Enable WaveAllJointReduction for 32-bit types

Turn on the WaveAllJointReduction optimization for 32-bit types (int and
float)
Add regkey that disables the optimization for 64-bit types

Author: bowenxue-intel

IGC/Compiler/Optimizer/WaveAllJointReduction.cpp

@@ -6,9 +6,12 @@ SPDX-License-Identifier: MIT
 
 ============================= end_copyright_notice ===========================*/
 
+#include <igc_regkeys.hpp>
+
 #include <GenISAIntrinsics/GenIntrinsicInst.h>
 #include "WaveAllJointReduction.hpp"
 #include "Compiler/IGCPassSupport.h"
+
 #include "common/LLVMWarningsPush.hpp"
 #include <llvm/IR/InstVisitor.h>
 #include <llvm/ADT/SmallVector.h>
@@ -93,6 +96,12 @@ void WaveAllJointReductionImpl::visitCallInst( CallInst& callInst )
             return;
         }
 
+        // Disable joint reduction optimization for >32-bit WaveAll operations
+        if (waveAllInst->getType()->getScalarSizeInBits() > 32) {
+            if (IGC_IS_FLAG_DISABLED(EnableWaveAllJointReduction64Bit))
+                return;
+        }
+
         SmallVector<WaveAllIntrinsic*, 16> mergeList{ waveAllInst };
 
         // For locality, only look at consecutive instructions since non-consecutive instructions may require sinking the final vector WaveAll instruction to where the last joined WaveAll is to satisfy proper domination of each WaveAll's Src

IGC/common/igc_flags.h

@@ -344,7 +344,8 @@ DECLARE_IGC_REGKEY(bool, DisableLoopSplitWidePHIs, false, "Disable splitting of
 DECLARE_IGC_REGKEY(bool, EnableBarrierControlFlowOptimizationPass, false, "Enable barrier control flow optimization pass", false)
 DECLARE_IGC_REGKEY(bool, EnableWaveShuffleIndexSinking, true, "Hoist identical instructions operating on WaveShuffleIndex instructions with the same source and a constant lane/channel", false)
 DECLARE_IGC_REGKEY(DWORD, WaveShuffleIndexSinkingMaxIterations, 3, "Max number of iterations to run iterative WaveShuffleIndexSinking", false)
-DECLARE_IGC_REGKEY(bool, EnableWaveAllJointReduction, false, "Enable Joint Reduction Optimization.", false)
+DECLARE_IGC_REGKEY(bool, EnableWaveAllJointReduction, true, "Enable Joint Reduction Optimization.", false)
+DECLARE_IGC_REGKEY(bool, EnableWaveAllJointReduction64Bit, false, "Enable Joint Reduction Optimization for 64-bit data types (i64, double).", false)
 DECLARE_IGC_REGKEY(bool, EnablePromoteToPredicatedMemoryAccess, false, "Enable predicated load/store if conversion.", true)
 DECLARE_IGC_REGKEY(bool, EnableIntDivRemIncrementReduction, true, "Enable consecutive Int DivRem increment by constant optimization", false)
 DECLARE_IGC_REGKEY(bool, DivRemIncrementCondBranchSimplify, false, "Create branches when simplifying consecutive udiv/urem groups increment dividend by constant greater than 1", false)