[Ffmpeg-devel-irc] ffmpeg-devel.log.20180430

burek burek021 at gmail.com
Tue May 1 03:05:03 EEST 2018


[00:05:22 CEST] <Chloe> durandal_1707: what you writing asm for?
[00:10:59 CEST] <durandal_1707> nlmeans filter function
[02:33:17 CEST] <atomnuker> oh look its another soname bump for libx265
[11:42:45 CEST] <Chloe> should I be using av_mallocz_array over just av_mallocz(size*count)?
[11:43:03 CEST] <Chloe> err, im being silly. I should just use calloc
[11:46:04 CEST] <Chloe> so why does av_mallocz_array exist
[11:48:26 CEST] <durandal_1707> Chloe: Libav NIH syndrome
[11:54:35 CEST] <iive> Chloe, i think all memory allocated by ffmpeg routines should be memory aligned to largest SIMD word
[11:54:59 CEST] <Chloe> iive: but why av_mallocz_array over av_calloc?
[11:55:35 CEST] <iive> that i have no idea.
[12:04:16 CEST] <Squeak> hello
[12:04:58 CEST] <Chloe> Squeak: guten Tag
[12:05:17 CEST] <Squeak> can tell me someone how can i use unicode version of function "avformat_open_input" ?
[12:05:46 CEST] <Squeak> exist any unicode version ?
[12:06:11 CEST] <Chloe> does it not just take utf8?
[12:06:30 CEST] <iive> do you mean windows ucs2 ?
[12:06:38 CEST] <Squeak> yes
[12:06:52 CEST] <iive> aka utf16.
[12:06:58 CEST] <Squeak> yes
[12:07:24 CEST] <Chloe> Squeak: I dont think ffmpeg has any utf16 functions
[12:07:41 CEST] <Squeak> :-)
[12:09:46 CEST] <SqueakPL> can i use "_wopen" to open FILE, ans then use FILE structure ?
[12:10:27 CEST] <nevcairiel> you can use whatever you want, but not with ffmpeg
[12:14:04 CEST] <iive> does windows support utf8 for the CreateFileA() ?
[12:14:21 CEST] <iive> or that would depend on locale settings?
[12:14:43 CEST] <nevcairiel> unicode with the A functions doesnt work, no
[12:14:52 CEST] <nevcairiel> the A functions use the system locale
[12:15:31 CEST] <nevcairiel> but ffmpeg automatically translates string to wide-character and uses the W functions if you use its file open API
[12:16:06 CEST] <Chloe> nevcairiel: so ffmpeg only takes in native or does it take utf8?
[12:17:08 CEST] <nevcairiel> ffmpeg understands utf8, it converts it to wchar for the Windows API, so you dont h ave to worry about it
[12:18:43 CEST] <Chloe> I guess that makes it annoying if you're working with wchar on windows since you have to convert to utf8 only for ffmpeg to convert it back >.<
[12:19:34 CEST] <nevcairiel> you're not supposed to care what ffmpeg does with it internally
[12:22:10 CEST] <Chloe> good point
[13:56:20 CEST] <jdarnley> Blame Windows and everyone else (like Lua) who thinks it is okay to have a NUL byte in a string!
[14:01:06 CEST] <nevcairiel> its a wchar, its not bytes, its words
[14:01:11 CEST] <nevcairiel> no null words in there
[14:28:40 CEST] <durandal_1707> how to store only first 4 bytes from xmm register into memory?
[14:35:57 CEST] <durandal_1707> pextrd?
[14:37:11 CEST] <nevcairiel> movd
[14:46:28 CEST] <BBB> movd
[14:46:32 CEST] <BBB> pextrd is very slow
[17:50:08 CEST] <kurosu> durandal_1707, even if you are doing integer maths, and it is said switching between float and int ops now has a penalty, I'd benchmark using movss too
[18:09:45 CEST] <atomnuker> there's really no penalty for e.g. using xorps on simd registers which were just used for integer ops
[18:09:58 CEST] <atomnuker> on intel CPUs since core 2 I think
[18:15:59 CEST] <jamrial> if you're only going to handle ints, then the only float instructions that make sense using if needed are shufps and some of the complex movs that have no int version
[18:17:05 CEST] <atomnuker> yeah, mainly the movs are useful
[19:18:32 CEST] <Gramner> shufpd, movmskps, movmskpd as well
[20:41:44 CEST] <cone-189> ffmpeg 03Michael Niedermayer 07master:34dbdcfc20d6: avformat/m4vdec: Fix detection of raw MPEG-4 ES Studio
[20:41:45 CEST] <cone-189> ffmpeg 03Michael Niedermayer 07master:0f176bb8e089: avformat/m4vdec: Use the same constant names as libavcodec
[20:41:46 CEST] <cone-189> ffmpeg 03Michael Niedermayer 07master:a47bd1cd1c71: avcodec/mpeg4video_parser: Fix incorrect spliting of MPEG-4 studio frames
[20:41:47 CEST] <cone-189> ffmpeg 03Michael Niedermayer 07master:c0aa89eeee2a: avcodec/mpeg4video_parser: Avoid litteral 0x1B6, use named constant instead
[20:46:53 CEST] <cone-189> ffmpeg 03Paul B Mahol 07master:da55304cafea: avfilter/vf_mix: initialize last to silence possible warning
[21:28:17 CEST] <durandal_1707> Gramner: you solution give extra boost
[21:29:31 CEST] <Gramner> durandal_1707: nice. also what's up with all the function parameters that aren't used?
[21:34:41 CEST] <Gramner> durandal_1707: oh, and in scalar code you can do movzx regd, byte [mem] instead of zeroing and doing a byte load
[21:36:23 CEST] <durandal_1707> Gramner: those are for cases for subsamples planes, when one averages alphas
[21:39:41 CEST] <durandal_1707> i thinking how to write yuv422 case for chroma planes
[21:39:52 CEST] <durandal_1707> and C version looks wrong
[22:33:27 CEST] <cone-189> ffmpeg 03Marton Balint 07master:c60a824ee87a: avformat/qtpalette: parse color table according to the QuickTime file format specs
[22:33:28 CEST] <cone-189> ffmpeg 03Marton Balint 07master:e894d958fce6: avcodec/anm: fix palette alpha
[22:33:29 CEST] <cone-189> ffmpeg 03Marton Balint 07master:4c501bafc08c: avcodec/hnm4video: fix palette alpha
[22:33:31 CEST] <cone-189> ffmpeg 03Marton Balint 07master:56b081da578f: avutil/pixdesc: add AV_PIX_FMT_FLAG_ALPHA to AV_PIX_FMT_PAL8
[22:33:32 CEST] <cone-189> ffmpeg 03Marton Balint 07master:7033654f7f3c: Use AV_PIX_FMT_FLAG_ALPHA for detecting transparency where nb_components was used
[23:04:27 CEST] <durandal_1707> Gramner: instruction to take every 2nd byte and other places to be zeroes? 01 00 02 00 03 ...
[23:05:45 CEST] <durandal_1707> use movu and pand ?
[23:06:05 CEST] <Gramner> pand to mask away unwanted bits, yes
[23:06:34 CEST] <Gramner> if you only care about avx then you can combine the movu and pand
[23:06:50 CEST] <Gramner> since avx memory args can be unaligned
[23:07:17 CEST] <Gramner> e.g. pand dst, const_mask, [mem]
[23:07:44 CEST] <iive> keep the mask in register and let the pand use memory argument?
[23:08:11 CEST] <Gramner> yes
[23:09:20 CEST] <Gramner> assuming the pand is done multiple times. if it's only done once then there's no point to having the mask in a register
[23:11:01 CEST] <jamrial> this is where avx512 k registers would come into play i guess?
[23:11:26 CEST] <Gramner> that's one use case of them, sure
[23:12:09 CEST] <Gramner> you can conditionally mask away elements or blend the result with the previous value of dst
[23:12:12 CEST] <Gramner> for free
[23:12:29 CEST] <Gramner> on basically every instruction
[23:13:18 CEST] <Gramner> or well, setting up the kmask requires some instructions I guess, but that can be done once and reused a million times in a loop
[23:16:19 CEST] <cone-189> ffmpeg 03Marton Balint 07master:75d1529c6e90: avdevice/decklink_dec: do not copy video data
[23:16:20 CEST] <cone-189> ffmpeg 03Marton Balint 07master:649087fa83a5: avdevice/decklink_dec: unref packets on avpacket_queue_put error
[23:27:56 CEST] <cone-189> ffmpeg 03James Almer 07master:e3866ea20d55: fftools/ffmpeg: fix mixed code and declarations
[00:00:00 CEST] --- Tue May  1 2018


More information about the Ffmpeg-devel-irc mailing list