Fix lowerBoundGRF value in VRT

IGC computes the lowerBoundGRF, but sets vISA_MinGRFNum too late for
vISA to retrieve in time. Move up the value setting to InitEncoder.

Author: iwwu

IGC/Compiler/CISACodeGen/CISABuilder.cpp

@@ -5335,7 +5335,7 @@ namespace IGC
         }
     }
 
-    void CEncoder::InitEncoder(bool canAbortOnSpill, bool hasStackCall, bool hasInlineAsmCall, bool hasAdditionalVisaAsmToLink, int numThreadsPerEU, VISAKernel* prevKernel)
+    void CEncoder::InitEncoder(bool canAbortOnSpill, bool hasStackCall, bool hasInlineAsmCall, bool hasAdditionalVisaAsmToLink, int numThreadsPerEU, uint lowerBoundGRF, VISAKernel* prevKernel)
     {
         m_aliasesMap.clear();
         m_encoderState.m_SubSpanDestination = false;
@@ -5399,6 +5399,11 @@ namespace IGC
             SaveOption(vISA_AutoGRFSelection, true);
         }
 
+        if (lowerBoundGRF > 0)
+        {
+            SaveOption(vISA_MinGRFNum, lowerBoundGRF);
+        }
+
         // Pass all build options to builder
         SetBuilderOptions(vbuilder);
 

IGC/Compiler/CISACodeGen/CISABuilder.hpp

@@ -115,7 +115,7 @@ namespace IGC
     class CEncoder
     {
     public:
-        void InitEncoder(bool canAbortOnSpill, bool hasStackCall, bool hasInlineAsmCall, bool hasAdditionalVisaAsmToLink, int numThreadsPerEU, VISAKernel* prevKernel);
+        void InitEncoder(bool canAbortOnSpill, bool hasStackCall, bool hasInlineAsmCall, bool hasAdditionalVisaAsmToLink, int numThreadsPerEU, uint lowerBoundGRF, VISAKernel* prevKernel);
         void InitBuildParams(llvm::SmallVector<std::unique_ptr< char, std::function<void(char*)>>, 10> & params);
         void InitVISABuilderOptions(TARGET_PLATFORM VISAPlatform, bool canAbortOnSpill, bool hasStackCall, bool enableVISA_IR);
         SEncoderState CopyEncoderState();

IGC/Compiler/CISACodeGen/EmitVISAPass.cpp

@@ -813,14 +813,17 @@ bool EmitPass::runOnFunction(llvm::Function& F)
                 hasAdditionalVisaAsmToLink = true;
             }
         }
+
+        uint lowerBoundGRF = 0;
         int numThreadsPerEU = -1;
         if (F.hasFnAttribute("num-thread-per-eu"))
         {
             numThreadsPerEU = std::stoi(
                 F.getFnAttribute("num-thread-per-eu").getValueAsString().str());
         }
         // call builder after pre-analysis pass where scratchspace offset to VISA is calculated
-        m_encoder->InitEncoder(m_canAbortOnSpill, m_currShader->HasStackCalls(), hasInlineAsmCall, hasAdditionalVisaAsmToLink, numThreadsPerEU, prevKernel);
+        m_encoder->InitEncoder(m_canAbortOnSpill, m_currShader->HasStackCalls(), hasInlineAsmCall, hasAdditionalVisaAsmToLink,
+            numThreadsPerEU, lowerBoundGRF, prevKernel);
 
         if (!m_encoder->IsCodePatchCandidate())
             createVMaskPred(m_vMaskPredForSubplane);