Repository navigation
AudioStreamConverter - #1551
Merged
Merged
AudioStreamConverter#1551
Conversation
Mysticial
marked this pull request as ready for review
October 11, 2026 06:42
This was referenced Oct 12, 2026
Closed
Mysticial
pushed a commit
that referenced
this pull request
Oct 13, 2026
…er (+ in_stride fix from #1568) (#1569) * AudioChannelMixer: honor in_stride in the 2-input Default fast paths; fix test comparing an implementation to itself. mix_channels_Default<2, 1> and <2, 2> loaded in[0]/in[1] instead of in[in_stride * 0]/in[in_stride * 1], so any strided input (the layout the AudioPipeline ChannelMixer always produces) read the wrong samples. The SSE/FMA3/AVX512 implementations all index with in_stride, and the generic Default fallback did too; only the two fast paths were wrong. The existing kernel tests never caught this because Kernels_AudioChannelMixer_TestInfra's Test constructor initialized m_convert0 with convert1, comparing an implementation against itself. Verified: the fixed Default implementation now matches mix_channels_GenericDefault bit-exactly for all 1-4 channel in/out combinations across packed and strided inputs (in_stride 1/5/512, sample counts 2/3/17/930). (cherry picked from commit 1efc832) * AudioPipeline: add unit tests for Raw<->Working round-trip and ChannelMixer. Mirrors the TestInfra pattern and sample-size progression from the AudioChannelMixer/AudioStreamConverter/AudioChannelInterleave kernels (#1551): for (size_t samples = 2; samples < 1000; samples = (size_t)(samples * 1.1) + 1). - Raw<->Working round-trip (Kernels_AudioPipeline_RoundTrip_TestInfra): the two converters are tested as a round-trip through both, since the test belongs to neither class. Random raw samples are pushed through RawAudio2WorkingAudio -> WorkingAudio2RawAudio (1.0f multiplier) into a recording RawAudioStreamListener and compared against the input, bit-exactly. Loops all sample formats (u8/s16/s32/f32), all channel counts 1-4, and the #1551 sample sizes. s32 input is masked to 0xfffff000 and all values are integer-valued so the round-trip is lossless (same convention as the AudioStreamConverter tests). - ChannelMixer (Kernels_ChannelMixer_TestInfra): - MixerIdentity: identity weight matrix for 1-4 channels; random integer samples must come out unchanged. - Mixer_1to2_2to1: a 1->2 mixer with the averaging matrix of a single channel (identity, weights 1.0) feeding a 2->1 averaging mixer (weights 0.5, exactly representable). Random integer inputs keep every intermediate value exactly representable, so the 1->2 output is identical to the input on both channels and the 2->1 output is identical to the original input. Registered via AudioPipeline::add_tests() from Kernels_Tests.cpp; ~1092 test instances in total. * AudioPipeline tests: 256-sample converter buffer; four mixer identity tests; one 1->2->1 round-trip. - The round-trip test used the default converter buffer (4096 + 256), which is larger than every sample size in the sweep, so the chunk loop never ran more than once. Override it to 256 so the sweep exercises multiple chunks. - The ChannelMixer tests are one identity test per channel count (1-4, 1000 samples each) and one 1->2->1 averaging round-trip (1000 samples), instead of sweeping the sample sizes. - Kernels_AudioPipeline.h: include <cstddef> for size_t, like the sibling kernel headers.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Adds 3 new audio processing modules:
AudioStreamConverter: Takes a raw audio stream and converts to/from thefloatformat. (this is a newer version ofAudioStreamConversionwith more SIMD)AudioChannelInterleave: Takes a stream of audio frames (interleaved channels) and splits them into separate single-channel streams. Also does the reverse direction.AudioChannelMixer: Takes an N-channel audio stream and converts it to an M-channel stream using a custom weight matrix.Thus our future audio pipeline will be:
floatwe convert it tofloatusingAudioStreamConverter.AudioChannelInterleave.For the audio playback path:
4. If the sample rate does not match the audio sink's sample rate, resample it using
AudioResampler.5. Re-interleave the channels using
AudioChannelInterleave.6. Convert to the audio sink's desired format using
AudioStreamConverter.For the FFT/inference path:
4. If the sample rate is not 48 kHz, we resample it to 48 kHz using
AudioResampler.5. We collapse the streams into 1 channel using
AudioChannelMixer.6. Perform FFTs and inference.
Step 4 on both branches may be optimized to one step if the conversion is identical.