performance about mlkernels.jl HOT 4 CLOSED

Thanks for pointing that out - I made the change.

I played around with syr2k and found that it tended to be a tiny bit slower than just iterating over all the entries in the upper triangle.

from mlkernels.jl.

I find the performance of MLKernels.jl very impressive and it is very flexible, great package!

syr2k has a multithreaded implementation and it also depends if you use a build for a certain architecture.

from mlkernels.jl.

Should this issue be closed? It seems that removing the redundant copying operation has been pushed to master. Or is the performance of using syr2k under question?

from mlkernels.jl.

For me the call to kernelmatrix! is ~0.3s faster than the gauss function in the above example and I think that the issue is resolved. I just gave the example to illustrate the speed increase if the extra copy was avoided (and because I stumbled over syrk2k which is not documented), now I think this would make a great example to show how julia can be faster than a specialized linear algebra library, while remaining super flexible.

from mlkernels.jl.