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/Dialect/GPU/GPUDialect.h"
18 #include "mlir/Dialect/LLVMIR/ROCDLDialect.h"
19 #include "mlir/Pass/Pass.h"
20 #include "mlir/Transforms/DialectConversion.h"
21 
22 #include "../GPUCommon/IndexIntrinsicsOpLowering.h"
23 #include "../GPUCommon/OpToFuncCallLowering.h"
24 
25 using namespace mlir;
26 
27 namespace {
28 
29 // A pass that replaces all occurrences of GPU device operations with their
30 // corresponding ROCDL equivalent.
31 //
32 // This pass only handles device code and is not meant to be run on GPU host
33 // code.
34 class LowerGpuOpsToROCDLOpsPass
35     : public OperationPass<LowerGpuOpsToROCDLOpsPass, gpu::GPUModuleOp> {
36 public:
37 /// Include the generated pass utilities.
38 #define GEN_PASS_ConvertGpuOpsToROCDLOps
39 #include "mlir/Conversion/Passes.h.inc"
40 
41   void runOnOperation() override {
42     gpu::GPUModuleOp m = getOperation();
43 
44     OwningRewritePatternList patterns;
45     LLVMTypeConverter converter(m.getContext());
46     populateStdToLLVMConversionPatterns(converter, patterns);
47     patterns.insert<
48         GPUIndexIntrinsicOpLowering<gpu::ThreadIdOp, ROCDL::ThreadIdXOp,
49                                     ROCDL::ThreadIdYOp, ROCDL::ThreadIdZOp>,
50         GPUIndexIntrinsicOpLowering<gpu::BlockDimOp, ROCDL::BlockDimXOp,
51                                     ROCDL::BlockDimYOp, ROCDL::BlockDimZOp>,
52         GPUIndexIntrinsicOpLowering<gpu::BlockIdOp, ROCDL::BlockIdXOp,
53                                     ROCDL::BlockIdYOp, ROCDL::BlockIdZOp>,
54         GPUIndexIntrinsicOpLowering<gpu::GridDimOp, ROCDL::GridDimXOp,
55                                     ROCDL::GridDimYOp, ROCDL::GridDimZOp>>(
56         converter);
57     patterns.insert<OpToFuncCallLowering<AbsFOp>>(converter, "__ocml_fabs_f32",
58                                                   "__ocml_fabs_f64");
59     patterns.insert<OpToFuncCallLowering<CeilFOp>>(converter, "__ocml_ceil_f32",
60                                                    "__ocml_ceil_f64");
61     patterns.insert<OpToFuncCallLowering<CosOp>>(converter, "__ocml_cos_f32",
62                                                  "__ocml_cos_f64");
63     patterns.insert<OpToFuncCallLowering<ExpOp>>(converter, "__ocml_exp_f32",
64                                                  "__ocml_exp_f64");
65     patterns.insert<OpToFuncCallLowering<LogOp>>(converter, "__ocml_log_f32",
66                                                  "__ocml_log_f64");
67     patterns.insert<OpToFuncCallLowering<Log10Op>>(
68         converter, "__ocml_log10_f32", "__ocml_log10_f64");
69     patterns.insert<OpToFuncCallLowering<Log2Op>>(converter, "__ocml_log2_f32",
70                                                   "__ocml_log2_f64");
71     patterns.insert<OpToFuncCallLowering<TanhOp>>(converter, "__ocml_tanh_f32",
72                                                   "__ocml_tanh_f64");
73 
74     ConversionTarget target(getContext());
75     target.addLegalDialect<LLVM::LLVMDialect, ROCDL::ROCDLDialect>();
76     target.addIllegalOp<LLVM::CosOp, LLVM::ExpOp, LLVM::FAbsOp, LLVM::FCeilOp,
77                         LLVM::LogOp, LLVM::Log10Op, LLVM::Log2Op>();
78     target.addIllegalOp<FuncOp>();
79     if (failed(applyPartialConversion(m, target, patterns, &converter)))
80       signalPassFailure();
81   }
82 };
83 
84 } // anonymous namespace
85 
86 std::unique_ptr<OpPassBase<gpu::GPUModuleOp>>
87 mlir::createLowerGpuOpsToROCDLOpsPass() {
88   return std::make_unique<LowerGpuOpsToROCDLOpsPass>();
89 }
90 
91