nih-plug/src/util/stft.rs

//! Utilities for buffering audio, likely used as part of a short-term Fourier transform.

use crate::buffer::Buffer;

/// Process the input buffer in equal sized blocks, running a callback on each block to transform
/// the block and then writing back the results from the previous block to the buffer. This
/// introduces latency equal to the size of the block.
///
/// Additional inputs can be processed by setting the `NUM_SIDECHAIN_INPUTS` constant. These buffers
/// will not be written to, so they are purely used for analysis. These sidechain inputs will have
/// the same number of channels as the main input.
///
/// TODO: Better name?
/// TODO: We may need something like this purely for analysis, e.g. for showing spectrums in a GUI.
///       Figure out the cleanest way to adapt this for the non-processing use case.
pub struct StftHelper<const NUM_SIDECHAIN_INPUTS: usize = 0> {
    // These ring buffers store the input samples and the already processed output produced by
    // adding overlapping windows. Whenever we reach a new overlapping window, we'll write the
    // already calculated outputs to the main buffer passed to the process function and then process
    // a new block.
    main_input_ring_buffers: Vec<Vec<f32>>,
    main_output_ring_buffers: Vec<Vec<f32>>,
    sidechain_ring_buffers: [Vec<Vec<f32>>; NUM_SIDECHAIN_INPUTS],

    /// Results from the ring buffers are copied to this scratch buffer before being passed to the
    /// plugin. Needed to handle overlap.
    scratch_buffer: Vec<f32>,

    /// The current position in our ring buffers. Whenever this wraps around to 0, we'll process
    /// a block.
    current_pos: usize,
}

impl<const NUM_SIDECHAIN_INPUTS: usize> StftHelper<NUM_SIDECHAIN_INPUTS> {
    /// Initialize the [`StftHelper`] for [`Buffer`]s with the specified number of channels and the
    /// given maximum block size. Call [`set_block_size()`][`Self::set_block_size()`] afterwards if
    /// you do not need the full capacity upfront.
    ///
    /// # Panics
    ///
    /// Panics if `num_channels == 0 || max_block_size == 0`.
    pub fn new(num_channels: usize, max_block_size: usize) -> Self {
        assert_ne!(num_channels, 0);
        assert_ne!(max_block_size, 0);

        Self {
            main_input_ring_buffers: vec![vec![0.0; max_block_size]; num_channels],
            main_output_ring_buffers: vec![vec![0.0; max_block_size]; num_channels],
            // Kinda hacky way to initialize an array of non-copy types
            sidechain_ring_buffers: [(); NUM_SIDECHAIN_INPUTS]
                .map(|_| vec![vec![0.0; max_block_size]; num_channels]),

            scratch_buffer: vec![0.0; max_block_size],

            current_pos: 0,
        }
    }

    /// Change the current block size. This will clear the buffers, causing the next block to output
    /// silence.
    ///
    /// # Panics
    ///
    /// WIll panic if `block_size > max_block_size`.
    pub fn set_block_size(&mut self, block_size: usize) {
        assert!(block_size <= self.main_input_ring_buffers[0].capacity());

        for main_ring_buffer in &mut self.main_input_ring_buffers {
            main_ring_buffer.resize(block_size, 0.0);
            main_ring_buffer.fill(0.0);
        }
        for main_ring_buffer in &mut self.main_output_ring_buffers {
            main_ring_buffer.resize(block_size, 0.0);
            main_ring_buffer.fill(0.0);
        }
        self.scratch_buffer.resize(block_size, 0.0);
        self.scratch_buffer.fill(0.0);
        for sidechain_ring_buffers in &mut self.sidechain_ring_buffers {
            for sidechain_ring_buffer in sidechain_ring_buffers {
                sidechain_ring_buffer.resize(block_size, 0.0);
                sidechain_ring_buffer.fill(0.0);
            }
        }

        self.current_pos = 0;
    }

    /// The amount of latency introduced when processing audio throug hthis [`StftHelper`].
    pub fn latency_samples(&self) -> u32 {
        self.main_input_ring_buffers[0].len() as u32
    }

    /// Process the audio in `main_buffer` and in any sidechain buffers in small overlapping blocks
    /// with a window function applied, adding up the results for the main buffer so they can be
    /// written back to the host. The window overlap amount is compensated automatically when adding
    /// up these samples. Whenever a new block is available, `process_cb()` gets called with a new
    /// audio block of the specified size with the windowing function already applied. The summed
    /// reults will then be written back to `main_buffer` exactly one block later, which means that
    /// this function will introduce one block of latency. This can be compensated by calling
    /// [`ProcessContext::set_latency()`][`crate::prelude::ProcessContext::set_latency()`] in your
    /// plugin's initialization function.
    ///
    /// For efficiency's sake this function will reuse the same vector for all calls to
    /// `process_cb`. This means you can only access a single channel's worth of windowed data at a
    /// time. The arguments to that function are `process_cb(channel_idx, sidechain_buffer_idx,
    /// data)`, where `sidechain_buffer_idx` will be `None` for the main buffer. If there are any
    /// sidechain buffers, then they will be processed before the main buffer.
    ///
    /// # Panics
    ///
    /// Panics if `main_buffer` or the buffers in `sidechain_buffers` do not have the same number of
    /// channels as this [`StftHelper`], if the sidechain buffers do not contain the same number of
    /// samples as the main buffer, or if the window function does not match the block size.
    ///
    /// TODO: Maybe introduce a trait here so this can be used with things that aren't whole buffers
    /// TODO: And also introduce that aforementioned read-only process function (`analyze()?`)
    pub fn process_overlap_add<F>(
        &mut self,
        main_buffer: &mut Buffer,
        sidechain_buffers: [&Buffer; NUM_SIDECHAIN_INPUTS],
        window_function: &[f32],
        overlap_times: usize,
        overlap_gain_compensation: f32,
        mut process_cb: F,
    ) where
        F: FnMut(usize, Option<usize>, &mut [f32]),
    {
        assert_eq!(main_buffer.channels(), self.main_input_ring_buffers.len());
        assert_eq!(window_function.len(), self.main_input_ring_buffers[0].len());
        assert!(overlap_times > 0);

        // We'll copy samples from `*_buffer` into `*_ring_buffers` while simultaneously copying
        // already processed samples from `main_ring_buffers` in into `main_buffer`
        let main_buffer_len = main_buffer.len();
        let num_channels = main_buffer.channels();
        let block_size = self.main_input_ring_buffers[0].len();
        let window_interval = (block_size / overlap_times) as i32;
        let mut already_processed_samples = 0;
        while already_processed_samples < main_buffer_len {
            let remaining_samples = main_buffer_len - already_processed_samples;
            let samples_until_next_window = ((window_interval - self.current_pos as i32 - 1)
                .rem_euclid(window_interval)
                + 1) as usize;
            let samples_to_process = samples_until_next_window.min(remaining_samples);

            // Copy the input from `main_buffer` to the ring buffer while copying last block's
            // result from the buffer to `main_buffer`
            // TODO: This might be able to be sped up a bit with SIMD
            {
                // For the main buffer
                let main_buffer = main_buffer.as_slice();
                for sample_offset in 0..samples_to_process {
                    for channel_idx in 0..num_channels {
                        let sample = unsafe {
                            main_buffer
                                .get_unchecked_mut(channel_idx)
                                .get_unchecked_mut(already_processed_samples + sample_offset)
                        };
                        let input_ring_buffer_sample = unsafe {
                            self.main_input_ring_buffers
                                .get_unchecked_mut(channel_idx)
                                .get_unchecked_mut(self.current_pos + sample_offset)
                        };
                        let output_ring_buffer_sample = unsafe {
                            self.main_output_ring_buffers
                                .get_unchecked_mut(channel_idx)
                                .get_unchecked_mut(self.current_pos + sample_offset)
                        };
                        *input_ring_buffer_sample = *sample;
                        *sample = *output_ring_buffer_sample;
                        // Very important, or else we'll overlap-add ourselves into a feedback hell
                        *output_ring_buffer_sample = 0.0;
                    }
                }

                // And for the sidechain buffers we only need to copy the inputs
                for (sidechain_buffer, sidechain_ring_buffers) in sidechain_buffers
                    .iter()
                    .zip(self.sidechain_ring_buffers.iter_mut())
                {
                    let sidechain_buffer = sidechain_buffer.as_slice_immutable();
                    for sample_offset in 0..samples_to_process {
                        for channel_idx in 0..num_channels {
                            let sample = unsafe {
                                sidechain_buffer
                                    .get_unchecked(channel_idx)
                                    .get_unchecked(already_processed_samples + sample_offset)
                            };
                            let ring_buffer_sample = unsafe {
                                sidechain_ring_buffers
                                    .get_unchecked_mut(channel_idx)
                                    .get_unchecked_mut(self.current_pos + sample_offset)
                            };
                            *ring_buffer_sample = *sample;
                        }
                    }
                }
            }

            // At this point we either have `already_processed_samples == main_buffer_len`, or
            // `self.current_pos % window_interval == 0`. If it's the latter, then we can process a
            // new block.
            if samples_to_process == samples_until_next_window {
                // Because we're processing in smaller windows, the input ring buffers sadly does
                // not always contain the full contiguous range we're interested in because they map
                // wrap around. Because premade FFT algorithms typically can't handle this, we'll
                // start with copying

                // TODO: Sdiechain

                for (channel_idx, (input_ring_buffer, output_ring_buffer)) in self
                    .main_input_ring_buffers
                    .iter()
                    .zip(self.main_output_ring_buffers.iter_mut())
                    .enumerate()
                {
                    copy_ring_to_scratch_buffer(
                        &mut self.scratch_buffer,
                        self.current_pos,
                        input_ring_buffer,
                    );
                    multiply_scratch_buffer(&mut self.scratch_buffer, window_function);
                    process_cb(channel_idx, None, &mut self.scratch_buffer);

                    // The actual overlap-add part of the equation
                    add_scratch_to_ring_buffer(
                        &self.scratch_buffer,
                        self.current_pos,
                        output_ring_buffer,
                        overlap_gain_compensation,
                    );
                }
            }

            // Do this after handling the block or else we'll copy the wrong samples.
            already_processed_samples += samples_to_process;
            self.current_pos = (self.current_pos + samples_to_process) % block_size;
        }
    }
}

/// Copy data from the the specified ring buffer (borrowed from `self`) to the scratch buffers at
/// the current position. This is a free function because you cannot pass an immutable reference to
/// a field from `&self` to a `&mut self` method.
#[inline]
fn copy_ring_to_scratch_buffer(
    scratch_buffer: &mut [f32],
    current_pos: usize,
    ring_buffer: &[f32],
) {
    let block_size = scratch_buffer.len();
    let num_copy_before_wrap = block_size - current_pos;
    scratch_buffer[0..num_copy_before_wrap].copy_from_slice(&ring_buffer[current_pos..block_size]);
    scratch_buffer[num_copy_before_wrap..block_size].copy_from_slice(&ring_buffer[0..current_pos]);
}

/// Multiply the scratch buffer by some window function. Also free function because you can't do
/// split borrows with methods.
#[inline]
fn multiply_scratch_buffer(scratch_buffer: &mut [f32], window_function: &[f32]) {
    for (sample, window_sample) in scratch_buffer.iter_mut().zip(window_function) {
        *sample *= window_sample;
    }
}

/// Add data from the scratch buffer to the specified ring buffer. When writing samples from this
/// ring buffer back to the host's outputs they must be cleared to prevent infinite feedback.
#[inline]
fn add_scratch_to_ring_buffer(
    scratch_buffer: &[f32],
    current_pos: usize,
    ring_buffer: &mut [f32],
    gain_compensation: f32,
) {
    // TODO: This could also use some SIMD
    let block_size = scratch_buffer.len();
    let num_copy_before_wrap = block_size - current_pos;
    for (scratch_sample, ring_sample) in scratch_buffer[0..num_copy_before_wrap]
        .iter()
        .zip(&mut ring_buffer[current_pos..block_size])
    {
        // TODO: Moving this gain compensation to the window is more efficient, but that makes the
        //       interface less nice to work with
        *ring_sample += *scratch_sample * gain_compensation;
    }
    for (scratch_sample, ring_sample) in scratch_buffer[num_copy_before_wrap..block_size]
        .iter()
        .zip(&mut ring_buffer[0..current_pos])
    {
        *ring_sample += *scratch_sample * gain_compensation;
    }
}