GFX: improve d3d12 performance. (#2408)

11 files changed, 92 insertions, 39 deletions

diff --git a/tools/gfx/d3d/d3d-util.cpp b/tools/gfx/d3d/d3d-util.cpp
index 942764d01..e50ddaa80 100644
--- a/tools/gfx/d3d/d3d-util.cpp
+++ b/tools/gfx/d3d/d3d-util.cpp
@@ -810,6 +810,10 @@ D3D12_RESOURCE_STATES D3DUtil::getResourceState(ResourceState state)
     case ResourceState::ShaderResource:
     case ResourceState::AccelerationStructureBuildInput:
         return D3D12_RESOURCE_STATE_PIXEL_SHADER_RESOURCE | D3D12_RESOURCE_STATE_NON_PIXEL_SHADER_RESOURCE;
+    case ResourceState::PixelShaderResource:
+        return D3D12_RESOURCE_STATE_PIXEL_SHADER_RESOURCE;
+    case ResourceState::NonPixelShaderResource:
+        return D3D12_RESOURCE_STATE_NON_PIXEL_SHADER_RESOURCE;
     case ResourceState::UnorderedAccess:
         return D3D12_RESOURCE_STATE_UNORDERED_ACCESS;
     case ResourceState::RenderTarget:
diff --git a/tools/gfx/d3d11/d3d11-helper-functions.cpp b/tools/gfx/d3d11/d3d11-helper-functions.cpp
index c4caab59b..c797979ce 100644
--- a/tools/gfx/d3d11/d3d11-helper-functions.cpp
+++ b/tools/gfx/d3d11/d3d11-helper-functions.cpp
@@ -43,6 +43,8 @@ namespace d3d11
         case ResourceState::UnorderedAccess:
             return D3D11_BIND_UNORDERED_ACCESS;
         case ResourceState::ShaderResource:
+        case ResourceState::PixelShaderResource:
+        case ResourceState::NonPixelShaderResource:
             return D3D11_BIND_SHADER_RESOURCE;
         default:
             return D3D11_BIND_FLAG(0);
diff --git a/tools/gfx/d3d12/d3d12-device.cpp b/tools/gfx/d3d12/d3d12-device.cpp
index f5c457809..efc1beb22 100644
--- a/tools/gfx/d3d12/d3d12-device.cpp
+++ b/tools/gfx/d3d12/d3d12-device.cpp
@@ -31,6 +31,8 @@ namespace d3d12
 
 using namespace Slang;
 
+static const uint32_t D3D_FEATURE_LEVEL_12_2 = 0xc200;
+
 Result DeviceImpl::createBuffer(
     const D3D12_RESOURCE_DESC& resourceDesc,
     const void* srcData,
@@ -416,6 +418,13 @@ Result DeviceImpl::initialize(const Desc& desc)
         return SLANG_FAIL;
     }
 
+    m_D3D12SerializeVersionedRootSignature =
+        (PFN_D3D12_SERIALIZE_VERSIONED_ROOT_SIGNATURE)loadProc(d3dModule, "D3D12SerializeVersionedRootSignature");
+    if (!m_D3D12SerializeVersionedRootSignature)
+    {
+        return SLANG_FAIL;
+    }
+
     HMODULE pixModule = LoadLibraryW(L"WinPixEventRuntime.dll");
     if (pixModule)
     {
@@ -472,21 +481,36 @@ Result DeviceImpl::initialize(const Desc& desc)
             DeviceCheckFlag::UseHardwareDevice,
             ChangeType::OnOff); ///< First try hardware, then reference
 
-        const D3D_FEATURE_LEVEL featureLevel = D3D_FEATURE_LEVEL_11_0;
 
-        const int numCombinations = combiner.getNumCombinations();
-        for (int i = 0; i < numCombinations; ++i)
+        const D3D_FEATURE_LEVEL featureLevels[] =
         {
-            if (SLANG_SUCCEEDED(_createDevice(
-                combiner.getCombination(i),
-                UnownedStringSlice(desc.adapter),
-                featureLevel,
-                m_deviceInfo)))
+            (D3D_FEATURE_LEVEL)D3D_FEATURE_LEVEL_12_2,
+            D3D_FEATURE_LEVEL_12_1,
+            D3D_FEATURE_LEVEL_12_0,
+            D3D_FEATURE_LEVEL_11_1,
+            D3D_FEATURE_LEVEL_11_0,
+            D3D_FEATURE_LEVEL_10_1,
+            D3D_FEATURE_LEVEL_10_0,
+            D3D_FEATURE_LEVEL_9_3,
+            D3D_FEATURE_LEVEL_9_2,
+            D3D_FEATURE_LEVEL_9_1
+        };
+        for (auto featureLevel : featureLevels)
+        {
+            const int numCombinations = combiner.getNumCombinations();
+            for (int i = 0; i < numCombinations; ++i)
             {
-                break;
+                if (SLANG_SUCCEEDED(_createDevice(
+                    combiner.getCombination(i),
+                    UnownedStringSlice(desc.adapter),
+                    featureLevel,
+                    m_deviceInfo)))
+                {
+                    goto succ;
+                }
             }
         }
-
+        succ:
         if (!m_deviceInfo.m_adapter)
         {
             // Couldn't find an adapter
diff --git a/tools/gfx/d3d12/d3d12-device.h b/tools/gfx/d3d12/d3d12-device.h
index 71daf9ffb..e02005ffd 100644
--- a/tools/gfx/d3d12/d3d12-device.h
+++ b/tools/gfx/d3d12/d3d12-device.h
@@ -78,7 +78,7 @@ public:
     PFN_D3D12_GET_DEBUG_INTERFACE m_D3D12GetDebugInterface = nullptr;
     PFN_D3D12_CREATE_DEVICE m_D3D12CreateDevice = nullptr;
     PFN_D3D12_SERIALIZE_ROOT_SIGNATURE m_D3D12SerializeRootSignature = nullptr;
-
+    PFN_D3D12_SERIALIZE_VERSIONED_ROOT_SIGNATURE m_D3D12SerializeVersionedRootSignature = nullptr;
     PFN_BeginEventOnCommandList m_BeginEventOnCommandList = nullptr;
     PFN_EndEventOnCommandList m_EndEventOnCommandList = nullptr;
 
diff --git a/tools/gfx/d3d12/d3d12-pipeline-state.cpp b/tools/gfx/d3d12/d3d12-pipeline-state.cpp
index 983348409..ec073bf44 100644
--- a/tools/gfx/d3d12/d3d12-pipeline-state.cpp
+++ b/tools/gfx/d3d12/d3d12-pipeline-state.cpp
@@ -330,7 +330,6 @@ Result RayTracingPipelineStateImpl::ensureAPIPipelineStateCreated()
     ChunkedList<ComPtr<ISlangBlob>> codeBlobs;
     ChunkedList<D3D12_EXPORT_DESC> exports;
     ChunkedList<const wchar_t*> strPtrs;
-
     ComPtr<ISlangBlob> diagnostics;
     ChunkedList<OSString> stringPool;
     auto getWStr = [&](const char* name)
@@ -339,6 +338,19 @@ Result RayTracingPipelineStateImpl::ensureAPIPipelineStateCreated()
         auto wstr = str.toWString();
         return stringPool.add(wstr)->begin();
     };
+
+    D3D12_RAYTRACING_PIPELINE_CONFIG1 pipelineConfig = {};
+    pipelineConfig.MaxTraceRecursionDepth = desc.rayTracing.maxRecursion;
+    if (desc.rayTracing.flags & RayTracingPipelineFlags::SkipTriangles)
+        pipelineConfig.Flags |= D3D12_RAYTRACING_PIPELINE_FLAG_SKIP_TRIANGLES;
+    if (desc.rayTracing.flags & RayTracingPipelineFlags::SkipProcedurals)
+        pipelineConfig.Flags |= D3D12_RAYTRACING_PIPELINE_FLAG_SKIP_PROCEDURAL_PRIMITIVES;
+
+    D3D12_STATE_SUBOBJECT pipelineConfigSubobject = {};
+    pipelineConfigSubobject.Type = D3D12_STATE_SUBOBJECT_TYPE_RAYTRACING_PIPELINE_CONFIG1;
+    pipelineConfigSubobject.pDesc = &pipelineConfig;
+    subObjects.add(pipelineConfigSubobject);
+
     auto compileShader = [&](slang::EntryPointLayout* entryPointInfo,
         slang::IComponentType* component,
         SlangInt entryPointIndex)
@@ -361,7 +373,7 @@ Result RayTracingPipelineStateImpl::ensureAPIPipelineStateCreated()
         library.NumExports = 1;
         D3D12_EXPORT_DESC exportDesc = {};
         exportDesc.Name = getWStr(entryPointInfo->getNameOverride());
-        exportDesc.ExportToRename = getWStr(entryPointInfo->getNameOverride());
+        exportDesc.ExportToRename = nullptr;
         exportDesc.Flags = D3D12_EXPORT_FLAG_NONE;
         library.pExports = exports.add(exportDesc);
 
@@ -435,13 +447,6 @@ Result RayTracingPipelineStateImpl::ensureAPIPipelineStateCreated()
     globalSignatureSubobject.pDesc = &globalSignatureDesc;
     subObjects.add(globalSignatureSubobject);
 
-    D3D12_RAYTRACING_PIPELINE_CONFIG pipelineConfig = {};
-    pipelineConfig.MaxTraceRecursionDepth = desc.rayTracing.maxRecursion;
-    D3D12_STATE_SUBOBJECT pipelineConfigSubobject = {};
-    pipelineConfigSubobject.Type = D3D12_STATE_SUBOBJECT_TYPE_RAYTRACING_PIPELINE_CONFIG;
-    pipelineConfigSubobject.pDesc = &pipelineConfig;
-    subObjects.add(pipelineConfigSubobject);
-
     if (m_device->m_pipelineCreationAPIDispatcher)
     {
         m_device->m_pipelineCreationAPIDispatcher->beforeCreateRayTracingState(
diff --git a/tools/gfx/d3d12/d3d12-shader-object-layout.cpp b/tools/gfx/d3d12/d3d12-shader-object-layout.cpp
index 3e8598d51..d90b638ba 100644
--- a/tools/gfx/d3d12/d3d12-shader-object-layout.cpp
+++ b/tools/gfx/d3d12/d3d12-shader-object-layout.cpp
@@ -500,7 +500,7 @@ Result RootShaderObjectLayoutImpl::RootSignatureDescBuilder::addDescriptorRange(
 {
     if (isRootParameter)
     {
-        D3D12_ROOT_PARAMETER rootParam = {};
+        D3D12_ROOT_PARAMETER1 rootParam = {};
         switch (rangeType)
         {
         case D3D12_DESCRIPTOR_RANGE_TYPE_SRV:
@@ -525,7 +525,7 @@ Result RootShaderObjectLayoutImpl::RootSignatureDescBuilder::addDescriptorRange(
 
     auto& descriptorSet = m_descriptorSets[physicalDescriptorSetIndex];
 
-    D3D12_DESCRIPTOR_RANGE range = {};
+    D3D12_DESCRIPTOR_RANGE1 range = {};
     range.RangeType = rangeType;
     range.NumDescriptors = count;
     range.BaseShaderRegister = registerIndex;
@@ -824,14 +824,14 @@ void RootShaderObjectLayoutImpl::RootSignatureDescBuilder::addAsValue(
     }
 }
 
-D3D12_ROOT_SIGNATURE_DESC& RootShaderObjectLayoutImpl::RootSignatureDescBuilder::build()
+D3D12_ROOT_SIGNATURE_DESC1& RootShaderObjectLayoutImpl::RootSignatureDescBuilder::build()
 {
     for (Index i = 0; i < m_descriptorSets.getCount(); i++)
     {
         auto& descriptorSet = m_descriptorSets[i];
         if (descriptorSet.m_resourceRanges.getCount())
         {
-            D3D12_ROOT_PARAMETER rootParam = {};
+            D3D12_ROOT_PARAMETER1 rootParam = {};
             rootParam.ParameterType = D3D12_ROOT_PARAMETER_TYPE_DESCRIPTOR_TABLE;
             rootParam.DescriptorTable.NumDescriptorRanges =
                 (UINT)descriptorSet.m_resourceRanges.getCount();
@@ -841,7 +841,7 @@ D3D12_ROOT_SIGNATURE_DESC& RootShaderObjectLayoutImpl::RootSignatureDescBuilder:
         }
         if (descriptorSet.m_samplerRanges.getCount())
         {
-            D3D12_ROOT_PARAMETER rootParam = {};
+            D3D12_ROOT_PARAMETER1 rootParam = {};
             rootParam.ParameterType = D3D12_ROOT_PARAMETER_TYPE_DESCRIPTOR_TABLE;
             rootParam.DescriptorTable.NumDescriptorRanges =
                 (UINT)descriptorSet.m_samplerRanges.getCount();
@@ -915,12 +915,13 @@ Result RootShaderObjectLayoutImpl::createRootSignatureFromSlang(
     }
 
     auto& rootSignatureDesc = builder.build();
-
+    D3D12_VERSIONED_ROOT_SIGNATURE_DESC versionedDesc = {};
+    versionedDesc.Version = D3D_ROOT_SIGNATURE_VERSION_1_1;
+    versionedDesc.Desc_1_1 = rootSignatureDesc;
     ComPtr<ID3DBlob> signature;
     ComPtr<ID3DBlob> error;
-    if (SLANG_FAILED(device->m_D3D12SerializeRootSignature(
-        &rootSignatureDesc,
-        D3D_ROOT_SIGNATURE_VERSION_1,
+    if (SLANG_FAILED(device->m_D3D12SerializeVersionedRootSignature(
+        &versionedDesc,
         signature.writeRef(),
         error.writeRef())))
     {
diff --git a/tools/gfx/d3d12/d3d12-shader-object-layout.h b/tools/gfx/d3d12/d3d12-shader-object-layout.h
index d70b194c8..8b72067f7 100644
--- a/tools/gfx/d3d12/d3d12-shader-object-layout.h
+++ b/tools/gfx/d3d12/d3d12-shader-object-layout.h
@@ -257,8 +257,8 @@ public:
 
     struct DescriptorSetLayout
     {
-        List<D3D12_DESCRIPTOR_RANGE> m_resourceRanges;
-        List<D3D12_DESCRIPTOR_RANGE> m_samplerRanges;
+        List<D3D12_DESCRIPTOR_RANGE1> m_resourceRanges;
+        List<D3D12_DESCRIPTOR_RANGE1> m_samplerRanges;
         uint32_t m_resourceCount = 0;
         uint32_t m_samplerCount = 0;
     };
@@ -275,8 +275,10 @@ public:
         // descriptor set for each `ParameterBlock` binding range in the shader object
         // hierarchy, regardless of the shader's `space` indices.
         List<DescriptorSetLayout> m_descriptorSets;
-        List<D3D12_ROOT_PARAMETER> m_rootParameters;
-        D3D12_ROOT_SIGNATURE_DESC m_rootSignatureDesc = {};
+        List<D3D12_ROOT_PARAMETER1> m_rootParameters;
+        List<D3D12_ROOT_PARAMETER1> m_rootDescTableParameters;
+
+        D3D12_ROOT_SIGNATURE_DESC1 m_rootSignatureDesc = {};
 
         static Result translateDescriptorRangeType(
             slang::BindingType c, D3D12_DESCRIPTOR_RANGE_TYPE* outType);
@@ -449,7 +451,7 @@ public:
             BindingRegisterOffsetPair const& containerOffset,
             BindingRegisterOffsetPair const& elementOffset);
 
-        D3D12_ROOT_SIGNATURE_DESC& build();
+        D3D12_ROOT_SIGNATURE_DESC1& build();
     };
 
     static Result createRootSignatureFromSlang(
diff --git a/tools/gfx/d3d12/d3d12-shader-table.cpp b/tools/gfx/d3d12/d3d12-shader-table.cpp
index 2773578b8..3e49350ab 100644
--- a/tools/gfx/d3d12/d3d12-shader-table.cpp
+++ b/tools/gfx/d3d12/d3d12-shader-table.cpp
@@ -31,6 +31,7 @@ RefPtr<BufferResource> ShaderTableImpl::createDeviceBuffer(
     IBufferResource::Desc bufferDesc = {};
     bufferDesc.memoryType = gfx::MemoryType::DeviceLocal;
     bufferDesc.defaultState = ResourceState::General;
+    bufferDesc.allowedStates.add(ResourceState::NonPixelShaderResource);
     bufferDesc.type = IResource::Type::Buffer;
     bufferDesc.sizeInBytes = tableSize;
     m_device->createBufferResource(bufferDesc, nullptr, bufferResource.writeRef());
@@ -57,10 +58,6 @@ RefPtr<BufferResource> ShaderTableImpl::createDeviceBuffer(
             void* shaderId = stateObjectProperties->GetShaderIdentifier(name.toWString().begin());
             memcpy(dest, shaderId, D3D12_SHADER_IDENTIFIER_SIZE_IN_BYTES);
         }
-        else
-        {
-            memset(dest, 0, D3D12_SHADER_IDENTIFIER_SIZE_IN_BYTES);
-        }
         if (overwrite.size)
         {
             memcpy((uint8_t*)dest + overwrite.offset, overwrite.data, overwrite.size);
@@ -68,6 +65,8 @@ RefPtr<BufferResource> ShaderTableImpl::createDeviceBuffer(
     };
 
     uint8_t* stagingBufferPtr = (uint8_t*)stagingPtr + stagingBufferOffset;
+    memset(stagingBufferPtr, 0, tableSize);
+
     for (uint32_t i = 0; i < m_rayGenShaderCount; i++)
     {
         copyShaderIdInto(
@@ -96,7 +95,7 @@ RefPtr<BufferResource> ShaderTableImpl::createDeviceBuffer(
         1,
         bufferResource.readRef(),
         gfx::ResourceState::CopyDestination,
-        gfx::ResourceState::ShaderResource);
+        gfx::ResourceState::NonPixelShaderResource);
     RefPtr<BufferResource> resultPtr = static_cast<BufferResource*>(bufferResource.get());
     return _Move(resultPtr);
 }
diff --git a/tools/gfx/renderer-shared.cpp b/tools/gfx/renderer-shared.cpp
index 69046d216..43629a0f2 100644
--- a/tools/gfx/renderer-shared.cpp
+++ b/tools/gfx/renderer-shared.cpp
@@ -386,6 +386,8 @@ Result RendererBase::getFormatSupportedResourceStates(Format format, ResourceSta
     outStates->add(ResourceState::ResolveDestination);
     outStates->add(ResourceState::ResolveSource);
     outStates->add(ResourceState::ShaderResource);
+    outStates->add(ResourceState::PixelShaderResource);
+    outStates->add(ResourceState::NonPixelShaderResource);
     outStates->add(ResourceState::StreamOutput);
     outStates->add(ResourceState::Undefined);
     outStates->add(ResourceState::UnorderedAccess);
diff --git a/tools/gfx/vulkan/vk-helper-functions.cpp b/tools/gfx/vulkan/vk-helper-functions.cpp
index 033096f8a..aa6c42ec5 100644
--- a/tools/gfx/vulkan/vk-helper-functions.cpp
+++ b/tools/gfx/vulkan/vk-helper-functions.cpp
@@ -83,6 +83,8 @@ VkImageLayout translateImageLayout(ResourceState state)
     case ResourceState::DepthWrite:
         return VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL;
     case ResourceState::ShaderResource:
+    case ResourceState::NonPixelShaderResource:
+    case ResourceState::PixelShaderResource:
         return VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
     case ResourceState::ResolveDestination:
     case ResourceState::CopyDestination:
@@ -116,6 +118,8 @@ VkAccessFlagBits calcAccessFlags(ResourceState state)
         return VkAccessFlagBits(
             VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT | VK_ACCESS_COLOR_ATTACHMENT_READ_BIT);
     case ResourceState::ShaderResource:
+    case ResourceState::NonPixelShaderResource:
+    case ResourceState::PixelShaderResource:
         return VK_ACCESS_INPUT_ATTACHMENT_READ_BIT;
     case ResourceState::UnorderedAccess:
         return VkAccessFlagBits(VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT);
@@ -167,6 +171,8 @@ VkPipelineStageFlagBits calcPipelineStageFlags(ResourceState state, bool src)
             VK_PIPELINE_STAGE_GEOMETRY_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT |
             VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT | VK_PIPELINE_STAGE_RAY_TRACING_SHADER_BIT_KHR);
     case ResourceState::ShaderResource:
+    case ResourceState::NonPixelShaderResource:
+    case ResourceState::PixelShaderResource:
         return VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT;
     case ResourceState::RenderTarget:
         return VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
@@ -237,6 +243,8 @@ VkBufferUsageFlagBits _calcBufferUsageFlags(ResourceState state)
         return (
             VkBufferUsageFlagBits)(VK_BUFFER_USAGE_STORAGE_TEXEL_BUFFER_BIT | VK_BUFFER_USAGE_STORAGE_BUFFER_BIT);
     case ResourceState::ShaderResource:
+    case ResourceState::NonPixelShaderResource:
+    case ResourceState::PixelShaderResource:
         return (
             VkBufferUsageFlagBits)(VK_BUFFER_USAGE_UNIFORM_TEXEL_BUFFER_BIT | VK_BUFFER_USAGE_STORAGE_BUFFER_BIT);
     case ResourceState::CopySource:
@@ -277,6 +285,8 @@ VkImageUsageFlagBits _calcImageUsageFlags(ResourceState state)
     case ResourceState::DepthRead:
         return VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT;
     case ResourceState::ShaderResource:
+    case ResourceState::NonPixelShaderResource:
+    case ResourceState::PixelShaderResource:
         return VK_IMAGE_USAGE_SAMPLED_BIT;
     case ResourceState::UnorderedAccess:
         return VK_IMAGE_USAGE_STORAGE_BIT;
diff --git a/tools/gfx/vulkan/vk-util.cpp b/tools/gfx/vulkan/vk-util.cpp
index 251e91cab..930e70439 100644
--- a/tools/gfx/vulkan/vk-util.cpp
+++ b/tools/gfx/vulkan/vk-util.cpp
@@ -201,6 +201,8 @@ VkImageLayout VulkanUtil::getImageLayoutFromState(ResourceState state)
     switch (state)
     {
     case ResourceState::ShaderResource:
+    case ResourceState::PixelShaderResource:
+    case ResourceState::NonPixelShaderResource:
         return VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
     case ResourceState::UnorderedAccess:
     case ResourceState::General:
@@ -555,6 +557,8 @@ VkImageLayout VulkanUtil::mapResourceStateToLayout(ResourceState state)
     case ResourceState::Undefined:
         return VK_IMAGE_LAYOUT_UNDEFINED;
     case ResourceState::ShaderResource:
+    case ResourceState::PixelShaderResource:
+    case ResourceState::NonPixelShaderResource:
         return VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
     case ResourceState::UnorderedAccess:
         return VK_IMAGE_LAYOUT_GENERAL;