Provide 5-arg mul!

[haampie]

Since Julia 1.3 we have a 5-arg mul! for BLAS, but somehow it has not made its way into CuArrays.

In principle this should not be a breaking change, however, I've commented out LinearAlgebra.lmul! because it is not provided by Julia base as far as I know. So that is breaking, but I don't think anybody uses it (?).

Finally it fixes an oversight in LinearAlgebra.mul!(C::CuMatrix{T}, adjA::Adjoint{<:Any, <:CuMatrix{T}}, adjB::Adjoint{<:Any, CuMatrix{T}}) where no <: was used 😅

[haampie]

Hm, there's ambigious methods when taking a and b as ::Number. Simple fix is to replace that type by the eltype T, but that's not such a nice api.

Edit, decided to go for ::T instead of ::Number. If we really want to be able to do mul!(y, A, x, 1, 1)` or something, then a whole lot of new methods have to be added to avoid ambiguity.

[maleadt]

Thanks! Next release is going to be breaking so go ahead and remove those methods.
bors try

[haampie]

Should there be a deprecation warning for lmul!?

[maleadt]

Should there be a deprecation warning for lmul!?

Seems weird to provide a deprecation for a method in another module? We're just implementing the interface here; if upstream has deprecated/removed something, shouldn't we just follow suit?

[haampie]

Ok, so I've worked my way through tons of ambiguity warnings and added tests for all combinations of eltypes, transposes and adjoints for mul!. Ready for review!

[maleadt]

bors try

[bors]

try

Build succeeded

• ci/gitlab/trying

[maleadt]

LGTM, coverage is perfect too. Would it make sense to guard the definitions for 1.3.0 in a if VERSION block, to make sure we can safely remove them once we bump the Julia requirement?

bord d+

[haampie]

Ok, done :)

[maleadt]

bors r+

[bors]

Build succeeded

• ci/gitlab/staging