noahbackus/godot
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 2 Wiki Activity

Overhaul compute shader based environment roughness calculation to improve performance and quality

Browse Source
This commit is contained in:
clayjohn
2026-01-12 15:18:32 -08:00
parent 2d0a589777
commit e33515c2e6
2 changed files with 53 additions and 31 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
servers/rendering/renderer_rd/effects/copy_effects.cpp
View File

@@ -1375,7 +1375,7 @@ void CopyEffects::octmap_roughness(RID p_source_rd_texture, RID p_dest_texture,
// Remap to perceptual-roughness^2 to create more detail in lower mips and match the mapping of octmap_filter.
roughness.push_constant.roughness = p_roughness * p_roughness;
roughness.push_constant.sample_count = p_sample_count;
roughness.push_constant.sample_count = MIN(p_sample_count, 64u);
roughness.push_constant.source_size = p_source_size;
roughness.push_constant.dest_size = p_dest_size;
roughness.push_constant.use_direct_write = p_roughness == 0.0;
@@ -1432,7 +1432,7 @@ void CopyEffects::octmap_roughness_raster(RID p_source_rd_texture, RID p_dest_fr
memset(&roughness.push_constant, 0, sizeof(OctmapRoughnessPushConstant));
roughness.push_constant.roughness = p_roughness * p_roughness; // Shader expects roughness, not perceptual roughness, so multiply before passing in.
roughness.push_constant.sample_count = p_sample_count;
roughness.push_constant.sample_count = MAX(uint32_t(float(p_sample_count * 4u) * roughness.push_constant.roughness), 4u);
roughness.push_constant.source_size = p_source_size;
roughness.push_constant.dest_size = p_dest_size;
roughness.push_constant.use_direct_write = p_roughness == 0.0;

80
servers/rendering/renderer_rd/shaders/effects/octmap_roughness.glsl
View File

@@ -8,6 +8,8 @@
layout(local_size_x = GROUP_SIZE, local_size_y = GROUP_SIZE, local_size_z = 1) in;
shared vec4 samples[GROUP_SIZE * GROUP_SIZE * 4]; // Up to 256 samples supported. Should never need more than that.
layout(set = 0, binding = 0) uniform sampler2D source_oct;
layout(OCTMAP_FORMAT, set = 1, binding = 0) uniform restrict writeonly image2D dest_octmap;
@@ -17,44 +19,64 @@ layout(OCTMAP_FORMAT, set = 1, binding = 0) uniform restrict writeonly image2D d
void main() {
uvec2 id = gl_GlobalInvocationID.xy;
if (id.x < params.dest_size && id.y < params.dest_size) {
vec2 inv_source_size = 1.0 / vec2(params.source_size);
vec2 inv_dest_size = 1.0 / vec2(params.dest_size);
vec2 uv = (vec2(id.xy) + 0.5) * inv_dest_size;
if (params.use_direct_write) {
vec2 inv_source_size = 1.0 / vec2(params.source_size);
vec2 inv_dest_size = 1.0 / vec2(params.dest_size);
vec2 uv = (vec2(id.xy) + 0.5) * inv_dest_size;
if (params.use_direct_write) {
if (id.x < params.dest_size && id.y < params.dest_size) {
imageStore(dest_octmap, ivec2(id), vec4(texture(source_oct, uv).rgb, 1.0));
} else {
}
} else {
float solid_angle_texel = 4.0 * M_PI / float(params.dest_size * params.dest_size);
float roughness2 = params.roughness * params.roughness;
float roughness4 = roughness2 * roughness2;
uint scaled_samples = max(uint(float(params.sample_count * 4) * params.roughness), 4);
// This effectively rounds the sample count up to the nearest (GROUP_SIZE * GROUP_SIZE).
uint samples_per_thread = max(1, ((scaled_samples) / (GROUP_SIZE * GROUP_SIZE)));
uint total_samples = samples_per_thread * (GROUP_SIZE * GROUP_SIZE);
for (uint local_sample = 0; local_sample < samples_per_thread; local_sample++) {
uint sample_idx = local_sample * (GROUP_SIZE * GROUP_SIZE) + gl_LocalInvocationIndex;
vec2 xi = Hammersley(sample_idx, total_samples);
vec3 H_local = ImportanceSampleGGX(xi, roughness4);
float NdotH = H_local.z;
vec3 L_local = 2.0 * NdotH * H_local - vec3(0.0, 0.0, 1.0);
float ndotl = L_local.z;
if (ndotl > 0.0) {
float D = DistributionGGX(NdotH, roughness4);
float pdf = D * NdotH / (4.0 * NdotH) + 0.0001;
float solid_angle_sample = 1.0 / (float(total_samples) * pdf + 0.0001);
float mipLevel = 0.5 * log2(solid_angle_sample / solid_angle_texel);
samples[sample_idx] = vec4(L_local, mipLevel);
} else {
samples[sample_idx] = vec4(-1.0);
}
}
memoryBarrierShared();
barrier();
if (id.x < params.dest_size && id.y < params.dest_size) {
vec3 N = oct_to_vec3_with_border(uv, params.border_size.y);
vec4 sum = vec4(0.0, 0.0, 0.0, 0.0);
float solid_angle_texel = 4.0 * M_PI / float(params.dest_size * params.dest_size);
float roughness2 = params.roughness * params.roughness;
float roughness4 = roughness2 * roughness2;
vec3 UpVector = abs(N.y) < 0.999 ? vec3(0.0, 1.0, 0.0) : vec3(0.0, 0.0, 1.0);
vec3 UpVector = abs(N.y) < 0.99999 ? vec3(0.0, 1.0, 0.0) : vec3(0.0, 0.0, 1.0);
mat3 T;
T[0] = normalize(cross(UpVector, N));
T[1] = cross(N, T[0]);
T[2] = N;
for (uint sampleNum = 0u; sampleNum < params.sample_count; sampleNum++) {
vec2 xi = Hammersley(sampleNum, params.sample_count);
vec3 H = T * ImportanceSampleGGX(xi, roughness4);
float NdotH = dot(N, H);
vec3 L = (2.0 * NdotH * H - N);
float ndotl = clamp(dot(N, L), 0.0, 1.0);
for (uint i = 0; i < total_samples; i++) {
vec4 s = samples[i];
float ndotl = s.z;
if (ndotl > 0.0) {
float D = DistributionGGX(NdotH, roughness4);
float pdf = D * NdotH / (4.0 * NdotH) + 0.0001;
float solid_angle_sample = 1.0 / (float(params.sample_count) * pdf + 0.0001);
float mipLevel = params.roughness == 0.0 ? 0.0 : 0.5 * log2(solid_angle_sample / solid_angle_texel);
vec2 sample_uv = vec3_to_oct_with_border(L, params.border_size);
sum.rgb += textureLod(source_oct, sample_uv, mipLevel).rgb * ndotl;
vec3 L_world = T * s.xyz;
vec2 sample_uv = vec3_to_oct_with_border(L_world, params.border_size);
sum.rgb += textureLod(source_oct, sample_uv, s.w).rgb * ndotl;
sum.a += ndotl;
}
}