NNlib batched_mul!

[mcabbott]

This hooks up FluxML/NNlib.jl#100 for CuArrays.

I've added an extremely simple test, but could find the right place to insert this into test/blas.jl perhaps if that would be better?

Tests fail locally, before this PR, but I'm not sure why.

I think that _GemmFloat == CUBLAS.CublasFloat, except that this isn't loaded yet. Is there a neater way?

[maleadt]

I've added an extremely simple test, but could find the right place to insert this into test/blas.jl perhaps if that would be better?

It's an nnlib specific interface, so that place seems fine.

Tests fail locally, before this PR, but I'm not sure why.

Please open an issue for that then. Let's see what CI does here:
bors try

I think that _GemmFloat == CUBLAS.CublasFloat, except that this isn't loaded yet. Is there a neater way?

Move the include of nnlib.jl until after the submodules (including CUBLAS)?

[bors]

try

Build succeeded

• ci/gitlab/trying

[mcabbott]

Thanks for having a look, I've moved the include("nnlib.jl") as suggested.

About tests, is there a list of what needs to be on master? I failed again, but can try more seriously once back in front of my desktop, probably in 2 weeks.

[maleadt]

About tests, is there a list of what needs to be on master? I failed again, but can try more seriously once back in front of my desktop, probably in 2 weeks.

Generally CuArrays, CUDAnative, CUDAdrv and GPUArrays. CuArrays also ships a Manifest with a known-good combination, so you can just clone CuArrays and run with julia --project in that dir.

[maleadt]

bors r+

[bors]

Build succeeded

• ci/gitlab/staging