Introduce __INTEL_SURFACE_STATE_SIZE relocation

smilczek · igcbot · commit b05b1c9c8d79 · 2026-06-02T14:37:05.000+02:00
This commit introduces a mechanism to allow for resolving of surface
state size by UMD.
diff --git a/IGC/AdaptorOCL/OCL/sp/zebin_builder.cpp b/IGC/AdaptorOCL/OCL/sp/zebin_builder.cpp
@@ -365,6 +365,9 @@ void ZEBinaryBuilder::addRuntimeSymbols(const IGC::SOpenCLProgramInfo &annotatio
   if (annotations.m_hasPerThreadOffsetRelocations)
     mBuilder.addSymbol(vISA::PER_THREAD_OFF_RELOCATION_NAME, /*addr*/ 0, /*size*/ 0, llvm::ELF::STB_GLOBAL,
                        llvm::ELF::STT_NOTYPE, /*sectionId*/ -1);
+  if (annotations.m_hasSurfaceStateSizeRelocations)
+    mBuilder.addSymbol(vISA::SURFACE_STATE_SIZE_RELOCATION_NAME, /*addr*/ 0, /*size*/ 0, llvm::ELF::STB_GLOBAL,
+                       llvm::ELF::STT_OBJECT, /*sectionId*/ -1);
 }
 
 void ZEBinaryBuilder::addProgramSymbols(const IGC::SOpenCLProgramInfo &annotations) {
diff --git a/IGC/Compiler/CISACodeGen/CISABuilder.cpp b/IGC/Compiler/CISACodeGen/CISABuilder.cpp
@@ -6125,6 +6125,9 @@ void CEncoder::createRelocationTables(VISAKernel &pMainKernel) {
       } else if (reloc.r_symbol == vISA::PER_THREAD_OFF_RELOCATION_NAME) {
         auto cl_context = static_cast<OpenCLProgramContext *>(context);
         cl_context->m_programInfo.m_hasPerThreadOffsetRelocations = true;
+      } else if (reloc.r_symbol == vISA::SURFACE_STATE_SIZE_RELOCATION_NAME) {
+        auto cl_context = static_cast<OpenCLProgramContext *>(context);
+        cl_context->m_programInfo.m_hasSurfaceStateSizeRelocations = true;
       }
     }
   }
diff --git a/IGC/Compiler/CISACodeGen/messageEncoding.hpp b/IGC/Compiler/CISACodeGen/messageEncoding.hpp
@@ -7,7 +7,6 @@ SPDX-License-Identifier: MIT
 ============================= end_copyright_notice ===========================*/
 
 #pragma once
-#include "Compiler/CodeGenPublic.h"
 #include "Compiler/CISACodeGen/helper.h"
 #include "visa_igc_common_header.h"
 
diff --git a/IGC/Compiler/CodeGenPublic.h b/IGC/Compiler/CodeGenPublic.h
@@ -636,6 +636,7 @@ struct SOpenCLProgramInfo {
   ZEBinGlobalHostAccessTable m_zebinGlobalHostAccessTable;
   bool m_hasCrossThreadOffsetRelocations = false;
   bool m_hasPerThreadOffsetRelocations = false;
+  bool m_hasSurfaceStateSizeRelocations = false;
 };
 
 class CBTILayout {
diff --git a/IGC/Compiler/Optimizer/OpenCLPasses/BindlessImage/ResolveImageImplicitArgsForBindless.cpp b/IGC/Compiler/Optimizer/OpenCLPasses/BindlessImage/ResolveImageImplicitArgsForBindless.cpp
@@ -8,7 +8,10 @@ SPDX-License-Identifier: MIT
 
 #include "ResolveImageImplicitArgsForBindless.hpp"
 
+#include "Compiler/CodeGenPublic.h"
 #include "Compiler/IGCPassSupport.h"
+#include "Compiler/CodeGenContextWrapper.hpp"
+#include "RelocationInfo.h"
 #include "Compiler/CISACodeGen/messageEncoding.hpp"
 #include "LLVMWarningsPush.hpp"
 #include <llvm/ADT/StringRef.h>
@@ -83,10 +86,10 @@ std::map<StringRef, ImageImplicitArgInfo> BuiltinToArgMap = {
 bool ResolveImageImplicitArgsForBindless::runOnModule(Module &M) {
   mChanged = false;
 
-  CodeGenContext *Ctx = getAnalysis<CodeGenContextWrapper>().getCodeGenContext();
-  if (!Ctx->getModuleMetaData()->extensions.spvINTELBindlessImages)
+  auto *Ctx = getAnalysis<CodeGenContextWrapper>().getCodeGenContext();
+  auto *modMD = Ctx->getModuleMetaData();
+  if (!modMD->extensions.spvINTELBindlessImages)
     return false; // Bindless exclusive pass.
-  mDriverInfo = &Ctx->m_DriverInfo;
 
   visit(M);
 
@@ -119,18 +122,31 @@ void ResolveImageImplicitArgsForBindless::visitCallInst(CallInst &CI) {
   Value *Img = CI.getOperand(0);
   Value *ImgToInt = isa<IntegerType>(Img->getType()) ? Builder.CreateZExtOrTrunc(Img, Builder.getInt64Ty())
                                                      : Builder.CreatePtrToInt(Img, Builder.getInt64Ty());
-  uint64_t SurfaceStateSize = mDriverInfo->getSurfaceStateSize();
-  auto *StateSizeValue = Builder.getInt64(SurfaceStateSize);
+
+  Value *StateSizeValue;
+  if (IGC_GET_FLAG_VALUE(ForceEnableSurfaceStateSizeReloc)) {
+    // When SurfaceStateSize is a relocation patched by the runtime, UMD can choose the surface state size without the
+    // need to recompile the module.
+    auto *PatchNameArg = ConstantDataArray::getString(CI.getContext(), vISA::SURFACE_STATE_SIZE_RELOCATION_NAME, false);
+    Type *PatchTys[] = {Builder.getInt32Ty(), PatchNameArg->getType()};
+    auto *PatchFunc = GenISAIntrinsic::getDeclaration(M, GenISAIntrinsic::GenISA_staticConstantPatchValue, PatchTys);
+    Value *SurfaceStateSizeI32 = Builder.CreateCall(PatchFunc, PatchNameArg, "surfaceStateSize");
+    StateSizeValue = Builder.CreateZExt(SurfaceStateSizeI32, Builder.getInt64Ty());
+  } else {
+    StateSizeValue = Builder.getInt64(64);
+  }
+
+  // Offset Img arg into ImageImplicitArgs bindless slot:
   auto *ImageImplicitArgsOffset = Builder.CreateAdd(ImgToInt, StateSizeValue);
-  Value *ImageImplicitArgs =
+  Value *ImplicitArgsPtr =
       Builder.CreateBitOrPointerCast(ImageImplicitArgsOffset, BindlessOffsetPtrTy, "imageImplicitArgs");
 
   // Create ldraw decl:
-  Type *const Tys[] = {CI.getType(), ImageImplicitArgs->getType()};
+  Type *const Tys[] = {CI.getType(), ImplicitArgsPtr->getType()};
   auto *Callee = GenISAIntrinsic::getDeclaration(M, GenISAIntrinsic::GenISA_ldraw_indexed, Tys);
 
   // Create ldraw call:
-  Value *const Args[] = {ImageImplicitArgs, Builder.getInt32(ArgOffsetInStruct), Builder.getInt32(AlignmentInBytes),
+  Value *const Args[] = {ImplicitArgsPtr, Builder.getInt32(ArgOffsetInStruct), Builder.getInt32(AlignmentInBytes),
                          Builder.getInt1(false)};
   CallInst *Result = Builder.CreateCall(Callee, Args);
   Result->setDebugLoc(CI.getDebugLoc());
diff --git a/IGC/Compiler/Optimizer/OpenCLPasses/BindlessImage/ResolveImageImplicitArgsForBindless.hpp b/IGC/Compiler/Optimizer/OpenCLPasses/BindlessImage/ResolveImageImplicitArgsForBindless.hpp
@@ -9,7 +9,7 @@ SPDX-License-Identifier: MIT
 ///
 /// \file
 /// This file contains the declaration of ResolveImageImplicitArgsForBindless
-/// llvm pass.
+/// LLVM pass.
 ///
 /// This pass searches for built-in calls querying for image properties and
 /// replaces them with calls to GenISA_ldraw_indexed intrinsic to fetch the
@@ -19,18 +19,14 @@ SPDX-License-Identifier: MIT
 
 #pragma once
 
-#include "Compiler/MetaDataUtilsWrapper.h"
 #include "Compiler/CodeGenContextWrapper.hpp"
-
 #include <common/LLVMWarningsPush.hpp>
 #include <llvm/ADT/SetVector.h>
 #include <llvm/Pass.h>
 #include <llvm/IR/InstVisitor.h>
 #include <common/LLVMWarningsPop.hpp>
 
 namespace IGC {
-class CPlatform;
-
 class ResolveImageImplicitArgsForBindless : public llvm::ModulePass,
                                             public llvm::InstVisitor<ResolveImageImplicitArgsForBindless> {
 public:
@@ -51,7 +47,6 @@ class ResolveImageImplicitArgsForBindless : public llvm::ModulePass,
   bool runOnModule(llvm::Module &M) override;
 
 private:
-  CDriverInfo const *mDriverInfo = nullptr;
   llvm::SmallSetVector<llvm::Instruction *, 16> mInstsToRemove;
 
   bool mChanged = false;
diff --git a/IGC/Compiler/Optimizer/OpenCLPasses/ImageFuncs/ResolveSampledImageBuiltins.cpp b/IGC/Compiler/Optimizer/OpenCLPasses/ImageFuncs/ResolveSampledImageBuiltins.cpp
@@ -8,8 +8,12 @@ SPDX-License-Identifier: MIT
 
 #include "Compiler/Optimizer/OpenCLPasses/ImageFuncs/ResolveSampledImageBuiltins.hpp"
 #include "Compiler/IGCPassSupport.h"
+#include "Compiler/CodeGenPublic.h"
+#include "RelocationInfo.h"
+#include "GenISAIntrinsics/GenIntrinsics.h"
 #include "common/MDFrameWork.h"
 #include "common/LLVMWarningsPush.hpp"
+#include <llvm/IR/Constants.h>
 #include <llvm/IR/Function.h>
 #include <llvm/IR/Instructions.h>
 #include "common/LLVMWarningsPop.hpp"
@@ -25,7 +29,7 @@ using namespace IGC;
 #define PASS_CFG_ONLY false
 #define PASS_ANALYSIS false
 IGC_INITIALIZE_PASS_BEGIN(ResolveSampledImageBuiltins, PASS_FLAG, PASS_DESCRIPTION, PASS_CFG_ONLY, PASS_ANALYSIS)
-IGC_INITIALIZE_PASS_DEPENDENCY(MetaDataUtilsWrapper)
+IGC_INITIALIZE_PASS_DEPENDENCY(CodeGenContextWrapper)
 IGC_INITIALIZE_PASS_END(ResolveSampledImageBuiltins, PASS_FLAG, PASS_DESCRIPTION, PASS_CFG_ONLY, PASS_ANALYSIS)
 
 char ResolveSampledImageBuiltins::ID = 0;
@@ -39,7 +43,7 @@ ResolveSampledImageBuiltins::ResolveSampledImageBuiltins() : ModulePass(ID) {
 
 bool ResolveSampledImageBuiltins::runOnModule(Module &M) {
   m_changed = false;
-  modMD = getAnalysis<MetaDataUtilsWrapper>().getModuleMetaData();
+  modMD = getAnalysis<CodeGenContextWrapper>().getCodeGenContext()->getModuleMetaData();
   visit(M);
 
   for (auto builtin : m_builtinsToRemove) {
@@ -138,6 +142,13 @@ Value *ResolveSampledImageBuiltins::lowerGetSampler(CallInst &CI) {
   //   %image_offset = ptrtoint %spirv.SampledImage._void_1_0_0_0_0_0_0 addrspace(1)* %image to i64
   //   %sampler_offset = add i64 %image_offset, 128
   //   %queried_sampler = or i64 %sampler_offset, 1
+  // Or:
+  //   %image_offset = ptrtoint %spirv.SampledImage._void_1_0_0_0_0_0_0 addrspace(1)* %image to i64
+  //   %surfaceStateSize = call i32 @__static_constant_patch_value
+  //   %surfaceStateSize64 = zext i32 %surfaceStateSize to i64
+  //   %twiceSurfaceStatesSize = shl i64 %surfaceStateSize64, 1
+  //   %sampler_offset = add i64 %image_offset, %twiceSurfaceStatesSize
+  //   %queried_sampler = or i64 %sampler_offset, 1
   if (!samplerArg) {
     IGC_ASSERT(modMD->UseBindlessImage);
     IGC_ASSERT(image->getType()->isPointerTy());
@@ -147,12 +158,25 @@ Value *ResolveSampledImageBuiltins::lowerGetSampler(CallInst &CI) {
     // | image state | image implicit args state | sampler state | redescribed image state | ...
     // Sampler state offset is addition of image state offset, size of
     // image state and size of image implicit args state.
-    // Both size of image state and image implicit args state are 64 bytes.
-    constexpr uint64_t surfaceStateSize = 64;
-    auto *stateSizeValue = ConstantInt::get(Int64Ty, surfaceStateSize * 2);
-    auto *samplerOffset = BinaryOperator::CreateAdd(imageOffset, stateSizeValue, "sampler_offset", &CI);
+    Module *M = CI.getParent()->getParent()->getParent();
+    Value *TwoStatesOffset;
+    if (IGC_GET_FLAG_VALUE(ForceEnableSurfaceStateSizeReloc)) {
+      // SurfaceStateSize is a relocation patched by the runtime.
+      auto *PatchNameArg =
+          ConstantDataArray::getString(CI.getContext(), vISA::SURFACE_STATE_SIZE_RELOCATION_NAME, false);
+      Type *PatchTys[] = {Type::getInt32Ty(CI.getContext()), PatchNameArg->getType()};
+      auto *PatchFunc = GenISAIntrinsic::getDeclaration(M, GenISAIntrinsic::GenISA_staticConstantPatchValue, PatchTys);
+      auto *SurfaceStateSizeI32 = CallInst::Create(PatchFunc, PatchNameArg, "surfaceStateSize", &CI);
+      auto *SurfaceStateSizeI64 = CastInst::Create(Instruction::ZExt, SurfaceStateSizeI32, Int64Ty, "", &CI);
+      // Sampler offset = image state offset + 2 * surfaceStateSize
+      TwoStatesOffset =
+          BinaryOperator::Create(Instruction::Shl, SurfaceStateSizeI64, ConstantInt::get(Int64Ty, 1), "", &CI);
+    } else {
+      TwoStatesOffset = ConstantInt::get(Int64Ty, 2 * 64);
+    }
+    auto *SamplerOffset = BinaryOperator::CreateAdd(imageOffset, TwoStatesOffset, "sampler_offset", &CI);
     // Set bit-field 0 to 1 to select Bindless Sampler State Base Address.
-    return BinaryOperator::CreateOr(samplerOffset, ConstantInt::get(Int64Ty, 1), "", &CI);
+    return BinaryOperator::CreateOr(SamplerOffset, ConstantInt::get(Int64Ty, 1), "", &CI);
   }
 
   // Transforms the following sequence:
diff --git a/IGC/Compiler/Optimizer/OpenCLPasses/ImageFuncs/ResolveSampledImageBuiltins.hpp b/IGC/Compiler/Optimizer/OpenCLPasses/ImageFuncs/ResolveSampledImageBuiltins.hpp
@@ -9,6 +9,7 @@ SPDX-License-Identifier: MIT
 #pragma once
 
 #include "Compiler/MetaDataUtilsWrapper.h"
+#include "Compiler/CodeGenContextWrapper.hpp"
 #include "common/LLVMWarningsPush.hpp"
 #include <llvm/Pass.h>
 #include <llvm/IR/InstVisitor.h>
@@ -32,7 +33,7 @@ class ResolveSampledImageBuiltins : public llvm::ModulePass, public llvm::InstVi
 
   virtual llvm::StringRef getPassName() const override { return "ResolveSampledImageBuiltins"; }
 
-  virtual void getAnalysisUsage(llvm::AnalysisUsage &AU) const override { AU.addRequired<MetaDataUtilsWrapper>(); }
+  virtual void getAnalysisUsage(llvm::AnalysisUsage &AU) const override { AU.addRequired<CodeGenContextWrapper>(); }
 
   virtual bool runOnModule(llvm::Module &M) override;
   void visitCallInst(llvm::CallInst &CI);
diff --git a/IGC/Compiler/tests/ResolveImageImplicitArgsForBindless/basic.ll b/IGC/Compiler/tests/ResolveImageImplicitArgsForBindless/basic.ll
diff --git a/IGC/Compiler/tests/ResolveImageImplicitArgsForBindless/exit-for-bindful.ll b/IGC/Compiler/tests/ResolveImageImplicitArgsForBindless/exit-for-bindful.ll
diff --git a/IGC/Compiler/tests/ResolveImageImplicitArgsForBindless/reloc.ll b/IGC/Compiler/tests/ResolveImageImplicitArgsForBindless/reloc.ll
@@ -0,0 +1,38 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+;=========================== begin_copyright_notice ============================
+;
+; Copyright (C) 2026 Intel Corporation
+;
+; SPDX-License-Identifier: MIT
+;
+;============================ end_copyright_notice =============================
+
+; REQUIRES: regkeys
+; RUN: igc_opt -resolve-image-implicit-args-for-bindless --opaque-pointers -regkey ForceEnableSurfaceStateSizeReloc=1  -S %s | FileCheck %s
+
+target triple = "spir64-unknown-unknown"
+
+%spirv.Sampler = type opaque
+
+
+define spir_func void @test_width(i64 %bindless_img) {
+; CHECK-LABEL: @test_width(
+; CHECK-NEXT:    [[SURFACESTATESIZE:%.*]] = call i32 @llvm.genx.GenISA.staticConstantPatchValue.i32.a26i8([26 x i8] c"__INTEL_SURFACE_STATE_SIZE")
+; CHECK-NEXT:    [[TMP1:%.*]] = zext i32 [[SURFACESTATESIZE]] to i64
+; CHECK-NEXT:    [[TMP2:%.*]] = add i64 [[BINDLESS_IMG:%.*]], [[TMP1]]
+; CHECK-NEXT:    [[IMAGEIMPLICITARGS:%.*]] = inttoptr i64 [[TMP2]] to ptr addrspace(393468)
+; CHECK-NEXT:    [[TMP3:%.*]] = call i32 @llvm.genx.GenISA.ldraw.indexed.i32.p393468(ptr addrspace(393468) [[IMAGEIMPLICITARGS]], i32 8, i32 8, i1 false)
+; CHECK-NEXT:    ret void
+;
+  %1 = call i32 @__builtin_IB_get_image_width(i64 %bindless_img)
+  ret void
+}
+
+declare spir_func i32 @__builtin_IB_get_image_width(i64 noundef) local_unnamed_addr
+
+!IGCMetadata = !{!0}
+
+!0 = !{!"ModuleMD", !1, !2}
+!1 = !{!"UseBindlessImage", i1 true}
+!2 = !{!"extensions", !3}
+!3 = !{!"spvINTELBindlessImages", i1 true}
diff --git a/IGC/Compiler/tests/ResolveImageImplicitArgsForBindless/typed.ll b/IGC/Compiler/tests/ResolveImageImplicitArgsForBindless/typed.ll
diff --git a/IGC/common/igc_flags.h b/IGC/common/igc_flags.h
@@ -1230,6 +1230,8 @@ DECLARE_IGC_REGKEY(bool, DisableStatefulFolding, false,
                    "Turns off all folding for stateful messages (imm offset, scaling, and surface state idx)", false)
 DECLARE_IGC_REGKEY(bool, EnableStatefulScaleFolding, false, "Enables folding of shl into the scale of a stateful send",
                    false)
+DECLARE_IGC_REGKEY(bool, ForceEnableSurfaceStateSizeReloc, false,
+                   "Forces emission of surface state size as a relocation in compute path.", true)
 DECLARE_IGC_REGKEY(bool, DisableLSCForTypedUAV, false,
                    "Forces legacy HDC messages for typed UAV read/write."
                    "Temporary knob for XE2 bringup.",
diff --git a/visa/include/RelocationInfo.h b/visa/include/RelocationInfo.h
@@ -27,6 +27,8 @@ constexpr const char *CROSS_THREAD_OFF_R0_RELOCATION_NAME =
     "__INTEL_PATCH_CROSS_THREAD_OFFSET_OFF_R0";
 constexpr const char *PER_THREAD_OFF_RELOCATION_NAME =
     "__INTEL_PER_THREAD_OFF";
+constexpr const char *SURFACE_STATE_SIZE_RELOCATION_NAME =
+    "__INTEL_SURFACE_STATE_SIZE";
 
 /// GenSymType - Specify the symbol's type
 enum GenSymType {

Original file line number	Diff line number	Diff line change
`@@ -365,6 +365,9 @@ void ZEBinaryBuilder::addRuntimeSymbols(const IGC::SOpenCLProgramInfo &annotatio`
`365`	`365`	`if (annotations.m_hasPerThreadOffsetRelocations)`
`366`	`366`	`mBuilder.addSymbol(vISA::PER_THREAD_OFF_RELOCATION_NAME, /addr/ 0, /size/ 0, llvm::ELF::STB_GLOBAL,`
`367`	`367`	`llvm::ELF::STT_NOTYPE, /sectionId/ -1);`
	`368`	`+ if (annotations.m_hasSurfaceStateSizeRelocations)`
	`369`	`+ mBuilder.addSymbol(vISA::SURFACE_STATE_SIZE_RELOCATION_NAME, /addr/ 0, /size/ 0, llvm::ELF::STB_GLOBAL,`
	`370`	`+ llvm::ELF::STT_OBJECT, /sectionId/ -1);`
`368`	`371`	`}`
`369`	`372`
`370`	`373`	`void ZEBinaryBuilder::addProgramSymbols(const IGC::SOpenCLProgramInfo &annotations) {`
Original file line number	Diff line number	Diff line change
`@@ -6125,6 +6125,9 @@ void CEncoder::createRelocationTables(VISAKernel &pMainKernel) {`
`6125`	`6125`	`} else if (reloc.r_symbol == vISA::PER_THREAD_OFF_RELOCATION_NAME) {`
`6126`	`6126`	`auto cl_context = static_cast<OpenCLProgramContext *>(context);`
`6127`	`6127`	`cl_context->m_programInfo.m_hasPerThreadOffsetRelocations = true;`
	`6128`	`+ } else if (reloc.r_symbol == vISA::SURFACE_STATE_SIZE_RELOCATION_NAME) {`
	`6129`	`+ auto cl_context = static_cast<OpenCLProgramContext *>(context);`
	`6130`	`+ cl_context->m_programInfo.m_hasSurfaceStateSizeRelocations = true;`
`6128`	`6131`	`}`
`6129`	`6132`	`}`
`6130`	`6133`	`}`