Open svenstaro opened 9 years ago
The benchmark code used for the blog can be seen here: https://github.com/arrayfire/valhalla. I am trying to make it a full blown benchmarking suite for all the libraries used in that blog post.
@kylelutz If you are thinking about doing this let me know if you need any help from my end.
Any update on this? And any tips for Boost.Compute how to make it faster?
@VincentSC One of the things I have noticed where ArrayFire is faster is because of the memory manager. If you can modify your local boost compute to implement even a rudimentary memory manager it can improve the performance quite a bit at smaller sizes.
Yeah, I'm hesitant to add more benchmark applications directly to Boost.Compute. I more like the valhalla
approach to providing a single, separate benchmark suite which compares a variety of GPU libraries. This seems preferable to each individual library maintaining their own benchmark suite for every other library.
And a memory manager would be good to have, especially for algorithms that could re-use the same "scratch" memory buffers over and over.
In that case maybe you should link to an external suite of benchmarks instead of providing your own and then work with the maintainer of that benchmark suite?
I think it would be interesting to see ArrayFire benchmarks especially because it claims to be (mostly) faster than boost.compute in their benchmarks. These could run as part of your other perf tests and give a better overall picture where boost.compute fits in performance-wise.