DFT should use aligned memory blocks for best results. Also allocate one
extra sample for DFT output, just in case DFT zop is as bad as zrop.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
DFT float happens to clobber one extra sample on forward translate, so
allocate one extra for every complex buffer.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
Work back to a vDSP implementation, this time using overlap-save instead
of overlap-add, also accumulating the results as complex values, only
inversing them once at the end, and finally, replacing the FFT method
with the newer DFT API.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
Only uninitialize the equalizer if sound output was successfully started
and the equalizer AudioUnit was successfully ininitialized.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
When leaving the workgroup, clear the token, as the join call requires
the token to be uninitialized.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
Errors should stop all attempts to further use the audio thread priority
code, so there won't be debug breakpoints called on older OSes.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
As the decimator has shown to be twice as loud as it should be, the
volume should be reduced by half when converting DSD to PCM with it.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
Pure downsampling is slower, but may or may not be more accurate. Though
probably not worth it. It did help me realize a minor error, though.
The decimator's volume is twice as loud as it should be.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
This prevents crashes where inputs were not returning either properties
or metadata blocks and the file open cache was attempting to cache the
resulting nil pointer as if it were valid. Also prevent the metadata
redundant string coalescing from processing nil objects as well, in case
it's used that way somewhere else.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
Set baseline real-time priority for audio threads even on old macOS,
since that API is available there. Only set it once, and do not attempt
again if it fails, only once per thread.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
I'm not sure about macOS Ventura, but stable releases of macOS, at
least on Intel, require that threads joining Audio Interval
workgroups already be set to run as real-time before joining. Not
doing this results in an uncaught exception and a crash.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
Now it allocates audio workgroups per thread, using work slices like the
Apple documentation describes for asynchronous threads.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
On Big Sur or newer, it is possible to join the audio threads to the
same OS workgroup as the audio output device, improving response.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
Update all project files with new upgrade version number, and add the
dead code stripping option. Don't touch MASShortcut because it's not my
project.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
Replace overlap-add vDSP/Accelerate implementation with a faster PFFFT
overlap-save implementation, using fewer FFT steps as well.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
Rewrite some of the output and a lot of the downmixer to use Accelerate
framework instead of dumb for loops.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
Apparently, PFFFT double is much faster than vDSP, and I didn't even
notice. Thanks to Aleksey Vaneev for testing this properly.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
When resampling the impulse according to the playback rate, it becomes
necessary to normalize the resulting impulse by the inverse of the
sample ratio, as resampling adds more or less loudness by virtue of
interpolating samples.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
This will be proper at least unless I get this commit merged upstream.
Squashed the changes to a single commit, and removed extraneous
whitespace that crept into the code.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
The function I added only works for non-interleaved real/imaginary pairs
and not the interleaved setup that r8brain expects. Fix that by removing
the multiply implementation and using the original one.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
The used fork of r8brain now uses the Accelerate vDSP FFT functions for
resampling, which should provide a slight speedup, or significant for
large sample ratios.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
Moved the string encoding guesser/converter to the Plugin.h header, so
it may be accessible from any plugin. I may make it a global member of
something eventually, but a static inline for such a simple function
should be fine for now.
This function facilitates converting arbitrary 8 bit encoded strings to
Unicode NSString objects. It should be used anywhere that UTF-8 is
expected, but not necessarily guaranteed, and where other 8-bit
encodings may also be supplied by a user's files.
Not using this setup for string inputs has already led to failed UTF-8
decoding resulting in nil NSStrings being passed to the inline array or
dictionary initializers, which results in crashes due to uncaught
exceptions.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
Update the Info.plist generator to emit file type definitions which use
system generated icons in place of the legacy icons in the app bundle.
Also include the new LSHandlerRank field. And also add a definition for
the scripting definition, which I accidentally added to the Info.plist
manually when I fixed scripting.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
Add safety check to check if a device is actually alive when enumerating
it, and also add nil pointer checks for the device name before trying to
CFRelease it. Fixes a rare crash on device add/remove cycle, such as
Bluetooth headphones.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
The resampler wasn't being given enough room to flush its final output,
so a function was added to determine the current output latency, and
more sample data is requested, allowing the full output flush to occur.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
These two changes fix playback issues with either starting in the middle
of the playlist on a really short file terminating immediately instead
of queueing more files (InputNode.m), and issues with starting playback
at all on the end of a playlist on a short file. (OutputCoreAudio.m)
Fixes#246
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
Just in case anything using the implementation ever needs to request
less sample data than would be returned by the resampler, it should be
able to return a remainder and keep extra remaining samples, if any.
However, the way Cog currently uses it, it would not be likely to run
into this scenario.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
Fixes to the resampler wrapper, such that it will survive some close
encounters with the edge of the buffer, if necessary. Also so it will
obey the buffer size limit for the output buffer.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
This rename is more in line with what R8Brain does in its example code.
No actual behavioral changes to the code, however.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
The cache thread should have an autoreleasepool around the release loop,
because it will be freeing Objective C objects periodically.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
Promote the Plugin Controller source file to Objective-C++, and add a
simple data cache that holds on to requests for up to 5 seconds after
their last access, for preventing spammed requests from hitting files
over and over. This is apparently really relevant to the CUESheet reader
and its embedded CUESheet handling, as that tends to reread the same
file over and over as it populates the playlist with tracks. The nested
reader can also lead to repeated reading even on files without CUESheets
embedded.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
This should improve performance slightly again, as there were some ARM
code paths that weren't being enabled for ARM64.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
Replace libsoxr dylib with a static library, and also build the two
architectures separately, to allow for platform-specific optimizations
to be employed for both. This also reduces the size of the CogAudio
framework by a few hundred kilobytes, as we eliminate unused code paths
better this way.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
When the option is enabled, and playback comes to a completion, the
player will quit on its own.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
IN A.D. 2101, WAR WAS BEGINNING. *boom*
Yeah, this was a dumb bug, I didn't realize that AUAudioUnit would just
arbitrarily ignore my configured block size and request a different one.
The AirPods Pro will just request 480 instead of the 512 I ask for, so
let's instead support variable block sizes, and only take up to the last
4096 samples of the chunk fed to the output device.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
vDSP functions expect their input and output pointers to be aligned to
an even four values. Correct this by aligning all pointers. The
allocated buffers used for one parameter should already be aligned
somewhat, but align the incremented positions used on some of them so
that the vDSP functions don't misbehave. Also align the volume scaler
input by doing scalar math until the pointer is aligned prior to calling
vDSP_vsmul. Also, change 16-bit and 32-bit scale to use vsdiv instead of
vsmul with a really small number already divided into one.
Fixes the test vectors that were sent in extrapolating incorrectly due
to their final blocks having uneven sample counts, resulting in
unaligned pointers.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
A bad sample scanner and cleaner will point out in the log whenever a
bad sample, such as infinity, or Not a Number, or even huge values over
±2.0, in case some piece of code, or a decoder, or even a bad file, has
taken over the output.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
The original didn't really handle backwards versus forwards differently,
as far as the predictor coefficients should have been, as they probably
should have been reversed for a different direction window.
This didn't fix my problem, though, but did possibly expose something
else to mess with.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
In the rare event that we're somehow playing decimated DSD at full
sample rate instead of resampling, only the start needs to be skipped,
and the end needs the input to the decimator padded to flush it, but
nothing needs to be truncated from the end of the output in that case.
Still, mostly pointless, since next to nobody will be outputting 384 kHz
from their Macs, in any case, much less unprocessed DSD.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
We should be extrapolating right over top of the DSD decimator latency,
rather than in front of it. Yeah, that'll do.
Signed-off-by: Christopher Snowhill <kode54@gmail.com>
Christopher Snowhill