topic Thanks Ruqiu. I have tested in Intel® oneAPI Math Kernel Library

AVX512 is slower than AVX2 when running CGESDD/SGESDD on Xeon Gold 6130

Ding__Jian — Fri, 21 Feb 2020 23:41:39 GMT

I am evaluating the performance of Intel MKL on Xeon Gold 6130 processors, which have two AVX512 FMA units. I see performance improvement with AVX512 for matrix multiplication and FFT. However, for matrix inversion, the performance of AVX512 is worse than AVX2. I tested complex float (CGESDD) and float (SGESDD).

My question is: what is the reason that cause the slowdown of AVX512 for CGESDD/SGESDD? Is it because these functions are not optimized for AVX512 or something I did wrong?

Below is the output when MKL_VERBOSE is enabled

MKL_VERBOSE Intel(R) MKL 2020.0 Product build 20191122 for Intel(R) 64 architecture Intel(R) Advanced Vector Extensions 512 (Intel(R) AVX-512) enabled processors, Lnx 2.10GHz lp64 sequential

I set MKL_ENABLE_INSTRUCTIONS to be AVX2 or AVX512 to compare their performance and set the library to be sequential.

-----------------------------------------------------------------

For SGESDD/CGESDD, AVX2 outperforms AVX512 in most cases

64x64 matrix:

SGESDD: AVX2: 536.91us AVX512: 703.39us
CGESDD: AVX2: 766.52us AVX512: 861.09us

1000x1000 matrix:

SGESDD: AVX2: 305.60ms AVX512: 360.65ms
CGESDD: AVX2: 744.38ms AVX512: 696.96ms (AVX512 is slightly better)

-----------------------------------------------------------------

For SGEMM/CGEMM, AVX512 outperforms AVX2

64x64 matrix:

SGEMM: AVX2: 8.58us AVX512: 7.08us
CGEMM: AVX2: 43.55us AVX512: 23.06us

1000x1000 matrix:

SGEMM: AVX2: 27.98ms AVX512: 18.40ms
CGEMM: AVX2: 109.17ms AVX512: 69.49ms

-----------------------------------------------------------------

Hello Ding, Jian,

Ruqiu_C_Intel — Tue, 25 Feb 2020 08:24:27 GMT

Hello Ding, Jian,

Thank you for raising the topic! We will investigate the problem and back to here once there is any update.

One quick question is based on your test, the performance issue is exist in MKL 2020.0 or other version also has the same problem?

Best Regards，

Ruqiu

Thanks Ruqiu. I have tested

Ding__Jian — Tue, 25 Feb 2020 08:37:43 GMT

Thanks Ruqiu. I have tested MKL 2019.0 as well and it has the same problem.

Best,

Jian

Jian, yes, we see the same

Gennady_F_Intel — Thu, 05 Mar 2020 05:00:39 GMT

Jian, yes, we see the same behavior with version 2020 and it looks like the code branch for this particular function is not well optimized. In the case, if this regression is important to you, I recommend submitting the problem via the official support channel - intel online service center against MKL product.

Jian,

Gennady_F_Intel — Sat, 04 Apr 2020 06:07:22 GMT

Jian,

The time spent in xGESDD highly depends on the distribution of singular values. Could you recheck the results by using exactly the same input matrix for calling ?GESDD on AVX2 and AVX512?

Jian, have you tried to check

Gennady_F_Intel — Wed, 22 Apr 2020 04:54:18 GMT

Jian, have you tried to check the problem with the same inputs?