[kshyatt]

I think our original approach here was wrong, because we can use the BLAS methods as long as the view that's passed is densely packed. So right now we aren't using BLAS where we could (a dense patch of a higher dimensional tensor, for example) and we are trying to use it in dangerous places. Hopefully this fixes both.

[lkdvos]

Good catch!

[kshyatt]

I got a little carried away here. I think we can get rid of the vendor specific extensions and just use LinearAlgebra.mul! which should dispatch to the vendor BLAS if it's available or fall back to GPUArrays' implementation if not.

[Jutho]

Why is this packed columns (essentially: contiguous data) necessary. I thought gemm supports this via the "leading dimension" argument?

[kshyatt]

The vendor gemms only support matrices whose first dimension (so rows for operation == N, cols for operation == T or C) has stride 1, right? It's the second dimension that can have lda > m.