topic I tested it on AVX512 setup - in Intel® oneAPI Math Kernel Library

cgemm3m, cgemm_compact AND cgemm give poor results for small problem 24*64

Yosef__Elad — Mon, 08 Jul 2019 09:33:58 GMT

Hi all,

I using sequential API and direct call to multilply matrices.

C = 1*conj(A')*A

A is 64*24 and C is 24*24 both are complex matrix (complex8).

I have arrays of matrices: A_ARR (filled with random values) and C_ARR (filled with zeros) both array have 1000 matrices.

My application is pinned to sinlge core and to corresponding RAM by NUMA id.

build cmd: icc -c -g -ipo -ipp -Ofast -DMKL_DIRECT_SEQ -xCORE-AVX2 *.c

Setup is Xeon E5-2699A v4, 64G ram on each numa

I run cblas_cgemm/cblas_cgemm3m/mkl_cgemm_compact in a loop over A_ARR and C_ARR (each time only 1 function) and I get really poor results (I'm measuring only the matrices multiplication time)

I'm aware to the MKL "warn-up" issue and running cblas_cgemm in advance with measuring it time

cblas_cgemm(CblasRowMajor, CblasConjTrans, CblasNoTrans, m, n, k, &alpha, &A_ARR, m, &A_ARR, n, &beta, &C_ARR, m)

Gives- AVG 6.5ms MAX 8.6ms MIN 6.3ms

cblas_cgemm3m(CblasRowMajor, CblasConjTrans, CblasNoTrans, m, n, k, &alpha, &A_ARR[, m, &A_ARR[, n, &beta, &C_ARR[, m)

Gives- AVG 7.5ms MAX 12ms MIN 7.3ms

mkl_cgemm_compact(CblasRowMajor, CblasConjTrans, CblasNoTrans, m, n, k, &alpha, &a_arr_compact[, m, &a_arr_compact, n, &beta, &c_arr_compact[, m, COMPACT_FORMAT, 1)

Gives- AVG 225ms MAX 231ms MIN 224ms

Note COMPACT_FORMAT is from mkl_get_format_compact();

Does any one can assist me with reducing with time it takes?

It is also not clear to me why the compact API that should mostly vectorize matrices multiplication it getting lowest results

Thanks

Elad

We need to check but probably

Gennady_F_Intel — Tue, 09 Jul 2019 07:21:26 GMT

We need to check but probably compact API has not been optimized for such "big" sizes. What version of MKL do you use?

MKL version is latest 2019.4

Yosef__Elad — Tue, 09 Jul 2019 08:47:43 GMT

MKL version is latest 2019.4.243

Another odd thing is that cblas_cgemm show better results than cblas_cgemm3m.

the latest should imporve by ~25% according to docs

Gennady_F_Intel — Tue, 09 Jul 2019 09:10:09 GMT

Could you share your benchmark to check these numbers on our side with the latest updates and CPU?

attached whole project

Yosef__Elad — Tue, 09 Jul 2019 09:53:02 GMT

attached whole project

make AVX / AVX2 / AVX512 =y

thanks for the project, we

Gennady_F_Intel — Wed, 10 Jul 2019 05:45:28 GMT

thanks for the project, we will check

I tested it on AVX512 setup -

Yosef__Elad — Wed, 10 Jul 2019 06:47:06 GMT

I tested it on AVX512 setup - Xeon Platinum 8176 2.10Gz

I can't see any improvments that comes from the AVX512.

Should I expect for any improvement against AVX2 on the above setup?

Can't find any info in release notes.

Elad

Closing this thread I fond

Yosef__Elad — Wed, 10 Jul 2019 13:36:35 GMT

Closing this thread I fond the issue in my timer function