[FFmpeg-devel] [PATCH] This patch addresses Trac ticket #5570. The optimized functions are in file libswscale/ppc/input_vsx.c. Each optimized function name is a concatenation of the corresponding name in libswscale/input.c with suffix _vsx. From bountysource https://www.bountysource.com/issues/34315232-power8-vsx-vectorization-libswscale-input-c (reverse_forever)

[Lauri Kasanen]

On Mon, 1 Apr 2019 09:07:48 +0300
slava <pestov.vyach at yandex.ru> wrote:

> Sorry for title. It is my first experience in git send-email. Can I make
> a benchmark with handwritten tests or have some standard tool in ffmeg?
> And will the benchmark on x86-64 be informative?

We have standard bench macros, START_TIMER and STOP_TIMER. Put those
around the function's callsite, then do some ffmpeg run that calls that
specific function. Then add "-cpuflags 0" to the call to get the C
results, and from the numbers you can calculate the speedup.

Both the C and VSX runs should be done on the POWER machine. A Qemu VM,
emulating POWER instructions on x86-64, would probably be useless for
benchmark purposes. There are free POWER VMs available for testing from
a few places.

- Lauri