Hey, After more thoughts, I'm wondering whether or not we should enable benchmarks for range and strides by default. These are not informative anyway because they depend quite a lot on the offsets used for the ranges, and the strides used for the slices. While, on the other hand, for big vectors, or big square matrices, the size shouldn't matter. So in the end, I'm just thinking about the short following output, that would deliver our peak performance.
[BLAS]: sCOPY : ...(GB/s sAXPY : ... (GB/s) sDOT : ... (GB/s) sGEMV : ...(GB/s) sGEMM-NN : ...(GFLOP/s) sGEMM-TN : ...(GFLOP/s) sGEMM-NT : ...(GFLOP/s) sGEMM-TT: ...(GFLOP/s) dCOPY : ...(GB/s dAXPY : ... (GB/s) dDOT : ... (GB/s) dGEMV : ...(GB/s) dGEMM-NN : ...(GFLOP/s) dGEMM-TN : ...(GFLOP/s) dGEMM-NT : ...(GFLOP/s) dGEMM-TT: ...(GFLOP/s) Philippe 2014-07-16 23:04 GMT+02:00 Karl Rupp <r...@iue.tuwien.ac.at>: > Hey, > > > I've noted that the console benchmarks for ViennaCL were quite outdated, > >> performance for AXPY are reported in FLOP/s, for example. I think it'd >> be great to have something compact, all incorporated in a single >> benchmarking executable: >> >> ======================= >> BLAS [float, full] >> --------------------- >> AXPY : ... (GB/s) >> DOT : ... (GB/s) >> GEMV : ...(GB/s) >> GEMM-NN : ...(GFLOP/s) >> GEMM-TN : ...(GFLOP/s) >> GEMM-NT : ...(GFLOP/s) >> GEMM-TT: ...(GFLOP/s) >> ... solver, perhaps some other things >> >> BLAS [float, ranges] >> --------------------- >> ... >> ======================= >> >> I can't really think of a case where one would be only interested in the >> performance of one single operation ! >> > > > I fully support these suggestions. In addition to running all the > benchmarks in one executable, it should also take a couple of command line > parameters such as the OpenCL device ID and matrix sizes. > > > > Do you have any other idea to make the benchmarks more > > concise/readable/informative ? > > One more thing we should set up for our own developer purposes is > something similar to the benchbot used by the FENICS guys: > http://fenicsproject.org/benchbot/ > This would help us in automatically tracking performance and in finding > performance regressions quickly. At the same time, it would provide a handy > skeleton for plots for use in presentations etc. Toby certainly knows what > I'm talking about ;-) > > Our summer students here in Vienna are very motivated, if they keep up > their pace they will have some time left by the end of the summer to > implement such a benchbot. Yes, we could do this ourselves within a couple > of days, but let's focus on getting 1.6.0 done. > > Best regards, > Karli > > >
------------------------------------------------------------------------------
_______________________________________________ ViennaCL-devel mailing list ViennaCL-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/viennacl-devel
