llvm/mlir/lib/Dialect/LLVMIR/Transforms/OptimizeForNVVM.cpp

//===- OptimizeForNVVM.cpp - Optimize LLVM IR for NVVM ---------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//

#include "mlir/Dialect/LLVMIR/Transforms/OptimizeForNVVM.h"

#include "mlir/Dialect/LLVMIR/NVVMDialect.h"
#include "mlir/IR/Builders.h"
#include "mlir/IR/PatternMatch.h"
#include "mlir/Pass/Pass.h"
#include "mlir/Transforms/GreedyPatternRewriteDriver.h"

namespace mlir {
namespace NVVM {
#define GEN_PASS_DEF_NVVMOPTIMIZEFORTARGET
#include "mlir/Dialect/LLVMIR/Transforms/Passes.h.inc"
} // namespace NVVM
} // namespace mlir

usingnamespacemlir;

namespace {
// Replaces fdiv on fp16 with fp32 multiplication with reciprocal plus one
// (conditional) Newton iteration.
//
// This as accurate as promoting the division to fp32 in the NVPTX backend, but
// faster because it performs less Newton iterations, avoids the slow path
// for e.g. denormals, and allows reuse of the reciprocal for multiple divisions
// by the same divisor.
struct ExpandDivF16 : public OpRewritePattern<LLVM::FDivOp> { … };

struct NVVMOptimizeForTarget
    : public NVVM::impl::NVVMOptimizeForTargetBase<NVVMOptimizeForTarget> { … };
} // namespace

LogicalResult ExpandDivF16::matchAndRewrite(LLVM::FDivOp op,
                                            PatternRewriter &rewriter) const { … }

void NVVMOptimizeForTarget::runOnOperation() { … }

std::unique_ptr<Pass> NVVM::createOptimizeForTargetPass() { … }