You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
RealtimeResampler::samples_outstanding() reports how many output samples the input written so far has yet to hand out, ready or not. Added to what a caller has read, it is where the next written sample will come out, to within a frame, so a caller can place a boundary such as a gapless track join without finalizing the stream.
Install ardftsrc-rs 0.4.4
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/phayes/ardftsrc-rs/releases/download/v0.4.4/ardftsrc-rs-installer.sh | sh
RealtimeResampler::write_span() and RealtimeResampler::read_span() replace write_rodio_buffer() and read_rodio_buffer(). They work without the rodio feature and allocate nothing. write_span() takes the samples from an iterator, starting a new span when their rate or channel count changes. read_span() hands back a SpanSamples iterator over one span's output, with its channel count and rate, so a caller can collect it into a buffer it reuses.
Removed
RealtimeResampler::write_rodio_buffer() and RealtimeResampler::read_rodio_buffer(), added in 0.4.2. Pass a rodio::buffer::SamplesBuffer's rate, channel count and samples to write_span(), and build one from a read_span() result's channels(), sample_rate() and collected samples.
Install ardftsrc-rs 0.4.3
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/phayes/ardftsrc-rs/releases/download/v0.4.3/ardftsrc-rs-installer.sh | sh
RealtimeResampler::write_rodio_buffer() and RealtimeResampler::read_rodio_buffer(), with the rodio feature, for push-style pipelines that resample finished packets as they are decoded and queue what comes out. A packet in another rate or channel count starts a new span by itself, and a read never straddles two spans, so each buffer has one channel count.
RealtimeResampler::is_empty() reports whether the resampler holds any audio, including the delayed tail of a stream not yet finalized, so a caller can send audio around it without reordering.
RealtimeResampler::pre() and RealtimeResampler::post() set previous- and next-track edge context for gapless material, as InterleavedResampler and PlanarResampler already do.
RealtimeResampler::write_iter() accepts interleaved samples from an iterator, without collecting them into a slice first.
Install ardftsrc-rs 0.4.2
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/phayes/ardftsrc-rs/releases/download/v0.4.2/ardftsrc-rs-installer.sh | sh
RealtimeResampler::is_input_span_primed() reports whether the span being written has been fed enough input to produce output. Unlike is_primed(), it goes false again after new_span(), so callers that fast-start can re-prime at a format change.
Fixed
RodioResampler no longer drops out for up to ~27 ms (at PRESET_GOOD) when the inner source changes sample rate or channel count mid-stream. Input pacing is now counted in frames, so a channel-count change no longer under-feeds the new span while the previous one drains, and with fast-start the new span is primed as soon as it starts.
RealtimeResampler::estimate_priming_duration() divides by the channel count, so it is no longer twice too long for stereo. RodioResampler::total_duration() without fast-start, which adds it, is corrected too.
The MIT licence names its copyright year and holder, and both licence files are now included in the published ardftsrc and ardftsrc-rs crates.
Install ardftsrc-rs 0.4.1
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/phayes/ardftsrc-rs/releases/download/v0.4.1/ardftsrc-rs-installer.sh | sh
High precision is now a sample type instead of a config option. With the high_precision feature, the new public types DoubleDouble, F128, and F256 can be used as any resampler's sample type, e.g. InterleavedResampler::<F128>. Convert samples in with From<f64>/From<f32> (exact) and out with to_f64().
Every stage now runs in the sample type: decimation, edge extrapolation, both FFTs, the spectral taper and phase, overlap-add, and output scaling. Previously only the FFTs ran in the wide type, and the spectrum was rounded to f64 between them, so high-precision output differs slightly from 0.3.1.
Removed Config::high_precision, Config::with_high_precision, HighPrecision, and Error::HighPrecisionIncompatibleWithF32.
InterleavedResampler, PlanarResampler, and RealtimeResampler are bounded by the new Sample trait instead of num_traits::Float + realfft::FftNum. Every Float + FftNum type implements Sample, so f32 and f64 code is unaffected unless it names the bounds.
ardftsrc-rs replaces --high-precision double-double|f128|f256 with --dd, --f128, and --f256, alongside --f32 and --f64. High-precision runs decode and encode in f64.
Added
RodioResampler::reset() discards buffered audio and resamples afresh from wherever the inner source is, without seeking it. Use it when the inner source jumps on its own, such as a cleared queue. With fast-start enabled it re-primes at once. try_seek() now calls it after seeking the inner source.
Changed
The CPU resampler now reuses per-core FFT scratch buffers instead of allocating them on every chunk.
Documentation
PERFORMANCE.md is regenerated. On one core, PRESET_HIGH converts 44.1 kHz → 48 kHz at 662× realtime (was 374×). rubato and PRESET_FAST are now within 3% of each other when upsampling.
Benchmarks
The throughput, chunk-time, and rubato benches share one input: 4 minutes of deterministic stereo white noise per input rate, replacing the test_wavs fixtures, which differed in length and channel count between rates.
Install ardftsrc-rs 0.4.0
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/phayes/ardftsrc-rs/releases/download/v0.4.0/ardftsrc-rs-installer.sh | sh
PERFORMANCE.md is regenerated for the 0.3.0 presets, with throughput, latency and per-chunk compute measurements, and a comparison with rubato.
PERFORMANCE.md has a new section on which presets suit realtime monitoring (5 ms), playback (100 ms) and offline use. No preset suits monitoring. PRESET_FAST and PRESET_GOOD suit playback from sources that can't be read ahead; every preset suits playback from files and other buffered sources.
The READMEs link to PERFORMANCE.md.
Benchmarks
New ardftsrc_latency bench. It measures latency through RealtimeResampler, the per-chunk compute time distribution and CPU load, and the latency floor at minimum quality.
ardftsrc_throughput includes PRESET_HIGH with HighPrecision::DoubleDouble when built with --features high_precision.
Install ardftsrc-rs 0.3.1
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/phayes/ardftsrc-rs/releases/download/v0.3.1/ardftsrc-rs-installer.sh | sh
All presets and the Config default have been retuned, so output differs from 0.2.0 for every preset. Each preset now scores higher on the HydrogenAudio SRC test suite and runs faster; PRESET_HIGH is much faster.
Preset
Quality
Bandwidth
Taper
PRESET_FAST
512
0.92
Kaiser(30.0)
PRESET_GOOD (default)
2304
0.975
Kaiser(30.0)
PRESET_HIGH
7056
0.996
Kaiser(34.0), alias floor 0 dB
PRESET_EXTREME is removed, along with ardftsrc-rs --preset extreme. PRESET_HIGH now permits aliasing at the lower Nyquist (alias floor 0 dB) and scores higher than PRESET_EXTREME did, with a quarter of its FFT size. Use PRESET_HIGH instead.
PRESET_HIGH now supports f32 processing (its quality is within the f32 limit of 8192).
TaperType::default() is now Kaiser(30.0) (was Cosine(3.4375)).
The bessel feature is removed. The Kaiser taper no longer depends on pxfm and is always available.
TaperType::Bessel(alpha) is replaced by TaperType::Kaiser(beta), which takes the Kaiser window's beta as MATLAB and SciPy do: Bessel(alpha) corresponds to Kaiser(PI * alpha). It integrates the window exactly and samples it at bin centres, so it is exactly amplitude-complementary and centred on its transition. When the transition straddles Nyquist it samples from unity to zero inclusive, putting −6 dB on Nyquist.
Kaiser accepts beta in (0, 300] and otherwise fails validation with the new Error::InvalidKaiserBeta.
TaperType::Kbd and TaperType::HalfKaiser are removed.
ardftsrc-rs: --taper-type bessel is now kaiser and takes --beta (default 30) instead of --alpha. kbd and half_kaiser are removed.
CpuCore is no longer exported. It could not be constructed outside the crate, since its constructor takes a crate-internal configuration type. Use PlanarResampler with channels = 1 for single-channel resampling.
Fixed
With a lowered alias floor, the transition is now centred on the lower Nyquist, so every pair of bins mirrored about it sums to unity gain and Nyquist sits at −6 dB. Taper samples that had underflowed to zero were trimmed, which pushed the transition above Nyquist: with the default taper and the widest floor at 44.1 kHz → 96 kHz, −6 dB sat 9 bins high and Nyquist was at −3.4 dB. This affects PRESET_HIGH, which now uses an alias floor of 0 dB. Without aliasing the layout is unchanged.
Taper profiles are now computed in f64 for every sample type. In f32, Cosine and Planck tails underflowed earlier and shifted the transition toward Nyquist (by up to 280 bins for Cosine(3.4375) over 5000 transition bins).
End-of-stream stop extrapolation no longer reads its input history from a buffer the forward FFT had already used as scratch. With the default FFT backend, the synthesized tail was built from corrupted samples, degrading the final chunk; outputs now match the high-precision backends.
With decimate enabled, output is now exactly time-aligned with the input. The decimation cascade's group delay was rounded to whole output samples, leaving a fractional-sample offset; the remainder is now compensated in the spectral stage.
Documentation
Quality reports regenerated for the new presets, with an f32 report for PRESET_FAST.
Added guidance on compiler optimization settings for realtime resampling.
Install ardftsrc-rs 0.3.0
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/phayes/ardftsrc-rs/releases/download/v0.3.0/ardftsrc-rs-installer.sh | sh
RodioResampler no longer over-reads input from packetized sources such as rodio's decoders. When upsampling, the buffered audio grew by about 88 ms per second of playback, delaying seeks by several seconds and allocating on the audio thread.
RodioResampler::try_seek() now discards audio buffered from before the seek and re-primes with fast-start, instead of playing up to two chunks of stale audio first (seconds at PRESET_HIGH and above).
RodioResampler no longer panics on sources whose current_span_len() is None.
RodioResampler no longer writes the first sample of a new span into the previous span when the source exposes the new format only after that sample is pulled, which panicked or swapped channels on format changes.
RealtimeResampler::reset() no longer allocates.
Install ardftsrc-rs 0.2.0
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/phayes/ardftsrc-rs/releases/download/v0.2.0/ardftsrc-rs-installer.sh | sh
