Check intel_reqd_sub_group_size attribute in SIMD16 drop

Check intel_reqd_sub_group_size attribute when deciding if we can early
drop to SIMD 16.

Author: stefan-il

IGC/Compiler/CISACodeGen/EmitVISAPass.cpp

@@ -558,33 +558,6 @@ bool EmitPass::shouldForceEarlyRecompile(MetaDataUtils *pMdUtils, llvm::Function
   return PassedThreshold;
 }
 
-bool EmitPass::shouldDropToSIMD16(MetaDataUtils *pMdUtils, llvm::Function *F) {
-  if (m_pCtx->type != ShaderType::OPENCL_SHADER || IGC_IS_FLAG_DISABLED(AllowEarlySIMD16DropForXE3)) {
-    return false;
-  }
-
-  if (!m_canAbortOnSpill || !m_pCtx->isAutoGRFSelectionEnabled() || !isEntryFunc(pMdUtils, F)) {
-    return false;
-  }
-
-  // If there are user set values for SIMD size or GRF number just return
-  if (IGC_IS_FLAG_ENABLED(ForceCSSIMD32) || m_pCtx->getModuleMetaData()->csInfo.forcedSIMDSize != 0 ||
-      m_pCtx->getNumGRFPerThread(false) != 0) {
-    return false;
-  }
-
-  // Currently, we do this optimization only for XE3 but we can relax this
-  // requirement to all platforms where abortOnSpills is enabled.
-  if (!m_pCtx->platform.isCoreXE3()) {
-    return false;
-  }
-
-  auto MaxRegPressure = getMaxRegPressureInFunctionGroup(F, pMdUtils);
-  auto Threshold = IGC_GET_FLAG_VALUE(EarlySIMD16DropForXE3Threshold);
-  bool shouldDrop = MaxRegPressure > Threshold;
-  return shouldDrop;
-}
-
 bool EmitPass::runOnFunction(llvm::Function &F) {
   m_currFuncHasSubroutine = false;
 
@@ -646,10 +619,6 @@ bool EmitPass::runOnFunction(llvm::Function &F) {
     return false;
   }
 
-  if (shouldDropToSIMD16(pMdUtils, &F)) {
-    return false;
-  }
-
   // Force SIMD8 on library compilations for non-OCL shaders
   if (m_pCtx->type != ShaderType::OPENCL_SHADER && m_pCtx->getCompilerOption().IsLibraryCompilation &&
       m_SimdMode != SIMDMode::SIMD8) {

IGC/Compiler/CISACodeGen/EmitVISAPass.hpp

@@ -979,8 +979,6 @@ class EmitPass : public llvm::FunctionPass {
 
   bool shouldForceEarlyRecompile(IGCMD::MetaDataUtils *pMdUtils, llvm::Function *F);
 
-  bool shouldDropToSIMD16(IGCMD::MetaDataUtils *pMdUtils, llvm::Function *F);
-
   bool isHalfGRFReturn(CVariable *dst, SIMDMode simdMode);
 
   void emitFeedbackEnable();

IGC/Compiler/CISACodeGen/OpenCLKernelCodeGen.cpp

@@ -2415,6 +2415,19 @@ bool COpenCLKernel::CompileSIMDSize(SIMDMode simdMode, EmitPass &EP, llvm::Funct
   return simdStatus == SIMDStatus::SIMD_PASS;
 }
 
+static bool shouldDropToSIMD16(uint32_t maxPressure, SIMDMode simdMode, CodeGenContext *pCtx, MetaDataUtils *pMdUtils,
+                               llvm::Function *F) {
+  if (simdMode != SIMDMode::SIMD32 || !isEntryFunc(pMdUtils, F)) {
+    return false;
+  }
+  if (!pCtx->isAutoGRFSelectionEnabled() || pCtx->getNumGRFPerThread(false) != 0) {
+    return false;
+  }
+  auto threshold = IGC_GET_FLAG_VALUE(EarlySIMD16DropForXE3Threshold);
+  bool shouldDrop = maxPressure > threshold;
+  return shouldDrop;
+}
+
 SIMDStatus COpenCLKernel::checkSIMDCompileCondsForMin16(SIMDMode simdMode, EmitPass &EP, llvm::Function &F,
                                                         bool hasSyncRTCalls) {
   if (simdMode == SIMDMode::SIMD8) {
@@ -2524,6 +2537,14 @@ SIMDStatus COpenCLKernel::checkSIMDCompileCondsForMin16(SIMDMode simdMode, EmitP
     }
   }
 
+  if (EP.m_canAbortOnSpill && pCtx->platform.isCoreXE3() && IGC_IS_FLAG_ENABLED(AllowEarlySIMD16DropForXE3)) {
+    bool shouldDrop = shouldDropToSIMD16(maxPressure, simdMode, pCtx, pMdUtils, &F);
+    if (shouldDrop) {
+      pCtx->SetSIMDInfo(SIMD_SKIP_PERF, simdMode, ShaderDispatchMode::NOT_APPLICABLE);
+      return SIMDStatus::SIMD_FUNC_FAIL;
+    }
+  }
+
   return SIMDStatus::SIMD_PASS;
 }
 

IGC/common/igc_flags.h

@@ -1540,7 +1540,7 @@ DECLARE_IGC_REGKEY(bool, EnableReusingXYZWStoreConstPayload, true, "Enable reusi
 DECLARE_IGC_REGKEY(bool, EnableReusingLSCStoreConstPayload, false, "Enable reusing LSC stores const payload", false)
 DECLARE_IGC_REGKEY(bool, AllowSIMD16DropForXE2, true, "Controls the switch for XE2 simd16 drop", false)
 DECLARE_IGC_REGKEY(bool, AllowSIMD16DropForXE3, true, "Controls the switch for XE3 simd16 drop", false)
-DECLARE_IGC_REGKEY(bool, AllowEarlySIMD16DropForXE3, false,
+DECLARE_IGC_REGKEY(bool, AllowEarlySIMD16DropForXE3, true,
                    "Controls the early drop to simd16 for XE3", false)
 DECLARE_IGC_REGKEY(DWORD, EarlySIMD16DropForXE3Threshold, 190, "Threshold for the early drop to simd16 for XE3", false)
 DECLARE_IGC_REGKEY(DWORD, RegPressureVerbocity, 2, "Different printing types", false)