use crate::mla::fmla;
use crate::util::{DctSample, try_vec};
use crate::{PxdctError, PxdctExecutor};
use num_complex::Complex;
use num_traits::AsPrimitive;
use std::sync::Arc;
pub(crate) struct SplitRadixDct3<T: DctSample> {
twiddles: Vec<Complex<T>>,
half_dct: Arc<dyn PxdctExecutor<T> + Send + Sync>,
quarter_dct: Arc<dyn PxdctExecutor<T> + Send + Sync>,
execution_length: usize,
}
impl<T: DctSample> SplitRadixDct3<T>
where
f64: AsPrimitive<T>,
{
pub(crate) fn new(
len: usize,
half_dct: Arc<dyn PxdctExecutor<T> + Send + Sync>,
quarter_dct: Arc<dyn PxdctExecutor<T> + Send + Sync>,
) -> Result<SplitRadixDct3<T>, PxdctError> {
assert_eq!(
half_dct.length(),
quarter_dct.length() * 2,
"Invalid DCT was received, quarter size is not multiple of half for Split-Radix DCT-III"
);
use crate::twiddles::compute_twiddle;
let mut twiddles = try_vec![Complex::<T>::default(); len / 4];
for (i, twiddle) in twiddles.iter_mut().enumerate() {
*twiddle = compute_twiddle::<T>(2 * i + 1, len * 4).conj();
}
Ok(SplitRadixDct3 {
twiddles,
half_dct,
quarter_dct,
execution_length: len,
})
}
}
impl<T: DctSample> PxdctExecutor<T> for SplitRadixDct3<T>
where
f64: AsPrimitive<T>,
{
fn execute(&self, data: &mut [T]) -> Result<(), PxdctError> {
if !data.len().is_multiple_of(self.execution_length) {
return Err(PxdctError::InvalidSizeMultiplier(
data.len(),
self.execution_length,
));
}
let mut scratch = try_vec![T::default(); self.execution_length];
let len = self.length();
let half_len = len / 2;
let quarter_len = len / 4;
for chunk in data.chunks_exact_mut(self.execution_length) {
let (recursive_input_evens, recursive_input_odds) = scratch.split_at_mut(half_len);
let (recursive_input_n1, recursive_input_n3) =
recursive_input_odds.split_at_mut(quarter_len);
recursive_input_evens[0] = chunk[0];
recursive_input_evens[1] = chunk[2];
recursive_input_n1[0] = chunk[1] * T::TWO;
recursive_input_n3[0] = chunk[len - 1] * T::TWO;
for i in 1..quarter_len {
let k = 4 * i;
unsafe {
*recursive_input_evens.get_unchecked_mut(i * 2) = *chunk.get_unchecked(k);
*recursive_input_evens.get_unchecked_mut(i * 2 + 1) =
*chunk.get_unchecked(k + 2);
*recursive_input_n1.get_unchecked_mut(i) =
*chunk.get_unchecked(k - 1) + *chunk.get_unchecked(k + 1);
*recursive_input_n3.get_unchecked_mut(quarter_len - i) =
*chunk.get_unchecked(k - 1) - *chunk.get_unchecked(k + 1);
}
}
self.half_dct.execute(recursive_input_evens)?;
self.quarter_dct.execute(recursive_input_n1)?;
self.quarter_dct.execute(recursive_input_n3)?;
for (i, (twiddle, &cosine_value)) in self
.twiddles
.iter()
.zip(recursive_input_n1.iter())
.enumerate()
{
let sine_value = if i % 2 == 0 {
recursive_input_n3[i]
} else {
-recursive_input_n3[i]
};
let lower_dct4 = fmla(cosine_value, twiddle.re, sine_value * twiddle.im);
let upper_dct4 = fmla(cosine_value, twiddle.im, -sine_value * twiddle.re);
unsafe {
let lower_dct3 = *recursive_input_evens.get_unchecked(i);
let upper_dct3 = *recursive_input_evens.get_unchecked(half_len - i - 1);
*chunk.get_unchecked_mut(i) = lower_dct3 + lower_dct4;
*chunk.get_unchecked_mut(len - i - 1) = lower_dct3 - lower_dct4;
*chunk.get_unchecked_mut(half_len - i - 1) = upper_dct3 + upper_dct4;
*chunk.get_unchecked_mut(half_len + i) = upper_dct3 - upper_dct4;
}
}
}
Ok(())
}
#[inline]
fn length(&self) -> usize {
self.execution_length
}
}
pub(crate) struct SplitRadixDst3<T: DctSample> {
split_radix_dct3: SplitRadixDct3<T>,
execution_length: usize,
}
impl<T: DctSample> SplitRadixDst3<T>
where
f64: AsPrimitive<T>,
{
pub(crate) fn new(
len: usize,
half_dct: Arc<dyn PxdctExecutor<T> + Send + Sync>,
quarter_dct: Arc<dyn PxdctExecutor<T> + Send + Sync>,
) -> Result<SplitRadixDst3<T>, PxdctError> {
assert_eq!(
half_dct.length(),
quarter_dct.length() * 2,
"Invalid DCT was received, quarter size is not multiple of half for Split-Radix DST-III"
);
Ok(SplitRadixDst3 {
split_radix_dct3: SplitRadixDct3::new(len, half_dct, quarter_dct)?,
execution_length: len,
})
}
}
impl<T: DctSample> PxdctExecutor<T> for SplitRadixDst3<T>
where
f64: AsPrimitive<T>,
{
fn execute(&self, data: &mut [T]) -> Result<(), PxdctError> {
if !data.len().is_multiple_of(self.execution_length) {
return Err(PxdctError::InvalidSizeMultiplier(
data.len(),
self.execution_length,
));
}
for chunk in data.chunks_exact_mut(self.execution_length) {
chunk.reverse();
self.split_radix_dct3.execute(chunk)?;
for dst in chunk.chunks_exact_mut(2) {
dst[1] = dst[1].neg();
}
}
Ok(())
}
#[inline]
fn length(&self) -> usize {
self.execution_length
}
}
#[cfg(test)]
mod tests {
use super::*;
use crate::dct3_butterflies::{Dct3Butterfly8, Dct3Butterfly16};
use crate::tests::{naive_dct3, naive_dst3};
use rand::Rng;
#[test]
fn test_split_dct3() {
let mut input = vec![0.; 32];
for z in input.iter_mut() {
*z = rand::rng().random_range(1.0..2.0);
}
let reference_input = input.clone();
let reference_input = naive_dct3(&reference_input);
let bf = SplitRadixDct3::new(
32,
Arc::new(Dct3Butterfly16::default()),
Arc::new(Dct3Butterfly8::default()),
)
.unwrap();
bf.execute(&mut input).unwrap();
input
.iter()
.zip(reference_input.iter())
.enumerate()
.for_each(|(i, (&src, &r0))| {
assert!(
(src - r0).abs() < 1e-7,
"Difference must be < {}, but it was {}, at position {i}",
1e-7,
(src - r0).abs()
)
});
}
#[test]
fn test_split_dst3() {
let mut input = vec![0.; 32];
for z in input.iter_mut() {
*z = rand::rng().random_range(1.0..2.0);
}
let reference_input = input.clone();
let reference_input = naive_dst3(&reference_input);
let bf = SplitRadixDst3::new(
32,
Arc::new(Dct3Butterfly16::default()),
Arc::new(Dct3Butterfly8::default()),
)
.unwrap();
bf.execute(&mut input).unwrap();
input
.iter()
.zip(reference_input.iter())
.enumerate()
.for_each(|(i, (&src, &r0))| {
assert!(
(src - r0).abs() < 1e-7,
"Difference must be < {}, but it was {}, at position {i}",
1e-7,
(src - r0).abs()
)
});
}
}