[llvm-branch-commits] [mlir] [mlir][acc] Translate #acc.offload_packager to LLVM IR (PR #242808)
Ivan R. Ivanov via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Sun Nov 15 03:55:27 PST 2026
https://github.com/ivanradanov created https://github.com/llvm/llvm-project/pull/242808
Add a translation library for the OpenACC dialect that implements the offloading translation interface of #acc.offload_packager. Each object of a gpu.binary is wrapped in an offloading binary of kind OpenACC, carrying the target triple and architecture, and the images are placed in the .llvm.offloading section of the host module for the offloading linker wrapper to extract.
Assisted-by: Claude Sonnet 5.5
>From 7d0debecf226c185fd2974215dbf3b1861cd4ccf Mon Sep 17 00:00:00 2001
From: Ivan Radanov Ivanov <iivanov at nvidia.com>
Date: Sun, 8 Nov 2026 17:32:30 -0800
Subject: [PATCH] [mlir][acc] Translate #acc.offload_packager to LLVM IR
Add a translation library for the OpenACC dialect that implements the
offloading translation interface of #acc.offload_packager. Each object
of a gpu.binary is wrapped in an offloading binary of kind OpenACC,
carrying the target triple and architecture, and the images are placed
in the .llvm.offloading section of the host module for the offloading
linker wrapper to extract.
Assisted-by: Claude Sonnet 5.5
---
mlir/include/mlir/Target/LLVMIR/Dialect/All.h | 2 +
.../OpenACC/OpenACCToLLVMIRTranslation.h | 29 +++
mlir/lib/Target/LLVMIR/CMakeLists.txt | 1 +
mlir/lib/Target/LLVMIR/Dialect/CMakeLists.txt | 1 +
.../LLVMIR/Dialect/OpenACC/CMakeLists.txt | 20 ++
.../OpenACC/OpenACCToLLVMIRTranslation.cpp | 173 ++++++++++++++++++
mlir/test/CMakeLists.txt | 1 +
.../LLVMIR/openacc-offload-objdump.mlir | 23 +++
mlir/test/Target/LLVMIR/openacc-offload.mlir | 43 +++++
mlir/test/lit.cfg.py | 1 +
10 files changed, 294 insertions(+)
create mode 100644 mlir/include/mlir/Target/LLVMIR/Dialect/OpenACC/OpenACCToLLVMIRTranslation.h
create mode 100644 mlir/lib/Target/LLVMIR/Dialect/OpenACC/CMakeLists.txt
create mode 100644 mlir/lib/Target/LLVMIR/Dialect/OpenACC/OpenACCToLLVMIRTranslation.cpp
create mode 100644 mlir/test/Target/LLVMIR/openacc-offload-objdump.mlir
create mode 100644 mlir/test/Target/LLVMIR/openacc-offload.mlir
diff --git a/mlir/include/mlir/Target/LLVMIR/Dialect/All.h b/mlir/include/mlir/Target/LLVMIR/Dialect/All.h
index 95e90789e9580b..05b66ace902ad3 100644
--- a/mlir/include/mlir/Target/LLVMIR/Dialect/All.h
+++ b/mlir/include/mlir/Target/LLVMIR/Dialect/All.h
@@ -23,6 +23,7 @@
#include "mlir/Target/LLVMIR/Dialect/LLVMIR/LLVMToLLVMIRTranslation.h"
#include "mlir/Target/LLVMIR/Dialect/NVVM/LLVMIRToNVVMTranslation.h"
#include "mlir/Target/LLVMIR/Dialect/NVVM/NVVMToLLVMIRTranslation.h"
+#include "mlir/Target/LLVMIR/Dialect/OpenACC/OpenACCToLLVMIRTranslation.h"
#include "mlir/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.h"
#include "mlir/Target/LLVMIR/Dialect/Ptr/PtrToLLVMIRTranslation.h"
#include "mlir/Target/LLVMIR/Dialect/ROCDL/ROCDLToLLVMIRTranslation.h"
@@ -43,6 +44,7 @@ static inline void registerAllToLLVMIRTranslations(DialectRegistry ®istry) {
registerGPUDialectTranslation(registry);
registerLLVMDialectTranslation(registry);
registerNVVMDialectTranslation(registry);
+ registerOpenACCDialectTranslation(registry);
registerOpenMPDialectTranslation(registry);
registerPtrDialectTranslation(registry);
registerROCDLDialectTranslation(registry);
diff --git a/mlir/include/mlir/Target/LLVMIR/Dialect/OpenACC/OpenACCToLLVMIRTranslation.h b/mlir/include/mlir/Target/LLVMIR/Dialect/OpenACC/OpenACCToLLVMIRTranslation.h
new file mode 100644
index 00000000000000..423a72651f4dca
--- /dev/null
+++ b/mlir/include/mlir/Target/LLVMIR/Dialect/OpenACC/OpenACCToLLVMIRTranslation.h
@@ -0,0 +1,29 @@
+//===- OpenACCToLLVMIRTranslation.h - OpenACC Dialect to LLVM IR -*- C++ -*-==//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This provides registration calls for OpenACC dialect to LLVM IR translation.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef MLIR_TARGET_LLVMIR_DIALECT_OPENACC_OPENACCTOLLVMIRTRANSLATION_H
+#define MLIR_TARGET_LLVMIR_DIALECT_OPENACC_OPENACCTOLLVMIRTRANSLATION_H
+
+namespace mlir {
+class DialectRegistry;
+class MLIRContext;
+
+/// Register the OpenACC dialect and the translation from it to the LLVM IR in
+/// the given registry.
+void registerOpenACCDialectTranslation(DialectRegistry ®istry);
+
+/// Register the OpenACC dialect and the translation from it in the registry
+/// associated with the given context.
+void registerOpenACCDialectTranslation(MLIRContext &context);
+} // namespace mlir
+
+#endif // MLIR_TARGET_LLVMIR_DIALECT_OPENACC_OPENACCTOLLVMIRTRANSLATION_H
diff --git a/mlir/lib/Target/LLVMIR/CMakeLists.txt b/mlir/lib/Target/LLVMIR/CMakeLists.txt
index 8f64d302c39fbb..585a70ebf9c7ef 100644
--- a/mlir/lib/Target/LLVMIR/CMakeLists.txt
+++ b/mlir/lib/Target/LLVMIR/CMakeLists.txt
@@ -56,6 +56,7 @@ add_mlir_translation_library(MLIRToLLVMIRTranslationRegistration
MLIRGPUToLLVMIRTranslation
MLIRLLVMToLLVMIRTranslation
MLIRNVVMToLLVMIRTranslation
+ MLIROpenACCToLLVMIRTranslation
MLIROpenMPToLLVMIRTranslation
MLIRROCDLToLLVMIRTranslation
MLIRPtrToLLVMIRTranslation
diff --git a/mlir/lib/Target/LLVMIR/Dialect/CMakeLists.txt b/mlir/lib/Target/LLVMIR/Dialect/CMakeLists.txt
index d590b236cfc915..a102c4323075b7 100644
--- a/mlir/lib/Target/LLVMIR/Dialect/CMakeLists.txt
+++ b/mlir/lib/Target/LLVMIR/Dialect/CMakeLists.txt
@@ -5,6 +5,7 @@ add_subdirectory(Builtin)
add_subdirectory(GPU)
add_subdirectory(LLVMIR)
add_subdirectory(NVVM)
+add_subdirectory(OpenACC)
add_subdirectory(OpenMP)
add_subdirectory(ROCDL)
add_subdirectory(Ptr)
diff --git a/mlir/lib/Target/LLVMIR/Dialect/OpenACC/CMakeLists.txt b/mlir/lib/Target/LLVMIR/Dialect/OpenACC/CMakeLists.txt
new file mode 100644
index 00000000000000..b7878da6e1f650
--- /dev/null
+++ b/mlir/lib/Target/LLVMIR/Dialect/OpenACC/CMakeLists.txt
@@ -0,0 +1,20 @@
+add_mlir_translation_library(MLIROpenACCToLLVMIRTranslation
+ OpenACCToLLVMIRTranslation.cpp
+
+ LINK_COMPONENTS
+ BinaryFormat
+ Core
+ Object
+ TransformUtils
+
+ LINK_LIBS PUBLIC
+ MLIRGPUDialect
+ MLIRIR
+ MLIRLLVMDialect
+ MLIRNVVMDialect
+ MLIROpenACCDialect
+ MLIRROCDLDialect
+ MLIRSupport
+ MLIRTargetLLVMIRExport
+ MLIRXeVMDialect
+ )
diff --git a/mlir/lib/Target/LLVMIR/Dialect/OpenACC/OpenACCToLLVMIRTranslation.cpp b/mlir/lib/Target/LLVMIR/Dialect/OpenACC/OpenACCToLLVMIRTranslation.cpp
new file mode 100644
index 00000000000000..b0af31a58a6d86
--- /dev/null
+++ b/mlir/lib/Target/LLVMIR/Dialect/OpenACC/OpenACCToLLVMIRTranslation.cpp
@@ -0,0 +1,173 @@
+//===- OpenACCToLLVMIRTranslation.cpp - Translate OpenACC to LLVM IR ------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements the OpenACC to LLVM IR translation.
+//
+//===----------------------------------------------------------------------===//
+
+#include "mlir/Target/LLVMIR/Dialect/OpenACC/OpenACCToLLVMIRTranslation.h"
+
+#include "mlir/Dialect/GPU/IR/CompilationInterfaces.h"
+#include "mlir/Dialect/GPU/IR/GPUDialect.h"
+#include "mlir/Dialect/LLVMIR/LLVMDialect.h"
+#include "mlir/Dialect/LLVMIR/NVVMDialect.h"
+#include "mlir/Dialect/LLVMIR/ROCDLDialect.h"
+#include "mlir/Dialect/LLVMIR/XeVMDialect.h"
+#include "mlir/Dialect/OpenACC/OpenACC.h"
+#include "mlir/IR/Operation.h"
+#include "mlir/Support/LLVM.h"
+#include "mlir/Target/LLVMIR/ModuleTranslation.h"
+
+#include "llvm/ADT/TypeSwitch.h"
+#include "llvm/IR/Constants.h"
+#include "llvm/IR/GlobalVariable.h"
+#include "llvm/IR/IRBuilder.h"
+#include "llvm/IR/Module.h"
+#include "llvm/Object/OffloadBinary.h"
+#include "llvm/Transforms/Utils/ModuleUtils.h"
+
+using namespace mlir;
+
+namespace {
+/// The extracted generic information for a target specific offload attribute.
+struct OffloadTargetInfo {
+ StringRef triple;
+ StringRef arch;
+};
+
+/// Returns the offloading information of a GPU target attribute, or
+/// `std::nullopt` if the target is not supported.
+static std::optional<OffloadTargetInfo> getOffloadTargetInfo(Attribute target) {
+ return llvm::TypeSwitch<Attribute, std::optional<OffloadTargetInfo>>(target)
+ .Case([](NVVM::NVVMTargetAttr nvvm) {
+ return OffloadTargetInfo{nvvm.getTriple(), nvvm.getChip()};
+ })
+ .Case([](ROCDL::ROCDLTargetAttr rocdl) {
+ return OffloadTargetInfo{rocdl.getTriple(), rocdl.getChip()};
+ })
+ .Case([](xevm::XeVMTargetAttr xevm) {
+ return OffloadTargetInfo{xevm.getTriple(), xevm.getChip()};
+ })
+ .Default([](Attribute) { return std::nullopt; });
+}
+
+/// Implementation of the `OffloadingLLVMTranslationAttrInterface` model for
+/// `#acc.offload_packager`.
+class OffloadPackagerAttrImpl
+ : public gpu::OffloadingLLVMTranslationAttrInterface::FallbackModel<
+ OffloadPackagerAttrImpl> {
+public:
+ /// Embeds the objects of a `gpu.binary` in the host module as offloading
+ /// binaries placed in the `.llvm.offloading` section.
+ LogicalResult embedBinary(Attribute attribute, Operation *operation,
+ llvm::IRBuilderBase &builder,
+ LLVM::ModuleTranslation &moduleTranslation) const;
+
+ LogicalResult launchKernel(Attribute attribute,
+ Operation *launchFuncOperation,
+ Operation *binaryOperation,
+ llvm::IRBuilderBase &builder,
+ LLVM::ModuleTranslation &moduleTranslation) const;
+};
+} // namespace
+
+LogicalResult OffloadPackagerAttrImpl::embedBinary(
+ Attribute attribute, Operation *operation, llvm::IRBuilderBase &builder,
+ LLVM::ModuleTranslation &moduleTranslation) const {
+ assert(operation && "The binary operation must be non null.");
+ auto op = mlir::dyn_cast<gpu::BinaryOp>(operation);
+ if (!op) {
+ operation->emitError("operation must be a GPU binary");
+ return failure();
+ }
+
+ using namespace llvm::object;
+
+ SmallVector<char> binaryData;
+ llvm::raw_svector_ostream os(binaryData);
+ for (Attribute attr : op.getObjectsAttr().getValue()) {
+ auto object = dyn_cast<gpu::ObjectAttr>(attr);
+ if (!object)
+ return op.emitError("expected a gpu.object");
+
+ // The section is chosen by the offloading binary, so a requested one would
+ // be ignored.
+ if (DictionaryAttr properties = object.getProperties())
+ if (properties.get(gpu::elfSectionName))
+ return op.emitError("the offload packager does not support custom "
+ "sections");
+
+ std::optional<OffloadTargetInfo> target =
+ getOffloadTargetInfo(object.getTarget());
+ if (!target)
+ return op.emitError("the offload packager does not support the target ")
+ << object.getTarget();
+
+ OffloadBinary::OffloadingImage image;
+ bool requiresNullTerminator = false;
+ switch (object.getFormat()) {
+ case gpu::CompilationTarget::Offload:
+ image.TheImageKind = IMG_Bitcode;
+ requiresNullTerminator = true;
+ break;
+ case gpu::CompilationTarget::Binary:
+ image.TheImageKind = IMG_Object;
+ requiresNullTerminator = false;
+ break;
+ case gpu::CompilationTarget::Assembly:
+ case gpu::CompilationTarget::Fatbin:
+ return op.emitError("the offloading linker wrapper cannot link the "
+ "assembly or fatbin formats");
+ }
+
+ StringRef serialized = object.getObject().getValue();
+ image.Image = llvm::MemoryBuffer::getMemBuffer(
+ serialized, /*BufferName=*/"", requiresNullTerminator);
+ image.TheOffloadKind = OFK_OpenACC;
+ image.StringData["arch"] = target->arch;
+ image.StringData["triple"] = target->triple;
+
+ llvm::SmallString<0> buffer = OffloadBinary::write(image);
+ if (buffer.size() % OffloadBinary::getAlignment() != 0)
+ return op.emitError("offload binary has invalid size alignment");
+ os << buffer;
+ }
+
+ llvm::Module &m = *moduleTranslation.getLLVMModule();
+ llvm::Constant *serializedCst = llvm::ConstantDataArray::getString(
+ m.getContext(), os.str(), /*AddNull=*/false);
+ auto *serializedObj =
+ new llvm::GlobalVariable(m, serializedCst->getType(), /*isConstant=*/true,
+ llvm::GlobalValue::InternalLinkage,
+ serializedCst, op.getName() + "_binary");
+ serializedObj->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
+ serializedObj->setSection(".llvm.offloading");
+ serializedObj->setAlignment(llvm::Align(OffloadBinary::getAlignment()));
+ llvm::appendToUsed(m, {serializedObj});
+ return success();
+}
+
+LogicalResult
+OffloadPackagerAttrImpl::launchKernel(Attribute, Operation *launchFuncOperation,
+ Operation *, llvm::IRBuilderBase &,
+ LLVM::ModuleTranslation &) const {
+ return launchFuncOperation->emitError("unexpected kernel launch");
+}
+
+void mlir::registerOpenACCDialectTranslation(DialectRegistry ®istry) {
+ registry.insert<acc::OpenACCDialect>();
+ registry.addExtension(+[](MLIRContext *ctx, acc::OpenACCDialect *dialect) {
+ acc::OffloadPackagerAttr::attachInterface<OffloadPackagerAttrImpl>(*ctx);
+ });
+}
+
+void mlir::registerOpenACCDialectTranslation(MLIRContext &context) {
+ DialectRegistry registry;
+ registerOpenACCDialectTranslation(registry);
+ context.appendDialectRegistry(registry);
+}
diff --git a/mlir/test/CMakeLists.txt b/mlir/test/CMakeLists.txt
index 0329b46b52a476..05704008c46126 100644
--- a/mlir/test/CMakeLists.txt
+++ b/mlir/test/CMakeLists.txt
@@ -149,6 +149,7 @@ if(NOT MLIR_STANDALONE_BUILD)
list(APPEND MLIR_TEST_DEPENDS
FileCheck
count
+ llvm-objdump
llvm-strings
not
split-file
diff --git a/mlir/test/Target/LLVMIR/openacc-offload-objdump.mlir b/mlir/test/Target/LLVMIR/openacc-offload-objdump.mlir
new file mode 100644
index 00000000000000..32833646e50117
--- /dev/null
+++ b/mlir/test/Target/LLVMIR/openacc-offload-objdump.mlir
@@ -0,0 +1,23 @@
+// REQUIRES: has_llc_binary
+// RUN: mlir-translate -mlir-to-llvmir %s | llc -filetype=obj -o %t.o
+// RUN: llvm-objdump --offloading %t.o | FileCheck %s
+
+// Check that the images embedded by the offload packager are read back as
+// offloading binaries from the object file of the host module.
+
+// CHECK: OFFLOADING IMAGE [0]:
+// CHECK-NEXT: kind llvm ir
+// CHECK-NEXT: arch sm_80
+// CHECK-NEXT: triple nvptx64-nvidia-cuda
+// CHECK-NEXT: producer openacc
+// CHECK: OFFLOADING IMAGE [1]:
+// CHECK-NEXT: kind elf
+// CHECK-NEXT: arch gfx90a
+// CHECK-NEXT: triple amdgcn-amd-amdhsa
+// CHECK-NEXT: producer openacc
+
+module attributes {gpu.container_module} {
+ gpu.binary @device_module <#acc.offload_packager> [
+ #gpu.object<#nvvm.target<chip = "sm_80">, offload = "BC">,
+ #gpu.object<#rocdl.target<chip = "gfx90a">, bin = "\177ELF">]
+}
diff --git a/mlir/test/Target/LLVMIR/openacc-offload.mlir b/mlir/test/Target/LLVMIR/openacc-offload.mlir
new file mode 100644
index 00000000000000..1db3cade46b1a4
--- /dev/null
+++ b/mlir/test/Target/LLVMIR/openacc-offload.mlir
@@ -0,0 +1,43 @@
+// RUN: mlir-translate -mlir-to-llvmir -split-input-file -verify-diagnostics %s | FileCheck %s
+
+// CHECK: @device_module_binary = internal unnamed_addr constant [{{[0-9]+}} x i8] c"\10\FF\10\AD
+// CHECK-SAME: section ".llvm.offloading"
+// CHECK: @llvm.used = appending global {{.*}} @device_module_binary
+
+module attributes {gpu.container_module} {
+ gpu.binary @device_module <#acc.offload_packager> [
+ #gpu.object<#nvvm.target<chip = "sm_80">, offload = "BC">]
+}
+
+// -----
+
+// CHECK: @nvidia_object_module_binary = internal unnamed_addr constant
+// CHECK: @amd_module_binary = internal unnamed_addr constant {{.*}}amdgcn-amd-amdhsa{{.*}}gfx90a
+// CHECK: @intel_module_binary = internal unnamed_addr constant {{.*}}spirv64-unknown-unknown{{.*}}bmg
+
+module attributes {gpu.container_module} {
+ gpu.binary @nvidia_object_module <#acc.offload_packager> [
+ #gpu.object<#nvvm.target<chip = "sm_80">, bin = "\177ELF">]
+ gpu.binary @amd_module <#acc.offload_packager> [
+ #gpu.object<#rocdl.target<chip = "gfx90a">, offload = "BC">]
+ gpu.binary @intel_module <#acc.offload_packager> [
+ #gpu.object<#xevm.target, offload = "BC">]
+}
+
+// -----
+
+module attributes {gpu.container_module} {
+ // expected-error @below {{cannot link the assembly or fatbin formats}}
+ // expected-error @below {{LLVM Translation failed for operation: gpu.binary}}
+ gpu.binary @ptx <#acc.offload_packager> [
+ #gpu.object<#nvvm.target<chip = "sm_80">, assembly = ".version 8.0">]
+}
+
+// -----
+
+module attributes {gpu.container_module} {
+ // expected-error @below {{cannot link the assembly or fatbin formats}}
+ // expected-error @below {{LLVM Translation failed for operation: gpu.binary}}
+ gpu.binary @fatbin <#acc.offload_packager> [
+ #gpu.object<#nvvm.target<chip = "sm_80">, fatbin = "">]
+}
diff --git a/mlir/test/lit.cfg.py b/mlir/test/lit.cfg.py
index ca070a37ebda50..998286b6380d61 100644
--- a/mlir/test/lit.cfg.py
+++ b/mlir/test/lit.cfg.py
@@ -178,6 +178,7 @@ def find_real_python_interpreter():
llvm_config.with_environment("PATH", dirs, append_path=True)
tools = [
+ "llvm-objdump",
"llvm-strings",
"mlir-tblgen",
"mlir-translate",
More information about the llvm-branch-commits
mailing list