anatoly:OpenCV vs LibJacket

阅读更多

关于LibJacket与OpenCV的对比，我看到网上有来自LibJacket团队mcclanahoochie大神的：OpenCV vs. LibJacket: GPU Sobel Filtering,以及来自anatoly大神的 OpenCV vs LibJacket，其中anatoly的来自blogspot，这里转贴一下，供大家参考讨论。

 

 

Recently mcclanahoochie noticed that our OpenCV GPU module shows not the best performance of Sobel filtering and DTF-based convolution. He posted simple compassion here. I want to thank him for pointing this. I appreciate any help to make OpenCV GPU better.

To the moment we’ve updated our code (available in svn trunk r6881) and done our own comparison using the same benchmark code with only difference that we use buffered version of OpenCV’s functions in order not to allocate GPU memory inside each time (http://pastebin.com/W41RwPnu).

Update: we tested using Lib­jacket v1.0.1 (build dd66add) by Accel­erEyes

Let’s look at plots. DFT based performance is approximately the same for Fermi.



For pre-Fermi:



Filtering performance. Now for small kernels OpenCV behaves much better than before.



Oh… definitely not primary target platform for us. But much better than before.



I omitted above comparison for convolve() for kernel sizes 27 and 32. Just we’ve got some strange results here. Maybe we did something wrong with LibJacket. mcclanahoochie’s numbers are much lower, and we asked what he thinks about this in comments of his post.