1//===- Utils.cpp - MLIR ROCDL target utils ----------------------*- C++ -*-===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This files defines ROCDL target related utility classes and functions.
10//
11//===----------------------------------------------------------------------===//
12
13#include "mlir/Target/LLVM/ROCDL/Utils.h"
14#include "mlir/Dialect/GPU/IR/GPUDialect.h"
15#include "mlir/Dialect/LLVMIR/LLVMModuleOps.h"
16
17#include "llvm/ADT/StringMap.h"
18#include "llvm/Frontend/Offloading/Utility.h"
19
20using namespace mlir;
21using namespace mlir::ROCDL;
22
23std::optional<DenseMap<StringAttr, NamedAttrList>>
24mlir::ROCDL::getAMDHSAKernelsELFMetadata(Builder &builder,
25 ArrayRef<char> elfData) {
26 uint16_t elfABIVersion;
27 llvm::StringMap<llvm::offloading::amdgpu::AMDGPUKernelMetaData> kernels;
28 llvm::MemoryBufferRef buffer(StringRef(elfData.data(), elfData.size()),
29 "buffer");
30 // Get the metadata.
31 llvm::Error error = llvm::offloading::amdgpu::getAMDGPUMetaDataFromImage(
32 MemBuffer: buffer, KernelInfoMap&: kernels, ELFABIVersion&: elfABIVersion);
33 // Return `nullopt` if the metadata couldn't be retrieved.
34 if (error) {
35 llvm::consumeError(Err: std::move(error));
36 return std::nullopt;
37 }
38 // Helper lambda for converting values.
39 auto getI32Array = [&builder](const uint32_t *array) {
40 return builder.getDenseI32ArrayAttr(values: {static_cast<int32_t>(array[0]),
41 static_cast<int32_t>(array[1]),
42 static_cast<int32_t>(array[2])});
43 };
44 DenseMap<StringAttr, NamedAttrList> kernelMD;
45 for (const auto &[name, kernel] : kernels) {
46 NamedAttrList attrs;
47 // Add kernel metadata.
48 attrs.append(name: "agpr_count", attr: builder.getI64IntegerAttr(value: kernel.AGPRCount));
49 attrs.append(name: "sgpr_count", attr: builder.getI64IntegerAttr(value: kernel.SGPRCount));
50 attrs.append(name: "vgpr_count", attr: builder.getI64IntegerAttr(value: kernel.VGPRCount));
51 attrs.append(name: "sgpr_spill_count",
52 attr: builder.getI64IntegerAttr(value: kernel.SGPRSpillCount));
53 attrs.append(name: "vgpr_spill_count",
54 attr: builder.getI64IntegerAttr(value: kernel.VGPRSpillCount));
55 attrs.append(name: "wavefront_size",
56 attr: builder.getI64IntegerAttr(value: kernel.WavefrontSize));
57 attrs.append(name: "max_flat_workgroup_size",
58 attr: builder.getI64IntegerAttr(value: kernel.MaxFlatWorkgroupSize));
59 attrs.append(name: "group_segment_fixed_size",
60 attr: builder.getI64IntegerAttr(value: kernel.GroupSegmentList));
61 attrs.append(name: "private_segment_fixed_size",
62 attr: builder.getI64IntegerAttr(value: kernel.PrivateSegmentSize));
63 attrs.append(name: "reqd_workgroup_size",
64 attr: getI32Array(kernel.RequestedWorkgroupSize));
65 attrs.append(name: "workgroup_size_hint", attr: getI32Array(kernel.WorkgroupSizeHint));
66 kernelMD[builder.getStringAttr(bytes: name)] = std::move(attrs);
67 }
68 return std::move(kernelMD);
69}
70
71gpu::KernelTableAttr mlir::ROCDL::getKernelMetadata(Operation *gpuModule,
72 ArrayRef<char> elfData) {
73 auto module = cast<gpu::GPUModuleOp>(Val: gpuModule);
74 Builder builder(module.getContext());
75 SmallVector<gpu::KernelMetadataAttr> kernels;
76 std::optional<DenseMap<StringAttr, NamedAttrList>> mdMapOrNull =
77 getAMDHSAKernelsELFMetadata(builder, elfData);
78 for (auto funcOp : module.getBody()->getOps<LLVM::LLVMFuncOp>()) {
79 if (!funcOp->getDiscardableAttr(name: "rocdl.kernel"))
80 continue;
81 kernels.push_back(Elt: gpu::KernelMetadataAttr::get(
82 kernel: funcOp, metadata: mdMapOrNull ? builder.getDictionaryAttr(
83 value: mdMapOrNull->lookup(Val: funcOp.getNameAttr()))
84 : nullptr));
85 }
86 return gpu::KernelTableAttr::get(context: gpuModule->getContext(), kernels);
87}
88