Use custom HLSL cross-compiler

sergcpp · Jan 27, 2024 · 3e68ec2 · 3e68ec2
1 parent 4ec4af1
commit 3e68ec2
Show file tree

Hide file tree

Showing 89 changed files with 383 additions and 384 deletions.
diff --git a/first-party/glslx/win32/glslx.exe b/first-party/glslx/win32/glslx.exe
diff --git a/first-party/linux/glslx b/first-party/linux/glslx
diff --git a/first-party/macos/glslx b/first-party/macos/glslx
diff --git a/internal/shaders/intersect_scene.comp.glsl b/internal/shaders/intersect_scene.comp.glsl
@@ -303,13 +303,18 @@ void main() {
         const float q = can_terminate_path ? max(0.05, 1.0 - lum) : 0.0;
         if (p < q || lum == 0.0 || get_transp_depth(g_rays[index].depth) + 1 >= g_params.max_transp_depth) {
             // terminate ray
-            g_rays[index].c[0] = g_rays[index].c[1] = g_rays[index].c[2] = 0.0;
+            g_rays[index].c[0] = 0.0;
+            g_rays[index].c[1] = 0.0;
+            g_rays[index].c[2] = 0.0;
             break;
         }
 
-        g_rays[index].c[0] *= mat.base_color[0] / (1.0 - q);
-        g_rays[index].c[1] *= mat.base_color[1] / (1.0 - q);
-        g_rays[index].c[2] *= mat.base_color[2] / (1.0 - q);
+        float c[3] = {g_rays[index].c[0] * mat.base_color[0] / (1.0 - q),
+                      g_rays[index].c[1] * mat.base_color[1] / (1.0 - q),
+                      g_rays[index].c[2] * mat.base_color[2] / (1.0 - q)};
+        g_rays[index].c[0] = c[0];
+        g_rays[index].c[1] = c[1];
+        g_rays[index].c[2] = c[2];
 
         const float t = inter.t + HIT_BIAS;
         ro += rd * t;
@@ -318,7 +323,8 @@ void main() {
         inter.v = -1.0;
         inter.t = t_val - inter.t;
 
-        g_rays[index].depth += pack_ray_depth(0, 0, 0, 1);
+        uint depth = g_rays[index].depth + pack_ray_depth(0, 0, 0, 1);
+        g_rays[index].depth = depth;
         rand_dim += RAND_DIM_BOUNCE_COUNT;
     }
 

diff --git a/internal/shaders/output/convolution_112_112_fp16.comp.cso.inl b/internal/shaders/output/convolution_112_112_fp16.comp.cso.inl
diff --git a/internal/shaders/output/convolution_112_112_fp32.comp.cso.inl b/internal/shaders/output/convolution_112_112_fp32.comp.cso.inl
diff --git a/internal/shaders/output/convolution_32_32_Downsample_fp16.comp.cso.inl b/internal/shaders/output/convolution_32_32_Downsample_fp16.comp.cso.inl
diff --git a/internal/shaders/output/convolution_32_32_Downsample_fp32.comp.cso.inl b/internal/shaders/output/convolution_32_32_Downsample_fp32.comp.cso.inl
diff --git a/internal/shaders/output/convolution_32_3_img_fp16.comp.cso.inl b/internal/shaders/output/convolution_32_3_img_fp16.comp.cso.inl
diff --git a/internal/shaders/output/convolution_32_3_img_fp32.comp.cso.inl b/internal/shaders/output/convolution_32_3_img_fp32.comp.cso.inl
diff --git a/internal/shaders/output/convolution_32_48_Downsample_fp16.comp.cso.inl b/internal/shaders/output/convolution_32_48_Downsample_fp16.comp.cso.inl
diff --git a/internal/shaders/output/convolution_32_48_Downsample_fp32.comp.cso.inl b/internal/shaders/output/convolution_32_48_Downsample_fp32.comp.cso.inl
diff --git a/internal/shaders/output/convolution_48_64_Downsample_fp16.comp.cso.inl b/internal/shaders/output/convolution_48_64_Downsample_fp16.comp.cso.inl
diff --git a/internal/shaders/output/convolution_48_64_Downsample_fp32.comp.cso.inl b/internal/shaders/output/convolution_48_64_Downsample_fp32.comp.cso.inl
diff --git a/internal/shaders/output/convolution_64_32_fp16.comp.cso.inl b/internal/shaders/output/convolution_64_32_fp16.comp.cso.inl
diff --git a/internal/shaders/output/convolution_64_32_fp32.comp.cso.inl b/internal/shaders/output/convolution_64_32_fp32.comp.cso.inl
diff --git a/internal/shaders/output/convolution_64_64_fp16.comp.cso.inl b/internal/shaders/output/convolution_64_64_fp16.comp.cso.inl
diff --git a/internal/shaders/output/convolution_64_64_fp32.comp.cso.inl b/internal/shaders/output/convolution_64_64_fp32.comp.cso.inl
diff --git a/internal/shaders/output/convolution_64_80_Downsample_fp16.comp.cso.inl b/internal/shaders/output/convolution_64_80_Downsample_fp16.comp.cso.inl
diff --git a/internal/shaders/output/convolution_64_80_Downsample_fp32.comp.cso.inl b/internal/shaders/output/convolution_64_80_Downsample_fp32.comp.cso.inl
diff --git a/internal/shaders/output/convolution_80_96_fp16.comp.cso.inl b/internal/shaders/output/convolution_80_96_fp16.comp.cso.inl
diff --git a/internal/shaders/output/convolution_80_96_fp32.comp.cso.inl b/internal/shaders/output/convolution_80_96_fp32.comp.cso.inl
diff --git a/internal/shaders/output/convolution_96_96_fp16.comp.cso.inl b/internal/shaders/output/convolution_96_96_fp16.comp.cso.inl
diff --git a/internal/shaders/output/convolution_96_96_fp32.comp.cso.inl b/internal/shaders/output/convolution_96_96_fp32.comp.cso.inl
diff --git a/internal/shaders/output/convolution_Img_9_32_fp16.comp.cso.inl b/internal/shaders/output/convolution_Img_9_32_fp16.comp.cso.inl
diff --git a/internal/shaders/output/convolution_Img_9_32_fp32.comp.cso.inl b/internal/shaders/output/convolution_Img_9_32_fp32.comp.cso.inl
diff --git a/internal/shaders/output/convolution_concat_112_48_96_fp16.comp.cso.inl b/internal/shaders/output/convolution_concat_112_48_96_fp16.comp.cso.inl
diff --git a/internal/shaders/output/convolution_concat_112_48_96_fp32.comp.cso.inl b/internal/shaders/output/convolution_concat_112_48_96_fp32.comp.cso.inl
diff --git a/internal/shaders/output/convolution_concat_64_3_64_fp16.comp.cso.inl b/internal/shaders/output/convolution_concat_64_3_64_fp16.comp.cso.inl
diff --git a/internal/shaders/output/convolution_concat_64_3_64_fp32.comp.cso.inl b/internal/shaders/output/convolution_concat_64_3_64_fp32.comp.cso.inl
diff --git a/internal/shaders/output/convolution_concat_64_6_64_fp16.comp.cso.inl b/internal/shaders/output/convolution_concat_64_6_64_fp16.comp.cso.inl
diff --git a/internal/shaders/output/convolution_concat_64_6_64_fp32.comp.cso.inl b/internal/shaders/output/convolution_concat_64_6_64_fp32.comp.cso.inl
diff --git a/internal/shaders/output/convolution_concat_64_9_64_fp16.comp.cso.inl b/internal/shaders/output/convolution_concat_64_9_64_fp16.comp.cso.inl
diff --git a/internal/shaders/output/convolution_concat_64_9_64_fp32.comp.cso.inl b/internal/shaders/output/convolution_concat_64_9_64_fp32.comp.cso.inl
diff --git a/internal/shaders/output/convolution_concat_96_32_64_fp16.comp.cso.inl b/internal/shaders/output/convolution_concat_96_32_64_fp16.comp.cso.inl
diff --git a/internal/shaders/output/convolution_concat_96_32_64_fp32.comp.cso.inl b/internal/shaders/output/convolution_concat_96_32_64_fp32.comp.cso.inl
diff --git a/internal/shaders/output/convolution_concat_96_64_112_fp16.comp.cso.inl b/internal/shaders/output/convolution_concat_96_64_112_fp16.comp.cso.inl
diff --git a/internal/shaders/output/convolution_concat_96_64_112_fp32.comp.cso.inl b/internal/shaders/output/convolution_concat_96_64_112_fp32.comp.cso.inl
diff --git a/internal/shaders/output/debug_rt.comp.cso.inl b/internal/shaders/output/debug_rt.comp.cso.inl
diff --git a/internal/shaders/output/filter_variance.comp.cso.inl b/internal/shaders/output/filter_variance.comp.cso.inl
diff --git a/internal/shaders/output/intersect_area_lights.comp.cso.inl b/internal/shaders/output/intersect_area_lights.comp.cso.inl
diff --git a/internal/shaders/output/intersect_scene_hwrt_atlas.comp.cso.inl b/internal/shaders/output/intersect_scene_hwrt_atlas.comp.cso.inl
diff --git a/internal/shaders/output/intersect_scene_hwrt_atlas.comp.spv.inl b/internal/shaders/output/intersect_scene_hwrt_atlas.comp.spv.inl
diff --git a/internal/shaders/output/intersect_scene_hwrt_bindless.comp.cso.inl b/internal/shaders/output/intersect_scene_hwrt_bindless.comp.cso.inl
diff --git a/internal/shaders/output/intersect_scene_hwrt_bindless.comp.spv.inl b/internal/shaders/output/intersect_scene_hwrt_bindless.comp.spv.inl
diff --git a/internal/shaders/output/intersect_scene_indirect_hwrt_atlas.comp.cso.inl b/internal/shaders/output/intersect_scene_indirect_hwrt_atlas.comp.cso.inl
diff --git a/internal/shaders/output/intersect_scene_indirect_hwrt_atlas.comp.spv.inl b/internal/shaders/output/intersect_scene_indirect_hwrt_atlas.comp.spv.inl
diff --git a/internal/shaders/output/intersect_scene_indirect_hwrt_bindless.comp.cso.inl b/internal/shaders/output/intersect_scene_indirect_hwrt_bindless.comp.cso.inl
diff --git a/internal/shaders/output/intersect_scene_indirect_hwrt_bindless.comp.spv.inl b/internal/shaders/output/intersect_scene_indirect_hwrt_bindless.comp.spv.inl
diff --git a/internal/shaders/output/intersect_scene_indirect_swrt_atlas.comp.cso.inl b/internal/shaders/output/intersect_scene_indirect_swrt_atlas.comp.cso.inl
diff --git a/internal/shaders/output/intersect_scene_indirect_swrt_atlas.comp.spv.inl b/internal/shaders/output/intersect_scene_indirect_swrt_atlas.comp.spv.inl
diff --git a/internal/shaders/output/intersect_scene_indirect_swrt_bindless.comp.cso.inl b/internal/shaders/output/intersect_scene_indirect_swrt_bindless.comp.cso.inl
diff --git a/internal/shaders/output/intersect_scene_indirect_swrt_bindless.comp.spv.inl b/internal/shaders/output/intersect_scene_indirect_swrt_bindless.comp.spv.inl
diff --git a/internal/shaders/output/intersect_scene_shadow_hwrt_atlas.comp.cso.inl b/internal/shaders/output/intersect_scene_shadow_hwrt_atlas.comp.cso.inl
diff --git a/internal/shaders/output/intersect_scene_shadow_hwrt_bindless.comp.cso.inl b/internal/shaders/output/intersect_scene_shadow_hwrt_bindless.comp.cso.inl
diff --git a/internal/shaders/output/intersect_scene_shadow_swrt_atlas.comp.cso.inl b/internal/shaders/output/intersect_scene_shadow_swrt_atlas.comp.cso.inl
diff --git a/internal/shaders/output/intersect_scene_shadow_swrt_bindless.comp.cso.inl b/internal/shaders/output/intersect_scene_shadow_swrt_bindless.comp.cso.inl
diff --git a/internal/shaders/output/intersect_scene_swrt_atlas.comp.cso.inl b/internal/shaders/output/intersect_scene_swrt_atlas.comp.cso.inl
diff --git a/internal/shaders/output/intersect_scene_swrt_atlas.comp.spv.inl b/internal/shaders/output/intersect_scene_swrt_atlas.comp.spv.inl
diff --git a/internal/shaders/output/intersect_scene_swrt_bindless.comp.cso.inl b/internal/shaders/output/intersect_scene_swrt_bindless.comp.cso.inl
diff --git a/internal/shaders/output/intersect_scene_swrt_bindless.comp.spv.inl b/internal/shaders/output/intersect_scene_swrt_bindless.comp.spv.inl
diff --git a/internal/shaders/output/mix_incremental.comp.cso.inl b/internal/shaders/output/mix_incremental.comp.cso.inl
diff --git a/internal/shaders/output/nlm_filter.comp.cso.inl b/internal/shaders/output/nlm_filter.comp.cso.inl
diff --git a/internal/shaders/output/postprocess.comp.cso.inl b/internal/shaders/output/postprocess.comp.cso.inl
diff --git a/internal/shaders/output/prepare_indir_args.comp.cso.inl b/internal/shaders/output/prepare_indir_args.comp.cso.inl
diff --git a/internal/shaders/output/primary_ray_gen_adaptive.comp.cso.inl b/internal/shaders/output/primary_ray_gen_adaptive.comp.cso.inl
diff --git a/internal/shaders/output/primary_ray_gen_simple.comp.cso.inl b/internal/shaders/output/primary_ray_gen_simple.comp.cso.inl
diff --git a/internal/shaders/output/shade_primary_atlas.comp.cso.inl b/internal/shaders/output/shade_primary_atlas.comp.cso.inl
diff --git a/internal/shaders/output/shade_primary_atlas.comp.spv.inl b/internal/shaders/output/shade_primary_atlas.comp.spv.inl
diff --git a/internal/shaders/output/shade_primary_bindless.comp.cso.inl b/internal/shaders/output/shade_primary_bindless.comp.cso.inl
diff --git a/internal/shaders/output/shade_primary_bindless.comp.spv.inl b/internal/shaders/output/shade_primary_bindless.comp.spv.inl
diff --git a/internal/shaders/output/shade_secondary_atlas.comp.cso.inl b/internal/shaders/output/shade_secondary_atlas.comp.cso.inl
diff --git a/internal/shaders/output/shade_secondary_atlas.comp.spv.inl b/internal/shaders/output/shade_secondary_atlas.comp.spv.inl
diff --git a/internal/shaders/output/shade_secondary_bindless.comp.cso.inl b/internal/shaders/output/shade_secondary_bindless.comp.cso.inl
diff --git a/internal/shaders/output/shade_secondary_bindless.comp.spv.inl b/internal/shaders/output/shade_secondary_bindless.comp.spv.inl
diff --git a/internal/shaders/output/sort_hash_rays.comp.cso.inl b/internal/shaders/output/sort_hash_rays.comp.cso.inl
diff --git a/internal/shaders/output/sort_init_count_table.comp.cso.inl b/internal/shaders/output/sort_init_count_table.comp.cso.inl
diff --git a/internal/shaders/output/sort_init_count_table.comp.spv.inl b/internal/shaders/output/sort_init_count_table.comp.spv.inl
diff --git a/internal/shaders/output/sort_reduce.comp.cso.inl b/internal/shaders/output/sort_reduce.comp.cso.inl
diff --git a/internal/shaders/output/sort_reorder_rays.comp.cso.inl b/internal/shaders/output/sort_reorder_rays.comp.cso.inl
diff --git a/internal/shaders/output/sort_scan.comp.cso.inl b/internal/shaders/output/sort_scan.comp.cso.inl
diff --git a/internal/shaders/output/sort_scan_add.comp.cso.inl b/internal/shaders/output/sort_scan_add.comp.cso.inl
diff --git a/internal/shaders/output/sort_scatter.comp.cso.inl b/internal/shaders/output/sort_scatter.comp.cso.inl
diff --git a/internal/shaders/output/sort_scatter.comp.spv.inl b/internal/shaders/output/sort_scatter.comp.spv.inl
diff --git a/internal/shaders/shade.comp.glsl b/internal/shaders/shade.comp.glsl
@@ -600,17 +600,25 @@ float pop_ior_stack(inout float stack[4], float default_value) {
 }
 
 float peek_ior_stack(float stack[4], bool skip_first, float default_value) {
-    if (stack[3] > 0.0 && !exchange(skip_first, false)) {
-        return stack[3];
+    if (stack[3] > 0.0) {
+        if (!exchange(skip_first, false)) {
+            return stack[3];
+        }
     }
-    if (stack[2] > 0.0 && !exchange(skip_first, false)) {
-        return stack[2];
+    if (stack[2] > 0.0) {
+        if (!exchange(skip_first, false)) {
+            return stack[2];
+        }
     }
-    if (stack[1] > 0.0 && !exchange(skip_first, false)) {
-        return stack[1];
+    if (stack[1] > 0.0) {
+        if (!exchange(skip_first, false)) {
+            return stack[1];
+        }
     }
-    if (stack[0] > 0.0 && !exchange(skip_first, false)) {
-        return stack[0];
+    if (stack[0] > 0.0) {
+        if (!exchange(skip_first, false)) {
+            return stack[0];
+        }
     }
     return default_value;
 }
@@ -909,10 +917,10 @@ void SampleLightSource(vec3 P, vec3 T, vec3 B, vec3 N, const float rand_pick_lig
 
 #if USE_HIERARCHICAL_NEE
     float factor = 1.0;
-    uint i = 0; // start from root
-    while ((g_light_wnodes[i].child[0] & LEAF_NODE_BIT) == 0) {
+    light_wbvh_node_t n = g_light_wnodes[0]; // start from root
+    while ((n.child[0] & LEAF_NODE_BIT) == 0) {
         float importance[8];
-        const float total_importance = calc_lnode_importance(g_light_wnodes[i], P, importance);
+        const float total_importance = calc_lnode_importance(n, P, importance);
         [[dont_flatten]] if (total_importance == 0.0) {
             // failed to find lightsource for sampling
             return;
@@ -943,10 +951,10 @@ void SampleLightSource(vec3 P, vec3 T, vec3 B, vec3 N, const float rand_pick_lig
         }
 
         u1 = fract((u1 - importance_cdf[next]) / importance[next]);
-        i = g_light_wnodes[i].child[next];
+        n = g_light_wnodes[n.child[next]];
         factor *= importance[next];
     }
-    const uint light_index = (g_light_wnodes[i].child[0] & PRIM_INDEX_BITS);
+    const uint light_index = (n.child[0] & PRIM_INDEX_BITS);
     factor = (1.0 / factor);
 #else
     uint light_index = min(uint(u1 * g_params.li_count), uint(g_params.li_count - 1));

diff --git a/internal/shaders/sort_init_count_table.comp.glsl b/internal/shaders/sort_init_count_table.comp.glsl
@@ -26,8 +26,8 @@ layout (local_size_x = SORT_THREADGROUP_SIZE, local_size_y = 1, local_size_z = 1
 
 void main() {
     const int li = int(gl_LocalInvocationID.x), wi = int(gl_WorkGroupID.x);
-    for (int i = 0; i < 0x10; ++i) {
-        g_shared_counters[i * SORT_THREADGROUP_SIZE + li] = 0;
+    for (int i1 = 0; i1 < 0x10; ++i1) {
+        g_shared_counters[i1 * SORT_THREADGROUP_SIZE + li] = 0;
     }
     groupMemoryBarrier(); barrier();
 
@@ -40,15 +40,15 @@ void main() {
     int data_index = block_start + li;
 
     uint hashes[SORT_ELEMENTS_PER_THREAD];
-    [[unroll]] for (int i = 0; i < SORT_ELEMENTS_PER_THREAD; ++i) {
-        if (data_index + i * SORT_THREADGROUP_SIZE < data_count) {
-            hashes[i] = g_hashes[data_index + i * SORT_THREADGROUP_SIZE].hash;
+    [[unroll]] for (int i2 = 0; i2 < SORT_ELEMENTS_PER_THREAD; ++i2) {
+        if (data_index + i2 * SORT_THREADGROUP_SIZE < data_count) {
+            hashes[i2] = g_hashes[data_index + i2 * SORT_THREADGROUP_SIZE].hash;
         }
     }
 
-    for (int i = 0; i < SORT_ELEMENTS_PER_THREAD; ++i) {
+    for (int i3 = 0; i3 < SORT_ELEMENTS_PER_THREAD; ++i3) {
         if (data_index < data_count) {
-            uint local_key = (hashes[i] >> g_params.shift) & 0xf;
+            uint local_key = (hashes[i3] >> g_params.shift) & 0xf;
             atomicAdd(g_shared_counters[local_key * SORT_THREADGROUP_SIZE + li], 1u);
             data_index += SORT_THREADGROUP_SIZE;
         }
@@ -58,8 +58,8 @@ void main() {
 
     if (li < 0x10) {
         uint sum = 0;
-        for (int i = 0; i < SORT_THREADGROUP_SIZE; ++i) {
-            sum += g_shared_counters[li * SORT_THREADGROUP_SIZE + i];
+        for (int i4 = 0; i4 < SORT_THREADGROUP_SIZE; ++i4) {
+            sum += g_shared_counters[li * SORT_THREADGROUP_SIZE + i4];
         }
         g_count_table[li * block_count + wi] = sum;
     }

diff --git a/internal/shaders/sort_scatter.comp.glsl b/internal/shaders/sort_scatter.comp.glsl
@@ -107,7 +107,10 @@ void main() {
             g_shared_local_counters[li] = 0;
         }
 
-        ray_hash_t local_key = (data_index < data_count) ? src_keys[i] : ray_hash_t(0xffffffff, 0xffffffff);
+        ray_hash_t local_key = ray_hash_t(0xffffffff, 0xffffffff);
+        if (data_index < data_count) {
+            local_key = src_keys[i];
+        }
 
         for (uint bit_shift = 0; bit_shift < SORT_BITS_PER_PASS; bit_shift += 2) {
             uint key_index = (local_key.hash >> g_params.shift) & 0xf;