Internal/6000.0/staging

Packages/com.unity.render-pipelines.core/Runtime/GPUDriven/InstanceCuller.cs

@@ -681,7 +681,7 @@ unsafe public void Execute()
                     allocInfo.drawCount = outIndirectCommandIndex;
                     allocInfo.instanceCount = outIndirectVisibleInstanceIndex;
 
-                    int drawAllocCount = allocInfo.drawCount + IndirectBufferContextStorage.kExtraDrawAllocationCount;
+                    int drawAllocCount = allocInfo.drawCount;
                     int drawAllocEnd = Interlocked.Add(ref UnsafeUtility.AsRef<int>(allocCounters + (int)IndirectAllocator.NextDrawIndex), drawAllocCount);
                     allocInfo.drawAllocIndex = drawAllocEnd - drawAllocCount;
 
@@ -1525,6 +1525,7 @@ private static class ShaderIDs
             public static readonly int InstanceOcclusionCullerShaderVariables = Shader.PropertyToID("InstanceOcclusionCullerShaderVariables");
             public static readonly int _DrawInfo = Shader.PropertyToID("_DrawInfo");
             public static readonly int _InstanceInfo = Shader.PropertyToID("_InstanceInfo");
+            public static readonly int _DispatchArgs = Shader.PropertyToID("_DispatchArgs");
             public static readonly int _DrawArgs = Shader.PropertyToID("_DrawArgs");
             public static readonly int _InstanceIndices = Shader.PropertyToID("_InstanceIndices");
             public static readonly int _InstanceDataBuffer = Shader.PropertyToID("_InstanceDataBuffer");
@@ -1800,7 +1801,7 @@ public unsafe JobHandle CreateCullJobTree(
                     cullingOutput = cullingOutput.drawCommands,
                     indirectBufferLimits = indirectBufferLimits,
                     visibleInstancesBufferHandle = m_IndirectStorage.visibleInstanceBufferHandle,
-                    indirectArgsBufferHandle = m_IndirectStorage.indirectArgsBufferHandle,
+                    indirectArgsBufferHandle = m_IndirectStorage.indirectDrawArgsBufferHandle,
                     indirectBufferAllocInfo = indirectBufferAllocInfo,
                     indirectInstanceInfoGlobalArray = m_IndirectStorage.instanceInfoGlobalArray,
                     indirectDrawInfoGlobalArray = m_IndirectStorage.drawInfoGlobalArray,
@@ -2092,7 +2093,7 @@ internal void EnsureValidOcclusionTestResults(int viewInstanceID)
                     int kernel = m_CopyInstancesKernel;
                     cmd.SetComputeBufferParam(cs, kernel, ShaderIDs._DrawInfo, m_IndirectStorage.drawInfoBuffer);
                     cmd.SetComputeBufferParam(cs, kernel, ShaderIDs._InstanceInfo, m_IndirectStorage.instanceInfoBuffer);
-                    cmd.SetComputeBufferParam(cs, kernel, ShaderIDs._DrawArgs, m_IndirectStorage.argsBuffer);
+                    cmd.SetComputeBufferParam(cs, kernel, ShaderIDs._DrawArgs, m_IndirectStorage.drawArgsBuffer);
                     cmd.SetComputeBufferParam(cs, kernel, ShaderIDs._InstanceIndices, m_IndirectStorage.instanceBuffer);
 
                     cmd.DispatchCompute(cs, kernel, (allocInfo.instanceCount + 63) / 64, 1, 1);
@@ -2245,10 +2246,10 @@ private void AddOcclusionCullingDispatch(
                         if (doCopyInstances)
                         {
                             int kernel = m_CopyInstancesKernel;
-                            cmd.SetComputeBufferParam(cs, kernel, ShaderIDs._DrawInfo, m_IndirectStorage.drawInfoBuffer);
-                            cmd.SetComputeBufferParam(cs, kernel, ShaderIDs._InstanceInfo, m_IndirectStorage.instanceInfoBuffer);
-                            cmd.SetComputeBufferParam(cs, kernel, ShaderIDs._DrawArgs, m_IndirectStorage.argsBuffer);
-                            cmd.SetComputeBufferParam(cs, kernel, ShaderIDs._InstanceIndices, m_IndirectStorage.instanceBuffer);
+                            cmd.SetComputeBufferParam(cs, kernel, ShaderIDs._DrawInfo, bufferHandles.drawInfoBuffer);
+                            cmd.SetComputeBufferParam(cs, kernel, ShaderIDs._InstanceInfo, bufferHandles.instanceInfoBuffer);
+                            cmd.SetComputeBufferParam(cs, kernel, ShaderIDs._DrawArgs, bufferHandles.drawArgsBuffer);
+                            cmd.SetComputeBufferParam(cs, kernel, ShaderIDs._InstanceIndices, bufferHandles.instanceBuffer);
 
                             cmd.DispatchCompute(cs, kernel, (allocInfo.instanceCount + 63) / 64, 1, 1);
                         }
@@ -2257,7 +2258,10 @@ private void AddOcclusionCullingDispatch(
                         {
                             int kernel = m_ResetDrawArgsKernel;
                             cmd.SetComputeBufferParam(cs, kernel, ShaderIDs._DrawInfo, bufferHandles.drawInfoBuffer);
-                            cmd.SetComputeBufferParam(cs, kernel, ShaderIDs._DrawArgs, bufferHandles.argsBuffer);
+                            cmd.SetComputeBufferParam(cs, kernel, ShaderIDs._DrawArgs, bufferHandles.drawArgsBuffer);
+                            if (isSecondPass)
+                                cmd.SetComputeBufferParam(cs, kernel, ShaderIDs._DispatchArgs, bufferHandles.dispatchArgsBuffer);
+
                             cmd.DispatchCompute(cs, kernel, (allocInfo.drawCount + 63) / 64, 1, 1);
                         }
 
@@ -2266,7 +2270,7 @@ private void AddOcclusionCullingDispatch(
                             int kernel = m_CullInstancesKernel;
                             cmd.SetComputeBufferParam(cs, kernel, ShaderIDs._DrawInfo, bufferHandles.drawInfoBuffer);
                             cmd.SetComputeBufferParam(cs, kernel, ShaderIDs._InstanceInfo, bufferHandles.instanceInfoBuffer);
-                            cmd.SetComputeBufferParam(cs, kernel, ShaderIDs._DrawArgs, bufferHandles.argsBuffer);
+                            cmd.SetComputeBufferParam(cs, kernel, ShaderIDs._DrawArgs, bufferHandles.drawArgsBuffer);
                             cmd.SetComputeBufferParam(cs, kernel, ShaderIDs._InstanceIndices, bufferHandles.instanceBuffer);
                             cmd.SetComputeBufferParam(cs, kernel, ShaderIDs._InstanceDataBuffer, batchersContext.gpuInstanceDataBuffer);
                             cmd.SetComputeBufferParam(cs, kernel, ShaderIDs._OcclusionDebugCounters, m_OcclusionEventDebugArray.CounterBuffer);
@@ -2278,7 +2282,7 @@ private void AddOcclusionCullingDispatch(
                                 OcclusionCullingCommon.SetDebugPyramid(cmd, m_OcclusionTestShader, kernel, occluderHandles);
 
                             if (isSecondPass)
-                                cmd.DispatchCompute(cs, kernel, bufferHandles.argsBuffer, (uint)(GraphicsBuffer.IndirectDrawIndexedArgs.size * allocInfo.GetExtraDrawInfoSlotIndex()));
+                                cmd.DispatchCompute(cs, kernel, bufferHandles.dispatchArgsBuffer, 0);
                             else
                                 cmd.DispatchCompute(cs, kernel, (allocInfo.instanceCount + 63) / 64, 1, 1);
                         }

Packages/com.unity.render-pipelines.core/Runtime/GPUDriven/InstanceOcclusionCuller.cs

@@ -560,22 +560,23 @@ internal struct IndirectBufferContextHandles
     {
         public BufferHandle instanceBuffer;
         public BufferHandle instanceInfoBuffer;
-        public BufferHandle argsBuffer;
+        public BufferHandle dispatchArgsBuffer;
+        public BufferHandle drawArgsBuffer;
         public BufferHandle drawInfoBuffer;
 
         public void UseForOcclusionTest(IBaseRenderGraphBuilder builder)
         {
             instanceBuffer = builder.UseBuffer(instanceBuffer, AccessFlags.ReadWrite);
             instanceInfoBuffer = builder.UseBuffer(instanceInfoBuffer, AccessFlags.Read);
-            argsBuffer = builder.UseBuffer(argsBuffer, AccessFlags.ReadWrite);
+            dispatchArgsBuffer = builder.UseBuffer(dispatchArgsBuffer, AccessFlags.ReadWrite);
+            drawArgsBuffer = builder.UseBuffer(drawArgsBuffer, AccessFlags.ReadWrite);
             drawInfoBuffer = builder.UseBuffer(drawInfoBuffer, AccessFlags.Read);
         }
     }
 
     internal struct IndirectBufferContextStorage : IDisposable
     {
         private const int kAllocatorCount = (int)IndirectAllocator.Count;
-        internal const int kExtraDrawAllocationCount = 1;           // over-allocate by one for indirect args scratch space GPU-side
         internal const int kInstanceInfoGpuOffsetMultiplier = 2;    // GPU side allocates storage for extra copy of instance list
 
         private IndirectBufferLimits m_BufferLimits;
@@ -584,7 +585,8 @@ internal struct IndirectBufferContextStorage : IDisposable
         private GraphicsBuffer m_InstanceInfoBuffer;
         private NativeArray<IndirectInstanceInfo> m_InstanceInfoStaging;
 
-        private GraphicsBuffer m_ArgsBuffer;
+        private GraphicsBuffer m_DispatchArgsBuffer;
+        private GraphicsBuffer m_DrawArgsBuffer;
         private GraphicsBuffer m_DrawInfoBuffer;
         private NativeArray<IndirectDrawInfo> m_DrawInfoStaging;
 
@@ -596,19 +598,21 @@ internal struct IndirectBufferContextStorage : IDisposable
 
         public GraphicsBuffer instanceBuffer { get { return m_InstanceBuffer; } }
         public GraphicsBuffer instanceInfoBuffer { get { return m_InstanceInfoBuffer; } }
-        public GraphicsBuffer argsBuffer { get { return m_ArgsBuffer; } }
+        public GraphicsBuffer dispatchArgsBuffer { get { return m_DispatchArgsBuffer; } }
+        public GraphicsBuffer drawArgsBuffer { get { return m_DrawArgsBuffer; } }
         public GraphicsBuffer drawInfoBuffer { get { return m_DrawInfoBuffer; } }
 
         public GraphicsBufferHandle visibleInstanceBufferHandle { get { return m_InstanceBuffer.bufferHandle; } }
-        public GraphicsBufferHandle indirectArgsBufferHandle { get { return m_ArgsBuffer.bufferHandle; } }
+        public GraphicsBufferHandle indirectDrawArgsBufferHandle { get { return m_DrawArgsBuffer.bufferHandle; } }
 
         public IndirectBufferContextHandles ImportBuffers(RenderGraph renderGraph)
         {
             return new IndirectBufferContextHandles()
             {
                 instanceBuffer = renderGraph.ImportBuffer(m_InstanceBuffer),
                 instanceInfoBuffer = renderGraph.ImportBuffer(m_InstanceInfoBuffer),
-                argsBuffer = renderGraph.ImportBuffer(m_ArgsBuffer),
+                dispatchArgsBuffer = renderGraph.ImportBuffer(m_DispatchArgsBuffer),
+                drawArgsBuffer = renderGraph.ImportBuffer(m_DrawArgsBuffer),
                 drawInfoBuffer = renderGraph.ImportBuffer(m_DrawInfoBuffer),
             };
         }
@@ -653,15 +657,18 @@ void FreeInstanceBuffers()
 
         void AllocateDrawBuffers(int maxDrawCount)
         {
-            m_ArgsBuffer = new GraphicsBuffer(GraphicsBuffer.Target.Structured | GraphicsBuffer.Target.IndirectArguments, (maxDrawCount + kExtraDrawAllocationCount) * (GraphicsBuffer.IndirectDrawIndexedArgs.size / sizeof(int)), sizeof(int));
+            // Compute dispatch arguments are number of thread groups in X,Y and Z dimensions, hence 3 integers for the size.
+            m_DispatchArgsBuffer = new GraphicsBuffer(GraphicsBuffer.Target.Structured | GraphicsBuffer.Target.IndirectArguments, 3, sizeof(int));
+            m_DrawArgsBuffer = new GraphicsBuffer(GraphicsBuffer.Target.Structured | GraphicsBuffer.Target.IndirectArguments, maxDrawCount * (GraphicsBuffer.IndirectDrawIndexedArgs.size / sizeof(int)), sizeof(int));
             m_DrawInfoBuffer = new GraphicsBuffer(GraphicsBuffer.Target.Structured, maxDrawCount, System.Runtime.InteropServices.Marshal.SizeOf<IndirectDrawInfo>());
             m_DrawInfoStaging = new NativeArray<IndirectDrawInfo>(maxDrawCount, Allocator.Persistent, NativeArrayOptions.UninitializedMemory);
             m_BufferLimits.maxDrawCount = maxDrawCount;
         }
 
         void FreeDrawBuffers()
         {
-            m_ArgsBuffer.Release();
+            m_DispatchArgsBuffer.Release();
+            m_DrawArgsBuffer.Release();
             m_DrawInfoBuffer.Release();
             m_DrawInfoStaging.Dispose();
             m_BufferLimits.maxDrawCount = 0;

Packages/com.unity.render-pipelines.core/Runtime/Lighting/ProbeVolume/ProbeVolume.hlsl

@@ -911,18 +911,24 @@ void EvaluateAdaptiveProbeVolume(in float3 posWS, in float3 normalWS, in float3
 void EvaluateAdaptiveProbeVolume(in float3 posWS, in float2 positionSS, out float3 bakeDiffuseLighting)
 {
     APVResources apvRes = FillAPVResources();
-
     posWS = AddNoiseToSamplingPosition(posWS, positionSS, 1);
     posWS -= _APVWorldOffset;
 
+    float3 ambientProbe = EvaluateAmbientProbe(0);
+
     float3 uvw;
     if (TryToGetPoolUVW(apvRes, posWS, 0, 0, uvw))
     {
         bakeDiffuseLighting = SAMPLE_TEXTURE3D_LOD(apvRes.L0_L1Rx, s_linear_clamp_sampler, uvw, 0).rgb;
+        if (_APVSkyOcclusionWeight > 0)
+        {
+            float skyOcclusionL0 = kSHBasis0 * SAMPLE_TEXTURE3D_LOD(apvRes.SkyOcclusionL0L1, s_linear_clamp_sampler, uvw, 0).x;
+            bakeDiffuseLighting += ambientProbe * skyOcclusionL0;
+        }
     }
     else
     {
-        bakeDiffuseLighting = EvaluateAmbientProbe(0);
+        bakeDiffuseLighting = ambientProbe;
     }
 }
 

Packages/com.unity.render-pipelines.core/Runtime/Lighting/ProbeVolume/ProbeVolumeBakingSet.cs

@@ -26,6 +26,7 @@ public LogarithmicAttribute(int min, int max)
     /// <summary>
     /// An Asset which holds a set of settings to use with a <see cref="Probe Reference Volume"/>.
     /// </summary>
+    [CoreRPHelpURL("probevolumes-usebakingsets", "com.unity.render-pipelines.high-definition")]
     public sealed partial class ProbeVolumeBakingSet : ScriptableObject, ISerializationCallbackReceiver
     {
         internal enum Version

Packages/com.unity.render-pipelines.core/Runtime/RenderGraph/Compiler/NativePassCompiler.cs

@@ -656,7 +656,7 @@ void DetectMemoryLessResources()
                 foreach (ref readonly var nativePass in contextData.NativePasses)
                 {
                     // Loop over all created resources by this nrp
-                    var graphPasses = nativePass.GraphPasses(contextData);
+                    var graphPasses = nativePass.GraphPasses(contextData, out var actualPasses);
                     foreach (ref readonly var subPass in graphPasses)
                     {
                         foreach (ref readonly var createdRes in subPass.FirstUsedResources(contextData))
@@ -698,6 +698,8 @@ void DetectMemoryLessResources()
                             }
                         }
                     }
+                    if (actualPasses.IsCreated)
+                        actualPasses.Dispose();
                 }
             }
         }
@@ -742,7 +744,8 @@ private void ExecuteInitializeResource(InternalRenderGraphContext rgContext, Ren
                     if (pass.mergeState == PassMergeState.Begin || pass.mergeState == PassMergeState.None)
                     {
                         ref var nativePass = ref contextData.nativePassData.ElementAt(pass.nativePassIndex);
-                        foreach (ref readonly var subPass in nativePass.GraphPasses(contextData))
+                        var graphPasses = nativePass.GraphPasses(contextData, out var actualPasses);
+                        foreach (ref readonly var subPass in graphPasses)
                         {
                             foreach (ref readonly var res in subPass.FirstUsedResources(contextData))
                             {
@@ -770,6 +773,8 @@ private void ExecuteInitializeResource(InternalRenderGraphContext rgContext, Ren
                                 }
                             }
                         }
+                        if (actualPasses.IsCreated)
+                            actualPasses.Dispose();
                     }
                 }
                 // Other passes just create them at the beginning of the individual pass
@@ -1290,7 +1295,8 @@ private void ExecuteDestroyResource(InternalRenderGraphContext rgContext, Render
                     if (pass.mergeState == PassMergeState.End || pass.mergeState == PassMergeState.None)
                     {
                         ref var nativePass = ref contextData.nativePassData.ElementAt(pass.nativePassIndex);
-                        foreach (ref readonly var subPass in nativePass.GraphPasses(contextData))
+                        var graphPasses = nativePass.GraphPasses(contextData, out var actualPasses);
+                        foreach (ref readonly var subPass in graphPasses)
                         {
                             foreach (ref readonly var res in subPass.LastUsedResources(contextData))
                             {
@@ -1301,6 +1307,8 @@ private void ExecuteDestroyResource(InternalRenderGraphContext rgContext, Render
                                 }
                             }
                         }
+                        if (actualPasses.IsCreated)
+                            actualPasses.Dispose();
                     }
                 }
                 else

Packages/com.unity.render-pipelines.core/Runtime/RenderGraph/Compiler/PassesData.cs

@@ -615,15 +615,16 @@ public readonly bool IsValid()
         }
 
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public readonly ReadOnlySpan<PassData> GraphPasses(CompilerContextData ctx)
+        public readonly ReadOnlySpan<PassData> GraphPasses(CompilerContextData ctx, out NativeArray<PassData> actualPasses)
         {
             // When there's no pass being culled, we can directly return a Span of the Native List
             if (lastGraphPass - firstGraphPass + 1 == numGraphPasses)
             {
+                actualPasses = default;
                 return ctx.passData.MakeReadOnlySpan(firstGraphPass, numGraphPasses);
             }
 
-            var actualPasses =
+            actualPasses =
                 new NativeArray<PassData>(numGraphPasses, Allocator.Temp,
                     NativeArrayOptions.UninitializedMemory);
 
@@ -642,10 +643,14 @@ public readonly ReadOnlySpan<PassData> GraphPasses(CompilerContextData ctx)
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
         public readonly void GetGraphPassNames(CompilerContextData ctx, DynamicArray<Name> dest)
         {
-            foreach (ref readonly var pass in GraphPasses(ctx))
+            var span = GraphPasses(ctx, out var actualPasses);
+            foreach (ref readonly var pass in span)
             {
                 dest.Add(pass.GetName(ctx));
             }
+
+            if (actualPasses.IsCreated)
+                actualPasses.Dispose();
         }
 
         // This function does not modify the current render graph state, it only evaluates and returns the correct PassBreakAudit