On jeudi 26 décembre 2019 04:02:35 CET Аl Воgnеr wrote:
> I wonder why an old pc with a GeForce 210, AMD Athlon II X2 270, 8G RAM
> is slower using opencl.
>
> darktable 2.6.3~git2.22c690a53
> Fresh installation of the pc, using default darktable-configuration
>
> $ darktable-cli bench.srw bench.srw.xmp bench.jpg --core -d perf -d
> opencl
>
> [opencl_init] device 0: GeForce 210
> GLOBAL_MEM_SIZE: 1024MB
> MAX_WORK_GROUP_SIZE: 512
> MAX_WORK_ITEM_DIMENSIONS: 3
> MAX_WORK_ITEM_SIZES: [ 512 512 64 ]
> DRIVER_VERSION: 340.107
> DEVICE_VERSION: OpenCL 1.0 CUDA
>
> 0.254144 [opencl_init] FINALLY: opencl is AVAILABLE on this system.
>
> 227,385925 [dev_process_export] pixel pipeline processing took 225,369
> secs (165,316 CPU)
>
> I do not find "blended on _G_PU" in the log. Log attached.
>
>
> $ darktable-cli bench.srw bench.srw.xmp bench.jpg --core
> --disable-opencl -d perf
>
> 76,490890 [dev_process_export] pixel pipeline processing took 74,577
> secs (144,522 CPU)
>
> BTW:
> GeForce GT 1030: 17,770 secs (26,050 CPU) / 37,168 secs (281,707 CPU),
> AMD FX-8320, 8G RAM
> So with a GT 1030 it is getting about 2 times faster using opencl.
>
https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units shows
that the Geforce 210 has very few processing units, and is slower than newer
cards (in single precision: ~40 GFlops for the 210, ~1000 Gflops for the 1030)
But, if you ran the tests only once and one after the other, the first will be
slower, as the image file has to be physically read from disk. Unless you fill
up the RAM completely, that file file will be cached in RAM for future use, so
the second run of darktable will be faster. And if there are background
processes that can run, it can give false results as well (it only through
increased task switching).
In other words, a single run for each setup is not reliable.
Note how the CPU times for the 210 don't differ all that much between the runs
with and without openCL, compared to the total times.
Remco
darktable user mailing list
to unsubscribe send a mail to darktable-user+unsubscr...@lists.darktable.org