|
MLIR-AIE 1.4.6.dev
|
Pattern to emulate an f32 multiply with bf16 inputs. More...
Public Member Functions | |
| LogicalResult | matchAndRewrite (arith::MulFOp op, PatternRewriter &rewriter) const override |
Pattern to emulate an f32 multiply with bf16 inputs.
The product of two bf16 values fits in f32 except at the ends of the exponent range, so only the inputs are rounded: r = arith.mulf a, b : vector<16xf32> becomes r = arith.mulf (extf (truncf a)), (extf (truncf b)) : vector<16xf32>
Definition at line 867 of file VectorToVectorConversions.cpp.
|
inlineoverride |
Definition at line 870 of file VectorToVectorConversions.cpp.