As far as I can see recent versions of OpenBLAS up to 0.3.5 have issues with
a) AVX512 kernels, specifically DGEMM has issues. b) Newer compilers such as GCC 8 and in particular GCC 9 which have uncovered bugs in the inline assembler. I don't think that means the code is doomed but it does unfortunately mean that even though 0.3.6 doesn't seem to have those issues any more that OpenBLAS 0.3.6 is not optimal on AVX512 since it needs to fall back to the AVX2 implementation of DGEMM, which only runs at ~70% of the speed on Xeon Gold and up chips for large matrix multiplications. (it would theoretically be 50% but AVX512 runs at a lower clock speed, and on Silver and Bronze chips there is only one FMA unit for AVX512 which makes it slower than AVX2). On Thu, 9 May 2019 at 10:41, Jakob Schiøtz <schi...@fysik.dtu.dk> wrote: > > > > On 8 May 2019, at 11:20, Jakob Schiøtz <schi...@fysik.dtu.dk> wrote: > > > > > > > >> On 7 May 2019, at 18:12, Thomas Eylenbosch < > thomas.eylenbosch....@agro.basf-se.com> wrote: > >> > >> > >> Can someone try to reproduce the error with the R/Python(foss/2018b) > modules. Or can someone give me feedback on this? > > > > I have tried with Python/3.6.6-foss-2018b both on a Skylake (Intel(R) > Xeon(R) Gold 6148 CPU @ 2.40GHz) and a Broadwell (Intel(R) Xeon(R) CPU > E5-2650 v4 @ 2.20GHz). I cannot reproduce the problem, the test (check.py) > succeeds on both machines. > > Does anyone know if this issue is specific to some versions of Skylake > only, or if I have done something wrong since I cannot reproduce the > problem? > > Jakob > > > -- > Jakob Schiøtz, professor, Ph.D. > Department of Physics > Technical University of Denmark > DK-2800 Kongens Lyngby, Denmark > http://www.fysik.dtu.dk/~schiotz/ > > > > -- Dr. Bart E. Oldeman | bart.olde...@mcgill.ca | bart.olde...@calculquebec.ca Scientific Computing Analyst / Analyste en calcul scientifique McGill HPC Centre / Centre de Calcul Haute Performance de McGill | http://www.hpc.mcgill.ca Calcul Québec | http://www.calculquebec.ca Compute/Calcul Canada | http://www.computecanada.ca Tel/Tél: 514-396-8926 | Fax/Télécopieur: 514-396-8934
