Reduce number of mul! methods

[dkarrasch]

Cf. JuliaGPU/GPUArrays.jl#472 and JuliaGPU/CUDA.jl#1904. Same here: if gemm_dispatch! is (sufficiently) internal, it might be actually removed.

[dkarrasch]

This requires the GPUArrays.jl PR to be merged and released first.

[maleadt]

Thanks for the PR! Note to self, we should also do this to Metal.jl (https://github.com/JuliaGPU/Metal.jl/blob/main/lib/mps/linalg.jl). I guess the change would be mostly similar.

This requires the GPUArrays.jl PR to be merged and released first.

We commit a Manifest here, so you can actually point it to your GPUArrays branch (or GPUArrays#master, once that PR would be merged).

[dkarrasch]

I couldn't find a generic_matvecmul!-like function that dispatches to gemv! and the symmetric/hermitian counterparts. Should there be one? It seems that currently mul!(::oneVector, ::oneMatrix, ::oneVector) falls back to GPUArrays.jl. Only if the matrix is Hermitian calls are caught and sent to symv! or hemv!.

[dkarrasch]

I wonder...

[maleadt]

Same difference in ArgumentError as with the CUDA.jl PR.

I couldn't find a generic_matvecmul!-like function that dispatches to gemv! and the symmetric/hermitian counterparts. Should there be one? It seems that currently mul!(::oneVector, ::oneMatrix, ::oneVector) falls back to GPUArrays.jl.

Yeah that looks like an oversight. oneMKL has gemv, so we should wrap that: https://spec.oneapi.io/versions/latest/elements/oneMKL/source/domains/blas/gemv.html

EDIT: actually, oneMKL.gemv and oneMKL.gemv! is available, it's just not wired op to LinearAlgebra.