Le mercredi 06 avril 2022 à 19:21 +0200, Jernej Škrabec a écrit : > Dne sreda, 06. april 2022 ob 08:54:07 CEST je Benjamin Gaignard napisal(a): > > Le 05/04/2022 à 20:40, Jernej Škrabec a écrit : > > > Hi Benjamin! > > > > > > Dne torek, 05. april 2022 ob 18:07:41 CEST je Benjamin Gaignard > napisal(a): > > > > Le 27/02/2022 à 15:49, Jernej Skrabec a écrit : > > > > > First two patches add 10-bit formats to UAPI, third extends filtering > > > > > mechanism, fourth fixes incorrect assumption, fifth moves register > > > > > configuration code to proper place, sixth and seventh enable 10-bit > > > > > VP9 decoding on Allwinner H6 and last increases core frequency on > > > > > Allwinner H6. > > > > > > > > > > I'm sending this as RFC to get some comments: > > > > > 1. format definitions - are fourcc's ok? are comments/descriptions ok? > > > > > 2. is extended filtering mechanism ok? > > > > > > > > > > I would also like if these patches are tested on some more HW. > > > > > Additionally, can someone test tiled P010? > > > > > > > > > > Please take a look. > > > > > > > > Hi Jernej, > > > > > > > > I have create a branch to test this series with VP9 and HEVC: > > > > https://gitlab.collabora.com/benjamin.gaignard/for-upstream/-/tree/10bit_ > > > > imx 8m Feel free to pick what I may need in it. > > > > > > > > That doesn't improve fluster scores, I think more dev are still needed in > > > > GST before getting something fully functional. > > > > Anyway I able to select P010 pixel format if the input is a 10bit > > > > bitstream.> > > > What kind of improvements do you expect? Actually, this series is designed > > > to change nothing for platforms, where 10-bit format is not added into > > > the list of supported formats. I think reasons are quite obvious. First, > > > not every device may support 10-bit output. Second, as you might already > > > figured it out, registers in this series are set only for legacy cores. I > > > have no idea, what needs to be done for newer ones, since I don't have > > > them. Anyway, I tested this with fluster and only one additional test > > > passes, because it is the only one for 10-bit YUV420. > > > > In this series you will find that I have added the registers for the new > > cores, fix hevc to be able to use 10-bit, and enable that in IMX8M. > > Your changes seems reasonable, but at this point I wouldn't bother with > fluster. Instead, try to test with one specific bitstream or even a sample video > file. I just tested with one random 10-bit VP9 video that I found when working > on this series. That way you avoid any corner cases which sometimes plaque > fluster testing (reference bitstreams smaller than min. supported size). > Anyway, re-check vendor lib if there is any other place to adjust something > for 10-bit. Just so we don't forget, there is a handful of 10bit tests that Daniel Almeida omitted when he added tests to fluster (though only 1 is 420). I will try and fix that later on. There is otherwise 5G worth of 10bit tests available. In fluster we decided to go for the same subset libvpx uses, otherwise no one would ever want to download these tests. https://storage.googleapis.com/downloads.webmproject.org/vp9/decoder-test-streams/Profile_2_10bit. About the "min supported", G2 VP9 scores is 157/303 here (in comparision rkvdec is 225, and MTK VCODEC 275). At this failure level this has no longer anything to do with the size of the render. There is likely couple of bugs hidden in the driver for the corner cases tested by the suite. Also, to illustrate that the size isn't the only variable in the failures, we have a vp90-2-02-size- 64x34.webm that pass (the driver pretends that 64x64 is the minimum). I didn't look at G2 output very closely, but on RKVDEC, in similar failures we have perfect keyframe, and a single corrupted tile on the following decode. My belief is that there is bugs in the drivers to be found and fixed. In absence of vendor support, or working reference it will be difficult / near impossible to fix, but I'm documenting this so we stop thinking this is just "not supported". cheers, Nicolas
