use super::ac_strategy::neon_log2p1_f32;
use crate::image::Plane;
use crate::inflated_cost::{
RateLog2Lut, ReconDistInput, ReconErrorKernels, ReconKernels, recon_dist_and_rate_with_kernels,
validate_ssim_inputs,
};
use std::arch::aarch64::*;
#[inline]
#[target_feature(enable = "neon")]
fn accumulate_gradient_vectors_x4(
left: float32x4_t,
right: float32x4_t,
sum: float32x4_t,
) -> float32x4_t {
let difference = vsubq_f32(right, left);
vfmaq_f32(sum, difference, difference)
}
#[inline]
#[target_feature(enable = "neon")]
fn accumulate_gradient_x4(left: &[f32; 4], right: &[f32; 4], sum: float32x4_t) -> float32x4_t {
let left = unsafe { vld1q_f32(left.as_ptr()) };
let right = unsafe { vld1q_f32(right.as_ptr()) };
accumulate_gradient_vectors_x4(left, right, sum)
}
#[target_feature(enable = "neon")]
pub(crate) fn error_gradient_energy_neon(error: &[f32], width: usize, height: usize) -> f32 {
let n = width
.checked_mul(height)
.expect("gradient plane size overflow");
assert!(error.len() >= n);
if width == 0 || height == 0 {
return 0.0;
}
let rows = error[..n].chunks_exact(width);
let mut sum = vdupq_n_f32(0.0);
for row in rows.clone() {
let (row4, row_tail) = row.as_chunks::<4>();
if row_tail.is_empty() {
let (right4, _) = row[1..].as_chunks::<4>();
for (left, right) in row4.iter().zip(right4) {
sum = accumulate_gradient_x4(left, right, sum);
}
let left = unsafe { vld1q_f32(row4.last().unwrap().as_ptr()) };
let right = vsetq_lane_f32::<3>(vgetq_lane_f32::<3>(left), vextq_f32::<1>(left, left));
sum = accumulate_gradient_vectors_x4(left, right, sum);
continue;
}
let (left4, tail) = row[..width - 1].as_chunks::<4>();
let (right4, right_tail) = row[1..].as_chunks::<4>();
for (left, right) in left4.iter().zip(right4) {
sum = accumulate_gradient_x4(left, right, sum);
}
if !tail.is_empty() {
let mut left = [0.0; 4];
let mut right = [0.0; 4];
left[..tail.len()].copy_from_slice(tail);
right[..right_tail.len()].copy_from_slice(right_tail);
sum = accumulate_gradient_x4(&left, &right, sum);
}
}
for (top, bottom) in rows.clone().zip(rows.skip(1)) {
let (top4, top_tail) = top.as_chunks::<4>();
let (bottom4, bottom_tail) = bottom.as_chunks::<4>();
for (top, bottom) in top4.iter().zip(bottom4) {
sum = accumulate_gradient_x4(top, bottom, sum);
}
if !top_tail.is_empty() {
let mut top = [0.0; 4];
let mut bottom = [0.0; 4];
top[..top_tail.len()].copy_from_slice(top_tail);
bottom[..bottom_tail.len()].copy_from_slice(bottom_tail);
sum = accumulate_gradient_x4(&top, &bottom, sum);
}
}
vaddvq_f32(sum)
}
#[inline]
#[target_feature(enable = "neon")]
fn combine_error_x4(spatial: &[f32; 4], luma: &[f32; 4], factor: f32, combined: &mut [f32; 4]) {
let spatial = unsafe { vld1q_f32(spatial.as_ptr()) };
let luma = unsafe { vld1q_f32(luma.as_ptr()) };
let value = vfmaq_n_f32(spatial, luma, factor);
unsafe { vst1q_f32(combined.as_mut_ptr(), value) };
}
#[target_feature(enable = "neon")]
pub(crate) fn combine_error_neon(spatial: &[f32], luma: &[f32], factor: f32, combined: &mut [f32]) {
debug_assert_eq!(spatial.len(), luma.len());
debug_assert_eq!(spatial.len(), combined.len());
let (spatial16, spatial_tail) = spatial.as_chunks::<16>();
let (luma16, luma_tail) = luma.as_chunks::<16>();
let (combined16, combined_tail) = combined.as_chunks_mut::<16>();
for ((spatial, luma), combined) in spatial16.iter().zip(luma16).zip(combined16) {
let [s0, s1, s2, s3] = spatial.as_chunks::<4>().0 else {
unreachable!()
};
let [l0, l1, l2, l3] = luma.as_chunks::<4>().0 else {
unreachable!()
};
let [c0, c1, c2, c3] = combined.as_chunks_mut::<4>().0 else {
unreachable!()
};
combine_error_x4(s0, l0, factor, c0);
combine_error_x4(s1, l1, factor, c1);
combine_error_x4(s2, l2, factor, c2);
combine_error_x4(s3, l3, factor, c3);
}
let (spatial4, spatial_remainder) = spatial_tail.as_chunks::<4>();
let (luma4, luma_remainder) = luma_tail.as_chunks::<4>();
let (combined4, combined_remainder) = combined_tail.as_chunks_mut::<4>();
for ((spatial, luma), combined) in spatial4.iter().zip(luma4).zip(combined4) {
combine_error_x4(spatial, luma, factor, combined);
}
if !spatial_remainder.is_empty() {
let mut spatial = [0.0; 4];
let mut luma = [0.0; 4];
let mut combined = [0.0; 4];
spatial[..spatial_remainder.len()].copy_from_slice(spatial_remainder);
luma[..luma_remainder.len()].copy_from_slice(luma_remainder);
combine_error_x4(&spatial, &luma, factor, &mut combined);
combined_remainder.copy_from_slice(&combined[..combined_remainder.len()]);
}
}
#[target_feature(enable = "neon")]
pub(crate) fn recon_dist_and_rate_neon(
scratch: &mut [[f32; 1024]; 8],
input: &ReconDistInput<'_>,
error: &ReconErrorKernels,
) -> (f32, f32) {
recon_dist_and_rate_with_kernels(
scratch,
input,
&ReconKernels {
quantize: recon_quantize_neon,
ssim: ssim_deficit_neon,
prepare: prepare_reconstruction_neon,
error,
},
)
}
#[allow(clippy::too_many_arguments)]
#[target_feature(enable = "neon")]
fn recon_quantize_neon(
coeff: &[f32],
inv: &[f32],
quant_scale: f32,
thresholds: &[f32; 4],
width: usize,
height: usize,
half: usize,
cx: usize,
cy: usize,
coeff_error: &mut [f32],
_rate_log2_lut: &RateLog2Lut,
scan_pos: &[u32],
) -> f32 {
let n = width
.checked_mul(height)
.expect("coefficient size overflow");
assert!(width.is_multiple_of(4) && half.is_multiple_of(4));
assert!(coeff.len() >= n && inv.len() >= n && coeff_error.len() >= n && scan_pos.len() >= n);
let scale = vdupq_n_f32(quant_scale);
let zero = vdupq_n_f32(0.0);
let all = vdupq_n_u32(u32::MAX);
let lanes = unsafe { vld1q_u32([0, 1, 2, 3].as_ptr()) };
let mut nonzero = 0usize;
let mut magnitude_bits = vdupq_n_f32(0.0);
let mut scan_acc = vdupq_n_u32(0);
for (y, (((coeff_row, inv_row), error_row), scan_row)) in coeff
.chunks_exact(width)
.zip(inv.chunks_exact(width))
.zip(coeff_error.chunks_exact_mut(width))
.zip(scan_pos.chunks_exact(width))
.take(height)
.enumerate()
{
let yfix = if y >= height / 2 { 2 } else { 0 };
for (chunk_x, (((coeff4, inv4), error4), scan4)) in coeff_row
.as_chunks::<4>()
.0
.iter()
.zip(inv_row.as_chunks::<4>().0.iter())
.zip(error_row.as_chunks_mut::<4>().0.iter_mut())
.zip(scan_row.as_chunks::<4>().0.iter())
.enumerate()
{
let x = chunk_x * 4;
let threshold = vdupq_n_f32(if x >= half {
thresholds[yfix + 1]
} else {
thresholds[yfix]
});
let coeff_v = unsafe { vld1q_f32(coeff4.as_ptr()) };
let inv_v = unsafe { vld1q_f32(inv4.as_ptr()) };
let denominator = vmulq_f32(inv_v, scale);
let scaled = vmulq_f32(denominator, coeff_v);
let keep = vcgeq_f32(vabsq_f32(scaled), threshold);
let quantized =
vreinterpretq_f32_u32(vandq_u32(vreinterpretq_u32_f32(vrndaq_f32(scaled)), keep));
let error = vdivq_f32(vsubq_f32(scaled, quantized), denominator);
let active = if y < cy && x < cx {
vcgeq_u32(
vaddq_u32(vdupq_n_u32(x as u32), lanes),
vdupq_n_u32(cx as u32),
)
} else {
all
};
unsafe { vst1q_f32(error4.as_mut_ptr(), vbslq_f32(active, error, zero)) };
let active_quantized = vbslq_f32(active, quantized, zero);
let absolute_q = vabsq_f32(active_quantized);
let nonzero_mask = vcgtq_f32(absolute_q, zero);
nonzero += vaddvq_u32(vshrq_n_u32::<31>(nonzero_mask)) as usize;
magnitude_bits = vaddq_f32(
magnitude_bits,
vbslq_f32(nonzero_mask, neon_log2p1_f32(absolute_q), zero),
);
let sv = unsafe { vld1q_u32(scan4.as_ptr()) };
scan_acc = vmaxq_u32(scan_acc, vandq_u32(sv, nonzero_mask));
}
}
let header = vgetq_lane_f32::<0>(neon_log2p1_f32(vsetq_lane_f32::<0>(
nonzero as f32,
vdupq_n_f32(0.0),
)));
nonzero as f32 * 1.6
+ vaddvq_f32(magnitude_bits)
+ 0.4 * header
+ crate::inflated_cost::R_ZERO
* crate::inflated_cost::visited_zeros(nonzero, vmaxvq_u32(scan_acc), cx, cy)
}
#[allow(clippy::too_many_arguments)]
#[target_feature(enable = "neon")]
fn prepare_reconstruction_neon(
plane: &Plane<f32>,
px: usize,
py: usize,
pixel_width: usize,
pixel_height: usize,
error: &[f32],
original: &mut [f32],
reconstructed: &mut [f32],
) {
let n = pixel_width
.checked_mul(pixel_height)
.expect("reconstruction size overflow");
assert!(pixel_width.is_multiple_of(4));
assert!(error.len() >= n && original.len() >= n && reconstructed.len() >= n);
let (image_width, image_height) = (plane.xsize(), plane.ysize());
assert!(image_width != 0 && image_height != 0);
let source_x = px.min(image_width - 1);
let copied = pixel_width.min(image_width - source_x);
for (y, ((original_row, reconstructed_row), error_row)) in original
.chunks_exact_mut(pixel_width)
.zip(reconstructed.chunks_exact_mut(pixel_width))
.zip(error.chunks_exact(pixel_width))
.take(pixel_height)
.enumerate()
{
let source = plane.row(py.saturating_add(y).min(image_height - 1));
original_row[..copied].copy_from_slice(&source[source_x..source_x + copied]);
original_row[copied..].fill(source[image_width - 1]);
for ((output, value), delta) in reconstructed_row
.as_chunks_mut::<4>()
.0
.iter_mut()
.zip(original_row.as_chunks::<4>().0.iter())
.zip(error_row.as_chunks::<4>().0.iter())
{
let values = unsafe { vld1q_f32(value.as_ptr()) };
let errors = unsafe { vld1q_f32(delta.as_ptr()) };
unsafe { vst1q_f32(output.as_mut_ptr(), vsubq_f32(values, errors)) };
}
}
}
#[target_feature(enable = "neon")]
pub(crate) fn ssim_deficit_neon(orig: &[f32], recon: &[f32], width: usize, height: usize) -> f32 {
validate_ssim_inputs(orig, recon, width, height);
const C1: f32 = 1e-4;
const C2: f32 = 9e-4;
const INV_64: f32 = 1.0 / 64.0;
let mut deficit = 0.0f32;
for wy in 0..height / 8 {
for wx in 0..width / 8 {
let (x0, y0) = (wx * 8, wy * 8);
let base_index = y0 * width + x0;
let base_o = orig[base_index];
let base_r = recon[base_index];
let base_ov = vdupq_n_f32(base_o);
let base_rv = vdupq_n_f32(base_r);
let (mut sum_o, mut sum_r) = (vdupq_n_f32(0.0), vdupq_n_f32(0.0));
for (orig_row, recon_row) in orig
.chunks_exact(width)
.zip(recon.chunks_exact(width))
.skip(y0)
.take(8)
{
let orig8 = &orig_row.as_chunks::<8>().0[wx];
let recon8 = &recon_row.as_chunks::<8>().0[wx];
for (orig4, recon4) in orig8
.as_chunks::<4>()
.0
.iter()
.zip(recon8.as_chunks::<4>().0)
{
sum_o = vaddq_f32(
sum_o,
vsubq_f32(unsafe { vld1q_f32(orig4.as_ptr()) }, base_ov),
);
sum_r = vaddq_f32(
sum_r,
vsubq_f32(unsafe { vld1q_f32(recon4.as_ptr()) }, base_rv),
);
}
}
let mean_o = f32::mul_add(vaddvq_f32(sum_o), INV_64, base_o);
let mean_r = f32::mul_add(vaddvq_f32(sum_r), INV_64, base_r);
let mean_ov = vdupq_n_f32(mean_o);
let mean_rv = vdupq_n_f32(mean_r);
let (mut var_o, mut var_r, mut cov) =
(vdupq_n_f32(0.0), vdupq_n_f32(0.0), vdupq_n_f32(0.0));
for (orig_row, recon_row) in orig
.chunks_exact(width)
.zip(recon.chunks_exact(width))
.skip(y0)
.take(8)
{
let orig8 = &orig_row.as_chunks::<8>().0[wx];
let recon8 = &recon_row.as_chunks::<8>().0[wx];
for (orig4, recon4) in orig8
.as_chunks::<4>()
.0
.iter()
.zip(recon8.as_chunks::<4>().0)
{
let o = vsubq_f32(unsafe { vld1q_f32(orig4.as_ptr()) }, mean_ov);
let r = vsubq_f32(unsafe { vld1q_f32(recon4.as_ptr()) }, mean_rv);
var_o = vfmaq_f32(var_o, o, o);
var_r = vfmaq_f32(var_r, r, r);
cov = vfmaq_f32(cov, o, r);
}
}
let vo = vaddvq_f32(var_o) * INV_64;
let vr = vaddvq_f32(var_r) * INV_64;
let covariance = vaddvq_f32(cov) * INV_64;
let luminance = (2.0 * mean_o * mean_r + C1) / (mean_o * mean_o + mean_r * mean_r + C1);
let structure = (2.0 * covariance + C2) / (vo + vr + C2);
deficit += (1.0 - luminance * structure) * 64.0;
}
}
deficit
}
#[cfg(test)]
mod tests {
use super::{
combine_error_neon, error_gradient_energy_neon, recon_dist_and_rate_neon, ssim_deficit_neon,
};
use crate::dc_group_data::{
STRATEGY_DCT, STRATEGY_DCT8X16, STRATEGY_DCT16X8, STRATEGY_DCT16X16, STRATEGY_DCT16X32,
STRATEGY_DCT32X16, STRATEGY_DCT32X32,
};
use crate::image::Image3F;
use crate::inflated_cost::{
ReconDistInput, ReconErrorKernels, ReconQuantization, ReconScoring, ReconSource,
ReconTransform, rate_log2_lut, recon_dist_and_rate_scalar, ssim_deficit_scalar,
};
#[test]
fn ssim_neon_matches_scalar() {
let mut state = 91u32;
for (width, height) in [(8, 8), (16, 8), (8, 16), (16, 16), (32, 32)] {
let n = width * height;
let mut orig = vec![0.0f32; n];
let mut recon = vec![0.0f32; n];
for i in 0..n {
state = state.wrapping_mul(1_664_525).wrapping_add(1_013_904_223);
orig[i] = (state >> 8) as f32 / (1u32 << 24) as f32;
recon[i] = orig[i] + ((i % 13) as f32 - 6.0) * 0.0007;
}
let scalar = ssim_deficit_scalar(&orig, &recon, width, height);
let simd = unsafe { ssim_deficit_neon(&orig, &recon, width, height) };
let tolerance = 2e-4f32.max(scalar.abs() * 2e-5);
assert!(
(simd - scalar).abs() <= tolerance,
"{width}x{height}: simd={simd} scalar={scalar}"
);
}
}
#[test]
fn recon_neon_matches_scalar_for_all_standard_transforms() {
let mut image = Image3F::new(40, 40);
for c in 0..3 {
for y in 0..40 {
for x in 0..40 {
image.plane_mut(c).row_mut(y)[x] =
0.15 * c as f32 + 0.003 * x as f32 + 0.002 * y as f32;
}
}
}
let mut coeffs = [[0.0f32; 1024]; 3];
let mut state = 17u32;
for channel in &mut coeffs {
for value in channel {
state = state.wrapping_mul(1_664_525).wrapping_add(1_013_904_223);
*value = ((state >> 8) as f32 / (1u32 << 24) as f32 - 0.5) * 2.0;
}
}
let inv_storage = [vec![1.0f32; 1024], vec![0.9f32; 1024], vec![1.1f32; 1024]];
for (strategy, cx, cy) in [
(STRATEGY_DCT, 1, 1),
(STRATEGY_DCT8X16, 2, 1),
(STRATEGY_DCT16X8, 2, 1),
(STRATEGY_DCT16X16, 2, 2),
(STRATEGY_DCT16X32, 4, 2),
(STRATEGY_DCT32X16, 4, 2),
(STRATEGY_DCT32X32, 4, 4),
] {
let inv = [
&inv_storage[0][..],
&inv_storage[1][..],
&inv_storage[2][..],
];
let input = ReconDistInput {
quantization: ReconQuantization {
rate_log2_lut: rate_log2_lut(),
coeffs: &coeffs,
inverse_matrices: inv,
qac: 7.0,
qm_mult_x: 1.2,
distance: 1.5,
},
transform: ReconTransform {
blocks_x: cx,
blocks_y: cy,
strategy,
},
source: ReconSource {
opsin: &image,
x: 3,
y: 5,
},
scoring: ReconScoring {
factor_x: 0.15,
factor_b: -0.1,
banding: true,
},
};
let mut scalar_scratch = [[0.0f32; 1024]; 8];
let scalar = recon_dist_and_rate_scalar(&mut scalar_scratch, &input);
let mut simd_scratch = [[0.0f32; 1024]; 8];
let simd = unsafe {
recon_dist_and_rate_neon(
&mut simd_scratch,
&input,
&ReconErrorKernels {
gradient_energy: |error, width, height| {
error_gradient_energy_neon(error, width, height)
},
combine: |spatial, luma, factor, combined| {
combine_error_neon(spatial, luma, factor, combined)
},
},
)
};
let rate_tolerance = 2e-4f32.max(scalar.1.abs() * 3e-6);
assert!(
(simd.1 - scalar.1).abs() <= rate_tolerance,
"strategy {strategy} rate: simd={} scalar={}",
simd.1,
scalar.1
);
let tolerance = 5e-4f32.max(scalar.0.abs() * 3e-5);
assert!(
(simd.0 - scalar.0).abs() <= tolerance,
"strategy {strategy}: simd={simd:?} scalar={scalar:?}"
);
}
}
#[test]
fn ssim_neon_rejects_short_slices_before_loading() {
let result =
std::panic::catch_unwind(|| unsafe { ssim_deficit_neon(&[0.0; 63], &[0.0; 64], 8, 8) });
assert!(result.is_err());
}
}