"Ronald S. Bultje" <[email protected]> writes: > Hi, > > On Sun, May 13, 2012 at 10:56 PM, Luca Barbato <[email protected]> wrote: >> On 13/05/12 22:19, Vitor Sessak wrote: >>> On 05/13/2012 11:40 PM, Diego Elio Pettenò wrote: >>>> Il 12/05/2012 01:15, Vitor Sessak ha scritto: >>>>> >>>>> Here is a version with 128-bit regs (again, completely untested). >>>> >>>> The patch seems to require a new series that factors "x86: use more >>>> standard construct for setting ASM funcitons in FFT code" into it. >>>> Probably you should resend the series. >>>> >>>> If you do I can test it as I have a xop-capable box at hand. >>> >>> Nice! Here is the full patchset. >> >> Fate test passed =) >> >> Do we have a documented way to benchmark it? > > You can decode a file that uses FFT (I suppose anything like mp3 or > aac will work), or run fft-test with START/STOP_TIMER around the
fft-test -s will run a speed test. > function of interest. To switch between optimizations on the same > host, use -cpuflags avx, -cpuflags xop, -cpuflags sse or -cpuflags 0 > (C) as arguments to avconv. I believe someone was working on adding > support for that argument to fft-test also but I don't know if that > was finished. It was, the flag is -c with the same arguments as avconv -cpuflags. -- Måns Rullgård [email protected] _______________________________________________ libav-devel mailing list [email protected] https://lists.libav.org/mailman/listinfo/libav-devel
