1
0
Fork 0

Optimize branching in Crisp

This commit is contained in:
Robbert van der Helm 2022-03-09 19:42:13 +01:00
parent 48ed39269d
commit 76394731a6

View file

@ -349,18 +349,17 @@ impl Plugin for Crisp {
let mut rm_outputs = [[0.0; NUM_CHANNELS as usize]; BLOCK_SIZE];
// Reduce per-sample branching a bit by iterating over smaller blocks and only then
// deciding what to dowith the output
for (channel_samples, rm_outputs) in block.iter_samples().zip(&mut rm_outputs) {
let amount = self.params.amount.smoothed.next() * AMOUNT_GAIN_MULTIPLIER;
// deciding what to do with the output. This version branches only once per sample (in
// `do_ring_mod()`) which can be trivially optimized to a masked min/max later.
// TODO: SIMD-ize this to process both channels at once
match self.params.stereo_mode.value() {
StereoMode::Mono => {
for (channel_samples, rm_outputs) in block.iter_samples().zip(&mut rm_outputs) {
let amount = self.params.amount.smoothed.next() * AMOUNT_GAIN_MULTIPLIER;
// Controls the pre-RM LPF and the HPF applied to the noise signal
self.maybe_update_filters();
// Controls the pre-RM LPF and the HPF applied to the noise signal
self.maybe_update_filters();
// TODO: SIMD-ize this to process both channels at once
// TODO: Avoid branching twice here. Modern branch predictors are pretty good at this
// though.
match self.params.stereo_mode.value() {
StereoMode::Mono => {
let noise = self.gen_noise(0);
for (channel_idx, (sample, rm_output)) in
channel_samples.into_iter().zip(rm_outputs).enumerate()
@ -368,7 +367,11 @@ impl Plugin for Crisp {
*rm_output = self.do_ring_mod(*sample, channel_idx, noise) * amount;
}
}
StereoMode::Stereo => {
}
StereoMode::Stereo => {
for (channel_samples, rm_outputs) in block.iter_samples().zip(&mut rm_outputs) {
let amount = self.params.amount.smoothed.next() * AMOUNT_GAIN_MULTIPLIER;
self.maybe_update_filters();
for (channel_idx, (sample, rm_output)) in
channel_samples.into_iter().zip(rm_outputs).enumerate()
{