[00:12] <ubitux> saste: mmh i don't remember [00:13] <ubitux> i think i didn't submit because i had some output issues [00:13] <ubitux> i'll have a look in a moment [00:14] <ubitux> we have almost the same code [00:15] <ubitux> i guess it's ok [00:15] <ubitux> huh that was my code maybe [00:16] <ubitux> let me re-bench it locally [00:18] <cone-941> ffmpeg.git 03Stefano Sabatini 07master:fbcc584d3abb: lavfi/unsharp: use named options, and add missing checks on matrix size values [00:18] <cone-941> ffmpeg.git 03Stefano Sabatini 07master:37b42bcb7db9: doc/filters: itemize unsharp examples [00:18] <cone-941> ffmpeg.git 03Stefano Sabatini 07master:2042cd37699f: lavfi/unsharp: extend range for amount values, to reflect the behavior of mp=unsharp [00:18] <cone-941> ffmpeg.git 03Clément BSsch 07master:51ba843f195c: lavfi/kerndeint: use aligned linesizes for the temporary buffer. [00:18] <cone-941> ffmpeg.git 03Stefano Sabatini 07master:f84ef19f3f8b: lavfi/kerndeint: move is_rgb_packed definition to the internal context [00:18] <saste> ubitux, uh I already pushed [00:19] <ubitux> haha it simplifies my test then [00:19] <ubitux> don't remove mp=kerndeint now [00:19] <ubitux> give me a few minutes :P [00:19] <saste> anyway, i think that the benchmarking differences may be due to the fact that i don't enable optims [00:19] <saste> no rush, take your time [00:19] <ubitux> what do you mean with the optim? [00:19] <saste> you can use my benchmarking patch on top of git master [00:20] <saste> --disable-optimizations [00:20] <ubitux> yes i have the same patch locally [00:20] <ubitux> yes but why would that affect the benchmarks diff? [00:21] <saste> yes, indeed it shouldn't [00:26] <ubitux> 68908034 decicycles in kerndeint, 2048 runs, 0 skips [00:26] <ubitux> 68417747 decicycles in mp=kerndeint, 2048 runs, 0 skips [00:27] <ubitux> 68329763 decicycles in mp=kerndeint, 2048 runs, 0 skips [00:28] <ubitux> 69982854 decicycles in kerndeint, 2048 runs, 0 skips [00:28] <ubitux> i'm doing ./ffmpeg -f lavfi -i testsrc=s=720x480 -vf tinterlace,kerndeint -f null - [00:28] <ubitux> (btw, surprise with s=720x479, with both filters) [00:29] <ubitux> saste ^ [00:30] <ubitux> anyway, i don't think the performance matters, but i'm actually very curious about the reason of that difference [00:31] <ubitux> i wasn't able to get it last time [00:34] <cone-941> ffmpeg.git 03James Darnley 07master:724061888be1: lavfi/yadif: remove unnecessary cast [00:34] <cone-941> ffmpeg.git 03James Darnley 07master:87d359ff8f55: lavfi/yadif: support 9, 12, and 14 bits per pixel [00:40] <saste> ubitux: you discovered a bug in testsrc [00:42] <ubitux> oh it's in testsrc? okay [00:42] <ubitux> i didn't dig [00:44] <saste> ubitux, I'm testing with 720x479, and still mp=kerndeint is slower [00:49] <ubitux> so with the command i gave, you get faster results? [00:49] <ubitux> maybe some magic somewhere [00:50] <ubitux> let's not bother too much about it then [00:51] <saste> 166831123 decicycles in mp=kerndeint, 1024 runs, 0 skips [00:51] <saste> 166170803 decicycles in kerndeint, 1024 runs, 0 skips [00:51] <saste> the difference is minimal, but yet it is consistently faster [00:52] <saste> i'll drop mp=kerndeint tomorrow, unless we have more benchmarks which prove the contrary [01:02] <saste> ping on "lavfi/unsharp: add check on matrix x/y size values oddity" [01:02] <saste> so i can finally drop mp=unsharp [03:13] <cone-941> ffmpeg.git 03Marton Balint 07master:b853cfe7eaf1: ffplay: remove refresh thread to achieve more precise frame delay [03:13] <cone-941> ffmpeg.git 03Marton Balint 07master:5534d8f75e27: ffplay: remove remains of video_clock [03:13] <cone-941> ffmpeg.git 03Marton Balint 07master:77ab702a848a: ffplay: only display frames if there are no pending events [03:13] <cone-941> ffmpeg.git 03Michael Niedermayer 07master:b9c544891b05: Merge remote-tracking branch 'cus/stable' [08:55] <cone-323> ffmpeg.git 03Anton Khirnov 07master:ccd70d9c1689: vf_yadif: factorize initializing the filtering callbacks [08:55] <cone-323> ffmpeg.git 03Michael Niedermayer 07master:fab8a89abeae: Merge commit 'ccd70d9c1689990e5aef2de383199bbc7cf60d13' [09:14] <cone-323> ffmpeg.git 03Anton Khirnov 07master:64ed397635ef: vf_yadif: fix out-of line reads [09:14] <cone-323> ffmpeg.git 03Michael Niedermayer 07master:446f7c62a2d1: Merge commit '64ed397635ef2666b0ca0c8d8c60a8bc44581d82' [11:09] <cone-323> ffmpeg.git 03Michael Niedermayer 07master:1e78679768b3: mpegvideo_enc: draw edges on input [12:47] <durandal_1707> michaelni: yadif have unused absrefs variable, probably included with merge [12:49] <durandal_1707> s/included with/left from/ [12:56] <cone-323> ffmpeg.git 03Paul B Mahol 07master:af882e1819ca: lavfi: port MP noise filter [13:05] <durandal_1707> michaelni: why ff_* is global in swr? [13:15] <michaelni> mplayer and our libmpcodecs used some of these i think [13:21] <cone-323> ffmpeg.git 03Anton Khirnov 07master:d2a25c4032ce: get_buffer(): do not initialize the data. [13:21] <cone-323> ffmpeg.git 03Michael Niedermayer 07master:a53b144ec041: avcodec: add avpriv_color_frame() [13:21] <cone-323> ffmpeg.git 03Michael Niedermayer 07master:1c5b8b916d3d: h264: color frames gray [13:21] <cone-323> ffmpeg.git 03Michael Niedermayer 07master:bb29ee62e9ef: Merge commit 'd2a25c4032ce6ceabb0f51b5c1e6ca865395a793' [13:21] <michaelni> durandal_1707, rand()/srand() is not good in a lib it can randomly affect applications that expect a repeatable sequence from rand() [13:24] <michaelni> in vf_noise that is [13:31] <durandal_1707> michaelni: so you wait until i commit this to comment? [13:32] <durandal_1707> also it was already in mp [13:33] <durandal_1707> it can be replaced by some fibonaci stuff saste mentioned [13:41] <michaelni> yes i wait until iam aware of a problem before mentioning, ill try to mention it next time before iam aware of it but iam not sure how to do that yet [13:41] <michaelni> mp is no real lib so it can do things like that [13:42] <durandal_1707> michaelni: i don't follow [13:43] <wm4> uh doesn't lavu have RNGs [13:43] <michaelni> it does [13:43] <michaelni> durandal_1707, i wasnt aware of the problem before noise was commited .... [13:45] <durandal_1707> michaelni: i see srand is used in lavu [13:46] <michaelni> its under #ifdef TEST [13:46] <durandal_1707> michaelni: so what should be used? av_lfg? [13:49] <durandal_1707> michaelni: ok to remove mp=noise it is slower that native one(for me, the SIMD stuff is not enabled at compilation, but this can be added later to native one) [13:50] <michaelni> av_lfg is fine, other generators would be fine too [13:50] <Compn> 26 cycles faster. :) [13:52] <Compn> BBB-work : dxva2 is windows hwaccel , chrome could take advantage of gpu speed up ... [13:55] <durandal_1707> really? [14:11] <cone-323> ffmpeg.git 03Michael Niedermayer 07master:b4bb996d6f41: yadif: remove unused variable [14:11] <cone-323> ffmpeg.git 03Michael Niedermayer 07master:0a707da37b41: cinepack: print an error if cinepak_decode() failed [14:20] <cone-323> ffmpeg.git 03Anton Khirnov 07master:68f930d2188a: h264: simplify calls to ff_er_add_slice(). [14:20] <cone-323> ffmpeg.git 03Michael Niedermayer 07master:131b9c924d6d: Merge commit '68f930d2188aba5b32624887dcbf688c23482834' [14:34] <cone-323> ffmpeg.git 03Anton Khirnov 07master:2491f9ee29ae: h264: remove a pointless if() [14:34] <cone-323> ffmpeg.git 03Anton Khirnov 07master:d9ebb00dcbaa: svq3: remove a pointless if() [14:34] <cone-323> ffmpeg.git 03Michael Niedermayer 07master:8b7568cc3482: Merge commit 'd9ebb00dcbaac3812b8b1fbc3d6e027506c11cbc' [15:58] <durandal_1707> what is max allowed value for color components in YUV 16 bit? [16:08] <cone-323> ffmpeg.git 03Stefano Sabatini 07master:2bac1535db97: Changelog: add missing empty line after the version line [16:24] <saste_> mp=unsharp is faster than unsharp [16:42] <durandal_1707> saste_: really? how much? [16:42] <saste_> durandal_1707, it's due to indirect variable access [16:43] <saste_> fp->x vs. x [16:43] <saste_> since this is in a loop it seems it is measurable [16:44] <saste> the other difference is due to a call to av_clip() [16:44] <saste> which is only relevant with no optims [16:44] <durandal_1707> look if it is different by 100 cycle I do not care [16:44] <saste> so all in all i'm going to push these changes and drop mp=unsharp [16:45] <durandal_1707> what compiler you are using? [16:45] <saste> 68524757 decicycles in unsharp, 4096 runs, 0 skips [16:45] <saste> 68856107 decicycles in mp=unsharp, 4096 runs, 0 skips [16:45] <saste> so with the "optims" unsharp is a bit faster now [16:46] <saste> durandal_1707, gcc [16:46] <durandal_1707> what version? [16:47] <saste> gcc (Ubuntu/Linaro 4.6.3-1ubuntu5) 4.6.3 [16:49] <durandal_1707> saste: it is expected for mp to be slow as it does some extra stuff... [16:49] <durandal_1707> also i think it is still using slices.... [16:50] <saste> durandal_1707, no, i'm just benchmarking the processing code [16:53] <durandal_1707> saste: also lavfi have problems when using mpegts as input in VV->V filter(nut works fine) [16:54] <durandal_1707> i guess it is related to probing [16:54] <saste> which kind of problems? [16:55] <saste> related to framerate or something? [16:55] <saste> there is a lavfi open ticket for that [16:56] <durandal_1707> [buffer @ 0x2940c3a0] Unable to parse option value "0x0" as image size [16:56] <durandal_1707> interestingly it works if input is 2nd [16:57] <durandal_1707> it have 2 video streams, but only one is actually ther [16:57] <durandal_1707> some mpegts bug [17:10] <mateo`> what is the consequences of setting AVFrame.interlaced_frame to 1 for a picture ? does it mean that the picture only contain odd or even fields ? [17:11] <durandal_1707> it means frame is interlaced and not progressive, nothing more - nothing less [17:12] <nevcairiel> an avframe typically always contains both fields weaved together, so yeah, it just indicates that this is the fact [17:13] <mateo`> thx [17:15] <durandal_1707> why would frame contain only odd/even fields? [17:16] <mateo`> it was a dumb question, i just wanted to be sure [17:17] <mateo`> by definition a frame contains both fields [17:18] <Daemon404> nevcairiel, from other channel: i feel like im taking crazy pills [17:22] <Shiz> Daemon404: sounds like guys are having fun [18:59] <cone-323> ffmpeg.git 03Anton Khirnov 07master:54974c62982a: error_resilience: decouple ER from MpegEncContext [18:59] <cone-323> ffmpeg.git 03Michael Niedermayer 07master:1fad547cefb4: Merge commit '54974c62982ae827becdbdb9b620b7ba75d079a0' [19:10] <cone-323> ffmpeg.git 03Anton Khirnov 07master:1d0feb5d1ac0: mpegvideo: split ff_draw_horiz_band(). [19:10] <cone-323> ffmpeg.git 03Michael Niedermayer 07master:ddbbfd95bee4: Merge commit '1d0feb5d1ac04d187b335f0e8d411c9f40b3a885' [22:36] <ubitux> mega accepts bitcoins @_@ [23:33] <cone-323> ffmpeg.git 03Anton Khirnov 07master:2c541554076c: h264: deMpegEncContextize [23:33] <cone-323> ffmpeg.git 03Michael Niedermayer 07master:b7fe35c9e50e: Merge remote-tracking branch 'qatar/master' [00:00] --- Sun Feb 17 2013
participants (1)
-
burek