| 1 |
use std::sync::Arc; |
| 2 |
|
| 3 |
use aligned::{Aligned, A64}; |
| 4 |
use arrayvec::ArrayVec; |
| 5 |
use num_rational::Rational32; |
| 6 |
use rayon::iter::{IntoParallelIterator, ParallelIterator}; |
| 7 |
use v_frame::{ |
| 8 |
frame::Frame, |
| 9 |
math::{clamp, ILog}, |
| 10 |
pixel::{ChromaSampling, Pixel}, |
| 11 |
plane::{Plane, PlaneConfig, PlaneOffset}, |
| 12 |
}; |
| 13 |
|
| 14 |
use super::importance::{ |
| 15 |
IMPORTANCE_BLOCK_SIZE, |
| 16 |
IMP_BLOCK_MV_UNITS_PER_PIXEL, |
| 17 |
IMP_BLOCK_SIZE_IN_MV_UNITS, |
| 18 |
}; |
| 19 |
use crate::{ |
| 20 |
cpu::CpuFeatureLevel, |
| 21 |
data::{ |
| 22 |
block::{BlockOffset, BlockSize, MIB_SIZE_LOG2}, |
| 23 |
frame::{FrameInvariants, FrameState, RefType, ALLOWED_REF_FRAMES}, |
| 24 |
motion::{ |
| 25 |
MEStats, |
| 26 |
MVSamplingMode, |
| 27 |
MotionEstimationSubsets, |
| 28 |
MotionVector, |
| 29 |
ReadGuardMEStats, |
| 30 |
RefMEStats, |
| 31 |
TileMEStats, |
| 32 |
MV_LOW, |
| 33 |
MV_UPP, |
| 34 |
}, |
| 35 |
plane::{Area, AsRegion, PlaneBlockOffset, PlaneRegion, PlaneRegionMut, Rect}, |
| 36 |
prediction::PredictionMode, |
| 37 |
sad::get_sad, |
| 38 |
satd::get_satd, |
| 39 |
superblock::{ |
| 40 |
SuperBlockOffset, |
| 41 |
TileSuperBlockOffset, |
| 42 |
MAX_SB_SIZE_LOG2, |
| 43 |
MI_SIZE, |
| 44 |
MI_SIZE_LOG2, |
| 45 |
SB_SIZE, |
| 46 |
}, |
| 47 |
tile::{TileBlockOffset, TileRect, TileStateMut, TilingInfo}, |
| 48 |
}, |
| 49 |
}; |
| 50 |
|
| 51 |
/// Declares an array of motion vectors in structure of arrays syntax. |
| 52 |
macro_rules! search_pattern_subpel { |
| 53 |
($field_a:ident: [$($ll_a:expr),*], $field_b:ident: [$($ll_b:expr),*]) => { |
| 54 |
[ $(MotionVector { $field_a: $ll_a, $field_b: $ll_b } ),*] |
| 55 |
}; |
| 56 |
} |
| 57 |
|
| 58 |
/// Declares an array of motion vectors in structure of arrays syntax. |
| 59 |
/// Compared to [`search_pattern_subpel`], this version creates motion vectors |
| 60 |
/// in fullpel resolution (x8). |
| 61 |
macro_rules! search_pattern { |
| 62 |
($field_a:ident: [$($ll_a:expr),*], $field_b:ident: [$($ll_b:expr),*]) => { |
| 63 |
[ $(MotionVector { $field_a: $ll_a << 3, $field_b: $ll_b << 3 } ),*] |
| 64 |
}; |
| 65 |
} |
| 66 |
|
| 67 |
/// Diamond pattern of radius 1 as shown below. For fullpel search, use |
| 68 |
/// `DIAMOND_R1_PATTERN_FULLPEL` since it has been scaled for fullpel search. |
| 69 |
/// ```text |
| 70 |
/// X |
| 71 |
/// XoX |
| 72 |
/// X |
| 73 |
/// ``` |
| 74 |
/// 'X's are motion candidates and the 'o' is the center. |
| 75 |
const DIAMOND_R1_PATTERN_SUBPEL: [MotionVector; 4] = search_pattern_subpel!( |
| 76 |
col: [ 0, 1, 0, -1], |
| 77 |
row: [ 1, 0, -1, 0] |
| 78 |
); |
| 79 |
|
| 80 |
/// Diamond pattern of radius 1 as shown below. Unlike `DIAMOND_R1_PATTERN`, the |
| 81 |
/// vectors have been shifted fullpel scale. |
| 82 |
/// ```text |
| 83 |
/// X |
| 84 |
/// XoX |
| 85 |
/// X |
| 86 |
/// ``` |
| 87 |
/// 'X's are motion candidates and the 'o' is the center. |
| 88 |
const DIAMOND_R1_PATTERN: [MotionVector; 4] = search_pattern!( |
| 89 |
col: [ 0, 1, 0, -1], |
| 90 |
row: [ 1, 0, -1, 0] |
| 91 |
); |
| 92 |
|
| 93 |
/// Uneven multi-hexagon search pattern around a center point. Used for locating |
| 94 |
/// irregular movement. |
| 95 |
/// ```text |
| 96 |
/// X |
| 97 |
/// X X |
| 98 |
/// X X |
| 99 |
/// X X |
| 100 |
/// X o X |
| 101 |
/// X X |
| 102 |
/// X X |
| 103 |
/// X X |
| 104 |
/// X |
| 105 |
/// ``` |
| 106 |
/// 'X's are motion candidates and the 'o' is the center. |
| 107 |
const UMH_PATTERN: [MotionVector; 16] = search_pattern!( |
| 108 |
col: [ -2, -1, 0, 1, 2, 3, 4, 3, 2, 1, 0, -1, -2, 3, -4, -3], |
| 109 |
row: [ 4, 4, 4, 4, 4, 2, 0, -2, -4, -4, -4, -4, -4, -2, 0, 2] |
| 110 |
); |
| 111 |
|
| 112 |
/// A hexagon pattern around a center point. The pattern is ordered so that the |
| 113 |
/// offsets circle around the center. This is done to allow pruning locations |
| 114 |
/// covered by the last iteration. |
| 115 |
/// ```text |
| 116 |
/// 21012 |
| 117 |
/// 2 X X |
| 118 |
/// 1 |
| 119 |
/// 0 X o X |
| 120 |
/// 1 |
| 121 |
/// 2 X X |
| 122 |
/// ``` |
| 123 |
/// 'X's are motion candidates and the 'o' is the center. |
| 124 |
/// |
| 125 |
/// The illustration below shows the process of a hexagon search. |
| 126 |
/// ```text |
| 127 |
/// Step 1 Step 2 |
| 128 |
/// 1 1 1 1 2 |
| 129 |
/// |
| 130 |
/// 1(0)1 => 1 0(1)2 |
| 131 |
/// |
| 132 |
/// 1 1 1 1 2 |
| 133 |
/// ``` |
| 134 |
/// The search above has gone through the following steps. |
| 135 |
/// 1. Search '1' elements for better candidates than the center '0'. |
| 136 |
/// 2. Recenter around the best candidate ('(1)') and hexagon candidates that |
| 137 |
/// don't overlap with the previous search step (labeled '2'). |
| 138 |
const HEXAGON_PATTERN: [MotionVector; 6] = search_pattern!( |
| 139 |
col: [ 0, 2, 2, 0, -2, -2], |
| 140 |
row: [ -2, -1, 1, 2, 1, -1] |
| 141 |
); |
| 142 |
|
| 143 |
/// A small square pattern around a center point. |
| 144 |
/// ```text |
| 145 |
/// 101 |
| 146 |
/// 1 XXX |
| 147 |
/// 0 XoX |
| 148 |
/// 1 XXX |
| 149 |
/// ``` |
| 150 |
/// 'X's are motion candidates and the 'o' is the center. |
| 151 |
const SQUARE_REFINE_PATTERN: [MotionVector; 8] = search_pattern!( |
| 152 |
col: [ -1, 0, 1, -1, 1, -1, 0, 1], |
| 153 |
row: [ 1, 1, 1, 0, 0, -1, -1, -1] |
| 154 |
); |
| 155 |
|
| 156 |
pub(crate) fn estimate_inter_costs<T: Pixel>( |
| 157 |
frame: Arc<Frame<T>>, |
| 158 |
ref_frame: Arc<Frame<T>>, |
| 159 |
bit_depth: usize, |
| 160 |
frame_rate: Rational32, |
| 161 |
chroma_sampling: ChromaSampling, |
| 162 |
buffer: RefMEStats, |
| 163 |
cpu_feature_level: CpuFeatureLevel, |
| 164 |
) -> f64 { |
| 165 |
let last_fi = |
| 166 |
FrameInvariants::new_key_frame(frame.planes[0].cfg.width, frame.planes[0].cfg.height); |
| 167 |
let fi = FrameInvariants::new_inter_frame(&last_fi, 1).unwrap(); |
| 168 |
|
| 169 |
// Compute the motion vectors. |
| 170 |
let mut fs = FrameState::new_with_frame_and_me_stats_and_rec(Arc::clone(&frame), buffer); |
| 171 |
let mut tiling = TilingInfo::from_target_tiles( |
| 172 |
frame.planes[0].cfg.width, |
| 173 |
frame.planes[0].cfg.height, |
| 174 |
*frame_rate.numer() as f64 / *frame_rate.denom() as f64, |
| 175 |
TilingInfo::tile_log2(1, 0).unwrap(), |
| 176 |
TilingInfo::tile_log2(1, 0).unwrap(), |
| 177 |
chroma_sampling == ChromaSampling::Cs422, |
| 178 |
); |
| 179 |
compute_motion_vectors(&fi, &mut fs, &mut tiling, bit_depth, cpu_feature_level); |
| 180 |
|
| 181 |
// Estimate inter costs |
| 182 |
let plane_org = &frame.planes[0]; |
| 183 |
let plane_ref = &ref_frame.planes[0]; |
| 184 |
let h_in_imp_b = plane_org.cfg.height / IMPORTANCE_BLOCK_SIZE; |
| 185 |
let w_in_imp_b = plane_org.cfg.width / IMPORTANCE_BLOCK_SIZE; |
| 186 |
let stats = &fs.frame_me_stats.read().expect("poisoned lock")[0]; |
| 187 |
let bsize = BlockSize::from_width_and_height(IMPORTANCE_BLOCK_SIZE, IMPORTANCE_BLOCK_SIZE); |
| 188 |
|
| 189 |
let mut inter_costs = 0; |
| 190 |
(0..h_in_imp_b).for_each(|y| { |
| 191 |
(0..w_in_imp_b).for_each(|x| { |
| 192 |
let mv = stats[y * 2][x * 2].mv; |
| 193 |
|
| 194 |
// Coordinates of the top-left corner of the reference block, in MV |
| 195 |
// units. |
| 196 |
let reference_x = x as i64 * IMP_BLOCK_SIZE_IN_MV_UNITS + mv.col as i64; |
| 197 |
let reference_y = y as i64 * IMP_BLOCK_SIZE_IN_MV_UNITS + mv.row as i64; |
| 198 |
|
| 199 |
let region_org = plane_org.region(Area::Rect(Rect { |
| 200 |
x: (x * IMPORTANCE_BLOCK_SIZE) as isize, |
| 201 |
y: (y * IMPORTANCE_BLOCK_SIZE) as isize, |
| 202 |
width: IMPORTANCE_BLOCK_SIZE, |
| 203 |
height: IMPORTANCE_BLOCK_SIZE, |
| 204 |
})); |
| 205 |
|
| 206 |
let region_ref = plane_ref.region(Area::Rect(Rect { |
| 207 |
x: reference_x as isize / IMP_BLOCK_MV_UNITS_PER_PIXEL as isize, |
| 208 |
y: reference_y as isize / IMP_BLOCK_MV_UNITS_PER_PIXEL as isize, |
| 209 |
width: IMPORTANCE_BLOCK_SIZE, |
| 210 |
height: IMPORTANCE_BLOCK_SIZE, |
| 211 |
})); |
| 212 |
|
| 213 |
inter_costs += get_satd( |
| 214 |
®ion_org, |
| 215 |
®ion_ref, |
| 216 |
bsize.width(), |
| 217 |
bsize.height(), |
| 218 |
bit_depth, |
| 219 |
cpu_feature_level, |
| 220 |
) as u64; |
| 221 |
}); |
| 222 |
}); |
| 223 |
inter_costs as f64 / (w_in_imp_b * h_in_imp_b) as f64 |
| 224 |
} |
| 225 |
|
| 226 |
fn compute_motion_vectors<T: Pixel>( |
| 227 |
fi: &FrameInvariants<T>, |
| 228 |
fs: &mut FrameState<T>, |
| 229 |
tiling_info: &mut TilingInfo, |
| 230 |
bit_depth: usize, |
| 231 |
cpu_feature_level: CpuFeatureLevel, |
| 232 |
) { |
| 233 |
tiling_info |
| 234 |
.tile_iter_mut(fs) |
| 235 |
.collect::<Vec<_>>() |
| 236 |
.into_par_iter() |
| 237 |
.for_each(|mut ctx| { |
| 238 |
let ts = &mut ctx.ts; |
| 239 |
estimate_tile_motion(fi, ts, bit_depth, cpu_feature_level); |
| 240 |
}); |
| 241 |
} |
| 242 |
|
| 243 |
fn estimate_tile_motion<T: Pixel>( |
| 244 |
fi: &FrameInvariants<T>, |
| 245 |
ts: &mut TileStateMut<'_, T>, |
| 246 |
bit_depth: usize, |
| 247 |
cpu_feature_level: CpuFeatureLevel, |
| 248 |
) { |
| 249 |
let init_size = MIB_SIZE_LOG2; |
| 250 |
|
| 251 |
let mut prev_ssdec: Option<u8> = None; |
| 252 |
for mv_size_in_b_log2 in (2..=init_size).rev() { |
| 253 |
let init = mv_size_in_b_log2 == init_size; |
| 254 |
|
| 255 |
// Choose subsampling. Pass one is quarter res and pass two is at half res. |
| 256 |
let ssdec = match init_size - mv_size_in_b_log2 { |
| 257 |
0 => 2, |
| 258 |
1 => 1, |
| 259 |
_ => 0, |
| 260 |
}; |
| 261 |
|
| 262 |
let new_subsampling = if let Some(prev) = prev_ssdec { |
| 263 |
prev != ssdec |
| 264 |
} else { |
| 265 |
false |
| 266 |
}; |
| 267 |
prev_ssdec = Some(ssdec); |
| 268 |
|
| 269 |
// 0.5 and 0.125 are a fudge factors |
| 270 |
let lambda = 0; |
| 271 |
|
| 272 |
for sby in 0..ts.sb_height { |
| 273 |
for sbx in 0..ts.sb_width { |
| 274 |
let mut tested_frames_flags = 0; |
| 275 |
for &ref_frame in ALLOWED_REF_FRAMES { |
| 276 |
let frame_flag = 1 << fi.ref_frames[ref_frame.to_index()]; |
| 277 |
if tested_frames_flags & frame_flag == frame_flag { |
| 278 |
continue; |
| 279 |
} |
| 280 |
tested_frames_flags |= frame_flag; |
| 281 |
|
| 282 |
let tile_bo = TileSuperBlockOffset(SuperBlockOffset { x: sbx, y: sby }) |
| 283 |
.block_offset(0, 0); |
| 284 |
|
| 285 |
if new_subsampling { |
| 286 |
refine_subsampled_sb_motion( |
| 287 |
fi, |
| 288 |
ts, |
| 289 |
ref_frame, |
| 290 |
mv_size_in_b_log2 + 1, |
| 291 |
tile_bo, |
| 292 |
ssdec, |
| 293 |
lambda, |
| 294 |
bit_depth, |
| 295 |
cpu_feature_level, |
| 296 |
); |
| 297 |
} |
| 298 |
|
| 299 |
estimate_sb_motion( |
| 300 |
fi, |
| 301 |
ts, |
| 302 |
ref_frame, |
| 303 |
mv_size_in_b_log2, |
| 304 |
tile_bo, |
| 305 |
init, |
| 306 |
ssdec, |
| 307 |
lambda, |
| 308 |
bit_depth, |
| 309 |
cpu_feature_level, |
| 310 |
); |
| 311 |
} |
| 312 |
} |
| 313 |
} |
| 314 |
} |
| 315 |
} |
| 316 |
|
| 317 |
#[allow(clippy::too_many_arguments)] |
| 318 |
fn refine_subsampled_sb_motion<T: Pixel>( |
| 319 |
fi: &FrameInvariants<T>, |
| 320 |
ts: &mut TileStateMut<'_, T>, |
| 321 |
ref_frame: RefType, |
| 322 |
mv_size_in_b_log2: usize, |
| 323 |
tile_bo: TileBlockOffset, |
| 324 |
ssdec: u8, |
| 325 |
lambda: u32, |
| 326 |
bit_depth: usize, |
| 327 |
cpu_feature_level: CpuFeatureLevel, |
| 328 |
) { |
| 329 |
let pix_offset = tile_bo.to_luma_plane_offset(); |
| 330 |
let sb_h: usize = SB_SIZE.min(ts.height - pix_offset.y as usize); |
| 331 |
let sb_w: usize = SB_SIZE.min(ts.width - pix_offset.x as usize); |
| 332 |
|
| 333 |
let mv_size = MI_SIZE << mv_size_in_b_log2; |
| 334 |
|
| 335 |
// Process in blocks, cropping at edges. |
| 336 |
for y in (0..sb_h).step_by(mv_size) { |
| 337 |
for x in (0..sb_w).step_by(mv_size) { |
| 338 |
let sub_bo = |
| 339 |
tile_bo.with_offset(x as isize >> MI_SIZE_LOG2, y as isize >> MI_SIZE_LOG2); |
| 340 |
|
| 341 |
// Clamp to frame edge, rounding up in the case of subsampling. |
| 342 |
// The rounding makes some assumptions about how subsampling is done. |
| 343 |
let w = mv_size.min(sb_w - x + (1 << ssdec) - 1) >> ssdec; |
| 344 |
let h = mv_size.min(sb_h - y + (1 << ssdec) - 1) >> ssdec; |
| 345 |
|
| 346 |
// Refine the existing motion estimate |
| 347 |
if let Some(results) = refine_subsampled_motion_estimate( |
| 348 |
fi, |
| 349 |
ts, |
| 350 |
w, |
| 351 |
h, |
| 352 |
sub_bo, |
| 353 |
ref_frame, |
| 354 |
ssdec, |
| 355 |
lambda, |
| 356 |
bit_depth, |
| 357 |
cpu_feature_level, |
| 358 |
) { |
| 359 |
// normalize sad to 128x128 block |
| 360 |
let sad = |
| 361 |
(((results.rd.sad as u64) << (MAX_SB_SIZE_LOG2 * 2)) / (w * h) as u64) as u32; |
| 362 |
save_me_stats(ts, mv_size_in_b_log2, sub_bo, ref_frame, MEStats { |
| 363 |
mv: results.mv, |
| 364 |
normalized_sad: sad, |
| 365 |
}); |
| 366 |
} |
| 367 |
} |
| 368 |
} |
| 369 |
} |
| 370 |
|
| 371 |
/// Refine motion estimation that was computed one level of subsampling up. |
| 372 |
#[allow(clippy::too_many_arguments)] |
| 373 |
fn refine_subsampled_motion_estimate<T: Pixel>( |
| 374 |
fi: &FrameInvariants<T>, |
| 375 |
ts: &TileStateMut<'_, T>, |
| 376 |
w: usize, |
| 377 |
h: usize, |
| 378 |
tile_bo: TileBlockOffset, |
| 379 |
ref_frame: RefType, |
| 380 |
ssdec: u8, |
| 381 |
lambda: u32, |
| 382 |
bit_depth: usize, |
| 383 |
cpu_feature_level: CpuFeatureLevel, |
| 384 |
) -> Option<MotionSearchResult> { |
| 385 |
if let Some(ref rec) = fi.rec_buffer.frames[fi.ref_frames[ref_frame.to_index()] as usize] { |
| 386 |
let frame_bo = ts.to_frame_block_offset(tile_bo); |
| 387 |
let (mvx_min, mvx_max, mvy_min, mvy_max) = |
| 388 |
get_mv_range(fi.w_in_b, fi.h_in_b, frame_bo, w << ssdec, h << ssdec); |
| 389 |
|
| 390 |
let pmv = [MotionVector { row: 0, col: 0 }; 2]; |
| 391 |
|
| 392 |
let po = frame_bo.to_luma_plane_offset(); |
| 393 |
let (mvx_min, mvx_max, mvy_min, mvy_max) = ( |
| 394 |
mvx_min >> ssdec, |
| 395 |
mvx_max >> ssdec, |
| 396 |
mvy_min >> ssdec, |
| 397 |
mvy_max >> ssdec, |
| 398 |
); |
| 399 |
let po = PlaneOffset { |
| 400 |
x: po.x >> ssdec, |
| 401 |
y: po.y >> ssdec, |
| 402 |
}; |
| 403 |
let p_ref = match ssdec { |
| 404 |
0 => &rec.frame.planes[0], |
| 405 |
1 => &rec.input_hres, |
| 406 |
2 => &rec.input_qres, |
| 407 |
_ => unimplemented!(), |
| 408 |
}; |
| 409 |
|
| 410 |
let org_region = &match ssdec { |
| 411 |
0 => ts.input_tile.planes[0].subregion(Area::BlockStartingAt { bo: tile_bo.0 }), |
| 412 |
1 => ts.input_hres.region(Area::StartingAt { x: po.x, y: po.y }), |
| 413 |
2 => ts.input_qres.region(Area::StartingAt { x: po.x, y: po.y }), |
| 414 |
_ => unimplemented!(), |
| 415 |
}; |
| 416 |
|
| 417 |
let mv = ts.me_stats[ref_frame.to_index()][tile_bo.0.y][tile_bo.0.x].mv >> ssdec; |
| 418 |
|
| 419 |
// Given a motion vector at 0 at higher subsampling: |
| 420 |
// | -1 | 0 | 1 | |
| 421 |
// then the vectors at -1 to 2 should be tested at the current subsampling. |
| 422 |
// |-------------| |
| 423 |
// | -2 -1 | 0 1 | 2 3 | |
| 424 |
// This corresponds to a 4x4 full search. |
| 425 |
let x_lo = po.x + (mv.col as isize / 8 - 1).max(mvx_min / 8); |
| 426 |
let x_hi = po.x + (mv.col as isize / 8 + 2).min(mvx_max / 8); |
| 427 |
let y_lo = po.y + (mv.row as isize / 8 - 1).max(mvy_min / 8); |
| 428 |
let y_hi = po.y + (mv.row as isize / 8 + 2).min(mvy_max / 8); |
| 429 |
let mut results = full_search( |
| 430 |
x_lo, |
| 431 |
x_hi, |
| 432 |
y_lo, |
| 433 |
y_hi, |
| 434 |
w, |
| 435 |
h, |
| 436 |
org_region, |
| 437 |
p_ref, |
| 438 |
po, |
| 439 |
1, |
| 440 |
lambda, |
| 441 |
pmv, |
| 442 |
bit_depth, |
| 443 |
cpu_feature_level, |
| 444 |
); |
| 445 |
|
| 446 |
// Scale motion vectors to full res size |
| 447 |
results.mv = results.mv << ssdec; |
| 448 |
|
| 449 |
Some(results) |
| 450 |
} else { |
| 451 |
None |
| 452 |
} |
| 453 |
} |
| 454 |
|
| 455 |
fn get_mv_range( |
| 456 |
w_in_b: usize, |
| 457 |
h_in_b: usize, |
| 458 |
bo: PlaneBlockOffset, |
| 459 |
blk_w: usize, |
| 460 |
blk_h: usize, |
| 461 |
) -> (isize, isize, isize, isize) { |
| 462 |
let border_w = 128 + blk_w as isize * 8; |
| 463 |
let border_h = 128 + blk_h as isize * 8; |
| 464 |
let mvx_min = -(bo.0.x as isize) * (8 * MI_SIZE) as isize - border_w; |
| 465 |
let mvx_max = ((w_in_b - bo.0.x) as isize - (blk_w / MI_SIZE) as isize) |
| 466 |
* (8 * MI_SIZE) as isize |
| 467 |
+ border_w; |
| 468 |
let mvy_min = -(bo.0.y as isize) * (8 * MI_SIZE) as isize - border_h; |
| 469 |
let mvy_max = ((h_in_b - bo.0.y) as isize - (blk_h / MI_SIZE) as isize) |
| 470 |
* (8 * MI_SIZE) as isize |
| 471 |
+ border_h; |
| 472 |
|
| 473 |
// <https://aomediacodec.github.io/av1-spec/#assign-mv-semantics> |
| 474 |
( |
| 475 |
mvx_min.max(MV_LOW as isize + 1), |
| 476 |
mvx_max.min(MV_UPP as isize - 1), |
| 477 |
mvy_min.max(MV_LOW as isize + 1), |
| 478 |
mvy_max.min(MV_UPP as isize - 1), |
| 479 |
) |
| 480 |
} |
| 481 |
|
| 482 |
#[allow(clippy::too_many_arguments)] |
| 483 |
fn full_search<T: Pixel>( |
| 484 |
x_lo: isize, |
| 485 |
x_hi: isize, |
| 486 |
y_lo: isize, |
| 487 |
y_hi: isize, |
| 488 |
w: usize, |
| 489 |
h: usize, |
| 490 |
org_region: &PlaneRegion<T>, |
| 491 |
p_ref: &Plane<T>, |
| 492 |
po: PlaneOffset, |
| 493 |
step: usize, |
| 494 |
lambda: u32, |
| 495 |
pmv: [MotionVector; 2], |
| 496 |
bit_depth: usize, |
| 497 |
cpu_feature_level: CpuFeatureLevel, |
| 498 |
) -> MotionSearchResult { |
| 499 |
let search_region = p_ref.region(Area::Rect(Rect { |
| 500 |
x: x_lo, |
| 501 |
y: y_lo, |
| 502 |
width: (x_hi - x_lo) as usize + w, |
| 503 |
height: (y_hi - y_lo) as usize + h, |
| 504 |
})); |
| 505 |
|
| 506 |
let mut best: MotionSearchResult = MotionSearchResult::empty(); |
| 507 |
|
| 508 |
// Select rectangular regions within search region with vert+horz windows |
| 509 |
for vert_window in search_region.vert_windows(h).step_by(step) { |
| 510 |
for ref_window in vert_window.horz_windows(w).step_by(step) { |
| 511 |
let &Rect { x, y, .. } = ref_window.rect(); |
| 512 |
|
| 513 |
let mv = MotionVector { |
| 514 |
row: 8 * (y as i16 - po.y as i16), |
| 515 |
col: 8 * (x as i16 - po.x as i16), |
| 516 |
}; |
| 517 |
|
| 518 |
let rd = compute_mv_rd( |
| 519 |
pmv, |
| 520 |
lambda, |
| 521 |
false, |
| 522 |
bit_depth, |
| 523 |
w, |
| 524 |
h, |
| 525 |
mv, |
| 526 |
org_region, |
| 527 |
&ref_window, |
| 528 |
cpu_feature_level, |
| 529 |
); |
| 530 |
|
| 531 |
if rd.cost < best.rd.cost { |
| 532 |
best.rd = rd; |
| 533 |
best.mv = mv; |
| 534 |
} |
| 535 |
} |
| 536 |
} |
| 537 |
|
| 538 |
best |
| 539 |
} |
| 540 |
|
| 541 |
/// Compute the rate distortion stats for a motion vector. |
| 542 |
#[allow(clippy::too_many_arguments)] |
| 543 |
fn compute_mv_rd<T: Pixel>( |
| 544 |
pmv: [MotionVector; 2], |
| 545 |
lambda: u32, |
| 546 |
use_satd: bool, |
| 547 |
bit_depth: usize, |
| 548 |
w: usize, |
| 549 |
h: usize, |
| 550 |
cand_mv: MotionVector, |
| 551 |
plane_org: &PlaneRegion<'_, T>, |
| 552 |
plane_ref: &PlaneRegion<'_, T>, |
| 553 |
cpu_feature_level: CpuFeatureLevel, |
| 554 |
) -> MVCandidateRD { |
| 555 |
let sad = if use_satd { |
| 556 |
get_satd(plane_org, plane_ref, w, h, bit_depth, cpu_feature_level) |
| 557 |
} else { |
| 558 |
get_sad(plane_org, plane_ref, w, h, bit_depth, cpu_feature_level) |
| 559 |
}; |
| 560 |
|
| 561 |
let rate1 = get_mv_rate(cand_mv, pmv[0]); |
| 562 |
let rate2 = get_mv_rate(cand_mv, pmv[1]); |
| 563 |
let rate = rate1.min(rate2 + 1); |
| 564 |
|
| 565 |
MVCandidateRD { |
| 566 |
cost: 256 * sad as u64 + rate as u64 * lambda as u64, |
| 567 |
sad, |
| 568 |
} |
| 569 |
} |
| 570 |
|
| 571 |
fn diff_to_rate(diff: i16) -> u32 { |
| 572 |
let d = diff >> 1; |
| 573 |
2 * ILog::ilog(d.abs()) as u32 |
| 574 |
} |
| 575 |
|
| 576 |
fn get_mv_rate(a: MotionVector, b: MotionVector) -> u32 { |
| 577 |
diff_to_rate(a.row - b.row) + diff_to_rate(a.col - b.col) |
| 578 |
} |
| 579 |
|
| 580 |
/// Result of motion search. |
| 581 |
#[derive(Debug, Copy, Clone)] |
| 582 |
pub struct MotionSearchResult { |
| 583 |
/// Motion vector chosen by the motion search. |
| 584 |
pub mv: MotionVector, |
| 585 |
/// Rate distortion data associated with `mv`. |
| 586 |
pub rd: MVCandidateRD, |
| 587 |
} |
| 588 |
|
| 589 |
impl MotionSearchResult { |
| 590 |
/// Creates an 'empty' value. |
| 591 |
/// |
| 592 |
/// To be considered empty, cost is set higher than any naturally occurring |
| 593 |
/// cost value. The idea is that comparing to any valid rd output, the |
| 594 |
/// search result will always be replaced. |
| 595 |
pub fn empty() -> MotionSearchResult { |
| 596 |
MotionSearchResult { |
| 597 |
mv: MotionVector::default(), |
| 598 |
rd: MVCandidateRD::empty(), |
| 599 |
} |
| 600 |
} |
| 601 |
|
| 602 |
/// Check if the value should be considered to be empty. |
| 603 |
const fn is_empty(&self) -> bool { |
| 604 |
self.rd.cost == u64::MAX |
| 605 |
} |
| 606 |
} |
| 607 |
|
| 608 |
/// Holds data from computing rate distortion of a motion vector. |
| 609 |
#[derive(Debug, Copy, Clone)] |
| 610 |
pub struct MVCandidateRD { |
| 611 |
/// Rate distortion cost of the motion vector. |
| 612 |
pub cost: u64, |
| 613 |
/// Distortion metric value for the motion vector. |
| 614 |
pub sad: u32, |
| 615 |
} |
| 616 |
|
| 617 |
impl MVCandidateRD { |
| 618 |
/// Creates an 'empty' value. |
| 619 |
/// |
| 620 |
/// To be considered empty, cost is set higher than any naturally occurring |
| 621 |
/// cost value. The idea is that comparing to any valid rd output, the |
| 622 |
/// search result will always be replaced. |
| 623 |
const fn empty() -> MVCandidateRD { |
| 624 |
MVCandidateRD { |
| 625 |
sad: u32::MAX, |
| 626 |
cost: u64::MAX, |
| 627 |
} |
| 628 |
} |
| 629 |
} |
| 630 |
|
| 631 |
fn save_me_stats<T: Pixel>( |
| 632 |
ts: &mut TileStateMut<'_, T>, |
| 633 |
mv_size_in_b_log2: usize, |
| 634 |
tile_bo: TileBlockOffset, |
| 635 |
ref_frame: RefType, |
| 636 |
stats: MEStats, |
| 637 |
) { |
| 638 |
let size_in_b = 1 << mv_size_in_b_log2; |
| 639 |
let tile_me_stats = &mut ts.me_stats[ref_frame.to_index()]; |
| 640 |
let tile_bo_x_end = (tile_bo.0.x + size_in_b).min(ts.mi_width); |
| 641 |
let tile_bo_y_end = (tile_bo.0.y + size_in_b).min(ts.mi_height); |
| 642 |
for mi_y in tile_bo.0.y..tile_bo_y_end { |
| 643 |
for a in tile_me_stats[mi_y][tile_bo.0.x..tile_bo_x_end].iter_mut() { |
| 644 |
*a = stats; |
| 645 |
} |
| 646 |
} |
| 647 |
} |
| 648 |
|
| 649 |
#[allow(clippy::too_many_arguments)] |
| 650 |
fn estimate_sb_motion<T: Pixel>( |
| 651 |
fi: &FrameInvariants<T>, |
| 652 |
ts: &mut TileStateMut<'_, T>, |
| 653 |
ref_frame: RefType, |
| 654 |
mv_size_in_b_log2: usize, |
| 655 |
tile_bo: TileBlockOffset, |
| 656 |
init: bool, |
| 657 |
ssdec: u8, |
| 658 |
lambda: u32, |
| 659 |
bit_depth: usize, |
| 660 |
cpu_feature_level: CpuFeatureLevel, |
| 661 |
) { |
| 662 |
let pix_offset = tile_bo.to_luma_plane_offset(); |
| 663 |
let sb_h: usize = SB_SIZE.min(ts.height - pix_offset.y as usize); |
| 664 |
let sb_w: usize = SB_SIZE.min(ts.width - pix_offset.x as usize); |
| 665 |
|
| 666 |
let mv_size = MI_SIZE << mv_size_in_b_log2; |
| 667 |
|
| 668 |
// Process in blocks, cropping at edges. |
| 669 |
for y in (0..sb_h).step_by(mv_size) { |
| 670 |
for x in (0..sb_w).step_by(mv_size) { |
| 671 |
let corner: MVSamplingMode = if init { |
| 672 |
MVSamplingMode::INIT |
| 673 |
} else { |
| 674 |
// Processing the block a size up produces data that can be used by |
| 675 |
// the right and bottom corners. |
| 676 |
MVSamplingMode::CORNER { |
| 677 |
right: x & mv_size == mv_size, |
| 678 |
bottom: y & mv_size == mv_size, |
| 679 |
} |
| 680 |
}; |
| 681 |
|
| 682 |
let sub_bo = |
| 683 |
tile_bo.with_offset(x as isize >> MI_SIZE_LOG2, y as isize >> MI_SIZE_LOG2); |
| 684 |
|
| 685 |
// Clamp to frame edge, rounding up in the case of subsampling. |
| 686 |
// The rounding makes some assumptions about how subsampling is done. |
| 687 |
let w = mv_size.min(sb_w - x + (1 << ssdec) - 1) >> ssdec; |
| 688 |
let h = mv_size.min(sb_h - y + (1 << ssdec) - 1) >> ssdec; |
| 689 |
|
| 690 |
// Run motion estimation. |
| 691 |
// Note that the initial search (init) instructs the called function to |
| 692 |
// perform a more extensive search. |
| 693 |
if let Some(results) = estimate_motion( |
| 694 |
fi, |
| 695 |
ts, |
| 696 |
w, |
| 697 |
h, |
| 698 |
sub_bo, |
| 699 |
ref_frame, |
| 700 |
None, |
| 701 |
corner, |
| 702 |
init, |
| 703 |
ssdec, |
| 704 |
Some(lambda), |
| 705 |
bit_depth, |
| 706 |
cpu_feature_level, |
| 707 |
) { |
| 708 |
// normalize sad to 128x128 block |
| 709 |
let sad = |
| 710 |
(((results.rd.sad as u64) << (MAX_SB_SIZE_LOG2 * 2)) / (w * h) as u64) as u32; |
| 711 |
save_me_stats(ts, mv_size_in_b_log2, sub_bo, ref_frame, MEStats { |
| 712 |
mv: results.mv, |
| 713 |
normalized_sad: sad, |
| 714 |
}); |
| 715 |
} |
| 716 |
} |
| 717 |
} |
| 718 |
} |
| 719 |
|
| 720 |
#[allow(clippy::too_many_arguments)] |
| 721 |
fn estimate_motion<T: Pixel>( |
| 722 |
fi: &FrameInvariants<T>, |
| 723 |
ts: &TileStateMut<'_, T>, |
| 724 |
w: usize, |
| 725 |
h: usize, |
| 726 |
tile_bo: TileBlockOffset, |
| 727 |
ref_frame: RefType, |
| 728 |
pmv: Option<[MotionVector; 2]>, |
| 729 |
corner: MVSamplingMode, |
| 730 |
extensive_search: bool, |
| 731 |
ssdec: u8, |
| 732 |
lambda: Option<u32>, |
| 733 |
bit_depth: usize, |
| 734 |
cpu_feature_level: CpuFeatureLevel, |
| 735 |
) -> Option<MotionSearchResult> { |
| 736 |
if let Some(ref rec) = fi.rec_buffer.frames[fi.ref_frames[ref_frame.to_index()] as usize] { |
| 737 |
let frame_bo = ts.to_frame_block_offset(tile_bo); |
| 738 |
let (mvx_min, mvx_max, mvy_min, mvy_max) = |
| 739 |
get_mv_range(fi.w_in_b, fi.h_in_b, frame_bo, w << ssdec, h << ssdec); |
| 740 |
|
| 741 |
let lambda = lambda.unwrap_or(0); |
| 742 |
|
| 743 |
let global_mv = [MotionVector { row: 0, col: 0 }; 2]; |
| 744 |
|
| 745 |
let po = frame_bo.to_luma_plane_offset(); |
| 746 |
let (mvx_min, mvx_max, mvy_min, mvy_max) = ( |
| 747 |
mvx_min >> ssdec, |
| 748 |
mvx_max >> ssdec, |
| 749 |
mvy_min >> ssdec, |
| 750 |
mvy_max >> ssdec, |
| 751 |
); |
| 752 |
let po = PlaneOffset { |
| 753 |
x: po.x >> ssdec, |
| 754 |
y: po.y >> ssdec, |
| 755 |
}; |
| 756 |
let p_ref = match ssdec { |
| 757 |
0 => &rec.frame.planes[0], |
| 758 |
1 => &rec.input_hres, |
| 759 |
2 => &rec.input_qres, |
| 760 |
_ => unimplemented!(), |
| 761 |
}; |
| 762 |
|
| 763 |
let org_region = &match ssdec { |
| 764 |
0 => ts.input_tile.planes[0].subregion(Area::BlockStartingAt { bo: tile_bo.0 }), |
| 765 |
1 => ts.input_hres.region(Area::StartingAt { x: po.x, y: po.y }), |
| 766 |
2 => ts.input_qres.region(Area::StartingAt { x: po.x, y: po.y }), |
| 767 |
_ => unimplemented!(), |
| 768 |
}; |
| 769 |
|
| 770 |
let mut best: MotionSearchResult = full_pixel_me( |
| 771 |
fi, |
| 772 |
ts, |
| 773 |
org_region, |
| 774 |
p_ref, |
| 775 |
tile_bo, |
| 776 |
po, |
| 777 |
lambda, |
| 778 |
pmv.unwrap_or(global_mv), |
| 779 |
w, |
| 780 |
h, |
| 781 |
mvx_min, |
| 782 |
mvx_max, |
| 783 |
mvy_min, |
| 784 |
mvy_max, |
| 785 |
ref_frame, |
| 786 |
corner, |
| 787 |
extensive_search, |
| 788 |
ssdec, |
| 789 |
bit_depth, |
| 790 |
cpu_feature_level, |
| 791 |
); |
| 792 |
|
| 793 |
if let Some(pmv) = pmv { |
| 794 |
best.rd = get_fullpel_mv_rd( |
| 795 |
po, |
| 796 |
org_region, |
| 797 |
p_ref, |
| 798 |
bit_depth, |
| 799 |
pmv, |
| 800 |
lambda, |
| 801 |
true, |
| 802 |
mvx_min, |
| 803 |
mvx_max, |
| 804 |
mvy_min, |
| 805 |
mvy_max, |
| 806 |
w, |
| 807 |
h, |
| 808 |
best.mv, |
| 809 |
cpu_feature_level, |
| 810 |
); |
| 811 |
|
| 812 |
sub_pixel_me( |
| 813 |
fi, |
| 814 |
po, |
| 815 |
org_region, |
| 816 |
p_ref, |
| 817 |
lambda, |
| 818 |
pmv, |
| 819 |
mvx_min, |
| 820 |
mvx_max, |
| 821 |
mvy_min, |
| 822 |
mvy_max, |
| 823 |
w, |
| 824 |
h, |
| 825 |
true, |
| 826 |
&mut best, |
| 827 |
ref_frame, |
| 828 |
bit_depth, |
| 829 |
cpu_feature_level, |
| 830 |
); |
| 831 |
} |
| 832 |
|
| 833 |
// Scale motion vectors to full res size |
| 834 |
best.mv = best.mv << ssdec; |
| 835 |
|
| 836 |
Some(best) |
| 837 |
} else { |
| 838 |
None |
| 839 |
} |
| 840 |
} |
| 841 |
|
| 842 |
#[allow(clippy::too_many_arguments)] |
| 843 |
fn full_pixel_me<T: Pixel>( |
| 844 |
fi: &FrameInvariants<T>, |
| 845 |
ts: &TileStateMut<'_, T>, |
| 846 |
org_region: &PlaneRegion<T>, |
| 847 |
p_ref: &Plane<T>, |
| 848 |
tile_bo: TileBlockOffset, |
| 849 |
po: PlaneOffset, |
| 850 |
lambda: u32, |
| 851 |
pmv: [MotionVector; 2], |
| 852 |
w: usize, |
| 853 |
h: usize, |
| 854 |
mvx_min: isize, |
| 855 |
mvx_max: isize, |
| 856 |
mvy_min: isize, |
| 857 |
mvy_max: isize, |
| 858 |
ref_frame: RefType, |
| 859 |
corner: MVSamplingMode, |
| 860 |
extensive_search: bool, |
| 861 |
ssdec: u8, |
| 862 |
bit_depth: usize, |
| 863 |
cpu_feature_level: CpuFeatureLevel, |
| 864 |
) -> MotionSearchResult { |
| 865 |
let ref_frame_id = ref_frame.to_index(); |
| 866 |
let tile_me_stats = &ts.me_stats[ref_frame_id].as_const(); |
| 867 |
let frame_ref = fi.rec_buffer.frames[fi.ref_frames[0] as usize] |
| 868 |
.as_ref() |
| 869 |
.map(|frame_ref| frame_ref.frame_me_stats.read().expect("poisoned lock")); |
| 870 |
let subsets = get_subset_predictors( |
| 871 |
tile_bo, |
| 872 |
tile_me_stats, |
| 873 |
frame_ref, |
| 874 |
ref_frame_id, |
| 875 |
w, |
| 876 |
h, |
| 877 |
mvx_min, |
| 878 |
mvx_max, |
| 879 |
mvy_min, |
| 880 |
mvy_max, |
| 881 |
corner, |
| 882 |
ssdec, |
| 883 |
); |
| 884 |
|
| 885 |
let try_cands = |predictors: &[MotionVector], best: &mut MotionSearchResult| { |
| 886 |
let mut results = get_best_predictor( |
| 887 |
po, |
| 888 |
org_region, |
| 889 |
p_ref, |
| 890 |
predictors, |
| 891 |
bit_depth, |
| 892 |
pmv, |
| 893 |
lambda, |
| 894 |
mvx_min, |
| 895 |
mvx_max, |
| 896 |
mvy_min, |
| 897 |
mvy_max, |
| 898 |
w, |
| 899 |
h, |
| 900 |
cpu_feature_level, |
| 901 |
); |
| 902 |
fullpel_diamond_search( |
| 903 |
po, |
| 904 |
org_region, |
| 905 |
p_ref, |
| 906 |
&mut results, |
| 907 |
bit_depth, |
| 908 |
pmv, |
| 909 |
lambda, |
| 910 |
mvx_min, |
| 911 |
mvx_max, |
| 912 |
mvy_min, |
| 913 |
mvy_max, |
| 914 |
w, |
| 915 |
h, |
| 916 |
cpu_feature_level, |
| 917 |
); |
| 918 |
|
| 919 |
if results.rd.cost < best.rd.cost { |
| 920 |
*best = results; |
| 921 |
} |
| 922 |
}; |
| 923 |
|
| 924 |
let mut best: MotionSearchResult = MotionSearchResult::empty(); |
| 925 |
if !extensive_search { |
| 926 |
try_cands(&subsets.all_mvs(), &mut best); |
| 927 |
best |
| 928 |
} else { |
| 929 |
// Perform a more thorough search before resorting to full search. |
| 930 |
// Search the median, the best mvs of neighboring blocks, and motion vectors |
| 931 |
// from the previous frame. Stop once a candidate with a sad less than a |
| 932 |
// threshold is found. |
| 933 |
|
| 934 |
let thresh = (subsets.min_sad as f32 * 1.2) as u32 + (((w * h) as u32) << (bit_depth - 8)); |
| 935 |
|
| 936 |
if let Some(median) = subsets.median { |
| 937 |
try_cands(&[median], &mut best); |
| 938 |
|
| 939 |
if best.rd.sad < thresh { |
| 940 |
return best; |
| 941 |
} |
| 942 |
} |
| 943 |
|
| 944 |
try_cands(&subsets.subset_b, &mut best); |
| 945 |
|
| 946 |
if best.rd.sad < thresh { |
| 947 |
return best; |
| 948 |
} |
| 949 |
|
| 950 |
try_cands(&subsets.subset_c, &mut best); |
| 951 |
|
| 952 |
if best.rd.sad < thresh { |
| 953 |
return best; |
| 954 |
} |
| 955 |
|
| 956 |
// Preform UMH search, either as the last possible search when full search |
| 957 |
// is disabled, or as the last search before resorting to full search. |
| 958 |
// Use 24 merange, since it is the largest range that x264 uses. |
| 959 |
uneven_multi_hex_search( |
| 960 |
po, |
| 961 |
org_region, |
| 962 |
p_ref, |
| 963 |
&mut best, |
| 964 |
bit_depth, |
| 965 |
pmv, |
| 966 |
lambda, |
| 967 |
mvx_min, |
| 968 |
mvx_max, |
| 969 |
mvy_min, |
| 970 |
mvy_max, |
| 971 |
w, |
| 972 |
h, |
| 973 |
24, |
| 974 |
cpu_feature_level, |
| 975 |
); |
| 976 |
|
| 977 |
best |
| 978 |
} |
| 979 |
} |
| 980 |
|
| 981 |
#[allow(clippy::too_many_arguments)] |
| 982 |
fn sub_pixel_me<T: Pixel>( |
| 983 |
fi: &FrameInvariants<T>, |
| 984 |
po: PlaneOffset, |
| 985 |
org_region: &PlaneRegion<T>, |
| 986 |
p_ref: &Plane<T>, |
| 987 |
lambda: u32, |
| 988 |
pmv: [MotionVector; 2], |
| 989 |
mvx_min: isize, |
| 990 |
mvx_max: isize, |
| 991 |
mvy_min: isize, |
| 992 |
mvy_max: isize, |
| 993 |
w: usize, |
| 994 |
h: usize, |
| 995 |
use_satd: bool, |
| 996 |
best: &mut MotionSearchResult, |
| 997 |
ref_frame: RefType, |
| 998 |
bit_depth: usize, |
| 999 |
cpu_feature_level: CpuFeatureLevel, |
| 1000 |
) { |
| 1001 |
subpel_diamond_search( |
| 1002 |
fi, |
| 1003 |
po, |
| 1004 |
org_region, |
| 1005 |
p_ref, |
| 1006 |
bit_depth, |
| 1007 |
pmv, |
| 1008 |
lambda, |
| 1009 |
mvx_min, |
| 1010 |
mvx_max, |
| 1011 |
mvy_min, |
| 1012 |
mvy_max, |
| 1013 |
w, |
| 1014 |
h, |
| 1015 |
use_satd, |
| 1016 |
best, |
| 1017 |
ref_frame, |
| 1018 |
cpu_feature_level, |
| 1019 |
); |
| 1020 |
} |
| 1021 |
|
| 1022 |
/// Run a subpixel diamond search. The search is run on multiple step sizes. |
| 1023 |
/// |
| 1024 |
/// For each step size, candidate motion vectors are examined for improvement |
| 1025 |
/// to the current search location. The search location is moved to the best |
| 1026 |
/// candidate (if any). This is repeated until the search location stops moving. |
| 1027 |
#[allow(clippy::too_many_arguments)] |
| 1028 |
fn subpel_diamond_search<T: Pixel>( |
| 1029 |
fi: &FrameInvariants<T>, |
| 1030 |
po: PlaneOffset, |
| 1031 |
org_region: &PlaneRegion<T>, |
| 1032 |
_p_ref: &Plane<T>, |
| 1033 |
bit_depth: usize, |
| 1034 |
pmv: [MotionVector; 2], |
| 1035 |
lambda: u32, |
| 1036 |
mvx_min: isize, |
| 1037 |
mvx_max: isize, |
| 1038 |
mvy_min: isize, |
| 1039 |
mvy_max: isize, |
| 1040 |
w: usize, |
| 1041 |
h: usize, |
| 1042 |
use_satd: bool, |
| 1043 |
current: &mut MotionSearchResult, |
| 1044 |
ref_frame: RefType, |
| 1045 |
cpu_feature_level: CpuFeatureLevel, |
| 1046 |
) { |
| 1047 |
// Motion compensation assembly has special requirements for edges |
| 1048 |
let mc_w = w.next_power_of_two(); |
| 1049 |
let mc_h = (h + 1) & !1; |
| 1050 |
|
| 1051 |
// Metadata for subpel scratch pad. |
| 1052 |
let cfg = PlaneConfig::new(mc_w, mc_h, 0, 0, 0, 0, std::mem::size_of::<T>()); |
| 1053 |
// Stack allocation for subpel scratch pad. |
| 1054 |
// SAFETY: We write to the array below before reading from it. |
| 1055 |
let mut buf: Aligned<A64, [T; 128 * 128]> = Aligned([T::cast_from(0); 128 * 128]); |
| 1056 |
let mut tmp_region = PlaneRegionMut::from_slice(buf.as_mut(), &cfg, Rect { |
| 1057 |
x: 0, |
| 1058 |
y: 0, |
| 1059 |
width: cfg.width, |
| 1060 |
height: cfg.height, |
| 1061 |
}); |
| 1062 |
|
| 1063 |
// start at 1/2 pel and end at 1/4 or 1/8 pel |
| 1064 |
let (mut diamond_radius_log2, diamond_radius_end_log2) = (2u8, 1u8); |
| 1065 |
|
| 1066 |
loop { |
| 1067 |
// Find the best candidate from the diamond pattern. |
| 1068 |
let mut best_cand: MotionSearchResult = MotionSearchResult::empty(); |
| 1069 |
for &offset in &DIAMOND_R1_PATTERN_SUBPEL { |
| 1070 |
let cand_mv = current.mv + (offset << diamond_radius_log2); |
| 1071 |
|
| 1072 |
let rd = get_subpel_mv_rd( |
| 1073 |
fi, |
| 1074 |
po, |
| 1075 |
org_region, |
| 1076 |
bit_depth, |
| 1077 |
pmv, |
| 1078 |
lambda, |
| 1079 |
use_satd, |
| 1080 |
mvx_min, |
| 1081 |
mvx_max, |
| 1082 |
mvy_min, |
| 1083 |
mvy_max, |
| 1084 |
w, |
| 1085 |
h, |
| 1086 |
cand_mv, |
| 1087 |
&mut tmp_region, |
| 1088 |
ref_frame, |
| 1089 |
cpu_feature_level, |
| 1090 |
); |
| 1091 |
|
| 1092 |
if rd.cost < best_cand.rd.cost { |
| 1093 |
best_cand.mv = cand_mv; |
| 1094 |
best_cand.rd = rd; |
| 1095 |
} |
| 1096 |
} |
| 1097 |
|
| 1098 |
// Continue the search at this scale until a better candidate isn't found. |
| 1099 |
if current.rd.cost <= best_cand.rd.cost { |
| 1100 |
if diamond_radius_log2 == diamond_radius_end_log2 { |
| 1101 |
break; |
| 1102 |
} else { |
| 1103 |
diamond_radius_log2 -= 1; |
| 1104 |
} |
| 1105 |
} else { |
| 1106 |
*current = best_cand; |
| 1107 |
} |
| 1108 |
} |
| 1109 |
|
| 1110 |
assert!(!current.is_empty()); |
| 1111 |
} |
| 1112 |
|
| 1113 |
#[allow(clippy::too_many_arguments)] |
| 1114 |
fn get_subpel_mv_rd<T: Pixel>( |
| 1115 |
fi: &FrameInvariants<T>, |
| 1116 |
po: PlaneOffset, |
| 1117 |
org_region: &PlaneRegion<T>, |
| 1118 |
bit_depth: usize, |
| 1119 |
pmv: [MotionVector; 2], |
| 1120 |
lambda: u32, |
| 1121 |
use_satd: bool, |
| 1122 |
mvx_min: isize, |
| 1123 |
mvx_max: isize, |
| 1124 |
mvy_min: isize, |
| 1125 |
mvy_max: isize, |
| 1126 |
w: usize, |
| 1127 |
h: usize, |
| 1128 |
cand_mv: MotionVector, |
| 1129 |
tmp_region: &mut PlaneRegionMut<T>, |
| 1130 |
ref_frame: RefType, |
| 1131 |
cpu_feature_level: CpuFeatureLevel, |
| 1132 |
) -> MVCandidateRD { |
| 1133 |
if (cand_mv.col as isize) < mvx_min |
| 1134 |
|| (cand_mv.col as isize) > mvx_max |
| 1135 |
|| (cand_mv.row as isize) < mvy_min |
| 1136 |
|| (cand_mv.row as isize) > mvy_max |
| 1137 |
{ |
| 1138 |
return MVCandidateRD::empty(); |
| 1139 |
} |
| 1140 |
|
| 1141 |
let tmp_width = tmp_region.rect().width; |
| 1142 |
let tmp_height = tmp_region.rect().height; |
| 1143 |
let tile_rect = TileRect { |
| 1144 |
x: 0, |
| 1145 |
y: 0, |
| 1146 |
width: tmp_width, |
| 1147 |
height: tmp_height, |
| 1148 |
}; |
| 1149 |
|
| 1150 |
PredictionMode::NEWMV.predict_inter_single( |
| 1151 |
fi, |
| 1152 |
tile_rect, |
| 1153 |
0, |
| 1154 |
po, |
| 1155 |
tmp_region, |
| 1156 |
// motion comp's w & h on edges can be different from distortion's |
| 1157 |
tmp_width, |
| 1158 |
tmp_height, |
| 1159 |
ref_frame, |
| 1160 |
cand_mv, |
| 1161 |
bit_depth, |
| 1162 |
cpu_feature_level, |
| 1163 |
); |
| 1164 |
let plane_ref = tmp_region.as_const(); |
| 1165 |
compute_mv_rd( |
| 1166 |
pmv, |
| 1167 |
lambda, |
| 1168 |
use_satd, |
| 1169 |
bit_depth, |
| 1170 |
w, |
| 1171 |
h, |
| 1172 |
cand_mv, |
| 1173 |
org_region, |
| 1174 |
&plane_ref, |
| 1175 |
cpu_feature_level, |
| 1176 |
) |
| 1177 |
} |
| 1178 |
|
| 1179 |
/// Perform an uneven multi-hexagon search. There are 4 stages: |
| 1180 |
/// 1. Unsymmetrical-cross search: Search the horizontal and vertical directions |
| 1181 |
/// for the general direction of the motion. |
| 1182 |
/// 2. A 5x5 full search is done to refine the current candidate. |
| 1183 |
/// 3. Uneven multi-hexagon search. See [`UMH_PATTERN`]. |
| 1184 |
/// 4. Refinement using standard hexagon search. |
| 1185 |
/// |
| 1186 |
/// `current` provides the initial search location and serves as |
| 1187 |
/// the output for the final search results. |
| 1188 |
/// |
| 1189 |
/// `me_range` parameter determines how far these stages can search. |
| 1190 |
#[allow(clippy::too_many_arguments)] |
| 1191 |
fn uneven_multi_hex_search<T: Pixel>( |
| 1192 |
po: PlaneOffset, |
| 1193 |
org_region: &PlaneRegion<T>, |
| 1194 |
p_ref: &Plane<T>, |
| 1195 |
current: &mut MotionSearchResult, |
| 1196 |
bit_depth: usize, |
| 1197 |
pmv: [MotionVector; 2], |
| 1198 |
lambda: u32, |
| 1199 |
mvx_min: isize, |
| 1200 |
mvx_max: isize, |
| 1201 |
mvy_min: isize, |
| 1202 |
mvy_max: isize, |
| 1203 |
w: usize, |
| 1204 |
h: usize, |
| 1205 |
me_range: i16, |
| 1206 |
cpu_feature_level: CpuFeatureLevel, |
| 1207 |
) { |
| 1208 |
assert!(!current.is_empty()); |
| 1209 |
|
| 1210 |
// Search in a cross pattern to obtain a rough approximate of motion. |
| 1211 |
// The cross is split into a horizontal and vertical component. Video content |
| 1212 |
// tends to have more horizontal motion, so the horizontal part of the cross |
| 1213 |
// is twice as large as the vertical half. |
| 1214 |
// X - |
| 1215 |
// | <- me_range/2 |
| 1216 |
// X | |
| 1217 |
// X X X XoX X X X - |
| 1218 |
// X |
| 1219 |
// |
| 1220 |
// X |
| 1221 |
// |------| |
| 1222 |
// \ |
| 1223 |
// me_range |
| 1224 |
let center = current.mv; |
| 1225 |
|
| 1226 |
// The larger, horizontal, part of the cross search. |
| 1227 |
for i in (1..=me_range).step_by(2) { |
| 1228 |
const HORIZONTAL_LINE: [MotionVector; 2] = search_pattern!( |
| 1229 |
col: [ 0, 0], |
| 1230 |
row: [-1, 1] |
| 1231 |
); |
| 1232 |
|
| 1233 |
for &offset in &HORIZONTAL_LINE { |
| 1234 |
let cand_mv = center + offset * i; |
| 1235 |
let rd = get_fullpel_mv_rd( |
| 1236 |
po, |
| 1237 |
org_region, |
| 1238 |
p_ref, |
| 1239 |
bit_depth, |
| 1240 |
pmv, |
| 1241 |
lambda, |
| 1242 |
false, |
| 1243 |
mvx_min, |
| 1244 |
mvx_max, |
| 1245 |
mvy_min, |
| 1246 |
mvy_max, |
| 1247 |
w, |
| 1248 |
h, |
| 1249 |
cand_mv, |
| 1250 |
cpu_feature_level, |
| 1251 |
); |
| 1252 |
|
| 1253 |
if rd.cost < current.rd.cost { |
| 1254 |
current.mv = cand_mv; |
| 1255 |
current.rd = rd; |
| 1256 |
} |
| 1257 |
} |
| 1258 |
} |
| 1259 |
|
| 1260 |
// The smaller, vertical, part of the cross search |
| 1261 |
for i in (1..=me_range >> 1).step_by(2) { |
| 1262 |
const VERTICAL_LINE: [MotionVector; 2] = search_pattern!( |
| 1263 |
col: [-1, 1], |
| 1264 |
row: [ 0, 0] |
| 1265 |
); |
| 1266 |
|
| 1267 |
for &offset in &VERTICAL_LINE { |
| 1268 |
let cand_mv = center + offset * i; |
| 1269 |
let rd = get_fullpel_mv_rd( |
| 1270 |
po, |
| 1271 |
org_region, |
| 1272 |
p_ref, |
| 1273 |
bit_depth, |
| 1274 |
pmv, |
| 1275 |
lambda, |
| 1276 |
false, |
| 1277 |
mvx_min, |
| 1278 |
mvx_max, |
| 1279 |
mvy_min, |
| 1280 |
mvy_max, |
| 1281 |
w, |
| 1282 |
h, |
| 1283 |
cand_mv, |
| 1284 |
cpu_feature_level, |
| 1285 |
); |
| 1286 |
|
| 1287 |
if rd.cost < current.rd.cost { |
| 1288 |
current.mv = cand_mv; |
| 1289 |
current.rd = rd; |
| 1290 |
} |
| 1291 |
} |
| 1292 |
} |
| 1293 |
|
| 1294 |
// 5x5 full search. Search a 5x5 square region around the current best mv. |
| 1295 |
let center = current.mv; |
| 1296 |
for row in -2..=2 { |
| 1297 |
for col in -2..=2 { |
| 1298 |
if row == 0 && col == 0 { |
| 1299 |
continue; |
| 1300 |
} |
| 1301 |
let cand_mv = center + MotionVector { row, col }; |
| 1302 |
let rd = get_fullpel_mv_rd( |
| 1303 |
po, |
| 1304 |
org_region, |
| 1305 |
p_ref, |
| 1306 |
bit_depth, |
| 1307 |
pmv, |
| 1308 |
lambda, |
| 1309 |
false, |
| 1310 |
mvx_min, |
| 1311 |
mvx_max, |
| 1312 |
mvy_min, |
| 1313 |
mvy_max, |
| 1314 |
w, |
| 1315 |
h, |
| 1316 |
cand_mv, |
| 1317 |
cpu_feature_level, |
| 1318 |
); |
| 1319 |
|
| 1320 |
if rd.cost < current.rd.cost { |
| 1321 |
current.mv = cand_mv; |
| 1322 |
current.rd = rd; |
| 1323 |
} |
| 1324 |
} |
| 1325 |
} |
| 1326 |
|
| 1327 |
// Run the hexagons in uneven multi-hexagon. The hexagonal pattern is tested |
| 1328 |
// around the best vector at multiple scales. |
| 1329 |
// Example of the UMH pattern run on a scale of 1 and 2: |
| 1330 |
// 2 - |
| 1331 |
// | <- me_range |
| 1332 |
// 2 2 | |
| 1333 |
// | |
| 1334 |
// 2 1 2 | |
| 1335 |
// 1 1 | |
| 1336 |
// 2 1 1 2 | |
| 1337 |
// 1 1 | |
| 1338 |
// 2 1 o 1 2 | |
| 1339 |
// 1 1 | |
| 1340 |
// 2 1 1 2 | |
| 1341 |
// 1 1 | |
| 1342 |
// 2 1 2 | |
| 1343 |
// | |
| 1344 |
// 2 2 | |
| 1345 |
// | |
| 1346 |
// 2 - |
| 1347 |
// |---------------| |
| 1348 |
// \ |
| 1349 |
// me_range |
| 1350 |
let center = current.mv; |
| 1351 |
|
| 1352 |
// Divide by 4, the radius of the UMH's hexagon. |
| 1353 |
let iterations = me_range >> 2; |
| 1354 |
for i in 1..=iterations { |
| 1355 |
for &offset in &UMH_PATTERN { |
| 1356 |
let cand_mv = center + offset * i; |
| 1357 |
let rd = get_fullpel_mv_rd( |
| 1358 |
po, |
| 1359 |
org_region, |
| 1360 |
p_ref, |
| 1361 |
bit_depth, |
| 1362 |
pmv, |
| 1363 |
lambda, |
| 1364 |
false, |
| 1365 |
mvx_min, |
| 1366 |
mvx_max, |
| 1367 |
mvy_min, |
| 1368 |
mvy_max, |
| 1369 |
w, |
| 1370 |
h, |
| 1371 |
cand_mv, |
| 1372 |
cpu_feature_level, |
| 1373 |
); |
| 1374 |
|
| 1375 |
if rd.cost < current.rd.cost { |
| 1376 |
current.mv = cand_mv; |
| 1377 |
current.rd = rd; |
| 1378 |
} |
| 1379 |
} |
| 1380 |
} |
| 1381 |
|
| 1382 |
// Refine the search results using a 'normal' hexagon search. |
| 1383 |
hexagon_search( |
| 1384 |
po, |
| 1385 |
org_region, |
| 1386 |
p_ref, |
| 1387 |
current, |
| 1388 |
bit_depth, |
| 1389 |
pmv, |
| 1390 |
lambda, |
| 1391 |
mvx_min, |
| 1392 |
mvx_max, |
| 1393 |
mvy_min, |
| 1394 |
mvy_max, |
| 1395 |
w, |
| 1396 |
h, |
| 1397 |
cpu_feature_level, |
| 1398 |
); |
| 1399 |
} |
| 1400 |
|
| 1401 |
#[allow(clippy::too_many_arguments)] |
| 1402 |
fn get_subset_predictors( |
| 1403 |
tile_bo: TileBlockOffset, |
| 1404 |
tile_me_stats: &TileMEStats<'_>, |
| 1405 |
frame_ref_opt: Option<ReadGuardMEStats<'_>>, |
| 1406 |
ref_frame_id: usize, |
| 1407 |
pix_w: usize, |
| 1408 |
pix_h: usize, |
| 1409 |
mvx_min: isize, |
| 1410 |
mvx_max: isize, |
| 1411 |
mvy_min: isize, |
| 1412 |
mvy_max: isize, |
| 1413 |
corner: MVSamplingMode, |
| 1414 |
ssdec: u8, |
| 1415 |
) -> MotionEstimationSubsets { |
| 1416 |
let mut min_sad: u32 = u32::MAX; |
| 1417 |
let mut subset_b = ArrayVec::<MotionVector, 5>::new(); |
| 1418 |
let mut subset_c = ArrayVec::<MotionVector, 5>::new(); |
| 1419 |
|
| 1420 |
// rounded up width in blocks |
| 1421 |
let w = ((pix_w << ssdec) + MI_SIZE - 1) >> MI_SIZE_LOG2; |
| 1422 |
let h = ((pix_h << ssdec) + MI_SIZE - 1) >> MI_SIZE_LOG2; |
| 1423 |
|
| 1424 |
// Get predictors from the same frame. |
| 1425 |
|
| 1426 |
let clipped_half_w = (w >> 1).min(tile_me_stats.cols() - 1 - tile_bo.0.x); |
| 1427 |
let clipped_half_h = (h >> 1).min(tile_me_stats.rows() - 1 - tile_bo.0.y); |
| 1428 |
|
| 1429 |
let mut process_cand = |stats: MEStats| -> MotionVector { |
| 1430 |
min_sad = min_sad.min(stats.normalized_sad); |
| 1431 |
let mv = stats.mv.quantize_to_fullpel(); |
| 1432 |
MotionVector { |
| 1433 |
col: clamp(mv.col as isize, mvx_min, mvx_max) as i16, |
| 1434 |
row: clamp(mv.row as isize, mvy_min, mvy_max) as i16, |
| 1435 |
} |
| 1436 |
}; |
| 1437 |
|
| 1438 |
// Sample the middle of all block edges bordering this one. |
| 1439 |
// Note: If motion vectors haven't been precomputed to a given blocksize, then |
| 1440 |
// the right and bottom edges will be duplicates of the center predictor when |
| 1441 |
// processing in raster order. |
| 1442 |
|
| 1443 |
// left |
| 1444 |
if tile_bo.0.x > 0 { |
| 1445 |
subset_b.push(process_cand( |
| 1446 |
tile_me_stats[tile_bo.0.y + clipped_half_h][tile_bo.0.x - 1], |
| 1447 |
)); |
| 1448 |
} |
| 1449 |
// top |
| 1450 |
if tile_bo.0.y > 0 { |
| 1451 |
subset_b.push(process_cand( |
| 1452 |
tile_me_stats[tile_bo.0.y - 1][tile_bo.0.x + clipped_half_w], |
| 1453 |
)); |
| 1454 |
} |
| 1455 |
|
| 1456 |
// Sampling far right and far bottom edges was tested, but had worse results |
| 1457 |
// without an extensive threshold test (with threshold being applied after |
| 1458 |
// checking median and the best of each subset). |
| 1459 |
|
| 1460 |
// right |
| 1461 |
if let MVSamplingMode::CORNER { |
| 1462 |
right: true, |
| 1463 |
bottom: _, |
| 1464 |
} = corner |
| 1465 |
{ |
| 1466 |
if tile_bo.0.x + w < tile_me_stats.cols() { |
| 1467 |
subset_b.push(process_cand( |
| 1468 |
tile_me_stats[tile_bo.0.y + clipped_half_h][tile_bo.0.x + w], |
| 1469 |
)); |
| 1470 |
} |
| 1471 |
} |
| 1472 |
// bottom |
| 1473 |
if let MVSamplingMode::CORNER { |
| 1474 |
right: _, |
| 1475 |
bottom: true, |
| 1476 |
} = corner |
| 1477 |
{ |
| 1478 |
if tile_bo.0.y + h < tile_me_stats.rows() { |
| 1479 |
subset_b.push(process_cand( |
| 1480 |
tile_me_stats[tile_bo.0.y + h][tile_bo.0.x + clipped_half_w], |
| 1481 |
)); |
| 1482 |
} |
| 1483 |
} |
| 1484 |
|
| 1485 |
let median = if corner != MVSamplingMode::INIT { |
| 1486 |
// Sample the center of the current block. |
| 1487 |
Some(process_cand( |
| 1488 |
tile_me_stats[tile_bo.0.y + clipped_half_h][tile_bo.0.x + clipped_half_w], |
| 1489 |
)) |
| 1490 |
} else if subset_b.len() != 3 { |
| 1491 |
None |
| 1492 |
} else { |
| 1493 |
let mut rows: ArrayVec<i16, 3> = subset_b.iter().map(|&a| a.row).collect(); |
| 1494 |
let mut cols: ArrayVec<i16, 3> = subset_b.iter().map(|&a| a.col).collect(); |
| 1495 |
rows.as_mut_slice().sort_unstable(); |
| 1496 |
cols.as_mut_slice().sort_unstable(); |
| 1497 |
Some(MotionVector { |
| 1498 |
row: rows[1], |
| 1499 |
col: cols[1], |
| 1500 |
}) |
| 1501 |
}; |
| 1502 |
|
| 1503 |
// Zero motion vector, don't use add_cand since it skips zero vectors. |
| 1504 |
subset_b.push(MotionVector::default()); |
| 1505 |
|
| 1506 |
// EPZS subset C predictors. |
| 1507 |
// Sample the middle of bordering side of the left, right, top and bottom |
| 1508 |
// blocks of the previous frame. |
| 1509 |
// Sample the middle of this block in the previous frame. |
| 1510 |
|
| 1511 |
if let Some(frame_me_stats) = frame_ref_opt { |
| 1512 |
let prev_frame = &frame_me_stats[ref_frame_id]; |
| 1513 |
|
| 1514 |
let frame_bo = PlaneBlockOffset(BlockOffset { |
| 1515 |
x: tile_me_stats.x() + tile_bo.0.x, |
| 1516 |
y: tile_me_stats.y() + tile_bo.0.y, |
| 1517 |
}); |
| 1518 |
let clipped_half_w = (w >> 1).min(prev_frame.cols - 1 - frame_bo.0.x); |
| 1519 |
let clipped_half_h = (h >> 1).min(prev_frame.rows - 1 - frame_bo.0.y); |
| 1520 |
|
| 1521 |
// left |
| 1522 |
if frame_bo.0.x > 0 { |
| 1523 |
subset_c.push(process_cand( |
| 1524 |
prev_frame[frame_bo.0.y + clipped_half_h][frame_bo.0.x - 1], |
| 1525 |
)); |
| 1526 |
} |
| 1527 |
// top |
| 1528 |
if frame_bo.0.y > 0 { |
| 1529 |
subset_c.push(process_cand( |
| 1530 |
prev_frame[frame_bo.0.y - 1][frame_bo.0.x + clipped_half_w], |
| 1531 |
)); |
| 1532 |
} |
| 1533 |
// right |
| 1534 |
if frame_bo.0.x + w < prev_frame.cols { |
| 1535 |
subset_c.push(process_cand( |
| 1536 |
prev_frame[frame_bo.0.y + clipped_half_h][frame_bo.0.x + w], |
| 1537 |
)); |
| 1538 |
} |
| 1539 |
// bottom |
| 1540 |
if frame_bo.0.y + h < prev_frame.rows { |
| 1541 |
subset_c.push(process_cand( |
| 1542 |
prev_frame[frame_bo.0.y + h][frame_bo.0.x + clipped_half_w], |
| 1543 |
)); |
| 1544 |
} |
| 1545 |
|
| 1546 |
subset_c.push(process_cand( |
| 1547 |
prev_frame[frame_bo.0.y + clipped_half_h][frame_bo.0.x + clipped_half_w], |
| 1548 |
)); |
| 1549 |
} |
| 1550 |
|
| 1551 |
// Undo normalization to 128x128 block size |
| 1552 |
let min_sad = ((min_sad as u64 * (pix_w * pix_h) as u64) >> (MAX_SB_SIZE_LOG2 * 2)) as u32; |
| 1553 |
|
| 1554 |
let dec_mv = |mv: MotionVector| MotionVector { |
| 1555 |
col: mv.col >> ssdec, |
| 1556 |
row: mv.row >> ssdec, |
| 1557 |
}; |
| 1558 |
let median = median.map(dec_mv); |
| 1559 |
for mv in subset_b.iter_mut() { |
| 1560 |
*mv = dec_mv(*mv); |
| 1561 |
} |
| 1562 |
for mv in subset_c.iter_mut() { |
| 1563 |
*mv = dec_mv(*mv); |
| 1564 |
} |
| 1565 |
|
| 1566 |
MotionEstimationSubsets { |
| 1567 |
min_sad, |
| 1568 |
median, |
| 1569 |
subset_b, |
| 1570 |
subset_c, |
| 1571 |
} |
| 1572 |
} |
| 1573 |
|
| 1574 |
#[allow(clippy::too_many_arguments)] |
| 1575 |
fn get_best_predictor<T: Pixel>( |
| 1576 |
po: PlaneOffset, |
| 1577 |
org_region: &PlaneRegion<T>, |
| 1578 |
p_ref: &Plane<T>, |
| 1579 |
predictors: &[MotionVector], |
| 1580 |
bit_depth: usize, |
| 1581 |
pmv: [MotionVector; 2], |
| 1582 |
lambda: u32, |
| 1583 |
mvx_min: isize, |
| 1584 |
mvx_max: isize, |
| 1585 |
mvy_min: isize, |
| 1586 |
mvy_max: isize, |
| 1587 |
w: usize, |
| 1588 |
h: usize, |
| 1589 |
cpu_feature_level: CpuFeatureLevel, |
| 1590 |
) -> MotionSearchResult { |
| 1591 |
let mut best: MotionSearchResult = MotionSearchResult::empty(); |
| 1592 |
|
| 1593 |
for &init_mv in predictors.iter() { |
| 1594 |
let rd = get_fullpel_mv_rd( |
| 1595 |
po, |
| 1596 |
org_region, |
| 1597 |
p_ref, |
| 1598 |
bit_depth, |
| 1599 |
pmv, |
| 1600 |
lambda, |
| 1601 |
false, |
| 1602 |
mvx_min, |
| 1603 |
mvx_max, |
| 1604 |
mvy_min, |
| 1605 |
mvy_max, |
| 1606 |
w, |
| 1607 |
h, |
| 1608 |
init_mv, |
| 1609 |
cpu_feature_level, |
| 1610 |
); |
| 1611 |
|
| 1612 |
if rd.cost < best.rd.cost { |
| 1613 |
best.mv = init_mv; |
| 1614 |
best.rd = rd; |
| 1615 |
} |
| 1616 |
} |
| 1617 |
|
| 1618 |
best |
| 1619 |
} |
| 1620 |
|
| 1621 |
#[allow(clippy::too_many_arguments)] |
| 1622 |
fn get_fullpel_mv_rd<T: Pixel>( |
| 1623 |
po: PlaneOffset, |
| 1624 |
org_region: &PlaneRegion<T>, |
| 1625 |
p_ref: &Plane<T>, |
| 1626 |
bit_depth: usize, |
| 1627 |
pmv: [MotionVector; 2], |
| 1628 |
lambda: u32, |
| 1629 |
use_satd: bool, |
| 1630 |
mvx_min: isize, |
| 1631 |
mvx_max: isize, |
| 1632 |
mvy_min: isize, |
| 1633 |
mvy_max: isize, |
| 1634 |
w: usize, |
| 1635 |
h: usize, |
| 1636 |
cand_mv: MotionVector, |
| 1637 |
cpu_feature_level: CpuFeatureLevel, |
| 1638 |
) -> MVCandidateRD { |
| 1639 |
if (cand_mv.col as isize) < mvx_min |
| 1640 |
|| (cand_mv.col as isize) > mvx_max |
| 1641 |
|| (cand_mv.row as isize) < mvy_min |
| 1642 |
|| (cand_mv.row as isize) > mvy_max |
| 1643 |
{ |
| 1644 |
return MVCandidateRD::empty(); |
| 1645 |
} |
| 1646 |
|
| 1647 |
// Convert the motion vector into an full pixel offset. |
| 1648 |
let plane_ref = p_ref.region(Area::StartingAt { |
| 1649 |
x: po.x + (cand_mv.col / 8) as isize, |
| 1650 |
y: po.y + (cand_mv.row / 8) as isize, |
| 1651 |
}); |
| 1652 |
compute_mv_rd( |
| 1653 |
pmv, |
| 1654 |
lambda, |
| 1655 |
use_satd, |
| 1656 |
bit_depth, |
| 1657 |
w, |
| 1658 |
h, |
| 1659 |
cand_mv, |
| 1660 |
org_region, |
| 1661 |
&plane_ref, |
| 1662 |
cpu_feature_level, |
| 1663 |
) |
| 1664 |
} |
| 1665 |
|
| 1666 |
/// Perform hexagon search and refine afterwards. |
| 1667 |
/// |
| 1668 |
/// In the hexagon search stage, candidate motion vectors are examined for |
| 1669 |
/// improvement to the current search location. The search location is moved to |
| 1670 |
/// the best candidate (if any). This is repeated until the search location |
| 1671 |
/// stops moving. |
| 1672 |
/// |
| 1673 |
/// Refinement uses a square pattern that fits between the hexagon candidates. |
| 1674 |
/// |
| 1675 |
/// The hexagon pattern is defined by [`HEXAGON_PATTERN`] and the refinement |
| 1676 |
/// is defined by [`SQUARE_REFINE_PATTERN`]. |
| 1677 |
/// |
| 1678 |
/// `current` provides the initial search location and serves as |
| 1679 |
/// the output for the final search results. |
| 1680 |
#[allow(clippy::too_many_arguments)] |
| 1681 |
fn hexagon_search<T: Pixel>( |
| 1682 |
po: PlaneOffset, |
| 1683 |
org_region: &PlaneRegion<T>, |
| 1684 |
p_ref: &Plane<T>, |
| 1685 |
current: &mut MotionSearchResult, |
| 1686 |
bit_depth: usize, |
| 1687 |
pmv: [MotionVector; 2], |
| 1688 |
lambda: u32, |
| 1689 |
mvx_min: isize, |
| 1690 |
mvx_max: isize, |
| 1691 |
mvy_min: isize, |
| 1692 |
mvy_max: isize, |
| 1693 |
w: usize, |
| 1694 |
h: usize, |
| 1695 |
cpu_feature_level: CpuFeatureLevel, |
| 1696 |
) { |
| 1697 |
// The first iteration of hexagon search is implemented separate from |
| 1698 |
// subsequent iterations, which overlap with previous iterations. |
| 1699 |
|
| 1700 |
// Holds what candidate is used (if any). This is used to determine which |
| 1701 |
// candidates have already been tested in a previous iteration and can be |
| 1702 |
// skipped. |
| 1703 |
let mut best_cand_idx: usize = 0; |
| 1704 |
let mut best_cand: MotionSearchResult = MotionSearchResult::empty(); |
| 1705 |
|
| 1706 |
// First iteration of hexagon search. There are six candidates to consider. |
| 1707 |
for (i, &pattern_mv) in HEXAGON_PATTERN.iter().enumerate() { |
| 1708 |
let cand_mv = current.mv + pattern_mv; |
| 1709 |
let rd = get_fullpel_mv_rd( |
| 1710 |
po, |
| 1711 |
org_region, |
| 1712 |
p_ref, |
| 1713 |
bit_depth, |
| 1714 |
pmv, |
| 1715 |
lambda, |
| 1716 |
false, |
| 1717 |
mvx_min, |
| 1718 |
mvx_max, |
| 1719 |
mvy_min, |
| 1720 |
mvy_max, |
| 1721 |
w, |
| 1722 |
h, |
| 1723 |
cand_mv, |
| 1724 |
cpu_feature_level, |
| 1725 |
); |
| 1726 |
|
| 1727 |
if rd.cost < best_cand.rd.cost { |
| 1728 |
best_cand_idx = i; |
| 1729 |
best_cand.mv = cand_mv; |
| 1730 |
best_cand.rd = rd; |
| 1731 |
} |
| 1732 |
} |
| 1733 |
|
| 1734 |
// Run additional iterations of hexagon search until the search location |
| 1735 |
// doesn't update. |
| 1736 |
while best_cand.rd.cost < current.rd.cost { |
| 1737 |
// Update the search location. |
| 1738 |
*current = best_cand; |
| 1739 |
best_cand = MotionSearchResult::empty(); |
| 1740 |
|
| 1741 |
// Save the index/direction taken in the previous iteration to the current |
| 1742 |
// search location. |
| 1743 |
let center_cand_idx = best_cand_idx; |
| 1744 |
|
| 1745 |
// Look only at candidates that don't overlap with previous iterations. This |
| 1746 |
// corresponds with the three offsets (2D) with the closest direction to |
| 1747 |
// that traveled by the previous iteration. HEXAGON_PATTERN has clockwise |
| 1748 |
// order, so the last direction -1, +0, and +1 (mod 6) give the indices for |
| 1749 |
// these offsets. |
| 1750 |
for idx_offset_mod6 in 5..=7 { |
| 1751 |
let i = (center_cand_idx + idx_offset_mod6) % 6; |
| 1752 |
let cand_mv = current.mv + HEXAGON_PATTERN[i]; |
| 1753 |
|
| 1754 |
let rd = get_fullpel_mv_rd( |
| 1755 |
po, |
| 1756 |
org_region, |
| 1757 |
p_ref, |
| 1758 |
bit_depth, |
| 1759 |
pmv, |
| 1760 |
lambda, |
| 1761 |
false, |
| 1762 |
mvx_min, |
| 1763 |
mvx_max, |
| 1764 |
mvy_min, |
| 1765 |
mvy_max, |
| 1766 |
w, |
| 1767 |
h, |
| 1768 |
cand_mv, |
| 1769 |
cpu_feature_level, |
| 1770 |
); |
| 1771 |
|
| 1772 |
if rd.cost < best_cand.rd.cost { |
| 1773 |
best_cand_idx = i; |
| 1774 |
best_cand.mv = cand_mv; |
| 1775 |
best_cand.rd = rd; |
| 1776 |
} |
| 1777 |
} |
| 1778 |
} |
| 1779 |
|
| 1780 |
// Refine the motion after completing hexagon search. |
| 1781 |
let mut best_cand: MotionSearchResult = MotionSearchResult::empty(); |
| 1782 |
for &offset in &SQUARE_REFINE_PATTERN { |
| 1783 |
let cand_mv = current.mv + offset; |
| 1784 |
let rd = get_fullpel_mv_rd( |
| 1785 |
po, |
| 1786 |
org_region, |
| 1787 |
p_ref, |
| 1788 |
bit_depth, |
| 1789 |
pmv, |
| 1790 |
lambda, |
| 1791 |
false, |
| 1792 |
mvx_min, |
| 1793 |
mvx_max, |
| 1794 |
mvy_min, |
| 1795 |
mvy_max, |
| 1796 |
w, |
| 1797 |
h, |
| 1798 |
cand_mv, |
| 1799 |
cpu_feature_level, |
| 1800 |
); |
| 1801 |
|
| 1802 |
if rd.cost < best_cand.rd.cost { |
| 1803 |
best_cand.mv = cand_mv; |
| 1804 |
best_cand.rd = rd; |
| 1805 |
} |
| 1806 |
} |
| 1807 |
if best_cand.rd.cost < current.rd.cost { |
| 1808 |
*current = best_cand; |
| 1809 |
} |
| 1810 |
|
| 1811 |
assert!(!current.is_empty()); |
| 1812 |
} |
| 1813 |
|
| 1814 |
/// Run a full pixel diamond search. The search is run on multiple step sizes. |
| 1815 |
/// |
| 1816 |
/// For each step size, candidate motion vectors are examined for improvement |
| 1817 |
/// to the current search location. The search location is moved to the best |
| 1818 |
/// candidate (if any). This is repeated until the search location stops moving. |
| 1819 |
#[allow(clippy::too_many_arguments)] |
| 1820 |
fn fullpel_diamond_search<T: Pixel>( |
| 1821 |
po: PlaneOffset, |
| 1822 |
org_region: &PlaneRegion<T>, |
| 1823 |
p_ref: &Plane<T>, |
| 1824 |
current: &mut MotionSearchResult, |
| 1825 |
bit_depth: usize, |
| 1826 |
pmv: [MotionVector; 2], |
| 1827 |
lambda: u32, |
| 1828 |
mvx_min: isize, |
| 1829 |
mvx_max: isize, |
| 1830 |
mvy_min: isize, |
| 1831 |
mvy_max: isize, |
| 1832 |
w: usize, |
| 1833 |
h: usize, |
| 1834 |
cpu_feature_level: CpuFeatureLevel, |
| 1835 |
) { |
| 1836 |
// Define the initial and the final scale (log2) of the diamond. |
| 1837 |
let (mut diamond_radius_log2, diamond_radius_end_log2) = (1u8, 0u8); |
| 1838 |
|
| 1839 |
loop { |
| 1840 |
// Find the best candidate from the diamond pattern. |
| 1841 |
let mut best_cand: MotionSearchResult = MotionSearchResult::empty(); |
| 1842 |
for &offset in &DIAMOND_R1_PATTERN { |
| 1843 |
let cand_mv = current.mv + (offset << diamond_radius_log2); |
| 1844 |
let rd = get_fullpel_mv_rd( |
| 1845 |
po, |
| 1846 |
org_region, |
| 1847 |
p_ref, |
| 1848 |
bit_depth, |
| 1849 |
pmv, |
| 1850 |
lambda, |
| 1851 |
false, |
| 1852 |
mvx_min, |
| 1853 |
mvx_max, |
| 1854 |
mvy_min, |
| 1855 |
mvy_max, |
| 1856 |
w, |
| 1857 |
h, |
| 1858 |
cand_mv, |
| 1859 |
cpu_feature_level, |
| 1860 |
); |
| 1861 |
|
| 1862 |
if rd.cost < best_cand.rd.cost { |
| 1863 |
best_cand.mv = cand_mv; |
| 1864 |
best_cand.rd = rd; |
| 1865 |
} |
| 1866 |
} |
| 1867 |
|
| 1868 |
// Continue the search at this scale until the can't find a better candidate |
| 1869 |
// to use. |
| 1870 |
if current.rd.cost <= best_cand.rd.cost { |
| 1871 |
if diamond_radius_log2 == diamond_radius_end_log2 { |
| 1872 |
break; |
| 1873 |
} else { |
| 1874 |
diamond_radius_log2 -= 1; |
| 1875 |
} |
| 1876 |
} else { |
| 1877 |
*current = best_cand; |
| 1878 |
} |
| 1879 |
} |
| 1880 |
|
| 1881 |
assert!(!current.is_empty()); |
| 1882 |
} |