1
0
Fork 0
nih-plug/src/buffer.rs

656 lines
23 KiB
Rust
Raw Normal View History

use std::marker::PhantomData;
#[cfg(feature = "simd")]
use std::simd::{LaneCount, Simd, SupportedLaneCount};
// TODO: Does adding `#[inline]` to the .next() functions make any difference?
/// The audio buffers used during processing. This contains the output audio output buffers with the
/// inputs already copied to the outputs. You can either use the iterator adapters to conveniently
/// and efficiently iterate over the samples, or you can do your own thing using the raw audio
/// buffers.
#[derive(Default)]
pub struct Buffer<'a> {
/// Contains slices for the plugin's outputs. You can't directly create a nested slice form
/// apointer to pointers, so this needs to be preallocated in the setup call and kept around
/// between process calls. And because storing a reference to this means a) that you need a lot
/// of lifetime annotations everywhere and b) that at some point you need unsound lifetime casts
/// because this `Buffers` either cannot have the same lifetime as the separately stored output
/// buffers, and it also cannot be stored in a field next to it because that would mean
/// containing mutable references to data stored in a mutex.
output_slices: Vec<&'a mut [f32]>,
}
2022-03-02 02:03:13 +11:00
// Per-sample per-channel iterators
/// An iterator over all samples in the buffer, yielding iterators over each channel for every
/// sample. This iteration order offers good cache locality for per-sample access.
pub struct SamplesIter<'slice, 'sample: 'slice> {
/// The raw output buffers.
pub(self) buffers: *mut [&'sample mut [f32]],
pub(self) current_sample: usize,
pub(self) _marker: PhantomData<&'slice mut [&'sample mut [f32]]>,
}
/// Can construct iterators over actual iterator over the channel data for a sample, yielded by
2022-03-02 02:03:13 +11:00
/// [Samples]. Can be turned into an iterator, or [Channels::iter_mut()] can be used to iterate over
/// the channel data multiple times, or more efficiently you can use [Channels::get_unchecked_mut()]
/// to do the same thing.
pub struct Channels<'slice, 'sample: 'slice> {
/// The raw output buffers.
pub(self) buffers: *mut [&'sample mut [f32]],
pub(self) current_sample: usize,
pub(self) _marker: PhantomData<&'slice mut [&'sample mut [f32]]>,
}
/// The actual iterator over the channel data for a sample, yielded by [Channels].
pub struct ChannelsIter<'slice, 'sample: 'slice> {
/// The raw output buffers.
2022-02-13 11:48:56 +11:00
pub(self) buffers: *mut [&'sample mut [f32]],
pub(self) current_sample: usize,
pub(self) current_channel: usize,
2022-02-13 11:58:22 +11:00
pub(self) _marker: PhantomData<&'slice mut [&'sample mut [f32]]>,
}
2022-03-02 02:03:13 +11:00
// Per-block per-channel per-sample iterators
/// An iterator over all samples in the buffer, slicing over the sample-dimension with a maximum
2022-03-02 03:46:34 +11:00
/// size of [Self::max_block_size]. See [Buffer::iter_blocks()]. Yields both the block and the
/// offset from the start of the buffer.
2022-03-02 02:03:13 +11:00
pub struct BlocksIter<'slice, 'sample: 'slice> {
/// The raw output buffers.
pub(self) buffers: *mut [&'sample mut [f32]],
pub(self) max_block_size: usize,
pub(self) current_block_start: usize,
pub(self) _marker: PhantomData<&'slice mut [&'sample mut [f32]]>,
}
/// A block yielded by [BlocksIter]. Can be iterated over once or multiple times, and also supports
/// direct access to the block's samples if needed.
pub struct Block<'slice, 'sample: 'slice> {
/// The raw output buffers.
pub(self) buffers: *mut [&'sample mut [f32]],
pub(self) current_block_start: usize,
pub(self) current_block_end: usize,
pub(self) _marker: PhantomData<&'slice mut [&'sample mut [f32]]>,
}
/// An iterator over all channels in a block yielded by [Block]. Analogous to [ChannelsIter] but for
/// blocks.
pub struct BlockChannelsIter<'slice, 'sample: 'slice> {
/// The raw output buffers.
pub(self) buffers: *mut [&'sample mut [f32]],
pub(self) current_block_start: usize,
pub(self) current_block_end: usize,
pub(self) current_channel: usize,
pub(self) _marker: PhantomData<&'slice mut [&'sample mut [f32]]>,
}
impl<'slice, 'sample> Iterator for SamplesIter<'slice, 'sample> {
2022-02-13 11:58:22 +11:00
type Item = Channels<'slice, 'sample>;
fn next(&mut self) -> Option<Self::Item> {
if self.current_sample < unsafe { (*self.buffers)[0].len() } {
let channels = Channels {
buffers: self.buffers,
current_sample: self.current_sample,
_marker: self._marker,
};
self.current_sample += 1;
Some(channels)
} else {
None
}
}
fn size_hint(&self) -> (usize, Option<usize>) {
let remaining = unsafe { (*self.buffers)[0].len() } - self.current_sample;
(remaining, Some(remaining))
}
}
2022-03-02 02:03:13 +11:00
impl<'slice, 'sample> Iterator for BlockChannelsIter<'slice, 'sample> {
type Item = &'sample mut [f32];
fn next(&mut self) -> Option<Self::Item> {
if self.current_channel < unsafe { (*self.buffers).len() } {
// SAFETY: These bounds have already been checked
// SAFETY: It is also not possible to have multiple mutable references to the same
// sample at the same time
let slice = unsafe {
(*self.buffers)
.get_unchecked_mut(self.current_channel)
.get_unchecked_mut(self.current_block_start..self.current_block_end)
};
self.current_channel += 1;
Some(slice)
} else {
None
}
}
fn size_hint(&self) -> (usize, Option<usize>) {
let remaining = unsafe { (*self.buffers).len() } - self.current_channel;
(remaining, Some(remaining))
}
}
impl<'slice, 'sample> Iterator for BlocksIter<'slice, 'sample> {
2022-03-02 03:46:34 +11:00
type Item = (usize, Block<'slice, 'sample>);
2022-03-02 02:03:13 +11:00
fn next(&mut self) -> Option<Self::Item> {
let buffer_len = unsafe { (*self.buffers)[0].len() };
if self.current_block_start < buffer_len {
2022-03-02 03:46:34 +11:00
let current_block_start = self.current_block_start;
2022-03-02 02:03:13 +11:00
let current_block_end =
(self.current_block_start + self.max_block_size).min(buffer_len);
let block = Block {
buffers: self.buffers,
2022-03-02 03:46:34 +11:00
current_block_start,
2022-03-02 02:03:13 +11:00
current_block_end,
_marker: self._marker,
};
self.current_block_start += self.max_block_size;
2022-03-02 03:46:34 +11:00
Some((current_block_start, block))
2022-03-02 02:03:13 +11:00
} else {
None
}
}
fn size_hint(&self) -> (usize, Option<usize>) {
let remaining = ((unsafe { (*self.buffers)[0].len() } - self.current_block_start) as f32
/ self.max_block_size as f32)
.ceil() as usize;
(remaining, Some(remaining))
}
}
2022-02-13 11:58:22 +11:00
impl<'slice, 'sample> Iterator for ChannelsIter<'slice, 'sample> {
2022-02-13 11:48:56 +11:00
type Item = &'sample mut f32;
fn next(&mut self) -> Option<Self::Item> {
if self.current_channel < unsafe { (*self.buffers).len() } {
// SAFETY: These bounds have already been checked
2022-02-13 11:58:22 +11:00
// SAFETY: It is also not possible to have multiple mutable references to the same
// sample at the same time
let sample = unsafe {
(*self.buffers)
.get_unchecked_mut(self.current_channel)
.get_unchecked_mut(self.current_sample)
};
self.current_channel += 1;
Some(sample)
} else {
None
}
}
fn size_hint(&self) -> (usize, Option<usize>) {
let remaining = unsafe { (*self.buffers).len() } - self.current_channel;
(remaining, Some(remaining))
}
}
impl<'slice, 'sample> IntoIterator for Channels<'slice, 'sample> {
type Item = &'sample mut f32;
type IntoIter = ChannelsIter<'slice, 'sample>;
fn into_iter(self) -> Self::IntoIter {
ChannelsIter {
buffers: self.buffers,
current_sample: self.current_sample,
current_channel: 0,
_marker: self._marker,
}
}
}
2022-03-02 02:03:13 +11:00
impl<'slice, 'sample> IntoIterator for Block<'slice, 'sample> {
type Item = &'sample mut [f32];
type IntoIter = BlockChannelsIter<'slice, 'sample>;
fn into_iter(self) -> Self::IntoIter {
BlockChannelsIter {
buffers: self.buffers,
current_block_start: self.current_block_start,
current_block_end: self.current_block_end,
current_channel: 0,
_marker: self._marker,
}
}
}
2022-03-02 02:03:13 +11:00
impl ExactSizeIterator for SamplesIter<'_, '_> {}
impl ExactSizeIterator for ChannelsIter<'_, '_> {}
2022-03-02 02:03:13 +11:00
impl ExactSizeIterator for BlocksIter<'_, '_> {}
impl ExactSizeIterator for BlockChannelsIter<'_, '_> {}
impl<'a> Buffer<'a> {
/// Returns true if this buffer does not contain any samples.
pub fn is_empty(&self) -> bool {
self.output_slices.is_empty() || self.output_slices[0].is_empty()
}
/// Obtain the raw audio buffers.
pub fn as_slice(&mut self) -> &mut [&'a mut [f32]] {
&mut self.output_slices
}
/// Iterate over the samples, returning a channel iterator for each sample.
pub fn iter_mut<'slice>(&'slice mut self) -> SamplesIter<'slice, 'a> {
SamplesIter {
buffers: self.output_slices.as_mut_slice(),
current_sample: 0,
_marker: PhantomData,
}
}
2022-03-02 02:03:13 +11:00
/// Iterate over the buffer in blocks with the specified maximum size. The ideal maximum block
/// size depends on the plugin in question, but 64 or 128 samples works for most plugins. Since
/// the buffer's total size may not be cleanly divisble by the maximum size, the returned
/// buffers may have any size in `[1, max_block_size]`. This is useful when using algorithms
/// that work on entire blocks of audio, like those that would otherwise need to perform
/// expensive per-sample branching or that can use per-sample SIMD as opposed to per-channel
/// SIMD.
///
/// The parameter smoothers can also produce smoothed values for an entire block using
/// [crate::Smoother::next_block()]. Before using this, you will need to call
/// [crate::Plugin::initialize_block_smoothers()] with the same `max_block_size` in your
/// initialization function first.
///
/// You can use this to obtain block-slices from a buffer so you can pass them to a libraryq:
///
/// ```ignore
/// for block in buffer.iter_blocks(128) {
/// let mut block_channels = block.into_iter();
/// let stereo_slice = &[
/// block_channels.next().unwrap(),
/// block_channels.next().unwrap(),
/// ];
///
/// // Do something cool with `stereo_slice`
/// }
/// ````
2022-03-02 02:03:13 +11:00
pub fn iter_blocks<'slice>(&'slice mut self, max_block_size: usize) -> BlocksIter<'slice, 'a> {
BlocksIter {
buffers: self.output_slices.as_mut_slice(),
max_block_size,
current_block_start: 0,
_marker: PhantomData,
}
}
/// Access the raw output slice vector. This needs to be resized to match the number of output
/// channels during the plugin's initialization. Then during audio processing, these slices
/// should be updated to point to the plugin's audio buffers.
///
/// # Safety
///
/// The stored slices must point to live data when this object is passed to the plugins' process
/// function. The rest of this object also assumes all channel lengths are equal. Panics will
/// likely occur if this is not the case.
2022-03-02 01:00:32 +11:00
pub unsafe fn with_raw_vec(&mut self, update: impl FnOnce(&mut Vec<&'a mut [f32]>)) {
update(&mut self.output_slices);
}
}
2022-02-13 11:58:22 +11:00
impl<'slice, 'sample> Channels<'slice, 'sample> {
/// Get the number of channels.
pub fn len(&self) -> usize {
unsafe { (*self.buffers).len() }
}
/// A resetting iterator. This lets you iterate over the same channels multiple times. Otherwise
/// you don't need to use this function as [Channels] already implements [Iterator].
2022-02-13 11:58:22 +11:00
pub fn iter_mut(&mut self) -> ChannelsIter<'slice, 'sample> {
ChannelsIter {
2022-02-13 11:58:22 +11:00
buffers: self.buffers,
current_sample: self.current_sample,
current_channel: 0,
2022-03-02 02:03:13 +11:00
_marker: self._marker,
}
}
2022-03-02 02:03:13 +11:00
/// Access a sample by index. Useful when you would otherwise iterate over this 'Channels'
/// iterator multiple times.
#[inline]
2022-02-13 06:28:58 +11:00
pub fn get_mut(&mut self, channel_index: usize) -> Option<&mut f32> {
2022-03-02 02:03:13 +11:00
// SAFETY: The sample bound has already been checked
unsafe {
2022-02-13 06:28:58 +11:00
Some(
(*self.buffers)
2022-02-13 06:28:58 +11:00
.get_mut(channel_index)?
.get_unchecked_mut(self.current_sample),
)
}
}
/// The same as [Self::get_mut], but without any bounds checking.
2022-02-13 06:28:58 +11:00
///
/// # Safety
///
/// `channel_index` must be in the range `0..Self::len()`.
#[inline]
2022-02-13 06:28:58 +11:00
pub unsafe fn get_unchecked_mut(&mut self, channel_index: usize) -> &mut f32 {
(*self.buffers)
2022-02-13 06:28:58 +11:00
.get_unchecked_mut(channel_index)
.get_unchecked_mut(self.current_sample)
}
/// Get a SIMD vector containing the channel data for this buffer. If `LANES > channels.len()`
/// then this will be padded with zeroes. If `LANES < channels.len()` then this won't contain
/// all values.
#[cfg(feature = "simd")]
#[inline]
pub fn to_simd<const LANES: usize>(&self) -> Simd<f32, LANES>
where
LaneCount<LANES>: SupportedLaneCount,
{
let used_lanes = self.len().max(LANES);
let mut values = [0.0; LANES];
for (channel_idx, value) in values.iter_mut().enumerate().take(used_lanes) {
*value = unsafe {
*(*self.buffers)
.get_unchecked(channel_idx)
.get_unchecked(self.current_sample)
};
}
Simd::from_array(values)
}
/// Get a SIMD vector containing the channel data for this buffer. Will always read exactly
/// `LANES` channels.
///
/// # Safety
///
/// Undefined behavior if `LANES > channels.len()`.
#[cfg(feature = "simd")]
#[inline]
pub unsafe fn to_simd_unchecked<const LANES: usize>(&self) -> Simd<f32, LANES>
where
LaneCount<LANES>: SupportedLaneCount,
{
let mut values = [0.0; LANES];
for (channel_idx, value) in values.iter_mut().enumerate() {
*value = *(*self.buffers)
.get_unchecked(channel_idx)
.get_unchecked(self.current_sample);
}
Simd::from_array(values)
}
/// Write data from a SIMD vector to this sample's channel data. This takes the padding added by
/// [Self::to_simd()] into account.
#[cfg(feature = "simd")]
#[allow(clippy::wrong_self_convention)]
#[inline]
pub fn from_simd<const LANES: usize>(&mut self, vector: Simd<f32, LANES>)
where
LaneCount<LANES>: SupportedLaneCount,
{
let used_lanes = self.len().max(LANES);
let values = vector.to_array();
for (channel_idx, value) in values.into_iter().enumerate().take(used_lanes) {
*unsafe {
(*self.buffers)
.get_unchecked_mut(channel_idx)
.get_unchecked_mut(self.current_sample)
} = value;
}
}
/// Write data from a SIMD vector to this sample's channel data. This assumes `LANES` matches
/// exactly with the number of channels in the buffer.
///
/// # Safety
///
/// Undefined behavior if `LANES > channels.len()`.
#[cfg(feature = "simd")]
#[allow(clippy::wrong_self_convention)]
#[inline]
pub unsafe fn from_simd_unchecked<const LANES: usize>(&mut self, vector: Simd<f32, LANES>)
where
LaneCount<LANES>: SupportedLaneCount,
{
let values = vector.to_array();
for (channel_idx, value) in values.into_iter().enumerate() {
*(*self.buffers)
.get_unchecked_mut(channel_idx)
.get_unchecked_mut(self.current_sample) = value;
}
}
}
2022-02-13 12:20:31 +11:00
2022-03-02 02:03:13 +11:00
impl<'slice, 'sample> Block<'slice, 'sample> {
2022-03-02 03:53:32 +11:00
/// Get the number of samples (not channels) in the block.
2022-03-02 02:03:13 +11:00
pub fn len(&self) -> usize {
2022-03-02 03:53:32 +11:00
self.current_block_end - self.current_block_start
2022-03-02 02:03:13 +11:00
}
/// A resetting iterator. This lets you iterate over the same block multiple times. Otherwise
/// you don't need to use this function as [Block] already implements [Iterator]. You can also
/// use the direct accessor functions on this block instead.
pub fn iter_mut(&mut self) -> BlockChannelsIter<'slice, 'sample> {
BlockChannelsIter {
buffers: self.buffers,
current_block_start: self.current_block_start,
current_block_end: self.current_block_end,
current_channel: 0,
_marker: self._marker,
}
}
/// Access a channel by index. Useful when you would otherwise iterate over this [Block]
/// multiple times.
#[inline]
pub fn get_mut(&mut self, channel_index: usize) -> Option<&mut [f32]> {
// SAFETY: The block bound has already been checked
unsafe {
Some(
(*self.buffers)
.get_mut(channel_index)?
.get_unchecked_mut(self.current_block_start..self.current_block_end),
)
}
}
/// The same as [Self::get_mut], but without any bounds checking.
///
/// # Safety
///
/// `channel_index` must be in the range `0..Self::len()`.
#[inline]
pub unsafe fn get_unchecked_mut(&mut self, channel_index: usize) -> &mut [f32] {
(*self.buffers)
.get_unchecked_mut(channel_index)
.get_unchecked_mut(self.current_block_start..self.current_block_end)
}
/// Get a SIMD vector containing the channel data for a specific sample in this block. If `LANES
/// > channels.len()` then this will be padded with zeroes. If `LANES < channels.len()` then
/// this won't contain all values.
///
/// Returns a `None` value if `sample_index` is out of bounds.
#[cfg(feature = "simd")]
#[inline]
pub fn to_simd<const LANES: usize>(&self, sample_index: usize) -> Option<Simd<f32, LANES>>
where
LaneCount<LANES>: SupportedLaneCount,
{
if sample_index > self.len() {
return None;
}
let used_lanes = self.len().max(LANES);
let mut values = [0.0; LANES];
for (channel_idx, value) in values.iter_mut().enumerate().take(used_lanes) {
*value = unsafe {
*(*self.buffers)
.get_unchecked(channel_idx)
.get_unchecked(self.current_block_start + sample_index)
};
}
Some(Simd::from_array(values))
}
/// Get a SIMD vector containing the channel data for a specific sample in this block. Will
/// always read exactly `LANES` channels, and does not perform bounds checks on `sample_index`.
///
/// # Safety
///
/// Undefined behavior if `LANES > block.len()` or if `sample_index > block.len()`.
#[cfg(feature = "simd")]
#[inline]
pub unsafe fn to_simd_unchecked<const LANES: usize>(
&self,
sample_index: usize,
) -> Simd<f32, LANES>
where
LaneCount<LANES>: SupportedLaneCount,
{
let mut values = [0.0; LANES];
for (channel_idx, value) in values.iter_mut().enumerate() {
*value = *(*self.buffers)
.get_unchecked(channel_idx)
.get_unchecked(self.current_block_start + sample_index);
}
Simd::from_array(values)
}
/// Write data from a SIMD vector to this sample's channel data for a specific sample in this
/// block. This takes the padding added by [Self::to_simd()] into account.
///
/// Returns `false` if `sample_index` is out of bounds.
#[cfg(feature = "simd")]
#[allow(clippy::wrong_self_convention)]
#[inline]
pub fn from_simd<const LANES: usize>(
&mut self,
sample_index: usize,
vector: Simd<f32, LANES>,
) -> bool
where
LaneCount<LANES>: SupportedLaneCount,
{
if sample_index > self.len() {
return false;
}
let used_lanes = self.len().max(LANES);
let values = vector.to_array();
for (channel_idx, value) in values.into_iter().enumerate().take(used_lanes) {
*unsafe {
(*self.buffers)
.get_unchecked_mut(channel_idx)
.get_unchecked_mut(self.current_block_start + sample_index)
} = value;
}
true
}
/// Write data from a SIMD vector to this sample's channel data for a specific sample in this
/// block.. This assumes `LANES` matches exactly with the number of channels in the buffer, and
/// does not perform bounds checks on `sample_index`.
///
/// # Safety
///
/// Undefined behavior if `LANES > block.len()` or if `sample_index > block.len()`.
#[cfg(feature = "simd")]
#[allow(clippy::wrong_self_convention)]
#[inline]
pub unsafe fn from_simd_unchecked<const LANES: usize>(
&mut self,
sample_index: usize,
vector: Simd<f32, LANES>,
) where
LaneCount<LANES>: SupportedLaneCount,
{
let values = vector.to_array();
for (channel_idx, value) in values.into_iter().enumerate() {
*(*self.buffers)
.get_unchecked_mut(channel_idx)
.get_unchecked_mut(self.current_block_start + sample_index) = value;
}
}
2022-03-02 02:03:13 +11:00
}
#[cfg(any(miri, test))]
2022-02-13 12:20:31 +11:00
mod miri {
use super::*;
#[test]
fn repeated_access() {
let mut real_buffers = vec![vec![0.0; 512]; 2];
let mut buffer = Buffer::default();
2022-03-02 01:00:32 +11:00
unsafe {
buffer.with_raw_vec(|output_slices| {
let (first_channel, other_channels) = real_buffers.split_at_mut(1);
*output_slices = vec![&mut first_channel[0], &mut other_channels[0]];
})
};
2022-02-13 12:20:31 +11:00
for samples in buffer.iter_mut() {
for sample in samples {
*sample += 0.001;
}
}
for mut samples in buffer.iter_mut() {
for _ in 0..2 {
2022-02-13 12:20:31 +11:00
for sample in samples.iter_mut() {
*sample += 0.001;
}
}
}
assert_eq!(real_buffers[0][0], 0.003);
2022-02-13 12:20:31 +11:00
}
2022-03-02 02:03:13 +11:00
#[test]
fn repeated_slices() {
let mut real_buffers = vec![vec![0.0; 512]; 2];
let mut buffer = Buffer::default();
unsafe {
buffer.with_raw_vec(|output_slices| {
let (first_channel, other_channels) = real_buffers.split_at_mut(1);
*output_slices = vec![&mut first_channel[0], &mut other_channels[0]];
})
};
// These iterators should not alias
let mut blocks = buffer.iter_blocks(16);
2022-03-02 03:46:34 +11:00
let (_block1_offset, block1) = blocks.next().unwrap();
let (_block2_offset, block2) = blocks.next().unwrap();
2022-03-02 02:03:13 +11:00
for channel in block1 {
for sample in channel.iter_mut() {
*sample += 0.001;
}
}
for channel in block2 {
for sample in channel.iter_mut() {
*sample += 0.001;
}
}
for i in 0..32 {
assert_eq!(real_buffers[0][i], 0.001);
}
for i in 32..48 {
assert_eq!(real_buffers[0][i], 0.0);
}
}
2022-02-13 12:20:31 +11:00
}