[Mlir-commits] [mlir] [mlir][tosa] Lower MATMUL_T to Linalg (PR #222661)

[Jack Frankland]

FranklandJack wrote:

> Codex pointed out:
> 
> Valid mixed-precision inputs generate invalid IR — mlir/lib/Conversion/TosaToLinalg/TosaToLinalgNamed.cpp:799. FP16 A × FP8 B → BF16 output passes strict TOSA validation. However, Linalg’s scalar conversion leaves FP16 unchanged because it has the same bit width as BF16. The resulting arith.mulf(f16, bf16) fails verification. This combination needs explicit FP16-to-BF16 conversion handling.

Good point, will add fix for this with some lit tests, thanks codex :) 

https://github.com/llvm/llvm-project/pull/222661