vesion 6.0.1 slower than 5.1

[delalaym]

Hello,

We were using the 5.1 version on our application. In that library, as far as I know, there was no optimization for the Core 2 Duo.
We are now trying the 6.0.1 version and it is slower then the 5.1 even though he now detects that I have a Core 2 Duo... How is that possible? The function that we use that takes the most CPU is : IppsMulc_32s_sfs
I've done also the performance test with ps_ipps.exe on both 5.1 and 6.0.1 and effectively, it sometimes takes longer on 6.0.1...
Any Idea?

Thanks.

[Vladimir_Dudnik]

How did you test the performance? Please note it may depend on the data you use. For example supplying NANs you will definetely drop the performance down.

Vladimir

[Ying_S_Intel]

Quoting - delalaym

Hello,

We were using the 5.1 version on our application. In that library, as far as I know, there was no optimization for the Core 2 Duo.
We are now trying the 6.0.1 version and it is slower then the 5.1 even though he now detects that I have a Core 2 Duo... How is that possible? The function that we use that takes the most CPU is : IppsMulc_32s_sfs
I've done also the performance test with ps_ipps.exe on both 5.1 and 6.0.1 and effectively, it sometimes takes longer on 6.0.1...
Any Idea?

Thanks.

Can you verify that what Intel IPP optimized libraries are loaded respectively from versions 5.1 and 6.0.1 on a Core 2 Duo? A simple way is to run ps_ipps.exe from /tools/perfsys from 5.1 and 6.0.1 folder and tell us what the outputs from first several lines, the data could help us understand the IPP optimized code used in these 2 versions.

Thanks,
Ying

[delalaym]

Quoting - YING S (Intel)

Can you verify that what Intel IPP optimized libraries are loaded respectively from versions 5.1 and 6.0.1 on a Core 2 Duo? A simple way is to run ps_ipps.exe from /tools/perfsys from 5.1 and 6.0.1 folder and tell us what the outputs from first several lines, the data could help us understand the IPP optimized code used in these 2 versions.

Thanks,
Ying

Hi,

I have attached a file. It contains a run for the ippsMulC_32s_Sfs function on both 5.1 and 6.0.1. It seems like the 6.0.1 is slower...

thanks.

[pvonkaenel]

Quoting - delalaym

Hi,

I have attached a file. It contains a run for the ippsMulC_32s_Sfs function on both 5.1 and 6.0.1. It seems like the 6.0.1 is slower...

thanks.

One possible difference could be if you're statically linking and using the threading layer in 6.0.1 (was not available in 5.1). Are you using the same linking/threading model in both tests?

Peter

[delalaym]

Quoting - pvonkaenel

One possible difference could be if you're statically linking and using the threading layer in 6.0.1 (was not available in 5.1). Are you using the same linking/threading model in both tests?

Peter

Yes, I am using the same thing...

Marc-Andre

[pvonkaenel]

Quoting - delalaym

Yes, I am using the same thing...

Marc-Andre

Since you have pretty good documentation about the slowdown, and it looks like you might even have a timing application that demopnstrates it, you might want to create an Issue in Premier support. I found a similar slowdown in the template matching migrating from IPP 4.1.3 to 5.? for a certain CPU. I submitted the information and they were able to fix the problem.

Peter