Remove tracking of bindless offset for images - PART#3

In case of bindless images, NEO runtime passes bindless offset in image
arg of kernel. This bindless offset doesn't need to be tracked because
arg tracking in OCLBIConverter pass is only needed to find image's BTI,
which isn't used with bindless as it's the bindless offset that is used
to find the image's surface state instead of values stored in the
Binding Table.

This commit is the third part of the tracking removal.

It implements handling of inline samplers for bindless mode in two new
passes. First, PrepareInlineSamplerForBindless visits instructions
calling __bindless_sampler_initializer(int32) and maps the callee's
argument for AddImplicitArgs pass and then creates inline sampler
metadata. Second, ResolveInlineSamplerForBindless replaces calls to
__bindless_sampler_initializer(int32) with corresponding inline sampler
implicit args created by AddImplicitArgs.

This commit also edits builtin __translate_sampler_initializer(int32) to
return __bindless_sampler_initializer(int32) for bindless.

Author: smilczek

IGC/AdaptorOCL/UnifyIROCL.cpp

@@ -60,6 +60,8 @@ SPDX-License-Identifier: MIT
 #include "Compiler/Optimizer/OpenCLPasses/ImageFuncs/ImageFuncsAnalysis.hpp"
 #include "Compiler/Optimizer/OpenCLPasses/ImageFuncs/ImageFuncResolution.hpp"
 #include "Compiler/Optimizer/OpenCLPasses/ImageFuncs/ResolveSampledImageBuiltins.hpp"
+#include "Compiler/Optimizer/OpenCLPasses/PrepareInlineSamplerForBindless/PrepareInlineSamplerForBindless.hpp"
+#include "Compiler/Optimizer/OpenCLPasses/ResolveInlineSamplerForBindless/ResolveInlineSamplerForBindless.hpp"
 #include "Compiler/Optimizer/OpenCLPasses/PrivateMemory/PrivateMemoryUsageAnalysis.hpp"
 #include "Compiler/Optimizer/OpenCLPasses/PrivateMemory/PrivateMemoryResolution.hpp"
 #include "Compiler/Optimizer/OpenCLPasses/ProgramScopeConstants/ProgramScopeConstantAnalysis.hpp"
@@ -560,9 +562,12 @@ static void CommonOCLBasedPasses(OpenCLProgramContext* pContext)
 
     mpm.add(new ResolveOCLRaytracingBuiltins());
     mpm.add(createRayTracingIntrinsicAnalysisPass());
+    mpm.add(new PrepareInlineSamplerForBindless());
+
     // Adding implicit args based on Analysis passes
     mpm.add(new AddImplicitArgs());
 
+    mpm.add(new ResolveInlineSamplerForBindless());
     if (IGC_IS_FLAG_ENABLED(BufferBoundsChecking) || pContext->isBufferBoundsChecking())
     {
         mpm.add(new BufferBoundsCheckingPatcher());

IGC/BiFModule/Headers/bif_flag_controls.h

@@ -38,4 +38,5 @@ BIF_FLAG_CONTROL(int, MaxHWThreadIDPerSubDevice)
 BIF_FLAG_CONTROL(bool, UseAssumeInGetGlobalId)
 BIF_FLAG_CONTROL(int, JointMatrixLoadStoreOpt)
 BIF_FLAG_CONTROL(bool, UseOOBChecks)
+BIF_FLAG_CONTROL(bool, UseBindlessImage)
 #endif // __BIF_FLAG_CONTROL_H__

IGC/BiFModule/Implementation/images.cl

@@ -22,8 +22,14 @@ SPDX-License-Identifier: MIT
 
 #include "../Headers/spirv.h"
 
+__spirv_Sampler __bindless_sampler_initializer(uint sampler);
+
 __spirv_Sampler __translate_sampler_initializer(uint sampler)
 {
+    if(BIF_FLAG_CTRL_GET(UseBindlessImage))
+    {
+        return __bindless_sampler_initializer(sampler);
+    }
     return __builtin_astype((ulong)sampler, __spirv_Sampler);
 }
 

IGC/Compiler/Builtins/BIFFlagCtrl/BIFFlagCtrlResolution.cpp

@@ -116,6 +116,10 @@ void BIFFlagCtrlResolution::FillFlagCtrl() {
   }
 
   BIF_FLAG_CTRL_SET(UseOOBChecks, PtrCGC->platform.needsOutOfBoundsBuiltinChecks());
+
+  // NOTE: No need to check for UseLegacyBindlessMode,
+  //       as it's unrelated to images.
+  BIF_FLAG_CTRL_SET(UseBindlessImage, PtrCGC->getModuleMetaData()->UseBindlessImage);
 }
 
 #undef BIF_FLAG_CTRL_SET

IGC/Compiler/InitializePasses.h

@@ -118,6 +118,7 @@ void initializeOpenCLPrintfAnalysisPass(llvm::PassRegistry&);
 void initializeOpenCLPrintfResolutionPass(llvm::PassRegistry&);
 void initializePeepholeTypeLegalizerPass(llvm::PassRegistry&);
 void initializePositionDepAnalysisPass(llvm::PassRegistry&);
+void initializePrepareInlineSamplerForBindlessPass(llvm::PassRegistry&);
 void initializePrivateMemoryResolutionPass(llvm::PassRegistry&);
 void initializePrivateMemoryToSLMPass(llvm::PassRegistry&);
 void initializePrivateMemoryUsageAnalysisPass(llvm::PassRegistry&);
@@ -133,6 +134,7 @@ void initializePullConstantHeuristicsPass(llvm::PassRegistry&);
 void initializeScalarizerCodeGenPass(llvm::PassRegistry&);
 void initializeReduceLocalPointersPass(llvm::PassRegistry&);
 void initializeReplaceUnsupportedIntrinsicsPass(llvm::PassRegistry&);
+void initializeResolveInlineSamplerForBindlessPass(llvm::PassRegistry&);
 void initializePreCompiledFuncImportPass(llvm::PassRegistry&);
 void initializePurgeMetaDataUtilsPass(llvm::PassRegistry&);
 void initializeRemoveNonPositionOutputPass(llvm::PassRegistry&);

IGC/Compiler/Optimizer/BuiltInFuncImport.cpp

@@ -649,7 +649,22 @@ bool BIImport::runOnModule(Module& M)
                     IRBuilder<> builder(CI);
                     Value* newV;
                     if (funcName.startswith("__builtin_IB_convert_sampler_to_int"))
-                        newV = builder.CreatePtrToInt(CI->getOperand(0), CI->getType());
+                    {
+                        Type* CIOpTy = CI->getOperand(0)->getType();
+                        if (isa<PointerType>(CIOpTy))
+                        {
+                            newV = builder.CreatePtrToInt(CI->getOperand(0), CI->getType());
+                        }
+                        else
+                        {
+                            CodeGenContext* ctx = getAnalysis<CodeGenContextWrapper>().getCodeGenContext();
+
+                            bool UseBindlessImage = ctx->getModuleMetaData()->UseBindlessImage;
+                            ctx->getModuleMetaData()->UseBindlessImageWithSamplerTracking = UseBindlessImage;
+                            // i32 to i64
+                            newV = builder.CreateZExt(CI->getOperand(0), CI->getType());
+                        }
+                    }
                     else
                         newV = builder.CreateBitOrPointerCast(CI->getOperand(0), CI->getType());
                     CI->replaceAllUsesWith(newV);