use crate::avx::dct2::dct2_split_radix_rotation_twiddles_avx;
use crate::avx::stored::AvxStoreD;
use crate::avx::util::fma;
use crate::util::{DctConstants, try_vec};
use crate::{PxdctError, PxdctExecutor};
use std::sync::Arc;
pub(crate) struct AvxSplitRadixDct3d {
twiddles: Vec<AvxStoreD>,
half_dct: Arc<dyn PxdctExecutor<f64> + Send + Sync>,
quarter_dct: Arc<dyn PxdctExecutor<f64> + Send + Sync>,
execution_length: usize,
}
impl AvxSplitRadixDct3d {
pub(crate) fn new(
len: usize,
half_dct: Arc<dyn PxdctExecutor<f64> + Send + Sync>,
quarter_dct: Arc<dyn PxdctExecutor<f64> + Send + Sync>,
) -> Result<AvxSplitRadixDct3d, 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"
);
Ok(AvxSplitRadixDct3d {
twiddles: unsafe { dct2_split_radix_rotation_twiddles_avx(len) },
half_dct,
quarter_dct,
execution_length: len,
})
}
}
impl PxdctExecutor<f64> for AvxSplitRadixDct3d {
fn execute(&self, data: &mut [f64]) -> Result<(), PxdctError> {
unsafe { self.execute_impl(data) }
}
#[inline]
fn length(&self) -> usize {
self.execution_length
}
}
impl AvxSplitRadixDct3d {
#[target_feature(enable = "avx2", enable = "fma")]
fn execute_impl(&self, data: &mut [f64]) -> 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![f64::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);
unsafe {
*recursive_input_evens.get_unchecked_mut(0) = *chunk.get_unchecked(0);
*recursive_input_evens.get_unchecked_mut(1) = *chunk.get_unchecked(2);
*recursive_input_n1.get_unchecked_mut(0) = *chunk.get_unchecked(1) * f64::TWO;
*recursive_input_n3.get_unchecked_mut(0) = *chunk.get_unchecked(len - 1) * f64::TWO;
}
let conj_odd = AvxStoreD::set_values(0.0, -0.0, 0.0, -0.0);
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);
let b = *chunk.get_unchecked(k - 1);
let f = *chunk.get_unchecked(k + 1);
*recursive_input_n1.get_unchecked_mut(i) = b + f;
*recursive_input_n3.get_unchecked_mut(quarter_len - i) = b - f;
}
}
self.half_dct.execute(recursive_input_evens)?;
self.quarter_dct.execute(recursive_input_n1)?;
self.quarter_dct.execute(recursive_input_n3)?;
let mut i = 0usize;
let mut tw_idx = 0usize;
while i + 4 <= quarter_len {
const S: usize = 4;
let sine_value =
AvxStoreD::load(unsafe { recursive_input_n3.get_unchecked(i..) }).xor(conj_odd);
let cos_value = AvxStoreD::load(unsafe { recursive_input_n1.get_unchecked(i..) });
let twiddle_re = unsafe { *self.twiddles.get_unchecked(tw_idx) };
let twiddle_im = unsafe { *self.twiddles.get_unchecked(tw_idx + 1) };
let lower_dct4 = fma(cos_value, twiddle_re, sine_value * twiddle_im);
let upper_dct4 = fma(cos_value, twiddle_im, -sine_value * twiddle_re);
unsafe {
let lower_dct3 = AvxStoreD::load(recursive_input_evens.get_unchecked(i..));
let upper_dct3 =
AvxStoreD::load(recursive_input_evens.get_unchecked(half_len - i - S..))
.reverse();
let v0 = lower_dct3 + lower_dct4;
let v1 = lower_dct3 - lower_dct4;
let v2 = upper_dct3 + upper_dct4;
let v3 = upper_dct3 - upper_dct4;
v0.write(chunk.get_unchecked_mut(i..));
v1.reverse().write(chunk.get_unchecked_mut(len - i - S..));
v2.reverse()
.write(chunk.get_unchecked_mut(half_len - i - S..));
v3.write(chunk.get_unchecked_mut(half_len + i..));
}
tw_idx += 2;
i += 4;
}
let rem = quarter_len - i;
if rem == 2 {
const S: usize = 2;
let sine_value = AvxStoreD::load2(unsafe { recursive_input_n3.get_unchecked(i..) })
.xor(conj_odd);
let cos_value = AvxStoreD::load2(unsafe { recursive_input_n1.get_unchecked(i..) });
let twiddle_re = unsafe { *self.twiddles.get_unchecked(tw_idx) };
let twiddle_im = unsafe { *self.twiddles.get_unchecked(tw_idx + 1) };
let lower_dct4 = fma(cos_value, twiddle_re, sine_value * twiddle_im);
let upper_dct4 = fma(cos_value, twiddle_im, -sine_value * twiddle_re);
unsafe {
let lower_dct3 = AvxStoreD::load2(recursive_input_evens.get_unchecked(i..));
let upper_dct3 =
AvxStoreD::load2(recursive_input_evens.get_unchecked(half_len - i - S..))
.reverse2();
let v0 = lower_dct3 + lower_dct4;
let v1 = lower_dct3 - lower_dct4;
let v2 = upper_dct3 + upper_dct4;
let v3 = upper_dct3 - upper_dct4;
v0.write2(chunk.get_unchecked_mut(i..));
v1.reverse2().write2(chunk.get_unchecked_mut(len - i - S..));
v2.reverse2()
.write2(chunk.get_unchecked_mut(half_len - i - S..));
v3.write2(chunk.get_unchecked_mut(half_len + i..));
}
}
}
Ok(())
}
}
#[cfg(test)]
mod tests {
use super::*;
use crate::dct3::{Dct3Butterfly8, Dct3Butterfly16};
use crate::tests::naive_dct3;
use crate::util::has_valid_avx;
use rand::Rng;
#[test]
fn test_split_dct3() {
if !has_valid_avx() {
return;
}
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 = AvxSplitRadixDct3d::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()
)
});
}
}