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

cvHaarDetectObjects + IPP = No speed-up !

boris_mansencal
Beginner
394 Views
Hi,

I would like to know why I do not get better performances for cvHaarDetectObjects if Intel Integrated Performance Primitives are installed ?

According to the following Intel presentation, I should expect a 1.5-2X speed-up.
http://fsa.ia.ac.cn/files/OpenCV_China_2007June9.pdf

However, on my installation, if I run the facedetect example provided with OpenCV, I get roughly the same time of execution (on lena.jpg image) with and without IPP !
Besides, cvGetModuleInfo gives me :
libippcv.so, libippi.so, libipps.so, libippvm.so, libippcc.so, libmkl_p4.so
so it seems that IPP is actually used.

I am on linux (Fedora 8) with gcc 4.1.2, on an Intel Core 2 Duo 6600 processor.
I have installed Intel IPP 5.3.2.068 & MKL 10.0.1.014 (free non-commercial versions).

Do you see something wrong ?
Should I also use Intel Compiler ?

Thanks a lot,

Boris.

PS: compiling OpenCV with gcc4.3 & enabling OpenMP seems to bring a speed up. But I hoped I could have both...

0 Kudos
1 Reply
Vladimir_Dudnik
Employee
394 Views

Hello,

we've investigated that issue and found our that IPP function called in OpenCV has not enough optimization (one with Ipp32s data type). We would recommend to use IPP function which support Ipp32f data type instead. It is optimized in IPP 5.3 and provides about 50% speedup over OpenCV without IPP.

Please check OpenCV sorce forge project for the latest update or contact with Vadim Pisarevsky directly on that.

From our side we will work on providing additional optimization in IPP function wich work with Ipp32s data type (in general it provides more detection accuracy)

Regards,
Vladimir

0 Kudos
Reply