Easy-click link: #15286.

LGTM. Andreas?

(aside: BLIS https://github.com/flame/blis would allow arbitrary strides here)

@tkelman We allow arbitrary strides so I'm wondering how much speedup BLIS can get on matrices with special strides.

Yeah, it depends on how well optimized the non unit stride case is in BLIS relative to the julia generic gemm. Probably not commonly benchmarked but worth trying.

When strides get so big that there's only 1 element per cache line, I suspect the best performance might be achieved by copying (which essentially compacts the data).

I believe BLIS does copy for the non unit stride case in order to still use optimized simd operations, but only one panel at a time rather than the entire array. Should reread their papers and code though. On typical dgemm their haswell kernels are quite competitive with openblas and mkl.

Interesting. I'd be surprised if we couldn't someday match them in pure julia with @threads and vectorization, but we aren't there yet.

Yeah, the basic code generation patterns they do would all translate naturally into Julia style code generation (nicer, actually, since they're leaning heavily on the c preprocessor), and we could use their kernels as inline llvm or asm. Some day.