mkl_scscmm performance problem

Ramakrishnan_K_ — Fri, 01 Jul 2016 20:06:40 GMT

Hello,

I am building the attached program in OLCF's RHEA supercomputer (https://www.olcf.ornl.gov/computing-resources/rhea/) with Intel compiler icc (ICC) 14.0.4 20140805. Armadillo has a naive implementation of cscmm. I run the program with MKL_NUM_THREADS=1 in rhea to multiply the sparse matrix of size 83328x124992 with a dense matrix of size 124992x50. The following is the output.

./a.out 83328 124992 50 0.00001

The output of the test code

::A::83328x124992
nnz::104153
::B::124992x50
mkl cscmm::162.13
::C::50x83328
arma ::0.06

I am seeing MKL_CSCMM to be really slow over armadillo naive implementation. Kindly let me know what am I doing wrong here.

Ramki

may you check if this gap

Gennady_F_Intel — Sun, 03 Jul 2016 05:14:40 GMT

may you check if this gap exists with the latest mkl 11.3.update 3? you may take evaluation package of intel complier v.16. May you also check CSR ( mkl_dcsrmm ) case?

topic mkl_scscmm performance problem in Intel® oneAPI Math Kernel Library

mkl_scscmm performance problem

may you check if this gap