Revisit BLAS tuning

Open fredrik-johansson opened this issue 2 years ago • 1 comments

The cutoffs for small matrices are not necessarily accurate, depending on the machine, especially when BLAS is using threads.

Apr 29 '23 10:04 fredrik-johansson

Directly related to this (not sure it's worth opening a new issue): this will also imply having some support for BLAS threads, at least for a few of the most popular BLAS libraries.

For example some profiling files just changed now (https://github.com/flintlib/flint/pull/2019) were assuming that if Flint uses BLAS then it can use openblas_set_num_threads. I'm not sure to which extent it is feasible e.g. to ask at configure time the name of a similar function to use for the provided choice of BLAS library? (I see e.g. BLIS has bli_thread_set_num_threads but I'm not sure all most popular BLAS libraries have this)

Jun 13 '24 16:06 vneiger