Further Optimizing 256x256 Tiling and Multithreading (OpenMP) for Large Image Processing with Intel

[klay1252]

Hello Intel IPP experts,

I am currently working on processing large images using Intel Integrated Performance Primitives (IPP) with a 256x256 tiling approach. This method has proven to be faster than processing the entire image at once. In addition to this, I have implemented multithreading using OpenMP to parallelize the processing of tiles.

While the performance is already improved, I am looking for ways to optimize it even further. Specifically, I would appreciate any advice or best practices on how to achieve better performance in the following areas:

1. Optimizing cache usage: Are there specific techniques or IPP functions that can help optimize cache utilization when processing multiple 256x256 tiles, especially in a multithreaded environment?
2. Reducing tile processing overhead: Are there advanced techniques to minimize overhead related to managing tile boundaries or transitions between tiles, even when using multithreading?
3. Optimizing OpenMP parallelization: Are there specific ways to fine-tune OpenMP usage or IPP functions to further enhance parallel performance when processing image tiles?
4. Additional IPP optimizations: Any other suggestions for maximizing the performance of large image processing, specifically when handling image tiles in combination with multithreading?

If there are any C++ code examples or references to further optimize this approach with IPP and OpenMP, that would be very helpful.

Thanks in advance for your insights and suggestions!

[Ruqiu_C_Intel · Accepted solution]

In general, IPP supports safety single thread​. Users take care of their multiple threads running in parallels, which means users take of OpenMP scheduling.

Several years ago, IPP introduced Integration Wrappers, which were designed to improve user experience with threading of Intel IPP functions and tiling. The Integration Wrappers document and examples are list in IPP install folder ./components/interfaces/iw/. Hopefully it helps you.