Support for uniform typed read

Author: srividyakarumuri

IGC/Compiler/CISACodeGen/EmitVISAPass.cpp

@@ -11171,7 +11171,7 @@ void EmitPass::JoinSIMD(CVariable* (&tempdst)[N], uint responseLength, SIMDMode
     }
 }
 
-CVariable* EmitPass::BroadcastIfUniform(CVariable* pVar)
+CVariable* EmitPass::BroadcastIfUniform(CVariable* pVar, bool nomask)
 {
     IGC_ASSERT_MESSAGE(nullptr != pVar, "pVar is null");
     VISA_Type VarT = pVar->GetType();
@@ -11204,12 +11204,16 @@ CVariable* EmitPass::BroadcastIfUniform(CVariable* pVar)
 
         for (uint i = 0; i < elts; ++i)
         {
+            if (nomask)
+                m_encoder->SetNoMask();
             m_encoder->SetSrcSubReg(0, i * Stride);
             if (Stride != 1) m_encoder->SetDstRegion(Stride);
             m_encoder->SetDstSubReg((i * Stride) * width);
             m_encoder->Copy(Dst, ImmLo ? ImmLo : Src);
             m_encoder->Push();
             if (Need64BitEmu) {
+                if (nomask)
+                    m_encoder->SetNoMask();
                 m_encoder->SetSrcSubReg(0, i * Stride + 1);
                 if (Stride != 1) m_encoder->SetDstRegion(Stride);
                 m_encoder->SetDstSubReg((i * Stride) * width + 1);
@@ -13341,18 +13345,42 @@ void EmitPass::emitTypedRead(llvm::Instruction* pInsn)
     CVariable* pV = (pR == nullptr && isUndefOrConstInt0(pllV)) ? nullptr : GetSymbol(pllV);
     CVariable* pU = GetSymbol(pllU);
 
-    pU = BroadcastIfUniform(pU);
-    pV = pV ? BroadcastIfUniform(pV) : nullptr;
-    pR = pR ? BroadcastIfUniform(pR) : nullptr;
-    pLOD = pLOD ? BroadcastIfUniform(pLOD) : nullptr;
+    pU = BroadcastIfUniform(pU, m_currShader->GetIsUniform(pInsn));
+    pV = pV ? BroadcastIfUniform(pV, m_currShader->GetIsUniform(pInsn)) : nullptr;
+    pR = pR ? BroadcastIfUniform(pR, m_currShader->GetIsUniform(pInsn)) : nullptr;
+    pLOD = pLOD ? BroadcastIfUniform(pLOD, m_currShader->GetIsUniform(pInsn)) : nullptr;
 
     ResourceDescriptor resource = GetResourceVariable(pllSrcBuffer);
 
     uint numChannels = iSTD::BitCount(writeMask);
 
     if (m_currShader->GetIsUniform(pInsn))
     {
-        IGC_ASSERT_MESSAGE(0, "Uniform ld_uav_typed not implemented yet");
+        SIMDMode nativeDispatchMode = m_currShader->m_Platform->getMinDispatchMode();
+        CVariable* tempdst = nullptr;
+        tempdst = m_currShader->GetNewVariable(
+            numChannels * numLanes(nativeDispatchMode),
+            ISA_TYPE_F,
+            EALIGN_GRF,
+            CName("tyReadDest"));
+        m_encoder->SetSimdSize(nativeDispatchMode);
+        m_encoder->SetPredicate(nullptr);
+        m_encoder->SetNoMask();
+        m_encoder->TypedRead4(resource, pU, pV, pR, pLOD, tempdst, writeMask);
+
+        m_encoder->Push();
+
+        // Mov the required channel values to m_destination
+        m_encoder->SetSimdSize(SIMDMode::SIMD1);
+        m_encoder->SetNoMask();
+
+        for (uint i = 0; i < numChannels; ++i)
+        {
+            m_encoder->SetSrcSubReg(0, i * numLanes(nativeDispatchMode));
+            m_encoder->SetDstSubReg(i);
+            m_encoder->Copy(m_destination, tempdst);
+            m_encoder->Push();
+        }
     }
     else
     {

IGC/Compiler/CISACodeGen/EmitVISAPass.hpp

@@ -461,7 +461,7 @@ class EmitPass : public llvm::FunctionPass
     void SplitSIMD(llvm::Instruction* inst, uint numSources, uint headerSize, CVariable* payload, SIMDMode mode, uint half);
     template<size_t N>
     void JoinSIMD(CVariable* (&tempdst)[N], uint responseLength, SIMDMode mode);
-    CVariable* BroadcastIfUniform(CVariable* pVar);
+    CVariable* BroadcastIfUniform(CVariable* pVar, bool nomask = false);
     uint DecideInstanceAndSlice(const llvm::BasicBlock& blk, SDAG& sdag, bool& slicing);
     bool IsUndefOrZeroImmediate(const llvm::Value* value);
     inline bool isUndefOrConstInt0(const llvm::Value* val)

IGC/Compiler/CISACodeGen/WIAnalysis.cpp

@@ -1252,6 +1252,7 @@ WIAnalysis::WIDependancy WIAnalysisRunner::calculate_dep(const CallInst* inst)
         intrinsic_name == llvm_waveAll ||
         intrinsic_name == llvm_waveClustered ||
         intrinsic_name == llvm_ld_ptr ||
+        (IGC_IS_FLAG_DISABLED(DisableUniformTypedAccess) && intrinsic_name == llvm_typed_read) ||
         intrinsic_name == llvm_add_pair ||
         intrinsic_name == llvm_sub_pair ||
         intrinsic_name == llvm_mul_pair ||

IGC/common/igc_flags.def

@@ -111,6 +111,7 @@ DECLARE_IGC_REGKEY(bool, DisableConstantCoalescing,     false, "Setting this to
 DECLARE_IGC_REGKEY(bool, DisableConstantCoalescingOutOfBoundsCheck,     false, "Setting this to 1/true adds a compiler switch to disable constant coalesing out of bounds check", false)
 DECLARE_IGC_REGKEY(bool, UseHDCTypedReadForAllTextures, false, "Setting this to use HDC message rather than sampler ld for texture read", false)
 DECLARE_IGC_REGKEY(bool, UseHDCTypedReadForAllTypedBuffers,  false, "Setting this to use HDC message rather than sampler ld for buffer read", false)
+DECLARE_IGC_REGKEY(bool, DisableUniformTypedAccess,     true, "Setting this will disable uniform typed access handling", false)
 DECLARE_IGC_REGKEY(bool, DisableURBWriteMerge,          false, "Setting this to 1/true adds a compiler switch to disable URB write merge", false)
 DECLARE_IGC_REGKEY(bool, DisableEmptyBlockRemoval,      false, "Setting this to 1/true adds a compiler switch to disable empty block optimization", false)
 DECLARE_IGC_REGKEY(bool, DisableSIMD32Slicing,          false, "Setting this to 1/true adds a compiler switch to disable emitting SIMD32 VISA code in slices", false)