A question about benchmark in mlirX

It’s a question mostly to Mr. Bondhugula.
And question is about benchmark test mlirx\mlir\benchmark\gemm-tiled-benchmark.mlir.
Please, why did you use a couple of operations
%3 = mulf %0, %1 : f64
%4 = addf %3, %2 : f64
instead of using single fmaf?
Is there any difference between these two cases?

fmaf didn’t exist at that time - it was only added in Feb 2021. Using fmaf does have an impact on precision — typically the accuracy/float-point stability increase with fmas, but there are also rare instances where it goes the other way. It should be good to use here.

Thanks. I thought it may affect the distribution of fma instructions between two CPU SIMD units and you used the couple of mpy+add instead of fma intentionally to give the compiler more freedom to schedule data flows.