Intel® Integrated Performance Primitives
Deliberate problems developing high-performance vision, signal, security, and storage applications.
6718 Discussions

Less optimization for functions in IPP 2021 when updated from IPP 5.3

BaffledCoder
New Contributor I
1,745 Views

I recently updated my code base from IPP 5.3 to IPP 2021, and optimization was not as expected.
I ran some performance tests on both applications, and the following are the results:

Function Percentage Decrease In Processing Time
DFTFwd_RToPack_32f_C1IR 18.4955%
FFTFwd_RToPack_32f_C1R 38.2696%

Maximum supported instruction set: AVX 2

I expected around 50% optimization as I updated from IPP 5.3, which does not even support AVX.
Are the above figures correct, or am I doing something wrong? 

Also, I built a custom dll from the intel IPP script, which only contains functions as declared in export.def file. Does it affect the result if I build a dll on a machine that only supports SSE?

Labels (1)
0 Kudos
13 Replies
ShanmukhS_Intel
Moderator
1,716 Views

Hi Shubham,

 

Thanks for posting on Intel Communities.

 

The deprecated functions in Intel IPP Legacy Libraries are moved outside of the Intel IPP product, and they were not optimized for future architectures (the latest optimization are targeted for Intel Advanced Vector Extensions 2 instruction set), and support for these legacy functions is also discontinued. Might be the custom dll. created was creating the impact. Could you please share with us a sample reproducer?

 

Best Regards,

Shanmukh.SS

 

0 Kudos
BaffledCoder
New Contributor I
1,673 Views

Hi @ShanmukhS_Intel 
Both of these functions are non-deprecated:

ippiDFTFwd_RToPack_32f_C1IR
ippiFFTFwd_RToPack_32f_C1R
and the above results are from comparing these functions from both IPP 5.3 and IPP 2021.6

0 Kudos
ShanmukhS_Intel
Moderator
1,635 Views

Hi Subham,

 

Could you please share with us a sample reproducer that you took as a reference in comparison, as it helps us in knowing the details better? 

Apologies for the inconvenience caused by the earlier reply.

 

Best Regards,

Shanmukh.SS

 

0 Kudos
ShanmukhS_Intel
Moderator
1,591 Views

Hi Subham Sharma,

 

A gentle reminder:

As we haven't heard back from you, Could you please get back to us with the details mentioned in the earlier thread, so that we could look it further?

 

Best Regards,

Shanmukh.SS

 

0 Kudos
BaffledCoder
New Contributor I
1,588 Views

Hi @ShanmukhS_Intel 
I don't think it is an issue. I wanted to confirm that I got the above-mentioned gain when shifting from IPP5.3 to IPP2021.6.
As IPP5.3 does not support AVX instruction, and there are not many resources for estimating the amount of gain when using the above-mentioned functions led me to ask this question as reassurance for properly doing everything.
Regards

Shubham Sharma

0 Kudos
ShanmukhS_Intel
Moderator
1,557 Views

Hi Subham,


Could you please get back to us with a sample reproducer that you took as a reference to calculate the processing time? It helps us in reproducing the issue at our environment and assist you further.


Best Regards,

Shanmukh.SS


0 Kudos
BaffledCoder
New Contributor I
1,549 Views

Hi,
I am sorry I can't, as this is not my priority, and I am already busy with other projects.
You can build 2 projects, one with IPP5.3 and the other with IPP2021.6, and get the same results.
Or you can create 2 DLLs from both IPP and use it in a performance calculator exe.
You can take an Idea from ippiMorphOpenBorder_8u_C1R is slower than older IPP version , In here, I have shared a reproducer.  
Regards,
Shubham Sharma

0 Kudos
ShanmukhS_Intel
Moderator
1,521 Views

Hi Subham,


We are working on your issue. We will get back to you if any details were needed.


Best Regards,

Shanmukh.SS


0 Kudos
ShanmukhS_Intel
Moderator
1,499 Views

Hi Subham Sharma,

 

We have tried compiling the shared reproducer using Visual Studio and oneAPI command prompt. However, we are facing issues wrt dll linking. Could you please let us know if any dependencies were missing?

Attached the screenshot for your reference.

 

Best Regards,

Shanmukh.SS

 

0 Kudos
ShanmukhS_Intel
Moderator
1,461 Views

Hi Subham Sharma,


Thanks for sharing the details. We are discussing your issue internally with our development team. We will get back to you soon with an update.


Best Regards,

Shanmukh.SS


0 Kudos
ShanmukhS_Intel
Moderator
1,391 Views

Hi Subham Sharma,


Thank you for your feedback. We have provided your feedback to the relevant team. At this moment there is no visibility when it will be implemented and available for use. Please let us know if we can go ahead and close this case?


Best Regards,

Shanmukh.SS


0 Kudos
BaffledCoder
New Contributor I
1,348 Views

Sure, close the ticket.

0 Kudos
ShanmukhS_Intel
Moderator
1,337 Views

Hi Subham Sharma,


Thanks for the confirmation. We are closing this case for now. If you need any other information, please post a new query as any further replies would be considered community only.


Best Regards,

Shanmukh.SS


0 Kudos
Reply