[01:13] <neg> koike: I dusted of my old hack patches and got it to work (patches on ML), I'm sure my patches can be improved as I'm not that familiar with vimc but it's a start. I'm not sure who tonyk is so I could not CC him, maybe you could let him know about the patches?
[02:32] <ndufresne> Kwiboo, in your branch, the hevc-ctrls.h don't match the cedrus 5.1 kernel
[05:13] <Kwiboo> ndufresne, timestamp 0 is not a problem, I just wanted to avoid using it to make it more easy to spot potential source of issue while debugging
[05:17] <Kwiboo> it was not that the code expected timestamp not to change, it was that the code to get a buffers reference timestamp relied on the timestamp being set/predicted for first frame as it contained a reference to itself
[05:20] <Kwiboo> but due to a code error on my part it was always 0 for all buffers until the first dqbuf for the buffer, resulted in a missmatch in timestamp_ts in "dpb" and the buffers actual timestamp
[05:23] <Kwiboo> ndufresne, the hevc-ctrls.h differs because jernej implemented missing features in cedrus hevc, see https://github.com/LibreELEC/LibreELEC.tv/blob/master/projects/Allwinner/patches/linux/0008-HEVC-improvements.patch for the additions
[05:23] <jernej> scaling list should be ok, but another one is a bit hackish
[07:36] <bbrezillon> Kwiboo: great!
[07:48] <bbrezillon> Kwiboo: not the the timestamp fix actually fixes a bug since the driver points to the current cap frame when it doesn't find a valid entry https://github.com/bbrezillon/linux/blob/rk-vpu-h264/drivers/staging/media/rockchip/vpu/rk3288_vpu_hw_h264_dec.c#L777
[07:48] <bbrezillon> AFAIU that's what you want
[07:50] <bbrezillon> note that capture buffers will anyway don't have a valid timestamp until they are used to store decoded content
[07:52] <bbrezillon> that's because vb2_find_timestamp() checks the ->copied_timestamp flag which is only set when v4l2_m2m_buf_copy_metadata() is called, and we call this when we're done decoding a slice/frame
[07:55] <bbrezillon> s/not the the timestamp fix actually fixes/note that the timestamp fix does not actually fix/
[08:17] <bbrezillon> Kwiboo: nevermind, I think I got what you fixed (B frame forward ref)
[08:25] <Kwiboo> bbrezillon, regarding the v4l2_m2m_buf_copy_metadata() call, it should probably be called before the decoding run is started (this is what is done in cedrus), that way reference timestamp can be picked up by vb2_find_timestamp() for self referenceing frames
[08:41] <bbrezillon> Kwiboo: yes, I was wondering if it was correctly placed when working on H264
[08:41] <bbrezillon> I'll move it
[18:49] <ndufresne> Kwiboo, thanks for the info, I actually never new if a frame could reference itself
[18:49] <ndufresne> I feel I am always doing miss-match each time I builde linux-cedurs / ffmpeg
[18:50] <ndufresne> things are moving, I guess that's normal
[18:51] <ndufresne> (don't worry about my clone ndufresne_kiwi, just giving that client a try)
[23:11] <ndufresne> paulk-leonov, I just tried to enabling dmabuf to KMS/DRM, but only got green frames