|
MLIR-AIE
|
Pattern to emulate f32 binary vector arithmetic ops in bf16. More...
Inheritance diagram for EmulateBinaryF32InBF16Pattern< OpTy >:
Collaboration diagram for EmulateBinaryF32InBF16Pattern< OpTy >:Public Member Functions | |
| LogicalResult | matchAndRewrite (OpTy op, PatternRewriter &rewriter) const override |
Pattern to emulate f32 binary vector arithmetic ops in bf16.
For an op like: r = arith.addf a, b : vector<16xf32> Produces: a_bf16 = arith.truncf a : vector<16xf32> to vector<16xbf16> b_bf16 = arith.truncf b : vector<16xf32> to vector<16xbf16> r_bf16 = arith.addf a_bf16, b_bf16 : vector<16xbf16> r = arith.extf r_bf16 : vector<16xbf16> to vector<16xf32>
Definition at line 922 of file VectorToVectorConversions.cpp.
|
inlineoverride |
Definition at line 925 of file VectorToVectorConversions.cpp.