xref: /llvm-project/mlir/lib/Dialect/LLVMIR/Transforms/OptimizeForNVVM.cpp (revision 09dfc5713d7e2342bea4c8447d1ed76c85eb8225)
1 //===- OptimizeForNVVM.cpp - Optimize LLVM IR for NVVM ---------===//
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/Dialect/LLVMIR/Transforms/OptimizeForNVVM.h"
10 
11 #include "mlir/Dialect/LLVMIR/NVVMDialect.h"
12 #include "mlir/IR/Builders.h"
13 #include "mlir/IR/PatternMatch.h"
14 #include "mlir/Pass/Pass.h"
15 #include "mlir/Transforms/GreedyPatternRewriteDriver.h"
16 
17 namespace mlir {
18 namespace NVVM {
19 #define GEN_PASS_DEF_NVVMOPTIMIZEFORTARGET
20 #include "mlir/Dialect/LLVMIR/Transforms/Passes.h.inc"
21 } // namespace NVVM
22 } // namespace mlir
23 
24 using namespace mlir;
25 
26 namespace {
27 // Replaces fdiv on fp16 with fp32 multiplication with reciprocal plus one
28 // (conditional) Newton iteration.
29 //
30 // This as accurate as promoting the division to fp32 in the NVPTX backend, but
31 // faster because it performs less Newton iterations, avoids the slow path
32 // for e.g. denormals, and allows reuse of the reciprocal for multiple divisions
33 // by the same divisor.
34 struct ExpandDivF16 : public OpRewritePattern<LLVM::FDivOp> {
35   using OpRewritePattern<LLVM::FDivOp>::OpRewritePattern;
36 
37 private:
38   LogicalResult matchAndRewrite(LLVM::FDivOp op,
39                                 PatternRewriter &rewriter) const override;
40 };
41 
42 struct NVVMOptimizeForTarget
43     : public NVVM::impl::NVVMOptimizeForTargetBase<NVVMOptimizeForTarget> {
44   void runOnOperation() override;
45 
46   void getDependentDialects(DialectRegistry &registry) const override {
47     registry.insert<NVVM::NVVMDialect>();
48   }
49 };
50 } // namespace
51 
52 LogicalResult ExpandDivF16::matchAndRewrite(LLVM::FDivOp op,
53                                             PatternRewriter &rewriter) const {
54   if (!op.getType().isF16())
55     return rewriter.notifyMatchFailure(op, "not f16");
56   Location loc = op.getLoc();
57 
58   Type f32Type = rewriter.getF32Type();
59   Type i32Type = rewriter.getI32Type();
60 
61   // Extend lhs and rhs to fp32.
62   Value lhs = rewriter.create<LLVM::FPExtOp>(loc, f32Type, op.getLhs());
63   Value rhs = rewriter.create<LLVM::FPExtOp>(loc, f32Type, op.getRhs());
64 
65   // float rcp = rcp.approx.ftz.f32(rhs), approx = lhs * rcp.
66   Value rcp = rewriter.create<NVVM::RcpApproxFtzF32Op>(loc, f32Type, rhs);
67   Value approx = rewriter.create<LLVM::FMulOp>(loc, lhs, rcp);
68 
69   // Refine the approximation with one Newton iteration:
70   // float refined = approx + (lhs - approx * rhs) * rcp;
71   Value err = rewriter.create<LLVM::FMAOp>(
72       loc, approx, rewriter.create<LLVM::FNegOp>(loc, rhs), lhs);
73   Value refined = rewriter.create<LLVM::FMAOp>(loc, err, rcp, approx);
74 
75   // Use refined value if approx is normal (exponent neither all 0 or all 1).
76   Value mask = rewriter.create<LLVM::ConstantOp>(
77       loc, i32Type, rewriter.getUI32IntegerAttr(0x7f800000));
78   Value cast = rewriter.create<LLVM::BitcastOp>(loc, i32Type, approx);
79   Value exp = rewriter.create<LLVM::AndOp>(loc, i32Type, cast, mask);
80   Value zero = rewriter.create<LLVM::ConstantOp>(
81       loc, i32Type, rewriter.getUI32IntegerAttr(0));
82   Value pred = rewriter.create<LLVM::OrOp>(
83       loc,
84       rewriter.create<LLVM::ICmpOp>(loc, LLVM::ICmpPredicate::eq, exp, zero),
85       rewriter.create<LLVM::ICmpOp>(loc, LLVM::ICmpPredicate::eq, exp, mask));
86   Value result =
87       rewriter.create<LLVM::SelectOp>(loc, f32Type, pred, approx, refined);
88 
89   // Replace with trucation back to fp16.
90   rewriter.replaceOpWithNewOp<LLVM::FPTruncOp>(op, op.getType(), result);
91 
92   return success();
93 }
94 
95 void NVVMOptimizeForTarget::runOnOperation() {
96   MLIRContext *ctx = getOperation()->getContext();
97   RewritePatternSet patterns(ctx);
98   patterns.add<ExpandDivF16>(ctx);
99   if (failed(applyPatternsGreedily(getOperation(), std::move(patterns))))
100     return signalPassFailure();
101 }
102 
103 std::unique_ptr<Pass> NVVM::createOptimizeForTargetPass() {
104   return std::make_unique<NVVMOptimizeForTarget>();
105 }
106