perf: fix skinning performance issues

PilotLightTech · Jan 24, 2025 · 06fb680 · 06fb680
1 parent 36cda69
commit 06fb680
Show file tree

Hide file tree

Showing 4 changed files with 20 additions and 11 deletions.
diff --git a/extensions/pl_renderer_ext.c b/extensions/pl_renderer_ext.c
@@ -3497,8 +3497,8 @@ pl_refr_render_scene(uint32_t uSceneHandle, const uint32_t* auViewHandles, const
             uJumpSteps = 1;
 
         const plDispatch tDispach = {
-            .uGroupCountX     = (uint32_t)tDimensions.x / 8,
-            .uGroupCountY     = (uint32_t)tDimensions.y / 8,
+            .uGroupCountX     = (uint32_t)ceilf(tDimensions.x / 8.0f),
+            .uGroupCountY     = (uint32_t)ceilf(tDimensions.y / 8.0f),
             .uGroupCountZ     = 1,
             .uThreadPerGroupX = 8,
             .uThreadPerGroupY = 8,
@@ -3522,7 +3522,7 @@ pl_refr_render_scene(uint32_t uSceneHandle, const uint32_t* auViewHandles, const
                 .tTexture = ptView->atUVMaskTexture1,
                 .uSlot    = 1,
                 .tType    = PL_TEXTURE_BINDING_TYPE_STORAGE,
-                    .tCurrentUsage = PL_TEXTURE_USAGE_STORAGE
+                .tCurrentUsage = PL_TEXTURE_USAGE_STORAGE
             }
         };
 
@@ -3577,7 +3577,9 @@ pl_refr_render_scene(uint32_t uSceneHandle, const uint32_t* auViewHandles, const
 
             plDynamicBinding tDynamicBinding = pl__allocate_dynamic_data(ptDevice);
             plVec4* ptJumpDistance = (plVec4*)tDynamicBinding.pcData;
-            ptJumpDistance->x = fJumpDistance;
+            ptJumpDistance->x = tDimensions.x;
+            ptJumpDistance->y = tDimensions.y;
+            ptJumpDistance->z = fJumpDistance;
 
             gptGfx->bind_compute_bind_groups(ptJumpEncoder, gptData->tJFAShader, 0, 1, &atJFABindGroups[i % 2], 1, &tDynamicBinding);
             gptGfx->dispatch(ptJumpEncoder, 1, &tDispach);

diff --git a/extensions/pl_renderer_internal.c b/extensions/pl_renderer_internal.c
@@ -746,7 +746,7 @@ pl_refr_perform_skinning(plCommandBuffer* ptCommandBuffer, uint32_t uSceneHandle
         int iSourceDataOffset;
         int iDestDataOffset;
         int iDestVertexOffset;
-        int iUnused;
+        uint32_t uMaxSize;
     } SkinDynamicData;
 
     if(uSkinCount)
@@ -772,12 +772,13 @@ pl_refr_perform_skinning(plCommandBuffer* ptCommandBuffer, uint32_t uSceneHandle
             ptDynamicData->iSourceDataOffset = ptScene->sbtSkinData[i].iSourceDataOffset;
             ptDynamicData->iDestDataOffset = ptScene->sbtSkinData[i].iDestDataOffset;
             ptDynamicData->iDestVertexOffset = ptScene->sbtSkinData[i].iDestVertexOffset;
+            ptDynamicData->uMaxSize = ptScene->sbtSkinData[i].uVertexCount;
 
             const plDispatch tDispach = {
-                .uGroupCountX     = ptScene->sbtSkinData[i].uVertexCount,
+                .uGroupCountX     = (uint32_t)ceilf((float)ptScene->sbtSkinData[i].uVertexCount / 64.0f),
                 .uGroupCountY     = 1,
                 .uGroupCountZ     = 1,
-                .uThreadPerGroupX = 1,
+                .uThreadPerGroupX = 64,
                 .uThreadPerGroupY = 1,
                 .uThreadPerGroupZ = 1
             };

diff --git a/shaders/jumpfloodalgo.comp b/shaders/jumpfloodalgo.comp
@@ -58,9 +58,12 @@ void main()
     const int iXCoord = int(gl_WorkGroupID.x * 8 + gl_LocalInvocationID.x);
     const int iYCoord = int(gl_WorkGroupID.y * 8 + gl_LocalInvocationID.y);
 
+    if(iXCoord >= tDynamicData.tJumpDistance.x || iYCoord >= tDynamicData.tJumpDistance.y)
+        return;
+
     vec2 outColor = vec2(0.0, 0.0);
 
-    ivec2 jumpDist = ivec2(tDynamicData.tJumpDistance.x);
+    ivec2 jumpDist = ivec2(tDynamicData.tJumpDistance.z);
 
     vec3 curr = vec3(1,1,9999999);
     const ivec2 tMin = ivec2(iXCoord, iYCoord);

diff --git a/shaders/skinning.comp b/shaders/skinning.comp
@@ -47,7 +47,7 @@ layout(set = 3, binding = 0) uniform PL_DYNAMIC_DATA
     int iSourceDataOffset;
     int iDestDataOffset;
     int iDestVertexOffset;
-    int iUnused;
+    uint uMaxSize;
 } tObjectInfo;
 
 //-----------------------------------------------------------------------------
@@ -108,13 +108,16 @@ get_skinning_matrix(vec4 inJoints0, vec4 inWeights0)
 // [SECTION] entry
 //-----------------------------------------------------------------------------
 
-layout (local_size_x = 16, local_size_y = 16, local_size_z = 1) in;
+layout (local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
 
 void
 main()
 {
 
-    const uint iVertexIndex = gl_WorkGroupID.x;
+    const uint iVertexIndex = gl_WorkGroupID.x * 64 + gl_LocalInvocationID.x;
+
+    if(iVertexIndex >= tObjectInfo.uMaxSize)
+        return;
 
     vec4 inPosition  = vec4(0.0, 0.0, 0.0, 1.0);
     vec3 inNormal    = vec3(0.0, 0.0, 0.0);