Resource usage (shader-slang#26)

* BufferUsage enum

* TextureUsage enum

* implement IDevice::getFormatSupport

* speed up format test

* refactor test-resource-state test

CHANGELOG

@@ -1,3 +1,6 @@
+- replace IDevice::getFormatSupportedResourceStates with IDevice::getFormatSupport
+- rename TextureDesc::allowedStates -> TextureDesc::usage and introduce TextureUsage enum
+- rename BufferDesc::allowedStates -> BufferDesc::usage and introduce BufferUsage enum
 - rename IQueryPool::Desc -> QueryPoolDesc, add label to QueryPoolDesc
 - rename IFence::Desc -> FenceDesc, add label to FenceDesc
 - implement dynamic rendering in Vulkan

YONGQUESTIONS.txt

@@ -0,0 +1,10 @@
+
+Pipeline refactor
+- can we remove runtime specialization? we have to!
+- specialize() doesn't lead to a program layout that can be used
+
+Shader program
+- can we always create shader program from a "linked program"
+- we cannot just have a single linked program passed in
+- d3d12 has support for "libraries" where we have individual entry points with different specializations
+- how should the interface to create shader programs look like?

include/slang-rhi.h

@@ -26,6 +26,18 @@
 #define SLANG_RHI_API
 #endif
 
+// clang-format off
+/// Implement logical operators on a class enum for making it usable as a flags enum.
+#define SLANG_RHI_ENUM_CLASS_OPERATORS(e_) \
+    inline e_ operator& (e_ a, e_ b) { return static_cast<e_>(static_cast<int>(a)& static_cast<int>(b)); } \
+    inline e_ operator| (e_ a, e_ b) { return static_cast<e_>(static_cast<int>(a)| static_cast<int>(b)); } \
+    inline e_& operator|= (e_& a, e_ b) { a = a | b; return a; }; \
+    inline e_& operator&= (e_& a, e_ b) { a = a & b; return a; }; \
+    inline e_  operator~ (e_ a) { return static_cast<e_>(~static_cast<int>(a)); } \
+    inline bool is_set(e_ val, e_ flag) { return (val & flag) != static_cast<e_>(0); } \
+    inline void flip_bit(e_& val, e_ flag) { val = is_set(val, flag) ? (val & (~flag)) : (val | flag); }
+// clang-format on
+
 // Needed for building on cygwin with gcc
 #undef Always
 #undef None
@@ -266,6 +278,27 @@ struct FormatInfo
     GfxCount blockHeight;
 };
 
+enum class FormatSupport
+{
+    None = 0x0,
+
+    Buffer = 0x1,
+    IndexBuffer = 0x2,
+    VertexBuffer = 0x4,
+
+    Texture = 0x8,
+    DepthStencil = 0x10,
+    RenderTarget = 0x20,
+    Blendable = 0x40,
+
+    ShaderLoad = 0x80,
+    ShaderSample = 0x100,
+    ShaderUavLoad = 0x200,
+    ShaderUavStore = 0x400,
+    ShaderAtomic = 0x800,
+};
+SLANG_RHI_ENUM_CLASS_OPERATORS(FormatSupport);
+
 enum class InputSlotClass
 {
     PerVertex,
@@ -313,42 +346,6 @@ enum class ResourceState
     AccelerationStructureBuildInput,
     PixelShaderResource,
     NonPixelShaderResource,
-    _Count
-};
-
-struct ResourceStateSet
-{
-public:
-    void add(ResourceState state) { m_bitFields |= (1LL << (uint32_t)state); }
-    template<typename... TResourceState>
-    void add(ResourceState s, TResourceState... states)
-    {
-        add(s);
-        add(states...);
-    }
-    bool contains(ResourceState state) const { return (m_bitFields & (1LL << (uint32_t)state)) != 0; }
-    ResourceStateSet()
-        : m_bitFields(0)
-    {
-    }
-    ResourceStateSet(const ResourceStateSet& other) = default;
-    ResourceStateSet(ResourceState state) { add(state); }
-    template<typename... TResourceState>
-    ResourceStateSet(TResourceState... states)
-    {
-        add(states...);
-    }
-
-    ResourceStateSet operator&(const ResourceStateSet& that) const
-    {
-        ResourceStateSet result;
-        result.m_bitFields = this->m_bitFields & that.m_bitFields;
-        return result;
-    }
-
-private:
-    uint64_t m_bitFields = 0;
-    void add() {}
 };
 
 /// Describes how memory for the resource should be allocated for CPU access.
@@ -467,6 +464,23 @@ struct MemoryRange
     uint64_t size;
 };
 
+enum class BufferUsage
+{
+    None = 0,
+    VertexBuffer = 0x1,
+    IndexBuffer = 0x2,
+    ConstantBuffer = 0x4,
+    ShaderResource = 0x8,
+    UnorderedAccess = 0x10,
+    IndirectArgument = 0x20,
+    CopySource = 0x40,
+    CopyDestination = 0x80,
+    AccelerationStructure = 0x100,
+    AccelerationStructureBuildInput = 0x200,
+    ShaderTable = 0x400,
+};
+SLANG_RHI_ENUM_CLASS_OPERATORS(BufferUsage);
+
 struct BufferDesc
 {
     /// Total size in bytes.
@@ -478,8 +492,8 @@ struct BufferDesc
 
     MemoryType memoryType = MemoryType::DeviceLocal;
 
+    BufferUsage usage = BufferUsage::None;
     ResourceState defaultState = ResourceState::Undefined;
-    ResourceStateSet allowedStates = ResourceStateSet();
 
     bool isShared = false;
 
@@ -522,6 +536,22 @@ struct BufferRange
     Size size;
 };
 
+enum class TextureUsage
+{
+    None = 0x0,
+    ShaderResource = 0x1,
+    UnorderedAccess = 0x2,
+    RenderTarget = 0x4,
+    DepthRead = 0x8,
+    DepthWrite = 0x10,
+    Present = 0x20,
+    CopySource = 0x40,
+    CopyDestination = 0x80,
+    ResolveSource = 0x100,
+    ResolveDestination = 0x200,
+};
+SLANG_RHI_ENUM_CLASS_OPERATORS(TextureUsage);
+
 enum class TextureType
 {
     Texture1D,   ///< A 1d texture
@@ -625,9 +655,11 @@ struct Extents
 struct TextureDesc
 {
     TextureType type = TextureType::Texture2D;
-    ResourceState defaultState = ResourceState::Undefined;
-    ResourceStateSet allowedStates = ResourceStateSet();
+
     MemoryType memoryType = MemoryType::DeviceLocal;
+
+    TextureUsage usage = TextureUsage::None;
+    ResourceState defaultState = ResourceState::Undefined;
     bool isShared = false;
 
     Extents size;
@@ -2122,8 +2154,7 @@ class IDevice : public ISlangUnknown
     virtual SLANG_NO_THROW Result SLANG_MCALL
     getFeatures(const char** outFeatures, Size bufferSize, GfxCount* outFeatureCount) = 0;
 
-    virtual SLANG_NO_THROW Result SLANG_MCALL
-    getFormatSupportedResourceStates(Format format, ResourceStateSet* outStates) = 0;
+    virtual SLANG_NO_THROW Result SLANG_MCALL getFormatSupport(Format format, FormatSupport* outFormatSupport) = 0;
 
     virtual SLANG_NO_THROW Result SLANG_MCALL getSlangSession(slang::ISession** outSlangSession) = 0;
 

src/cuda/cuda-device.cpp

@@ -700,7 +700,7 @@ DeviceImpl::createTexture(const TextureDesc& desc, const SubresourceData* initDa
         // time we create a resource, and then allocate the surface or
         // texture objects as part of view creation.
         //
-        if (desc.allowedStates.contains(ResourceState::UnorderedAccess))
+        if (is_set(desc.usage, TextureUsage::UnorderedAccess))
         {
             // On CUDA surfaces only support a single MIP map
             SLANG_RHI_ASSERT(desc.numMipLevels == 1);

src/d3d11/d3d11-device.cpp

@@ -496,7 +496,7 @@ Result DeviceImpl::createTexture(const TextureDesc& descIn, const SubresourceDat
         return SLANG_FAIL;
     }
 
-    const int bindFlags = _calcResourceBindFlags(srcDesc.allowedStates);
+    const int bindFlags = _calcResourceBindFlags(srcDesc.usage);
 
     // Set up the initialize data
     std::vector<D3D11_SUBRESOURCE_DATA> subRes;
@@ -609,7 +609,7 @@ Result DeviceImpl::createBuffer(const BufferDesc& descIn, const void* initData,
 {
     BufferDesc srcDesc = fixupBufferDesc(descIn);
 
-    auto d3dBindFlags = _calcResourceBindFlags(srcDesc.allowedStates);
+    auto d3dBindFlags = _calcResourceBindFlags(srcDesc.usage);
 
     size_t alignedSizeInBytes = srcDesc.size;
 
@@ -636,7 +636,7 @@ Result DeviceImpl::createBuffer(const BufferDesc& descIn, const void* initData,
     bufferDesc.Usage = D3D11_USAGE_DEFAULT;
 
     // If written by CPU, make it dynamic
-    if (descIn.memoryType == MemoryType::Upload && !descIn.allowedStates.contains(ResourceState::UnorderedAccess))
+    if (descIn.memoryType == MemoryType::Upload && !is_set(descIn.usage, BufferUsage::UnorderedAccess))
     {
         bufferDesc.Usage = D3D11_USAGE_DYNAMIC;
     }

src/d3d11/d3d11-helper-functions.cpp

@@ -18,44 +18,36 @@ bool isSupportedNVAPIOp(IUnknown* dev, uint32_t op)
 #endif
 }
 
-D3D11_BIND_FLAG calcResourceFlag(ResourceState state)
+int _calcResourceBindFlags(BufferUsage usage)
 {
-    switch (state)
-    {
-    case ResourceState::VertexBuffer:
-        return D3D11_BIND_VERTEX_BUFFER;
-    case ResourceState::IndexBuffer:
-        return D3D11_BIND_INDEX_BUFFER;
-    case ResourceState::ConstantBuffer:
-        return D3D11_BIND_CONSTANT_BUFFER;
-    case ResourceState::StreamOutput:
-        return D3D11_BIND_STREAM_OUTPUT;
-    case ResourceState::RenderTarget:
-        return D3D11_BIND_RENDER_TARGET;
-    case ResourceState::DepthRead:
-    case ResourceState::DepthWrite:
-        return D3D11_BIND_DEPTH_STENCIL;
-    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);
-    }
+    int flags = 0;
+    if (is_set(usage, BufferUsage::VertexBuffer))
+        flags |= D3D11_BIND_VERTEX_BUFFER;
+    if (is_set(usage, BufferUsage::IndexBuffer))
+        flags |= D3D11_BIND_INDEX_BUFFER;
+    if (is_set(usage, BufferUsage::ConstantBuffer))
+        flags |= D3D11_BIND_CONSTANT_BUFFER;
+    if (is_set(usage, BufferUsage::ShaderResource))
+        flags |= D3D11_BIND_SHADER_RESOURCE;
+    if (is_set(usage, BufferUsage::UnorderedAccess))
+        flags |= D3D11_BIND_UNORDERED_ACCESS;
+    return flags;
 }
 
-int _calcResourceBindFlags(ResourceStateSet allowedStates)
+int _calcResourceBindFlags(TextureUsage usage)
 {
-    int dstFlags = 0;
-    for (uint32_t i = 0; i < (uint32_t)ResourceState::_Count; i++)
-    {
-        auto state = (ResourceState)i;
-        if (allowedStates.contains(state))
-            dstFlags |= calcResourceFlag(state);
-    }
-    return dstFlags;
+    int flags = 0;
+    if (is_set(usage, TextureUsage::RenderTarget))
+        flags |= D3D11_BIND_RENDER_TARGET;
+    if (is_set(usage, TextureUsage::DepthRead))
+        flags |= D3D11_BIND_DEPTH_STENCIL;
+    if (is_set(usage, TextureUsage::DepthWrite))
+        flags |= D3D11_BIND_DEPTH_STENCIL;
+    if (is_set(usage, TextureUsage::ShaderResource))
+        flags |= D3D11_BIND_SHADER_RESOURCE;
+    if (is_set(usage, TextureUsage::UnorderedAccess))
+        flags |= D3D11_BIND_UNORDERED_ACCESS;
+    return flags;
 }
 
 int _calcResourceAccessFlags(MemoryType memType)

src/d3d11/d3d11-helper-functions.h

@@ -242,8 +242,8 @@ struct BindingOffset : SimpleBindingOffset
 
 bool isSupportedNVAPIOp(IUnknown* dev, uint32_t op);
 
-D3D11_BIND_FLAG calcResourceFlag(ResourceState state);
-int _calcResourceBindFlags(ResourceStateSet allowedStates);
+int _calcResourceBindFlags(BufferUsage usage);
+int _calcResourceBindFlags(TextureUsage usage);
 int _calcResourceAccessFlags(MemoryType memType);
 
 D3D11_FILTER_TYPE translateFilterMode(TextureFilteringMode mode);

src/d3d11/d3d11-shader-object.cpp

@@ -245,8 +245,8 @@ Result ShaderObjectImpl::_ensureOrdinaryDataBufferCreatedIfNeeded(
         ComPtr<IBuffer> buffer;
         BufferDesc bufferDesc = {};
         bufferDesc.size = specializedOrdinaryDataSize;
+        bufferDesc.usage = BufferUsage::ConstantBuffer | BufferUsage::CopyDestination;
         bufferDesc.defaultState = ResourceState::ConstantBuffer;
-        bufferDesc.allowedStates = ResourceStateSet(ResourceState::ConstantBuffer, ResourceState::CopyDestination);
         bufferDesc.memoryType = MemoryType::Upload;
         SLANG_RETURN_ON_FAIL(device->createBuffer(bufferDesc, nullptr, buffer.writeRef()));
         m_ordinaryDataBuffer = static_cast<BufferImpl*>(buffer.get());

src/d3d11/d3d11-swap-chain.cpp

@@ -27,9 +27,8 @@ void SwapchainImpl::createSwapchainBufferImages()
     imageDesc.size.height = m_desc.height;
     imageDesc.size.depth = 1;
     imageDesc.format = m_desc.format;
+    imageDesc.usage = TextureUsage::Present | TextureUsage::CopyDestination | TextureUsage::RenderTarget;
     imageDesc.defaultState = ResourceState::Present;
-    imageDesc.allowedStates =
-        ResourceStateSet(ResourceState::Present, ResourceState::CopyDestination, ResourceState::RenderTarget);
     RefPtr<TextureImpl> image = new TextureImpl(imageDesc);
     image->m_resource = d3dResource;
     for (GfxIndex i = 0; i < m_desc.imageCount; i++)