Revert [mlir][Target] Improve ROCDL gpu serialization API #95790
+136
−193
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This reverts commit 954cb5f.
|
@llvm/pr-subscribers-mlir-llvm @llvm/pr-subscribers-mlir-gpu Author: Fabian Mora (fabianmcg) ChangesReverts llvm/llvm-project#95456 Patch is 20.39 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/95790.diff 4 Files Affected:
diff --git a/mlir/include/mlir/Target/LLVM/ROCDL/Utils.h b/mlir/include/mlir/Target/LLVM/ROCDL/Utils.h
index 44c9ded317fa5..374fa65bd02e3 100644
--- a/mlir/include/mlir/Target/LLVM/ROCDL/Utils.h
+++ b/mlir/include/mlir/Target/LLVM/ROCDL/Utils.h
@@ -27,19 +27,6 @@ namespace ROCDL {
/// 5. Returns an empty string.
StringRef getROCMPath();
-/// Helper enum for specifying the AMD GCN device libraries required for
-/// compilation.
-enum class AMDGCNLibraries : uint32_t {
- None = 0,
- Ockl = 1,
- Ocml = 2,
- OpenCL = 4,
- Hip = 8,
- LastLib = Hip,
- LLVM_MARK_AS_BITMASK_ENUM(LastLib),
- All = (LastLib << 1) - 1
-};
-
/// Base class for all ROCDL serializations from GPU modules into binary
/// strings. By default this class serializes into LLVM bitcode.
class SerializeGPUModuleBase : public LLVM::ModuleToObject {
@@ -62,8 +49,8 @@ class SerializeGPUModuleBase : public LLVM::ModuleToObject {
/// Returns the bitcode files to be loaded.
ArrayRef<std::string> getFileList() const;
- /// Appends standard ROCm device libraries to `fileList`.
- LogicalResult appendStandardLibs(AMDGCNLibraries libs);
+ /// Appends standard ROCm device libraries like `ocml.bc`, `ockl.bc`, etc.
+ LogicalResult appendStandardLibs();
/// Loads the bitcode files in `fileList`.
virtual std::optional<SmallVector<std::unique_ptr<llvm::Module>>>
@@ -76,20 +63,15 @@ class SerializeGPUModuleBase : public LLVM::ModuleToObject {
LogicalResult handleBitcodeFile(llvm::Module &module) override;
protected:
- /// Adds `oclc` control variables to the LLVM module.
- void addControlVariables(llvm::Module &module, AMDGCNLibraries libs,
- bool wave64, bool daz, bool finiteOnly,
- bool unsafeMath, bool fastMath, bool correctSqrt,
- StringRef abiVer);
+ /// Appends the paths of common ROCm device libraries to `libs`.
+ LogicalResult getCommonBitcodeLibs(llvm::SmallVector<std::string> &libs,
+ SmallVector<char, 256> &libPath,
+ StringRef isaVersion);
- /// Compiles assembly to a binary.
- virtual std::optional<SmallVector<char, 0>>
- compileToBinary(const std::string &serializedISA);
-
- /// Default implementation of `ModuleToObject::moduleToObject`.
- std::optional<SmallVector<char, 0>>
- moduleToObjectImpl(const gpu::TargetOptions &targetOptions,
- llvm::Module &llvmModule);
+ /// Adds `oclc` control variables to the LLVM module.
+ void addControlVariables(llvm::Module &module, bool wave64, bool daz,
+ bool finiteOnly, bool unsafeMath, bool fastMath,
+ bool correctSqrt, StringRef abiVer);
/// Returns the assembled ISA.
std::optional<SmallVector<char, 0>> assembleIsa(StringRef isa);
@@ -102,9 +84,6 @@ class SerializeGPUModuleBase : public LLVM::ModuleToObject {
/// List of LLVM bitcode files to link to.
SmallVector<std::string> fileList;
-
- /// AMD GCN libraries to use when linking, the default is using none.
- AMDGCNLibraries deviceLibs = AMDGCNLibraries::None;
};
} // namespace ROCDL
} // namespace mlir
diff --git a/mlir/lib/Dialect/GPU/CMakeLists.txt b/mlir/lib/Dialect/GPU/CMakeLists.txt
index 08c8aea36fac9..61ab298ebfb98 100644
--- a/mlir/lib/Dialect/GPU/CMakeLists.txt
+++ b/mlir/lib/Dialect/GPU/CMakeLists.txt
@@ -106,7 +106,7 @@ if(MLIR_ENABLE_ROCM_CONVERSIONS)
"Building mlir with ROCm support requires the AMDGPU backend")
endif()
- set(DEFAULT_ROCM_PATH "" CACHE PATH "Fallback path to search for ROCm installs")
+ set(DEFAULT_ROCM_PATH "/opt/rocm" CACHE PATH "Fallback path to search for ROCm installs")
target_compile_definitions(obj.MLIRGPUTransforms
PRIVATE
__DEFAULT_ROCM_PATH__="${DEFAULT_ROCM_PATH}"
diff --git a/mlir/lib/Target/LLVM/CMakeLists.txt b/mlir/lib/Target/LLVM/CMakeLists.txt
index 4393ff1775ef9..5a3fa160850b4 100644
--- a/mlir/lib/Target/LLVM/CMakeLists.txt
+++ b/mlir/lib/Target/LLVM/CMakeLists.txt
@@ -123,12 +123,17 @@ add_mlir_dialect_library(MLIRROCDLTarget
)
if(MLIR_ENABLE_ROCM_CONVERSIONS)
+ if (NOT ("AMDGPU" IN_LIST LLVM_TARGETS_TO_BUILD))
+ message(SEND_ERROR
+ "Building mlir with ROCm support requires the AMDGPU backend")
+ endif()
+
if (DEFINED ROCM_PATH)
set(DEFAULT_ROCM_PATH "${ROCM_PATH}" CACHE PATH "Fallback path to search for ROCm installs")
elseif(DEFINED ENV{ROCM_PATH})
set(DEFAULT_ROCM_PATH "$ENV{ROCM_PATH}" CACHE PATH "Fallback path to search for ROCm installs")
else()
- set(DEFAULT_ROCM_PATH "" CACHE PATH "Fallback path to search for ROCm installs")
+ set(DEFAULT_ROCM_PATH "/opt/rocm" CACHE PATH "Fallback path to search for ROCm installs")
endif()
message(VERBOSE "MLIR Default ROCM toolkit path: ${DEFAULT_ROCM_PATH}")
diff --git a/mlir/lib/Target/LLVM/ROCDL/Target.cpp b/mlir/lib/Target/LLVM/ROCDL/Target.cpp
index fdf1c93b37210..cc13e5b7436ea 100644
--- a/mlir/lib/Target/LLVM/ROCDL/Target.cpp
+++ b/mlir/lib/Target/LLVM/ROCDL/Target.cpp
@@ -17,6 +17,9 @@
#include "mlir/Dialect/LLVMIR/ROCDLDialect.h"
#include "mlir/Support/FileUtilities.h"
#include "mlir/Target/LLVM/ROCDL/Utils.h"
+#include "mlir/Target/LLVMIR/Dialect/GPU/GPUToLLVMIRTranslation.h"
+#include "mlir/Target/LLVMIR/Dialect/LLVMIR/LLVMToLLVMIRTranslation.h"
+#include "mlir/Target/LLVMIR/Dialect/ROCDL/ROCDLToLLVMIRTranslation.h"
#include "mlir/Target/LLVMIR/Export.h"
#include "llvm/IR/Constants.h"
@@ -109,9 +112,8 @@ SerializeGPUModuleBase::SerializeGPUModuleBase(
if (auto file = dyn_cast<StringAttr>(attr))
fileList.push_back(file.str());
- // By default add all libraries if the toolkit path is not empty.
- if (!getToolkitPath().empty())
- deviceLibs = AMDGCNLibraries::All;
+ // Append standard ROCm device bitcode libraries to the files to be loaded.
+ (void)appendStandardLibs();
}
void SerializeGPUModuleBase::init() {
@@ -136,67 +138,29 @@ ArrayRef<std::string> SerializeGPUModuleBase::getFileList() const {
return fileList;
}
-LogicalResult SerializeGPUModuleBase::appendStandardLibs(AMDGCNLibraries libs) {
- if (libs == AMDGCNLibraries::None)
- return success();
+LogicalResult SerializeGPUModuleBase::appendStandardLibs() {
StringRef pathRef = getToolkitPath();
- // Fail if the toolkit is empty.
- if (pathRef.empty())
- return failure();
-
- // Get the path for the device libraries
- SmallString<256> path;
- path.insert(path.begin(), pathRef.begin(), pathRef.end());
- llvm::sys::path::append(path, "amdgcn", "bitcode");
- pathRef = StringRef(path.data(), path.size());
-
- // Fail if the path is invalid.
- if (!llvm::sys::fs::is_directory(pathRef)) {
- getOperation().emitRemark() << "ROCm amdgcn bitcode path: " << pathRef
- << " does not exist or is not a directory.";
- return failure();
- }
-
- // Get the ISA version.
- StringRef isaVersion =
- llvm::AMDGPU::getArchNameAMDGCN(llvm::AMDGPU::parseArchAMDGCN(chip));
- isaVersion.consume_front("gfx");
-
- // Helper function for adding a library.
- auto addLib = [&](const Twine &lib) -> bool {
- auto baseSize = path.size();
- llvm::sys::path::append(path, lib);
- StringRef pathRef(path.data(), path.size());
- if (!llvm::sys::fs::is_regular_file(pathRef)) {
- getOperation().emitRemark() << "Bitcode library path: " << pathRef
- << " does not exist or is not a file.\n";
- return true;
+ if (!pathRef.empty()) {
+ SmallVector<char, 256> path;
+ path.insert(path.begin(), pathRef.begin(), pathRef.end());
+ llvm::sys::path::append(path, "amdgcn", "bitcode");
+ pathRef = StringRef(path.data(), path.size());
+ if (!llvm::sys::fs::is_directory(pathRef)) {
+ getOperation().emitRemark() << "ROCm amdgcn bitcode path: " << pathRef
+ << " does not exist or is not a directory.";
+ return failure();
}
- fileList.push_back(pathRef.str());
- path.truncate(baseSize);
- return false;
- };
-
- // Add ROCm device libraries. Fail if any of the libraries is not found, ie.
- // if any of the `addLib` failed.
- if ((any(libs & AMDGCNLibraries::Ocml) && addLib("ocml.bc")) ||
- (any(libs & AMDGCNLibraries::Ockl) && addLib("ockl.bc")) ||
- (any(libs & AMDGCNLibraries::Hip) && addLib("hip.bc")) ||
- (any(libs & AMDGCNLibraries::OpenCL) && addLib("opencl.bc")) ||
- (any(libs & (AMDGCNLibraries::Ocml | AMDGCNLibraries::Ockl)) &&
- addLib("oclc_isa_version_" + isaVersion + ".bc")))
- return failure();
+ StringRef isaVersion =
+ llvm::AMDGPU::getArchNameAMDGCN(llvm::AMDGPU::parseArchAMDGCN(chip));
+ isaVersion.consume_front("gfx");
+ return getCommonBitcodeLibs(fileList, path, isaVersion);
+ }
return success();
}
std::optional<SmallVector<std::unique_ptr<llvm::Module>>>
SerializeGPUModuleBase::loadBitcodeFiles(llvm::Module &module) {
SmallVector<std::unique_ptr<llvm::Module>> bcFiles;
- // Return if there are no libs to load.
- if (deviceLibs == AMDGCNLibraries::None && fileList.empty())
- return bcFiles;
- if (failed(appendStandardLibs(deviceLibs)))
- return std::nullopt;
if (failed(loadBitcodeFilesFromList(module.getContext(), fileList, bcFiles,
true)))
return std::nullopt;
@@ -210,79 +174,80 @@ LogicalResult SerializeGPUModuleBase::handleBitcodeFile(llvm::Module &module) {
// Stop spamming us with clang version numbers
if (auto *ident = module.getNamedMetadata("llvm.ident"))
module.eraseNamedMetadata(ident);
- // Override the libModules datalayout and target triple with the compiler's
- // data layout should there be a discrepency.
- setDataLayoutAndTriple(module);
return success();
}
void SerializeGPUModuleBase::handleModulePreLink(llvm::Module &module) {
- std::optional<llvm::TargetMachine *> targetMachine =
+ [[maybe_unused]] std::optional<llvm::TargetMachine *> targetMachine =
getOrCreateTargetMachine();
assert(targetMachine && "expect a TargetMachine");
- // If all libraries are not set, traverse the module to determine which
- // libraries are required.
- if (deviceLibs != AMDGCNLibraries::All) {
- for (llvm::Function &f : module.functions()) {
- if (f.hasExternalLinkage() && f.hasName() && !f.hasExactDefinition()) {
- StringRef funcName = f.getName();
- if ("printf" == funcName)
- deviceLibs |= AMDGCNLibraries::OpenCL | AMDGCNLibraries::Ockl |
- AMDGCNLibraries::Ocml;
- if (funcName.starts_with("__ockl_"))
- deviceLibs |= AMDGCNLibraries::Ockl;
- if (funcName.starts_with("__ocml_"))
- deviceLibs |= AMDGCNLibraries::Ocml;
- }
- }
- }
- addControlVariables(module, deviceLibs, target.hasWave64(), target.hasDaz(),
+ addControlVariables(module, target.hasWave64(), target.hasDaz(),
target.hasFiniteOnly(), target.hasUnsafeMath(),
target.hasFastMath(), target.hasCorrectSqrt(),
target.getAbi());
}
-void SerializeGPUModuleBase::addControlVariables(
- llvm::Module &module, AMDGCNLibraries libs, bool wave64, bool daz,
- bool finiteOnly, bool unsafeMath, bool fastMath, bool correctSqrt,
- StringRef abiVer) {
- // Return if no device libraries are required.
- if (libs == AMDGCNLibraries::None)
- return;
- // Helper function for adding control variables.
- auto addControlVariable = [&module](StringRef name, uint32_t value,
- uint32_t bitwidth) {
- if (module.getNamedGlobal(name)) {
- return;
+// Get the paths of ROCm device libraries.
+LogicalResult SerializeGPUModuleBase::getCommonBitcodeLibs(
+ llvm::SmallVector<std::string> &libs, SmallVector<char, 256> &libPath,
+ StringRef isaVersion) {
+ auto addLib = [&](StringRef path) -> bool {
+ if (!llvm::sys::fs::is_regular_file(path)) {
+ getOperation().emitRemark() << "Bitcode library path: " << path
+ << " does not exist or is not a file.\n";
+ return true;
}
- llvm::IntegerType *type =
- llvm::IntegerType::getIntNTy(module.getContext(), bitwidth);
+ libs.push_back(path.str());
+ return false;
+ };
+ auto getLibPath = [&libPath](Twine lib) {
+ auto baseSize = libPath.size();
+ llvm::sys::path::append(libPath, lib + ".bc");
+ std::string path(StringRef(libPath.data(), libPath.size()).str());
+ libPath.truncate(baseSize);
+ return path;
+ };
+
+ // Add ROCm device libraries. Fail if any of the libraries is not found.
+ if (addLib(getLibPath("ocml")) || addLib(getLibPath("ockl")) ||
+ addLib(getLibPath("hip")) || addLib(getLibPath("opencl")) ||
+ addLib(getLibPath("oclc_isa_version_" + isaVersion)))
+ return failure();
+ return success();
+}
+
+void SerializeGPUModuleBase::addControlVariables(
+ llvm::Module &module, bool wave64, bool daz, bool finiteOnly,
+ bool unsafeMath, bool fastMath, bool correctSqrt, StringRef abiVer) {
+ llvm::Type *i8Ty = llvm::Type::getInt8Ty(module.getContext());
+ auto addControlVariable = [i8Ty, &module](StringRef name, bool enable) {
llvm::GlobalVariable *controlVariable = new llvm::GlobalVariable(
- module, /*isConstant=*/type, true,
- llvm::GlobalValue::LinkageTypes::LinkOnceODRLinkage,
- llvm::ConstantInt::get(type, value), name, /*before=*/nullptr,
- /*threadLocalMode=*/llvm::GlobalValue::ThreadLocalMode::NotThreadLocal,
- /*addressSpace=*/4);
+ module, i8Ty, true, llvm::GlobalValue::LinkageTypes::LinkOnceODRLinkage,
+ llvm::ConstantInt::get(i8Ty, enable), name, nullptr,
+ llvm::GlobalValue::ThreadLocalMode::NotThreadLocal, 4);
controlVariable->setVisibility(
llvm::GlobalValue::VisibilityTypes::ProtectedVisibility);
- controlVariable->setAlignment(llvm::MaybeAlign(bitwidth / 8));
+ controlVariable->setAlignment(llvm::MaybeAlign(1));
controlVariable->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Local);
};
- // Add ocml related control variables.
- if (any(libs & AMDGCNLibraries::Ocml)) {
- addControlVariable("__oclc_finite_only_opt", finiteOnly || fastMath, 8);
- addControlVariable("__oclc_daz_opt", daz || fastMath, 8);
- addControlVariable("__oclc_correctly_rounded_sqrt32",
- correctSqrt && !fastMath, 8);
- addControlVariable("__oclc_unsafe_math_opt", unsafeMath || fastMath, 8);
- }
- // Add ocml or ockl related control variables.
- if (any(libs & (AMDGCNLibraries::Ocml | AMDGCNLibraries::Ockl))) {
- addControlVariable("__oclc_wavefrontsize64", wave64, 8);
- int abi = 500;
- abiVer.getAsInteger(0, abi);
- addControlVariable("__oclc_ABI_version", abi, 32);
- }
+ addControlVariable("__oclc_finite_only_opt", finiteOnly || fastMath);
+ addControlVariable("__oclc_unsafe_math_opt", unsafeMath || fastMath);
+ addControlVariable("__oclc_daz_opt", daz || fastMath);
+ addControlVariable("__oclc_correctly_rounded_sqrt32",
+ correctSqrt && !fastMath);
+ addControlVariable("__oclc_wavefrontsize64", wave64);
+
+ llvm::Type *i32Ty = llvm::Type::getInt32Ty(module.getContext());
+ int abi = 500;
+ abiVer.getAsInteger(0, abi);
+ llvm::GlobalVariable *abiVersion = new llvm::GlobalVariable(
+ module, i32Ty, true, llvm::GlobalValue::LinkageTypes::LinkOnceODRLinkage,
+ llvm::ConstantInt::get(i32Ty, abi), "__oclc_ABI_version", nullptr,
+ llvm::GlobalValue::ThreadLocalMode::NotThreadLocal, 4);
+ abiVersion->setVisibility(
+ llvm::GlobalValue::VisibilityTypes::ProtectedVisibility);
+ abiVersion->setAlignment(llvm::MaybeAlign(4));
+ abiVersion->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Local);
}
std::optional<SmallVector<char, 0>>
@@ -347,11 +312,43 @@ SerializeGPUModuleBase::assembleIsa(StringRef isa) {
parser->setTargetParser(*tap);
parser->Run(false);
+
return result;
}
+#if MLIR_ENABLE_ROCM_CONVERSIONS
+namespace {
+class AMDGPUSerializer : public SerializeGPUModuleBase {
+public:
+ AMDGPUSerializer(Operation &module, ROCDLTargetAttr target,
+ const gpu::TargetOptions &targetOptions);
+
+ gpu::GPUModuleOp getOperation();
+
+ // Compile to HSA.
+ std::optional<SmallVector<char, 0>>
+ compileToBinary(const std::string &serializedISA);
+
+ std::optional<SmallVector<char, 0>>
+ moduleToObject(llvm::Module &llvmModule) override;
+
+private:
+ // Target options.
+ gpu::TargetOptions targetOptions;
+};
+} // namespace
+
+AMDGPUSerializer::AMDGPUSerializer(Operation &module, ROCDLTargetAttr target,
+ const gpu::TargetOptions &targetOptions)
+ : SerializeGPUModuleBase(module, target, targetOptions),
+ targetOptions(targetOptions) {}
+
+gpu::GPUModuleOp AMDGPUSerializer::getOperation() {
+ return dyn_cast<gpu::GPUModuleOp>(&SerializeGPUModuleBase::getOperation());
+}
+
std::optional<SmallVector<char, 0>>
-SerializeGPUModuleBase::compileToBinary(const std::string &serializedISA) {
+AMDGPUSerializer::compileToBinary(const std::string &serializedISA) {
// Assemble the ISA.
std::optional<SmallVector<char, 0>> isaBinary = assembleIsa(serializedISA);
@@ -410,13 +407,13 @@ SerializeGPUModuleBase::compileToBinary(const std::string &serializedISA) {
return SmallVector<char, 0>(buffer.begin(), buffer.end());
}
-std::optional<SmallVector<char, 0>> SerializeGPUModuleBase::moduleToObjectImpl(
- const gpu::TargetOptions &targetOptions, llvm::Module &llvmModule) {
+std::optional<SmallVector<char, 0>>
+AMDGPUSerializer::moduleToObject(llvm::Module &llvmModule) {
// Return LLVM IR if the compilation target is offload.
#define DEBUG_TYPE "serialize-to-llvm"
LLVM_DEBUG({
- llvm::dbgs() << "LLVM IR for module: "
- << cast<gpu::GPUModuleOp>(getOperation()).getNameAttr() << "\n"
+ llvm::dbgs() << "LLVM IR for module: " << getOperation().getNameAttr()
+ << "\n"
<< llvmModule << "\n";
});
#undef DEBUG_TYPE
@@ -440,8 +437,7 @@ std::optional<SmallVector<char, 0>> SerializeGPUModuleBase::moduleToObjectImpl(
}
#define DEBUG_TYPE "serialize-to-isa"
LLVM_DEBUG({
- llvm::dbgs() << "ISA for module: "
- << cast<gpu::GPUModuleOp>(getOperation()).getNameAttr() << "\n"
+ llvm::dbgs() << "ISA for module: " << getOperation().getNameAttr() << "\n"
<< *serializedISA << "\n";
});
#undef DEBUG_TYPE
@@ -452,38 +448,6 @@ std::optional<SmallVector<char, 0>> SerializeGPUModuleBase::moduleToObjectImpl(
// Compile to binary.
return compileToBinary(*serializedISA);
}
-
-#if MLIR_ENABLE_ROCM_CONVERSIONS
-namespace {
-class AMDGPUSerializer : public SerializeGPUModuleBase {
-public:
- AMDGPUSerializer(Operation &module, ROCDLTargetAttr target,
- const gpu::TargetOptions &targetOptions);
-
- gpu::GPUModuleOp getOperation();
-
- std::optional<SmallVector<char, 0>>
- moduleToObject(llvm::Module &llvmModule) override;
-
-private:
- // Target options.
- gpu::TargetOptions targetOptions;
-};
-} // namespace
-
-AMDGPUSerializer::AMDGPUSerializer(Operation &module, ROCDLTargetAttr target,
- const gpu::TargetOptions &targetOptions)
- : SerializeGPUModuleBase(module, target, targetOptions),
- targetOptions(targetOptions) {}
-
-gpu::GPUModuleOp AMDGPUSerializer::getOperation() {
- return dyn_cast<gpu::GPUModuleOp>(&SerializeGPUModuleBase::getOperation());
-}
-
-std::optional<SmallVector<char, 0>>
-AMDGPUSerializer::moduleToObject(llvm::Module &llvmModule) {
- return moduleToObjectImpl(targetOptions, llvmModule);
-}
#endif // MLIR_ENABLE_ROCM_CONVERSIONS
std::optional<SmallVector<char, 0>> ROCDLTargetAttrImpl::serializeToObject(
@@ -513,15 +477,10 @@ ROCDLTargetAttrImpl::createObject(Attribute attribute,
const SmallVector<char, 0> &object,
const gpu::TargetOptions &options) const {
gpu::CompilationTarget format = options.getCompilationTarget();
- // If format is `fatbin` transform it to binary as `fatbin` is not yet
- // supported.
- if (format > gpu::CompilationTarget::Binary)
- format = gpu::CompilationTarget::Binary;
-
- DictionaryAttr properties{};
Builder builder(attribute.getContext());...
[truncated]
|
|
@llvm/pr-subscribers-mlir Author: Fabian Mora (fabianmcg) ChangesReverts llvm/llvm-project#95456 Patch is 20.39 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/95790.diff 4 Files Affected:
diff --git a/mlir/include/mlir/Target/LLVM/ROCDL/Utils.h b/mlir/include/mlir/Target/LLVM/ROCDL/Utils.h
index 44c9ded317fa5..374fa65bd02e3 100644
--- a/mlir/include/mlir/Target/LLVM/ROCDL/Utils.h
+++ b/mlir/include/mlir/Target/LLVM/ROCDL/Utils.h
@@ -27,19 +27,6 @@ namespace ROCDL {
/// 5. Returns an empty string.
StringRef getROCMPath();
-/// Helper enum for specifying the AMD GCN device libraries required for
-/// compilation.
-enum class AMDGCNLibraries : uint32_t {
- None = 0,
- Ockl = 1,
- Ocml = 2,
- OpenCL = 4,
- Hip = 8,
- LastLib = Hip,
- LLVM_MARK_AS_BITMASK_ENUM(LastLib),
- All = (LastLib << 1) - 1
-};
-
/// Base class for all ROCDL serializations from GPU modules into binary
/// strings. By default this class serializes into LLVM bitcode.
class SerializeGPUModuleBase : public LLVM::ModuleToObject {
@@ -62,8 +49,8 @@ class SerializeGPUModuleBase : public LLVM::ModuleToObject {
/// Returns the bitcode files to be loaded.
ArrayRef<std::string> getFileList() const;
- /// Appends standard ROCm device libraries to `fileList`.
- LogicalResult appendStandardLibs(AMDGCNLibraries libs);
+ /// Appends standard ROCm device libraries like `ocml.bc`, `ockl.bc`, etc.
+ LogicalResult appendStandardLibs();
/// Loads the bitcode files in `fileList`.
virtual std::optional<SmallVector<std::unique_ptr<llvm::Module>>>
@@ -76,20 +63,15 @@ class SerializeGPUModuleBase : public LLVM::ModuleToObject {
LogicalResult handleBitcodeFile(llvm::Module &module) override;
protected:
- /// Adds `oclc` control variables to the LLVM module.
- void addControlVariables(llvm::Module &module, AMDGCNLibraries libs,
- bool wave64, bool daz, bool finiteOnly,
- bool unsafeMath, bool fastMath, bool correctSqrt,
- StringRef abiVer);
+ /// Appends the paths of common ROCm device libraries to `libs`.
+ LogicalResult getCommonBitcodeLibs(llvm::SmallVector<std::string> &libs,
+ SmallVector<char, 256> &libPath,
+ StringRef isaVersion);
- /// Compiles assembly to a binary.
- virtual std::optional<SmallVector<char, 0>>
- compileToBinary(const std::string &serializedISA);
-
- /// Default implementation of `ModuleToObject::moduleToObject`.
- std::optional<SmallVector<char, 0>>
- moduleToObjectImpl(const gpu::TargetOptions &targetOptions,
- llvm::Module &llvmModule);
+ /// Adds `oclc` control variables to the LLVM module.
+ void addControlVariables(llvm::Module &module, bool wave64, bool daz,
+ bool finiteOnly, bool unsafeMath, bool fastMath,
+ bool correctSqrt, StringRef abiVer);
/// Returns the assembled ISA.
std::optional<SmallVector<char, 0>> assembleIsa(StringRef isa);
@@ -102,9 +84,6 @@ class SerializeGPUModuleBase : public LLVM::ModuleToObject {
/// List of LLVM bitcode files to link to.
SmallVector<std::string> fileList;
-
- /// AMD GCN libraries to use when linking, the default is using none.
- AMDGCNLibraries deviceLibs = AMDGCNLibraries::None;
};
} // namespace ROCDL
} // namespace mlir
diff --git a/mlir/lib/Dialect/GPU/CMakeLists.txt b/mlir/lib/Dialect/GPU/CMakeLists.txt
index 08c8aea36fac9..61ab298ebfb98 100644
--- a/mlir/lib/Dialect/GPU/CMakeLists.txt
+++ b/mlir/lib/Dialect/GPU/CMakeLists.txt
@@ -106,7 +106,7 @@ if(MLIR_ENABLE_ROCM_CONVERSIONS)
"Building mlir with ROCm support requires the AMDGPU backend")
endif()
- set(DEFAULT_ROCM_PATH "" CACHE PATH "Fallback path to search for ROCm installs")
+ set(DEFAULT_ROCM_PATH "/opt/rocm" CACHE PATH "Fallback path to search for ROCm installs")
target_compile_definitions(obj.MLIRGPUTransforms
PRIVATE
__DEFAULT_ROCM_PATH__="${DEFAULT_ROCM_PATH}"
diff --git a/mlir/lib/Target/LLVM/CMakeLists.txt b/mlir/lib/Target/LLVM/CMakeLists.txt
index 4393ff1775ef9..5a3fa160850b4 100644
--- a/mlir/lib/Target/LLVM/CMakeLists.txt
+++ b/mlir/lib/Target/LLVM/CMakeLists.txt
@@ -123,12 +123,17 @@ add_mlir_dialect_library(MLIRROCDLTarget
)
if(MLIR_ENABLE_ROCM_CONVERSIONS)
+ if (NOT ("AMDGPU" IN_LIST LLVM_TARGETS_TO_BUILD))
+ message(SEND_ERROR
+ "Building mlir with ROCm support requires the AMDGPU backend")
+ endif()
+
if (DEFINED ROCM_PATH)
set(DEFAULT_ROCM_PATH "${ROCM_PATH}" CACHE PATH "Fallback path to search for ROCm installs")
elseif(DEFINED ENV{ROCM_PATH})
set(DEFAULT_ROCM_PATH "$ENV{ROCM_PATH}" CACHE PATH "Fallback path to search for ROCm installs")
else()
- set(DEFAULT_ROCM_PATH "" CACHE PATH "Fallback path to search for ROCm installs")
+ set(DEFAULT_ROCM_PATH "/opt/rocm" CACHE PATH "Fallback path to search for ROCm installs")
endif()
message(VERBOSE "MLIR Default ROCM toolkit path: ${DEFAULT_ROCM_PATH}")
diff --git a/mlir/lib/Target/LLVM/ROCDL/Target.cpp b/mlir/lib/Target/LLVM/ROCDL/Target.cpp
index fdf1c93b37210..cc13e5b7436ea 100644
--- a/mlir/lib/Target/LLVM/ROCDL/Target.cpp
+++ b/mlir/lib/Target/LLVM/ROCDL/Target.cpp
@@ -17,6 +17,9 @@
#include "mlir/Dialect/LLVMIR/ROCDLDialect.h"
#include "mlir/Support/FileUtilities.h"
#include "mlir/Target/LLVM/ROCDL/Utils.h"
+#include "mlir/Target/LLVMIR/Dialect/GPU/GPUToLLVMIRTranslation.h"
+#include "mlir/Target/LLVMIR/Dialect/LLVMIR/LLVMToLLVMIRTranslation.h"
+#include "mlir/Target/LLVMIR/Dialect/ROCDL/ROCDLToLLVMIRTranslation.h"
#include "mlir/Target/LLVMIR/Export.h"
#include "llvm/IR/Constants.h"
@@ -109,9 +112,8 @@ SerializeGPUModuleBase::SerializeGPUModuleBase(
if (auto file = dyn_cast<StringAttr>(attr))
fileList.push_back(file.str());
- // By default add all libraries if the toolkit path is not empty.
- if (!getToolkitPath().empty())
- deviceLibs = AMDGCNLibraries::All;
+ // Append standard ROCm device bitcode libraries to the files to be loaded.
+ (void)appendStandardLibs();
}
void SerializeGPUModuleBase::init() {
@@ -136,67 +138,29 @@ ArrayRef<std::string> SerializeGPUModuleBase::getFileList() const {
return fileList;
}
-LogicalResult SerializeGPUModuleBase::appendStandardLibs(AMDGCNLibraries libs) {
- if (libs == AMDGCNLibraries::None)
- return success();
+LogicalResult SerializeGPUModuleBase::appendStandardLibs() {
StringRef pathRef = getToolkitPath();
- // Fail if the toolkit is empty.
- if (pathRef.empty())
- return failure();
-
- // Get the path for the device libraries
- SmallString<256> path;
- path.insert(path.begin(), pathRef.begin(), pathRef.end());
- llvm::sys::path::append(path, "amdgcn", "bitcode");
- pathRef = StringRef(path.data(), path.size());
-
- // Fail if the path is invalid.
- if (!llvm::sys::fs::is_directory(pathRef)) {
- getOperation().emitRemark() << "ROCm amdgcn bitcode path: " << pathRef
- << " does not exist or is not a directory.";
- return failure();
- }
-
- // Get the ISA version.
- StringRef isaVersion =
- llvm::AMDGPU::getArchNameAMDGCN(llvm::AMDGPU::parseArchAMDGCN(chip));
- isaVersion.consume_front("gfx");
-
- // Helper function for adding a library.
- auto addLib = [&](const Twine &lib) -> bool {
- auto baseSize = path.size();
- llvm::sys::path::append(path, lib);
- StringRef pathRef(path.data(), path.size());
- if (!llvm::sys::fs::is_regular_file(pathRef)) {
- getOperation().emitRemark() << "Bitcode library path: " << pathRef
- << " does not exist or is not a file.\n";
- return true;
+ if (!pathRef.empty()) {
+ SmallVector<char, 256> path;
+ path.insert(path.begin(), pathRef.begin(), pathRef.end());
+ llvm::sys::path::append(path, "amdgcn", "bitcode");
+ pathRef = StringRef(path.data(), path.size());
+ if (!llvm::sys::fs::is_directory(pathRef)) {
+ getOperation().emitRemark() << "ROCm amdgcn bitcode path: " << pathRef
+ << " does not exist or is not a directory.";
+ return failure();
}
- fileList.push_back(pathRef.str());
- path.truncate(baseSize);
- return false;
- };
-
- // Add ROCm device libraries. Fail if any of the libraries is not found, ie.
- // if any of the `addLib` failed.
- if ((any(libs & AMDGCNLibraries::Ocml) && addLib("ocml.bc")) ||
- (any(libs & AMDGCNLibraries::Ockl) && addLib("ockl.bc")) ||
- (any(libs & AMDGCNLibraries::Hip) && addLib("hip.bc")) ||
- (any(libs & AMDGCNLibraries::OpenCL) && addLib("opencl.bc")) ||
- (any(libs & (AMDGCNLibraries::Ocml | AMDGCNLibraries::Ockl)) &&
- addLib("oclc_isa_version_" + isaVersion + ".bc")))
- return failure();
+ StringRef isaVersion =
+ llvm::AMDGPU::getArchNameAMDGCN(llvm::AMDGPU::parseArchAMDGCN(chip));
+ isaVersion.consume_front("gfx");
+ return getCommonBitcodeLibs(fileList, path, isaVersion);
+ }
return success();
}
std::optional<SmallVector<std::unique_ptr<llvm::Module>>>
SerializeGPUModuleBase::loadBitcodeFiles(llvm::Module &module) {
SmallVector<std::unique_ptr<llvm::Module>> bcFiles;
- // Return if there are no libs to load.
- if (deviceLibs == AMDGCNLibraries::None && fileList.empty())
- return bcFiles;
- if (failed(appendStandardLibs(deviceLibs)))
- return std::nullopt;
if (failed(loadBitcodeFilesFromList(module.getContext(), fileList, bcFiles,
true)))
return std::nullopt;
@@ -210,79 +174,80 @@ LogicalResult SerializeGPUModuleBase::handleBitcodeFile(llvm::Module &module) {
// Stop spamming us with clang version numbers
if (auto *ident = module.getNamedMetadata("llvm.ident"))
module.eraseNamedMetadata(ident);
- // Override the libModules datalayout and target triple with the compiler's
- // data layout should there be a discrepency.
- setDataLayoutAndTriple(module);
return success();
}
void SerializeGPUModuleBase::handleModulePreLink(llvm::Module &module) {
- std::optional<llvm::TargetMachine *> targetMachine =
+ [[maybe_unused]] std::optional<llvm::TargetMachine *> targetMachine =
getOrCreateTargetMachine();
assert(targetMachine && "expect a TargetMachine");
- // If all libraries are not set, traverse the module to determine which
- // libraries are required.
- if (deviceLibs != AMDGCNLibraries::All) {
- for (llvm::Function &f : module.functions()) {
- if (f.hasExternalLinkage() && f.hasName() && !f.hasExactDefinition()) {
- StringRef funcName = f.getName();
- if ("printf" == funcName)
- deviceLibs |= AMDGCNLibraries::OpenCL | AMDGCNLibraries::Ockl |
- AMDGCNLibraries::Ocml;
- if (funcName.starts_with("__ockl_"))
- deviceLibs |= AMDGCNLibraries::Ockl;
- if (funcName.starts_with("__ocml_"))
- deviceLibs |= AMDGCNLibraries::Ocml;
- }
- }
- }
- addControlVariables(module, deviceLibs, target.hasWave64(), target.hasDaz(),
+ addControlVariables(module, target.hasWave64(), target.hasDaz(),
target.hasFiniteOnly(), target.hasUnsafeMath(),
target.hasFastMath(), target.hasCorrectSqrt(),
target.getAbi());
}
-void SerializeGPUModuleBase::addControlVariables(
- llvm::Module &module, AMDGCNLibraries libs, bool wave64, bool daz,
- bool finiteOnly, bool unsafeMath, bool fastMath, bool correctSqrt,
- StringRef abiVer) {
- // Return if no device libraries are required.
- if (libs == AMDGCNLibraries::None)
- return;
- // Helper function for adding control variables.
- auto addControlVariable = [&module](StringRef name, uint32_t value,
- uint32_t bitwidth) {
- if (module.getNamedGlobal(name)) {
- return;
+// Get the paths of ROCm device libraries.
+LogicalResult SerializeGPUModuleBase::getCommonBitcodeLibs(
+ llvm::SmallVector<std::string> &libs, SmallVector<char, 256> &libPath,
+ StringRef isaVersion) {
+ auto addLib = [&](StringRef path) -> bool {
+ if (!llvm::sys::fs::is_regular_file(path)) {
+ getOperation().emitRemark() << "Bitcode library path: " << path
+ << " does not exist or is not a file.\n";
+ return true;
}
- llvm::IntegerType *type =
- llvm::IntegerType::getIntNTy(module.getContext(), bitwidth);
+ libs.push_back(path.str());
+ return false;
+ };
+ auto getLibPath = [&libPath](Twine lib) {
+ auto baseSize = libPath.size();
+ llvm::sys::path::append(libPath, lib + ".bc");
+ std::string path(StringRef(libPath.data(), libPath.size()).str());
+ libPath.truncate(baseSize);
+ return path;
+ };
+
+ // Add ROCm device libraries. Fail if any of the libraries is not found.
+ if (addLib(getLibPath("ocml")) || addLib(getLibPath("ockl")) ||
+ addLib(getLibPath("hip")) || addLib(getLibPath("opencl")) ||
+ addLib(getLibPath("oclc_isa_version_" + isaVersion)))
+ return failure();
+ return success();
+}
+
+void SerializeGPUModuleBase::addControlVariables(
+ llvm::Module &module, bool wave64, bool daz, bool finiteOnly,
+ bool unsafeMath, bool fastMath, bool correctSqrt, StringRef abiVer) {
+ llvm::Type *i8Ty = llvm::Type::getInt8Ty(module.getContext());
+ auto addControlVariable = [i8Ty, &module](StringRef name, bool enable) {
llvm::GlobalVariable *controlVariable = new llvm::GlobalVariable(
- module, /*isConstant=*/type, true,
- llvm::GlobalValue::LinkageTypes::LinkOnceODRLinkage,
- llvm::ConstantInt::get(type, value), name, /*before=*/nullptr,
- /*threadLocalMode=*/llvm::GlobalValue::ThreadLocalMode::NotThreadLocal,
- /*addressSpace=*/4);
+ module, i8Ty, true, llvm::GlobalValue::LinkageTypes::LinkOnceODRLinkage,
+ llvm::ConstantInt::get(i8Ty, enable), name, nullptr,
+ llvm::GlobalValue::ThreadLocalMode::NotThreadLocal, 4);
controlVariable->setVisibility(
llvm::GlobalValue::VisibilityTypes::ProtectedVisibility);
- controlVariable->setAlignment(llvm::MaybeAlign(bitwidth / 8));
+ controlVariable->setAlignment(llvm::MaybeAlign(1));
controlVariable->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Local);
};
- // Add ocml related control variables.
- if (any(libs & AMDGCNLibraries::Ocml)) {
- addControlVariable("__oclc_finite_only_opt", finiteOnly || fastMath, 8);
- addControlVariable("__oclc_daz_opt", daz || fastMath, 8);
- addControlVariable("__oclc_correctly_rounded_sqrt32",
- correctSqrt && !fastMath, 8);
- addControlVariable("__oclc_unsafe_math_opt", unsafeMath || fastMath, 8);
- }
- // Add ocml or ockl related control variables.
- if (any(libs & (AMDGCNLibraries::Ocml | AMDGCNLibraries::Ockl))) {
- addControlVariable("__oclc_wavefrontsize64", wave64, 8);
- int abi = 500;
- abiVer.getAsInteger(0, abi);
- addControlVariable("__oclc_ABI_version", abi, 32);
- }
+ addControlVariable("__oclc_finite_only_opt", finiteOnly || fastMath);
+ addControlVariable("__oclc_unsafe_math_opt", unsafeMath || fastMath);
+ addControlVariable("__oclc_daz_opt", daz || fastMath);
+ addControlVariable("__oclc_correctly_rounded_sqrt32",
+ correctSqrt && !fastMath);
+ addControlVariable("__oclc_wavefrontsize64", wave64);
+
+ llvm::Type *i32Ty = llvm::Type::getInt32Ty(module.getContext());
+ int abi = 500;
+ abiVer.getAsInteger(0, abi);
+ llvm::GlobalVariable *abiVersion = new llvm::GlobalVariable(
+ module, i32Ty, true, llvm::GlobalValue::LinkageTypes::LinkOnceODRLinkage,
+ llvm::ConstantInt::get(i32Ty, abi), "__oclc_ABI_version", nullptr,
+ llvm::GlobalValue::ThreadLocalMode::NotThreadLocal, 4);
+ abiVersion->setVisibility(
+ llvm::GlobalValue::VisibilityTypes::ProtectedVisibility);
+ abiVersion->setAlignment(llvm::MaybeAlign(4));
+ abiVersion->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Local);
}
std::optional<SmallVector<char, 0>>
@@ -347,11 +312,43 @@ SerializeGPUModuleBase::assembleIsa(StringRef isa) {
parser->setTargetParser(*tap);
parser->Run(false);
+
return result;
}
+#if MLIR_ENABLE_ROCM_CONVERSIONS
+namespace {
+class AMDGPUSerializer : public SerializeGPUModuleBase {
+public:
+ AMDGPUSerializer(Operation &module, ROCDLTargetAttr target,
+ const gpu::TargetOptions &targetOptions);
+
+ gpu::GPUModuleOp getOperation();
+
+ // Compile to HSA.
+ std::optional<SmallVector<char, 0>>
+ compileToBinary(const std::string &serializedISA);
+
+ std::optional<SmallVector<char, 0>>
+ moduleToObject(llvm::Module &llvmModule) override;
+
+private:
+ // Target options.
+ gpu::TargetOptions targetOptions;
+};
+} // namespace
+
+AMDGPUSerializer::AMDGPUSerializer(Operation &module, ROCDLTargetAttr target,
+ const gpu::TargetOptions &targetOptions)
+ : SerializeGPUModuleBase(module, target, targetOptions),
+ targetOptions(targetOptions) {}
+
+gpu::GPUModuleOp AMDGPUSerializer::getOperation() {
+ return dyn_cast<gpu::GPUModuleOp>(&SerializeGPUModuleBase::getOperation());
+}
+
std::optional<SmallVector<char, 0>>
-SerializeGPUModuleBase::compileToBinary(const std::string &serializedISA) {
+AMDGPUSerializer::compileToBinary(const std::string &serializedISA) {
// Assemble the ISA.
std::optional<SmallVector<char, 0>> isaBinary = assembleIsa(serializedISA);
@@ -410,13 +407,13 @@ SerializeGPUModuleBase::compileToBinary(const std::string &serializedISA) {
return SmallVector<char, 0>(buffer.begin(), buffer.end());
}
-std::optional<SmallVector<char, 0>> SerializeGPUModuleBase::moduleToObjectImpl(
- const gpu::TargetOptions &targetOptions, llvm::Module &llvmModule) {
+std::optional<SmallVector<char, 0>>
+AMDGPUSerializer::moduleToObject(llvm::Module &llvmModule) {
// Return LLVM IR if the compilation target is offload.
#define DEBUG_TYPE "serialize-to-llvm"
LLVM_DEBUG({
- llvm::dbgs() << "LLVM IR for module: "
- << cast<gpu::GPUModuleOp>(getOperation()).getNameAttr() << "\n"
+ llvm::dbgs() << "LLVM IR for module: " << getOperation().getNameAttr()
+ << "\n"
<< llvmModule << "\n";
});
#undef DEBUG_TYPE
@@ -440,8 +437,7 @@ std::optional<SmallVector<char, 0>> SerializeGPUModuleBase::moduleToObjectImpl(
}
#define DEBUG_TYPE "serialize-to-isa"
LLVM_DEBUG({
- llvm::dbgs() << "ISA for module: "
- << cast<gpu::GPUModuleOp>(getOperation()).getNameAttr() << "\n"
+ llvm::dbgs() << "ISA for module: " << getOperation().getNameAttr() << "\n"
<< *serializedISA << "\n";
});
#undef DEBUG_TYPE
@@ -452,38 +448,6 @@ std::optional<SmallVector<char, 0>> SerializeGPUModuleBase::moduleToObjectImpl(
// Compile to binary.
return compileToBinary(*serializedISA);
}
-
-#if MLIR_ENABLE_ROCM_CONVERSIONS
-namespace {
-class AMDGPUSerializer : public SerializeGPUModuleBase {
-public:
- AMDGPUSerializer(Operation &module, ROCDLTargetAttr target,
- const gpu::TargetOptions &targetOptions);
-
- gpu::GPUModuleOp getOperation();
-
- std::optional<SmallVector<char, 0>>
- moduleToObject(llvm::Module &llvmModule) override;
-
-private:
- // Target options.
- gpu::TargetOptions targetOptions;
-};
-} // namespace
-
-AMDGPUSerializer::AMDGPUSerializer(Operation &module, ROCDLTargetAttr target,
- const gpu::TargetOptions &targetOptions)
- : SerializeGPUModuleBase(module, target, targetOptions),
- targetOptions(targetOptions) {}
-
-gpu::GPUModuleOp AMDGPUSerializer::getOperation() {
- return dyn_cast<gpu::GPUModuleOp>(&SerializeGPUModuleBase::getOperation());
-}
-
-std::optional<SmallVector<char, 0>>
-AMDGPUSerializer::moduleToObject(llvm::Module &llvmModule) {
- return moduleToObjectImpl(targetOptions, llvmModule);
-}
#endif // MLIR_ENABLE_ROCM_CONVERSIONS
std::optional<SmallVector<char, 0>> ROCDLTargetAttrImpl::serializeToObject(
@@ -513,15 +477,10 @@ ROCDLTargetAttrImpl::createObject(Attribute attribute,
const SmallVector<char, 0> &object,
const gpu::TargetOptions &options) const {
gpu::CompilationTarget format = options.getCompilationTarget();
- // If format is `fatbin` transform it to binary as `fatbin` is not yet
- // supported.
- if (format > gpu::CompilationTarget::Binary)
- format = gpu::CompilationTarget::Binary;
-
- DictionaryAttr properties{};
Builder builder(attribute.getContext());...
[truncated]
|
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Reverts #95456