Re: [flac-dev] libFLAC optimizations request
Hi, On Sat, Feb 3, 2018 at 10:10 AM, Erik de Castro Lopo wrote: > Gabriel Corneanu wrote: > >> Is this interesting for the project? > > I would theoretically accept a patch that allows direct access to the > functionality you seek, as long as it doesn't change the existing API. > >> If yes, how should I share it? As a patch against current git? > > Probably easiest as as PR against: > >https://github.com/xiph/flac/ I create a PR. The existing API should not have been changed, all tests are passed. >> >> Other small details: >> - I commented out by default the vorbis comment; if this is important >> to be kept, I would like an option (e.g. I could save ~100MB out of >> ~1.1GB compressed data). > > I'm pretty sure this default vorbis comment is part of the official > specfication of FLAC and hence should probably not even be made optional. Here https://xiph.org/flac/format.html#format_overview only the streaminfo is mandatory. > You can probably avoid it by using the library directly instead of using > the flac command line program. It is the other way around :) . I could NOT avoid generating the comment with the library. Even after my change, the command line tool is still generating it (I don't really mind). My proposal is to not generate it by default (with libFLAC) for a non-ogg target. > Erik Gabriel ___ flac-dev mailing list flac-dev@xiph.org http://lists.xiph.org/mailman/listinfo/flac-dev
Re: [flac-dev] libFLAC optimizations request
Hi, On Sun, Feb 4, 2018 at 11:50 PM, Brian Willoughby wrote: > I wasn’t suggesting that you run metaflac, but that you examine its source to > see how it creates new FLAC files without the Vorbis comment. As far as I > know, metaflac uses the standard libFLAC and creates files without the Vorbis > overhead. Without reading the sources, I simply assume that metaflac is parsing the metadata blocks and filtering as required. Of course it is possible with libFLAC, but the overhead is huge (parse stream, filter, save back, etc) compared to simply not generating it. > My quick review of the source seemed to indicate that calling > FLAC__metadata_object_new(FLAC__METADATA_TYPE_VORBIS_COMMENT) will create the > comment, but I assumed that you could pass in NULL. I mistakenly thought that > was FLAC__stream_encoder_new(). The comment is auto-generated in "init_stream_internal_", called from all _init functions. I am NOT explicitly creating any metadata object. > There is a comment saying that the Ogg FLAC mapping requires a > VORBIS_COMMENT, but that doesn’t necessarily mean that standard FLAC also > requires the Vorbis comment. That's what I thought, maybe if the target is NOT Ogg we could by default not generate the comment; the user can still explicitly generate it. > Looking around metaflac sources, I see > FLAC__metadata_iterator_delete_block(iterator, options->use_padding), which > might be the code responsible for removing the Vorbis comment. > > In other words, perhaps there is no existing way to avoid having the encoder > create the Vorbis comment metadata, but there might be a way to delete that > particular metadata before the file is saved. A closer look at how metaflac > creates a file without Vorbis comments should help. > > Brian > I will create a PR with my suggestions. Gabriel ___ flac-dev mailing list flac-dev@xiph.org http://lists.xiph.org/mailman/listinfo/flac-dev
Re: [flac-dev] libFLAC optimizations request
The problem is really as I wrote: 1. Metaflac is no option for me, I use libFLAC.dll 2. There is no way (at least how I read the code) to avoid saving comment with libFLAC; I would appreciate an extra option to avoid it, which can default to old behavior if compatibility is important. 3. I have a high speed application, where re-initializing an encoder is really significant. On corner cases it causes an 25% overhead! Of course I don't expect it to be that significant in normal cases. Thanks for all replies, but I don't have the code at home. I will create a patch with my changes for review. Regards, Gabriel On Sun, Feb 4, 2018 at 9:57 AM, Brian Willoughby wrote: > Correction, the flac command-line does create a 40-byte Vorbis comment by > default. I just never noticed it before. I’ve been using —no-padding all > these years for minimal file sizes without realizing that I could save > another 40 bytes. > > Anyway, since metaflac can remove the Vorbis comment using the standard > library, then you should be able to create a solution without modifying > libFLAC. > > Brian > > > On Feb 4, 2018, at 12:43 AM, Brian Willoughby wrote: >> Gabriel, >> >> metadata_has_vorbis_comment is a FLAC__bool which defaults to false. There >> should be no reason to modify stream_encoder.c, but just modify the caller. >> >> The following command: >> >> metaflac —remove —block-type=VORBIS_COMMENT —don’t-use-padding >> >> … will remove Vorbis comments from existing files, so is must be legal >> without modifying the library. metaflac can clearly create a new FLAC file >> without the Vorbis comment. >> >> I use the flac command-line, and I never get Vorbis comments in the files >> that I create. Perhaps you are using another tool which assumes Vorbis >> comments are needed. >> >> >> The FLAC algorithm is not dependent upon sample rate. AIFF has an 80-bit >> floating point type for sample rate, so it should be able to handle 40 MHz. >> I assume that any AIFF can be converted to FLAC losslessly, but I have not >> tested whether certain sample rates are rejected. FLAC itself only supports >> sample rates up to 655,350 Hz, so you may have a problem there unless you >> “lie” about the sample rate when creating the file. Maybe you could just >> establish a private convention to divide the sample rate by 100 to make >> yours fit. 40 MHz would map to 400 kHz, 10 MHz would map to 100 kHz, and 5 >> MHz would map to 50 kHz. >> >> >> You’re probably asking for trouble if you try to reuse an encoder. It seems >> like there would always be some risk that details from the previous file >> would bleed through into the next. Have you benchmarked allocation and >> initialization? Is it really that slow? In order to reuse an encoder, you’ll >> need to overwrite all state variables, and I don’t see how that could be >> much faster than simply allocating them anew. Perhaps you could allocate >> groups of encoders at once, if that would speed the process. >> >> >> On Feb 1, 2018, at 4:29 AM, Gabriel Corneanu >> wrote: >>> Hello all >>> >>> I am using libFLAC in a corner application, compressing a lot of small >>> signals. >>> First is a general question: in our application we have signals in range >>> 5-10 MHz, potentially 40MHz! Is there any potential problem with that?? The >>> mac sample rate is limited in flac, but it doesn't really seem to be a >>> problem. >>> The output is stored as blob in a sqlite database, it never needs to be a >>> valid audio file outside our application. >>> In my tests, the signals are compressed very well, much better than general >>> compression libraries like zlib, zstd, etc. >>> >>> Now other small issues; I also made some tickets about them, but I thought >>> asking here might be better. >>> >>> 1. I would like to avoid saving vorbis comment, by default ~40 bytes. Right >>> now the only option is to modify stream_encoder.c, see >>> "metadata_has_vorbis_comment". >>> >>> 2. Speed is very important, therefore I would like to reuse an encoder >>> without re-initializing everything. >>> Ideally I would like 2 (exported) functions: "flush" and "restart". >>> "Flush" is self-explanatory, should properly end the encoding. I could >>> split myself "flush" from "finish", it looks relatively simple. >>> "Restart" should keep all current settings, generate a new stream header >>> and clear everything for encoding a new signal. >>> It' clear that current settings, re-creating windows, cpu-dependent >>> functions, etc could be kept around. >>> I was not quickly able to extract all the necessary initialization from >>> "init_stream_internal_" into a new "FLAC__stream_encoder_restart" function. >>> >>> Regards, >>> Gabriel Corneanu > ___ flac-dev mailing list flac-dev@xiph.org http://lists.xiph.org/mailman/listinfo/flac-dev
Re: [flac-dev] libFLAC optimizations request
Hello again, After some time, I succeeded in splitting "flush" and "restart" functions. A quick measurement gives me ~25% speed improvement, which is quite a lot. Of course this optimization is specific for this use case (many short signals). Is this interesting for the project? I would gladly contribute it... If yes, how should I share it? As a patch against current git? Other small details: - I commented out by default the vorbis comment; if this is important to be kept, I would like an option (e.g. I could save ~100MB out of ~1.1GB compressed data). - I allowed ".._set_total_samples_estimate" in initialized state (intended to be used between flush and restart), in order to generate a proper streaminfo header; it seems to be otherwise harmless even if misused. Regards, Gabriel ___ flac-dev mailing list flac-dev@xiph.org http://lists.xiph.org/mailman/listinfo/flac-dev
[flac-dev] libFLAC optimizations request
Hello all I am using libFLAC in a corner application, compressing *a lot* of small signals. First is a general question: in our application we have signals in range 5-10 MHz, potentially 40MHz! Is there any potential problem with that?? The mac sample rate is limited in flac, but it doesn't really seem to be a problem. The output is stored as blob in a sqlite database, it *never *needs to be a valid audio file outside our application. In my tests, the signals are compressed very well, much better than general compression libraries like zlib, zstd, etc. Now other small issues; I also made some tickets about them, but I thought asking here might be better. 1. I would like to avoid saving vorbis comment, by default ~40 bytes. Right now the only option is to modify stream_encoder.c, see "metadata_has_vorbis_comment". 2. Speed is very important, therefore I would like to reuse an encoder without re-initializing everything. Ideally I would like 2 (exported) functions: "flush" and "restart". "Flush" is self-explanatory, should properly end the encoding. I could split myself "flush" from "finish", it looks relatively simple. "Restart" should keep all current settings, generate a new stream header and clear everything for encoding a new signal. It' clear that current settings, re-creating windows, cpu-dependent functions, etc could be kept around. I was not quickly able to extract all the necessary initialization from "init_stream_internal_" into a new "FLAC__stream_encoder_restart" function. Regards, Gabriel Corneanu ___ flac-dev mailing list flac-dev@xiph.org http://lists.xiph.org/mailman/listinfo/flac-dev
