| 1 |
/* |
| 2 |
* Copyright (c) 2023. |
| 3 |
* |
| 4 |
* This software is free software; |
| 5 |
* |
| 6 |
* You can redistribute it or modify it under terms of the MIT, Apache License or Zlib license |
| 7 |
*/ |
| 8 |
|
| 9 |
#![cfg(any(target_arch = "x86", target_arch = "x86_64"))] |
| 10 |
//! AVX optimised IDCT. |
| 11 |
//! |
| 12 |
//! Okay not thaat optimised. |
| 13 |
//! |
| 14 |
//! |
| 15 |
//! # The implementation |
| 16 |
//! The implementation is neatly broken down into two operations. |
| 17 |
//! |
| 18 |
//! 1. Test for zeroes |
| 19 |
//! > There is a shortcut method for idct where when all AC values are zero, we can get the answer really quickly. |
| 20 |
//! by scaling the 1/8th of the DCT coefficient of the block to the whole block and level shifting. |
| 21 |
//! |
| 22 |
//! 2. If above fails, we proceed to carry out IDCT as a two pass one dimensional algorithm. |
| 23 |
//! IT does two whole scans where it carries out IDCT on all items |
| 24 |
//! After each successive scan, data is transposed in register(thank you x86 SIMD powers). and the second |
| 25 |
//! pass is carried out. |
| 26 |
//! |
| 27 |
//! The code is not super optimized, it produces bit identical results with scalar code hence it's |
| 28 |
//! `mm256_add_epi16` |
| 29 |
//! and it also has the advantage of making this implementation easy to maintain. |
| 30 |
|
| 31 |
#![cfg(feature = "x86")] |
| 32 |
#![allow(dead_code)] |
| 33 |
|
| 34 |
#[cfg(target_arch = "x86")] |
| 35 |
use core::arch::x86::*; |
| 36 |
#[cfg(target_arch = "x86_64")] |
| 37 |
use core::arch::x86_64::*; |
| 38 |
|
| 39 |
use crate::unsafe_utils::{transpose, YmmRegister}; |
| 40 |
|
| 41 |
const SCALE_BITS: i32 = 512 + 65536 + (128 << 17); |
| 42 |
|
| 43 |
// Pack i32 to i16's, |
| 44 |
// clamp them to be between 0-255 |
| 45 |
// Undo shuffling |
| 46 |
// Store back to array |
| 47 |
macro_rules! permute_store { |
| 48 |
($x:tt,$y:tt,$index:tt,$out:tt,$stride:tt) => { |
| 49 |
let a = _mm256_packs_epi32($x, $y); |
| 50 |
|
| 51 |
// Clamp the values after packing, we can clamp more values at once |
| 52 |
let b = clamp_avx(a); |
| 53 |
|
| 54 |
// /Undo shuffling |
| 55 |
let c = _mm256_permute4x64_epi64(b, shuffle(3, 1, 2, 0)); |
| 56 |
|
| 57 |
// store first vector |
| 58 |
_mm_storeu_si128( |
| 59 |
($out) |
| 60 |
.get_mut($index..$index + 8) |
| 61 |
.unwrap() |
| 62 |
.as_mut_ptr() |
| 63 |
.cast(), |
| 64 |
_mm256_extractf128_si256::<0>(c), |
| 65 |
); |
| 66 |
$index += $stride; |
| 67 |
// second vector |
| 68 |
_mm_storeu_si128( |
| 69 |
($out) |
| 70 |
.get_mut($index..$index + 8) |
| 71 |
.unwrap() |
| 72 |
.as_mut_ptr() |
| 73 |
.cast(), |
| 74 |
_mm256_extractf128_si256::<1>(c), |
| 75 |
); |
| 76 |
$index += $stride; |
| 77 |
}; |
| 78 |
} |
| 79 |
|
| 80 |
#[target_feature(enable = "avx2")] |
| 81 |
#[allow( |
| 82 |
clippy::too_many_lines, |
| 83 |
clippy::cast_possible_truncation, |
| 84 |
clippy::similar_names, |
| 85 |
clippy::op_ref, |
| 86 |
unused_assignments, |
| 87 |
clippy::zero_prefixed_literal |
| 88 |
)] |
| 89 |
pub unsafe fn idct_avx2( |
| 90 |
in_vector: &mut [i32; 64], out_vector: &mut [i16], stride: usize, |
| 91 |
) { |
| 92 |
let mut pos = 0; |
| 93 |
|
| 94 |
// load into registers |
| 95 |
// |
| 96 |
// We sign extend i16's to i32's and calculate them with extended precision and |
| 97 |
// later reduce them to i16's when we are done carrying out IDCT |
| 98 |
|
| 99 |
let rw0 = _mm256_loadu_si256(in_vector[00..].as_ptr().cast()); |
| 100 |
let rw1 = _mm256_loadu_si256(in_vector[08..].as_ptr().cast()); |
| 101 |
let rw2 = _mm256_loadu_si256(in_vector[16..].as_ptr().cast()); |
| 102 |
let rw3 = _mm256_loadu_si256(in_vector[24..].as_ptr().cast()); |
| 103 |
let rw4 = _mm256_loadu_si256(in_vector[32..].as_ptr().cast()); |
| 104 |
let rw5 = _mm256_loadu_si256(in_vector[40..].as_ptr().cast()); |
| 105 |
let rw6 = _mm256_loadu_si256(in_vector[48..].as_ptr().cast()); |
| 106 |
let rw7 = _mm256_loadu_si256(in_vector[56..].as_ptr().cast()); |
| 107 |
|
| 108 |
// Forward DCT and quantization may cause all the AC terms to be zero, for such |
| 109 |
// cases we can try to accelerate it |
| 110 |
|
| 111 |
// Basically the poop is that whenever the array has 63 zeroes, its idct is |
| 112 |
// (arr[0]>>3)or (arr[0]/8) propagated to all the elements. |
| 113 |
// We first test to see if the array contains zero elements and if it does, we go the |
| 114 |
// short way. |
| 115 |
// |
| 116 |
// This reduces IDCT overhead from about 39% to 18 %, almost half |
| 117 |
|
| 118 |
// Do another load for the first row, we don't want to check DC value, because |
| 119 |
// we only care about AC terms |
| 120 |
let rw8 = _mm256_loadu_si256(in_vector[1..].as_ptr().cast()); |
| 121 |
|
| 122 |
let mut bitmap = _mm256_or_si256(rw1, rw2); |
| 123 |
bitmap = _mm256_or_si256(bitmap, rw3); |
| 124 |
bitmap = _mm256_or_si256(bitmap, rw4); |
| 125 |
bitmap = _mm256_or_si256(bitmap, rw5); |
| 126 |
bitmap = _mm256_or_si256(bitmap, rw6); |
| 127 |
bitmap = _mm256_or_si256(bitmap, rw7); |
| 128 |
bitmap = _mm256_or_si256(bitmap, rw8); |
| 129 |
|
| 130 |
if _mm256_testz_si256(bitmap, bitmap) == 1 { |
| 131 |
// AC terms all zero, idct of the block is ( coeff[0] * qt[0] )/8 + 128 (bias) |
| 132 |
// (and clamped to 255) |
| 133 |
// Round by adding 0.5 * (1 << 3) and offset by adding (128 << 3) before scaling |
| 134 |
let coeff = ((in_vector[0] + 4 + 1024) >> 3).clamp(0, 255) as i16; |
| 135 |
let idct_value = _mm_set1_epi16(coeff); |
| 136 |
|
| 137 |
macro_rules! store { |
| 138 |
($pos:tt,$value:tt) => { |
| 139 |
// store |
| 140 |
_mm_storeu_si128( |
| 141 |
out_vector |
| 142 |
.get_mut($pos..$pos + 8) |
| 143 |
.unwrap() |
| 144 |
.as_mut_ptr() |
| 145 |
.cast(), |
| 146 |
$value, |
| 147 |
); |
| 148 |
$pos += stride; |
| 149 |
}; |
| 150 |
} |
| 151 |
store!(pos, idct_value); |
| 152 |
store!(pos, idct_value); |
| 153 |
store!(pos, idct_value); |
| 154 |
store!(pos, idct_value); |
| 155 |
|
| 156 |
store!(pos, idct_value); |
| 157 |
store!(pos, idct_value); |
| 158 |
store!(pos, idct_value); |
| 159 |
store!(pos, idct_value); |
| 160 |
|
| 161 |
return; |
| 162 |
} |
| 163 |
|
| 164 |
let mut row0 = YmmRegister { mm256: rw0 }; |
| 165 |
let mut row1 = YmmRegister { mm256: rw1 }; |
| 166 |
let mut row2 = YmmRegister { mm256: rw2 }; |
| 167 |
let mut row3 = YmmRegister { mm256: rw3 }; |
| 168 |
|
| 169 |
let mut row4 = YmmRegister { mm256: rw4 }; |
| 170 |
let mut row5 = YmmRegister { mm256: rw5 }; |
| 171 |
let mut row6 = YmmRegister { mm256: rw6 }; |
| 172 |
let mut row7 = YmmRegister { mm256: rw7 }; |
| 173 |
|
| 174 |
macro_rules! dct_pass { |
| 175 |
($SCALE_BITS:tt,$scale:tt) => { |
| 176 |
// There are a lot of ways to do this |
| 177 |
// but to keep it simple(and beautiful), ill make a direct translation of the |
| 178 |
// scalar code to also make this code fully transparent(this version and the non |
| 179 |
// avx one should produce identical code.) |
| 180 |
|
| 181 |
// even part |
| 182 |
let p1 = (row2 + row6) * 2217; |
| 183 |
|
| 184 |
let mut t2 = p1 + row6 * -7567; |
| 185 |
let mut t3 = p1 + row2 * 3135; |
| 186 |
|
| 187 |
let mut t0 = YmmRegister { |
| 188 |
mm256: _mm256_slli_epi32((row0 + row4).mm256, 12), |
| 189 |
}; |
| 190 |
let mut t1 = YmmRegister { |
| 191 |
mm256: _mm256_slli_epi32((row0 - row4).mm256, 12), |
| 192 |
}; |
| 193 |
|
| 194 |
let x0 = t0 + t3 + $SCALE_BITS; |
| 195 |
let x3 = t0 - t3 + $SCALE_BITS; |
| 196 |
let x1 = t1 + t2 + $SCALE_BITS; |
| 197 |
let x2 = t1 - t2 + $SCALE_BITS; |
| 198 |
|
| 199 |
let p3 = row7 + row3; |
| 200 |
let p4 = row5 + row1; |
| 201 |
let p1 = row7 + row1; |
| 202 |
let p2 = row5 + row3; |
| 203 |
let p5 = (p3 + p4) * 4816; |
| 204 |
|
| 205 |
t0 = row7 * 1223; |
| 206 |
t1 = row5 * 8410; |
| 207 |
t2 = row3 * 12586; |
| 208 |
t3 = row1 * 6149; |
| 209 |
|
| 210 |
let p1 = p5 + p1 * -3685; |
| 211 |
let p2 = p5 + (p2 * -10497); |
| 212 |
let p3 = p3 * -8034; |
| 213 |
let p4 = p4 * -1597; |
| 214 |
|
| 215 |
t3 += p1 + p4; |
| 216 |
t2 += p2 + p3; |
| 217 |
t1 += p2 + p4; |
| 218 |
t0 += p1 + p3; |
| 219 |
|
| 220 |
row0.mm256 = _mm256_srai_epi32((x0 + t3).mm256, $scale); |
| 221 |
row1.mm256 = _mm256_srai_epi32((x1 + t2).mm256, $scale); |
| 222 |
row2.mm256 = _mm256_srai_epi32((x2 + t1).mm256, $scale); |
| 223 |
row3.mm256 = _mm256_srai_epi32((x3 + t0).mm256, $scale); |
| 224 |
|
| 225 |
row4.mm256 = _mm256_srai_epi32((x3 - t0).mm256, $scale); |
| 226 |
row5.mm256 = _mm256_srai_epi32((x2 - t1).mm256, $scale); |
| 227 |
row6.mm256 = _mm256_srai_epi32((x1 - t2).mm256, $scale); |
| 228 |
row7.mm256 = _mm256_srai_epi32((x0 - t3).mm256, $scale); |
| 229 |
}; |
| 230 |
} |
| 231 |
|
| 232 |
// Process rows |
| 233 |
dct_pass!(512, 10); |
| 234 |
transpose( |
| 235 |
&mut row0, &mut row1, &mut row2, &mut row3, &mut row4, &mut row5, &mut row6, &mut row7, |
| 236 |
); |
| 237 |
|
| 238 |
// process columns |
| 239 |
dct_pass!(SCALE_BITS, 17); |
| 240 |
transpose( |
| 241 |
&mut row0, &mut row1, &mut row2, &mut row3, &mut row4, &mut row5, &mut row6, &mut row7, |
| 242 |
); |
| 243 |
// Pack and write the values back to the array |
| 244 |
permute_store!((row0.mm256), (row1.mm256), pos, out_vector, stride); |
| 245 |
permute_store!((row2.mm256), (row3.mm256), pos, out_vector, stride); |
| 246 |
permute_store!((row4.mm256), (row5.mm256), pos, out_vector, stride); |
| 247 |
permute_store!((row6.mm256), (row7.mm256), pos, out_vector, stride); |
| 248 |
} |
| 249 |
|
| 250 |
|
| 251 |
#[target_feature(enable = "avx2")] |
| 252 |
#[allow( |
| 253 |
clippy::too_many_lines, |
| 254 |
clippy::cast_possible_truncation, |
| 255 |
clippy::similar_names, |
| 256 |
clippy::op_ref, |
| 257 |
unused_assignments, |
| 258 |
clippy::zero_prefixed_literal |
| 259 |
)] |
| 260 |
pub unsafe fn idct_avx2_4x4( |
| 261 |
in_vector: &mut [i32; 64], out_vector: &mut [i16], stride: usize, |
| 262 |
) { |
| 263 |
let rw0 = _mm256_loadu_si256(in_vector[00..].as_ptr().cast()); |
| 264 |
let rw1 = _mm256_loadu_si256(in_vector[08..].as_ptr().cast()); |
| 265 |
let rw2 = _mm256_loadu_si256(in_vector[16..].as_ptr().cast()); |
| 266 |
let rw3 = _mm256_loadu_si256(in_vector[24..].as_ptr().cast()); |
| 267 |
|
| 268 |
let mut row0 = YmmRegister { mm256: rw0 }; |
| 269 |
let mut row1 = YmmRegister { mm256: rw1 }; |
| 270 |
let mut row2 = YmmRegister { mm256: rw2 }; |
| 271 |
let mut row3 = YmmRegister { mm256: rw3 }; |
| 272 |
|
| 273 |
let mut row4 = YmmRegister { mm256: rw0 }; |
| 274 |
let mut row5 = YmmRegister { mm256: rw0 }; |
| 275 |
let mut row6 = YmmRegister { mm256: rw0 }; |
| 276 |
let mut row7 = YmmRegister { mm256: rw0 }; |
| 277 |
|
| 278 |
{ |
| 279 |
row0.mm256 = _mm256_slli_epi32(row0.mm256, 12); |
| 280 |
row0 += 512; |
| 281 |
|
| 282 |
let i2 = row2; |
| 283 |
|
| 284 |
let p1 = i2 * 2217; |
| 285 |
let p3 = i2 * 5352; |
| 286 |
|
| 287 |
let x0 = row0 + p3; |
| 288 |
let x1 = row0 + p1; |
| 289 |
let x2 = row0 - p1; |
| 290 |
let x3 = row0 - p3; |
| 291 |
|
| 292 |
// odd part |
| 293 |
let i4 = row3; |
| 294 |
let i3 = row1; |
| 295 |
|
| 296 |
let p5 = (i4 + i3) * 4816; |
| 297 |
|
| 298 |
let p1 = p5 + i3 * -3685; |
| 299 |
let p2 = p5 + i4 * -10497; |
| 300 |
|
| 301 |
let t3 = p5 + i3 * 867; |
| 302 |
let t2 = p5 + i4 * -5945; |
| 303 |
|
| 304 |
let t1 = p2 + i3 * -1597; |
| 305 |
let t0 = p1 + i4 * -8034; |
| 306 |
|
| 307 |
row0.mm256 = _mm256_srai_epi32((x0 + t3).mm256, 10); |
| 308 |
row1.mm256 = _mm256_srai_epi32((x1 + t2).mm256, 10); |
| 309 |
row2.mm256 = _mm256_srai_epi32((x2 + t1).mm256, 10); |
| 310 |
row3.mm256 = _mm256_srai_epi32((x3 + t0).mm256, 10); |
| 311 |
|
| 312 |
row4.mm256 = _mm256_srai_epi32((x3 - t0).mm256, 10); |
| 313 |
row5.mm256 = _mm256_srai_epi32((x2 - t1).mm256, 10); |
| 314 |
row6.mm256 = _mm256_srai_epi32((x1 - t2).mm256, 10); |
| 315 |
row7.mm256 = _mm256_srai_epi32((x0 - t3).mm256, 10); |
| 316 |
} |
| 317 |
|
| 318 |
transpose( |
| 319 |
&mut row0, &mut row1, &mut row2, &mut row3, &mut row4, &mut row5, &mut row6, &mut row7, |
| 320 |
); |
| 321 |
|
| 322 |
{ |
| 323 |
let i2 = row2; |
| 324 |
let i0 = row0; |
| 325 |
|
| 326 |
row0.mm256 = _mm256_slli_epi32(i0.mm256, 12); |
| 327 |
let t0 = row0 + SCALE_BITS; |
| 328 |
|
| 329 |
let t2 = i2 * 2217; |
| 330 |
let t3 = i2 * 5352; |
| 331 |
|
| 332 |
// constants scaled things up by 1<<12, plus we had 1<<2 from first |
| 333 |
// loop, plus horizontal and vertical each scale by sqrt(8) so together |
| 334 |
// we've got an extra 1<<3, so 1<<17 total we need to remove. |
| 335 |
// so we want to round that, which means adding 0.5 * 1<<17, |
| 336 |
// aka 65536. Also, we'll end up with -128 to 127 that we want |
| 337 |
// to encode as 0..255 by adding 128, so we'll add that before the shift |
| 338 |
// Rounding constant is already added into `t0` |
| 339 |
let x0 = t0 + t3; |
| 340 |
let x3 = t0 - t3; |
| 341 |
let x1 = t0 + t2; |
| 342 |
let x2 = t0 - t2; |
| 343 |
|
| 344 |
// odd part |
| 345 |
let i3 = row3; |
| 346 |
let i1 = row1; |
| 347 |
|
| 348 |
let p5 = (i3 + i1) * 4816; |
| 349 |
|
| 350 |
let p1 = p5 + i1 * -3685; |
| 351 |
let p2 = p5 + i3 * -10497; |
| 352 |
|
| 353 |
let t3 = p5 + i1 * 867; |
| 354 |
let t2 = p5 + i3 * -5945; |
| 355 |
|
| 356 |
let t1 = p2 + i1 * -1597; |
| 357 |
let t0 = p1 + i3 * -8034; |
| 358 |
|
| 359 |
row0.mm256 = _mm256_srai_epi32((x0 + t3).mm256, 17); |
| 360 |
row1.mm256 = _mm256_srai_epi32((x1 + t2).mm256, 17); |
| 361 |
row2.mm256 = _mm256_srai_epi32((x2 + t1).mm256, 17); |
| 362 |
row3.mm256 = _mm256_srai_epi32((x3 + t0).mm256, 17); |
| 363 |
row4.mm256 = _mm256_srai_epi32((x3 - t0).mm256, 17); |
| 364 |
row5.mm256 = _mm256_srai_epi32((x2 - t1).mm256, 17); |
| 365 |
row6.mm256 = _mm256_srai_epi32((x1 - t2).mm256, 17); |
| 366 |
row7.mm256 = _mm256_srai_epi32((x0 - t3).mm256, 17); |
| 367 |
} |
| 368 |
|
| 369 |
transpose( |
| 370 |
&mut row0, &mut row1, &mut row2, &mut row3, &mut row4, &mut row5, &mut row6, &mut row7, |
| 371 |
); |
| 372 |
|
| 373 |
let mut pos = 0; |
| 374 |
|
| 375 |
// Pack and write the values back to the array |
| 376 |
permute_store!((row0.mm256), (row1.mm256), pos, out_vector, stride); |
| 377 |
permute_store!((row2.mm256), (row3.mm256), pos, out_vector, stride); |
| 378 |
permute_store!((row4.mm256), (row5.mm256), pos, out_vector, stride); |
| 379 |
permute_store!((row6.mm256), (row7.mm256), pos, out_vector, stride); |
| 380 |
} |
| 381 |
|
| 382 |
#[inline] |
| 383 |
#[target_feature(enable = "avx2")] |
| 384 |
unsafe fn clamp_avx(reg: __m256i) -> __m256i { |
| 385 |
let min_s = _mm256_set1_epi16(0); |
| 386 |
let max_s = _mm256_set1_epi16(255); |
| 387 |
|
| 388 |
let max_v = _mm256_max_epi16(reg, min_s); //max(a,0) |
| 389 |
let min_v = _mm256_min_epi16(max_v, max_s); //min(max(a,0),255) |
| 390 |
return min_v; |
| 391 |
} |
| 392 |
|
| 393 |
/// A copy of `_MM_SHUFFLE()` that doesn't require |
| 394 |
/// a nightly compiler |
| 395 |
#[inline] |
| 396 |
const fn shuffle(z: i32, y: i32, x: i32, w: i32) -> i32 { |
| 397 |
((z << 6) | (y << 4) | (x << 2) | w) |
| 398 |
} |