[Mlir-commits] [mlir] [MLIR][Transforms] add eliminate-explicit-rounding pass (PR #93443)

[Ivy Zhang]

================
@@ -0,0 +1,86 @@
+//===- EliminateExplicitRounding.cpp - Remove intermediate extf/truncf pairs
+//-===//
+//
+// 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
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements removing intermediate extf/truncf pairs inserted from
+// type conversion.
+//
+//===----------------------------------------------------------------------===//
+#include "mlir/Dialect/Arith/Transforms/Passes.h"
+
+#include "mlir/Dialect/Arith/IR/Arith.h"
+#include "mlir/IR/PatternMatch.h"
+#include "mlir/IR/TypeUtilities.h"
+#include "mlir/Transforms/GreedyPatternRewriteDriver.h"
+
+namespace mlir {
+namespace arith {
+#define GEN_PASS_DEF_ELIMINATEEXPLICITROUNDING
+#include "mlir/Dialect/Arith/Transforms/Passes.h.inc"
+} // namespace arith
+} // namespace mlir
+
+using namespace mlir;
+
+namespace {
+
+struct EliminateExplicitRoundingRewritePattern final
+    : OpRewritePattern<arith::ExtFOp> {
+  using OpRewritePattern::OpRewritePattern;
+  using FilterFunction = std::function<bool(Operation *)>;
+
+  EliminateExplicitRoundingRewritePattern(MLIRContext *context,
+                                          FilterFunction filterFunc = nullptr)
+      : OpRewritePattern(context), filterFunc(filterFunc) {}
+
+  LogicalResult matchAndRewrite(arith::ExtFOp extFOp,
+                                PatternRewriter &rewriter) const final {
+    // Check whether match `truncF->extF` pair.
+    auto truncFOp = extFOp.getOperand().getDefiningOp<arith::TruncFOp>();
+    if (!truncFOp)
+      return failure();
+
+    // Check whether need to filter out.
+    if (filterFunc && filterFunc(extFOp))
+      return failure();
+
+    // Check whether the rounding pair's input and output data type are the
+    // same. Currently only consider to eliminate rounding pairs for (bf16 / f16
+    // <-> f32).
+    auto input = truncFOp.getOperand();
+    auto inTy = input.getType();
+    auto outTy = extFOp.getType();
+    auto shortTy = getElementTypeOrSelf(truncFOp.getType());
+    if (inTy == outTy && getElementTypeOrSelf(inTy).isF32() &&
+        (shortTy.isF16() || shortTy.isBF16())) {
+      rewriter.replaceOp(extFOp, {input});
+      return success();
+    }
+
+    return failure();
+  }
+
+private:
+  FilterFunction filterFunc;
+};
+
+struct EliminateExplicitRounding final
+    : arith::impl::EliminateExplicitRoundingBase<EliminateExplicitRounding> {
+  using EliminateExplicitRoundingBase::EliminateExplicitRoundingBase;
+  void runOnOperation() override {
+    RewritePatternSet patterns(&getContext());
+    patterns.insert<EliminateExplicitRoundingRewritePattern>(&getContext());
+    FrozenRewritePatternSet patternSet(std::move(patterns));
+    SmallVector<Operation *> ops;
+    getOperation()->walk([&](arith::ExtFOp op) { ops.push_back(op); });
+    if (failed(applyOpPatternsAndFold(ops, patternSet)))
+      signalPassFailure();
----------------
crazydemo wrote:

If there are cases with nested trunc / ext op pairs, then a greedy pattern rewriter would be needed. However, in the current design, this pass is executed only after all f32 promotions are completed, so nested cases do not occur. The code snippet above is an example of a simple IR walk and rewrite.

https://github.com/llvm/llvm-project/pull/93443