Allow StaticArray eltype in matmat{vec,mul}

[lcw]

Here we avoid promotion since it is not defined between scalar numbers
and static array types.

Fixes #1953

Co-authored-by: Mason McCallum masonamccallum@gmail.com

[maleadt]

I'll let @dkarrasch chime in, who wrote this code.

[dkarrasch]

Sorry for the breakage. Indeed, that promotion did not exist prior to my PR. I took it from LinearAlgebra. I believe it is required there because BLAS needs the constant scalars to be of the same type as eltype? I wonder what happens when you pass an integer 1 as alpha with float eltypes. Do gemv! and gemmEx! and gemm! accept that? Or is there some internal type promotion happening?

[lcw]

We just tried

julia> mul!(C, A, B, 3, 0)
3×1 CuArray{SVector{2, Float32}, 2, CUDA.Mem.DeviceBuffer}:
 [2.441199, 1.6931901]
 [2.1284773, 3.0013726]
 [4.8450527, 4.7269773]

julia> A = CuArray(rand(SVector{2, Float32}, 3, 3))
3×3 CuArray{SVector{2, Float32}, 2, CUDA.Mem.DeviceBuffer}:
 [0.327501, 0.991028]  [0.328234, 0.0684084]  [0.426921, 0.892227]
 [0.069393, 0.921621]  [0.74181, 0.508899]    [0.318923, 0.0873129]
 [0.152261, 0.157193]  [0.139696, 0.0202893]  [0.67665, 0.207444]

julia> B = CuArray(rand(Float32, 3, 1))
3×1 CuArray{Float32, 2, CUDA.Mem.DeviceBuffer}:
 0.7732126
 0.28225046
 0.848472

julia> C = similar(A, 3, 1)
3×1 CuArray{SVector{2, Float32}, 2, CUDA.Mem.DeviceBuffer}:
 [0.0, 0.0]
 [0.0, 0.0]
 [0.0, 0.0]

julia> mul!(C, A, B, 3, 0)
3×1 CuArray{SVector{2, Float32}, 2, CUDA.Mem.DeviceBuffer}:
 [2.124308, 4.62784]
 [1.6008871, 2.790985]
 [2.193833, 0.909842]

julia> all(C .≈ 3*A*B)
true

and it seemed to work. I don't know where the promotion is happening.