Fixed ambient probe convolution warning

com.unity.render-pipelines.high-definition/Runtime/Sky/AmbientProbeConvolution.compute

@@ -24,6 +24,12 @@ RWStructuredBuffer<float> _AmbientProbeOutputBuffer;
 RWStructuredBuffer<float4> _VolumetricAmbientProbeOutputBuffer;
 RWStructuredBuffer<float4> _DiffuseAmbientProbeOutputBuffer;
 
+// If we use local VGPRs as a scratch buffer we end up using too many register
+// To avoid that we go through memory.
+// This is quite messy and bad for performance but this shader should never be critical so it should be fine.
+// Uint is used as it's the only format supported everywhere as read/write from same thread.
+RWStructuredBuffer<uint> _ScratchBuffer;
+
 uniform float4 _FogParameters;
 
 #define _FogDimmer _FogParameters.x
@@ -40,6 +46,45 @@ uniform float4 _FogParameters;
     groupshared float outputSHCoeffsLDS[SH_COEFF_COUNT * SAMPLE_COUNT / 2];
 #endif
 
+    void PackSHFromScratchBuffer(RWStructuredBuffer<float4> buffer)
+    {
+        int c = 0;
+        for (c = 0; c < 3; c++)
+
+        {
+            buffer[c] = float4(asfloat(_ScratchBuffer[c * 9 + 3]), asfloat(_ScratchBuffer[c * 9 + 1]), asfloat(_ScratchBuffer[c * 9 + 2]), asfloat(_ScratchBuffer[c * 9 + 0]) - asfloat(_ScratchBuffer[c * 9 + 6]));
+        }
+
+        // Quadratic (4/5)
+        for (c = 0; c < 3; c++)
+        {
+            buffer[3 + c] = float4(asfloat(_ScratchBuffer[c * 9 + 4]), asfloat(_ScratchBuffer[c * 9 + 5]), asfloat(_ScratchBuffer[c * 9 + 6]) * 3.0f, asfloat(_ScratchBuffer[c * 9 + 7]));
+        }
+
+        // Quadratic (5)
+        buffer[6] = float4(asfloat(_ScratchBuffer[0 * 9 + 8]), asfloat(_ScratchBuffer[1 * 9 + 8]), asfloat(_ScratchBuffer[2 * 9 + 8]), 1.0f);
+    }
+
+    void ConvolveZonalFromScratchBuffer(float3 zh)
+    {
+        for (int l = 0; l <= 2; l++)
+        {
+            float n = sqrt((4.0f * PI) / (2 * l + 1));
+            float k = zh[l];
+            float p = n * k;
+
+            for (int m = -l; m <= l; m++)
+            {
+                int i = l * (l + 1) + m;
+
+                for (int c = 0; c < 3; c++)
+                {
+                    _ScratchBuffer[c * 9 + i] = asuint(asfloat(_ScratchBuffer[c * 9 + i]) * p);
+                }
+            }
+        }
+    }
+
 [numthreads(SAMPLE_COUNT, 1, 1)]
 void KERNEL_NAME(uint dispatchThreadId : SV_DispatchThreadID)
 {
@@ -173,16 +218,14 @@ void KERNEL_NAME(uint dispatchThreadId : SV_DispatchThreadID)
             _AmbientProbeOutputBuffer[i] = outputSHCoeffs[i] * kClampedCosineCoefs[i % 9] * kSHBasisCoef[i % 9];
         }
 
-        float result[27];
-
 #if OUTPUT_DIFFUSE
         for (i = 0; i < SH_COEFF_COUNT; ++i)
         {
-            result[i] = _AmbientProbeOutputBuffer[i];
+            _ScratchBuffer[i] = asuint(_AmbientProbeOutputBuffer[i]);
         }
 
         // Diffuse convolution packed to be ready for shader consumption
-        PackSH(_DiffuseAmbientProbeOutputBuffer, result);
+        PackSHFromScratchBuffer(_DiffuseAmbientProbeOutputBuffer);
 #endif
 
 #if OUTPUT_VOLUMETRIC
@@ -191,24 +234,24 @@ void KERNEL_NAME(uint dispatchThreadId : SV_DispatchThreadID)
         // Apply FogDimmer
         for (i = 0; i < SH_COEFF_COUNT; ++i)
         {
-            result[i] = outputSHCoeffs[i] *_FogDimmer;
+            _ScratchBuffer[i] = asuint(outputSHCoeffs[i] *_FogDimmer);
         }
 
         // Apply CornetteShank phase function
         float3 zh;
         GetCornetteShanksPhaseFunction(zh, _FogAnisotropy);
-        ConvolveZonal(result, zh);
+        ConvolveZonalFromScratchBuffer(zh);
 
         // Premultiplies the SH with the polynomial coefficients of SH basis functions,
         // which avoids using any constants during SH evaluation.
         // The resulting evaluation takes the form:
         // (c_0 - c_6) + c_1 y + c_2 z + c_3 x + c_4 x y + c_5 y z + c_6 (3 z^2) + c_7 x z + c_8 (x^2 - y^2)
         for (i = 0; i < SH_COEFF_COUNT; ++i)
         {
-            result[i] = result[i] * kSHBasisCoef[i % 9];
+            _ScratchBuffer[i] = asuint(asfloat(_ScratchBuffer[i]) * kSHBasisCoef[i % 9]);
         }
 
-        PackSH(_VolumetricAmbientProbeOutputBuffer, result);
+        PackSHFromScratchBuffer(_VolumetricAmbientProbeOutputBuffer);
 #endif
     }
 }

com.unity.render-pipelines.high-definition/Runtime/Sky/SkyManager.cs

@@ -195,6 +195,7 @@ class SkyManager
         static readonly int s_AmbientProbeOutputBufferParam = Shader.PropertyToID("_AmbientProbeOutputBuffer");
         static readonly int s_VolumetricAmbientProbeOutputBufferParam = Shader.PropertyToID("_VolumetricAmbientProbeOutputBuffer");
         static readonly int s_DiffuseAmbientProbeOutputBufferParam = Shader.PropertyToID("_DiffuseAmbientProbeOutputBuffer");
+        static readonly int s_ScratchBufferParam = Shader.PropertyToID("_ScratchBuffer");
         static readonly int s_AmbientProbeInputCubemap = Shader.PropertyToID("_AmbientProbeInputCubemap");
         static readonly int s_FogParameters = Shader.PropertyToID("_FogParameters");
         int m_ComputeAmbientProbeKernel;
@@ -721,6 +722,7 @@ class UpdateAmbientProbePassData
             public ComputeBuffer ambientProbeResult;
             public ComputeBuffer diffuseAmbientProbeResult;
             public ComputeBuffer volumetricAmbientProbeResult;
+            public ComputeBufferHandle scratchBuffer;
             public Vector4 fogParameters;
             public Action<AsyncGPUReadbackRequest> callback;
         }
@@ -738,6 +740,7 @@ internal void UpdateAmbientProbe(RenderGraph renderGraph, TextureHandle skyCubem
                 passData.skyCubemap = builder.ReadTexture(skyCubemap);
                 passData.ambientProbeResult = ambientProbeResult;
                 passData.diffuseAmbientProbeResult = diffuseAmbientProbeResult;
+                passData.scratchBuffer = builder.CreateTransientComputeBuffer(new ComputeBufferDesc(27, sizeof(uint))); // L2 = 9 channel per component
                 passData.volumetricAmbientProbeResult = volumetricAmbientProbeResult;
                 passData.fogParameters = fog != null ? new Vector4(fog.globalLightProbeDimmer.value, fog.anisotropy.value, 0.0f, 0.0f) : Vector4.zero;
                 passData.callback = callback;
@@ -746,6 +749,7 @@ internal void UpdateAmbientProbe(RenderGraph renderGraph, TextureHandle skyCubem
                 (UpdateAmbientProbePassData data, RenderGraphContext ctx) =>
                 {
                     ctx.cmd.SetComputeBufferParam(data.computeAmbientProbeCS, data.computeAmbientProbeKernel, s_AmbientProbeOutputBufferParam, data.ambientProbeResult);
+                    ctx.cmd.SetComputeBufferParam(data.computeAmbientProbeCS, data.computeAmbientProbeKernel, s_ScratchBufferParam, data.scratchBuffer);
                     ctx.cmd.SetComputeTextureParam(data.computeAmbientProbeCS, data.computeAmbientProbeKernel, s_AmbientProbeInputCubemap, data.skyCubemap);
                     if (data.diffuseAmbientProbeResult != null)
                         ctx.cmd.SetComputeBufferParam(data.computeAmbientProbeCS, data.computeAmbientProbeKernel, s_DiffuseAmbientProbeOutputBufferParam, data.diffuseAmbientProbeResult);