Fix calculation of URB read entry size and max number of input registers in pull model.

Author: mmerecki

IGC/Compiler/CISACodeGen/VertexShaderCodeGen.cpp

@@ -246,31 +246,33 @@ namespace IGC
         return m_R1;
     }
 
-
     /// Returns VS URB allocation size.
     /// This is the size of VS URB entry consisting of the header data and attribute data.
     OctEltUnit CVertexShader::GetURBAllocationSize() const
     {
-        const size_t payloadSize = m_Platform->getWATable().Wa_16011983264 ?
-            std::max(setup.size(), size_t(1)) :
-            setup.size();
-        // max index of the variables in the payload
-        const EltUnit maxSetupVarNum(isInputsPulled ? size_t(132) : payloadSize);
-        const OctEltUnit maxSetupOct = round_up<OctElement>(maxSetupVarNum);
-        // URB allocation size is the maximum of the input and ouput entry size.
-        return std::max(round_up<OctElement>(m_properties.m_URBOutputLength), maxSetupOct);
+        QuadEltUnit urbInputLength = isInputsPulled ?
+            GetMaxNumInputRegister() : GetNumInputRegistersPushed();
+        if (m_Platform->getWATable().Wa_16011983264)
+        {
+            const QuadEltUnit one(1);
+            urbInputLength = std::max(urbInputLength, one);
+        }
+
+        // URB allocation size is the maximum of the input and output entry size.
+        return round_up<OctElement>(
+            std::max(m_properties.m_URBOutputLength, urbInputLength));
     }
 
     OctEltUnit CVertexShader::GetVertexURBEntryReadLength() const
     {
-        // max index of the variables in the payload
-        const EltUnit maxSetupVarNum(
-            m_Platform->getWATable().Wa_16011983264 ?
-             std::max(setup.size(), size_t(1)) :
-             setup.size());
-
-        // rounded up to 8-element size
-        return round_up<OctElement>(maxSetupVarNum);
+        QuadEltUnit numInputsPushed = GetNumInputRegistersPushed();
+        if (m_Platform->getWATable().Wa_16011983264)
+        {
+            const QuadEltUnit one(1);
+            numInputsPushed = std::max(numInputsPushed, one);
+        }
+
+        return round_up<OctElement>(numInputsPushed);
     }
 
     OctEltUnit CVertexShader::GetVertexURBEntryReadOffset() const
@@ -300,9 +302,29 @@ namespace IGC
 
     QuadEltUnit CVertexShader::GetMaxNumInputRegister() const
     {
-        // max index of the variables in the payload
-        // if there are any pulled inputs set max num input register to max possible inputs 33 * 4
-        const EltUnit maxSetupVarNum(isInputsPulled ? 132 : setup.size());
+        const EltUnit maxSetupVarNum(isInputsPulled ?
+            m_properties.m_MaxUsedInputSlots : setup.size());
+        return round_up<QuadElement>(maxSetupVarNum);
+    }
+
+    QuadEltUnit CVertexShader::GetNumInputRegistersPushed() const
+    {
+        uint numInputComponents = 0;
+        if (m_ElementComponentPackingEnabled)
+        {
+            // This code does not expect inputs with index >= 32 to be pushed.
+            // see PushAnalysis pass.
+            IGC_ASSERT(setup.size() <= MAX_VSHADER_INPUT_REGISTERS_PACKAGEABLE);
+            for (int i = 0; i < MAX_VSHADER_INPUT_REGISTERS_PACKAGEABLE; ++i)
+            {
+                numInputComponents += iSTD::BitCount(m_ElementComponentEnableMask[i]);
+            }
+        }
+        else
+        {
+            numInputComponents = setup.size();
+        }
+        const EltUnit maxSetupVarNum(numInputComponents);
         return round_up<QuadElement>(maxSetupVarNum);
     }
 

IGC/Compiler/CISACodeGen/VertexShaderCodeGen.hpp

@@ -64,10 +64,12 @@ namespace IGC
 
         /// Returns the offset that SBE or GS should apply when reading the URB entries
         /// output by Vertex Shader.
-        /// This value is used to set the corresponding field in 3DSTATE_GS.
+        /// This value is used to set the corresponding field in 3DSTATE_VS.
         OctEltUnit        GetVertexURBEntryOutputReadOffset() const;
 
-        /// Returns the size of the vertex entry read used to load payload registers.
+        /// Returns the size of the vertex entry read used to load payload
+        /// registers. The value is calculated with vertex element component
+        /// packing applied.
         OctEltUnit        GetVertexURBEntryReadLength() const;
 
         /// Returns the offset that should be applied to vertex entry read used to load payload registers.
@@ -76,9 +78,17 @@ namespace IGC
         /// Returns the overall URB Allocation Size that should be used for handling of vertex data.
         OctEltUnit        GetURBAllocationSize() const;
 
-        /// Returns
+        /// Returns the maximum input register index (vertex element) accessed
+        /// in the shader plus 1. The value returned is calculated without
+        /// vertex element component packing applied. Driver uses this value to
+        /// to configure VF.
         QuadEltUnit       GetMaxNumInputRegister() const;
 
+        /// Returns the number of pushed input registers, in 4*DWORD units.
+        /// The value returned is calculated with vertex element component
+        /// packing applied.
+        QuadEltUnit       GetNumInputRegistersPushed() const;
+
         void              AddEpilogue(llvm::ReturnInst* pRet);
 
         /// Helper to compact inputs

IGC/Compiler/CISACodeGen/VertexShaderLowering.cpp

@@ -331,7 +331,7 @@ namespace IGC
                         case XP0:
                         case XP1:
                         case XP2:
-                            IGC_ASSERT(m_context->platform.supportsDrawParametersSGVs() || 
+                            IGC_ASSERT(m_context->platform.supportsDrawParametersSGVs() ||
                                 m_context->m_DriverInfo.UsesVertexBuffersToSendShaderDrawParameters());
                             vertexFetchSGVExtendedParameters.at(usage - XP0) = inst;
                             break;
@@ -355,7 +355,7 @@ namespace IGC
 
             if (baseVertex || baseInstance || drawIndex || vertexId || InstanceId)
             {
-                // Find first free location at the end i.e. after all user inputs. 
+                // Find first free location at the end i.e. after all user inputs.
                 unsigned int drawParametersIndex = (ARRAY_COUNT(m_inputUsed) - 1);
                 for (int index = drawParametersIndex; index >= 0; --index)
                 {
@@ -370,7 +370,7 @@ namespace IGC
                     }
                 }
 
-                /// UMD has to limit the number of user inputs in order 
+                /// UMD has to limit the number of user inputs in order
                 /// to be sure there are 2 free inputs at the end.
                 assert(drawParametersIndex < (ARRAY_COUNT(m_inputUsed) - 8));
 
@@ -442,6 +442,16 @@ namespace IGC
             }
         }
 
+        for (unsigned int slot = 0;
+            slot < ARRAY_COUNT(m_inputUsed);
+            ++slot)
+        {
+            if (m_inputUsed[slot])
+            {
+                m_vsPropsPass->SetInputSlotUsed(slot);
+            }
+        }
+
         //URB padding to 32Byte offsets
         bool addURBPaddingTo32Bytes = true;
         for (unsigned int i = 0; addURBPaddingTo32Bytes && i < MaxNumOfOutput + m_headerSize.Count(); i++)
@@ -696,6 +706,7 @@ namespace IGC
     }
 
     VertexShaderProperties::VertexShaderProperties() :
+        m_MaxUsedInputSlots(0),
         m_HasVertexID(false),
         m_VID(0),
         m_HasInstanceID(false),
@@ -750,6 +761,12 @@ namespace IGC
         m_vsProps.m_VID = VIDSlot;
     }
 
+    void CollectVertexShaderProperties::SetInputSlotUsed(unsigned int slot)
+    {
+        m_vsProps.m_MaxUsedInputSlots =
+            std::max(m_vsProps.m_MaxUsedInputSlots, slot + 1);
+    }
+
     void CollectVertexShaderProperties::SetShaderDrawParameter(size_t paramIndex, unsigned int slot)
     {
         IGC_ASSERT(paramIndex < ARRAY_COUNT(m_vsProps.m_VertexFetchSGVExtendedParameters.extendedParameters));

IGC/Compiler/CISACodeGen/VertexShaderLowering.hpp

@@ -37,6 +37,8 @@ namespace IGC
     {
     public:
         VertexShaderProperties();
+        // Maximum used input vertex element component +1.
+        unsigned int m_MaxUsedInputSlots;
         bool m_HasVertexID;
         unsigned int m_VID;
         bool m_HasInstanceID;
@@ -53,6 +55,7 @@ namespace IGC
     public:
         CollectVertexShaderProperties();
         static char ID;
+        void SetInputSlotUsed(unsigned int slot);
         void SetVertexIdSlot(unsigned int VIDslot);
         void SetInstanceIdSlot(unsigned int IIDslot);
         void SetShaderDrawParameter(size_t paramIndex, unsigned int slot);