1 /* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
2
3 Licensed under the Apache License, Version 2.0 (the "License");
4 you may not use this file except in compliance with the License.
5 You may obtain a copy of the License at
6
7 http://www.apache.org/licenses/LICENSE-2.0
8
9 Unless required by applicable law or agreed to in writing, software
10 distributed under the License is distributed on an "AS IS" BASIS,
11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 See the License for the specific language governing permissions and
13 limitations under the License.
14 ==============================================================================*/
15
16 #include "mlir/Conversion/ComplexToLLVM/ComplexToLLVM.h" // from @llvm-project
17 #include "mlir/Conversion/GPUToNVVM/GPUToNVVMPass.h" // from @llvm-project
18 #include "mlir/Conversion/GPUToROCDL/GPUToROCDLPass.h" // from @llvm-project
19 #include "mlir/Conversion/LLVMCommon/LoweringOptions.h" // from @llvm-project
20 #include "mlir/Conversion/LLVMCommon/TypeConverter.h" // from @llvm-project
21 #include "mlir/Conversion/MemRefToLLVM/MemRefToLLVM.h" // from @llvm-project
22 #include "mlir/Conversion/StandardToLLVM/ConvertStandardToLLVM.h" // from @llvm-project
23 #include "mlir/Dialect/GPU/GPUDialect.h" // from @llvm-project
24 #include "mlir/Dialect/LLVMIR/LLVMDialect.h" // from @llvm-project
25 #include "mlir/Dialect/LLVMIR/LLVMTypes.h" // from @llvm-project
26 #include "mlir/Dialect/LLVMIR/NVVMDialect.h" // from @llvm-project
27 #include "mlir/Dialect/LLVMIR/ROCDLDialect.h" // from @llvm-project
28 #include "mlir/Interfaces/DataLayoutInterfaces.h" // from @llvm-project
29 #include "mlir/Transforms/DialectConversion.h" // from @llvm-project
30 #include "tensorflow/compiler/mlir/tools/kernel_gen/transforms/passes.h"
31
32 namespace mlir {
33 namespace kernel_gen {
34 namespace transforms {
35
36 using gpu::GPUModuleOp;
37
38 namespace {
39
40 #define GEN_PASS_CLASSES
41 #include "tensorflow/compiler/mlir/tools/kernel_gen/transforms/kernel_gen_passes.h.inc"
42
43 /// A pass that does the final lowering to NVVM. It collects all the patterns
44 /// that are currently required, currently mixing std, linalg and gpu.
45 class GpuKernelToNVVMPass
46 : public GpuKernelToNVVMPassBase<GpuKernelToNVVMPass> {
getDependentDialects(mlir::DialectRegistry & registry) const47 void getDependentDialects(mlir::DialectRegistry& registry) const override {
48 registry.insert<mlir::NVVM::NVVMDialect, mlir::LLVM::LLVMDialect>();
49 }
50
51 public:
runOnOperation()52 void runOnOperation() override {
53 GPUModuleOp m = getOperation();
54
55 RewritePatternSet patterns(&getContext());
56 mlir::LowerToLLVMOptions llvm_opts(
57 m.getContext(),
58 DataLayout(cast<DataLayoutOpInterface>(m.getOperation())));
59 llvm_opts.overrideIndexBitwidth(32);
60 LLVMTypeConverter converter(m.getContext(), llvm_opts);
61 populateMemRefToLLVMConversionPatterns(converter, patterns);
62 populateStdToLLVMConversionPatterns(converter, patterns);
63 populateGpuToNVVMConversionPatterns(converter, patterns);
64 populateComplexToLLVMConversionPatterns(converter, patterns);
65 ConversionTarget target(getContext());
66 configureGpuToNVVMConversionLegality(target);
67 if (failed(mlir::applyFullConversion(m, target, std::move(patterns)))) {
68 signalPassFailure();
69 }
70 }
71 };
72
73 /// A pass that does the final lowering to ROCDL. It collects all the patterns
74 /// that are currently required, currently mixing std, linalg and gpu.
75 class GpuKernelToROCDLPass
76 : public GpuKernelToROCDLPassBase<GpuKernelToROCDLPass> {
getDependentDialects(mlir::DialectRegistry & registry) const77 void getDependentDialects(mlir::DialectRegistry& registry) const override {
78 registry.insert<mlir::ROCDL::ROCDLDialect, mlir::LLVM::LLVMDialect>();
79 }
80
81 public:
runOnOperation()82 void runOnOperation() override {
83 gpu::GPUModuleOp m = getOperation();
84
85 RewritePatternSet patterns(&getContext());
86 LLVMTypeConverter converter(m.getContext());
87 populateMemRefToLLVMConversionPatterns(converter, patterns);
88 populateStdToLLVMConversionPatterns(converter, patterns);
89 populateGpuToROCDLConversionPatterns(converter, patterns);
90 populateComplexToLLVMConversionPatterns(converter, patterns);
91 ConversionTarget target(getContext());
92 configureGpuToROCDLConversionLegality(target);
93 if (failed(mlir::applyFullConversion(m, target, std::move(patterns)))) {
94 signalPassFailure();
95 }
96 }
97 };
98
99 } // namespace
100
CreateGpuKernelToNvvmPass()101 std::unique_ptr<OperationPass<GPUModuleOp> > CreateGpuKernelToNvvmPass() {
102 return std::make_unique<GpuKernelToNVVMPass>();
103 }
104
CreateGpuKernelToRocdlPass()105 std::unique_ptr<OperationPass<GPUModuleOp> > CreateGpuKernelToRocdlPass() {
106 return std::make_unique<GpuKernelToROCDLPass>();
107 }
108
109 } // namespace transforms
110 } // namespace kernel_gen
111 } // namespace mlir
112