Re: [libav-devel] dca xll: Working audio
ni...@lysator.liu.se (Niels Möller) writes: It seems to give reasonable output for the Master Audio 7.1.dts file, except that the LFE channel is messed up in some places. I suspect this is not a bug in my code... Let's I look at one sample where there's a large error, index 461360. For the reference flac file, I have sample values -134990 -39862 -1678929 -248163 -116804 -112124 -1656 -34272 With avconv master I get -95623 -28254 -1187410 -159055 -83522 -103577 This is the core channels only, downmix version of the 7.1 data, and without applying any lossless residuals. The inverse downmix scales all channels by sqrt(2), and then undoes some mixing of the last two surround channels. Ignore the surround stuff, and just scale the first four channels. This gives -135231 -39959 -1679253 -224938 with errors -241-97 -324 23225 So the final channel, the LFE, gets a much worse error than the three normal channels. In dcadec.c, there's a comment /* Generate LFE samples for this subsubframe FIXME!!! */ which looks like it predates the libav/ffmpeg fork. Does anyone remember what the problem was? Regards, /Niels -- Niels Möller. PGP-encrypted email is preferred. Keyid C0B98E26. Internet email is subject to wholesale government surveillance. ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] dca xll: Working audio
On Thu, May 08, 2014 at 02:25:40PM +0200, Niels Möller wrote: ni...@lysator.liu.se (Niels Möller) writes: It seems to give reasonable output for the Master Audio 7.1.dts file, except that the LFE channel is messed up in some places. I suspect this is not a bug in my code... Let's I look at one sample where there's a large error, index 461360. For the reference flac file, I have sample values -134990 -39862 -1678929 -248163 -116804 -112124 -1656 -34272 With avconv master I get -95623 -28254 -1187410 -159055 -83522 -103577 This is the core channels only, downmix version of the 7.1 data, and without applying any lossless residuals. The inverse downmix scales all channels by sqrt(2), and then undoes some mixing of the last two surround channels. Ignore the surround stuff, and just scale the first four channels. This gives -135231 -39959 -1679253 -224938 with errors -241-97 -324 23225 So the final channel, the LFE, gets a much worse error than the three normal channels. In dcadec.c, there's a comment /* Generate LFE samples for this subsubframe FIXME!!! */ which looks like it predates the libav/ffmpeg fork. Does anyone remember what the problem was? It comes from the original libdts (now known as libdca) so no. But IIRC D** decided to use IIR-based interpolation for LFE in case it's not core-only so the problem lies in Calabasas, California. ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] dca xll: Working audio
Kostya Shishkov kostya.shish...@gmail.com writes: But IIRC D** decided to use IIR-based interpolation for LFE in case it's not core-only so the problem lies in Calabasas, California. Sounds like a mess... Any information about what that filter is like? Regards, /Niels -- Niels Möller. PGP-encrypted email is preferred. Keyid C0B98E26. Internet email is subject to wholesale government surveillance. ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] dca xll: Working audio
ni...@lysator.liu.se (Niels Möller) writes: I'm now looking at the Master Audio 7.1.dts file. There are 5.1 core channels, and the xll extension has two channel sets. Channel set 0 contain 6 residual channels, and from the channel mask, these are, in order, C, L, R, LFE1, Lss, Rss (surround side). Channel set 1 contain 2 non-residual channels, Lsr, Rsr (surround rear). And it has the downmix coefficients, and the BHierChSet flag set. Implemented this now, mostly as described in previous mail. For residual channels, also the core channels need to be scaled. For now, that's no problem, because I do one channel set at a time, and by the time I get to chset 1 which implies an inverse downmix, core and residual are already added together for the first channel set. It seems to give reasonable output for the Master Audio 7.1.dts file, except that the LFE channel is messed up in some places. I had to add a couple of additional hacks for channel order, and I also had to implement fixed prediction, since that was used for one of the frames in this file. Available at my wip repo as usual. Regards, /Niels -- Niels Möller. PGP-encrypted email is preferred. Keyid C0B98E26. Internet email is subject to wholesale government surveillance. ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] dca xll: Working audio
Kostya Shishkov kostya.shish...@gmail.com writes: Maybe someone can give you a sample with lossless-only tracks? IIRC it's quite common to have 7.1 lossless audio with 5.1 core channels with residue plus two fully lossless channels. And you'll have fun with remixing channel sets then. Fun, indeed... I'm now looking at the Master Audio 7.1.dts file. There are 5.1 core channels, and the xll extension has two channel sets. Channel set 0 contain 6 residual channels, and from the channel mask, these are, in order, C, L, R, LFE1, Lss, Rss (surround side). Channel set 1 contain 2 non-residual channels, Lsr, Rsr (surround rear). And it has the downmix coefficients, and the BHierChSet flag set. There appear to be 18 coefficients (2+1) * 6. All either zero or 1/sqrt(2) (nine bit codes 256 and 473, respectively). To make sense of them, I can arrange them in the following matrix: C L R LFE Lss Rss self? 0.7 0.7 0.7 0.7 0.7 0.7 Lsr 0 0 0 0 0.7 0 Rsr 0 0 0 0 0 0.7 Ordered in the stream starting with the first column, i.e., 0.7, 0, 0, 0.7, 0, 0, ..., 0.7, 0, 0.7. I guess that this means that the downmix was done as C' = 0.7 C L' = 0.7 L R' = 0.7 R LFE' = 0.7 LFE Lss' = 0.7 Lss + 0.7 Lsr Rss' = 0.7 Rss + 0.7 Rsr where C, L, ..., Rsr are the intended 7.1 audio, and C', L',..., Rss' is the downmixed 5.1 audio coded by the first channel set (+ core). Then, to undo this downmix and recover 7.1 audio, I have to compute C = C' / 0.7 L = L' / 0.7 R = R' / 0.7 LFE = LFE' / 0.7 L = L' / 0.7 Lss = Lss' / 0.7 - Lsr Rss = Rss' / 0.7 - Lsr Or in general, if column X is x_0, x_1, x_2, downmix is X' = x_0 X + x_1 Lsr + x_2 Rsr and the inverse is X = X' / x_0 + (x_1 / x_0) Lsr + (x_1 / x_0) Rsr And the inversion of x_0 is what the invDmixTab (D.11 in the spec) is for. Do I get this right? Regards, /Niels -- Niels Möller. PGP-encrypted email is preferred. Keyid C0B98E26. Internet email is subject to wholesale government surveillance. ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] dca xll: Working audio
On 28 Apr 2014, at 15:01, Hendrik Leppkes h.lepp...@gmail.com wrote: I also used to have a sample somewhere where all audible audio is in the lossless extension for all channels, playing just the core yielded nothing but silence. Maybe I can dig it up if its of interest. - Hendrik Please do (more out of curiosity than any other reason though). Tim ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] dca xll: Working audio
On 28 Apr 2014, at 14:44, Kostya Shishkov kostya.shish...@gmail.com wrote: On Mon, Apr 28, 2014 at 02:35:36PM +0200, Niels Möller wrote: After implementing proper channel reorder, I think the audio is pretty close to correct. Peak sample error for Master Audio 5.0 96khz.dts is now down to 115 (for 24-bit integer output). To be compared to 7183 if the core channels are upmixed but the xll residual data is ignored. Code in my public wip repo at http://www.southpole.se/~nisse/libav.git, branch xll. What's the next steps? I guess I should rebase or merge with the master branch, maybe remove or clean up debug output, and then post the patch here for review? (It might be possible to split it in pieces, but I doubt the history in my repo is very useful). Or you can work on integer mode for DTS. Indeed, this will be needed for lossless eventually, so if you're not sure what to work on you might as well tackle it now :-) Tim ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
[libav-devel] dca xll: Working audio
After implementing proper channel reorder, I think the audio is pretty close to correct. Peak sample error for Master Audio 5.0 96khz.dts is now down to 115 (for 24-bit integer output). To be compared to 7183 if the core channels are upmixed but the xll residual data is ignored. Code in my public wip repo at http://www.southpole.se/~nisse/libav.git, branch xll. What's the next steps? I guess I should rebase or merge with the master branch, maybe remove or clean up debug output, and then post the patch here for review? (It might be possible to split it in pieces, but I doubt the history in my repo is very useful). It would also be good with some command line option to disable xll processing (don't know how codec-specific options are done). Regards, /Niels -- Niels Möller. PGP-encrypted email is preferred. Keyid C0B98E26. Internet email is subject to wholesale government surveillance. ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] dca xll: Working audio
On Mon, Apr 28, 2014 at 02:35:36PM +0200, Niels Möller wrote: After implementing proper channel reorder, I think the audio is pretty close to correct. Peak sample error for Master Audio 5.0 96khz.dts is now down to 115 (for 24-bit integer output). To be compared to 7183 if the core channels are upmixed but the xll residual data is ignored. Code in my public wip repo at http://www.southpole.se/~nisse/libav.git, branch xll. What's the next steps? I guess I should rebase or merge with the master branch, maybe remove or clean up debug output, and then post the patch here for review? (It might be possible to split it in pieces, but I doubt the history in my repo is very useful). Maybe someone can give you a sample with lossless-only tracks? IIRC it's quite common to have 7.1 lossless audio with 5.1 core channels with residue plus two fully lossless channels. And you'll have fun with remixing channel sets then. Or you can work on integer mode for DTS. It would also be good with some command line option to disable xll processing (don't know how codec-specific options are done). There's an option for disable_xch already there, just follow the example. ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] dca xll: Working audio
On Mon, Apr 28, 2014 at 2:44 PM, Kostya Shishkov kostya.shish...@gmail.com wrote: On Mon, Apr 28, 2014 at 02:35:36PM +0200, Niels Möller wrote: After implementing proper channel reorder, I think the audio is pretty close to correct. Peak sample error for Master Audio 5.0 96khz.dts is now down to 115 (for 24-bit integer output). To be compared to 7183 if the core channels are upmixed but the xll residual data is ignored. Code in my public wip repo at http://www.southpole.se/~nisse/libav.git, branch xll. What's the next steps? I guess I should rebase or merge with the master branch, maybe remove or clean up debug output, and then post the patch here for review? (It might be possible to split it in pieces, but I doubt the history in my repo is very useful). Maybe someone can give you a sample with lossless-only tracks? IIRC it's quite common to have 7.1 lossless audio with 5.1 core channels with residue plus two fully lossless channels. And you'll have fun with remixing channel sets then. Samples for that are available in the usual places, ie this one should do the trick: http://samples.ffmpeg.org/A-codecs/DTS/dts/Master%20Audio%207.1%2024bit.dts I also used to have a sample somewhere where all audible audio is in the lossless extension for all channels, playing just the core yielded nothing but silence. Maybe I can dig it up if its of interest. - Hendrik ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] dca xll: Working audio
Kostya Shishkov kostya.shish...@gmail.com writes: Maybe someone can give you a sample with lossless-only tracks? IIRC it's quite common to have 7.1 lossless audio with 5.1 core channels with residue plus two fully lossless channels. And you'll have fun with remixing channel sets then. I think I have two 7.1 samples, which I haven't dared trying yet. About the ch_mask in the channel header, you said it uses the conventional order C, L, R, (for the 7.1 samples I have, according to eac3to, they contain the channels C L R Ls Rs LFE Lw Rw ($40f) and C L R LFE Lsr Rsr Lss Rss ($84b). Is there some authoritative spec for the order used in xll? If I have understood correctly, libav uses the same order as specified for .wav (which is different from xll), (http://msdn.microsoft.com/en-us/library/windows/hardware/dn653308%28v=vs.85%29.aspx). There's an option for disable_xch already there, just follow the example. Ok, shouldn't be too hard. Regards, /Niels -- Niels Möller. PGP-encrypted email is preferred. Keyid C0B98E26. Internet email is subject to wholesale government surveillance. ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
