OpenCL benchmarking ....
Hi all, and happy editing Not VERY useful for CinelerraGG yet, still I found whole topic attractive enough :} http://www.bealto.com/gpu-benchmarks_mem.html I compiled benchmarks for my machine, with two simple fixes, and uploaded result to github: https://github.com/Randrianasulu/MPBenchmarks IMO my current devices quite fast: ===== Host to device copy 223 424 771 1182 1770 2147 2403 2378 2506 2501 1490 1828 756 743 777 771 776 708 683 0 Device to host copy 325 613 1155 1971 3320 4811 6282 7188 8039 8316 7039 5471 2817 2695 3107 3256 3221 2634 2716 0 copyN 327 651 1272 2189 3300 4486 5419 5932 6193 6266 6299 6320 6330 6345 6348 6349 6346 6335 0 0 For fastest NVA3 gpu (reclocked manually) GPU<->GPU copy still slow ..... https://gitlab.freedesktop.org/mesa/mesa/-/issues/3479 This is with very hacked mix of two experimental mesa branches ... Master (mesa3d) doesn't support compute on nv50-era cards. This bench requires images(2D) support, and works with OpenCL 1.1 driver
Andrew,
Not VERY useful for CinelerraGG yet, still I found whole topic attractive enough :}
Interesting benchmarks. What I found a little puzzling was in the first graph where the Read and Write throughput is the same as I always thought a Read would be faster.
В сообщении от Saturday 24 October 2020 03:51:42 Phyllis Smith via Cin написал(а):
Andrew,
Not VERY useful for CinelerraGG yet, still I found whole topic attractive enough :}
Interesting benchmarks. What I found a little puzzling was in the first graph where the Read and Write throughput is the same as I always thought a Read would be faster.
Yeah, on hdd/ usb flash/ssd and related memory devices it usually like this - read is fast, but writing is less fast .. But this specific benchmark sometimes works with smaller sizes, so they fit into cache and thus work much faster than usual.. I also was quite sure read and write speeds even on RAM should be a bit different, but I guess I need to read more on this ... In the meantime I found another becnhmark (apparently also requring OpenCL 1.1 + images), it doesn't work correctly for me on hacked nv92, but might be of some interest: https://github.com/CRVI/OpenCLIPP "OpenCLIPP - OpenCL Integrated Performance Primitives" - mostly from 2014 and last activity was in 2017 .... I found it via https://www.iwocl.org/resources/opencl-libraries-and-toolkits/ As far as I understand scanline-based image processing doesn't fit very well to opencCL, but may be some plugins will benefit from it?
participants (2)
-
Andrew Randrianasulu -
Phyllis Smith