performance regressions in 2021.9

[Wobo]

In the release notes for 2021.9 Intel wrote:

2021.9

...

This release of IPP has been built with updated compilers.  As a result, we’ve noted some performance regressions in our testing.  If these regressions are thought to be impacting your application in a negative way or you have specific questions, please contact Intel® oneAPI Products Support.

...

So, my question is: Do these performance regressions still exist in version 2021.11 or have these problems fixed by Intel for 2021.11?

If these performance regressions do still exist, which functions are affected (or all)?

[Ruqiu_C_Intel]

Hi Wolfram,

Thank you for the posting.

The performance regression is improving in every new  version, target to fix all in coming major release.

Regards,

Ruqiu

[Ruqiu_C_Intel]

Hi Wolfram,

If you can help provide out the functions adopted in your project, it would help us prioritize the ​functions if they have performance impact.

Regards,

Ruqiu

[Wobo]

I can't write down all the used functions, these are some of the typically used ones:

ippsMinMax_64f
ippsAddC_64f
ippsSubC_64f
ippsMulC_64f
ippsDivC_64f
ippsMul_64f
ippsMul_32f
ippsSum_64f
ippsSqr_64f
ippsMinMax_64f
ippsMaxAbs_64f
ippsSampleDown_64f

ippsCopy_64f
ippsCopy_32fc
ippsSampleUp_16s

ippsAdd_16u_ISfs

ippsZero_16s

ippsMalloc_...
ippsFree

ippiDFTFwd_RToPack_32f_C1R
ippiDFTInv_CToC_32fc_C1R
FFT-Functions too
ippiFilterSobel_32f_C1R
ippiCopy_32f_C1R
ippiRemap_32f_C1R
ippiPackToCplxExtend_32f32fc_C1R

Very important for us would be superfast ippiRemap functions and we had the wish that these(all) functions can handle arrays with more than 2GiB data size, Intel calls it platform aware, afaik.

Kind Regards

[Wobo]

In my previous posts, I forgot to mention: we use/need Intel IPP especially on Windows platform...and hence, we are especially interested in IPP performance optimizations for the Intel/Windows platform.

Kind Regards

[Ruqiu_C_Intel]

Hi Wolfram,

Thank you for the information, we will check all the list functions, and put them into our performance improvement backlogs for Windows as well as platform aware feature enabling if there are gaps here.

Regards,

Ruqiu