Merge pull request microsoft#403 from Microsoft/develop-dxr

Adding initial Fallback Layer support for cheap AABB refitting when update build flag is used Triangle intersection fixes for close geometry Fallback layer support for geometry flags Small sample fixes
Krebrov001 · Jun 15, 2018 · 8f68bf0 · 8f68bf0
2 parents ea065cc + 2dcfc70
commit 8f68bf0
Show file tree

Hide file tree

Showing 37 changed files with 1,463 additions and 569 deletions.
diff --git a/Libraries/D3D12RaytracingFallback/src/BottomLevelLoadTriangles.hlsli b/Libraries/D3D12RaytracingFallback/src/BottomLevelLoadTriangles.hlsli
@@ -98,8 +98,6 @@ void main( uint3 DTid : SV_DispatchThreadID )
     uint NumberOfAABBs = NumberOfInternalNodes + Constants.TotalPrimitiveCount;
 
     uint3 indicies = GetIndex(localTriangleIndex);
-
-    uint globalTriangleIndex = localTriangleIndex + Constants.PrimitiveOffset;
 
     Triangle tri;
     tri.v0 = GetVertex(elementBuffer, indicies[0], Constants.ElementBufferStride);
@@ -120,6 +118,9 @@ void main( uint3 DTid : SV_DispatchThreadID )
         tri.v2 = TransformVertex(tri.v2, transform);
     }
 
-    PrimitiveBuffer[globalTriangleIndex] = CreateTrianglePrimitive(tri);
-    StorePrimitiveMetadata(globalTriangleIndex, localTriangleIndex);
+    uint globalTriangleIndex = localTriangleIndex + Constants.PrimitiveOffset;
+    uint outputIndex = GetOutputIndex(globalTriangleIndex);
+
+    PrimitiveBuffer[outputIndex] = CreateTrianglePrimitive(tri);
+    StorePrimitiveMetadata(outputIndex, localTriangleIndex);
 }
diff --git a/Libraries/D3D12RaytracingFallback/src/ComObject.h b/Libraries/D3D12RaytracingFallback/src/ComObject.h
@@ -37,3 +37,35 @@ public:                                                                        \
     }                                                                          \
 private:                                                                       \
     UINT m_ReferenceCount = 1;                                                 
+
+#define COM_IMPLEMENTATION_WITH_QUERYINTERFACE(pParentObject)                  \
+public:                                                                        \
+    virtual HRESULT STDMETHODCALLTYPE QueryInterface(                          \
+        REFIID riid,                                                           \
+        void **ppvObject)                                                      \
+    {                                                                          \
+        if (riid == __uuidof(*pParentObject))                                  \
+        {                                                                      \
+            pParentObject->AddRef();                                           \
+            *ppvObject = pParentObject;                                        \
+            return S_OK;                                                       \
+        }                                                                      \
+        return pParentObject->QueryInterface(riid, ppvObject);                 \
+    }                                                                          \
+                                                                               \
+    virtual ULONG STDMETHODCALLTYPE AddRef(void)                               \
+    {                                                                          \
+        return InterlockedIncrement(&m_ReferenceCount);                        \
+    }                                                                          \
+                                                                               \
+    virtual ULONG STDMETHODCALLTYPE Release(void)                              \
+    {                                                                          \
+        ULONG currentRefCount = InterlockedDecrement(&m_ReferenceCount);       \
+        if (currentRefCount == 0)                                              \
+        {                                                                      \
+            delete this;                                                       \
+        }                                                                      \
+        return currentRefCount;                                                \
+    }                                                                          \
+private:                                                                       \
+    UINT m_ReferenceCount = 1;                                                 
diff --git a/Libraries/D3D12RaytracingFallback/src/ComputeAABBs.hlsli b/Libraries/D3D12RaytracingFallback/src/ComputeAABBs.hlsli
@@ -23,9 +23,47 @@ uint GetLeafCount(uint boxIndex)
 uint4 GetNodeData(uint boxIndex, out uint2 flags)
 {
     uint nodeAddress = GetBoxAddress(offsetToBoxes, boxIndex);
-    uint4 data = outputBVH.Load4(nodeAddress);
-    flags = data.wz;
-    return data;
+    uint4 data1 = outputBVH.Load4(nodeAddress);
+    uint4 data2 = outputBVH.Load4(nodeAddress + 16);
+    flags = float2(data1.w, data2.w);
+    return data1;
+}
+
+uint2 GetNodeFlags(uint boxIndex) 
+{
+    uint2 flags;
+    GetNodeData(boxIndex, flags);
+    return flags;
+}
+
+uint GetLeftChildIndex(uint boxIndex) 
+{
+    if (ShouldPerformUpdate)
+    {
+        return GetLeftNodeIndex(GetNodeFlags(boxIndex));
+    }
+
+    return hierarchyBuffer[boxIndex].LeftChildIndex;
+}
+
+uint GetRightChildIndex(uint boxIndex) 
+{
+    if (ShouldPerformUpdate)
+    {
+        return GetRightNodeIndex(GetNodeFlags(boxIndex));
+    }
+
+    return hierarchyBuffer[boxIndex].RightChildIndex;
+}
+
+uint GetParentIndex(uint boxIndex) 
+{
+    if (ShouldPerformUpdate)
+    {
+        return aabbParentBuffer[boxIndex];
+    }
+
+    return hierarchyBuffer[boxIndex].ParentIndex;
 }
 
 [numthreads(THREAD_GROUP_1D_WIDTH, 1, 1)]
@@ -63,8 +101,8 @@ void main(uint3 DTid : SV_DispatchThreadID)
         }
         else
         {
-            uint leftNodeIndex = hierarchyBuffer[nodeIndex].LeftChildIndex;
-            uint rightNodeIndex = hierarchyBuffer[nodeIndex].RightChildIndex;
+            uint leftNodeIndex = GetLeftChildIndex(nodeIndex);
+            uint rightNodeIndex = GetRightChildIndex(nodeIndex);
             if (swapChildIndices)
             {
                 uint temp = leftNodeIndex;
@@ -104,21 +142,26 @@ void main(uint3 DTid : SV_DispatchThreadID)
             break;
         }
 
-        uint parentNodeIndex = hierarchyBuffer[nodeIndex].ParentIndex;
+        uint parentNodeIndex = GetParentIndex(nodeIndex);
         uint trianglesFromOtherChild;
         // If this counter was already incremented, that means both children for the parent 
         // node have computed their AABB, and the parent is ready to be processed
         childNodesProcessedCounter.InterlockedAdd(parentNodeIndex * SizeOfUINT32, numTriangles, trianglesFromOtherChild);
         if (trianglesFromOtherChild != 0)
         {
-
             // Prioritize having the smaller nodes on the left
             // TODO: Investigate better heuristics for node ordering
-            bool IsLeft = hierarchyBuffer[parentNodeIndex].LeftChildIndex == nodeIndex;
+            bool IsLeft = GetLeftChildIndex(parentNodeIndex) == nodeIndex;
             bool IsOtherSideSmaller = numTriangles > trianglesFromOtherChild;
             swapChildIndices = IsLeft ? IsOtherSideSmaller : !IsOtherSideSmaller;
             nodeIndex = parentNodeIndex;
             numTriangles += trianglesFromOtherChild;
+
+            if (ShouldPrepareUpdate)
+            {
+                aabbParentBuffer[GetLeftChildIndex(nodeIndex)] = nodeIndex;
+                aabbParentBuffer[GetRightChildIndex(nodeIndex)] = nodeIndex;
+            }
         }
         else
         {

diff --git a/Libraries/D3D12RaytracingFallback/src/ConstructAABBBindings.h b/Libraries/D3D12RaytracingFallback/src/ConstructAABBBindings.h
@@ -17,27 +17,33 @@
 struct InputConstants
 {
     uint NumberOfElements;
+    uint UpdateFlags;
 };
 
 // UAVs
 #define OutputBVHRegister 0
 #define ScratchBufferRegister 1
 #define ChildNodesProcessedBufferRegister 2
-#define MortonCodesBufferRegister 3
-#define HierarchyBufferRegister 4
+#define HierarchyBufferRegister 3
+#define AABBParentBufferRegister 4
 
 #define GlobalDescriptorHeapRegister 0
 #define GlobalDescriptorHeapRegisterSpace 1
 
 // CBVs
 #define InputConstantsRegister 0
 
+#define PREPARE_UPDATE_FLAG 0x1
+#define PERFORM_UPDATE_FLAG 0x2
+#define ShouldPrepareUpdate Constants.UpdateFlags & PREPARE_UPDATE_FLAG
+#define ShouldPerformUpdate Constants.UpdateFlags & PERFORM_UPDATE_FLAG
+
 #ifdef HLSL
 globallycoherent RWByteAddressBuffer outputBVH : UAV_REGISTER(OutputBVHRegister);
 RWByteAddressBuffer scratchMemory : UAV_REGISTER(ScratchBufferRegister);
 RWByteAddressBuffer childNodesProcessedCounter : UAV_REGISTER(ChildNodesProcessedBufferRegister);
-RWStructuredBuffer<uint> mortonCodes : UAV_REGISTER(MortonCodesBufferRegister);
 RWStructuredBuffer<HierarchyNode> hierarchyBuffer : UAV_REGISTER(HierarchyBufferRegister);
+RWStructuredBuffer<uint> aabbParentBuffer : UAV_REGISTER(AABBParentBufferRegister);
 RWByteAddressBuffer DescriptorHeapBufferTable[] : UAV_REGISTER_SPACE(GlobalDescriptorHeapRegister, GlobalDescriptorHeapRegisterSpace);
 
 cbuffer ConstructHierarchyConstants : CONSTANT_REGISTER(InputConstantsRegister)

diff --git a/Libraries/D3D12RaytracingFallback/src/ConstructAABBPass.cpp b/Libraries/D3D12RaytracingFallback/src/ConstructAABBPass.cpp
@@ -24,8 +24,8 @@ namespace FallbackLayer
         rootParameters[OutputBVHRootUAVParam].InitAsUnorderedAccessView(OutputBVHRegister);
         rootParameters[ScratchUAVParam].InitAsUnorderedAccessView(ScratchBufferRegister);
         rootParameters[HierarchyUAVParam].InitAsUnorderedAccessView(HierarchyBufferRegister);
+        rootParameters[AABBParentBufferParam].InitAsUnorderedAccessView(AABBParentBufferRegister);
         rootParameters[ChildNodesProcessedCountBufferParam].InitAsUnorderedAccessView(ChildNodesProcessedBufferRegister);
-        rootParameters[MortonCodesBufferParam].InitAsUnorderedAccessView(MortonCodesBufferRegister);
         rootParameters[InputRootConstants].InitAsConstants(SizeOfInUint32(InputConstants), InputConstantsRegister);
         rootParameters[GlobalDescriptorHeap].InitAsDescriptorTable(1, &globalDescriptorHeapRange);
 
@@ -44,15 +44,19 @@ namespace FallbackLayer
         D3D12_GPU_VIRTUAL_ADDRESS outputVH,
         D3D12_GPU_VIRTUAL_ADDRESS scratchBuffer,
         D3D12_GPU_VIRTUAL_ADDRESS childNodesProcessedCountBuffer,
-        D3D12_GPU_VIRTUAL_ADDRESS mortonCodeBuffer,
         D3D12_GPU_VIRTUAL_ADDRESS hierarchyBuffer,
+        D3D12_GPU_VIRTUAL_ADDRESS outputAABBParentBuffer,
         D3D12_GPU_DESCRIPTOR_HANDLE globalDescriptorHeap,
+        const bool prepareUpdate,
+        const bool performUpdate,
         UINT numElements)
     {
         bool isEmptyAccelerationStructure = numElements == 0;
         Level level = (sceneType == SceneType::Triangles) ? Level::Bottom : Level::Top;
 
-        InputConstants constants = { numElements };
+        InputConstants constants = {};
+        constants.NumberOfElements = numElements;
+        constants.UpdateFlags = ((UINT) prepareUpdate) | (performUpdate << 1);
 
         pCommandList->SetComputeRootSignature(m_pRootSignature);
         pCommandList->SetComputeRoot32BitConstants(InputRootConstants, SizeOfInUint32(InputConstants), &constants, 0);
@@ -61,10 +65,14 @@ namespace FallbackLayer
         {
             pCommandList->SetComputeRootUnorderedAccessView(ScratchUAVParam, scratchBuffer);
             pCommandList->SetComputeRootUnorderedAccessView(ChildNodesProcessedCountBufferParam, childNodesProcessedCountBuffer);
-            pCommandList->SetComputeRootUnorderedAccessView(MortonCodesBufferParam, mortonCodeBuffer);
             pCommandList->SetComputeRootUnorderedAccessView(HierarchyUAVParam, hierarchyBuffer);
         }
 
+        if (prepareUpdate || performUpdate)
+        {
+            pCommandList->SetComputeRootUnorderedAccessView(AABBParentBufferParam, outputAABBParentBuffer);
+        }
+
         if (level == Top)
         {
             pCommandList->SetComputeRootDescriptorTable(GlobalDescriptorHeap, globalDescriptorHeap);

diff --git a/Libraries/D3D12RaytracingFallback/src/ConstructAABBPass.h b/Libraries/D3D12RaytracingFallback/src/ConstructAABBPass.h
@@ -21,18 +21,20 @@ namespace FallbackLayer
             D3D12_GPU_VIRTUAL_ADDRESS outputVH,
             D3D12_GPU_VIRTUAL_ADDRESS scratchBuffer,
             D3D12_GPU_VIRTUAL_ADDRESS childNodesProcessedCountBuffer,
-            D3D12_GPU_VIRTUAL_ADDRESS mortonCodeBuffer,
             D3D12_GPU_VIRTUAL_ADDRESS hierarchyBuffer,
+            D3D12_GPU_VIRTUAL_ADDRESS outputAABBParentBuffer,
             D3D12_GPU_DESCRIPTOR_HANDLE globalDescriptorHeap,
+            const bool prepareUpdate,
+            const bool performUpdate,
             UINT numElements);
     private:
         enum RootParameterSlot
         {
             OutputBVHRootUAVParam = 0,
             ScratchUAVParam,
             HierarchyUAVParam,
+            AABBParentBufferParam,
             ChildNodesProcessedCountBufferParam,
-            MortonCodesBufferParam,
             InputRootConstants,
             GlobalDescriptorHeap,
             NumRootParameters,

diff --git a/Libraries/D3D12RaytracingFallback/src/CpuBVH2Builder.cpp b/Libraries/D3D12RaytracingFallback/src/CpuBVH2Builder.cpp
@@ -623,6 +623,7 @@ namespace FallbackLayer
                 PrimitiveMetaData metadata;
                 metadata.GeometryContributionToHitGroupIndex = i;
                 metadata.PrimitiveIndex = triangleIndex;
+                metadata.GeometryFlags = geometry.Flags;
                 primitiveMetaData[triangleIndex] = metadata;
 
                 // Next triangle

diff --git a/Libraries/D3D12RaytracingFallback/src/FallbackLayer.cpp b/Libraries/D3D12RaytracingFallback/src/FallbackLayer.cpp
@@ -231,7 +231,7 @@ namespace FallbackLayer
                 {
                     if (range.RangeType != D3D12_DESCRIPTOR_RANGE_TYPE_SAMPLER)
                     {
-                        range.Flags = D3D12_DESCRIPTOR_RANGE_FLAG_DATA_VOLATILE;
+                        range.Flags = D3D12_DESCRIPTOR_RANGE_FLAG_DATA_VOLATILE | D3D12_DESCRIPTOR_RANGE_FLAG_DESCRIPTORS_VOLATILE;
                     }
                 }
 

diff --git a/Libraries/D3D12RaytracingFallback/src/FallbackLayer.h b/Libraries/D3D12RaytracingFallback/src/FallbackLayer.h
@@ -78,7 +78,7 @@ namespace FallbackLayer
         ID3D12DescriptorHeap *m_pBoundDescriptorHeaps[DescriptorHeapType::NumTypes] = {};
         ATL::CComPtr<ID3D12GraphicsCommandList> m_pCommandList;
         RaytracingDevice &m_device;
-        COM_IMPLEMENTATION();
+        COM_IMPLEMENTATION_WITH_QUERYINTERFACE(m_pCommandList.p);
     };
 
     class RaytracingStateObject : public ID3D12RaytracingFallbackStateObject
@@ -190,7 +190,7 @@ namespace FallbackLayer
         RaytracingProgramFactory m_RaytracingProgramFactory;
         DWORD m_flags;
 
-        COM_IMPLEMENTATION();
+        COM_IMPLEMENTATION_WITH_QUERYINTERFACE(m_pDevice.p)
 
 #if ENABLE_ACCELERATION_STRUCTURE_VISUALIZATION
     public:

diff --git a/Libraries/D3D12RaytracingFallback/src/FallbackLayerUnitTests/FallbackLayerUnitTests.vcxproj b/Libraries/D3D12RaytracingFallback/src/FallbackLayerUnitTests/FallbackLayerUnitTests.vcxproj
@@ -332,6 +332,12 @@
       <AdditionalOptions Condition="'$(Configuration)|$(Platform)'=='Release|Win32'"> -T lib_6_1 -HV 2017 -O4 %(AdditionalOptions)</AdditionalOptions>
       <AdditionalOptions Condition="'$(Configuration)|$(Platform)'=='Debug|x64'"> -T lib_6_1 -HV 2017 -O4 %(AdditionalOptions)</AdditionalOptions>
       <AdditionalOptions Condition="'$(Configuration)|$(Platform)'=='Release|x64'"> -T lib_6_1 -HV 2017 -O4 %(AdditionalOptions)</AdditionalOptions>
+      <VariableName Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">g_p%(Filename)</VariableName>
+      <VariableName Condition="'$(Configuration)|$(Platform)'=='Release|x64'">g_p%(Filename)</VariableName>
+      <HeaderFileOutput Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">$(SolutionDir)..\Build_VS15\$(Platform)\$(Configuration)\Output\$(ProjectName)\CompiledShaders\%(Filename).h</HeaderFileOutput>
+      <HeaderFileOutput Condition="'$(Configuration)|$(Platform)'=='Release|x64'">$(SolutionDir)..\Build_VS15\$(Platform)\$(Configuration)\Output\$(ProjectName)\CompiledShaders\%(Filename).h</HeaderFileOutput>
+      <ObjectFileOutput Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">$(SolutionDir)..\Build_VS15\$(Platform)\$(Configuration)\Output\$(ProjectName)\Shaders\%(Filename).cso</ObjectFileOutput>
+      <ObjectFileOutput Condition="'$(Configuration)|$(Platform)'=='Release|x64'">$(SolutionDir)..\Build_VS15\$(Platform)\$(Configuration)\Output\$(ProjectName)\Shaders\%(Filename).cso</ObjectFileOutput>
     </FxCompile>
     <FxCompile Include="SimpleRayTracing.hlsl">
       <ShaderType Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">

diff --git a/...D3D12RaytracingFallback/src/FallbackLayerUnitTests/FallbackLayerUnitTests.vcxproj.filters b/...D3D12RaytracingFallback/src/FallbackLayerUnitTests/FallbackLayerUnitTests.vcxproj.filters
@@ -204,17 +204,6 @@
     <FxCompile Include="ReadDataTexture1DArray.hlsl" />
     <FxCompile Include="ReadDataTexture2D.hlsl" />
     <FxCompile Include="ReadDataTexture2DArray.hlsl" />
-    <FxCompile Include="ReadDataRWBuffer.hlsl" />
-    <FxCompile Include="ReadDataRWByteAddressBuffer.hlsl" />
-    <FxCompile Include="ReadDataRWTexture1D.hlsl" />
-    <FxCompile Include="ReadDataRWTexture1DArray.hlsl" />
-    <FxCompile Include="ReadDataRWTexture2D.hlsl" />
-    <FxCompile Include="ReadDataRWTexture2DArray.hlsl" />
-    <FxCompile Include="ReadDataRWTexture3D.hlsl" />
-    <FxCompile Include="ReadDataTexture1D.hlsl" />
-    <FxCompile Include="ReadDataTexture1DArray.hlsl" />
-    <FxCompile Include="ReadDataTexture2D.hlsl" />
-    <FxCompile Include="ReadDataTexture2DArray.hlsl" />
     <FxCompile Include="ReadData*.hlsl">
       <Filter>Shaders</Filter>
     </FxCompile>
@@ -282,6 +271,17 @@
     <FxCompile Include="ReadData*.hlsl" />
     <FxCompile Include="ReadData*.hlsl" />
     <FxCompile Include="ReadData*.hlsl" />
+    <FxCompile Include="ReadData*.hlsl" />
+    <FxCompile Include="ReadData*.hlsl" />
+    <FxCompile Include="ReadData*.hlsl" />
+    <FxCompile Include="ReadData*.hlsl" />
+    <FxCompile Include="ReadData*.hlsl" />
+    <FxCompile Include="ReadData*.hlsl" />
+    <FxCompile Include="ReadData*.hlsl" />
+    <FxCompile Include="ReadData*.hlsl" />
+    <FxCompile Include="ReadData*.hlsl" />
+    <FxCompile Include="ReadData*.hlsl" />
+    <FxCompile Include="ReadData*.hlsl" />
   </ItemGroup>
   <ItemGroup>
     <None Include="SharedReadData.hlsli">