Use GLSL scalar layout for constant buffers. (#2147)

7 files changed, 94 insertions, 9 deletions

diff --git a/tools/gfx/d3d/d3d-util.cpp b/tools/gfx/d3d/d3d-util.cpp
index 5d7aa4dd8..942764d01 100644
--- a/tools/gfx/d3d/d3d-util.cpp
+++ b/tools/gfx/d3d/d3d-util.cpp
@@ -808,6 +808,7 @@ D3D12_RESOURCE_STATES D3DUtil::getResourceState(ResourceState state)
     case ResourceState::StreamOutput:
         return D3D12_RESOURCE_STATE_STREAM_OUT;
     case ResourceState::ShaderResource:
+    case ResourceState::AccelerationStructureBuildInput:
         return D3D12_RESOURCE_STATE_PIXEL_SHADER_RESOURCE | D3D12_RESOURCE_STATE_NON_PIXEL_SHADER_RESOURCE;
     case ResourceState::UnorderedAccess:
         return D3D12_RESOURCE_STATE_UNORDERED_ACCESS;
diff --git a/tools/gfx/debug-layer.cpp b/tools/gfx/debug-layer.cpp
index 0926c5273..ec51dbee3 100644
--- a/tools/gfx/debug-layer.cpp
+++ b/tools/gfx/debug-layer.cpp
@@ -669,6 +669,42 @@ Result DebugDevice::createMutableRootShaderObject(
     return result;
 }
 
+Result DebugDevice::createShaderObjectFromTypeLayout(
+    slang::TypeLayoutReflection* typeLayout, IShaderObject** outShaderObject)
+{
+    SLANG_GFX_API_FUNC;
+
+    RefPtr<DebugShaderObject> outObject = new DebugShaderObject();
+    auto result = baseObject->createShaderObjectFromTypeLayout(typeLayout, outObject->baseObject.writeRef());
+    auto type = typeLayout->getType();
+    auto typeName = type->getName();
+    outObject->m_typeName = typeName;
+    outObject->m_device = this;
+    outObject->m_slangType = type;
+    if (SLANG_FAILED(result))
+        return result;
+    returnComPtr(outShaderObject, outObject);
+    return result;
+}
+
+Result DebugDevice::createMutableShaderObjectFromTypeLayout(
+    slang::TypeLayoutReflection* typeLayout, IShaderObject** outShaderObject)
+{
+    SLANG_GFX_API_FUNC;
+    RefPtr<DebugShaderObject> outObject = new DebugShaderObject();
+    auto result = baseObject->createMutableShaderObjectFromTypeLayout(
+        typeLayout, outObject->baseObject.writeRef());
+    if (SLANG_FAILED(result))
+        return result;
+    auto type = typeLayout->getType();
+    auto typeName = type->getName();
+    outObject->m_typeName = typeName;
+    outObject->m_device = this;
+    outObject->m_slangType = type;
+    returnComPtr(outShaderObject, outObject);
+    return result;
+}
+
 Result DebugDevice::createProgram(
     const IShaderProgram::Desc& desc, IShaderProgram** outProgram, ISlangBlob** outDiagnostics)
 {
diff --git a/tools/gfx/debug-layer.h b/tools/gfx/debug-layer.h
index a8cb0e849..5a139fd10 100644
--- a/tools/gfx/debug-layer.h
+++ b/tools/gfx/debug-layer.h
@@ -120,6 +120,10 @@ public:
         slang::TypeReflection* type,
         ShaderObjectContainerType container,
         IShaderObject** outObject) override;
+    virtual SLANG_NO_THROW Result SLANG_MCALL createShaderObjectFromTypeLayout(
+        slang::TypeLayoutReflection* typeLayout, IShaderObject** outObject) override;
+    virtual SLANG_NO_THROW Result SLANG_MCALL createMutableShaderObjectFromTypeLayout(
+        slang::TypeLayoutReflection* typeLayout, IShaderObject** outObject) override;
     virtual SLANG_NO_THROW Result SLANG_MCALL
         createMutableRootShaderObject(IShaderProgram* program, IShaderObject** outObject) override;
     virtual SLANG_NO_THROW Result SLANG_MCALL
diff --git a/tools/gfx/renderer-shared.cpp b/tools/gfx/renderer-shared.cpp
index 555186840..83fc9223d 100644
--- a/tools/gfx/renderer-shared.cpp
+++ b/tools/gfx/renderer-shared.cpp
@@ -366,6 +366,7 @@ Result RendererBase::getFormatSupportedResourceStates(Format format, ResourceSta
 {
     SLANG_UNUSED(format);
     outStates->add(ResourceState::AccelerationStructure);
+    outStates->add(ResourceState::AccelerationStructureBuildInput);
     outStates->add(ResourceState::ConstantBuffer);
     outStates->add(ResourceState::CopyDestination);
     outStates->add(ResourceState::CopySource);
@@ -459,6 +460,22 @@ SLANG_NO_THROW Result SLANG_MCALL RendererBase::createMutableShaderObject(
     return createMutableShaderObject(shaderObjectLayout, outObject);
 }
 
+SLANG_NO_THROW Result SLANG_MCALL RendererBase::createShaderObjectFromTypeLayout(
+    slang::TypeLayoutReflection* typeLayout, IShaderObject** outObject)
+{
+    RefPtr<ShaderObjectLayoutBase> shaderObjectLayout;
+    SLANG_RETURN_ON_FAIL(getShaderObjectLayout(typeLayout, shaderObjectLayout.writeRef()));
+    return createShaderObject(shaderObjectLayout, outObject);
+}
+
+SLANG_NO_THROW Result SLANG_MCALL RendererBase::createMutableShaderObjectFromTypeLayout(
+    slang::TypeLayoutReflection* typeLayout, IShaderObject** outObject)
+{
+    RefPtr<ShaderObjectLayoutBase> shaderObjectLayout;
+    SLANG_RETURN_ON_FAIL(getShaderObjectLayout(typeLayout, shaderObjectLayout.writeRef()));
+    return createMutableShaderObject(shaderObjectLayout, outObject);
+}
+
 Result RendererBase::getAccelerationStructurePrebuildInfo(
     const IAccelerationStructure::BuildInputs& buildInputs,
     IAccelerationStructure::PrebuildInfo* outPrebuildInfo)
@@ -531,7 +548,6 @@ Result RendererBase::getShaderObjectLayout(
     ShaderObjectContainerType container,
     ShaderObjectLayoutBase** outLayout)
 {
-    RefPtr<ShaderObjectLayoutBase> shaderObjectLayout;
     switch (container)
     {
     case ShaderObjectContainerType::StructuredBuffer:
@@ -544,11 +560,18 @@ Result RendererBase::getShaderObjectLayout(
         break;
     }
 
-    if( !m_shaderObjectLayoutCache.TryGetValue(type, shaderObjectLayout) )
+    auto typeLayout = slangContext.session->getTypeLayout(type);
+    return getShaderObjectLayout(typeLayout, outLayout);
+}
+
+Result RendererBase::getShaderObjectLayout(
+    slang::TypeLayoutReflection* typeLayout, ShaderObjectLayoutBase** outLayout)
+{
+    RefPtr<ShaderObjectLayoutBase> shaderObjectLayout;
+    if (!m_shaderObjectLayoutCache.TryGetValue(typeLayout, shaderObjectLayout))
     {
-        auto typeLayout = slangContext.session->getTypeLayout(type);
         SLANG_RETURN_ON_FAIL(createShaderObjectLayout(typeLayout, shaderObjectLayout.writeRef()));
-        m_shaderObjectLayoutCache.Add(type, shaderObjectLayout);
+        m_shaderObjectLayoutCache.Add(typeLayout, shaderObjectLayout);
     }
     *outLayout = shaderObjectLayout.detach();
     return SLANG_OK;
diff --git a/tools/gfx/renderer-shared.h b/tools/gfx/renderer-shared.h
index e82bc83d0..0462b802b 100644
--- a/tools/gfx/renderer-shared.h
+++ b/tools/gfx/renderer-shared.h
@@ -1240,6 +1240,12 @@ public:
         ShaderObjectContainerType containerType,
         IShaderObject** outObject) SLANG_OVERRIDE;
 
+    virtual SLANG_NO_THROW Result SLANG_MCALL createShaderObjectFromTypeLayout(
+        slang::TypeLayoutReflection* typeLayout, IShaderObject** outObject) override;
+
+    virtual SLANG_NO_THROW Result SLANG_MCALL createMutableShaderObjectFromTypeLayout(
+        slang::TypeLayoutReflection* typeLayout, IShaderObject** outObject) override;
+
     // Provides a default implementation that returns SLANG_E_NOT_AVAILABLE for platforms
     // without ray tracing support.
     virtual SLANG_NO_THROW Result SLANG_MCALL getAccelerationStructurePrebuildInfo(
@@ -1283,6 +1289,10 @@ public:
         ShaderObjectContainerType   container,
         ShaderObjectLayoutBase**    outLayout);
 
+    Result getShaderObjectLayout(
+        slang::TypeLayoutReflection* typeLayout,
+        ShaderObjectLayoutBase** outLayout);
+
 public:
     ExtendedShaderObjectTypeList specializationArgs;
     // Given current pipeline and root shader object binding, generate and bind a specialized pipeline if necessary.
@@ -1316,7 +1326,7 @@ public:
     SlangContext slangContext;
     ShaderCache shaderCache;
 
-    Slang::Dictionary<slang::TypeReflection*, Slang::RefPtr<ShaderObjectLayoutBase>> m_shaderObjectLayoutCache;
+    Slang::Dictionary<slang::TypeLayoutReflection*, Slang::RefPtr<ShaderObjectLayoutBase>> m_shaderObjectLayoutCache;
 };
 
 bool isDepthFormat(Format format);
diff --git a/tools/gfx/vulkan/render-vk.cpp b/tools/gfx/vulkan/render-vk.cpp
index 7cfdf06c8..8cec9d4a0 100644
--- a/tools/gfx/vulkan/render-vk.cpp
+++ b/tools/gfx/vulkan/render-vk.cpp
@@ -4252,6 +4252,8 @@ public:
                     return VkAccessFlagBits(
                         VK_ACCESS_ACCELERATION_STRUCTURE_READ_BIT_KHR |
                         VK_ACCESS_ACCELERATION_STRUCTURE_WRITE_BIT_KHR);
+                case ResourceState::AccelerationStructureBuildInput:
+                    return VkAccessFlagBits(VK_ACCESS_ACCELERATION_STRUCTURE_READ_BIT_KHR);
                 case ResourceState::General:
                     return VkAccessFlagBits(VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT);
                 default:
@@ -4314,6 +4316,8 @@ public:
                         VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT |
                         VK_PIPELINE_STAGE_RAY_TRACING_SHADER_BIT_KHR |
                         VK_PIPELINE_STAGE_ACCELERATION_STRUCTURE_BUILD_BIT_KHR);
+                case ResourceState::AccelerationStructureBuildInput:
+                    return VkPipelineStageFlagBits(VK_PIPELINE_STAGE_ACCELERATION_STRUCTURE_BUILD_BIT_KHR);
                 default:
                     assert(!"Unsupported");
                     return VkPipelineStageFlagBits(0);
@@ -4666,7 +4670,8 @@ public:
                             region.bufferImageHeight = 0;
 
                             region.imageSubresource.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
-                            region.imageSubresource.mipLevel = uint32_t(j);
+                            region.imageSubresource.mipLevel =
+                                subResourceRange.mipLevel + uint32_t(j);
                             region.imageSubresource.baseArrayLayer =
                                 subResourceRange.baseArrayLayer + i;
                             region.imageSubresource.layerCount = 1;
@@ -6393,7 +6398,7 @@ public:
                 m_desc.format = Format::B8G8R8A8_UNORM;
             }
 
-            SLANG_RETURN_ON_FAIL(createSwapchainAndImages());
+            createSwapchainAndImages();
             return SLANG_OK;
         }
 
@@ -7771,7 +7776,8 @@ static VkBufferUsageFlagBits _calcBufferUsageFlags(ResourceState state)
     case ResourceState::UnorderedAccess:
         return (VkBufferUsageFlagBits)(VK_BUFFER_USAGE_STORAGE_TEXEL_BUFFER_BIT | VK_BUFFER_USAGE_STORAGE_BUFFER_BIT);
     case ResourceState::ShaderResource:
-        return (VkBufferUsageFlagBits)(VK_BUFFER_USAGE_UNIFORM_TEXEL_BUFFER_BIT | VK_BUFFER_USAGE_STORAGE_BUFFER_BIT);
+        return (VkBufferUsageFlagBits)(VK_BUFFER_USAGE_UNIFORM_TEXEL_BUFFER_BIT |
+                                       VK_BUFFER_USAGE_STORAGE_BUFFER_BIT);
     case ResourceState::CopySource:
         return VK_BUFFER_USAGE_TRANSFER_SRC_BIT;
     case ResourceState::CopyDestination:
@@ -7780,6 +7786,8 @@ static VkBufferUsageFlagBits _calcBufferUsageFlags(ResourceState state)
         return VK_BUFFER_USAGE_ACCELERATION_STRUCTURE_STORAGE_BIT_KHR;
     case ResourceState::IndirectArgument:
         return VK_BUFFER_USAGE_INDIRECT_BUFFER_BIT;
+    case ResourceState::AccelerationStructureBuildInput:
+        return VK_BUFFER_USAGE_ACCELERATION_STRUCTURE_BUILD_INPUT_READ_ONLY_BIT_KHR;
     default:
         return VkBufferUsageFlagBits(0);
     }
@@ -8740,7 +8748,10 @@ Result VKDevice::getFormatSupportedResourceStates(Format format, ResourceStateSe
     }
     // AccelerationStructure
     if (bufferFeatures & VK_FORMAT_FEATURE_ACCELERATION_STRUCTURE_VERTEX_BUFFER_BIT_KHR)
+    {
         allowedStates.add(ResourceState::AccelerationStructure);
+        allowedStates.add(ResourceState::AccelerationStructureBuildInput);
+    }
 
     *outStates = allowedStates;
     return SLANG_OK;
diff --git a/tools/render-test/render-test-main.cpp b/tools/render-test/render-test-main.cpp
index a62cc7a9b..7046e2fda 100644
--- a/tools/render-test/render-test-main.cpp
+++ b/tools/render-test/render-test-main.cpp
@@ -804,7 +804,7 @@ void RenderTestApp::_initializeAccelerationStructure()
         instanceBufferDesc.type = IResource::Type::Buffer;
         instanceBufferDesc.sizeInBytes =
             instanceDescs.getCount() * sizeof(IAccelerationStructure::InstanceDesc);
-        instanceBufferDesc.defaultState = ResourceState::ShaderResource;
+        instanceBufferDesc.defaultState = ResourceState::AccelerationStructureBuildInput;
         ComPtr<IBufferResource> instanceBuffer =
             m_device->createBufferResource(instanceBufferDesc, instanceDescs.getBuffer());