Provide RT, SRV, UAV usage mask

Change-Id: I11c8c4de7ec39f145ab8cf859663673f38688e17

Author: fftzeng

IGC/Compiler/CISACodeGen/CShader.cpp

@@ -65,6 +65,9 @@ CShader::CShader(Function* pFunc, CShaderProgram* pProgram)
     m_shaderStats = nullptr;
     m_constantBufferMask = 0;
     m_constantBufferLoaded = 0;
+    m_uavLoaded = 0;
+    m_shaderResourceLoaded = 0;
+    m_renderTargetLoaded = 0;
     isInputsPulled = false;
     m_cbSlot = -1;
     m_statelessCBPushedSize = 0;
@@ -3096,11 +3099,26 @@ void CShader::SampleHeader(CVariable* payload, uint offset, uint writeMask, uint
     encoder.Push();
 }
 
+void CShader::UavAccesed(uint index)
+{
+    m_uavLoaded |= BIT(index);
+}
+
 void CShader::ConstantBufferAccesed(uint index)
 {
     m_constantBufferLoaded |= BIT(index);
 }
 
+void CShader::ShaderResourceAccesed(uint index)
+{
+    m_shaderResourceLoaded |= BIT(index);
+}
+
+void CShader::RenderTargetAccesed(uint index)
+{
+    m_renderTargetLoaded |= BIT(index);
+}
+
 // Helper function to copy raw register
 void CShader::CopyVariable(
     CVariable* dst,

IGC/Compiler/CISACodeGen/ComputeShaderCodeGen.cpp

@@ -481,6 +481,10 @@ namespace IGC
         CreateConstantBufferOutput(pKernelProgram);
 
         pKernelProgram->ConstantBufferLoaded = m_constantBufferLoaded;
+        pKernelProgram->UavLoaded = m_uavLoaded;
+        pKernelProgram->ShaderResourceLoaded = m_shaderResourceLoaded;
+        pKernelProgram->RenderTargetLoaded = m_renderTargetLoaded;
+
         pKernelProgram->hasControlFlow = m_numBlocks > 1 ? true : false;
 
         pKernelProgram->MaxNumberOfThreads = m_Platform->getMaxGPGPUShaderThreads();

IGC/Compiler/CISACodeGen/DomainShaderCodeGen.cpp

@@ -276,6 +276,10 @@ namespace IGC
         pKernelProgram->VertexURBEntryReadOffset = OctEltUnit(0);
         pKernelProgram->VertexURBEntryOutputReadOffset = GetURBHeaderSize();
         pKernelProgram->ConstantBufferLoaded = m_constantBufferLoaded;
+        pKernelProgram->UavLoaded = m_uavLoaded;
+        pKernelProgram->ShaderResourceLoaded = m_shaderResourceLoaded;
+        pKernelProgram->RenderTargetLoaded = m_renderTargetLoaded;
+
         pKernelProgram->DeclaresRTAIndex = m_properties.m_isRTAIndexDeclared;
         pKernelProgram->DeclaresVPAIndex = m_properties.m_isVPAIndexDeclared;
         pKernelProgram->HasClipCullAsOutput = m_properties.m_hasClipDistance;

IGC/Compiler/CISACodeGen/EmitVISAPass.cpp

@@ -4328,6 +4328,7 @@ void EmitPass::emitRenderTargetWrite(llvm::RTWritIntrinsic* inst, bool fromRet)
     if (RTIndex != -1)
     {
         bindingTableIndex = m_currShader->m_pBtiLayout->GetRenderTargetIndex(RTIndex);
+        m_currShader->RenderTargetAccesed(RTIndex);
     }
     else
     {
@@ -15198,17 +15199,20 @@ ResourceDescriptor EmitPass::GetResourceVariable(Value* resourcePtr)
             {
             case UAV:
                 bti = m_currShader->m_pBtiLayout->GetUavIndex(bufferIndex);
+                m_currShader->UavAccesed(bufferIndex);  // return to driver
                 break;
             case CONSTANT_BUFFER:
                 bti = m_currShader->m_pBtiLayout->GetConstantBufferIndex(bufferIndex);
                 m_currShader->ConstantBufferAccesed(bufferIndex);  // return to driver
                 break;
             case RESOURCE:
                 bti = m_currShader->m_pBtiLayout->GetTextureIndex(bufferIndex);
+                m_currShader->ShaderResourceAccesed(bufferIndex);  // return to driver
                 break;
             case RENDER_TARGET:
                 assert(m_currShader->GetShaderType() == ShaderType::PIXEL_SHADER);
                 bti = m_currShader->m_pBtiLayout->GetRenderTargetIndex(bufferIndex);
+                m_currShader->RenderTargetAccesed(bufferIndex);  // return to driver
                 break;
             case SLM:
                 bti = 254;  // \todo, remove hard-coding

IGC/Compiler/CISACodeGen/GeometryShaderCodeGen.cpp

@@ -350,6 +350,10 @@ namespace IGC
         CreateConstantBufferOutput(pKernelProgram);
 
         pKernelProgram->ConstantBufferLoaded = m_constantBufferLoaded;
+        pKernelProgram->UavLoaded = m_uavLoaded;
+        pKernelProgram->ShaderResourceLoaded = m_shaderResourceLoaded;
+        pKernelProgram->RenderTargetLoaded = m_renderTargetLoaded;
+
         pKernelProgram->NOSBufferSize = m_NOSBufferSize / getGRFSize(); // in 256 bits
 
         pKernelProgram->MaxNumberOfThreads = m_Platform->getMaxGeometryShaderThreads();

IGC/Compiler/CISACodeGen/HullShaderCodeGen.cpp

@@ -277,6 +277,10 @@ namespace IGC
         CreateConstantBufferOutput(pKernelProgram);
 
         pKernelProgram->ConstantBufferLoaded = m_constantBufferLoaded;
+        pKernelProgram->UavLoaded = m_uavLoaded;
+        pKernelProgram->ShaderResourceLoaded = m_shaderResourceLoaded;
+        pKernelProgram->RenderTargetLoaded = m_renderTargetLoaded;
+
         pKernelProgram->NOSBufferSize = m_NOSBufferSize / getGRFSize(); // in 256 bits
 
         pKernelProgram->MaxNumberOfThreads = m_Platform->getMaxHullShaderThreads();

IGC/Compiler/CISACodeGen/PixelShaderCodeGen.cpp

@@ -713,6 +713,10 @@ namespace IGC
         pKernelProgram->hasSampleOffset = m_SampleOffsetX || m_SampleOffsetY;
         pKernelProgram->hasZWDelta = m_ZWDelta;
         pKernelProgram->ConstantBufferLoaded = m_constantBufferLoaded;
+        pKernelProgram->UavLoaded = m_uavLoaded;
+        pKernelProgram->ShaderResourceLoaded = m_shaderResourceLoaded;
+        pKernelProgram->RenderTargetLoaded = m_renderTargetLoaded;
+
         pKernelProgram->hasControlFlow = m_numBlocks > 1 ? true : false;
         pKernelProgram->MaxNumberOfThreads = m_Platform->getMaxPixelShaderThreads();
         pKernelProgram->needPerspectiveBary = m_PerspectivePixel ? true : false;

IGC/Compiler/CISACodeGen/ShaderCodeGen.hpp

@@ -206,6 +206,9 @@ namespace IGC
         ShaderType  GetShaderType() const { return GetContext()->type; }
         bool        IsValueCoalesced(llvm::Value* v);
         void        ConstantBufferAccesed(uint index);
+        void        UavAccesed(uint index);
+        void        ShaderResourceAccesed(uint index);
+        void        RenderTargetAccesed(uint index);
 
         void        SampleHeader(CVariable* payload, uint offset, uint writeMask, uint rti);
 
@@ -456,6 +459,10 @@ namespace IGC
         uint m_ConstantBufferLength;
         uint m_constantBufferMask;
         uint m_constantBufferLoaded;
+        uint m_uavLoaded;
+        uint m_shaderResourceLoaded;
+        uint m_renderTargetLoaded;
+
         int  m_cbSlot;
         uint m_statelessCBPushedSize;
         uint m_NOSBufferSize;

IGC/Compiler/CISACodeGen/VertexShaderCodeGen.cpp

@@ -173,6 +173,10 @@ namespace IGC
         pKernelProgram->hasControlFlow = m_numBlocks > 1 ? true : false;
         pKernelProgram->MaxNumberOfThreads = m_Platform->getMaxVertexShaderThreads(isPositionOnlyShader);
         pKernelProgram->ConstantBufferLoaded = m_constantBufferLoaded;
+        pKernelProgram->UavLoaded = m_uavLoaded;
+        pKernelProgram->ShaderResourceLoaded = m_shaderResourceLoaded;
+        pKernelProgram->RenderTargetLoaded = m_renderTargetLoaded;
+
         pKernelProgram->hasVertexID = m_properties.m_HasVertexID;
         pKernelProgram->vertexIdLocation = m_properties.m_VID;
         pKernelProgram->hasInstanceID = m_properties.m_HasInstanceID;

IGC/Compiler/CodeGenPublic.h

@@ -262,6 +262,10 @@ namespace IGC
 
         unsigned int NOSBufferSize;
         unsigned int ConstantBufferLoaded;
+        unsigned int UavLoaded;
+        unsigned int ShaderResourceLoaded;
+        unsigned int RenderTargetLoaded;
+
         bool         hasControlFlow;
         unsigned int bufferSlot;
         unsigned int statelessCBPushedSize;