[SYCL][Fusion] Add HIP support

clang/lib/Driver/Driver.cpp

@@ -5827,8 +5827,9 @@ class OffloadingActionBuilder final {
             DA.add(*DeviceWrappingAction, *TC, BoundArch, Action::OFK_SYCL);
             continue;
           }
-          if (isNVPTX && Args.hasArg(options::OPT_fsycl_embed_ir)) {
-            // When compiling for Nvidia/CUDA devices and the user requested the
+          if ((isNVPTX || isAMDGCN) &&
+              Args.hasArg(options::OPT_fsycl_embed_ir)) {
+            // When compiling for Nvidia/AMD devices and the user requested the
             // IR to be embedded in the application (via option), run the output
             // of sycl-post-link (filetable referencing LLVM Bitcode + symbols)
             // through the offload wrapper and link the resulting object to the

sycl-fusion/common/include/Kernel.h

@@ -20,6 +20,30 @@ namespace jit_compiler {
 
 using BinaryAddress = const uint8_t *;
 
+/// Possible barrier flags
+enum class BarrierFlags : uint32_t {
+  None = 0,   // Do not insert barrier
+  Local = 1,  // Ensure correct ordering of memory operations to local memory
+  Global = 2, // Ensure correct ordering of memory operations to global memory
+  LocalAndGlobal = Local | Global
+};
+
+constexpr BarrierFlags getNoBarrierFlag() { return BarrierFlags::None; }
+constexpr BarrierFlags getLocalAndGlobalBarrierFlag() {
+  return BarrierFlags::LocalAndGlobal;
+}
+constexpr bool isNoBarrierFlag(BarrierFlags Flag) {
+  return Flag == BarrierFlags::None;
+}
+constexpr bool hasLocalBarrierFlag(BarrierFlags Flag) {
+  return static_cast<uint32_t>(Flag) &
+         static_cast<uint32_t>(BarrierFlags::Local);
+}
+constexpr bool hasGlobalBarrierFlag(BarrierFlags Flag) {
+  return static_cast<uint32_t>(Flag) &
+         static_cast<uint32_t>(BarrierFlags::Global);
+}
+
 ///
 /// Enumerate possible kinds of parameters.
 /// 1:1 correspondence with the definition in kernel_desc.hpp in the DPC++ SYCL
@@ -35,7 +59,7 @@ enum class ParameterKind : uint32_t {
 };
 
 /// Different binary formats supported as input to the JIT compiler.
-enum class BinaryFormat : uint32_t { INVALID, LLVM, SPIRV, PTX };
+enum class BinaryFormat : uint32_t { INVALID, LLVM, SPIRV, PTX, AMDGCN };
 
 /// Information about a device intermediate representation module (e.g., SPIR-V,
 /// LLVM IR) from DPC++.

sycl-fusion/common/lib/KernelIO.h

@@ -48,6 +48,7 @@ template <> struct ScalarEnumerationTraits<jit_compiler::BinaryFormat> {
     IO.enumCase(BF, "LLVM", jit_compiler::BinaryFormat::LLVM);
     IO.enumCase(BF, "SPIRV", jit_compiler::BinaryFormat::SPIRV);
     IO.enumCase(BF, "PTX", jit_compiler::BinaryFormat::PTX);
+    IO.enumCase(BF, "AMDGCN", jit_compiler::BinaryFormat::AMDGCN);
     IO.enumCase(BF, "INVALID", jit_compiler::BinaryFormat::INVALID);
   }
 };

sycl-fusion/jit-compiler/CMakeLists.txt

@@ -9,6 +9,9 @@ add_llvm_library(sycl-fusion
    lib/fusion/ModuleHelper.cpp
    lib/helper/ConfigHelper.cpp
 
+   DEPENDS
+   intrinsics_gen
+
    LINK_COMPONENTS
    BitReader
    Core
@@ -50,6 +53,10 @@ if("NVPTX" IN_LIST LLVM_TARGETS_TO_BUILD)
   target_compile_definitions(sycl-fusion PRIVATE FUSION_JIT_SUPPORT_PTX)
 endif()
 
+if("AMDGPU" IN_LIST LLVM_TARGETS_TO_BUILD)
+  target_compile_definitions(sycl-fusion PRIVATE FUSION_JIT_SUPPORT_AMDGCN)
+endif()
+
 if (BUILD_SHARED_LIBS)
   if(NOT MSVC AND NOT APPLE)
     # Manage symbol visibility through the linker to make sure no LLVM symbols

sycl-fusion/jit-compiler/include/JITContext.h

@@ -19,6 +19,7 @@
 
 #include "Hashing.h"
 #include "Kernel.h"
+#include "Options.h"
 #include "Parameter.h"
 
 namespace llvm {
@@ -28,7 +29,7 @@ class LLVMContext;
 namespace jit_compiler {
 
 using CacheKeyT =
-    std::tuple<std::vector<std::string>, ParamIdentList, int,
+    std::tuple<std::vector<std::string>, ParamIdentList, BarrierFlags,
                std::vector<ParameterInternalization>, std::vector<JITConstant>,
                // This field of the cache is optional because, if all of the
                // ranges are equal, we will perform no remapping, so that fused

sycl-fusion/jit-compiler/include/KernelFusion.h

@@ -60,7 +60,8 @@ class KernelFusion {
               const std::vector<SYCLKernelInfo> &KernelInformation,
               const std::vector<std::string> &KernelsToFuse,
               const std::string &FusedKernelName,
-              jit_compiler::ParamIdentList &Identities, int BarriersFlags,
+              jit_compiler::ParamIdentList &Identities,
+              BarrierFlags BarriersFlags,
               const std::vector<jit_compiler::ParameterInternalization>
                   &Internalization,
               const std::vector<jit_compiler::JITConstant> &JITConstants);

sycl-fusion/jit-compiler/lib/KernelFusion.cpp

@@ -58,6 +58,13 @@ static bool isTargetFormatSupported(BinaryFormat TargetFormat) {
 #else  // FUSION_JIT_SUPPORT_PTX
     return false;
 #endif // FUSION_JIT_SUPPORT_PTX
+  }
+  case BinaryFormat::AMDGCN: {
+#ifdef FUSION_JIT_SUPPORT_AMDGCN
+    return true;
+#else  // FUSION_JIT_SUPPORT_AMDGCN
+    return false;
+#endif // FUSION_JIT_SUPPORT_AMDGCN
   }
   default:
     return false;
@@ -69,7 +76,7 @@ FusionResult KernelFusion::fuseKernels(
     const std::vector<SYCLKernelInfo> &KernelInformation,
     const std::vector<std::string> &KernelsToFuse,
     const std::string &FusedKernelName, ParamIdentList &Identities,
-    int BarriersFlags,
+    BarrierFlags BarriersFlags,
     const std::vector<jit_compiler::ParameterInternalization> &Internalization,
     const std::vector<jit_compiler::JITConstant> &Constants) {
   // Initialize the configuration helper to make the options for this invocation
@@ -93,8 +100,11 @@ FusionResult KernelFusion::fuseKernels(
         "Fusion output target format not supported by this build");
   }
 
-  if (TargetFormat == BinaryFormat::PTX && IsHeterogeneousList) {
-    return FusionResult{"Heterogeneous ND ranges not supported for CUDA"};
+  if ((TargetFormat == BinaryFormat::PTX ||
+       TargetFormat == BinaryFormat::AMDGCN) &&
+      IsHeterogeneousList) {
+    return FusionResult{
+        "Heterogeneous ND ranges not supported for CUDA and HIP"};
   }
 
   bool CachingEnabled = ConfigHelper::get<option::JITEnableCaching>();

sycl-fusion/jit-compiler/lib/fusion/FusionPipeline.cpp

@@ -40,7 +40,7 @@ static unsigned getFlatAddressSpace(Module &Mod) {
   // Ideally, we could get this information from the TargetTransformInfo, but
   // the SPIR-V backend does not yet seem to have an implementation for that.
   llvm::Triple Tri(Mod.getTargetTriple());
-  if (Tri.isNVPTX()) {
+  if (Tri.isNVPTX() || Tri.isAMDGCN()) {
     return 0;
   }
   if (Tri.isSPIRV() || Tri.isSPIR()) {
@@ -53,7 +53,7 @@ static unsigned getFlatAddressSpace(Module &Mod) {
 
 std::unique_ptr<SYCLModuleInfo>
 FusionPipeline::runFusionPasses(Module &Mod, SYCLModuleInfo &InputInfo,
-                                int BarriersFlags) {
+                                BarrierFlags BarriersFlags) {
   // Perform the actual kernel fusion, i.e., generate a kernel function for the
   // fused kernel from the kernel functions of the input kernels. This is done
   // by the SYCLKernelFusion LLVM pass, which is run here through a custom LLVM

sycl-fusion/jit-compiler/lib/fusion/FusionPipeline.h

@@ -25,7 +25,7 @@ class FusionPipeline {
   /// contain an entry for the fused kernel.
   static std::unique_ptr<SYCLModuleInfo>
   runFusionPasses(llvm::Module &Mod, SYCLModuleInfo &InputInfo,
-                  int BarriersFlags);
+                  BarrierFlags BarriersFlags);
 };
 } // namespace fusion
 } // namespace jit_compiler

sycl-fusion/jit-compiler/lib/translation/KernelTranslation.cpp

@@ -191,6 +191,15 @@ llvm::Error KernelTranslator::translateKernel(SYCLKernelInfo &Kernel,
     KernelBin = *BinaryOrError;
     break;
   }
+  case BinaryFormat::AMDGCN: {
+    llvm::Expected<KernelBinary *> BinaryOrError =
+        translateToAMDGCN(Kernel, Mod, JITCtx);
+    if (auto Error = BinaryOrError.takeError()) {
+      return Error;
+    }
+    KernelBin = *BinaryOrError;
+    break;
+  }
   default: {
     return createStringError(
         inconvertibleErrorCode(),
@@ -287,3 +296,76 @@ KernelTranslator::translateToPTX(SYCLKernelInfo &KernelInfo, llvm::Module &Mod,
   return &JITCtx.emplaceKernelBinary(std::move(PTXASM), BinaryFormat::PTX);
 #endif // FUSION_JIT_SUPPORT_PTX
 }
+
+llvm::Expected<KernelBinary *>
+KernelTranslator::translateToAMDGCN(SYCLKernelInfo &KernelInfo,
+                                    llvm::Module &Mod, JITContext &JITCtx) {
+#ifndef FUSION_JIT_SUPPORT_AMDGCN
+  (void)KernelInfo;
+  (void)Mod;
+  (void)JITCtx;
+  return createStringError(inconvertibleErrorCode(),
+                           "AMDGPU translation not supported in this build");
+#else  // FUSION_JIT_SUPPORT_AMDGCN
+
+  LLVMInitializeAMDGPUTargetInfo();
+  LLVMInitializeAMDGPUTarget();
+  LLVMInitializeAMDGPUAsmPrinter();
+  LLVMInitializeAMDGPUTargetMC();
+
+  static const char *TARGET_CPU_ATTRIBUTE = "target-cpu";
+  static const char *TARGET_FEATURE_ATTRIBUTE = "target-features";
+
+  std::string TargetTriple{"amdgcn-amd-amdhsa"};
+
+  std::string ErrorMessage;
+  const auto *Target =
+      llvm::TargetRegistry::lookupTarget(TargetTriple, ErrorMessage);
+
+  if (!Target) {
+    return createStringError(
+        inconvertibleErrorCode(),
+        "Failed to load and translate AMDGCN LLVM IR module with error %s",
+        ErrorMessage.c_str());
+  }
+
+  // Set to the lowest tested target according to the GetStartedGuide, section
+  // "Build DPC++ toolchain with support for HIP AMD"
+  llvm::StringRef TargetCPU{"gfx906"};
+  llvm::StringRef TargetFeatures{""};
+  if (auto *KernelFunc = Mod.getFunction(KernelInfo.Name)) {
+    if (KernelFunc->hasFnAttribute(TARGET_CPU_ATTRIBUTE)) {
+      TargetCPU =
+          KernelFunc->getFnAttribute(TARGET_CPU_ATTRIBUTE).getValueAsString();
+    }
+    if (KernelFunc->hasFnAttribute(TARGET_FEATURE_ATTRIBUTE)) {
+      TargetFeatures = KernelFunc->getFnAttribute(TARGET_FEATURE_ATTRIBUTE)
+                           .getValueAsString();
+    }
+  }
+
+  // FIXME: Check whether we can provide more accurate target information here
+  auto *TargetMachine = Target->createTargetMachine(
+      TargetTriple, TargetCPU, TargetFeatures, {}, llvm::Reloc::PIC_,
+      std::nullopt, llvm::CodeGenOpt::Default);
+
+  std::string AMDObj;
+  {
+    llvm::legacy::PassManager PM;
+    llvm::raw_string_ostream OBJStream{AMDObj};
+    llvm::buffer_ostream BufferedOBJ{OBJStream};
+
+    if (TargetMachine->addPassesToEmitFile(PM, BufferedOBJ, nullptr,
+                                           llvm::CGFT_ObjectFile)) {
+      return createStringError(
+          inconvertibleErrorCode(),
+          "Failed to construct pass pipeline to emit output");
+    }
+
+    PM.run(Mod);
+    OBJStream.flush();
+  }
+
+  return &JITCtx.emplaceKernelBinary(std::move(AMDObj), BinaryFormat::AMDGCN);
+#endif // FUSION_JIT_SUPPORT_AMDGCN
+}

sycl-fusion/jit-compiler/lib/translation/KernelTranslation.h

@@ -43,6 +43,10 @@ class KernelTranslator {
 
   static llvm::Expected<KernelBinary *>
   translateToPTX(SYCLKernelInfo &Kernel, llvm::Module &Mod, JITContext &JITCtx);
+
+  static llvm::Expected<KernelBinary *>
+  translateToAMDGCN(SYCLKernelInfo &KernelInfo, llvm::Module &Mod,
+                    JITContext &JITCtx);
 };
 } // namespace translation
 } // namespace jit_compiler

sycl-fusion/passes/CMakeLists.txt

@@ -30,6 +30,10 @@ if("NVPTX" IN_LIST LLVM_TARGETS_TO_BUILD)
   target_compile_definitions(SYCLKernelFusion PRIVATE FUSION_JIT_SUPPORT_PTX)
 endif()
 
+if("AMDGPU" IN_LIST LLVM_TARGETS_TO_BUILD)
+  target_compile_definitions(SYCLKernelFusion PRIVATE FUSION_JIT_SUPPORT_AMDGCN)
+endif()
+
 # Static library for linking with the jit_compiler
 add_llvm_library(SYCLKernelFusionPasses
   SYCLFusionPasses.cpp
@@ -68,3 +72,7 @@ target_link_libraries(SYCLKernelFusionPasses
 if("NVPTX" IN_LIST LLVM_TARGETS_TO_BUILD)
   target_compile_definitions(SYCLKernelFusionPasses PRIVATE FUSION_JIT_SUPPORT_PTX)
 endif()
+
+if("AMDGPU" IN_LIST LLVM_TARGETS_TO_BUILD)
+  target_compile_definitions(SYCLKernelFusionPasses PRIVATE FUSION_JIT_SUPPORT_AMDGCN)
+endif()

sycl-fusion/passes/SYCLFusionPasses.cpp

@@ -9,12 +9,15 @@
 #include "llvm/Passes/PassBuilder.h"
 #include "llvm/Passes/PassPlugin.h"
 
+#include "Kernel.h"
+
 #include "internalization/Internalization.h"
 #include "kernel-fusion/SYCLKernelFusion.h"
 #include "kernel-info/SYCLKernelInfo.h"
 #include "syclcp/SYCLCP.h"
 
 using namespace llvm;
+using namespace jit_compiler;
 
 cl::opt<bool>
     NoBarriers("sycl-kernel-fusion-no-barriers",
@@ -28,8 +31,9 @@ llvm::PassPluginLibraryInfo getSYCLKernelFusionPluginInfo() {
             [](StringRef Name, ModulePassManager &MPM,
                ArrayRef<PassBuilder::PipelineElement>) {
               if (Name == "sycl-kernel-fusion") {
-                int BarrierFlag =
-                    (NoBarriers) ? -1 : SYCLKernelFusion::DefaultBarriersFlags;
+                BarrierFlags BarrierFlag =
+                    (NoBarriers) ? getNoBarrierFlag()
+                                 : SYCLKernelFusion::DefaultBarriersFlags;
                 MPM.addPass(SYCLKernelFusion(BarrierFlag));
                 return true;
               }

sycl-fusion/passes/kernel-fusion/SYCLKernelFusion.cpp

@@ -230,13 +230,12 @@ static FusionInsertPoints addGuard(IRBuilderBase &Builder,
   return {Entry, CallInsertion, Exit};
 }
 
-static Expected<CallInst *>
-createFusionCall(IRBuilderBase &Builder, Function *F,
-                 ArrayRef<Value *> CallArgs,
-                 const jit_compiler::NDRange &SrcNDRange,
-                 const jit_compiler::NDRange &FusedNDRange, bool IsLast,
-                 int BarriersFlags, jit_compiler::Remapper &Remapper,
-                 bool ShouldRemap, TargetFusionInfo &TargetInfo) {
+static Expected<CallInst *> createFusionCall(
+    IRBuilderBase &Builder, Function *F, ArrayRef<Value *> CallArgs,
+    const jit_compiler::NDRange &SrcNDRange,
+    const jit_compiler::NDRange &FusedNDRange, bool IsLast,
+    jit_compiler::BarrierFlags BarriersFlags, jit_compiler::Remapper &Remapper,
+    bool ShouldRemap, TargetFusionInfo &TargetInfo) {
   const auto IPs = addGuard(Builder, SrcNDRange, FusedNDRange, IsLast);
 
   if (ShouldRemap) {
@@ -261,7 +260,7 @@ createFusionCall(IRBuilderBase &Builder, Function *F,
   Builder.SetInsertPoint(IPs.Exit);
 
   // Insert barrier if needed
-  if (!IsLast && BarriersFlags > 0) {
+  if (!IsLast && !jit_compiler::isNoBarrierFlag(BarriersFlags)) {
     TargetInfo.createBarrierCall(Builder, BarriersFlags);
   }
 

sycl-fusion/passes/kernel-fusion/SYCLKernelFusion.h

@@ -35,7 +35,7 @@ class SYCLKernelFusion : public llvm::PassInfoMixin<SYCLKernelFusion> {
   constexpr static llvm::StringLiteral NDRangesMDKey{"sycl.kernel.nd-ranges"};
 
   constexpr SYCLKernelFusion() = default;
-  constexpr explicit SYCLKernelFusion(int BarriersFlags)
+  constexpr explicit SYCLKernelFusion(jit_compiler::BarrierFlags BarriersFlags)
       : BarriersFlags{BarriersFlags} {}
 
   llvm::PreservedAnalyses run(llvm::Module &M, llvm::ModuleAnalysisManager &AM);
@@ -45,7 +45,8 @@ class SYCLKernelFusion : public llvm::PassInfoMixin<SYCLKernelFusion> {
   ///
   /// By default, correct ordering of memory operations to global memory is
   /// ensured.
-  constexpr static int DefaultBarriersFlags{3};
+  constexpr static jit_compiler::BarrierFlags DefaultBarriersFlags{
+      jit_compiler::getLocalAndGlobalBarrierFlag()};
 
 private:
   // This needs to be in sync with the metadata kind
@@ -155,11 +156,7 @@ class SYCLKernelFusion : public llvm::PassInfoMixin<SYCLKernelFusion> {
   ///
   /// Flags to apply to the barrier to be introduced between fused kernels.
   ///
-  /// Possible values:
-  /// - -1: Do not insert barrier
-  /// - 1: ensure correct ordering of memory operations to local memory
-  /// - 2: ensure correct ordering of memory operations to global memory
-  const int BarriersFlags{DefaultBarriersFlags};
+  const jit_compiler::BarrierFlags BarriersFlags{DefaultBarriersFlags};
 
   ///
   /// Merge the content of Other into Attributes, adding, removing or updating