| 1 |
// Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved |
| 2 |
// Copyright (c) 2017-2022, The rav1e contributors. All rights reserved |
| 3 |
// |
| 4 |
// This source code is subject to the terms of the BSD 2 Clause License and |
| 5 |
// the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License |
| 6 |
// was not distributed with this source code in the LICENSE file, you can |
| 7 |
// obtain it at www.aomedia.org/license/software. If the Alliance for Open |
| 8 |
// Media Patent License 1.0 was not distributed with this source code in the |
| 9 |
// PATENTS file, you can obtain it at www.aomedia.org/license/patent. |
| 10 |
|
| 11 |
#![allow(non_camel_case_types)] |
| 12 |
|
| 13 |
use std::fmt; |
| 14 |
use std::mem::MaybeUninit; |
| 15 |
|
| 16 |
use arrayvec::*; |
| 17 |
use itertools::izip; |
| 18 |
|
| 19 |
use crate::api::*; |
| 20 |
use crate::cdef::*; |
| 21 |
use crate::context::*; |
| 22 |
use crate::cpu_features::CpuFeatureLevel; |
| 23 |
use crate::deblock::*; |
| 24 |
use crate::dist::*; |
| 25 |
use crate::ec::{Writer, WriterCounter, OD_BITRES}; |
| 26 |
use crate::encode_block_with_modes; |
| 27 |
use crate::encoder::{FrameInvariants, IMPORTANCE_BLOCK_SIZE}; |
| 28 |
use crate::frame::*; |
| 29 |
use crate::header::ReferenceMode; |
| 30 |
use crate::lrf::*; |
| 31 |
use crate::mc::MotionVector; |
| 32 |
use crate::me::estimate_motion; |
| 33 |
use crate::me::MVSamplingMode; |
| 34 |
use crate::me::MotionSearchResult; |
| 35 |
use crate::motion_compensate; |
| 36 |
use crate::partition::PartitionType::*; |
| 37 |
use crate::partition::RefType::*; |
| 38 |
use crate::partition::*; |
| 39 |
use crate::predict::{ |
| 40 |
luma_ac, AngleDelta, IntraEdgeFilterParameters, IntraParam, PredictionMode, |
| 41 |
RAV1E_INTER_COMPOUND_MODES, RAV1E_INTER_MODES_MINIMAL, RAV1E_INTRA_MODES, |
| 42 |
}; |
| 43 |
use crate::rdo_tables::*; |
| 44 |
use crate::tiling::*; |
| 45 |
use crate::transform::{TxSet, TxSize, TxType, RAV1E_TX_TYPES}; |
| 46 |
use crate::util::{init_slice_repeat_mut, Aligned, Pixel}; |
| 47 |
use crate::write_tx_blocks; |
| 48 |
use crate::write_tx_tree; |
| 49 |
use crate::Tune; |
| 50 |
use crate::{encode_block_post_cdef, encode_block_pre_cdef}; |
| 51 |
|
| 52 |
#[derive(Copy, Clone, PartialEq, Eq)] |
| 53 |
pub enum RDOType { |
| 54 |
PixelDistRealRate, |
| 55 |
TxDistRealRate, |
| 56 |
TxDistEstRate, |
| 57 |
} |
| 58 |
|
| 59 |
impl RDOType { |
| 60 |
#[inline] |
| 61 |
pub const fn needs_tx_dist(self) -> bool { |
| 62 |
match self { |
| 63 |
// Pixel-domain distortion and exact ec rate |
| 64 |
RDOType::PixelDistRealRate => false, |
| 65 |
// Tx-domain distortion and exact ec rate |
| 66 |
RDOType::TxDistRealRate => true, |
| 67 |
// Tx-domain distortion and txdist-based rate |
| 68 |
RDOType::TxDistEstRate => true, |
| 69 |
} |
| 70 |
} |
| 71 |
#[inline] |
| 72 |
pub const fn needs_coeff_rate(self) -> bool { |
| 73 |
match self { |
| 74 |
RDOType::PixelDistRealRate => true, |
| 75 |
RDOType::TxDistRealRate => true, |
| 76 |
RDOType::TxDistEstRate => false, |
| 77 |
} |
| 78 |
} |
| 79 |
} |
| 80 |
|
| 81 |
#[derive(Clone)] |
| 82 |
pub struct PartitionGroupParameters { |
| 83 |
pub rd_cost: f64, |
| 84 |
pub part_type: PartitionType, |
| 85 |
pub part_modes: ArrayVec<PartitionParameters, 4>, |
| 86 |
} |
| 87 |
|
| 88 |
#[derive(Clone, Debug)] |
| 89 |
pub struct PartitionParameters { |
| 90 |
pub rd_cost: f64, |
| 91 |
pub bo: TileBlockOffset, |
| 92 |
pub bsize: BlockSize, |
| 93 |
pub pred_mode_luma: PredictionMode, |
| 94 |
pub pred_mode_chroma: PredictionMode, |
| 95 |
pub pred_cfl_params: CFLParams, |
| 96 |
pub angle_delta: AngleDelta, |
| 97 |
pub ref_frames: [RefType; 2], |
| 98 |
pub mvs: [MotionVector; 2], |
| 99 |
pub skip: bool, |
| 100 |
pub has_coeff: bool, |
| 101 |
pub tx_size: TxSize, |
| 102 |
pub tx_type: TxType, |
| 103 |
pub sidx: u8, |
| 104 |
} |
| 105 |
|
| 106 |
impl Default for PartitionParameters { |
| 107 |
fn default() -> Self { |
| 108 |
PartitionParameters { |
| 109 |
rd_cost: f64::MAX, |
| 110 |
bo: TileBlockOffset::default(), |
| 111 |
bsize: BlockSize::BLOCK_32X32, |
| 112 |
pred_mode_luma: PredictionMode::default(), |
| 113 |
pred_mode_chroma: PredictionMode::default(), |
| 114 |
pred_cfl_params: CFLParams::default(), |
| 115 |
angle_delta: AngleDelta::default(), |
| 116 |
ref_frames: [RefType::INTRA_FRAME, RefType::NONE_FRAME], |
| 117 |
mvs: [MotionVector::default(); 2], |
| 118 |
skip: false, |
| 119 |
has_coeff: true, |
| 120 |
tx_size: TxSize::TX_4X4, |
| 121 |
tx_type: TxType::DCT_DCT, |
| 122 |
sidx: 0, |
| 123 |
} |
| 124 |
} |
| 125 |
} |
| 126 |
|
| 127 |
pub fn estimate_rate(qindex: u8, ts: TxSize, fast_distortion: u64) -> u64 { |
| 128 |
let bs_index = ts as usize; |
| 129 |
let q_bin_idx = (qindex as usize) / RDO_QUANT_DIV; |
| 130 |
let bin_idx_down = |
| 131 |
((fast_distortion) / RATE_EST_BIN_SIZE).min((RDO_NUM_BINS - 2) as u64); |
| 132 |
let bin_idx_up = (bin_idx_down + 1).min((RDO_NUM_BINS - 1) as u64); |
| 133 |
let x0 = (bin_idx_down * RATE_EST_BIN_SIZE) as i64; |
| 134 |
let x1 = (bin_idx_up * RATE_EST_BIN_SIZE) as i64; |
| 135 |
let y0 = RDO_RATE_TABLE[q_bin_idx][bs_index][bin_idx_down as usize] as i64; |
| 136 |
let y1 = RDO_RATE_TABLE[q_bin_idx][bs_index][bin_idx_up as usize] as i64; |
| 137 |
let slope = ((y1 - y0) << 8) / (x1 - x0); |
| 138 |
(y0 + (((fast_distortion as i64 - x0) * slope) >> 8)).max(0) as u64 |
| 139 |
} |
| 140 |
|
| 141 |
#[allow(unused)] |
| 142 |
pub fn cdef_dist_wxh<T: Pixel, F: Fn(Area, BlockSize) -> DistortionScale>( |
| 143 |
src1: &PlaneRegion<'_, T>, src2: &PlaneRegion<'_, T>, w: usize, h: usize, |
| 144 |
bit_depth: usize, compute_bias: F, cpu: CpuFeatureLevel, |
| 145 |
) -> Distortion { |
| 146 |
debug_assert!(src1.plane_cfg.xdec == 0); |
| 147 |
debug_assert!(src1.plane_cfg.ydec == 0); |
| 148 |
debug_assert!(src2.plane_cfg.xdec == 0); |
| 149 |
debug_assert!(src2.plane_cfg.ydec == 0); |
| 150 |
|
| 151 |
let mut sum = Distortion::zero(); |
| 152 |
for y in (0..h).step_by(8) { |
| 153 |
for x in (0..w).step_by(8) { |
| 154 |
let kernel_h = (h - y).min(8); |
| 155 |
let kernel_w = (w - x).min(8); |
| 156 |
let area = Area::StartingAt { x: x as isize, y: y as isize }; |
| 157 |
|
| 158 |
let value = RawDistortion(cdef_dist_kernel( |
| 159 |
&src1.subregion(area), |
| 160 |
&src2.subregion(area), |
| 161 |
kernel_w, |
| 162 |
kernel_h, |
| 163 |
bit_depth, |
| 164 |
cpu, |
| 165 |
) as u64); |
| 166 |
|
| 167 |
// cdef is always called on non-subsampled planes, so BLOCK_8X8 is |
| 168 |
// correct here. |
| 169 |
sum += value * compute_bias(area, BlockSize::BLOCK_8X8); |
| 170 |
} |
| 171 |
} |
| 172 |
sum |
| 173 |
} |
| 174 |
|
| 175 |
/// Sum of Squared Error for a wxh block |
| 176 |
/// Currently limited to w and h of valid blocks |
| 177 |
pub fn sse_wxh<T: Pixel, F: Fn(Area, BlockSize) -> DistortionScale>( |
| 178 |
src1: &PlaneRegion<'_, T>, src2: &PlaneRegion<'_, T>, w: usize, h: usize, |
| 179 |
compute_bias: F, bit_depth: usize, cpu: CpuFeatureLevel, |
| 180 |
) -> Distortion { |
| 181 |
// See get_weighted_sse in src/dist.rs. |
| 182 |
// Provide a scale to get_weighted_sse for each square region of this size. |
| 183 |
const CHUNK_SIZE: usize = IMPORTANCE_BLOCK_SIZE >> 1; |
| 184 |
|
| 185 |
// To bias the distortion correctly, compute it in blocks up to the size |
| 186 |
// importance block size in a non-subsampled plane. |
| 187 |
let imp_block_w = CHUNK_SIZE << src1.plane_cfg.xdec; |
| 188 |
let imp_block_h = CHUNK_SIZE << src1.plane_cfg.ydec; |
| 189 |
|
| 190 |
let imp_bsize = BlockSize::from_width_and_height(imp_block_w, imp_block_h); |
| 191 |
|
| 192 |
let n_imp_blocks_w = w.div_ceil(CHUNK_SIZE); |
| 193 |
let n_imp_blocks_h = h.div_ceil(CHUNK_SIZE); |
| 194 |
|
| 195 |
// TODO: Copying biases into a buffer is slow. It would be best if biases were |
| 196 |
// passed directly. To do this, we would need different versions of the |
| 197 |
// weighted sse function for decimated/subsampled data. Also requires |
| 198 |
// eliminating use of unbiased sse. |
| 199 |
// It should also be noted that the current copy code does not auto-vectorize. |
| 200 |
|
| 201 |
// Copy biases into a buffer. |
| 202 |
let mut buf_storage = Aligned::new( |
| 203 |
[MaybeUninit::<u32>::uninit(); 128 / CHUNK_SIZE * 128 / CHUNK_SIZE], |
| 204 |
); |
| 205 |
let buf_stride = n_imp_blocks_w.next_power_of_two(); |
| 206 |
let buf = init_slice_repeat_mut( |
| 207 |
&mut buf_storage.data[..buf_stride * n_imp_blocks_h], |
| 208 |
0, |
| 209 |
); |
| 210 |
|
| 211 |
for block_y in 0..n_imp_blocks_h { |
| 212 |
for block_x in 0..n_imp_blocks_w { |
| 213 |
let block = Area::StartingAt { |
| 214 |
x: (block_x * CHUNK_SIZE) as isize, |
| 215 |
y: (block_y * CHUNK_SIZE) as isize, |
| 216 |
}; |
| 217 |
buf[block_y * buf_stride + block_x] = compute_bias(block, imp_bsize).0; |
| 218 |
} |
| 219 |
} |
| 220 |
|
| 221 |
Distortion(get_weighted_sse( |
| 222 |
src1, src2, buf, buf_stride, w, h, bit_depth, cpu, |
| 223 |
)) |
| 224 |
} |
| 225 |
|
| 226 |
// TODO consider saturating_sub later |
| 227 |
#[allow(clippy::implicit_saturating_sub)] |
| 228 |
pub const fn clip_visible_bsize( |
| 229 |
frame_w: usize, frame_h: usize, bsize: BlockSize, x: usize, y: usize, |
| 230 |
) -> (usize, usize) { |
| 231 |
let blk_w = bsize.width(); |
| 232 |
let blk_h = bsize.height(); |
| 233 |
|
| 234 |
let visible_w: usize = if x + blk_w <= frame_w { |
| 235 |
blk_w |
| 236 |
} else if x >= frame_w { |
| 237 |
0 |
| 238 |
} else { |
| 239 |
frame_w - x |
| 240 |
}; |
| 241 |
|
| 242 |
let visible_h: usize = if y + blk_h <= frame_h { |
| 243 |
blk_h |
| 244 |
} else if y >= frame_h { |
| 245 |
0 |
| 246 |
} else { |
| 247 |
frame_h - y |
| 248 |
}; |
| 249 |
|
| 250 |
(visible_w, visible_h) |
| 251 |
} |
| 252 |
|
| 253 |
// Compute the pixel-domain distortion for an encode |
| 254 |
fn compute_distortion<T: Pixel>( |
| 255 |
fi: &FrameInvariants<T>, ts: &TileStateMut<'_, T>, bsize: BlockSize, |
| 256 |
is_chroma_block: bool, tile_bo: TileBlockOffset, luma_only: bool, |
| 257 |
) -> ScaledDistortion { |
| 258 |
let area = Area::BlockStartingAt { bo: tile_bo.0 }; |
| 259 |
let input_region = ts.input_tile.planes[0].subregion(area); |
| 260 |
let rec_region = ts.rec.planes[0].subregion(area); |
| 261 |
|
| 262 |
// clip a block to have visible pixles only |
| 263 |
let frame_bo = ts.to_frame_block_offset(tile_bo); |
| 264 |
let (visible_w, visible_h) = clip_visible_bsize( |
| 265 |
fi.width, |
| 266 |
fi.height, |
| 267 |
bsize, |
| 268 |
frame_bo.0.x << MI_SIZE_LOG2, |
| 269 |
frame_bo.0.y << MI_SIZE_LOG2, |
| 270 |
); |
| 271 |
|
| 272 |
if visible_w == 0 || visible_h == 0 { |
| 273 |
return ScaledDistortion::zero(); |
| 274 |
} |
| 275 |
|
| 276 |
let mut distortion = match fi.config.tune { |
| 277 |
Tune::Psychovisual => cdef_dist_wxh( |
| 278 |
&input_region, |
| 279 |
&rec_region, |
| 280 |
visible_w, |
| 281 |
visible_h, |
| 282 |
fi.sequence.bit_depth, |
| 283 |
|bias_area, bsize| { |
| 284 |
distortion_scale( |
| 285 |
fi, |
| 286 |
input_region.subregion(bias_area).frame_block_offset(), |
| 287 |
bsize, |
| 288 |
) |
| 289 |
}, |
| 290 |
fi.cpu_feature_level, |
| 291 |
), |
| 292 |
Tune::Psnr => sse_wxh( |
| 293 |
&input_region, |
| 294 |
&rec_region, |
| 295 |
visible_w, |
| 296 |
visible_h, |
| 297 |
|bias_area, bsize| { |
| 298 |
distortion_scale( |
| 299 |
fi, |
| 300 |
input_region.subregion(bias_area).frame_block_offset(), |
| 301 |
bsize, |
| 302 |
) |
| 303 |
}, |
| 304 |
fi.sequence.bit_depth, |
| 305 |
fi.cpu_feature_level, |
| 306 |
), |
| 307 |
} * fi.dist_scale[0]; |
| 308 |
|
| 309 |
if is_chroma_block |
| 310 |
&& !luma_only |
| 311 |
&& fi.sequence.chroma_sampling != ChromaSampling::Cs400 |
| 312 |
{ |
| 313 |
let PlaneConfig { xdec, ydec, .. } = ts.input.planes[1].cfg; |
| 314 |
let chroma_w = if bsize.width() >= 8 || xdec == 0 { |
| 315 |
(visible_w + xdec) >> xdec |
| 316 |
} else { |
| 317 |
(4 + visible_w + xdec) >> xdec |
| 318 |
}; |
| 319 |
let chroma_h = if bsize.height() >= 8 || ydec == 0 { |
| 320 |
(visible_h + ydec) >> ydec |
| 321 |
} else { |
| 322 |
(4 + visible_h + ydec) >> ydec |
| 323 |
}; |
| 324 |
|
| 325 |
for p in 1..3 { |
| 326 |
let input_region = ts.input_tile.planes[p].subregion(area); |
| 327 |
let rec_region = ts.rec.planes[p].subregion(area); |
| 328 |
distortion += sse_wxh( |
| 329 |
&input_region, |
| 330 |
&rec_region, |
| 331 |
chroma_w, |
| 332 |
chroma_h, |
| 333 |
|bias_area, bsize| { |
| 334 |
distortion_scale( |
| 335 |
fi, |
| 336 |
input_region.subregion(bias_area).frame_block_offset(), |
| 337 |
bsize, |
| 338 |
) |
| 339 |
}, |
| 340 |
fi.sequence.bit_depth, |
| 341 |
fi.cpu_feature_level, |
| 342 |
) * fi.dist_scale[p]; |
| 343 |
} |
| 344 |
} |
| 345 |
distortion |
| 346 |
} |
| 347 |
|
| 348 |
// Compute the transform-domain distortion for an encode |
| 349 |
fn compute_tx_distortion<T: Pixel>( |
| 350 |
fi: &FrameInvariants<T>, ts: &TileStateMut<'_, T>, bsize: BlockSize, |
| 351 |
is_chroma_block: bool, tile_bo: TileBlockOffset, tx_dist: ScaledDistortion, |
| 352 |
skip: bool, luma_only: bool, |
| 353 |
) -> ScaledDistortion { |
| 354 |
assert!(fi.config.tune == Tune::Psnr); |
| 355 |
let area = Area::BlockStartingAt { bo: tile_bo.0 }; |
| 356 |
let input_region = ts.input_tile.planes[0].subregion(area); |
| 357 |
let rec_region = ts.rec.planes[0].subregion(area); |
| 358 |
|
| 359 |
let (visible_w, visible_h) = if !skip { |
| 360 |
(bsize.width(), bsize.height()) |
| 361 |
} else { |
| 362 |
let frame_bo = ts.to_frame_block_offset(tile_bo); |
| 363 |
clip_visible_bsize( |
| 364 |
fi.width, |
| 365 |
fi.height, |
| 366 |
bsize, |
| 367 |
frame_bo.0.x << MI_SIZE_LOG2, |
| 368 |
frame_bo.0.y << MI_SIZE_LOG2, |
| 369 |
) |
| 370 |
}; |
| 371 |
|
| 372 |
if visible_w == 0 || visible_h == 0 { |
| 373 |
return ScaledDistortion::zero(); |
| 374 |
} |
| 375 |
|
| 376 |
let mut distortion = if skip { |
| 377 |
sse_wxh( |
| 378 |
&input_region, |
| 379 |
&rec_region, |
| 380 |
visible_w, |
| 381 |
visible_h, |
| 382 |
|bias_area, bsize| { |
| 383 |
distortion_scale( |
| 384 |
fi, |
| 385 |
input_region.subregion(bias_area).frame_block_offset(), |
| 386 |
bsize, |
| 387 |
) |
| 388 |
}, |
| 389 |
fi.sequence.bit_depth, |
| 390 |
fi.cpu_feature_level, |
| 391 |
) * fi.dist_scale[0] |
| 392 |
} else { |
| 393 |
tx_dist |
| 394 |
}; |
| 395 |
|
| 396 |
if is_chroma_block |
| 397 |
&& !luma_only |
| 398 |
&& skip |
| 399 |
&& fi.sequence.chroma_sampling != ChromaSampling::Cs400 |
| 400 |
{ |
| 401 |
let PlaneConfig { xdec, ydec, .. } = ts.input.planes[1].cfg; |
| 402 |
let chroma_w = if bsize.width() >= 8 || xdec == 0 { |
| 403 |
(visible_w + xdec) >> xdec |
| 404 |
} else { |
| 405 |
(4 + visible_w + xdec) >> xdec |
| 406 |
}; |
| 407 |
let chroma_h = if bsize.height() >= 8 || ydec == 0 { |
| 408 |
(visible_h + ydec) >> ydec |
| 409 |
} else { |
| 410 |
(4 + visible_h + ydec) >> ydec |
| 411 |
}; |
| 412 |
|
| 413 |
for p in 1..3 { |
| 414 |
let input_region = ts.input_tile.planes[p].subregion(area); |
| 415 |
let rec_region = ts.rec.planes[p].subregion(area); |
| 416 |
distortion += sse_wxh( |
| 417 |
&input_region, |
| 418 |
&rec_region, |
| 419 |
chroma_w, |
| 420 |
chroma_h, |
| 421 |
|bias_area, bsize| { |
| 422 |
distortion_scale( |
| 423 |
fi, |
| 424 |
input_region.subregion(bias_area).frame_block_offset(), |
| 425 |
bsize, |
| 426 |
) |
| 427 |
}, |
| 428 |
fi.sequence.bit_depth, |
| 429 |
fi.cpu_feature_level, |
| 430 |
) * fi.dist_scale[p]; |
| 431 |
} |
| 432 |
} |
| 433 |
distortion |
| 434 |
} |
| 435 |
|
| 436 |
/// Compute a scaling factor to multiply the distortion of a block by, |
| 437 |
/// this factor is determined using temporal RDO. |
| 438 |
/// |
| 439 |
/// # Panics |
| 440 |
/// |
| 441 |
/// - If called with `bsize` of 8x8 or smaller |
| 442 |
/// - If the coded frame data doesn't exist on the `FrameInvariants` |
| 443 |
pub fn distortion_scale<T: Pixel>( |
| 444 |
fi: &FrameInvariants<T>, frame_bo: PlaneBlockOffset, bsize: BlockSize, |
| 445 |
) -> DistortionScale { |
| 446 |
if !fi.config.temporal_rdo() { |
| 447 |
return DistortionScale::default(); |
| 448 |
} |
| 449 |
// EncoderConfig::temporal_rdo() should always return false in situations |
| 450 |
// where distortion is computed on > 8x8 blocks, so we should never hit this |
| 451 |
// assert. |
| 452 |
assert!(bsize <= BlockSize::BLOCK_8X8); |
| 453 |
|
| 454 |
let x = frame_bo.0.x >> IMPORTANCE_BLOCK_TO_BLOCK_SHIFT; |
| 455 |
let y = frame_bo.0.y >> IMPORTANCE_BLOCK_TO_BLOCK_SHIFT; |
| 456 |
|
| 457 |
let coded_data = fi.coded_frame_data.as_ref().unwrap(); |
| 458 |
coded_data.distortion_scales[y * coded_data.w_in_imp_b + x] |
| 459 |
} |
| 460 |
|
| 461 |
/// # Panics |
| 462 |
/// |
| 463 |
/// - If the coded frame data doesn't exist on the `FrameInvariants` |
| 464 |
pub fn spatiotemporal_scale<T: Pixel>( |
| 465 |
fi: &FrameInvariants<T>, frame_bo: PlaneBlockOffset, bsize: BlockSize, |
| 466 |
) -> DistortionScale { |
| 467 |
if !fi.config.temporal_rdo() && fi.config.tune != Tune::Psychovisual { |
| 468 |
return DistortionScale::default(); |
| 469 |
} |
| 470 |
|
| 471 |
let coded_data = fi.coded_frame_data.as_ref().unwrap(); |
| 472 |
|
| 473 |
let x0 = frame_bo.0.x >> IMPORTANCE_BLOCK_TO_BLOCK_SHIFT; |
| 474 |
let y0 = frame_bo.0.y >> IMPORTANCE_BLOCK_TO_BLOCK_SHIFT; |
| 475 |
let x1 = (x0 + bsize.width_imp_b()).min(coded_data.w_in_imp_b); |
| 476 |
let y1 = (y0 + bsize.height_imp_b()).min(coded_data.h_in_imp_b); |
| 477 |
let den = (((x1 - x0) * (y1 - y0)) as u64) << DistortionScale::SHIFT; |
| 478 |
|
| 479 |
// calling this on each slice individually improves autovectorization |
| 480 |
// compared to using `Iterator::take` |
| 481 |
#[inline(always)] |
| 482 |
fn take_slice<T>(slice: &[T], n: usize) -> &[T] { |
| 483 |
slice.get(..n).unwrap_or(slice) |
| 484 |
} |
| 485 |
|
| 486 |
let mut sum = 0; |
| 487 |
for y in y0..y1 { |
| 488 |
sum += take_slice( |
| 489 |
&coded_data.distortion_scales[y * coded_data.w_in_imp_b..][x0..x1], |
| 490 |
MAX_SB_IN_IMP_B, |
| 491 |
) |
| 492 |
.iter() |
| 493 |
.zip( |
| 494 |
take_slice( |
| 495 |
&coded_data.activity_scales[y * coded_data.w_in_imp_b..][x0..x1], |
| 496 |
MAX_SB_IN_IMP_B, |
| 497 |
) |
| 498 |
.iter(), |
| 499 |
) |
| 500 |
.map(|(d, a)| d.0 as u64 * a.0 as u64) |
| 501 |
.sum::<u64>(); |
| 502 |
} |
| 503 |
DistortionScale(((sum + (den >> 1)) / den) as u32) |
| 504 |
} |
| 505 |
|
| 506 |
pub fn distortion_scale_for( |
| 507 |
propagate_cost: f64, intra_cost: f64, |
| 508 |
) -> DistortionScale { |
| 509 |
// The mbtree paper \cite{mbtree} uses the following formula: |
| 510 |
// |
| 511 |
// QP_delta = -strength * log2(1 + (propagate_cost / intra_cost)) |
| 512 |
// |
| 513 |
// Since this is H.264, this corresponds to the following quantizer: |
| 514 |
// |
| 515 |
// Q' = Q * 2^(QP_delta/6) |
| 516 |
// |
| 517 |
// Since lambda is proportial to Q^2, this means we want to minimize: |
| 518 |
// |
| 519 |
// D + lambda' * R |
| 520 |
// = D + 2^(QP_delta / 3) * lambda * R |
| 521 |
// |
| 522 |
// If we want to keep lambda fixed, we can instead scale distortion and |
| 523 |
// minimize: |
| 524 |
// |
| 525 |
// D * scale + lambda * R |
| 526 |
// |
| 527 |
// where: |
| 528 |
// |
| 529 |
// scale = 2^(QP_delta / -3) |
| 530 |
// = (1 + (propagate_cost / intra_cost))^(strength / 3) |
| 531 |
// |
| 532 |
// The original paper empirically chooses strength = 2.0, but strength = 1.0 |
| 533 |
// seems to work best in rav1e currently, this may have something to do with |
| 534 |
// the fact that they use 16x16 blocks whereas our "importance blocks" are |
| 535 |
// 8x8, but everything should be scale invariant here so that's weird. |
| 536 |
// |
| 537 |
// @article{mbtree, |
| 538 |
// title={A novel macroblock-tree algorithm for high-performance |
| 539 |
// optimization of dependent video coding in H.264/AVC}, |
| 540 |
// author={Garrett-Glaser, Jason}, |
| 541 |
// journal={Tech. Rep.}, |
| 542 |
// year={2009}, |
| 543 |
// url={https://pdfs.semanticscholar.org/032f/1ab7d9db385780a02eb2d579af8303b266d2.pdf} |
| 544 |
// } |
| 545 |
|
| 546 |
if intra_cost == 0. { |
| 547 |
return DistortionScale::default(); // no scaling |
| 548 |
} |
| 549 |
|
| 550 |
let strength = 1.0; // empirical, see comment above |
| 551 |
let frac = (intra_cost + propagate_cost) / intra_cost; |
| 552 |
frac.powf(strength / 3.0).into() |
| 553 |
} |
| 554 |
|
| 555 |
/// Fixed point arithmetic version of distortion scale |
| 556 |
#[repr(transparent)] |
| 557 |
#[derive(Copy, Clone)] |
| 558 |
pub struct DistortionScale(pub u32); |
| 559 |
|
| 560 |
#[repr(transparent)] |
| 561 |
pub struct RawDistortion(u64); |
| 562 |
|
| 563 |
#[repr(transparent)] |
| 564 |
pub struct Distortion(pub u64); |
| 565 |
|
| 566 |
#[repr(transparent)] |
| 567 |
pub struct ScaledDistortion(u64); |
| 568 |
|
| 569 |
impl DistortionScale { |
| 570 |
/// Bits past the radix point |
| 571 |
const SHIFT: u32 = 14; |
| 572 |
/// Number of bits used. Determines the max value. |
| 573 |
/// 28 bits is quite excessive. |
| 574 |
const BITS: u32 = 28; |
| 575 |
/// Maximum internal value |
| 576 |
const MAX: u64 = (1 << Self::BITS) - 1; |
| 577 |
|
| 578 |
#[inline] |
| 579 |
pub const fn new(num: u64, den: u64) -> Self { |
| 580 |
let raw = (num << Self::SHIFT).saturating_add(den / 2) / den; |
| 581 |
let mask = (raw <= Self::MAX) as u64; |
| 582 |
Self((mask * raw + (1 - mask) * Self::MAX) as u32) |
| 583 |
} |
| 584 |
|
| 585 |
pub fn inv_mean(slice: &[Self]) -> Self { |
| 586 |
use crate::util::{bexp64, blog32_q11}; |
| 587 |
let sum = slice.iter().map(|&s| blog32_q11(s.0) as i64).sum::<i64>(); |
| 588 |
let log_inv_mean_q11 = |
| 589 |
(Self::SHIFT << 11) as i64 - sum / slice.len() as i64; |
| 590 |
Self( |
| 591 |
bexp64((log_inv_mean_q11 + (Self::SHIFT << 11) as i64) << (57 - 11)) |
| 592 |
.clamp(1, (1 << Self::BITS) - 1) as u32, |
| 593 |
) |
| 594 |
} |
| 595 |
|
| 596 |
/// Binary logarithm in Q11 |
| 597 |
#[inline] |
| 598 |
pub const fn blog16(self) -> i16 { |
| 599 |
use crate::util::blog32_q11; |
| 600 |
(blog32_q11(self.0) - ((Self::SHIFT as i32) << 11)) as i16 |
| 601 |
} |
| 602 |
|
| 603 |
/// Binary logarithm in Q57 |
| 604 |
#[inline] |
| 605 |
pub const fn blog64(self) -> i64 { |
| 606 |
use crate::util::{blog64, q57}; |
| 607 |
blog64(self.0 as i64) - q57(Self::SHIFT as i32) |
| 608 |
} |
| 609 |
|
| 610 |
/// Multiply, round and shift |
| 611 |
/// Internal implementation, so don't use multiply trait. |
| 612 |
#[inline] |
| 613 |
pub const fn mul_u64(self, dist: u64) -> u64 { |
| 614 |
(self.0 as u64 * dist + (1 << Self::SHIFT >> 1)) >> Self::SHIFT |
| 615 |
} |
| 616 |
} |
| 617 |
|
| 618 |
impl std::ops::Mul for DistortionScale { |
| 619 |
type Output = Self; |
| 620 |
|
| 621 |
/// Multiply, round and shift |
| 622 |
#[inline] |
| 623 |
fn mul(self, rhs: Self) -> Self { |
| 624 |
Self( |
| 625 |
(((self.0 as u64 * rhs.0 as u64) + (1 << (Self::SHIFT - 1))) |
| 626 |
>> Self::SHIFT) |
| 627 |
.clamp(1, (1 << Self::BITS) - 1) as u32, |
| 628 |
) |
| 629 |
} |
| 630 |
} |
| 631 |
|
| 632 |
impl std::ops::MulAssign for DistortionScale { |
| 633 |
fn mul_assign(&mut self, rhs: Self) { |
| 634 |
*self = *self * rhs; |
| 635 |
} |
| 636 |
} |
| 637 |
|
| 638 |
// Default value for DistortionScale is a fixed point 1 |
| 639 |
impl Default for DistortionScale { |
| 640 |
#[inline] |
| 641 |
fn default() -> Self { |
| 642 |
Self(1 << Self::SHIFT) |
| 643 |
} |
| 644 |
} |
| 645 |
|
| 646 |
impl fmt::Debug for DistortionScale { |
| 647 |
fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result { |
| 648 |
write!(f, "{}", f64::from(*self)) |
| 649 |
} |
| 650 |
} |
| 651 |
|
| 652 |
impl From<f64> for DistortionScale { |
| 653 |
#[inline] |
| 654 |
fn from(scale: f64) -> Self { |
| 655 |
let den = 1 << (Self::SHIFT + 1); |
| 656 |
Self::new((scale * den as f64) as u64, den) |
| 657 |
} |
| 658 |
} |
| 659 |
|
| 660 |
impl From<DistortionScale> for f64 { |
| 661 |
#[inline] |
| 662 |
fn from(scale: DistortionScale) -> Self { |
| 663 |
scale.0 as f64 / (1 << DistortionScale::SHIFT) as f64 |
| 664 |
} |
| 665 |
} |
| 666 |
|
| 667 |
impl RawDistortion { |
| 668 |
#[inline] |
| 669 |
pub const fn new(dist: u64) -> Self { |
| 670 |
Self(dist) |
| 671 |
} |
| 672 |
} |
| 673 |
|
| 674 |
impl std::ops::Mul<DistortionScale> for RawDistortion { |
| 675 |
type Output = Distortion; |
| 676 |
#[inline] |
| 677 |
fn mul(self, rhs: DistortionScale) -> Distortion { |
| 678 |
Distortion(rhs.mul_u64(self.0)) |
| 679 |
} |
| 680 |
} |
| 681 |
|
| 682 |
impl Distortion { |
| 683 |
#[inline] |
| 684 |
pub const fn zero() -> Self { |
| 685 |
Self(0) |
| 686 |
} |
| 687 |
} |
| 688 |
|
| 689 |
impl std::ops::Mul<DistortionScale> for Distortion { |
| 690 |
type Output = ScaledDistortion; |
| 691 |
#[inline] |
| 692 |
fn mul(self, rhs: DistortionScale) -> ScaledDistortion { |
| 693 |
ScaledDistortion(rhs.mul_u64(self.0)) |
| 694 |
} |
| 695 |
} |
| 696 |
|
| 697 |
impl std::ops::AddAssign for Distortion { |
| 698 |
#[inline] |
| 699 |
fn add_assign(&mut self, other: Self) { |
| 700 |
self.0 += other.0; |
| 701 |
} |
| 702 |
} |
| 703 |
|
| 704 |
impl ScaledDistortion { |
| 705 |
#[inline] |
| 706 |
pub const fn zero() -> Self { |
| 707 |
Self(0) |
| 708 |
} |
| 709 |
} |
| 710 |
|
| 711 |
impl std::ops::AddAssign for ScaledDistortion { |
| 712 |
#[inline] |
| 713 |
fn add_assign(&mut self, other: Self) { |
| 714 |
self.0 += other.0; |
| 715 |
} |
| 716 |
} |
| 717 |
|
| 718 |
pub fn compute_rd_cost<T: Pixel>( |
| 719 |
fi: &FrameInvariants<T>, rate: u32, distortion: ScaledDistortion, |
| 720 |
) -> f64 { |
| 721 |
let rate_in_bits = (rate as f64) / ((1 << OD_BITRES) as f64); |
| 722 |
fi.lambda.mul_add(rate_in_bits, distortion.0 as f64) |
| 723 |
} |
| 724 |
|
| 725 |
pub fn rdo_tx_size_type<T: Pixel>( |
| 726 |
fi: &FrameInvariants<T>, ts: &mut TileStateMut<'_, T>, |
| 727 |
cw: &mut ContextWriter, bsize: BlockSize, tile_bo: TileBlockOffset, |
| 728 |
luma_mode: PredictionMode, ref_frames: [RefType; 2], mvs: [MotionVector; 2], |
| 729 |
skip: bool, |
| 730 |
) -> (TxSize, TxType) { |
| 731 |
let is_inter = !luma_mode.is_intra(); |
| 732 |
let mut tx_size = max_txsize_rect_lookup[bsize as usize]; |
| 733 |
|
| 734 |
if fi.enable_inter_txfm_split && is_inter && !skip { |
| 735 |
tx_size = sub_tx_size_map[tx_size as usize]; // Always choose one level split size |
| 736 |
} |
| 737 |
|
| 738 |
let mut best_tx_type = TxType::DCT_DCT; |
| 739 |
let mut best_tx_size = tx_size; |
| 740 |
let mut best_rd = f64::MAX; |
| 741 |
|
| 742 |
let do_rdo_tx_size = fi.tx_mode_select |
| 743 |
&& fi.config.speed_settings.transform.rdo_tx_decision |
| 744 |
&& !is_inter; |
| 745 |
let rdo_tx_depth = if do_rdo_tx_size { 2 } else { 0 }; |
| 746 |
let mut cw_checkpoint: Option<ContextWriterCheckpoint> = None; |
| 747 |
|
| 748 |
for _ in 0..=rdo_tx_depth { |
| 749 |
let tx_set = get_tx_set(tx_size, is_inter, fi.use_reduced_tx_set); |
| 750 |
|
| 751 |
let do_rdo_tx_type = tx_set > TxSet::TX_SET_DCTONLY |
| 752 |
&& fi.config.speed_settings.transform.rdo_tx_decision |
| 753 |
&& !is_inter |
| 754 |
&& !skip; |
| 755 |
|
| 756 |
if !do_rdo_tx_size && !do_rdo_tx_type { |
| 757 |
return (best_tx_size, best_tx_type); |
| 758 |
}; |
| 759 |
|
| 760 |
let tx_types = |
| 761 |
if do_rdo_tx_type { RAV1E_TX_TYPES } else { &[TxType::DCT_DCT] }; |
| 762 |
|
| 763 |
// Luma plane transform type decision |
| 764 |
let (tx_type, rd_cost) = rdo_tx_type_decision( |
| 765 |
fi, |
| 766 |
ts, |
| 767 |
cw, |
| 768 |
&mut cw_checkpoint, |
| 769 |
luma_mode, |
| 770 |
ref_frames, |
| 771 |
mvs, |
| 772 |
bsize, |
| 773 |
tile_bo, |
| 774 |
tx_size, |
| 775 |
tx_set, |
| 776 |
tx_types, |
| 777 |
best_rd, |
| 778 |
); |
| 779 |
|
| 780 |
if rd_cost < best_rd { |
| 781 |
best_tx_size = tx_size; |
| 782 |
best_tx_type = tx_type; |
| 783 |
best_rd = rd_cost; |
| 784 |
} |
| 785 |
|
| 786 |
debug_assert!(tx_size.width_log2() <= bsize.width_log2()); |
| 787 |
debug_assert!(tx_size.height_log2() <= bsize.height_log2()); |
| 788 |
debug_assert!( |
| 789 |
tx_size.sqr() <= TxSize::TX_32X32 || tx_type == TxType::DCT_DCT |
| 790 |
); |
| 791 |
|
| 792 |
let next_tx_size = sub_tx_size_map[tx_size as usize]; |
| 793 |
|
| 794 |
if next_tx_size == tx_size { |
| 795 |
break; |
| 796 |
} else { |
| 797 |
tx_size = next_tx_size; |
| 798 |
}; |
| 799 |
} |
| 800 |
|
| 801 |
(best_tx_size, best_tx_type) |
| 802 |
} |
| 803 |
|
| 804 |
#[inline] |
| 805 |
const fn dmv_in_range(mv: MotionVector, ref_mv: MotionVector) -> bool { |
| 806 |
let diff_row = mv.row as i32 - ref_mv.row as i32; |
| 807 |
let diff_col = mv.col as i32 - ref_mv.col as i32; |
| 808 |
diff_row >= MV_LOW |
| 809 |
&& diff_row <= MV_UPP |
| 810 |
&& diff_col >= MV_LOW |
| 811 |
&& diff_col <= MV_UPP |
| 812 |
} |
| 813 |
|
| 814 |
#[inline] |
| 815 |
#[profiling::function] |
| 816 |
fn luma_chroma_mode_rdo<T: Pixel>( |
| 817 |
luma_mode: PredictionMode, fi: &FrameInvariants<T>, bsize: BlockSize, |
| 818 |
tile_bo: TileBlockOffset, ts: &mut TileStateMut<'_, T>, |
| 819 |
cw: &mut ContextWriter, rdo_type: RDOType, |
| 820 |
cw_checkpoint: &ContextWriterCheckpoint, best: &mut PartitionParameters, |
| 821 |
mvs: [MotionVector; 2], ref_frames: [RefType; 2], |
| 822 |
mode_set_chroma: &[PredictionMode], luma_mode_is_intra: bool, |
| 823 |
mode_context: usize, mv_stack: &ArrayVec<CandidateMV, 9>, |
| 824 |
angle_delta: AngleDelta, |
| 825 |
) { |
| 826 |
let PlaneConfig { xdec, ydec, .. } = ts.input.planes[1].cfg; |
| 827 |
|
| 828 |
let is_chroma_block = |
| 829 |
has_chroma(tile_bo, bsize, xdec, ydec, fi.sequence.chroma_sampling); |
| 830 |
|
| 831 |
if !luma_mode_is_intra { |
| 832 |
let ref_mvs = if mv_stack.is_empty() { |
| 833 |
[MotionVector::default(); 2] |
| 834 |
} else { |
| 835 |
[mv_stack[0].this_mv, mv_stack[0].comp_mv] |
| 836 |
}; |
| 837 |
|
| 838 |
if (luma_mode == PredictionMode::NEWMV |
| 839 |
|| luma_mode == PredictionMode::NEW_NEWMV |
| 840 |
|| luma_mode == PredictionMode::NEW_NEARESTMV) |
| 841 |
&& !dmv_in_range(mvs[0], ref_mvs[0]) |
| 842 |
{ |
| 843 |
return; |
| 844 |
} |
| 845 |
|
| 846 |
if (luma_mode == PredictionMode::NEW_NEWMV |
| 847 |
|| luma_mode == PredictionMode::NEAREST_NEWMV) |
| 848 |
&& !dmv_in_range(mvs[1], ref_mvs[1]) |
| 849 |
{ |
| 850 |
return; |
| 851 |
} |
| 852 |
} |
| 853 |
|
| 854 |
// Find the best chroma prediction mode for the current luma prediction mode |
| 855 |
let mut chroma_rdo = |skip: bool| -> bool { |
| 856 |
use crate::segmentation::select_segment; |
| 857 |
|
| 858 |
let mut zero_distortion = false; |
| 859 |
|
| 860 |
for sidx in select_segment(fi, ts, tile_bo, bsize, skip) { |
| 861 |
cw.bc.blocks.set_segmentation_idx(tile_bo, bsize, sidx); |
| 862 |
|
| 863 |
let (tx_size, tx_type) = rdo_tx_size_type( |
| 864 |
fi, ts, cw, bsize, tile_bo, luma_mode, ref_frames, mvs, skip, |
| 865 |
); |
| 866 |
for &chroma_mode in mode_set_chroma.iter() { |
| 867 |
let wr = &mut WriterCounter::new(); |
| 868 |
let tell = wr.tell_frac(); |
| 869 |
|
| 870 |
if bsize >= BlockSize::BLOCK_8X8 && bsize.is_sqr() { |
| 871 |
cw.write_partition( |
| 872 |
wr, |
| 873 |
tile_bo, |
| 874 |
PartitionType::PARTITION_NONE, |
| 875 |
bsize, |
| 876 |
); |
| 877 |
} |
| 878 |
|
| 879 |
// TODO(yushin): luma and chroma would have different decision based on chroma format |
| 880 |
let need_recon_pixel = |
| 881 |
luma_mode_is_intra && tx_size.block_size() != bsize; |
| 882 |
|
| 883 |
encode_block_pre_cdef(&fi.sequence, ts, cw, wr, bsize, tile_bo, skip); |
| 884 |
let (has_coeff, tx_dist) = encode_block_post_cdef( |
| 885 |
fi, |
| 886 |
ts, |
| 887 |
cw, |
| 888 |
wr, |
| 889 |
luma_mode, |
| 890 |
chroma_mode, |
| 891 |
angle_delta, |
| 892 |
ref_frames, |
| 893 |
mvs, |
| 894 |
bsize, |
| 895 |
tile_bo, |
| 896 |
skip, |
| 897 |
CFLParams::default(), |
| 898 |
tx_size, |
| 899 |
tx_type, |
| 900 |
mode_context, |
| 901 |
mv_stack, |
| 902 |
rdo_type, |
| 903 |
need_recon_pixel, |
| 904 |
None, |
| 905 |
); |
| 906 |
|
| 907 |
let rate = wr.tell_frac() - tell; |
| 908 |
let distortion = if fi.use_tx_domain_distortion && !need_recon_pixel { |
| 909 |
compute_tx_distortion( |
| 910 |
fi, |
| 911 |
ts, |
| 912 |
bsize, |
| 913 |
is_chroma_block, |
| 914 |
tile_bo, |
| 915 |
tx_dist, |
| 916 |
skip, |
| 917 |
false, |
| 918 |
) |
| 919 |
} else { |
| 920 |
compute_distortion(fi, ts, bsize, is_chroma_block, tile_bo, false) |
| 921 |
}; |
| 922 |
let is_zero_dist = distortion.0 == 0; |
| 923 |
let rd = compute_rd_cost(fi, rate, distortion); |
| 924 |
if rd < best.rd_cost { |
| 925 |
//if rd < best.rd_cost || luma_mode == PredictionMode::NEW_NEWMV { |
| 926 |
best.rd_cost = rd; |
| 927 |
best.pred_mode_luma = luma_mode; |
| 928 |
best.pred_mode_chroma = chroma_mode; |
| 929 |
best.angle_delta = angle_delta; |
| 930 |
best.ref_frames = ref_frames; |
| 931 |
best.mvs = mvs; |
| 932 |
best.skip = skip; |
| 933 |
best.has_coeff = has_coeff; |
| 934 |
best.tx_size = tx_size; |
| 935 |
best.tx_type = tx_type; |
| 936 |
best.sidx = sidx; |
| 937 |
zero_distortion = is_zero_dist; |
| 938 |
} |
| 939 |
|
| 940 |
cw.rollback(cw_checkpoint); |
| 941 |
} |
| 942 |
} |
| 943 |
|
| 944 |
zero_distortion |
| 945 |
}; |
| 946 |
|
| 947 |
// Don't skip when using intra modes |
| 948 |
let zero_distortion = |
| 949 |
if !luma_mode_is_intra { chroma_rdo(true) } else { false }; |
| 950 |
// early skip |
| 951 |
if !zero_distortion { |
| 952 |
chroma_rdo(false); |
| 953 |
} |
| 954 |
} |
| 955 |
|
| 956 |
/// RDO-based mode decision |
| 957 |
/// |
| 958 |
/// # Panics |
| 959 |
/// |
| 960 |
/// - If the best RD found is negative. |
| 961 |
/// This should never happen and indicates a development error. |
| 962 |
#[profiling::function] |
| 963 |
pub fn rdo_mode_decision<T: Pixel>( |
| 964 |
fi: &FrameInvariants<T>, ts: &mut TileStateMut<'_, T>, |
| 965 |
cw: &mut ContextWriter, bsize: BlockSize, tile_bo: TileBlockOffset, |
| 966 |
inter_cfg: &InterConfig, |
| 967 |
) -> PartitionParameters { |
| 968 |
let PlaneConfig { xdec, ydec, .. } = ts.input.planes[1].cfg; |
| 969 |
let cw_checkpoint = cw.checkpoint(&tile_bo, fi.sequence.chroma_sampling); |
| 970 |
|
| 971 |
let rdo_type = if fi.use_tx_domain_rate { |
| 972 |
RDOType::TxDistEstRate |
| 973 |
} else if fi.use_tx_domain_distortion { |
| 974 |
RDOType::TxDistRealRate |
| 975 |
} else { |
| 976 |
RDOType::PixelDistRealRate |
| 977 |
}; |
| 978 |
|
| 979 |
let mut best = if fi.frame_type.has_inter() { |
| 980 |
assert!(fi.frame_type != FrameType::KEY); |
| 981 |
|
| 982 |
inter_frame_rdo_mode_decision( |
| 983 |
fi, |
| 984 |
ts, |
| 985 |
cw, |
| 986 |
bsize, |
| 987 |
tile_bo, |
| 988 |
inter_cfg, |
| 989 |
&cw_checkpoint, |
| 990 |
rdo_type, |
| 991 |
) |
| 992 |
} else { |
| 993 |
PartitionParameters::default() |
| 994 |
}; |
| 995 |
|
| 996 |
let is_chroma_block = |
| 997 |
has_chroma(tile_bo, bsize, xdec, ydec, fi.sequence.chroma_sampling); |
| 998 |
|
| 999 |
if !best.skip { |
| 1000 |
best = intra_frame_rdo_mode_decision( |
| 1001 |
fi, |
| 1002 |
ts, |
| 1003 |
cw, |
| 1004 |
bsize, |
| 1005 |
tile_bo, |
| 1006 |
&cw_checkpoint, |
| 1007 |
rdo_type, |
| 1008 |
best, |
| 1009 |
is_chroma_block, |
| 1010 |
); |
| 1011 |
} |
| 1012 |
|
| 1013 |
if best.pred_mode_luma.is_intra() && is_chroma_block && bsize.cfl_allowed() { |
| 1014 |
cw.bc.blocks.set_segmentation_idx(tile_bo, bsize, best.sidx); |
| 1015 |
|
| 1016 |
let chroma_mode = PredictionMode::UV_CFL_PRED; |
| 1017 |
let cw_checkpoint = cw.checkpoint(&tile_bo, fi.sequence.chroma_sampling); |
| 1018 |
let mut wr = WriterCounter::new(); |
| 1019 |
let angle_delta = AngleDelta { y: best.angle_delta.y, uv: 0 }; |
| 1020 |
|
| 1021 |
write_tx_blocks( |
| 1022 |
fi, |
| 1023 |
ts, |
| 1024 |
cw, |
| 1025 |
&mut wr, |
| 1026 |
best.pred_mode_luma, |
| 1027 |
best.pred_mode_luma, |
| 1028 |
angle_delta, |
| 1029 |
tile_bo, |
| 1030 |
bsize, |
| 1031 |
best.tx_size, |
| 1032 |
best.tx_type, |
| 1033 |
false, |
| 1034 |
CFLParams::default(), |
| 1035 |
true, |
| 1036 |
rdo_type, |
| 1037 |
true, |
| 1038 |
); |
| 1039 |
cw.rollback(&cw_checkpoint); |
| 1040 |
if fi.sequence.chroma_sampling != ChromaSampling::Cs400 { |
| 1041 |
if let Some(cfl) = rdo_cfl_alpha(ts, tile_bo, bsize, best.tx_size, fi) { |
| 1042 |
let mut wr = WriterCounter::new(); |
| 1043 |
let tell = wr.tell_frac(); |
| 1044 |
|
| 1045 |
encode_block_pre_cdef( |
| 1046 |
&fi.sequence, |
| 1047 |
ts, |
| 1048 |
cw, |
| 1049 |
&mut wr, |
| 1050 |
bsize, |
| 1051 |
tile_bo, |
| 1052 |
best.skip, |
| 1053 |
); |
| 1054 |
let (has_coeff, _) = encode_block_post_cdef( |
| 1055 |
fi, |
| 1056 |
ts, |
| 1057 |
cw, |
| 1058 |
&mut wr, |
| 1059 |
best.pred_mode_luma, |
| 1060 |
chroma_mode, |
| 1061 |
angle_delta, |
| 1062 |
best.ref_frames, |
| 1063 |
best.mvs, |
| 1064 |
bsize, |
| 1065 |
tile_bo, |
| 1066 |
best.skip, |
| 1067 |
cfl, |
| 1068 |
best.tx_size, |
| 1069 |
best.tx_type, |
| 1070 |
0, |
| 1071 |
&[], |
| 1072 |
rdo_type, |
| 1073 |
true, // For CFL, luma should be always reconstructed. |
| 1074 |
None, |
| 1075 |
); |
| 1076 |
|
| 1077 |
let rate = wr.tell_frac() - tell; |
| 1078 |
|
| 1079 |
// For CFL, tx-domain distortion is not an option. |
| 1080 |
let distortion = |
| 1081 |
compute_distortion(fi, ts, bsize, is_chroma_block, tile_bo, false); |
| 1082 |
let rd = compute_rd_cost(fi, rate, distortion); |
| 1083 |
if rd < best.rd_cost { |
| 1084 |
best.rd_cost = rd; |
| 1085 |
best.pred_mode_chroma = chroma_mode; |
| 1086 |
best.angle_delta = angle_delta; |
| 1087 |
best.has_coeff = has_coeff; |
| 1088 |
best.pred_cfl_params = cfl; |
| 1089 |
} |
| 1090 |
|
| 1091 |
cw.rollback(&cw_checkpoint); |
| 1092 |
} |
| 1093 |
} |
| 1094 |
} |
| 1095 |
|
| 1096 |
cw.bc.blocks.set_mode(tile_bo, bsize, best.pred_mode_luma); |
| 1097 |
cw.bc.blocks.set_ref_frames(tile_bo, bsize, best.ref_frames); |
| 1098 |
cw.bc.blocks.set_motion_vectors(tile_bo, bsize, best.mvs); |
| 1099 |
|
| 1100 |
assert!(best.rd_cost >= 0_f64); |
| 1101 |
|
| 1102 |
PartitionParameters { |
| 1103 |
bo: tile_bo, |
| 1104 |
bsize, |
| 1105 |
pred_mode_luma: best.pred_mode_luma, |
| 1106 |
pred_mode_chroma: best.pred_mode_chroma, |
| 1107 |
pred_cfl_params: best.pred_cfl_params, |
| 1108 |
angle_delta: best.angle_delta, |
| 1109 |
ref_frames: best.ref_frames, |
| 1110 |
mvs: best.mvs, |
| 1111 |
rd_cost: best.rd_cost, |
| 1112 |
skip: best.skip, |
| 1113 |
has_coeff: best.has_coeff, |
| 1114 |
tx_size: best.tx_size, |
| 1115 |
tx_type: best.tx_type, |
| 1116 |
sidx: best.sidx, |
| 1117 |
} |
| 1118 |
} |
| 1119 |
|
| 1120 |
#[profiling::function] |
| 1121 |
fn inter_frame_rdo_mode_decision<T: Pixel>( |
| 1122 |
fi: &FrameInvariants<T>, ts: &mut TileStateMut<'_, T>, |
| 1123 |
cw: &mut ContextWriter, bsize: BlockSize, tile_bo: TileBlockOffset, |
| 1124 |
inter_cfg: &InterConfig, cw_checkpoint: &ContextWriterCheckpoint, |
| 1125 |
rdo_type: RDOType, |
| 1126 |
) -> PartitionParameters { |
| 1127 |
let mut best = PartitionParameters::default(); |
| 1128 |
|
| 1129 |
// we can never have more than 7 reference frame sets |
| 1130 |
let mut ref_frames_set = ArrayVec::<_, 7>::new(); |
| 1131 |
// again, max of 7 ref slots |
| 1132 |
let mut ref_slot_set = ArrayVec::<_, 7>::new(); |
| 1133 |
// our implementation never returns more than 3 at the moment |
| 1134 |
let mut mvs_from_me = ArrayVec::<_, 3>::new(); |
| 1135 |
let mut fwdref = None; |
| 1136 |
let mut bwdref = None; |
| 1137 |
|
| 1138 |
for i in inter_cfg.allowed_ref_frames().iter().copied() { |
| 1139 |
// Don't search LAST3 since it's used only for probs |
| 1140 |
if i == LAST3_FRAME { |
| 1141 |
continue; |
| 1142 |
} |
| 1143 |
|
| 1144 |
if !ref_slot_set.contains(&fi.ref_frames[i.to_index()]) { |
| 1145 |
if fwdref.is_none() && i.is_fwd_ref() { |
| 1146 |
fwdref = Some(ref_frames_set.len()); |
| 1147 |
} |
| 1148 |
if bwdref.is_none() && i.is_bwd_ref() { |
| 1149 |
bwdref = Some(ref_frames_set.len()); |
| 1150 |
} |
| 1151 |
ref_frames_set.push([i, NONE_FRAME]); |
| 1152 |
let slot_idx = fi.ref_frames[i.to_index()]; |
| 1153 |
ref_slot_set.push(slot_idx); |
| 1154 |
} |
| 1155 |
} |
| 1156 |
assert!(!ref_frames_set.is_empty()); |
| 1157 |
|
| 1158 |
let mut inter_mode_set = ArrayVec::<(PredictionMode, usize), 20>::new(); |
| 1159 |
let mut mvs_set = ArrayVec::<[MotionVector; 2], 20>::new(); |
| 1160 |
let mut satds = ArrayVec::<u32, 20>::new(); |
| 1161 |
let mut mv_stacks = ArrayVec::<_, 20>::new(); |
| 1162 |
let mut mode_contexts = ArrayVec::<_, 7>::new(); |
| 1163 |
|
| 1164 |
for (i, &ref_frames) in ref_frames_set.iter().enumerate() { |
| 1165 |
let mut mv_stack = ArrayVec::<CandidateMV, 9>::new(); |
| 1166 |
mode_contexts.push(cw.find_mvrefs( |
| 1167 |
tile_bo, |
| 1168 |
ref_frames, |
| 1169 |
&mut mv_stack, |
| 1170 |
bsize, |
| 1171 |
fi, |
| 1172 |
false, |
| 1173 |
)); |
| 1174 |
|
| 1175 |
let mut pmv = [MotionVector::default(); 2]; |
| 1176 |
if !mv_stack.is_empty() { |
| 1177 |
pmv[0] = mv_stack[0].this_mv; |
| 1178 |
} |
| 1179 |
if mv_stack.len() > 1 { |
| 1180 |
pmv[1] = mv_stack[1].this_mv; |
| 1181 |
} |
| 1182 |
|
| 1183 |
let res = estimate_motion( |
| 1184 |
fi, |
| 1185 |
ts, |
| 1186 |
bsize.width(), |
| 1187 |
bsize.height(), |
| 1188 |
tile_bo, |
| 1189 |
ref_frames[0], |
| 1190 |
Some(pmv), |
| 1191 |
MVSamplingMode::CORNER { right: true, bottom: true }, |
| 1192 |
false, |
| 1193 |
0, |
| 1194 |
None, |
| 1195 |
) |
| 1196 |
.unwrap_or_else(MotionSearchResult::empty); |
| 1197 |
let b_me = res.mv; |
| 1198 |
|
| 1199 |
mvs_from_me.push([b_me, MotionVector::default()]); |
| 1200 |
|
| 1201 |
for &x in RAV1E_INTER_MODES_MINIMAL { |
| 1202 |
inter_mode_set.push((x, i)); |
| 1203 |
} |
| 1204 |
if !mv_stack.is_empty() { |
| 1205 |
inter_mode_set.push((PredictionMode::NEAR0MV, i)); |
| 1206 |
} |
| 1207 |
if mv_stack.len() >= 2 { |
| 1208 |
inter_mode_set.push((PredictionMode::GLOBALMV, i)); |
| 1209 |
} |
| 1210 |
let include_near_mvs = fi.config.speed_settings.motion.include_near_mvs; |
| 1211 |
if include_near_mvs { |
| 1212 |
if mv_stack.len() >= 3 { |
| 1213 |
inter_mode_set.push((PredictionMode::NEAR1MV, i)); |
| 1214 |
} |
| 1215 |
if mv_stack.len() >= 4 { |
| 1216 |
inter_mode_set.push((PredictionMode::NEAR2MV, i)); |
| 1217 |
} |
| 1218 |
} |
| 1219 |
let same_row_col = |x: &CandidateMV| { |
| 1220 |
x.this_mv.row == mvs_from_me[i][0].row |
| 1221 |
&& x.this_mv.col == mvs_from_me[i][0].col |
| 1222 |
}; |
| 1223 |
if !mv_stack |
| 1224 |
.iter() |
| 1225 |
.take(if include_near_mvs { 4 } else { 2 }) |
| 1226 |
.any(same_row_col) |
| 1227 |
&& (mvs_from_me[i][0].row != 0 || mvs_from_me[i][0].col != 0) |
| 1228 |
{ |
| 1229 |
inter_mode_set.push((PredictionMode::NEWMV, i)); |
| 1230 |
} |
| 1231 |
|
| 1232 |
mv_stacks.push(mv_stack); |
| 1233 |
} |
| 1234 |
|
| 1235 |
let sz = bsize.width_mi().min(bsize.height_mi()); |
| 1236 |
|
| 1237 |
// To use non single reference modes, block width and height must be greater than 4. |
| 1238 |
if fi.reference_mode != ReferenceMode::SINGLE && sz >= 2 { |
| 1239 |
// Adding compound candidate |
| 1240 |
if let Some(r0) = fwdref { |
| 1241 |
if let Some(r1) = bwdref { |
| 1242 |
let ref_frames = [ref_frames_set[r0][0], ref_frames_set[r1][0]]; |
| 1243 |
ref_frames_set.push(ref_frames); |
| 1244 |
let mv0 = mvs_from_me[r0][0]; |
| 1245 |
let mv1 = mvs_from_me[r1][0]; |
| 1246 |
mvs_from_me.push([mv0, mv1]); |
| 1247 |
let mut mv_stack = ArrayVec::<CandidateMV, 9>::new(); |
| 1248 |
mode_contexts.push(cw.find_mvrefs( |
| 1249 |
tile_bo, |
| 1250 |
ref_frames, |
| 1251 |
&mut mv_stack, |
| 1252 |
bsize, |
| 1253 |
fi, |
| 1254 |
true, |
| 1255 |
)); |
| 1256 |
for &x in RAV1E_INTER_COMPOUND_MODES { |
| 1257 |
// exclude any NEAR mode based on speed setting |
| 1258 |
if fi.config.speed_settings.motion.include_near_mvs |
| 1259 |
|| !x.has_nearmv() |
| 1260 |
{ |
| 1261 |
let mv_stack_idx = ref_frames_set.len() - 1; |
| 1262 |
// exclude NEAR modes if the mv_stack is too short |
| 1263 |
if !(x.has_nearmv() && x.ref_mv_idx() >= mv_stack.len()) { |
| 1264 |
inter_mode_set.push((x, mv_stack_idx)); |
| 1265 |
} |
| 1266 |
} |
| 1267 |
} |
| 1268 |
mv_stacks.push(mv_stack); |
| 1269 |
} |
| 1270 |
} |
| 1271 |
} |
| 1272 |
|
| 1273 |
let num_modes_rdo = if fi.config.speed_settings.prediction.prediction_modes |
| 1274 |
>= PredictionModesSetting::ComplexAll |
| 1275 |
{ |
| 1276 |
inter_mode_set.len() |
| 1277 |
} else { |
| 1278 |
9 // This number is determined by AWCY test |
| 1279 |
}; |
| 1280 |
|
| 1281 |
inter_mode_set.iter().for_each(|&(luma_mode, i)| { |
| 1282 |
let mvs = match luma_mode { |
| 1283 |
PredictionMode::NEWMV | PredictionMode::NEW_NEWMV => mvs_from_me[i], |
| 1284 |
PredictionMode::NEARESTMV | PredictionMode::NEAREST_NEARESTMV => { |
| 1285 |
if !mv_stacks[i].is_empty() { |
| 1286 |
[mv_stacks[i][0].this_mv, mv_stacks[i][0].comp_mv] |
| 1287 |
} else { |
| 1288 |
[MotionVector::default(); 2] |
| 1289 |
} |
| 1290 |
} |
| 1291 |
PredictionMode::NEAR0MV | PredictionMode::NEAR_NEAR0MV => { |
| 1292 |
if mv_stacks[i].len() > 1 { |
| 1293 |
[mv_stacks[i][1].this_mv, mv_stacks[i][1].comp_mv] |
| 1294 |
} else { |
| 1295 |
[MotionVector::default(); 2] |
| 1296 |
} |
| 1297 |
} |
| 1298 |
PredictionMode::NEAR1MV |
| 1299 |
| PredictionMode::NEAR2MV |
| 1300 |
| PredictionMode::NEAR_NEAR1MV |
| 1301 |
| PredictionMode::NEAR_NEAR2MV => [ |
| 1302 |
mv_stacks[i][luma_mode.ref_mv_idx()].this_mv, |
| 1303 |
mv_stacks[i][luma_mode.ref_mv_idx()].comp_mv, |
| 1304 |
], |
| 1305 |
PredictionMode::NEAREST_NEWMV => { |
| 1306 |
[mv_stacks[i][0].this_mv, mvs_from_me[i][1]] |
| 1307 |
} |
| 1308 |
PredictionMode::NEW_NEARESTMV => { |
| 1309 |
[mvs_from_me[i][0], mv_stacks[i][0].comp_mv] |
| 1310 |
} |
| 1311 |
PredictionMode::GLOBALMV | PredictionMode::GLOBAL_GLOBALMV => { |
| 1312 |
[MotionVector::default(); 2] |
| 1313 |
} |
| 1314 |
_ => { |
| 1315 |
unimplemented!(); |
| 1316 |
} |
| 1317 |
}; |
| 1318 |
mvs_set.push(mvs); |
| 1319 |
|
| 1320 |
// Calculate SATD for each mode |
| 1321 |
if num_modes_rdo != inter_mode_set.len() { |
| 1322 |
let tile_rect = ts.tile_rect(); |
| 1323 |
let rec = &mut ts.rec.planes[0]; |
| 1324 |
let po = tile_bo.plane_offset(rec.plane_cfg); |
| 1325 |
let mut rec_region = |
| 1326 |
rec.subregion_mut(Area::BlockStartingAt { bo: tile_bo.0 }); |
| 1327 |
|
| 1328 |
luma_mode.predict_inter( |
| 1329 |
fi, |
| 1330 |
tile_rect, |
| 1331 |
0, |
| 1332 |
po, |
| 1333 |
&mut rec_region, |
| 1334 |
bsize.width(), |
| 1335 |
bsize.height(), |
| 1336 |
ref_frames_set[i], |
| 1337 |
mvs, |
| 1338 |
&mut ts.inter_compound_buffers, |
| 1339 |
); |
| 1340 |
|
| 1341 |
let plane_org = ts.input_tile.planes[0] |
| 1342 |
.subregion(Area::BlockStartingAt { bo: tile_bo.0 }); |
| 1343 |
let plane_ref = rec_region.as_const(); |
| 1344 |
|
| 1345 |
let satd = get_satd( |
| 1346 |
&plane_org, |
| 1347 |
&plane_ref, |
| 1348 |
bsize.width(), |
| 1349 |
bsize.height(), |
| 1350 |
fi.sequence.bit_depth, |
| 1351 |
fi.cpu_feature_level, |
| 1352 |
); |
| 1353 |
satds.push(satd); |
| 1354 |
} else { |
| 1355 |
satds.push(0); |
| 1356 |
} |
| 1357 |
}); |
| 1358 |
|
| 1359 |
let mut sorted = |
| 1360 |
izip!(inter_mode_set, mvs_set, satds).collect::<ArrayVec<_, 20>>(); |
| 1361 |
if num_modes_rdo != sorted.len() { |
| 1362 |
sorted.sort_by_key(|((_mode, _i), _mvs, satd)| *satd); |
| 1363 |
} |
| 1364 |
|
| 1365 |
sorted.iter().take(num_modes_rdo).for_each( |
| 1366 |
|&((luma_mode, i), mvs, _satd)| { |
| 1367 |
let mode_set_chroma = ArrayVec::from([luma_mode]); |
| 1368 |
|
| 1369 |
luma_chroma_mode_rdo( |
| 1370 |
luma_mode, |
| 1371 |
fi, |
| 1372 |
bsize, |
| 1373 |
tile_bo, |
| 1374 |
ts, |
| 1375 |
cw, |
| 1376 |
rdo_type, |
| 1377 |
cw_checkpoint, |
| 1378 |
&mut best, |
| 1379 |
mvs, |
| 1380 |
ref_frames_set[i], |
| 1381 |
&mode_set_chroma, |
| 1382 |
false, |
| 1383 |
mode_contexts[i], |
| 1384 |
&mv_stacks[i], |
| 1385 |
AngleDelta::default(), |
| 1386 |
); |
| 1387 |
}, |
| 1388 |
); |
| 1389 |
|
| 1390 |
best |
| 1391 |
} |
| 1392 |
|
| 1393 |
#[profiling::function] |
| 1394 |
fn intra_frame_rdo_mode_decision<T: Pixel>( |
| 1395 |
fi: &FrameInvariants<T>, ts: &mut TileStateMut<'_, T>, |
| 1396 |
cw: &mut ContextWriter, bsize: BlockSize, tile_bo: TileBlockOffset, |
| 1397 |
cw_checkpoint: &ContextWriterCheckpoint, rdo_type: RDOType, |
| 1398 |
mut best: PartitionParameters, is_chroma_block: bool, |
| 1399 |
) -> PartitionParameters { |
| 1400 |
let mut modes = ArrayVec::<_, INTRA_MODES>::new(); |
| 1401 |
|
| 1402 |
// Reduce number of prediction modes at higher speed levels |
| 1403 |
let num_modes_rdo = if (fi.frame_type == FrameType::KEY |
| 1404 |
&& fi.config.speed_settings.prediction.prediction_modes |
| 1405 |
>= PredictionModesSetting::ComplexKeyframes) |
| 1406 |
|| (fi.frame_type.has_inter() |
| 1407 |
&& fi.config.speed_settings.prediction.prediction_modes |
| 1408 |
>= PredictionModesSetting::ComplexAll) |
| 1409 |
{ |
| 1410 |
7 |
| 1411 |
} else { |
| 1412 |
3 |
| 1413 |
}; |
| 1414 |
|
| 1415 |
let intra_mode_set = RAV1E_INTRA_MODES; |
| 1416 |
|
| 1417 |
// Find mode with lowest rate cost |
| 1418 |
{ |
| 1419 |
use crate::ec::cdf_to_pdf; |
| 1420 |
|
| 1421 |
let probs_all = cdf_to_pdf(if fi.frame_type.has_inter() { |
| 1422 |
cw.get_cdf_intra_mode(bsize) |
| 1423 |
} else { |
| 1424 |
cw.get_cdf_intra_mode_kf(tile_bo) |
| 1425 |
}); |
| 1426 |
|
| 1427 |
modes.try_extend_from_slice(intra_mode_set).unwrap(); |
| 1428 |
modes.sort_by_key(|&a| !probs_all[a as usize]); |
| 1429 |
} |
| 1430 |
|
| 1431 |
// If tx partition (i.e. fi.tx_mode_select) is enabled, the below intra prediction screening |
| 1432 |
// may be improved by emulating prediction for each tx block. |
| 1433 |
{ |
| 1434 |
let satds = { |
| 1435 |
// FIXME: If tx partition is used, this whole sads block should be fixed |
| 1436 |
let tx_size = bsize.tx_size(); |
| 1437 |
let mut edge_buf = Aligned::uninit_array(); |
| 1438 |
let edge_buf = { |
| 1439 |
let rec = &ts.rec.planes[0].as_const(); |
| 1440 |
let po = tile_bo.plane_offset(rec.plane_cfg); |
| 1441 |
// FIXME: If tx partition is used, get_intra_edges() should be called for each tx block |
| 1442 |
get_intra_edges( |
| 1443 |
&mut edge_buf, |
| 1444 |
rec, |
| 1445 |
tile_bo, |
| 1446 |
0, |
| 1447 |
0, |
| 1448 |
bsize, |
| 1449 |
po, |
| 1450 |
tx_size, |
| 1451 |
fi.sequence.bit_depth, |
| 1452 |
None, |
| 1453 |
fi.sequence.enable_intra_edge_filter, |
| 1454 |
IntraParam::None, |
| 1455 |
) |
| 1456 |
}; |
| 1457 |
|
| 1458 |
let ief_params = if fi.sequence.enable_intra_edge_filter { |
| 1459 |
let above_block_info = ts.above_block_info(tile_bo, 0, 0); |
| 1460 |
let left_block_info = ts.left_block_info(tile_bo, 0, 0); |
| 1461 |
Some(IntraEdgeFilterParameters::new( |
| 1462 |
0, |
| 1463 |
above_block_info, |
| 1464 |
left_block_info, |
| 1465 |
)) |
| 1466 |
} else { |
| 1467 |
None |
| 1468 |
}; |
| 1469 |
|
| 1470 |
let mut satds_all = [0; INTRA_MODES]; |
| 1471 |
for &luma_mode in modes.iter().skip(num_modes_rdo / 2) { |
| 1472 |
let tile_rect = ts.tile_rect(); |
| 1473 |
let rec = &mut ts.rec.planes[0]; |
| 1474 |
let mut rec_region = |
| 1475 |
rec.subregion_mut(Area::BlockStartingAt { bo: tile_bo.0 }); |
| 1476 |
// FIXME: If tx partition is used, luma_mode.predict_intra() should be called for each tx block |
| 1477 |
luma_mode.predict_intra( |
| 1478 |
tile_rect, |
| 1479 |
&mut rec_region, |
| 1480 |
tx_size, |
| 1481 |
fi.sequence.bit_depth, |
| 1482 |
&[0i16; 2], |
| 1483 |
IntraParam::None, |
| 1484 |
if luma_mode.is_directional() { ief_params } else { None }, |
| 1485 |
&edge_buf, |
| 1486 |
fi.cpu_feature_level, |
| 1487 |
); |
| 1488 |
|
| 1489 |
let plane_org = ts.input_tile.planes[0] |
| 1490 |
.subregion(Area::BlockStartingAt { bo: tile_bo.0 }); |
| 1491 |
let plane_ref = rec_region.as_const(); |
| 1492 |
|
| 1493 |
satds_all[luma_mode as usize] = get_satd( |
| 1494 |
&plane_org, |
| 1495 |
&plane_ref, |
| 1496 |
tx_size.width(), |
| 1497 |
tx_size.height(), |
| 1498 |
fi.sequence.bit_depth, |
| 1499 |
fi.cpu_feature_level, |
| 1500 |
); |
| 1501 |
} |
| 1502 |
satds_all |
| 1503 |
}; |
| 1504 |
|
| 1505 |
modes[num_modes_rdo / 2..].sort_by_key(|&a| satds[a as usize]); |
| 1506 |
} |
| 1507 |
|
| 1508 |
debug_assert!(num_modes_rdo >= 1); |
| 1509 |
|
| 1510 |
modes.iter().take(num_modes_rdo).for_each(|&luma_mode| { |
| 1511 |
let mvs = [MotionVector::default(); 2]; |
| 1512 |
let ref_frames = [INTRA_FRAME, NONE_FRAME]; |
| 1513 |
let mut mode_set_chroma = ArrayVec::<_, 2>::new(); |
| 1514 |
mode_set_chroma.push(luma_mode); |
| 1515 |
if is_chroma_block && luma_mode != PredictionMode::DC_PRED { |
| 1516 |
mode_set_chroma.push(PredictionMode::DC_PRED); |
| 1517 |
} |
| 1518 |
luma_chroma_mode_rdo( |
| 1519 |
luma_mode, |
| 1520 |
fi, |
| 1521 |
bsize, |
| 1522 |
tile_bo, |
| 1523 |
ts, |
| 1524 |
cw, |
| 1525 |
rdo_type, |
| 1526 |
cw_checkpoint, |
| 1527 |
&mut best, |
| 1528 |
mvs, |
| 1529 |
ref_frames, |
| 1530 |
&mode_set_chroma, |
| 1531 |
true, |
| 1532 |
0, |
| 1533 |
&ArrayVec::<CandidateMV, 9>::new(), |
| 1534 |
AngleDelta::default(), |
| 1535 |
); |
| 1536 |
}); |
| 1537 |
|
| 1538 |
if fi.config.speed_settings.prediction.fine_directional_intra |
| 1539 |
&& bsize >= BlockSize::BLOCK_8X8 |
| 1540 |
{ |
| 1541 |
// Find the best angle delta for the current best prediction mode |
| 1542 |
let luma_deltas = best.pred_mode_luma.angle_delta_count(); |
| 1543 |
let chroma_deltas = best.pred_mode_chroma.angle_delta_count(); |
| 1544 |
|
| 1545 |
let mvs = [MotionVector::default(); 2]; |
| 1546 |
let ref_frames = [INTRA_FRAME, NONE_FRAME]; |
| 1547 |
let mode_set_chroma = [best.pred_mode_chroma]; |
| 1548 |
let mv_stack = ArrayVec::<_, 9>::new(); |
| 1549 |
let mut best_angle_delta = best.angle_delta; |
| 1550 |
let mut angle_delta_rdo = |y, uv| -> AngleDelta { |
| 1551 |
if best.angle_delta.y != y || best.angle_delta.uv != uv { |
| 1552 |
luma_chroma_mode_rdo( |
| 1553 |
best.pred_mode_luma, |
| 1554 |
fi, |
| 1555 |
bsize, |
| 1556 |
tile_bo, |
| 1557 |
ts, |
| 1558 |
cw, |
| 1559 |
rdo_type, |
| 1560 |
cw_checkpoint, |
| 1561 |
&mut best, |
| 1562 |
mvs, |
| 1563 |
ref_frames, |
| 1564 |
&mode_set_chroma, |
| 1565 |
true, |
| 1566 |
0, |
| 1567 |
&mv_stack, |
| 1568 |
AngleDelta { y, uv }, |
| 1569 |
); |
| 1570 |
} |
| 1571 |
best.angle_delta |
| 1572 |
}; |
| 1573 |
|
| 1574 |
for i in 0..luma_deltas { |
| 1575 |
let angle_delta_y = |
| 1576 |
if luma_deltas == 1 { 0 } else { i - MAX_ANGLE_DELTA as i8 }; |
| 1577 |
best_angle_delta = angle_delta_rdo(angle_delta_y, best_angle_delta.uv); |
| 1578 |
} |
| 1579 |
for j in 0..chroma_deltas { |
| 1580 |
let angle_delta_uv = |
| 1581 |
if chroma_deltas == 1 { 0 } else { j - MAX_ANGLE_DELTA as i8 }; |
| 1582 |
best_angle_delta = angle_delta_rdo(best_angle_delta.y, angle_delta_uv); |
| 1583 |
} |
| 1584 |
} |
| 1585 |
|
| 1586 |
best |
| 1587 |
} |
| 1588 |
|
| 1589 |
/// # Panics |
| 1590 |
/// |
| 1591 |
/// - If the block size is invalid for subsampling. |
| 1592 |
#[profiling::function] |
| 1593 |
pub fn rdo_cfl_alpha<T: Pixel>( |
| 1594 |
ts: &mut TileStateMut<'_, T>, tile_bo: TileBlockOffset, bsize: BlockSize, |
| 1595 |
luma_tx_size: TxSize, fi: &FrameInvariants<T>, |
| 1596 |
) -> Option<CFLParams> { |
| 1597 |
let PlaneConfig { xdec, ydec, .. } = ts.input.planes[1].cfg; |
| 1598 |
let uv_tx_size = bsize.largest_chroma_tx_size(xdec, ydec); |
| 1599 |
debug_assert!( |
| 1600 |
bsize.subsampled_size(xdec, ydec).unwrap() == uv_tx_size.block_size() |
| 1601 |
); |
| 1602 |
|
| 1603 |
let frame_bo = ts.to_frame_block_offset(tile_bo); |
| 1604 |
let (visible_tx_w, visible_tx_h) = clip_visible_bsize( |
| 1605 |
(fi.width + xdec) >> xdec, |
| 1606 |
(fi.height + ydec) >> ydec, |
| 1607 |
uv_tx_size.block_size(), |
| 1608 |
(frame_bo.0.x << MI_SIZE_LOG2) >> xdec, |
| 1609 |
(frame_bo.0.y << MI_SIZE_LOG2) >> ydec, |
| 1610 |
); |
| 1611 |
|
| 1612 |
if visible_tx_w == 0 || visible_tx_h == 0 { |
| 1613 |
return None; |
| 1614 |
}; |
| 1615 |
let mut ac = Aligned::<[MaybeUninit<i16>; 32 * 32]>::uninit_array(); |
| 1616 |
let ac = luma_ac(&mut ac.data, ts, tile_bo, bsize, luma_tx_size, fi); |
| 1617 |
let best_alpha: ArrayVec<i16, 2> = (1..3) |
| 1618 |
.map(|p| { |
| 1619 |
let &PlaneConfig { xdec, ydec, .. } = ts.rec.planes[p].plane_cfg; |
| 1620 |
let tile_rect = ts.tile_rect().decimated(xdec, ydec); |
| 1621 |
let rec = &mut ts.rec.planes[p]; |
| 1622 |
let input = &ts.input_tile.planes[p]; |
| 1623 |
let po = tile_bo.plane_offset(rec.plane_cfg); |
| 1624 |
let mut edge_buf = Aligned::uninit_array(); |
| 1625 |
let edge_buf = get_intra_edges( |
| 1626 |
&mut edge_buf, |
| 1627 |
&rec.as_const(), |
| 1628 |
tile_bo, |
| 1629 |
0, |
| 1630 |
0, |
| 1631 |
bsize, |
| 1632 |
po, |
| 1633 |
uv_tx_size, |
| 1634 |
fi.sequence.bit_depth, |
| 1635 |
Some(PredictionMode::UV_CFL_PRED), |
| 1636 |
fi.sequence.enable_intra_edge_filter, |
| 1637 |
IntraParam::None, |
| 1638 |
); |
| 1639 |
let mut alpha_cost = |alpha: i16| -> u64 { |
| 1640 |
let mut rec_region = |
| 1641 |
rec.subregion_mut(Area::BlockStartingAt { bo: tile_bo.0 }); |
| 1642 |
PredictionMode::UV_CFL_PRED.predict_intra( |
| 1643 |
tile_rect, |
| 1644 |
&mut rec_region, |
| 1645 |
uv_tx_size, |
| 1646 |
fi.sequence.bit_depth, |
| 1647 |
ac, |
| 1648 |
IntraParam::Alpha(alpha), |
| 1649 |
None, |
| 1650 |
&edge_buf, |
| 1651 |
fi.cpu_feature_level, |
| 1652 |
); |
| 1653 |
sse_wxh( |
| 1654 |
&input.subregion(Area::BlockStartingAt { bo: tile_bo.0 }), |
| 1655 |
&rec_region.as_const(), |
| 1656 |
visible_tx_w, |
| 1657 |
visible_tx_h, |
| 1658 |
|_, _| DistortionScale::default(), // We're not doing RDO here. |
| 1659 |
fi.sequence.bit_depth, |
| 1660 |
fi.cpu_feature_level, |
| 1661 |
) |
| 1662 |
.0 |
| 1663 |
}; |
| 1664 |
let mut best = (alpha_cost(0), 0); |
| 1665 |
let mut count = 2; |
| 1666 |
for alpha in 1i16..=16i16 { |
| 1667 |
let cost = (alpha_cost(alpha), alpha_cost(-alpha)); |
| 1668 |
if cost.0 < best.0 { |
| 1669 |
best = (cost.0, alpha); |
| 1670 |
count += 2; |
| 1671 |
} |
| 1672 |
if cost.1 < best.0 { |
| 1673 |
best = (cost.1, -alpha); |
| 1674 |
count += 2; |
| 1675 |
} |
| 1676 |
if count < alpha { |
| 1677 |
break; |
| 1678 |
} |
| 1679 |
} |
| 1680 |
best.1 |
| 1681 |
}) |
| 1682 |
.collect(); |
| 1683 |
|
| 1684 |
if best_alpha[0] == 0 && best_alpha[1] == 0 { |
| 1685 |
None |
| 1686 |
} else { |
| 1687 |
Some(CFLParams::from_alpha(best_alpha[0], best_alpha[1])) |
| 1688 |
} |
| 1689 |
} |
| 1690 |
|
| 1691 |
/// RDO-based transform type decision |
| 1692 |
/// If `cw_checkpoint` is `None`, a checkpoint for cw's (`ContextWriter`) current |
| 1693 |
/// state is created and stored for later use. |
| 1694 |
/// |
| 1695 |
/// # Panics |
| 1696 |
/// |
| 1697 |
/// - If a writer checkpoint is never created before or within the function. |
| 1698 |
/// This should never happen and indicates a development error. |
| 1699 |
/// - If the best RD found is negative. |
| 1700 |
/// This should never happen and indicates a development error. |
| 1701 |
pub fn rdo_tx_type_decision<T: Pixel>( |
| 1702 |
fi: &FrameInvariants<T>, ts: &mut TileStateMut<'_, T>, |
| 1703 |
cw: &mut ContextWriter, cw_checkpoint: &mut Option<ContextWriterCheckpoint>, |
| 1704 |
mode: PredictionMode, ref_frames: [RefType; 2], mvs: [MotionVector; 2], |
| 1705 |
bsize: BlockSize, tile_bo: TileBlockOffset, tx_size: TxSize, tx_set: TxSet, |
| 1706 |
tx_types: &[TxType], cur_best_rd: f64, |
| 1707 |
) -> (TxType, f64) { |
| 1708 |
let mut best_type = TxType::DCT_DCT; |
| 1709 |
let mut best_rd = f64::MAX; |
| 1710 |
|
| 1711 |
let PlaneConfig { xdec, ydec, .. } = ts.input.planes[1].cfg; |
| 1712 |
let is_chroma_block = |
| 1713 |
has_chroma(tile_bo, bsize, xdec, ydec, fi.sequence.chroma_sampling); |
| 1714 |
|
| 1715 |
let is_inter = !mode.is_intra(); |
| 1716 |
|
| 1717 |
if cw_checkpoint.is_none() { |
| 1718 |
// Only run the first call |
| 1719 |
// Prevents creating multiple checkpoints for own version of cw |
| 1720 |
*cw_checkpoint = |
| 1721 |
Some(cw.checkpoint(&tile_bo, fi.sequence.chroma_sampling)); |
| 1722 |
} |
| 1723 |
|
| 1724 |
let rdo_type = if fi.use_tx_domain_distortion { |
| 1725 |
RDOType::TxDistRealRate |
| 1726 |
} else { |
| 1727 |
RDOType::PixelDistRealRate |
| 1728 |
}; |
| 1729 |
let need_recon_pixel = tx_size.block_size() != bsize && !is_inter; |
| 1730 |
|
| 1731 |
let mut first_iteration = true; |
| 1732 |
for &tx_type in tx_types { |
| 1733 |
// Skip unsupported transform types |
| 1734 |
if av1_tx_used[tx_set as usize][tx_type as usize] == 0 { |
| 1735 |
continue; |
| 1736 |
} |
| 1737 |
|
| 1738 |
if is_inter { |
| 1739 |
motion_compensate( |
| 1740 |
fi, ts, cw, mode, ref_frames, mvs, bsize, tile_bo, true, |
| 1741 |
); |
| 1742 |
} |
| 1743 |
|
| 1744 |
let mut wr = WriterCounter::new(); |
| 1745 |
let tell = wr.tell_frac(); |
| 1746 |
let (_, tx_dist) = if is_inter { |
| 1747 |
write_tx_tree( |
| 1748 |
fi, |
| 1749 |
ts, |
| 1750 |
cw, |
| 1751 |
&mut wr, |
| 1752 |
mode, |
| 1753 |
0, |
| 1754 |
tile_bo, |
| 1755 |
bsize, |
| 1756 |
tx_size, |
| 1757 |
tx_type, |
| 1758 |
false, |
| 1759 |
true, |
| 1760 |
rdo_type, |
| 1761 |
need_recon_pixel, |
| 1762 |
) |
| 1763 |
} else { |
| 1764 |
write_tx_blocks( |
| 1765 |
fi, |
| 1766 |
ts, |
| 1767 |
cw, |
| 1768 |
&mut wr, |
| 1769 |
mode, |
| 1770 |
mode, |
| 1771 |
AngleDelta::default(), |
| 1772 |
tile_bo, |
| 1773 |
bsize, |
| 1774 |
tx_size, |
| 1775 |
tx_type, |
| 1776 |
false, |
| 1777 |
CFLParams::default(), // Unused. |
| 1778 |
true, |
| 1779 |
rdo_type, |
| 1780 |
need_recon_pixel, |
| 1781 |
) |
| 1782 |
}; |
| 1783 |
|
| 1784 |
let rate = wr.tell_frac() - tell; |
| 1785 |
let distortion = if fi.use_tx_domain_distortion { |
| 1786 |
compute_tx_distortion( |
| 1787 |
fi, |
| 1788 |
ts, |
| 1789 |
bsize, |
| 1790 |
is_chroma_block, |
| 1791 |
tile_bo, |
| 1792 |
tx_dist, |
| 1793 |
false, |
| 1794 |
true, |
| 1795 |
) |
| 1796 |
} else { |
| 1797 |
compute_distortion(fi, ts, bsize, is_chroma_block, tile_bo, true) |
| 1798 |
}; |
| 1799 |
cw.rollback(cw_checkpoint.as_ref().unwrap()); |
| 1800 |
|
| 1801 |
let rd = compute_rd_cost(fi, rate, distortion); |
| 1802 |
|
| 1803 |
if first_iteration { |
| 1804 |
// We use an optimization to early exit after testing the first |
| 1805 |
// transform type if the cost is higher than the existing best. |
| 1806 |
// The idea is that if this transform size is not better than he |
| 1807 |
// previous size, it is not worth testing remaining modes for this size. |
| 1808 |
if rd > cur_best_rd { |
| 1809 |
break; |
| 1810 |
} |
| 1811 |
first_iteration = false; |
| 1812 |
} |
| 1813 |
|
| 1814 |
if rd < best_rd { |
| 1815 |
best_rd = rd; |
| 1816 |
best_type = tx_type; |
| 1817 |
} |
| 1818 |
} |
| 1819 |
|
| 1820 |
assert!(best_rd >= 0_f64); |
| 1821 |
|
| 1822 |
(best_type, best_rd) |
| 1823 |
} |
| 1824 |
|
| 1825 |
pub fn get_sub_partitions( |
| 1826 |
four_partitions: &[TileBlockOffset; 4], partition: PartitionType, |
| 1827 |
) -> ArrayVec<TileBlockOffset, 4> { |
| 1828 |
let mut partition_offsets = ArrayVec::<TileBlockOffset, 4>::new(); |
| 1829 |
|
| 1830 |
partition_offsets.push(four_partitions[0]); |
| 1831 |
|
| 1832 |
if partition == PARTITION_NONE { |
| 1833 |
return partition_offsets; |
| 1834 |
} |
| 1835 |
if partition == PARTITION_VERT || partition == PARTITION_SPLIT { |
| 1836 |
partition_offsets.push(four_partitions[1]); |
| 1837 |
}; |
| 1838 |
if partition == PARTITION_HORZ || partition == PARTITION_SPLIT { |
| 1839 |
partition_offsets.push(four_partitions[2]); |
| 1840 |
}; |
| 1841 |
if partition == PARTITION_SPLIT { |
| 1842 |
partition_offsets.push(four_partitions[3]); |
| 1843 |
}; |
| 1844 |
|
| 1845 |
partition_offsets |
| 1846 |
} |
| 1847 |
|
| 1848 |
#[inline(always)] |
| 1849 |
fn rdo_partition_none<T: Pixel>( |
| 1850 |
fi: &FrameInvariants<T>, ts: &mut TileStateMut<'_, T>, |
| 1851 |
cw: &mut ContextWriter, bsize: BlockSize, tile_bo: TileBlockOffset, |
| 1852 |
inter_cfg: &InterConfig, child_modes: &mut ArrayVec<PartitionParameters, 4>, |
| 1853 |
) -> f64 { |
| 1854 |
debug_assert!(tile_bo.0.x < ts.mi_width && tile_bo.0.y < ts.mi_height); |
| 1855 |
|
| 1856 |
let mode = rdo_mode_decision(fi, ts, cw, bsize, tile_bo, inter_cfg); |
| 1857 |
let cost = mode.rd_cost; |
| 1858 |
|
| 1859 |
child_modes.push(mode); |
| 1860 |
|
| 1861 |
cost |
| 1862 |
} |
| 1863 |
|
| 1864 |
// VERTICAL, HORIZONTAL or simple SPLIT |
| 1865 |
#[inline(always)] |
| 1866 |
fn rdo_partition_simple<T: Pixel, W: Writer>( |
| 1867 |
fi: &FrameInvariants<T>, ts: &mut TileStateMut<'_, T>, |
| 1868 |
cw: &mut ContextWriter, w_pre_cdef: &mut W, w_post_cdef: &mut W, |
| 1869 |
bsize: BlockSize, tile_bo: TileBlockOffset, inter_cfg: &InterConfig, |
| 1870 |
partition: PartitionType, rdo_type: RDOType, best_rd: f64, |
| 1871 |
child_modes: &mut ArrayVec<PartitionParameters, 4>, |
| 1872 |
) -> Option<f64> { |
| 1873 |
debug_assert!(tile_bo.0.x < ts.mi_width && tile_bo.0.y < ts.mi_height); |
| 1874 |
let subsize = bsize.subsize(partition).unwrap(); |
| 1875 |
|
| 1876 |
let cost = if bsize >= BlockSize::BLOCK_8X8 { |
| 1877 |
let w: &mut W = if cw.bc.cdef_coded { w_post_cdef } else { w_pre_cdef }; |
| 1878 |
let tell = w.tell_frac(); |
| 1879 |
cw.write_partition(w, tile_bo, partition, bsize); |
| 1880 |
compute_rd_cost(fi, w.tell_frac() - tell, ScaledDistortion::zero()) |
| 1881 |
} else { |
| 1882 |
0.0 |
| 1883 |
}; |
| 1884 |
|
| 1885 |
let hbsw = subsize.width_mi(); // Half the block size width in blocks |
| 1886 |
let hbsh = subsize.height_mi(); // Half the block size height in blocks |
| 1887 |
let four_partitions = [ |
| 1888 |
tile_bo, |
| 1889 |
TileBlockOffset(BlockOffset { x: tile_bo.0.x + hbsw, y: tile_bo.0.y }), |
| 1890 |
TileBlockOffset(BlockOffset { x: tile_bo.0.x, y: tile_bo.0.y + hbsh }), |
| 1891 |
TileBlockOffset(BlockOffset { |
| 1892 |
x: tile_bo.0.x + hbsw, |
| 1893 |
y: tile_bo.0.y + hbsh, |
| 1894 |
}), |
| 1895 |
]; |
| 1896 |
|
| 1897 |
let partitions = get_sub_partitions(&four_partitions, partition); |
| 1898 |
|
| 1899 |
let mut rd_cost_sum = 0.0; |
| 1900 |
|
| 1901 |
for offset in partitions { |
| 1902 |
let hbs = subsize.width_mi() >> 1; |
| 1903 |
let has_cols = offset.0.x + hbs < ts.mi_width; |
| 1904 |
let has_rows = offset.0.y + hbs < ts.mi_height; |
| 1905 |
|
| 1906 |
if has_cols && has_rows { |
| 1907 |
let mode_decision = |
| 1908 |
rdo_mode_decision(fi, ts, cw, subsize, offset, inter_cfg); |
| 1909 |
|
| 1910 |
rd_cost_sum += mode_decision.rd_cost; |
| 1911 |
|
| 1912 |
if fi.enable_early_exit && rd_cost_sum > best_rd { |
| 1913 |
return None; |
| 1914 |
} |
| 1915 |
if subsize >= BlockSize::BLOCK_8X8 && subsize.is_sqr() { |
| 1916 |
let w: &mut W = |
| 1917 |
if cw.bc.cdef_coded { w_post_cdef } else { w_pre_cdef }; |
| 1918 |
cw.write_partition(w, offset, PartitionType::PARTITION_NONE, subsize); |
| 1919 |
} |
| 1920 |
encode_block_with_modes( |
| 1921 |
fi, |
| 1922 |
ts, |
| 1923 |
cw, |
| 1924 |
w_pre_cdef, |
| 1925 |
w_post_cdef, |
| 1926 |
subsize, |
| 1927 |
offset, |
| 1928 |
&mode_decision, |
| 1929 |
rdo_type, |
| 1930 |
None, |
| 1931 |
); |
| 1932 |
child_modes.push(mode_decision); |
| 1933 |
} else { |
| 1934 |
//rd_cost_sum += f64::MAX; |
| 1935 |
return None; |
| 1936 |
} |
| 1937 |
} |
| 1938 |
|
| 1939 |
Some(cost + rd_cost_sum) |
| 1940 |
} |
| 1941 |
|
| 1942 |
/// RDO-based single level partitioning decision |
| 1943 |
/// |
| 1944 |
/// # Panics |
| 1945 |
/// |
| 1946 |
/// - If the best RD found is negative. |
| 1947 |
/// This should never happen, and indicates a development error. |
| 1948 |
#[profiling::function] |
| 1949 |
pub fn rdo_partition_decision<T: Pixel, W: Writer>( |
| 1950 |
fi: &FrameInvariants<T>, ts: &mut TileStateMut<'_, T>, |
| 1951 |
cw: &mut ContextWriter, w_pre_cdef: &mut W, w_post_cdef: &mut W, |
| 1952 |
bsize: BlockSize, tile_bo: TileBlockOffset, |
| 1953 |
cached_block: &PartitionGroupParameters, partition_types: &[PartitionType], |
| 1954 |
rdo_type: RDOType, inter_cfg: &InterConfig, |
| 1955 |
) -> PartitionGroupParameters { |
| 1956 |
let mut best_partition = cached_block.part_type; |
| 1957 |
let mut best_rd = cached_block.rd_cost; |
| 1958 |
let mut best_pred_modes = cached_block.part_modes.clone(); |
| 1959 |
|
| 1960 |
let cw_checkpoint = cw.checkpoint(&tile_bo, fi.sequence.chroma_sampling); |
| 1961 |
let w_pre_checkpoint = w_pre_cdef.checkpoint(); |
| 1962 |
let w_post_checkpoint = w_post_cdef.checkpoint(); |
| 1963 |
|
| 1964 |
for &partition in partition_types { |
| 1965 |
// Do not re-encode results we already have |
| 1966 |
if partition == cached_block.part_type { |
| 1967 |
continue; |
| 1968 |
} |
| 1969 |
|
| 1970 |
let mut child_modes = ArrayVec::<_, 4>::new(); |
| 1971 |
|
| 1972 |
let cost = match partition { |
| 1973 |
PARTITION_NONE if bsize <= BlockSize::BLOCK_64X64 => { |
| 1974 |
Some(rdo_partition_none( |
| 1975 |
fi, |
| 1976 |
ts, |
| 1977 |
cw, |
| 1978 |
bsize, |
| 1979 |
tile_bo, |
| 1980 |
inter_cfg, |
| 1981 |
&mut child_modes, |
| 1982 |
)) |
| 1983 |
} |
| 1984 |
PARTITION_SPLIT | PARTITION_HORZ | PARTITION_VERT => { |
| 1985 |
rdo_partition_simple( |
| 1986 |
fi, |
| 1987 |
ts, |
| 1988 |
cw, |
| 1989 |
w_pre_cdef, |
| 1990 |
w_post_cdef, |
| 1991 |
bsize, |
| 1992 |
tile_bo, |
| 1993 |
inter_cfg, |
| 1994 |
partition, |
| 1995 |
rdo_type, |
| 1996 |
best_rd, |
| 1997 |
&mut child_modes, |
| 1998 |
) |
| 1999 |
} |
| 2000 |
_ => { |
| 2001 |
unreachable!(); |
| 2002 |
} |
| 2003 |
}; |
| 2004 |
|
| 2005 |
if let Some(rd) = cost { |
| 2006 |
if rd < best_rd { |
| 2007 |
best_rd = rd; |
| 2008 |
best_partition = partition; |
| 2009 |
best_pred_modes.clone_from(&child_modes); |
| 2010 |
} |
| 2011 |
} |
| 2012 |
cw.rollback(&cw_checkpoint); |
| 2013 |
w_pre_cdef.rollback(&w_pre_checkpoint); |
| 2014 |
w_post_cdef.rollback(&w_post_checkpoint); |
| 2015 |
} |
| 2016 |
|
| 2017 |
assert!(best_rd >= 0_f64); |
| 2018 |
|
| 2019 |
PartitionGroupParameters { |
| 2020 |
rd_cost: best_rd, |
| 2021 |
part_type: best_partition, |
| 2022 |
part_modes: best_pred_modes, |
| 2023 |
} |
| 2024 |
} |
| 2025 |
|
| 2026 |
#[profiling::function] |
| 2027 |
fn rdo_loop_plane_error<T: Pixel>( |
| 2028 |
base_sbo: TileSuperBlockOffset, offset_sbo: TileSuperBlockOffset, |
| 2029 |
sb_w: usize, sb_h: usize, fi: &FrameInvariants<T>, ts: &TileStateMut<'_, T>, |
| 2030 |
blocks: &TileBlocks<'_>, test: &Frame<T>, src: &Tile<'_, T>, pli: usize, |
| 2031 |
) -> ScaledDistortion { |
| 2032 |
let sb_w_blocks = |
| 2033 |
if fi.sequence.use_128x128_superblock { 16 } else { 8 } * sb_w; |
| 2034 |
let sb_h_blocks = |
| 2035 |
if fi.sequence.use_128x128_superblock { 16 } else { 8 } * sb_h; |
| 2036 |
// Each direction block is 8x8 in y, potentially smaller if subsampled in chroma |
| 2037 |
// accumulating in-frame and unpadded |
| 2038 |
let mut err = Distortion::zero(); |
| 2039 |
for by in 0..sb_h_blocks { |
| 2040 |
for bx in 0..sb_w_blocks { |
| 2041 |
let loop_bo = offset_sbo.block_offset(bx << 1, by << 1); |
| 2042 |
if loop_bo.0.x < blocks.cols() && loop_bo.0.y < blocks.rows() { |
| 2043 |
let src_plane = &src.planes[pli]; |
| 2044 |
let test_plane = &test.planes[pli]; |
| 2045 |
let PlaneConfig { xdec, ydec, .. } = *src_plane.plane_cfg; |
| 2046 |
debug_assert_eq!(xdec, test_plane.cfg.xdec); |
| 2047 |
debug_assert_eq!(ydec, test_plane.cfg.ydec); |
| 2048 |
|
| 2049 |
// Unfortunately, our distortion biases are only available via |
| 2050 |
// Frame-absolute addressing, so we need a block offset |
| 2051 |
// relative to the full frame origin (not the tile or analysis |
| 2052 |
// area) |
| 2053 |
let frame_bo = (base_sbo + offset_sbo).block_offset(bx << 1, by << 1); |
| 2054 |
let bias = distortion_scale( |
| 2055 |
fi, |
| 2056 |
ts.to_frame_block_offset(frame_bo), |
| 2057 |
BlockSize::BLOCK_8X8, |
| 2058 |
); |
| 2059 |
|
| 2060 |
let src_region = |
| 2061 |
src_plane.subregion(Area::BlockStartingAt { bo: loop_bo.0 }); |
| 2062 |
let test_region = |
| 2063 |
test_plane.region(Area::BlockStartingAt { bo: loop_bo.0 }); |
| 2064 |
|
| 2065 |
err += if pli == 0 { |
| 2066 |
// For loop filters, We intentionally use cdef_dist even with |
| 2067 |
// `--tune Psnr`. Using SSE instead gives no PSNR gain but has a |
| 2068 |
// significant negative impact on other metrics and visual quality. |
| 2069 |
RawDistortion(cdef_dist_kernel( |
| 2070 |
&src_region, |
| 2071 |
&test_region, |
| 2072 |
8, |
| 2073 |
8, |
| 2074 |
fi.sequence.bit_depth, |
| 2075 |
fi.cpu_feature_level, |
| 2076 |
) as u64) |
| 2077 |
* bias |
| 2078 |
} else { |
| 2079 |
sse_wxh( |
| 2080 |
&src_region, |
| 2081 |
&test_region, |
| 2082 |
8 >> xdec, |
| 2083 |
8 >> ydec, |
| 2084 |
|_, _| bias, |
| 2085 |
fi.sequence.bit_depth, |
| 2086 |
fi.cpu_feature_level, |
| 2087 |
) |
| 2088 |
}; |
| 2089 |
} |
| 2090 |
} |
| 2091 |
} |
| 2092 |
err * fi.dist_scale[pli] |
| 2093 |
} |
| 2094 |
|
| 2095 |
/// Passed in a superblock offset representing the upper left corner of |
| 2096 |
/// the LRU area we're optimizing. This area covers the largest LRU in |
| 2097 |
/// any of the present planes, but may consist of a number of |
| 2098 |
/// superblocks and full, smaller LRUs in the other planes |
| 2099 |
/// |
| 2100 |
/// # Panics |
| 2101 |
/// |
| 2102 |
/// - If both CDEF and LRF are disabled. |
| 2103 |
#[profiling::function] |
| 2104 |
pub fn rdo_loop_decision<T: Pixel, W: Writer>( |
| 2105 |
base_sbo: TileSuperBlockOffset, fi: &FrameInvariants<T>, |
| 2106 |
ts: &mut TileStateMut<'_, T>, cw: &mut ContextWriter, w: &mut W, |
| 2107 |
deblock_p: bool, |
| 2108 |
) { |
| 2109 |
let planes = if fi.sequence.chroma_sampling == ChromaSampling::Cs400 { |
| 2110 |
1 |
| 2111 |
} else { |
| 2112 |
MAX_PLANES |
| 2113 |
}; |
| 2114 |
assert!(fi.sequence.enable_cdef || fi.sequence.enable_restoration); |
| 2115 |
// Determine area of optimization: Which plane has the largest LRUs? |
| 2116 |
// How many LRUs for each? |
| 2117 |
let mut sb_w = 1; // how many superblocks wide the largest LRU |
| 2118 |
// is/how many SBs we're processing (same thing) |
| 2119 |
let mut sb_h = 1; // how many superblocks wide the largest LRU |
| 2120 |
// is/how many SBs we're processing (same thing) |
| 2121 |
let mut lru_w = [0; MAX_PLANES]; // how many LRUs we're processing |
| 2122 |
let mut lru_h = [0; MAX_PLANES]; // how many LRUs we're processing |
| 2123 |
for pli in 0..planes { |
| 2124 |
let sb_h_shift = ts.restoration.planes[pli].rp_cfg.sb_h_shift; |
| 2125 |
let sb_v_shift = ts.restoration.planes[pli].rp_cfg.sb_v_shift; |
| 2126 |
if sb_w < (1 << sb_h_shift) { |
| 2127 |
sb_w = 1 << sb_h_shift; |
| 2128 |
} |
| 2129 |
if sb_h < (1 << sb_v_shift) { |
| 2130 |
sb_h = 1 << sb_v_shift; |
| 2131 |
} |
| 2132 |
} |
| 2133 |
for pli in 0..planes { |
| 2134 |
let sb_h_shift = ts.restoration.planes[pli].rp_cfg.sb_h_shift; |
| 2135 |
let sb_v_shift = ts.restoration.planes[pli].rp_cfg.sb_v_shift; |
| 2136 |
lru_w[pli] = sb_w / (1 << sb_h_shift); |
| 2137 |
lru_h[pli] = sb_h / (1 << sb_v_shift); |
| 2138 |
} |
| 2139 |
|
| 2140 |
// The superblock width/height determinations may be calling for us |
| 2141 |
// to compute over superblocks that do not actually exist in the |
| 2142 |
// frame (off the right or lower edge). Trim sb width/height down |
| 2143 |
// to actual superblocks. Note that these last superblocks on the |
| 2144 |
// right/bottom may themselves still span the edge of the frame, but |
| 2145 |
// they do hold at least some visible pixels. |
| 2146 |
sb_w = sb_w.min(ts.sb_width - base_sbo.0.x); |
| 2147 |
sb_h = sb_h.min(ts.sb_height - base_sbo.0.y); |
| 2148 |
|
| 2149 |
// We have need to know the Y visible pixel limits as well (the |
| 2150 |
// sb_w/sb_h figures above can be used to determine how many |
| 2151 |
// allocated pixels, possibly beyond the visible frame, exist). |
| 2152 |
let crop_w = |
| 2153 |
fi.width - ((ts.sbo.0.x + base_sbo.0.x) << SUPERBLOCK_TO_PLANE_SHIFT); |
| 2154 |
let crop_h = |
| 2155 |
fi.height - ((ts.sbo.0.y + base_sbo.0.y) << SUPERBLOCK_TO_PLANE_SHIFT); |
| 2156 |
let pixel_w = crop_w.min(sb_w << SUPERBLOCK_TO_PLANE_SHIFT); |
| 2157 |
let pixel_h = crop_h.min(sb_h << SUPERBLOCK_TO_PLANE_SHIFT); |
| 2158 |
|
| 2159 |
// Based on `RestorationState::new` |
| 2160 |
const MAX_SB_SHIFT: usize = 4; |
| 2161 |
const MAX_SB_SIZE: usize = 1 << MAX_SB_SHIFT; |
| 2162 |
const MAX_LRU_SIZE: usize = MAX_SB_SIZE; |
| 2163 |
|
| 2164 |
// Static allocation relies on the "minimal LRU area for all N planes" invariant. |
| 2165 |
let mut best_index = [-1; MAX_SB_SIZE * MAX_SB_SIZE]; |
| 2166 |
let mut best_lrf = |
| 2167 |
[[RestorationFilter::None; MAX_PLANES]; MAX_LRU_SIZE * MAX_LRU_SIZE]; |
| 2168 |
|
| 2169 |
// due to imprecision in the reconstruction parameter solver, we |
| 2170 |
// need to make sure we don't fall into a limit cycle. Track our |
| 2171 |
// best cost at LRF so that we can break if we get a solution that doesn't |
| 2172 |
// improve at the reconstruction stage. |
| 2173 |
let mut best_lrf_cost = [[-1.0; MAX_PLANES]; MAX_LRU_SIZE * MAX_LRU_SIZE]; |
| 2174 |
|
| 2175 |
// sub-setted region of the TileBlocks for our working frame area. |
| 2176 |
// Note that the size of this subset is what signals CDEF as to the |
| 2177 |
// actual coded size. |
| 2178 |
let mut tileblocks_subset = cw.bc.blocks.subregion_mut( |
| 2179 |
base_sbo.block_offset(0, 0).0.x, |
| 2180 |
base_sbo.block_offset(0, 0).0.y, |
| 2181 |
sb_w << SUPERBLOCK_TO_BLOCK_SHIFT, |
| 2182 |
sb_h << SUPERBLOCK_TO_BLOCK_SHIFT, |
| 2183 |
); |
| 2184 |
|
| 2185 |
// cdef doesn't run on superblocks that are completely skipped. |
| 2186 |
// Determine which super blocks are marked as skipped so we can avoid running |
| 2187 |
// them. If all blocks are skipped, we can avoid some of the overhead related |
| 2188 |
// to setting up for cdef. |
| 2189 |
let mut cdef_skip = [true; MAX_SB_SIZE * MAX_SB_SIZE]; |
| 2190 |
let mut cdef_skip_all = true; |
| 2191 |
if fi.sequence.enable_cdef { |
| 2192 |
for sby in 0..sb_h { |
| 2193 |
for sbx in 0..sb_w { |
| 2194 |
let blocks = tileblocks_subset.subregion(16 * sbx, 16 * sby, 16, 16); |
| 2195 |
let mut skip = true; |
| 2196 |
for y in 0..blocks.rows() { |
| 2197 |
for block in blocks[y].iter() { |
| 2198 |
skip &= block.skip; |
| 2199 |
} |
| 2200 |
} |
| 2201 |
cdef_skip[sby * MAX_SB_SIZE + sbx] = skip; |
| 2202 |
cdef_skip_all &= skip; |
| 2203 |
} |
| 2204 |
} |
| 2205 |
} |
| 2206 |
|
| 2207 |
// Unlike cdef, loop restoration will run regardless of whether blocks are |
| 2208 |
// skipped or not. At the same time, the most significant improvement will |
| 2209 |
// generally be from un-skipped blocks, so lru is only performed if there are |
| 2210 |
// un-skipped blocks. |
| 2211 |
// This should be the same as `cdef_skip_all`, except when cdef is disabled. |
| 2212 |
let mut lru_skip_all = true; |
| 2213 |
let mut lru_skip = [[true; MAX_PLANES]; MAX_LRU_SIZE * MAX_LRU_SIZE]; |
| 2214 |
if fi.sequence.enable_restoration { |
| 2215 |
if fi.config.speed_settings.lru_on_skip { |
| 2216 |
lru_skip_all = false; |
| 2217 |
lru_skip = [[false; MAX_PLANES]; MAX_LRU_SIZE * MAX_LRU_SIZE]; |
| 2218 |
} else { |
| 2219 |
for pli in 0..planes { |
| 2220 |
// width, in sb, of an LRU in this plane |
| 2221 |
let lru_sb_w = 1 << ts.restoration.planes[pli].rp_cfg.sb_h_shift; |
| 2222 |
// height, in sb, of an LRU in this plane |
| 2223 |
let lru_sb_h = 1 << ts.restoration.planes[pli].rp_cfg.sb_v_shift; |
| 2224 |
for lru_y in 0..lru_h[pli] { |
| 2225 |
// number of LRUs vertically |
| 2226 |
for lru_x in 0..lru_w[pli] { |
| 2227 |
// number of LRUs horizontally |
| 2228 |
|
| 2229 |
let loop_sbo = TileSuperBlockOffset(SuperBlockOffset { |
| 2230 |
x: lru_x * lru_sb_w, |
| 2231 |
y: lru_y * lru_sb_h, |
| 2232 |
}); |
| 2233 |
|
| 2234 |
if !ts.restoration.has_restoration_unit( |
| 2235 |
base_sbo + loop_sbo, |
| 2236 |
pli, |
| 2237 |
false, |
| 2238 |
) { |
| 2239 |
continue; |
| 2240 |
} |
| 2241 |
|
| 2242 |
let start = loop_sbo.block_offset(0, 0).0; |
| 2243 |
let size = TileSuperBlockOffset(SuperBlockOffset { |
| 2244 |
x: lru_sb_w, |
| 2245 |
y: lru_sb_h, |
| 2246 |
}) |
| 2247 |
.block_offset(0, 0) |
| 2248 |
.0; |
| 2249 |
|
| 2250 |
let blocks = |
| 2251 |
tileblocks_subset.subregion(start.x, start.y, size.x, size.y); |
| 2252 |
let mut skip = true; |
| 2253 |
for y in 0..blocks.rows() { |
| 2254 |
for block in blocks[y].iter() { |
| 2255 |
skip &= block.skip; |
| 2256 |
} |
| 2257 |
} |
| 2258 |
lru_skip[lru_y * MAX_LRU_SIZE + lru_x][pli] = skip; |
| 2259 |
lru_skip_all &= skip; |
| 2260 |
} |
| 2261 |
} |
| 2262 |
} |
| 2263 |
} |
| 2264 |
} |
| 2265 |
|
| 2266 |
// Return early if all blocks are skipped for lru and cdef. |
| 2267 |
if lru_skip_all && cdef_skip_all { |
| 2268 |
return; |
| 2269 |
} |
| 2270 |
|
| 2271 |
// Loop filter RDO is an iterative process and we need temporary |
| 2272 |
// scratch data to hold the results of deblocking, cdef, and the |
| 2273 |
// loop reconstruction filter so that each can be partially updated |
| 2274 |
// without recomputing the entire stack. Construct |
| 2275 |
// largest-LRU-sized frames for each, accounting for padding |
| 2276 |
// required by deblocking, cdef and [optionally] LR. |
| 2277 |
let mut rec_subset = ts |
| 2278 |
.rec |
| 2279 |
.subregion(Area::BlockRect { |
| 2280 |
bo: base_sbo.block_offset(0, 0).0, |
| 2281 |
width: (pixel_w + 7) >> 3 << 3, |
| 2282 |
height: (pixel_h + 7) >> 3 << 3, |
| 2283 |
}) |
| 2284 |
.scratch_copy(); |
| 2285 |
|
| 2286 |
// const, no need to copy, just need the subregion (but do zero the |
| 2287 |
// origin to match the other copies/new backing frames). |
| 2288 |
let src_subset = ts |
| 2289 |
.input_tile |
| 2290 |
.subregion(Area::BlockRect { |
| 2291 |
bo: base_sbo.block_offset(0, 0).0, |
| 2292 |
width: (pixel_w + 7) >> 3 << 3, |
| 2293 |
height: (pixel_h + 7) >> 3 << 3, |
| 2294 |
}) |
| 2295 |
.home(); |
| 2296 |
|
| 2297 |
if deblock_p { |
| 2298 |
// Find a good deblocking filter solution for the passed in area. |
| 2299 |
// This is not RDO of deblocking itself, merely a solution to get |
| 2300 |
// better results from CDEF/LRF RDO. |
| 2301 |
let deblock_levels = deblock_filter_optimize( |
| 2302 |
fi, |
| 2303 |
&rec_subset.as_tile(), |
| 2304 |
&src_subset, |
| 2305 |
&tileblocks_subset.as_const(), |
| 2306 |
crop_w, |
| 2307 |
crop_h, |
| 2308 |
); |
| 2309 |
|
| 2310 |
// Deblock the contents of our reconstruction copy. |
| 2311 |
if deblock_levels[0] != 0 || deblock_levels[1] != 0 { |
| 2312 |
// copy ts.deblock because we need to set some of our own values here |
| 2313 |
let mut deblock_copy = *ts.deblock; |
| 2314 |
deblock_copy.levels = deblock_levels; |
| 2315 |
|
| 2316 |
// finally, deblock the temp frame |
| 2317 |
deblock_filter_frame( |
| 2318 |
&deblock_copy, |
| 2319 |
&mut rec_subset.as_tile_mut(), |
| 2320 |
&tileblocks_subset.as_const(), |
| 2321 |
crop_w, |
| 2322 |
crop_h, |
| 2323 |
fi.sequence.bit_depth, |
| 2324 |
planes, |
| 2325 |
); |
| 2326 |
} |
| 2327 |
} |
| 2328 |
|
| 2329 |
let mut cdef_work = |
| 2330 |
if !cdef_skip_all { Some(rec_subset.clone()) } else { None }; |
| 2331 |
let mut lrf_work = if !lru_skip_all { |
| 2332 |
Some(Frame { |
| 2333 |
planes: { |
| 2334 |
let new_plane = |pli: usize| { |
| 2335 |
let PlaneConfig { xdec, ydec, width, height, .. } = |
| 2336 |
rec_subset.planes[pli].cfg; |
| 2337 |
Plane::new(width, height, xdec, ydec, 0, 0) |
| 2338 |
}; |
| 2339 |
[new_plane(0), new_plane(1), new_plane(2)] |
| 2340 |
}, |
| 2341 |
}) |
| 2342 |
} else { |
| 2343 |
None |
| 2344 |
}; |
| 2345 |
|
| 2346 |
// Precompute directional analysis for CDEF |
| 2347 |
let cdef_data = { |
| 2348 |
if cdef_work.is_some() { |
| 2349 |
Some(( |
| 2350 |
&rec_subset, |
| 2351 |
cdef_analyze_superblock_range( |
| 2352 |
fi, |
| 2353 |
&rec_subset, |
| 2354 |
&tileblocks_subset.as_const(), |
| 2355 |
sb_w, |
| 2356 |
sb_h, |
| 2357 |
), |
| 2358 |
)) |
| 2359 |
} else { |
| 2360 |
None |
| 2361 |
} |
| 2362 |
}; |
| 2363 |
|
| 2364 |
// CDEF/LRF decision iteration |
| 2365 |
// Start with a default of CDEF 0 and RestorationFilter::None |
| 2366 |
// Try all CDEF options for each sb with current LRF; if new CDEF+LRF choice is better, select it. |
| 2367 |
// Then try all LRF options with current CDEFs; if new CDEFs+LRF choice is better, select it. |
| 2368 |
// If LRF choice changed for any plane, repeat until no changes |
| 2369 |
// Limit iterations and where we break based on speed setting (in the TODO list ;-) |
| 2370 |
let mut cdef_change = true; |
| 2371 |
let mut lrf_change = true; |
| 2372 |
while cdef_change || lrf_change { |
| 2373 |
// search for improved cdef indices, superblock by superblock, if cdef is enabled. |
| 2374 |
if let (Some((rec_copy, cdef_dirs)), Some(cdef_ref)) = |
| 2375 |
(&cdef_data, &mut cdef_work.as_mut()) |
| 2376 |
{ |
| 2377 |
for sby in 0..sb_h { |
| 2378 |
for sbx in 0..sb_w { |
| 2379 |
// determine whether this superblock can be skipped |
| 2380 |
if cdef_skip[sby * MAX_SB_SIZE + sbx] { |
| 2381 |
continue; |
| 2382 |
} |
| 2383 |
|
| 2384 |
let prev_best_index = best_index[sby * sb_w + sbx]; |
| 2385 |
let mut best_cost = -1.; |
| 2386 |
let mut best_new_index = -1i8; |
| 2387 |
|
| 2388 |
/* offset of the superblock we're currently testing within the larger |
| 2389 |
analysis area */ |
| 2390 |
let loop_sbo = |
| 2391 |
TileSuperBlockOffset(SuperBlockOffset { x: sbx, y: sby }); |
| 2392 |
|
| 2393 |
/* cdef index testing loop */ |
| 2394 |
for cdef_index in 0..(1 << fi.cdef_bits) { |
| 2395 |
let mut err = ScaledDistortion::zero(); |
| 2396 |
let mut rate = 0; |
| 2397 |
|
| 2398 |
cdef_filter_superblock( |
| 2399 |
fi, |
| 2400 |
&rec_subset, |
| 2401 |
&mut cdef_ref.as_tile_mut(), |
| 2402 |
&tileblocks_subset.as_const(), |
| 2403 |
loop_sbo, |
| 2404 |
cdef_index, |
| 2405 |
&cdef_dirs[sby * sb_w + sbx], |
| 2406 |
); |
| 2407 |
// apply LRF if any |
| 2408 |
for pli in 0..planes { |
| 2409 |
// We need the cropped-to-visible-frame area of this SB |
| 2410 |
let wh = |
| 2411 |
if fi.sequence.use_128x128_superblock { 128 } else { 64 }; |
| 2412 |
let PlaneConfig { xdec, ydec, .. } = cdef_ref.planes[pli].cfg; |
| 2413 |
let vis_width = (wh >> xdec).min( |
| 2414 |
(crop_w >> xdec) |
| 2415 |
- loop_sbo.plane_offset(&cdef_ref.planes[pli].cfg).x |
| 2416 |
as usize, |
| 2417 |
); |
| 2418 |
let vis_height = (wh >> ydec).min( |
| 2419 |
(crop_h >> ydec) |
| 2420 |
- loop_sbo.plane_offset(&cdef_ref.planes[pli].cfg).y |
| 2421 |
as usize, |
| 2422 |
); |
| 2423 |
// which LRU are we currently testing against? |
| 2424 |
if let (Some((lru_x, lru_y)), Some(lrf_ref)) = { |
| 2425 |
let rp = &ts.restoration.planes[pli]; |
| 2426 |
( |
| 2427 |
rp.restoration_unit_offset(base_sbo, loop_sbo, false), |
| 2428 |
&mut lrf_work, |
| 2429 |
) |
| 2430 |
} { |
| 2431 |
// We have a valid LRU, apply LRF, compute error |
| 2432 |
match best_lrf[lru_y * lru_w[pli] + lru_x][pli] { |
| 2433 |
RestorationFilter::None => { |
| 2434 |
err += rdo_loop_plane_error( |
| 2435 |
base_sbo, |
| 2436 |
loop_sbo, |
| 2437 |
1, |
| 2438 |
1, |
| 2439 |
fi, |
| 2440 |
ts, |
| 2441 |
&tileblocks_subset.as_const(), |
| 2442 |
cdef_ref, |
| 2443 |
&src_subset, |
| 2444 |
pli, |
| 2445 |
); |
| 2446 |
rate += if fi.sequence.enable_restoration { |
| 2447 |
cw.fc.count_lrf_switchable( |
| 2448 |
w, |
| 2449 |
&ts.restoration.as_const(), |
| 2450 |
best_lrf[lru_y * lru_w[pli] + lru_x][pli], |
| 2451 |
pli, |
| 2452 |
) |
| 2453 |
} else { |
| 2454 |
0 // no relative cost differeneces to different |
| 2455 |
// CDEF params. If cdef is on, it's a wash. |
| 2456 |
}; |
| 2457 |
} |
| 2458 |
RestorationFilter::Sgrproj { set, xqd } => { |
| 2459 |
// only run on this single superblock |
| 2460 |
let loop_po = |
| 2461 |
loop_sbo.plane_offset(&cdef_ref.planes[pli].cfg); |
| 2462 |
// todo: experiment with borrowing border pixels |
| 2463 |
// rather than edge-extending. Right now this is |
| 2464 |
// hard-clipping to the superblock boundary. |
| 2465 |
setup_integral_image( |
| 2466 |
&mut ts.integral_buffer, |
| 2467 |
SOLVE_IMAGE_STRIDE, |
| 2468 |
vis_width, |
| 2469 |
vis_height, |
| 2470 |
vis_width, |
| 2471 |
vis_height, |
| 2472 |
&cdef_ref.planes[pli].slice(loop_po), |
| 2473 |
&cdef_ref.planes[pli].slice(loop_po), |
| 2474 |
); |
| 2475 |
sgrproj_stripe_filter( |
| 2476 |
set, |
| 2477 |
xqd, |
| 2478 |
fi, |
| 2479 |
&ts.integral_buffer, |
| 2480 |
SOLVE_IMAGE_STRIDE, |
| 2481 |
&cdef_ref.planes[pli].slice(loop_po), |
| 2482 |
&mut lrf_ref.planes[pli].region_mut(Area::Rect { |
| 2483 |
x: loop_po.x, |
| 2484 |
y: loop_po.y, |
| 2485 |
width: vis_width, |
| 2486 |
height: vis_height, |
| 2487 |
}), |
| 2488 |
); |
| 2489 |
err += rdo_loop_plane_error( |
| 2490 |
base_sbo, |
| 2491 |
loop_sbo, |
| 2492 |
1, |
| 2493 |
1, |
| 2494 |
fi, |
| 2495 |
ts, |
| 2496 |
&tileblocks_subset.as_const(), |
| 2497 |
lrf_ref, |
| 2498 |
&src_subset, |
| 2499 |
pli, |
| 2500 |
); |
| 2501 |
rate += cw.fc.count_lrf_switchable( |
| 2502 |
w, |
| 2503 |
&ts.restoration.as_const(), |
| 2504 |
best_lrf[lru_y * lru_w[pli] + lru_x][pli], |
| 2505 |
pli, |
| 2506 |
); |
| 2507 |
} |
| 2508 |
RestorationFilter::Wiener { .. } => unreachable!(), // coming soon |
| 2509 |
} |
| 2510 |
} else { |
| 2511 |
// No actual LRU here, compute error directly from CDEF output. |
| 2512 |
err += rdo_loop_plane_error( |
| 2513 |
base_sbo, |
| 2514 |
loop_sbo, |
| 2515 |
1, |
| 2516 |
1, |
| 2517 |
fi, |
| 2518 |
ts, |
| 2519 |
&tileblocks_subset.as_const(), |
| 2520 |
cdef_ref, |
| 2521 |
&src_subset, |
| 2522 |
pli, |
| 2523 |
); |
| 2524 |
// no relative cost differeneces to different |
| 2525 |
// CDEF params. If cdef is on, it's a wash. |
| 2526 |
// rate += 0; |
| 2527 |
} |
| 2528 |
} |
| 2529 |
|
| 2530 |
let cost = compute_rd_cost(fi, rate, err); |
| 2531 |
if best_cost < 0. || cost < best_cost { |
| 2532 |
best_cost = cost; |
| 2533 |
best_new_index = cdef_index as i8; |
| 2534 |
} |
| 2535 |
} |
| 2536 |
|
| 2537 |
// Did we change any preexisting choices? |
| 2538 |
if best_new_index != prev_best_index { |
| 2539 |
cdef_change = true; |
| 2540 |
best_index[sby * sb_w + sbx] = best_new_index; |
| 2541 |
tileblocks_subset.set_cdef(loop_sbo, best_new_index as u8); |
| 2542 |
} |
| 2543 |
|
| 2544 |
let mut cdef_ref_tm = TileMut::new( |
| 2545 |
cdef_ref, |
| 2546 |
TileRect { |
| 2547 |
x: 0, |
| 2548 |
y: 0, |
| 2549 |
width: cdef_ref.planes[0].cfg.width, |
| 2550 |
height: cdef_ref.planes[0].cfg.height, |
| 2551 |
}, |
| 2552 |
); |
| 2553 |
|
| 2554 |
// Keep cdef output up to date; we need it for restoration |
| 2555 |
// both below and above (padding) |
| 2556 |
cdef_filter_superblock( |
| 2557 |
fi, |
| 2558 |
rec_copy, |
| 2559 |
&mut cdef_ref_tm, |
| 2560 |
&tileblocks_subset.as_const(), |
| 2561 |
loop_sbo, |
| 2562 |
best_index[sby * sb_w + sbx] as u8, |
| 2563 |
&cdef_dirs[sby * sb_w + sbx], |
| 2564 |
); |
| 2565 |
} |
| 2566 |
} |
| 2567 |
} |
| 2568 |
|
| 2569 |
if !cdef_change { |
| 2570 |
break; |
| 2571 |
} |
| 2572 |
cdef_change = false; |
| 2573 |
lrf_change = false; |
| 2574 |
|
| 2575 |
// search for improved restoration filter parameters if restoration is enabled |
| 2576 |
if let Some(lrf_ref) = &mut lrf_work.as_mut() { |
| 2577 |
let lrf_input = if cdef_work.is_some() { |
| 2578 |
// When CDEF is enabled, we pull from the CDEF output |
| 2579 |
cdef_work.as_ref().unwrap() |
| 2580 |
} else { |
| 2581 |
// When CDEF is disabled, we pull from the [optionally |
| 2582 |
// deblocked] reconstruction |
| 2583 |
&rec_subset |
| 2584 |
}; |
| 2585 |
for pli in 0..planes { |
| 2586 |
// Nominal size of LRU in pixels before clipping to visible frame |
| 2587 |
let unit_size = ts.restoration.planes[pli].rp_cfg.unit_size; |
| 2588 |
// width, in sb, of an LRU in this plane |
| 2589 |
let lru_sb_w = 1 << ts.restoration.planes[pli].rp_cfg.sb_h_shift; |
| 2590 |
// height, in sb, of an LRU in this plane |
| 2591 |
let lru_sb_h = 1 << ts.restoration.planes[pli].rp_cfg.sb_v_shift; |
| 2592 |
let PlaneConfig { xdec, ydec, .. } = lrf_ref.planes[pli].cfg; |
| 2593 |
for lru_y in 0..lru_h[pli] { |
| 2594 |
// number of LRUs vertically |
| 2595 |
for lru_x in 0..lru_w[pli] { |
| 2596 |
// number of LRUs horizontally |
| 2597 |
|
| 2598 |
// determine whether this lru should be skipped |
| 2599 |
if lru_skip[lru_y * MAX_LRU_SIZE + lru_x][pli] { |
| 2600 |
continue; |
| 2601 |
} |
| 2602 |
|
| 2603 |
let loop_sbo = TileSuperBlockOffset(SuperBlockOffset { |
| 2604 |
x: lru_x * lru_sb_w, |
| 2605 |
y: lru_y * lru_sb_h, |
| 2606 |
}); |
| 2607 |
if ts.restoration.has_restoration_unit( |
| 2608 |
base_sbo + loop_sbo, |
| 2609 |
pli, |
| 2610 |
false, |
| 2611 |
) { |
| 2612 |
let src_plane = &src_subset.planes[pli]; // uncompressed input for reference |
| 2613 |
let lrf_in_plane = &lrf_input.planes[pli]; |
| 2614 |
let lrf_po = loop_sbo.plane_offset(src_plane.plane_cfg); |
| 2615 |
let mut best_new_lrf = best_lrf[lru_y * lru_w[pli] + lru_x][pli]; |
| 2616 |
let mut best_cost = |
| 2617 |
best_lrf_cost[lru_y * lru_w[pli] + lru_x][pli]; |
| 2618 |
|
| 2619 |
// Check the no filter option |
| 2620 |
{ |
| 2621 |
let err = rdo_loop_plane_error( |
| 2622 |
base_sbo, |
| 2623 |
loop_sbo, |
| 2624 |
lru_sb_w, |
| 2625 |
lru_sb_h, |
| 2626 |
fi, |
| 2627 |
ts, |
| 2628 |
&tileblocks_subset.as_const(), |
| 2629 |
lrf_input, |
| 2630 |
&src_subset, |
| 2631 |
pli, |
| 2632 |
); |
| 2633 |
let rate = cw.fc.count_lrf_switchable( |
| 2634 |
w, |
| 2635 |
&ts.restoration.as_const(), |
| 2636 |
best_new_lrf, |
| 2637 |
pli, |
| 2638 |
); |
| 2639 |
|
| 2640 |
let cost = compute_rd_cost(fi, rate, err); |
| 2641 |
// Was this choice actually an improvement? |
| 2642 |
if best_cost < 0. || cost < best_cost { |
| 2643 |
best_cost = cost; |
| 2644 |
best_lrf_cost[lru_y * lru_w[pli] + lru_x][pli] = cost; |
| 2645 |
best_new_lrf = RestorationFilter::None; |
| 2646 |
} |
| 2647 |
} |
| 2648 |
|
| 2649 |
// Look for a self guided filter |
| 2650 |
// We need the cropped-to-visible-frame computation area of this LRU |
| 2651 |
let vis_width = unit_size.min( |
| 2652 |
(crop_w >> xdec) |
| 2653 |
- loop_sbo.plane_offset(&lrf_ref.planes[pli].cfg).x as usize, |
| 2654 |
); |
| 2655 |
let vis_height = unit_size.min( |
| 2656 |
(crop_h >> ydec) |
| 2657 |
- loop_sbo.plane_offset(&lrf_ref.planes[pli].cfg).y as usize, |
| 2658 |
); |
| 2659 |
|
| 2660 |
// todo: experiment with borrowing border pixels |
| 2661 |
// rather than edge-extending. Right now this is |
| 2662 |
// hard-clipping to the superblock boundary. |
| 2663 |
setup_integral_image( |
| 2664 |
&mut ts.integral_buffer, |
| 2665 |
SOLVE_IMAGE_STRIDE, |
| 2666 |
vis_width, |
| 2667 |
vis_height, |
| 2668 |
vis_width, |
| 2669 |
vis_height, |
| 2670 |
&lrf_in_plane.slice(lrf_po), |
| 2671 |
&lrf_in_plane.slice(lrf_po), |
| 2672 |
); |
| 2673 |
|
| 2674 |
for &set in get_sgr_sets(fi.config.speed_settings.sgr_complexity) |
| 2675 |
{ |
| 2676 |
let (xqd0, xqd1) = sgrproj_solve( |
| 2677 |
set, |
| 2678 |
fi, |
| 2679 |
&ts.integral_buffer, |
| 2680 |
&src_plane |
| 2681 |
.subregion(Area::StartingAt { x: lrf_po.x, y: lrf_po.y }), |
| 2682 |
&lrf_in_plane.slice(lrf_po), |
| 2683 |
vis_width, |
| 2684 |
vis_height, |
| 2685 |
); |
| 2686 |
let current_lrf = |
| 2687 |
RestorationFilter::Sgrproj { set, xqd: [xqd0, xqd1] }; |
| 2688 |
if let RestorationFilter::Sgrproj { set, xqd } = current_lrf { |
| 2689 |
sgrproj_stripe_filter( |
| 2690 |
set, |
| 2691 |
xqd, |
| 2692 |
fi, |
| 2693 |
&ts.integral_buffer, |
| 2694 |
SOLVE_IMAGE_STRIDE, |
| 2695 |
&lrf_in_plane.slice(lrf_po), |
| 2696 |
&mut lrf_ref.planes[pli].region_mut(Area::Rect { |
| 2697 |
x: lrf_po.x, |
| 2698 |
y: lrf_po.y, |
| 2699 |
width: vis_width, |
| 2700 |
height: vis_height, |
| 2701 |
}), |
| 2702 |
); |
| 2703 |
} |
| 2704 |
let err = rdo_loop_plane_error( |
| 2705 |
base_sbo, |
| 2706 |
loop_sbo, |
| 2707 |
lru_sb_w, |
| 2708 |
lru_sb_h, |
| 2709 |
fi, |
| 2710 |
ts, |
| 2711 |
&tileblocks_subset.as_const(), |
| 2712 |
lrf_ref, |
| 2713 |
&src_subset, |
| 2714 |
pli, |
| 2715 |
); |
| 2716 |
let rate = cw.fc.count_lrf_switchable( |
| 2717 |
w, |
| 2718 |
&ts.restoration.as_const(), |
| 2719 |
current_lrf, |
| 2720 |
pli, |
| 2721 |
); |
| 2722 |
let cost = compute_rd_cost(fi, rate, err); |
| 2723 |
if cost < best_cost { |
| 2724 |
best_cost = cost; |
| 2725 |
best_lrf_cost[lru_y * lru_w[pli] + lru_x][pli] = cost; |
| 2726 |
best_new_lrf = current_lrf; |
| 2727 |
} |
| 2728 |
} |
| 2729 |
|
| 2730 |
if best_lrf[lru_y * lru_w[pli] + lru_x][pli] |
| 2731 |
.notequal(best_new_lrf) |
| 2732 |
{ |
| 2733 |
best_lrf[lru_y * lru_w[pli] + lru_x][pli] = best_new_lrf; |
| 2734 |
lrf_change = true; |
| 2735 |
if let Some(ru) = ts.restoration.planes[pli] |
| 2736 |
.restoration_unit_mut(base_sbo + loop_sbo) |
| 2737 |
{ |
| 2738 |
ru.filter = best_new_lrf; |
| 2739 |
} |
| 2740 |
} |
| 2741 |
} |
| 2742 |
} |
| 2743 |
} |
| 2744 |
} |
| 2745 |
} |
| 2746 |
} |
| 2747 |
} |
| 2748 |
|
| 2749 |
#[test] |
| 2750 |
fn estimate_rate_test() { |
| 2751 |
assert_eq!(estimate_rate(0, TxSize::TX_4X4, 0), RDO_RATE_TABLE[0][0][0]); |
| 2752 |
} |