Dr. Thomas Orgis

This is somewhat related. We should not forget that the headers for netlib CBLAS are not only provided by netlib … NumPy always uses its own header: https://github.com/numpy/numpy/blob/main/numpy/core/src/common/npy_cblas.h And in...

And how dangerous is it, actually, to link numpy with reference cblas?

Sure. It's just that OpenBLAS or MKL are what people use in practice and both seem to have settled on ``` #define CBLAS_INDEX size_t /* this may vary between platforms...

Thanks for the pointer. So the relevant part seems to be B.2.2 in that spec, which says that `BLAS_INDEX` usually is `size_t`, but might also be chosen to be identical...

@martin-frbg Good to know. Can you point out a code path for, say, x86-64 that shows how the size_t is passed around to the actual computation for cblas_isamax()? I found...

Ah, good. So the one case which is actually problematic is depending projects using a copy of cblas.h that does not fit the library. I don't find actual usage of...

Any progress on this? I am trying to get a grip on https://github.com/DrTimothyAldenDavis/SuiteSparse/issues/1 to decide if it is worth it packaging Octave with sparse matrix computations via SuiteSparse and OpenBLAS...

So you actually need some formular or table to use 1, 2, 4, … $many cores depending on data size? Lots of heuristics have to go into that:-/ Right now...

My issue right now is that I'd like to see an actual little test that checks how badly we are actually affected by such. I am no user of cholmod...

> My initial tests with the cholmod "demo" files suggest that current MKL (plus Intel compiler optimizations, Intel libc and whatnot) is up to ten percent faster than current, gcc-compiled...