Recommendations on how to evaluate openCV with Intel's Integrated Performance Primitives?

Question

We have a fair amount of openCV code, and we plan on investigating Intel's IPP, and seeing what kind of speed ups we can achieve in our application.

Intel has a white paper on how to link OpenCV and IPP. Using Intel IPP with OpenCV It looks like the link may be automatic in some cases, but the amount of support is unclear.

We're a windows shop, and am interested in any war stories, or problems/issues to watch out for when performing this kind of evaluation.

Thanks

Answer 1

More and more less OpenCV has something to do with IPP. Later years IPP has difficulty to provide something great compared to results of GCC or any other decent compiler on Intel platforms. In old times IPP provided up to 50% of speedup but not anymore if we consider all GCC-s fancy features like TBB.

I read OpenCV lists for years and I cannot remind anyone speaking about IPP for long time. Search over my OpenCV mailbox also didn't expose anything. I suspect IPP support is something not so tested anymore and can even be very broken.

I have seen huge improvements by manual tuning OpenCV on non-Intel platforms. For example on IOS I got big boost replacing cos() and sin() by lookup tables. But IPP is for Intel only.

New trend seems to be GPU processing and there you can see 10 or 20 times speedup more likely. Engineers time is also expensive and instead of looking for hard work to get 20% of benefit, better invest into GPU if applicable.