[Mlir-commits] [mlir] [mlir][x86vector] AVX512-BF16 Dot op (PR #124800)

[Renato Golin]

rengolin wrote:

> 1. IREE-wise, we are not directly concerned because we use microkernels for matrix multiplication kernels on CPU, e.g. for avx512bf16: https://github.com/iree-org/iree/blob/main/runtime/src/iree/builtins/ukernel/arch/x86_64/mmt4d_x86_64_avx512_bf16.c). Just mentioning why I don't have strong opinions here as a downstream.

Indeed, our position too. 

We're now looking into generating simple kernels directly and why we want to resurrect this dialect. The plan is to have micro-kernel quality "special lowering" for particular patterns but still use micro-kernels for the more complex stuff, and slowly cover the space.

Once we have some minimal prototype we'll have lessons learned on both dialect and vector level transform, which we'll start working with you guys to upstream. We're hoping this will directly benefit IREE too.

https://github.com/llvm/llvm-project/pull/124800