Optimized Depth Pyramid Generation with Gather

ScriptableRenderPipeline/HDRenderPipeline/HDRenderPipeline.cs

-using System.Collections.Generic;
+using System.Collections.Generic;
 using UnityEngine.Rendering;
 using System;
 using System.Diagnostics;
                 var mipSize = size;
                 for (int i = 0; i < lodCount; i++)
                 {
+                    int srcMipSize = mipSize;
                     mipSize >>= 1;
 
                     cmd.ReleaseTemporaryRT(HDShaderIDs._DepthPyramidMips[i + 1]);
-                    cmd.SetComputeVectorParam(m_DepthPyramidCS, "_Size", new Vector4(mipSize, mipSize, 1f / mipSize, 1f / mipSize));
+                    cmd.SetComputeVectorParam(m_DepthPyramidCS, "_SrcSize", new Vector4(srcMipSize, srcMipSize, 1f / srcMipSize, 1f / srcMipSize));
                     cmd.DispatchCompute(m_DepthPyramidCS, m_DepthPyramidKernel, mipSize / 8, mipSize / 8, 1);
                     cmd.CopyTexture(HDShaderIDs._DepthPyramidMips[i + 1], 0, 0, m_DepthPyramidBufferRT, 0, i + 1);
                 }

ScriptableRenderPipeline/HDRenderPipeline/RenderPipelineResources/DepthDownsample.compute

-#include "ShaderLibrary/Common.hlsl"
-
-Texture2D<float> _Source;
-RWTexture2D<float> _Result;
-
-SamplerState sampler_LinearClamp;
-
-CBUFFER_START(cb)
-    float4 _Size;
-CBUFFER_END
-
-#pragma kernel KMain
-[numthreads(8, 8, 1)]
-void KMain(uint2 groupId : SV_GroupID, uint2 groupThreadId : SV_GroupThreadID, uint2 dispatchThreadId : SV_DispatchThreadID)
-{
-    // Upper-left pixel coordinate of quad that this thread will read
-    int2 threadUL = dispatchThreadId;
-
-    // Downsample the block
-    float2 offset = float2(threadUL);
-    float p00 = _Source.SampleLevel(sampler_LinearClamp, (offset                   ) * _Size.zw, 0.0).x;
-    float p10 = _Source.SampleLevel(sampler_LinearClamp, (offset + float2(1.0, 0.0)) * _Size.zw, 0.0).x;
-    float p01 = _Source.SampleLevel(sampler_LinearClamp, (offset + float2(0.0, 1.0)) * _Size.zw, 0.0).x;
-    float p11 = _Source.SampleLevel(sampler_LinearClamp, (offset + float2(1.0, 1.0)) * _Size.zw, 0.0).x;
-
-    float depth = min(min(min(p00, p01), p10), p11);
-
-    // Write to the final target
-    _Result[dispatchThreadId] = depth;
+#include "ShaderLibrary/Common.hlsl"
+
+Texture2D<float> _Source;
+RWTexture2D<float> _Result;
+
+SamplerState sampler_PointClamp;    //TODO: could we use min-sampler instead of using ALU?
+
+CBUFFER_START(cb)
+float4 _SrcSize;
+CBUFFER_END
+
+#pragma kernel KMain
+[numthreads(8, 8, 1)]
+void KMain(uint2 groupId : SV_GroupID, uint2 groupThreadId : SV_GroupThreadID, uint2 dispatchThreadId : SV_DispatchThreadID)
+{
+    // Upper-left pixel coordinate of quad that this thread will read
+    int2 threadUL = dispatchThreadId;
+
+    // Downsample the block
+    float2 offset = float2(threadUL) * 2.0f + 1.0f;
+    float4 depths = _Source.GatherRed(sampler_PointClamp, offset * _SrcSize.zw, 0.0);
+
+    float minDepth = min(min(depths.x, depths.y), min(depths.z, depths.w));
+
+    // Write to the final target
+    _Result[dispatchThreadId] = minDepth;
 }