OpenMathLib / OpenBLAS

OpenBLAS is an optimized BLAS library based on GotoBLAS2 1.13 BSD version.
http://www.openblas.net
BSD 3-Clause "New" or "Revised" License
6.38k stars 1.5k forks source link

ELAPS experiment and benchmarking #693

Closed bhack closed 4 years ago

bhack commented 8 years ago

What do you think of elaps?

martin-frbg commented 8 years ago

Seems to be a framework for performance testing of BLAS/LAPACK implementations, could you provide some background or own opinion of what advantages it would offer over the benchmarks used so far?

bhack commented 8 years ago

http://arxiv.org/abs/1504.08035

martin-frbg commented 8 years ago

Thank you. This does look like it would be a valuable tool in the right hands (though not mine probably). As an aside I assume that xianyi will be quite familiar with the libflame code that was used in their comparison demo run.

bhack commented 8 years ago

/cc @elmar-peise

elmar-peise commented 8 years ago

ELAPS is indeed a framework or experimentation with dense linear algebra kernels with BLAS/LAPACK like interfaces. Regarding OpenBLAS, I imagine ELAPS might come in handy when interactively testing and optimizing the kernels and various blocking factors. On the other hand, OpenBLAS's existing benchmarks serve their purpose well and introducing ELAPS into the project only to replace these static benchmarks would be overkill.

Do you have a specific problem or task in mind concerning ELAPS?

bhack commented 8 years ago

Probably It can have a role in kernel optimization and testing. @xianyi What do you think?

xianyi commented 8 years ago

@bhack , ELAPS is a good tool on desktop. However, I usually develop OpenBLAS on a server by ssh. Therefore, I prefer cmd line tools on server. I uses OpenBLAS/benchmark and https://github.com/xianyi/BLAS-Tester

bhack commented 8 years ago

I don't think that it is embedded with the GUI. I read that:

Experiments are designed either through Python scripts or a specialized GUI, and run on the whole spectrum of architectures, ranging from laptops to clusters, accelerators, and supercomputers.