Experimental SYCL unmasked call feature.

Change-Id: Iaaee3a6626a64faa83fa746a9d048f758b396340

Author: mbelicki

IGC/AdaptorOCL/UnifyIROCL.cpp

@@ -97,6 +97,7 @@ SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 #include "Compiler/Optimizer/OpenCLPasses/SubGroupFuncs/SubGroupFuncsResolution.hpp"
 #include "Compiler/Optimizer/OpenCLPasses/BIFTransforms/BIFTransforms.hpp"
 #include "Compiler/Optimizer/OpenCLPasses/BreakdownIntrinsic.h"
+#include "Compiler/Optimizer/OpenCLPasses/TransformUnmaskedFunctionsPass.h"
 #include "Compiler/Optimizer/OpenCLPasses/StatelessToStatefull/StatelessToStatefull.hpp"
 #include "Compiler/Optimizer/OpenCLPasses/KernelFunctionCloning.h"
 #include "Compiler/Legalizer/TypeLegalizerPass.h"
@@ -279,6 +280,8 @@ static void CommonOCLBasedPasses(
     mpm.add(new MetaDataUtilsWrapper(pMdUtils, pContext->getModuleMetaData()));
     mpm.add(new CodeGenContextWrapper(pContext));
 
+    mpm.add(new TransformUnmaskedFunctionsPass());
+
     mpm.add(new ClampLoopUnroll(256));
 
     mpm.add(new MoveStaticAllocas());

IGC/Compiler/CISACodeGen/CISABuilder.cpp

@@ -344,6 +344,10 @@ namespace IGC
         m_encoderState.m_noMask = false;
         m_encoderState.m_simdSize = m_program->m_SIMDSize;
         m_encoderState.m_uniformSIMDSize = SIMDMode::SIMD1;
+
+        if (m_insideForcedNoMaskRegion) {
+          m_encoderState.m_noMask = true;
+        }
     }
 
     CEncoder::CEncoder()
@@ -4332,6 +4336,7 @@ namespace IGC
         CodeGenContext* context = m_program->GetContext();
         m_encoderState.m_secondHalf = false;
         m_enableVISAdump = false;
+        m_insideForcedNoMaskRegion = false;
         labelMap.clear();
         labelMap.resize(m_program->entry->size(), nullptr);
         labelCounter = 0;

IGC/Compiler/CISACodeGen/CISABuilder.hpp

@@ -348,6 +348,9 @@ namespace IGC
         inline void SetSecondHalf(bool secondHalf);
         inline bool IsSecondHalf();
 
+        inline void BeginForcedNoMaskRegion();
+        inline void EndForcedNoMaskRegion();
+
         void Wait();
 
         VISAKernel* GetVISAKernel() { return vKernel; }
@@ -561,6 +564,8 @@ namespace IGC
         VISABuilder* vbuilder;
         VISABuilder* vAsmTextBuilder;
 
+        bool m_insideForcedNoMaskRegion;
+
         bool m_enableVISAdump;
         bool m_hasInlineAsm;
         std::vector<VISA_LabelOpnd*> labelMap;
@@ -873,6 +878,17 @@ namespace IGC
         Arithmetic(ISA_DP4A, dst, src0, src1, src2);
     }
 
+    inline void CEncoder::BeginForcedNoMaskRegion()
+    {
+        m_insideForcedNoMaskRegion = true;
+        m_encoderState.m_noMask = true;
+    }
+
+    inline void CEncoder::EndForcedNoMaskRegion()
+    {
+        m_insideForcedNoMaskRegion = false;
+    }
+
     inline void CEncoder::SetNoMask()
     {
         m_encoderState.m_noMask = true;

IGC/Compiler/CISACodeGen/EmitVISAPass.cpp

@@ -16817,6 +16817,15 @@ void EmitPass::emitDP4A(GenIntrinsicInst* GII) {
     m_encoder->Push();
 }
 
+void EmitPass::emitUnmaskedRegionBoundary(bool start)
+{
+    if (start) {
+      m_encoder->BeginForcedNoMaskRegion();
+    } else {
+      m_encoder->EndForcedNoMaskRegion();
+    }
+}
+
 void EmitPass::emitDebugPlaceholder(llvm::GenIntrinsicInst* I)
 {
     m_encoder->Loc(I->getDebugLoc().getLine());

IGC/Compiler/CISACodeGen/EmitVISAPass.hpp

@@ -428,6 +428,7 @@ namespace IGC
 
         void emitLLVMbswap(llvm::IntrinsicInst* inst);
         void emitDP4A(llvm::GenIntrinsicInst* GII);
+        void emitUnmaskedRegionBoundary(bool start);
         // Debug Built-Ins
         void emitStateRegID(uint64_t and_imm, uint64_t shr_imm);
         void emitThreadPause(llvm::GenIntrinsicInst* inst);

IGC/Compiler/CISACodeGen/PatternMatchPass.cpp

@@ -1126,6 +1126,12 @@ namespace IGC
                 match = MatchURBRead(*CI) ||
                     MatchSingleInstruction(*CI);
                 break;
+            case GenISAIntrinsic::GenISA_UnmaskedRegionBegin:
+                match = MatchUnmaskedRegionBoundary(I, true);
+                break;
+            case GenISAIntrinsic::GenISA_UnmaskedRegionEnd:
+                match = MatchUnmaskedRegionBoundary(I, false);
+                break;
             default:
                 match = MatchSingleInstruction(I);
                 // no pattern for the rest of the intrinsics
@@ -3725,6 +3731,24 @@ namespace IGC
         return true;
     }
 
+    bool CodeGenPatternMatch::MatchUnmaskedRegionBoundary(Instruction &I, bool start)
+    {
+        struct UnmaskedBoundaryPattern : public Pattern
+        {
+            bool start;
+            virtual void Emit(EmitPass* pass, const DstModifier& modifier)
+            {
+                (void) modifier;
+                pass->emitUnmaskedRegionBoundary(start);
+            }
+        };
+
+        UnmaskedBoundaryPattern* pattern = new (m_allocator) UnmaskedBoundaryPattern();
+        pattern->start = start;
+        AddPattern(pattern);
+        return true;
+    }
+
 
     bool CodeGenPatternMatch::MatchLogicAlu(llvm::BinaryOperator& I)
     {

IGC/Compiler/CISACodeGen/PatternMatchPass.hpp

@@ -242,6 +242,8 @@ namespace IGC
         bool MatchCopyToStruct(llvm::InsertValueInst*);
         bool MatchCopyFromStruct(llvm::ExtractValueInst*);
 
+        bool MatchUnmaskedRegionBoundary(llvm::Instruction& I, bool start);
+
         void AddPattern(Pattern* P)
         {
             m_currentPattern = P;

IGC/Compiler/CISACodeGen/opCode.h

@@ -256,3 +256,6 @@ DECLARE_OPCODE(GenISA_WavePrefix, GenISAIntrinsic, llvm_wavePrefix, false, false
 DECLARE_OPCODE(GenISA_QuadPrefix, GenISAIntrinsic, llvm_quadPrefix, false, false, false, false, false, false, false)
 DECLARE_OPCODE(GenISA_WaveShuffleIndex, GenISAIntrinsic, llvm_waveShuffleIndex, false, false, false, false, false, false, false)
 
+// Unmasked region
+DECLARE_OPCODE(GenISA_UnmaskedRegionBegin, GenISAIntrinsic, llvm_unmaskedBegin, false, false, false, false, false, false, false)
+DECLARE_OPCODE(GenISA_UnmaskedRegionEnd, GenISAIntrinsic, llvm_unmaskedEnd, false, false, false, false, false, false, false)

IGC/Compiler/Optimizer/OpenCLPasses/CMakeLists.txt

@@ -39,6 +39,7 @@ add_subdirectory(UnreachableHandling)
 set(IGC_BUILD__SRC__OpenCLPasses
     "${CMAKE_CURRENT_SOURCE_DIR}/KernelArgs.cpp"
     "${CMAKE_CURRENT_SOURCE_DIR}/BreakdownIntrinsic.cpp"
+    "${CMAKE_CURRENT_SOURCE_DIR}/TransformUnmaskedFunctionsPass.cpp"
     "${CMAKE_CURRENT_SOURCE_DIR}/KernelFunctionCloning.cpp"
     "${CMAKE_CURRENT_SOURCE_DIR}/ErrorCheckPass.cpp"
   )
@@ -86,6 +87,7 @@ set(IGC_BUILD__SRC__Optimizer_OpenCLPasses
 set(IGC_BUILD__HDR__OpenCLPasses
     "${CMAKE_CURRENT_SOURCE_DIR}/KernelArgs.hpp"
     "${CMAKE_CURRENT_SOURCE_DIR}/BreakdownIntrinsic.h"
+    "${CMAKE_CURRENT_SOURCE_DIR}/TransformUnmaskedFunctionsPass.h"
     "${CMAKE_CURRENT_SOURCE_DIR}/KernelFunctionCloning.h"
     "${CMAKE_CURRENT_SOURCE_DIR}/ErrorCheckPass.h"
   )