1 //===- LowerGpuOpsToROCDLOps.cpp - MLIR GPU to ROCDL lowering passes ------===//
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 file implements a pass to generate ROCDLIR operations for higher-level
10 // GPU operations.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "mlir/Conversion/GPUToROCDL/GPUToROCDLPass.h"
15 
16 #include "mlir/Conversion/StandardToLLVM/ConvertStandardToLLVMPass.h"
17 #include "mlir/Conversion/VectorToLLVM/ConvertVectorToLLVM.h"
18 #include "mlir/Conversion/VectorToROCDL/VectorToROCDL.h"
19 #include "mlir/Dialect/GPU/GPUDialect.h"
20 #include "mlir/Dialect/GPU/Passes.h"
21 #include "mlir/Dialect/LLVMIR/ROCDLDialect.h"
22 #include "mlir/Dialect/Vector/VectorOps.h"
23 #include "mlir/Pass/Pass.h"
24 #include "mlir/Transforms/DialectConversion.h"
25 #include "mlir/Transforms/GreedyPatternRewriteDriver.h"
26 #include "llvm/Support/FormatVariadic.h"
27 
28 #include "../GPUCommon/GPUOpsLowering.h"
29 #include "../GPUCommon/IndexIntrinsicsOpLowering.h"
30 #include "../GPUCommon/OpToFuncCallLowering.h"
31 #include "../PassDetail.h"
32 
33 using namespace mlir;
34 
35 namespace {
36 
37 /// Import the GPU Ops to ROCDL Patterns.
38 #include "GPUToROCDL.cpp.inc"
39 
40 // A pass that replaces all occurrences of GPU device operations with their
41 // corresponding ROCDL equivalent.
42 //
43 // This pass only handles device code and is not meant to be run on GPU host
44 // code.
45 struct LowerGpuOpsToROCDLOpsPass
46     : public ConvertGpuOpsToROCDLOpsBase<LowerGpuOpsToROCDLOpsPass> {
47   LowerGpuOpsToROCDLOpsPass() = default;
LowerGpuOpsToROCDLOpsPass__anon844d57300111::LowerGpuOpsToROCDLOpsPass48   LowerGpuOpsToROCDLOpsPass(unsigned indexBitwidth) {
49     this->indexBitwidth = indexBitwidth;
50   }
51 
runOnOperation__anon844d57300111::LowerGpuOpsToROCDLOpsPass52   void runOnOperation() override {
53     gpu::GPUModuleOp m = getOperation();
54 
55     /// Customize the bitwidth used for the device side index computations.
56     LowerToLLVMOptions options = {/*useBarePtrCallConv =*/false,
57                                   /*emitCWrappers =*/true,
58                                   /*indexBitwidth =*/indexBitwidth,
59                                   /*useAlignedAlloc =*/false};
60     LLVMTypeConverter converter(m.getContext(), options);
61 
62     OwningRewritePatternList patterns, llvmPatterns;
63 
64     populateGpuRewritePatterns(m.getContext(), patterns);
65     applyPatternsAndFoldGreedily(m, std::move(patterns));
66 
67     populateVectorToLLVMConversionPatterns(converter, llvmPatterns);
68     populateVectorToROCDLConversionPatterns(converter, llvmPatterns);
69     populateStdToLLVMConversionPatterns(converter, llvmPatterns);
70     populateGpuToROCDLConversionPatterns(converter, llvmPatterns);
71     LLVMConversionTarget target(getContext());
72     target.addIllegalDialect<gpu::GPUDialect>();
73     target.addIllegalOp<LLVM::CosOp, LLVM::ExpOp, LLVM::FAbsOp, LLVM::FCeilOp,
74                         LLVM::FFloorOp, LLVM::LogOp, LLVM::Log10Op,
75                         LLVM::Log2Op, LLVM::SinOp>();
76     target.addIllegalOp<FuncOp>();
77     target.addLegalDialect<ROCDL::ROCDLDialect>();
78     // TODO: Remove once we support replacing non-root ops.
79     target.addLegalOp<gpu::YieldOp, gpu::GPUModuleOp, gpu::ModuleEndOp>();
80     if (failed(applyPartialConversion(m, target, std::move(llvmPatterns))))
81       signalPassFailure();
82   }
83 };
84 
85 } // anonymous namespace
86 
populateGpuToROCDLConversionPatterns(LLVMTypeConverter & converter,OwningRewritePatternList & patterns)87 void mlir::populateGpuToROCDLConversionPatterns(
88     LLVMTypeConverter &converter, OwningRewritePatternList &patterns) {
89   populateWithGenerated(converter.getDialect()->getContext(), patterns);
90   patterns.insert<
91       GPUIndexIntrinsicOpLowering<gpu::ThreadIdOp, ROCDL::ThreadIdXOp,
92                                   ROCDL::ThreadIdYOp, ROCDL::ThreadIdZOp>,
93       GPUIndexIntrinsicOpLowering<gpu::BlockDimOp, ROCDL::BlockDimXOp,
94                                   ROCDL::BlockDimYOp, ROCDL::BlockDimZOp>,
95       GPUIndexIntrinsicOpLowering<gpu::BlockIdOp, ROCDL::BlockIdXOp,
96                                   ROCDL::BlockIdYOp, ROCDL::BlockIdZOp>,
97       GPUIndexIntrinsicOpLowering<gpu::GridDimOp, ROCDL::GridDimXOp,
98                                   ROCDL::GridDimYOp, ROCDL::GridDimZOp>,
99       GPUFuncOpLowering<5>, GPUReturnOpLowering>(converter);
100   patterns.insert<OpToFuncCallLowering<AbsFOp>>(converter, "__ocml_fabs_f32",
101                                                 "__ocml_fabs_f64");
102   patterns.insert<OpToFuncCallLowering<CeilFOp>>(converter, "__ocml_ceil_f32",
103                                                  "__ocml_ceil_f64");
104   patterns.insert<OpToFuncCallLowering<CosOp>>(converter, "__ocml_cos_f32",
105                                                "__ocml_cos_f64");
106   patterns.insert<OpToFuncCallLowering<ExpOp>>(converter, "__ocml_exp_f32",
107                                                "__ocml_exp_f64");
108   patterns.insert<OpToFuncCallLowering<FloorFOp>>(converter, "__ocml_floor_f32",
109                                                   "__ocml_floor_f64");
110   patterns.insert<OpToFuncCallLowering<LogOp>>(converter, "__ocml_log_f32",
111                                                "__ocml_log_f64");
112   patterns.insert<OpToFuncCallLowering<Log10Op>>(converter, "__ocml_log10_f32",
113                                                  "__ocml_log10_f64");
114   patterns.insert<OpToFuncCallLowering<Log2Op>>(converter, "__ocml_log2_f32",
115                                                 "__ocml_log2_f64");
116   patterns.insert<OpToFuncCallLowering<SinOp>>(converter, "__ocml_sin_f32",
117                                                "__ocml_sin_f64");
118   patterns.insert<OpToFuncCallLowering<TanhOp>>(converter, "__ocml_tanh_f32",
119                                                 "__ocml_tanh_f64");
120 }
121 
122 std::unique_ptr<OperationPass<gpu::GPUModuleOp>>
createLowerGpuOpsToROCDLOpsPass(unsigned indexBitwidth)123 mlir::createLowerGpuOpsToROCDLOpsPass(unsigned indexBitwidth) {
124   return std::make_unique<LowerGpuOpsToROCDLOpsPass>(indexBitwidth);
125 }
126