[Ffmpeg-devel-irc] ffmpeg-devel.log.20190101

[burek]

[00:05:36 CET] <thardin> happy GNU+year!
[11:59:11 CET] <cone-113> ffmpeg 03Paul B Mahol 07master:f7745edeaaee: doc/filters: improve afir description a little
[12:30:08 CET] <cone-113> ffmpeg 03Paul B Mahol 07master:e1f7881d926f: afilter/af_sofalizer: check explicitly other type
[13:35:15 CET] <cone-113> ffmpeg 03Paul B Mahol 07master:35a8179149c1: avfilter/af_adelay: allow setting delays in seconds
[18:12:06 CET] <j-b> HappyNewYear!
[19:13:03 CET] <durandal_1707> i always forgot what to call when counting number of instructions
[19:15:06 CET] <durandal_1707> so what code count cycles?
[19:16:50 CET] <atomnuker> libavutil/timer.h
[19:17:12 CET] <atomnuker> START_TIMER <code> STOP_TIMER("name");
[19:27:36 CET] <cone-597> ffmpeg 03James Almer 07master:3209d7b3930b: configure: bump year
[19:33:13 CET] <kurosu> and you can't intermingle several
[19:52:19 CET] <durandal_1707> i get slightly more decycles but overall speed is about same, and perf reports smaller usage of time in fcmul_add
[19:53:00 CET] <durandal_1707> that is probably avx overhead
[19:53:22 CET] <kierank> durandal_1707: turn cpu govenor to perofmrance
[19:53:30 CET] <kierank> also set affinity to one cpu
[19:53:47 CET] <durandal_1707> how?
[19:54:52 CET] <kierank> google
[20:08:30 CET] <durandal_1707> yes, avx is slower somehow on this cpu
[20:10:42 CET] <atomnuker> compared to sse?
[20:12:02 CET] <durandal_1707> yes,   23,20%  ffmpeg   libavfilter.so.7.46.101   [.] ff_fcmul_add_avx.loop VS   23,09%  ffmpeg   libavfilter.so.7.46.101   [.] ff_fcmul_add_sse3.loop
[20:16:28 CET] <atomnuker> meh, seems like the difference is down in the noise
[20:16:56 CET] <atomnuker> though avx should definitely be faster than sse due to register sizes
[20:21:21 CET] <durandal_1707> https://pastebin.com/pgQJMz6K
[20:32:54 CET] <atomnuker> can you post timer cycles instead of relative perf percentages?
[20:33:51 CET] <atomnuker> also be sure the amount of data per call doesn't change or you'll be introducing noise
[20:36:18 CET] <kierank> 19:16:53 <"atomnuker> though avx should definitely be faster than sse due to register sizes
[20:36:21 CET] <kierank> not if it's memory bound
[20:38:21 CET] <atomnuker> should still be faster, movs take the same amount of time for either register size
[20:38:35 CET] <atomnuker> and page size >> register size
[20:47:28 CET] <jamrial> durandal_1707: write a checkasm module for it. much better than timer.h benchmarks, as those include function call overhead
[20:48:15 CET] <jamrial> should be trivial, just look at other float based modules, like float_dsp and synth_filter
[20:48:24 CET] <durandal_1707>  703500 decicycles in fcmul_add,  261908 runs,    236 skips   VS  687032 decicycles in fcmul_add,  261985 runs,    159 skips
[21:09:52 CET] <durandal_1707> maybe asm can be overwritten to get better performance
[21:16:50 CET] <iive> is the sse3 and avx assembly the same, with just different register sizes?
[21:18:27 CET] <nevcairiel> I don't even see avx assembly in the source
[21:18:29 CET] <nevcairiel> did you  make it?
[21:22:17 CET] <kurosu> from the name of the function, avx downclock may also be a thing
[21:22:18 CET] <iive> durandal_1707, btw, on intel avx 256 bit processing is not enabled right away, the module boot up time is quite significant.
[21:22:40 CET] <cone-597> ffmpeg 03Marton Balint 07release/4.1:ddc284300eb0: avfilter/af_asetnsamples: fix last frame props
[21:22:41 CET] <iive> you may either boot it up during cpuid or use banchmark that runs for longer than second.
[21:22:47 CET] <kurosu> also, what cpu? amd are almost all 128b AU, so no benefit of AVX vs xmm
[21:23:08 CET] <kurosu> (except code size, sometimes)
[21:23:29 CET] <iive> isn't xmm like mostly bigger?
[21:23:42 CET] <kurosu> not always, iirc
[21:23:42 CET] <iive> (as code size)
[21:23:48 CET] <kurosu> different encoding of insn
[21:24:01 CET] <iive> i thought it uses more prefixes.
[21:27:34 CET] <kurosu> I don't think so, the opposite actually? evx prefix makes the insn bigger
[21:41:12 CET] <durandal_1707> nevcairiel: what avx instructions i could use?
[21:58:10 CET] <durandal_1707> are thy going to remove FMA support in CPUs?
[21:58:48 CET] <jamrial> no
[00:00:00 CET] --- Wed Jan  2 2019