct to clear.*/
+extern void th_comment_clear(th_comment *_tc);
+/*@}*/
+/*@}*/
+
+
+
+#if defined(__cplusplus)
+}
+#endif
+
+#endif /* OGG_THEORA_CODEC_HEADER */

thirdparty/libtheora/theora/theora.h

@@ -0,0 +1,785 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation https://www.xiph.org/                 *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+#ifndef OGG_THEORA_HEADER
+#define OGG_THEORA_HEADER
+
+#ifdef __cplusplus
+extern "C"
+{
+#endif /* __cplusplus */
+
+#include <stddef.h>	/* for size_t */
+
+#include <ogg/ogg.h>
+
+/** \file
+ * The libtheora pre-1.0 legacy C API.
+ *
+ * \ingroup oldfuncs
+ *
+ * \section intro Introduction
+ *
+ * This is the documentation for the libtheora legacy C API, declared in
+ * the theora.h header, which describes the old interface used before
+ * the 1.0 release. This API was widely deployed for several years and
+ * remains supported, but for new code we recommend the cleaner API
+ * declared in theoradec.h and theoraenc.h.
+ *
+ * libtheora is the reference implementation for
+ * <a href="https://www.theora.org/">Theora</a>, a free video codec.
+ * Theora is derived from On2's VP3 codec with improved integration with
+ * Ogg multimedia formats by <a href="https://www.xiph.org/">Xiph.Org</a>.
+ *
+ * \section overview Overview
+ *
+ * This library will both decode and encode theora packets to/from raw YUV
+ * frames.  In either case, the packets will most likely either come from or
+ * need to be embedded in an Ogg stream.  Use
+ * <a href="https://www.xiph.org/ogg/">libogg</a> or
+ * <a href="http://www.annodex.net/software/liboggz/index.html">liboggz</a>
+ * to extract/package these packets.
+ *
+ * \section decoding Decoding Process
+ *
+ * Decoding can be separated into the following steps:
+ * -# initialise theora_info and theora_comment structures using
+ *    theora_info_init() and theora_comment_init():
+ \verbatim
+ theora_info     info;
+ theora_comment  comment;
+
+ theora_info_init(&info);
+ theora_comment_init(&comment);
+ \endverbatim
+ * -# retrieve header packets from Ogg stream (there should be 3) and decode
+ *    into theora_info and theora_comment structures using
+ *    theora_decode_header().  See \ref identification for more information on
+ *    identifying which packets are theora packets.
+ \verbatim
+ int i;
+ for (i = 0; i < 3; i++)
+ {
+   (get a theora packet "op" from the Ogg stream)
+   theora_decode_header(&info, &comment, op);
+ }
+ \endverbatim
+ * -# initialise the decoder based on the information retrieved into the
+ *    theora_info struct by theora_decode_header().  You will need a
+ *    theora_state struct.
+ \verbatim
+ theora_state state;
+
+ theora_decode_init(&state, &info);
+ \endverbatim
+ * -# pass in packets and retrieve decoded frames!  See the yuv_buffer
+ *    documentation for information on how to retrieve raw YUV data.
+ \verbatim
+ yuf_buffer buffer;
+ while (last packet was not e_o_s) {
+   (get a theora packet "op" from the Ogg stream)
+   theora_decode_packetin(&state, op);
+   theora_decode_YUVout(&state, &buffer);
+ }
+ \endverbatim
+ *
+ *
+ * \subsection identification Identifying Theora Packets
+ *
+ * All streams inside an Ogg file have a unique serial_no attached to the
+ * stream.  Typically, you will want to
+ *  - retrieve the serial_no for each b_o_s (beginning of stream) page
+ *    encountered within the Ogg file;
+ *  - test the first (only) packet on that page to determine if it is a theora
+ *    packet;
+ *  - once you have found a theora b_o_s page then use the retrieved serial_no
+ *    to identify future packets belonging to the same theora stream.
+ *
+ * Note that you \e cannot use theora_packet_isheader() to determine if a
+ * packet is a theora packet or not, as this function does not perform any
+ * checking beyond whether a header bit is present.  Instead, use the
+ * theora_decode_header() function and check the return value; or examine the
+ * header bytes at the beginning of the Ogg page.
+ */
+
+
+/** \defgroup oldfuncs Legacy pre-1.0 C API */
+/*  @{ */
+
+/**
+ * A YUV buffer for passing uncompressed frames to and from the codec.
+ * This holds a Y'CbCr frame in planar format. The CbCr planes can be
+ * subsampled and have their own separate dimensions and row stride
+ * offsets. Note that the strides may be negative in some
+ * configurations. For theora the width and height of the largest plane
+ * must be a multiple of 16. The actual meaningful picture size and
+ * offset are stored in the theora_info structure; frames returned by
+ * the decoder may need to be cropped for display.
+ *
+ * All samples are 8 bits. Within each plane samples are ordered by
+ * row from the top of the frame to the bottom. Within each row samples
+ * are ordered from left to right.
+ *
+ * During decode, the yuv_buffer struct is allocated by the user, but all
+ * fields (including luma and chroma pointers) are filled by the library.
+ * These pointers address library-internal memory and their contents should
+ * not be modified.
+ *
+ * Conversely, during encode the user allocates the struct and fills out all
+ * fields.  The user also manages the data addressed by the luma and chroma
+ * pointers.  See the encoder_example.c and dump_video.c example files in
+ * theora/examples/ for more information.
+ */
+typedef struct {
+    int   y_width;      /**< Width of the Y' luminance plane */
+    int   y_height;     /**< Height of the luminance plane */
+    int   y_stride;     /**< Offset in bytes between successive rows */
+
+    int   uv_width;     /**< Width of the Cb and Cr chroma planes */
+    int   uv_height;    /**< Height of the chroma planes */
+    int   uv_stride;    /**< Offset between successive chroma rows */
+    unsigned char *y;   /**< Pointer to start of luminance data */
+    unsigned char *u;   /**< Pointer to start of Cb data */
+    unsigned char *v;   /**< Pointer to start of Cr data */
+
+} yuv_buffer;
+
+/**
+ * A Colorspace.
+ */
+typedef enum {
+  OC_CS_UNSPECIFIED,    /**< The colorspace is unknown or unspecified */
+  OC_CS_ITU_REC_470M,   /**< This is the best option for 'NTSC' content */
+  OC_CS_ITU_REC_470BG,  /**< This is the best option for 'PAL' content */
+  OC_CS_NSPACES         /**< This marks the end of the defined colorspaces */
+} theora_colorspace;
+
+/**
+ * A Chroma subsampling
+ *
+ * These enumerate the available chroma subsampling options supported
+ * by the theora format. See Section 4.4 of the specification for
+ * exact definitions.
+ */
+typedef enum {
+  OC_PF_420,    /**< Chroma subsampling by 2 in each direction (4:2:0) */
+  OC_PF_RSVD,   /**< Reserved value */
+  OC_PF_422,    /**< Horizonatal chroma subsampling by 2 (4:2:2) */
+  OC_PF_444     /**< No chroma subsampling at all (4:4:4) */
+} theora_pixelformat;
+
+/**
+ * Theora bitstream info.
+ * Contains the basic playback parameters for a stream,
+ * corresponding to the initial 'info' header packet.
+ *
+ * Encoded theora frames must be a multiple of 16 in width and height.
+ * To handle other frame sizes, a crop rectangle is specified in
+ * frame_height and frame_width, offset_x and * offset_y. The offset
+ * and size should still be a multiple of 2 to avoid chroma sampling
+ * shifts. Offset values in this structure are measured from the
+ * upper left of the image.
+ *
+ * Frame rate, in frames per second, is stored as a rational
+ * fraction. Aspect ratio is also stored as a rational fraction, and
+ * refers to the aspect ratio of the frame pixels, not of the
+ * overall frame itself.
+ *
+ * See <a href="http://svn.xiph.org/trunk/theora/examples/encoder_example.c">
+ * examples/encoder_example.c</a> for usage examples of the
+ * other parameters and good default settings for the encoder parameters.
+ */
+typedef struct {
+  ogg_uint32_t  width;		/**< encoded frame width  */
+  ogg_uint32_t  height;		/**< encoded frame height */
+  ogg_uint32_t  frame_width;	/**< display frame width  */
+  ogg_uint32_t  frame_height;	/**< display frame height */
+  ogg_uint32_t  offset_x;	/**< horizontal offset of the displayed frame */
+  ogg_uint32_t  offset_y;	/**< vertical offset of the displayed frame */
+  ogg_uint32_t  fps_numerator;	    /**< frame rate numerator **/
+  ogg_uint32_t  fps_denominator;    /**< frame rate denominator **/
+  ogg_uint32_t  aspect_numerator;   /**< pixel aspect ratio numerator */
+  ogg_uint32_t  aspect_denominator; /**< pixel aspect ratio denominator */
+  theora_colorspace colorspace;	    /**< colorspace */
+  int           target_bitrate;	    /**< nominal bitrate in bits per second */
+  int           quality;  /**< Nominal quality setting, 0-63 */
+  int           quick_p;  /**< Quick encode/decode */
+
+  /* decode only */
+  unsigned char version_major;
+  unsigned char version_minor;
+  unsigned char version_subminor;
+
+  void *codec_setup;
+
+  /* encode only */
+  int           dropframes_p;
+  int           keyframe_auto_p;
+  ogg_uint32_t  keyframe_frequency;
+  ogg_uint32_t  keyframe_frequency_force;  /* also used for decode init to
+                                              get granpos shift correct */
+  ogg_uint32_t  keyframe_data_target_bitrate;
+  ogg_int32_t   keyframe_auto_threshold;
+  ogg_uint32_t  keyframe_mindistance;
+  ogg_int32_t   noise_sensitivity;
+  ogg_int32_t   sharpness;
+
+  theora_pixelformat pixelformat;	/**< chroma subsampling mode to expect */
+
+} theora_info;
+
+/** Codec internal state and context.
+ */
+typedef struct{
+  theora_info *i;
+  ogg_int64_t granulepos;
+
+  void *internal_encode;
+  void *internal_decode;
+
+} theora_state;
+
+/**
+ * Comment header metadata.
+ *
+ * This structure holds the in-stream metadata corresponding to
+ * the 'comment' header packet.
+ *
+ * Meta data is stored as a series of (tag, value) pairs, in
+ * length-encoded string vectors. The first occurrence of the
+ * '=' character delimits the tag and value. A particular tag
+ * may occur more than once. The character set encoding for
+ * the strings is always UTF-8, but the tag names are limited
+ * to case-insensitive ASCII. See the spec for details.
+ *
+ * In filling in this structure, theora_decode_header() will
+ * null-terminate the user_comment strings for safety. However,
+ * the bitstream format itself treats them as 8-bit clean,
+ * and so the length array should be treated as authoritative
+ * for their length.
+ */
+typedef struct theora_comment{
+  char **user_comments;         /**< An array of comment string vectors */
+  int   *comment_lengths;       /**< An array of corresponding string vector lengths in bytes */
+  int    comments;              /**< The total number of comment string vectors */
+  char  *vendor;                /**< The vendor string identifying the encoder, null terminated */
+
+} theora_comment;
+
+
+/**\name theora_control() codes */
+/* \anchor decctlcodes_old
+ * These are the available request codes for theora_control()
+ * when called with a decoder instance.
+ * By convention decoder control codes are odd, to distinguish
+ * them from \ref encctlcodes_old "encoder control codes" which
+ * are even.
+ *
+ * Note that since the 1.0 release, both the legacy and the final
+ * implementation accept all the same control codes, but only the
+ * final API declares the newer codes.
+ *
+ * Keep any experimental or vendor-specific values above \c 0x8000.*/
+
+/*@{*/
+
+/**Get the maximum post-processing level.
+ * The decoder supports a post-processing filter that can improve
+ * the appearance of the decoded images. This returns the highest
+ * level setting for this post-processor, corresponding to maximum
+ * improvement and computational expense.
+ */
+#define TH_DECCTL_GET_PPLEVEL_MAX (1)
+
+/**Set the post-processing level.
+ * Sets the level of post-processing to use when decoding the
+ * compressed stream. This must be a value between zero (off)
+ * and the maximum returned by TH_DECCTL_GET_PPLEVEL_MAX.
+ */
+#define TH_DECCTL_SET_PPLEVEL (3)
+
+/**Sets the maximum distance between key frames.
+ * This can be changed during an encode, but will be bounded by
+ *  <tt>1<<th_info#keyframe_granule_shift</tt>.
+ * If it is set before encoding begins, th_info#keyframe_granule_shift will
+ *  be enlarged appropriately.
+ *
+ * \param[in]  buf <tt>ogg_uint32_t</tt>: The maximum distance between key
+ *                   frames.
+ * \param[out] buf <tt>ogg_uint32_t</tt>: The actual maximum distance set.
+ * \retval OC_FAULT  \a theora_state or \a buf is <tt>NULL</tt>.
+ * \retval OC_EINVAL \a buf_sz is not <tt>sizeof(ogg_uint32_t)</tt>.
+ * \retval OC_IMPL   Not supported by this implementation.*/
+#define TH_ENCCTL_SET_KEYFRAME_FREQUENCY_FORCE (4)
+
+/**Set the granule position.
+ * Call this after a seek, to update the internal granulepos
+ * in the decoder, to insure that subsequent frames are marked
+ * properly. If you track timestamps yourself and do not use
+ * the granule position returned by the decoder, then you do
+ * not need to use this control.
+ */
+#define TH_DECCTL_SET_GRANPOS (5)
+
+/**\anchor encctlcodes_old */
+
+/**Sets the quantization parameters to use.
+ * The parameters are copied, not stored by reference, so they can be freed
+ *  after this call.
+ * <tt>NULL</tt> may be specified to revert to the default parameters.
+ *
+ * \param[in] buf #th_quant_info
+ * \retval OC_FAULT  \a theora_state is <tt>NULL</tt>.
+ * \retval OC_EINVAL Encoding has already begun, the quantization parameters
+ *                    are not acceptable to this version of the encoder,
+ *                    \a buf is <tt>NULL</tt> and \a buf_sz is not zero,
+ *                    or \a buf is non-<tt>NULL</tt> and \a buf_sz is
+ *                    not <tt>sizeof(#th_quant_info)</tt>.
+ * \retval OC_IMPL   Not supported by this implementation.*/
+#define TH_ENCCTL_SET_QUANT_PARAMS (2)
+
+/**Disables any encoder features that would prevent lossless transcoding back
+ *  to VP3.
+ * This primarily means disabling block-level QI values and not using 4MV mode
+ *  when any of the luma blocks in a macro block are not coded.
+ * It also includes using the VP3 quantization tables and Huffman codes; if you
+ *  set them explicitly after calling this function, the resulting stream will
+ *  not be VP3-compatible.
+ * If you enable VP3-compatibility when encoding 4:2:2 or 4:4:4 source
+ *  material, or when using a picture region smaller than the full frame (e.g.
+ *  a non-multiple-of-16 width or height), then non-VP3 bitstream features will
+ *  still be disabled, but the stream will still not be VP3-compatible, as VP3
+ *  was not capable of encoding such formats.
+ * If you call this after encoding has already begun, then the quantization
+ *  tables and codebooks cannot be changed, but the frame-level features will
+ *  be enabled or disabled as requested.
+ *
+ * \param[in]  buf <tt>int</tt>: a non-zero value to enable VP3 compatibility,
+ *                   or 0 to disable it (the default).
+ * \param[out] buf <tt>int</tt>: 1 if all bitstream features required for
+ *                   VP3-compatibility could be set, and 0 otherwise.
+ *                  The latter will be returned if the pixel format is not
+ *                   4:2:0, the picture region is smaller than the full frame,
+ *                   or if encoding has begun, preventing the quantization
+ *                   tables and codebooks from being set.
+ * \retval OC_FAULT  \a theora_state or \a buf is <tt>NULL</tt>.
+ * \retval OC_EINVAL \a buf_sz is not <tt>sizeof(int)</tt>.
+ * \retval OC_IMPL   Not supported by this implementation.*/
+#define TH_ENCCTL_SET_VP3_COMPATIBLE (10)
+
+/**Gets the maximum speed level.
+ * Higher speed levels favor quicker encoding over better quality per bit.
+ * Depending on the encoding mode, and the internal algorithms used, quality
+ *  may actually improve, but in this case bitrate will also likely increase.
+ * In any case, overall rate/distortion performance will probably decrease.
+ * The maximum value, and the meaning of each value, may change depending on
+ *  the current encoding mode (VBR vs. CQI, etc.).
+ *
+ * \param[out] buf int: The maximum encoding speed level.
+ * \retval OC_FAULT  \a theora_state or \a buf is <tt>NULL</tt>.
+ * \retval OC_EINVAL \a buf_sz is not <tt>sizeof(int)</tt>.
+ * \retval OC_IMPL   Not supported by this implementation in the current
+ *                    encoding mode.*/
+#define TH_ENCCTL_GET_SPLEVEL_MAX (12)
+
+/**Sets the speed level.
+ * By default a speed value of 1 is used.
+ *
+ * \param[in] buf int: The new encoding speed level.
+ *                      0 is slowest, larger values use less CPU.
+ * \retval OC_FAULT  \a theora_state or \a buf is <tt>NULL</tt>.
+ * \retval OC_EINVAL \a buf_sz is not <tt>sizeof(int)</tt>, or the
+ *                    encoding speed level is out of bounds.
+ *                   The maximum encoding speed level may be
+ *                    implementation- and encoding mode-specific, and can be
+ *                    obtained via #TH_ENCCTL_GET_SPLEVEL_MAX.
+ * \retval OC_IMPL   Not supported by this implementation in the current
+ *                    encoding mode.*/
+#define TH_ENCCTL_SET_SPLEVEL (14)
+
+/*@}*/
+
+#define OC_FAULT       -1       /**< General failure */
+#define OC_EINVAL      -10      /**< Library encountered invalid internal data */
+#define OC_DISABLED    -11      /**< Requested action is disabled */
+#define OC_BADHEADER   -20      /**< Header packet was corrupt/invalid */
+#define OC_NOTFORMAT   -21      /**< Packet is not a theora packet */
+#define OC_VERSION     -22      /**< Bitstream version is not handled */
+#define OC_IMPL        -23      /**< Feature or action not implemented */
+#define OC_BADPACKET   -24      /**< Packet is corrupt */
+#define OC_NEWPACKET   -25      /**< Packet is an (ignorable) unhandled extension */
+#define OC_DUPFRAME    1        /**< Packet is a dropped frame */
+
+/**
+ * Retrieve a human-readable string to identify the encoder vendor and version.
+ * \returns A version string.
+ */
+extern const char *theora_version_string(void);
+
+/**
+ * Retrieve a 32-bit version number.
+ * This number is composed of a 16-bit major version, 8-bit minor version
+ * and 8 bit sub-version, composed as follows:
+<pre>
+   (VERSION_MAJOR<<16) + (VERSION_MINOR<<8) + (VERSION_SUB)
+</pre>
+* \returns The version number.
+*/
+extern ogg_uint32_t theora_version_number(void);
+
+/**
+ * Initialize the theora encoder.
+ * \param th The theora_state handle to initialize for encoding.
+ * \param ti A theora_info struct filled with the desired encoding parameters.
+ * \retval 0 Success
+ */
+extern int theora_encode_init(theora_state *th, theora_info *ti);
+
+/**
+ * Submit a YUV buffer to the theora encoder.
+ * \param t A theora_state handle previously initialized for encoding.
+ * \param yuv A buffer of YUV data to encode.  Note that both the yuv_buffer
+ *            struct and the luma/chroma buffers within should be allocated by
+ *            the user.
+ * \retval OC_EINVAL Encoder is not ready, or is finished.
+ * \retval -1 The size of the given frame differs from those previously input
+ * \retval 0 Success
+ */
+extern int theora_encode_YUVin(theora_state *t, yuv_buffer *yuv);
+
+/**
+ * Request the next packet of encoded video.
+ * The encoded data is placed in a user-provided ogg_packet structure.
+ * \param t A theora_state handle previously initialized for encoding.
+ * \param last_p whether this is the last packet the encoder should produce.
+ * \param op An ogg_packet structure to fill. libtheora will set all
+ *           elements of this structure, including a pointer to encoded
+ *           data. The memory for the encoded data is owned by libtheora.
+ * \retval 0 No internal storage exists OR no packet is ready
+ * \retval -1 The encoding process has completed
+ * \retval 1 Success
+ */
+extern int theora_encode_packetout( theora_state *t, int last_p,
+                                    ogg_packet *op);
+
+/**
+ * Request a packet containing the initial header.
+ * A pointer to the header data is placed in a user-provided ogg_packet
+ * structure.
+ * \param t A theora_state handle previously initialized for encoding.
+ * \param op An ogg_packet structure to fill. libtheora will set all
+ *           elements of this structure, including a pointer to the header
+ *           data. The memory for the header data is owned by libtheora.
+ * \retval 0 Success
+ */
+extern int theora_encode_header(theora_state *t, ogg_packet *op);
+
+/**
+ * Request a comment header packet from provided metadata.
+ * A pointer to the comment data is placed in a user-provided ogg_packet
+ * structure.
+ * \param tc A theora_comment structure filled with the desired metadata
+ * \param op An ogg_packet structure to fill. libtheora will set all
+ *           elements of this structure, including a pointer to the encoded
+ *           comment data. The memory for the comment data is owned by
+ *           the application, and must be freed by it using _ogg_free().
+ *           On some systems (such as Windows when using dynamic linking), this
+ *           may mean the free is executed in a different module from the
+ *           malloc, which will crash; there is no way to free this memory on
+ *           such systems.
+ * \retval 0 Success
+ */
+extern int theora_encode_comment(theora_comment *tc, ogg_packet *op);
+
+/**
+ * Request a packet containing the codebook tables for the stream.
+ * A pointer to the codebook data is placed in a user-provided ogg_packet
+ * structure.
+ * \param t A theora_state handle previously initialized for encoding.
+ * \param op An ogg_packet structure to fill. libtheora will set all
+ *           elements of this structure, including a pointer to the codebook
+ *           data. The memory for the header data is owned by libtheora.
+ * \retval 0 Success
+ */
+extern int theora_encode_tables(theora_state *t, ogg_packet *op);
+
+/**
+ * Decode an Ogg packet, with the expectation that the packet contains
+ * an initial header, comment data or codebook tables.
+ *
+ * \param ci A theora_info structure to fill. This must have been previously
+ *           initialized with theora_info_init(). If \a op contains an initial
+ *           header, theora_decode_header() will fill \a ci with the
+ *           parsed header values. If \a op contains codebook tables,
+ *           theora_decode_header() will parse these and attach an internal
+ *           representation to \a ci->codec_setup.
+ * \param cc A theora_comment structure to fill. If \a op contains comment
+ *           data, theora_decode_header() will fill \a cc with the parsed
+ *           comments.
+ * \param op An ogg_packet structure which you expect contains an initial
+ *           header, comment data or codebook tables.
+ *
+ * \retval OC_BADHEADER \a op is NULL; OR the first byte of \a op->packet
+ *                      has the signature of an initial packet, but op is
+ *                      not a b_o_s packet; OR this packet has the signature
+ *                      of an initial header packet, but an initial header
+ *                      packet has already been seen; OR this packet has the
+ *                      signature of a comment packet, but the initial header
+ *                      has not yet been seen; OR this packet has the signature
+ *                      of a comment packet, but contains invalid data; OR
+ *                      this packet has the signature of codebook tables,
+ *                      but the initial header or comments have not yet
+ *                      been seen; OR this packet has the signature of codebook
+ *                      tables, but contains invalid data;
+ *                      OR the stream being decoded has a compatible version
+ *                      but this packet does not have the signature of a
+ *                      theora initial header, comments, or codebook packet
+ * \retval OC_VERSION   The packet data of \a op is an initial header with
+ *                      a version which is incompatible with this version of
+ *                      libtheora.
+ * \retval OC_NEWPACKET the stream being decoded has an incompatible (future)
+ *                      version and contains an unknown signature.
+ * \retval 0            Success
+ *
+ * \note The normal usage is that theora_decode_header() be called on the
+ *       first three packets of a theora logical bitstream in succession.
+ */
+extern int theora_decode_header(theora_info *ci, theora_comment *cc,
+                                ogg_packet *op);
+
+/**
+ * Initialize a theora_state handle for decoding.
+ * \param th The theora_state handle to initialize.
+ * \param c  A theora_info struct filled with the desired decoding parameters.
+ *           This is of course usually obtained from a previous call to
+ *           theora_decode_header().
+ * \retval 0 Success
+ */
+extern int theora_decode_init(theora_state *th, theora_info *c);
+
+/**
+ * Input a packet containing encoded data into the theora decoder.
+ * \param th A theora_state handle previously initialized for decoding.
+ * \param op An ogg_packet containing encoded theora data.
+ * \retval 0 Success
+ * \retval OC_BADPACKET \a op does not contain encoded video data
+ */
+extern int theora_decode_packetin(theora_state *th,ogg_packet *op);
+
+/**
+ * Output the next available frame of decoded YUV data.
+ * \param th A theora_state handle previously initialized for decoding.
+ * \param yuv A yuv_buffer in which libtheora should place the decoded data.
+ *            Note that the buffer struct itself is allocated by the user, but
+ *            that the luma and chroma pointers will be filled in by the
+ *            library.  Also note that these luma and chroma regions should be
+ *            considered read-only by the user.
+ * \retval 0 Success
+ */
+extern int theora_decode_YUVout(theora_state *th,yuv_buffer *yuv);
+
+/**
+ * Report whether a theora packet is a header or not
+ * This function does no verification beyond checking the header
+ * flag bit so it should not be used for bitstream identification;
+ * use theora_decode_header() for that.
+ *
+ * \param op An ogg_packet containing encoded theora data.
+ * \retval 1 The packet is a header packet
+ * \retval 0 The packet is not a header packet (and so contains frame data)
+ *
+ * Thus function was added in the 1.0alpha4 release.
+ */
+extern int theora_packet_isheader(ogg_packet *op);
+
+/**
+ * Report whether a theora packet is a keyframe or not
+ *
+ * \param op An ogg_packet containing encoded theora data.
+ * \retval 1 The packet contains a keyframe image
+ * \retval 0 The packet is contains an interframe delta
+ * \retval -1 The packet is not an image data packet at all
+ *
+ * Thus function was added in the 1.0alpha4 release.
+ */
+extern int theora_packet_iskeyframe(ogg_packet *op);
+
+/**
+ * Report the granulepos shift radix
+ *
+ * When embedded in Ogg, Theora uses a two-part granulepos,
+ * splitting the 64-bit field into two pieces. The more-significant
+ * section represents the frame count at the last keyframe,
+ * and the less-significant section represents the count of
+ * frames since the last keyframe. In this way the overall
+ * field is still non-decreasing with time, but usefully encodes
+ * a pointer to the last keyframe, which is necessary for
+ * correctly restarting decode after a seek.
+ *
+ * This function reports the number of bits used to represent
+ * the distance to the last keyframe, and thus how the granulepos
+ * field must be shifted or masked to obtain the two parts.
+ *
+ * Since libtheora returns compressed data in an ogg_packet
+ * structure, this may be generally useful even if the Theora
+ * packets are not being used in an Ogg container.
+ *
+ * \param ti A previously initialized theora_info struct
+ * \returns The bit shift dividing the two granulepos fields
+ *
+ * This function was added in the 1.0alpha5 release.
+ */
+int theora_granule_shift(theora_info *ti);
+
+/**
+ * Convert a granulepos to an absolute frame index, starting at 0.
+ * The granulepos is interpreted in the context of a given theora_state handle.
+ *
+ * Note that while the granulepos encodes the frame count (i.e. starting
+ * from 1) this call returns the frame index, starting from zero. Thus
+ * One can calculate the presentation time by multiplying the index by
+ * the rate.
+ *
+ * \param th A previously initialized theora_state handle (encode or decode)
+ * \param granulepos The granulepos to convert.
+ * \returns The frame index corresponding to \a granulepos.
+ * \retval -1 The given granulepos is undefined (i.e. negative)
+ *
+ * Thus function was added in the 1.0alpha4 release.
+ */
+extern ogg_int64_t theora_granule_frame(theora_state *th,ogg_int64_t granulepos);
+
+/**
+ * Convert a granulepos to absolute time in seconds. The granulepos is
+ * interpreted in the context of a given theora_state handle, and gives
+ * the end time of a frame's presentation as used in Ogg mux ordering.
+ *
+ * \param th A previously initialized theora_state handle (encode or decode)
+ * \param granulepos The granulepos to convert.
+ * \returns The absolute time in seconds corresponding to \a granulepos.
+ *          This is the "end time" for the frame, or the latest time it should
+ *           be displayed.
+ *          It is not the presentation time.
+ * \retval -1. The given granulepos is undefined (i.e. negative).
+ */
+extern double theora_granule_time(theora_state *th,ogg_int64_t granulepos);
+
+/**
+ * Initialize a theora_info structure. All values within the given theora_info
+ * structure are initialized, and space is allocated within libtheora for
+ * internal codec setup data.
+ * \param c A theora_info struct to initialize.
+ */
+extern void theora_info_init(theora_info *c);
+
+/**
+ * Clear a theora_info structure. All values within the given theora_info
+ * structure are cleared, and associated internal codec setup data is freed.
+ * \param c A theora_info struct to initialize.
+ */
+extern void theora_info_clear(theora_info *c);
+
+/**
+ * Free all internal data associated with a theora_state handle.
+ * \param t A theora_state handle.
+ */
+extern void theora_clear(theora_state *t);
+
+/**
+ * Initialize an allocated theora_comment structure
+ * \param tc An allocated theora_comment structure
+ **/
+extern void theora_comment_init(theora_comment *tc);
+
+/**
+ * Add a comment to an initialized theora_comment structure
+ * \param tc A previously initialized theora comment structure
+ * \param comment A null-terminated string encoding the comment in the form
+ *                "TAG=the value"
+ *
+ * Neither theora_comment_add() nor theora_comment_add_tag() support
+ * comments containing null values, although the bitstream format
+ * supports this. To add such comments you will need to manipulate
+ * the theora_comment structure directly.
+ **/
+
+extern void theora_comment_add(theora_comment *tc, char *comment);
+
+/**
+ * Add a comment to an initialized theora_comment structure.
+ * \param tc A previously initialized theora comment structure
+ * \param tag A null-terminated string containing the tag
+ *            associated with the comment.
+ * \param value The corresponding value as a null-terminated string
+ *
+ * Neither theora_comment_add() nor theora_comment_add_tag() support
+ * comments containing null values, although the bitstream format
+ * supports this. To add such comments you will need to manipulate
+ * the theora_comment structure directly.
+ **/
+extern void theora_comment_add_tag(theora_comment *tc,
+                                       char *tag, char *value);
+
+/**
+ * Look up a comment value by tag.
+ * \param tc Tn initialized theora_comment structure
+ * \param tag The tag to look up
+ * \param count The instance of the tag. The same tag can appear multiple
+ *              times, each with a distinct and ordered value, so an index
+ *              is required to retrieve them all.
+ * \returns A pointer to the queried tag's value
+ * \retval NULL No matching tag is found
+ *
+ * \note Use theora_comment_query_count() to get the legal range for the
+ * count parameter.
+ **/
+
+extern char *theora_comment_query(theora_comment *tc, char *tag, int count);
+
+/** Look up the number of instances of a tag.
+ *  \param tc An initialized theora_comment structure
+ *  \param tag The tag to look up
+ *  \returns The number on instances of a particular tag.
+ *
+ *  Call this first when querying for a specific tag and then iterate
+ *  over the number of instances with separate calls to
+ *  theora_comment_query() to retrieve all instances in order.
+ **/
+extern int   theora_comment_query_count(theora_comment *tc, char *tag);
+
+/**
+ * Clear an allocated theora_comment struct so that it can be freed.
+ * \param tc An allocated theora_comment structure.
+ **/
+extern void  theora_comment_clear(theora_comment *tc);
+
+/**Encoder control function.
+ * This is used to provide advanced control the encoding process.
+ * \param th     A #theora_state handle.
+ * \param req    The control code to process.
+ *                See \ref encctlcodes_old "the list of available
+ *			control codes" for details.
+ * \param buf    The parameters for this control code.
+ * \param buf_sz The size of the parameter buffer.*/
+extern int theora_control(theora_state *th,int req,void *buf,size_t buf_sz);
+
+/* @} */ /* end oldfuncs doxygen group */
+
+#ifdef __cplusplus
+}
+#endif /* __cplusplus */
+
+#endif /* OGG_THEORA_HEADER */

thirdparty/libtheora/theora/theoradec.h

@@ -0,0 +1,332 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation https://www.xiph.org/                 *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+/**\file
+ * The <tt>libtheoradec</tt> C decoding API.*/
+
+#if !defined(OGG_THEORA_THEORADEC_HEADER)
+# define OGG_THEORA_THEORADEC_HEADER (1)
+# include <stddef.h>
+# include <ogg/ogg.h>
+# include "codec.h"
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+
+
+/**\name th_decode_ctl() codes
+ * \anchor decctlcodes
+ * These are the available request codes for th_decode_ctl().
+ * By convention, these are odd, to distinguish them from the
+ *  \ref encctlcodes "encoder control codes".
+ * Keep any experimental or vendor-specific values above \c 0x8000.*/
+/*@{*/
+/**Gets the maximum post-processing level.
+ * The decoder supports a post-processing filter that can improve
+ * the appearance of the decoded images. This returns the highest
+ * level setting for this post-processor, corresponding to maximum
+ * improvement and computational expense.
+ *
+ * \param[out] _buf int: The maximum post-processing level.
+ * \retval TH_EFAULT  \a _dec_ctx or \a _buf is <tt>NULL</tt>.
+ * \retval TH_EINVAL  \a _buf_sz is not <tt>sizeof(int)</tt>.
+ * \retval TH_EIMPL   Not supported by this implementation.*/
+#define TH_DECCTL_GET_PPLEVEL_MAX (1)
+/**Sets the post-processing level.
+ * By default, post-processing is disabled.
+ *
+ * Sets the level of post-processing to use when decoding the
+ * compressed stream. This must be a value between zero (off)
+ * and the maximum returned by TH_DECCTL_GET_PPLEVEL_MAX.
+ *
+ * \param[in] _buf int: The new post-processing level.
+ *                      0 to disable; larger values use more CPU.
+ * \retval TH_EFAULT  \a _dec_ctx or \a _buf is <tt>NULL</tt>.
+ * \retval TH_EINVAL  \a _buf_sz is not <tt>sizeof(int)</tt>, or the
+ *                     post-processing level is out of bounds.
+ *                    The maximum post-processing level may be
+ *                     implementation-specific, and can be obtained via
+ *                     #TH_DECCTL_GET_PPLEVEL_MAX.
+ * \retval TH_EIMPL   Not supported by this implementation.*/
+#define TH_DECCTL_SET_PPLEVEL (3)
+/**Sets the granule position.
+ * Call this after a seek, before decoding the first frame, to ensure that the
+ *  proper granule position is returned for all subsequent frames.
+ * If you track timestamps yourself and do not use the granule position
+ *  returned by the decoder, then you need not call this function.
+ *
+ * \param[in] _buf <tt>ogg_int64_t</tt>: The granule position of the next
+ *                  frame.
+ * \retval TH_EFAULT  \a _dec_ctx or \a _buf is <tt>NULL</tt>.
+ * \retval TH_EINVAL  \a _buf_sz is not <tt>sizeof(ogg_int64_t)</tt>, or the
+ *                     granule position is negative.*/
+#define TH_DECCTL_SET_GRANPOS (5)
+/**Sets the striped decode callback function.
+ * If set, this function will be called as each piece of a frame is fully
+ *  decoded in th_decode_packetin().
+ * You can pass in a #th_stripe_callback with
+ *  th_stripe_callback#stripe_decoded set to <tt>NULL</tt> to disable the
+ *  callbacks at any point.
+ * Enabling striped decode does not prevent you from calling
+ *  th_decode_ycbcr_out() after the frame is fully decoded.
+ *
+ * \param[in]  _buf #th_stripe_callback: The callback parameters.
+ * \retval TH_EFAULT  \a _dec_ctx or \a _buf is <tt>NULL</tt>.
+ * \retval TH_EINVAL  \a _buf_sz is not
+ *                     <tt>sizeof(th_stripe_callback)</tt>.*/
+#define TH_DECCTL_SET_STRIPE_CB (7)
+
+/**Sets the macroblock display mode. Set to 0 to disable displaying
+ * macroblocks.*/
+#define TH_DECCTL_SET_TELEMETRY_MBMODE (9)
+/**Sets the motion vector display mode. Set to 0 to disable displaying motion
+ * vectors.*/
+#define TH_DECCTL_SET_TELEMETRY_MV (11)
+/**Sets the adaptive quantization display mode. Set to 0 to disable displaying
+ * adaptive quantization. */
+#define TH_DECCTL_SET_TELEMETRY_QI (13)
+/**Sets the bitstream breakdown visualization mode. Set to 0 to disable
+ * displaying bitstream breakdown.*/
+#define TH_DECCTL_SET_TELEMETRY_BITS (15)
+/*@}*/
+
+
+
+/**A callback function for striped decode.
+ * This is a function pointer to an application-provided function that will be
+ *  called each time a section of the image is fully decoded in
+ *  th_decode_packetin().
+ * This allows the application to process the section immediately, while it is
+ *  still in cache.
+ * Note that the frame is decoded bottom to top, so \a _yfrag0 will steadily
+ *  decrease with each call until it reaches 0, at which point the full frame
+ *  is decoded.
+ * The number of fragment rows made available in each call depends on the pixel
+ *  format and the number of post-processing filters enabled, and may not even
+ *  be constant for the entire frame.
+ * If a non-<tt>NULL</tt> \a _granpos pointer is passed to
+ *  th_decode_packetin(), the granule position for the frame will be stored
+ *  in it before the first callback is made.
+ * If an entire frame is dropped (a 0-byte packet), then no callbacks will be
+ *  made at all for that frame.
+ * \param _ctx       An application-provided context pointer.
+ * \param _buf       The image buffer for the decoded frame.
+ * \param _yfrag0    The Y coordinate of the first row of 8x8 fragments
+ *                    decoded.
+ *                   Multiply this by 8 to obtain the pixel row number in the
+ *                    luma plane.
+ *                   If the chroma planes are subsampled in the Y direction,
+ *                    this will always be divisible by two.
+ * \param _yfrag_end The Y coordinate of the first row of 8x8 fragments past
+ *                    the newly decoded section.
+ *                   If the chroma planes are subsampled in the Y direction,
+ *                    this will always be divisible by two.
+ *                   I.e., this section contains fragment rows
+ *                    <tt>\a _yfrag0 ...\a _yfrag_end -1</tt>.*/
+typedef void (*th_stripe_decoded_func)(void *_ctx,th_ycbcr_buffer _buf,
+ int _yfrag0,int _yfrag_end);
+
+/**The striped decode callback data to pass to #TH_DECCTL_SET_STRIPE_CB.*/
+typedef struct{
+  /**An application-provided context pointer.
+   * This will be passed back verbatim to the application.*/
+  void                   *ctx;
+  /**The callback function pointer.*/
+  th_stripe_decoded_func  stripe_decoded;
+}th_stripe_callback;
+
+
+
+/**\name Decoder state
+   The following data structures are opaque, and their contents are not
+    publicly defined by this API.
+   Referring to their internals directly is unsupported, and may break without
+    warning.*/
+/*@{*/
+/**The decoder context.*/
+typedef struct th_dec_ctx    th_dec_ctx;
+/**Setup information.
+   This contains auxiliary information (Huffman tables and quantization
+    parameters) decoded from the setup header by th_decode_headerin() to be
+    passed to th_decode_alloc().
+   It can be re-used to initialize any number of decoders, and can be freed
+    via th_setup_free() at any time.*/
+typedef struct th_setup_info th_setup_info;
+/*@}*/
+
+
+
+/**\defgroup decfuncs Functions for Decoding*/
+/*@{*/
+/**\name Functions for decoding
+ * You must link to <tt>libtheoradec</tt> if you use any of the
+ * functions in this section.
+ *
+ * The functions are listed in the order they are used in a typical decode.
+ * The basic steps are:
+ * - Parse the header packets by repeatedly calling th_decode_headerin().
+ * - Allocate a #th_dec_ctx handle with th_decode_alloc().
+ * - Call th_setup_free() to free any memory used for codec setup
+ *    information.
+ * - Perform any additional decoder configuration with th_decode_ctl().
+ * - For each video data packet:
+ *   - Submit the packet to the decoder via th_decode_packetin().
+ *   - Retrieve the uncompressed video data via th_decode_ycbcr_out().
+ * - Call th_decode_free() to release all decoder memory.*/
+/*@{*/
+/**Decodes the header packets of a Theora stream.
+ * This should be called on the initial packets of the stream, in succession,
+ *  until it returns <tt>0</tt>, indicating that all headers have been
+ *  processed, or an error is encountered.
+ * At least three header packets are required, and additional optional header
+ *  packets may follow.
+ * This can be used on the first packet of any logical stream to determine if
+ *  that stream is a Theora stream.
+ * \param _info  A #th_info structure to fill in.
+ *               This must have been previously initialized with
+ *                th_info_init().
+ *               The application may immediately begin using the contents of
+ *                this structure after the first header is decoded, though it
+ *                must continue to be passed in on all subsequent calls.
+ * \param _tc    A #th_comment structure to fill in.
+ *               The application may immediately begin using the contents of
+ *                this structure after the second header is decoded, though it
+ *                must continue to be passed in on all subsequent calls.
+ * \param _setup Returns a pointer to additional, private setup information
+ *                needed by the decoder.
+ *               The contents of this pointer must be initialized to
+ *                <tt>NULL</tt> on the first call, and the returned value must
+ *                continue to be passed in on all subsequent calls.
+ * \param _op    An <tt>ogg_packet</tt> structure which contains one of the
+ *                initial packets of an Ogg logical stream.
+ * \return A positive value indicates that a Theora header was successfully
+ *          processed.
+ * \retval 0             The first video data packet was encountered after all
+ *                        required header packets were parsed.
+ *                       The packet just passed in on this call should be saved
+ *                        and fed to th_decode_packetin() to begin decoding
+ *                        video data.
+ * \retval TH_EFAULT     One of \a _info, \a _tc, or \a _setup was
+ *                        <tt>NULL</tt>.
+ * \retval TH_EBADHEADER \a _op was <tt>NULL</tt>, the packet was not the next
+ *                        header packet in the expected sequence, or the format
+ *                        of the header data was invalid.
+ * \retval TH_EVERSION   The packet data was a Theora info header, but for a
+ *                        bitstream version not decodable with this version of
+ *                        <tt>libtheoradec</tt>.
+ * \retval TH_ENOTFORMAT The packet was not a Theora header.
+ */
+extern int th_decode_headerin(th_info *_info,th_comment *_tc,
+ th_setup_info **_setup,ogg_packet *_op);
+/**Allocates a decoder instance.
+ *
+ * <b>Security Warning:</b> The Theora format supports very large frame sizes,
+ *  potentially even larger than the address space of a 32-bit machine, and
+ *  creating a decoder context allocates the space for several frames of data.
+ * If the allocation fails here, your program will crash, possibly at some
+ *  future point because the OS kernel returned a valid memory range and will
+ *  only fail when it tries to map the pages in it the first time they are
+ *  used.
+ * Even if it succeeds, you may experience a denial of service if the frame
+ *  size is large enough to cause excessive paging.
+ * If you are integrating libtheora in a larger application where such things
+ *  are undesirable, it is highly recommended that you check the frame size in
+ *  \a _info before calling this function and refuse to decode streams where it
+ *  is larger than some reasonable maximum.
+ * libtheora will not check this for you, because there may be machines that
+ *  can handle such streams and applications that wish to.
+ * \param _info  A #th_info struct filled via th_decode_headerin().
+ * \param _setup A #th_setup_info handle returned via
+ *                th_decode_headerin().
+ * \return The initialized #th_dec_ctx handle.
+ * \retval NULL If the decoding parameters were invalid.*/
+extern th_dec_ctx *th_decode_alloc(const th_info *_info,
+ const th_setup_info *_setup);
+/**Releases all storage used for the decoder setup information.
+ * This should be called after you no longer want to create any decoders for
+ *  a stream whose headers you have parsed with th_decode_headerin().
+ * \param _setup The setup information to free.
+ *               This can safely be <tt>NULL</tt>.*/
+extern void th_setup_free(th_setup_info *_setup);
+/**Decoder control function.
+ * This is used to provide advanced control of the decoding process.
+ * \param _dec    A #th_dec_ctx handle.
+ * \param _req    The control code to process.
+ *                See \ref decctlcodes "the list of available control codes"
+ *                 for details.
+ * \param _buf    The parameters for this control code.
+ * \param _buf_sz The size of the parameter buffer.
+ * \return Possible return values depend on the control code used.
+ *          See \ref decctlcodes "the list of control codes" for
+ *          specific values. Generally 0 indicates success.*/
+extern int th_decode_ctl(th_dec_ctx *_dec,int _req,void *_buf,
+ size_t _buf_sz);
+/**Submits a packet containing encoded video data to the decoder.
+ * \param _dec     A #th_dec_ctx handle.
+ * \param _op      An <tt>ogg_packet</tt> containing encoded video data.
+ * \param _granpos Returns the granule position of the decoded packet.
+ *                 If non-<tt>NULL</tt>, the granule position for this specific
+ *                  packet is stored in this location.
+ *                 This is computed incrementally from previously decoded
+ *                  packets.
+ *                 After a seek, the correct granule position must be set via
+ *                  #TH_DECCTL_SET_GRANPOS for this to work properly.
+ * \retval 0             Success.
+ *                       A new decoded frame can be retrieved by calling
+ *                        th_decode_ycbcr_out().
+ * \retval TH_DUPFRAME   The packet represented a dropped frame (either a
+ *                        0-byte frame or an INTER frame with no coded blocks).
+ *                       The player can skip the call to th_decode_ycbcr_out(),
+ *                        as the contents of the decoded frame buffer have not
+ *                        changed.
+ * \retval TH_EFAULT     \a _dec or \a _op was <tt>NULL</tt>.
+ * \retval TH_EBADPACKET \a _op does not contain encoded video data.
+ * \retval TH_EIMPL      The video data uses bitstream features which this
+ *                        library does not support.*/
+extern int th_decode_packetin(th_dec_ctx *_dec,const ogg_packet *_op,
+ ogg_int64_t *_granpos);
+/**Outputs the next available frame of decoded Y'CbCr data.
+ * If a striped decode callback has been set with #TH_DECCTL_SET_STRIPE_CB,
+ *  then the application does not need to call this function.
+ * \param _dec   A #th_dec_ctx handle.
+ * \param _ycbcr A video buffer structure to fill in.
+ *               <tt>libtheoradec</tt> will fill in all the members of this
+ *                structure, including the pointers to the uncompressed video
+ *                data.
+ *               The memory for this video data is owned by
+ *                <tt>libtheoradec</tt>.
+ *               It may be freed or overwritten without notification when
+ *                subsequent frames are decoded.
+ * \retval 0 Success
+ * \retval TH_EFAULT     \a _dec or \a _ycbcr was <tt>NULL</tt>.
+ */
+extern int th_decode_ycbcr_out(th_dec_ctx *_dec,
+ th_ycbcr_buffer _ycbcr);
+/**Frees an allocated decoder instance.
+ * \param _dec A #th_dec_ctx handle.*/
+extern void th_decode_free(th_dec_ctx *_dec);
+/*@}*/
+/*@}*/
+
+
+
+#if defined(__cplusplus)
+}
+#endif
+
+#endif /* OGG_THEORA_THEORADEC_HEADER */

thirdparty/libtheora/theora/theoraenc.h

@@ -0,0 +1,547 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation https://www.xiph.org/                 *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+/**\file
+ * The <tt>libtheoraenc</tt> C encoding API.*/
+
+#if !defined(OGG_THEORA_THEORAENC_HEADER)
+# define OGG_THEORA_THEORAENC_HEADER (1)
+# include <stddef.h>
+# include <ogg/ogg.h>
+# include "codec.h"
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+
+
+/**\name th_encode_ctl() codes
+ * \anchor encctlcodes
+ * These are the available request codes for th_encode_ctl().
+ * By convention, these are even, to distinguish them from the
+ *  \ref decctlcodes "decoder control codes".
+ * Keep any experimental or vendor-specific values above \c 0x8000.*/
+/*@{*/
+/**Sets the Huffman tables to use.
+ * The tables are copied, not stored by reference, so they can be freed after
+ *  this call.
+ * <tt>NULL</tt> may be specified to revert to the default tables.
+ *
+ * \param[in] _buf <tt>#th_huff_code[#TH_NHUFFMAN_TABLES][#TH_NDCT_TOKENS]</tt>
+ * \retval TH_EFAULT \a _enc is <tt>NULL</tt>.
+ * \retval TH_EINVAL Encoding has already begun or one or more of the given
+ *                     tables is not full or prefix-free, \a _buf is
+ *                     <tt>NULL</tt> and \a _buf_sz is not zero, or \a _buf is
+ *                     non-<tt>NULL</tt> and \a _buf_sz is not
+ *                     <tt>sizeof(#th_huff_code)*#TH_NHUFFMAN_TABLES*#TH_NDCT_TOKENS</tt>.
+ * \retval TH_EIMPL   Not supported by this implementation.*/
+#define TH_ENCCTL_SET_HUFFMAN_CODES (0)
+/**Sets the quantization parameters to use.
+ * The parameters are copied, not stored by reference, so they can be freed
+ *  after this call.
+ * <tt>NULL</tt> may be specified to revert to the default parameters.
+ *
+ * \param[in] _buf #th_quant_info
+ * \retval TH_EFAULT \a _enc is <tt>NULL</tt>.
+ * \retval TH_EINVAL Encoding has already begun, \a _buf is
+ *                    <tt>NULL</tt> and \a _buf_sz is not zero,
+ *                    or \a _buf is non-<tt>NULL</tt> and
+ *                    \a _buf_sz is not <tt>sizeof(#th_quant_info)</tt>.
+ * \retval TH_EIMPL   Not supported by this implementation.*/
+#define TH_ENCCTL_SET_QUANT_PARAMS (2)
+/**Sets the maximum distance between key frames.
+ * This can be changed during an encode, but will be bounded by
+ *  <tt>1<<th_info#keyframe_granule_shift</tt>.
+ * If it is set before encoding begins, th_info#keyframe_granule_shift will
+ *  be enlarged appropriately.
+ *
+ * \param[in]  _buf <tt>ogg_uint32_t</tt>: The maximum distance between key
+ *                   frames.
+ * \param[out] _buf <tt>ogg_uint32_t</tt>: The actual maximum distance set.
+ * \retval TH_EFAULT \a _enc or \a _buf is <tt>NULL</tt>.
+ * \retval TH_EINVAL \a _buf_sz is not <tt>sizeof(ogg_uint32_t)</tt>.
+ * \retval TH_EIMPL   Not supported by this implementation.*/
+#define TH_ENCCTL_SET_KEYFRAME_FREQUENCY_FORCE (4)
+/**Disables any encoder features that would prevent lossless transcoding back
+ *  to VP3.
+ * This primarily means disabling block-adaptive quantization and always coding
+ *  all four luma blocks in a macro block when 4MV is used.
+ * It also includes using the VP3 quantization tables and Huffman codes; if you
+ *  set them explicitly after calling this function, the resulting stream will
+ *  not be VP3-compatible.
+ * If you enable VP3-compatibility when encoding 4:2:2 or 4:4:4 source
+ *  material, or when using a picture region smaller than the full frame (e.g.
+ *  a non-multiple-of-16 width or height), then non-VP3 bitstream features will
+ *  still be disabled, but the stream will still not be VP3-compatible, as VP3
+ *  was not capable of encoding such formats.
+ * If you call this after encoding has already begun, then the quantization
+ *  tables and codebooks cannot be changed, but the frame-level features will
+ *  be enabled or disabled as requested.
+ *
+ * \param[in]  _buf <tt>int</tt>: a non-zero value to enable VP3 compatibility,
+ *                   or 0 to disable it (the default).
+ * \param[out] _buf <tt>int</tt>: 1 if all bitstream features required for
+ *                   VP3-compatibility could be set, and 0 otherwise.
+ *                  The latter will be returned if the pixel format is not
+ *                   4:2:0, the picture region is smaller than the full frame,
+ *                   or if encoding has begun, preventing the quantization
+ *                   tables and codebooks from being set.
+ * \retval TH_EFAULT \a _enc or \a _buf is <tt>NULL</tt>.
+ * \retval TH_EINVAL \a _buf_sz is not <tt>sizeof(int)</tt>.
+ * \retval TH_EIMPL   Not supported by this implementation.*/
+#define TH_ENCCTL_SET_VP3_COMPATIBLE (10)
+/**Gets the maximum speed level.
+ * Higher speed levels favor quicker encoding over better quality per bit.
+ * Depending on the encoding mode, and the internal algorithms used, quality
+ *  may actually improve, but in this case bitrate will also likely increase.
+ * In any case, overall rate/distortion performance will probably decrease.
+ * The maximum value, and the meaning of each value, may change depending on
+ *  the current encoding mode (VBR vs. constant quality, etc.).
+ *
+ * \param[out] _buf <tt>int</tt>: The maximum encoding speed level.
+ * \retval TH_EFAULT \a _enc or \a _buf is <tt>NULL</tt>.
+ * \retval TH_EINVAL \a _buf_sz is not <tt>sizeof(int)</tt>.
+ * \retval TH_EIMPL   Not supported by this implementation in the current
+ *                    encoding mode.*/
+#define TH_ENCCTL_GET_SPLEVEL_MAX (12)
+/**Sets the speed level.
+ * The current speed level may be retrieved using #TH_ENCCTL_GET_SPLEVEL.
+ *
+ * \param[in] _buf <tt>int</tt>: The new encoding speed level.
+ *                 0 is slowest, larger values use less CPU.
+ * \retval TH_EFAULT \a _enc or \a _buf is <tt>NULL</tt>.
+ * \retval TH_EINVAL \a _buf_sz is not <tt>sizeof(int)</tt>, or the
+ *                    encoding speed level is out of bounds.
+ *                   The maximum encoding speed level may be
+ *                    implementation- and encoding mode-specific, and can be
+ *                    obtained via #TH_ENCCTL_GET_SPLEVEL_MAX.
+ * \retval TH_EIMPL   Not supported by this implementation in the current
+ *                    encoding mode.*/
+#define TH_ENCCTL_SET_SPLEVEL (14)
+/**Gets the current speed level.
+ * The default speed level may vary according to encoder implementation, but if
+ *  this control code is not supported (it returns #TH_EIMPL), the default may
+ *  be assumed to be the slowest available speed (0).
+ * The maximum encoding speed level may be implementation- and encoding
+ *  mode-specific, and can be obtained via #TH_ENCCTL_GET_SPLEVEL_MAX.
+ *
+ * \param[out] _buf <tt>int</tt>: The current encoding speed level.
+ *                  0 is slowest, larger values use less CPU.
+ * \retval TH_EFAULT \a _enc or \a _buf is <tt>NULL</tt>.
+ * \retval TH_EINVAL \a _buf_sz is not <tt>sizeof(int)</tt>.
+ * \retval TH_EIMPL   Not supported by this implementation in the current
+ *                    encoding mode.*/
+#define TH_ENCCTL_GET_SPLEVEL (16)
+/**Sets the number of duplicates of the next frame to produce.
+ * Although libtheora can encode duplicate frames very cheaply, it costs some
+ *  amount of CPU to detect them, and a run of duplicates cannot span a
+ *  keyframe boundary.
+ * This control code tells the encoder to produce the specified number of extra
+ *  duplicates of the next frame.
+ * This allows the encoder to make smarter keyframe placement decisions and
+ *  rate control decisions, and reduces CPU usage as well, when compared to
+ *  just submitting the same frame for encoding multiple times.
+ * This setting only applies to the next frame submitted for encoding.
+ * You MUST call th_encode_packetout() repeatedly until it returns 0, or the
+ *  extra duplicate frames will be lost.
+ *
+ * \param[in] _buf <tt>int</tt>: The number of duplicates to produce.
+ *                 If this is negative or zero, no duplicates will be produced.
+ * \retval TH_EFAULT \a _enc or \a _buf is <tt>NULL</tt>.
+ * \retval TH_EINVAL \a _buf_sz is not <tt>sizeof(int)</tt>, or the
+ *                    number of duplicates is greater than or equal to the
+ *                    maximum keyframe interval.
+ *                   In the latter case, NO duplicate frames will be produced.
+ *                   You must ensure that the maximum keyframe interval is set
+ *                    larger than the maximum number of duplicates you will
+ *                    ever wish to insert prior to encoding.
+ * \retval TH_EIMPL   Not supported by this implementation in the current
+ *                    encoding mode.*/
+#define TH_ENCCTL_SET_DUP_COUNT (18)
+/**Modifies the default bitrate management behavior.
+ * Use to allow or disallow frame dropping, and to enable or disable capping
+ *  bit reservoir overflows and underflows.
+ * See \ref encctlcodes "the list of available flags".
+ * The flags are set by default to
+ *  <tt>#TH_RATECTL_DROP_FRAMES|#TH_RATECTL_CAP_OVERFLOW</tt>.
+ *
+ * \param[in] _buf <tt>int</tt>: Any combination of
+ *                  \ref ratectlflags "the available flags":
+ *                 - #TH_RATECTL_DROP_FRAMES: Enable frame dropping.
+ *                 - #TH_RATECTL_CAP_OVERFLOW: Don't bank excess bits for later
+ *                    use.
+ *                 - #TH_RATECTL_CAP_UNDERFLOW: Don't try to make up shortfalls
+ *                    later.
+ * \retval TH_EFAULT \a _enc or \a _buf is <tt>NULL</tt>.
+ * \retval TH_EINVAL \a _buf_sz is not <tt>sizeof(int)</tt> or rate control
+ *                    is not enabled.
+ * \retval TH_EIMPL   Not supported by this implementation in the current
+ *                    encoding mode.*/
+#define TH_ENCCTL_SET_RATE_FLAGS (20)
+/**Sets the size of the bitrate management bit reservoir as a function
+ *  of number of frames.
+ * The reservoir size affects how quickly bitrate management reacts to
+ *  instantaneous changes in the video complexity.
+ * Larger reservoirs react more slowly, and provide better overall quality, but
+ *  require more buffering by a client, adding more latency to live streams.
+ * By default, libtheora sets the reservoir to the maximum distance between
+ *  keyframes, subject to a minimum and maximum limit.
+ * This call may be used to increase or decrease the reservoir, increasing or
+ *  decreasing the allowed temporary variance in bitrate.
+ * An implementation may impose some limits on the size of a reservoir it can
+ *  handle, in which case the actual reservoir size may not be exactly what was
+ *  requested.
+ * The actual value set will be returned.
+ *
+ * \param[in]  _buf <tt>int</tt>: Requested size of the reservoir measured in
+ *                   frames.
+ * \param[out] _buf <tt>int</tt>: The actual size of the reservoir set.
+ * \retval TH_EFAULT \a _enc or \a _buf is <tt>NULL</tt>.
+ * \retval TH_EINVAL \a _buf_sz is not <tt>sizeof(int)</tt>, or rate control
+ *                    is not enabled.  The buffer has an implementation
+ *                    defined minimum and maximum size and the value in _buf
+ *                    will be adjusted to match the actual value set.
+ * \retval TH_EIMPL   Not supported by this implementation in the current
+ *                    encoding mode.*/
+#define TH_ENCCTL_SET_RATE_BUFFER (22)
+/**Enable pass 1 of two-pass encoding mode and retrieve the first pass metrics.
+ * Pass 1 mode must be enabled before the first frame is encoded, and a target
+ *  bitrate must have already been specified to the encoder.
+ * Although this does not have to be the exact rate that will be used in the
+ *  second pass, closer values may produce better results.
+ * The first call returns the size of the two-pass header data, along with some
+ *  placeholder content, and sets the encoder into pass 1 mode implicitly.
+ * This call sets the encoder to pass 1 mode implicitly.
+ * Then, a subsequent call must be made after each call to
+ *  th_encode_ycbcr_in() to retrieve the metrics for that frame.
+ * An additional, final call must be made to retrieve the summary data,
+ *  containing such information as the total number of frames, etc.
+ * This must be stored in place of the placeholder data that was returned
+ *  in the first call, before the frame metrics data.
+ * All of this data must be presented back to the encoder during pass 2 using
+ *  #TH_ENCCTL_2PASS_IN.
+ *
+ * \param[out] <tt>char *</tt>_buf: Returns a pointer to internal storage
+ *              containing the two pass metrics data.
+ *             This storage is only valid until the next call, or until the
+ *              encoder context is freed, and must be copied by the
+ *              application.
+ * \retval >=0       The number of bytes of metric data available in the
+ *                    returned buffer.
+ * \retval TH_EFAULT \a _enc or \a _buf is <tt>NULL</tt>.
+ * \retval TH_EINVAL \a _buf_sz is not <tt>sizeof(char *)</tt>, no target
+ *                    bitrate has been set, or the first call was made after
+ *                    the first frame was submitted for encoding.
+ * \retval TH_EIMPL   Not supported by this implementation.*/
+#define TH_ENCCTL_2PASS_OUT (24)
+/**Submits two-pass encoding metric data collected the first encoding pass to
+ *  the second pass.
+ * The first call must be made before the first frame is encoded, and a target
+ *  bitrate must have already been specified to the encoder.
+ * It sets the encoder to pass 2 mode implicitly; this cannot be disabled.
+ * The encoder may require reading data from some or all of the frames in
+ *  advance, depending on, e.g., the reservoir size used in the second pass.
+ * You must call this function repeatedly before each frame to provide data
+ *  until either a) it fails to consume all of the data presented or b) all of
+ *  the pass 1 data has been consumed.
+ * In the first case, you must save the remaining data to be presented after
+ *  the next frame.
+ * You can call this function with a NULL argument to get an upper bound on
+ *  the number of bytes that will be required before the next frame.
+ *
+ * When pass 2 is first enabled, the default bit reservoir is set to the entire
+ *  file; this gives maximum flexibility but can lead to very high peak rates.
+ * You can subsequently set it to another value with #TH_ENCCTL_SET_RATE_BUFFER
+ *  (e.g., to set it to the keyframe interval for non-live streaming), however,
+ *  you may then need to provide more data before the next frame.
+ *
+ * \param[in] _buf <tt>char[]</tt>: A buffer containing the data returned by
+ *                  #TH_ENCCTL_2PASS_OUT in pass 1.
+ *                 You may pass <tt>NULL</tt> for \a _buf to return an upper
+ *                  bound on the number of additional bytes needed before the
+ *                  next frame.
+ *                 The summary data returned at the end of pass 1 must be at
+ *                  the head of the buffer on the first call with a
+ *                  non-<tt>NULL</tt> \a _buf, and the placeholder data
+ *                  returned at the start of pass 1 should be omitted.
+ *                 After each call you should advance this buffer by the number
+ *                  of bytes consumed.
+ * \retval >0            The number of bytes of metric data required/consumed.
+ * \retval 0             No more data is required before the next frame.
+ * \retval TH_EFAULT     \a _enc is <tt>NULL</tt>.
+ * \retval TH_EINVAL     No target bitrate has been set, or the first call was
+ *                        made after the first frame was submitted for
+ *                        encoding.
+ * \retval TH_ENOTFORMAT The data did not appear to be pass 1 from a compatible
+ *                        implementation of this library.
+ * \retval TH_EBADHEADER The data was invalid; this may be returned when
+ *                        attempting to read an aborted pass 1 file that still
+ *                        has the placeholder data in place of the summary
+ *                        data.
+ * \retval TH_EIMPL       Not supported by this implementation.*/
+#define TH_ENCCTL_2PASS_IN (26)
+/**Sets the current encoding quality.
+ * This is only valid so long as no bitrate has been specified, either through
+ *  the #th_info struct used to initialize the encoder or through
+ *  #TH_ENCCTL_SET_BITRATE (this restriction may be relaxed in a future
+ *  version).
+ * If it is set before the headers are emitted, the target quality encoded in
+ *  them will be updated.
+ *
+ * \param[in] _buf <tt>int</tt>: The new target quality, in the range 0...63,
+ *                  inclusive.
+ * \retval 0             Success.
+ * \retval TH_EFAULT     \a _enc or \a _buf is <tt>NULL</tt>.
+ * \retval TH_EINVAL     A target bitrate has already been specified, or the
+ *                        quality index was not in the range 0...63.
+ * \retval TH_EIMPL       Not supported by this implementation.*/
+#define TH_ENCCTL_SET_QUALITY (28)
+/**Sets the current encoding bitrate.
+ * Once a bitrate is set, the encoder must use a rate-controlled mode for all
+ *  future frames (this restriction may be relaxed in a future version).
+ * If it is set before the headers are emitted, the target bitrate encoded in
+ *  them will be updated.
+ * Due to the buffer delay, the exact bitrate of each section of the encode is
+ *  not guaranteed.
+ * The encoder may have already used more bits than allowed for the frames it
+ *  has encoded, expecting to make them up in future frames, or it may have
+ *  used fewer, holding the excess in reserve.
+ * The exact transition between the two bitrates is not well-defined by this
+ *  API, but may be affected by flags set with #TH_ENCCTL_SET_RATE_FLAGS.
+ * After a number of frames equal to the buffer delay, one may expect further
+ *  output to average at the target bitrate.
+ *
+ * \param[in] _buf <tt>long</tt>: The new target bitrate, in bits per second.
+ * \retval 0             Success.
+ * \retval TH_EFAULT     \a _enc or \a _buf is <tt>NULL</tt>.
+ * \retval TH_EINVAL     The target bitrate was not positive.
+ *                       A future version of this library may allow passing 0
+ *                        to disabled rate-controlled mode and return to a
+ *                        quality-based mode, in which case this function will
+ *                        not return an error for that value.
+ * \retval TH_EIMPL      Not supported by this implementation.*/
+#define TH_ENCCTL_SET_BITRATE (30)
+/**Sets the configuration to be compatible with that from the given setup
+ *  header.
+ * This sets the Huffman codebooks and quantization parameters to match those
+ *  found in the given setup header.
+ * This guarantees that packets encoded by this encoder will be decodable using
+ *  a decoder configured with the passed-in setup header.
+ * It does <em>not</em> guarantee that th_encode_flushheader() will produce a
+ *  bit-identical setup header, only that they will be compatible.
+ * If you need a bit-identical setup header, then use the one you passed into
+ *  this command, and not the one returned by th_encode_flushheader().
+ *
+ * This also does <em>not</em> enable or disable VP3 compatibility; that is not
+ *  signaled in the setup header (or anywhere else in the encoded stream), and
+ *  is controlled independently by the #TH_ENCCTL_SET_VP3_COMPATIBLE function.
+ * If you wish to enable VP3 compatibility mode <em>and</em> want the codebooks
+ *  and quantization parameters to match the given setup header, you should
+ *  enable VP3 compatibility before invoking this command, otherwise the
+ *  codebooks and quantization parameters will be reset to the VP3 defaults.
+ *
+ * The current encoder does not support Huffman codebooks which do not contain
+ *  codewords for all 32 tokens.
+ * Such codebooks are legal, according to the specification, but cannot be
+ *  configured with this function.
+ *
+ * \param[in] _buf <tt>unsigned char[]</tt>: The encoded setup header to copy
+ *                                            the configuration from.
+ *                                           This should be the original,
+ *                                            undecoded setup header packet,
+ *                                            and <em>not</em> a #th_setup_info
+ *                                            structure filled in by
+ *                                            th_decode_headerin().
+ * \retval TH_EFAULT     \a _enc or \a _buf is <tt>NULL</tt>.
+ * \retval TH_EINVAL     Encoding has already begun, so the codebooks and
+ *                        quantization parameters cannot be changed, or the
+ *                        data in the setup header was not supported by this
+ *                        encoder.
+ * \retval TH_EBADHEADER \a _buf did not contain a valid setup header packet.
+ * \retval TH_ENOTFORMAT \a _buf did not contain a Theora header at all.
+ * \retval TH_EIMPL   Not supported by this implementation.*/
+#define TH_ENCCTL_SET_COMPAT_CONFIG (32)
+
+/*@}*/
+
+
+/**\name TH_ENCCTL_SET_RATE_FLAGS flags
+ * \anchor ratectlflags
+ * These are the flags available for use with #TH_ENCCTL_SET_RATE_FLAGS.*/
+/*@{*/
+/**Drop frames to keep within bitrate buffer constraints.
+ * This can have a severe impact on quality, but is the only way to ensure that
+ *  bitrate targets are met at low rates during sudden bursts of activity.
+ * It is enabled by default.*/
+#define TH_RATECTL_DROP_FRAMES   (0x1)
+/**Ignore bitrate buffer overflows.
+ * If the encoder uses so few bits that the reservoir of available bits
+ *  overflows, ignore the excess.
+ * The encoder will not try to use these extra bits in future frames.
+ * At high rates this may cause the result to be undersized, but allows a
+ *  client to play the stream using a finite buffer; it should normally be
+ *  enabled, which is the default.*/
+#define TH_RATECTL_CAP_OVERFLOW  (0x2)
+/**Ignore bitrate buffer underflows.
+ * If the encoder uses so many bits that the reservoir of available bits
+ *  underflows, ignore the deficit.
+ * The encoder will not try to make up these extra bits in future frames.
+ * At low rates this may cause the result to be oversized; it should normally
+ *  be disabled, which is the default.*/
+#define TH_RATECTL_CAP_UNDERFLOW (0x4)
+/*@}*/
+
+
+
+/**The quantization parameters used by VP3.*/
+extern const th_quant_info TH_VP31_QUANT_INFO;
+
+/**The Huffman tables used by VP3.*/
+extern const th_huff_code
+ TH_VP31_HUFF_CODES[TH_NHUFFMAN_TABLES][TH_NDCT_TOKENS];
+
+
+
+/**\name Encoder state
+   The following data structure is opaque, and its contents are not publicly
+    defined by this API.
+   Referring to its internals directly is unsupported, and may break without
+    warning.*/
+/*@{*/
+/**The encoder context.*/
+typedef struct th_enc_ctx    th_enc_ctx;
+/*@}*/
+
+
+
+/**\defgroup encfuncs Functions for Encoding*/
+/*@{*/
+/**\name Functions for encoding
+ * You must link to <tt>libtheoraenc</tt> and <tt>libtheoradec</tt>
+ *  if you use any of the functions in this section.
+ *
+ * The functions are listed in the order they are used in a typical encode.
+ * The basic steps are:
+ * - Fill in a #th_info structure with details on the format of the video you
+ *    wish to encode.
+ * - Allocate a #th_enc_ctx handle with th_encode_alloc().
+ * - Perform any additional encoder configuration required with
+ *    th_encode_ctl().
+ * - Repeatedly call th_encode_flushheader() to retrieve all the header
+ *    packets.
+ * - For each uncompressed frame:
+ *   - Submit the uncompressed frame via th_encode_ycbcr_in()
+ *   - Repeatedly call th_encode_packetout() to retrieve any video
+ *      data packets that are ready.
+ * - Call th_encode_free() to release all encoder memory.*/
+/*@{*/
+/**Allocates an encoder instance.
+ * \param _info A #th_info struct filled with the desired encoding parameters.
+ * \return The initialized #th_enc_ctx handle.
+ * \retval NULL If the encoding parameters were invalid.*/
+extern th_enc_ctx *th_encode_alloc(const th_info *_info);
+/**Encoder control function.
+ * This is used to provide advanced control the encoding process.
+ * \param _enc    A #th_enc_ctx handle.
+ * \param _req    The control code to process.
+ *                See \ref encctlcodes "the list of available control codes"
+ *                 for details.
+ * \param _buf    The parameters for this control code.
+ * \param _buf_sz The size of the parameter buffer.
+ * \return Possible return values depend on the control code used.
+ *          See \ref encctlcodes "the list of control codes" for
+ *          specific values. Generally 0 indicates success.*/
+extern int th_encode_ctl(th_enc_ctx *_enc,int _req,void *_buf,size_t _buf_sz);
+/**Outputs the next header packet.
+ * This should be called repeatedly after encoder initialization until it
+ *  returns 0 in order to get all of the header packets, in order, before
+ *  encoding actual video data.
+ * \param _enc      A #th_enc_ctx handle.
+ * \param _comments The metadata to place in the comment header, when it is
+ *                   encoded.
+ * \param _op       An <tt>ogg_packet</tt> structure to fill.
+ *                  All of the elements of this structure will be set,
+ *                   including a pointer to the header data.
+ *                  The memory for the header data is owned by
+ *                   <tt>libtheoraenc</tt>, and may be invalidated when the
+ *                   next encoder function is called.
+ * \return A positive value indicates that a header packet was successfully
+ *          produced.
+ * \retval 0         No packet was produced, and no more header packets remain.
+ * \retval TH_EFAULT \a _enc, \a _comments, or \a _op was <tt>NULL</tt>.*/
+extern int th_encode_flushheader(th_enc_ctx *_enc,
+ th_comment *_comments,ogg_packet *_op);
+/**Submits an uncompressed frame to the encoder.
+ * \param _enc   A #th_enc_ctx handle.
+ * \param _ycbcr A buffer of Y'CbCr data to encode.
+ *               If the width and height of the buffer matches the frame size
+ *                the encoder was initialized with, the encoder will only
+ *                reference the portion inside the picture region.
+ *               Any data outside this region will be ignored, and need not map
+ *                to a valid address.
+ *               Alternatively, you can pass a buffer equal to the size of the
+ *                picture region, if this is less than the full frame size.
+ *               When using subsampled chroma planes, odd picture sizes or odd
+ *                picture offsets may require an unexpected chroma plane size,
+ *                and their use is generally discouraged, as they will not be
+ *                well-supported by players and other media frameworks.
+ *               See Section 4.4 of
+ *                <a href="https://www.theora.org/doc/Theora.pdf">the Theora
+ *                specification</a> for details if you wish to use them anyway.
+ * \retval 0         Success.
+ * \retval TH_EFAULT \a _enc or \a _ycbcr is <tt>NULL</tt>.
+ * \retval TH_EINVAL The buffer size matches neither the frame size nor the
+ *                    picture size the encoder was initialized with, or
+ *                    encoding has already completed.*/
+extern int th_encode_ycbcr_in(th_enc_ctx *_enc,th_ycbcr_buffer _ycbcr);
+/**Retrieves encoded video data packets.
+ * This should be called repeatedly after each frame is submitted to flush any
+ *  encoded packets, until it returns 0.
+ * The encoder will not buffer these packets as subsequent frames are
+ *  compressed, so a failure to do so will result in lost video data.
+ * \note Currently the encoder operates in a one-frame-in, one-packet-out
+ *        manner.
+ *       However, this may be changed in the future.
+ * \param _enc  A #th_enc_ctx handle.
+ * \param _last Set this flag to a non-zero value if no more uncompressed
+ *               frames will be submitted.
+ *              This ensures that a proper EOS flag is set on the last packet.
+ * \param _op   An <tt>ogg_packet</tt> structure to fill.
+ *              All of the elements of this structure will be set, including a
+ *               pointer to the video data.
+ *              The memory for the video data is owned by
+ *               <tt>libtheoraenc</tt>, and may be invalidated when the next
+ *               encoder function is called.
+ * \return A positive value indicates that a video data packet was successfully
+ *          produced.
+ * \retval 0         No packet was produced, and no more encoded video data
+ *                    remains.
+ * \retval TH_EFAULT \a _enc or \a _op was <tt>NULL</tt>.*/
+extern int th_encode_packetout(th_enc_ctx *_enc,int _last,ogg_packet *_op);
+/**Frees an allocated encoder instance.
+ * \param _enc A #th_enc_ctx handle.*/
+extern void th_encode_free(th_enc_ctx *_enc);
+/*@}*/
+/*@}*/
+
+
+
+#if defined(__cplusplus)
+}
+#endif
+
+#endif /* OGG_THEORA_THEORAENC_HEADER */

thirdparty/libtheora/x86/mmxencfrag.c

@@ -0,0 +1,903 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation https://www.xiph.org/                 *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+#include <stddef.h>
+#include "x86enc.h"
+
+#if defined(OC_X86_ASM)
+
+unsigned oc_enc_frag_sad_mmxext(const unsigned char *_src,
+ const unsigned char *_ref,int _ystride){
+  ptrdiff_t ystride3;
+  ptrdiff_t ret;
+  __asm__ __volatile__(
+    /*Load the first 4 rows of each block.*/
+    "movq (%[src]),%%mm0\n\t"
+    "movq (%[ref]),%%mm1\n\t"
+    "movq (%[src],%[ystride]),%%mm2\n\t"
+    "movq (%[ref],%[ystride]),%%mm3\n\t"
+    "lea (%[ystride],%[ystride],2),%[ystride3]\n\t"
+    "movq (%[src],%[ystride],2),%%mm4\n\t"
+    "movq (%[ref],%[ystride],2),%%mm5\n\t"
+    "movq (%[src],%[ystride3]),%%mm6\n\t"
+    "movq (%[ref],%[ystride3]),%%mm7\n\t"
+    /*Compute their SADs and add them in %%mm0*/
+    "psadbw %%mm1,%%mm0\n\t"
+    "psadbw %%mm3,%%mm2\n\t"
+    "lea (%[src],%[ystride],4),%[src]\n\t"
+    "paddw %%mm2,%%mm0\n\t"
+    "lea (%[ref],%[ystride],4),%[ref]\n\t"
+    /*Load the next 3 rows as registers become available.*/
+    "movq (%[src]),%%mm2\n\t"
+    "movq (%[ref]),%%mm3\n\t"
+    "psadbw %%mm5,%%mm4\n\t"
+    "psadbw %%mm7,%%mm6\n\t"
+    "paddw %%mm4,%%mm0\n\t"
+    "movq (%[ref],%[ystride]),%%mm5\n\t"
+    "movq (%[src],%[ystride]),%%mm4\n\t"
+    "paddw %%mm6,%%mm0\n\t"
+    "movq (%[ref],%[ystride],2),%%mm7\n\t"
+    "movq (%[src],%[ystride],2),%%mm6\n\t"
+    /*Start adding their SADs to %%mm0*/
+    "psadbw %%mm3,%%mm2\n\t"
+    "psadbw %%mm5,%%mm4\n\t"
+    "paddw %%mm2,%%mm0\n\t"
+    "psadbw %%mm7,%%mm6\n\t"
+    /*Load last row as registers become available.*/
+    "movq (%[src],%[ystride3]),%%mm2\n\t"
+    "movq (%[ref],%[ystride3]),%%mm3\n\t"
+    /*And finish adding up their SADs.*/
+    "paddw %%mm4,%%mm0\n\t"
+    "psadbw %%mm3,%%mm2\n\t"
+    "paddw %%mm6,%%mm0\n\t"
+    "paddw %%mm2,%%mm0\n\t"
+    "movd %%mm0,%[ret]\n\t"
+    :[ret]"=a"(ret),[src]"+r"(_src),[ref]"+r"(_ref),[ystride3]"=&r"(ystride3)
+    :[ystride]"r"((ptrdiff_t)_ystride)
+  );
+  return (unsigned)ret;
+}
+
+unsigned oc_enc_frag_sad_thresh_mmxext(const unsigned char *_src,
+ const unsigned char *_ref,int _ystride,unsigned _thresh){
+  /*Early termination is for suckers.*/
+  return oc_enc_frag_sad_mmxext(_src,_ref,_ystride);
+}
+
+/*Assumes the first two rows of %[ref1] and %[ref2] are in %%mm0...%%mm3, the
+   first two rows of %[src] are in %%mm4,%%mm5, and {1}x8 is in %%mm7.
+  We pre-load the next two rows of data as registers become available.*/
+#define OC_SAD2_LOOP \
+ "#OC_SAD2_LOOP\n\t" \
+ /*We want to compute (%%mm0+%%mm1>>1) on unsigned bytes without overflow, but \
+    pavgb computes (%%mm0+%%mm1+1>>1). \
+   The latter is exactly 1 too large when the low bit of two corresponding \
+    bytes is only set in one of them. \
+   Therefore we pxor the operands, pand to mask out the low bits, and psubb to \
+    correct the output of pavgb. \
+   TODO: This should be rewritten to compute ~pavgb(~a,~b) instead, which \
+    schedules better; currently, however, this function is unused.*/ \
+ "movq %%mm0,%%mm6\n\t" \
+ "lea (%[ref1],%[ystride],2),%[ref1]\n\t" \
+ "pxor %%mm1,%%mm0\n\t" \
+ "pavgb %%mm1,%%mm6\n\t" \
+ "lea (%[ref2],%[ystride],2),%[ref2]\n\t" \
+ "movq %%mm2,%%mm1\n\t" \
+ "pand %%mm7,%%mm0\n\t" \
+ "pavgb %%mm3,%%mm2\n\t" \
+ "pxor %%mm3,%%mm1\n\t" \
+ "movq (%[ref2],%[ystride]),%%mm3\n\t" \
+ "psubb %%mm0,%%mm6\n\t" \
+ "movq (%[ref1]),%%mm0\n\t" \
+ "pand %%mm7,%%mm1\n\t" \
+ "psadbw %%mm6,%%mm4\n\t" \
+ "movd %[ret],%%mm6\n\t" \
+ "psubb %%mm1,%%mm2\n\t" \
+ "movq (%[ref2]),%%mm1\n\t" \
+ "lea (%[src],%[ystride],2),%[src]\n\t" \
+ "psadbw %%mm2,%%mm5\n\t" \
+ "movq (%[ref1],%[ystride]),%%mm2\n\t" \
+ "paddw %%mm4,%%mm5\n\t" \
+ "movq (%[src]),%%mm4\n\t" \
+ "paddw %%mm5,%%mm6\n\t" \
+ "movq (%[src],%[ystride]),%%mm5\n\t" \
+ "movd %%mm6,%[ret]\n\t" \
+
+/*Same as above, but does not pre-load the next two rows.*/
+#define OC_SAD2_TAIL \
+ "#OC_SAD2_TAIL\n\t" \
+ "movq %%mm0,%%mm6\n\t" \
+ "pavgb %%mm1,%%mm0\n\t" \
+ "pxor %%mm1,%%mm6\n\t" \
+ "movq %%mm2,%%mm1\n\t" \
+ "pand %%mm7,%%mm6\n\t" \
+ "pavgb %%mm3,%%mm2\n\t" \
+ "pxor %%mm3,%%mm1\n\t" \
+ "psubb %%mm6,%%mm0\n\t" \
+ "pand %%mm7,%%mm1\n\t" \
+ "psadbw %%mm0,%%mm4\n\t" \
+ "psubb %%mm1,%%mm2\n\t" \
+ "movd %[ret],%%mm6\n\t" \
+ "psadbw %%mm2,%%mm5\n\t" \
+ "paddw %%mm4,%%mm5\n\t" \
+ "paddw %%mm5,%%mm6\n\t" \
+ "movd %%mm6,%[ret]\n\t" \
+
+unsigned oc_enc_frag_sad2_thresh_mmxext(const unsigned char *_src,
+ const unsigned char *_ref1,const unsigned char *_ref2,int _ystride,
+ unsigned _thresh){
+  ptrdiff_t ret;
+  __asm__ __volatile__(
+    "movq (%[ref1]),%%mm0\n\t"
+    "movq (%[ref2]),%%mm1\n\t"
+    "movq (%[ref1],%[ystride]),%%mm2\n\t"
+    "movq (%[ref2],%[ystride]),%%mm3\n\t"
+    "xor %[ret],%[ret]\n\t"
+    "movq (%[src]),%%mm4\n\t"
+    "pxor %%mm7,%%mm7\n\t"
+    "pcmpeqb %%mm6,%%mm6\n\t"
+    "movq (%[src],%[ystride]),%%mm5\n\t"
+    "psubb %%mm6,%%mm7\n\t"
+    OC_SAD2_LOOP
+    OC_SAD2_LOOP
+    OC_SAD2_LOOP
+    OC_SAD2_TAIL
+    :[ret]"=&a"(ret),[src]"+r"(_src),[ref1]"+r"(_ref1),[ref2]"+r"(_ref2)
+    :[ystride]"r"((ptrdiff_t)_ystride)
+  );
+  return (unsigned)ret;
+}
+
+/*Load an 8x4 array of pixel values from %[src] and %[ref] and compute their
+   16-bit difference in %%mm0...%%mm7.*/
+#define OC_LOAD_SUB_8x4(_off) \
+ "#OC_LOAD_SUB_8x4\n\t" \
+ "movd "#_off"(%[src]),%%mm0\n\t" \
+ "movd "#_off"(%[ref]),%%mm4\n\t" \
+ "movd "#_off"(%[src],%[src_ystride]),%%mm1\n\t" \
+ "lea (%[src],%[src_ystride],2),%[src]\n\t" \
+ "movd "#_off"(%[ref],%[ref_ystride]),%%mm5\n\t" \
+ "lea (%[ref],%[ref_ystride],2),%[ref]\n\t" \
+ "movd "#_off"(%[src]),%%mm2\n\t" \
+ "movd "#_off"(%[ref]),%%mm7\n\t" \
+ "movd "#_off"(%[src],%[src_ystride]),%%mm3\n\t" \
+ "movd "#_off"(%[ref],%[ref_ystride]),%%mm6\n\t" \
+ "punpcklbw %%mm4,%%mm0\n\t" \
+ "lea (%[src],%[src_ystride],2),%[src]\n\t" \
+ "punpcklbw %%mm4,%%mm4\n\t" \
+ "lea (%[ref],%[ref_ystride],2),%[ref]\n\t" \
+ "psubw %%mm4,%%mm0\n\t" \
+ "movd "#_off"(%[src]),%%mm4\n\t" \
+ "movq %%mm0,"OC_MEM_OFFS(_off*2,buf)"\n\t" \
+ "movd "#_off"(%[ref]),%%mm0\n\t" \
+ "punpcklbw %%mm5,%%mm1\n\t" \
+ "punpcklbw %%mm5,%%mm5\n\t" \
+ "psubw %%mm5,%%mm1\n\t" \
+ "movd "#_off"(%[src],%[src_ystride]),%%mm5\n\t" \
+ "punpcklbw %%mm7,%%mm2\n\t" \
+ "punpcklbw %%mm7,%%mm7\n\t" \
+ "psubw %%mm7,%%mm2\n\t" \
+ "movd "#_off"(%[ref],%[ref_ystride]),%%mm7\n\t" \
+ "punpcklbw %%mm6,%%mm3\n\t" \
+ "lea (%[src],%[src_ystride],2),%[src]\n\t" \
+ "punpcklbw %%mm6,%%mm6\n\t" \
+ "psubw %%mm6,%%mm3\n\t" \
+ "movd "#_off"(%[src]),%%mm6\n\t" \
+ "punpcklbw %%mm0,%%mm4\n\t" \
+ "lea (%[ref],%[ref_ystride],2),%[ref]\n\t" \
+ "punpcklbw %%mm0,%%mm0\n\t" \
+ "lea (%[src],%[src_ystride],2),%[src]\n\t" \
+ "psubw %%mm0,%%mm4\n\t" \
+ "movd "#_off"(%[ref]),%%mm0\n\t" \
+ "punpcklbw %%mm7,%%mm5\n\t" \
+ "neg %[src_ystride]\n\t" \
+ "punpcklbw %%mm7,%%mm7\n\t" \
+ "psubw %%mm7,%%mm5\n\t" \
+ "movd "#_off"(%[src],%[src_ystride]),%%mm7\n\t" \
+ "punpcklbw %%mm0,%%mm6\n\t" \
+ "lea (%[ref],%[ref_ystride],2),%[ref]\n\t" \
+ "punpcklbw %%mm0,%%mm0\n\t" \
+ "neg %[ref_ystride]\n\t" \
+ "psubw %%mm0,%%mm6\n\t" \
+ "movd "#_off"(%[ref],%[ref_ystride]),%%mm0\n\t" \
+ "lea (%[src],%[src_ystride],8),%[src]\n\t" \
+ "punpcklbw %%mm0,%%mm7\n\t" \
+ "neg %[src_ystride]\n\t" \
+ "punpcklbw %%mm0,%%mm0\n\t" \
+ "lea (%[ref],%[ref_ystride],8),%[ref]\n\t" \
+ "psubw %%mm0,%%mm7\n\t" \
+ "neg %[ref_ystride]\n\t" \
+ "movq "OC_MEM_OFFS(_off*2,buf)",%%mm0\n\t" \
+
+/*Load an 8x4 array of pixel values from %[src] into %%mm0...%%mm7.*/
+#define OC_LOAD_8x4(_off) \
+ "#OC_LOAD_8x4\n\t" \
+ "movd "#_off"(%[src]),%%mm0\n\t" \
+ "movd "#_off"(%[src],%[ystride]),%%mm1\n\t" \
+ "movd "#_off"(%[src],%[ystride],2),%%mm2\n\t" \
+ "pxor %%mm7,%%mm7\n\t" \
+ "movd "#_off"(%[src],%[ystride3]),%%mm3\n\t" \
+ "punpcklbw %%mm7,%%mm0\n\t" \
+ "movd "#_off"(%[src4]),%%mm4\n\t" \
+ "punpcklbw %%mm7,%%mm1\n\t" \
+ "movd "#_off"(%[src4],%[ystride]),%%mm5\n\t" \
+ "punpcklbw %%mm7,%%mm2\n\t" \
+ "movd "#_off"(%[src4],%[ystride],2),%%mm6\n\t" \
+ "punpcklbw %%mm7,%%mm3\n\t" \
+ "movd "#_off"(%[src4],%[ystride3]),%%mm7\n\t" \
+ "punpcklbw %%mm4,%%mm4\n\t" \
+ "punpcklbw %%mm5,%%mm5\n\t" \
+ "psrlw $8,%%mm4\n\t" \
+ "psrlw $8,%%mm5\n\t" \
+ "punpcklbw %%mm6,%%mm6\n\t" \
+ "punpcklbw %%mm7,%%mm7\n\t" \
+ "psrlw $8,%%mm6\n\t" \
+ "psrlw $8,%%mm7\n\t" \
+
+/*Performs the first two stages of an 8-point 1-D Hadamard transform.
+  The transform is performed in place, except that outputs 0-3 are swapped with
+   outputs 4-7.
+  Outputs 2, 3, 6, and 7 from the second stage are negated (which allows us to
+   perform this stage in place with no temporary registers).*/
+#define OC_HADAMARD_AB_8x4 \
+ "#OC_HADAMARD_AB_8x4\n\t" \
+ /*Stage A: \
+   Outputs 0-3 are swapped with 4-7 here.*/ \
+ "paddw %%mm1,%%mm5\n\t" \
+ "paddw %%mm2,%%mm6\n\t" \
+ "paddw %%mm1,%%mm1\n\t" \
+ "paddw %%mm2,%%mm2\n\t" \
+ "psubw %%mm5,%%mm1\n\t" \
+ "psubw %%mm6,%%mm2\n\t" \
+ "paddw %%mm3,%%mm7\n\t" \
+ "paddw %%mm0,%%mm4\n\t" \
+ "paddw %%mm3,%%mm3\n\t" \
+ "paddw %%mm0,%%mm0\n\t" \
+ "psubw %%mm7,%%mm3\n\t" \
+ "psubw %%mm4,%%mm0\n\t" \
+ /*Stage B:*/ \
+ "paddw %%mm2,%%mm0\n\t" \
+ "paddw %%mm3,%%mm1\n\t" \
+ "paddw %%mm6,%%mm4\n\t" \
+ "paddw %%mm7,%%mm5\n\t" \
+ "paddw %%mm2,%%mm2\n\t" \
+ "paddw %%mm3,%%mm3\n\t" \
+ "paddw %%mm6,%%mm6\n\t" \
+ "paddw %%mm7,%%mm7\n\t" \
+ "psubw %%mm0,%%mm2\n\t" \
+ "psubw %%mm1,%%mm3\n\t" \
+ "psubw %%mm4,%%mm6\n\t" \
+ "psubw %%mm5,%%mm7\n\t" \
+
+/*Performs the last stage of an 8-point 1-D Hadamard transform in place.
+  Outputs 1, 3, 5, and 7 are negated (which allows us to perform this stage in
+   place with no temporary registers).*/
+#define OC_HADAMARD_C_8x4 \
+ "#OC_HADAMARD_C_8x4\n\t" \
+ /*Stage C:*/ \
+ "paddw %%mm1,%%mm0\n\t" \
+ "paddw %%mm3,%%mm2\n\t" \
+ "paddw %%mm5,%%mm4\n\t" \
+ "paddw %%mm7,%%mm6\n\t" \
+ "paddw %%mm1,%%mm1\n\t" \
+ "paddw %%mm3,%%mm3\n\t" \
+ "paddw %%mm5,%%mm5\n\t" \
+ "paddw %%mm7,%%mm7\n\t" \
+ "psubw %%mm0,%%mm1\n\t" \
+ "psubw %%mm2,%%mm3\n\t" \
+ "psubw %%mm4,%%mm5\n\t" \
+ "psubw %%mm6,%%mm7\n\t" \
+
+/*Performs an 8-point 1-D Hadamard transform.
+  The transform is performed in place, except that outputs 0-3 are swapped with
+   outputs 4-7.
+  Outputs 1, 2, 5 and 6 are negated (which allows us to perform the transform
+   in place with no temporary registers).*/
+#define OC_HADAMARD_8x4 \
+ OC_HADAMARD_AB_8x4 \
+ OC_HADAMARD_C_8x4 \
+
+/*Performs the first part of the final stage of the Hadamard transform and
+   summing of absolute values.
+  At the end of this part, %%mm1 will contain the DC coefficient of the
+   transform.*/
+#define OC_HADAMARD_C_ABS_ACCUM_A_8x4(_r6,_r7) \
+ /*We use the fact that \
+     (abs(a+b)+abs(a-b))/2=max(abs(a),abs(b)) \
+    to merge the final butterfly with the abs and the first stage of \
+    accumulation. \
+   Thus we can avoid using pabsw, which is not available until SSSE3. \
+   Emulating pabsw takes 3 instructions, so the straightforward MMXEXT \
+    implementation would be (3+3)*8+7=55 instructions (+4 for spilling \
+    registers). \
+   Even with pabsw, it would be (3+1)*8+7=39 instructions (with no spills). \
+   This implementation is only 26 (+4 for spilling registers).*/ \
+ "#OC_HADAMARD_C_ABS_ACCUM_A_8x4\n\t" \
+ "movq %%mm7,"OC_MEM_OFFS(_r7,buf)"\n\t" \
+ "movq %%mm6,"OC_MEM_OFFS(_r6,buf)"\n\t" \
+ /*mm7={0x7FFF}x4 \
+   mm0=max(abs(mm0),abs(mm1))-0x7FFF*/ \
+ "pcmpeqb %%mm7,%%mm7\n\t" \
+ "movq %%mm0,%%mm6\n\t" \
+ "psrlw $1,%%mm7\n\t" \
+ "paddw %%mm1,%%mm6\n\t" \
+ "pmaxsw %%mm1,%%mm0\n\t" \
+ "paddsw %%mm7,%%mm6\n\t" \
+ "psubw %%mm6,%%mm0\n\t" \
+ /*mm2=max(abs(mm2),abs(mm3))-0x7FFF \
+   mm4=max(abs(mm4),abs(mm5))-0x7FFF*/ \
+ "movq %%mm2,%%mm6\n\t" \
+ "movq %%mm4,%%mm1\n\t" \
+ "pmaxsw %%mm3,%%mm2\n\t" \
+ "pmaxsw %%mm5,%%mm4\n\t" \
+ "paddw %%mm3,%%mm6\n\t" \
+ "paddw %%mm5,%%mm1\n\t" \
+ "movq "OC_MEM_OFFS(_r7,buf)",%%mm3\n\t" \
+
+/*Performs the second part of the final stage of the Hadamard transform and
+   summing of absolute values.*/
+#define OC_HADAMARD_C_ABS_ACCUM_B_8x4(_r6,_r7) \
+ "#OC_HADAMARD_C_ABS_ACCUM_B_8x4\n\t" \
+ "paddsw %%mm7,%%mm6\n\t" \
+ "movq "OC_MEM_OFFS(_r6,buf)",%%mm5\n\t" \
+ "paddsw %%mm7,%%mm1\n\t" \
+ "psubw %%mm6,%%mm2\n\t" \
+ "psubw %%mm1,%%mm4\n\t" \
+ /*mm7={1}x4 (needed for the horizontal add that follows) \
+   mm0+=mm2+mm4+max(abs(mm3),abs(mm5))-0x7FFF*/ \
+ "movq %%mm3,%%mm6\n\t" \
+ "pmaxsw %%mm5,%%mm3\n\t" \
+ "paddw %%mm2,%%mm0\n\t" \
+ "paddw %%mm5,%%mm6\n\t" \
+ "paddw %%mm4,%%mm0\n\t" \
+ "paddsw %%mm7,%%mm6\n\t" \
+ "paddw %%mm3,%%mm0\n\t" \
+ "psrlw $14,%%mm7\n\t" \
+ "psubw %%mm6,%%mm0\n\t" \
+
+/*Performs the last stage of an 8-point 1-D Hadamard transform, takes the
+   absolute value of each component, and accumulates everything into mm0.
+  This is the only portion of SATD which requires MMXEXT (we could use plain
+   MMX, but it takes 4 instructions and an extra register to work around the
+   lack of a pmaxsw, which is a pretty serious penalty).*/
+#define OC_HADAMARD_C_ABS_ACCUM_8x4(_r6,_r7) \
+ OC_HADAMARD_C_ABS_ACCUM_A_8x4(_r6,_r7) \
+ OC_HADAMARD_C_ABS_ACCUM_B_8x4(_r6,_r7) \
+
+/*Performs an 8-point 1-D Hadamard transform, takes the absolute value of each
+   component, and accumulates everything into mm0.
+  Note that mm0 will have an extra 4 added to each column, and that after
+   removing this value, the remainder will be half the conventional value.*/
+#define OC_HADAMARD_ABS_ACCUM_8x4(_r6,_r7) \
+ OC_HADAMARD_AB_8x4 \
+ OC_HADAMARD_C_ABS_ACCUM_8x4(_r6,_r7)
+
+/*Performs two 4x4 transposes (mostly) in place.
+  On input, {mm0,mm1,mm2,mm3} contains rows {e,f,g,h}, and {mm4,mm5,mm6,mm7}
+   contains rows {a,b,c,d}.
+  On output, {0x40,0x50,0x60,0x70}+_off(%[buf]) contains {e,f,g,h}^T, and
+   {mm4,mm5,mm6,mm7} contains the transposed rows {a,b,c,d}^T.*/
+#define OC_TRANSPOSE_4x4x2(_off) \
+ "#OC_TRANSPOSE_4x4x2\n\t" \
+ /*First 4x4 transpose:*/ \
+ "movq %%mm5,"OC_MEM_OFFS(0x10+(_off),buf)"\n\t" \
+ /*mm0 = e3 e2 e1 e0 \
+   mm1 = f3 f2 f1 f0 \
+   mm2 = g3 g2 g1 g0 \
+   mm3 = h3 h2 h1 h0*/ \
+ "movq %%mm2,%%mm5\n\t" \
+ "punpcklwd %%mm3,%%mm2\n\t" \
+ "punpckhwd %%mm3,%%mm5\n\t" \
+ "movq %%mm0,%%mm3\n\t" \
+ "punpcklwd %%mm1,%%mm0\n\t" \
+ "punpckhwd %%mm1,%%mm3\n\t" \
+ /*mm0 = f1 e1 f0 e0 \
+   mm3 = f3 e3 f2 e2 \
+   mm2 = h1 g1 h0 g0 \
+   mm5 = h3 g3 h2 g2*/ \
+ "movq %%mm0,%%mm1\n\t" \
+ "punpckldq %%mm2,%%mm0\n\t" \
+ "punpckhdq %%mm2,%%mm1\n\t" \
+ "movq %%mm3,%%mm2\n\t" \
+ "punpckhdq %%mm5,%%mm3\n\t" \
+ "movq %%mm0,"OC_MEM_OFFS(0x40+(_off),buf)"\n\t" \
+ "punpckldq %%mm5,%%mm2\n\t" \
+ /*mm0 = h0 g0 f0 e0 \
+   mm1 = h1 g1 f1 e1 \
+   mm2 = h2 g2 f2 e2 \
+   mm3 = h3 g3 f3 e3*/ \
+ "movq "OC_MEM_OFFS(0x10+(_off),buf)",%%mm5\n\t" \
+ /*Second 4x4 transpose:*/ \
+ /*mm4 = a3 a2 a1 a0 \
+   mm5 = b3 b2 b1 b0 \
+   mm6 = c3 c2 c1 c0 \
+   mm7 = d3 d2 d1 d0*/ \
+ "movq %%mm6,%%mm0\n\t" \
+ "punpcklwd %%mm7,%%mm6\n\t" \
+ "movq %%mm1,"OC_MEM_OFFS(0x50+(_off),buf)"\n\t" \
+ "punpckhwd %%mm7,%%mm0\n\t" \
+ "movq %%mm4,%%mm7\n\t" \
+ "punpcklwd %%mm5,%%mm4\n\t" \
+ "movq %%mm2,"OC_MEM_OFFS(0x60+(_off),buf)"\n\t" \
+ "punpckhwd %%mm5,%%mm7\n\t" \
+ /*mm4 = b1 a1 b0 a0 \
+   mm7 = b3 a3 b2 a2 \
+   mm6 = d1 c1 d0 c0 \
+   mm0 = d3 c3 d2 c2*/ \
+ "movq %%mm4,%%mm5\n\t" \
+ "punpckldq %%mm6,%%mm4\n\t" \
+ "movq %%mm3,"OC_MEM_OFFS(0x70+(_off),buf)"\n\t" \
+ "punpckhdq %%mm6,%%mm5\n\t" \
+ "movq %%mm7,%%mm6\n\t" \
+ "punpckhdq %%mm0,%%mm7\n\t" \
+ "punpckldq %%mm0,%%mm6\n\t" \
+ /*mm4 = d0 c0 b0 a0 \
+   mm5 = d1 c1 b1 a1 \
+   mm6 = d2 c2 b2 a2 \
+   mm7 = d3 c3 b3 a3*/ \
+
+static unsigned oc_int_frag_satd_mmxext(int *_dc,
+ const unsigned char *_src,int _src_ystride,
+ const unsigned char *_ref,int _ref_ystride){
+  OC_ALIGN8(ogg_int16_t buf[64]);
+  unsigned ret;
+  unsigned ret2;
+  int      dc;
+  __asm__ __volatile__(
+    OC_LOAD_SUB_8x4(0x00)
+    OC_HADAMARD_8x4
+    OC_TRANSPOSE_4x4x2(0x00)
+    /*Finish swapping out this 8x4 block to make room for the next one.
+      mm0...mm3 have been swapped out already.*/
+    "movq %%mm4,"OC_MEM_OFFS(0x00,buf)"\n\t"
+    "movq %%mm5,"OC_MEM_OFFS(0x10,buf)"\n\t"
+    "movq %%mm6,"OC_MEM_OFFS(0x20,buf)"\n\t"
+    "movq %%mm7,"OC_MEM_OFFS(0x30,buf)"\n\t"
+    OC_LOAD_SUB_8x4(0x04)
+    OC_HADAMARD_8x4
+    OC_TRANSPOSE_4x4x2(0x08)
+    /*Here the first 4x4 block of output from the last transpose is the second
+       4x4 block of input for the next transform.
+      We have cleverly arranged that it already be in the appropriate place, so
+       we only have to do half the loads.*/
+    "movq "OC_MEM_OFFS(0x10,buf)",%%mm1\n\t"
+    "movq "OC_MEM_OFFS(0x20,buf)",%%mm2\n\t"
+    "movq "OC_MEM_OFFS(0x30,buf)",%%mm3\n\t"
+    "movq "OC_MEM_OFFS(0x00,buf)",%%mm0\n\t"
+    /*We split out the stages here so we can save the DC coefficient in the
+       middle.*/
+    OC_HADAMARD_AB_8x4
+    OC_HADAMARD_C_ABS_ACCUM_A_8x4(0x28,0x38)
+    "movd %%mm1,%[dc]\n\t"
+    OC_HADAMARD_C_ABS_ACCUM_B_8x4(0x28,0x38)
+    /*Up to this point, everything fit in 16 bits (8 input + 1 for the
+       difference + 2*3 for the two 8-point 1-D Hadamards - 1 for the abs - 1
+       for the factor of two we dropped + 3 for the vertical accumulation).
+      Now we finally have to promote things to dwords.
+      We break this part out of OC_HADAMARD_ABS_ACCUM_8x4 to hide the long
+       latency of pmaddwd by starting the next series of loads now.*/
+    "pmaddwd %%mm7,%%mm0\n\t"
+    "movq "OC_MEM_OFFS(0x50,buf)",%%mm1\n\t"
+    "movq "OC_MEM_OFFS(0x58,buf)",%%mm5\n\t"
+    "movq %%mm0,%%mm4\n\t"
+    "movq "OC_MEM_OFFS(0x60,buf)",%%mm2\n\t"
+    "punpckhdq %%mm0,%%mm0\n\t"
+    "movq "OC_MEM_OFFS(0x68,buf)",%%mm6\n\t"
+    "paddd %%mm0,%%mm4\n\t"
+    "movq "OC_MEM_OFFS(0x70,buf)",%%mm3\n\t"
+    "movd %%mm4,%[ret2]\n\t"
+    "movq "OC_MEM_OFFS(0x78,buf)",%%mm7\n\t"
+    "movq "OC_MEM_OFFS(0x40,buf)",%%mm0\n\t"
+    "movq "OC_MEM_OFFS(0x48,buf)",%%mm4\n\t"
+    OC_HADAMARD_ABS_ACCUM_8x4(0x68,0x78)
+    "pmaddwd %%mm7,%%mm0\n\t"
+    /*Subtract abs(dc) from 2*ret2.*/
+    "movsx %w[dc],%[dc]\n\t"
+    "cdq\n\t"
+    "lea (%[ret],%[ret2],2),%[ret2]\n\t"
+    "movq %%mm0,%%mm4\n\t"
+    "punpckhdq %%mm0,%%mm0\n\t"
+    "xor %[dc],%[ret]\n\t"
+    "paddd %%mm0,%%mm4\n\t"
+    /*The sums produced by OC_HADAMARD_ABS_ACCUM_8x4 each have an extra 4
+       added to them, a factor of two removed, and the DC value included;
+       correct the final sum here.*/
+    "sub %[ret],%[ret2]\n\t"
+    "movd %%mm4,%[ret]\n\t"
+    "lea -64(%[ret2],%[ret],2),%[ret]\n\t"
+    /*Although it looks like we're using 8 registers here, gcc can alias %[ret]
+       and %[ret2] with some of the inputs, since for once we don't write to
+       them until after we're done using everything but %[buf].*/
+    /*Note that _src_ystride and _ref_ystride must be given non-overlapping
+       constraints, otherwise if gcc can prove they're equal it will allocate
+       them to the same register (which is bad); _src and _ref face a similar
+       problem, though those are never actually the same.*/
+    :[ret]"=d"(ret),[ret2]"=r"(ret2),[dc]"=a"(dc),
+     [buf]"=m"(OC_ARRAY_OPERAND(ogg_int16_t,buf,64))
+    :[src]"r"(_src),[src_ystride]"c"((ptrdiff_t)_src_ystride),
+     [ref]"r"(_ref),[ref_ystride]"d"((ptrdiff_t)_ref_ystride)
+    /*We have to use neg, so we actually clobber the condition codes for once
+       (not to mention cmp, sub, and add).*/
+    :"cc"
+  );
+  *_dc=dc;
+  return ret;
+}
+
+unsigned oc_enc_frag_satd_mmxext(int *_dc,const unsigned char *_src,
+ const unsigned char *_ref,int _ystride){
+  return oc_int_frag_satd_mmxext(_dc,_src,_ystride,_ref,_ystride);
+}
+
+/*Our internal implementation of frag_copy2 takes an extra stride parameter so
+   we can share code with oc_enc_frag_satd2_mmxext().*/
+void oc_int_frag_copy2_mmxext(unsigned char *_dst,int _dst_ystride,
+ const unsigned char *_src1,const unsigned char *_src2,int _src_ystride){
+  __asm__ __volatile__(
+    /*Load the first 3 rows.*/
+    "movq (%[src1]),%%mm0\n\t"
+    "movq (%[src2]),%%mm1\n\t"
+    "movq (%[src1],%[src_ystride]),%%mm2\n\t"
+    "lea (%[src1],%[src_ystride],2),%[src1]\n\t"
+    "movq (%[src2],%[src_ystride]),%%mm3\n\t"
+    "lea (%[src2],%[src_ystride],2),%[src2]\n\t"
+    "pxor %%mm7,%%mm7\n\t"
+    "movq (%[src1]),%%mm4\n\t"
+    "pcmpeqb %%mm6,%%mm6\n\t"
+    "movq (%[src2]),%%mm5\n\t"
+    /*mm7={1}x8.*/
+    "psubb %%mm6,%%mm7\n\t"
+    /*Start averaging %%mm0 and %%mm1 into %%mm6.*/
+    "movq %%mm0,%%mm6\n\t"
+    "pxor %%mm1,%%mm0\n\t"
+    "pavgb %%mm1,%%mm6\n\t"
+    /*%%mm1 is free, start averaging %%mm3 into %%mm2 using %%mm1.*/
+    "movq %%mm2,%%mm1\n\t"
+    "pand %%mm7,%%mm0\n\t"
+    "pavgb %%mm3,%%mm2\n\t"
+    "pxor %%mm3,%%mm1\n\t"
+    /*%%mm3 is free.*/
+    "psubb %%mm0,%%mm6\n\t"
+    /*%%mm0 is free, start loading the next row.*/
+    "movq (%[src1],%[src_ystride]),%%mm0\n\t"
+    /*Start averaging %%mm5 and %%mm4 using %%mm3.*/
+    "movq %%mm4,%%mm3\n\t"
+    /*%%mm6 (row 0) is done; write it out.*/
+    "movq %%mm6,(%[dst])\n\t"
+    "pand %%mm7,%%mm1\n\t"
+    "pavgb %%mm5,%%mm4\n\t"
+    "psubb %%mm1,%%mm2\n\t"
+    /*%%mm1 is free, continue loading the next row.*/
+    "movq (%[src2],%[src_ystride]),%%mm1\n\t"
+    "pxor %%mm5,%%mm3\n\t"
+    "lea (%[src1],%[src_ystride],2),%[src1]\n\t"
+    /*%%mm2 (row 1) is done; write it out.*/
+    "movq %%mm2,(%[dst],%[dst_ystride])\n\t"
+    "pand %%mm7,%%mm3\n\t"
+    /*Start loading the next row.*/
+    "movq (%[src1]),%%mm2\n\t"
+    "lea (%[dst],%[dst_ystride],2),%[dst]\n\t"
+    "psubb %%mm3,%%mm4\n\t"
+    "lea (%[src2],%[src_ystride],2),%[src2]\n\t"
+    /*%%mm4 (row 2) is done; write it out.*/
+    "movq %%mm4,(%[dst])\n\t"
+    /*Continue loading the next row.*/
+    "movq (%[src2]),%%mm3\n\t"
+    /*Start averaging %%mm0 and %%mm1 into %%mm6.*/
+    "movq %%mm0,%%mm6\n\t"
+    "pxor %%mm1,%%mm0\n\t"
+    /*Start loading the next row.*/
+    "movq (%[src1],%[src_ystride]),%%mm4\n\t"
+    "pavgb %%mm1,%%mm6\n\t"
+    /*%%mm1 is free; start averaging %%mm3 into %%mm2 using %%mm1.*/
+    "movq %%mm2,%%mm1\n\t"
+    "pand %%mm7,%%mm0\n\t"
+    /*Continue loading the next row.*/
+    "movq (%[src2],%[src_ystride]),%%mm5\n\t"
+    "pavgb %%mm3,%%mm2\n\t"
+    "lea (%[src1],%[src_ystride],2),%[src1]\n\t"
+    "pxor %%mm3,%%mm1\n\t"
+    /*%%mm3 is free.*/
+    "psubb %%mm0,%%mm6\n\t"
+    /*%%mm0 is free, start loading the next row.*/
+    "movq (%[src1]),%%mm0\n\t"
+    /*Start averaging %%mm5 into %%mm4 using %%mm3.*/
+    "movq %%mm4,%%mm3\n\t"
+    /*%%mm6 (row 3) is done; write it out.*/
+    "movq %%mm6,(%[dst],%[dst_ystride])\n\t"
+    "pand %%mm7,%%mm1\n\t"
+    "lea (%[src2],%[src_ystride],2),%[src2]\n\t"
+    "pavgb %%mm5,%%mm4\n\t"
+    "lea (%[dst],%[dst_ystride],2),%[dst]\n\t"
+    "psubb %%mm1,%%mm2\n\t"
+    /*%%mm1 is free; continue loading the next row.*/
+    "movq (%[src2]),%%mm1\n\t"
+    "pxor %%mm5,%%mm3\n\t"
+    /*%%mm2 (row 4) is done; write it out.*/
+    "movq %%mm2,(%[dst])\n\t"
+    "pand %%mm7,%%mm3\n\t"
+    /*Start loading the next row.*/
+    "movq (%[src1],%[src_ystride]),%%mm2\n\t"
+    "psubb %%mm3,%%mm4\n\t"
+    /*Start averaging %%mm0 and %%mm1 into %%mm6.*/
+    "movq %%mm0,%%mm6\n\t"
+    /*Continue loading the next row.*/
+    "movq (%[src2],%[src_ystride]),%%mm3\n\t"
+    /*%%mm4 (row 5) is done; write it out.*/
+    "movq %%mm4,(%[dst],%[dst_ystride])\n\t"
+    "pxor %%mm1,%%mm0\n\t"
+    "pavgb %%mm1,%%mm6\n\t"
+    /*%%mm4 is free; start averaging %%mm3 into %%mm2 using %%mm4.*/
+    "movq %%mm2,%%mm4\n\t"
+    "pand %%mm7,%%mm0\n\t"
+    "pavgb %%mm3,%%mm2\n\t"
+    "pxor %%mm3,%%mm4\n\t"
+    "lea (%[dst],%[dst_ystride],2),%[dst]\n\t"
+    "psubb %%mm0,%%mm6\n\t"
+    "pand %%mm7,%%mm4\n\t"
+    /*%%mm6 (row 6) is done, write it out.*/
+    "movq %%mm6,(%[dst])\n\t"
+    "psubb %%mm4,%%mm2\n\t"
+    /*%%mm2 (row 7) is done, write it out.*/
+    "movq %%mm2,(%[dst],%[dst_ystride])\n\t"
+    :[dst]"+r"(_dst),[src1]"+r"(_src1),[src2]"+r"(_src2)
+    :[dst_ystride]"r"((ptrdiff_t)_dst_ystride),
+     [src_ystride]"r"((ptrdiff_t)_src_ystride)
+    :"memory"
+  );
+}
+
+unsigned oc_enc_frag_satd2_mmxext(int *_dc,const unsigned char *_src,
+ const unsigned char *_ref1,const unsigned char *_ref2,int _ystride){
+  OC_ALIGN8(unsigned char ref[64]);
+  oc_int_frag_copy2_mmxext(ref,8,_ref1,_ref2,_ystride);
+  return oc_int_frag_satd_mmxext(_dc,_src,_ystride,ref,8);
+}
+
+unsigned oc_enc_frag_intra_satd_mmxext(int *_dc,
+ const unsigned char *_src,int _ystride){
+  OC_ALIGN8(ogg_int16_t buf[64]);
+  unsigned ret;
+  unsigned ret2;
+  int      dc;
+  __asm__ __volatile__(
+    OC_LOAD_8x4(0x00)
+    OC_HADAMARD_8x4
+    OC_TRANSPOSE_4x4x2(0x00)
+    /*Finish swapping out this 8x4 block to make room for the next one.
+      mm0...mm3 have been swapped out already.*/
+    "movq %%mm4,"OC_MEM_OFFS(0x00,buf)"\n\t"
+    "movq %%mm5,"OC_MEM_OFFS(0x10,buf)"\n\t"
+    "movq %%mm6,"OC_MEM_OFFS(0x20,buf)"\n\t"
+    "movq %%mm7,"OC_MEM_OFFS(0x30,buf)"\n\t"
+    OC_LOAD_8x4(0x04)
+    OC_HADAMARD_8x4
+    OC_TRANSPOSE_4x4x2(0x08)
+    /*Here the first 4x4 block of output from the last transpose is the second
+       4x4 block of input for the next transform.
+      We have cleverly arranged that it already be in the appropriate place, so
+       we only have to do half the loads.*/
+    "movq "OC_MEM_OFFS(0x10,buf)",%%mm1\n\t"
+    "movq "OC_MEM_OFFS(0x20,buf)",%%mm2\n\t"
+    "movq "OC_MEM_OFFS(0x30,buf)",%%mm3\n\t"
+    "movq "OC_MEM_OFFS(0x00,buf)",%%mm0\n\t"
+    /*We split out the stages here so we can save the DC coefficient in the
+       middle.*/
+    OC_HADAMARD_AB_8x4
+    OC_HADAMARD_C_ABS_ACCUM_A_8x4(0x28,0x38)
+    "movd %%mm1,%[dc]\n\t"
+    OC_HADAMARD_C_ABS_ACCUM_B_8x4(0x28,0x38)
+    /*Up to this point, everything fit in 16 bits (8 input + 1 for the
+       difference + 2*3 for the two 8-point 1-D Hadamards - 1 for the abs - 1
+       for the factor of two we dropped + 3 for the vertical accumulation).
+      Now we finally have to promote things to dwords.
+      We break this part out of OC_HADAMARD_ABS_ACCUM_8x4 to hide the long
+       latency of pmaddwd by starting the next series of loads now.*/
+    "pmaddwd %%mm7,%%mm0\n\t"
+    "movq "OC_MEM_OFFS(0x50,buf)",%%mm1\n\t"
+    "movq "OC_MEM_OFFS(0x58,buf)",%%mm5\n\t"
+    "movq "OC_MEM_OFFS(0x60,buf)",%%mm2\n\t"
+    "movq %%mm0,%%mm4\n\t"
+    "movq "OC_MEM_OFFS(0x68,buf)",%%mm6\n\t"
+    "punpckhdq %%mm0,%%mm0\n\t"
+    "movq "OC_MEM_OFFS(0x70,buf)",%%mm3\n\t"
+    "paddd %%mm0,%%mm4\n\t"
+    "movq "OC_MEM_OFFS(0x78,buf)",%%mm7\n\t"
+    "movd %%mm4,%[ret]\n\t"
+    "movq "OC_MEM_OFFS(0x40,buf)",%%mm0\n\t"
+    "movq "OC_MEM_OFFS(0x48,buf)",%%mm4\n\t"
+    OC_HADAMARD_ABS_ACCUM_8x4(0x68,0x78)
+    "pmaddwd %%mm7,%%mm0\n\t"
+    /*We assume that the DC coefficient is always positive (which is true,
+       because the input to the INTRA transform was not a difference).*/
+    "movzx %w[dc],%[dc]\n\t"
+    "add %[ret],%[ret]\n\t"
+    "sub %[dc],%[ret]\n\t"
+    "movq %%mm0,%%mm4\n\t"
+    "punpckhdq %%mm0,%%mm0\n\t"
+    "paddd %%mm0,%%mm4\n\t"
+    "movd %%mm4,%[ret2]\n\t"
+    "lea -64(%[ret],%[ret2],2),%[ret]\n\t"
+    /*Although it looks like we're using 8 registers here, gcc can alias %[ret]
+       and %[ret2] with some of the inputs, since for once we don't write to
+       them until after we're done using everything but %[buf] (which is also
+       listed as an output to ensure gcc _doesn't_ alias them against it).*/
+    :[ret]"=a"(ret),[ret2]"=r"(ret2),[dc]"=r"(dc),
+     [buf]"=m"(OC_ARRAY_OPERAND(ogg_int16_t,buf,64))
+    :[src]"r"(_src),[src4]"r"(_src+4*_ystride),
+     [ystride]"r"((ptrdiff_t)_ystride),[ystride3]"r"((ptrdiff_t)3*_ystride)
+    /*We have to use sub, so we actually clobber the condition codes for once
+       (not to mention add).*/
+    :"cc"
+  );
+  *_dc=dc;
+  return ret;
+}
+
+void oc_enc_frag_sub_mmx(ogg_int16_t _residue[64],
+ const unsigned char *_src,const unsigned char *_ref,int _ystride){
+  int i;
+  __asm__ __volatile__("pxor %%mm7,%%mm7\n\t"::);
+  for(i=4;i-->0;){
+    __asm__ __volatile__(
+      /*mm0=[src]*/
+      "movq (%[src]),%%mm0\n\t"
+      /*mm1=[ref]*/
+      "movq (%[ref]),%%mm1\n\t"
+      /*mm4=[src+ystride]*/
+      "movq (%[src],%[ystride]),%%mm4\n\t"
+      /*mm5=[ref+ystride]*/
+      "movq (%[ref],%[ystride]),%%mm5\n\t"
+      /*Compute [src]-[ref].*/
+      "movq %%mm0,%%mm2\n\t"
+      "punpcklbw %%mm7,%%mm0\n\t"
+      "movq %%mm1,%%mm3\n\t"
+      "punpckhbw %%mm7,%%mm2\n\t"
+      "punpcklbw %%mm7,%%mm1\n\t"
+      "punpckhbw %%mm7,%%mm3\n\t"
+      "psubw %%mm1,%%mm0\n\t"
+      "psubw %%mm3,%%mm2\n\t"
+      /*Compute [src+ystride]-[ref+ystride].*/
+      "movq %%mm4,%%mm1\n\t"
+      "punpcklbw %%mm7,%%mm4\n\t"
+      "movq %%mm5,%%mm3\n\t"
+      "punpckhbw %%mm7,%%mm1\n\t"
+      "lea (%[src],%[ystride],2),%[src]\n\t"
+      "punpcklbw %%mm7,%%mm5\n\t"
+      "lea (%[ref],%[ystride],2),%[ref]\n\t"
+      "punpckhbw %%mm7,%%mm3\n\t"
+      "psubw %%mm5,%%mm4\n\t"
+      "psubw %%mm3,%%mm1\n\t"
+      /*Write the answer out.*/
+      "movq %%mm0,0x00(%[residue])\n\t"
+      "movq %%mm2,0x08(%[residue])\n\t"
+      "movq %%mm4,0x10(%[residue])\n\t"
+      "movq %%mm1,0x18(%[residue])\n\t"
+      "lea 0x20(%[residue]),%[residue]\n\t"
+      :[residue]"+r"(_residue),[src]"+r"(_src),[ref]"+r"(_ref)
+      :[ystride]"r"((ptrdiff_t)_ystride)
+      :"memory"
+    );
+  }
+}
+
+void oc_enc_frag_sub_128_mmx(ogg_int16_t _residue[64],
+ const unsigned char *_src,int _ystride){
+  ptrdiff_t ystride3;
+  __asm__ __volatile__(
+    /*mm0=[src]*/
+    "movq (%[src]),%%mm0\n\t"
+    /*mm1=[src+ystride]*/
+    "movq (%[src],%[ystride]),%%mm1\n\t"
+    /*mm6={-1}x4*/
+    "pcmpeqw %%mm6,%%mm6\n\t"
+    /*mm2=[src+2*ystride]*/
+    "movq (%[src],%[ystride],2),%%mm2\n\t"
+    /*[ystride3]=3*[ystride]*/
+    "lea (%[ystride],%[ystride],2),%[ystride3]\n\t"
+    /*mm6={1}x4*/
+    "psllw $15,%%mm6\n\t"
+    /*mm3=[src+3*ystride]*/
+    "movq (%[src],%[ystride3]),%%mm3\n\t"
+    /*mm6={128}x4*/
+    "psrlw $8,%%mm6\n\t"
+    /*mm7=0*/
+    "pxor %%mm7,%%mm7\n\t"
+    /*[src]=[src]+4*[ystride]*/
+    "lea (%[src],%[ystride],4),%[src]\n\t"
+    /*Compute [src]-128 and [src+ystride]-128*/
+    "movq %%mm0,%%mm4\n\t"
+    "punpcklbw %%mm7,%%mm0\n\t"
+    "movq %%mm1,%%mm5\n\t"
+    "punpckhbw %%mm7,%%mm4\n\t"
+    "psubw %%mm6,%%mm0\n\t"
+    "punpcklbw %%mm7,%%mm1\n\t"
+    "psubw %%mm6,%%mm4\n\t"
+    "punpckhbw %%mm7,%%mm5\n\t"
+    "psubw %%mm6,%%mm1\n\t"
+    "psubw %%mm6,%%mm5\n\t"
+    /*Write the answer out.*/
+    "movq %%mm0,0x00(%[residue])\n\t"
+    "movq %%mm4,0x08(%[residue])\n\t"
+    "movq %%mm1,0x10(%[residue])\n\t"
+    "movq %%mm5,0x18(%[residue])\n\t"
+    /*mm0=[src+4*ystride]*/
+    "movq (%[src]),%%mm0\n\t"
+    /*mm1=[src+5*ystride]*/
+    "movq (%[src],%[ystride]),%%mm1\n\t"
+    /*Compute [src+2*ystride]-128 and [src+3*ystride]-128*/
+    "movq %%mm2,%%mm4\n\t"
+    "punpcklbw %%mm7,%%mm2\n\t"
+    "movq %%mm3,%%mm5\n\t"
+    "punpckhbw %%mm7,%%mm4\n\t"
+    "psubw %%mm6,%%mm2\n\t"
+    "punpcklbw %%mm7,%%mm3\n\t"
+    "psubw %%mm6,%%mm4\n\t"
+    "punpckhbw %%mm7,%%mm5\n\t"
+    "psubw %%mm6,%%mm3\n\t"
+    "psubw %%mm6,%%mm5\n\t"
+    /*Write the answer out.*/
+    "movq %%mm2,0x20(%[residue])\n\t"
+    "movq %%mm4,0x28(%[residue])\n\t"
+    "movq %%mm3,0x30(%[residue])\n\t"
+    "movq %%mm5,0x38(%[residue])\n\t"
+    /*mm2=[src+6*ystride]*/
+    "movq (%[src],%[ystride],2),%%mm2\n\t"
+    /*mm3=[src+7*ystride]*/
+    "movq (%[src],%[ystride3]),%%mm3\n\t"
+    /*Compute [src+4*ystride]-128 and [src+5*ystride]-128*/
+    "movq %%mm0,%%mm4\n\t"
+    "punpcklbw %%mm7,%%mm0\n\t"
+    "movq %%mm1,%%mm5\n\t"
+    "punpckhbw %%mm7,%%mm4\n\t"
+    "psubw %%mm6,%%mm0\n\t"
+    "punpcklbw %%mm7,%%mm1\n\t"
+    "psubw %%mm6,%%mm4\n\t"
+    "punpckhbw %%mm7,%%mm5\n\t"
+    "psubw %%mm6,%%mm1\n\t"
+    "psubw %%mm6,%%mm5\n\t"
+    /*Write the answer out.*/
+    "movq %%mm0,0x40(%[residue])\n\t"
+    "movq %%mm4,0x48(%[residue])\n\t"
+    "movq %%mm1,0x50(%[residue])\n\t"
+    "movq %%mm5,0x58(%[residue])\n\t"
+    /*Compute [src+6*ystride]-128 and [src+7*ystride]-128*/
+    "movq %%mm2,%%mm4\n\t"
+    "punpcklbw %%mm7,%%mm2\n\t"
+    "movq %%mm3,%%mm5\n\t"
+    "punpckhbw %%mm7,%%mm4\n\t"
+    "psubw %%mm6,%%mm2\n\t"
+    "punpcklbw %%mm7,%%mm3\n\t"
+    "psubw %%mm6,%%mm4\n\t"
+    "punpckhbw %%mm7,%%mm5\n\t"
+    "psubw %%mm6,%%mm3\n\t"
+    "psubw %%mm6,%%mm5\n\t"
+    /*Write the answer out.*/
+    "movq %%mm2,0x60(%[residue])\n\t"
+    "movq %%mm4,0x68(%[residue])\n\t"
+    "movq %%mm3,0x70(%[residue])\n\t"
+    "movq %%mm5,0x78(%[residue])\n\t"
+    :[src]"+r"(_src),[ystride3]"=&r"(ystride3)
+    :[residue]"r"(_residue),[ystride]"r"((ptrdiff_t)_ystride)
+    :"memory"
+  );
+}
+
+void oc_enc_frag_copy2_mmxext(unsigned char *_dst,
+ const unsigned char *_src1,const unsigned char *_src2,int _ystride){
+  oc_int_frag_copy2_mmxext(_dst,_ystride,_src1,_src2,_ystride);
+}
+
+#endif

thirdparty/libtheora/x86/mmxfdct.c

@@ -0,0 +1,678 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 1999-2006                *
+ * by the Xiph.Org Foundation https://www.xiph.org/                 *
+ *                                                                  *
+ ********************************************************************/
+/*MMX fDCT implementation for x86_32*/
+/*$Id: fdct_ses2.c 14579 2008-03-12 06:42:40Z xiphmont $*/
+#include "x86enc.h"
+#include "x86zigzag.h"
+
+#if defined(OC_X86_ASM)
+
+# define OC_FDCT_STAGE1_8x4 \
+ "#OC_FDCT_STAGE1_8x4\n\t" \
+ /*Stage 1:*/ \
+ /*mm0=t7'=t0-t7*/ \
+ "psubw %%mm7,%%mm0\n\t" \
+ "paddw %%mm7,%%mm7\n\t" \
+ /*mm1=t6'=t1-t6*/ \
+ "psubw %%mm6,%%mm1\n\t" \
+ "paddw %%mm6,%%mm6\n\t" \
+ /*mm2=t5'=t2-t5*/ \
+ "psubw %%mm5,%%mm2\n\t" \
+ "paddw %%mm5,%%mm5\n\t" \
+ /*mm3=t4'=t3-t4*/ \
+ "psubw %%mm4,%%mm3\n\t" \
+ "paddw %%mm4,%%mm4\n\t" \
+ /*mm7=t0'=t0+t7*/ \
+ "paddw %%mm0,%%mm7\n\t" \
+ /*mm6=t1'=t1+t6*/ \
+ "paddw %%mm1,%%mm6\n\t" \
+ /*mm5=t2'=t2+t5*/ \
+ "paddw %%mm2,%%mm5\n\t" \
+ /*mm4=t3'=t3+t4*/ \
+ "paddw %%mm3,%%mm4\n\t" \
+
+# define OC_FDCT8x4(_r0,_r1,_r2,_r3,_r4,_r5,_r6,_r7) \
+ "#OC_FDCT8x4\n\t" \
+ /*Stage 2:*/ \
+ /*mm7=t3''=t0'-t3'*/ \
+ "psubw %%mm4,%%mm7\n\t" \
+ "paddw %%mm4,%%mm4\n\t" \
+ /*mm6=t2''=t1'-t2'*/ \
+ "psubw %%mm5,%%mm6\n\t" \
+ "movq %%mm7,"_r6"(%[y])\n\t" \
+ "paddw %%mm5,%%mm5\n\t" \
+ /*mm1=t5''=t6'-t5'*/ \
+ "psubw %%mm2,%%mm1\n\t" \
+ "movq %%mm6,"_r2"(%[y])\n\t" \
+ /*mm4=t0''=t0'+t3'*/ \
+ "paddw %%mm7,%%mm4\n\t" \
+ "paddw %%mm2,%%mm2\n\t" \
+ /*mm5=t1''=t1'+t2'*/ \
+ "movq %%mm4,"_r0"(%[y])\n\t" \
+ "paddw %%mm6,%%mm5\n\t" \
+ /*mm2=t6''=t6'+t5'*/ \
+ "paddw %%mm1,%%mm2\n\t" \
+ "movq %%mm5,"_r4"(%[y])\n\t" \
+ /*mm0=t7', mm1=t5'', mm2=t6'', mm3=t4'.*/ \
+ /*mm4, mm5, mm6, mm7 are free.*/ \
+ /*Stage 3:*/ \
+ /*mm6={2}x4, mm7={27146,0xB500>>1}x2*/ \
+ "mov $0x5A806A0A,%[a]\n\t" \
+ "pcmpeqb %%mm6,%%mm6\n\t" \
+ "movd %[a],%%mm7\n\t" \
+ "psrlw $15,%%mm6\n\t" \
+ "punpckldq %%mm7,%%mm7\n\t" \
+ "paddw %%mm6,%%mm6\n\t" \
+ /*mm0=0, m2={-1}x4 \
+   mm5:mm4=t5''*27146+0xB500*/ \
+ "movq %%mm1,%%mm4\n\t" \
+ "movq %%mm1,%%mm5\n\t" \
+ "punpcklwd %%mm6,%%mm4\n\t" \
+ "movq %%mm2,"_r3"(%[y])\n\t" \
+ "pmaddwd %%mm7,%%mm4\n\t" \
+ "movq %%mm0,"_r7"(%[y])\n\t" \
+ "punpckhwd %%mm6,%%mm5\n\t" \
+ "pxor %%mm0,%%mm0\n\t" \
+ "pmaddwd %%mm7,%%mm5\n\t" \
+ "pcmpeqb %%mm2,%%mm2\n\t" \
+ /*mm2=t6'', mm1=t5''+(t5''!=0) \
+   mm4=(t5''*27146+0xB500>>16)*/ \
+ "pcmpeqw %%mm1,%%mm0\n\t" \
+ "psrad $16,%%mm4\n\t" \
+ "psubw %%mm2,%%mm0\n\t" \
+ "movq "_r3"(%[y]),%%mm2\n\t" \
+ "psrad $16,%%mm5\n\t" \
+ "paddw %%mm0,%%mm1\n\t" \
+ "packssdw %%mm5,%%mm4\n\t" \
+ /*mm4=s=(t5''*27146+0xB500>>16)+t5''+(t5''!=0)>>1*/ \
+ "paddw %%mm1,%%mm4\n\t" \
+ "movq "_r7"(%[y]),%%mm0\n\t" \
+ "psraw $1,%%mm4\n\t" \
+ "movq %%mm3,%%mm1\n\t" \
+ /*mm3=t4''=t4'+s*/ \
+ "paddw %%mm4,%%mm3\n\t" \
+ /*mm1=t5'''=t4'-s*/ \
+ "psubw %%mm4,%%mm1\n\t" \
+ /*mm1=0, mm3={-1}x4 \
+   mm5:mm4=t6''*27146+0xB500*/ \
+ "movq %%mm2,%%mm4\n\t" \
+ "movq %%mm2,%%mm5\n\t" \
+ "punpcklwd %%mm6,%%mm4\n\t" \
+ "movq %%mm1,"_r5"(%[y])\n\t" \
+ "pmaddwd %%mm7,%%mm4\n\t" \
+ "movq %%mm3,"_r1"(%[y])\n\t" \
+ "punpckhwd %%mm6,%%mm5\n\t" \
+ "pxor %%mm1,%%mm1\n\t" \
+ "pmaddwd %%mm7,%%mm5\n\t" \
+ "pcmpeqb %%mm3,%%mm3\n\t" \
+ /*mm2=t6''+(t6''!=0), mm4=(t6''*27146+0xB500>>16)*/ \
+ "psrad $16,%%mm4\n\t" \
+ "pcmpeqw %%mm2,%%mm1\n\t" \
+ "psrad $16,%%mm5\n\t" \
+ "psubw %%mm3,%%mm1\n\t" \
+ "packssdw %%mm5,%%mm4\n\t" \
+ "paddw %%mm1,%%mm2\n\t" \
+ /*mm1=t1'' \
+   mm4=s=(t6''*27146+0xB500>>16)+t6''+(t6''!=0)>>1*/ \
+ "paddw %%mm2,%%mm4\n\t" \
+ "movq "_r4"(%[y]),%%mm1\n\t" \
+ "psraw $1,%%mm4\n\t" \
+ "movq %%mm0,%%mm2\n\t" \
+ /*mm7={54491-0x7FFF,0x7FFF}x2 \
+   mm0=t7''=t7'+s*/ \
+ "paddw %%mm4,%%mm0\n\t" \
+ /*mm2=t6'''=t7'-s*/ \
+ "psubw %%mm4,%%mm2\n\t" \
+ /*Stage 4:*/ \
+ /*mm0=0, mm2=t0'' \
+   mm5:mm4=t1''*27146+0xB500*/ \
+ "movq %%mm1,%%mm4\n\t" \
+ "movq %%mm1,%%mm5\n\t" \
+ "punpcklwd %%mm6,%%mm4\n\t" \
+ "movq %%mm2,"_r3"(%[y])\n\t" \
+ "pmaddwd %%mm7,%%mm4\n\t" \
+ "movq "_r0"(%[y]),%%mm2\n\t" \
+ "punpckhwd %%mm6,%%mm5\n\t" \
+ "movq %%mm0,"_r7"(%[y])\n\t" \
+ "pmaddwd %%mm7,%%mm5\n\t" \
+ "pxor %%mm0,%%mm0\n\t" \
+ /*mm7={27146,0x4000>>1}x2 \
+   mm0=s=(t1''*27146+0xB500>>16)+t1''+(t1''!=0)*/ \
+ "psrad $16,%%mm4\n\t" \
+ "mov $0x20006A0A,%[a]\n\t" \
+ "pcmpeqw %%mm1,%%mm0\n\t" \
+ "movd %[a],%%mm7\n\t" \
+ "psrad $16,%%mm5\n\t" \
+ "psubw %%mm3,%%mm0\n\t" \
+ "packssdw %%mm5,%%mm4\n\t" \
+ "paddw %%mm1,%%mm0\n\t" \
+ "punpckldq %%mm7,%%mm7\n\t" \
+ "paddw %%mm4,%%mm0\n\t" \
+ /*mm6={0x00000E3D}x2 \
+   mm1=-(t0''==0), mm5:mm4=t0''*27146+0x4000*/ \
+ "movq %%mm2,%%mm4\n\t" \
+ "movq %%mm2,%%mm5\n\t" \
+ "punpcklwd %%mm6,%%mm4\n\t" \
+ "mov $0x0E3D,%[a]\n\t" \
+ "pmaddwd %%mm7,%%mm4\n\t" \
+ "punpckhwd %%mm6,%%mm5\n\t" \
+ "movd %[a],%%mm6\n\t" \
+ "pmaddwd %%mm7,%%mm5\n\t" \
+ "pxor %%mm1,%%mm1\n\t" \
+ "punpckldq %%mm6,%%mm6\n\t" \
+ "pcmpeqw %%mm2,%%mm1\n\t" \
+ /*mm4=r=(t0''*27146+0x4000>>16)+t0''+(t0''!=0)*/ \
+ "psrad $16,%%mm4\n\t" \
+ "psubw %%mm3,%%mm1\n\t" \
+ "psrad $16,%%mm5\n\t" \
+ "paddw %%mm1,%%mm2\n\t" \
+ "packssdw %%mm5,%%mm4\n\t" \
+ "movq "_r5"(%[y]),%%mm1\n\t" \
+ "paddw %%mm2,%%mm4\n\t" \
+ /*mm2=t6'', mm0=_y[0]=u=r+s>>1 \
+   The naive implementation could cause overflow, so we use \
+    u=(r&s)+((r^s)>>1).*/ \
+ "movq "_r3"(%[y]),%%mm2\n\t" \
+ "movq %%mm0,%%mm7\n\t" \
+ "pxor %%mm4,%%mm0\n\t" \
+ "pand %%mm4,%%mm7\n\t" \
+ "psraw $1,%%mm0\n\t" \
+ "mov $0x7FFF54DC,%[a]\n\t" \
+ "paddw %%mm7,%%mm0\n\t" \
+ "movd %[a],%%mm7\n\t" \
+ /*mm7={54491-0x7FFF,0x7FFF}x2 \
+   mm4=_y[4]=v=r-u*/ \
+ "psubw %%mm0,%%mm4\n\t" \
+ "punpckldq %%mm7,%%mm7\n\t" \
+ "movq %%mm4,"_r4"(%[y])\n\t" \
+ /*mm0=0, mm7={36410}x4 \
+   mm1=(t5'''!=0), mm5:mm4=54491*t5'''+0x0E3D*/ \
+ "movq %%mm1,%%mm4\n\t" \
+ "movq %%mm1,%%mm5\n\t" \
+ "punpcklwd %%mm1,%%mm4\n\t" \
+ "mov $0x8E3A8E3A,%[a]\n\t" \
+ "pmaddwd %%mm7,%%mm4\n\t" \
+ "movq %%mm0,"_r0"(%[y])\n\t" \
+ "punpckhwd %%mm1,%%mm5\n\t" \
+ "pxor %%mm0,%%mm0\n\t" \
+ "pmaddwd %%mm7,%%mm5\n\t" \
+ "pcmpeqw %%mm0,%%mm1\n\t" \
+ "movd %[a],%%mm7\n\t" \
+ "psubw %%mm3,%%mm1\n\t" \
+ "punpckldq %%mm7,%%mm7\n\t" \
+ "paddd %%mm6,%%mm4\n\t" \
+ "paddd %%mm6,%%mm5\n\t" \
+ /*mm0=0 \
+   mm3:mm1=36410*t6'''+((t5'''!=0)<<16)*/ \
+ "movq %%mm2,%%mm6\n\t" \
+ "movq %%mm2,%%mm3\n\t" \
+ "pmulhw %%mm7,%%mm6\n\t" \
+ "paddw %%mm2,%%mm1\n\t" \
+ "pmullw %%mm7,%%mm3\n\t" \
+ "pxor %%mm0,%%mm0\n\t" \
+ "paddw %%mm1,%%mm6\n\t" \
+ "movq %%mm3,%%mm1\n\t" \
+ "punpckhwd %%mm6,%%mm3\n\t" \
+ "punpcklwd %%mm6,%%mm1\n\t" \
+ /*mm3={-1}x4, mm6={1}x4 \
+   mm4=_y[5]=u=(54491*t5'''+36410*t6'''+0x0E3D>>16)+(t5'''!=0)*/ \
+ "paddd %%mm3,%%mm5\n\t" \
+ "paddd %%mm1,%%mm4\n\t" \
+ "psrad $16,%%mm5\n\t" \
+ "pxor %%mm6,%%mm6\n\t" \
+ "psrad $16,%%mm4\n\t" \
+ "pcmpeqb %%mm3,%%mm3\n\t" \
+ "packssdw %%mm5,%%mm4\n\t" \
+ "psubw %%mm3,%%mm6\n\t" \
+ /*mm1=t7'', mm7={26568,0x3400}x2 \
+   mm2=s=t6'''-(36410*u>>16)*/ \
+ "movq %%mm4,%%mm1\n\t" \
+ "mov $0x340067C8,%[a]\n\t" \
+ "pmulhw %%mm7,%%mm4\n\t" \
+ "movd %[a],%%mm7\n\t" \
+ "movq %%mm1,"_r5"(%[y])\n\t" \
+ "punpckldq %%mm7,%%mm7\n\t" \
+ "paddw %%mm1,%%mm4\n\t" \
+ "movq "_r7"(%[y]),%%mm1\n\t" \
+ "psubw %%mm4,%%mm2\n\t" \
+ /*mm6={0x00007B1B}x2 \
+   mm0=(s!=0), mm5:mm4=s*26568+0x3400*/ \
+ "movq %%mm2,%%mm4\n\t" \
+ "movq %%mm2,%%mm5\n\t" \
+ "punpcklwd %%mm6,%%mm4\n\t" \
+ "pcmpeqw %%mm2,%%mm0\n\t" \
+ "pmaddwd %%mm7,%%mm4\n\t" \
+ "mov $0x7B1B,%[a]\n\t" \
+ "punpckhwd %%mm6,%%mm5\n\t" \
+ "movd %[a],%%mm6\n\t" \
+ "pmaddwd %%mm7,%%mm5\n\t" \
+ "psubw %%mm3,%%mm0\n\t" \
+ "punpckldq %%mm6,%%mm6\n\t" \
+ /*mm7={64277-0x7FFF,0x7FFF}x2 \
+   mm2=_y[3]=v=(s*26568+0x3400>>17)+s+(s!=0)*/ \
+ "psrad $17,%%mm4\n\t" \
+ "paddw %%mm0,%%mm2\n\t" \
+ "psrad $17,%%mm5\n\t" \
+ "mov $0x7FFF7B16,%[a]\n\t" \
+ "packssdw %%mm5,%%mm4\n\t" \
+ "movd %[a],%%mm7\n\t" \
+ "paddw %%mm4,%%mm2\n\t" \
+ "punpckldq %%mm7,%%mm7\n\t" \
+ /*mm0=0, mm7={12785}x4 \
+   mm1=(t7''!=0), mm2=t4'', mm5:mm4=64277*t7''+0x7B1B*/ \
+ "movq %%mm1,%%mm4\n\t" \
+ "movq %%mm1,%%mm5\n\t" \
+ "movq %%mm2,"_r3"(%[y])\n\t" \
+ "punpcklwd %%mm1,%%mm4\n\t" \
+ "movq "_r1"(%[y]),%%mm2\n\t" \
+ "pmaddwd %%mm7,%%mm4\n\t" \
+ "mov $0x31F131F1,%[a]\n\t" \
+ "punpckhwd %%mm1,%%mm5\n\t" \
+ "pxor %%mm0,%%mm0\n\t" \
+ "pmaddwd %%mm7,%%mm5\n\t" \
+ "pcmpeqw %%mm0,%%mm1\n\t" \
+ "movd %[a],%%mm7\n\t" \
+ "psubw %%mm3,%%mm1\n\t" \
+ "punpckldq %%mm7,%%mm7\n\t" \
+ "paddd %%mm6,%%mm4\n\t" \
+ "paddd %%mm6,%%mm5\n\t" \
+ /*mm3:mm1=12785*t4'''+((t7''!=0)<<16)*/ \
+ "movq %%mm2,%%mm6\n\t" \
+ "movq %%mm2,%%mm3\n\t" \
+ "pmulhw %%mm7,%%mm6\n\t" \
+ "pmullw %%mm7,%%mm3\n\t" \
+ "paddw %%mm1,%%mm6\n\t" \
+ "movq %%mm3,%%mm1\n\t" \
+ "punpckhwd %%mm6,%%mm3\n\t" \
+ "punpcklwd %%mm6,%%mm1\n\t" \
+ /*mm3={-1}x4, mm6={1}x4 \
+   mm4=_y[1]=u=(12785*t4'''+64277*t7''+0x7B1B>>16)+(t7''!=0)*/ \
+ "paddd %%mm3,%%mm5\n\t" \
+ "paddd %%mm1,%%mm4\n\t" \
+ "psrad $16,%%mm5\n\t" \
+ "pxor %%mm6,%%mm6\n\t" \
+ "psrad $16,%%mm4\n\t" \
+ "pcmpeqb %%mm3,%%mm3\n\t" \
+ "packssdw %%mm5,%%mm4\n\t" \
+ "psubw %%mm3,%%mm6\n\t" \
+ /*mm1=t3'', mm7={20539,0x3000}x2 \
+   mm4=s=(12785*u>>16)-t4''*/ \
+ "movq %%mm4,"_r1"(%[y])\n\t" \
+ "pmulhw %%mm7,%%mm4\n\t" \
+ "mov $0x3000503B,%[a]\n\t" \
+ "movq "_r6"(%[y]),%%mm1\n\t" \
+ "movd %[a],%%mm7\n\t" \
+ "psubw %%mm2,%%mm4\n\t" \
+ "punpckldq %%mm7,%%mm7\n\t" \
+ /*mm6={0x00006CB7}x2 \
+   mm0=(s!=0), mm5:mm4=s*20539+0x3000*/ \
+ "movq %%mm4,%%mm5\n\t" \
+ "movq %%mm4,%%mm2\n\t" \
+ "punpcklwd %%mm6,%%mm4\n\t" \
+ "pcmpeqw %%mm2,%%mm0\n\t" \
+ "pmaddwd %%mm7,%%mm4\n\t" \
+ "mov $0x6CB7,%[a]\n\t" \
+ "punpckhwd %%mm6,%%mm5\n\t" \
+ "movd %[a],%%mm6\n\t" \
+ "pmaddwd %%mm7,%%mm5\n\t" \
+ "psubw %%mm3,%%mm0\n\t" \
+ "punpckldq %%mm6,%%mm6\n\t" \
+ /*mm7={60547-0x7FFF,0x7FFF}x2 \
+   mm2=_y[7]=v=(s*20539+0x3000>>20)+s+(s!=0)*/ \
+ "psrad $20,%%mm4\n\t" \
+ "paddw %%mm0,%%mm2\n\t" \
+ "psrad $20,%%mm5\n\t" \
+ "mov $0x7FFF6C84,%[a]\n\t" \
+ "packssdw %%mm5,%%mm4\n\t" \
+ "movd %[a],%%mm7\n\t" \
+ "paddw %%mm4,%%mm2\n\t" \
+ "punpckldq %%mm7,%%mm7\n\t" \
+ /*mm0=0, mm7={25080}x4 \
+   mm2=t2'', mm5:mm4=60547*t3''+0x6CB7*/ \
+ "movq %%mm1,%%mm4\n\t" \
+ "movq %%mm1,%%mm5\n\t" \
+ "movq %%mm2,"_r7"(%[y])\n\t" \
+ "punpcklwd %%mm1,%%mm4\n\t" \
+ "movq "_r2"(%[y]),%%mm2\n\t" \
+ "pmaddwd %%mm7,%%mm4\n\t" \
+ "mov $0x61F861F8,%[a]\n\t" \
+ "punpckhwd %%mm1,%%mm5\n\t" \
+ "pxor %%mm0,%%mm0\n\t" \
+ "pmaddwd %%mm7,%%mm5\n\t" \
+ "movd %[a],%%mm7\n\t" \
+ "pcmpeqw %%mm0,%%mm1\n\t" \
+ "psubw %%mm3,%%mm1\n\t" \
+ "punpckldq %%mm7,%%mm7\n\t" \
+ "paddd %%mm6,%%mm4\n\t" \
+ "paddd %%mm6,%%mm5\n\t" \
+ /*mm3:mm1=25080*t2''+((t3''!=0)<<16)*/ \
+ "movq %%mm2,%%mm6\n\t" \
+ "movq %%mm2,%%mm3\n\t" \
+ "pmulhw %%mm7,%%mm6\n\t" \
+ "pmullw %%mm7,%%mm3\n\t" \
+ "paddw %%mm1,%%mm6\n\t" \
+ "movq %%mm3,%%mm1\n\t" \
+ "punpckhwd %%mm6,%%mm3\n\t" \
+ "punpcklwd %%mm6,%%mm1\n\t" \
+ /*mm1={-1}x4 \
+   mm4=u=(25080*t2''+60547*t3''+0x6CB7>>16)+(t3''!=0)*/ \
+ "paddd %%mm3,%%mm5\n\t" \
+ "paddd %%mm1,%%mm4\n\t" \
+ "psrad $16,%%mm5\n\t" \
+ "mov $0x28005460,%[a]\n\t" \
+ "psrad $16,%%mm4\n\t" \
+ "pcmpeqb %%mm1,%%mm1\n\t" \
+ "packssdw %%mm5,%%mm4\n\t" \
+ /*mm5={1}x4, mm6=_y[2]=u, mm7={21600,0x2800}x2 \
+   mm4=s=(25080*u>>16)-t2''*/ \
+ "movq %%mm4,%%mm6\n\t" \
+ "pmulhw %%mm7,%%mm4\n\t" \
+ "pxor %%mm5,%%mm5\n\t" \
+ "movd %[a],%%mm7\n\t" \
+ "psubw %%mm1,%%mm5\n\t" \
+ "punpckldq %%mm7,%%mm7\n\t" \
+ "psubw %%mm2,%%mm4\n\t" \
+ /*mm2=s+(s!=0) \
+   mm4:mm3=s*21600+0x2800*/ \
+ "movq %%mm4,%%mm3\n\t" \
+ "movq %%mm4,%%mm2\n\t" \
+ "punpckhwd %%mm5,%%mm4\n\t" \
+ "pcmpeqw %%mm2,%%mm0\n\t" \
+ "pmaddwd %%mm7,%%mm4\n\t" \
+ "psubw %%mm1,%%mm0\n\t" \
+ "punpcklwd %%mm5,%%mm3\n\t" \
+ "paddw %%mm0,%%mm2\n\t" \
+ "pmaddwd %%mm7,%%mm3\n\t" \
+ /*mm0=_y[4], mm1=_y[7], mm4=_y[0], mm5=_y[5] \
+   mm3=_y[6]=v=(s*21600+0x2800>>18)+s+(s!=0)*/ \
+ "movq "_r4"(%[y]),%%mm0\n\t" \
+ "psrad $18,%%mm4\n\t" \
+ "movq "_r5"(%[y]),%%mm5\n\t" \
+ "psrad $18,%%mm3\n\t" \
+ "movq "_r7"(%[y]),%%mm1\n\t" \
+ "packssdw %%mm4,%%mm3\n\t" \
+ "movq "_r0"(%[y]),%%mm4\n\t" \
+ "paddw %%mm2,%%mm3\n\t" \
+
+/*On input, mm4=_y[0], mm6=_y[2], mm0=_y[4], mm5=_y[5], mm3=_y[6], mm1=_y[7].
+  On output, {_y[4],mm1,mm2,mm3} contains the transpose of _y[4...7] and
+   {mm4,mm5,mm6,mm7} contains the transpose of _y[0...3].*/
+# define OC_TRANSPOSE8x4(_r0,_r1,_r2,_r3,_r4,_r5,_r6,_r7) \
+ "#OC_TRANSPOSE8x4\n\t" \
+ /*First 4x4 transpose:*/ \
+ /*mm0 = e3 e2 e1 e0 \
+   mm5 = f3 f2 f1 f0 \
+   mm3 = g3 g2 g1 g0 \
+   mm1 = h3 h2 h1 h0*/ \
+ "movq %%mm0,%%mm2\n\t" \
+ "punpcklwd %%mm5,%%mm0\n\t" \
+ "punpckhwd %%mm5,%%mm2\n\t" \
+ "movq %%mm3,%%mm5\n\t" \
+ "punpcklwd %%mm1,%%mm3\n\t" \
+ "punpckhwd %%mm1,%%mm5\n\t" \
+ /*mm0 = f1 e1 f0 e0 \
+   mm2 = f3 e3 f2 e2 \
+   mm3 = h1 g1 h0 g0 \
+   mm5 = h3 g3 h2 g2*/ \
+ "movq %%mm0,%%mm1\n\t" \
+ "punpckldq %%mm3,%%mm0\n\t" \
+ "movq %%mm0,"_r4"(%[y])\n\t" \
+ "punpckhdq %%mm3,%%mm1\n\t" \
+ "movq "_r1"(%[y]),%%mm0\n\t" \
+ "movq %%mm2,%%mm3\n\t" \
+ "punpckldq %%mm5,%%mm2\n\t" \
+ "punpckhdq %%mm5,%%mm3\n\t" \
+ "movq "_r3"(%[y]),%%mm5\n\t" \
+ /*_y[4] = h0 g0 f0 e0 \
+    mm1  = h1 g1 f1 e1 \
+    mm2  = h2 g2 f2 e2 \
+    mm3  = h3 g3 f3 e3*/ \
+ /*Second 4x4 transpose:*/ \
+ /*mm4 = a3 a2 a1 a0 \
+   mm0 = b3 b2 b1 b0 \
+   mm6 = c3 c2 c1 c0 \
+   mm5 = d3 d2 d1 d0*/ \
+ "movq %%mm4,%%mm7\n\t" \
+ "punpcklwd %%mm0,%%mm4\n\t" \
+ "punpckhwd %%mm0,%%mm7\n\t" \
+ "movq %%mm6,%%mm0\n\t" \
+ "punpcklwd %%mm5,%%mm6\n\t" \
+ "punpckhwd %%mm5,%%mm0\n\t" \
+ /*mm4 = b1 a1 b0 a0 \
+   mm7 = b3 a3 b2 a2 \
+   mm6 = d1 c1 d0 c0 \
+   mm0 = d3 c3 d2 c2*/ \
+ "movq %%mm4,%%mm5\n\t" \
+ "punpckldq %%mm6,%%mm4\n\t" \
+ "punpckhdq %%mm6,%%mm5\n\t" \
+ "movq %%mm7,%%mm6\n\t" \
+ "punpckhdq %%mm0,%%mm7\n\t" \
+ "punpckldq %%mm0,%%mm6\n\t" \
+ /*mm4 = d0 c0 b0 a0 \
+   mm5 = d1 c1 b1 a1 \
+   mm6 = d2 c2 b2 a2 \
+   mm7 = d3 c3 b3 a3*/ \
+
+/*MMX implementation of the fDCT.*/
+void oc_enc_fdct8x8_mmxext(ogg_int16_t _y[64],const ogg_int16_t _x[64]){
+  OC_ALIGN8(ogg_int16_t buf[64]);
+  ptrdiff_t   a;
+  __asm__ __volatile__(
+    /*Add two extra bits of working precision to improve accuracy; any more and
+       we could overflow.*/
+    /*We also add biases to correct for some systematic error that remains in
+       the full fDCT->iDCT round trip.*/
+    "movq 0x00(%[x]),%%mm0\n\t"
+    "movq 0x10(%[x]),%%mm1\n\t"
+    "movq 0x20(%[x]),%%mm2\n\t"
+    "movq 0x30(%[x]),%%mm3\n\t"
+    "pcmpeqb %%mm4,%%mm4\n\t"
+    "pxor %%mm7,%%mm7\n\t"
+    "movq %%mm0,%%mm5\n\t"
+    "psllw $2,%%mm0\n\t"
+    "pcmpeqw %%mm7,%%mm5\n\t"
+    "movq 0x70(%[x]),%%mm7\n\t"
+    "psllw $2,%%mm1\n\t"
+    "psubw %%mm4,%%mm5\n\t"
+    "psllw $2,%%mm2\n\t"
+    "mov $1,%[a]\n\t"
+    "pslld $16,%%mm5\n\t"
+    "movd %[a],%%mm6\n\t"
+    "psllq $16,%%mm5\n\t"
+    "mov $0x10001,%[a]\n\t"
+    "psllw $2,%%mm3\n\t"
+    "movd %[a],%%mm4\n\t"
+    "punpckhwd %%mm6,%%mm5\n\t"
+    "psubw %%mm6,%%mm1\n\t"
+    "movq 0x60(%[x]),%%mm6\n\t"
+    "paddw %%mm5,%%mm0\n\t"
+    "movq 0x50(%[x]),%%mm5\n\t"
+    "paddw %%mm4,%%mm0\n\t"
+    "movq 0x40(%[x]),%%mm4\n\t"
+    /*We inline stage1 of the transform here so we can get better instruction
+       scheduling with the shifts.*/
+    /*mm0=t7'=t0-t7*/
+    "psllw $2,%%mm7\n\t"
+    "psubw %%mm7,%%mm0\n\t"
+    "psllw $2,%%mm6\n\t"
+    "paddw %%mm7,%%mm7\n\t"
+    /*mm1=t6'=t1-t6*/
+    "psllw $2,%%mm5\n\t"
+    "psubw %%mm6,%%mm1\n\t"
+    "psllw $2,%%mm4\n\t"
+    "paddw %%mm6,%%mm6\n\t"
+    /*mm2=t5'=t2-t5*/
+    "psubw %%mm5,%%mm2\n\t"
+    "paddw %%mm5,%%mm5\n\t"
+    /*mm3=t4'=t3-t4*/
+    "psubw %%mm4,%%mm3\n\t"
+    "paddw %%mm4,%%mm4\n\t"
+    /*mm7=t0'=t0+t7*/
+    "paddw %%mm0,%%mm7\n\t"
+    /*mm6=t1'=t1+t6*/
+    "paddw %%mm1,%%mm6\n\t"
+    /*mm5=t2'=t2+t5*/
+    "paddw %%mm2,%%mm5\n\t"
+    /*mm4=t3'=t3+t4*/
+    "paddw %%mm3,%%mm4\n\t"
+    OC_FDCT8x4("0x00","0x10","0x20","0x30","0x40","0x50","0x60","0x70")
+    OC_TRANSPOSE8x4("0x00","0x10","0x20","0x30","0x40","0x50","0x60","0x70")
+    /*Swap out this 8x4 block for the next one.*/
+    "movq 0x08(%[x]),%%mm0\n\t"
+    "movq %%mm7,0x30(%[y])\n\t"
+    "movq 0x78(%[x]),%%mm7\n\t"
+    "movq %%mm1,0x50(%[y])\n\t"
+    "movq 0x18(%[x]),%%mm1\n\t"
+    "movq %%mm6,0x20(%[y])\n\t"
+    "movq 0x68(%[x]),%%mm6\n\t"
+    "movq %%mm2,0x60(%[y])\n\t"
+    "movq 0x28(%[x]),%%mm2\n\t"
+    "movq %%mm5,0x10(%[y])\n\t"
+    "movq 0x58(%[x]),%%mm5\n\t"
+    "movq %%mm3,0x70(%[y])\n\t"
+    "movq 0x38(%[x]),%%mm3\n\t"
+    /*And increase its working precision, too.*/
+    "psllw $2,%%mm0\n\t"
+    "movq %%mm4,0x00(%[y])\n\t"
+    "psllw $2,%%mm7\n\t"
+    "movq 0x48(%[x]),%%mm4\n\t"
+    /*We inline stage1 of the transform here so we can get better instruction
+       scheduling with the shifts.*/
+    /*mm0=t7'=t0-t7*/
+    "psubw %%mm7,%%mm0\n\t"
+    "psllw $2,%%mm1\n\t"
+    "paddw %%mm7,%%mm7\n\t"
+    "psllw $2,%%mm6\n\t"
+    /*mm1=t6'=t1-t6*/
+    "psubw %%mm6,%%mm1\n\t"
+    "psllw $2,%%mm2\n\t"
+    "paddw %%mm6,%%mm6\n\t"
+    "psllw $2,%%mm5\n\t"
+    /*mm2=t5'=t2-t5*/
+    "psubw %%mm5,%%mm2\n\t"
+    "psllw $2,%%mm3\n\t"
+    "paddw %%mm5,%%mm5\n\t"
+    "psllw $2,%%mm4\n\t"
+    /*mm3=t4'=t3-t4*/
+    "psubw %%mm4,%%mm3\n\t"
+    "paddw %%mm4,%%mm4\n\t"
+    /*mm7=t0'=t0+t7*/
+    "paddw %%mm0,%%mm7\n\t"
+    /*mm6=t1'=t1+t6*/
+    "paddw %%mm1,%%mm6\n\t"
+    /*mm5=t2'=t2+t5*/
+    "paddw %%mm2,%%mm5\n\t"
+    /*mm4=t3'=t3+t4*/
+    "paddw %%mm3,%%mm4\n\t"
+    OC_FDCT8x4("0x08","0x18","0x28","0x38","0x48","0x58","0x68","0x78")
+    OC_TRANSPOSE8x4("0x08","0x18","0x28","0x38","0x48","0x58","0x68","0x78")
+    /*Here the first 4x4 block of output from the last transpose is the second
+       4x4 block of input for the next transform.
+      We have cleverly arranged that it already be in the appropriate place,
+       so we only have to do half the stores and loads.*/
+    "movq 0x00(%[y]),%%mm0\n\t"
+    "movq %%mm1,0x58(%[y])\n\t"
+    "movq 0x10(%[y]),%%mm1\n\t"
+    "movq %%mm2,0x68(%[y])\n\t"
+    "movq 0x20(%[y]),%%mm2\n\t"
+    "movq %%mm3,0x78(%[y])\n\t"
+    "movq 0x30(%[y]),%%mm3\n\t"
+    OC_FDCT_STAGE1_8x4
+    OC_FDCT8x4("0x00","0x10","0x20","0x30","0x08","0x18","0x28","0x38")
+    /*mm2={-2}x4*/
+    "pcmpeqw %%mm2,%%mm2\n\t"
+    "paddw %%mm2,%%mm2\n\t"
+    /*Round and store the results (no transpose).*/
+    "movq 0x10(%[y]),%%mm7\n\t"
+    "psubw %%mm2,%%mm4\n\t"
+    "psubw %%mm2,%%mm6\n\t"
+    "psraw $2,%%mm4\n\t"
+    "psubw %%mm2,%%mm0\n\t"
+    "movq %%mm4,"OC_MEM_OFFS(0x00,buf)"\n\t"
+    "movq 0x30(%[y]),%%mm4\n\t"
+    "psraw $2,%%mm6\n\t"
+    "psubw %%mm2,%%mm5\n\t"
+    "movq %%mm6,"OC_MEM_OFFS(0x20,buf)"\n\t"
+    "psraw $2,%%mm0\n\t"
+    "psubw %%mm2,%%mm3\n\t"
+    "movq %%mm0,"OC_MEM_OFFS(0x40,buf)"\n\t"
+    "psraw $2,%%mm5\n\t"
+    "psubw %%mm2,%%mm1\n\t"
+    "movq %%mm5,"OC_MEM_OFFS(0x50,buf)"\n\t"
+    "psraw $2,%%mm3\n\t"
+    "psubw %%mm2,%%mm7\n\t"
+    "movq %%mm3,"OC_MEM_OFFS(0x60,buf)"\n\t"
+    "psraw $2,%%mm1\n\t"
+    "psubw %%mm2,%%mm4\n\t"
+    "movq %%mm1,"OC_MEM_OFFS(0x70,buf)"\n\t"
+    "psraw $2,%%mm7\n\t"
+    "movq %%mm7,"OC_MEM_OFFS(0x10,buf)"\n\t"
+    "psraw $2,%%mm4\n\t"
+    "movq %%mm4,"OC_MEM_OFFS(0x30,buf)"\n\t"
+    /*Load the next block.*/
+    "movq 0x40(%[y]),%%mm0\n\t"
+    "movq 0x78(%[y]),%%mm7\n\t"
+    "movq 0x50(%[y]),%%mm1\n\t"
+    "movq 0x68(%[y]),%%mm6\n\t"
+    "movq 0x60(%[y]),%%mm2\n\t"
+    "movq 0x58(%[y]),%%mm5\n\t"
+    "movq 0x70(%[y]),%%mm3\n\t"
+    "movq 0x48(%[y]),%%mm4\n\t"
+    OC_FDCT_STAGE1_8x4
+    OC_FDCT8x4("0x40","0x50","0x60","0x70","0x48","0x58","0x68","0x78")
+    /*mm2={-2}x4*/
+    "pcmpeqw %%mm2,%%mm2\n\t"
+    "paddw %%mm2,%%mm2\n\t"
+    /*Round and store the results (no transpose).*/
+    "movq 0x50(%[y]),%%mm7\n\t"
+    "psubw %%mm2,%%mm4\n\t"
+    "psubw %%mm2,%%mm6\n\t"
+    "psraw $2,%%mm4\n\t"
+    "psubw %%mm2,%%mm0\n\t"
+    "movq %%mm4,"OC_MEM_OFFS(0x08,buf)"\n\t"
+    "movq 0x70(%[y]),%%mm4\n\t"
+    "psraw $2,%%mm6\n\t"
+    "psubw %%mm2,%%mm5\n\t"
+    "movq %%mm6,"OC_MEM_OFFS(0x28,buf)"\n\t"
+    "psraw $2,%%mm0\n\t"
+    "psubw %%mm2,%%mm3\n\t"
+    "movq %%mm0,"OC_MEM_OFFS(0x48,buf)"\n\t"
+    "psraw $2,%%mm5\n\t"
+    "psubw %%mm2,%%mm1\n\t"
+    "movq %%mm5,"OC_MEM_OFFS(0x58,buf)"\n\t"
+    "psraw $2,%%mm3\n\t"
+    "psubw %%mm2,%%mm7\n\t"
+    "movq %%mm3,"OC_MEM_OFFS(0x68,buf)"\n\t"
+    "psraw $2,%%mm1\n\t"
+    "psubw %%mm2,%%mm4\n\t"
+    "movq %%mm1,"OC_MEM_OFFS(0x78,buf)"\n\t"
+    "psraw $2,%%mm7\n\t"
+    "movq %%mm7,"OC_MEM_OFFS(0x18,buf)"\n\t"
+    "psraw $2,%%mm4\n\t"
+    "movq %%mm4,"OC_MEM_OFFS(0x38,buf)"\n\t"
+    /*Final transpose and zig-zag.*/
+#define OC_ZZ_LOAD_ROW_LO(_row,_reg) \
+    "movq "OC_MEM_OFFS(16*_row,buf)","_reg"\n\t" \
+
+#define OC_ZZ_LOAD_ROW_HI(_row,_reg) \
+    "movq "OC_MEM_OFFS(16*_row+8,buf)","_reg"\n\t" \
+
+    OC_TRANSPOSE_ZIG_ZAG_MMXEXT
+#undef OC_ZZ_LOAD_ROW_LO
+#undef OC_ZZ_LOAD_ROW_HI
+    :[a]"=&r"(a),[buf]"=m"(OC_ARRAY_OPERAND(ogg_int16_t,buf,64))
+    :[y]"r"(_y),[x]"r"(_x)
+    :"memory"
+  );
+}
+
+#endif

thirdparty/libtheora/x86/mmxfrag.c

@@ -0,0 +1,368 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009,2025           *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+/*MMX acceleration of fragment reconstruction for motion compensation.
+  Originally written by Rudolf Marek.
+  Additional optimization by Nils Pipenbrinck.
+  Note: Loops are unrolled for best performance.
+  The iteration each instruction belongs to is marked in the comments as #i.*/
+#include <stddef.h>
+#include "x86int.h"
+
+#if defined(OC_X86_ASM)
+
+/*Copies an 8x8 block of pixels from _src to _dst, assuming _ystride bytes
+   between rows.*/
+# define OC_FRAG_COPY_MMX(_dst,_src,_ystride) \
+  do{ \
+    const unsigned char *src; \
+    unsigned char       *dst; \
+    ptrdiff_t            ystride3; \
+    src=(_src); \
+    dst=(_dst); \
+    __asm__ __volatile__( \
+      /*src+0*ystride*/ \
+      "movq (%[src]),%%mm0\n\t" \
+      /*src+1*ystride*/ \
+      "movq (%[src],%[ystride]),%%mm1\n\t" \
+      /*ystride3=ystride*3*/ \
+      "lea (%[ystride],%[ystride],2),%[ystride3]\n\t" \
+      /*src+2*ystride*/ \
+      "movq (%[src],%[ystride],2),%%mm2\n\t" \
+      /*src+3*ystride*/ \
+      "movq (%[src],%[ystride3]),%%mm3\n\t" \
+      /*dst+0*ystride*/ \
+      "movq %%mm0,(%[dst])\n\t" \
+      /*dst+1*ystride*/ \
+      "movq %%mm1,(%[dst],%[ystride])\n\t" \
+      /*Pointer to next 4.*/ \
+      "lea (%[src],%[ystride],4),%[src]\n\t" \
+      /*dst+2*ystride*/ \
+      "movq %%mm2,(%[dst],%[ystride],2)\n\t" \
+      /*dst+3*ystride*/ \
+      "movq %%mm3,(%[dst],%[ystride3])\n\t" \
+      /*Pointer to next 4.*/ \
+      "lea (%[dst],%[ystride],4),%[dst]\n\t" \
+      /*src+0*ystride*/ \
+      "movq (%[src]),%%mm0\n\t" \
+      /*src+1*ystride*/ \
+      "movq (%[src],%[ystride]),%%mm1\n\t" \
+      /*src+2*ystride*/ \
+      "movq (%[src],%[ystride],2),%%mm2\n\t" \
+      /*src+3*ystride*/ \
+      "movq (%[src],%[ystride3]),%%mm3\n\t" \
+      /*dst+0*ystride*/ \
+      "movq %%mm0,(%[dst])\n\t" \
+      /*dst+1*ystride*/ \
+      "movq %%mm1,(%[dst],%[ystride])\n\t" \
+      /*dst+2*ystride*/ \
+      "movq %%mm2,(%[dst],%[ystride],2)\n\t" \
+      /*dst+3*ystride*/ \
+      "movq %%mm3,(%[dst],%[ystride3])\n\t" \
+      :[dst]"+r"(dst),[src]"+r"(src),[ystride3]"=&r"(ystride3) \
+      :[ystride]"r"((ptrdiff_t)(_ystride)) \
+      :"memory" \
+    ); \
+  } \
+  while(0)
+
+/*Copies an 8x8 block of pixels from _src to _dst, assuming _ystride bytes
+   between rows.*/
+void oc_frag_copy_mmx(unsigned char *_dst,
+ const unsigned char *_src,int _ystride){
+  OC_FRAG_COPY_MMX(_dst,_src,_ystride);
+}
+
+/*Copies the fragments specified by the lists of fragment indices from one
+   frame to another.
+  _dst_frame:     The reference frame to copy to.
+  _src_frame:     The reference frame to copy from.
+  _ystride:       The row stride of the reference frames.
+  _fragis:        A pointer to a list of fragment indices.
+  _nfragis:       The number of fragment indices to copy.
+  _frag_buf_offs: The offsets of fragments in the reference frames.*/
+void oc_frag_copy_list_mmx(unsigned char *_dst_frame,
+ const unsigned char *_src_frame,int _ystride,
+ const ptrdiff_t *_fragis,ptrdiff_t _nfragis,const ptrdiff_t *_frag_buf_offs){
+  ptrdiff_t fragii;
+  for(fragii=0;fragii<_nfragis;fragii++){
+    ptrdiff_t frag_buf_off;
+    frag_buf_off=_frag_buf_offs[_fragis[fragii]];
+    OC_FRAG_COPY_MMX(_dst_frame+frag_buf_off,
+     _src_frame+frag_buf_off,_ystride);
+  }
+}
+
+
+void oc_frag_recon_intra_mmx(unsigned char *_dst,int _ystride,
+ const ogg_int16_t *_residue){
+  __asm__ __volatile__(
+    /*Set mm0 to 0xFFFFFFFFFFFFFFFF.*/
+    "pcmpeqw %%mm0,%%mm0\n\t"
+    /*#0 Load low residue.*/
+    "movq 0*8(%[residue]),%%mm1\n\t"
+    /*#0 Load high residue.*/
+    "movq 1*8(%[residue]),%%mm2\n\t"
+    /*Set mm0 to 0x8000800080008000.*/
+    "psllw $15,%%mm0\n\t"
+    /*#1 Load low residue.*/
+    "movq 2*8(%[residue]),%%mm3\n\t"
+    /*#1 Load high residue.*/
+    "movq 3*8(%[residue]),%%mm4\n\t"
+    /*Set mm0 to 0x0080008000800080.*/
+    "psrlw $8,%%mm0\n\t"
+    /*#2 Load low residue.*/
+    "movq 4*8(%[residue]),%%mm5\n\t"
+    /*#2 Load high residue.*/
+    "movq 5*8(%[residue]),%%mm6\n\t"
+    /*#0 Bias low  residue.*/
+    "paddsw %%mm0,%%mm1\n\t"
+    /*#0 Bias high residue.*/
+    "paddsw %%mm0,%%mm2\n\t"
+    /*#0 Pack to byte.*/
+    "packuswb %%mm2,%%mm1\n\t"
+    /*#1 Bias low  residue.*/
+    "paddsw %%mm0,%%mm3\n\t"
+    /*#1 Bias high residue.*/
+    "paddsw %%mm0,%%mm4\n\t"
+    /*#1 Pack to byte.*/
+    "packuswb %%mm4,%%mm3\n\t"
+    /*#2 Bias low  residue.*/
+    "paddsw %%mm0,%%mm5\n\t"
+    /*#2 Bias high residue.*/
+    "paddsw %%mm0,%%mm6\n\t"
+    /*#2 Pack to byte.*/
+    "packuswb %%mm6,%%mm5\n\t"
+    /*#0 Write row.*/
+    "movq %%mm1,(%[dst])\n\t"
+    /*#1 Write row.*/
+    "movq %%mm3,(%[dst],%[ystride])\n\t"
+    /*#2 Write row.*/
+    "movq %%mm5,(%[dst],%[ystride],2)\n\t"
+    /*#3 Load low residue.*/
+    "movq 6*8(%[residue]),%%mm1\n\t"
+    /*#3 Load high residue.*/
+    "movq 7*8(%[residue]),%%mm2\n\t"
+    /*#4 Load high residue.*/
+    "movq 8*8(%[residue]),%%mm3\n\t"
+    /*#4 Load high residue.*/
+    "movq 9*8(%[residue]),%%mm4\n\t"
+    /*#5 Load high residue.*/
+    "movq 10*8(%[residue]),%%mm5\n\t"
+    /*#5 Load high residue.*/
+    "movq 11*8(%[residue]),%%mm6\n\t"
+    /*#3 Bias low  residue.*/
+    "paddsw %%mm0,%%mm1\n\t"
+    /*#3 Bias high residue.*/
+    "paddsw %%mm0,%%mm2\n\t"
+    /*#3 Pack to byte.*/
+    "packuswb %%mm2,%%mm1\n\t"
+    /*#4 Bias low  residue.*/
+    "paddsw %%mm0,%%mm3\n\t"
+    /*#4 Bias high residue.*/
+    "paddsw %%mm0,%%mm4\n\t"
+    /*#4 Pack to byte.*/
+    "packuswb %%mm4,%%mm3\n\t"
+    /*#5 Bias low  residue.*/
+    "paddsw %%mm0,%%mm5\n\t"
+    /*#5 Bias high residue.*/
+    "paddsw %%mm0,%%mm6\n\t"
+    /*#5 Pack to byte.*/
+    "packuswb %%mm6,%%mm5\n\t"
+    /*#3 Write row.*/
+    "movq %%mm1,(%[dst],%[ystride3])\n\t"
+    /*#4 Write row.*/
+    "movq %%mm3,(%[dst4])\n\t"
+    /*#5 Write row.*/
+    "movq %%mm5,(%[dst4],%[ystride])\n\t"
+    /*#6 Load low residue.*/
+    "movq 12*8(%[residue]),%%mm1\n\t"
+    /*#6 Load high residue.*/
+    "movq 13*8(%[residue]),%%mm2\n\t"
+    /*#7 Load low residue.*/
+    "movq 14*8(%[residue]),%%mm3\n\t"
+    /*#7 Load high residue.*/
+    "movq 15*8(%[residue]),%%mm4\n\t"
+    /*#6 Bias low  residue.*/
+    "paddsw %%mm0,%%mm1\n\t"
+    /*#6 Bias high residue.*/
+    "paddsw %%mm0,%%mm2\n\t"
+    /*#6 Pack to byte.*/
+    "packuswb %%mm2,%%mm1\n\t"
+    /*#7 Bias low  residue.*/
+    "paddsw %%mm0,%%mm3\n\t"
+    /*#7 Bias high residue.*/
+    "paddsw %%mm0,%%mm4\n\t"
+    /*#7 Pack to byte.*/
+    "packuswb %%mm4,%%mm3\n\t"
+    /*#6 Write row.*/
+    "movq %%mm1,(%[dst4],%[ystride],2)\n\t"
+    /*#7 Write row.*/
+    "movq %%mm3,(%[dst4],%[ystride3])\n\t"
+    :
+    :[residue]"r"(_residue),
+     [dst]"r"(_dst),
+     [dst4]"r"(_dst+(_ystride*4)),
+     [ystride]"r"((ptrdiff_t)_ystride),
+     [ystride3]"r"((ptrdiff_t)_ystride*3)
+    :"memory"
+  );
+}
+
+void oc_frag_recon_inter_mmx(unsigned char *_dst,const unsigned char *_src,
+ int _ystride,const ogg_int16_t *_residue){
+  int i;
+  /*Zero mm0.*/
+  __asm__ __volatile__("pxor %%mm0,%%mm0\n\t"::);
+  for(i=4;i-->0;){
+    __asm__ __volatile__(
+      /*#0 Load source.*/
+      "movq (%[src]),%%mm3\n\t"
+      /*#1 Load source.*/
+      "movq (%[src],%[ystride]),%%mm7\n\t"
+      /*#0 Get copy of src.*/
+      "movq %%mm3,%%mm4\n\t"
+      /*#0 Expand high source.*/
+      "punpckhbw %%mm0,%%mm4\n\t"
+      /*#0 Expand low  source.*/
+      "punpcklbw %%mm0,%%mm3\n\t"
+      /*#0 Add residue high.*/
+      "paddsw 8(%[residue]),%%mm4\n\t"
+      /*#1 Get copy of src.*/
+      "movq %%mm7,%%mm2\n\t"
+      /*#0 Add residue low.*/
+      "paddsw (%[residue]), %%mm3\n\t"
+      /*#1 Expand high source.*/
+      "punpckhbw %%mm0,%%mm2\n\t"
+      /*#0 Pack final row pixels.*/
+      "packuswb %%mm4,%%mm3\n\t"
+      /*#1 Expand low  source.*/
+      "punpcklbw %%mm0,%%mm7\n\t"
+      /*#1 Add residue low.*/
+      "paddsw 16(%[residue]),%%mm7\n\t"
+      /*#1 Add residue high.*/
+      "paddsw 24(%[residue]),%%mm2\n\t"
+      /*Advance residue.*/
+      "lea 32(%[residue]),%[residue]\n\t"
+      /*#1 Pack final row pixels.*/
+      "packuswb %%mm2,%%mm7\n\t"
+      /*Advance src.*/
+      "lea (%[src],%[ystride],2),%[src]\n\t"
+      /*#0 Write row.*/
+      "movq %%mm3,(%[dst])\n\t"
+      /*#1 Write row.*/
+      "movq %%mm7,(%[dst],%[ystride])\n\t"
+      /*Advance dst.*/
+      "lea (%[dst],%[ystride],2),%[dst]\n\t"
+      :[residue]"+r"(_residue),[dst]"+r"(_dst),[src]"+r"(_src)
+      :[ystride]"r"((ptrdiff_t)_ystride)
+      :"memory"
+    );
+  }
+}
+
+void oc_frag_recon_inter2_mmx(unsigned char *_dst,const unsigned char *_src1,
+ const unsigned char *_src2,int _ystride,const ogg_int16_t *_residue){
+  int i;
+  /*Zero mm7.*/
+  __asm__ __volatile__("pxor %%mm7,%%mm7\n\t"::);
+  for(i=4;i-->0;){
+    __asm__ __volatile__(
+      /*#0 Load src1.*/
+      "movq (%[src1]),%%mm0\n\t"
+      /*#0 Load src2.*/
+      "movq (%[src2]),%%mm2\n\t"
+      /*#0 Copy src1.*/
+      "movq %%mm0,%%mm1\n\t"
+      /*#0 Copy src2.*/
+      "movq %%mm2,%%mm3\n\t"
+      /*#1 Load src1.*/
+      "movq (%[src1],%[ystride]),%%mm4\n\t"
+      /*#0 Unpack lower src1.*/
+      "punpcklbw %%mm7,%%mm0\n\t"
+      /*#1 Load src2.*/
+      "movq (%[src2],%[ystride]),%%mm5\n\t"
+      /*#0 Unpack higher src1.*/
+      "punpckhbw %%mm7,%%mm1\n\t"
+      /*#0 Unpack lower src2.*/
+      "punpcklbw %%mm7,%%mm2\n\t"
+      /*#0 Unpack higher src2.*/
+      "punpckhbw %%mm7,%%mm3\n\t"
+      /*Advance src1 ptr.*/
+      "lea (%[src1],%[ystride],2),%[src1]\n\t"
+      /*Advance src2 ptr.*/
+      "lea (%[src2],%[ystride],2),%[src2]\n\t"
+      /*#0 Lower src1+src2.*/
+      "paddsw %%mm2,%%mm0\n\t"
+      /*#0 Higher src1+src2.*/
+      "paddsw %%mm3,%%mm1\n\t"
+      /*#1 Copy src1.*/
+      "movq %%mm4,%%mm2\n\t"
+      /*#0 Build lo average.*/
+      "psraw $1,%%mm0\n\t"
+      /*#1 Copy src2.*/
+      "movq %%mm5,%%mm3\n\t"
+      /*#1 Unpack lower src1.*/
+      "punpcklbw %%mm7,%%mm4\n\t"
+      /*#0 Build hi average.*/
+      "psraw $1,%%mm1\n\t"
+      /*#1 Unpack higher src1.*/
+      "punpckhbw %%mm7,%%mm2\n\t"
+      /*#0 low+=residue.*/
+      "paddsw (%[residue]),%%mm0\n\t"
+      /*#1 Unpack lower src2.*/
+      "punpcklbw %%mm7,%%mm5\n\t"
+      /*#0 high+=residue.*/
+      "paddsw 8(%[residue]),%%mm1\n\t"
+      /*#1 Unpack higher src2.*/
+      "punpckhbw %%mm7,%%mm3\n\t"
+      /*#1 Lower src1+src2.*/
+      "paddsw %%mm4,%%mm5\n\t"
+      /*#0 Pack and saturate.*/
+      "packuswb %%mm1,%%mm0\n\t"
+      /*#1 Higher src1+src2.*/
+      "paddsw %%mm2,%%mm3\n\t"
+      /*#0 Write row.*/
+      "movq %%mm0,(%[dst])\n\t"
+      /*#1 Build lo average.*/
+      "psraw $1,%%mm5\n\t"
+      /*#1 Build hi average.*/
+      "psraw $1,%%mm3\n\t"
+      /*#1 low+=residue.*/
+      "paddsw 16(%[residue]),%%mm5\n\t"
+      /*#1 high+=residue.*/
+      "paddsw 24(%[residue]),%%mm3\n\t"
+      /*#1 Pack and saturate.*/
+      "packuswb  %%mm3,%%mm5\n\t"
+      /*#1 Write row ptr.*/
+      "movq %%mm5,(%[dst],%[ystride])\n\t"
+      /*Advance residue ptr.*/
+      "add $32,%[residue]\n\t"
+      /*Advance dest ptr.*/
+      "lea (%[dst],%[ystride],2),%[dst]\n\t"
+     :[dst]"+r"(_dst),[residue]"+r"(_residue),
+      [src1]"+r"(_src1),[src2]"+r"(_src2)
+     :[ystride]"r"((ptrdiff_t)_ystride)
+     :"memory"
+    );
+  }
+}
+
+void oc_restore_fpu_mmx(void){
+  __asm__ __volatile__("emms\n\t");
+}
+#endif

thirdparty/libtheora/x86/mmxidct.c

@@ -0,0 +1,558 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009,2025           *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+/*MMX acceleration of Theora's iDCT.
+  Originally written by Rudolf Marek, based on code from On2's VP3.*/
+#include "x86int.h"
+#include "../dct.h"
+
+#if defined(OC_X86_ASM)
+
+/*These are offsets into the table of constants below.*/
+/*7 rows of cosines, in order: pi/16 * (1 ... 7).*/
+#define OC_COSINE_OFFSET (0)
+/*A row of 8's.*/
+#define OC_EIGHT_OFFSET  (56)
+
+
+
+/*38 cycles*/
+#define OC_IDCT_BEGIN(_y,_x) \
+  "#OC_IDCT_BEGIN\n\t" \
+  "movq "OC_I(3,_x)",%%mm2\n\t" \
+  "movq "OC_MEM_OFFS(0x30,c)",%%mm6\n\t" \
+  "movq %%mm2,%%mm4\n\t" \
+  "movq "OC_J(5,_x)",%%mm7\n\t" \
+  "pmulhw %%mm6,%%mm4\n\t" \
+  "movq "OC_MEM_OFFS(0x50,c)",%%mm1\n\t" \
+  "pmulhw %%mm7,%%mm6\n\t" \
+  "movq %%mm1,%%mm5\n\t" \
+  "pmulhw %%mm2,%%mm1\n\t" \
+  "movq "OC_I(1,_x)",%%mm3\n\t" \
+  "pmulhw %%mm7,%%mm5\n\t" \
+  "movq "OC_MEM_OFFS(0x10,c)",%%mm0\n\t" \
+  "paddw %%mm2,%%mm4\n\t" \
+  "paddw %%mm7,%%mm6\n\t" \
+  "paddw %%mm1,%%mm2\n\t" \
+  "movq "OC_J(7,_x)",%%mm1\n\t" \
+  "paddw %%mm5,%%mm7\n\t" \
+  "movq %%mm0,%%mm5\n\t" \
+  "pmulhw %%mm3,%%mm0\n\t" \
+  "paddw %%mm7,%%mm4\n\t" \
+  "pmulhw %%mm1,%%mm5\n\t" \
+  "movq "OC_MEM_OFFS(0x70,c)",%%mm7\n\t" \
+  "psubw %%mm2,%%mm6\n\t" \
+  "paddw %%mm3,%%mm0\n\t" \
+  "pmulhw %%mm7,%%mm3\n\t" \
+  "movq "OC_I(2,_x)",%%mm2\n\t" \
+  "pmulhw %%mm1,%%mm7\n\t" \
+  "paddw %%mm1,%%mm5\n\t" \
+  "movq %%mm2,%%mm1\n\t" \
+  "pmulhw "OC_MEM_OFFS(0x20,c)",%%mm2\n\t" \
+  "psubw %%mm5,%%mm3\n\t" \
+  "movq "OC_J(6,_x)",%%mm5\n\t" \
+  "paddw %%mm7,%%mm0\n\t" \
+  "movq %%mm5,%%mm7\n\t" \
+  "psubw %%mm4,%%mm0\n\t" \
+  "pmulhw "OC_MEM_OFFS(0x20,c)",%%mm5\n\t" \
+  "paddw %%mm1,%%mm2\n\t" \
+  "pmulhw "OC_MEM_OFFS(0x60,c)",%%mm1\n\t" \
+  "paddw %%mm4,%%mm4\n\t" \
+  "paddw %%mm0,%%mm4\n\t" \
+  "psubw %%mm6,%%mm3\n\t" \
+  "paddw %%mm7,%%mm5\n\t" \
+  "paddw %%mm6,%%mm6\n\t" \
+  "pmulhw "OC_MEM_OFFS(0x60,c)",%%mm7\n\t" \
+  "paddw %%mm3,%%mm6\n\t" \
+  "movq %%mm4,"OC_I(1,_y)"\n\t" \
+  "psubw %%mm5,%%mm1\n\t" \
+  "movq "OC_MEM_OFFS(0x40,c)",%%mm4\n\t" \
+  "movq %%mm3,%%mm5\n\t" \
+  "pmulhw %%mm4,%%mm3\n\t" \
+  "paddw %%mm2,%%mm7\n\t" \
+  "movq %%mm6,"OC_I(2,_y)"\n\t" \
+  "movq %%mm0,%%mm2\n\t" \
+  "movq "OC_I(0,_x)",%%mm6\n\t" \
+  "pmulhw %%mm4,%%mm0\n\t" \
+  "paddw %%mm3,%%mm5\n\t" \
+  "movq "OC_J(4,_x)",%%mm3\n\t" \
+  "psubw %%mm1,%%mm5\n\t" \
+  "paddw %%mm0,%%mm2\n\t" \
+  "psubw %%mm3,%%mm6\n\t" \
+  "movq %%mm6,%%mm0\n\t" \
+  "pmulhw %%mm4,%%mm6\n\t" \
+  "paddw %%mm3,%%mm3\n\t" \
+  "paddw %%mm1,%%mm1\n\t" \
+  "paddw %%mm0,%%mm3\n\t" \
+  "paddw %%mm5,%%mm1\n\t" \
+  "pmulhw %%mm3,%%mm4\n\t" \
+  "paddw %%mm0,%%mm6\n\t" \
+  "psubw %%mm2,%%mm6\n\t" \
+  "paddw %%mm2,%%mm2\n\t" \
+  "movq "OC_I(1,_y)",%%mm0\n\t" \
+  "paddw %%mm6,%%mm2\n\t" \
+  "paddw %%mm3,%%mm4\n\t" \
+  "psubw %%mm1,%%mm2\n\t" \
+  "#end OC_IDCT_BEGIN\n\t" \
+
+/*38+8=46 cycles.*/
+#define OC_ROW_IDCT(_y,_x) \
+  "#OC_ROW_IDCT\n" \
+  OC_IDCT_BEGIN(_y,_x) \
+  /*r3=D'*/ \
+  "movq "OC_I(2,_y)",%%mm3\n\t" \
+  /*r4=E'=E-G*/ \
+  "psubw %%mm7,%%mm4\n\t" \
+  /*r1=H'+H'*/ \
+  "paddw %%mm1,%%mm1\n\t" \
+  /*r7=G+G*/ \
+  "paddw %%mm7,%%mm7\n\t" \
+  /*r1=R1=A''+H'*/ \
+  "paddw %%mm2,%%mm1\n\t" \
+  /*r7=G'=E+G*/ \
+  "paddw %%mm4,%%mm7\n\t" \
+  /*r4=R4=E'-D'*/ \
+  "psubw %%mm3,%%mm4\n\t" \
+  "paddw %%mm3,%%mm3\n\t" \
+  /*r6=R6=F'-B''*/ \
+  "psubw %%mm5,%%mm6\n\t" \
+  "paddw %%mm5,%%mm5\n\t" \
+  /*r3=R3=E'+D'*/ \
+  "paddw %%mm4,%%mm3\n\t" \
+  /*r5=R5=F'+B''*/ \
+  "paddw %%mm6,%%mm5\n\t" \
+  /*r7=R7=G'-C'*/ \
+  "psubw %%mm0,%%mm7\n\t" \
+  "paddw %%mm0,%%mm0\n\t" \
+  /*Save R1.*/ \
+  "movq %%mm1,"OC_I(1,_y)"\n\t" \
+  /*r0=R0=G.+C.*/ \
+  "paddw %%mm7,%%mm0\n\t" \
+  "#end OC_ROW_IDCT\n\t" \
+
+/*The following macro does two 4x4 transposes in place.
+  At entry, we assume:
+    r0 = a3 a2 a1 a0
+  I(1) = b3 b2 b1 b0
+    r2 = c3 c2 c1 c0
+    r3 = d3 d2 d1 d0
+
+    r4 = e3 e2 e1 e0
+    r5 = f3 f2 f1 f0
+    r6 = g3 g2 g1 g0
+    r7 = h3 h2 h1 h0
+
+  At exit, we have:
+  I(0) = d0 c0 b0 a0
+  I(1) = d1 c1 b1 a1
+  I(2) = d2 c2 b2 a2
+  I(3) = d3 c3 b3 a3
+
+  J(4) = h0 g0 f0 e0
+  J(5) = h1 g1 f1 e1
+  J(6) = h2 g2 f2 e2
+  J(7) = h3 g3 f3 e3
+
+  I(0) I(1) I(2) I(3) is the transpose of r0 I(1) r2 r3.
+  J(4) J(5) J(6) J(7) is the transpose of r4  r5  r6 r7.
+
+  Since r1 is free at entry, we calculate the Js first.*/
+/*19 cycles.*/
+#define OC_TRANSPOSE(_y) \
+  "#OC_TRANSPOSE\n\t" \
+  "movq %%mm4,%%mm1\n\t" \
+  "punpcklwd %%mm5,%%mm4\n\t" \
+  "movq %%mm0,"OC_I(0,_y)"\n\t" \
+  "punpckhwd %%mm5,%%mm1\n\t" \
+  "movq %%mm6,%%mm0\n\t" \
+  "punpcklwd %%mm7,%%mm6\n\t" \
+  "movq %%mm4,%%mm5\n\t" \
+  "punpckldq %%mm6,%%mm4\n\t" \
+  "punpckhdq %%mm6,%%mm5\n\t" \
+  "movq %%mm1,%%mm6\n\t" \
+  "movq %%mm4,"OC_J(4,_y)"\n\t" \
+  "punpckhwd %%mm7,%%mm0\n\t" \
+  "movq %%mm5,"OC_J(5,_y)"\n\t" \
+  "punpckhdq %%mm0,%%mm6\n\t" \
+  "movq "OC_I(0,_y)",%%mm4\n\t" \
+  "punpckldq %%mm0,%%mm1\n\t" \
+  "movq "OC_I(1,_y)",%%mm5\n\t" \
+  "movq %%mm4,%%mm0\n\t" \
+  "movq %%mm6,"OC_J(7,_y)"\n\t" \
+  "punpcklwd %%mm5,%%mm0\n\t" \
+  "movq %%mm1,"OC_J(6,_y)"\n\t" \
+  "punpckhwd %%mm5,%%mm4\n\t" \
+  "movq %%mm2,%%mm5\n\t" \
+  "punpcklwd %%mm3,%%mm2\n\t" \
+  "movq %%mm0,%%mm1\n\t" \
+  "punpckldq %%mm2,%%mm0\n\t" \
+  "punpckhdq %%mm2,%%mm1\n\t" \
+  "movq %%mm4,%%mm2\n\t" \
+  "movq %%mm0,"OC_I(0,_y)"\n\t" \
+  "punpckhwd %%mm3,%%mm5\n\t" \
+  "movq %%mm1,"OC_I(1,_y)"\n\t" \
+  "punpckhdq %%mm5,%%mm4\n\t" \
+  "punpckldq %%mm5,%%mm2\n\t" \
+  "movq %%mm4,"OC_I(3,_y)"\n\t" \
+  "movq %%mm2,"OC_I(2,_y)"\n\t" \
+  "#end OC_TRANSPOSE\n\t" \
+
+/*38+19=57 cycles.*/
+#define OC_COLUMN_IDCT(_y) \
+  "#OC_COLUMN_IDCT\n" \
+  OC_IDCT_BEGIN(_y,_y) \
+  "paddw "OC_MEM_OFFS(0x00,c)",%%mm2\n\t" \
+  /*r1=H'+H'*/ \
+  "paddw %%mm1,%%mm1\n\t" \
+  /*r1=R1=A''+H'*/ \
+  "paddw %%mm2,%%mm1\n\t" \
+  /*r2=NR2*/ \
+  "psraw $4,%%mm2\n\t" \
+  /*r4=E'=E-G*/ \
+  "psubw %%mm7,%%mm4\n\t" \
+  /*r1=NR1*/ \
+  "psraw $4,%%mm1\n\t" \
+  /*r3=D'*/ \
+  "movq "OC_I(2,_y)",%%mm3\n\t" \
+  /*r7=G+G*/ \
+  "paddw %%mm7,%%mm7\n\t" \
+  /*Store NR2 at I(2).*/ \
+  "movq %%mm2,"OC_I(2,_y)"\n\t" \
+  /*r7=G'=E+G*/ \
+  "paddw %%mm4,%%mm7\n\t" \
+  /*Store NR1 at I(1).*/ \
+  "movq %%mm1,"OC_I(1,_y)"\n\t" \
+  /*r4=R4=E'-D'*/ \
+  "psubw %%mm3,%%mm4\n\t" \
+  "paddw "OC_MEM_OFFS(0x00,c)",%%mm4\n\t" \
+  /*r3=D'+D'*/ \
+  "paddw %%mm3,%%mm3\n\t" \
+  /*r3=R3=E'+D'*/ \
+  "paddw %%mm4,%%mm3\n\t" \
+  /*r4=NR4*/ \
+  "psraw $4,%%mm4\n\t" \
+  /*r6=R6=F'-B''*/ \
+  "psubw %%mm5,%%mm6\n\t" \
+  /*r3=NR3*/ \
+  "psraw $4,%%mm3\n\t" \
+  "paddw "OC_MEM_OFFS(0x00,c)",%%mm6\n\t" \
+  /*r5=B''+B''*/ \
+  "paddw %%mm5,%%mm5\n\t" \
+  /*r5=R5=F'+B''*/ \
+  "paddw %%mm6,%%mm5\n\t" \
+  /*r6=NR6*/ \
+  "psraw $4,%%mm6\n\t" \
+  /*Store NR4 at J(4).*/ \
+  "movq %%mm4,"OC_J(4,_y)"\n\t" \
+  /*r5=NR5*/ \
+  "psraw $4,%%mm5\n\t" \
+  /*Store NR3 at I(3).*/ \
+  "movq %%mm3,"OC_I(3,_y)"\n\t" \
+  /*r7=R7=G'-C'*/ \
+  "psubw %%mm0,%%mm7\n\t" \
+  "paddw "OC_MEM_OFFS(0x00,c)",%%mm7\n\t" \
+  /*r0=C'+C'*/ \
+  "paddw %%mm0,%%mm0\n\t" \
+  /*r0=R0=G'+C'*/ \
+  "paddw %%mm7,%%mm0\n\t" \
+  /*r7=NR7*/ \
+  "psraw $4,%%mm7\n\t" \
+  /*Store NR6 at J(6).*/ \
+  "movq %%mm6,"OC_J(6,_y)"\n\t" \
+  /*r0=NR0*/ \
+  "psraw $4,%%mm0\n\t" \
+  /*Store NR5 at J(5).*/ \
+  "movq %%mm5,"OC_J(5,_y)"\n\t" \
+  /*Store NR7 at J(7).*/ \
+  "movq %%mm7,"OC_J(7,_y)"\n\t" \
+  /*Store NR0 at I(0).*/ \
+  "movq %%mm0,"OC_I(0,_y)"\n\t" \
+  "#end OC_COLUMN_IDCT\n\t" \
+
+static void oc_idct8x8_slow_mmx(ogg_int16_t _y[64],ogg_int16_t _x[64]){
+  int i;
+  /*This routine accepts an 8x8 matrix, but in partially transposed form.
+    Every 4x4 block is transposed.*/
+  __asm__ __volatile__(
+#define OC_I(_k,_y)   OC_MEM_OFFS((_k)*16,_y)
+#define OC_J(_k,_y)   OC_MEM_OFFS(((_k)-4)*16+8,_y)
+    OC_ROW_IDCT(y,x)
+    OC_TRANSPOSE(y)
+#undef  OC_I
+#undef  OC_J
+#define OC_I(_k,_y)   OC_MEM_OFFS((_k)*16+64,_y)
+#define OC_J(_k,_y)   OC_MEM_OFFS(((_k)-4)*16+72,_y)
+    OC_ROW_IDCT(y,x)
+    OC_TRANSPOSE(y)
+#undef  OC_I
+#undef  OC_J
+#define OC_I(_k,_y)   OC_MEM_OFFS((_k)*16,_y)
+#define OC_J(_k,_y)   OC_I(_k,_y)
+    OC_COLUMN_IDCT(y)
+#undef  OC_I
+#undef  OC_J
+#define OC_I(_k,_y)   OC_MEM_OFFS((_k)*16+8,_y)
+#define OC_J(_k,_y)   OC_I(_k,_y)
+    OC_COLUMN_IDCT(y)
+#undef  OC_I
+#undef  OC_J
+    :[y]"=m"OC_ARRAY_OPERAND(ogg_int16_t,_y,64)
+    :[x]"m"OC_CONST_ARRAY_OPERAND(ogg_int16_t,_x,64),
+     [c]"m"OC_CONST_ARRAY_OPERAND(ogg_int16_t,OC_IDCT_CONSTS,64)
+  );
+  __asm__ __volatile__("pxor %%mm0,%%mm0\n\t"::);
+  for(i=0;i<4;i++){
+    __asm__ __volatile__(
+      "movq %%mm0,"OC_MEM_OFFS(0x00,x)"\n\t"
+      "movq %%mm0,"OC_MEM_OFFS(0x08,x)"\n\t"
+      "movq %%mm0,"OC_MEM_OFFS(0x10,x)"\n\t"
+      "movq %%mm0,"OC_MEM_OFFS(0x18,x)"\n\t"
+      :[x]"=m"OC_ARRAY_OPERAND(ogg_int16_t,_x+16*i,16)
+    );
+  }
+}
+
+/*25 cycles.*/
+#define OC_IDCT_BEGIN_10(_y,_x) \
+ "#OC_IDCT_BEGIN_10\n\t" \
+ "movq "OC_I(3,_x)",%%mm2\n\t" \
+ "nop\n\t" \
+ "movq "OC_MEM_OFFS(0x30,c)",%%mm6\n\t" \
+ "movq %%mm2,%%mm4\n\t" \
+ "movq "OC_MEM_OFFS(0x50,c)",%%mm1\n\t" \
+ "pmulhw %%mm6,%%mm4\n\t" \
+ "movq "OC_I(1,_x)",%%mm3\n\t" \
+ "pmulhw %%mm2,%%mm1\n\t" \
+ "movq "OC_MEM_OFFS(0x10,c)",%%mm0\n\t" \
+ "paddw %%mm2,%%mm4\n\t" \
+ "pxor %%mm6,%%mm6\n\t" \
+ "paddw %%mm1,%%mm2\n\t" \
+ "movq "OC_I(2,_x)",%%mm5\n\t" \
+ "pmulhw %%mm3,%%mm0\n\t" \
+ "movq %%mm5,%%mm1\n\t" \
+ "paddw %%mm3,%%mm0\n\t" \
+ "pmulhw "OC_MEM_OFFS(0x70,c)",%%mm3\n\t" \
+ "psubw %%mm2,%%mm6\n\t" \
+ "pmulhw "OC_MEM_OFFS(0x20,c)",%%mm5\n\t" \
+ "psubw %%mm4,%%mm0\n\t" \
+ "movq "OC_I(2,_x)",%%mm7\n\t" \
+ "paddw %%mm4,%%mm4\n\t" \
+ "paddw %%mm5,%%mm7\n\t" \
+ "paddw %%mm0,%%mm4\n\t" \
+ "pmulhw "OC_MEM_OFFS(0x60,c)",%%mm1\n\t" \
+ "psubw %%mm6,%%mm3\n\t" \
+ "movq %%mm4,"OC_I(1,_y)"\n\t" \
+ "paddw %%mm6,%%mm6\n\t" \
+ "movq "OC_MEM_OFFS(0x40,c)",%%mm4\n\t" \
+ "paddw %%mm3,%%mm6\n\t" \
+ "movq %%mm3,%%mm5\n\t" \
+ "pmulhw %%mm4,%%mm3\n\t" \
+ "movq %%mm6,"OC_I(2,_y)"\n\t" \
+ "movq %%mm0,%%mm2\n\t" \
+ "movq "OC_I(0,_x)",%%mm6\n\t" \
+ "pmulhw %%mm4,%%mm0\n\t" \
+ "paddw %%mm3,%%mm5\n\t" \
+ "paddw %%mm0,%%mm2\n\t" \
+ "psubw %%mm1,%%mm5\n\t" \
+ "pmulhw %%mm4,%%mm6\n\t" \
+ "paddw "OC_I(0,_x)",%%mm6\n\t" \
+ "paddw %%mm1,%%mm1\n\t" \
+ "movq %%mm6,%%mm4\n\t" \
+ "paddw %%mm5,%%mm1\n\t" \
+ "psubw %%mm2,%%mm6\n\t" \
+ "paddw %%mm2,%%mm2\n\t" \
+ "movq "OC_I(1,_y)",%%mm0\n\t" \
+ "paddw %%mm6,%%mm2\n\t" \
+ "psubw %%mm1,%%mm2\n\t" \
+ "nop\n\t" \
+ "#end OC_IDCT_BEGIN_10\n\t" \
+
+/*25+8=33 cycles.*/
+#define OC_ROW_IDCT_10(_y,_x) \
+ "#OC_ROW_IDCT_10\n\t" \
+ OC_IDCT_BEGIN_10(_y,_x) \
+ /*r3=D'*/ \
+ "movq "OC_I(2,_y)",%%mm3\n\t" \
+ /*r4=E'=E-G*/ \
+ "psubw %%mm7,%%mm4\n\t" \
+ /*r1=H'+H'*/ \
+ "paddw %%mm1,%%mm1\n\t" \
+ /*r7=G+G*/ \
+ "paddw %%mm7,%%mm7\n\t" \
+ /*r1=R1=A''+H'*/ \
+ "paddw %%mm2,%%mm1\n\t" \
+ /*r7=G'=E+G*/ \
+ "paddw %%mm4,%%mm7\n\t" \
+ /*r4=R4=E'-D'*/ \
+ "psubw %%mm3,%%mm4\n\t" \
+ "paddw %%mm3,%%mm3\n\t" \
+ /*r6=R6=F'-B''*/ \
+ "psubw %%mm5,%%mm6\n\t" \
+ "paddw %%mm5,%%mm5\n\t" \
+ /*r3=R3=E'+D'*/ \
+ "paddw %%mm4,%%mm3\n\t" \
+ /*r5=R5=F'+B''*/ \
+ "paddw %%mm6,%%mm5\n\t" \
+ /*r7=R7=G'-C'*/ \
+ "psubw %%mm0,%%mm7\n\t" \
+ "paddw %%mm0,%%mm0\n\t" \
+ /*Save R1.*/ \
+ "movq %%mm1,"OC_I(1,_y)"\n\t" \
+ /*r0=R0=G'+C'*/ \
+ "paddw %%mm7,%%mm0\n\t" \
+ "#end OC_ROW_IDCT_10\n\t" \
+
+/*25+19=44 cycles'*/
+#define OC_COLUMN_IDCT_10(_y) \
+ "#OC_COLUMN_IDCT_10\n\t" \
+ OC_IDCT_BEGIN_10(_y,_y) \
+ "paddw "OC_MEM_OFFS(0x00,c)",%%mm2\n\t" \
+ /*r1=H'+H'*/ \
+ "paddw %%mm1,%%mm1\n\t" \
+ /*r1=R1=A''+H'*/ \
+ "paddw %%mm2,%%mm1\n\t" \
+ /*r2=NR2*/ \
+ "psraw $4,%%mm2\n\t" \
+ /*r4=E'=E-G*/ \
+ "psubw %%mm7,%%mm4\n\t" \
+ /*r1=NR1*/ \
+ "psraw $4,%%mm1\n\t" \
+ /*r3=D'*/ \
+ "movq "OC_I(2,_y)",%%mm3\n\t" \
+ /*r7=G+G*/ \
+ "paddw %%mm7,%%mm7\n\t" \
+ /*Store NR2 at I(2).*/ \
+ "movq %%mm2,"OC_I(2,_y)"\n\t" \
+ /*r7=G'=E+G*/ \
+ "paddw %%mm4,%%mm7\n\t" \
+ /*Store NR1 at I(1).*/ \
+ "movq %%mm1,"OC_I(1,_y)"\n\t" \
+ /*r4=R4=E'-D'*/ \
+ "psubw %%mm3,%%mm4\n\t" \
+ "paddw "OC_MEM_OFFS(0x00,c)",%%mm4\n\t" \
+ /*r3=D'+D'*/ \
+ "paddw %%mm3,%%mm3\n\t" \
+ /*r3=R3=E'+D'*/ \
+ "paddw %%mm4,%%mm3\n\t" \
+ /*r4=NR4*/ \
+ "psraw $4,%%mm4\n\t" \
+ /*r6=R6=F'-B''*/ \
+ "psubw %%mm5,%%mm6\n\t" \
+ /*r3=NR3*/ \
+ "psraw $4,%%mm3\n\t" \
+ "paddw "OC_MEM_OFFS(0x00,c)",%%mm6\n\t" \
+ /*r5=B''+B''*/ \
+ "paddw %%mm5,%%mm5\n\t" \
+ /*r5=R5=F'+B''*/ \
+ "paddw %%mm6,%%mm5\n\t" \
+ /*r6=NR6*/ \
+ "psraw $4,%%mm6\n\t" \
+ /*Store NR4 at J(4).*/ \
+ "movq %%mm4,"OC_J(4,_y)"\n\t" \
+ /*r5=NR5*/ \
+ "psraw $4,%%mm5\n\t" \
+ /*Store NR3 at I(3).*/ \
+ "movq %%mm3,"OC_I(3,_y)"\n\t" \
+ /*r7=R7=G'-C'*/ \
+ "psubw %%mm0,%%mm7\n\t" \
+ "paddw "OC_MEM_OFFS(0x00,c)",%%mm7\n\t" \
+ /*r0=C'+C'*/ \
+ "paddw %%mm0,%%mm0\n\t" \
+ /*r0=R0=G'+C'*/ \
+ "paddw %%mm7,%%mm0\n\t" \
+ /*r7=NR7*/ \
+ "psraw $4,%%mm7\n\t" \
+ /*Store NR6 at J(6).*/ \
+ "movq %%mm6,"OC_J(6,_y)"\n\t" \
+ /*r0=NR0*/ \
+ "psraw $4,%%mm0\n\t" \
+ /*Store NR5 at J(5).*/ \
+ "movq %%mm5,"OC_J(5,_y)"\n\t" \
+ /*Store NR7 at J(7).*/ \
+ "movq %%mm7,"OC_J(7,_y)"\n\t" \
+ /*Store NR0 at I(0).*/ \
+ "movq %%mm0,"OC_I(0,_y)"\n\t" \
+ "#end OC_COLUMN_IDCT_10\n\t" \
+
+static void oc_idct8x8_10_mmx(ogg_int16_t _y[64],ogg_int16_t _x[64]){
+  __asm__ __volatile__(
+#define OC_I(_k,_y) OC_MEM_OFFS((_k)*16,_y)
+#define OC_J(_k,_y) OC_MEM_OFFS(((_k)-4)*16+8,_y)
+    /*Done with dequant, descramble, and partial transpose.
+      Now do the iDCT itself.*/
+    OC_ROW_IDCT_10(y,x)
+    OC_TRANSPOSE(y)
+#undef  OC_I
+#undef  OC_J
+#define OC_I(_k,_y) OC_MEM_OFFS((_k)*16,_y)
+#define OC_J(_k,_y) OC_I(_k,_y)
+    OC_COLUMN_IDCT_10(y)
+#undef  OC_I
+#undef  OC_J
+#define OC_I(_k,_y) OC_MEM_OFFS((_k)*16+8,_y)
+#define OC_J(_k,_y) OC_I(_k,_y)
+    OC_COLUMN_IDCT_10(y)
+#undef  OC_I
+#undef  OC_J
+    :[y]"=m"OC_ARRAY_OPERAND(ogg_int16_t,_y,64)
+    :[x]"m"OC_CONST_ARRAY_OPERAND(ogg_int16_t,_x,64),
+     [c]"m"OC_CONST_ARRAY_OPERAND(ogg_int16_t,OC_IDCT_CONSTS,64)
+  );
+  __asm__ __volatile__(
+    "pxor %%mm0,%%mm0\n\t"
+    "movq %%mm0,"OC_MEM_OFFS(0x00,x)"\n\t"
+    "movq %%mm0,"OC_MEM_OFFS(0x10,x)"\n\t"
+    "movq %%mm0,"OC_MEM_OFFS(0x20,x)"\n\t"
+    "movq %%mm0,"OC_MEM_OFFS(0x30,x)"\n\t"
+    :[x]"+m"OC_ARRAY_OPERAND(ogg_int16_t,_x,28)
+  );
+}
+
+/*Performs an inverse 8x8 Type-II DCT transform.
+  The input is assumed to be scaled by a factor of 4 relative to orthonormal
+   version of the transform.*/
+void oc_idct8x8_mmx(ogg_int16_t _y[64],ogg_int16_t _x[64],int _last_zzi){
+  /*_last_zzi is subtly different from an actual count of the number of
+     coefficients we decoded for this block.
+    It contains the value of zzi BEFORE the final token in the block was
+     decoded.
+    In most cases this is an EOB token (the continuation of an EOB run from a
+     previous block counts), and so this is the same as the coefficient count.
+    However, in the case that the last token was NOT an EOB token, but filled
+     the block up with exactly 64 coefficients, _last_zzi will be less than 64.
+    Provided the last token was not a pure zero run, the minimum value it can
+     be is 46, and so that doesn't affect any of the cases in this routine.
+    However, if the last token WAS a pure zero run of length 63, then _last_zzi
+     will be 1 while the number of coefficients decoded is 64.
+    Thus, we will trigger the following special case, where the real
+     coefficient count would not.
+    Note also that a zero run of length 64 will give _last_zzi a value of 0,
+     but we still process the DC coefficient, which might have a non-zero value
+     due to DC prediction.
+    Although convoluted, this is arguably the correct behavior: it allows us to
+     use a smaller transform when the block ends with a long zero run instead
+     of a normal EOB token.
+    It could be smarter... multiple separate zero runs at the end of a block
+     will fool it, but an encoder that generates these really deserves what it
+     gets.
+    Needless to say we inherited this approach from VP3.*/
+  /*Then perform the iDCT.*/
+  if(_last_zzi<=10)oc_idct8x8_10_mmx(_y,_x);
+  else oc_idct8x8_slow_mmx(_y,_x);
+}
+
+#endif

thirdparty/libtheora/x86/mmxloop.h

@@ -0,0 +1,318 @@
+#if !defined(_x86_mmxloop_H)
+# define _x86_mmxloop_H (1)
+# include <stddef.h>
+# include "x86int.h"
+
+#if defined(OC_X86_ASM)
+
+/*On entry, mm0={a0,...,a7}, mm1={b0,...,b7}, mm2={c0,...,c7}, mm3={d0,...d7}.
+  On exit, mm1={b0+lflim(R_0,L),...,b7+lflim(R_7,L)} and
+   mm2={c0-lflim(R_0,L),...,c7-lflim(R_7,L)}; mm0 and mm3 are clobbered.*/
+#define OC_LOOP_FILTER8_MMX \
+  "#OC_LOOP_FILTER8_MMX\n\t" \
+  /*mm7=0*/ \
+  "pxor %%mm7,%%mm7\n\t" \
+  /*mm6:mm0={a0,...,a7}*/ \
+  "movq %%mm0,%%mm6\n\t" \
+  "punpcklbw %%mm7,%%mm0\n\t" \
+  "punpckhbw %%mm7,%%mm6\n\t" \
+  /*mm3:mm5={d0,...,d7}*/ \
+  "movq %%mm3,%%mm5\n\t" \
+  "punpcklbw %%mm7,%%mm3\n\t" \
+  "punpckhbw %%mm7,%%mm5\n\t" \
+  /*mm6:mm0={a0-d0,...,a7-d7}*/ \
+  "psubw %%mm3,%%mm0\n\t" \
+  "psubw %%mm5,%%mm6\n\t" \
+  /*mm3:mm1={b0,...,b7}*/ \
+  "movq %%mm1,%%mm3\n\t" \
+  "punpcklbw %%mm7,%%mm1\n\t" \
+  "movq %%mm2,%%mm4\n\t" \
+  "punpckhbw %%mm7,%%mm3\n\t" \
+  /*mm5:mm4={c0,...,c7}*/ \
+  "movq %%mm2,%%mm5\n\t" \
+  "punpcklbw %%mm7,%%mm4\n\t" \
+  "punpckhbw %%mm7,%%mm5\n\t" \
+  /*mm7={3}x4 \
+    mm5:mm4={c0-b0,...,c7-b7}*/ \
+  "pcmpeqw %%mm7,%%mm7\n\t" \
+  "psubw %%mm1,%%mm4\n\t" \
+  "psrlw $14,%%mm7\n\t" \
+  "psubw %%mm3,%%mm5\n\t" \
+  /*Scale by 3.*/ \
+  "pmullw %%mm7,%%mm4\n\t" \
+  "pmullw %%mm7,%%mm5\n\t" \
+  /*mm7={4}x4 \
+    mm5:mm4=f={a0-d0+3*(c0-b0),...,a7-d7+3*(c7-b7)}*/ \
+  "psrlw $1,%%mm7\n\t" \
+  "paddw %%mm0,%%mm4\n\t" \
+  "psllw $2,%%mm7\n\t" \
+  "movq (%[ll]),%%mm0\n\t" \
+  "paddw %%mm6,%%mm5\n\t" \
+  /*R_i has the range [-127,128], so we compute -R_i instead. \
+    mm4=-R_i=-(f+4>>3)=0xFF^(f-4>>3)*/ \
+  "psubw %%mm7,%%mm4\n\t" \
+  "psubw %%mm7,%%mm5\n\t" \
+  "psraw $3,%%mm4\n\t" \
+  "psraw $3,%%mm5\n\t" \
+  "pcmpeqb %%mm7,%%mm7\n\t" \
+  "packsswb %%mm5,%%mm4\n\t" \
+  "pxor %%mm6,%%mm6\n\t" \
+  "pxor %%mm7,%%mm4\n\t" \
+  "packuswb %%mm3,%%mm1\n\t" \
+  /*Now compute lflim of -mm4 cf. Section 7.10 of the sepc.*/ \
+  /*There's no unsigned byte+signed byte with unsigned saturation op code, so \
+     we have to split things by sign (the other option is to work in 16 bits, \
+     but working in 8 bits gives much better parallelism). \
+    We compute abs(R_i), but save a mask of which terms were negative in mm6. \
+    Then we compute mm4=abs(lflim(R_i,L))=min(abs(R_i),max(2*L-abs(R_i),0)). \
+    Finally, we split mm4 into positive and negative pieces using the mask in \
+     mm6, and add and subtract them as appropriate.*/ \
+  /*mm4=abs(-R_i)*/ \
+  /*mm7=255-2*L*/ \
+  "pcmpgtb %%mm4,%%mm6\n\t" \
+  "psubb %%mm0,%%mm7\n\t" \
+  "pxor %%mm6,%%mm4\n\t" \
+  "psubb %%mm0,%%mm7\n\t" \
+  "psubb %%mm6,%%mm4\n\t" \
+  /*mm7=255-max(2*L-abs(R_i),0)*/ \
+  "paddusb %%mm4,%%mm7\n\t" \
+  /*mm4=min(abs(R_i),max(2*L-abs(R_i),0))*/ \
+  "paddusb %%mm7,%%mm4\n\t" \
+  "psubusb %%mm7,%%mm4\n\t" \
+  /*Now split mm4 by the original sign of -R_i.*/ \
+  "movq %%mm4,%%mm5\n\t" \
+  "pand %%mm6,%%mm4\n\t" \
+  "pandn %%mm5,%%mm6\n\t" \
+  /*mm1={b0+lflim(R_0,L),...,b7+lflim(R_7,L)}*/ \
+  /*mm2={c0-lflim(R_0,L),...,c7-lflim(R_7,L)}*/ \
+  "paddusb %%mm4,%%mm1\n\t" \
+  "psubusb %%mm4,%%mm2\n\t" \
+  "psubusb %%mm6,%%mm1\n\t" \
+  "paddusb %%mm6,%%mm2\n\t" \
+
+/*On entry, mm0={a0,...,a7}, mm1={b0,...,b7}, mm2={c0,...,c7}, mm3={d0,...d7}.
+  On exit, mm1={b0+lflim(R_0,L),...,b7+lflim(R_7,L)} and
+   mm2={c0-lflim(R_0,L),...,c7-lflim(R_7,L)}.
+  All other MMX registers are clobbered.*/
+#define OC_LOOP_FILTER8_MMXEXT \
+  "#OC_LOOP_FILTER8_MMXEXT\n\t" \
+  /*R_i=(a_i-3*b_i+3*c_i-d_i+4>>3) has the range [-127,128], so we compute \
+     -R_i=(-a_i+3*b_i-3*c_i+d_i+3>>3) instead.*/ \
+  /*This first part is based on the transformation \
+      f = -(3*(c-b)+a-d+4>>3) \
+        = -(3*(c+255-b)+(a+255-d)+4-1020>>3) \
+        = -(3*(c+~b)+(a+~d)-1016>>3) \
+        = 127-(3*(c+~b)+(a+~d)>>3) \
+        = 128+~(3*(c+~b)+(a+~d)>>3) (mod 256). \
+    Although pavgb(a,b) = (a+b+1>>1) (biased up), we rely heavily on the \
+     fact that ~pavgb(~a,~b) = (a+b>>1) (biased down). \
+    Using this, the last expression above can be computed in 8 bits of working \
+     precision via: \
+      u = ~pavgb(~b,c); \
+      v = pavgb(b,~c); \
+      This mask is 0 or 0xFF, and controls whether t is biased up or down: \
+      m = u-v; \
+      t = m^pavgb(m^~a,m^d); \
+      f = 128+pavgb(pavgb(t,u),v); \
+    This required some careful analysis to ensure that carries are propagated \
+     correctly in all cases, but has been checked exhaustively.*/ \
+  /*input (a, b, c, d, ., ., ., .)*/ \
+  /*ff=0xFF; \
+    u=b; \
+    v=c; \
+    ll=255-2*L;*/ \
+  "pcmpeqb %%mm7,%%mm7\n\t" \
+  "movq %%mm1,%%mm4\n\t" \
+  "movq %%mm2,%%mm5\n\t" \
+  "movq (%[ll]),%%mm6\n\t" \
+  /*allocated u, v, ll, ff: (a, b, c, d, u, v, ll, ff)*/ \
+  /*u^=ff; \
+    v^=ff;*/ \
+  "pxor %%mm7,%%mm4\n\t" \
+  "pxor %%mm7,%%mm5\n\t" \
+  /*allocated ll: (a, b, c, d, u, v, ll, ff)*/ \
+  /*u=pavgb(u,c); \
+    v=pavgb(v,b);*/ \
+  "pavgb %%mm2,%%mm4\n\t" \
+  "pavgb %%mm1,%%mm5\n\t" \
+  /*u^=ff; \
+    a^=ff;*/ \
+  "pxor %%mm7,%%mm4\n\t" \
+  "pxor %%mm7,%%mm0\n\t" \
+  /*m=u-v;*/ \
+  "psubb %%mm5,%%mm4\n\t" \
+  /*freed u, allocated m: (a, b, c, d, m, v, ll, ff)*/ \
+  /*a^=m; \
+    d^=m;*/ \
+  "pxor %%mm4,%%mm0\n\t" \
+  "pxor %%mm4,%%mm3\n\t" \
+  /*t=pavgb(a,d);*/ \
+  "pavgb %%mm3,%%mm0\n\t" \
+  "psllw $7,%%mm7\n\t" \
+  /*freed a, d, ff, allocated t, of: (t, b, c, ., m, v, ll, of)*/ \
+  /*t^=m; \
+    u=m+v;*/ \
+  "pxor %%mm4,%%mm0\n\t" \
+  "paddb %%mm5,%%mm4\n\t" \
+  /*freed t, m, allocated f, u: (f, b, c, ., u, v, ll, of)*/ \
+  /*f=pavgb(f,u); \
+    of=128;*/ \
+  "pavgb %%mm4,%%mm0\n\t" \
+  "packsswb %%mm7,%%mm7\n\t" \
+  /*freed u, ff, allocated ll: (f, b, c, ., ll, v, ll, of)*/ \
+  /*f=pavgb(f,v);*/ \
+  "pavgb %%mm5,%%mm0\n\t" \
+  "movq %%mm7,%%mm3\n\t" \
+  "movq %%mm6,%%mm4\n\t" \
+  /*freed v, allocated of: (f, b, c, of, ll, ., ll, of)*/ \
+  /*Now compute lflim of R_i=-(128+mm0) cf. Section 7.10 of the sepc.*/ \
+  /*There's no unsigned byte+signed byte with unsigned saturation op code, so \
+     we have to split things by sign (the other option is to work in 16 bits, \
+     but staying in 8 bits gives much better parallelism).*/ \
+  /*Instead of adding the offset of 128 in mm3, we use it to split mm0. \
+    This is the same number of instructions as computing a mask and splitting \
+     after the lflim computation, but has shorter dependency chains.*/ \
+  /*mm0=R_i<0?-R_i:0 (denoted abs(R_i<0))\
+    mm3=R_i>0?R_i:0* (denoted abs(R_i>0))*/ \
+  "psubusb %%mm0,%%mm3\n\t" \
+  "psubusb %%mm7,%%mm0\n\t" \
+  /*mm6=255-max(2*L-abs(R_i<0),0) \
+    mm4=255-max(2*L-abs(R_i>0),0)*/ \
+  "paddusb %%mm3,%%mm4\n\t" \
+  "paddusb %%mm0,%%mm6\n\t" \
+  /*mm0=min(abs(R_i<0),max(2*L-abs(R_i<0),0)) \
+    mm3=min(abs(R_i>0),max(2*L-abs(R_i>0),0))*/ \
+  "paddusb %%mm4,%%mm3\n\t" \
+  "paddusb %%mm6,%%mm0\n\t" \
+  "psubusb %%mm4,%%mm3\n\t" \
+  "psubusb %%mm6,%%mm0\n\t" \
+  /*mm1={b0+lflim(R_0,L),...,b7+lflim(R_7,L)}*/ \
+  /*mm2={c0-lflim(R_0,L),...,c7-lflim(R_7,L)}*/ \
+  "paddusb %%mm3,%%mm1\n\t" \
+  "psubusb %%mm3,%%mm2\n\t" \
+  "psubusb %%mm0,%%mm1\n\t" \
+  "paddusb %%mm0,%%mm2\n\t" \
+
+#define OC_LOOP_FILTER_V(_filter,_pix,_ystride,_ll) \
+  do{ \
+    ptrdiff_t ystride3__; \
+    __asm__ __volatile__( \
+      /*mm0={a0,...,a7}*/ \
+      "movq (%[pix]),%%mm0\n\t" \
+      /*ystride3=_ystride*3*/ \
+      "lea (%[ystride],%[ystride],2),%[ystride3]\n\t" \
+      /*mm3={d0,...,d7}*/ \
+      "movq (%[pix],%[ystride3]),%%mm3\n\t" \
+      /*mm1={b0,...,b7}*/ \
+      "movq (%[pix],%[ystride]),%%mm1\n\t" \
+      /*mm2={c0,...,c7}*/ \
+      "movq (%[pix],%[ystride],2),%%mm2\n\t" \
+      _filter \
+      /*Write it back out.*/ \
+      "movq %%mm1,(%[pix],%[ystride])\n\t" \
+      "movq %%mm2,(%[pix],%[ystride],2)\n\t" \
+      :[ystride3]"=&r"(ystride3__) \
+      :[pix]"r"(_pix-_ystride*2),[ystride]"r"((ptrdiff_t)(_ystride)), \
+       [ll]"r"(_ll) \
+      :"memory" \
+    ); \
+  } \
+  while(0)
+
+#define OC_LOOP_FILTER_H(_filter,_pix,_ystride,_ll) \
+  do{ \
+    unsigned char *pix__; \
+    ptrdiff_t      ystride3__; \
+    ptrdiff_t      d__; \
+    pix__=(_pix)-2; \
+    __asm__ __volatile__( \
+      /*x x x x d0 c0 b0 a0*/ \
+      "movd (%[pix]),%%mm0\n\t" \
+      /*x x x x d1 c1 b1 a1*/ \
+      "movd (%[pix],%[ystride]),%%mm1\n\t" \
+      /*ystride3=_ystride*3*/ \
+      "lea (%[ystride],%[ystride],2),%[ystride3]\n\t" \
+      /*x x x x d2 c2 b2 a2*/ \
+      "movd (%[pix],%[ystride],2),%%mm2\n\t" \
+      /*x x x x d3 c3 b3 a3*/ \
+      "lea (%[pix],%[ystride],4),%[d]\n\t" \
+      "movd (%[pix],%[ystride3]),%%mm3\n\t" \
+      /*x x x x d4 c4 b4 a4*/ \
+      "movd (%[d]),%%mm4\n\t" \
+      /*x x x x d5 c5 b5 a5*/ \
+      "movd (%[d],%[ystride]),%%mm5\n\t" \
+      /*x x x x d6 c6 b6 a6*/ \
+      "movd (%[d],%[ystride],2),%%mm6\n\t" \
+      /*x x x x d7 c7 b7 a7*/ \
+      "movd (%[d],%[ystride3]),%%mm7\n\t" \
+      /*mm0=d1 d0 c1 c0 b1 b0 a1 a0*/ \
+      "punpcklbw %%mm1,%%mm0\n\t" \
+      /*mm2=d3 d2 c3 c2 b3 b2 a3 a2*/ \
+      "punpcklbw %%mm3,%%mm2\n\t" \
+      /*mm3=d1 d0 c1 c0 b1 b0 a1 a0*/ \
+      "movq %%mm0,%%mm3\n\t" \
+      /*mm0=b3 b2 b1 b0 a3 a2 a1 a0*/ \
+      "punpcklwd %%mm2,%%mm0\n\t" \
+      /*mm3=d3 d2 d1 d0 c3 c2 c1 c0*/ \
+      "punpckhwd %%mm2,%%mm3\n\t" \
+      /*mm1=b3 b2 b1 b0 a3 a2 a1 a0*/ \
+      "movq %%mm0,%%mm1\n\t" \
+      /*mm4=d5 d4 c5 c4 b5 b4 a5 a4*/ \
+      "punpcklbw %%mm5,%%mm4\n\t" \
+      /*mm6=d7 d6 c7 c6 b7 b6 a7 a6*/ \
+      "punpcklbw %%mm7,%%mm6\n\t" \
+      /*mm5=d5 d4 c5 c4 b5 b4 a5 a4*/ \
+      "movq %%mm4,%%mm5\n\t" \
+      /*mm4=b7 b6 b5 b4 a7 a6 a5 a4*/ \
+      "punpcklwd %%mm6,%%mm4\n\t" \
+      /*mm5=d7 d6 d5 d4 c7 c6 c5 c4*/ \
+      "punpckhwd %%mm6,%%mm5\n\t" \
+      /*mm2=d3 d2 d1 d0 c3 c2 c1 c0*/ \
+      "movq %%mm3,%%mm2\n\t" \
+      /*mm0=a7 a6 a5 a4 a3 a2 a1 a0*/ \
+      "punpckldq %%mm4,%%mm0\n\t" \
+      /*mm1=b7 b6 b5 b4 b3 b2 b1 b0*/ \
+      "punpckhdq %%mm4,%%mm1\n\t" \
+      /*mm2=c7 c6 c5 c4 c3 c2 c1 c0*/ \
+      "punpckldq %%mm5,%%mm2\n\t" \
+      /*mm3=d7 d6 d5 d4 d3 d2 d1 d0*/ \
+      "punpckhdq %%mm5,%%mm3\n\t" \
+      _filter \
+      /*mm2={b0+R_0'',...,b7+R_7''}*/ \
+      "movq %%mm1,%%mm0\n\t" \
+      /*mm1={b0+R_0'',c0-R_0'',...,b3+R_3'',c3-R_3''}*/ \
+      "punpcklbw %%mm2,%%mm1\n\t" \
+      /*mm2={b4+R_4'',c4-R_4'',...,b7+R_7'',c7-R_7''}*/ \
+      "punpckhbw %%mm2,%%mm0\n\t" \
+      /*[d]=c1 b1 c0 b0*/ \
+      "movd %%mm1,%[d]\n\t" \
+      "movw %w[d],1(%[pix])\n\t" \
+      "psrlq $32,%%mm1\n\t" \
+      "shr $16,%[d]\n\t" \
+      "movw %w[d],1(%[pix],%[ystride])\n\t" \
+      /*[d]=c3 b3 c2 b2*/ \
+      "movd %%mm1,%[d]\n\t" \
+      "movw %w[d],1(%[pix],%[ystride],2)\n\t" \
+      "shr $16,%[d]\n\t" \
+      "movw %w[d],1(%[pix],%[ystride3])\n\t" \
+      "lea (%[pix],%[ystride],4),%[pix]\n\t" \
+      /*[d]=c5 b5 c4 b4*/ \
+      "movd %%mm0,%[d]\n\t" \
+      "movw %w[d],1(%[pix])\n\t" \
+      "psrlq $32,%%mm0\n\t" \
+      "shr $16,%[d]\n\t" \
+      "movw %w[d],1(%[pix],%[ystride])\n\t" \
+      /*[d]=c7 b7 c6 b6*/ \
+      "movd %%mm0,%[d]\n\t" \
+      "movw %w[d],1(%[pix],%[ystride],2)\n\t" \
+      "shr $16,%[d]\n\t" \
+      "movw %w[d],1(%[pix],%[ystride3])\n\t" \
+      :[pix]"+r"(pix__),[ystride3]"=&r"(ystride3__),[d]"=&r"(d__) \
+      :[ystride]"r"((ptrdiff_t)(_ystride)),[ll]"r"(_ll) \
+      :"memory" \
+    ); \
+  } \
+  while(0)
+
+# endif
+#endif

thirdparty/libtheora/x86/mmxstate.c

@@ -0,0 +1,226 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+/*MMX acceleration of complete fragment reconstruction algorithm.
+  Originally written by Rudolf Marek.*/
+#include <string.h>
+#include "x86int.h"
+#include "mmxloop.h"
+
+#if defined(OC_X86_ASM)
+
+void oc_state_frag_recon_mmx(const oc_theora_state *_state,ptrdiff_t _fragi,
+ int _pli,ogg_int16_t _dct_coeffs[128],int _last_zzi,ogg_uint16_t _dc_quant){
+  unsigned char *dst;
+  ptrdiff_t      frag_buf_off;
+  int            ystride;
+  int            refi;
+  /*Apply the inverse transform.*/
+  /*Special case only having a DC component.*/
+  if(_last_zzi<2){
+    /*Note that this value must be unsigned, to keep the __asm__ block from
+       sign-extending it when it puts it in a register.*/
+    ogg_uint16_t p;
+    int          i;
+    /*We round this dequant product (and not any of the others) because there's
+       no iDCT rounding.*/
+    p=(ogg_int16_t)(_dct_coeffs[0]*(ogg_int32_t)_dc_quant+15>>5);
+    /*Fill _dct_coeffs with p.*/
+    __asm__ __volatile__(
+      /*mm0=0000 0000 0000 AAAA*/
+      "movd %[p],%%mm0\n\t"
+      /*mm0=0000 0000 AAAA AAAA*/
+      "punpcklwd %%mm0,%%mm0\n\t"
+      /*mm0=AAAA AAAA AAAA AAAA*/
+      "punpckldq %%mm0,%%mm0\n\t"
+      :
+      :[p]"r"((unsigned)p)
+    );
+    for(i=0;i<4;i++){
+      __asm__ __volatile__(
+        "movq %%mm0,"OC_MEM_OFFS(0x00,y)"\n\t"
+        "movq %%mm0,"OC_MEM_OFFS(0x08,y)"\n\t"
+        "movq %%mm0,"OC_MEM_OFFS(0x10,y)"\n\t"
+        "movq %%mm0,"OC_MEM_OFFS(0x18,y)"\n\t"
+        :[y]"=m"OC_ARRAY_OPERAND(ogg_int16_t,_dct_coeffs+64+16*i,16)
+      );
+    }
+  }
+  else{
+    /*Dequantize the DC coefficient.*/
+    _dct_coeffs[0]=(ogg_int16_t)(_dct_coeffs[0]*(int)_dc_quant);
+    oc_idct8x8(_state,_dct_coeffs+64,_dct_coeffs,_last_zzi);
+  }
+  /*Fill in the target buffer.*/
+  frag_buf_off=_state->frag_buf_offs[_fragi];
+  refi=_state->frags[_fragi].refi;
+  ystride=_state->ref_ystride[_pli];
+  dst=_state->ref_frame_data[OC_FRAME_SELF]+frag_buf_off;
+  if(refi==OC_FRAME_SELF)oc_frag_recon_intra_mmx(dst,ystride,_dct_coeffs+64);
+  else{
+    const unsigned char *ref;
+    int                  mvoffsets[2];
+    ref=_state->ref_frame_data[refi]+frag_buf_off;
+    if(oc_state_get_mv_offsets(_state,mvoffsets,_pli,
+     _state->frag_mvs[_fragi])>1){
+      oc_frag_recon_inter2_mmx(dst,ref+mvoffsets[0],ref+mvoffsets[1],ystride,
+       _dct_coeffs+64);
+    }
+    else oc_frag_recon_inter_mmx(dst,ref+mvoffsets[0],ystride,_dct_coeffs+64);
+  }
+}
+
+/*We copy these entire function to inline the actual MMX routines so that we
+   use only a single indirect call.*/
+
+void oc_loop_filter_init_mmx(signed char _bv[256],int _flimit){
+  memset(_bv,_flimit,8);
+}
+
+/*Apply the loop filter to a given set of fragment rows in the given plane.
+  The filter may be run on the bottom edge, affecting pixels in the next row of
+   fragments, so this row also needs to be available.
+  _bv:        The bounding values array.
+  _refi:      The index of the frame buffer to filter.
+  _pli:       The color plane to filter.
+  _fragy0:    The Y coordinate of the first fragment row to filter.
+  _fragy_end: The Y coordinate of the fragment row to stop filtering at.*/
+void oc_state_loop_filter_frag_rows_mmx(const oc_theora_state *_state,
+ signed char _bv[256],int _refi,int _pli,int _fragy0,int _fragy_end){
+  OC_ALIGN8(unsigned char   ll[8]);
+  const oc_fragment_plane *fplane;
+  const oc_fragment       *frags;
+  const ptrdiff_t         *frag_buf_offs;
+  unsigned char           *ref_frame_data;
+  ptrdiff_t                fragi_top;
+  ptrdiff_t                fragi_bot;
+  ptrdiff_t                fragi0;
+  ptrdiff_t                fragi0_end;
+  int                      ystride;
+  int                      nhfrags;
+  memset(ll,_state->loop_filter_limits[_state->qis[0]],sizeof(ll));
+  fplane=_state->fplanes+_pli;
+  nhfrags=fplane->nhfrags;
+  fragi_top=fplane->froffset;
+  fragi_bot=fragi_top+fplane->nfrags;
+  fragi0=fragi_top+_fragy0*(ptrdiff_t)nhfrags;
+  fragi0_end=fragi0+(_fragy_end-_fragy0)*(ptrdiff_t)nhfrags;
+  ystride=_state->ref_ystride[_pli];
+  frags=_state->frags;
+  frag_buf_offs=_state->frag_buf_offs;
+  ref_frame_data=_state->ref_frame_data[_refi];
+  /*The following loops are constructed somewhat non-intuitively on purpose.
+    The main idea is: if a block boundary has at least one coded fragment on
+     it, the filter is applied to it.
+    However, the order that the filters are applied in matters, and VP3 chose
+     the somewhat strange ordering used below.*/
+  while(fragi0<fragi0_end){
+    ptrdiff_t fragi;
+    ptrdiff_t fragi_end;
+    fragi=fragi0;
+    fragi_end=fragi+nhfrags;
+    while(fragi<fragi_end){
+      if(frags[fragi].coded){
+        unsigned char *ref;
+        ref=ref_frame_data+frag_buf_offs[fragi];
+        if(fragi>fragi0){
+          OC_LOOP_FILTER_H(OC_LOOP_FILTER8_MMX,ref,ystride,ll);
+        }
+        if(fragi0>fragi_top){
+          OC_LOOP_FILTER_V(OC_LOOP_FILTER8_MMX,ref,ystride,ll);
+        }
+        if(fragi+1<fragi_end&&!frags[fragi+1].coded){
+          OC_LOOP_FILTER_H(OC_LOOP_FILTER8_MMX,ref+8,ystride,ll);
+        }
+        if(fragi+nhfrags<fragi_bot&&!frags[fragi+nhfrags].coded){
+          OC_LOOP_FILTER_V(OC_LOOP_FILTER8_MMX,ref+(ystride*8),ystride,ll);
+        }
+      }
+      fragi++;
+    }
+    fragi0+=nhfrags;
+  }
+}
+
+void oc_loop_filter_init_mmxext(signed char _bv[256],int _flimit){
+  memset(_bv,~(_flimit<<1),8);
+}
+
+/*Apply the loop filter to a given set of fragment rows in the given plane.
+  The filter may be run on the bottom edge, affecting pixels in the next row of
+   fragments, so this row also needs to be available.
+  _bv:        The bounding values array.
+  _refi:      The index of the frame buffer to filter.
+  _pli:       The color plane to filter.
+  _fragy0:    The Y coordinate of the first fragment row to filter.
+  _fragy_end: The Y coordinate of the fragment row to stop filtering at.*/
+void oc_state_loop_filter_frag_rows_mmxext(const oc_theora_state *_state,
+ signed char _bv[256],int _refi,int _pli,int _fragy0,int _fragy_end){
+  const oc_fragment_plane *fplane;
+  const oc_fragment       *frags;
+  const ptrdiff_t         *frag_buf_offs;
+  unsigned char           *ref_frame_data;
+  ptrdiff_t                fragi_top;
+  ptrdiff_t                fragi_bot;
+  ptrdiff_t                fragi0;
+  ptrdiff_t                fragi0_end;
+  int                      ystride;
+  int                      nhfrags;
+  fplane=_state->fplanes+_pli;
+  nhfrags=fplane->nhfrags;
+  fragi_top=fplane->froffset;
+  fragi_bot=fragi_top+fplane->nfrags;
+  fragi0=fragi_top+_fragy0*(ptrdiff_t)nhfrags;
+  fragi0_end=fragi_top+_fragy_end*(ptrdiff_t)nhfrags;
+  ystride=_state->ref_ystride[_pli];
+  frags=_state->frags;
+  frag_buf_offs=_state->frag_buf_offs;
+  ref_frame_data=_state->ref_frame_data[_refi];
+  /*The following loops are constructed somewhat non-intuitively on purpose.
+    The main idea is: if a block boundary has at least one coded fragment on
+     it, the filter is applied to it.
+    However, the order that the filters are applied in matters, and VP3 chose
+     the somewhat strange ordering used below.*/
+  while(fragi0<fragi0_end){
+    ptrdiff_t fragi;
+    ptrdiff_t fragi_end;
+    fragi=fragi0;
+    fragi_end=fragi+nhfrags;
+    while(fragi<fragi_end){
+      if(frags[fragi].coded){
+        unsigned char *ref;
+        ref=ref_frame_data+frag_buf_offs[fragi];
+        if(fragi>fragi0){
+          OC_LOOP_FILTER_H(OC_LOOP_FILTER8_MMXEXT,ref,ystride,_bv);
+        }
+        if(fragi0>fragi_top){
+          OC_LOOP_FILTER_V(OC_LOOP_FILTER8_MMXEXT,ref,ystride,_bv);
+        }
+        if(fragi+1<fragi_end&&!frags[fragi+1].coded){
+          OC_LOOP_FILTER_H(OC_LOOP_FILTER8_MMXEXT,ref+8,ystride,_bv);
+        }
+        if(fragi+nhfrags<fragi_bot&&!frags[fragi+nhfrags].coded){
+          OC_LOOP_FILTER_V(OC_LOOP_FILTER8_MMXEXT,ref+(ystride*8),ystride,_bv);
+        }
+      }
+      fragi++;
+    }
+    fragi0+=nhfrags;
+  }
+}
+
+#endif

thirdparty/libtheora/x86/sse2encfrag.c

@@ -0,0 +1,500 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation https://www.xiph.org/                 *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+#include <stddef.h>
+#include "x86enc.h"
+#include "sse2trans.h"
+
+#if defined(OC_X86_ASM)
+
+/*Load a 4x8 array of pixels values from %[src] and %[ref] and compute their
+   16-bit differences.
+  On output, these are stored in _m0, xmm1, xmm2, and xmm3.
+  xmm4 and xmm5 are clobbered.*/
+#define OC_LOAD_SUB_4x8(_m0) \
+ "#OC_LOAD_SUB_4x8\n\t" \
+ /*Load the first three rows.*/ \
+ "movq (%[src]),"_m0"\n\t" \
+ "movq (%[ref]),%%xmm4\n\t" \
+ "movq (%[src],%[ystride]),%%xmm1\n\t" \
+ "movq (%[ref],%[ystride]),%%xmm3\n\t" \
+ "movq (%[src],%[ystride],2),%%xmm2\n\t" \
+ "movq (%[ref],%[ystride],2),%%xmm5\n\t" \
+ /*Unpack and subtract.*/ \
+ "punpcklbw %%xmm4,"_m0"\n\t" \
+ "punpcklbw %%xmm4,%%xmm4\n\t" \
+ "punpcklbw %%xmm3,%%xmm1\n\t" \
+ "punpcklbw %%xmm3,%%xmm3\n\t" \
+ "psubw %%xmm4,"_m0"\n\t" \
+ "psubw %%xmm3,%%xmm1\n\t" \
+ /*Load the last row.*/ \
+ "movq (%[src],%[ystride3]),%%xmm3\n\t" \
+ "movq (%[ref],%[ystride3]),%%xmm4\n\t" \
+ /*Unpack, subtract, and advance the pointers.*/ \
+ "punpcklbw %%xmm5,%%xmm2\n\t" \
+ "punpcklbw %%xmm5,%%xmm5\n\t" \
+ "lea (%[src],%[ystride],4),%[src]\n\t" \
+ "psubw %%xmm5,%%xmm2\n\t" \
+ "punpcklbw %%xmm4,%%xmm3\n\t" \
+ "punpcklbw %%xmm4,%%xmm4\n\t" \
+ "lea (%[ref],%[ystride],4),%[ref]\n\t" \
+ "psubw %%xmm4,%%xmm3\n\t" \
+
+/*Square and accumulate four rows of differences in _m0, xmm1, xmm2, and xmm3.
+  On output, xmm0 contains the sum of two of the rows, and the other two are
+   added to xmm7.*/
+#define OC_SSD_4x8(_m0) \
+ "pmaddwd "_m0","_m0"\n\t" \
+ "pmaddwd %%xmm1,%%xmm1\n\t" \
+ "pmaddwd %%xmm2,%%xmm2\n\t" \
+ "pmaddwd %%xmm3,%%xmm3\n\t" \
+ "paddd %%xmm1,"_m0"\n\t" \
+ "paddd %%xmm3,%%xmm2\n\t" \
+ "paddd %%xmm2,%%xmm7\n\t" \
+
+unsigned oc_enc_frag_ssd_sse2(const unsigned char *_src,
+ const unsigned char *_ref,int _ystride){
+  unsigned ret;
+  __asm__ __volatile__(
+    OC_LOAD_SUB_4x8("%%xmm7")
+    OC_SSD_4x8("%%xmm7")
+    OC_LOAD_SUB_4x8("%%xmm0")
+    OC_SSD_4x8("%%xmm0")
+    "paddd %%xmm0,%%xmm7\n\t"
+    "movdqa %%xmm7,%%xmm6\n\t"
+    "punpckhqdq %%xmm7,%%xmm7\n\t"
+    "paddd %%xmm6,%%xmm7\n\t"
+    "pshufd $1,%%xmm7,%%xmm6\n\t"
+    "paddd %%xmm6,%%xmm7\n\t"
+    "movd %%xmm7,%[ret]\n\t"
+    :[ret]"=a"(ret)
+    :[src]"r"(_src),[ref]"r"(_ref),[ystride]"r"((ptrdiff_t)_ystride),
+     [ystride3]"r"((ptrdiff_t)_ystride*3)
+  );
+  return ret;
+}
+
+static const unsigned char __attribute__((aligned(16))) OC_MASK_CONSTS[8]={
+  0x01,0x02,0x04,0x08,0x10,0x20,0x40,0x80
+};
+
+/*Load a 2x8 array of pixels values from %[src] and %[ref] and compute their
+   horizontal sums as well as their 16-bit differences subject to a mask.
+  %%xmm5 must contain OC_MASK_CONSTS[0...7] and %%xmm6 must contain 0.*/
+#define OC_LOAD_SUB_MASK_2x8 \
+ "#OC_LOAD_SUB_MASK_2x8\n\t" \
+ /*Start the loads and expand the next 8 bits of the mask.*/ \
+ "shl $8,%[m]\n\t" \
+ "movq (%[src]),%%xmm0\n\t" \
+ "mov %h[m],%b[m]\n\t" \
+ "movq (%[ref]),%%xmm2\n\t" \
+ "movd %[m],%%xmm4\n\t" \
+ "shr $8,%[m]\n\t" \
+ "pshuflw $0x00,%%xmm4,%%xmm4\n\t" \
+ "mov %h[m],%b[m]\n\t" \
+ "pand %%xmm6,%%xmm4\n\t" \
+ "pcmpeqb %%xmm6,%%xmm4\n\t" \
+ /*Perform the masking.*/ \
+ "pand %%xmm4,%%xmm0\n\t" \
+ "pand %%xmm4,%%xmm2\n\t" \
+ /*Finish the loads while unpacking the first set of rows, and expand the next
+    8 bits of the mask.*/ \
+ "movd %[m],%%xmm4\n\t" \
+ "movq (%[src],%[ystride]),%%xmm1\n\t" \
+ "pshuflw $0x00,%%xmm4,%%xmm4\n\t" \
+ "movq (%[ref],%[ystride]),%%xmm3\n\t" \
+ "pand %%xmm6,%%xmm4\n\t" \
+ "punpcklbw %%xmm2,%%xmm0\n\t" \
+ "pcmpeqb %%xmm6,%%xmm4\n\t" \
+ "punpcklbw %%xmm2,%%xmm2\n\t" \
+ /*Mask and unpack the second set of rows.*/ \
+ "pand %%xmm4,%%xmm1\n\t" \
+ "pand %%xmm4,%%xmm3\n\t" \
+ "punpcklbw %%xmm3,%%xmm1\n\t" \
+ "punpcklbw %%xmm3,%%xmm3\n\t" \
+ "psubw %%xmm2,%%xmm0\n\t" \
+ "psubw %%xmm3,%%xmm1\n\t" \
+
+unsigned oc_enc_frag_border_ssd_sse2(const unsigned char *_src,
+ const unsigned char *_ref,int _ystride,ogg_int64_t _mask){
+  ptrdiff_t ystride;
+  unsigned  ret;
+  int       i;
+  ystride=_ystride;
+  __asm__ __volatile__(
+    "pxor %%xmm7,%%xmm7\n\t"
+    "movq %[c],%%xmm6\n\t"
+    :
+    :[c]"m"(OC_CONST_ARRAY_OPERAND(unsigned char,OC_MASK_CONSTS,8))
+  );
+  for(i=0;i<4;i++){
+    unsigned m;
+    m=_mask&0xFFFF;
+    _mask>>=16;
+    if(m){
+      __asm__ __volatile__(
+        OC_LOAD_SUB_MASK_2x8
+        "pmaddwd %%xmm0,%%xmm0\n\t"
+        "pmaddwd %%xmm1,%%xmm1\n\t"
+        "paddd %%xmm0,%%xmm7\n\t"
+        "paddd %%xmm1,%%xmm7\n\t"
+        :[src]"+r"(_src),[ref]"+r"(_ref),[ystride]"+r"(ystride),[m]"+Q"(m)
+      );
+    }
+    _src+=2*ystride;
+    _ref+=2*ystride;
+  }
+  __asm__ __volatile__(
+    "movdqa %%xmm7,%%xmm6\n\t"
+    "punpckhqdq %%xmm7,%%xmm7\n\t"
+    "paddd %%xmm6,%%xmm7\n\t"
+    "pshufd $1,%%xmm7,%%xmm6\n\t"
+    "paddd %%xmm6,%%xmm7\n\t"
+    "movd %%xmm7,%[ret]\n\t"
+    :[ret]"=a"(ret)
+  );
+  return ret;
+}
+
+
+/*Load an 8x8 array of pixel values from %[src] and %[ref] and compute their
+   16-bit difference in %%xmm0...%%xmm7.*/
+#define OC_LOAD_SUB_8x8 \
+ "#OC_LOAD_SUB_8x8\n\t" \
+ "movq (%[src]),%%xmm0\n\t" \
+ "movq (%[ref]),%%xmm4\n\t" \
+ "movq (%[src],%[src_ystride]),%%xmm1\n\t" \
+ "lea (%[src],%[src_ystride],2),%[src]\n\t" \
+ "movq (%[ref],%[ref_ystride]),%%xmm5\n\t" \
+ "lea (%[ref],%[ref_ystride],2),%[ref]\n\t" \
+ "movq (%[src]),%%xmm2\n\t" \
+ "movq (%[ref]),%%xmm7\n\t" \
+ "movq (%[src],%[src_ystride]),%%xmm3\n\t" \
+ "movq (%[ref],%[ref_ystride]),%%xmm6\n\t" \
+ "punpcklbw %%xmm4,%%xmm0\n\t" \
+ "lea (%[src],%[src_ystride],2),%[src]\n\t" \
+ "punpcklbw %%xmm4,%%xmm4\n\t" \
+ "lea (%[ref],%[ref_ystride],2),%[ref]\n\t" \
+ "psubw %%xmm4,%%xmm0\n\t" \
+ "movq (%[src]),%%xmm4\n\t" \
+ "movdqa %%xmm0,"OC_MEM_OFFS(0x00,buf)"\n\t" \
+ "movq (%[ref]),%%xmm0\n\t" \
+ "punpcklbw %%xmm5,%%xmm1\n\t" \
+ "punpcklbw %%xmm5,%%xmm5\n\t" \
+ "psubw %%xmm5,%%xmm1\n\t" \
+ "movq (%[src],%[src_ystride]),%%xmm5\n\t" \
+ "punpcklbw %%xmm7,%%xmm2\n\t" \
+ "punpcklbw %%xmm7,%%xmm7\n\t" \
+ "psubw %%xmm7,%%xmm2\n\t" \
+ "movq (%[ref],%[ref_ystride]),%%xmm7\n\t" \
+ "punpcklbw %%xmm6,%%xmm3\n\t" \
+ "lea (%[src],%[src_ystride],2),%[src]\n\t" \
+ "punpcklbw %%xmm6,%%xmm6\n\t" \
+ "psubw %%xmm6,%%xmm3\n\t" \
+ "movq (%[src]),%%xmm6\n\t" \
+ "punpcklbw %%xmm0,%%xmm4\n\t" \
+ "lea (%[ref],%[ref_ystride],2),%[ref]\n\t" \
+ "punpcklbw %%xmm0,%%xmm0\n\t" \
+ "lea (%[src],%[src_ystride],2),%[src]\n\t" \
+ "psubw %%xmm0,%%xmm4\n\t" \
+ "movq (%[ref]),%%xmm0\n\t" \
+ "punpcklbw %%xmm7,%%xmm5\n\t" \
+ "neg %[src_ystride]\n\t" \
+ "punpcklbw %%xmm7,%%xmm7\n\t" \
+ "psubw %%xmm7,%%xmm5\n\t" \
+ "movq (%[src],%[src_ystride]),%%xmm7\n\t" \
+ "punpcklbw %%xmm0,%%xmm6\n\t" \
+ "lea (%[ref],%[ref_ystride],2),%[ref]\n\t" \
+ "punpcklbw %%xmm0,%%xmm0\n\t" \
+ "neg %[ref_ystride]\n\t" \
+ "psubw %%xmm0,%%xmm6\n\t" \
+ "movq (%[ref],%[ref_ystride]),%%xmm0\n\t" \
+ "punpcklbw %%xmm0,%%xmm7\n\t" \
+ "punpcklbw %%xmm0,%%xmm0\n\t" \
+ "psubw %%xmm0,%%xmm7\n\t" \
+ "movdqa "OC_MEM_OFFS(0x00,buf)",%%xmm0\n\t" \
+
+/*Load an 8x8 array of pixel values from %[src] into %%xmm0...%%xmm7.*/
+#define OC_LOAD_8x8 \
+ "#OC_LOAD_8x8\n\t" \
+ "movq (%[src]),%%xmm0\n\t" \
+ "movq (%[src],%[ystride]),%%xmm1\n\t" \
+ "movq (%[src],%[ystride],2),%%xmm2\n\t" \
+ "pxor %%xmm7,%%xmm7\n\t" \
+ "movq (%[src],%[ystride3]),%%xmm3\n\t" \
+ "punpcklbw %%xmm7,%%xmm0\n\t" \
+ "movq (%[src4]),%%xmm4\n\t" \
+ "punpcklbw %%xmm7,%%xmm1\n\t" \
+ "movq (%[src4],%[ystride]),%%xmm5\n\t" \
+ "punpcklbw %%xmm7,%%xmm2\n\t" \
+ "movq (%[src4],%[ystride],2),%%xmm6\n\t" \
+ "punpcklbw %%xmm7,%%xmm3\n\t" \
+ "movq (%[src4],%[ystride3]),%%xmm7\n\t" \
+ "punpcklbw %%xmm4,%%xmm4\n\t" \
+ "punpcklbw %%xmm5,%%xmm5\n\t" \
+ "psrlw $8,%%xmm4\n\t" \
+ "psrlw $8,%%xmm5\n\t" \
+ "punpcklbw %%xmm6,%%xmm6\n\t" \
+ "punpcklbw %%xmm7,%%xmm7\n\t" \
+ "psrlw $8,%%xmm6\n\t" \
+ "psrlw $8,%%xmm7\n\t" \
+
+/*Performs the first two stages of an 8-point 1-D Hadamard transform in place.
+  Outputs 1, 3, 4, and 5 from the second stage are negated (which allows us to
+   perform this stage in place with no temporary registers).*/
+#define OC_HADAMARD_AB_8x8 \
+ "#OC_HADAMARD_AB_8x8\n\t" \
+ /*Stage A:*/ \
+ "paddw %%xmm5,%%xmm1\n\t" \
+ "paddw %%xmm6,%%xmm2\n\t" \
+ "paddw %%xmm5,%%xmm5\n\t" \
+ "paddw %%xmm6,%%xmm6\n\t" \
+ "psubw %%xmm1,%%xmm5\n\t" \
+ "psubw %%xmm2,%%xmm6\n\t" \
+ "paddw %%xmm7,%%xmm3\n\t" \
+ "paddw %%xmm4,%%xmm0\n\t" \
+ "paddw %%xmm7,%%xmm7\n\t" \
+ "paddw %%xmm4,%%xmm4\n\t" \
+ "psubw %%xmm3,%%xmm7\n\t" \
+ "psubw %%xmm0,%%xmm4\n\t" \
+ /*Stage B:*/ \
+ "paddw %%xmm2,%%xmm0\n\t" \
+ "paddw %%xmm3,%%xmm1\n\t" \
+ "paddw %%xmm6,%%xmm4\n\t" \
+ "paddw %%xmm7,%%xmm5\n\t" \
+ "paddw %%xmm2,%%xmm2\n\t" \
+ "paddw %%xmm3,%%xmm3\n\t" \
+ "paddw %%xmm6,%%xmm6\n\t" \
+ "paddw %%xmm7,%%xmm7\n\t" \
+ "psubw %%xmm0,%%xmm2\n\t" \
+ "psubw %%xmm1,%%xmm3\n\t" \
+ "psubw %%xmm4,%%xmm6\n\t" \
+ "psubw %%xmm5,%%xmm7\n\t" \
+
+/*Performs the last stage of an 8-point 1-D Hadamard transform in place.
+  Outputs 1, 3, 5, and 7 are negated (which allows us to perform this stage in
+   place with no temporary registers).*/
+#define OC_HADAMARD_C_8x8 \
+ "#OC_HADAMARD_C_8x8\n\t" \
+ /*Stage C:*/ \
+ "paddw %%xmm1,%%xmm0\n\t" \
+ "paddw %%xmm3,%%xmm2\n\t" \
+ "paddw %%xmm5,%%xmm4\n\t" \
+ "paddw %%xmm7,%%xmm6\n\t" \
+ "paddw %%xmm1,%%xmm1\n\t" \
+ "paddw %%xmm3,%%xmm3\n\t" \
+ "paddw %%xmm5,%%xmm5\n\t" \
+ "paddw %%xmm7,%%xmm7\n\t" \
+ "psubw %%xmm0,%%xmm1\n\t" \
+ "psubw %%xmm2,%%xmm3\n\t" \
+ "psubw %%xmm4,%%xmm5\n\t" \
+ "psubw %%xmm6,%%xmm7\n\t" \
+
+/*Performs an 8-point 1-D Hadamard transform in place.
+  Outputs 1, 2, 4, and 7 are negated (which allows us to perform the transform
+   in place with no temporary registers).*/
+#define OC_HADAMARD_8x8 \
+ OC_HADAMARD_AB_8x8 \
+ OC_HADAMARD_C_8x8 \
+
+/*Performs the first part of the final stage of the Hadamard transform and
+   summing of absolute values.
+  At the end of this part, %%xmm1 will contain the DC coefficient of the
+   transform.*/
+#define OC_HADAMARD_C_ABS_ACCUM_A_8x8 \
+ /*We use the fact that \
+     (abs(a+b)+abs(a-b))/2=max(abs(a),abs(b)) \
+    to merge the final butterfly with the abs and the first stage of \
+    accumulation. \
+   Thus we can avoid using pabsw, which is not available until SSSE3. \
+   Emulating pabsw takes 3 instructions, so the straightforward SSE2 \
+    implementation would be (3+3)*8+7=55 instructions (+4 for spilling \
+    registers). \
+   Even with pabsw, it would be (3+1)*8+7=39 instructions (with no spills). \
+   This implementation is only 26 (+4 for spilling registers).*/ \
+ "#OC_HADAMARD_C_ABS_ACCUM_A_8x8\n\t" \
+ "movdqa %%xmm7,"OC_MEM_OFFS(0x10,buf)"\n\t" \
+ "movdqa %%xmm6,"OC_MEM_OFFS(0x00,buf)"\n\t" \
+ /*xmm7={0x7FFF}x4 \
+   xmm4=max(abs(xmm4),abs(xmm5))-0x7FFF*/ \
+ "pcmpeqb %%xmm7,%%xmm7\n\t" \
+ "movdqa %%xmm4,%%xmm6\n\t" \
+ "psrlw $1,%%xmm7\n\t" \
+ "paddw %%xmm5,%%xmm6\n\t" \
+ "pmaxsw %%xmm5,%%xmm4\n\t" \
+ "paddsw %%xmm7,%%xmm6\n\t" \
+ "psubw %%xmm6,%%xmm4\n\t" \
+ /*xmm2=max(abs(xmm2),abs(xmm3))-0x7FFF \
+   xmm0=max(abs(xmm0),abs(xmm1))-0x7FFF*/ \
+ "movdqa %%xmm2,%%xmm6\n\t" \
+ "movdqa %%xmm0,%%xmm5\n\t" \
+ "pmaxsw %%xmm3,%%xmm2\n\t" \
+ "pmaxsw %%xmm1,%%xmm0\n\t" \
+ "paddw %%xmm3,%%xmm6\n\t" \
+ "movdqa "OC_MEM_OFFS(0x10,buf)",%%xmm3\n\t" \
+ "paddw %%xmm5,%%xmm1\n\t" \
+ "movdqa "OC_MEM_OFFS(0x00,buf)",%%xmm5\n\t" \
+
+/*Performs the second part of the final stage of the Hadamard transform and
+   summing of absolute values.*/
+#define OC_HADAMARD_C_ABS_ACCUM_B_8x8 \
+ "#OC_HADAMARD_C_ABS_ACCUM_B_8x8\n\t" \
+ "paddsw %%xmm7,%%xmm6\n\t" \
+ "paddsw %%xmm7,%%xmm1\n\t" \
+ "psubw %%xmm6,%%xmm2\n\t" \
+ "psubw %%xmm1,%%xmm0\n\t" \
+ /*xmm7={1}x4 (needed for the horizontal add that follows) \
+   xmm0+=xmm2+xmm4+max(abs(xmm3),abs(xmm5))-0x7FFF*/ \
+ "movdqa %%xmm3,%%xmm6\n\t" \
+ "pmaxsw %%xmm5,%%xmm3\n\t" \
+ "paddw %%xmm2,%%xmm0\n\t" \
+ "paddw %%xmm5,%%xmm6\n\t" \
+ "paddw %%xmm4,%%xmm0\n\t" \
+ "paddsw %%xmm7,%%xmm6\n\t" \
+ "paddw %%xmm3,%%xmm0\n\t" \
+ "psrlw $14,%%xmm7\n\t" \
+ "psubw %%xmm6,%%xmm0\n\t" \
+
+/*Performs the last stage of an 8-point 1-D Hadamard transform, takes the
+   absolute value of each component, and accumulates everything into xmm0.*/
+#define OC_HADAMARD_C_ABS_ACCUM_8x8 \
+ OC_HADAMARD_C_ABS_ACCUM_A_8x8 \
+ OC_HADAMARD_C_ABS_ACCUM_B_8x8 \
+
+/*Performs an 8-point 1-D Hadamard transform, takes the absolute value of each
+   component, and accumulates everything into xmm0.
+  Note that xmm0 will have an extra 4 added to each column, and that after
+   removing this value, the remainder will be half the conventional value.*/
+#define OC_HADAMARD_ABS_ACCUM_8x8 \
+ OC_HADAMARD_AB_8x8 \
+ OC_HADAMARD_C_ABS_ACCUM_8x8
+
+static unsigned oc_int_frag_satd_sse2(int *_dc,
+ const unsigned char *_src,int _src_ystride,
+ const unsigned char *_ref,int _ref_ystride){
+  OC_ALIGN16(ogg_int16_t buf[16]);
+  unsigned ret;
+  unsigned ret2;
+  int      dc;
+  __asm__ __volatile__(
+    OC_LOAD_SUB_8x8
+    OC_HADAMARD_8x8
+    OC_TRANSPOSE_8x8
+    /*We split out the stages here so we can save the DC coefficient in the
+       middle.*/
+    OC_HADAMARD_AB_8x8
+    OC_HADAMARD_C_ABS_ACCUM_A_8x8
+    "movd %%xmm1,%[dc]\n\t"
+    OC_HADAMARD_C_ABS_ACCUM_B_8x8
+    /*Up to this point, everything fit in 16 bits (8 input + 1 for the
+       difference + 2*3 for the two 8-point 1-D Hadamards - 1 for the abs - 1
+       for the factor of two we dropped + 3 for the vertical accumulation).
+      Now we finally have to promote things to dwords.
+      We break this part out of OC_HADAMARD_ABS_ACCUM_8x8 to hide the long
+       latency of pmaddwd by starting to compute abs(dc) here.*/
+    "pmaddwd %%xmm7,%%xmm0\n\t"
+    "movsx %w[dc],%[dc]\n\t"
+    "cdq\n\t"
+    "movdqa %%xmm0,%%xmm1\n\t"
+    "punpckhqdq %%xmm0,%%xmm0\n\t"
+    "paddd %%xmm1,%%xmm0\n\t"
+    "pshuflw $0xE,%%xmm0,%%xmm1\n\t"
+    "paddd %%xmm1,%%xmm0\n\t"
+    "movd %%xmm0,%[ret]\n\t"
+    /*The sums produced by OC_HADAMARD_ABS_ACCUM_8x8 each have an extra 4
+       added to them, a factor of two removed, and the DC value included;
+       correct the final sum here.*/
+    "lea -64(%[ret2],%[ret],2),%[ret]\n\t"
+    "xor %[dc],%[ret2]\n\t"
+    "sub %[ret2],%[ret]\n\t"
+    /*Although it looks like we're using 7 registers here, gcc can alias %[ret]
+       and %[dc] with some of the inputs, since for once we don't write to
+       them until after we're done using everything but %[buf].*/
+    /*Note that _src_ystride and _ref_ystride must be given non-overlapping
+       constraints, otherwise if gcc can prove they're equal it will allocate
+       them to the same register (which is bad); _src and _ref face a similar
+       problem.
+      All four are destructively modified, but if we list them as output
+       constraints, gcc can't alias them with other outputs.*/
+    :[ret]"=r"(ret),[ret2]"=d"(ret2),[dc]"=a"(dc),
+     [buf]"=m"(OC_ARRAY_OPERAND(ogg_int16_t,buf,16))
+    :[src]"S"(_src),[src_ystride]"c"((ptrdiff_t)_src_ystride),
+     [ref]"a"(_ref),[ref_ystride]"d"((ptrdiff_t)_ref_ystride)
+    /*We have to use neg, so we actually clobber the condition codes for once
+       (not to mention sub, and add).*/
+    :"cc"
+  );
+  *_dc=dc;
+  return ret;
+}
+
+unsigned oc_enc_frag_satd_sse2(int *_dc,const unsigned char *_src,
+ const unsigned char *_ref,int _ystride){
+  return oc_int_frag_satd_sse2(_dc,_src,_ystride,_ref,_ystride);
+}
+
+unsigned oc_enc_frag_satd2_sse2(int *_dc,const unsigned char *_src,
+ const unsigned char *_ref1,const unsigned char *_ref2,int _ystride){
+  OC_ALIGN8(unsigned char ref[64]);
+  oc_int_frag_copy2_mmxext(ref,8,_ref1,_ref2,_ystride);
+  return oc_int_frag_satd_sse2(_dc,_src,_ystride,ref,8);
+}
+
+unsigned oc_enc_frag_intra_satd_sse2(int *_dc,
+ const unsigned char *_src,int _ystride){
+  OC_ALIGN16(ogg_int16_t buf[16]);
+  unsigned ret;
+  int      dc;
+  __asm__ __volatile__(
+    OC_LOAD_8x8
+    OC_HADAMARD_8x8
+    OC_TRANSPOSE_8x8
+    /*We split out the stages here so we can save the DC coefficient in the
+       middle.*/
+    OC_HADAMARD_AB_8x8
+    OC_HADAMARD_C_ABS_ACCUM_A_8x8
+    "movd %%xmm1,%[dc]\n\t"
+    OC_HADAMARD_C_ABS_ACCUM_B_8x8
+    /*Up to this point, everything fit in 16 bits (8 input + 1 for the
+       difference + 2*3 for the two 8-point 1-D Hadamards - 1 for the abs - 1
+       for the factor of two we dropped + 3 for the vertical accumulation).
+      Now we finally have to promote things to dwords.*/
+    "pmaddwd %%xmm7,%%xmm0\n\t"
+    /*We assume that the DC coefficient is always positive (which is true,
+       because the input to the INTRA transform was not a difference).*/
+    "movzx %w[dc],%[dc]\n\t"
+    "movdqa %%xmm0,%%xmm1\n\t"
+    "punpckhqdq %%xmm0,%%xmm0\n\t"
+    "paddd %%xmm1,%%xmm0\n\t"
+    "pshuflw $0xE,%%xmm0,%%xmm1\n\t"
+    "paddd %%xmm1,%%xmm0\n\t"
+    "movd %%xmm0,%[ret]\n\t"
+    "lea -64(%[ret],%[ret]),%[ret]\n\t"
+    "sub %[dc],%[ret]\n\t"
+    /*Although it looks like we're using 7 registers here, gcc can alias %[ret]
+       and %[dc] with some of the inputs, since for once we don't write to
+       them until after we're done using everything but %[buf].*/
+    :[ret]"=a"(ret),[dc]"=r"(dc),
+     [buf]"=m"(OC_ARRAY_OPERAND(ogg_int16_t,buf,16))
+    :[src]"r"(_src),[src4]"r"(_src+4*_ystride),
+     [ystride]"r"((ptrdiff_t)_ystride),[ystride3]"r"((ptrdiff_t)3*_ystride)
+    /*We have to use sub, so we actually clobber the condition codes for once.*/
+    :"cc"
+  );
+  *_dc=dc;
+  return ret;
+}
+
+#endif

thirdparty/libtheora/x86/sse2fdct.c

@@ -0,0 +1,452 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 1999-2006                *
+ * by the Xiph.Org Foundation https://www.xiph.org/                 *
+ *                                                                  *
+ ********************************************************************/
+/*SSE2 fDCT implementation for x86_64.*/
+/*$Id: fdct_ses2.c 14579 2008-03-12 06:42:40Z xiphmont $*/
+#include <stddef.h>
+#include "x86enc.h"
+#include "x86zigzag.h"
+#include "sse2trans.h"
+
+#if defined(OC_X86_64_ASM)
+
+# define OC_FDCT_8x8 \
+ /*Note: xmm15={0}x8 and xmm14={-1}x8.*/ \
+ "#OC_FDCT_8x8\n\t" \
+ /*Stage 1:*/ \
+ "movdqa %%xmm0,%%xmm11\n\t" \
+ "movdqa %%xmm1,%%xmm10\n\t" \
+ "movdqa %%xmm2,%%xmm9\n\t" \
+ "movdqa %%xmm3,%%xmm8\n\t" \
+ /*xmm11=t7'=t0-t7*/ \
+ "psubw %%xmm7,%%xmm11\n\t" \
+ /*xmm10=t6'=t1-t6*/ \
+ "psubw %%xmm6,%%xmm10\n\t" \
+ /*xmm9=t5'=t2-t5*/ \
+ "psubw %%xmm5,%%xmm9\n\t" \
+ /*xmm8=t4'=t3-t4*/ \
+ "psubw %%xmm4,%%xmm8\n\t" \
+ /*xmm0=t0'=t0+t7*/ \
+ "paddw %%xmm7,%%xmm0\n\t" \
+ /*xmm1=t1'=t1+t6*/ \
+ "paddw %%xmm6,%%xmm1\n\t" \
+ /*xmm5=t2'=t2+t5*/ \
+ "paddw %%xmm2,%%xmm5\n\t" \
+ /*xmm4=t3'=t3+t4*/ \
+ "paddw %%xmm3,%%xmm4\n\t" \
+ /*xmm2,3,6,7 are now free.*/ \
+ /*Stage 2:*/ \
+ "movdqa %%xmm0,%%xmm3\n\t" \
+ "mov $0x5A806A0A,%[a]\n\t" \
+ "movdqa %%xmm1,%%xmm2\n\t" \
+ "movd %[a],%%xmm13\n\t" \
+ "movdqa %%xmm10,%%xmm6\n\t" \
+ "pshufd $00,%%xmm13,%%xmm13\n\t" \
+ /*xmm2=t2''=t1'-t2'*/ \
+ "psubw %%xmm5,%%xmm2\n\t" \
+ "pxor %%xmm12,%%xmm12\n\t" \
+ /*xmm3=t3''=t0'-t3'*/ \
+ "psubw %%xmm4,%%xmm3\n\t" \
+ "psubw %%xmm14,%%xmm12\n\t" \
+ /*xmm10=t5''=t6'-t5'*/ \
+ "psubw %%xmm9,%%xmm10\n\t" \
+ "paddw %%xmm12,%%xmm12\n\t" \
+ /*xmm4=t0''=t0'+t3'*/ \
+ "paddw %%xmm0,%%xmm4\n\t" \
+ /*xmm1=t1''=t1'+t2'*/ \
+ "paddw %%xmm5,%%xmm1\n\t" \
+ /*xmm6=t6''=t6'+t5'*/ \
+ "paddw %%xmm9,%%xmm6\n\t" \
+ /*xmm0,xmm5,xmm9 are now free.*/ \
+ /*Stage 3:*/ \
+ /*xmm10:xmm5=t5''*27146+0xB500 \
+   xmm0=t5''*/ \
+ "movdqa %%xmm10,%%xmm5\n\t" \
+ "movdqa %%xmm10,%%xmm0\n\t" \
+ "punpckhwd %%xmm12,%%xmm10\n\t" \
+ "pmaddwd %%xmm13,%%xmm10\n\t" \
+ "punpcklwd %%xmm12,%%xmm5\n\t" \
+ "pmaddwd %%xmm13,%%xmm5\n\t" \
+ /*xmm5=(t5''*27146+0xB500>>16)+t5''*/ \
+ "psrad $16,%%xmm10\n\t" \
+ "psrad $16,%%xmm5\n\t" \
+ "packssdw %%xmm10,%%xmm5\n\t" \
+ "paddw %%xmm0,%%xmm5\n\t" \
+ /*xmm0=s=(t5''*27146+0xB500>>16)+t5''+(t5''!=0)>>1*/ \
+ "pcmpeqw %%xmm15,%%xmm0\n\t" \
+ "psubw %%xmm14,%%xmm0\n\t" \
+ "paddw %%xmm5,%%xmm0\n\t" \
+ "movdqa %%xmm8,%%xmm5\n\t" \
+ "psraw $1,%%xmm0\n\t" \
+ /*xmm5=t5'''=t4'-s*/ \
+ "psubw %%xmm0,%%xmm5\n\t" \
+ /*xmm8=t4''=t4'+s*/ \
+ "paddw %%xmm0,%%xmm8\n\t" \
+ /*xmm0,xmm7,xmm9,xmm10 are free.*/ \
+ /*xmm7:xmm9=t6''*27146+0xB500*/ \
+ "movdqa %%xmm6,%%xmm7\n\t" \
+ "movdqa %%xmm6,%%xmm9\n\t" \
+ "punpckhwd %%xmm12,%%xmm7\n\t" \
+ "pmaddwd %%xmm13,%%xmm7\n\t" \
+ "punpcklwd %%xmm12,%%xmm9\n\t" \
+ "pmaddwd %%xmm13,%%xmm9\n\t" \
+ /*xmm9=(t6''*27146+0xB500>>16)+t6''*/ \
+ "psrad $16,%%xmm7\n\t" \
+ "psrad $16,%%xmm9\n\t" \
+ "packssdw %%xmm7,%%xmm9\n\t" \
+ "paddw %%xmm6,%%xmm9\n\t" \
+ /*xmm9=s=(t6''*27146+0xB500>>16)+t6''+(t6''!=0)>>1*/ \
+ "pcmpeqw %%xmm15,%%xmm6\n\t" \
+ "psubw %%xmm14,%%xmm6\n\t" \
+ "paddw %%xmm6,%%xmm9\n\t" \
+ "movdqa %%xmm11,%%xmm7\n\t" \
+ "psraw $1,%%xmm9\n\t" \
+ /*xmm7=t6'''=t7'-s*/ \
+ "psubw %%xmm9,%%xmm7\n\t" \
+ /*xmm9=t7''=t7'+s*/ \
+ "paddw %%xmm11,%%xmm9\n\t" \
+ /*xmm0,xmm6,xmm10,xmm11 are free.*/ \
+ /*Stage 4:*/ \
+ /*xmm10:xmm0=t1''*27146+0xB500*/ \
+ "movdqa %%xmm1,%%xmm0\n\t" \
+ "movdqa %%xmm1,%%xmm10\n\t" \
+ "punpcklwd %%xmm12,%%xmm0\n\t" \
+ "pmaddwd %%xmm13,%%xmm0\n\t" \
+ "punpckhwd %%xmm12,%%xmm10\n\t" \
+ "pmaddwd %%xmm13,%%xmm10\n\t" \
+ /*xmm0=(t1''*27146+0xB500>>16)+t1''*/ \
+ "psrad $16,%%xmm0\n\t" \
+ "psrad $16,%%xmm10\n\t" \
+ "mov $0x20006A0A,%[a]\n\t" \
+ "packssdw %%xmm10,%%xmm0\n\t" \
+ "movd %[a],%%xmm13\n\t" \
+ "paddw %%xmm1,%%xmm0\n\t" \
+ /*xmm0=s=(t1''*27146+0xB500>>16)+t1''+(t1''!=0)*/ \
+ "pcmpeqw %%xmm15,%%xmm1\n\t" \
+ "pshufd $00,%%xmm13,%%xmm13\n\t" \
+ "psubw %%xmm14,%%xmm1\n\t" \
+ "paddw %%xmm1,%%xmm0\n\t" \
+ /*xmm10:xmm4=t0''*27146+0x4000*/ \
+ "movdqa %%xmm4,%%xmm1\n\t" \
+ "movdqa %%xmm4,%%xmm10\n\t" \
+ "punpcklwd %%xmm12,%%xmm4\n\t" \
+ "pmaddwd %%xmm13,%%xmm4\n\t" \
+ "punpckhwd %%xmm12,%%xmm10\n\t" \
+ "pmaddwd %%xmm13,%%xmm10\n\t" \
+ /*xmm4=(t0''*27146+0x4000>>16)+t0''*/ \
+ "psrad $16,%%xmm4\n\t" \
+ "psrad $16,%%xmm10\n\t" \
+ "mov $0x6CB7,%[a]\n\t" \
+ "packssdw %%xmm10,%%xmm4\n\t" \
+ "movd %[a],%%xmm12\n\t" \
+ "paddw %%xmm1,%%xmm4\n\t" \
+ /*xmm4=r=(t0''*27146+0x4000>>16)+t0''+(t0''!=0)*/ \
+ "pcmpeqw %%xmm15,%%xmm1\n\t" \
+ "pshufd $00,%%xmm12,%%xmm12\n\t" \
+ "psubw %%xmm14,%%xmm1\n\t" \
+ "mov $0x7FFF6C84,%[a]\n\t" \
+ "paddw %%xmm1,%%xmm4\n\t" \
+ /*xmm0=_y[0]=u=r+s>>1 \
+   The naive implementation could cause overflow, so we use \
+    u=(r&s)+((r^s)>>1).*/ \
+ "movdqa %%xmm0,%%xmm6\n\t" \
+ "pxor %%xmm4,%%xmm0\n\t" \
+ "pand %%xmm4,%%xmm6\n\t" \
+ "psraw $1,%%xmm0\n\t" \
+ "movd %[a],%%xmm13\n\t" \
+ "paddw %%xmm6,%%xmm0\n\t" \
+ /*xmm4=_y[4]=v=r-u*/ \
+ "pshufd $00,%%xmm13,%%xmm13\n\t" \
+ "psubw %%xmm0,%%xmm4\n\t" \
+ /*xmm1,xmm6,xmm10,xmm11 are free.*/ \
+ /*xmm6:xmm10=60547*t3''+0x6CB7*/ \
+ "movdqa %%xmm3,%%xmm10\n\t" \
+ "movdqa %%xmm3,%%xmm6\n\t" \
+ "punpcklwd %%xmm3,%%xmm10\n\t" \
+ "pmaddwd %%xmm13,%%xmm10\n\t" \
+ "mov $0x61F861F8,%[a]\n\t" \
+ "punpckhwd %%xmm3,%%xmm6\n\t" \
+ "pmaddwd %%xmm13,%%xmm6\n\t" \
+ "movd %[a],%%xmm13\n\t" \
+ "paddd %%xmm12,%%xmm10\n\t" \
+ "pshufd $00,%%xmm13,%%xmm13\n\t" \
+ "paddd %%xmm12,%%xmm6\n\t" \
+ /*xmm1:xmm2=25080*t2'' \
+   xmm12=t2''*/ \
+ "movdqa %%xmm2,%%xmm11\n\t" \
+ "movdqa %%xmm2,%%xmm12\n\t" \
+ "pmullw %%xmm13,%%xmm2\n\t" \
+ "pmulhw %%xmm13,%%xmm11\n\t" \
+ "movdqa %%xmm2,%%xmm1\n\t" \
+ "punpcklwd %%xmm11,%%xmm2\n\t" \
+ "punpckhwd %%xmm11,%%xmm1\n\t" \
+ /*xmm10=u=(25080*t2''+60547*t3''+0x6CB7>>16)+(t3''!=0)*/ \
+ "paddd %%xmm2,%%xmm10\n\t" \
+ "paddd %%xmm1,%%xmm6\n\t" \
+ "psrad $16,%%xmm10\n\t" \
+ "pcmpeqw %%xmm15,%%xmm3\n\t" \
+ "psrad $16,%%xmm6\n\t" \
+ "psubw %%xmm14,%%xmm3\n\t" \
+ "packssdw %%xmm6,%%xmm10\n\t" \
+ "paddw %%xmm3,%%xmm10\n\t" \
+ /*xmm2=_y[2]=u \
+   xmm10=s=(25080*u>>16)-t2''*/ \
+ "movdqa %%xmm10,%%xmm2\n\t" \
+ "pmulhw %%xmm13,%%xmm10\n\t" \
+ "psubw %%xmm12,%%xmm10\n\t" \
+ /*xmm1:xmm6=s*21600+0x2800*/ \
+ "pxor %%xmm12,%%xmm12\n\t" \
+ "psubw %%xmm14,%%xmm12\n\t" \
+ "mov $0x28005460,%[a]\n\t" \
+ "movd %[a],%%xmm13\n\t" \
+ "pshufd $00,%%xmm13,%%xmm13\n\t" \
+ "movdqa %%xmm10,%%xmm6\n\t" \
+ "movdqa %%xmm10,%%xmm1\n\t" \
+ "punpcklwd %%xmm12,%%xmm6\n\t" \
+ "pmaddwd %%xmm13,%%xmm6\n\t" \
+ "mov $0x0E3D,%[a]\n\t" \
+ "punpckhwd %%xmm12,%%xmm1\n\t" \
+ "pmaddwd %%xmm13,%%xmm1\n\t" \
+ /*xmm6=(s*21600+0x2800>>18)+s*/ \
+ "psrad $18,%%xmm6\n\t" \
+ "psrad $18,%%xmm1\n\t" \
+ "movd %[a],%%xmm12\n\t" \
+ "packssdw %%xmm1,%%xmm6\n\t" \
+ "pshufd $00,%%xmm12,%%xmm12\n\t" \
+ "paddw %%xmm10,%%xmm6\n\t" \
+ /*xmm6=_y[6]=v=(s*21600+0x2800>>18)+s+(s!=0)*/ \
+ "mov $0x7FFF54DC,%[a]\n\t" \
+ "pcmpeqw %%xmm15,%%xmm10\n\t" \
+ "movd %[a],%%xmm13\n\t" \
+ "psubw %%xmm14,%%xmm10\n\t" \
+ "pshufd $00,%%xmm13,%%xmm13\n\t" \
+ "paddw %%xmm10,%%xmm6\n\t " \
+ /*xmm1,xmm3,xmm10,xmm11 are free.*/ \
+ /*xmm11:xmm10=54491*t5'''+0x0E3D*/ \
+ "movdqa %%xmm5,%%xmm10\n\t" \
+ "movdqa %%xmm5,%%xmm11\n\t" \
+ "punpcklwd %%xmm5,%%xmm10\n\t" \
+ "pmaddwd %%xmm13,%%xmm10\n\t" \
+ "mov $0x8E3A8E3A,%[a]\n\t" \
+ "punpckhwd %%xmm5,%%xmm11\n\t" \
+ "pmaddwd %%xmm13,%%xmm11\n\t" \
+ "movd %[a],%%xmm13\n\t" \
+ "paddd %%xmm12,%%xmm10\n\t" \
+ "pshufd $00,%%xmm13,%%xmm13\n\t" \
+ "paddd %%xmm12,%%xmm11\n\t" \
+ /*xmm7:xmm12=36410*t6''' \
+   xmm1=t6'''*/ \
+ "movdqa %%xmm7,%%xmm3\n\t" \
+ "movdqa %%xmm7,%%xmm1\n\t" \
+ "pmulhw %%xmm13,%%xmm3\n\t" \
+ "pmullw %%xmm13,%%xmm7\n\t" \
+ "paddw %%xmm1,%%xmm3\n\t" \
+ "movdqa %%xmm7,%%xmm12\n\t" \
+ "punpckhwd %%xmm3,%%xmm7\n\t" \
+ "punpcklwd %%xmm3,%%xmm12\n\t" \
+ /*xmm10=u=(54491*t5'''+36410*t6'''+0x0E3D>>16)+(t5'''!=0)*/ \
+ "paddd %%xmm12,%%xmm10\n\t" \
+ "paddd %%xmm7,%%xmm11\n\t" \
+ "psrad $16,%%xmm10\n\t" \
+ "pcmpeqw %%xmm15,%%xmm5\n\t" \
+ "psrad $16,%%xmm11\n\t" \
+ "psubw %%xmm14,%%xmm5\n\t" \
+ "packssdw %%xmm11,%%xmm10\n\t" \
+ "pxor %%xmm12,%%xmm12\n\t" \
+ "paddw %%xmm5,%%xmm10\n\t" \
+ /*xmm5=_y[5]=u \
+   xmm1=s=t6'''-(36410*u>>16)*/ \
+ "psubw %%xmm14,%%xmm12\n\t" \
+ "movdqa %%xmm10,%%xmm5\n\t" \
+ "mov $0x340067C8,%[a]\n\t" \
+ "pmulhw %%xmm13,%%xmm10\n\t" \
+ "movd %[a],%%xmm13\n\t" \
+ "paddw %%xmm5,%%xmm10\n\t" \
+ "pshufd $00,%%xmm13,%%xmm13\n\t" \
+ "psubw %%xmm10,%%xmm1\n\t" \
+ /*xmm11:xmm3=s*26568+0x3400*/ \
+ "movdqa %%xmm1,%%xmm3\n\t" \
+ "movdqa %%xmm1,%%xmm11\n\t" \
+ "punpcklwd %%xmm12,%%xmm3\n\t" \
+ "pmaddwd %%xmm13,%%xmm3\n\t" \
+ "mov $0x7B1B,%[a]\n\t" \
+ "punpckhwd %%xmm12,%%xmm11\n\t" \
+ "pmaddwd %%xmm13,%%xmm11\n\t" \
+ /*xmm3=(s*26568+0x3400>>17)+s*/ \
+ "psrad $17,%%xmm3\n\t" \
+ "psrad $17,%%xmm11\n\t" \
+ "movd %[a],%%xmm12\n\t" \
+ "packssdw %%xmm11,%%xmm3\n\t" \
+ "pshufd $00,%%xmm12,%%xmm12\n\t" \
+ "paddw %%xmm1,%%xmm3\n\t" \
+ /*xmm3=_y[3]=v=(s*26568+0x3400>>17)+s+(s!=0)*/ \
+ "mov $0x7FFF7B16,%[a]\n\t" \
+ "pcmpeqw %%xmm15,%%xmm1\n\t" \
+ "movd %[a],%%xmm13\n\t" \
+ "psubw %%xmm14,%%xmm1\n\t" \
+ "pshufd $00,%%xmm13,%%xmm13\n\t" \
+ "paddw %%xmm1,%%xmm3\n\t " \
+ /*xmm1,xmm7,xmm10,xmm11 are free.*/ \
+ /*xmm11:xmm10=64277*t7''+0x7B1B*/ \
+ "movdqa %%xmm9,%%xmm10\n\t" \
+ "movdqa %%xmm9,%%xmm11\n\t" \
+ "punpcklwd %%xmm9,%%xmm10\n\t" \
+ "pmaddwd %%xmm13,%%xmm10\n\t" \
+ "mov $0x31F131F1,%[a]\n\t" \
+ "punpckhwd %%xmm9,%%xmm11\n\t" \
+ "pmaddwd %%xmm13,%%xmm11\n\t" \
+ "movd %[a],%%xmm13\n\t" \
+ "paddd %%xmm12,%%xmm10\n\t" \
+ "pshufd $00,%%xmm13,%%xmm13\n\t" \
+ "paddd %%xmm12,%%xmm11\n\t" \
+ /*xmm12:xmm7=12785*t4''*/ \
+ "movdqa %%xmm8,%%xmm7\n\t" \
+ "movdqa %%xmm8,%%xmm1\n\t" \
+ "pmullw %%xmm13,%%xmm7\n\t" \
+ "pmulhw %%xmm13,%%xmm1\n\t" \
+ "movdqa %%xmm7,%%xmm12\n\t" \
+ "punpcklwd %%xmm1,%%xmm7\n\t" \
+ "punpckhwd %%xmm1,%%xmm12\n\t" \
+ /*xmm10=u=(12785*t4''+64277*t7''+0x7B1B>>16)+(t7''!=0)*/ \
+ "paddd %%xmm7,%%xmm10\n\t" \
+ "paddd %%xmm12,%%xmm11\n\t" \
+ "psrad $16,%%xmm10\n\t" \
+ "pcmpeqw %%xmm15,%%xmm9\n\t" \
+ "psrad $16,%%xmm11\n\t" \
+ "psubw %%xmm14,%%xmm9\n\t" \
+ "packssdw %%xmm11,%%xmm10\n\t" \
+ "pxor %%xmm12,%%xmm12\n\t" \
+ "paddw %%xmm9,%%xmm10\n\t" \
+ /*xmm1=_y[1]=u \
+   xmm10=s=(12785*u>>16)-t4''*/ \
+ "psubw %%xmm14,%%xmm12\n\t" \
+ "movdqa %%xmm10,%%xmm1\n\t" \
+ "mov $0x3000503B,%[a]\n\t" \
+ "pmulhw %%xmm13,%%xmm10\n\t" \
+ "movd %[a],%%xmm13\n\t" \
+ "psubw %%xmm8,%%xmm10\n\t" \
+ "pshufd $00,%%xmm13,%%xmm13\n\t" \
+ /*xmm8:xmm7=s*20539+0x3000*/ \
+ "movdqa %%xmm10,%%xmm7\n\t" \
+ "movdqa %%xmm10,%%xmm8\n\t" \
+ "punpcklwd %%xmm12,%%xmm7\n\t" \
+ "pmaddwd %%xmm13,%%xmm7\n\t" \
+ "punpckhwd %%xmm12,%%xmm8\n\t" \
+ "pmaddwd %%xmm13,%%xmm8\n\t" \
+ /*xmm7=(s*20539+0x3000>>20)+s*/ \
+ "psrad $20,%%xmm7\n\t" \
+ "psrad $20,%%xmm8\n\t" \
+ "packssdw %%xmm8,%%xmm7\n\t" \
+ "paddw %%xmm10,%%xmm7\n\t" \
+ /*xmm7=_y[7]=v=(s*20539+0x3000>>20)+s+(s!=0)*/ \
+ "pcmpeqw %%xmm15,%%xmm10\n\t" \
+ "psubw %%xmm14,%%xmm10\n\t" \
+ "paddw %%xmm10,%%xmm7\n\t " \
+
+/*SSE2 implementation of the fDCT for x86-64 only.
+  Because of the 8 extra XMM registers on x86-64, this version can operate
+   without any temporary stack access at all.*/
+void oc_enc_fdct8x8_x86_64sse2(ogg_int16_t _y[64],const ogg_int16_t _x[64]){
+  ptrdiff_t a;
+  __asm__ __volatile__(
+    /*Load the input.*/
+    "movdqa 0x00(%[x]),%%xmm0\n\t"
+    "movdqa 0x10(%[x]),%%xmm1\n\t"
+    "movdqa 0x20(%[x]),%%xmm2\n\t"
+    "movdqa 0x30(%[x]),%%xmm3\n\t"
+    "movdqa 0x40(%[x]),%%xmm4\n\t"
+    "movdqa 0x50(%[x]),%%xmm5\n\t"
+    "movdqa 0x60(%[x]),%%xmm6\n\t"
+    "movdqa 0x70(%[x]),%%xmm7\n\t"
+    /*Add two extra bits of working precision to improve accuracy; any more and
+       we could overflow.*/
+    /*We also add a few biases to correct for some systematic error that
+       remains in the full fDCT->iDCT round trip.*/
+    /*xmm15={0}x8*/
+    "pxor %%xmm15,%%xmm15\n\t"
+    /*xmm14={-1}x8*/
+    "pcmpeqb %%xmm14,%%xmm14\n\t"
+    "psllw $2,%%xmm0\n\t"
+    /*xmm8=xmm0*/
+    "movdqa %%xmm0,%%xmm8\n\t"
+    "psllw $2,%%xmm1\n\t"
+    /*xmm8={_x[7...0]==0}*/
+    "pcmpeqw %%xmm15,%%xmm8\n\t"
+    "psllw $2,%%xmm2\n\t"
+    /*xmm8={_x[7...0]!=0}*/
+    "psubw %%xmm14,%%xmm8\n\t"
+    "psllw $2,%%xmm3\n\t"
+    /*%[a]=1*/
+    "mov $1,%[a]\n\t"
+    /*xmm8={_x[6]!=0,0,_x[4]!=0,0,_x[2]!=0,0,_x[0]!=0,0}*/
+    "pslld $16,%%xmm8\n\t"
+    "psllw $2,%%xmm4\n\t"
+    /*xmm9={0,0,0,0,0,0,0,1}*/
+    "movd %[a],%%xmm9\n\t"
+    /*xmm8={0,0,_x[2]!=0,0,_x[0]!=0,0}*/
+    "pshufhw $0x00,%%xmm8,%%xmm8\n\t"
+    "psllw $2,%%xmm5\n\t"
+    /*%[a]={1}x2*/
+    "mov $0x10001,%[a]\n\t"
+    /*xmm8={0,0,0,0,0,0,0,_x[0]!=0}*/
+    "pshuflw $0x01,%%xmm8,%%xmm8\n\t"
+    "psllw $2,%%xmm6\n\t"
+    /*xmm10={0,0,0,0,0,0,1,1}*/
+    "movd %[a],%%xmm10\n\t"
+    /*xmm0=_x[7...0]+{0,0,0,0,0,0,0,_x[0]!=0}*/
+    "paddw %%xmm8,%%xmm0\n\t"
+    "psllw $2,%%xmm7\n\t"
+    /*xmm0=_x[7...0]+{0,0,0,0,0,0,1,(_x[0]!=0)+1}*/
+    "paddw %%xmm10,%%xmm0\n\t"
+    /*xmm1=_x[15...8]-{0,0,0,0,0,0,0,1}*/
+    "psubw %%xmm9,%%xmm1\n\t"
+    /*Transform columns.*/
+    OC_FDCT_8x8
+    /*Transform rows.*/
+    OC_TRANSPOSE_8x8
+    OC_FDCT_8x8
+    /*xmm14={-2,-2,-2,-2,-2,-2,-2,-2}*/
+    "paddw %%xmm14,%%xmm14\n\t"
+    "psubw %%xmm14,%%xmm0\n\t"
+    "psubw %%xmm14,%%xmm1\n\t"
+    "psraw $2,%%xmm0\n\t"
+    "psubw %%xmm14,%%xmm2\n\t"
+    "psraw $2,%%xmm1\n\t"
+    "psubw %%xmm14,%%xmm3\n\t"
+    "psraw $2,%%xmm2\n\t"
+    "psubw %%xmm14,%%xmm4\n\t"
+    "psraw $2,%%xmm3\n\t"
+    "psubw %%xmm14,%%xmm5\n\t"
+    "psraw $2,%%xmm4\n\t"
+    "psubw %%xmm14,%%xmm6\n\t"
+    "psraw $2,%%xmm5\n\t"
+    "psubw %%xmm14,%%xmm7\n\t"
+    "psraw $2,%%xmm6\n\t"
+    "psraw $2,%%xmm7\n\t"
+    /*Transpose, zig-zag, and store the result.*/
+    /*We could probably do better using SSSE3's palignr, but re-using MMXEXT
+       version will do for now.*/
+#define OC_ZZ_LOAD_ROW_LO(_row,_reg) \
+    "movdq2q %%xmm"#_row","_reg"\n\t" \
+
+#define OC_ZZ_LOAD_ROW_HI(_row,_reg) \
+    "punpckhqdq %%xmm"#_row",%%xmm"#_row"\n\t" \
+    "movdq2q %%xmm"#_row","_reg"\n\t" \
+
+    OC_TRANSPOSE_ZIG_ZAG_MMXEXT
+#undef OC_ZZ_LOAD_ROW_LO
+#undef OC_ZZ_LOAD_ROW_HI
+    :[a]"=&r"(a)
+    :[y]"r"(_y),[x]"r"(_x)
+    :"memory"
+  );
+}
+#endif

thirdparty/libtheora/x86/sse2idct.c

@@ -0,0 +1,456 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009,2025           *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+/*SSE2 acceleration of Theora's iDCT.*/
+#include "x86int.h"
+#include "sse2trans.h"
+#include "../dct.h"
+
+#if defined(OC_X86_ASM)
+
+/*A table of constants used by the MMX routines.*/
+const unsigned short __attribute__((aligned(16),used)) OC_IDCT_CONSTS[64]={
+        8,      8,      8,      8,      8,      8,      8,      8,
+  OC_C1S7,OC_C1S7,OC_C1S7,OC_C1S7,OC_C1S7,OC_C1S7,OC_C1S7,OC_C1S7,
+  OC_C2S6,OC_C2S6,OC_C2S6,OC_C2S6,OC_C2S6,OC_C2S6,OC_C2S6,OC_C2S6,
+  OC_C3S5,OC_C3S5,OC_C3S5,OC_C3S5,OC_C3S5,OC_C3S5,OC_C3S5,OC_C3S5,
+  OC_C4S4,OC_C4S4,OC_C4S4,OC_C4S4,OC_C4S4,OC_C4S4,OC_C4S4,OC_C4S4,
+  OC_C5S3,OC_C5S3,OC_C5S3,OC_C5S3,OC_C5S3,OC_C5S3,OC_C5S3,OC_C5S3,
+  OC_C6S2,OC_C6S2,OC_C6S2,OC_C6S2,OC_C6S2,OC_C6S2,OC_C6S2,OC_C6S2,
+  OC_C7S1,OC_C7S1,OC_C7S1,OC_C7S1,OC_C7S1,OC_C7S1,OC_C7S1,OC_C7S1
+};
+
+
+/*Performs the first three stages of the iDCT.
+  xmm2, xmm6, xmm3, and xmm5 must contain the corresponding rows of the input
+   (accessed in that order).
+  The remaining rows must be in _x at their corresponding locations.
+  On output, xmm7 down to xmm4 contain rows 0 through 3, and xmm0 up to xmm3
+   contain rows 4 through 7.*/
+#define OC_IDCT_8x8_ABC(_x) \
+  "#OC_IDCT_8x8_ABC\n\t" \
+  /*Stage 1:*/ \
+  /*2-3 rotation by 6pi/16. \
+    xmm4=xmm7=C6, xmm0=xmm1=C2, xmm2=X2, xmm6=X6.*/ \
+  "movdqa "OC_MEM_OFFS(0x20,c)",%%xmm1\n\t" \
+  "movdqa "OC_MEM_OFFS(0x60,c)",%%xmm4\n\t" \
+  "movdqa %%xmm1,%%xmm0\n\t" \
+  "pmulhw %%xmm2,%%xmm1\n\t" \
+  "movdqa %%xmm4,%%xmm7\n\t" \
+  "pmulhw %%xmm6,%%xmm0\n\t" \
+  "pmulhw %%xmm2,%%xmm7\n\t" \
+  "pmulhw %%xmm6,%%xmm4\n\t" \
+  "paddw %%xmm6,%%xmm0\n\t" \
+  "movdqa "OC_MEM_OFFS(0x30,c)",%%xmm6\n\t" \
+  "paddw %%xmm1,%%xmm2\n\t" \
+  "psubw %%xmm0,%%xmm7\n\t" \
+  "movdqa %%xmm7,"OC_MEM_OFFS(0x00,buf)"\n\t" \
+  "paddw %%xmm4,%%xmm2\n\t" \
+  "movdqa "OC_MEM_OFFS(0x50,c)",%%xmm4\n\t" \
+  "movdqa %%xmm2,"OC_MEM_OFFS(0x10,buf)"\n\t" \
+  /*5-6 rotation by 3pi/16. \
+    xmm4=xmm2=C5, xmm1=xmm6=C3, xmm3=X3, xmm5=X5.*/ \
+  "movdqa %%xmm4,%%xmm2\n\t" \
+  "movdqa %%xmm6,%%xmm1\n\t" \
+  "pmulhw %%xmm3,%%xmm4\n\t" \
+  "pmulhw %%xmm5,%%xmm1\n\t" \
+  "pmulhw %%xmm3,%%xmm6\n\t" \
+  "pmulhw %%xmm5,%%xmm2\n\t" \
+  "paddw %%xmm3,%%xmm4\n\t" \
+  "paddw %%xmm5,%%xmm3\n\t" \
+  "paddw %%xmm6,%%xmm3\n\t" \
+  "movdqa "OC_MEM_OFFS(0x70,_x)",%%xmm6\n\t" \
+  "paddw %%xmm5,%%xmm1\n\t" \
+  "movdqa "OC_MEM_OFFS(0x10,_x)",%%xmm5\n\t" \
+  "paddw %%xmm3,%%xmm2\n\t" \
+  "movdqa "OC_MEM_OFFS(0x70,c)",%%xmm3\n\t" \
+  "psubw %%xmm4,%%xmm1\n\t" \
+  "movdqa "OC_MEM_OFFS(0x10,c)",%%xmm4\n\t" \
+  /*4-7 rotation by 7pi/16. \
+    xmm4=xmm7=C1, xmm3=xmm0=C7, xmm5=X1, xmm6=X7.*/ \
+  "movdqa %%xmm3,%%xmm0\n\t" \
+  "movdqa %%xmm4,%%xmm7\n\t" \
+  "pmulhw %%xmm5,%%xmm3\n\t" \
+  "pmulhw %%xmm5,%%xmm7\n\t" \
+  "pmulhw %%xmm6,%%xmm4\n\t" \
+  "pmulhw %%xmm6,%%xmm0\n\t" \
+  "paddw %%xmm6,%%xmm4\n\t" \
+  "movdqa "OC_MEM_OFFS(0x40,_x)",%%xmm6\n\t" \
+  "paddw %%xmm5,%%xmm7\n\t" \
+  "psubw %%xmm4,%%xmm3\n\t" \
+  "movdqa "OC_MEM_OFFS(0x40,c)",%%xmm4\n\t" \
+  "paddw %%xmm7,%%xmm0\n\t" \
+  "movdqa "OC_MEM_OFFS(0x00,_x)",%%xmm7\n\t" \
+  /*0-1 butterfly. \
+    xmm4=xmm5=C4, xmm7=X0, xmm6=X4.*/ \
+  "paddw %%xmm7,%%xmm6\n\t" \
+  "movdqa %%xmm4,%%xmm5\n\t" \
+  "pmulhw %%xmm6,%%xmm4\n\t" \
+  "paddw %%xmm7,%%xmm7\n\t" \
+  "psubw %%xmm6,%%xmm7\n\t" \
+  "paddw %%xmm6,%%xmm4\n\t" \
+  /*Stage 2:*/ \
+  /*4-5 butterfly: xmm3=t[4], xmm1=t[5] \
+    7-6 butterfly: xmm2=t[6], xmm0=t[7]*/ \
+  "movdqa %%xmm3,%%xmm6\n\t" \
+  "paddw %%xmm1,%%xmm3\n\t" \
+  "psubw %%xmm1,%%xmm6\n\t" \
+  "movdqa %%xmm5,%%xmm1\n\t" \
+  "pmulhw %%xmm7,%%xmm5\n\t" \
+  "paddw %%xmm7,%%xmm5\n\t" \
+  "movdqa %%xmm0,%%xmm7\n\t" \
+  "paddw %%xmm2,%%xmm0\n\t" \
+  "psubw %%xmm2,%%xmm7\n\t" \
+  "movdqa %%xmm1,%%xmm2\n\t" \
+  "pmulhw %%xmm6,%%xmm1\n\t" \
+  "pmulhw %%xmm7,%%xmm2\n\t" \
+  "paddw %%xmm6,%%xmm1\n\t" \
+  "movdqa "OC_MEM_OFFS(0x00,buf)",%%xmm6\n\t" \
+  "paddw %%xmm7,%%xmm2\n\t" \
+  "movdqa "OC_MEM_OFFS(0x10,buf)",%%xmm7\n\t" \
+  /*Stage 3: \
+    6-5 butterfly: xmm1=t[5], xmm2=t[6] -> xmm1=t[6]+t[5], xmm2=t[6]-t[5] \
+    0-3 butterfly: xmm4=t[0], xmm7=t[3] -> xmm7=t[0]+t[3], xmm4=t[0]-t[3] \
+    1-2 butterfly: xmm5=t[1], xmm6=t[2] -> xmm6=t[1]+t[2], xmm5=t[1]-t[2]*/ \
+  "paddw %%xmm2,%%xmm1\n\t" \
+  "paddw %%xmm5,%%xmm6\n\t" \
+  "paddw %%xmm4,%%xmm7\n\t" \
+  "paddw %%xmm2,%%xmm2\n\t" \
+  "paddw %%xmm4,%%xmm4\n\t" \
+  "paddw %%xmm5,%%xmm5\n\t" \
+  "psubw %%xmm1,%%xmm2\n\t" \
+  "psubw %%xmm7,%%xmm4\n\t" \
+  "psubw %%xmm6,%%xmm5\n\t" \
+
+/*Performs the last stage of the iDCT.
+  On input, xmm7 down to xmm4 contain rows 0 through 3, and xmm0 up to xmm3
+   contain rows 4 through 7.
+  On output, xmm0 through xmm7 contain the corresponding rows.*/
+#define OC_IDCT_8x8_D \
+  "#OC_IDCT_8x8_D\n\t" \
+  /*Stage 4: \
+    0-7 butterfly: xmm7=t[0], xmm0=t[7] -> xmm0=t[0]+t[7], xmm7=t[0]-t[7] \
+    1-6 butterfly: xmm6=t[1], xmm1=t[6] -> xmm1=t[1]+t[6], xmm6=t[1]-t[6] \
+    2-5 butterfly: xmm5=t[2], xmm2=t[5] -> xmm2=t[2]+t[5], xmm5=t[2]-t[5] \
+    3-4 butterfly: xmm4=t[3], xmm3=t[4] -> xmm3=t[3]+t[4], xmm4=t[3]-t[4]*/ \
+  "psubw %%xmm0,%%xmm7\n\t" \
+  "psubw %%xmm1,%%xmm6\n\t" \
+  "psubw %%xmm2,%%xmm5\n\t" \
+  "psubw %%xmm3,%%xmm4\n\t" \
+  "paddw %%xmm0,%%xmm0\n\t" \
+  "paddw %%xmm1,%%xmm1\n\t" \
+  "paddw %%xmm2,%%xmm2\n\t" \
+  "paddw %%xmm3,%%xmm3\n\t" \
+  "paddw %%xmm7,%%xmm0\n\t" \
+  "paddw %%xmm6,%%xmm1\n\t" \
+  "paddw %%xmm5,%%xmm2\n\t" \
+  "paddw %%xmm4,%%xmm3\n\t" \
+
+/*Performs the last stage of the iDCT.
+  On input, xmm7 down to xmm4 contain rows 0 through 3, and xmm0 up to xmm3
+   contain rows 4 through 7.
+  On output, xmm0 through xmm7 contain the corresponding rows.*/
+#define OC_IDCT_8x8_D_STORE \
+  "#OC_IDCT_8x8_D_STORE\n\t" \
+  /*Stage 4: \
+    0-7 butterfly: xmm7=t[0], xmm0=t[7] -> xmm0=t[0]+t[7], xmm7=t[0]-t[7] \
+    1-6 butterfly: xmm6=t[1], xmm1=t[6] -> xmm1=t[1]+t[6], xmm6=t[1]-t[6] \
+    2-5 butterfly: xmm5=t[2], xmm2=t[5] -> xmm2=t[2]+t[5], xmm5=t[2]-t[5] \
+    3-4 butterfly: xmm4=t[3], xmm3=t[4] -> xmm3=t[3]+t[4], xmm4=t[3]-t[4]*/ \
+  "psubw %%xmm3,%%xmm4\n\t" \
+  "movdqa %%xmm4,"OC_MEM_OFFS(0x40,y)"\n\t" \
+  "movdqa "OC_MEM_OFFS(0x00,c)",%%xmm4\n\t" \
+  "psubw %%xmm0,%%xmm7\n\t" \
+  "psubw %%xmm1,%%xmm6\n\t" \
+  "psubw %%xmm2,%%xmm5\n\t" \
+  "paddw %%xmm4,%%xmm7\n\t" \
+  "paddw %%xmm4,%%xmm6\n\t" \
+  "paddw %%xmm4,%%xmm5\n\t" \
+  "paddw "OC_MEM_OFFS(0x40,y)",%%xmm4\n\t" \
+  "paddw %%xmm0,%%xmm0\n\t" \
+  "paddw %%xmm1,%%xmm1\n\t" \
+  "paddw %%xmm2,%%xmm2\n\t" \
+  "paddw %%xmm3,%%xmm3\n\t" \
+  "paddw %%xmm7,%%xmm0\n\t" \
+  "paddw %%xmm6,%%xmm1\n\t" \
+  "psraw $4,%%xmm0\n\t" \
+  "paddw %%xmm5,%%xmm2\n\t" \
+  "movdqa %%xmm0,"OC_MEM_OFFS(0x00,y)"\n\t" \
+  "psraw $4,%%xmm1\n\t" \
+  "paddw %%xmm4,%%xmm3\n\t" \
+  "movdqa %%xmm1,"OC_MEM_OFFS(0x10,y)"\n\t" \
+  "psraw $4,%%xmm2\n\t" \
+  "movdqa %%xmm2,"OC_MEM_OFFS(0x20,y)"\n\t" \
+  "psraw $4,%%xmm3\n\t" \
+  "movdqa %%xmm3,"OC_MEM_OFFS(0x30,y)"\n\t" \
+  "psraw $4,%%xmm4\n\t" \
+  "movdqa %%xmm4,"OC_MEM_OFFS(0x40,y)"\n\t" \
+  "psraw $4,%%xmm5\n\t" \
+  "movdqa %%xmm5,"OC_MEM_OFFS(0x50,y)"\n\t" \
+  "psraw $4,%%xmm6\n\t" \
+  "movdqa %%xmm6,"OC_MEM_OFFS(0x60,y)"\n\t" \
+  "psraw $4,%%xmm7\n\t" \
+  "movdqa %%xmm7,"OC_MEM_OFFS(0x70,y)"\n\t" \
+
+static void oc_idct8x8_slow_sse2(ogg_int16_t _y[64],ogg_int16_t _x[64]){
+  OC_ALIGN16(ogg_int16_t buf[16]);
+  int i;
+  /*This routine accepts an 8x8 matrix pre-transposed.*/
+  __asm__ __volatile__(
+    /*Load rows 2, 3, 5, and 6 for the first stage of the iDCT.*/
+    "movdqa "OC_MEM_OFFS(0x20,x)",%%xmm2\n\t"
+    "movdqa "OC_MEM_OFFS(0x60,x)",%%xmm6\n\t"
+    "movdqa "OC_MEM_OFFS(0x30,x)",%%xmm3\n\t"
+    "movdqa "OC_MEM_OFFS(0x50,x)",%%xmm5\n\t"
+    OC_IDCT_8x8_ABC(x)
+    OC_IDCT_8x8_D
+    OC_TRANSPOSE_8x8
+    /*Clear out rows 0, 1, 4, and 7 for the first stage of the iDCT.*/
+    "movdqa %%xmm7,"OC_MEM_OFFS(0x70,y)"\n\t"
+    "movdqa %%xmm4,"OC_MEM_OFFS(0x40,y)"\n\t"
+    "movdqa %%xmm1,"OC_MEM_OFFS(0x10,y)"\n\t"
+    "movdqa %%xmm0,"OC_MEM_OFFS(0x00,y)"\n\t"
+    OC_IDCT_8x8_ABC(y)
+    OC_IDCT_8x8_D_STORE
+    :[buf]"=m"(OC_ARRAY_OPERAND(ogg_int16_t,buf,16)),
+     [y]"=m"(OC_ARRAY_OPERAND(ogg_int16_t,_y,64))
+    :[x]"m"(OC_CONST_ARRAY_OPERAND(ogg_int16_t,_x,64)),
+     [c]"m"(OC_CONST_ARRAY_OPERAND(ogg_int16_t,OC_IDCT_CONSTS,64))
+  );
+  __asm__ __volatile__("pxor %%xmm0,%%xmm0\n\t"::);
+  /*Clear input data for next block (decoder only).*/
+  for(i=0;i<2;i++){
+    __asm__ __volatile__(
+      "movdqa %%xmm0,"OC_MEM_OFFS(0x00,x)"\n\t"
+      "movdqa %%xmm0,"OC_MEM_OFFS(0x10,x)"\n\t"
+      "movdqa %%xmm0,"OC_MEM_OFFS(0x20,x)"\n\t"
+      "movdqa %%xmm0,"OC_MEM_OFFS(0x30,x)"\n\t"
+      :[x]"=m"(OC_ARRAY_OPERAND(ogg_int16_t,_x+i*32,32))
+    );
+  }
+}
+
+/*For the first step of the 10-coefficient version of the 8x8 iDCT, we only
+   need to work with four columns at a time.
+  Doing this in MMX is faster on processors with a 64-bit data path.*/
+#define OC_IDCT_8x8_10_MMX \
+  "#OC_IDCT_8x8_10_MMX\n\t" \
+  /*Stage 1:*/ \
+  /*2-3 rotation by 6pi/16. \
+    mm7=C6, mm6=C2, mm2=X2, X6=0.*/ \
+  "movq "OC_MEM_OFFS(0x60,c)",%%mm7\n\t" \
+  "movq "OC_MEM_OFFS(0x20,c)",%%mm6\n\t" \
+  "pmulhw %%mm2,%%mm6\n\t" \
+  "pmulhw %%mm2,%%mm7\n\t" \
+  "movq "OC_MEM_OFFS(0x50,c)",%%mm5\n\t" \
+  "paddw %%mm6,%%mm2\n\t" \
+  "movq %%mm2,"OC_MEM_OFFS(0x10,buf)"\n\t" \
+  "movq "OC_MEM_OFFS(0x30,c)",%%mm2\n\t" \
+  "movq %%mm7,"OC_MEM_OFFS(0x00,buf)"\n\t" \
+  /*5-6 rotation by 3pi/16. \
+    mm5=C5, mm2=C3, mm3=X3, X5=0.*/ \
+  "pmulhw %%mm3,%%mm5\n\t" \
+  "pmulhw %%mm3,%%mm2\n\t" \
+  "movq "OC_MEM_OFFS(0x10,c)",%%mm7\n\t" \
+  "paddw %%mm3,%%mm5\n\t" \
+  "paddw %%mm3,%%mm2\n\t" \
+  "movq "OC_MEM_OFFS(0x70,c)",%%mm3\n\t" \
+  /*4-7 rotation by 7pi/16. \
+    mm7=C1, mm3=C7, mm1=X1, X7=0.*/ \
+  "pmulhw %%mm1,%%mm3\n\t" \
+  "pmulhw %%mm1,%%mm7\n\t" \
+  "movq "OC_MEM_OFFS(0x40,c)",%%mm4\n\t" \
+  "movq %%mm3,%%mm6\n\t" \
+  "paddw %%mm1,%%mm7\n\t" \
+  /*0-1 butterfly. \
+    mm4=C4, mm0=X0, X4=0.*/ \
+  /*Stage 2:*/ \
+  /*4-5 butterfly: mm3=t[4], mm5=t[5] \
+    7-6 butterfly: mm2=t[6], mm7=t[7]*/ \
+  "psubw %%mm5,%%mm3\n\t" \
+  "paddw %%mm5,%%mm6\n\t" \
+  "movq %%mm4,%%mm1\n\t" \
+  "pmulhw %%mm0,%%mm4\n\t" \
+  "paddw %%mm0,%%mm4\n\t" \
+  "movq %%mm7,%%mm0\n\t" \
+  "movq %%mm4,%%mm5\n\t" \
+  "paddw %%mm2,%%mm0\n\t" \
+  "psubw %%mm2,%%mm7\n\t" \
+  "movq %%mm1,%%mm2\n\t" \
+  "pmulhw %%mm6,%%mm1\n\t" \
+  "pmulhw %%mm7,%%mm2\n\t" \
+  "paddw %%mm6,%%mm1\n\t" \
+  "movq "OC_MEM_OFFS(0x00,buf)",%%mm6\n\t" \
+  "paddw %%mm7,%%mm2\n\t" \
+  "movq "OC_MEM_OFFS(0x10,buf)",%%mm7\n\t" \
+  /*Stage 3: \
+    6-5 butterfly: mm1=t[5], mm2=t[6] -> mm1=t[6]+t[5], mm2=t[6]-t[5] \
+    0-3 butterfly: mm4=t[0], mm7=t[3] -> mm7=t[0]+t[3], mm4=t[0]-t[3] \
+    1-2 butterfly: mm5=t[1], mm6=t[2] -> mm6=t[1]+t[2], mm5=t[1]-t[2]*/ \
+  "paddw %%mm2,%%mm1\n\t" \
+  "paddw %%mm5,%%mm6\n\t" \
+  "paddw %%mm4,%%mm7\n\t" \
+  "paddw %%mm2,%%mm2\n\t" \
+  "paddw %%mm4,%%mm4\n\t" \
+  "paddw %%mm5,%%mm5\n\t" \
+  "psubw %%mm1,%%mm2\n\t" \
+  "psubw %%mm7,%%mm4\n\t" \
+  "psubw %%mm6,%%mm5\n\t" \
+  /*Stage 4: \
+    0-7 butterfly: mm7=t[0], mm0=t[7] -> mm0=t[0]+t[7], mm7=t[0]-t[7] \
+    1-6 butterfly: mm6=t[1], mm1=t[6] -> mm1=t[1]+t[6], mm6=t[1]-t[6] \
+    2-5 butterfly: mm5=t[2], mm2=t[5] -> mm2=t[2]+t[5], mm5=t[2]-t[5] \
+    3-4 butterfly: mm4=t[3], mm3=t[4] -> mm3=t[3]+t[4], mm4=t[3]-t[4]*/ \
+  "psubw %%mm0,%%mm7\n\t" \
+  "psubw %%mm1,%%mm6\n\t" \
+  "psubw %%mm2,%%mm5\n\t" \
+  "psubw %%mm3,%%mm4\n\t" \
+  "paddw %%mm0,%%mm0\n\t" \
+  "paddw %%mm1,%%mm1\n\t" \
+  "paddw %%mm2,%%mm2\n\t" \
+  "paddw %%mm3,%%mm3\n\t" \
+  "paddw %%mm7,%%mm0\n\t" \
+  "paddw %%mm6,%%mm1\n\t" \
+  "paddw %%mm5,%%mm2\n\t" \
+  "paddw %%mm4,%%mm3\n\t" \
+
+#define OC_IDCT_8x8_10_ABC \
+  "#OC_IDCT_8x8_10_ABC\n\t" \
+  /*Stage 1:*/ \
+  /*2-3 rotation by 6pi/16. \
+    xmm7=C6, xmm6=C2, xmm2=X2, X6=0.*/ \
+  "movdqa "OC_MEM_OFFS(0x60,c)",%%xmm7\n\t" \
+  "movdqa "OC_MEM_OFFS(0x20,c)",%%xmm6\n\t" \
+  "pmulhw %%xmm2,%%xmm6\n\t" \
+  "pmulhw %%xmm2,%%xmm7\n\t" \
+  "movdqa "OC_MEM_OFFS(0x50,c)",%%xmm5\n\t" \
+  "paddw %%xmm6,%%xmm2\n\t" \
+  "movdqa %%xmm2,"OC_MEM_OFFS(0x10,buf)"\n\t" \
+  "movdqa "OC_MEM_OFFS(0x30,c)",%%xmm2\n\t" \
+  "movdqa %%xmm7,"OC_MEM_OFFS(0x00,buf)"\n\t" \
+  /*5-6 rotation by 3pi/16. \
+    xmm5=C5, xmm2=C3, xmm3=X3, X5=0.*/ \
+  "pmulhw %%xmm3,%%xmm5\n\t" \
+  "pmulhw %%xmm3,%%xmm2\n\t" \
+  "movdqa "OC_MEM_OFFS(0x10,c)",%%xmm7\n\t" \
+  "paddw %%xmm3,%%xmm5\n\t" \
+  "paddw %%xmm3,%%xmm2\n\t" \
+  "movdqa "OC_MEM_OFFS(0x70,c)",%%xmm3\n\t" \
+  /*4-7 rotation by 7pi/16. \
+    xmm7=C1, xmm3=C7, xmm1=X1, X7=0.*/ \
+  "pmulhw %%xmm1,%%xmm3\n\t" \
+  "pmulhw %%xmm1,%%xmm7\n\t" \
+  "movdqa "OC_MEM_OFFS(0x40,c)",%%xmm4\n\t" \
+  "movdqa %%xmm3,%%xmm6\n\t" \
+  "paddw %%xmm1,%%xmm7\n\t" \
+  /*0-1 butterfly. \
+    xmm4=C4, xmm0=X0, X4=0.*/ \
+  /*Stage 2:*/ \
+  /*4-5 butterfly: xmm3=t[4], xmm5=t[5] \
+    7-6 butterfly: xmm2=t[6], xmm7=t[7]*/ \
+  "psubw %%xmm5,%%xmm3\n\t" \
+  "paddw %%xmm5,%%xmm6\n\t" \
+  "movdqa %%xmm4,%%xmm1\n\t" \
+  "pmulhw %%xmm0,%%xmm4\n\t" \
+  "paddw %%xmm0,%%xmm4\n\t" \
+  "movdqa %%xmm7,%%xmm0\n\t" \
+  "movdqa %%xmm4,%%xmm5\n\t" \
+  "paddw %%xmm2,%%xmm0\n\t" \
+  "psubw %%xmm2,%%xmm7\n\t" \
+  "movdqa %%xmm1,%%xmm2\n\t" \
+  "pmulhw %%xmm6,%%xmm1\n\t" \
+  "pmulhw %%xmm7,%%xmm2\n\t" \
+  "paddw %%xmm6,%%xmm1\n\t" \
+  "movdqa "OC_MEM_OFFS(0x00,buf)",%%xmm6\n\t" \
+  "paddw %%xmm7,%%xmm2\n\t" \
+  "movdqa "OC_MEM_OFFS(0x10,buf)",%%xmm7\n\t" \
+  /*Stage 3: \
+    6-5 butterfly: xmm1=t[5], xmm2=t[6] -> xmm1=t[6]+t[5], xmm2=t[6]-t[5] \
+    0-3 butterfly: xmm4=t[0], xmm7=t[3] -> xmm7=t[0]+t[3], xmm4=t[0]-t[3] \
+    1-2 butterfly: xmm5=t[1], xmm6=t[2] -> xmm6=t[1]+t[2], xmm5=t[1]-t[2]*/ \
+  "paddw %%xmm2,%%xmm1\n\t" \
+  "paddw %%xmm5,%%xmm6\n\t" \
+  "paddw %%xmm4,%%xmm7\n\t" \
+  "paddw %%xmm2,%%xmm2\n\t" \
+  "paddw %%xmm4,%%xmm4\n\t" \
+  "paddw %%xmm5,%%xmm5\n\t" \
+  "psubw %%xmm1,%%xmm2\n\t" \
+  "psubw %%xmm7,%%xmm4\n\t" \
+  "psubw %%xmm6,%%xmm5\n\t" \
+
+static void oc_idct8x8_10_sse2(ogg_int16_t _y[64],ogg_int16_t _x[64]){
+  OC_ALIGN16(ogg_int16_t buf[16]);
+  /*This routine accepts an 8x8 matrix pre-transposed.*/
+  __asm__ __volatile__(
+    "movq "OC_MEM_OFFS(0x20,x)",%%mm2\n\t"
+    "movq "OC_MEM_OFFS(0x30,x)",%%mm3\n\t"
+    "movq "OC_MEM_OFFS(0x10,x)",%%mm1\n\t"
+    "movq "OC_MEM_OFFS(0x00,x)",%%mm0\n\t"
+    OC_IDCT_8x8_10_MMX
+    OC_TRANSPOSE_8x4_MMX2SSE
+    OC_IDCT_8x8_10_ABC
+    OC_IDCT_8x8_D_STORE
+    :[buf]"=m"(OC_ARRAY_OPERAND(short,buf,16)),
+     [y]"=m"(OC_ARRAY_OPERAND(ogg_int16_t,_y,64))
+    :[x]"m"OC_CONST_ARRAY_OPERAND(ogg_int16_t,_x,64),
+     [c]"m"(OC_CONST_ARRAY_OPERAND(ogg_int16_t,OC_IDCT_CONSTS,64))
+  );
+  /*Clear input data for next block (decoder only).*/
+  __asm__ __volatile__(
+    "pxor %%mm0,%%mm0\n\t"
+    "movq %%mm0,"OC_MEM_OFFS(0x00,x)"\n\t"
+    "movq %%mm0,"OC_MEM_OFFS(0x10,x)"\n\t"
+    "movq %%mm0,"OC_MEM_OFFS(0x20,x)"\n\t"
+    "movq %%mm0,"OC_MEM_OFFS(0x30,x)"\n\t"
+    :[x]"+m"(OC_ARRAY_OPERAND(ogg_int16_t,_x,28))
+  );
+}
+
+/*Performs an inverse 8x8 Type-II DCT transform.
+  The input is assumed to be scaled by a factor of 4 relative to orthonormal
+   version of the transform.*/
+void oc_idct8x8_sse2(ogg_int16_t _y[64],ogg_int16_t _x[64],int _last_zzi){
+  /*_last_zzi is subtly different from an actual count of the number of
+     coefficients we decoded for this block.
+    It contains the value of zzi BEFORE the final token in the block was
+     decoded.
+    In most cases this is an EOB token (the continuation of an EOB run from a
+     previous block counts), and so this is the same as the coefficient count.
+    However, in the case that the last token was NOT an EOB token, but filled
+     the block up with exactly 64 coefficients, _last_zzi will be less than 64.
+    Provided the last token was not a pure zero run, the minimum value it can
+     be is 46, and so that doesn't affect any of the cases in this routine.
+    However, if the last token WAS a pure zero run of length 63, then _last_zzi
+     will be 1 while the number of coefficients decoded is 64.
+    Thus, we will trigger the following special case, where the real
+     coefficient count would not.
+    Note also that a zero run of length 64 will give _last_zzi a value of 0,
+     but we still process the DC coefficient, which might have a non-zero value
+     due to DC prediction.
+    Although convoluted, this is arguably the correct behavior: it allows us to
+     use a smaller transform when the block ends with a long zero run instead
+     of a normal EOB token.
+    It could be smarter... multiple separate zero runs at the end of a block
+     will fool it, but an encoder that generates these really deserves what it
+     gets.
+    Needless to say we inherited this approach from VP3.*/
+  /*Then perform the iDCT.*/
+  if(_last_zzi<=10)oc_idct8x8_10_sse2(_y,_x);
+  else oc_idct8x8_slow_sse2(_y,_x);
+}
+
+#endif

thirdparty/libtheora/x86/sse2trans.h

@@ -0,0 +1,242 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+#if !defined(_x86_sse2trans_H)
+# define _x86_sse2trans_H (1)
+# include "x86int.h"
+
+# if defined(OC_X86_64_ASM)
+/*On x86-64 we can transpose in-place without spilling registers.
+  By clever choices of the order to apply the butterflies and the order of
+   their outputs, we can take the rows in order and output the columns in order
+   without any extra operations and using just one temporary register.*/
+#  define OC_TRANSPOSE_8x8 \
+ "#OC_TRANSPOSE_8x8\n\t" \
+ "movdqa %%xmm4,%%xmm8\n\t" \
+ /*xmm4 = f3 e3 f2 e2 f1 e1 f0 e0*/ \
+ "punpcklwd %%xmm5,%%xmm4\n\t" \
+ /*xmm8 = f7 e7 f6 e6 f5 e5 f4 e4*/ \
+ "punpckhwd %%xmm5,%%xmm8\n\t" \
+ /*xmm5 is free.*/ \
+ "movdqa %%xmm0,%%xmm5\n\t" \
+ /*xmm0 = b3 a3 b2 a2 b1 a1 b0 a0*/ \
+ "punpcklwd %%xmm1,%%xmm0\n\t" \
+ /*xmm5 = b7 a7 b6 a6 b5 a5 b4 a4*/ \
+ "punpckhwd %%xmm1,%%xmm5\n\t" \
+ /*xmm1 is free.*/ \
+ "movdqa %%xmm6,%%xmm1\n\t" \
+ /*xmm6 = h3 g3 h2 g2 h1 g1 h0 g0*/ \
+ "punpcklwd %%xmm7,%%xmm6\n\t" \
+ /*xmm1 = h7 g7 h6 g6 h5 g5 h4 g4*/ \
+ "punpckhwd %%xmm7,%%xmm1\n\t" \
+ /*xmm7 is free.*/ \
+ "movdqa %%xmm2,%%xmm7\n\t" \
+ /*xmm2 = d7 c7 d6 c6 d5 c5 d4 c4*/ \
+ "punpckhwd %%xmm3,%%xmm2\n\t" \
+ /*xmm7 = d3 c3 d2 c2 d1 c1 d0 c0*/ \
+ "punpcklwd %%xmm3,%%xmm7\n\t" \
+ /*xmm3 is free.*/ \
+ "movdqa %%xmm0,%%xmm3\n\t" \
+ /*xmm0 = d1 c1 b1 a1 d0 c0 b0 a0*/ \
+ "punpckldq %%xmm7,%%xmm0\n\t" \
+ /*xmm3 = d3 c3 b3 a3 d2 c2 b2 a2*/ \
+ "punpckhdq %%xmm7,%%xmm3\n\t" \
+ /*xmm7 is free.*/ \
+ "movdqa %%xmm5,%%xmm7\n\t" \
+ /*xmm5 = d5 c5 b5 a5 d4 c4 b4 a4*/ \
+ "punpckldq %%xmm2,%%xmm5\n\t" \
+ /*xmm7 = d7 c7 b7 a7 d6 c6 b6 a6*/ \
+ "punpckhdq %%xmm2,%%xmm7\n\t" \
+ /*xmm2 is free.*/ \
+ "movdqa %%xmm4,%%xmm2\n\t" \
+ /*xmm4 = h3 g3 f3 e3 h2 g2 f2 e2*/ \
+ "punpckhdq %%xmm6,%%xmm4\n\t" \
+ /*xmm2 = h1 g1 f1 e1 h0 g0 f0 e0*/ \
+ "punpckldq %%xmm6,%%xmm2\n\t" \
+ /*xmm6 is free.*/ \
+ "movdqa %%xmm8,%%xmm6\n\t" \
+ /*xmm6 = h5 g5 f5 e5 h4 g4 f4 e4*/ \
+ "punpckldq %%xmm1,%%xmm6\n\t" \
+ /*xmm8 = h7 g7 f7 e7 h6 g6 f6 e6*/ \
+ "punpckhdq %%xmm1,%%xmm8\n\t" \
+ /*xmm1 is free.*/ \
+ "movdqa %%xmm0,%%xmm1\n\t" \
+ /*xmm0 = h0 g0 f0 e0 d0 c0 b0 a0*/ \
+ "punpcklqdq %%xmm2,%%xmm0\n\t" \
+ /*xmm1 = h1 g1 f1 e1 d1 c1 b1 a1*/ \
+ "punpckhqdq %%xmm2,%%xmm1\n\t" \
+ /*xmm2 is free.*/ \
+ "movdqa %%xmm3,%%xmm2\n\t" \
+ /*xmm3 = h3 g3 f3 e3 d3 c3 b3 a3*/ \
+ "punpckhqdq %%xmm4,%%xmm3\n\t" \
+ /*xmm2 = h2 g2 f2 e2 d2 c2 b2 a2*/ \
+ "punpcklqdq %%xmm4,%%xmm2\n\t" \
+ /*xmm4 is free.*/ \
+ "movdqa %%xmm5,%%xmm4\n\t" \
+ /*xmm5 = h5 g5 f5 e5 d5 c5 b5 a5*/ \
+ "punpckhqdq %%xmm6,%%xmm5\n\t" \
+ /*xmm4 = h4 g4 f4 e4 d4 c4 b4 a4*/ \
+ "punpcklqdq %%xmm6,%%xmm4\n\t" \
+ /*xmm6 is free.*/ \
+ "movdqa %%xmm7,%%xmm6\n\t" \
+ /*xmm7 = h7 g7 f7 e7 d7 c7 b7 a7*/ \
+ "punpckhqdq %%xmm8,%%xmm7\n\t" \
+ /*xmm6 = h6 g6 f6 e6 d6 c6 b6 a6*/ \
+ "punpcklqdq %%xmm8,%%xmm6\n\t" \
+ /*xmm8 is free.*/ \
+
+# else
+/*Otherwise, we need to spill some values to %[buf] temporarily.
+  Again, the butterflies are carefully arranged to get the columns to come out
+   in order, minimizing register spills and maximizing the delay between a load
+   and when the value loaded is actually used.*/
+#  define OC_TRANSPOSE_8x8 \
+ "#OC_TRANSPOSE_8x8\n\t" \
+ /*buf[0] = a7 a6 a5 a4 a3 a2 a1 a0*/ \
+ "movdqa %%xmm0,"OC_MEM_OFFS(0x00,buf)"\n\t" \
+ /*xmm0 is free.*/ \
+ "movdqa %%xmm2,%%xmm0\n\t" \
+ /*xmm2 = d7 c7 d6 c6 d5 c5 d4 c4*/ \
+ "punpckhwd %%xmm3,%%xmm2\n\t" \
+ /*xmm0 = d3 c3 d2 c2 d1 c1 d0 c0*/ \
+ "punpcklwd %%xmm3,%%xmm0\n\t" \
+ /*xmm3 = a7 a6 a5 a4 a3 a2 a1 a0*/ \
+ "movdqa "OC_MEM_OFFS(0x00,buf)",%%xmm3\n\t" \
+ /*buf[1] = d7 c7 d6 c6 d5 c5 d4 c4*/ \
+ "movdqa %%xmm2,"OC_MEM_OFFS(0x10,buf)"\n\t" \
+ /*xmm2 is free.*/ \
+ "movdqa %%xmm6,%%xmm2\n\t" \
+ /*xmm6 = h3 g3 h2 g2 h1 g1 h0 g0*/ \
+ "punpcklwd %%xmm7,%%xmm6\n\t" \
+ /*xmm2 = h7 g7 h6 g6 h5 g5 h4 g4*/ \
+ "punpckhwd %%xmm7,%%xmm2\n\t" \
+ /*xmm7 is free.*/ \
+ "movdqa %%xmm4,%%xmm7\n\t" \
+ /*xmm4 = f3 e3 f2 e2 f1 e1 f0 e0*/ \
+ "punpcklwd %%xmm5,%%xmm4\n\t" \
+ /*xmm7 = f7 e7 f6 e6 f5 e5 f4 e4*/ \
+ "punpckhwd %%xmm5,%%xmm7\n\t" \
+ /*xmm5 is free.*/ \
+ "movdqa %%xmm3,%%xmm5\n\t" \
+ /*xmm3 = b3 a3 b2 a2 b1 a1 b0 a0*/ \
+ "punpcklwd %%xmm1,%%xmm3\n\t" \
+ /*xmm5 = b7 a7 b6 a6 b5 a5 b4 a4*/ \
+ "punpckhwd %%xmm1,%%xmm5\n\t" \
+ /*xmm1 is free.*/ \
+ "movdqa %%xmm7,%%xmm1\n\t" \
+ /*xmm7 = h5 g5 f5 e5 h4 g4 f4 e4*/ \
+ "punpckldq %%xmm2,%%xmm7\n\t" \
+ /*xmm1 = h7 g7 f7 e7 h6 g6 f6 e6*/ \
+ "punpckhdq %%xmm2,%%xmm1\n\t" \
+ /*xmm2 = d7 c7 d6 c6 d5 c5 d4 c4*/ \
+ "movdqa "OC_MEM_OFFS(0x10,buf)",%%xmm2\n\t" \
+ /*buf[0] = h7 g7 f7 e7 h6 g6 f6 e6*/ \
+ "movdqa %%xmm1,"OC_MEM_OFFS(0x00,buf)"\n\t" \
+ /*xmm1 is free.*/ \
+ "movdqa %%xmm3,%%xmm1\n\t" \
+ /*xmm3 = d3 c3 b3 a3 d2 c2 b2 a2*/ \
+ "punpckhdq %%xmm0,%%xmm3\n\t" \
+ /*xmm1 = d1 c1 b1 a1 d0 c0 b0 a0*/ \
+ "punpckldq %%xmm0,%%xmm1\n\t" \
+ /*xmm0 is free.*/ \
+ "movdqa %%xmm4,%%xmm0\n\t" \
+ /*xmm4 = h3 g3 f3 e3 h2 g2 f2 e2*/ \
+ "punpckhdq %%xmm6,%%xmm4\n\t" \
+ /*xmm0 = h1 g1 f1 e1 h0 g0 f0 e0*/ \
+ "punpckldq %%xmm6,%%xmm0\n\t" \
+ /*xmm6 is free.*/ \
+ "movdqa %%xmm5,%%xmm6\n\t" \
+ /*xmm5 = d5 c5 b5 a5 d4 c4 b4 a4*/ \
+ "punpckldq %%xmm2,%%xmm5\n\t" \
+ /*xmm6 = d7 c7 b7 a7 d6 c6 b6 a6*/ \
+ "punpckhdq %%xmm2,%%xmm6\n\t" \
+ /*xmm2 is free.*/ \
+ "movdqa %%xmm1,%%xmm2\n\t" \
+ /*xmm1 = h1 g1 f1 e1 d1 c1 b1 a1*/ \
+ "punpckhqdq %%xmm0,%%xmm1\n\t" \
+ /*xmm2 = h0 g0 f0 e0 d0 c0 b0 a0*/ \
+ "punpcklqdq %%xmm0,%%xmm2\n\t" \
+ /*xmm0 = h7 g7 f7 e7 h6 g6 f6 e6*/ \
+ "movdqa "OC_MEM_OFFS(0x00,buf)",%%xmm0\n\t" \
+ /*buf[1] = h0 g0 f0 e0 d0 c0 b0 a0*/ \
+ "movdqa %%xmm2,"OC_MEM_OFFS(0x10,buf)"\n\t" \
+ /*xmm2 is free.*/ \
+ "movdqa %%xmm3,%%xmm2\n\t" \
+ /*xmm3 = h3 g3 f3 e3 d3 c3 b3 a3*/ \
+ "punpckhqdq %%xmm4,%%xmm3\n\t" \
+ /*xmm2 = h2 g2 f2 e2 d2 c2 b2 a2*/ \
+ "punpcklqdq %%xmm4,%%xmm2\n\t" \
+ /*xmm4 is free.*/ \
+ "movdqa %%xmm5,%%xmm4\n\t" \
+ /*xmm5 = h5 g5 f5 e5 d5 c5 b5 a5*/ \
+ "punpckhqdq %%xmm7,%%xmm5\n\t" \
+ /*xmm4 = h4 g4 f4 e4 d4 c4 b4 a4*/ \
+ "punpcklqdq %%xmm7,%%xmm4\n\t" \
+ /*xmm7 is free.*/ \
+ "movdqa %%xmm6,%%xmm7\n\t" \
+ /*xmm6 = h6 g6 f6 e6 d6 c6 b6 a6*/ \
+ "punpcklqdq %%xmm0,%%xmm6\n\t" \
+ /*xmm7 = h7 g7 f7 e7 d7 c7 b7 a7*/ \
+ "punpckhqdq %%xmm0,%%xmm7\n\t" \
+ /*xmm0 = h0 g0 f0 e0 d0 c0 b0 a0*/ \
+ "movdqa "OC_MEM_OFFS(0x10,buf)",%%xmm0\n\t" \
+
+# endif
+
+/*Transpose 4 values in each of 8 MMX registers into 8 values in the first
+   four SSE registers.
+  No need to be clever here; we have plenty of room.*/
+#  define OC_TRANSPOSE_8x4_MMX2SSE \
+ "#OC_TRANSPOSE_8x4_MMX2SSE\n\t" \
+ "movq2dq %%mm0,%%xmm0\n\t" \
+ "movq2dq %%mm1,%%xmm1\n\t" \
+ /*xmmA = b3 a3 b2 a2 b1 a1 b0 a0*/ \
+ "punpcklwd %%xmm1,%%xmm0\n\t" \
+ "movq2dq %%mm2,%%xmm3\n\t" \
+ "movq2dq %%mm3,%%xmm2\n\t" \
+ /*xmmC = d3 c3 d2 c2 d1 c1 d0 c0*/ \
+ "punpcklwd %%xmm2,%%xmm3\n\t" \
+ "movq2dq %%mm4,%%xmm4\n\t" \
+ "movq2dq %%mm5,%%xmm5\n\t" \
+ /*xmmE = f3 e3 f2 e2 f1 e1 f0 e0*/ \
+ "punpcklwd %%xmm5,%%xmm4\n\t" \
+ "movq2dq %%mm6,%%xmm7\n\t" \
+ "movq2dq %%mm7,%%xmm6\n\t" \
+ /*xmmG = h3 g3 h2 g2 h1 g1 h0 g0*/ \
+ "punpcklwd %%xmm6,%%xmm7\n\t" \
+ "movdqa %%xmm0,%%xmm2\n\t" \
+ /*xmm0 = d1 c1 b1 a1 d0 c0 b0 a0*/ \
+ "punpckldq %%xmm3,%%xmm0\n\t" \
+ /*xmm2 = d3 c3 b3 a3 d2 c2 b2 a2*/ \
+ "punpckhdq %%xmm3,%%xmm2\n\t" \
+ "movdqa %%xmm4,%%xmm5\n\t" \
+ /*xmm4 = h1 g1 f1 e1 h0 g0 f0 e0*/ \
+ "punpckldq %%xmm7,%%xmm4\n\t" \
+ /*xmm3 = h3 g3 f3 e3 h2 g2 f2 e2*/ \
+ "punpckhdq %%xmm7,%%xmm5\n\t" \
+ "movdqa %%xmm0,%%xmm1\n\t" \
+ /*xmm0 = h0 g0 f0 e0 d0 c0 b0 a0*/ \
+ "punpcklqdq %%xmm4,%%xmm0\n\t" \
+ /*xmm1 = h1 g1 f1 e1 d1 c1 b1 a1*/ \
+ "punpckhqdq %%xmm4,%%xmm1\n\t" \
+ "movdqa %%xmm2,%%xmm3\n\t" \
+ /*xmm2 = h2 g2 f2 e2 d2 c2 b2 a2*/ \
+ "punpcklqdq %%xmm5,%%xmm2\n\t" \
+ /*xmm3 = h3 g3 f3 e3 d3 c3 b3 a3*/ \
+ "punpckhqdq %%xmm5,%%xmm3\n\t" \
+
+#endif

thirdparty/libtheora/x86/x86cpu.c

@@ -0,0 +1,182 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+ CPU capability detection for x86 processors.
+  Originally written by Rudolf Marek.
+
+ function:
+
+ ********************************************************************/
+
+#include "x86cpu.h"
+
+#if !defined(OC_X86_ASM)
+ogg_uint32_t oc_cpu_flags_get(void){
+  return 0;
+}
+#else
+# if defined(__amd64__)||defined(__x86_64__)
+/*On x86-64, gcc seems to be able to figure out how to save %rbx for us when
+   compiling with -fPIC.*/
+#  define cpuid(_op,_eax,_ebx,_ecx,_edx) \
+  __asm__ __volatile__( \
+   "cpuid\n\t" \
+   :[eax]"=a"(_eax),[ebx]"=b"(_ebx),[ecx]"=c"(_ecx),[edx]"=d"(_edx) \
+   :"a"(_op) \
+   :"cc" \
+  )
+# else
+/*On x86-32, not so much.*/
+#  define cpuid(_op,_eax,_ebx,_ecx,_edx) \
+  __asm__ __volatile__( \
+   "xchgl %%ebx,%[ebx]\n\t" \
+   "cpuid\n\t" \
+   "xchgl %%ebx,%[ebx]\n\t" \
+   :[eax]"=a"(_eax),[ebx]"=r"(_ebx),[ecx]"=c"(_ecx),[edx]"=d"(_edx) \
+   :"a"(_op) \
+   :"cc" \
+  )
+# endif
+
+static ogg_uint32_t oc_parse_intel_flags(ogg_uint32_t _edx,ogg_uint32_t _ecx){
+  ogg_uint32_t flags;
+  /*If there isn't even MMX, give up.*/
+  if(!(_edx&0x00800000))return 0;
+  flags=OC_CPU_X86_MMX;
+  if(_edx&0x02000000)flags|=OC_CPU_X86_MMXEXT|OC_CPU_X86_SSE;
+  if(_edx&0x04000000)flags|=OC_CPU_X86_SSE2;
+  if(_ecx&0x00000001)flags|=OC_CPU_X86_PNI;
+  if(_ecx&0x00000100)flags|=OC_CPU_X86_SSSE3;
+  if(_ecx&0x00080000)flags|=OC_CPU_X86_SSE4_1;
+  if(_ecx&0x00100000)flags|=OC_CPU_X86_SSE4_2;
+  return flags;
+}
+
+static ogg_uint32_t oc_parse_amd_flags(ogg_uint32_t _edx,ogg_uint32_t _ecx){
+  ogg_uint32_t flags;
+  /*If there isn't even MMX, give up.*/
+  if(!(_edx&0x00800000))return 0;
+  flags=OC_CPU_X86_MMX;
+  if(_edx&0x00400000)flags|=OC_CPU_X86_MMXEXT;
+  if(_edx&0x80000000)flags|=OC_CPU_X86_3DNOW;
+  if(_edx&0x40000000)flags|=OC_CPU_X86_3DNOWEXT;
+  if(_ecx&0x00000040)flags|=OC_CPU_X86_SSE4A;
+  if(_ecx&0x00000800)flags|=OC_CPU_X86_SSE5;
+  return flags;
+}
+
+ogg_uint32_t oc_cpu_flags_get(void){
+  ogg_uint32_t flags;
+  ogg_uint32_t eax;
+  ogg_uint32_t ebx;
+  ogg_uint32_t ecx;
+  ogg_uint32_t edx;
+# if !defined(__amd64__)&&!defined(__x86_64__)
+  /*Not all x86-32 chips support cpuid, so we have to check.*/
+  __asm__ __volatile__(
+   "pushfl\n\t"
+   "pushfl\n\t"
+   "popl %[a]\n\t"
+   "movl %[a],%[b]\n\t"
+   "xorl $0x200000,%[a]\n\t"
+   "pushl %[a]\n\t"
+   "popfl\n\t"
+   "pushfl\n\t"
+   "popl %[a]\n\t"
+   "popfl\n\t"
+   :[a]"=r"(eax),[b]"=r"(ebx)
+   :
+   :"cc"
+  );
+  /*No cpuid.*/
+  if(eax==ebx)return 0;
+# endif
+  cpuid(0,eax,ebx,ecx,edx);
+  /*         l e t n          I e n i          u n e G*/
+  if(ecx==0x6C65746E&&edx==0x49656E69&&ebx==0x756E6547||
+   /*      6 8 x M          T e n i          u n e G*/
+   ecx==0x3638784D&&edx==0x54656E69&&ebx==0x756E6547){
+    int family;
+    int model;
+    /*Intel, Transmeta (tested with Crusoe TM5800):*/
+    cpuid(1,eax,ebx,ecx,edx);
+    flags=oc_parse_intel_flags(edx,ecx);
+    family=(eax>>8)&0xF;
+    model=(eax>>4)&0xF;
+    /*The SSE unit on the Pentium M and Core Duo is much slower than the MMX
+       unit, so don't use it.*/
+    if(family==6&&(model==9||model==13||model==14)){
+      flags&=~(OC_CPU_X86_SSE2|OC_CPU_X86_PNI);
+    }
+  }
+  /*              D M A c          i t n e          h t u A*/
+  else if(ecx==0x444D4163&&edx==0x69746E65&&ebx==0x68747541||
+   /*      C S N            y b   e          d o e G*/
+   ecx==0x43534e20&&edx==0x79622065&&ebx==0x646f6547){
+    /*AMD, Geode:*/
+    cpuid(0x80000000,eax,ebx,ecx,edx);
+    if(eax<0x80000001)flags=0;
+    else{
+      cpuid(0x80000001,eax,ebx,ecx,edx);
+      flags=oc_parse_amd_flags(edx,ecx);
+    }
+    /*Also check for SSE.*/
+    cpuid(1,eax,ebx,ecx,edx);
+    flags|=oc_parse_intel_flags(edx,ecx);
+  }
+  /*Technically some VIA chips can be configured in the BIOS to return any
+     string here the user wants.
+    There is a special detection method that can be used to identify such
+     processors, but in my opinion, if the user really wants to change it, they
+     deserve what they get.*/
+  /*              s l u a          H r u a          t n e C*/
+  else if(ecx==0x736C7561&&edx==0x48727561&&ebx==0x746E6543){
+    /*VIA:*/
+    /*I only have documentation for the C7 (Esther) and Isaiah (forthcoming)
+       chips (thanks to the engineers from Centaur Technology who provided it).
+      These chips support Intel-like cpuid info.
+      The C3-2 (Nehemiah) cores appear to, as well.*/
+    cpuid(1,eax,ebx,ecx,edx);
+    flags=oc_parse_intel_flags(edx,ecx);
+    if(eax>=0x80000001){
+      /*The (non-Nehemiah) C3 processors support AMD-like cpuid info.
+        We need to check this even if the Intel test succeeds to pick up 3DNow!
+         support on these processors.
+        Unlike actual AMD processors, we cannot _rely_ on this info, since
+         some cores (e.g., the 693 stepping of the Nehemiah) claim to support
+         this function, yet return edx=0, despite the Intel test indicating
+         MMX support.
+        Therefore the features detected here are strictly added to those
+         detected by the Intel test.*/
+      /*TODO: How about earlier chips?*/
+      cpuid(0x80000001,eax,ebx,ecx,edx);
+      /*Note: As of the C7, this function returns Intel-style extended feature
+         flags, not AMD-style.
+        Currently, this only defines bits 11, 20, and 29 (0x20100800), which
+         do not conflict with any of the AMD flags we inspect.
+        For the remaining bits, Intel tells us, "Do not count on their value",
+         but VIA assures us that they will all be zero (at least on the C7 and
+         Isaiah chips).
+        In the (unlikely) event a future processor uses bits 18, 19, 30, or 31
+         (0xC0C00000) for something else, we will have to add code to detect
+         the model to decide when it is appropriate to inspect them.*/
+      flags|=oc_parse_amd_flags(edx,ecx);
+    }
+  }
+  else{
+    /*Implement me.*/
+    flags=0;
+  }
+  return flags;
+}
+#endif

thirdparty/libtheora/x86/x86cpu.h

@@ -0,0 +1,36 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+ function:
+
+ ********************************************************************/
+
+#if !defined(_x86_x86cpu_H)
+# define _x86_x86cpu_H (1)
+#include "../internal.h"
+
+#define OC_CPU_X86_MMX      (1<<0)
+#define OC_CPU_X86_3DNOW    (1<<1)
+#define OC_CPU_X86_3DNOWEXT (1<<2)
+#define OC_CPU_X86_MMXEXT   (1<<3)
+#define OC_CPU_X86_SSE      (1<<4)
+#define OC_CPU_X86_SSE2     (1<<5)
+#define OC_CPU_X86_PNI      (1<<6)
+#define OC_CPU_X86_SSSE3    (1<<7)
+#define OC_CPU_X86_SSE4_1   (1<<8)
+#define OC_CPU_X86_SSE4_2   (1<<9)
+#define OC_CPU_X86_SSE4A    (1<<10)
+#define OC_CPU_X86_SSE5     (1<<11)
+
+ogg_uint32_t oc_cpu_flags_get(void);
+
+#endif

thirdparty/libtheora/x86/x86enc.c

@@ -0,0 +1,63 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+#include "x86enc.h"
+
+#if defined(OC_X86_ASM)
+
+void oc_enc_accel_init_x86(oc_enc_ctx *_enc){
+  ogg_uint32_t cpu_flags;
+  cpu_flags=_enc->state.cpu_flags;
+  oc_enc_accel_init_c(_enc);
+# if defined(OC_ENC_USE_VTABLE)
+  if(cpu_flags&OC_CPU_X86_MMX){
+    _enc->opt_vtable.frag_sub=oc_enc_frag_sub_mmx;
+    _enc->opt_vtable.frag_sub_128=oc_enc_frag_sub_128_mmx;
+    _enc->opt_vtable.frag_recon_intra=oc_frag_recon_intra_mmx;
+    _enc->opt_vtable.frag_recon_inter=oc_frag_recon_inter_mmx;
+  }
+  if(cpu_flags&OC_CPU_X86_MMXEXT){
+    _enc->opt_vtable.frag_sad=oc_enc_frag_sad_mmxext;
+    _enc->opt_vtable.frag_sad_thresh=oc_enc_frag_sad_thresh_mmxext;
+    _enc->opt_vtable.frag_sad2_thresh=oc_enc_frag_sad2_thresh_mmxext;
+    _enc->opt_vtable.frag_satd=oc_enc_frag_satd_mmxext;
+    _enc->opt_vtable.frag_satd2=oc_enc_frag_satd2_mmxext;
+    _enc->opt_vtable.frag_intra_satd=oc_enc_frag_intra_satd_mmxext;
+    _enc->opt_vtable.frag_copy2=oc_enc_frag_copy2_mmxext;
+    _enc->opt_vtable.fdct8x8=oc_enc_fdct8x8_mmxext;
+  }
+  if(cpu_flags&OC_CPU_X86_SSE2){
+#  if defined(OC_X86_64_ASM)
+    _enc->opt_vtable.fdct8x8=oc_enc_fdct8x8_x86_64sse2;
+#  endif
+    _enc->opt_vtable.frag_ssd=oc_enc_frag_ssd_sse2;
+    _enc->opt_vtable.frag_border_ssd=oc_enc_frag_border_ssd_sse2;
+    _enc->opt_vtable.frag_satd=oc_enc_frag_satd_sse2;
+    _enc->opt_vtable.frag_satd2=oc_enc_frag_satd2_sse2;
+    _enc->opt_vtable.frag_intra_satd=oc_enc_frag_intra_satd_sse2;
+    _enc->opt_vtable.enquant_table_init=oc_enc_enquant_table_init_x86;
+    _enc->opt_vtable.enquant_table_fixup=oc_enc_enquant_table_fixup_x86;
+    _enc->opt_vtable.quantize=oc_enc_quantize_sse2;
+# else
+    (void) cpu_flags;
+# endif
+    _enc->opt_data.enquant_table_size=128*sizeof(ogg_uint16_t);
+    _enc->opt_data.enquant_table_alignment=16;
+# if defined(OC_ENC_USE_VTABLE)
+  }
+# endif
+}
+#endif

thirdparty/libtheora/x86/x86enc.h

@@ -0,0 +1,114 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+#if !defined(_x86_x86enc_H)
+# define _x86_x86enc_H (1)
+# include "x86int.h"
+
+# if defined(OC_X86_ASM)
+#  define oc_enc_accel_init oc_enc_accel_init_x86
+#  if defined(OC_X86_64_ASM)
+/*x86-64 guarantees SIMD support up through at least SSE2.
+  If the best routine we have available only needs SSE2 (which at the moment
+   covers all of them), then we can avoid runtime detection and the indirect
+   call.*/
+#   define oc_enc_frag_sub(_enc,_diff,_x,_y,_stride) \
+  oc_enc_frag_sub_mmx(_diff,_x,_y,_stride)
+#   define oc_enc_frag_sub_128(_enc,_diff,_x,_stride) \
+  oc_enc_frag_sub_128_mmx(_diff,_x,_stride)
+#   define oc_enc_frag_sad(_enc,_src,_ref,_ystride) \
+  oc_enc_frag_sad_mmxext(_src,_ref,_ystride)
+#   define oc_enc_frag_sad_thresh(_enc,_src,_ref,_ystride,_thresh) \
+  oc_enc_frag_sad_thresh_mmxext(_src,_ref,_ystride,_thresh)
+#   define oc_enc_frag_sad2_thresh(_enc,_src,_ref1,_ref2,_ystride,_thresh) \
+  oc_enc_frag_sad2_thresh_mmxext(_src,_ref1,_ref2,_ystride,_thresh)
+#   define oc_enc_frag_satd(_enc,_dc,_src,_ref,_ystride) \
+  oc_enc_frag_satd_sse2(_dc,_src,_ref,_ystride)
+#   define oc_enc_frag_satd2(_enc,_dc,_src,_ref1,_ref2,_ystride) \
+  oc_enc_frag_satd2_sse2(_dc,_src,_ref1,_ref2,_ystride)
+#   define oc_enc_frag_intra_satd(_enc,_dc,_src,_ystride) \
+  oc_enc_frag_intra_satd_sse2(_dc,_src,_ystride)
+#   define oc_enc_frag_ssd(_enc,_src,_ref,_ystride) \
+  oc_enc_frag_ssd_sse2(_src,_ref,_ystride)
+#   define oc_enc_frag_border_ssd(_enc,_src,_ref,_ystride,_mask) \
+  oc_enc_frag_border_ssd_sse2(_src,_ref,_ystride,_mask)
+#   define oc_enc_frag_copy2(_enc,_dst,_src1,_src2,_ystride) \
+  oc_int_frag_copy2_mmxext(_dst,_ystride,_src1,_src2,_ystride)
+#   define oc_enc_enquant_table_init(_enc,_enquant,_dequant) \
+  oc_enc_enquant_table_init_x86(_enquant,_dequant)
+#   define oc_enc_enquant_table_fixup(_enc,_enquant,_nqis) \
+  oc_enc_enquant_table_fixup_x86(_enquant,_nqis)
+#  define oc_enc_quantize(_enc,_qdct,_dct,_dequant,_enquant) \
+  oc_enc_quantize_sse2(_qdct,_dct,_dequant,_enquant)
+#   define oc_enc_frag_recon_intra(_enc,_dst,_ystride,_residue) \
+  oc_frag_recon_intra_mmx(_dst,_ystride,_residue)
+#   define oc_enc_frag_recon_inter(_enc,_dst,_src,_ystride,_residue) \
+  oc_frag_recon_inter_mmx(_dst,_src,_ystride,_residue)
+#   define oc_enc_fdct8x8(_enc,_y,_x) \
+  oc_enc_fdct8x8_x86_64sse2(_y,_x)
+#  else
+#   define OC_ENC_USE_VTABLE (1)
+#  endif
+# endif
+
+# include "../encint.h"
+
+void oc_enc_accel_init_x86(oc_enc_ctx *_enc);
+
+void oc_enc_frag_sub_mmx(ogg_int16_t _diff[64],
+ const unsigned char *_x,const unsigned char *_y,int _stride);
+void oc_enc_frag_sub_128_mmx(ogg_int16_t _diff[64],
+ const unsigned char *_x,int _stride);
+unsigned oc_enc_frag_sad_mmxext(const unsigned char *_src,
+ const unsigned char *_ref,int _ystride);
+unsigned oc_enc_frag_sad_thresh_mmxext(const unsigned char *_src,
+ const unsigned char *_ref,int _ystride,unsigned _thresh);
+unsigned oc_enc_frag_sad2_thresh_mmxext(const unsigned char *_src,
+ const unsigned char *_ref1,const unsigned char *_ref2,int _ystride,
+ unsigned _thresh);
+unsigned oc_enc_frag_satd_mmxext(int *_dc,const unsigned char *_src,
+ const unsigned char *_ref,int _ystride);
+unsigned oc_enc_frag_satd_sse2(int *_dc,const unsigned char *_src,
+ const unsigned char *_ref,int _ystride);
+unsigned oc_enc_frag_satd2_mmxext(int *_dc,const unsigned char *_src,
+ const unsigned char *_ref1,const unsigned char *_ref2,int _ystride);
+unsigned oc_enc_frag_satd2_sse2(int *_dc,const unsigned char *_src,
+ const unsigned char *_ref1,const unsigned char *_ref2,int _ystride);
+unsigned oc_enc_frag_intra_satd_mmxext(int *_dc,
+ const unsigned char *_src,int _ystride);
+unsigned oc_enc_frag_intra_satd_sse2(int *_dc,
+ const unsigned char *_src,int _ystride);
+unsigned oc_enc_frag_ssd_sse2(const unsigned char *_src,
+ const unsigned char *_ref,int _ystride);
+unsigned oc_enc_frag_border_ssd_sse2(const unsigned char *_src,
+ const unsigned char *_ref,int _ystride,ogg_int64_t _mask);
+void oc_int_frag_copy2_mmxext(unsigned char *_dst,int _dst_ystride,
+ const unsigned char *_src1,const unsigned char *_src2,int _src_ystride);
+void oc_enc_frag_copy2_mmxext(unsigned char *_dst,
+ const unsigned char *_src1,const unsigned char *_src2,int _ystride);
+void oc_enc_enquant_table_init_x86(void *_enquant,
+ const ogg_uint16_t _dequant[64]);
+void oc_enc_enquant_table_fixup_x86(void *_enquant[3][3][2],int _nqis);
+int oc_enc_quantize_sse2(ogg_int16_t _qdct[64],const ogg_int16_t _dct[64],
+ const ogg_uint16_t _dequant[64],const void *_enquant);
+void oc_enc_fdct8x8_mmxext(ogg_int16_t _y[64],const ogg_int16_t _x[64]);
+
+# if defined(OC_X86_64_ASM)
+void oc_enc_fdct8x8_x86_64sse2(ogg_int16_t _y[64],const ogg_int16_t _x[64]);
+# endif
+
+#endif

thirdparty/libtheora/x86/x86enquant.c

@@ -0,0 +1,149 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+#include "x86enc.h"
+
+#if defined(OC_X86_ASM)
+
+
+
+/*The default enquant table is not quite suitable for SIMD purposes.
+  First, the m and l parameters need to be separated so that an entire row full
+   of m's or l's can be loaded at a time.
+  Second, x86 SIMD has no element-wise arithmetic right-shift, so we have to
+   emulate one with a multiply.
+  Therefore we translate the shift count into a scale factor.*/
+void oc_enc_enquant_table_init_x86(void *_enquant,
+ const ogg_uint16_t _dequant[64]){
+  ogg_int16_t *m;
+  ogg_int16_t *l;
+  int          zzi;
+  m=(ogg_int16_t *)_enquant;
+  l=m+64;
+  for(zzi=0;zzi<64;zzi++){
+    oc_iquant q;
+    oc_iquant_init(&q,_dequant[zzi]);
+    m[zzi]=q.m;
+    /*q.l must be at least 2 for this to work; fortunately, once all the scale
+       factors are baked in, the minimum quantizer is much larger than that.*/
+    l[zzi]=1<<16-q.l;
+  }
+}
+
+void oc_enc_enquant_table_fixup_x86(void *_enquant[3][3][2],int _nqis){
+  int pli;
+  int qii;
+  int qti;
+  for(pli=0;pli<3;pli++)for(qii=1;qii<_nqis;qii++)for(qti=0;qti<2;qti++){
+    ((ogg_int16_t *)_enquant[pli][qii][qti])[0]=
+     ((ogg_int16_t *)_enquant[pli][0][qti])[0];
+    ((ogg_int16_t *)_enquant[pli][qii][qti])[64]=
+     ((ogg_int16_t *)_enquant[pli][0][qti])[64];
+  }
+}
+
+int oc_enc_quantize_sse2(ogg_int16_t _qdct[64],const ogg_int16_t _dct[64],
+ const ogg_uint16_t _dequant[64],const void *_enquant){
+  ptrdiff_t r;
+  __asm__ __volatile__(
+    "xor %[r],%[r]\n\t"
+    /*Loop through two rows at a time.*/
+    ".p2align 4\n\t"
+    "0:\n\t"
+    /*Load the first two rows of the data and the quant matrices.*/
+    "movdqa 0x00(%[dct],%[r]),%%xmm0\n\t"
+    "movdqa 0x10(%[dct],%[r]),%%xmm1\n\t"
+    "movdqa 0x00(%[dq],%[r]),%%xmm2\n\t"
+    "movdqa 0x10(%[dq],%[r]),%%xmm3\n\t"
+    "movdqa 0x00(%[q],%[r]),%%xmm4\n\t"
+    "movdqa 0x10(%[q],%[r]),%%xmm5\n\t"
+    /*Double the input and propagate its sign to the rounding factor.
+      Using SSSE3's psignw would help here, but we need the mask later anyway.*/
+    "movdqa %%xmm0,%%xmm6\n\t"
+    "psraw $15,%%xmm0\n\t"
+    "movdqa %%xmm1,%%xmm7\n\t"
+    "paddw %%xmm6,%%xmm6\n\t"
+    "psraw $15,%%xmm1\n\t"
+    "paddw %%xmm7,%%xmm7\n\t"
+    "paddw %%xmm0,%%xmm2\n\t"
+    "paddw %%xmm1,%%xmm3\n\t"
+    "pxor %%xmm0,%%xmm2\n\t"
+    "pxor %%xmm1,%%xmm3\n\t"
+    /*Add the rounding factor and perform the first multiply.*/
+    "paddw %%xmm2,%%xmm6\n\t"
+    "paddw %%xmm3,%%xmm7\n\t"
+    "pmulhw %%xmm6,%%xmm4\n\t"
+    "pmulhw %%xmm7,%%xmm5\n\t"
+    "movdqa 0x80(%[q],%[r]),%%xmm2\n\t"
+    "movdqa 0x90(%[q],%[r]),%%xmm3\n\t"
+    "paddw %%xmm4,%%xmm6\n\t"
+    "paddw %%xmm5,%%xmm7\n\t"
+    /*Emulate an element-wise right-shift via a second multiply.*/
+    "pmulhw %%xmm2,%%xmm6\n\t"
+    "pmulhw %%xmm3,%%xmm7\n\t"
+    "add $32,%[r]\n\t"
+    "cmp $96,%[r]\n\t"
+    /*Correct for the sign.*/
+    "psubw %%xmm0,%%xmm6\n\t"
+    "psubw %%xmm1,%%xmm7\n\t"
+    /*Save the result.*/
+    "movdqa %%xmm6,-0x20(%[qdct],%[r])\n\t"
+    "movdqa %%xmm7,-0x10(%[qdct],%[r])\n\t"
+    "jle 0b\n\t"
+    /*Now find the location of the last non-zero value.*/
+    "movdqa 0x50(%[qdct]),%%xmm5\n\t"
+    "movdqa 0x40(%[qdct]),%%xmm4\n\t"
+    "packsswb %%xmm7,%%xmm6\n\t"
+    "packsswb %%xmm5,%%xmm4\n\t"
+    "pxor %%xmm0,%%xmm0\n\t"
+    "mov $-1,%k[dq]\n\t"
+    "pcmpeqb %%xmm0,%%xmm6\n\t"
+    "pcmpeqb %%xmm0,%%xmm4\n\t"
+    "pmovmskb %%xmm6,%k[q]\n\t"
+    "pmovmskb %%xmm4,%k[r]\n\t"
+    "shl $16,%k[q]\n\t"
+    "or %k[r],%k[q]\n\t"
+    "mov $32,%[r]\n\t"
+    /*We have to use xor here instead of not in order to set the flags.*/
+    "xor %k[dq],%k[q]\n\t"
+    "jnz 1f\n\t"
+    "movdqa 0x30(%[qdct]),%%xmm7\n\t"
+    "movdqa 0x20(%[qdct]),%%xmm6\n\t"
+    "movdqa 0x10(%[qdct]),%%xmm5\n\t"
+    "movdqa 0x00(%[qdct]),%%xmm4\n\t"
+    "packsswb %%xmm7,%%xmm6\n\t"
+    "packsswb %%xmm5,%%xmm4\n\t"
+    "pcmpeqb %%xmm0,%%xmm6\n\t"
+    "pcmpeqb %%xmm0,%%xmm4\n\t"
+    "pmovmskb %%xmm6,%k[q]\n\t"
+    "pmovmskb %%xmm4,%k[r]\n\t"
+    "shl $16,%k[q]\n\t"
+    "or %k[r],%k[q]\n\t"
+    "xor %[r],%[r]\n\t"
+    "not %k[q]\n\t"
+    "or $1,%k[q]\n\t"
+    "1:\n\t"
+    "bsr %k[q],%k[q]\n\t"
+    "add %k[q],%k[r]\n\t"
+    :[r]"=&a"(r),[q]"+r"(_enquant),[dq]"+r"(_dequant)
+    :[dct]"r"(_dct),[qdct]"r"(_qdct)
+    :"cc","memory"
+  );
+  return (int)r;
+}
+
+#endif

thirdparty/libtheora/x86/x86int.h

@@ -0,0 +1,122 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+#if !defined(_x86_x86int_H)
+# define _x86_x86int_H (1)
+# include "../internal.h"
+
+# if defined(OC_X86_ASM)
+#  define oc_state_accel_init oc_state_accel_init_x86
+#  if defined(OC_X86_64_ASM)
+/*x86-64 guarantees SIMD support up through at least SSE2.
+  If the best routine we have available only needs SSE2 (which at the moment
+   covers all of them), then we can avoid runtime detection and the indirect
+   call.*/
+#   define oc_frag_copy(_state,_dst,_src,_ystride) \
+  oc_frag_copy_mmx(_dst,_src,_ystride)
+#   define oc_frag_copy_list(_state,_dst_frame,_src_frame,_ystride, \
+ _fragis,_nfragis,_frag_buf_offs) \
+  oc_frag_copy_list_mmx(_dst_frame,_src_frame,_ystride, \
+   _fragis,_nfragis,_frag_buf_offs)
+#   define oc_frag_recon_intra(_state,_dst,_ystride,_residue) \
+  oc_frag_recon_intra_mmx(_dst,_ystride,_residue)
+#   define oc_frag_recon_inter(_state,_dst,_src,_ystride,_residue) \
+  oc_frag_recon_inter_mmx(_dst,_src,_ystride,_residue)
+#   define oc_frag_recon_inter2(_state,_dst,_src1,_src2,_ystride,_residue) \
+  oc_frag_recon_inter2_mmx(_dst,_src1,_src2,_ystride,_residue)
+#   define oc_idct8x8(_state,_y,_x,_last_zzi) \
+  oc_idct8x8_sse2(_y,_x,_last_zzi)
+#   define oc_state_frag_recon oc_state_frag_recon_mmx
+#   define oc_loop_filter_init(_state,_bv,_flimit) \
+  oc_loop_filter_init_mmxext(_bv,_flimit)
+#   define oc_state_loop_filter_frag_rows oc_state_loop_filter_frag_rows_mmxext
+#   define oc_restore_fpu(_state) \
+  oc_restore_fpu_mmx()
+#  else
+#   define OC_STATE_USE_VTABLE (1)
+#  endif
+# endif
+
+# include "../state.h"
+# include "x86cpu.h"
+
+/*Converts the expression in the argument to a string.*/
+#define OC_M2STR(_s) #_s
+
+/*Memory operands do not always include an offset.
+  To avoid warnings, we force an offset with %H (which adds 8).*/
+# if __GNUC_PREREQ(4,0)
+#  define OC_MEM_OFFS(_offs,_name) \
+  OC_M2STR(_offs-8+%H[_name])
+# endif
+/*If your gcc version doesn't support %H, then you get to suffer the warnings.
+  Note that Apple's gas breaks on things like _offs+(%esp): it throws away the
+   whole offset, instead of substituting in 0 for the missing operand to +.*/
+# if !defined(OC_MEM_OFFS)
+#  define OC_MEM_OFFS(_offs,_name) \
+  OC_M2STR(_offs+%[_name])
+# endif
+
+/*Declare an array operand with an exact size.
+  This tells gcc we're going to clobber this memory region, without having to
+   clobber all of "memory" and lets us access local buffers directly using the
+   stack pointer, without allocating a separate register to point to them.*/
+#define OC_ARRAY_OPERAND(_type,_ptr,_size) \
+  (*({ \
+    struct{_type array_value__[(_size)];} *array_addr__=(void *)(_ptr); \
+    array_addr__; \
+  }))
+
+/*Declare an array operand with an exact size.
+  This tells gcc we're going to clobber this memory region, without having to
+   clobber all of "memory" and lets us access local buffers directly using the
+   stack pointer, without allocating a separate register to point to them.*/
+#define OC_CONST_ARRAY_OPERAND(_type,_ptr,_size) \
+  (*({ \
+    const struct{_type array_value__[(_size)];} *array_addr__= \
+     (const void *)(_ptr); \
+    array_addr__; \
+  }))
+
+extern const unsigned short __attribute__((aligned(16))) OC_IDCT_CONSTS[64];
+
+void oc_state_accel_init_x86(oc_theora_state *_state);
+
+void oc_frag_copy_mmx(unsigned char *_dst,
+ const unsigned char *_src,int _ystride);
+void oc_frag_copy_list_mmx(unsigned char *_dst_frame,
+ const unsigned char *_src_frame,int _ystride,
+ const ptrdiff_t *_fragis,ptrdiff_t _nfragis,const ptrdiff_t *_frag_buf_offs);
+void oc_frag_recon_intra_mmx(unsigned char *_dst,int _ystride,
+ const ogg_int16_t *_residue);
+void oc_frag_recon_inter_mmx(unsigned char *_dst,
+ const unsigned char *_src,int _ystride,const ogg_int16_t *_residue);
+void oc_frag_recon_inter2_mmx(unsigned char *_dst,const unsigned char *_src1,
+ const unsigned char *_src2,int _ystride,const ogg_int16_t *_residue);
+void oc_idct8x8_mmx(ogg_int16_t _y[64],ogg_int16_t _x[64],int _last_zzi);
+void oc_idct8x8_sse2(ogg_int16_t _y[64],ogg_int16_t _x[64],int _last_zzi);
+void oc_state_frag_recon_mmx(const oc_theora_state *_state,ptrdiff_t _fragi,
+ int _pli,ogg_int16_t _dct_coeffs[128],int _last_zzi,ogg_uint16_t _dc_quant);
+void oc_loop_filter_init_mmx(signed char _bv[256],int _flimit);
+void oc_loop_filter_init_mmxext(signed char _bv[256],int _flimit);
+void oc_state_loop_filter_frag_rows_mmx(const oc_theora_state *_state,
+ signed char _bv[256],int _refi,int _pli,int _fragy0,int _fragy_end);
+void oc_state_loop_filter_frag_rows_mmxext(const oc_theora_state *_state,
+ signed char _bv[256],int _refi,int _pli,int _fragy0,int _fragy_end);
+void oc_restore_fpu_mmx(void);
+
+#endif

thirdparty/libtheora/x86/x86state.c

@@ -0,0 +1,97 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+#include "x86int.h"
+
+#if defined(OC_X86_ASM)
+
+#if defined(OC_STATE_USE_VTABLE)
+/*This table has been modified from OC_FZIG_ZAG by baking a 4x4 transpose into
+   each quadrant of the destination.*/
+static const unsigned char OC_FZIG_ZAG_MMX[128]={
+   0, 8, 1, 2, 9,16,24,17,
+  10, 3,32,11,18,25, 4,12,
+   5,26,19,40,33,34,41,48,
+  27, 6,13,20,28,21,14, 7,
+  56,49,42,35,43,50,57,36,
+  15,22,29,30,23,44,37,58,
+  51,59,38,45,52,31,60,53,
+  46,39,47,54,61,62,55,63,
+  64,64,64,64,64,64,64,64,
+  64,64,64,64,64,64,64,64,
+  64,64,64,64,64,64,64,64,
+  64,64,64,64,64,64,64,64,
+  64,64,64,64,64,64,64,64,
+  64,64,64,64,64,64,64,64,
+  64,64,64,64,64,64,64,64,
+  64,64,64,64,64,64,64,64
+};
+#endif
+
+/*This table has been modified from OC_FZIG_ZAG by baking an 8x8 transpose into
+   the destination.*/
+static const unsigned char OC_FZIG_ZAG_SSE2[128]={
+   0, 8, 1, 2, 9,16,24,17,
+  10, 3, 4,11,18,25,32,40,
+  33,26,19,12, 5, 6,13,20,
+  27,34,41,48,56,49,42,35,
+  28,21,14, 7,15,22,29,36,
+  43,50,57,58,51,44,37,30,
+  23,31,38,45,52,59,60,53,
+  46,39,47,54,61,62,55,63,
+  64,64,64,64,64,64,64,64,
+  64,64,64,64,64,64,64,64,
+  64,64,64,64,64,64,64,64,
+  64,64,64,64,64,64,64,64,
+  64,64,64,64,64,64,64,64,
+  64,64,64,64,64,64,64,64,
+  64,64,64,64,64,64,64,64,
+  64,64,64,64,64,64,64,64
+};
+
+void oc_state_accel_init_x86(oc_theora_state *_state){
+  oc_state_accel_init_c(_state);
+  _state->cpu_flags=oc_cpu_flags_get();
+# if defined(OC_STATE_USE_VTABLE)
+  if(_state->cpu_flags&OC_CPU_X86_MMX){
+    _state->opt_vtable.frag_copy=oc_frag_copy_mmx;
+    _state->opt_vtable.frag_copy_list=oc_frag_copy_list_mmx;
+    _state->opt_vtable.frag_recon_intra=oc_frag_recon_intra_mmx;
+    _state->opt_vtable.frag_recon_inter=oc_frag_recon_inter_mmx;
+    _state->opt_vtable.frag_recon_inter2=oc_frag_recon_inter2_mmx;
+    _state->opt_vtable.idct8x8=oc_idct8x8_mmx;
+    _state->opt_vtable.state_frag_recon=oc_state_frag_recon_mmx;
+    _state->opt_vtable.loop_filter_init=oc_loop_filter_init_mmx;
+    _state->opt_vtable.state_loop_filter_frag_rows=
+     oc_state_loop_filter_frag_rows_mmx;
+    _state->opt_vtable.restore_fpu=oc_restore_fpu_mmx;
+    _state->opt_data.dct_fzig_zag=OC_FZIG_ZAG_MMX;
+  }
+  if(_state->cpu_flags&OC_CPU_X86_MMXEXT){
+    _state->opt_vtable.loop_filter_init=oc_loop_filter_init_mmxext;
+    _state->opt_vtable.state_loop_filter_frag_rows=
+     oc_state_loop_filter_frag_rows_mmxext;
+  }
+  if(_state->cpu_flags&OC_CPU_X86_SSE2){
+    _state->opt_vtable.idct8x8=oc_idct8x8_sse2;
+# endif
+    _state->opt_data.dct_fzig_zag=OC_FZIG_ZAG_SSE2;
+# if defined(OC_STATE_USE_VTABLE)
+  }
+# endif
+}
+#endif

thirdparty/libtheora/x86/x86zigzag.h

@@ -0,0 +1,244 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+#if !defined(_x86_x86zigzag_H)
+# define _x86_x86zigzag_H (1)
+# include "x86enc.h"
+
+
+/*Converts DCT coefficients from transposed order into zig-zag scan order and
+   stores them in %[y].
+  This relies on two macros to load the contents of each row:
+   OC_ZZ_LOAD_ROW_LO(row,"reg") and OC_ZZ_LOAD_ROW_HI(row,"reg"), which load
+   the first four and second four entries of each row into the specified
+   register, respectively.
+  OC_ZZ_LOAD_ROW_LO must be called before OC_ZZ_LOAD_ROW_HI for the same row
+   (because when the rows are already in SSE2 registers, loading the high half
+   destructively modifies the register).
+  The index of each output element in the original 64-element array should wind
+   up in the following 8x8 matrix (the letters indicate the order we compute
+   each 4-tuple below):
+    A  0  8  1  2   9 16 24 17 B
+    C 10  3  4 11  18 25 32 40 E
+    F 33 26 19 12   5  6 13 20 D
+    G 27 34 41 48  56 49 42 35 I
+    L 28 21 14  7  15 22 29 36 M
+    H 43 50 57 58  51 44 37 30 O
+    N 23 31 38 45  52 59 60 53 J
+    P 46 39 47 54  61 62 55 63 K
+  The order of the coefficients within each tuple is reversed in the comments
+   below to reflect the usual MSB to LSB notation.*/
+#define OC_TRANSPOSE_ZIG_ZAG_MMXEXT \
+  OC_ZZ_LOAD_ROW_LO(0,"%%mm0")   /*mm0=03 02 01 00*/ \
+  OC_ZZ_LOAD_ROW_LO(1,"%%mm1")   /*mm1=11 10 09 08*/ \
+  OC_ZZ_LOAD_ROW_LO(2,"%%mm2")   /*mm2=19 18 17 16*/ \
+  OC_ZZ_LOAD_ROW_LO(3,"%%mm3")   /*mm3=27 26 25 24*/ \
+  OC_ZZ_LOAD_ROW_HI(0,"%%mm4")   /*mm4=07 06 05 04*/ \
+  OC_ZZ_LOAD_ROW_HI(1,"%%mm5")   /*mm5=15 14 13 12*/ \
+  OC_ZZ_LOAD_ROW_HI(2,"%%mm6")   /*mm6=23 22 21 20*/ \
+  "movq %%mm0,%%mm7\n\t"         /*mm7=03 02 01 00*/ \
+  "punpckhdq %%mm1,%%mm0\n\t"    /*mm0=11 10 03 02*/ \
+  "pshufw $0x39,%%mm4,%%mm4\n\t" /*mm4=04 07 06 05*/ \
+  "punpcklwd %%mm0,%%mm1\n\t"    /*mm1=03 09 02 08*/ \
+  "pshufw $0x39,%%mm5,%%mm5\n\t" /*mm5=12 15 14 13*/ \
+  "punpcklwd %%mm1,%%mm7\n\t"    /*mm7=02 01 08 00 *A*/ \
+  "movq %%mm7,0x00(%[y])\n\t" \
+  "punpckhwd %%mm4,%%mm1\n\t"    /*mm1=04 03 07 09*/ \
+  "movq %%mm2,%%mm7\n\t"         /*mm7=19 18 17 16*/ \
+  "punpckhdq %%mm1,%%mm0\n\t"    /*mm0=04 03 11 10*/ \
+  "punpckhwd %%mm5,%%mm7\n\t"    /*mm7=12 19 15 18*/ \
+  "punpcklwd %%mm3,%%mm1\n\t"    /*mm1=25 07 24 09*/ \
+  "punpcklwd %%mm6,%%mm5\n\t"    /*mm5=21 14 20 13*/ \
+  "punpcklwd %%mm2,%%mm1\n\t"    /*mm1=17 24 16 09 *B*/ \
+  OC_ZZ_LOAD_ROW_LO(4,"%%mm2")   /*mm2=35 34 33 32*/ \
+  "movq %%mm1,0x08(%[y])\n\t" \
+  OC_ZZ_LOAD_ROW_LO(5,"%%mm1")   /*mm1=43 42 41 40*/ \
+  "pshufw $0x78,%%mm0,%%mm0\n\t" /*mm0=11 04 03 10 *C*/ \
+  "movq %%mm0,0x10(%[y])\n\t" \
+  "punpckhdq %%mm4,%%mm6\n\t"    /*mm6=?? 07 23 22*/ \
+  "punpckldq %%mm5,%%mm4\n\t"    /*mm4=20 13 06 05 *D*/ \
+  "movq %%mm4,0x28(%[y])\n\t" \
+  "psrlq $16,%%mm3\n\t"          /*mm3=.. 27 26 25*/ \
+  "pshufw $0x0E,%%mm2,%%mm0\n\t" /*mm0=?? ?? 35 34*/ \
+  "movq %%mm7,%%mm4\n\t"         /*mm4=12 19 15 18*/ \
+  "punpcklwd %%mm3,%%mm2\n\t"    /*mm2=26 33 25 32*/ \
+  "punpcklwd %%mm1,%%mm4\n\t"    /*mm4=41 15 40 18*/ \
+  "punpckhwd %%mm1,%%mm3\n\t"    /*mm3=43 .. 42 27*/ \
+  "punpckldq %%mm2,%%mm4\n\t"    /*mm4=25 32 40 18*/ \
+  "punpcklwd %%mm0,%%mm3\n\t"    /*mm3=35 42 34 27*/ \
+  OC_ZZ_LOAD_ROW_LO(6,"%%mm0")   /*mm0=51 50 49 48*/ \
+  "pshufw $0x6C,%%mm4,%%mm4\n\t" /*mm4=40 32 25 18 *E*/ \
+  "movq %%mm4,0x18(%[y])\n\t" \
+  OC_ZZ_LOAD_ROW_LO(7,"%%mm4")   /*mm4=59 58 57 56*/ \
+  "punpckhdq %%mm7,%%mm2\n\t"    /*mm2=12 19 26 33 *F*/ \
+  "movq %%mm2,0x20(%[y])\n\t" \
+  "pshufw $0xD0,%%mm1,%%mm1\n\t" /*mm1=43 41 ?? ??*/ \
+  "pshufw $0x87,%%mm0,%%mm0\n\t" /*mm0=50 48 49 51*/ \
+  "movq %%mm3,%%mm2\n\t"         /*mm2=35 42 34 27*/ \
+  "punpckhwd %%mm0,%%mm1\n\t"    /*mm1=50 43 48 41*/ \
+  "pshufw $0x93,%%mm4,%%mm4\n\t" /*mm4=58 57 56 59*/ \
+  "punpckldq %%mm1,%%mm3\n\t"    /*mm3=48 41 34 27 *G*/ \
+  "movq %%mm3,0x30(%[y])\n\t" \
+  "punpckhdq %%mm4,%%mm1\n\t"    /*mm1=58 57 50 43 *H*/ \
+  "movq %%mm1,0x50(%[y])\n\t" \
+  OC_ZZ_LOAD_ROW_HI(7,"%%mm1")   /*mm1=63 62 61 60*/ \
+  "punpcklwd %%mm0,%%mm4\n\t"    /*mm4=49 56 51 59*/ \
+  OC_ZZ_LOAD_ROW_HI(6,"%%mm0")   /*mm0=55 54 53 52*/ \
+  "psllq $16,%%mm6\n\t"          /*mm6=07 23 22 ..*/ \
+  "movq %%mm4,%%mm3\n\t"         /*mm3=49 56 51 59*/ \
+  "punpckhdq %%mm2,%%mm4\n\t"    /*mm4=35 42 49 56 *I*/ \
+  OC_ZZ_LOAD_ROW_HI(3,"%%mm2")   /*mm2=31 30 29 28*/ \
+  "movq %%mm4,0x38(%[y])\n\t" \
+  "punpcklwd %%mm1,%%mm3\n\t"    /*mm3=61 51 60 59*/ \
+  "punpcklwd %%mm6,%%mm7\n\t"    /*mm7=22 15 .. ??*/ \
+  "movq %%mm3,%%mm4\n\t"         /*mm4=61 51 60 59*/ \
+  "punpcklwd %%mm0,%%mm3\n\t"    /*mm3=53 60 52 59*/ \
+  "punpckhwd %%mm0,%%mm4\n\t"    /*mm4=55 61 54 51*/ \
+  OC_ZZ_LOAD_ROW_HI(4,"%%mm0")   /*mm0=39 38 37 36*/ \
+  "pshufw $0xE1,%%mm3,%%mm3\n\t" /*mm3=53 60 59 52 *J*/ \
+  "movq %%mm3,0x68(%[y])\n\t" \
+  "movq %%mm4,%%mm3\n\t"         /*mm3=?? ?? 54 51*/ \
+  "pshufw $0x39,%%mm2,%%mm2\n\t" /*mm2=28 31 30 29*/ \
+  "punpckhwd %%mm1,%%mm4\n\t"    /*mm4=63 55 62 61 *K*/ \
+  OC_ZZ_LOAD_ROW_HI(5,"%%mm1")   /*mm1=47 46 45 44*/ \
+  "movq %%mm4,0x78(%[y])\n\t" \
+  "punpckhwd %%mm2,%%mm6\n\t"    /*mm6=28 07 31 23*/ \
+  "punpcklwd %%mm0,%%mm2\n\t"    /*mm2=37 30 36 29*/ \
+  "punpckhdq %%mm6,%%mm5\n\t"    /*mm5=28 07 21 14*/ \
+  "pshufw $0x4B,%%mm2,%%mm2\n\t" /*mm2=36 29 30 37*/ \
+  "pshufw $0x87,%%mm5,%%mm5\n\t" /*mm5=07 14 21 28 *L*/ \
+  "movq %%mm5,0x40(%[y])\n\t" \
+  "punpckhdq %%mm2,%%mm7\n\t"    /*mm7=36 29 22 15 *M*/ \
+  "movq %%mm7,0x48(%[y])\n\t" \
+  "pshufw $0x9C,%%mm1,%%mm1\n\t" /*mm1=46 45 47 44*/ \
+  "punpckhwd %%mm1,%%mm0\n\t"    /*mm0=46 39 45 38*/ \
+  "punpcklwd %%mm1,%%mm3\n\t"    /*mm3=47 54 44 51*/ \
+  "punpckldq %%mm0,%%mm6\n\t"    /*mm6=45 38 31 23 *N*/ \
+  "movq %%mm6,0x60(%[y])\n\t" \
+  "punpckhdq %%mm3,%%mm0\n\t"    /*mm0=47 54 46 39*/ \
+  "punpckldq %%mm2,%%mm3\n\t"    /*mm3=30 37 44 51 *O*/ \
+  "movq %%mm3,0x58(%[y])\n\t" \
+  "pshufw $0xB1,%%mm0,%%mm0\n\t" /*mm0=54 47 39 46 *P*/ \
+  "movq %%mm0,0x70(%[y])\n\t" \
+
+/*Converts DCT coefficients in %[dct] from natural order into zig-zag scan
+   order and stores them in %[qdct].
+  The index of each output element in the original 64-element array should wind
+   up in the following 8x8 matrix (the letters indicate the order we compute
+   each 4-tuple below):
+    A  0  1  8 16   9  2  3 10 B
+    C 17 24 32 25  18 11  4  5 D
+    E 12 19 26 33  40 48 41 34 I
+    H 27 20 13  6   7 14 21 28 G
+    K 35 42 49 56  57 50 43 36 J
+    F 29 22 15 23  30 37 44 51 M
+    P 58 59 52 45  38 31 39 46 L
+    N 53 60 61 54  47 55 62 63 O
+  The order of the coefficients within each tuple is reversed in the comments
+   below to reflect the usual MSB to LSB notation.*/
+#define OC_ZIG_ZAG_MMXEXT \
+  "movq 0x00(%[dct]),%%mm0\n\t"  /*mm0=03 02 01 00*/ \
+  "movq 0x08(%[dct]),%%mm1\n\t"  /*mm1=07 06 05 04*/ \
+  "movq 0x10(%[dct]),%%mm2\n\t"  /*mm2=11 10 09 08*/ \
+  "movq 0x20(%[dct]),%%mm3\n\t"  /*mm3=19 18 17 16*/ \
+  "movq 0x30(%[dct]),%%mm4\n\t"  /*mm4=27 26 25 24*/ \
+  "movq 0x40(%[dct]),%%mm5\n\t"  /*mm5=35 34 33 32*/ \
+  "movq %%mm2,%%mm7\n\t"         /*mm7=11 10 09 08*/ \
+  "punpcklwd %%mm3,%%mm2\n\t"    /*mm2=17 09 16 08*/ \
+  "movq %%mm0,%%mm6\n\t"         /*mm6=03 02 01 00*/ \
+  "punpckldq %%mm2,%%mm0\n\t"    /*mm0=16 08 01 00 *A*/ \
+  "movq %%mm0,0x00(%[qdct])\n\t" \
+  "movq 0x18(%[dct]),%%mm0\n\t"  /*mm0=15 14 13 12*/ \
+  "punpckhdq %%mm6,%%mm6\n\t"    /*mm6=03 02 03 02*/ \
+  "psrlq $16,%%mm7\n\t"          /*mm7=.. 11 10 09*/ \
+  "punpckldq %%mm7,%%mm6\n\t"    /*mm6=10 09 03 02*/ \
+  "punpckhwd %%mm7,%%mm3\n\t"    /*mm3=.. 19 11 18*/ \
+  "pshufw $0xD2,%%mm6,%%mm6\n\t" /*mm6=10 03 02 09 *B*/ \
+  "movq %%mm6,0x08(%[qdct])\n\t" \
+  "psrlq $48,%%mm2\n\t"          /*mm2=.. .. .. 17*/ \
+  "movq %%mm1,%%mm6\n\t"         /*mm6=07 06 05 04*/ \
+  "punpcklwd %%mm5,%%mm2\n\t"    /*mm2=33 .. 32 17*/ \
+  "movq %%mm3,%%mm7\n\t"         /*mm7=.. 19 11 18*/ \
+  "punpckldq %%mm1,%%mm3\n\t"    /*mm3=05 04 11 18 *C*/ \
+  "por %%mm2,%%mm7\n\t"          /*mm7=33 19 ?? ??*/ \
+  "punpcklwd %%mm4,%%mm2\n\t"    /*mm2=25 32 24 17 *D**/ \
+  "movq %%mm2,0x10(%[qdct])\n\t" \
+  "movq %%mm3,0x18(%[qdct])\n\t" \
+  "movq 0x28(%[dct]),%%mm2\n\t"  /*mm2=23 22 21 20*/ \
+  "movq 0x38(%[dct]),%%mm1\n\t"  /*mm1=31 30 29 28*/ \
+  "pshufw $0x9C,%%mm0,%%mm3\n\t" /*mm3=14 13 15 12*/ \
+  "punpckhdq %%mm7,%%mm7\n\t"    /*mm7=33 19 33 19*/ \
+  "punpckhwd %%mm3,%%mm6\n\t"    /*mm6=14 07 13 06*/ \
+  "punpckldq %%mm0,%%mm0\n\t"    /*mm0=13 12 13 12*/ \
+  "punpcklwd %%mm1,%%mm3\n\t"    /*mm3=29 15 28 12*/ \
+  "punpckhwd %%mm4,%%mm0\n\t"    /*mm0=27 13 26 12*/ \
+  "pshufw $0xB4,%%mm3,%%mm3\n\t" /*mm3=15 29 28 12*/ \
+  "psrlq $48,%%mm4\n\t"          /*mm4=.. .. .. 27*/ \
+  "punpcklwd %%mm7,%%mm0\n\t"    /*mm0=33 26 19 12 *E*/ \
+  "punpcklwd %%mm1,%%mm4\n\t"    /*mm4=29 .. 28 27*/ \
+  "punpckhwd %%mm2,%%mm3\n\t"    /*mm3=23 15 22 29 *F*/ \
+  "movq %%mm0,0x20(%[qdct])\n\t" \
+  "movq %%mm3,0x50(%[qdct])\n\t" \
+  "movq 0x60(%[dct]),%%mm3\n\t"  /*mm3=51 50 49 48*/ \
+  "movq 0x70(%[dct]),%%mm7\n\t"  /*mm7=59 58 57 56*/ \
+  "movq 0x50(%[dct]),%%mm0\n\t"  /*mm0=43 42 41 40*/ \
+  "punpcklwd %%mm4,%%mm2\n\t"    /*mm2=28 21 27 20*/ \
+  "psrlq $32,%%mm5\n\t"          /*mm5=.. .. 35 34*/ \
+  "movq %%mm2,%%mm4\n\t"         /*mm4=28 21 27 20*/ \
+  "punpckldq %%mm6,%%mm2\n\t"    /*mm2=13 06 27 20*/ \
+  "punpckhdq %%mm4,%%mm6\n\t"    /*mm6=28 21 14 07 *G*/ \
+  "movq %%mm3,%%mm4\n\t"         /*mm4=51 50 49 48*/ \
+  "pshufw $0xB1,%%mm2,%%mm2\n\t" /*mm2=06 13 20 27 *H*/ \
+  "movq %%mm2,0x30(%[qdct])\n\t" \
+  "movq %%mm6,0x38(%[qdct])\n\t" \
+  "movq 0x48(%[dct]),%%mm2\n\t"  /*mm2=39 38 37 36*/ \
+  "punpcklwd %%mm5,%%mm4\n\t"    /*mm4=35 49 34 48*/ \
+  "movq 0x58(%[dct]),%%mm5\n\t"  /*mm5=47 46 45 44*/ \
+  "punpckldq %%mm7,%%mm6\n\t"    /*mm6=57 56 14 07*/ \
+  "psrlq $32,%%mm3\n\t"          /*mm3=.. .. 51 50*/ \
+  "punpckhwd %%mm0,%%mm6\n\t"    /*mm6=43 57 42 56*/ \
+  "punpcklwd %%mm4,%%mm0\n\t"    /*mm0=34 41 48 40 *I*/ \
+  "pshufw $0x4E,%%mm6,%%mm6\n\t" /*mm6=42 56 43 57*/ \
+  "movq %%mm0,0x28(%[qdct])\n\t" \
+  "punpcklwd %%mm2,%%mm3\n\t"    /*mm3=37 51 36 50*/ \
+  "punpckhwd %%mm6,%%mm4\n\t"    /*mm4=42 35 56 49*/ \
+  "punpcklwd %%mm3,%%mm6\n\t"    /*mm6=36 43 50 57 *J*/ \
+  "pshufw $0x4E,%%mm4,%%mm4\n\t" /*mm4=56 49 42 35 *K*/ \
+  "movq %%mm4,0x40(%[qdct])\n\t" \
+  "movq %%mm6,0x48(%[qdct])\n\t" \
+  "movq 0x68(%[dct]),%%mm6\n\t"  /*mm6=55 54 53 52*/ \
+  "movq 0x78(%[dct]),%%mm0\n\t"  /*mm0=63 62 61 60*/ \
+  "psrlq $32,%%mm1\n\t"          /*mm1=.. .. 31 30*/ \
+  "pshufw $0xD8,%%mm5,%%mm5\n\t" /*mm5=47 45 46 44*/ \
+  "pshufw $0x0B,%%mm3,%%mm3\n\t" /*mm3=50 50 51 37*/ \
+  "punpcklwd %%mm5,%%mm1\n\t"    /*mm1=46 31 44 30*/ \
+  "pshufw $0xC9,%%mm6,%%mm6\n\t" /*mm6=55 52 54 53*/ \
+  "punpckhwd %%mm1,%%mm2\n\t"    /*mm2=46 39 31 38 *L*/ \
+  "punpcklwd %%mm3,%%mm1\n\t"    /*mm1=51 44 37 30 *M*/ \
+  "movq %%mm2,0x68(%[qdct])\n\t" \
+  "movq %%mm1,0x58(%[qdct])\n\t" \
+  "punpckhwd %%mm6,%%mm5\n\t"    /*mm5=55 47 52 45*/ \
+  "punpckldq %%mm0,%%mm6\n\t"    /*mm6=61 60 54 53*/ \
+  "pshufw $0x10,%%mm5,%%mm4\n\t" /*mm4=45 52 45 45*/ \
+  "pshufw $0x78,%%mm6,%%mm6\n\t" /*mm6=53 60 61 54 *N*/ \
+  "punpckhdq %%mm0,%%mm5\n\t"    /*mm5=63 62 55 47 *O*/ \
+  "punpckhdq %%mm4,%%mm7\n\t"    /*mm7=45 52 59 58 *P*/ \
+  "movq %%mm6,0x70(%[qdct])\n\t" \
+  "movq %%mm5,0x78(%[qdct])\n\t" \
+  "movq %%mm7,0x60(%[qdct])\n\t" \
+
+#endif

thirdparty/libtheora/x86_vc/mmxencfrag.c

@@ -0,0 +1,982 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation https://www.xiph.org/                 *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+#include <stddef.h>
+#include "x86enc.h"
+
+#if defined(OC_X86_ASM)
+
+unsigned oc_enc_frag_sad_mmxext(const unsigned char *_src,
+ const unsigned char *_ref,int _ystride){
+  ptrdiff_t ret;
+  __asm{
+#define SRC esi
+#define REF edx
+#define YSTRIDE ecx
+#define YSTRIDE3 edi
+    mov YSTRIDE,_ystride
+    mov SRC,_src
+    mov REF,_ref
+    /*Load the first 4 rows of each block.*/
+    movq mm0,[SRC]
+    movq mm1,[REF]
+    movq mm2,[SRC][YSTRIDE]
+    movq mm3,[REF][YSTRIDE]
+    lea YSTRIDE3,[YSTRIDE+YSTRIDE*2]
+    movq mm4,[SRC+YSTRIDE*2]
+    movq mm5,[REF+YSTRIDE*2]
+    movq mm6,[SRC+YSTRIDE3]
+    movq mm7,[REF+YSTRIDE3]
+    /*Compute their SADs and add them in mm0*/
+    psadbw mm0,mm1
+    psadbw mm2,mm3
+    lea SRC,[SRC+YSTRIDE*4]
+    paddw mm0,mm2
+    lea REF,[REF+YSTRIDE*4]
+    /*Load the next 3 rows as registers become available.*/
+    movq mm2,[SRC]
+    movq mm3,[REF]
+    psadbw mm4,mm5
+    psadbw mm6,mm7
+    paddw mm0,mm4
+    movq mm5,[REF+YSTRIDE]
+    movq mm4,[SRC+YSTRIDE]
+    paddw mm0,mm6
+    movq mm7,[REF+YSTRIDE*2]
+    movq mm6,[SRC+YSTRIDE*2]
+    /*Start adding their SADs to mm0*/
+    psadbw mm2,mm3
+    psadbw mm4,mm5
+    paddw mm0,mm2
+    psadbw mm6,mm7
+    /*Load last row as registers become available.*/
+    movq mm2,[SRC+YSTRIDE3]
+    movq mm3,[REF+YSTRIDE3]
+    /*And finish adding up their SADs.*/
+    paddw mm0,mm4
+    psadbw mm2,mm3
+    paddw mm0,mm6
+    paddw mm0,mm2
+    movd [ret],mm0
+#undef SRC
+#undef REF
+#undef YSTRIDE
+#undef YSTRIDE3
+  }
+  return (unsigned)ret;
+}
+
+unsigned oc_enc_frag_sad_thresh_mmxext(const unsigned char *_src,
+ const unsigned char *_ref,int _ystride,unsigned _thresh){
+  /*Early termination is for suckers.*/
+  return oc_enc_frag_sad_mmxext(_src,_ref,_ystride);
+}
+
+#define OC_SAD2_LOOP __asm{ \
+  /*We want to compute (mm0+mm1>>1) on unsigned bytes without overflow, but \
+     pavgb computes (mm0+mm1+1>>1). \
+   The latter is exactly 1 too large when the low bit of two corresponding \
+    bytes is only set in one of them. \
+   Therefore we pxor the operands, pand to mask out the low bits, and psubb to \
+    correct the output of pavgb.*/ \
+  __asm  movq mm6,mm0 \
+  __asm  lea REF1,[REF1+YSTRIDE*2] \
+  __asm  pxor mm0,mm1 \
+  __asm  pavgb mm6,mm1 \
+  __asm  lea REF2,[REF2+YSTRIDE*2] \
+  __asm  movq mm1,mm2 \
+  __asm  pand mm0,mm7 \
+  __asm  pavgb mm2,mm3 \
+  __asm  pxor mm1,mm3 \
+  __asm  movq mm3,[REF2+YSTRIDE] \
+  __asm  psubb mm6,mm0 \
+  __asm  movq mm0,[REF1] \
+  __asm  pand mm1,mm7 \
+  __asm  psadbw mm4,mm6 \
+  __asm  movd mm6,RET \
+  __asm  psubb mm2,mm1 \
+  __asm  movq mm1,[REF2] \
+  __asm  lea SRC,[SRC+YSTRIDE*2] \
+  __asm  psadbw mm5,mm2 \
+  __asm  movq mm2,[REF1+YSTRIDE] \
+  __asm  paddw mm5,mm4 \
+  __asm  movq mm4,[SRC] \
+  __asm  paddw mm6,mm5 \
+  __asm  movq mm5,[SRC+YSTRIDE] \
+  __asm  movd RET,mm6 \
+}
+
+/*Same as above, but does not pre-load the next two rows.*/
+#define OC_SAD2_TAIL __asm{ \
+  __asm  movq mm6,mm0 \
+  __asm  pavgb mm0,mm1 \
+  __asm  pxor mm6,mm1 \
+  __asm  movq mm1,mm2 \
+  __asm  pand mm6,mm7 \
+  __asm  pavgb mm2,mm3 \
+  __asm  pxor mm1,mm3 \
+  __asm  psubb mm0,mm6 \
+  __asm  pand mm1,mm7 \
+  __asm  psadbw mm4,mm0 \
+  __asm  psubb mm2,mm1 \
+  __asm  movd mm6,RET \
+  __asm  psadbw mm5,mm2 \
+  __asm  paddw mm5,mm4 \
+  __asm  paddw mm6,mm5 \
+  __asm  movd RET,mm6 \
+}
+
+unsigned oc_enc_frag_sad2_thresh_mmxext(const unsigned char *_src,
+ const unsigned char *_ref1,const unsigned char *_ref2,int _ystride,
+ unsigned _thresh){
+  ptrdiff_t ret;
+  __asm{
+#define REF1 ecx
+#define REF2 edi
+#define YSTRIDE esi
+#define SRC edx
+#define RET eax
+    mov YSTRIDE,_ystride
+    mov SRC,_src
+    mov REF1,_ref1
+    mov REF2,_ref2
+    movq mm0,[REF1]
+    movq mm1,[REF2]
+    movq mm2,[REF1+YSTRIDE]
+    movq mm3,[REF2+YSTRIDE]
+    xor RET,RET
+    movq mm4,[SRC]
+    pxor mm7,mm7
+    pcmpeqb mm6,mm6
+    movq mm5,[SRC+YSTRIDE]
+    psubb mm7,mm6
+    OC_SAD2_LOOP
+    OC_SAD2_LOOP
+    OC_SAD2_LOOP
+    OC_SAD2_TAIL
+    mov [ret],RET
+#undef REF1
+#undef REF2
+#undef YSTRIDE
+#undef SRC
+#undef RET
+  }
+  return (unsigned)ret;
+}
+
+/*Load an 8x4 array of pixel values from %[src] and %[ref] and compute their
+  16-bit difference in mm0...mm7.*/
+#define OC_LOAD_SUB_8x4(_off) __asm{ \
+  __asm  movd mm0,[_off+SRC] \
+  __asm  movd mm4,[_off+REF] \
+  __asm  movd mm1,[_off+SRC+SRC_YSTRIDE] \
+  __asm  lea SRC,[SRC+SRC_YSTRIDE*2] \
+  __asm  movd mm5,[_off+REF+REF_YSTRIDE] \
+  __asm  lea REF,[REF+REF_YSTRIDE*2] \
+  __asm  movd mm2,[_off+SRC] \
+  __asm  movd mm7,[_off+REF] \
+  __asm  movd mm3,[_off+SRC+SRC_YSTRIDE] \
+  __asm  movd mm6,[_off+REF+REF_YSTRIDE] \
+  __asm  punpcklbw mm0,mm4 \
+  __asm  lea SRC,[SRC+SRC_YSTRIDE*2] \
+  __asm  punpcklbw mm4,mm4 \
+  __asm  lea REF,[REF+REF_YSTRIDE*2] \
+  __asm  psubw mm0,mm4 \
+  __asm  movd mm4,[_off+SRC] \
+  __asm  movq [_off*2+BUF],mm0 \
+  __asm  movd mm0,[_off+REF] \
+  __asm  punpcklbw mm1,mm5 \
+  __asm  punpcklbw mm5,mm5 \
+  __asm  psubw mm1,mm5 \
+  __asm  movd mm5,[_off+SRC+SRC_YSTRIDE] \
+  __asm  punpcklbw mm2,mm7 \
+  __asm  punpcklbw mm7,mm7 \
+  __asm  psubw mm2,mm7 \
+  __asm  movd mm7,[_off+REF+REF_YSTRIDE] \
+  __asm  punpcklbw mm3,mm6 \
+  __asm  lea SRC,[SRC+SRC_YSTRIDE*2] \
+  __asm  punpcklbw mm6,mm6 \
+  __asm  psubw mm3,mm6 \
+  __asm  movd mm6,[_off+SRC] \
+  __asm  punpcklbw mm4,mm0 \
+  __asm  lea REF,[REF+REF_YSTRIDE*2] \
+  __asm  punpcklbw mm0,mm0 \
+  __asm  lea SRC,[SRC+SRC_YSTRIDE*2] \
+  __asm  psubw mm4,mm0 \
+  __asm  movd mm0,[_off+REF] \
+  __asm  punpcklbw mm5,mm7 \
+  __asm  neg SRC_YSTRIDE \
+  __asm  punpcklbw mm7,mm7 \
+  __asm  psubw mm5,mm7 \
+  __asm  movd mm7,[_off+SRC+SRC_YSTRIDE] \
+  __asm  punpcklbw mm6,mm0 \
+  __asm  lea REF,[REF+REF_YSTRIDE*2] \
+  __asm  punpcklbw mm0,mm0 \
+  __asm  neg REF_YSTRIDE \
+  __asm  psubw mm6,mm0 \
+  __asm  movd mm0,[_off+REF+REF_YSTRIDE] \
+  __asm  lea SRC,[SRC+SRC_YSTRIDE*8] \
+  __asm  punpcklbw mm7,mm0 \
+  __asm  neg SRC_YSTRIDE \
+  __asm  punpcklbw mm0,mm0 \
+  __asm  lea REF,[REF+REF_YSTRIDE*8] \
+  __asm  psubw mm7,mm0 \
+  __asm  neg REF_YSTRIDE \
+  __asm  movq mm0,[_off*2+BUF] \
+}
+
+/*Load an 8x4 array of pixel values from %[src] into %%mm0...%%mm7.*/
+#define OC_LOAD_8x4(_off) __asm{ \
+  __asm  movd mm0,[_off+SRC] \
+  __asm  movd mm1,[_off+SRC+YSTRIDE] \
+  __asm  movd mm2,[_off+SRC+YSTRIDE*2] \
+  __asm  pxor mm7,mm7 \
+  __asm  movd mm3,[_off+SRC+YSTRIDE3] \
+  __asm  punpcklbw mm0,mm7 \
+  __asm  movd mm4,[_off+SRC4] \
+  __asm  punpcklbw mm1,mm7 \
+  __asm  movd mm5,[_off+SRC4+YSTRIDE] \
+  __asm  punpcklbw mm2,mm7 \
+  __asm  movd mm6,[_off+SRC4+YSTRIDE*2] \
+  __asm  punpcklbw mm3,mm7 \
+  __asm  movd mm7,[_off+SRC4+YSTRIDE3] \
+  __asm  punpcklbw mm4,mm4 \
+  __asm  punpcklbw mm5,mm5 \
+  __asm  psrlw mm4,8 \
+  __asm  psrlw mm5,8 \
+  __asm  punpcklbw mm6,mm6 \
+  __asm  punpcklbw mm7,mm7 \
+  __asm  psrlw mm6,8 \
+  __asm  psrlw mm7,8 \
+}
+
+/*Performs the first two stages of an 8-point 1-D Hadamard transform.
+  The transform is performed in place, except that outputs 0-3 are swapped with
+   outputs 4-7.
+  Outputs 2, 3, 6, and 7 from the second stage are negated (which allows us to
+   perform this stage in place with no temporary registers).*/
+#define OC_HADAMARD_AB_8x4 __asm{ \
+  /*Stage A: \
+    Outputs 0-3 are swapped with 4-7 here.*/ \
+  __asm  paddw mm5,mm1 \
+  __asm  paddw mm6,mm2 \
+  __asm  paddw mm1,mm1 \
+  __asm  paddw mm2,mm2 \
+  __asm  psubw mm1,mm5 \
+  __asm  psubw mm2,mm6 \
+  __asm  paddw mm7,mm3 \
+  __asm  paddw mm4,mm0 \
+  __asm  paddw mm3,mm3 \
+  __asm  paddw mm0,mm0 \
+  __asm  psubw mm3,mm7 \
+  __asm  psubw mm0,mm4 \
+   /*Stage B:*/ \
+  __asm  paddw mm0,mm2 \
+  __asm  paddw mm1,mm3 \
+  __asm  paddw mm4,mm6 \
+  __asm  paddw mm5,mm7 \
+  __asm  paddw mm2,mm2 \
+  __asm  paddw mm3,mm3 \
+  __asm  paddw mm6,mm6 \
+  __asm  paddw mm7,mm7 \
+  __asm  psubw mm2,mm0 \
+  __asm  psubw mm3,mm1 \
+  __asm  psubw mm6,mm4 \
+  __asm  psubw mm7,mm5 \
+}
+
+/*Performs the last stage of an 8-point 1-D Hadamard transform in place.
+  Outputs 1, 3, 5, and 7 are negated (which allows us to perform this stage in
+   place with no temporary registers).*/
+#define OC_HADAMARD_C_8x4 __asm{ \
+  /*Stage C:*/ \
+  __asm  paddw mm0,mm1 \
+  __asm  paddw mm2,mm3 \
+  __asm  paddw mm4,mm5 \
+  __asm  paddw mm6,mm7 \
+  __asm  paddw mm1,mm1 \
+  __asm  paddw mm3,mm3 \
+  __asm  paddw mm5,mm5 \
+  __asm  paddw mm7,mm7 \
+  __asm  psubw mm1,mm0 \
+  __asm  psubw mm3,mm2 \
+  __asm  psubw mm5,mm4 \
+  __asm  psubw mm7,mm6 \
+}
+
+/*Performs an 8-point 1-D Hadamard transform.
+  The transform is performed in place, except that outputs 0-3 are swapped with
+   outputs 4-7.
+  Outputs 1, 2, 5 and 6 are negated (which allows us to perform the transform
+   in place with no temporary registers).*/
+#define OC_HADAMARD_8x4 __asm{ \
+  OC_HADAMARD_AB_8x4 \
+  OC_HADAMARD_C_8x4 \
+}
+
+/*Performs the first part of the final stage of the Hadamard transform and
+   summing of absolute values.
+  At the end of this part, mm1 will contain the DC coefficient of the
+   transform.*/
+#define OC_HADAMARD_C_ABS_ACCUM_A_8x4(_r6,_r7) __asm{ \
+  /*We use the fact that \
+      (abs(a+b)+abs(a-b))/2=max(abs(a),abs(b)) \
+     to merge the final butterfly with the abs and the first stage of \
+     accumulation. \
+    Thus we can avoid using pabsw, which is not available until SSSE3. \
+    Emulating pabsw takes 3 instructions, so the straightforward MMXEXT \
+     implementation would be (3+3)*8+7=55 instructions (+4 for spilling \
+     registers). \
+    Even with pabsw, it would be (3+1)*8+7=39 instructions (with no spills). \
+    This implementation is only 26 (+4 for spilling registers).*/ \
+  __asm  movq [_r7+BUF],mm7 \
+  __asm  movq [_r6+BUF],mm6 \
+  /*mm7={0x7FFF}x4 \
+    mm0=max(abs(mm0),abs(mm1))-0x7FFF*/ \
+  __asm  pcmpeqb mm7,mm7 \
+  __asm  movq mm6,mm0 \
+  __asm  psrlw mm7,1 \
+  __asm  paddw mm6,mm1 \
+  __asm  pmaxsw mm0,mm1 \
+  __asm  paddsw mm6,mm7 \
+  __asm  psubw mm0,mm6 \
+  /*mm2=max(abs(mm2),abs(mm3))-0x7FFF \
+    mm4=max(abs(mm4),abs(mm5))-0x7FFF*/ \
+  __asm  movq mm6,mm2 \
+  __asm  movq mm1,mm4 \
+  __asm  pmaxsw mm2,mm3 \
+  __asm  pmaxsw mm4,mm5 \
+  __asm  paddw mm6,mm3 \
+  __asm  paddw mm1,mm5 \
+  __asm  movq mm3,[_r7+BUF] \
+}
+
+/*Performs the second part of the final stage of the Hadamard transform and
+   summing of absolute values.*/
+#define OC_HADAMARD_C_ABS_ACCUM_B_8x4(_r6,_r7) __asm{ \
+  __asm  paddsw mm6,mm7 \
+  __asm  movq mm5,[_r6+BUF] \
+  __asm  paddsw mm1,mm7 \
+  __asm  psubw mm2,mm6 \
+  __asm  psubw mm4,mm1 \
+  /*mm7={1}x4 (needed for the horizontal add that follows) \
+    mm0+=mm2+mm4+max(abs(mm3),abs(mm5))-0x7FFF*/ \
+  __asm  movq mm6,mm3 \
+  __asm  pmaxsw mm3,mm5 \
+  __asm  paddw mm0,mm2 \
+  __asm  paddw mm6,mm5 \
+  __asm  paddw mm0,mm4 \
+  __asm  paddsw mm6,mm7 \
+  __asm  paddw mm0,mm3 \
+  __asm  psrlw mm7,14 \
+  __asm  psubw mm0,mm6 \
+}
+
+/*Performs the last stage of an 8-point 1-D Hadamard transform, takes the
+   absolute value of each component, and accumulates everything into mm0.
+  This is the only portion of SATD which requires MMXEXT (we could use plain
+   MMX, but it takes 4 instructions and an extra register to work around the
+   lack of a pmaxsw, which is a pretty serious penalty).*/
+#define OC_HADAMARD_C_ABS_ACCUM_8x4(_r6,_r7) __asm{ \
+  OC_HADAMARD_C_ABS_ACCUM_A_8x4(_r6,_r7) \
+  OC_HADAMARD_C_ABS_ACCUM_B_8x4(_r6,_r7) \
+}
+
+/*Performs an 8-point 1-D Hadamard transform, takes the absolute value of each
+   component, and accumulates everything into mm0.
+  Note that mm0 will have an extra 4 added to each column, and that after
+   removing this value, the remainder will be half the conventional value.*/
+#define OC_HADAMARD_ABS_ACCUM_8x4(_r6,_r7) __asm{ \
+  OC_HADAMARD_AB_8x4 \
+  OC_HADAMARD_C_ABS_ACCUM_8x4(_r6,_r7) \
+}
+
+/*Performs two 4x4 transposes (mostly) in place.
+  On input, {mm0,mm1,mm2,mm3} contains rows {e,f,g,h}, and {mm4,mm5,mm6,mm7}
+   contains rows {a,b,c,d}.
+  On output, {0x40,0x50,0x60,0x70}+_off+BUF contains {e,f,g,h}^T, and
+   {mm4,mm5,mm6,mm7} contains the transposed rows {a,b,c,d}^T.*/
+#define OC_TRANSPOSE_4x4x2(_off) __asm{ \
+  /*First 4x4 transpose:*/ \
+  __asm  movq [0x10+_off+BUF],mm5 \
+  /*mm0 = e3 e2 e1 e0 \
+    mm1 = f3 f2 f1 f0 \
+    mm2 = g3 g2 g1 g0 \
+    mm3 = h3 h2 h1 h0*/ \
+  __asm  movq mm5,mm2 \
+  __asm  punpcklwd mm2,mm3 \
+  __asm  punpckhwd mm5,mm3 \
+  __asm  movq mm3,mm0 \
+  __asm  punpcklwd mm0,mm1 \
+  __asm  punpckhwd mm3,mm1 \
+  /*mm0 = f1 e1 f0 e0 \
+    mm3 = f3 e3 f2 e2 \
+    mm2 = h1 g1 h0 g0 \
+    mm5 = h3 g3 h2 g2*/ \
+  __asm  movq mm1,mm0 \
+  __asm  punpckldq mm0,mm2 \
+  __asm  punpckhdq mm1,mm2 \
+  __asm  movq mm2,mm3 \
+  __asm  punpckhdq mm3,mm5 \
+  __asm  movq [0x40+_off+BUF],mm0 \
+  __asm  punpckldq mm2,mm5 \
+  /*mm0 = h0 g0 f0 e0 \
+    mm1 = h1 g1 f1 e1 \
+    mm2 = h2 g2 f2 e2 \
+    mm3 = h3 g3 f3 e3*/ \
+  __asm  movq mm5,[0x10+_off+BUF] \
+  /*Second 4x4 transpose:*/ \
+  /*mm4 = a3 a2 a1 a0 \
+    mm5 = b3 b2 b1 b0 \
+    mm6 = c3 c2 c1 c0 \
+    mm7 = d3 d2 d1 d0*/ \
+  __asm  movq mm0,mm6 \
+  __asm  punpcklwd mm6,mm7 \
+  __asm  movq [0x50+_off+BUF],mm1 \
+  __asm  punpckhwd mm0,mm7 \
+  __asm  movq mm7,mm4 \
+  __asm  punpcklwd mm4,mm5 \
+  __asm  movq [0x60+_off+BUF],mm2 \
+  __asm  punpckhwd mm7,mm5 \
+  /*mm4 = b1 a1 b0 a0 \
+    mm7 = b3 a3 b2 a2 \
+    mm6 = d1 c1 d0 c0 \
+    mm0 = d3 c3 d2 c2*/ \
+  __asm  movq mm5,mm4 \
+  __asm  punpckldq mm4,mm6 \
+  __asm  movq [0x70+_off+BUF],mm3 \
+  __asm  punpckhdq mm5,mm6 \
+  __asm  movq mm6,mm7 \
+  __asm  punpckhdq mm7,mm0 \
+  __asm  punpckldq mm6,mm0 \
+  /*mm4 = d0 c0 b0 a0 \
+    mm5 = d1 c1 b1 a1 \
+    mm6 = d2 c2 b2 a2 \
+    mm7 = d3 c3 b3 a3*/ \
+}
+
+static unsigned oc_int_frag_satd_mmxext(int *_dc,
+ const unsigned char *_src,int _src_ystride,
+ const unsigned char *_ref,int _ref_ystride){
+  OC_ALIGN8(ogg_int16_t buf[64]);
+  ogg_int16_t *bufp;
+  unsigned     ret;
+  unsigned     ret2;
+  int          dc;
+  bufp=buf;
+  __asm{
+#define SRC esi
+#define REF eax
+#define SRC_YSTRIDE ecx
+#define REF_YSTRIDE edx
+#define BUF edi
+#define RET edx
+#define RET2 ecx
+#define DC eax
+#define DC_WORD ax
+    mov SRC,_src
+    mov SRC_YSTRIDE,_src_ystride
+    mov REF,_ref
+    mov REF_YSTRIDE,_ref_ystride
+    mov BUF,bufp
+    OC_LOAD_SUB_8x4(0x00)
+    OC_HADAMARD_8x4
+    OC_TRANSPOSE_4x4x2(0x00)
+    /*Finish swapping out this 8x4 block to make room for the next one.
+      mm0...mm3 have been swapped out already.*/
+    movq [0x00+BUF],mm4
+    movq [0x10+BUF],mm5
+    movq [0x20+BUF],mm6
+    movq [0x30+BUF],mm7
+    OC_LOAD_SUB_8x4(0x04)
+    OC_HADAMARD_8x4
+    OC_TRANSPOSE_4x4x2(0x08)
+    /*Here the first 4x4 block of output from the last transpose is the second
+       4x4 block of input for the next transform.
+      We have cleverly arranged that it already be in the appropriate place, so
+       we only have to do half the loads.*/
+    movq mm1,[0x10+BUF]
+    movq mm2,[0x20+BUF]
+    movq mm3,[0x30+BUF]
+    movq mm0,[0x00+BUF]
+    /*We split out the stages here so we can save the DC coefficient in the
+       middle.*/
+    OC_HADAMARD_AB_8x4
+    OC_HADAMARD_C_ABS_ACCUM_A_8x4(0x28,0x38)
+    movd DC,mm1
+    OC_HADAMARD_C_ABS_ACCUM_B_8x4(0x28,0x38)
+    /*Up to this point, everything fit in 16 bits (8 input + 1 for the
+       difference + 2*3 for the two 8-point 1-D Hadamards - 1 for the abs - 1
+       for the factor of two we dropped + 3 for the vertical accumulation).
+      Now we finally have to promote things to dwords.
+      We break this part out of OC_HADAMARD_ABS_ACCUM_8x4 to hide the long
+       latency of pmaddwd by starting the next series of loads now.*/
+    pmaddwd mm0,mm7
+    movq mm1,[0x50+BUF]
+    movq mm5,[0x58+BUF]
+    movq mm4,mm0
+    movq mm2,[0x60+BUF]
+    punpckhdq mm0,mm0
+    movq mm6,[0x68+BUF]
+    paddd mm4,mm0
+    movq mm3,[0x70+BUF]
+    movd RET2,mm4
+    movq mm7,[0x78+BUF]
+    movq mm0,[0x40+BUF]
+    movq mm4,[0x48+BUF]
+    OC_HADAMARD_ABS_ACCUM_8x4(0x68,0x78)
+    pmaddwd mm0,mm7
+    /*Subtract abs(dc) from 2*ret2.*/
+    movsx DC,DC_WORD
+    cdq
+    lea RET2,[RET+RET2*2]
+    movq mm4,mm0
+    punpckhdq mm0,mm0
+    xor RET,DC
+    paddd mm4,mm0
+    /*The sums produced by OC_HADAMARD_ABS_ACCUM_8x4 each have an extra 4
+       added to them, a factor of two removed, and the DC value included;
+       correct the final sum here.*/
+    sub RET2,RET
+    movd RET,mm4
+    lea RET,[RET2+RET*2-64]
+    mov ret,RET
+    mov dc,DC
+#undef SRC
+#undef REF
+#undef SRC_YSTRIDE
+#undef REF_YSTRIDE
+#undef BUF
+#undef RET
+#undef RET2
+#undef DC
+#undef DC_WORD
+  }
+  *_dc=dc;
+  return ret;
+}
+
+unsigned oc_enc_frag_satd_mmxext(int *_dc,const unsigned char *_src,
+ const unsigned char *_ref,int _ystride){
+  return oc_int_frag_satd_mmxext(_dc,_src,_ystride,_ref,_ystride);
+}
+
+
+/*Our internal implementation of frag_copy2 takes an extra stride parameter so
+   we can share code with oc_enc_frag_satd2_mmxext().*/
+static void oc_int_frag_copy2_mmxext(unsigned char *_dst,int _dst_ystride,
+ const unsigned char *_src1,const unsigned char *_src2,int _src_ystride){
+  __asm{
+    /*Load the first 3 rows.*/
+#define DST_YSTRIDE edi
+#define SRC_YSTRIDE esi
+#define DST eax
+#define SRC1 edx
+#define SRC2 ecx
+    mov DST_YSTRIDE,_dst_ystride
+    mov SRC_YSTRIDE,_src_ystride
+    mov DST,_dst
+    mov SRC1,_src1
+    mov SRC2,_src2
+    movq mm0,[SRC1]
+    movq mm1,[SRC2]
+    movq mm2,[SRC1+SRC_YSTRIDE]
+    lea SRC1,[SRC1+SRC_YSTRIDE*2]
+    movq mm3,[SRC2+SRC_YSTRIDE]
+    lea SRC2,[SRC2+SRC_YSTRIDE*2]
+    pxor mm7,mm7
+    movq mm4,[SRC1]
+    pcmpeqb mm6,mm6
+    movq mm5,[SRC2]
+    /*mm7={1}x8.*/
+    psubb mm7,mm6
+    /*Start averaging mm0 and mm1 into mm6.*/
+    movq mm6,mm0
+    pxor mm0,mm1
+    pavgb mm6,mm1
+    /*mm1 is free, start averaging mm3 into mm2 using mm1.*/
+    movq mm1,mm2
+    pand mm0,mm7
+    pavgb mm2,mm3
+    pxor mm1,mm3
+    /*mm3 is free.*/
+    psubb mm6,mm0
+    /*mm0 is free, start loading the next row.*/
+    movq mm0,[SRC1+SRC_YSTRIDE]
+    /*Start averaging mm5 and mm4 using mm3.*/
+    movq mm3,mm4
+    /*mm6 [row 0] is done; write it out.*/
+    movq [DST],mm6
+    pand mm1,mm7
+    pavgb mm4,mm5
+    psubb mm2,mm1
+    /*mm1 is free, continue loading the next row.*/
+    movq mm1,[SRC2+SRC_YSTRIDE]
+    pxor mm3,mm5
+    lea SRC1,[SRC1+SRC_YSTRIDE*2]
+    /*mm2 [row 1] is done; write it out.*/
+    movq [DST+DST_YSTRIDE],mm2
+    pand mm3,mm7
+    /*Start loading the next row.*/
+    movq mm2,[SRC1]
+    lea DST,[DST+DST_YSTRIDE*2]
+    psubb mm4,mm3
+    lea SRC2,[SRC2+SRC_YSTRIDE*2]
+    /*mm4 [row 2] is done; write it out.*/
+    movq [DST],mm4
+    /*Continue loading the next row.*/
+    movq mm3,[SRC2]
+    /*Start averaging mm0 and mm1 into mm6.*/
+    movq mm6,mm0
+    pxor mm0,mm1
+    /*Start loading the next row.*/
+    movq mm4,[SRC1+SRC_YSTRIDE]
+    pavgb mm6,mm1
+    /*mm1 is free; start averaging mm3 into mm2 using mm1.*/
+    movq mm1,mm2
+    pand mm0,mm7
+    /*Continue loading the next row.*/
+    movq mm5,[SRC2+SRC_YSTRIDE]
+    pavgb mm2,mm3
+    lea SRC1,[SRC1+SRC_YSTRIDE*2]
+    pxor mm1,mm3
+    /*mm3 is free.*/
+    psubb mm6,mm0
+    /*mm0 is free, start loading the next row.*/
+    movq mm0,[SRC1]
+    /*Start averaging mm5 into mm4 using mm3.*/
+    movq mm3,mm4
+    /*mm6 [row 3] is done; write it out.*/
+    movq [DST+DST_YSTRIDE],mm6
+    pand mm1,mm7
+    lea SRC2,[SRC2+SRC_YSTRIDE*2]
+    pavgb mm4,mm5
+    lea DST,[DST+DST_YSTRIDE*2]
+    psubb mm2,mm1
+    /*mm1 is free; continue loading the next row.*/
+    movq mm1,[SRC2]
+    pxor mm3,mm5
+    /*mm2 [row 4] is done; write it out.*/
+    movq [DST],mm2
+    pand mm3,mm7
+    /*Start loading the next row.*/
+    movq mm2,[SRC1+SRC_YSTRIDE]
+    psubb mm4,mm3
+    /*Start averaging mm0 and mm1 into mm6.*/
+    movq mm6,mm0
+    /*Continue loading the next row.*/
+    movq mm3,[SRC2+SRC_YSTRIDE]
+    /*mm4 [row 5] is done; write it out.*/
+    movq [DST+DST_YSTRIDE],mm4
+    pxor mm0,mm1
+    pavgb mm6,mm1
+    /*mm4 is free; start averaging mm3 into mm2 using mm4.*/
+    movq mm4,mm2
+    pand mm0,mm7
+    pavgb mm2,mm3
+    pxor mm4,mm3
+    lea DST,[DST+DST_YSTRIDE*2]
+    psubb mm6,mm0
+    pand mm4,mm7
+    /*mm6 [row 6] is done, write it out.*/
+    movq [DST],mm6
+    psubb mm2,mm4
+    /*mm2 [row 7] is done, write it out.*/
+    movq [DST+DST_YSTRIDE],mm2
+#undef SRC1
+#undef SRC2
+#undef SRC_YSTRIDE
+#undef DST_YSTRIDE
+#undef DST
+  }
+}
+
+unsigned oc_enc_frag_satd2_mmxext(int *_dc,const unsigned char *_src,
+ const unsigned char *_ref1,const unsigned char *_ref2,int _ystride){
+  OC_ALIGN8(unsigned char ref[64]);
+  oc_int_frag_copy2_mmxext(ref,8,_ref1,_ref2,_ystride);
+  return oc_int_frag_satd_mmxext(_dc,_src,_ystride,ref,8);
+}
+
+unsigned oc_enc_frag_intra_satd_mmxext(int *_dc,const unsigned char *_src,
+ int _ystride){
+  OC_ALIGN8(ogg_int16_t buf[64]);
+  ogg_int16_t *bufp;
+  unsigned     ret1;
+  unsigned     ret2;
+  int          dc;
+  bufp=buf;
+  __asm{
+#define SRC eax
+#define SRC4 esi
+#define BUF edi
+#define YSTRIDE edx
+#define YSTRIDE3 ecx
+#define RET eax
+#define RET2 ecx
+#define DC edx
+#define DC_WORD dx
+    mov SRC,_src
+    mov BUF,bufp
+    mov YSTRIDE,_ystride
+    /* src4 = src+4*ystride */
+    lea SRC4,[SRC+YSTRIDE*4]
+    /* ystride3 = 3*ystride */
+    lea YSTRIDE3,[YSTRIDE+YSTRIDE*2]
+    OC_LOAD_8x4(0x00)
+    OC_HADAMARD_8x4
+    OC_TRANSPOSE_4x4x2(0x00)
+    /*Finish swapping out this 8x4 block to make room for the next one.
+      mm0...mm3 have been swapped out already.*/
+    movq [0x00+BUF],mm4
+    movq [0x10+BUF],mm5
+    movq [0x20+BUF],mm6
+    movq [0x30+BUF],mm7
+    OC_LOAD_8x4(0x04)
+    OC_HADAMARD_8x4
+    OC_TRANSPOSE_4x4x2(0x08)
+    /*Here the first 4x4 block of output from the last transpose is the second
+      4x4 block of input for the next transform.
+      We have cleverly arranged that it already be in the appropriate place, so
+      we only have to do half the loads.*/
+    movq mm1,[0x10+BUF]
+    movq mm2,[0x20+BUF]
+    movq mm3,[0x30+BUF]
+    movq mm0,[0x00+BUF]
+    /*We split out the stages here so we can save the DC coefficient in the
+      middle.*/
+    OC_HADAMARD_AB_8x4
+    OC_HADAMARD_C_ABS_ACCUM_A_8x4(0x28,0x38)
+    movd DC,mm1
+    OC_HADAMARD_C_ABS_ACCUM_B_8x4(0x28,0x38)
+    /*Up to this point, everything fit in 16 bits (8 input + 1 for the
+      difference + 2*3 for the two 8-point 1-D Hadamards - 1 for the abs - 1
+      for the factor of two we dropped + 3 for the vertical accumulation).
+      Now we finally have to promote things to dwords.
+      We break this part out of OC_HADAMARD_ABS_ACCUM_8x4 to hide the long
+      latency of pmaddwd by starting the next series of loads now.*/
+    pmaddwd mm0,mm7
+    movq mm1,[0x50+BUF]
+    movq mm5,[0x58+BUF]
+    movq mm2,[0x60+BUF]
+    movq mm4,mm0
+    movq mm6,[0x68+BUF]
+    punpckhdq mm0,mm0
+    movq mm3,[0x70+BUF]
+    paddd mm4,mm0
+    movq mm7,[0x78+BUF]
+    movd RET,mm4
+    movq mm0,[0x40+BUF]
+    movq mm4,[0x48+BUF]
+    OC_HADAMARD_ABS_ACCUM_8x4(0x68,0x78)
+    pmaddwd mm0,mm7
+    /*We assume that the DC coefficient is always positive (which is true,
+    because the input to the INTRA transform was not a difference).*/
+    movzx DC,DC_WORD
+    add RET,RET
+    sub RET,DC
+    movq mm4,mm0
+    punpckhdq mm0,mm0
+    paddd mm4,mm0
+    movd RET2,mm4
+    lea RET,[-64+RET+RET2*2]
+    mov [dc],DC
+    mov [ret1],RET
+#undef SRC
+#undef SRC4
+#undef BUF
+#undef YSTRIDE
+#undef YSTRIDE3
+#undef RET
+#undef RET2
+#undef DC
+#undef DC_WORD
+  }
+  *_dc=dc;
+  return ret1;
+}
+
+void oc_enc_frag_sub_mmx(ogg_int16_t _residue[64],
+ const unsigned char *_src, const unsigned char *_ref,int _ystride){
+  int i;
+  __asm  pxor mm7,mm7
+  for(i=4;i-->0;){
+    __asm{
+#define SRC edx
+#define YSTRIDE esi
+#define RESIDUE eax
+#define REF ecx
+      mov YSTRIDE,_ystride
+      mov RESIDUE,_residue
+      mov SRC,_src
+      mov REF,_ref
+      /*mm0=[src]*/
+      movq mm0,[SRC]
+      /*mm1=[ref]*/
+      movq mm1,[REF]
+      /*mm4=[src+ystride]*/
+      movq mm4,[SRC+YSTRIDE]
+      /*mm5=[ref+ystride]*/
+      movq mm5,[REF+YSTRIDE]
+      /*Compute [src]-[ref].*/
+      movq mm2,mm0
+      punpcklbw mm0,mm7
+      movq mm3,mm1
+      punpckhbw mm2,mm7
+      punpcklbw mm1,mm7
+      punpckhbw mm3,mm7
+      psubw mm0,mm1
+      psubw mm2,mm3
+      /*Compute [src+ystride]-[ref+ystride].*/
+      movq mm1,mm4
+      punpcklbw mm4,mm7
+      movq mm3,mm5
+      punpckhbw mm1,mm7
+      lea SRC,[SRC+YSTRIDE*2]
+      punpcklbw mm5,mm7
+      lea REF,[REF+YSTRIDE*2]
+      punpckhbw mm3,mm7
+      psubw mm4,mm5
+      psubw mm1,mm3
+      /*Write the answer out.*/
+      movq [RESIDUE+0x00],mm0
+      movq [RESIDUE+0x08],mm2
+      movq [RESIDUE+0x10],mm4
+      movq [RESIDUE+0x18],mm1
+      lea RESIDUE,[RESIDUE+0x20]
+      mov _residue,RESIDUE
+      mov _src,SRC
+      mov _ref,REF
+#undef SRC
+#undef YSTRIDE
+#undef RESIDUE
+#undef REF
+    }
+  }
+}
+
+void oc_enc_frag_sub_128_mmx(ogg_int16_t _residue[64],
+ const unsigned char *_src,int _ystride){
+   __asm{
+#define YSTRIDE edx
+#define YSTRIDE3 edi
+#define RESIDUE ecx
+#define SRC eax
+    mov YSTRIDE,_ystride
+    mov RESIDUE,_residue
+    mov SRC,_src
+    /*mm0=[src]*/
+    movq mm0,[SRC]
+    /*mm1=[src+ystride]*/
+    movq mm1,[SRC+YSTRIDE]
+    /*mm6={-1}x4*/
+    pcmpeqw mm6,mm6
+    /*mm2=[src+2*ystride]*/
+    movq mm2,[SRC+YSTRIDE*2]
+    /*[ystride3]=3*[ystride]*/
+    lea YSTRIDE3,[YSTRIDE+YSTRIDE*2]
+    /*mm6={1}x4*/
+    psllw mm6,15
+    /*mm3=[src+3*ystride]*/
+    movq mm3,[SRC+YSTRIDE3]
+    /*mm6={128}x4*/
+    psrlw mm6,8
+    /*mm7=0*/ 
+    pxor mm7,mm7
+    /*[src]=[src]+4*[ystride]*/
+    lea SRC,[SRC+YSTRIDE*4]
+    /*Compute [src]-128 and [src+ystride]-128*/
+    movq mm4,mm0
+    punpcklbw mm0,mm7
+    movq mm5,mm1
+    punpckhbw mm4,mm7
+    psubw mm0,mm6
+    punpcklbw mm1,mm7
+    psubw mm4,mm6
+    punpckhbw mm5,mm7
+    psubw mm1,mm6
+    psubw mm5,mm6
+    /*Write the answer out.*/
+    movq [RESIDUE+0x00],mm0
+    movq [RESIDUE+0x08],mm4
+    movq [RESIDUE+0x10],mm1
+    movq [RESIDUE+0x18],mm5
+    /*mm0=[src+4*ystride]*/
+    movq mm0,[SRC]
+    /*mm1=[src+5*ystride]*/
+    movq mm1,[SRC+YSTRIDE]
+    /*Compute [src+2*ystride]-128 and [src+3*ystride]-128*/
+    movq mm4,mm2
+    punpcklbw mm2,mm7
+    movq mm5,mm3
+    punpckhbw mm4,mm7
+    psubw mm2,mm6
+    punpcklbw mm3,mm7
+    psubw mm4,mm6
+    punpckhbw mm5,mm7
+    psubw mm3,mm6
+    psubw mm5,mm6
+    /*Write the answer out.*/
+    movq [RESIDUE+0x20],mm2
+    movq [RESIDUE+0x28],mm4
+    movq [RESIDUE+0x30],mm3
+    movq [RESIDUE+0x38],mm5
+    /*Compute [src+6*ystride]-128 and [src+7*ystride]-128*/
+    movq mm2,[SRC+YSTRIDE*2]
+    movq mm3,[SRC+YSTRIDE3]
+    movq mm4,mm0
+    punpcklbw mm0,mm7
+    movq mm5,mm1
+    punpckhbw mm4,mm7
+    psubw mm0,mm6
+    punpcklbw mm1,mm7
+    psubw mm4,mm6
+    punpckhbw mm5,mm7
+    psubw mm1,mm6
+    psubw mm5,mm6
+    /*Write the answer out.*/
+    movq [RESIDUE+0x40],mm0
+    movq [RESIDUE+0x48],mm4
+    movq [RESIDUE+0x50],mm1
+    movq [RESIDUE+0x58],mm5
+    /*Compute [src+6*ystride]-128 and [src+7*ystride]-128*/
+    movq mm4,mm2
+    punpcklbw mm2,mm7
+    movq mm5,mm3
+    punpckhbw mm4,mm7
+    psubw mm2,mm6
+    punpcklbw mm3,mm7
+    psubw mm4,mm6
+    punpckhbw mm5,mm7
+    psubw mm3,mm6
+    psubw mm5,mm6
+    /*Write the answer out.*/
+    movq [RESIDUE+0x60],mm2
+    movq [RESIDUE+0x68],mm4
+    movq [RESIDUE+0x70],mm3
+    movq [RESIDUE+0x78],mm5
+#undef YSTRIDE
+#undef YSTRIDE3
+#undef RESIDUE
+#undef SRC
+  }
+}
+
+void oc_enc_frag_copy2_mmxext(unsigned char *_dst,
+ const unsigned char *_src1,const unsigned char *_src2,int _ystride){
+  oc_int_frag_copy2_mmxext(_dst,_ystride,_src1,_src2,_ystride);
+}
+
+#endif

thirdparty/libtheora/x86_vc/mmxfdct.c

@@ -0,0 +1,686 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 1999-2006                *
+ * by the Xiph.Org Foundation https://www.xiph.org/                 *
+ *                                                                  *
+ ********************************************************************/ 
+ /*MMX fDCT implementation for x86_32*/
+/*$Id: fdct_ses2.c 14579 2008-03-12 06:42:40Z xiphmont $*/
+#include "x86enc.h"
+#include "x86zigzag.h"
+
+#if defined(OC_X86_ASM)
+
+#define OC_FDCT_STAGE1_8x4  __asm{ \
+  /*Stage 1:*/ \
+  /*mm0=t7'=t0-t7*/ \
+  __asm  psubw mm0,mm7 \
+  __asm  paddw mm7,mm7 \
+  /*mm1=t6'=t1-t6*/ \
+  __asm  psubw mm1, mm6 \
+  __asm  paddw mm6,mm6 \
+  /*mm2=t5'=t2-t5*/ \
+  __asm  psubw mm2,mm5 \
+  __asm  paddw mm5,mm5 \
+  /*mm3=t4'=t3-t4*/ \
+  __asm  psubw mm3,mm4 \
+  __asm  paddw mm4,mm4 \
+  /*mm7=t0'=t0+t7*/ \
+  __asm  paddw mm7,mm0 \
+  /*mm6=t1'=t1+t6*/  \
+  __asm  paddw mm6,mm1 \
+  /*mm5=t2'=t2+t5*/ \
+  __asm  paddw mm5,mm2 \
+  /*mm4=t3'=t3+t4*/ \
+  __asm  paddw mm4,mm3\
+}
+
+#define OC_FDCT8x4(_r0,_r1,_r2,_r3,_r4,_r5,_r6,_r7) __asm{ \
+  /*Stage 2:*/ \
+  /*mm7=t3''=t0'-t3'*/ \
+  __asm  psubw mm7,mm4 \
+  __asm  paddw mm4,mm4 \
+  /*mm6=t2''=t1'-t2'*/ \
+  __asm  psubw mm6,mm5 \
+  __asm  movq [Y+_r6],mm7 \
+  __asm  paddw mm5,mm5 \
+  /*mm1=t5''=t6'-t5'*/ \
+  __asm  psubw mm1,mm2 \
+  __asm  movq [Y+_r2],mm6 \
+  /*mm4=t0''=t0'+t3'*/ \
+  __asm  paddw mm4,mm7 \
+  __asm  paddw mm2,mm2 \
+  /*mm5=t1''=t1'+t2'*/ \
+  __asm  movq [Y+_r0],mm4 \
+  __asm  paddw mm5,mm6 \
+  /*mm2=t6''=t6'+t5'*/ \
+  __asm  paddw mm2,mm1 \
+  __asm  movq [Y+_r4],mm5 \
+  /*mm0=t7', mm1=t5'', mm2=t6'', mm3=t4'.*/ \
+  /*mm4, mm5, mm6, mm7 are free.*/ \
+  /*Stage 3:*/ \
+  /*mm6={2}x4, mm7={27146,0xB500>>1}x2*/ \
+  __asm  mov A,0x5A806A0A \
+  __asm  pcmpeqb mm6,mm6 \
+  __asm  movd mm7,A \
+  __asm  psrlw mm6,15 \
+  __asm  punpckldq mm7,mm7 \
+  __asm  paddw mm6,mm6 \
+  /*mm0=0, m2={-1}x4 \
+    mm5:mm4=t5''*27146+0xB500*/ \
+  __asm  movq mm4,mm1 \
+  __asm  movq mm5,mm1 \
+  __asm  punpcklwd mm4,mm6 \
+  __asm  movq [Y+_r3],mm2 \
+  __asm  pmaddwd mm4,mm7 \
+  __asm  movq [Y+_r7],mm0 \
+  __asm  punpckhwd mm5,mm6 \
+  __asm  pxor mm0,mm0 \
+  __asm  pmaddwd mm5,mm7 \
+  __asm  pcmpeqb mm2,mm2 \
+  /*mm2=t6'', mm1=t5''+(t5''!=0) \
+    mm4=(t5''*27146+0xB500>>16)*/ \
+  __asm  pcmpeqw mm0,mm1 \
+  __asm  psrad mm4,16 \
+  __asm  psubw mm0,mm2 \
+  __asm  movq mm2, [Y+_r3] \
+  __asm  psrad mm5,16 \
+  __asm  paddw mm1,mm0 \
+  __asm  packssdw mm4,mm5 \
+  /*mm4=s=(t5''*27146+0xB500>>16)+t5''+(t5''!=0)>>1*/ \
+  __asm  paddw mm4,mm1 \
+  __asm  movq mm0, [Y+_r7] \
+  __asm  psraw mm4,1 \
+  __asm  movq mm1,mm3 \
+  /*mm3=t4''=t4'+s*/ \
+  __asm  paddw mm3,mm4 \
+  /*mm1=t5'''=t4'-s*/ \
+  __asm  psubw mm1,mm4 \
+  /*mm1=0, mm3={-1}x4 \
+    mm5:mm4=t6''*27146+0xB500*/ \
+  __asm  movq mm4,mm2 \
+  __asm  movq mm5,mm2 \
+  __asm  punpcklwd mm4,mm6 \
+  __asm  movq [Y+_r5],mm1 \
+  __asm  pmaddwd mm4,mm7 \
+  __asm  movq [Y+_r1],mm3 \
+  __asm  punpckhwd mm5,mm6 \
+  __asm  pxor mm1,mm1 \
+  __asm  pmaddwd mm5,mm7 \
+  __asm  pcmpeqb mm3,mm3 \
+  /*mm2=t6''+(t6''!=0), mm4=(t6''*27146+0xB500>>16)*/ \
+  __asm  psrad mm4,16 \
+  __asm  pcmpeqw mm1,mm2 \
+  __asm  psrad mm5,16 \
+  __asm  psubw mm1,mm3 \
+  __asm  packssdw mm4,mm5 \
+  __asm  paddw mm2,mm1 \
+  /*mm1=t1'' \
+    mm4=s=(t6''*27146+0xB500>>16)+t6''+(t6''!=0)>>1*/ \
+  __asm  paddw mm4,mm2 \
+  __asm  movq mm1,[Y+_r4] \
+  __asm  psraw mm4,1 \
+  __asm  movq mm2,mm0 \
+  /*mm7={54491-0x7FFF,0x7FFF}x2 \
+    mm0=t7''=t7'+s*/ \
+  __asm  paddw mm0,mm4 \
+  /*mm2=t6'''=t7'-s*/ \
+  __asm  psubw mm2,mm4 \
+  /*Stage 4:*/ \
+  /*mm0=0, mm2=t0'' \
+    mm5:mm4=t1''*27146+0xB500*/ \
+  __asm  movq mm4,mm1 \
+  __asm  movq mm5,mm1 \
+  __asm  punpcklwd mm4,mm6 \
+  __asm  movq [Y+_r3],mm2 \
+  __asm  pmaddwd mm4,mm7 \
+  __asm  movq mm2,[Y+_r0] \
+  __asm  punpckhwd mm5,mm6 \
+  __asm  movq [Y+_r7],mm0 \
+  __asm  pmaddwd mm5,mm7 \
+  __asm  pxor mm0,mm0 \
+  /*mm7={27146,0x4000>>1}x2 \
+    mm0=s=(t1''*27146+0xB500>>16)+t1''+(t1''!=0)*/ \
+  __asm  psrad mm4,16 \
+  __asm  mov A,0x20006A0A \
+  __asm  pcmpeqw mm0,mm1 \
+  __asm  movd mm7,A \
+  __asm  psrad mm5,16 \
+  __asm  psubw mm0,mm3 \
+  __asm  packssdw mm4,mm5 \
+  __asm  paddw mm0,mm1 \
+  __asm  punpckldq mm7,mm7 \
+  __asm  paddw mm0,mm4 \
+  /*mm6={0x00000E3D}x2 \
+    mm1=-(t0''==0), mm5:mm4=t0''*27146+0x4000*/ \
+  __asm  movq mm4,mm2 \
+  __asm  movq mm5,mm2 \
+  __asm  punpcklwd mm4,mm6 \
+  __asm  mov A,0x0E3D \
+  __asm  pmaddwd mm4,mm7 \
+  __asm  punpckhwd mm5,mm6 \
+  __asm  movd mm6,A \
+  __asm  pmaddwd mm5,mm7 \
+  __asm  pxor mm1,mm1 \
+  __asm  punpckldq mm6,mm6 \
+  __asm  pcmpeqw mm1,mm2 \
+  /*mm4=r=(t0''*27146+0x4000>>16)+t0''+(t0''!=0)*/ \
+  __asm  psrad mm4,16 \
+  __asm  psubw mm1,mm3 \
+  __asm  psrad mm5,16 \
+  __asm  paddw mm2,mm1 \
+  __asm  packssdw mm4,mm5 \
+  __asm  movq mm1,[Y+_r5] \
+  __asm  paddw mm4,mm2 \
+  /*mm2=t6'', mm0=_y[0]=u=r+s>>1 \
+    The naive implementation could cause overflow, so we use \
+     u=(r&s)+((r^s)>>1).*/ \
+  __asm  movq mm2,[Y+_r3] \
+  __asm  movq mm7,mm0 \
+  __asm  pxor mm0,mm4 \
+  __asm  pand mm7,mm4 \
+  __asm  psraw mm0,1 \
+  __asm  mov A,0x7FFF54DC \
+  __asm  paddw mm0,mm7 \
+  __asm  movd mm7,A \
+  /*mm7={54491-0x7FFF,0x7FFF}x2 \
+    mm4=_y[4]=v=r-u*/ \
+  __asm  psubw mm4,mm0 \
+  __asm  punpckldq mm7,mm7 \
+  __asm  movq [Y+_r4],mm4 \
+  /*mm0=0, mm7={36410}x4 \
+    mm1=(t5'''!=0), mm5:mm4=54491*t5'''+0x0E3D*/ \
+  __asm  movq mm4,mm1 \
+  __asm  movq mm5,mm1 \
+  __asm  punpcklwd mm4,mm1 \
+  __asm  mov A,0x8E3A8E3A \
+  __asm  pmaddwd mm4,mm7 \
+  __asm  movq [Y+_r0],mm0 \
+  __asm  punpckhwd mm5,mm1 \
+  __asm  pxor mm0,mm0 \
+  __asm  pmaddwd mm5,mm7 \
+  __asm  pcmpeqw mm1,mm0 \
+  __asm  movd mm7,A \
+  __asm  psubw mm1,mm3 \
+  __asm  punpckldq mm7,mm7 \
+  __asm  paddd mm4,mm6 \
+  __asm  paddd mm5,mm6 \
+  /*mm0=0 \
+    mm3:mm1=36410*t6'''+((t5'''!=0)<<16)*/ \
+  __asm  movq mm6,mm2 \
+  __asm  movq mm3,mm2 \
+  __asm  pmulhw mm6,mm7 \
+  __asm  paddw mm1,mm2 \
+  __asm  pmullw mm3,mm7 \
+  __asm  pxor mm0,mm0 \
+  __asm  paddw mm6,mm1 \
+  __asm  movq mm1,mm3 \
+  __asm  punpckhwd mm3,mm6 \
+  __asm  punpcklwd mm1,mm6 \
+  /*mm3={-1}x4, mm6={1}x4 \
+    mm4=_y[5]=u=(54491*t5'''+36410*t6'''+0x0E3D>>16)+(t5'''!=0)*/ \
+  __asm  paddd mm5,mm3 \
+  __asm  paddd mm4,mm1 \
+  __asm  psrad mm5,16 \
+  __asm  pxor mm6,mm6 \
+  __asm  psrad mm4,16 \
+  __asm  pcmpeqb mm3,mm3 \
+  __asm  packssdw mm4,mm5 \
+  __asm  psubw mm6,mm3 \
+  /*mm1=t7'', mm7={26568,0x3400}x2 \
+    mm2=s=t6'''-(36410*u>>16)*/ \
+  __asm  movq mm1,mm4 \
+  __asm  mov A,0x340067C8 \
+  __asm  pmulhw mm4,mm7 \
+  __asm  movd mm7,A \
+  __asm  movq [Y+_r5],mm1 \
+  __asm  punpckldq mm7,mm7 \
+  __asm  paddw mm4,mm1 \
+  __asm  movq mm1,[Y+_r7] \
+  __asm  psubw mm2,mm4 \
+  /*mm6={0x00007B1B}x2 \
+    mm0=(s!=0), mm5:mm4=s*26568+0x3400*/ \
+  __asm  movq mm4,mm2 \
+  __asm  movq mm5,mm2 \
+  __asm  punpcklwd mm4,mm6 \
+  __asm  pcmpeqw mm0,mm2 \
+  __asm  pmaddwd mm4,mm7 \
+  __asm  mov A,0x7B1B \
+  __asm  punpckhwd mm5,mm6 \
+  __asm  movd mm6,A \
+  __asm  pmaddwd mm5,mm7 \
+  __asm  psubw mm0,mm3 \
+  __asm  punpckldq mm6,mm6 \
+  /*mm7={64277-0x7FFF,0x7FFF}x2 \
+    mm2=_y[3]=v=(s*26568+0x3400>>17)+s+(s!=0)*/ \
+  __asm  psrad mm4,17 \
+  __asm  paddw mm2,mm0 \
+  __asm  psrad mm5,17 \
+  __asm  mov A,0x7FFF7B16 \
+  __asm  packssdw mm4,mm5 \
+  __asm  movd mm7,A \
+  __asm  paddw mm2,mm4 \
+  __asm  punpckldq mm7,mm7 \
+  /*mm0=0, mm7={12785}x4 \
+    mm1=(t7''!=0), mm2=t4'', mm5:mm4=64277*t7''+0x7B1B*/ \
+  __asm  movq mm4,mm1 \
+  __asm  movq mm5,mm1 \
+  __asm  movq [Y+_r3],mm2 \
+  __asm  punpcklwd mm4,mm1 \
+  __asm  movq mm2,[Y+_r1] \
+  __asm  pmaddwd mm4,mm7 \
+  __asm  mov A,0x31F131F1 \
+  __asm  punpckhwd mm5,mm1 \
+  __asm  pxor mm0,mm0 \
+  __asm  pmaddwd mm5,mm7 \
+  __asm  pcmpeqw mm1,mm0 \
+  __asm  movd mm7,A \
+  __asm  psubw mm1,mm3 \
+  __asm  punpckldq mm7,mm7 \
+  __asm  paddd mm4,mm6 \
+  __asm  paddd mm5,mm6 \
+  /*mm3:mm1=12785*t4'''+((t7''!=0)<<16)*/ \
+  __asm  movq mm6,mm2 \
+  __asm  movq mm3,mm2 \
+  __asm  pmulhw mm6,mm7 \
+  __asm  pmullw mm3,mm7 \
+  __asm  paddw mm6,mm1 \
+  __asm  movq mm1,mm3 \
+  __asm  punpckhwd mm3,mm6 \
+  __asm  punpcklwd mm1,mm6 \
+  /*mm3={-1}x4, mm6={1}x4 \
+    mm4=_y[1]=u=(12785*t4'''+64277*t7''+0x7B1B>>16)+(t7''!=0)*/ \
+  __asm  paddd mm5,mm3 \
+  __asm  paddd mm4,mm1 \
+  __asm  psrad mm5,16 \
+  __asm  pxor mm6,mm6 \
+  __asm  psrad mm4,16 \
+  __asm  pcmpeqb mm3,mm3 \
+  __asm  packssdw mm4,mm5 \
+  __asm  psubw mm6,mm3 \
+  /*mm1=t3'', mm7={20539,0x3000}x2 \
+    mm4=s=(12785*u>>16)-t4''*/ \
+  __asm  movq [Y+_r1],mm4 \
+  __asm  pmulhw mm4,mm7 \
+  __asm  mov A,0x3000503B \
+  __asm  movq mm1,[Y+_r6] \
+  __asm  movd mm7,A \
+  __asm  psubw mm4,mm2 \
+  __asm  punpckldq mm7,mm7 \
+  /*mm6={0x00006CB7}x2 \
+    mm0=(s!=0), mm5:mm4=s*20539+0x3000*/ \
+  __asm  movq mm5,mm4 \
+  __asm  movq mm2,mm4 \
+  __asm  punpcklwd mm4,mm6 \
+  __asm  pcmpeqw mm0,mm2 \
+  __asm  pmaddwd mm4,mm7 \
+  __asm  mov A,0x6CB7 \
+  __asm  punpckhwd mm5,mm6 \
+  __asm  movd mm6,A \
+  __asm  pmaddwd mm5,mm7 \
+  __asm  psubw mm0,mm3 \
+  __asm  punpckldq mm6,mm6 \
+  /*mm7={60547-0x7FFF,0x7FFF}x2 \
+    mm2=_y[7]=v=(s*20539+0x3000>>20)+s+(s!=0)*/ \
+  __asm  psrad mm4,20 \
+  __asm  paddw mm2,mm0 \
+  __asm  psrad mm5,20 \
+  __asm  mov A,0x7FFF6C84 \
+  __asm  packssdw mm4,mm5 \
+  __asm  movd mm7,A \
+  __asm  paddw mm2,mm4 \
+  __asm  punpckldq mm7,mm7 \
+  /*mm0=0, mm7={25080}x4 \
+    mm2=t2'', mm5:mm4=60547*t3''+0x6CB7*/ \
+  __asm  movq mm4,mm1 \
+  __asm  movq mm5,mm1 \
+  __asm  movq [Y+_r7],mm2 \
+  __asm  punpcklwd mm4,mm1 \
+  __asm  movq mm2,[Y+_r2] \
+  __asm  pmaddwd mm4,mm7 \
+  __asm  mov A,0x61F861F8 \
+  __asm  punpckhwd mm5,mm1 \
+  __asm  pxor mm0,mm0 \
+  __asm  pmaddwd mm5,mm7 \
+  __asm  movd mm7,A \
+  __asm  pcmpeqw mm1,mm0 \
+  __asm  psubw mm1,mm3 \
+  __asm  punpckldq mm7,mm7 \
+  __asm  paddd mm4,mm6 \
+  __asm  paddd mm5,mm6 \
+  /*mm3:mm1=25080*t2''+((t3''!=0)<<16)*/ \
+  __asm  movq mm6,mm2 \
+  __asm  movq mm3,mm2 \
+  __asm  pmulhw mm6,mm7 \
+  __asm  pmullw mm3,mm7 \
+  __asm  paddw mm6,mm1 \
+  __asm  movq mm1,mm3 \
+  __asm  punpckhwd mm3,mm6 \
+  __asm  punpcklwd mm1,mm6 \
+  /*mm1={-1}x4 \
+    mm4=u=(25080*t2''+60547*t3''+0x6CB7>>16)+(t3''!=0)*/ \
+  __asm  paddd mm5,mm3 \
+  __asm  paddd mm4,mm1 \
+  __asm  psrad mm5,16 \
+  __asm  mov A,0x28005460 \
+  __asm  psrad mm4,16 \
+  __asm  pcmpeqb mm1,mm1 \
+  __asm  packssdw mm4,mm5 \
+  /*mm5={1}x4, mm6=_y[2]=u, mm7={21600,0x2800}x2 \
+    mm4=s=(25080*u>>16)-t2''*/ \
+  __asm  movq mm6,mm4 \
+  __asm  pmulhw mm4,mm7 \
+  __asm  pxor mm5,mm5 \
+  __asm  movd mm7,A \
+  __asm  psubw mm5,mm1 \
+  __asm  punpckldq mm7,mm7 \
+  __asm  psubw mm4,mm2 \
+  /*mm2=s+(s!=0) \
+    mm4:mm3=s*21600+0x2800*/ \
+  __asm  movq mm3,mm4 \
+  __asm  movq mm2,mm4 \
+  __asm  punpckhwd mm4,mm5 \
+  __asm  pcmpeqw mm0,mm2 \
+  __asm  pmaddwd mm4,mm7 \
+  __asm  psubw mm0,mm1 \
+  __asm  punpcklwd mm3,mm5 \
+  __asm  paddw mm2,mm0 \
+  __asm  pmaddwd mm3,mm7 \
+  /*mm0=_y[4], mm1=_y[7], mm4=_y[0], mm5=_y[5] \
+    mm3=_y[6]=v=(s*21600+0x2800>>18)+s+(s!=0)*/ \
+  __asm  movq mm0,[Y+_r4] \
+  __asm  psrad mm4,18 \
+  __asm  movq mm5,[Y+_r5] \
+  __asm  psrad mm3,18 \
+  __asm  movq mm1,[Y+_r7] \
+  __asm  packssdw mm3,mm4 \
+  __asm  movq mm4,[Y+_r0] \
+  __asm  paddw mm3,mm2 \
+}
+
+/*On input, mm4=_y[0], mm6=_y[2], mm0=_y[4], mm5=_y[5], mm3=_y[6], mm1=_y[7].
+  On output, {_y[4],mm1,mm2,mm3} contains the transpose of _y[4...7] and
+   {mm4,mm5,mm6,mm7} contains the transpose of _y[0...3].*/
+#define OC_TRANSPOSE8x4(_r0,_r1,_r2,_r3,_r4,_r5,_r6,_r7) __asm{ \
+  /*First 4x4 transpose:*/ \
+  /*mm0 = e3 e2 e1 e0 \
+    mm5 = f3 f2 f1 f0 \
+    mm3 = g3 g2 g1 g0 \
+    mm1 = h3 h2 h1 h0*/ \
+  __asm  movq mm2,mm0 \
+  __asm  punpcklwd mm0,mm5 \
+  __asm  punpckhwd mm2,mm5 \
+  __asm  movq mm5,mm3 \
+  __asm  punpcklwd mm3,mm1 \
+  __asm  punpckhwd mm5,mm1 \
+  /*mm0 = f1 e1 f0 e0 \
+    mm2 = f3 e3 f2 e2 \
+    mm3 = h1 g1 h0 g0 \
+    mm5 = h3 g3 h2 g2*/ \
+  __asm  movq mm1,mm0 \
+  __asm  punpckldq mm0,mm3 \
+  __asm  movq [Y+_r4],mm0 \
+  __asm  punpckhdq mm1,mm3 \
+  __asm  movq mm0,[Y+_r1] \
+  __asm  movq mm3,mm2 \
+  __asm  punpckldq mm2,mm5 \
+  __asm  punpckhdq mm3,mm5 \
+  __asm  movq mm5,[Y+_r3] \
+  /*_y[4] = h0 g0 f0 e0 \
+   mm1  = h1 g1 f1 e1 \
+   mm2  = h2 g2 f2 e2 \
+   mm3  = h3 g3 f3 e3*/ \
+  /*Second 4x4 transpose:*/ \
+  /*mm4 = a3 a2 a1 a0 \
+    mm0 = b3 b2 b1 b0 \
+    mm6 = c3 c2 c1 c0 \
+    mm5 = d3 d2 d1 d0*/ \
+  __asm  movq mm7,mm4 \
+  __asm  punpcklwd mm4,mm0 \
+  __asm  punpckhwd mm7,mm0 \
+  __asm  movq mm0,mm6 \
+  __asm  punpcklwd mm6,mm5 \
+  __asm  punpckhwd mm0,mm5 \
+  /*mm4 = b1 a1 b0 a0 \
+    mm7 = b3 a3 b2 a2 \
+    mm6 = d1 c1 d0 c0 \
+    mm0 = d3 c3 d2 c2*/ \
+  __asm  movq mm5,mm4 \
+  __asm  punpckldq mm4,mm6 \
+  __asm  punpckhdq mm5,mm6 \
+  __asm  movq mm6,mm7 \
+  __asm  punpckhdq mm7,mm0 \
+  __asm  punpckldq mm6,mm0 \
+  /*mm4 = d0 c0 b0 a0 \
+    mm5 = d1 c1 b1 a1 \
+    mm6 = d2 c2 b2 a2 \
+    mm7 = d3 c3 b3 a3*/ \
+}
+
+/*MMX implementation of the fDCT.*/
+void oc_enc_fdct8x8_mmxext(ogg_int16_t _y[64],const ogg_int16_t _x[64]){
+  OC_ALIGN8(ogg_int16_t buf[64]);
+  ogg_int16_t *bufp;
+  bufp=buf;
+  __asm{
+#define X edx
+#define Y eax
+#define A ecx
+#define BUF esi
+    /*Add two extra bits of working precision to improve accuracy; any more and
+       we could overflow.*/
+    /*We also add biases to correct for some systematic error that remains in
+       the full fDCT->iDCT round trip.*/
+    mov X, _x
+    mov Y, _y
+	mov BUF, bufp
+    movq mm0,[0x00+X]
+    movq mm1,[0x10+X]
+    movq mm2,[0x20+X]
+    movq mm3,[0x30+X]
+    pcmpeqb mm4,mm4
+    pxor mm7,mm7
+    movq mm5,mm0
+    psllw mm0,2
+    pcmpeqw mm5,mm7
+    movq mm7,[0x70+X]
+    psllw mm1,2
+    psubw mm5,mm4
+    psllw mm2,2
+    mov A,1
+    pslld mm5,16
+    movd mm6,A
+    psllq mm5,16
+    mov A,0x10001
+    psllw mm3,2
+    movd mm4,A
+    punpckhwd mm5,mm6
+    psubw mm1,mm6
+    movq mm6,[0x60+X]
+    paddw mm0,mm5
+    movq mm5,[0x50+X]
+    paddw mm0,mm4
+    movq mm4,[0x40+X]
+    /*We inline stage1 of the transform here so we can get better instruction
+       scheduling with the shifts.*/
+    /*mm0=t7'=t0-t7*/
+    psllw mm7,2
+    psubw mm0,mm7
+    psllw mm6,2
+    paddw mm7,mm7
+    /*mm1=t6'=t1-t6*/
+    psllw mm5,2
+    psubw mm1,mm6
+    psllw mm4,2
+    paddw mm6,mm6
+    /*mm2=t5'=t2-t5*/
+    psubw mm2,mm5
+    paddw mm5,mm5
+    /*mm3=t4'=t3-t4*/
+    psubw mm3,mm4
+    paddw mm4,mm4
+    /*mm7=t0'=t0+t7*/
+    paddw mm7,mm0
+    /*mm6=t1'=t1+t6*/
+    paddw mm6,mm1
+    /*mm5=t2'=t2+t5*/
+    paddw mm5,mm2
+    /*mm4=t3'=t3+t4*/
+    paddw mm4,mm3
+    OC_FDCT8x4(0x00,0x10,0x20,0x30,0x40,0x50,0x60,0x70)
+    OC_TRANSPOSE8x4(0x00,0x10,0x20,0x30,0x40,0x50,0x60,0x70)
+    /*Swap out this 8x4 block for the next one.*/
+    movq mm0,[0x08+X]
+    movq [0x30+Y],mm7
+    movq mm7,[0x78+X]
+    movq [0x50+Y],mm1
+    movq mm1,[0x18+X]
+    movq [0x20+Y],mm6
+    movq mm6,[0x68+X]
+    movq [0x60+Y],mm2
+    movq mm2,[0x28+X]
+    movq [0x10+Y],mm5
+    movq mm5,[0x58+X]
+    movq [0x70+Y],mm3
+    movq mm3,[0x38+X]
+    /*And increase its working precision, too.*/
+    psllw mm0,2
+    movq [0x00+Y],mm4
+    psllw mm7,2
+    movq mm4,[0x48+X]
+    /*We inline stage1 of the transform here so we can get better instruction
+       scheduling with the shifts.*/
+    /*mm0=t7'=t0-t7*/
+    psubw mm0,mm7
+    psllw mm1,2
+    paddw mm7,mm7
+    psllw mm6,2
+    /*mm1=t6'=t1-t6*/
+    psubw mm1,mm6
+    psllw mm2,2
+    paddw mm6,mm6
+    psllw mm5,2
+    /*mm2=t5'=t2-t5*/
+    psubw mm2,mm5
+    psllw mm3,2
+    paddw mm5,mm5
+    psllw mm4,2
+    /*mm3=t4'=t3-t4*/
+    psubw mm3,mm4
+    paddw mm4,mm4
+    /*mm7=t0'=t0+t7*/
+    paddw mm7,mm0
+    /*mm6=t1'=t1+t6*/
+    paddw mm6,mm1
+    /*mm5=t2'=t2+t5*/
+    paddw mm5,mm2
+    /*mm4=t3'=t3+t4*/
+    paddw mm4,mm3
+    OC_FDCT8x4(0x08,0x18,0x28,0x38,0x48,0x58,0x68,0x78)
+    OC_TRANSPOSE8x4(0x08,0x18,0x28,0x38,0x48,0x58,0x68,0x78)
+    /*Here the first 4x4 block of output from the last transpose is the second
+       4x4 block of input for the next transform.
+      We have cleverly arranged that it already be in the appropriate place,
+       so we only have to do half the stores and loads.*/
+    movq mm0,[0x00+Y]
+    movq [0x58+Y],mm1
+    movq mm1,[0x10+Y]
+    movq [0x68+Y],mm2
+    movq mm2,[0x20+Y]
+    movq [0x78+Y],mm3
+    movq mm3,[0x30+Y]
+    OC_FDCT_STAGE1_8x4
+    OC_FDCT8x4(0x00,0x10,0x20,0x30,0x08,0x18,0x28,0x38)
+    /*mm0={-2}x4*/
+    pcmpeqw mm2,mm2
+    paddw mm2,mm2
+    /*Round and store the results (no transpose).*/
+    movq mm7,[Y+0x10]
+    psubw mm4,mm2
+    psubw mm6,mm2
+    psraw mm4,2
+    psubw mm0,mm2
+    movq [BUF+0x00],mm4
+    movq mm4,[Y+0x30]
+    psraw mm6,2
+    psubw mm5,mm2
+    movq [BUF+0x20],mm6
+    psraw mm0,2
+    psubw mm3,mm2
+    movq [BUF+0x40],mm0
+    psraw mm5,2
+    psubw mm1,mm2
+    movq [BUF+0x50],mm5
+    psraw mm3,2
+    psubw mm7,mm2
+    movq [BUF+0x60],mm3
+    psraw mm1,2
+    psubw mm4,mm2
+    movq [BUF+0x70],mm1
+    psraw mm7,2
+    movq [BUF+0x10],mm7
+    psraw mm4,2
+    movq [BUF+0x30],mm4
+    /*Load the next block.*/
+    movq mm0,[0x40+Y]
+    movq mm7,[0x78+Y]
+    movq mm1,[0x50+Y]
+    movq mm6,[0x68+Y]
+    movq mm2,[0x60+Y]
+    movq mm5,[0x58+Y]
+    movq mm3,[0x70+Y]
+    movq mm4,[0x48+Y]
+    OC_FDCT_STAGE1_8x4
+    OC_FDCT8x4(0x40,0x50,0x60,0x70,0x48,0x58,0x68,0x78)
+    /*mm0={-2}x4*/
+    pcmpeqw mm2,mm2
+    paddw mm2,mm2
+    /*Round and store the results (no transpose).*/
+    movq mm7,[Y+0x50]
+    psubw mm4,mm2
+    psubw mm6,mm2
+    psraw mm4,2
+    psubw mm0,mm2
+    movq [BUF+0x08],mm4
+    movq mm4,[Y+0x70]
+    psraw mm6,2
+    psubw mm5,mm2
+    movq [BUF+0x28],mm6
+    psraw mm0,2
+    psubw mm3,mm2
+    movq [BUF+0x48],mm0
+    psraw mm5,2
+    psubw mm1,mm2
+    movq [BUF+0x58],mm5
+    psraw mm3,2
+    psubw mm7,mm2
+    movq [BUF+0x68],mm3
+    psraw mm1,2
+    psubw mm4,mm2
+    movq [BUF+0x78],mm1
+    psraw mm7,2
+    movq [BUF+0x18],mm7
+    psraw mm4,2
+    movq [BUF+0x38],mm4
+#define OC_ZZ_LOAD_ROW_LO(_row,_reg) \
+    __asm movq _reg,[BUF+16*(_row)] \
+
+#define OC_ZZ_LOAD_ROW_HI(_row,_reg) \
+    __asm movq _reg,[BUF+16*(_row)+8] \
+
+    OC_TRANSPOSE_ZIG_ZAG_MMXEXT
+#undef OC_ZZ_LOAD_ROW_LO
+#undef OC_ZZ_LOAD_ROW_HI
+#undef X
+#undef Y
+#undef A
+#undef BUF
+  }
+}
+
+#endif

thirdparty/libtheora/x86_vc/mmxfrag.c

@@ -0,0 +1,416 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+/*MMX acceleration of fragment reconstruction for motion compensation.
+  Originally written by Rudolf Marek.
+  Additional optimization by Nils Pipenbrinck.
+  Note: Loops are unrolled for best performance.
+  The iteration each instruction belongs to is marked in the comments as #i.*/
+#include <stddef.h>
+#include "x86int.h"
+
+#if defined(OC_X86_ASM)
+
+/*Copies an 8x8 block of pixels from _src to _dst, assuming _ystride bytes
+   between rows.*/
+# define OC_FRAG_COPY_MMX(_dst,_src,_ystride) \
+  do{ \
+    const unsigned char *src; \
+    unsigned char       *dst; \
+    src=(_src); \
+    dst=(_dst); \
+    __asm  mov SRC,src \
+    __asm  mov DST,dst \
+    __asm  mov YSTRIDE,_ystride \
+    /*src+0*ystride*/ \
+    __asm  movq mm0,[SRC] \
+    /*src+1*ystride*/ \
+    __asm  movq mm1,[SRC+YSTRIDE] \
+    /*ystride3=ystride*3*/ \
+    __asm  lea YSTRIDE3,[YSTRIDE+YSTRIDE*2] \
+    /*src+2*ystride*/ \
+    __asm  movq mm2,[SRC+YSTRIDE*2] \
+    /*src+3*ystride*/ \
+    __asm  movq mm3,[SRC+YSTRIDE3] \
+    /*dst+0*ystride*/ \
+    __asm  movq [DST],mm0 \
+    /*dst+1*ystride*/ \
+    __asm  movq [DST+YSTRIDE],mm1 \
+    /*Pointer to next 4.*/ \
+    __asm  lea SRC,[SRC+YSTRIDE*4] \
+    /*dst+2*ystride*/ \
+    __asm  movq [DST+YSTRIDE*2],mm2 \
+    /*dst+3*ystride*/ \
+    __asm  movq [DST+YSTRIDE3],mm3 \
+    /*Pointer to next 4.*/ \
+    __asm  lea DST,[DST+YSTRIDE*4] \
+    /*src+0*ystride*/ \
+    __asm  movq mm0,[SRC] \
+    /*src+1*ystride*/ \
+    __asm  movq mm1,[SRC+YSTRIDE] \
+    /*src+2*ystride*/ \
+    __asm  movq mm2,[SRC+YSTRIDE*2] \
+    /*src+3*ystride*/ \
+    __asm  movq mm3,[SRC+YSTRIDE3] \
+    /*dst+0*ystride*/ \
+    __asm  movq [DST],mm0 \
+    /*dst+1*ystride*/ \
+    __asm  movq [DST+YSTRIDE],mm1 \
+    /*dst+2*ystride*/ \
+    __asm  movq [DST+YSTRIDE*2],mm2 \
+    /*dst+3*ystride*/ \
+    __asm  movq [DST+YSTRIDE3],mm3 \
+  } \
+  while(0)
+
+/*Copies an 8x8 block of pixels from _src to _dst, assuming _ystride bytes
+   between rows.*/
+void oc_frag_copy_mmx(unsigned char *_dst,
+ const unsigned char *_src,int _ystride){
+#define SRC edx
+#define DST eax
+#define YSTRIDE ecx
+#define YSTRIDE3 esi
+  OC_FRAG_COPY_MMX(_dst,_src,_ystride);
+#undef SRC
+#undef DST
+#undef YSTRIDE
+#undef YSTRIDE3
+}
+
+/*Copies the fragments specified by the lists of fragment indices from one
+   frame to another.
+  _dst_frame:     The reference frame to copy to.
+  _src_frame:     The reference frame to copy from.
+  _ystride:       The row stride of the reference frames.
+  _fragis:        A pointer to a list of fragment indices.
+  _nfragis:       The number of fragment indices to copy.
+  _frag_buf_offs: The offsets of fragments in the reference frames.*/
+void oc_frag_copy_list_mmx(unsigned char *_dst_frame,
+ const unsigned char *_src_frame,int _ystride,
+ const ptrdiff_t *_fragis,ptrdiff_t _nfragis,const ptrdiff_t *_frag_buf_offs){
+  ptrdiff_t fragii;
+  for(fragii=0;fragii<_nfragis;fragii++){
+    ptrdiff_t frag_buf_off;
+    frag_buf_off=_frag_buf_offs[_fragis[fragii]];
+#define SRC edx
+#define DST eax
+#define YSTRIDE ecx
+#define YSTRIDE3 edi
+    OC_FRAG_COPY_MMX(_dst_frame+frag_buf_off,
+     _src_frame+frag_buf_off,_ystride);
+#undef SRC
+#undef DST
+#undef YSTRIDE
+#undef YSTRIDE3
+  }
+}
+
+void oc_frag_recon_intra_mmx(unsigned char *_dst,int _ystride,
+ const ogg_int16_t *_residue){
+  __asm{
+#define DST edx
+#define DST4 esi
+#define YSTRIDE eax
+#define YSTRIDE3 edi
+#define RESIDUE ecx
+    mov DST,_dst
+    mov YSTRIDE,_ystride
+    mov RESIDUE,_residue
+    lea DST4,[DST+YSTRIDE*4]
+    lea YSTRIDE3,[YSTRIDE+YSTRIDE*2]
+    /*Set mm0 to 0xFFFFFFFFFFFFFFFF.*/
+    pcmpeqw mm0,mm0
+    /*#0 Load low residue.*/
+    movq mm1,[0*8+RESIDUE]
+    /*#0 Load high residue.*/
+    movq mm2,[1*8+RESIDUE]
+    /*Set mm0 to 0x8000800080008000.*/
+    psllw mm0,15
+    /*#1 Load low residue.*/
+    movq mm3,[2*8+RESIDUE]
+    /*#1 Load high residue.*/
+    movq mm4,[3*8+RESIDUE]
+    /*Set mm0 to 0x0080008000800080.*/
+    psrlw mm0,8
+    /*#2 Load low residue.*/
+    movq mm5,[4*8+RESIDUE]
+    /*#2 Load high residue.*/
+    movq mm6,[5*8+RESIDUE]
+    /*#0 Bias low  residue.*/
+    paddsw mm1,mm0
+    /*#0 Bias high residue.*/
+    paddsw mm2,mm0
+    /*#0 Pack to byte.*/
+    packuswb mm1,mm2
+    /*#1 Bias low  residue.*/
+    paddsw mm3,mm0
+    /*#1 Bias high residue.*/
+    paddsw mm4,mm0
+    /*#1 Pack to byte.*/
+    packuswb mm3,mm4
+    /*#2 Bias low  residue.*/
+    paddsw mm5,mm0
+    /*#2 Bias high residue.*/
+    paddsw mm6,mm0
+    /*#2 Pack to byte.*/
+    packuswb mm5,mm6
+    /*#0 Write row.*/
+    movq [DST],mm1
+    /*#1 Write row.*/
+    movq [DST+YSTRIDE],mm3
+    /*#2 Write row.*/
+    movq [DST+YSTRIDE*2],mm5
+    /*#3 Load low residue.*/
+    movq mm1,[6*8+RESIDUE]
+    /*#3 Load high residue.*/
+    movq mm2,[7*8+RESIDUE]
+    /*#4 Load high residue.*/
+    movq mm3,[8*8+RESIDUE]
+    /*#4 Load high residue.*/
+    movq mm4,[9*8+RESIDUE]
+    /*#5 Load high residue.*/
+    movq mm5,[10*8+RESIDUE]
+    /*#5 Load high residue.*/
+    movq mm6,[11*8+RESIDUE]
+    /*#3 Bias low  residue.*/
+    paddsw mm1,mm0
+    /*#3 Bias high residue.*/
+    paddsw mm2,mm0
+    /*#3 Pack to byte.*/
+    packuswb mm1,mm2
+    /*#4 Bias low  residue.*/
+    paddsw mm3,mm0
+    /*#4 Bias high residue.*/
+    paddsw mm4,mm0
+    /*#4 Pack to byte.*/
+    packuswb mm3,mm4
+    /*#5 Bias low  residue.*/
+    paddsw mm5,mm0
+    /*#5 Bias high residue.*/
+    paddsw mm6,mm0
+    /*#5 Pack to byte.*/
+    packuswb mm5,mm6
+    /*#3 Write row.*/
+    movq [DST+YSTRIDE3],mm1
+    /*#4 Write row.*/
+    movq [DST4],mm3
+    /*#5 Write row.*/
+    movq [DST4+YSTRIDE],mm5
+    /*#6 Load low residue.*/
+    movq mm1,[12*8+RESIDUE]
+    /*#6 Load high residue.*/
+    movq mm2,[13*8+RESIDUE]
+    /*#7 Load low residue.*/
+    movq mm3,[14*8+RESIDUE]
+    /*#7 Load high residue.*/
+    movq mm4,[15*8+RESIDUE]
+    /*#6 Bias low  residue.*/
+    paddsw mm1,mm0
+    /*#6 Bias high residue.*/
+    paddsw mm2,mm0
+    /*#6 Pack to byte.*/
+    packuswb mm1,mm2
+    /*#7 Bias low  residue.*/
+    paddsw mm3,mm0
+    /*#7 Bias high residue.*/
+    paddsw mm4,mm0
+    /*#7 Pack to byte.*/
+    packuswb mm3,mm4
+    /*#6 Write row.*/
+    movq [DST4+YSTRIDE*2],mm1
+    /*#7 Write row.*/
+    movq [DST4+YSTRIDE3],mm3
+#undef DST
+#undef DST4
+#undef YSTRIDE
+#undef YSTRIDE3
+#undef RESIDUE
+  }
+}
+
+void oc_frag_recon_inter_mmx(unsigned char *_dst,const unsigned char *_src,
+ int _ystride,const ogg_int16_t *_residue){
+  int i;
+  /*Zero mm0.*/
+  __asm pxor mm0,mm0;
+  for(i=4;i-->0;){
+    __asm{
+#define DST edx
+#define SRC ecx
+#define YSTRIDE edi
+#define RESIDUE eax
+      mov DST,_dst
+      mov SRC,_src
+      mov YSTRIDE,_ystride
+      mov RESIDUE,_residue
+      /*#0 Load source.*/
+      movq mm3,[SRC]
+      /*#1 Load source.*/
+      movq mm7,[SRC+YSTRIDE]
+      /*#0 Get copy of src.*/
+      movq mm4,mm3
+      /*#0 Expand high source.*/
+      punpckhbw mm4,mm0
+      /*#0 Expand low  source.*/
+      punpcklbw mm3,mm0
+      /*#0 Add residue high.*/
+      paddsw mm4,[8+RESIDUE]
+      /*#1 Get copy of src.*/
+      movq mm2,mm7
+      /*#0 Add residue low.*/
+      paddsw  mm3,[RESIDUE]
+      /*#1 Expand high source.*/
+      punpckhbw mm2,mm0
+      /*#0 Pack final row pixels.*/
+      packuswb mm3,mm4
+      /*#1 Expand low  source.*/
+      punpcklbw mm7,mm0
+      /*#1 Add residue low.*/
+      paddsw mm7,[16+RESIDUE]
+      /*#1 Add residue high.*/
+      paddsw mm2,[24+RESIDUE]
+      /*Advance residue.*/
+      lea RESIDUE,[32+RESIDUE]
+      /*#1 Pack final row pixels.*/
+      packuswb mm7,mm2
+      /*Advance src.*/
+      lea SRC,[SRC+YSTRIDE*2]
+      /*#0 Write row.*/
+      movq [DST],mm3
+      /*#1 Write row.*/
+      movq [DST+YSTRIDE],mm7
+      /*Advance dst.*/
+      lea DST,[DST+YSTRIDE*2]
+      mov _residue,RESIDUE
+      mov _dst,DST
+      mov _src,SRC
+#undef DST
+#undef SRC
+#undef YSTRIDE
+#undef RESIDUE
+    }
+  }
+}
+
+void oc_frag_recon_inter2_mmx(unsigned char *_dst,const unsigned char *_src1,
+ const unsigned char *_src2,int _ystride,const ogg_int16_t *_residue){
+  int i;
+  /*Zero mm7.*/
+  __asm pxor mm7,mm7;
+  for(i=4;i-->0;){
+    __asm{
+#define SRC1 ecx
+#define SRC2 edi
+#define YSTRIDE esi
+#define RESIDUE edx
+#define DST eax
+      mov YSTRIDE,_ystride
+      mov DST,_dst
+      mov RESIDUE,_residue
+      mov SRC1,_src1
+      mov SRC2,_src2
+      /*#0 Load src1.*/
+      movq mm0,[SRC1]
+      /*#0 Load src2.*/
+      movq mm2,[SRC2]
+      /*#0 Copy src1.*/
+      movq mm1,mm0
+      /*#0 Copy src2.*/
+      movq mm3,mm2
+      /*#1 Load src1.*/
+      movq mm4,[SRC1+YSTRIDE]
+      /*#0 Unpack lower src1.*/
+      punpcklbw mm0,mm7
+      /*#1 Load src2.*/
+      movq mm5,[SRC2+YSTRIDE]
+      /*#0 Unpack higher src1.*/
+      punpckhbw mm1,mm7
+      /*#0 Unpack lower src2.*/
+      punpcklbw mm2,mm7
+      /*#0 Unpack higher src2.*/
+      punpckhbw mm3,mm7
+      /*Advance src1 ptr.*/
+      lea SRC1,[SRC1+YSTRIDE*2]
+      /*Advance src2 ptr.*/
+      lea SRC2,[SRC2+YSTRIDE*2]
+      /*#0 Lower src1+src2.*/
+      paddsw mm0,mm2
+      /*#0 Higher src1+src2.*/
+      paddsw mm1,mm3
+      /*#1 Copy src1.*/
+      movq mm2,mm4
+      /*#0 Build lo average.*/
+      psraw mm0,1
+      /*#1 Copy src2.*/
+      movq mm3,mm5
+      /*#1 Unpack lower src1.*/
+      punpcklbw mm4,mm7
+      /*#0 Build hi average.*/
+      psraw mm1,1
+      /*#1 Unpack higher src1.*/
+      punpckhbw mm2,mm7
+      /*#0 low+=residue.*/
+      paddsw mm0,[RESIDUE]
+      /*#1 Unpack lower src2.*/
+      punpcklbw mm5,mm7
+      /*#0 high+=residue.*/
+      paddsw mm1,[8+RESIDUE]
+      /*#1 Unpack higher src2.*/
+      punpckhbw mm3,mm7
+      /*#1 Lower src1+src2.*/
+      paddsw mm5,mm4
+      /*#0 Pack and saturate.*/
+      packuswb mm0,mm1
+      /*#1 Higher src1+src2.*/
+      paddsw mm3,mm2
+      /*#0 Write row.*/
+      movq [DST],mm0
+      /*#1 Build lo average.*/
+      psraw mm5,1
+      /*#1 Build hi average.*/
+      psraw mm3,1
+      /*#1 low+=residue.*/
+      paddsw mm5,[16+RESIDUE]
+      /*#1 high+=residue.*/
+      paddsw mm3,[24+RESIDUE]
+      /*#1 Pack and saturate.*/
+      packuswb  mm5,mm3
+      /*#1 Write row ptr.*/
+      movq [DST+YSTRIDE],mm5
+      /*Advance residue ptr.*/
+      add RESIDUE,32
+      /*Advance dest ptr.*/
+      lea DST,[DST+YSTRIDE*2]
+      mov _dst,DST
+      mov _residue,RESIDUE
+      mov _src1,SRC1
+      mov _src2,SRC2
+#undef SRC1
+#undef SRC2
+#undef YSTRIDE
+#undef RESIDUE
+#undef DST
+    }
+  }
+}
+
+void oc_restore_fpu_mmx(void){
+  __asm emms;
+}
+
+#endif

thirdparty/libtheora/x86_vc/mmxidct.c

@@ -0,0 +1,592 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+/*MMX acceleration of Theora's iDCT.
+  Originally written by Rudolf Marek, based on code from On2's VP3.*/
+#include "x86int.h"
+#include "../dct.h"
+
+#if defined(OC_X86_ASM)
+
+/*These are offsets into the table of constants below.*/
+/*7 rows of cosines, in order: pi/16 * (1 ... 7).*/
+#define OC_COSINE_OFFSET (8)
+/*A row of 8's.*/
+#define OC_EIGHT_OFFSET  (0)
+
+
+
+/*A table of constants used by the MMX routines.*/
+static const OC_ALIGN16(ogg_uint16_t) OC_IDCT_CONSTS[(1+7)*4]={
+      8,    8,    8,    8,
+  (ogg_uint16_t)OC_C1S7,(ogg_uint16_t)OC_C1S7,
+  (ogg_uint16_t)OC_C1S7,(ogg_uint16_t)OC_C1S7,
+  (ogg_uint16_t)OC_C2S6,(ogg_uint16_t)OC_C2S6,
+  (ogg_uint16_t)OC_C2S6,(ogg_uint16_t)OC_C2S6,
+  (ogg_uint16_t)OC_C3S5,(ogg_uint16_t)OC_C3S5,
+  (ogg_uint16_t)OC_C3S5,(ogg_uint16_t)OC_C3S5,
+  (ogg_uint16_t)OC_C4S4,(ogg_uint16_t)OC_C4S4,
+  (ogg_uint16_t)OC_C4S4,(ogg_uint16_t)OC_C4S4,
+  (ogg_uint16_t)OC_C5S3,(ogg_uint16_t)OC_C5S3,
+  (ogg_uint16_t)OC_C5S3,(ogg_uint16_t)OC_C5S3,
+  (ogg_uint16_t)OC_C6S2,(ogg_uint16_t)OC_C6S2,
+  (ogg_uint16_t)OC_C6S2,(ogg_uint16_t)OC_C6S2,
+  (ogg_uint16_t)OC_C7S1,(ogg_uint16_t)OC_C7S1,
+  (ogg_uint16_t)OC_C7S1,(ogg_uint16_t)OC_C7S1
+};
+
+/*38 cycles*/
+#define OC_IDCT_BEGIN(_y,_x) __asm{ \
+  __asm movq mm2,OC_I(3,_x) \
+  __asm movq mm6,OC_C(3) \
+  __asm movq mm4,mm2 \
+  __asm movq mm7,OC_J(5,_x) \
+  __asm pmulhw mm4,mm6 \
+  __asm movq mm1,OC_C(5) \
+  __asm pmulhw mm6,mm7 \
+  __asm movq mm5,mm1 \
+  __asm pmulhw mm1,mm2 \
+  __asm movq mm3,OC_I(1,_x) \
+  __asm pmulhw mm5,mm7 \
+  __asm movq mm0,OC_C(1) \
+  __asm paddw mm4,mm2 \
+  __asm paddw mm6,mm7 \
+  __asm paddw mm2,mm1 \
+  __asm movq mm1,OC_J(7,_x) \
+  __asm paddw mm7,mm5 \
+  __asm movq mm5,mm0 \
+  __asm pmulhw mm0,mm3 \
+  __asm paddw mm4,mm7 \
+  __asm pmulhw mm5,mm1 \
+  __asm movq mm7,OC_C(7) \
+  __asm psubw mm6,mm2 \
+  __asm paddw mm0,mm3 \
+  __asm pmulhw mm3,mm7 \
+  __asm movq mm2,OC_I(2,_x) \
+  __asm pmulhw mm7,mm1 \
+  __asm paddw mm5,mm1 \
+  __asm movq mm1,mm2 \
+  __asm pmulhw mm2,OC_C(2) \
+  __asm psubw mm3,mm5 \
+  __asm movq mm5,OC_J(6,_x) \
+  __asm paddw mm0,mm7 \
+  __asm movq mm7,mm5 \
+  __asm psubw mm0,mm4 \
+  __asm pmulhw mm5,OC_C(2) \
+  __asm paddw mm2,mm1 \
+  __asm pmulhw mm1,OC_C(6) \
+  __asm paddw mm4,mm4 \
+  __asm paddw mm4,mm0 \
+  __asm psubw mm3,mm6 \
+  __asm paddw mm5,mm7 \
+  __asm paddw mm6,mm6 \
+  __asm pmulhw mm7,OC_C(6) \
+  __asm paddw mm6,mm3 \
+  __asm movq OC_I(1,_y),mm4 \
+  __asm psubw mm1,mm5 \
+  __asm movq mm4,OC_C(4) \
+  __asm movq mm5,mm3 \
+  __asm pmulhw mm3,mm4 \
+  __asm paddw mm7,mm2 \
+  __asm movq OC_I(2,_y),mm6 \
+  __asm movq mm2,mm0 \
+  __asm movq mm6,OC_I(0,_x) \
+  __asm pmulhw mm0,mm4 \
+  __asm paddw mm5,mm3 \
+  __asm movq mm3,OC_J(4,_x) \
+  __asm psubw mm5,mm1 \
+  __asm paddw mm2,mm0 \
+  __asm psubw mm6,mm3 \
+  __asm movq mm0,mm6 \
+  __asm pmulhw mm6,mm4 \
+  __asm paddw mm3,mm3 \
+  __asm paddw mm1,mm1 \
+  __asm paddw mm3,mm0 \
+  __asm paddw mm1,mm5 \
+  __asm pmulhw mm4,mm3 \
+  __asm paddw mm6,mm0 \
+  __asm psubw mm6,mm2 \
+  __asm paddw mm2,mm2 \
+  __asm movq mm0,OC_I(1,_y) \
+  __asm paddw mm2,mm6 \
+  __asm paddw mm4,mm3 \
+  __asm psubw mm2,mm1 \
+}
+
+/*38+8=46 cycles.*/
+#define OC_ROW_IDCT(_y,_x) __asm{ \
+  OC_IDCT_BEGIN(_y,_x) \
+  /*r3=D'*/ \
+  __asm  movq mm3,OC_I(2,_y) \
+  /*r4=E'=E-G*/ \
+  __asm  psubw mm4,mm7 \
+  /*r1=H'+H'*/ \
+  __asm  paddw mm1,mm1 \
+  /*r7=G+G*/ \
+  __asm  paddw mm7,mm7 \
+  /*r1=R1=A''+H'*/ \
+  __asm  paddw mm1,mm2 \
+  /*r7=G'=E+G*/ \
+  __asm  paddw mm7,mm4 \
+  /*r4=R4=E'-D'*/ \
+  __asm  psubw mm4,mm3 \
+  __asm  paddw mm3,mm3 \
+  /*r6=R6=F'-B''*/ \
+  __asm  psubw mm6,mm5 \
+  __asm  paddw mm5,mm5 \
+  /*r3=R3=E'+D'*/ \
+  __asm  paddw mm3,mm4 \
+  /*r5=R5=F'+B''*/ \
+  __asm  paddw mm5,mm6 \
+  /*r7=R7=G'-C'*/ \
+  __asm  psubw mm7,mm0 \
+  __asm  paddw mm0,mm0 \
+  /*Save R1.*/ \
+  __asm  movq OC_I(1,_y),mm1 \
+  /*r0=R0=G.+C.*/ \
+  __asm  paddw mm0,mm7 \
+}
+
+/*The following macro does two 4x4 transposes in place.
+  At entry, we assume:
+    r0 = a3 a2 a1 a0
+  I(1) = b3 b2 b1 b0
+    r2 = c3 c2 c1 c0
+    r3 = d3 d2 d1 d0
+
+    r4 = e3 e2 e1 e0
+    r5 = f3 f2 f1 f0
+    r6 = g3 g2 g1 g0
+    r7 = h3 h2 h1 h0
+
+  At exit, we have:
+  I(0) = d0 c0 b0 a0
+  I(1) = d1 c1 b1 a1
+  I(2) = d2 c2 b2 a2
+  I(3) = d3 c3 b3 a3
+
+  J(4) = h0 g0 f0 e0
+  J(5) = h1 g1 f1 e1
+  J(6) = h2 g2 f2 e2
+  J(7) = h3 g3 f3 e3
+
+  I(0) I(1) I(2) I(3) is the transpose of r0 I(1) r2 r3.
+  J(4) J(5) J(6) J(7) is the transpose of r4  r5  r6 r7.
+
+  Since r1 is free at entry, we calculate the Js first.*/
+/*19 cycles.*/
+#define OC_TRANSPOSE(_y) __asm{ \
+  __asm movq mm1,mm4 \
+  __asm punpcklwd mm4,mm5 \
+  __asm movq OC_I(0,_y),mm0 \
+  __asm punpckhwd mm1,mm5 \
+  __asm movq mm0,mm6 \
+  __asm punpcklwd mm6,mm7 \
+  __asm movq mm5,mm4 \
+  __asm punpckldq mm4,mm6 \
+  __asm punpckhdq mm5,mm6 \
+  __asm movq mm6,mm1 \
+  __asm movq OC_J(4,_y),mm4 \
+  __asm punpckhwd mm0,mm7 \
+  __asm movq OC_J(5,_y),mm5 \
+  __asm punpckhdq mm6,mm0 \
+  __asm movq mm4,OC_I(0,_y) \
+  __asm punpckldq mm1,mm0 \
+  __asm movq mm5,OC_I(1,_y) \
+  __asm movq mm0,mm4 \
+  __asm movq OC_J(7,_y),mm6 \
+  __asm punpcklwd mm0,mm5 \
+  __asm movq OC_J(6,_y),mm1 \
+  __asm punpckhwd mm4,mm5 \
+  __asm movq mm5,mm2 \
+  __asm punpcklwd mm2,mm3 \
+  __asm movq mm1,mm0 \
+  __asm punpckldq mm0,mm2 \
+  __asm punpckhdq mm1,mm2 \
+  __asm movq mm2,mm4 \
+  __asm movq OC_I(0,_y),mm0 \
+  __asm punpckhwd mm5,mm3 \
+  __asm movq OC_I(1,_y),mm1 \
+  __asm punpckhdq mm4,mm5 \
+  __asm punpckldq mm2,mm5 \
+  __asm movq OC_I(3,_y),mm4 \
+  __asm movq OC_I(2,_y),mm2 \
+}
+
+/*38+19=57 cycles.*/
+#define OC_COLUMN_IDCT(_y) __asm{ \
+  OC_IDCT_BEGIN(_y,_y) \
+  __asm paddw mm2,OC_8 \
+  /*r1=H'+H'*/ \
+  __asm paddw mm1,mm1 \
+  /*r1=R1=A''+H'*/ \
+  __asm paddw mm1,mm2 \
+  /*r2=NR2*/ \
+  __asm psraw mm2,4 \
+  /*r4=E'=E-G*/ \
+  __asm psubw mm4,mm7 \
+  /*r1=NR1*/ \
+  __asm psraw mm1,4 \
+  /*r3=D'*/ \
+  __asm movq mm3,OC_I(2,_y) \
+  /*r7=G+G*/ \
+  __asm paddw mm7,mm7 \
+  /*Store NR2 at I(2).*/ \
+  __asm movq OC_I(2,_y),mm2 \
+  /*r7=G'=E+G*/ \
+  __asm paddw mm7,mm4 \
+  /*Store NR1 at I(1).*/ \
+  __asm movq OC_I(1,_y),mm1 \
+  /*r4=R4=E'-D'*/ \
+  __asm psubw mm4,mm3 \
+  __asm paddw mm4,OC_8 \
+  /*r3=D'+D'*/ \
+  __asm paddw mm3,mm3 \
+  /*r3=R3=E'+D'*/ \
+  __asm paddw mm3,mm4 \
+  /*r4=NR4*/ \
+  __asm psraw mm4,4 \
+  /*r6=R6=F'-B''*/ \
+  __asm psubw mm6,mm5 \
+  /*r3=NR3*/ \
+  __asm psraw mm3,4 \
+  __asm paddw mm6,OC_8 \
+  /*r5=B''+B''*/ \
+  __asm paddw mm5,mm5 \
+  /*r5=R5=F'+B''*/ \
+  __asm paddw mm5,mm6 \
+  /*r6=NR6*/ \
+  __asm psraw mm6,4 \
+  /*Store NR4 at J(4).*/ \
+  __asm movq OC_J(4,_y),mm4 \
+  /*r5=NR5*/ \
+  __asm psraw mm5,4 \
+  /*Store NR3 at I(3).*/ \
+  __asm movq OC_I(3,_y),mm3 \
+  /*r7=R7=G'-C'*/ \
+  __asm psubw mm7,mm0 \
+  __asm paddw mm7,OC_8 \
+  /*r0=C'+C'*/ \
+  __asm paddw mm0,mm0 \
+  /*r0=R0=G'+C'*/ \
+  __asm paddw mm0,mm7 \
+  /*r7=NR7*/ \
+  __asm psraw mm7,4 \
+  /*Store NR6 at J(6).*/ \
+  __asm movq OC_J(6,_y),mm6 \
+  /*r0=NR0*/ \
+  __asm psraw mm0,4 \
+  /*Store NR5 at J(5).*/ \
+  __asm movq OC_J(5,_y),mm5 \
+  /*Store NR7 at J(7).*/ \
+  __asm movq OC_J(7,_y),mm7 \
+  /*Store NR0 at I(0).*/ \
+  __asm movq OC_I(0,_y),mm0 \
+}
+
+#define OC_MID(_m,_i) [CONSTS+_m+(_i)*8]
+#define OC_C(_i)      OC_MID(OC_COSINE_OFFSET,_i-1)
+#define OC_8          OC_MID(OC_EIGHT_OFFSET,0)
+
+static void oc_idct8x8_slow(ogg_int16_t _y[64],ogg_int16_t _x[64]){
+  int i;
+  /*This routine accepts an 8x8 matrix, but in partially transposed form.
+    Every 4x4 block is transposed.*/
+  __asm{
+#define CONSTS eax
+#define Y edx
+#define X ecx
+    mov CONSTS,offset OC_IDCT_CONSTS
+    mov Y,_y
+    mov X,_x
+#define OC_I(_k,_y)   [(_y)+(_k)*16]
+#define OC_J(_k,_y)   [(_y)+((_k)-4)*16+8]
+    OC_ROW_IDCT(Y,X)
+    OC_TRANSPOSE(Y)
+#undef  OC_I
+#undef  OC_J
+#define OC_I(_k,_y)   [(_y)+(_k)*16+64]
+#define OC_J(_k,_y)   [(_y)+((_k)-4)*16+72]
+    OC_ROW_IDCT(Y,X)
+    OC_TRANSPOSE(Y)
+#undef  OC_I
+#undef  OC_J
+#define OC_I(_k,_y)   [(_y)+(_k)*16]
+#define OC_J(_k,_y)   OC_I(_k,_y)
+    OC_COLUMN_IDCT(Y)
+#undef  OC_I
+#undef  OC_J
+#define OC_I(_k,_y)   [(_y)+(_k)*16+8]
+#define OC_J(_k,_y)   OC_I(_k,_y)
+    OC_COLUMN_IDCT(Y)
+#undef  OC_I
+#undef  OC_J
+#undef  CONSTS
+#undef  Y
+#undef  X
+  }
+  __asm pxor mm0,mm0;
+  for(i=0;i<4;i++){
+    ogg_int16_t *x;
+    x=_x+16*i;
+#define X ecx
+    __asm{
+      mov X,x
+      movq [X+0x00],mm0
+      movq [X+0x08],mm0
+      movq [X+0x10],mm0
+      movq [X+0x18],mm0
+    }
+#undef  X
+  }
+}
+
+/*25 cycles.*/
+#define OC_IDCT_BEGIN_10(_y,_x) __asm{ \
+  __asm movq mm2,OC_I(3,_x) \
+  __asm nop \
+  __asm movq mm6,OC_C(3) \
+  __asm movq mm4,mm2 \
+  __asm movq mm1,OC_C(5) \
+  __asm pmulhw mm4,mm6 \
+  __asm movq mm3,OC_I(1,_x) \
+  __asm pmulhw mm1,mm2 \
+  __asm movq mm0,OC_C(1) \
+  __asm paddw mm4,mm2 \
+  __asm pxor mm6,mm6 \
+  __asm paddw mm2,mm1 \
+  __asm movq mm5,OC_I(2,_x) \
+  __asm pmulhw mm0,mm3 \
+  __asm movq mm1,mm5 \
+  __asm paddw mm0,mm3 \
+  __asm pmulhw mm3,OC_C(7) \
+  __asm psubw mm6,mm2 \
+  __asm pmulhw mm5,OC_C(2) \
+  __asm psubw mm0,mm4 \
+  __asm movq mm7,OC_I(2,_x) \
+  __asm paddw mm4,mm4 \
+  __asm paddw mm7,mm5 \
+  __asm paddw mm4,mm0 \
+  __asm pmulhw mm1,OC_C(6) \
+  __asm psubw mm3,mm6 \
+  __asm movq OC_I(1,_y),mm4 \
+  __asm paddw mm6,mm6 \
+  __asm movq mm4,OC_C(4) \
+  __asm paddw mm6,mm3 \
+  __asm movq mm5,mm3 \
+  __asm pmulhw mm3,mm4 \
+  __asm movq OC_I(2,_y),mm6 \
+  __asm movq mm2,mm0 \
+  __asm movq mm6,OC_I(0,_x) \
+  __asm pmulhw mm0,mm4 \
+  __asm paddw mm5,mm3 \
+  __asm paddw mm2,mm0 \
+  __asm psubw mm5,mm1 \
+  __asm pmulhw mm6,mm4 \
+  __asm paddw mm6,OC_I(0,_x) \
+  __asm paddw mm1,mm1 \
+  __asm movq mm4,mm6 \
+  __asm paddw mm1,mm5 \
+  __asm psubw mm6,mm2 \
+  __asm paddw mm2,mm2 \
+  __asm movq mm0,OC_I(1,_y) \
+  __asm paddw mm2,mm6 \
+  __asm psubw mm2,mm1 \
+  __asm nop \
+}
+
+/*25+8=33 cycles.*/
+#define OC_ROW_IDCT_10(_y,_x) __asm{ \
+  OC_IDCT_BEGIN_10(_y,_x) \
+  /*r3=D'*/ \
+   __asm movq mm3,OC_I(2,_y) \
+  /*r4=E'=E-G*/ \
+   __asm psubw mm4,mm7 \
+  /*r1=H'+H'*/ \
+   __asm paddw mm1,mm1 \
+  /*r7=G+G*/ \
+   __asm paddw mm7,mm7 \
+  /*r1=R1=A''+H'*/ \
+   __asm paddw mm1,mm2 \
+  /*r7=G'=E+G*/ \
+   __asm paddw mm7,mm4 \
+  /*r4=R4=E'-D'*/ \
+   __asm psubw mm4,mm3 \
+   __asm paddw mm3,mm3 \
+  /*r6=R6=F'-B''*/ \
+   __asm psubw mm6,mm5 \
+   __asm paddw mm5,mm5 \
+  /*r3=R3=E'+D'*/ \
+   __asm paddw mm3,mm4 \
+  /*r5=R5=F'+B''*/ \
+   __asm paddw mm5,mm6 \
+  /*r7=R7=G'-C'*/ \
+   __asm psubw mm7,mm0 \
+   __asm paddw mm0,mm0 \
+  /*Save R1.*/ \
+   __asm movq OC_I(1,_y),mm1 \
+  /*r0=R0=G'+C'*/ \
+   __asm paddw mm0,mm7 \
+}
+
+/*25+19=44 cycles'*/
+#define OC_COLUMN_IDCT_10(_y) __asm{ \
+  OC_IDCT_BEGIN_10(_y,_y) \
+  __asm paddw mm2,OC_8 \
+  /*r1=H'+H'*/ \
+  __asm paddw mm1,mm1 \
+  /*r1=R1=A''+H'*/ \
+  __asm paddw mm1,mm2 \
+  /*r2=NR2*/ \
+  __asm psraw mm2,4 \
+  /*r4=E'=E-G*/ \
+  __asm psubw mm4,mm7 \
+  /*r1=NR1*/ \
+  __asm psraw mm1,4 \
+  /*r3=D'*/ \
+  __asm movq mm3,OC_I(2,_y) \
+  /*r7=G+G*/ \
+  __asm paddw mm7,mm7 \
+  /*Store NR2 at I(2).*/ \
+  __asm movq OC_I(2,_y),mm2 \
+  /*r7=G'=E+G*/ \
+  __asm paddw mm7,mm4 \
+  /*Store NR1 at I(1).*/ \
+  __asm movq OC_I(1,_y),mm1 \
+  /*r4=R4=E'-D'*/ \
+  __asm psubw mm4,mm3 \
+  __asm paddw mm4,OC_8 \
+  /*r3=D'+D'*/ \
+  __asm paddw mm3,mm3 \
+  /*r3=R3=E'+D'*/ \
+  __asm paddw mm3,mm4 \
+  /*r4=NR4*/ \
+  __asm psraw mm4,4 \
+  /*r6=R6=F'-B''*/ \
+  __asm psubw mm6,mm5 \
+  /*r3=NR3*/ \
+  __asm psraw mm3,4 \
+  __asm paddw mm6,OC_8 \
+  /*r5=B''+B''*/ \
+  __asm paddw mm5,mm5 \
+  /*r5=R5=F'+B''*/ \
+  __asm paddw mm5,mm6 \
+  /*r6=NR6*/ \
+  __asm psraw mm6,4 \
+  /*Store NR4 at J(4).*/ \
+  __asm movq OC_J(4,_y),mm4 \
+  /*r5=NR5*/ \
+  __asm psraw mm5,4 \
+  /*Store NR3 at I(3).*/ \
+  __asm movq OC_I(3,_y),mm3 \
+  /*r7=R7=G'-C'*/ \
+  __asm psubw mm7,mm0 \
+  __asm paddw mm7,OC_8 \
+  /*r0=C'+C'*/ \
+  __asm paddw mm0,mm0 \
+  /*r0=R0=G'+C'*/ \
+  __asm paddw mm0,mm7 \
+  /*r7=NR7*/ \
+  __asm psraw mm7,4 \
+  /*Store NR6 at J(6).*/ \
+  __asm movq OC_J(6,_y),mm6 \
+  /*r0=NR0*/ \
+  __asm psraw mm0,4 \
+  /*Store NR5 at J(5).*/ \
+  __asm movq OC_J(5,_y),mm5 \
+  /*Store NR7 at J(7).*/ \
+  __asm movq OC_J(7,_y),mm7 \
+  /*Store NR0 at I(0).*/ \
+  __asm movq OC_I(0,_y),mm0 \
+}
+
+static void oc_idct8x8_10(ogg_int16_t _y[64],ogg_int16_t _x[64]){
+  __asm{
+#define CONSTS eax
+#define Y edx
+#define X ecx
+    mov CONSTS,offset OC_IDCT_CONSTS
+    mov Y,_y
+    mov X,_x
+#define OC_I(_k,_y) [(_y)+(_k)*16]
+#define OC_J(_k,_y) [(_y)+((_k)-4)*16+8]
+    /*Done with dequant, descramble, and partial transpose.
+      Now do the iDCT itself.*/
+    OC_ROW_IDCT_10(Y,X)
+    OC_TRANSPOSE(Y)
+#undef  OC_I
+#undef  OC_J
+#define OC_I(_k,_y) [(_y)+(_k)*16]
+#define OC_J(_k,_y) OC_I(_k,_y)
+    OC_COLUMN_IDCT_10(Y)
+#undef  OC_I
+#undef  OC_J
+#define OC_I(_k,_y) [(_y)+(_k)*16+8]
+#define OC_J(_k,_y) OC_I(_k,_y)
+    OC_COLUMN_IDCT_10(Y)
+#undef  OC_I
+#undef  OC_J
+#undef  CONSTS
+#undef  Y
+#undef  X
+  }
+#define X ecx
+  __asm{
+    pxor mm0,mm0;
+    mov X,_x
+    movq [X+0x00],mm0
+    movq [X+0x10],mm0
+    movq [X+0x20],mm0
+    movq [X+0x30],mm0
+  }
+#undef  X
+}
+
+/*Performs an inverse 8x8 Type-II DCT transform.
+  The input is assumed to be scaled by a factor of 4 relative to orthonormal
+   version of the transform.*/
+void oc_idct8x8_mmx(ogg_int16_t _y[64],ogg_int16_t _x[64],int _last_zzi){
+  /*_last_zzi is subtly different from an actual count of the number of
+     coefficients we decoded for this block.
+    It contains the value of zzi BEFORE the final token in the block was
+     decoded.
+    In most cases this is an EOB token (the continuation of an EOB run from a
+     previous block counts), and so this is the same as the coefficient count.
+    However, in the case that the last token was NOT an EOB token, but filled
+     the block up with exactly 64 coefficients, _last_zzi will be less than 64.
+    Provided the last token was not a pure zero run, the minimum value it can
+     be is 46, and so that doesn't affect any of the cases in this routine.
+    However, if the last token WAS a pure zero run of length 63, then _last_zzi
+     will be 1 while the number of coefficients decoded is 64.
+    Thus, we will trigger the following special case, where the real
+     coefficient count would not.
+    Note also that a zero run of length 64 will give _last_zzi a value of 0,
+     but we still process the DC coefficient, which might have a non-zero value
+     due to DC prediction.
+    Although convoluted, this is arguably the correct behavior: it allows us to
+     use a smaller transform when the block ends with a long zero run instead
+     of a normal EOB token.
+    It could be smarter... multiple separate zero runs at the end of a block
+     will fool it, but an encoder that generates these really deserves what it
+     gets.
+    Needless to say we inherited this approach from VP3.*/
+  /*Perform the iDCT.*/
+  if(_last_zzi<=10)oc_idct8x8_10(_y,_x);
+  else oc_idct8x8_slow(_y,_x);
+}
+
+#endif

thirdparty/libtheora/x86_vc/mmxloop.h

@@ -0,0 +1,219 @@
+#if !defined(_x86_vc_mmxloop_H)
+# define _x86_vc_mmxloop_H (1)
+# include <stddef.h>
+# include "x86int.h"
+
+#if defined(OC_X86_ASM)
+
+/*On entry, mm0={a0,...,a7}, mm1={b0,...,b7}, mm2={c0,...,c7}, mm3={d0,...d7}.
+  On exit, mm1={b0+lflim(R_0,L),...,b7+lflim(R_7,L)} and
+   mm2={c0-lflim(R_0,L),...,c7-lflim(R_7,L)}; mm0 and mm3 are clobbered.*/
+#define OC_LOOP_FILTER8_MMX __asm{ \
+  /*mm7=0*/ \
+  __asm pxor mm7,mm7 \
+  /*mm6:mm0={a0,...,a7}*/ \
+  __asm movq mm6,mm0 \
+  __asm punpcklbw mm0,mm7 \
+  __asm punpckhbw mm6,mm7 \
+  /*mm3:mm5={d0,...,d7}*/ \
+  __asm movq mm5,mm3 \
+  __asm punpcklbw mm3,mm7 \
+  __asm punpckhbw mm5,mm7 \
+  /*mm6:mm0={a0-d0,...,a7-d7}*/ \
+  __asm psubw mm0,mm3 \
+  __asm psubw mm6,mm5 \
+  /*mm3:mm1={b0,...,b7}*/ \
+  __asm movq mm3,mm1 \
+  __asm punpcklbw mm1,mm7 \
+  __asm movq mm4,mm2 \
+  __asm punpckhbw mm3,mm7 \
+  /*mm5:mm4={c0,...,c7}*/ \
+  __asm movq mm5,mm2 \
+  __asm punpcklbw mm4,mm7 \
+  __asm punpckhbw mm5,mm7 \
+  /*mm7={3}x4 \
+    mm5:mm4={c0-b0,...,c7-b7}*/ \
+  __asm pcmpeqw mm7,mm7 \
+  __asm psubw mm4,mm1 \
+  __asm psrlw mm7,14 \
+  __asm psubw mm5,mm3 \
+  /*Scale by 3.*/ \
+  __asm pmullw mm4,mm7 \
+  __asm pmullw mm5,mm7 \
+  /*mm7={4}x4 \
+    mm5:mm4=f={a0-d0+3*(c0-b0),...,a7-d7+3*(c7-b7)}*/ \
+  __asm psrlw mm7,1 \
+  __asm paddw mm4,mm0 \
+  __asm psllw mm7,2 \
+  __asm movq mm0,[LL] \
+  __asm paddw mm5,mm6 \
+  /*R_i has the range [-127,128], so we compute -R_i instead. \
+    mm4=-R_i=-(f+4>>3)=0xFF^(f-4>>3)*/ \
+  __asm psubw mm4,mm7 \
+  __asm psubw mm5,mm7 \
+  __asm psraw mm4,3 \
+  __asm psraw mm5,3 \
+  __asm pcmpeqb mm7,mm7 \
+  __asm packsswb mm4,mm5 \
+  __asm pxor mm6,mm6 \
+  __asm pxor mm4,mm7 \
+  __asm packuswb mm1,mm3 \
+  /*Now compute lflim of -mm4 cf. Section 7.10 of the sepc.*/ \
+  /*There's no unsigned byte+signed byte with unsigned saturation op code, so \
+     we have to split things by sign (the other option is to work in 16 bits, \
+     but working in 8 bits gives much better parallelism). \
+    We compute abs(R_i), but save a mask of which terms were negative in mm6. \
+    Then we compute mm4=abs(lflim(R_i,L))=min(abs(R_i),max(2*L-abs(R_i),0)). \
+    Finally, we split mm4 into positive and negative pieces using the mask in \
+     mm6, and add and subtract them as appropriate.*/ \
+  /*mm4=abs(-R_i)*/ \
+  /*mm7=255-2*L*/ \
+  __asm pcmpgtb mm6,mm4 \
+  __asm psubb mm7,mm0 \
+  __asm pxor mm4,mm6 \
+  __asm psubb mm7,mm0 \
+  __asm psubb mm4,mm6 \
+  /*mm7=255-max(2*L-abs(R_i),0)*/ \
+  __asm paddusb mm7,mm4 \
+  /*mm4=min(abs(R_i),max(2*L-abs(R_i),0))*/ \
+  __asm paddusb mm4,mm7 \
+  __asm psubusb mm4,mm7 \
+  /*Now split mm4 by the original sign of -R_i.*/ \
+  __asm movq mm5,mm4 \
+  __asm pand mm4,mm6 \
+  __asm pandn mm6,mm5 \
+  /*mm1={b0+lflim(R_0,L),...,b7+lflim(R_7,L)}*/ \
+  /*mm2={c0-lflim(R_0,L),...,c7-lflim(R_7,L)}*/ \
+  __asm paddusb mm1,mm4 \
+  __asm psubusb mm2,mm4 \
+  __asm psubusb mm1,mm6 \
+  __asm paddusb mm2,mm6 \
+}
+
+#define OC_LOOP_FILTER_V_MMX(_pix,_ystride,_ll) \
+  do{ \
+    /*Used local variable pix__ in order to fix compilation errors like: \
+       "error C2425: 'SHL' : non-constant expression in 'second operand'".*/ \
+    unsigned char *pix__; \
+    unsigned char *ll__; \
+    ll__=(_ll); \
+    pix__=(_pix); \
+    __asm mov YSTRIDE,_ystride \
+    __asm mov LL,ll__ \
+    __asm mov PIX,pix__ \
+    __asm sub PIX,YSTRIDE \
+    __asm sub PIX,YSTRIDE \
+    /*mm0={a0,...,a7}*/ \
+    __asm movq mm0,[PIX] \
+    /*ystride3=_ystride*3*/ \
+    __asm lea YSTRIDE3,[YSTRIDE+YSTRIDE*2] \
+    /*mm3={d0,...,d7}*/ \
+    __asm movq mm3,[PIX+YSTRIDE3] \
+    /*mm1={b0,...,b7}*/ \
+    __asm movq mm1,[PIX+YSTRIDE] \
+    /*mm2={c0,...,c7}*/ \
+    __asm movq mm2,[PIX+YSTRIDE*2] \
+    OC_LOOP_FILTER8_MMX \
+    /*Write it back out.*/ \
+    __asm movq [PIX+YSTRIDE],mm1 \
+    __asm movq [PIX+YSTRIDE*2],mm2 \
+  } \
+  while(0)
+
+#define OC_LOOP_FILTER_H_MMX(_pix,_ystride,_ll) \
+  do{ \
+    /*Used local variable ll__ in order to fix compilation errors like: \
+       "error C2443: operand size conflict".*/ \
+    unsigned char *ll__; \
+    unsigned char *pix__; \
+    ll__=(_ll); \
+    pix__=(_pix)-2; \
+    __asm mov PIX,pix__ \
+    __asm mov YSTRIDE,_ystride \
+    __asm mov LL,ll__ \
+    /*x x x x d0 c0 b0 a0*/ \
+    __asm movd mm0,[PIX] \
+    /*x x x x d1 c1 b1 a1*/ \
+    __asm movd mm1,[PIX+YSTRIDE] \
+    /*ystride3=_ystride*3*/ \
+    __asm lea YSTRIDE3,[YSTRIDE+YSTRIDE*2] \
+    /*x x x x d2 c2 b2 a2*/ \
+    __asm movd mm2,[PIX+YSTRIDE*2] \
+    /*x x x x d3 c3 b3 a3*/ \
+    __asm lea D,[PIX+YSTRIDE*4] \
+    __asm movd mm3,[PIX+YSTRIDE3] \
+    /*x x x x d4 c4 b4 a4*/ \
+    __asm movd mm4,[D] \
+    /*x x x x d5 c5 b5 a5*/ \
+    __asm movd mm5,[D+YSTRIDE] \
+    /*x x x x d6 c6 b6 a6*/ \
+    __asm movd mm6,[D+YSTRIDE*2] \
+    /*x x x x d7 c7 b7 a7*/ \
+    __asm movd mm7,[D+YSTRIDE3] \
+    /*mm0=d1 d0 c1 c0 b1 b0 a1 a0*/ \
+    __asm punpcklbw mm0,mm1 \
+    /*mm2=d3 d2 c3 c2 b3 b2 a3 a2*/ \
+    __asm punpcklbw mm2,mm3 \
+    /*mm3=d1 d0 c1 c0 b1 b0 a1 a0*/ \
+    __asm movq mm3,mm0 \
+    /*mm0=b3 b2 b1 b0 a3 a2 a1 a0*/ \
+    __asm punpcklwd mm0,mm2 \
+    /*mm3=d3 d2 d1 d0 c3 c2 c1 c0*/ \
+    __asm punpckhwd mm3,mm2 \
+    /*mm1=b3 b2 b1 b0 a3 a2 a1 a0*/ \
+    __asm movq mm1,mm0 \
+    /*mm4=d5 d4 c5 c4 b5 b4 a5 a4*/ \
+    __asm punpcklbw mm4,mm5 \
+    /*mm6=d7 d6 c7 c6 b7 b6 a7 a6*/ \
+    __asm punpcklbw mm6,mm7 \
+    /*mm5=d5 d4 c5 c4 b5 b4 a5 a4*/ \
+    __asm movq mm5,mm4 \
+    /*mm4=b7 b6 b5 b4 a7 a6 a5 a4*/ \
+    __asm punpcklwd mm4,mm6 \
+    /*mm5=d7 d6 d5 d4 c7 c6 c5 c4*/ \
+    __asm punpckhwd mm5,mm6 \
+    /*mm2=d3 d2 d1 d0 c3 c2 c1 c0*/ \
+    __asm movq mm2,mm3 \
+    /*mm0=a7 a6 a5 a4 a3 a2 a1 a0*/ \
+    __asm punpckldq mm0,mm4 \
+    /*mm1=b7 b6 b5 b4 b3 b2 b1 b0*/ \
+    __asm punpckhdq mm1,mm4 \
+    /*mm2=c7 c6 c5 c4 c3 c2 c1 c0*/ \
+    __asm punpckldq mm2,mm5 \
+    /*mm3=d7 d6 d5 d4 d3 d2 d1 d0*/ \
+    __asm punpckhdq mm3,mm5 \
+    OC_LOOP_FILTER8_MMX \
+    /*mm2={b0+R_0'',...,b7+R_7''}*/ \
+    __asm movq mm0,mm1 \
+    /*mm1={b0+R_0'',c0-R_0'',...,b3+R_3'',c3-R_3''}*/ \
+    __asm punpcklbw mm1,mm2 \
+    /*mm2={b4+R_4'',c4-R_4'',...,b7+R_7'',c7-R_7''}*/ \
+    __asm punpckhbw mm0,mm2 \
+    /*[d]=c1 b1 c0 b0*/ \
+    __asm movd D,mm1 \
+    __asm mov [PIX+1],D_WORD \
+    __asm psrlq mm1,32 \
+    __asm shr D,16 \
+    __asm mov [PIX+YSTRIDE+1],D_WORD \
+    /*[d]=c3 b3 c2 b2*/ \
+    __asm movd D,mm1 \
+    __asm mov [PIX+YSTRIDE*2+1],D_WORD \
+    __asm shr D,16 \
+    __asm mov [PIX+YSTRIDE3+1],D_WORD \
+    __asm lea PIX,[PIX+YSTRIDE*4] \
+    /*[d]=c5 b5 c4 b4*/ \
+    __asm movd D,mm0 \
+    __asm mov [PIX+1],D_WORD \
+    __asm psrlq mm0,32 \
+    __asm shr D,16 \
+    __asm mov [PIX+YSTRIDE+1],D_WORD \
+    /*[d]=c7 b7 c6 b6*/ \
+    __asm movd D,mm0 \
+    __asm mov [PIX+YSTRIDE*2+1],D_WORD \
+    __asm shr D,16 \
+    __asm mov [PIX+YSTRIDE3+1],D_WORD \
+  } \
+  while(0)
+
+# endif
+#endif

thirdparty/libtheora/x86_vc/mmxstate.c

@@ -0,0 +1,176 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+/*MMX acceleration of complete fragment reconstruction algorithm.
+  Originally written by Rudolf Marek.*/
+#include <string.h>
+#include "x86int.h"
+#include "mmxloop.h"
+
+#if defined(OC_X86_ASM)
+
+void oc_state_frag_recon_mmx(const oc_theora_state *_state,ptrdiff_t _fragi,
+ int _pli,ogg_int16_t _dct_coeffs[128],int _last_zzi,ogg_uint16_t _dc_quant){
+  unsigned char *dst;
+  ptrdiff_t      frag_buf_off;
+  int            ystride;
+  int            refi;
+  /*Apply the inverse transform.*/
+  /*Special case only having a DC component.*/
+  if(_last_zzi<2){
+    /*Note that this value must be unsigned, to keep the __asm__ block from
+       sign-extending it when it puts it in a register.*/
+    ogg_uint16_t p;
+    /*We round this dequant product (and not any of the others) because there's
+       no iDCT rounding.*/
+    p=(ogg_int16_t)(_dct_coeffs[0]*(ogg_int32_t)_dc_quant+15>>5);
+    /*Fill _dct_coeffs with p.*/
+    __asm{
+#define Y eax
+#define P ecx
+      mov Y,_dct_coeffs
+      movzx P,p
+      lea Y,[Y+128]
+      /*mm0=0000 0000 0000 AAAA*/
+      movd mm0,P
+      /*mm0=0000 0000 AAAA AAAA*/
+      punpcklwd mm0,mm0
+      /*mm0=AAAA AAAA AAAA AAAA*/
+      punpckldq mm0,mm0
+      movq [Y],mm0
+      movq [8+Y],mm0
+      movq [16+Y],mm0
+      movq [24+Y],mm0
+      movq [32+Y],mm0
+      movq [40+Y],mm0
+      movq [48+Y],mm0
+      movq [56+Y],mm0
+      movq [64+Y],mm0
+      movq [72+Y],mm0
+      movq [80+Y],mm0
+      movq [88+Y],mm0
+      movq [96+Y],mm0
+      movq [104+Y],mm0
+      movq [112+Y],mm0
+      movq [120+Y],mm0
+#undef Y
+#undef P
+    }
+  }
+  else{
+    /*Dequantize the DC coefficient.*/
+    _dct_coeffs[0]=(ogg_int16_t)(_dct_coeffs[0]*(int)_dc_quant);
+    oc_idct8x8_mmx(_dct_coeffs+64,_dct_coeffs,_last_zzi);
+  }
+  /*Fill in the target buffer.*/
+  frag_buf_off=_state->frag_buf_offs[_fragi];
+  refi=_state->frags[_fragi].refi;
+  ystride=_state->ref_ystride[_pli];
+  dst=_state->ref_frame_data[OC_FRAME_SELF]+frag_buf_off;
+  if(refi==OC_FRAME_SELF)oc_frag_recon_intra_mmx(dst,ystride,_dct_coeffs+64);
+  else{
+    const unsigned char *ref;
+    int                  mvoffsets[2];
+    ref=_state->ref_frame_data[refi]+frag_buf_off;
+    if(oc_state_get_mv_offsets(_state,mvoffsets,_pli,
+     _state->frag_mvs[_fragi])>1){
+      oc_frag_recon_inter2_mmx(dst,ref+mvoffsets[0],ref+mvoffsets[1],ystride,
+       _dct_coeffs+64);
+    }
+    else oc_frag_recon_inter_mmx(dst,ref+mvoffsets[0],ystride,_dct_coeffs+64);
+  }
+}
+
+/*We copy these entire function to inline the actual MMX routines so that we
+   use only a single indirect call.*/
+
+void oc_loop_filter_init_mmx(signed char _bv[256],int _flimit){
+  memset(_bv,~(_flimit<<1),8);
+}
+
+/*Apply the loop filter to a given set of fragment rows in the given plane.
+  The filter may be run on the bottom edge, affecting pixels in the next row of
+   fragments, so this row also needs to be available.
+  _bv:        The bounding values array.
+  _refi:      The index of the frame buffer to filter.
+  _pli:       The color plane to filter.
+  _fragy0:    The Y coordinate of the first fragment row to filter.
+  _fragy_end: The Y coordinate of the fragment row to stop filtering at.*/
+void oc_state_loop_filter_frag_rows_mmx(const oc_theora_state *_state,
+ signed char _bv[256],int _refi,int _pli,int _fragy0,int _fragy_end){
+  const oc_fragment_plane *fplane;
+  const oc_fragment       *frags;
+  const ptrdiff_t         *frag_buf_offs;
+  unsigned char           *ref_frame_data;
+  ptrdiff_t                fragi_top;
+  ptrdiff_t                fragi_bot;
+  ptrdiff_t                fragi0;
+  ptrdiff_t                fragi0_end;
+  int                      ystride;
+  int                      nhfrags;
+  fplane=_state->fplanes+_pli;
+  nhfrags=fplane->nhfrags;
+  fragi_top=fplane->froffset;
+  fragi_bot=fragi_top+fplane->nfrags;
+  fragi0=fragi_top+_fragy0*(ptrdiff_t)nhfrags;
+  fragi0_end=fragi_top+_fragy_end*(ptrdiff_t)nhfrags;
+  ystride=_state->ref_ystride[_pli];
+  frags=_state->frags;
+  frag_buf_offs=_state->frag_buf_offs;
+  ref_frame_data=_state->ref_frame_data[_refi];
+  /*The following loops are constructed somewhat non-intuitively on purpose.
+    The main idea is: if a block boundary has at least one coded fragment on
+     it, the filter is applied to it.
+    However, the order that the filters are applied in matters, and VP3 chose
+     the somewhat strange ordering used below.*/
+  while(fragi0<fragi0_end){
+    ptrdiff_t fragi;
+    ptrdiff_t fragi_end;
+    fragi=fragi0;
+    fragi_end=fragi+nhfrags;
+    while(fragi<fragi_end){
+      if(frags[fragi].coded){
+        unsigned char *ref;
+        ref=ref_frame_data+frag_buf_offs[fragi];
+#define PIX eax
+#define YSTRIDE3 edi
+#define YSTRIDE ecx
+#define LL edx
+#define D esi
+#define D_WORD si
+        if(fragi>fragi0)OC_LOOP_FILTER_H_MMX(ref,ystride,_bv);
+        if(fragi0>fragi_top)OC_LOOP_FILTER_V_MMX(ref,ystride,_bv);
+        if(fragi+1<fragi_end&&!frags[fragi+1].coded){
+          OC_LOOP_FILTER_H_MMX(ref+8,ystride,_bv);
+        }
+        if(fragi+nhfrags<fragi_bot&&!frags[fragi+nhfrags].coded){
+          OC_LOOP_FILTER_V_MMX(ref+(ystride*8),ystride,_bv);
+        }
+#undef PIX
+#undef YSTRIDE3
+#undef YSTRIDE
+#undef LL
+#undef D
+#undef D_WORD
+      }
+      fragi++;
+    }
+    fragi0+=nhfrags;
+  }
+}
+
+#endif

thirdparty/libtheora/x86_vc/x86cpu.c

@@ -0,0 +1,192 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+ CPU capability detection for x86 processors.
+  Originally written by Rudolf Marek.
+
+ function:
+
+ ********************************************************************/
+
+#include "x86cpu.h"
+
+#if !defined(OC_X86_ASM)
+ogg_uint32_t oc_cpu_flags_get(void){
+  return 0;
+}
+#else
+/*Why does MSVC need this complicated rigamarole?
+  At this point I honestly do not care.*/
+
+/*Visual C cpuid helper function.
+  For VS2005 we could as well use the _cpuid builtin, but that wouldn't work
+   for VS2003 users, so we do it in inline assembler.*/
+static void oc_cpuid_helper(ogg_uint32_t _cpu_info[4],ogg_uint32_t _op){
+  _asm{
+    mov eax,[_op]
+    mov esi,_cpu_info
+    cpuid
+    mov [esi+0],eax
+    mov [esi+4],ebx
+    mov [esi+8],ecx
+    mov [esi+12],edx
+  }
+}
+
+#  define cpuid(_op,_eax,_ebx,_ecx,_edx) \
+  do{ \
+    ogg_uint32_t cpu_info[4]; \
+    oc_cpuid_helper(cpu_info,_op); \
+    (_eax)=cpu_info[0]; \
+    (_ebx)=cpu_info[1]; \
+    (_ecx)=cpu_info[2]; \
+    (_edx)=cpu_info[3]; \
+  }while(0)
+
+static void oc_detect_cpuid_helper(ogg_uint32_t *_eax,ogg_uint32_t *_ebx){
+  _asm{
+    pushfd
+    pushfd
+    pop eax
+    mov ebx,eax
+    xor eax,200000h
+    push eax
+    popfd
+    pushfd
+    pop eax
+    popfd
+    mov ecx,_eax
+    mov [ecx],eax
+    mov ecx,_ebx
+    mov [ecx],ebx
+  }
+}
+
+static ogg_uint32_t oc_parse_intel_flags(ogg_uint32_t _edx,ogg_uint32_t _ecx){
+  ogg_uint32_t flags;
+  /*If there isn't even MMX, give up.*/
+  if(!(_edx&0x00800000))return 0;
+  flags=OC_CPU_X86_MMX;
+  if(_edx&0x02000000)flags|=OC_CPU_X86_MMXEXT|OC_CPU_X86_SSE;
+  if(_edx&0x04000000)flags|=OC_CPU_X86_SSE2;
+  if(_ecx&0x00000001)flags|=OC_CPU_X86_PNI;
+  if(_ecx&0x00000100)flags|=OC_CPU_X86_SSSE3;
+  if(_ecx&0x00080000)flags|=OC_CPU_X86_SSE4_1;
+  if(_ecx&0x00100000)flags|=OC_CPU_X86_SSE4_2;
+  return flags;
+}
+
+static ogg_uint32_t oc_parse_amd_flags(ogg_uint32_t _edx,ogg_uint32_t _ecx){
+  ogg_uint32_t flags;
+  /*If there isn't even MMX, give up.*/
+  if(!(_edx&0x00800000))return 0;
+  flags=OC_CPU_X86_MMX;
+  if(_edx&0x00400000)flags|=OC_CPU_X86_MMXEXT;
+  if(_edx&0x80000000)flags|=OC_CPU_X86_3DNOW;
+  if(_edx&0x40000000)flags|=OC_CPU_X86_3DNOWEXT;
+  if(_ecx&0x00000040)flags|=OC_CPU_X86_SSE4A;
+  if(_ecx&0x00000800)flags|=OC_CPU_X86_SSE5;
+  return flags;
+}
+
+ogg_uint32_t oc_cpu_flags_get(void){
+  ogg_uint32_t flags;
+  ogg_uint32_t eax;
+  ogg_uint32_t ebx;
+  ogg_uint32_t ecx;
+  ogg_uint32_t edx;
+# if !defined(__amd64__)&&!defined(__x86_64__)
+  /*Not all x86-32 chips support cpuid, so we have to check.*/
+  oc_detect_cpuid_helper(&eax,&ebx);
+  /*No cpuid.*/
+  if(eax==ebx)return 0;
+# endif
+  cpuid(0,eax,ebx,ecx,edx);
+  /*         l e t n          I e n i          u n e G*/
+  if(ecx==0x6C65746E&&edx==0x49656E69&&ebx==0x756E6547||
+   /*      6 8 x M          T e n i          u n e G*/
+   ecx==0x3638784D&&edx==0x54656E69&&ebx==0x756E6547){
+    int family;
+    int model;
+    /*Intel, Transmeta (tested with Crusoe TM5800):*/
+    cpuid(1,eax,ebx,ecx,edx);
+    flags=oc_parse_intel_flags(edx,ecx);
+    family=(eax>>8)&0xF;
+    model=(eax>>4)&0xF;
+    /*The SSE unit on the Pentium M and Core Duo is much slower than the MMX
+       unit, so don't use it.*/
+    if(family==6&&(model==9||model==13||model==14)){
+      flags&=~(OC_CPU_X86_SSE2|OC_CPU_X86_PNI);
+    }
+  }
+  /*              D M A c          i t n e          h t u A*/
+  else if(ecx==0x444D4163&&edx==0x69746E65&&ebx==0x68747541||
+   /*      C S N            y b   e          d o e G*/
+   ecx==0x43534e20&&edx==0x79622065&&ebx==0x646f6547){
+    /*AMD, Geode:*/
+    cpuid(0x80000000,eax,ebx,ecx,edx);
+    if(eax<0x80000001)flags=0;
+    else{
+      cpuid(0x80000001,eax,ebx,ecx,edx);
+      flags=oc_parse_amd_flags(edx,ecx);
+    }
+    /*Also check for SSE.*/
+    cpuid(1,eax,ebx,ecx,edx);
+    flags|=oc_parse_intel_flags(edx,ecx);
+  }
+  /*Technically some VIA chips can be configured in the BIOS to return any
+     string here the user wants.
+    There is a special detection method that can be used to identify such
+     processors, but in my opinion, if the user really wants to change it, they
+     deserve what they get.*/
+  /*              s l u a          H r u a          t n e C*/
+  else if(ecx==0x736C7561&&edx==0x48727561&&ebx==0x746E6543){
+    /*VIA:*/
+    /*I only have documentation for the C7 (Esther) and Isaiah (forthcoming)
+       chips (thanks to the engineers from Centaur Technology who provided it).
+      These chips support Intel-like cpuid info.
+      The C3-2 (Nehemiah) cores appear to, as well.*/
+    cpuid(1,eax,ebx,ecx,edx);
+    flags=oc_parse_intel_flags(edx,ecx);
+    if(eax>=0x80000001){
+      /*The (non-Nehemiah) C3 processors support AMD-like cpuid info.
+        We need to check this even if the Intel test succeeds to pick up 3DNow!
+         support on these processors.
+        Unlike actual AMD processors, we cannot _rely_ on this info, since
+         some cores (e.g., the 693 stepping of the Nehemiah) claim to support
+         this function, yet return edx=0, despite the Intel test indicating
+         MMX support.
+        Therefore the features detected here are strictly added to those
+         detected by the Intel test.*/
+      /*TODO: How about earlier chips?*/
+      cpuid(0x80000001,eax,ebx,ecx,edx);
+      /*Note: As of the C7, this function returns Intel-style extended feature
+         flags, not AMD-style.
+        Currently, this only defines bits 11, 20, and 29 (0x20100800), which
+         do not conflict with any of the AMD flags we inspect.
+        For the remaining bits, Intel tells us, "Do not count on their value",
+         but VIA assures us that they will all be zero (at least on the C7 and
+         Isaiah chips).
+        In the (unlikely) event a future processor uses bits 18, 19, 30, or 31
+         (0xC0C00000) for something else, we will have to add code to detect
+         the model to decide when it is appropriate to inspect them.*/
+      flags|=oc_parse_amd_flags(edx,ecx);
+    }
+  }
+  else{
+    /*Implement me.*/
+    flags=0;
+  }
+  return flags;
+}
+#endif

thirdparty/libtheora/x86_vc/x86cpu.h

@@ -0,0 +1,36 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+ function:
+
+ ********************************************************************/
+
+#if !defined(_x86_vc_x86cpu_H)
+# define _x86_vc_x86cpu_H (1)
+#include "../internal.h"
+
+#define OC_CPU_X86_MMX      (1<<0)
+#define OC_CPU_X86_3DNOW    (1<<1)
+#define OC_CPU_X86_3DNOWEXT (1<<2)
+#define OC_CPU_X86_MMXEXT   (1<<3)
+#define OC_CPU_X86_SSE      (1<<4)
+#define OC_CPU_X86_SSE2     (1<<5)
+#define OC_CPU_X86_PNI      (1<<6)
+#define OC_CPU_X86_SSSE3    (1<<7)
+#define OC_CPU_X86_SSE4_1   (1<<8)
+#define OC_CPU_X86_SSE4_2   (1<<9)
+#define OC_CPU_X86_SSE4A    (1<<10)
+#define OC_CPU_X86_SSE5     (1<<11)
+
+ogg_uint32_t oc_cpu_flags_get(void);
+
+#endif

thirdparty/libtheora/x86_vc/x86enc.c

@@ -0,0 +1,47 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+#include "x86enc.h"
+
+#if defined(OC_X86_ASM)
+
+void oc_enc_accel_init_x86(oc_enc_ctx *_enc){
+  ogg_uint32_t cpu_flags;
+  cpu_flags=_enc->state.cpu_flags;
+  oc_enc_accel_init_c(_enc);
+  if(cpu_flags&OC_CPU_X86_MMX){
+    _enc->opt_vtable.frag_sub=oc_enc_frag_sub_mmx;
+    _enc->opt_vtable.frag_sub_128=oc_enc_frag_sub_128_mmx;
+    _enc->opt_vtable.frag_recon_intra=oc_frag_recon_intra_mmx;
+    _enc->opt_vtable.frag_recon_inter=oc_frag_recon_inter_mmx;
+  }
+  if(cpu_flags&OC_CPU_X86_MMXEXT){
+    _enc->opt_vtable.frag_sad=oc_enc_frag_sad_mmxext;
+    _enc->opt_vtable.frag_sad_thresh=oc_enc_frag_sad_thresh_mmxext;
+    _enc->opt_vtable.frag_sad2_thresh=oc_enc_frag_sad2_thresh_mmxext;
+    _enc->opt_vtable.frag_satd=oc_enc_frag_satd_mmxext;
+    _enc->opt_vtable.frag_satd2=oc_enc_frag_satd2_mmxext;
+    _enc->opt_vtable.frag_intra_satd=oc_enc_frag_intra_satd_mmxext;
+    _enc->opt_vtable.frag_copy2=oc_enc_frag_copy2_mmxext;
+    _enc->opt_vtable.fdct8x8=oc_enc_fdct8x8_mmxext;
+  }
+  if(cpu_flags&OC_CPU_X86_SSE2){
+# if defined(OC_X86_64_ASM)
+    _enc->opt_vtable.fdct8x8=oc_enc_fdct8x8_x86_64sse2;
+# endif
+  }
+}
+#endif

thirdparty/libtheora/x86_vc/x86enc.h

@@ -0,0 +1,51 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+#if !defined(_x86_vc_x86enc_H)
+# define _x86_vc_x86enc_H (1)
+# include "x86int.h"
+# if defined(OC_X86_ASM)
+#  define oc_enc_accel_init oc_enc_accel_init_x86
+#  define OC_ENC_USE_VTABLE (1)
+# endif
+# include "../encint.h"
+
+void oc_enc_accel_init_x86(oc_enc_ctx *_enc);
+
+unsigned oc_enc_frag_sad_mmxext(const unsigned char *_src,
+ const unsigned char *_ref,int _ystride);
+unsigned oc_enc_frag_sad_thresh_mmxext(const unsigned char *_src,
+ const unsigned char *_ref,int _ystride,unsigned _thresh);
+unsigned oc_enc_frag_sad2_thresh_mmxext(const unsigned char *_src,
+ const unsigned char *_ref1,const unsigned char *_ref2,int _ystride,
+ unsigned _thresh);
+unsigned oc_enc_frag_satd_mmxext(unsigned *_dc,const unsigned char *_src,
+ const unsigned char *_ref,int _ystride);
+unsigned oc_enc_frag_satd2_mmxext(unsigned *_dc,const unsigned char *_src,
+ const unsigned char *_ref1,const unsigned char *_ref2,int _ystride);
+unsigned oc_enc_frag_intra_satd_mmxext(unsigned *_dc,
+ const unsigned char *_src,int _ystride);
+void oc_enc_frag_sub_mmx(ogg_int16_t _diff[64],
+ const unsigned char *_x,const unsigned char *_y,int _stride);
+void oc_enc_frag_sub_128_mmx(ogg_int16_t _diff[64],
+ const unsigned char *_x,int _stride);
+void oc_enc_frag_copy2_mmxext(unsigned char *_dst,
+ const unsigned char *_src1,const unsigned char *_src2,int _ystride);
+void oc_enc_fdct8x8_mmxext(ogg_int16_t _y[64],const ogg_int16_t _x[64]);
+void oc_enc_fdct8x8_x86_64sse2(ogg_int16_t _y[64],const ogg_int16_t _x[64]);
+
+#endif

thirdparty/libtheora/x86_vc/x86int.h

@@ -0,0 +1,49 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+#if !defined(_x86_vc_x86int_H)
+# define _x86_vc_x86int_H (1)
+# include "../internal.h"
+# if defined(OC_X86_ASM)
+#  define oc_state_accel_init oc_state_accel_init_x86
+#  define OC_STATE_USE_VTABLE (1)
+# endif
+# include "../state.h"
+# include "x86cpu.h"
+
+void oc_state_accel_init_x86(oc_theora_state *_state);
+
+void oc_frag_copy_mmx(unsigned char *_dst,
+ const unsigned char *_src,int _ystride);
+void oc_frag_copy_list_mmx(unsigned char *_dst_frame,
+ const unsigned char *_src_frame,int _ystride,
+ const ptrdiff_t *_fragis,ptrdiff_t _nfragis,const ptrdiff_t *_frag_buf_offs);
+void oc_frag_recon_intra_mmx(unsigned char *_dst,int _ystride,
+ const ogg_int16_t *_residue);
+void oc_frag_recon_inter_mmx(unsigned char *_dst,
+ const unsigned char *_src,int _ystride,const ogg_int16_t *_residue);
+void oc_frag_recon_inter2_mmx(unsigned char *_dst,const unsigned char *_src1,
+ const unsigned char *_src2,int _ystride,const ogg_int16_t *_residue);
+void oc_idct8x8_mmx(ogg_int16_t _y[64],ogg_int16_t _x[64],int _last_zzi);
+void oc_state_frag_recon_mmx(const oc_theora_state *_state,ptrdiff_t _fragi,
+ int _pli,ogg_int16_t _dct_coeffs[128],int _last_zzi,ogg_uint16_t _dc_quant);
+void oc_loop_filter_init_mmx(signed char _bv[256],int _flimit);
+void oc_state_loop_filter_frag_rows_mmx(const oc_theora_state *_state,
+ signed char _bv[256],int _refi,int _pli,int _fragy0,int _fragy_end);
+void oc_restore_fpu_mmx(void);
+
+#endif

thirdparty/libtheora/x86_vc/x86state.c

@@ -0,0 +1,61 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+#include "x86int.h"
+
+#if defined(OC_X86_ASM)
+
+/*This table has been modified from OC_FZIG_ZAG by baking a 4x4 transpose into
+   each quadrant of the destination.*/
+static const unsigned char OC_FZIG_ZAG_MMX[128]={
+   0, 8, 1, 2, 9,16,24,17,
+  10, 3,32,11,18,25, 4,12,
+   5,26,19,40,33,34,41,48,
+  27, 6,13,20,28,21,14, 7,
+  56,49,42,35,43,50,57,36,
+  15,22,29,30,23,44,37,58,
+  51,59,38,45,52,31,60,53,
+  46,39,47,54,61,62,55,63,
+  64,64,64,64,64,64,64,64,
+  64,64,64,64,64,64,64,64,
+  64,64,64,64,64,64,64,64,
+  64,64,64,64,64,64,64,64,
+  64,64,64,64,64,64,64,64,
+  64,64,64,64,64,64,64,64,
+  64,64,64,64,64,64,64,64,
+  64,64,64,64,64,64,64,64,
+};
+
+void oc_state_accel_init_x86(oc_theora_state *_state){
+  _state->cpu_flags=oc_cpu_flags_get();
+  if(_state->cpu_flags&OC_CPU_X86_MMX){
+    _state->opt_vtable.frag_copy=oc_frag_copy_mmx;
+    _state->opt_vtable.frag_copy_list=oc_frag_copy_list_mmx;
+    _state->opt_vtable.frag_recon_intra=oc_frag_recon_intra_mmx;
+    _state->opt_vtable.frag_recon_inter=oc_frag_recon_inter_mmx;
+    _state->opt_vtable.frag_recon_inter2=oc_frag_recon_inter2_mmx;
+    _state->opt_vtable.idct8x8=oc_idct8x8_mmx;
+    _state->opt_vtable.state_frag_recon=oc_state_frag_recon_mmx;
+    _state->opt_vtable.loop_filter_init=oc_loop_filter_init_mmx;
+    _state->opt_vtable.state_loop_filter_frag_rows=
+     oc_state_loop_filter_frag_rows_mmx;
+    _state->opt_vtable.restore_fpu=oc_restore_fpu_mmx;
+    _state->opt_data.dct_fzig_zag=OC_FZIG_ZAG_MMX;
+  }
+  else oc_state_accel_init_c(_state);
+}
+#endif

thirdparty/libtheora/x86_vc/x86zigzag.h

@@ -0,0 +1,244 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009                *
+ * by the Xiph.Org Foundation and contributors                      *
+ * https://www.xiph.org/                                            *
+ *                                                                  *
+ ********************************************************************
+
+  function:
+
+ ********************************************************************/
+
+#if !defined(_x86_vc_x86zigzag_H)
+# define _x86_vc_x86zigzag_H (1)
+# include "x86enc.h"
+
+
+/*Converts DCT coefficients from transposed order into zig-zag scan order and
+   stores them in Y.
+  This relies on two macros to load the contents of each row:
+   OC_ZZ_LOAD_ROW_LO(row,reg) and OC_ZZ_LOAD_ROW_HI(row,reg), which load the
+   first four and second four entries of each row into the specified register,
+   respectively.
+  OC_ZZ_LOAD_ROW_LO must be called before OC_ZZ_LOAD_ROW_HI for the same row
+   (because when the rows are already in SSE2 registers, loading the high half
+   destructively modifies the register).
+  The index of each output element in the original 64-element array should wind
+   up in the following 8x8 matrix (the letters indicate the order we compute
+   each 4-tuple below):
+    A  0  8  1  2   9 16 24 17 B
+    C 10  3  4 11  18 25 32 40 E
+    F 33 26 19 12   5  6 13 20 D
+    G 27 34 41 48  56 49 42 35 I
+    L 28 21 14  7  15 22 29 36 M
+    H 43 50 57 58  51 44 37 30 O
+    N 23 31 38 45  52 59 60 53 J
+    P 46 39 47 54  61 62 55 63 K
+  The order of the coefficients within each tuple is reversed in the comments
+   below to reflect the usual MSB to LSB notation.*/
+#define OC_TRANSPOSE_ZIG_ZAG_MMXEXT \
+  OC_ZZ_LOAD_ROW_LO(0,mm0)  /*mm0=03 02 01 00*/ \
+  OC_ZZ_LOAD_ROW_LO(1,mm1)  /*mm1=11 10 09 08*/ \
+  OC_ZZ_LOAD_ROW_LO(2,mm2)  /*mm2=19 18 17 16*/ \
+  OC_ZZ_LOAD_ROW_LO(3,mm3)  /*mm3=27 26 25 24*/ \
+  OC_ZZ_LOAD_ROW_HI(0,mm4)  /*mm4=07 06 05 04*/ \
+  OC_ZZ_LOAD_ROW_HI(1,mm5)  /*mm5=15 14 13 12*/ \
+  OC_ZZ_LOAD_ROW_HI(2,mm6)  /*mm6=23 22 21 20*/ \
+  __asm movq mm7,mm0        /*mm7=03 02 01 00*/ \
+  __asm punpckhdq mm0,mm1   /*mm0=11 10 03 02*/ \
+  __asm pshufw mm4,mm4,0x39 /*mm4=04 07 06 05*/ \
+  __asm punpcklwd mm1,mm0   /*mm1=03 09 02 08*/ \
+  __asm pshufw mm5,mm5,0x39 /*mm5=12 15 14 13*/ \
+  __asm punpcklwd mm7,mm1   /*mm7=02 01 08 00 *A*/ \
+  __asm movq [Y+0x00],mm7 \
+  __asm punpckhwd mm1,mm4   /*mm1=04 03 07 09*/ \
+  __asm movq mm7,mm2        /*mm7=19 18 17 16*/ \
+  __asm punpckhdq mm0,mm1   /*mm0=04 03 11 10*/ \
+  __asm punpckhwd mm7,mm5   /*mm7=12 19 15 18*/ \
+  __asm punpcklwd mm1,mm3   /*mm1=25 07 24 09*/ \
+  __asm punpcklwd mm5,mm6   /*mm5=21 14 20 13*/ \
+  __asm punpcklwd mm1,mm2   /*mm1=17 24 16 09 *B*/ \
+  OC_ZZ_LOAD_ROW_LO(4,mm2)  /*mm2=35 34 33 32*/ \
+  __asm movq [Y+0x08],mm1 \
+  OC_ZZ_LOAD_ROW_LO(5,mm1)  /*mm1=43 42 41 40*/ \
+  __asm pshufw mm0,mm0,0x78 /*mm0=11 04 03 10 *C*/ \
+  __asm movq [Y+0x10],mm0 \
+  __asm punpckhdq mm6,mm4   /*mm6=?? 07 23 22*/ \
+  __asm punpckldq mm4,mm5   /*mm4=20 13 06 05 *D*/ \
+  __asm movq [Y+0x28],mm4 \
+  __asm psrlq mm3,16        /*mm3=.. 27 26 25*/ \
+  __asm pshufw mm0,mm2,0x0E /*mm0=?? ?? 35 34*/ \
+  __asm movq mm4,mm7        /*mm4=12 19 15 18*/ \
+  __asm punpcklwd mm2,mm3   /*mm2=26 33 25 32*/ \
+  __asm punpcklwd mm4,mm1   /*mm4=41 15 40 18*/ \
+  __asm punpckhwd mm3,mm1   /*mm3=43 .. 42 27*/ \
+  __asm punpckldq mm4,mm2   /*mm4=25 32 40 18*/ \
+  __asm punpcklwd mm3,mm0   /*mm3=35 42 34 27*/ \
+  OC_ZZ_LOAD_ROW_LO(6,mm0)  /*mm0=51 50 49 48*/ \
+  __asm pshufw mm4,mm4,0x6C /*mm4=40 32 25 18 *E*/ \
+  __asm movq [Y+0x18],mm4 \
+  OC_ZZ_LOAD_ROW_LO(7,mm4)  /*mm4=59 58 57 56*/ \
+  __asm punpckhdq mm2,mm7   /*mm2=12 19 26 33 *F*/ \
+  __asm movq [Y+0x20],mm2 \
+  __asm pshufw mm1,mm1,0xD0 /*mm1=43 41 ?? ??*/ \
+  __asm pshufw mm0,mm0,0x87 /*mm0=50 48 49 51*/ \
+  __asm movq mm2,mm3        /*mm2=35 42 34 27*/ \
+  __asm punpckhwd mm1,mm0   /*mm1=50 43 48 41*/ \
+  __asm pshufw mm4,mm4,0x93 /*mm4=58 57 56 59*/ \
+  __asm punpckldq mm3,mm1   /*mm3=48 41 34 27 *G*/ \
+  __asm movq [Y+0x30],mm3 \
+  __asm punpckhdq mm1,mm4   /*mm1=58 57 50 43 *H*/ \
+  __asm movq [Y+0x50],mm1 \
+  OC_ZZ_LOAD_ROW_HI(7,mm1)  /*mm1=63 62 61 60*/ \
+  __asm punpcklwd mm4,mm0   /*mm4=49 56 51 59*/ \
+  OC_ZZ_LOAD_ROW_HI(6,mm0)  /*mm0=55 54 53 52*/ \
+  __asm psllq mm6,16        /*mm6=07 23 22 ..*/ \
+  __asm movq mm3,mm4        /*mm3=49 56 51 59*/ \
+  __asm punpckhdq mm4,mm2   /*mm4=35 42 49 56 *I*/ \
+  OC_ZZ_LOAD_ROW_HI(3,mm2)  /*mm2=31 30 29 28*/ \
+  __asm movq [Y+0x38],mm4 \
+  __asm punpcklwd mm3,mm1   /*mm3=61 51 60 59*/ \
+  __asm punpcklwd mm7,mm6   /*mm7=22 15 .. ??*/ \
+  __asm movq mm4,mm3        /*mm4=61 51 60 59*/ \
+  __asm punpcklwd mm3,mm0   /*mm3=53 60 52 59*/ \
+  __asm punpckhwd mm4,mm0   /*mm4=55 61 54 51*/ \
+  OC_ZZ_LOAD_ROW_HI(4,mm0)  /*mm0=39 38 37 36*/ \
+  __asm pshufw mm3,mm3,0xE1 /*mm3=53 60 59 52 *J*/ \
+  __asm movq [Y+0x68],mm3 \
+  __asm movq mm3,mm4        /*mm3=?? ?? 54 51*/ \
+  __asm pshufw mm2,mm2,0x39 /*mm2=28 31 30 29*/ \
+  __asm punpckhwd mm4,mm1   /*mm4=63 55 62 61 *K*/ \
+  OC_ZZ_LOAD_ROW_HI(5,mm1)  /*mm1=47 46 45 44*/ \
+  __asm movq [Y+0x78],mm4 \
+  __asm punpckhwd mm6,mm2   /*mm6=28 07 31 23*/ \
+  __asm punpcklwd mm2,mm0   /*mm2=37 30 36 29*/ \
+  __asm punpckhdq mm5,mm6   /*mm5=28 07 21 14*/ \
+  __asm pshufw mm2,mm2,0x4B /*mm2=36 29 30 37*/ \
+  __asm pshufw mm5,mm5,0x87 /*mm5=07 14 21 28 *L*/ \
+  __asm movq [Y+0x40],mm5 \
+  __asm punpckhdq mm7,mm2   /*mm7=36 29 22 15 *M*/ \
+  __asm movq [Y+0x48],mm7 \
+  __asm pshufw mm1,mm1,0x9C /*mm1=46 45 47 44*/ \
+  __asm punpckhwd mm0,mm1   /*mm0=46 39 45 38*/ \
+  __asm punpcklwd mm3,mm1   /*mm3=47 54 44 51*/ \
+  __asm punpckldq mm6,mm0   /*mm6=45 38 31 23 *N*/ \
+  __asm movq [Y+0x60],mm6 \
+  __asm punpckhdq mm0,mm3   /*mm0=47 54 46 39*/ \
+  __asm punpckldq mm3,mm2   /*mm3=30 37 44 51 *O*/ \
+  __asm movq [Y+0x58],mm3 \
+  __asm pshufw mm0,mm0,0xB1 /*mm0=54 47 39 46 *P*/ \
+  __asm movq [Y+0x70],mm0 \
+
+/*Converts DCT coefficients in %[dct] from natural order into zig-zag scan
+   order and stores them in %[qdct].
+  The index of each output element in the original 64-element array should wind
+   up in the following 8x8 matrix (the letters indicate the order we compute
+   each 4-tuple below):
+    A  0  1  8 16   9  2  3 10 B
+    C 17 24 32 25  18 11  4  5 D
+    E 12 19 26 33  40 48 41 34 I
+    H 27 20 13  6   7 14 21 28 G
+    K 35 42 49 56  57 50 43 36 J
+    F 29 22 15 23  30 37 44 51 M
+    P 58 59 52 45  38 31 39 46 L
+    N 53 60 61 54  47 55 62 63 O
+  The order of the coefficients within each tuple is reversed in the comments
+   below to reflect the usual MSB to LSB notation.*/
+#define OC_ZIG_ZAG_MMXEXT \
+  "movq 0x00(%[dct]),%%mm0\n\t"  /*mm0=03 02 01 00*/ \
+  "movq 0x08(%[dct]),%%mm1\n\t"  /*mm1=07 06 05 04*/ \
+  "movq 0x10(%[dct]),%%mm2\n\t"  /*mm2=11 10 09 08*/ \
+  "movq 0x20(%[dct]),%%mm3\n\t"  /*mm3=19 18 17 16*/ \
+  "movq 0x30(%[dct]),%%mm4\n\t"  /*mm4=27 26 25 24*/ \
+  "movq 0x40(%[dct]),%%mm5\n\t"  /*mm5=35 34 33 32*/ \
+  "movq %%mm2,%%mm7\n\t"         /*mm7=11 10 09 08*/ \
+  "punpcklwd %%mm3,%%mm2\n\t"    /*mm2=17 09 16 08*/ \
+  "movq %%mm0,%%mm6\n\t"         /*mm6=03 02 01 00*/ \
+  "punpckldq %%mm2,%%mm0\n\t"    /*mm0=16 08 01 00 *A*/ \
+  "movq %%mm0,0x00(%[qdct])\n\t" \
+  "movq 0x18(%[dct]),%%mm0\n\t"  /*mm0=15 14 13 12*/ \
+  "punpckhdq %%mm6,%%mm6\n\t"    /*mm6=03 02 03 02*/ \
+  "psrlq $16,%%mm7\n\t"          /*mm7=.. 11 10 09*/ \
+  "punpckldq %%mm7,%%mm6\n\t"    /*mm6=10 09 03 02*/ \
+  "punpckhwd %%mm7,%%mm3\n\t"    /*mm3=.. 19 11 18*/ \
+  "pshufw $0xD2,%%mm6,%%mm6\n\t" /*mm6=10 03 02 09 *B*/ \
+  "movq %%mm6,0x08(%[qdct])\n\t" \
+  "psrlq $48,%%mm2\n\t"          /*mm2=.. .. .. 17*/ \
+  "movq %%mm1,%%mm6\n\t"         /*mm6=07 06 05 04*/ \
+  "punpcklwd %%mm5,%%mm2\n\t"    /*mm2=33 .. 32 17*/ \
+  "movq %%mm3,%%mm7\n\t"         /*mm7=.. 19 11 18*/ \
+  "punpckldq %%mm1,%%mm3\n\t"    /*mm3=05 04 11 18 *C*/ \
+  "por %%mm2,%%mm7\n\t"          /*mm7=33 19 ?? ??*/ \
+  "punpcklwd %%mm4,%%mm2\n\t"    /*mm2=25 32 24 17 *D**/ \
+  "movq %%mm2,0x10(%[qdct])\n\t" \
+  "movq %%mm3,0x18(%[qdct])\n\t" \
+  "movq 0x28(%[dct]),%%mm2\n\t"  /*mm2=23 22 21 20*/ \
+  "movq 0x38(%[dct]),%%mm1\n\t"  /*mm1=31 30 29 28*/ \
+  "pshufw $0x9C,%%mm0,%%mm3\n\t" /*mm3=14 13 15 12*/ \
+  "punpckhdq %%mm7,%%mm7\n\t"    /*mm7=33 19 33 19*/ \
+  "punpckhwd %%mm3,%%mm6\n\t"    /*mm6=14 07 13 06*/ \
+  "punpckldq %%mm0,%%mm0\n\t"    /*mm0=13 12 13 12*/ \
+  "punpcklwd %%mm1,%%mm3\n\t"    /*mm3=29 15 28 12*/ \
+  "punpckhwd %%mm4,%%mm0\n\t"    /*mm0=27 13 26 12*/ \
+  "pshufw $0xB4,%%mm3,%%mm3\n\t" /*mm3=15 29 28 12*/ \
+  "psrlq $48,%%mm4\n\t"          /*mm4=.. .. .. 27*/ \
+  "punpcklwd %%mm7,%%mm0\n\t"    /*mm0=33 26 19 12 *E*/ \
+  "punpcklwd %%mm1,%%mm4\n\t"    /*mm4=29 .. 28 27*/ \
+  "punpckhwd %%mm2,%%mm3\n\t"    /*mm3=23 15 22 29 *F*/ \
+  "movq %%mm0,0x20(%[qdct])\n\t" \
+  "movq %%mm3,0x50(%[qdct])\n\t" \
+  "movq 0x60(%[dct]),%%mm3\n\t"  /*mm3=51 50 49 48*/ \
+  "movq 0x70(%[dct]),%%mm7\n\t"  /*mm7=59 58 57 56*/ \
+  "movq 0x50(%[dct]),%%mm0\n\t"  /*mm0=43 42 41 40*/ \
+  "punpcklwd %%mm4,%%mm2\n\t"    /*mm2=28 21 27 20*/ \
+  "psrlq $32,%%mm5\n\t"          /*mm5=.. .. 35 34*/ \
+  "movq %%mm2,%%mm4\n\t"         /*mm4=28 21 27 20*/ \
+  "punpckldq %%mm6,%%mm2\n\t"    /*mm2=13 06 27 20*/ \
+  "punpckhdq %%mm4,%%mm6\n\t"    /*mm6=28 21 14 07 *G*/ \
+  "movq %%mm3,%%mm4\n\t"         /*mm4=51 50 49 48*/ \
+  "pshufw $0xB1,%%mm2,%%mm2\n\t" /*mm2=06 13 20 27 *H*/ \
+  "movq %%mm2,0x30(%[qdct])\n\t" \
+  "movq %%mm6,0x38(%[qdct])\n\t" \
+  "movq 0x48(%[dct]),%%mm2\n\t"  /*mm2=39 38 37 36*/ \
+  "punpcklwd %%mm5,%%mm4\n\t"    /*mm4=35 49 34 48*/ \
+  "movq 0x58(%[dct]),%%mm5\n\t"  /*mm5=47 46 45 44*/ \
+  "punpckldq %%mm7,%%mm6\n\t"    /*mm6=57 56 14 07*/ \
+  "psrlq $32,%%mm3\n\t"          /*mm3=.. .. 51 50*/ \
+  "punpckhwd %%mm0,%%mm6\n\t"    /*mm6=43 57 42 56*/ \
+  "punpcklwd %%mm4,%%mm0\n\t"    /*mm0=34 41 48 40 *I*/ \
+  "pshufw $0x4E,%%mm6,%%mm6\n\t" /*mm6=42 56 43 57*/ \
+  "movq %%mm0,0x28(%[qdct])\n\t" \
+  "punpcklwd %%mm2,%%mm3\n\t"    /*mm3=37 51 36 50*/ \
+  "punpckhwd %%mm6,%%mm4\n\t"    /*mm4=42 35 56 49*/ \
+  "punpcklwd %%mm3,%%mm6\n\t"    /*mm6=36 43 50 57 *J*/ \
+  "pshufw $0x4E,%%mm4,%%mm4\n\t" /*mm4=56 49 42 35 *K*/ \
+  "movq %%mm4,0x40(%[qdct])\n\t" \
+  "movq %%mm6,0x48(%[qdct])\n\t" \
+  "movq 0x68(%[dct]),%%mm6\n\t"  /*mm6=55 54 53 52*/ \
+  "movq 0x78(%[dct]),%%mm0\n\t"  /*mm0=63 62 61 60*/ \
+  "psrlq $32,%%mm1\n\t"          /*mm1=.. .. 31 30*/ \
+  "pshufw $0xD8,%%mm5,%%mm5\n\t" /*mm5=47 45 46 44*/ \
+  "pshufw $0x0B,%%mm3,%%mm3\n\t" /*mm3=50 50 51 37*/ \
+  "punpcklwd %%mm5,%%mm1\n\t"    /*mm1=46 31 44 30*/ \
+  "pshufw $0xC9,%%mm6,%%mm6\n\t" /*mm6=55 52 54 53*/ \
+  "punpckhwd %%mm1,%%mm2\n\t"    /*mm2=46 39 31 38 *L*/ \
+  "punpcklwd %%mm3,%%mm1\n\t"    /*mm1=51 44 37 30 *M*/ \
+  "movq %%mm2,0x68(%[qdct])\n\t" \
+  "movq %%mm1,0x58(%[qdct])\n\t" \
+  "punpckhwd %%mm6,%%mm5\n\t"    /*mm5=55 47 52 45*/ \
+  "punpckldq %%mm0,%%mm6\n\t"    /*mm6=61 60 54 53*/ \
+  "pshufw $0x10,%%mm5,%%mm4\n\t" /*mm4=45 52 45 45*/ \
+  "pshufw $0x78,%%mm6,%%mm6\n\t" /*mm6=53 60 61 54 *N*/ \
+  "punpckhdq %%mm0,%%mm5\n\t"    /*mm5=63 62 55 47 *O*/ \
+  "punpckhdq %%mm4,%%mm7\n\t"    /*mm7=45 52 59 58 *P*/ \
+  "movq %%mm6,0x70(%[qdct])\n\t" \
+  "movq %%mm5,0x78(%[qdct])\n\t" \
+  "movq %%mm7,0x60(%[qdct])\n\t" \
+
+#endif