| 1 | //===-- MathToROCDL.cpp - conversion from Math to rocdl calls -------------===// |
| 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 | #include "mlir/Conversion/MathToROCDL/MathToROCDL.h" |
| 10 | #include "mlir/Conversion/GPUCommon/GPUCommonPass.h" |
| 11 | #include "mlir/Conversion/LLVMCommon/LoweringOptions.h" |
| 12 | #include "mlir/Conversion/LLVMCommon/TypeConverter.h" |
| 13 | #include "mlir/Conversion/LLVMCommon/VectorPattern.h" |
| 14 | #include "mlir/Dialect/Arith/IR/Arith.h" |
| 15 | #include "mlir/Dialect/Func/IR/FuncDialect.h" |
| 16 | #include "mlir/Dialect/LLVMIR/LLVMArithOps.h" |
| 17 | #include "mlir/Dialect/LLVMIR/LLVMFloatIntrinsicOps.h" |
| 18 | #include "mlir/Dialect/LLVMIR/ROCDLDialect.h" |
| 19 | #include "mlir/Dialect/LLVMIR/ROCDLTargetInfo.h" |
| 20 | #include "mlir/Dialect/Math/IR/Math.h" |
| 21 | #include "mlir/Dialect/Vector/IR/VectorOps.h" |
| 22 | #include "mlir/IR/BuiltinDialect.h" |
| 23 | #include "mlir/IR/PatternMatch.h" |
| 24 | #include "mlir/Pass/Pass.h" |
| 25 | #include "mlir/Transforms/DialectConversion.h" |
| 26 | #include "llvm/Support/DebugLog.h" |
| 27 | |
| 28 | #include "../GPUCommon/GPUOpsLowering.h" |
| 29 | #include "../GPUCommon/OpToFuncCallLowering.h" |
| 30 | |
| 31 | namespace mlir { |
| 32 | #define GEN_PASS_DEF_CONVERTMATHTOROCDL |
| 33 | #include "mlir/Conversion/Passes.h.inc" |
| 34 | } // namespace mlir |
| 35 | |
| 36 | using namespace mlir; |
| 37 | |
| 38 | #define DEBUG_TYPE "math-to-rocdl" |
| 39 | |
| 40 | template <typename OpTy> |
| 41 | static void populateOpPatterns(const LLVMTypeConverter &converter, |
| 42 | RewritePatternSet &patterns, StringRef f32Func, |
| 43 | StringRef f64Func, StringRef f16Func, |
| 44 | StringRef f32ApproxFunc = "" ) { |
| 45 | patterns.add<ScalarizeVectorOpLowering<OpTy>>(converter); |
| 46 | patterns.add<OpToFuncCallLowering<OpTy>>(converter, f32Func, f64Func, |
| 47 | f32ApproxFunc, f16Func); |
| 48 | } |
| 49 | |
| 50 | struct ClampFOpConversion final |
| 51 | : public ConvertOpToLLVMPattern<math::ClampFOp> { |
| 52 | using ConvertOpToLLVMPattern::ConvertOpToLLVMPattern; |
| 53 | |
| 54 | LogicalResult |
| 55 | matchAndRewrite(math::ClampFOp op, OpAdaptor adaptor, |
| 56 | ConversionPatternRewriter &rewriter) const override { |
| 57 | // Only f16 and f32 types are supported by fmed3 |
| 58 | Type opTy = op.getType(); |
| 59 | Type resultType = getTypeConverter()->convertType(t: opTy); |
| 60 | |
| 61 | if (auto vectorType = dyn_cast<VectorType>(Val&: opTy)) |
| 62 | opTy = vectorType.getElementType(); |
| 63 | |
| 64 | if (!isa<Float16Type, Float32Type>(Val: opTy)) |
| 65 | return rewriter.notifyMatchFailure( |
| 66 | arg&: op, msg: "fmed3 only supports f16 and f32 types" ); |
| 67 | |
| 68 | // Handle multi-dimensional vectors (converted to LLVM arrays) |
| 69 | if (auto arrayType = dyn_cast<LLVM::LLVMArrayType>(Val&: resultType)) |
| 70 | return LLVM::detail::handleMultidimensionalVectors( |
| 71 | op: op.getOperation(), operands: adaptor.getOperands(), typeConverter: *getTypeConverter(), |
| 72 | createOperand: [&](Type llvm1DVectorTy, ValueRange operands) -> Value { |
| 73 | math::ClampFOp::Adaptor adaptor(operands); |
| 74 | return ROCDL::FMed3Op::create(builder&: rewriter, location: op.getLoc(), res: llvm1DVectorTy, |
| 75 | src0: adaptor.getValue(), src1: adaptor.getMin(), |
| 76 | src2: adaptor.getMax()); |
| 77 | }, |
| 78 | rewriter); |
| 79 | |
| 80 | // Handle 1D vectors and scalars directly |
| 81 | rewriter.replaceOpWithNewOp<ROCDL::FMed3Op>(op, args: op.getType(), args: op.getValue(), |
| 82 | args: op.getMin(), args: op.getMax()); |
| 83 | return success(); |
| 84 | } |
| 85 | }; |
| 86 | |
| 87 | void mlir::populateMathToROCDLConversionPatterns( |
| 88 | const LLVMTypeConverter &converter, RewritePatternSet &patterns, |
| 89 | std::optional<ROCDL::TargetInfo> target) { |
| 90 | // Handled by mathToLLVM: math::AbsIOp |
| 91 | // Handled by mathToLLVM: math::AbsFOp |
| 92 | // Handled by mathToLLVM: math::CopySignOp |
| 93 | // Handled by mathToLLVM: math::CountLeadingZerosOp |
| 94 | // Handled by mathToLLVM: math::CountTrailingZerosOp |
| 95 | // Handled by mathToLLVM: math::CgPopOp |
| 96 | // Handled by mathToLLVM: math::ExpOp (32-bit only) |
| 97 | // Handled by mathToLLVM: math::FmaOp |
| 98 | // Handled by mathToLLVM: math::LogOp (32-bit only) |
| 99 | // FIXME: math::IPowIOp |
| 100 | // Handled by mathToLLVM: math::RoundEvenOp |
| 101 | // Handled by mathToLLVM: math::RoundOp |
| 102 | // Handled by mathToLLVM: math::SqrtOp |
| 103 | // Handled by mathToLLVM: math::TruncOp |
| 104 | populateOpPatterns<math::AcosOp>(converter, patterns, f32Func: "__ocml_acos_f32" , |
| 105 | f64Func: "__ocml_acos_f64" , f16Func: "__ocml_acos_f16" ); |
| 106 | populateOpPatterns<math::AcoshOp>(converter, patterns, f32Func: "__ocml_acosh_f32" , |
| 107 | f64Func: "__ocml_acosh_f64" , f16Func: "__ocml_acosh_f16" ); |
| 108 | populateOpPatterns<math::AsinOp>(converter, patterns, f32Func: "__ocml_asin_f32" , |
| 109 | f64Func: "__ocml_asin_f64" , f16Func: "__ocml_asin_f16" ); |
| 110 | populateOpPatterns<math::AsinhOp>(converter, patterns, f32Func: "__ocml_asinh_f32" , |
| 111 | f64Func: "__ocml_asinh_f64" , f16Func: "__ocml_asinh_f16" ); |
| 112 | populateOpPatterns<math::AtanOp>(converter, patterns, f32Func: "__ocml_atan_f32" , |
| 113 | f64Func: "__ocml_atan_f64" , f16Func: "__ocml_atan_f16" ); |
| 114 | populateOpPatterns<math::AtanhOp>(converter, patterns, f32Func: "__ocml_atanh_f32" , |
| 115 | f64Func: "__ocml_atanh_f64" , f16Func: "__ocml_atanh_f16" ); |
| 116 | populateOpPatterns<math::Atan2Op>(converter, patterns, f32Func: "__ocml_atan2_f32" , |
| 117 | f64Func: "__ocml_atan2_f64" , f16Func: "__ocml_atan2_f16" ); |
| 118 | populateOpPatterns<math::CbrtOp>(converter, patterns, f32Func: "__ocml_cbrt_f32" , |
| 119 | f64Func: "__ocml_cbrt_f64" , f16Func: "__ocml_cbrt_f16" ); |
| 120 | populateOpPatterns<math::CeilOp>(converter, patterns, f32Func: "__ocml_ceil_f32" , |
| 121 | f64Func: "__ocml_ceil_f64" , f16Func: "__ocml_ceil_f16" ); |
| 122 | populateOpPatterns<math::CosOp>(converter, patterns, f32Func: "__ocml_cos_f32" , |
| 123 | f64Func: "__ocml_cos_f64" , f16Func: "__ocml_cos_f16" ); |
| 124 | populateOpPatterns<math::CoshOp>(converter, patterns, f32Func: "__ocml_cosh_f32" , |
| 125 | f64Func: "__ocml_cosh_f64" , f16Func: "__ocml_cosh_f16" ); |
| 126 | populateOpPatterns<math::SinhOp>(converter, patterns, f32Func: "__ocml_sinh_f32" , |
| 127 | f64Func: "__ocml_sinh_f64" , f16Func: "__ocml_sinh_f16" ); |
| 128 | populateOpPatterns<math::ExpOp>(converter, patterns, f32Func: "" , f64Func: "__ocml_exp_f64" , |
| 129 | f16Func: "__ocml_exp_f16" ); |
| 130 | populateOpPatterns<math::Exp2Op>(converter, patterns, f32Func: "__ocml_exp2_f32" , |
| 131 | f64Func: "__ocml_exp2_f64" , f16Func: "__ocml_exp2_f16" ); |
| 132 | populateOpPatterns<math::ExpM1Op>(converter, patterns, f32Func: "__ocml_expm1_f32" , |
| 133 | f64Func: "__ocml_expm1_f64" , f16Func: "__ocml_expm1_f16" ); |
| 134 | populateOpPatterns<math::FloorOp>(converter, patterns, f32Func: "__ocml_floor_f32" , |
| 135 | f64Func: "__ocml_floor_f64" , f16Func: "__ocml_floor_f16" ); |
| 136 | populateOpPatterns<math::LogOp>(converter, patterns, f32Func: "" , f64Func: "__ocml_log_f64" , |
| 137 | f16Func: "__ocml_log_f16" ); |
| 138 | populateOpPatterns<math::Log10Op>(converter, patterns, f32Func: "__ocml_log10_f32" , |
| 139 | f64Func: "__ocml_log10_f64" , f16Func: "__ocml_log10_f16" ); |
| 140 | populateOpPatterns<math::Log1pOp>(converter, patterns, f32Func: "__ocml_log1p_f32" , |
| 141 | f64Func: "__ocml_log1p_f64" , f16Func: "__ocml_log1p_f16" ); |
| 142 | populateOpPatterns<math::Log2Op>(converter, patterns, f32Func: "__ocml_log2_f32" , |
| 143 | f64Func: "__ocml_log2_f64" , f16Func: "__ocml_log2_f16" ); |
| 144 | populateOpPatterns<math::PowFOp>(converter, patterns, f32Func: "__ocml_pow_f32" , |
| 145 | f64Func: "__ocml_pow_f64" , f16Func: "__ocml_pow_f16" ); |
| 146 | populateOpPatterns<math::RsqrtOp>(converter, patterns, f32Func: "__ocml_rsqrt_f32" , |
| 147 | f64Func: "__ocml_rsqrt_f64" , f16Func: "__ocml_rsqrt_f16" ); |
| 148 | populateOpPatterns<math::SinOp>(converter, patterns, f32Func: "__ocml_sin_f32" , |
| 149 | f64Func: "__ocml_sin_f64" , f16Func: "__ocml_sin_f16" ); |
| 150 | populateOpPatterns<math::TanhOp>(converter, patterns, f32Func: "__ocml_tanh_f32" , |
| 151 | f64Func: "__ocml_tanh_f64" , f16Func: "__ocml_tanh_f16" ); |
| 152 | populateOpPatterns<math::TanOp>(converter, patterns, f32Func: "__ocml_tan_f32" , |
| 153 | f64Func: "__ocml_tan_f64" , f16Func: "__ocml_tan_f16" ); |
| 154 | populateOpPatterns<math::ErfOp>(converter, patterns, f32Func: "__ocml_erf_f32" , |
| 155 | f64Func: "__ocml_erf_f64" , f16Func: "__ocml_erf_f16" ); |
| 156 | populateOpPatterns<math::ErfcOp>(converter, patterns, f32Func: "__ocml_erfc_f32" , |
| 157 | f64Func: "__ocml_erfc_f64" , f16Func: "__ocml_erfc_f16" ); |
| 158 | populateOpPatterns<math::FPowIOp>(converter, patterns, f32Func: "__ocml_pown_f32" , |
| 159 | f64Func: "__ocml_pown_f64" , f16Func: "__ocml_pown_f16" ); |
| 160 | // Single arith pattern that needs a ROCDL call, probably not |
| 161 | // worth creating a separate pass for it. |
| 162 | populateOpPatterns<arith::RemFOp>(converter, patterns, f32Func: "__ocml_fmod_f32" , |
| 163 | f64Func: "__ocml_fmod_f64" , f16Func: "__ocml_fmod_f16" ); |
| 164 | |
| 165 | if (target && target->has(feature: llvm::AMDGPU::FEAT_GFX9_INSTS)) { |
| 166 | patterns.add<ClampFOpConversion>(arg: converter); |
| 167 | } else { |
| 168 | LDBG() << "Target dependent patterns were not added" ; |
| 169 | } |
| 170 | } |
| 171 | |
| 172 | struct ConvertMathToROCDLPass final |
| 173 | : impl::ConvertMathToROCDLBase<ConvertMathToROCDLPass> { |
| 174 | using impl::ConvertMathToROCDLBase< |
| 175 | ConvertMathToROCDLPass>::ConvertMathToROCDLBase; |
| 176 | |
| 177 | void runOnOperation() override; |
| 178 | }; |
| 179 | |
| 180 | void ConvertMathToROCDLPass::runOnOperation() { |
| 181 | auto m = getOperation(); |
| 182 | MLIRContext *ctx = m.getContext(); |
| 183 | |
| 184 | RewritePatternSet patterns(&getContext()); |
| 185 | LowerToLLVMOptions options(ctx, DataLayout(m)); |
| 186 | LLVMTypeConverter converter(ctx, options); |
| 187 | |
| 188 | // An empty architecture means "no target", in which case the |
| 189 | // target-dependent patterns are simply not added. |
| 190 | std::optional<ROCDL::TargetInfo> resolved; |
| 191 | StringRef resolvedArch = ROCDL::resolveArchOption(arch, deprecatedAlias: chipset); |
| 192 | if (!resolvedArch.empty()) { |
| 193 | FailureOr<ROCDL::TargetInfo> targetInfo = |
| 194 | ROCDL::TargetInfo::get(arch: resolvedArch, /*waveSize=*/0, emitError: [&] { |
| 195 | return emitError(loc: UnknownLoc::get(context: &getContext())); |
| 196 | }); |
| 197 | if (failed(Result: targetInfo)) |
| 198 | return signalPassFailure(); |
| 199 | resolved = *targetInfo; |
| 200 | } |
| 201 | populateMathToROCDLConversionPatterns(converter, patterns, target: resolved); |
| 202 | |
| 203 | ConversionTarget target(getContext()); |
| 204 | target |
| 205 | .addLegalDialect<BuiltinDialect, func::FuncDialect, vector::VectorDialect, |
| 206 | LLVM::LLVMDialect, ROCDL::ROCDLDialect>(); |
| 207 | target.addIllegalOp<LLVM::CosOp, LLVM::ExpOp, LLVM::Exp2Op, LLVM::FAbsOp, |
| 208 | LLVM::FCeilOp, LLVM::FFloorOp, LLVM::FRemOp, LLVM::LogOp, |
| 209 | LLVM::Log10Op, LLVM::Log2Op, LLVM::PowOp, LLVM::SinOp, |
| 210 | LLVM::SqrtOp>(); |
| 211 | if (failed(Result: applyPartialConversion(op: m, target, patterns: std::move(patterns)))) |
| 212 | signalPassFailure(); |
| 213 | } |
| 214 | |