04:24 <Lynne> yay, I can encode raptorq now!

04:24 <Lynne> it was so simple

04:24 <Lynne> still need to write a proper finite-field matrix implementation, but hacking something up to test worked

04:27 <Lynne> the spec helpfully provides most of the annoying arithmetic needed for GF(256) like exp and log

10:34 <mkver> Are the old issues (from before the switch to track) available somewhere?

10:52 <elenril> I have some stuff from there

10:52 <elenril> something specific you need?

10:53 <Daemon404> TIL there was something before trac

10:54 <elenril> wow, it's even still being developed

10:54 <Daemon404> it == ?

10:55 <elenril> roundup

10:55 <Daemon404> never heard of ir

10:55 <Daemon404> it

10:56 <Daemon404> lol well it doesnt render properly at all on mobile

10:56 <Daemon404> guess that is not surprising

10:58 <elenril> mkver: seems I have files and messages as standalone files, but not their links to issues

11:12 <mkver> elenril: issue 1890

11:16 <__monty__> Am I reading the contributing docs correctly? (Specifically https://ffmpeg.org/developer.html#toc-Miscellaneous-conventions) So an explicit cast for this test wouldn't be welcome as a patch, https://hydra.nixos.org/build/263535062/nixlog/1/tail ?

11:26 <elenril> mkver: https://ups.khirnov.net/4d5d0cd558f2e0597256eb32d0c19e42f548018bdc802b121b5dfe7bc414ea7a/msg10191 ?

11:26 <mkver> elenril: Do you happen to have the sample?

11:27 <elenril> probably, but figuring out which one it is will be nontrivial

11:34 <elenril> mkver: it's probably https://ups.khirnov.net/b8ea9e58401cd0828544876a9ddb552db4b145ea0a29faf5c2c1f1a22a00aec3/file866

11:40 <cone-445> ffmpeg Michael Niedermayer master:c69e6cccd7e1: avformat/iamf_parse: consider nb_substreams when accessing substreams array

11:40 <cone-445> ffmpeg Michael Niedermayer master:2b2ced61eba0: avcodec/libvpxenc: Cleanup on error

11:40 <cone-445> ffmpeg Michael Niedermayer master:97ecfb5a193c: MAINTAINERS: Update the entries for the release maintainer for FFmpeg

11:40 <cone-445> ffmpeg Michael Niedermayer master:7fab9b97613e: avformat/iamf_parse: 0 layers are not allowed

11:40 <cone-445> ffmpeg Michael Niedermayer master:4cab028bd0e3: avformat/mxfdec: Check container_ul->desc before use

11:40 <cone-445> ffmpeg Michael Niedermayer master:0da6865ce292: avfilter/qsvvpp: Remove unreachable code

11:48 <elenril> __monty__: it might be ok in this case

12:40 <Lynne> this may be the first time anyone's gotten to use the gf2p8mult instruction for anything that isn't aes

13:13 <cone-445> ffmpeg James Almer master:5191339f7465: avformat/evc: remove useless struct field

13:13 <cone-445> ffmpeg James Almer master:4e608e90eb9a: avformat/evc: fix writing reserved bits

13:13 <cone-445> ffmpeg James Almer master:0ae157b3603f: avformat/iamf_parse: add missing padding to AAC extradata

13:44 <jdarnley> Would it be a bug for AV_FRAME_DATA_S12M_TIMECODE to not start with a count of the number of timecodes that follow?

23:05 <IndecisiveTurtle> Lynne: Some more progress today. Wanted to ask before doing more deeply if you have any preferences/or want to recommend how to implement the encoding aspect.I was thinking of following the logical steps of the encoder. Make a few image views for the different planes and perform haar wavelet in compute shader for each plane.As far as optimization goes I thought about using subgroup quads to minimize memory loads on

23:05 <IndecisiveTurtle> Then have an execution barrier and deinterleave in place in the image or to a host visible device local buffer (might be more convenient to read it back later)The second step is admittedly the hardest one for me to grasp, but after reading the spec a few times again I believe it's trying to find the best quantinization index so that each slicewill get encoded to a byte count between some min and max it has configure

23:06 <IndecisiveTurtle> At each step it counts how many bits the slice takes using the Exp-Golomb Codes (not 100% sure on that).This should probably go inside the compute shader as it looks expensive and relatively easily parallelizable.The encoding part looks simple enough, it iterates over each subband, of each slice of each plane quantizes it and encodes it.

23:06 <IndecisiveTurtle> Another question would be if we want all this process in a single compute dispatch or want to split it in parts.Sorry for the wall of text, wanted to inform you first in case you have something more specific in mind

23:08 <Lynne> keep it simple for now, one slice per invocation, write the main encode shader that takes some big slice size that won't be exceeded no matter what, transforms, quantizes and encodes

23:10 <Lynne> once you have that running, it will make it much easier to build on top, and add a shader before it to calculate the slice size for a quantizer (or a few quantizers)

23:11 <IndecisiveTurtle> Ok so I will start by implementing a single shader that does all 3 steps in a way similar to the C code.

23:13 <IndecisiveTurtle> Skipping slice calculation as you said though

23:20 <IndecisiveTurtle> Btw I saw that for getting the output buffer for the encoder there is ff_get_encode_buffer. So could either use memory import to make into a vulkan buffer and write encoded data directly into it, or use a device local host visible buffer and copy result (maybe faster). Which one do you think is better?

23:23 <Lynne> allocate a new device local buffer via ff_vk_create_buf(), map it, then just call av_buffer_create to create an AVBufferRef for it, using the FFVkBuffer * as the data ptr, and the avvkdevicecontext as the opaque pointer, and a small free function that just unmaps and frees the buffer

23:24 <Lynne> oh... I already wrote a function that does exactly that

23:24 <Lynne> ff_vk_create_avbuf()

23:25 <IndecisiveTurtle> Oh cool thanks

23:26 <Lynne> you don't even have to map the buffer yourself, if you use the VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT flag, the mapping and unmapping at the end is all done for you

23:28 <Lynne> oh, and, use ff_vk_get_pooled_buffer() instead

23:28 <Lynne> that gives you pooling for free

23:29 <Lynne> then just set avpkt->buf to the returned avbufferref, avpkt->size and avpkt->data to the size and mapped address and it should work

23:30 <IndecisiveTurtle> Will do

23:31 <IndecisiveTurtle> Another thing. For shader inclusion is there a way to use separate .comp files for them, or need to use the GLSL macros exposed by avfilter?

23:42 <Lynne> IndecisiveTurtle: you don't have to use the macros, copy the "VULKAN =" part at the end of libavfilter/Makefile into libavcodec's makefile, s/libavfilter/libavcodec, then you can place your .comp shaders in libavcodec/vulkan

23:44 <Lynne> then just add "extern const char *ff_source_<shader_name>_comp;" in the encoder and use it

23:46 <IndecisiveTurtle> Great, because those macros would make formatting a challenge :p