1
0
Fork 0

Add a non-functional STFT to Spectral Compressor

This commit is contained in:
Robbert van der Helm 2022-07-14 15:23:43 +02:00
parent 85e52b8ab7
commit 12aa8544cd
4 changed files with 236 additions and 5 deletions

1
Cargo.lock generated
View file

@ -3623,6 +3623,7 @@ version = "0.1.0"
dependencies = [
"nih_plug",
"nih_plug_vizia",
"realfft",
]
[[package]]

View file

@ -11,3 +11,5 @@ crate-type = ["cdylib"]
[dependencies]
nih_plug = { path = "../../", features = ["assert_process_allocs"] }
nih_plug_vizia = { path = "../../nih_plug_vizia" }
realfft = "3.0"

View file

@ -16,23 +16,124 @@
use nih_plug::prelude::*;
use nih_plug_vizia::ViziaState;
use realfft::num_complex::Complex32;
use realfft::{ComplexToReal, RealFftPlanner, RealToComplex};
use std::sync::Arc;
mod editor;
const MIN_WINDOW_ORDER: usize = 6;
#[allow(dead_code)]
const MIN_WINDOW_SIZE: usize = 1 << MIN_WINDOW_ORDER; // 64
const DEFAULT_WINDOW_ORDER: usize = 12;
#[allow(dead_code)]
const DEFAULT_WINDOW_SIZE: usize = 1 << DEFAULT_WINDOW_ORDER; // 4096
const MAX_WINDOW_ORDER: usize = 15;
const MAX_WINDOW_SIZE: usize = 1 << MAX_WINDOW_ORDER; // 32768
const MIN_OVERLAP_ORDER: usize = 2;
#[allow(dead_code)]
const MIN_OVERLAP_TIMES: usize = 2 << MIN_OVERLAP_ORDER; // 4
const DEFAULT_OVERLAP_ORDER: usize = 3;
#[allow(dead_code)]
const DEFAULT_OVERLAP_TIMES: usize = 1 << DEFAULT_OVERLAP_ORDER; // 4
const MAX_OVERLAP_ORDER: usize = 5;
#[allow(dead_code)]
const MAX_OVERLAP_TIMES: usize = 1 << MAX_OVERLAP_ORDER; // 32
/// This is a port of <https://github.com/robbert-vdh/spectral-compressor/>.
struct SpectralCompressor {
params: Arc<SpectralCompressorParams>,
editor_state: Arc<ViziaState>,
/// An adapter that performs most of the overlap-add algorithm for us.
stft: util::StftHelper,
/// Contains a Hann window function of the current window length, passed to the overlap-add
/// helper. Allocated with a `MAX_WINDOW_SIZE` initial capacity.
window_function: Vec<f32>,
/// The algorithms for the FFT and IFFT operations, for each supported order so we can switch
/// between them without replanning or allocations. Initialized during `initialize()`.
plan_for_order: Option<[Plan; MAX_WINDOW_ORDER - MIN_WINDOW_ORDER + 1]>,
/// The output of our real->complex FFT.
complex_fft_buffer: Vec<Complex32>,
}
#[derive(Params, Default)]
struct SpectralCompressorParams {}
/// An FFT plan for a specific window size, all of which will be precomputed during initilaization.
struct Plan {
/// The algorithm for the FFT operation.
r2c_plan: Arc<dyn RealToComplex<f32>>,
/// The algorithm for the IFFT operation.
c2r_plan: Arc<dyn ComplexToReal<f32>>,
}
#[derive(Params)]
struct SpectralCompressorParams {
/// Gain applied just before the DFT as part of the STFT process.
#[id = "input_db"]
input_gain_db: FloatParam,
/// Makeup gain applied after the IDFT in the STFT process. If automatic makeup gain is enabled,
/// then this acts as an offset on top of that.
#[id = "output_db"]
output_gain_db: FloatParam,
/// Try to automatically compensate for low thresholds. Doesn't do anything when sidechaining is
/// active.
#[id = "auto_makeup"]
auto_makeup_gain: BoolParam,
/// How much of the dry signal to mix in with the processed signal. The mixing is done after
/// applying the output gain. In other words, the dry signal is not gained in any way.
#[id = "dry_wet"]
dry_wet_ratio: FloatParam,
}
impl Default for SpectralCompressor {
fn default() -> Self {
Self {
params: Arc::new(SpectralCompressorParams::default()),
editor_state: editor::default_state(),
// These two will be set to the correct values in the initialize function
stft: util::StftHelper::new(Self::DEFAULT_NUM_OUTPUTS as usize, MAX_WINDOW_SIZE, 0),
window_function: Vec::with_capacity(MAX_WINDOW_SIZE),
// This is initialized later since we don't want to do non-trivial computations before
// the plugin is initialized
plan_for_order: None,
complex_fft_buffer: Vec::with_capacity(MAX_WINDOW_SIZE / 2 + 1),
}
}
}
impl Default for SpectralCompressorParams {
fn default() -> Self {
Self {
// We don't need any smoothing for these parameters as the overlap-add process will
// already act as a form of smoothing
input_gain_db: FloatParam::new(
"Input Gain",
0.0,
FloatRange::Linear {
min: -50.0,
max: 50.0,
},
)
.with_unit(" dB")
.with_step_size(0.1),
output_gain_db: FloatParam::new(
"Output Gain",
0.0,
FloatRange::Linear {
min: -50.0,
max: 50.0,
},
)
.with_unit(" dB")
.with_step_size(0.1),
auto_makeup_gain: BoolParam::new("Auto Makeup Gain", true),
dry_wet_ratio: FloatParam::new("Mix", 1.0, FloatRange::Linear { min: 0.0, max: 1.0 })
.with_unit("%")
.with_value_to_string(formatters::v2s_f32_percentage(0))
.with_string_to_value(formatters::s2v_f32_percentage()),
}
}
}
@ -63,17 +164,134 @@ impl Plugin for SpectralCompressor {
config.num_input_channels == config.num_output_channels && config.num_input_channels > 0
}
fn initialize(
&mut self,
bus_config: &BusConfig,
_buffer_config: &BufferConfig,
context: &mut impl InitContext,
) -> bool {
// This plugin can accept any number of channels, so we need to resize channel-dependent
// data structures accordinly
if self.stft.num_channels() != bus_config.num_output_channels as usize {
self.stft = util::StftHelper::new(self.stft.num_channels(), MAX_WINDOW_SIZE, 0);
}
// Planning with RustFFT is very fast, but it will still allocate we we'll plan all of the
// FFTs we might need in advance
if self.plan_for_order.is_none() {
let mut planner = RealFftPlanner::new();
let plan_for_order: Vec<Plan> = (MIN_WINDOW_ORDER..=MAX_WINDOW_ORDER)
.map(|order| Plan {
r2c_plan: planner.plan_fft_forward(1 << order),
c2r_plan: planner.plan_fft_inverse(1 << order),
})
.collect();
self.plan_for_order = Some(
plan_for_order
.try_into()
.unwrap_or_else(|_| panic!("Mismatched plan orders")),
);
}
// TODO: Fetch from a parameter
let window_size = DEFAULT_WINDOW_SIZE;
self.resize_for_window(window_size);
context.set_latency_samples(self.stft.latency_samples());
true
}
fn process(
&mut self,
_buffer: &mut Buffer,
buffer: &mut Buffer,
_aux: &mut AuxiliaryBuffers,
_context: &mut impl ProcessContext,
context: &mut impl ProcessContext,
) -> ProcessStatus {
// TODO: Do the thing
// If the window size has changed since the last process call, reset the buffers and chance
// our latency. All of these buffers already have enough capacity so this won't allocate.
// TODO: Fetch from a parameter
let overlap_times = DEFAULT_OVERLAP_TIMES;
// TODO: Fetch from a parameter
let window_size = DEFAULT_WINDOW_SIZE;
if self.window_function.len() != window_size {
self.resize_for_window(window_size);
context.set_latency_samples(self.stft.latency_samples());
}
// These plans have already been made during initialization we can switch between versions
// without reallocating
let fft_plan = &mut self.plan_for_order.as_mut().unwrap()
// FIXME: Use the parameter
// [self.params.window_size_order.value as usize - MIN_WINDOW_ORDER];
[DEFAULT_WINDOW_ORDER - MIN_WINDOW_ORDER];
// The overlap gain compensation is based on a squared Hann window, which will sum perfectly
// at four times overlap or higher. We'll apply a regular Hann window before the analysis
// and after the synthesis.
let gain_compensation: f32 =
((overlap_times as f32 / 4.0) * 1.5).recip() / window_size as f32;
// We'll apply the square root of the total gain compensation at the DFT and the IDFT
// stages. That way the compressor threshold values make much more sense.
let input_gain =
util::db_to_gain(self.params.input_gain_db.value) * gain_compensation.sqrt();
let output_gain =
util::db_to_gain(self.params.output_gain_db.value) * gain_compensation.sqrt();
// TODO: Mix in the dry signal
self.stft
.process_overlap_add(buffer, overlap_times, |_channel_idx, real_fft_buffer| {
// We'll window the input with a Hann function to avoid spectral leakage. The input
// gain here also contains a compensation factor for the forward FFT to make the
// compressor thresholds make more sense.
for (sample, window_sample) in real_fft_buffer.iter_mut().zip(&self.window_function)
{
*sample *= window_sample * input_gain;
}
// RustFFT doesn't actually need a scratch buffer here, so we'll pass an empty
// buffer instead
fft_plan
.r2c_plan
.process_with_scratch(real_fft_buffer, &mut self.complex_fft_buffer, &mut [])
.unwrap();
// TODO: Do the thing
// Inverse FFT back into the scratch buffer. This will be added to a ring buffer
// which gets written back to the host at a one block delay.
fft_plan
.c2r_plan
.process_with_scratch(&mut self.complex_fft_buffer, real_fft_buffer, &mut [])
.unwrap();
// Apply the window function once more to reduce time domain aliasing. The gain
// compensation compensates for the squared Hann window that would be applied if we
// didn't do any processing at all as well as the FFT+IFFT itself.
for (sample, window_sample) in real_fft_buffer.iter_mut().zip(&self.window_function)
{
*sample *= window_sample * output_gain;
}
});
ProcessStatus::Normal
}
}
impl SpectralCompressor {
/// `window_size` should not exceed `MAX_WINDOW_SIZE` or this will allocate.
fn resize_for_window(&mut self, window_size: usize) {
// The FFT algorithms for this window size have already been planned in
// `self.plan_for_order`, and all of these data structures already have enough capacity, so
// we just need to change some sizes.
self.stft.set_block_size(window_size);
self.window_function.resize(window_size, 0.0);
util::window::hann_in_place(&mut self.window_function);
self.complex_fft_buffer
.resize(window_size / 2 + 1, Complex32::default());
}
}
impl ClapPlugin for SpectralCompressor {
const CLAP_ID: &'static str = "nl.robbertvanderhelm.spectral-compressor";
const CLAP_DESCRIPTION: Option<&'static str> = Some("Turn things into pink noise on demand");

View file

@ -245,6 +245,16 @@ impl<const NUM_SIDECHAIN_INPUTS: usize> StftHelper<NUM_SIDECHAIN_INPUTS> {
self.current_pos = 0;
}
/// The number of channels this `StftHelper` was configured for
pub fn num_channels(&self) -> usize {
self.main_input_ring_buffers.len()
}
/// The maximum block size supported by this `StftHelper`.
pub fn max_block_size(&self) -> usize {
self.main_input_ring_buffers.capacity()
}
/// 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