1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236
|
// Copyright 2025 The Chromium Authors
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.
use std::simd::prelude::*;
use std::simd::Simd;
use crate::{Reg, Reg32, UReg};
/// Subblock sums and averages, used in eval_quant_err.
#[derive(Default, Copy, Clone)]
pub struct SubblockStats {
pub sum: [Reg; 3],
pub avg: [Reg; 3],
}
#[inline]
pub fn fast_div_255_round(x: Reg) -> Reg {
let r = x + Simd::splat(128);
(r + ((r + Simd::splat(257)) >> 8)) >> 8
}
/// Compute subblock (2x4 or 4x2) sums and averages.
///
/// Returns: subblock averages in order of top, bottom, left, right.
#[inline]
pub fn prepare_averages(data: &[[[Reg; 3]; 4]; 4]) -> [SubblockStats; 4] {
let mut sum_2x2 = [[Reg::default(); 3]; 4];
for y in 0..2 {
for x in 0..2 {
for ch in 0..3 {
sum_2x2[y * 2 + x][ch] = data[y * 2][x * 2][ch]
+ data[y * 2 + 1][x * 2][ch]
+ data[y * 2][x * 2 + 1][ch]
+ data[y * 2 + 1][x * 2 + 1][ch];
}
}
}
let mut out = [SubblockStats::default(); 4];
for (i, (s0, s1)) in [(0, 1), (2, 3), (0, 2), (1, 3)].into_iter().enumerate() {
for ch in 0..3 {
out[i].sum[ch] = sum_2x2[s0][ch] + sum_2x2[s1][ch];
out[i].avg[ch] = (out[i].sum[ch] + Simd::splat(4)) >> 3;
}
}
out
}
struct QuantResultWithErr {
/// Bit 47..32 of ETC1 codeword
lo: UReg,
/// Bit 63..48 of ETC1 codeword
hi: UReg,
/// Value of endpoint 0, scaled to `0..=255``
scaled0: [Reg; 3],
/// Value of endpoint 1, scaled to `0..=255``
scaled1: [Reg; 3],
/// Error metric, see [`eval_quant_err`]
err: Reg32,
}
pub struct QuantResult {
/// Bit 47..32 of ETC1 codeword
pub lo: UReg,
/// Bit 63..48 of ETC1 codeword
pub hi: UReg,
/// Value of endpoint 0, scaled to `0..=255``
pub scaled0: [Reg; 3],
/// Value of endpoint 1, scaled to `0..=255``
pub scaled1: [Reg; 3],
}
#[inline]
fn quant_444(
avg0: [Reg; 3],
avg1: [Reg; 3],
sum0: [Reg; 3],
sum1: [Reg; 3],
flip: bool,
) -> QuantResultWithErr {
#[inline]
fn quant(avg: [Reg; 3]) -> [Reg; 3] {
avg.map(|x| fast_div_255_round(x * Simd::splat(15)))
}
#[inline]
fn scale(q: [Reg; 3]) -> [Reg; 3] {
q.map(|x| (x << 4) | x)
}
#[inline]
fn encode(q0: [Reg; 3], q1: [Reg; 3], flip: bool) -> (UReg, UReg) {
let flip = if flip { UReg::splat(1) } else { UReg::splat(0) };
let diff = UReg::splat(0);
let base1_b = q1[2].cast::<u16>() << 8;
let base0_b = q0[2].cast::<u16>() << 12;
let lo = flip | diff | base1_b | base0_b;
let base1_g = q1[1].cast::<u16>();
let base0_g = q0[1].cast::<u16>() << 4;
let base1_r = q1[0].cast::<u16>() << 8;
let base0_r = q0[0].cast::<u16>() << 12;
let hi = base1_g | base0_g | base1_r | base0_r;
(lo, hi)
}
let q0 = quant(avg0);
let q1 = quant(avg1);
let scaled0 = scale(q0);
let scaled1 = scale(q1);
let err = eval_quant_err(scaled0, scaled1, sum0, sum1);
let (lo, hi) = encode(q0, q1, flip);
QuantResultWithErr { lo, hi, scaled0, scaled1, err }
}
#[inline]
fn quant_555(
avg0: [Reg; 3],
avg1: [Reg; 3],
sum0: [Reg; 3],
sum1: [Reg; 3],
flip: bool,
) -> QuantResultWithErr {
#[inline]
fn quant(avg: [Reg; 3]) -> [Reg; 3] {
avg.map(|x| fast_div_255_round(x * Simd::splat(31)))
}
#[inline]
fn scale(q: [Reg; 3]) -> [Reg; 3] {
// Per ETC1 spec, the "five-bit codewords are extended to eight bits by
// replicating the top three highest-order bits to the three lowest
// order bits".
q.map(|x| (x << 3) | (x >> 2))
}
#[inline]
fn encode(q0: [Reg; 3], delta: [Reg; 3], flip: bool) -> (UReg, UReg) {
#[inline]
fn encode_delta(d: Reg) -> UReg {
d.cast::<u16>() & Simd::splat(0b111)
}
let flip = if flip { UReg::splat(1) } else { UReg::splat(0) };
let diff = UReg::splat(1 << 1);
let delta_b = encode_delta(delta[2]) << 8;
let base_b = q0[2].cast::<u16>() << 11;
let lo = flip | diff | delta_b | base_b;
let delta_g = encode_delta(delta[1]);
let base_g = q0[1].cast::<u16>() << 3;
let delta_r = encode_delta(delta[0]) << 8;
let base_r = q0[0].cast::<u16>() << 11;
let hi = delta_g | base_g | delta_r | base_r;
(lo, hi)
}
let q0 = quant(avg0);
let q1 = quant(avg1);
let delta = [0, 1, 2].map(|i| (q1[i] - q0[i]).simd_clamp(Simd::splat(-4), Simd::splat(3)));
let q1 = [0, 1, 2].map(|i| q0[i] + delta[i]);
let scaled0 = scale(q0);
let scaled1 = scale(q1);
let err = eval_quant_err(scaled0, scaled1, sum0, sum1);
let (lo, hi) = encode(q0, delta, flip);
QuantResultWithErr { lo, hi, scaled0, scaled1, err }
}
#[inline]
fn eval_quant_err(q0: [Reg; 3], q1: [Reg; 3], sum0: [Reg; 3], sum1: [Reg; 3]) -> Reg32 {
// Target error metric:
// sum((x - q) ** 2) (for each pixel)
// where x is the original pixel value, and
// q is the quantized average of the block
// This can be rewritten as:
// sum(x ** 2) - 2 * sum(x * q) + sum(q ** 2)
// For relative comparisons, sum(x ** 2) is constant and can be omitted.
// With this and more simplification:
// q * sum(q - 2 * x)
// Assuming that we are computing the sum for 8 pixels within a subblock:
// q * (8 * q - 2 * sum(x))
// Dividing by 2:
// q * ((q << 2) - sum(x))
(0..3).fold(Reg32::splat(0), |mut acc, i| {
let q0 = q0[i].cast::<i32>();
let q1 = q1[i].cast::<i32>();
let sum0 = sum0[i].cast::<i32>();
let sum1 = sum1[i].cast::<i32>();
acc += q0 * ((q0 << 2) - sum0);
acc += q1 * ((q1 << 2) - sum1);
acc
})
}
#[inline]
fn quantize_endpoint_pairs(
avg0: [Reg; 3],
avg1: [Reg; 3],
sum0: [Reg; 3],
sum1: [Reg; 3],
flip: bool,
) -> QuantResultWithErr {
let q444 = quant_444(avg0, avg1, sum0, sum1, flip);
let q555 = quant_555(avg0, avg1, sum0, sum1, flip);
let prefer555_32 = q555.err.simd_lt(q444.err);
let prefer555 = prefer555_32.cast::<i16>();
QuantResultWithErr {
lo: prefer555.select(q555.lo, q444.lo),
hi: prefer555.select(q555.hi, q444.hi),
scaled0: [0, 1, 2].map(|i| prefer555.select(q555.scaled0[i], q444.scaled0[i])),
scaled1: [0, 1, 2].map(|i| prefer555.select(q555.scaled1[i], q444.scaled1[i])),
err: prefer555_32.select(q555.err, q444.err),
}
}
#[inline]
/// Search through flip / no-flip and individual / differential modes, and
/// return the result with the least MSE from original pixels.
pub fn quantize_averages(data: &[[[Reg; 3]; 4]; 4]) -> QuantResult {
let stats = prepare_averages(&data);
let flip =
quantize_endpoint_pairs(stats[0].avg, stats[1].avg, stats[0].sum, stats[1].sum, true);
let no_flip =
quantize_endpoint_pairs(stats[2].avg, stats[3].avg, stats[2].sum, stats[3].sum, false);
let prefer_flip = flip.err.simd_lt(no_flip.err).cast::<i16>();
QuantResult {
lo: prefer_flip.select(flip.lo, no_flip.lo),
hi: prefer_flip.select(flip.hi, no_flip.hi),
scaled0: [0, 1, 2].map(|i| prefer_flip.select(flip.scaled0[i], no_flip.scaled0[i])),
scaled1: [0, 1, 2].map(|i| prefer_flip.select(flip.scaled1[i], no_flip.scaled1[i])),
}
}
|