FoldBF16RoundTripPattern Struct Reference

Pattern to drop bf16 -> f32 -> bf16 round trips. More...

Inheritance diagram for FoldBF16RoundTripPattern:

Collaboration diagram for FoldBF16RoundTripPattern:

Public Member Functions
LogicalResult	matchAndRewrite (arith::TruncFOp op, PatternRewriter &rewriter) const override

Detailed Description

Pattern to drop bf16 -> f32 -> bf16 round trips.

Arith used to fold these but no longer does: widening a signaling NaN quiets it, so the original bit pattern cannot be recovered. This pass has always made that trade anyway – smartTruncF32ToBF16 collapses the same pair on sight – so fold here to keep the pipeline's pre-existing behavior rather than leaving the round trips for the backend.

Definition at line 755 of file VectorToVectorConversions.cpp.

Member Function Documentation

matchAndRewrite()

LogicalResult FoldBF16RoundTripPattern::matchAndRewrite ( arith::TruncFOp op, PatternRewriter & rewriter ) const

inlineoverride

Definition at line 758 of file VectorToVectorConversions.cpp.

---

The documentation for this struct was generated from the following file:

• lib/Dialect/AIEVec/Transforms/VectorToVectorConversions.cpp