use std::arch::x86_64::*;
use super::ac_strategy::{avx_dequantized_level_f32, avx2_log2p1_f32, hsum256};
use crate::color_correlation::{CFL_MAGNITUDE_BITS, CFL_NONZERO_BITS, cfl_rd_cost_scalar};
#[target_feature(enable = "avx2,fma")]
pub(crate) fn cfl_rd_cost_avx2(m: &[f32], s: &[f32], factor: f32, weight: f32, lambda: f32) -> f32 {
let len = m.len().min(s.len());
let (m_chunks, m_tail) = m[..len].as_chunks::<8>();
let (s_chunks, s_tail) = s[..len].as_chunks::<8>();
let factor_v = _mm256_set1_ps(factor);
let weight_v = _mm256_set1_ps(weight);
let zero = _mm256_setzero_ps();
let one = _mm256_set1_ps(1.0);
let half = _mm256_set1_ps(0.5);
let sign = _mm256_set1_ps(-0.0);
let nonzero_bits = _mm256_set1_ps(lambda * CFL_NONZERO_BITS);
let magnitude_bits = _mm256_set1_ps(lambda * CFL_MAGNITUDE_BITS);
let mut acc = zero;
for (m, s) in m_chunks.iter().zip(s_chunks) {
let mv = unsafe { _mm256_loadu_ps(m.as_ptr()) };
let sv = unsafe { _mm256_loadu_ps(s.as_ptr()) };
let residual = _mm256_sub_ps(sv, _mm256_mul_ps(factor_v, mv));
let zero_cost = _mm256_mul_ps(_mm256_mul_ps(weight_v, residual), residual);
let truncated = _mm256_round_ps::<{ _MM_FROUND_TO_ZERO | _MM_FROUND_NO_EXC }>(residual);
let fraction = _mm256_andnot_ps(sign, _mm256_sub_ps(residual, truncated));
let at_least_half = _mm256_cmp_ps::<_CMP_GE_OQ>(fraction, half);
let signed_one = _mm256_or_ps(one, _mm256_and_ps(sign, residual));
let level = _mm256_add_ps(truncated, _mm256_and_ps(signed_one, at_least_half));
let coded_err = _mm256_sub_ps(residual, avx_dequantized_level_f32(level));
let abs_level = _mm256_andnot_ps(sign, level);
let log2_level = avx2_log2p1_f32(_mm256_max_ps(_mm256_sub_ps(abs_level, one), zero));
let bits = _mm256_fmadd_ps(magnitude_bits, log2_level, nonzero_bits);
let coded = _mm256_add_ps(
_mm256_mul_ps(_mm256_mul_ps(weight_v, coded_err), coded_err),
bits,
);
let nonzero = _mm256_cmp_ps::<_CMP_GT_OQ>(abs_level, zero);
let best = _mm256_min_ps(zero_cost, _mm256_blendv_ps(zero_cost, coded, nonzero));
acc = _mm256_add_ps(acc, best);
}
hsum256(acc) + cfl_rd_cost_scalar(m_tail, s_tail, factor, weight, lambda)
}
#[inline]
#[target_feature(enable = "avx2")]
fn reduce_transposed_4x8(a: __m256, b: __m256, c: __m256, d: __m256) -> [f32; 4] {
let ab_lo = _mm256_unpacklo_ps(a, b);
let ab_hi = _mm256_unpackhi_ps(a, b);
let cd_lo = _mm256_unpacklo_ps(c, d);
let cd_hi = _mm256_unpackhi_ps(c, d);
let lane_0_4 = _mm256_shuffle_ps::<0x44>(ab_lo, cd_lo);
let lane_1_5 = _mm256_shuffle_ps::<0xee>(ab_lo, cd_lo);
let lane_2_6 = _mm256_shuffle_ps::<0x44>(ab_hi, cd_hi);
let lane_3_7 = _mm256_shuffle_ps::<0xee>(ab_hi, cd_hi);
let sums_0_3_and_4_7 = _mm256_add_ps(
_mm256_add_ps(lane_0_4, lane_1_5),
_mm256_add_ps(lane_2_6, lane_3_7),
);
let low = _mm256_castps256_ps128(sums_0_3_and_4_7);
let high = _mm256_extractf128_ps::<1>(sums_0_3_and_4_7);
let sums = _mm_add_ps(low, high);
let mut out = [0.0f32; 4];
unsafe { _mm_storeu_ps(out.as_mut_ptr(), sums) };
out
}
#[target_feature(enable = "avx2,fma")]
pub(crate) fn cfl_regression_avx2(
y: &[f32; 64],
x: &[f32; 64],
b: &[f32; 64],
qm_x: &[f32; 64],
qm_b: &[f32; 64],
) -> [f32; 4] {
let inv_color = _mm256_set1_ps(1.0 / 84.0);
let mut ca_x = _mm256_setzero_ps();
let mut cb_x = _mm256_setzero_ps();
let mut ca_b = _mm256_setzero_ps();
let mut cb_b = _mm256_setzero_ps();
let y_chunks = y.as_chunks::<8>().0;
let x_chunks = x.as_chunks::<8>().0;
let b_chunks = b.as_chunks::<8>().0;
let qm_x_chunks = qm_x.as_chunks::<8>().0;
let qm_b_chunks = qm_b.as_chunks::<8>().0;
for ((((y, x), b), qm_x), qm_b) in y_chunks
.iter()
.zip(x_chunks)
.zip(b_chunks)
.zip(qm_x_chunks)
.zip(qm_b_chunks)
{
let yv = unsafe { _mm256_loadu_ps(y.as_ptr()) };
let xv = unsafe { _mm256_loadu_ps(x.as_ptr()) };
let bv = unsafe { _mm256_loadu_ps(b.as_ptr()) };
let qx = unsafe { _mm256_loadu_ps(qm_x.as_ptr()) };
let qb = unsafe { _mm256_loadu_ps(qm_b.as_ptr()) };
let mx = _mm256_mul_ps(yv, qx);
let sx = _mm256_mul_ps(xv, qx);
let ax = _mm256_mul_ps(inv_color, mx);
ca_x = _mm256_fmadd_ps(ax, ax, ca_x);
cb_x = _mm256_fnmadd_ps(ax, sx, cb_x);
let mb = _mm256_mul_ps(yv, qb);
let sb = _mm256_mul_ps(bv, qb);
let ab = _mm256_mul_ps(inv_color, mb);
let residual_b = _mm256_sub_ps(mb, sb);
ca_b = _mm256_fmadd_ps(ab, ab, ca_b);
cb_b = _mm256_fmadd_ps(ab, residual_b, cb_b);
}
reduce_transposed_4x8(ca_x, cb_x, ca_b, cb_b)
}
#[target_feature(enable = "avx2")]
pub(crate) fn cfl_rdo_block_avx2(
m_x: &mut [f32; 63],
s_x: &mut [f32; 63],
m_b: &mut [f32; 63],
s_b: &mut [f32; 63],
block_y: &[f32; 64],
block_x: &[f32; 64],
block_b: &[f32; 64],
qm_x: &[f32; 64],
qm_b: &[f32; 64],
q_block: f32,
) {
let q_block_v = _mm256_set1_ps(q_block);
for i in (0..56).step_by(8) {
let coeff = i + 1;
let y = unsafe { _mm256_loadu_ps(block_y.as_ptr().add(coeff)) };
let x = unsafe { _mm256_loadu_ps(block_x.as_ptr().add(coeff)) };
let b = unsafe { _mm256_loadu_ps(block_b.as_ptr().add(coeff)) };
let qx = unsafe { _mm256_loadu_ps(qm_x.as_ptr().add(coeff)) };
let qb = unsafe { _mm256_loadu_ps(qm_b.as_ptr().add(coeff)) };
let qx = _mm256_mul_ps(qx, q_block_v);
let qb = _mm256_mul_ps(qb, q_block_v);
unsafe {
_mm256_storeu_ps(m_x.as_mut_ptr().add(i), _mm256_mul_ps(y, qx));
_mm256_storeu_ps(s_x.as_mut_ptr().add(i), _mm256_mul_ps(x, qx));
_mm256_storeu_ps(m_b.as_mut_ptr().add(i), _mm256_mul_ps(y, qb));
_mm256_storeu_ps(s_b.as_mut_ptr().add(i), _mm256_mul_ps(b, qb));
}
}
let i = 56;
let coeff = i + 1;
let mask = _mm256_setr_epi32(-1, -1, -1, -1, -1, -1, -1, 0);
let y = unsafe { _mm256_maskload_ps(block_y.as_ptr().add(coeff), mask) };
let x = unsafe { _mm256_maskload_ps(block_x.as_ptr().add(coeff), mask) };
let b = unsafe { _mm256_maskload_ps(block_b.as_ptr().add(coeff), mask) };
let qx = unsafe { _mm256_maskload_ps(qm_x.as_ptr().add(coeff), mask) };
let qb = unsafe { _mm256_maskload_ps(qm_b.as_ptr().add(coeff), mask) };
let qx = _mm256_mul_ps(qx, q_block_v);
let qb = _mm256_mul_ps(qb, q_block_v);
unsafe {
_mm256_maskstore_ps(m_x.as_mut_ptr().add(i), mask, _mm256_mul_ps(y, qx));
_mm256_maskstore_ps(s_x.as_mut_ptr().add(i), mask, _mm256_mul_ps(x, qx));
_mm256_maskstore_ps(m_b.as_mut_ptr().add(i), mask, _mm256_mul_ps(y, qb));
_mm256_maskstore_ps(s_b.as_mut_ptr().add(i), mask, _mm256_mul_ps(b, qb));
}
}
#[target_feature(enable = "avx2,fma")]
pub(crate) fn cfl_rdo_stats_avx2(m: &[f32], s: &[f32]) -> [f32; 4] {
let len = m.len().min(s.len());
let (m_chunks, m_tail) = m[..len].as_chunks::<8>();
let (s_chunks, s_tail) = s[..len].as_chunks::<8>();
let mut dot_ms = _mm256_setzero_ps();
let mut dot_mm = _mm256_setzero_ps();
let mut dot_ss = _mm256_setzero_ps();
let mut sum_abs_s = _mm256_setzero_ps();
let sign = _mm256_set1_ps(-0.0);
for (m, s) in m_chunks.iter().zip(s_chunks) {
let mv = unsafe { _mm256_loadu_ps(m.as_ptr()) };
let sv = unsafe { _mm256_loadu_ps(s.as_ptr()) };
dot_ms = _mm256_fmadd_ps(mv, sv, dot_ms);
dot_mm = _mm256_fmadd_ps(mv, mv, dot_mm);
dot_ss = _mm256_fmadd_ps(sv, sv, dot_ss);
sum_abs_s = _mm256_add_ps(sum_abs_s, _mm256_andnot_ps(sign, sv));
}
let tail_len = m_tail.len();
if tail_len != 0 {
let lanes = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
let mask = _mm256_cmpgt_epi32(_mm256_set1_epi32(tail_len as i32), lanes);
let mv = unsafe { _mm256_maskload_ps(m_tail.as_ptr(), mask) };
let sv = unsafe { _mm256_maskload_ps(s_tail.as_ptr(), mask) };
dot_ms = _mm256_fmadd_ps(mv, sv, dot_ms);
dot_mm = _mm256_fmadd_ps(mv, mv, dot_mm);
dot_ss = _mm256_fmadd_ps(sv, sv, dot_ss);
sum_abs_s = _mm256_add_ps(sum_abs_s, _mm256_andnot_ps(sign, sv));
}
reduce_transposed_4x8(dot_ms, dot_mm, dot_ss, sum_abs_s)
}
#[target_feature(enable = "avx2,fma")]
pub(crate) fn apply_cfl_avx2(x: &mut [f32], y: &[f32], b: &mut [f32], cmap_factor: [f32; 3]) {
assert_eq!(x.len(), y.len());
assert_eq!(x.len(), b.len());
let cx = _mm256_set1_ps(-cmap_factor[0]);
let cb = _mm256_set1_ps(-cmap_factor[2]);
let (x_chunks, x_tail) = x.as_chunks_mut::<8>();
let (y_chunks, y_tail) = y.as_chunks::<8>();
let (b_chunks, b_tail) = b.as_chunks_mut::<8>();
let (x_groups, x_chunks_tail) = x_chunks.as_chunks_mut::<4>();
let (y_groups, y_chunks_tail) = y_chunks.as_chunks::<4>();
let (b_groups, b_chunks_tail) = b_chunks.as_chunks_mut::<4>();
for ((x, y), b) in x_groups.iter_mut().zip(y_groups).zip(b_groups) {
let y0 = unsafe { _mm256_loadu_ps(y[0].as_ptr()) };
let y1 = unsafe { _mm256_loadu_ps(y[1].as_ptr()) };
let y2 = unsafe { _mm256_loadu_ps(y[2].as_ptr()) };
let y3 = unsafe { _mm256_loadu_ps(y[3].as_ptr()) };
let x0 = unsafe { _mm256_loadu_ps(x[0].as_ptr()) };
let x1 = unsafe { _mm256_loadu_ps(x[1].as_ptr()) };
let x2 = unsafe { _mm256_loadu_ps(x[2].as_ptr()) };
let x3 = unsafe { _mm256_loadu_ps(x[3].as_ptr()) };
let b0 = unsafe { _mm256_loadu_ps(b[0].as_ptr()) };
let b1 = unsafe { _mm256_loadu_ps(b[1].as_ptr()) };
let b2 = unsafe { _mm256_loadu_ps(b[2].as_ptr()) };
let b3 = unsafe { _mm256_loadu_ps(b[3].as_ptr()) };
let x0 = _mm256_fmadd_ps(cx, y0, x0);
let x1 = _mm256_fmadd_ps(cx, y1, x1);
let x2 = _mm256_fmadd_ps(cx, y2, x2);
let x3 = _mm256_fmadd_ps(cx, y3, x3);
let b0 = _mm256_fmadd_ps(cb, y0, b0);
let b1 = _mm256_fmadd_ps(cb, y1, b1);
let b2 = _mm256_fmadd_ps(cb, y2, b2);
let b3 = _mm256_fmadd_ps(cb, y3, b3);
unsafe {
_mm256_storeu_ps(x[0].as_mut_ptr(), x0);
_mm256_storeu_ps(x[1].as_mut_ptr(), x1);
_mm256_storeu_ps(x[2].as_mut_ptr(), x2);
_mm256_storeu_ps(x[3].as_mut_ptr(), x3);
_mm256_storeu_ps(b[0].as_mut_ptr(), b0);
_mm256_storeu_ps(b[1].as_mut_ptr(), b1);
_mm256_storeu_ps(b[2].as_mut_ptr(), b2);
_mm256_storeu_ps(b[3].as_mut_ptr(), b3);
}
}
for ((x, y), b) in x_chunks_tail
.iter_mut()
.zip(y_chunks_tail)
.zip(b_chunks_tail)
{
let yv = unsafe { _mm256_loadu_ps(y.as_ptr()) };
let xv = unsafe { _mm256_loadu_ps(x.as_ptr()) };
let bv = unsafe { _mm256_loadu_ps(b.as_ptr()) };
unsafe {
_mm256_storeu_ps(x.as_mut_ptr(), _mm256_fmadd_ps(cx, yv, xv));
_mm256_storeu_ps(b.as_mut_ptr(), _mm256_fmadd_ps(cb, yv, bv));
}
}
for ((x, &y), b) in x_tail.iter_mut().zip(y_tail).zip(b_tail) {
*x = (-cmap_factor[0]).mul_add(y, *x);
*b = (-cmap_factor[2]).mul_add(y, *b);
}
}
#[cfg(test)]
mod tests {
#[test]
fn apply_cfl_avx2_matches_scalar() {
if !is_x86_feature_detected!("avx2") || !is_x86_feature_detected!("fma") {
return;
}
for size in [0, 1, 7, 8, 15, 32, 64, 1024, 4096] {
let y: Vec<f32> = (0..size)
.map(|i| ((i * 37 % 101) as f32 - 50.0) * 0.03125)
.collect();
let mut x: Vec<f32> = (0..size).map(|i| i as f32 * 0.015625 - 2.0).collect();
let mut b: Vec<f32> = (0..size).map(|i| 3.0 - i as f32 * 0.0078125).collect();
let mut want_x = x.clone();
let mut want_b = b.clone();
let factors: [f32; 3] = [0.3125, 0.0, -0.1875];
for i in 0..size {
want_x[i] = (-factors[0]).mul_add(y[i], want_x[i]);
want_b[i] = (-factors[2]).mul_add(y[i], want_b[i]);
}
unsafe { super::apply_cfl_avx2(&mut x, &y, &mut b, factors) };
assert_eq!(x, want_x, "X mismatch at size {size}");
assert_eq!(b, want_b, "B mismatch at size {size}");
}
}
}