Add explicit GPU resource transitions, memory and compute barriers

Source/Engine/GraphicsDevice/DirectX/DX11/GPUContextDX11.cpp

@@ -297,7 +297,7 @@ void GPUContextDX11::SetRenderTarget(GPUTextureView* depthBuffer, const Span<GPU
     __declspec(align(16)) ID3D11RenderTargetView* rtvs[GPU_MAX_RT_BINDED];
     for (int32 i = 0; i < rts.Length(); i++)
     {
-        auto rtDX11 = reinterpret_cast<GPUTextureViewDX11*>(rts[i]);
+        auto rtDX11 = reinterpret_cast<GPUTextureViewDX11*>(rts.Get()[i]);
         rtvs[i] = rtDX11 ? rtDX11->RTV() : nullptr;
     }
     int32 rtvsSize = sizeof(ID3D11RenderTargetView*) * rts.Length();
@@ -431,7 +431,7 @@ void GPUContextDX11::BindVB(const Span<GPUBuffer*>& vertexBuffers, const uint32*
     bool vbEdited = false;
     for (int32 i = 0; i < vertexBuffers.Length(); i++)
     {
-        const auto vbDX11 = static_cast<GPUBufferDX11*>(vertexBuffers[i]);
+        const auto vbDX11 = static_cast<GPUBufferDX11*>(vertexBuffers.Get()[i]);
         const auto vb = vbDX11 ? vbDX11->GetBuffer() : nullptr;
         vbEdited |= vb != _vbHandles[i];
         _vbHandles[i] = vb;

Source/Engine/GraphicsDevice/DirectX/DX12/GPUContextDX12.cpp

@@ -35,6 +35,7 @@
 #include "GPUShaderProgramDX12.h"
 #include "CommandSignatureDX12.h"
 #include "Engine/Profiler/RenderStats.h"
+#include "Engine/Graphics/GPUResourceAccess.h"
 #include "Engine/Graphics/Shaders/GPUShader.h"
 #include "Engine/Threading/Threading.h"
 
@@ -51,6 +52,47 @@ inline bool operator!=(const D3D12_INDEX_BUFFER_VIEW& l, const D3D12_INDEX_BUFFE
     return l.SizeInBytes != r.SizeInBytes || l.Format != r.Format || l.BufferLocation != r.BufferLocation;
 }
 
+FORCE_INLINE D3D12_RESOURCE_STATES GetResourceState(GPUResourceAccess access)
+{
+    switch (access)
+    {
+    case GPUResourceAccess::None:
+        return D3D12_RESOURCE_STATE_COMMON;
+    case GPUResourceAccess::CopyRead:
+        return D3D12_RESOURCE_STATE_COPY_SOURCE;
+    case GPUResourceAccess::CopyWrite:
+        return D3D12_RESOURCE_STATE_COPY_DEST;
+    case GPUResourceAccess::CpuRead:
+        return D3D12_RESOURCE_STATE_GENERIC_READ;
+    case GPUResourceAccess::CpuWrite:
+        return D3D12_RESOURCE_STATE_COMMON;
+    case GPUResourceAccess::DepthRead:
+        return D3D12_RESOURCE_STATE_DEPTH_READ;
+    case GPUResourceAccess::DepthWrite:
+        return D3D12_RESOURCE_STATE_DEPTH_WRITE;
+    case GPUResourceAccess::DepthBuffer:
+        return D3D12_RESOURCE_STATE_DEPTH_READ | D3D12_RESOURCE_STATE_DEPTH_WRITE;
+    case GPUResourceAccess::RenderTarget:
+        return D3D12_RESOURCE_STATE_RENDER_TARGET;
+    case GPUResourceAccess::UnorderedAccess:
+        return D3D12_RESOURCE_STATE_UNORDERED_ACCESS;
+    case GPUResourceAccess::IndirectArgs:
+        return D3D12_RESOURCE_STATE_INDIRECT_ARGUMENT;
+    case GPUResourceAccess::ShaderReadPixel:
+        //return D3D12_RESOURCE_STATE_PIXEL_SHADER_RESOURCE; // TODO: optimize SRV states in flushSRVs to be based on current binding usage slots
+    case GPUResourceAccess::ShaderReadCompute:
+    case GPUResourceAccess::ShaderReadNonPixel:
+        //return D3D12_RESOURCE_STATE_NON_PIXEL_SHADER_RESOURCE; // TODO: optimize SRV states in flushSRVs to be based on current binding usage slots
+    case GPUResourceAccess::ShaderReadGraphics:
+        return D3D12_RESOURCE_STATE_NON_PIXEL_SHADER_RESOURCE | D3D12_RESOURCE_STATE_PIXEL_SHADER_RESOURCE;
+#if !BUILD_RELEASE
+    default:
+        LOG(Error, "Unsupported GPU Resource Access: {}", (uint32)access);
+#endif
+    }
+    return D3D12_RESOURCE_STATE_COMMON;
+}
+
 // Ensure to match the indirect commands arguments layout
 static_assert(sizeof(GPUDispatchIndirectArgs) == sizeof(D3D12_DISPATCH_ARGUMENTS), "Wrong size of GPUDrawIndirectArgs.");
 static_assert(OFFSET_OF(GPUDispatchIndirectArgs, ThreadGroupCountX) == OFFSET_OF(D3D12_DISPATCH_ARGUMENTS, ThreadGroupCountX), "Wrong offset for GPUDrawIndirectArgs::ThreadGroupCountX");
@@ -1124,7 +1166,8 @@ void GPUContextDX12::Dispatch(GPUShaderProgramCS* shader, uint32 threadGroupCoun
     _psDirtyFlag = true;
 
     // Insert UAV barrier to ensure proper memory access for multiple sequential dispatches
-    AddUAVBarrier();
+    if (_pass == 0)
+        AddUAVBarrier();
 }
 
 void GPUContextDX12::DispatchIndirect(GPUShaderProgramCS* shader, GPUBuffer* bufferForArgs, uint32 offsetForArgs)
@@ -1158,7 +1201,8 @@ void GPUContextDX12::DispatchIndirect(GPUShaderProgramCS* shader, GPUBuffer* buf
     _psDirtyFlag = true;
 
     // Insert UAV barrier to ensure proper memory access for multiple sequential dispatches
-    AddUAVBarrier();
+    if (_pass == 0)
+        AddUAVBarrier();
 }
 
 void GPUContextDX12::ResolveMultisample(GPUTexture* sourceMultisampleTexture, GPUTexture* destTexture, int32 sourceSubResource, int32 destSubResource, PixelFormat format)
@@ -1549,4 +1593,15 @@ void GPUContextDX12::ForceRebindDescriptors()
     _commandList->SetDescriptorHeaps(ARRAY_COUNT(ppHeaps), ppHeaps);
 }
 
+void GPUContextDX12::Transition(GPUResource* resource, GPUResourceAccess access)
+{
+    SetResourceState(dynamic_cast<ResourceOwnerDX12*>(resource), GetResourceState(access));
+}
+
+void GPUContextDX12::OverlapUA(bool end)
+{
+    if (end)
+        AddUAVBarrier();
+}
+
 #endif

Source/Engine/GraphicsDevice/DirectX/DX12/GPUContextDX12.h

@@ -21,7 +21,7 @@ class GPUVertexLayoutDX12;
 /// <summary>
 /// Size of the resource barriers buffer size (will be flushed on overflow)
 /// </summary>
-#define DX12_RB_BUFFER_SIZE 16
+#define DX12_RB_BUFFER_SIZE 64
 
 /// <summary>
 /// GPU Commands Context implementation for DirectX 12
@@ -214,6 +214,8 @@ public:
     void CopySubresource(GPUResource* dstResource, uint32 dstSubresource, GPUResource* srcResource, uint32 srcSubresource) override;
     void SetResourceState(GPUResource* resource, uint64 state, int32 subresource) override;
     void ForceRebindDescriptors() override;
+    void Transition(GPUResource* resource, GPUResourceAccess access) override;
+    void OverlapUA(bool end) override;
 };
 
 #endif