I am doing a postdoc research work related to the quality and performance assessments of various demosaicing algorithms. I'm interesting in evaluating the performance of the IPP ippiCFAToRGB() and ippiDemosaicAHD() demosaicing functionality, and would therefore like to know if and how these two IPP functions exploit the thread-level parallelism. I would like to perform scalability tests with various numbers of threads (1, 2, 4, 8) and obtain how much speedup can be achieved on a given platform.
First of all, I checked to see if both functions are listed in the ThreadedFunctionsList.txt. I found ippiCFAToRGB() is only there, which probably means that there is not threading support for ippiDemosaicAHD(). Is it so?
Is there a way to configure the number of threads used by Intel IPP and how? I found that the IPP threading control can be established using ippSetNumThreads(n) and ippGetNumThreads(). For me it is unclear if these two functions are still working, because the multithreading version of the new library releases is deprecated. My test results shows that there is no effect when trying to disable internal parallelization with ippSetNumThreads(1). Also there is not any improvement, if the number of thread is given to be 2, 4 and etc.
Is it possible somehow to be able to compile the project without optimizations (internal threading turned off) and then with specific CPU optimizations and internal threading turned on?
I have both single-threaded and multi-threaded libraries installed and I tried to compile my test using each one of them. I didn’t see any improvement in the execution times obtained with these versions. Furthermore, when I compiled the project using the multi-threaded library, ippiCFAToRGB() doesn’t produce well reconstructed image. The image is reconstructed by half (the above part contains the successfully reconstructed image and the bottom is black). I suppose that there is a bug. With the single-treaded library everything is OK. With ippiDemosaicAHD() I haven’t these problems. The raw image file reconstructs properly.
I am using ippIP AVX (e9), 2017.0.1 (r53196) since Oct 4 2016 running on Intel core i5-2400 under Win64.
I hope to receive your assistance. Thanks in advance.
Sincerely yours, Iva
- Parallel Computing