imageproc
imageproc copied to clipboard
image-rs
Interest for a parallel version of match_template?
Hello,
For a personal project, I wrote a parallel version of match_template, keeping essentially the same algorithm but using a Rayon parallel iterator for the y coordinate and collecting the result.
If there is interest here, I'm happy to do a PR with my changes. It would be good to know whether this should be an addition or a replacement, and if some tests are needed.
kind regards
I'll do a benchmark and report the results tomorrow, but I would say, close to the number of cores.
On an AMD Ryzen 9 7950X 16-Core Processor (running Ubuntu 22.04.2 LTS), looking for a 259 x 415 template in a 750 x 750 image took 10.4s with the current algorithm, and 974ms with my new version, so a speed-up of more than a factor of 10. This is lower than the core count, but not unexpected due to the thermal limitations of the CPU preventing all cores running at the same speed as a single core.
Funnily, running the parallel version but restricted to one cpu with cpulimit takes 9.98s, so less that the non-parallel version. This is reproducible. I do not know whether this is due to the usage of iterators in the parallel version or somehow an artifact of cpulimit.
In any case, I would say that the speed-up is worth considering integration, do you agree? If so, do you want the new function as a replacement or as an additional function? I can then prepare a PR.
Okay, sounds good.
Funnily, running the parallel version but restricted to one cpu with cpulimit takes 9.98s, so less that the non-parallel version
I see
In any case, I would say that the speed-up is worth considering integration
For sure
do you want the new function as a replacement or as an additional function?
Additional is what i would prefer..
Thanks a lot for writing this parallel version btw...i was looking for it and found this issue in just the right time. I hope your PR gets merged asap, otherwise i would like to ask you to share the code with me..let's see, thanks again!
The PR is ready: #527.
I went for simplicity, if we are all happy about this new feature, it might be possible to de-duplicate some code between the original and the parallel version, or even, only use the parallel version by just changing the loop type depending on the availability of rayon.
