topic Re: IPP 2019 version 1 in Intel® Integrated Performance Primitives

IPP 2019 version 1

Mopu — Fri, 27 Nov 2020 13:41:35 GMT

I am using IPP lib 2019 version1.

Implemented one feature in C++ tested in two set of machines.

1) Intel Xeon Platinum 8160 processor took 1.5 secs to complete the flow.

2) Intel Xeon CPU E5-2699 v4 processor took 70 secs to complete the flow.

Is there any dependencies to use IPP library. If yes, Could you please anyone let me know.

Re: IPP 2019 version 1

Andrey_B_Intel — Fri, 27 Nov 2020 13:56:28 GMT

Hi Mopu.

Could you please provide more information about your flow?

IPP functions list
Output of ippGetLibVersion

May be some reproducer?

At least AVX512 instruction set is available in Intel Xeon Platinum 8160.

Thanks.

Re: IPP 2019 version 1

Mopu — Fri, 27 Nov 2020 14:05:40 GMT

Hi,

IPP function list: Add, Sub, Division, Multiplication, Transpose and Threshold. All operations are related to Floating point images.

IPP lib version:

#define IPP_VERSION_STR "2019.0.1"

CPU E5-2699:

Instruction Set Extensions Intel® AVX2

Platinum 8160 Processor:

Instruction Set Extensions Intel® SSE4.2, Intel® AVX, Intel® AVX2, Intel® AVX-512

Thanks

Re: IPP 2019 version 1

Andrey_B_Intel — Sun, 29 Nov 2020 21:16:32 GMT

Hi Mopu.

There is small benchmark in attachment with ippiAdd_32f_C1R function.

Could you please compile it and provide it's output at 2 your systems? And may be redesign it according with your workload.

Thanks,

Andrey.

Re: IPP 2019 version 1

Mopu — Wed, 02 Dec 2020 04:45:30 GMT

Hi Andrey,

I have tried to compile that piece of code. It's not working.

Then I have searched in Google. They were telling if we use namespace 'Ipp" lib it will restrict for single thread processing. if we use namespace "IppST" it will handle multithread processing.

Now, Faster machine didn't change the TPT improvement. But Slower machine improved a bit higher TPT and reduces (70 secs to 10 secs).

Re:IPP 2019 version 1

AbhishekD_Intel — Wed, 23 Dec 2020 06:51:06 GMT

Hi,

We are forwarding this issue to the SME.

Thank You