| 1 |
// Copyright (c) 2018-2022, The rav1e contributors. All rights reserved |
| 2 |
// |
| 3 |
// This source code is subject to the terms of the BSD 2 Clause License and |
| 4 |
// the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License |
| 5 |
// was not distributed with this source code in the LICENSE file, you can |
| 6 |
// obtain it at www.aomedia.org/license/software. If the Alliance for Open |
| 7 |
// Media Patent License 1.0 was not distributed with this source code in the |
| 8 |
// PATENTS file, you can obtain it at www.aomedia.org/license/patent. |
| 9 |
#![deny(missing_docs)] |
| 10 |
|
| 11 |
use crate::activity::ActivityMask; |
| 12 |
use crate::api::lookahead::*; |
| 13 |
use crate::api::{ |
| 14 |
EncoderConfig, EncoderStatus, FrameType, Opaque, Packet, T35, |
| 15 |
}; |
| 16 |
use crate::color::ChromaSampling::Cs400; |
| 17 |
use crate::dist::get_satd; |
| 18 |
use crate::encoder::*; |
| 19 |
use crate::frame::*; |
| 20 |
use crate::partition::*; |
| 21 |
use crate::rate::{ |
| 22 |
RCState, FRAME_NSUBTYPES, FRAME_SUBTYPE_I, FRAME_SUBTYPE_P, |
| 23 |
FRAME_SUBTYPE_SEF, |
| 24 |
}; |
| 25 |
use crate::stats::EncoderStats; |
| 26 |
use crate::tiling::Area; |
| 27 |
use crate::util::Pixel; |
| 28 |
use arrayvec::ArrayVec; |
| 29 |
use av_scenechange::SceneChangeDetector; |
| 30 |
use std::cmp; |
| 31 |
use std::collections::{BTreeMap, BTreeSet}; |
| 32 |
use std::env; |
| 33 |
use std::fs; |
| 34 |
use std::path::PathBuf; |
| 35 |
use std::sync::Arc; |
| 36 |
|
| 37 |
/// The set of options that controls frame re-ordering and reference picture |
| 38 |
/// selection. |
| 39 |
/// The options stored here are invariant over the whole encode. |
| 40 |
#[derive(Debug, Clone, Copy)] |
| 41 |
pub struct InterConfig { |
| 42 |
/// Whether frame re-ordering is enabled. |
| 43 |
reorder: bool, |
| 44 |
/// Whether P-frames can use multiple references. |
| 45 |
pub(crate) multiref: bool, |
| 46 |
/// The depth of the re-ordering pyramid. |
| 47 |
/// The current code cannot support values larger than 2. |
| 48 |
pub(crate) pyramid_depth: u64, |
| 49 |
/// Number of input frames in group. |
| 50 |
pub(crate) group_input_len: u64, |
| 51 |
/// Number of output frames in group. |
| 52 |
/// This includes both hidden frames and "show existing frame" frames. |
| 53 |
group_output_len: u64, |
| 54 |
/// Interval between consecutive S-frames. |
| 55 |
/// Keyframes reset this interval. |
| 56 |
/// This MUST be a multiple of `group_input_len`. |
| 57 |
pub(crate) switch_frame_interval: u64, |
| 58 |
} |
| 59 |
|
| 60 |
impl InterConfig { |
| 61 |
pub(crate) fn new(enc_config: &EncoderConfig) -> InterConfig { |
| 62 |
let reorder = !enc_config.low_latency; |
| 63 |
// A group always starts with (group_output_len - group_input_len) hidden |
| 64 |
// frames, followed by group_input_len shown frames. |
| 65 |
// The shown frames iterate over the input frames in order, with frames |
| 66 |
// already encoded as hidden frames now displayed with Show Existing |
| 67 |
// Frame. |
| 68 |
// For example, for a pyramid depth of 2, the group is as follows: |
| 69 |
// |TU |TU |TU |TU |
| 70 |
// idx_in_group_output: 0 1 2 3 4 5 |
| 71 |
// input_frameno: 4 2 1 SEF 3 SEF |
| 72 |
// output_frameno: 1 2 3 4 5 6 |
| 73 |
// level: 0 1 2 1 2 0 |
| 74 |
// ^^^^^ ^^^^^^^^^^^^^ |
| 75 |
// hidden shown |
| 76 |
// TODO: This only works for pyramid_depth <= 2 --- after that we need |
| 77 |
// more hidden frames in the middle of the group. |
| 78 |
let pyramid_depth = if reorder { 2 } else { 0 }; |
| 79 |
let group_input_len = 1 << pyramid_depth; |
| 80 |
let group_output_len = group_input_len + pyramid_depth; |
| 81 |
let switch_frame_interval = enc_config.switch_frame_interval; |
| 82 |
assert!(switch_frame_interval % group_input_len == 0); |
| 83 |
InterConfig { |
| 84 |
reorder, |
| 85 |
multiref: reorder || enc_config.speed_settings.multiref, |
| 86 |
pyramid_depth, |
| 87 |
group_input_len, |
| 88 |
group_output_len, |
| 89 |
switch_frame_interval, |
| 90 |
} |
| 91 |
} |
| 92 |
|
| 93 |
/// Get the index of an output frame in its re-ordering group given the output |
| 94 |
/// frame number of the frame in the current keyframe gop. |
| 95 |
/// When re-ordering is disabled, this always returns 0. |
| 96 |
pub(crate) fn get_idx_in_group_output( |
| 97 |
&self, output_frameno_in_gop: u64, |
| 98 |
) -> u64 { |
| 99 |
// The first frame in the GOP should be a keyframe and is not re-ordered, |
| 100 |
// so we should not be calling this function on it. |
| 101 |
debug_assert!(output_frameno_in_gop > 0); |
| 102 |
(output_frameno_in_gop - 1) % self.group_output_len |
| 103 |
} |
| 104 |
|
| 105 |
/// Get the order-hint of an output frame given the output frame number of the |
| 106 |
/// frame in the current keyframe gop and the index of that output frame |
| 107 |
/// in its re-ordering gorup. |
| 108 |
pub(crate) fn get_order_hint( |
| 109 |
&self, output_frameno_in_gop: u64, idx_in_group_output: u64, |
| 110 |
) -> u32 { |
| 111 |
// The first frame in the GOP should be a keyframe, but currently this |
| 112 |
// function only handles inter frames. |
| 113 |
// We could return 0 for keyframes if keyframe support is needed. |
| 114 |
debug_assert!(output_frameno_in_gop > 0); |
| 115 |
// Which P-frame group in the current gop is this output frame in? |
| 116 |
// Subtract 1 because the first frame in the gop is always a keyframe. |
| 117 |
let group_idx = (output_frameno_in_gop - 1) / self.group_output_len; |
| 118 |
// Get the offset to the corresponding input frame. |
| 119 |
// TODO: This only works with pyramid_depth <= 2. |
| 120 |
let offset = if idx_in_group_output < self.pyramid_depth { |
| 121 |
self.group_input_len >> idx_in_group_output |
| 122 |
} else { |
| 123 |
idx_in_group_output - self.pyramid_depth + 1 |
| 124 |
}; |
| 125 |
// Construct the final order hint relative to the start of the group. |
| 126 |
(self.group_input_len * group_idx + offset) as u32 |
| 127 |
} |
| 128 |
|
| 129 |
/// Get the level of the current frame in the pyramid. |
| 130 |
pub(crate) const fn get_level(&self, idx_in_group_output: u64) -> u64 { |
| 131 |
if !self.reorder { |
| 132 |
0 |
| 133 |
} else if idx_in_group_output < self.pyramid_depth { |
| 134 |
// Hidden frames are output first (to be shown in the future). |
| 135 |
idx_in_group_output |
| 136 |
} else { |
| 137 |
// Shown frames |
| 138 |
// TODO: This only works with pyramid_depth <= 2. |
| 139 |
pos_to_lvl( |
| 140 |
idx_in_group_output - self.pyramid_depth + 1, |
| 141 |
self.pyramid_depth, |
| 142 |
) |
| 143 |
} |
| 144 |
} |
| 145 |
|
| 146 |
pub(crate) const fn get_slot_idx(&self, level: u64, order_hint: u32) -> u32 { |
| 147 |
// Frames with level == 0 are stored in slots 0..4, and frames with higher |
| 148 |
// values of level in slots 4..8 |
| 149 |
if level == 0 { |
| 150 |
(order_hint >> self.pyramid_depth) & 3 |
| 151 |
} else { |
| 152 |
// This only works with pyramid_depth <= 4. |
| 153 |
3 + level as u32 |
| 154 |
} |
| 155 |
} |
| 156 |
|
| 157 |
pub(crate) const fn get_show_frame(&self, idx_in_group_output: u64) -> bool { |
| 158 |
idx_in_group_output >= self.pyramid_depth |
| 159 |
} |
| 160 |
|
| 161 |
pub(crate) const fn get_show_existing_frame( |
| 162 |
&self, idx_in_group_output: u64, |
| 163 |
) -> bool { |
| 164 |
// The self.reorder test here is redundant, but short-circuits the rest, |
| 165 |
// avoiding a bunch of work when it's false. |
| 166 |
self.reorder |
| 167 |
&& self.get_show_frame(idx_in_group_output) |
| 168 |
&& (idx_in_group_output - self.pyramid_depth + 1).count_ones() == 1 |
| 169 |
&& idx_in_group_output != self.pyramid_depth |
| 170 |
} |
| 171 |
|
| 172 |
pub(crate) fn get_input_frameno( |
| 173 |
&self, output_frameno_in_gop: u64, gop_input_frameno_start: u64, |
| 174 |
) -> u64 { |
| 175 |
if output_frameno_in_gop == 0 { |
| 176 |
gop_input_frameno_start |
| 177 |
} else { |
| 178 |
let idx_in_group_output = |
| 179 |
self.get_idx_in_group_output(output_frameno_in_gop); |
| 180 |
let order_hint = |
| 181 |
self.get_order_hint(output_frameno_in_gop, idx_in_group_output); |
| 182 |
gop_input_frameno_start + order_hint as u64 |
| 183 |
} |
| 184 |
} |
| 185 |
|
| 186 |
const fn max_reordering_latency(&self) -> u64 { |
| 187 |
self.group_input_len |
| 188 |
} |
| 189 |
|
| 190 |
pub(crate) const fn keyframe_lookahead_distance(&self) -> u64 { |
| 191 |
self.max_reordering_latency() + 1 |
| 192 |
} |
| 193 |
|
| 194 |
pub(crate) const fn allowed_ref_frames(&self) -> &[RefType] { |
| 195 |
use crate::partition::RefType::*; |
| 196 |
if self.reorder { |
| 197 |
&ALL_INTER_REFS |
| 198 |
} else if self.multiref { |
| 199 |
&[LAST_FRAME, LAST2_FRAME, LAST3_FRAME, GOLDEN_FRAME] |
| 200 |
} else { |
| 201 |
&[LAST_FRAME] |
| 202 |
} |
| 203 |
} |
| 204 |
} |
| 205 |
|
| 206 |
// Thin wrapper for frame-related data |
| 207 |
// that gets cached and reused throughout the life of a frame. |
| 208 |
#[derive(Clone)] |
| 209 |
pub(crate) struct FrameData<T: Pixel> { |
| 210 |
pub(crate) fi: FrameInvariants<T>, |
| 211 |
pub(crate) fs: FrameState<T>, |
| 212 |
} |
| 213 |
|
| 214 |
impl<T: Pixel> FrameData<T> { |
| 215 |
pub(crate) fn new(fi: FrameInvariants<T>, frame: Arc<Frame<T>>) -> Self { |
| 216 |
let fs = FrameState::new_with_frame(&fi, frame); |
| 217 |
FrameData { fi, fs } |
| 218 |
} |
| 219 |
} |
| 220 |
|
| 221 |
type FrameQueue<T> = BTreeMap<u64, Option<Arc<Frame<T>>>>; |
| 222 |
type FrameDataQueue<T> = BTreeMap<u64, Option<FrameData<T>>>; |
| 223 |
|
| 224 |
// the fields pub(super) are accessed only by the tests |
| 225 |
pub(crate) struct ContextInner<T: Pixel> { |
| 226 |
pub(crate) frame_count: u64, |
| 227 |
pub(crate) limit: Option<u64>, |
| 228 |
pub(crate) output_frameno: u64, |
| 229 |
pub(super) inter_cfg: InterConfig, |
| 230 |
pub(super) frames_processed: u64, |
| 231 |
/// Maps *`input_frameno`* to frames |
| 232 |
pub(super) frame_q: FrameQueue<T>, |
| 233 |
/// Maps *`output_frameno`* to frame data |
| 234 |
pub(super) frame_data: FrameDataQueue<T>, |
| 235 |
/// A list of the `input_frameno` for keyframes in this encode. |
| 236 |
/// Needed so that we don't need to keep all of the `frame_invariants` in |
| 237 |
/// memory for the whole life of the encode. |
| 238 |
// TODO: Is this needed at all? |
| 239 |
keyframes: BTreeSet<u64>, |
| 240 |
// TODO: Is this needed at all? |
| 241 |
keyframes_forced: BTreeSet<u64>, |
| 242 |
/// A storage space for reordered frames. |
| 243 |
packet_data: Vec<u8>, |
| 244 |
/// Maps `output_frameno` to `gop_output_frameno_start`. |
| 245 |
gop_output_frameno_start: BTreeMap<u64, u64>, |
| 246 |
/// Maps `output_frameno` to `gop_input_frameno_start`. |
| 247 |
pub(crate) gop_input_frameno_start: BTreeMap<u64, u64>, |
| 248 |
keyframe_detector: SceneChangeDetector<T>, |
| 249 |
pub(crate) config: Arc<EncoderConfig>, |
| 250 |
seq: Arc<Sequence>, |
| 251 |
pub(crate) rc_state: RCState, |
| 252 |
maybe_prev_log_base_q: Option<i64>, |
| 253 |
/// The next `input_frameno` to be processed by lookahead. |
| 254 |
next_lookahead_frame: u64, |
| 255 |
/// The next `output_frameno` to be computed by lookahead. |
| 256 |
next_lookahead_output_frameno: u64, |
| 257 |
/// Optional opaque to be sent back to the user |
| 258 |
opaque_q: BTreeMap<u64, Opaque>, |
| 259 |
/// Optional T35 metadata per frame |
| 260 |
t35_q: BTreeMap<u64, Box<[T35]>>, |
| 261 |
} |
| 262 |
|
| 263 |
impl<T: Pixel> ContextInner<T> { |
| 264 |
pub fn new(enc: &EncoderConfig) -> Self { |
| 265 |
// initialize with temporal delimiter |
| 266 |
let packet_data = TEMPORAL_DELIMITER.to_vec(); |
| 267 |
let mut keyframes = BTreeSet::new(); |
| 268 |
keyframes.insert(0); |
| 269 |
|
| 270 |
let maybe_ac_qi_max = |
| 271 |
if enc.quantizer < 255 { Some(enc.quantizer as u8) } else { None }; |
| 272 |
|
| 273 |
let seq = Arc::new(Sequence::new(enc)); |
| 274 |
let inter_cfg = InterConfig::new(enc); |
| 275 |
let lookahead_distance = inter_cfg.keyframe_lookahead_distance() as usize; |
| 276 |
let mut keyframe_detector = SceneChangeDetector::new( |
| 277 |
(enc.width, enc.height), |
| 278 |
enc.bit_depth, |
| 279 |
av_scenechange::Rational32::new( |
| 280 |
enc.time_base.den as i32, |
| 281 |
enc.time_base.num as i32, |
| 282 |
), |
| 283 |
enc.chroma_sampling, |
| 284 |
lookahead_distance, |
| 285 |
match enc.speed_settings.scene_detection_mode { |
| 286 |
super::SceneDetectionSpeed::Fast => { |
| 287 |
av_scenechange::SceneDetectionSpeed::Fast |
| 288 |
} |
| 289 |
super::SceneDetectionSpeed::Standard => { |
| 290 |
av_scenechange::SceneDetectionSpeed::Standard |
| 291 |
} |
| 292 |
super::SceneDetectionSpeed::None => { |
| 293 |
av_scenechange::SceneDetectionSpeed::None |
| 294 |
} |
| 295 |
}, |
| 296 |
enc.min_key_frame_interval as usize, |
| 297 |
enc.max_key_frame_interval as usize, |
| 298 |
av_scenechange::CpuFeatureLevel::default(), |
| 299 |
); |
| 300 |
keyframe_detector.enable_cache(); |
| 301 |
|
| 302 |
ContextInner { |
| 303 |
frame_count: 0, |
| 304 |
limit: None, |
| 305 |
inter_cfg, |
| 306 |
output_frameno: 0, |
| 307 |
frames_processed: 0, |
| 308 |
frame_q: BTreeMap::new(), |
| 309 |
frame_data: BTreeMap::new(), |
| 310 |
keyframes, |
| 311 |
keyframes_forced: BTreeSet::new(), |
| 312 |
packet_data, |
| 313 |
gop_output_frameno_start: BTreeMap::new(), |
| 314 |
gop_input_frameno_start: BTreeMap::new(), |
| 315 |
keyframe_detector, |
| 316 |
config: Arc::new(enc.clone()), |
| 317 |
seq, |
| 318 |
rc_state: RCState::new( |
| 319 |
enc.width as i32, |
| 320 |
enc.height as i32, |
| 321 |
enc.time_base.den as i64, |
| 322 |
enc.time_base.num as i64, |
| 323 |
enc.bitrate, |
| 324 |
maybe_ac_qi_max, |
| 325 |
enc.min_quantizer, |
| 326 |
enc.max_key_frame_interval as i32, |
| 327 |
enc.reservoir_frame_delay, |
| 328 |
), |
| 329 |
maybe_prev_log_base_q: None, |
| 330 |
next_lookahead_frame: 1, |
| 331 |
next_lookahead_output_frameno: 0, |
| 332 |
opaque_q: BTreeMap::new(), |
| 333 |
t35_q: BTreeMap::new(), |
| 334 |
} |
| 335 |
} |
| 336 |
|
| 337 |
#[profiling::function] |
| 338 |
pub fn send_frame( |
| 339 |
&mut self, mut frame: Option<Arc<Frame<T>>>, |
| 340 |
params: Option<FrameParameters>, |
| 341 |
) -> Result<(), EncoderStatus> { |
| 342 |
if let Some(ref mut frame) = frame { |
| 343 |
use crate::api::color::ChromaSampling; |
| 344 |
let EncoderConfig { width, height, chroma_sampling, .. } = *self.config; |
| 345 |
let planes = |
| 346 |
if chroma_sampling == ChromaSampling::Cs400 { 1 } else { 3 }; |
| 347 |
// Try to add padding |
| 348 |
if let Some(ref mut frame) = Arc::get_mut(frame) { |
| 349 |
for plane in frame.planes[..planes].iter_mut() { |
| 350 |
plane.pad(width, height); |
| 351 |
} |
| 352 |
} |
| 353 |
// Enforce that padding is added |
| 354 |
for (p, plane) in frame.planes[..planes].iter().enumerate() { |
| 355 |
assert!( |
| 356 |
plane.probe_padding(width, height), |
| 357 |
"Plane {p} was not padded before passing Frame to send_frame()." |
| 358 |
); |
| 359 |
} |
| 360 |
} |
| 361 |
|
| 362 |
let input_frameno = self.frame_count; |
| 363 |
let is_flushing = frame.is_none(); |
| 364 |
if !is_flushing { |
| 365 |
self.frame_count += 1; |
| 366 |
} |
| 367 |
self.frame_q.insert(input_frameno, frame); |
| 368 |
|
| 369 |
if let Some(params) = params { |
| 370 |
if params.frame_type_override == FrameTypeOverride::Key { |
| 371 |
self.keyframes_forced.insert(input_frameno); |
| 372 |
} |
| 373 |
if let Some(op) = params.opaque { |
| 374 |
self.opaque_q.insert(input_frameno, op); |
| 375 |
} |
| 376 |
self.t35_q.insert(input_frameno, params.t35_metadata); |
| 377 |
} |
| 378 |
|
| 379 |
if !self.needs_more_frame_q_lookahead(self.next_lookahead_frame) { |
| 380 |
let lookahead_frames = self |
| 381 |
.frame_q |
| 382 |
.range(self.next_lookahead_frame - 1..) |
| 383 |
.filter_map(|(&_input_frameno, frame)| frame.as_ref()) |
| 384 |
.collect::<Vec<&Arc<Frame<T>>>>(); |
| 385 |
|
| 386 |
if is_flushing { |
| 387 |
// This is the last time send_frame is called, process all the |
| 388 |
// remaining frames. |
| 389 |
for cur_lookahead_frames in |
| 390 |
std::iter::successors(Some(&lookahead_frames[..]), |s| s.get(1..)) |
| 391 |
{ |
| 392 |
if cur_lookahead_frames.len() < 2 { |
| 393 |
// All frames have been processed |
| 394 |
break; |
| 395 |
} |
| 396 |
|
| 397 |
Self::compute_keyframe_placement( |
| 398 |
cur_lookahead_frames, |
| 399 |
&self.keyframes_forced, |
| 400 |
&mut self.keyframe_detector, |
| 401 |
&mut self.next_lookahead_frame, |
| 402 |
&mut self.keyframes, |
| 403 |
); |
| 404 |
} |
| 405 |
} else { |
| 406 |
Self::compute_keyframe_placement( |
| 407 |
&lookahead_frames, |
| 408 |
&self.keyframes_forced, |
| 409 |
&mut self.keyframe_detector, |
| 410 |
&mut self.next_lookahead_frame, |
| 411 |
&mut self.keyframes, |
| 412 |
); |
| 413 |
} |
| 414 |
} |
| 415 |
|
| 416 |
self.compute_frame_invariants(); |
| 417 |
|
| 418 |
Ok(()) |
| 419 |
} |
| 420 |
|
| 421 |
/// Indicates whether more frames need to be read into the frame queue |
| 422 |
/// in order for frame queue lookahead to be full. |
| 423 |
fn needs_more_frame_q_lookahead(&self, input_frameno: u64) -> bool { |
| 424 |
let lookahead_end = self.frame_q.keys().last().cloned().unwrap_or(0); |
| 425 |
let frames_needed = |
| 426 |
input_frameno + self.inter_cfg.keyframe_lookahead_distance() + 1; |
| 427 |
lookahead_end < frames_needed && self.needs_more_frames(lookahead_end) |
| 428 |
} |
| 429 |
|
| 430 |
/// Indicates whether more frames need to be processed into `FrameInvariants` |
| 431 |
/// in order for FI lookahead to be full. |
| 432 |
pub fn needs_more_fi_lookahead(&self) -> bool { |
| 433 |
let ready_frames = self.get_rdo_lookahead_frames().count(); |
| 434 |
ready_frames < self.config.speed_settings.rdo_lookahead_frames + 1 |
| 435 |
&& self.needs_more_frames(self.next_lookahead_frame) |
| 436 |
} |
| 437 |
|
| 438 |
pub fn needs_more_frames(&self, frame_count: u64) -> bool { |
| 439 |
self.limit.map(|limit| frame_count < limit).unwrap_or(true) |
| 440 |
} |
| 441 |
|
| 442 |
fn get_rdo_lookahead_frames( |
| 443 |
&self, |
| 444 |
) -> impl Iterator<Item = (&u64, &FrameData<T>)> { |
| 445 |
self |
| 446 |
.frame_data |
| 447 |
.iter() |
| 448 |
.skip_while(move |(&output_frameno, _)| { |
| 449 |
output_frameno < self.output_frameno |
| 450 |
}) |
| 451 |
.filter_map(|(fno, data)| data.as_ref().map(|data| (fno, data))) |
| 452 |
.filter(|(_, data)| !data.fi.is_show_existing_frame()) |
| 453 |
.take(self.config.speed_settings.rdo_lookahead_frames + 1) |
| 454 |
} |
| 455 |
|
| 456 |
fn next_keyframe_input_frameno( |
| 457 |
&self, gop_input_frameno_start: u64, ignore_limit: bool, |
| 458 |
) -> u64 { |
| 459 |
let next_detected = self |
| 460 |
.keyframes |
| 461 |
.iter() |
| 462 |
.find(|&&input_frameno| input_frameno > gop_input_frameno_start) |
| 463 |
.cloned(); |
| 464 |
let mut next_limit = |
| 465 |
gop_input_frameno_start + self.config.max_key_frame_interval; |
| 466 |
if !ignore_limit && self.limit.is_some() { |
| 467 |
next_limit = next_limit.min(self.limit.unwrap()); |
| 468 |
} |
| 469 |
if next_detected.is_none() { |
| 470 |
return next_limit; |
| 471 |
} |
| 472 |
cmp::min(next_detected.unwrap(), next_limit) |
| 473 |
} |
| 474 |
|
| 475 |
fn set_frame_properties( |
| 476 |
&mut self, output_frameno: u64, |
| 477 |
) -> Result<(), EncoderStatus> { |
| 478 |
let fi = self.build_frame_properties(output_frameno)?; |
| 479 |
|
| 480 |
self.frame_data.insert( |
| 481 |
output_frameno, |
| 482 |
fi.map(|fi| { |
| 483 |
let frame = self |
| 484 |
.frame_q |
| 485 |
.get(&fi.input_frameno) |
| 486 |
.as_ref() |
| 487 |
.unwrap() |
| 488 |
.as_ref() |
| 489 |
.unwrap(); |
| 490 |
FrameData::new(fi, frame.clone()) |
| 491 |
}), |
| 492 |
); |
| 493 |
|
| 494 |
Ok(()) |
| 495 |
} |
| 496 |
|
| 497 |
#[allow(unused)] |
| 498 |
pub fn build_dump_properties() -> PathBuf { |
| 499 |
let mut data_location = PathBuf::new(); |
| 500 |
if env::var_os("RAV1E_DATA_PATH").is_some() { |
| 501 |
data_location.push(env::var_os("RAV1E_DATA_PATH").unwrap()); |
| 502 |
} else { |
| 503 |
data_location.push(env::current_dir().unwrap()); |
| 504 |
data_location.push(".lookahead_data"); |
| 505 |
} |
| 506 |
fs::create_dir_all(&data_location).unwrap(); |
| 507 |
data_location |
| 508 |
} |
| 509 |
|
| 510 |
fn build_frame_properties( |
| 511 |
&mut self, output_frameno: u64, |
| 512 |
) -> Result<Option<FrameInvariants<T>>, EncoderStatus> { |
| 513 |
let (prev_gop_output_frameno_start, prev_gop_input_frameno_start) = |
| 514 |
if output_frameno == 0 { |
| 515 |
(0, 0) |
| 516 |
} else { |
| 517 |
( |
| 518 |
self.gop_output_frameno_start[&(output_frameno - 1)], |
| 519 |
self.gop_input_frameno_start[&(output_frameno - 1)], |
| 520 |
) |
| 521 |
}; |
| 522 |
|
| 523 |
self |
| 524 |
.gop_output_frameno_start |
| 525 |
.insert(output_frameno, prev_gop_output_frameno_start); |
| 526 |
self |
| 527 |
.gop_input_frameno_start |
| 528 |
.insert(output_frameno, prev_gop_input_frameno_start); |
| 529 |
|
| 530 |
let output_frameno_in_gop = |
| 531 |
output_frameno - self.gop_output_frameno_start[&output_frameno]; |
| 532 |
let mut input_frameno = self.inter_cfg.get_input_frameno( |
| 533 |
output_frameno_in_gop, |
| 534 |
self.gop_input_frameno_start[&output_frameno], |
| 535 |
); |
| 536 |
|
| 537 |
if self.needs_more_frame_q_lookahead(input_frameno) { |
| 538 |
return Err(EncoderStatus::NeedMoreData); |
| 539 |
} |
| 540 |
|
| 541 |
let t35_metadata = if let Some(t35) = self.t35_q.remove(&input_frameno) { |
| 542 |
t35 |
| 543 |
} else { |
| 544 |
Box::new([]) |
| 545 |
}; |
| 546 |
|
| 547 |
if output_frameno_in_gop > 0 { |
| 548 |
let next_keyframe_input_frameno = self.next_keyframe_input_frameno( |
| 549 |
self.gop_input_frameno_start[&output_frameno], |
| 550 |
false, |
| 551 |
); |
| 552 |
let prev_input_frameno = |
| 553 |
self.get_previous_fi(output_frameno).input_frameno; |
| 554 |
if input_frameno >= next_keyframe_input_frameno { |
| 555 |
if !self.inter_cfg.reorder |
| 556 |
|| ((output_frameno_in_gop - 1) % self.inter_cfg.group_output_len |
| 557 |
== 0 |
| 558 |
&& prev_input_frameno == (next_keyframe_input_frameno - 1)) |
| 559 |
{ |
| 560 |
input_frameno = next_keyframe_input_frameno; |
| 561 |
|
| 562 |
// If we'll return early, do it before modifying the state. |
| 563 |
match self.frame_q.get(&input_frameno) { |
| 564 |
Some(Some(_)) => {} |
| 565 |
_ => { |
| 566 |
return Err(EncoderStatus::NeedMoreData); |
| 567 |
} |
| 568 |
} |
| 569 |
|
| 570 |
*self.gop_output_frameno_start.get_mut(&output_frameno).unwrap() = |
| 571 |
output_frameno; |
| 572 |
*self.gop_input_frameno_start.get_mut(&output_frameno).unwrap() = |
| 573 |
next_keyframe_input_frameno; |
| 574 |
} else { |
| 575 |
let fi = FrameInvariants::new_inter_frame( |
| 576 |
self.get_previous_coded_fi(output_frameno), |
| 577 |
&self.inter_cfg, |
| 578 |
self.gop_input_frameno_start[&output_frameno], |
| 579 |
output_frameno_in_gop, |
| 580 |
next_keyframe_input_frameno, |
| 581 |
self.config.error_resilient, |
| 582 |
t35_metadata, |
| 583 |
); |
| 584 |
assert!(fi.is_none()); |
| 585 |
return Ok(fi); |
| 586 |
} |
| 587 |
} |
| 588 |
} |
| 589 |
|
| 590 |
match self.frame_q.get(&input_frameno) { |
| 591 |
Some(Some(_)) => {} |
| 592 |
_ => { |
| 593 |
return Err(EncoderStatus::NeedMoreData); |
| 594 |
} |
| 595 |
} |
| 596 |
|
| 597 |
// Now that we know the input_frameno, look up the correct frame type |
| 598 |
let frame_type = if self.keyframes.contains(&input_frameno) { |
| 599 |
FrameType::KEY |
| 600 |
} else { |
| 601 |
FrameType::INTER |
| 602 |
}; |
| 603 |
if frame_type == FrameType::KEY { |
| 604 |
*self.gop_output_frameno_start.get_mut(&output_frameno).unwrap() = |
| 605 |
output_frameno; |
| 606 |
*self.gop_input_frameno_start.get_mut(&output_frameno).unwrap() = |
| 607 |
input_frameno; |
| 608 |
} |
| 609 |
|
| 610 |
let output_frameno_in_gop = |
| 611 |
output_frameno - self.gop_output_frameno_start[&output_frameno]; |
| 612 |
if output_frameno_in_gop == 0 { |
| 613 |
let fi = FrameInvariants::new_key_frame( |
| 614 |
self.config.clone(), |
| 615 |
self.seq.clone(), |
| 616 |
self.gop_input_frameno_start[&output_frameno], |
| 617 |
t35_metadata, |
| 618 |
); |
| 619 |
Ok(Some(fi)) |
| 620 |
} else { |
| 621 |
let next_keyframe_input_frameno = self.next_keyframe_input_frameno( |
| 622 |
self.gop_input_frameno_start[&output_frameno], |
| 623 |
false, |
| 624 |
); |
| 625 |
let fi = FrameInvariants::new_inter_frame( |
| 626 |
self.get_previous_coded_fi(output_frameno), |
| 627 |
&self.inter_cfg, |
| 628 |
self.gop_input_frameno_start[&output_frameno], |
| 629 |
output_frameno_in_gop, |
| 630 |
next_keyframe_input_frameno, |
| 631 |
self.config.error_resilient, |
| 632 |
t35_metadata, |
| 633 |
); |
| 634 |
assert!(fi.is_some()); |
| 635 |
Ok(fi) |
| 636 |
} |
| 637 |
} |
| 638 |
|
| 639 |
fn get_previous_fi(&self, output_frameno: u64) -> &FrameInvariants<T> { |
| 640 |
let res = self |
| 641 |
.frame_data |
| 642 |
.iter() |
| 643 |
.filter(|(fno, _)| **fno < output_frameno) |
| 644 |
.rfind(|(_, fd)| fd.is_some()) |
| 645 |
.unwrap(); |
| 646 |
&res.1.as_ref().unwrap().fi |
| 647 |
} |
| 648 |
|
| 649 |
fn get_previous_coded_fi(&self, output_frameno: u64) -> &FrameInvariants<T> { |
| 650 |
let res = self |
| 651 |
.frame_data |
| 652 |
.iter() |
| 653 |
.filter(|(fno, _)| **fno < output_frameno) |
| 654 |
.rfind(|(_, fd)| { |
| 655 |
fd.as_ref().map(|fd| !fd.fi.is_show_existing_frame()).unwrap_or(false) |
| 656 |
}) |
| 657 |
.unwrap(); |
| 658 |
&res.1.as_ref().unwrap().fi |
| 659 |
} |
| 660 |
|
| 661 |
pub(crate) fn done_processing(&self) -> bool { |
| 662 |
self.limit.map(|limit| self.frames_processed == limit).unwrap_or(false) |
| 663 |
} |
| 664 |
|
| 665 |
/// Computes lookahead motion vectors and fills in `lookahead_mvs`, |
| 666 |
/// `rec_buffer` and `lookahead_rec_buffer` on the `FrameInvariants`. This |
| 667 |
/// function must be called after every new `FrameInvariants` is initially |
| 668 |
/// computed. |
| 669 |
#[profiling::function] |
| 670 |
fn compute_lookahead_motion_vectors(&mut self, output_frameno: u64) { |
| 671 |
let frame_data = self.frame_data.get(&output_frameno).unwrap(); |
| 672 |
|
| 673 |
// We're only interested in valid frames which are not show-existing-frame. |
| 674 |
// Those two don't modify the rec_buffer so there's no need to do anything |
| 675 |
// special about it either, it'll propagate on its own. |
| 676 |
if frame_data |
| 677 |
.as_ref() |
| 678 |
.map(|fd| fd.fi.is_show_existing_frame()) |
| 679 |
.unwrap_or(true) |
| 680 |
{ |
| 681 |
return; |
| 682 |
} |
| 683 |
|
| 684 |
let qps = { |
| 685 |
let fti = frame_data.as_ref().unwrap().fi.get_frame_subtype(); |
| 686 |
self.rc_state.select_qi( |
| 687 |
self, |
| 688 |
output_frameno, |
| 689 |
fti, |
| 690 |
self.maybe_prev_log_base_q, |
| 691 |
0, |
| 692 |
) |
| 693 |
}; |
| 694 |
|
| 695 |
let frame_data = |
| 696 |
self.frame_data.get_mut(&output_frameno).unwrap().as_mut().unwrap(); |
| 697 |
let fs = &mut frame_data.fs; |
| 698 |
let fi = &mut frame_data.fi; |
| 699 |
let coded_data = fi.coded_frame_data.as_mut().unwrap(); |
| 700 |
|
| 701 |
#[cfg(feature = "dump_lookahead_data")] |
| 702 |
{ |
| 703 |
let data_location = Self::build_dump_properties(); |
| 704 |
let plane = &fs.input_qres; |
| 705 |
let mut file_name = format!("{:010}-qres", fi.input_frameno); |
| 706 |
let buf: Vec<_> = plane.iter().map(|p| p.as_()).collect(); |
| 707 |
image::GrayImage::from_vec( |
| 708 |
plane.cfg.width as u32, |
| 709 |
plane.cfg.height as u32, |
| 710 |
buf, |
| 711 |
) |
| 712 |
.unwrap() |
| 713 |
.save(data_location.join(file_name).with_extension("png")) |
| 714 |
.unwrap(); |
| 715 |
let plane = &fs.input_hres; |
| 716 |
file_name = format!("{:010}-hres", fi.input_frameno); |
| 717 |
let buf: Vec<_> = plane.iter().map(|p| p.as_()).collect(); |
| 718 |
image::GrayImage::from_vec( |
| 719 |
plane.cfg.width as u32, |
| 720 |
plane.cfg.height as u32, |
| 721 |
buf, |
| 722 |
) |
| 723 |
.unwrap() |
| 724 |
.save(data_location.join(file_name).with_extension("png")) |
| 725 |
.unwrap(); |
| 726 |
} |
| 727 |
|
| 728 |
// Do not modify the next output frame's FrameInvariants. |
| 729 |
if self.output_frameno == output_frameno { |
| 730 |
// We do want to propagate the lookahead_rec_buffer though. |
| 731 |
let rfs = Arc::new(ReferenceFrame { |
| 732 |
order_hint: fi.order_hint, |
| 733 |
width: fi.width as u32, |
| 734 |
height: fi.height as u32, |
| 735 |
render_width: fi.render_width, |
| 736 |
render_height: fi.render_height, |
| 737 |
// Use the original frame contents. |
| 738 |
frame: fs.input.clone(), |
| 739 |
input_hres: fs.input_hres.clone(), |
| 740 |
input_qres: fs.input_qres.clone(), |
| 741 |
cdfs: fs.cdfs, |
| 742 |
frame_me_stats: fs.frame_me_stats.clone(), |
| 743 |
output_frameno, |
| 744 |
segmentation: fs.segmentation, |
| 745 |
}); |
| 746 |
for i in 0..REF_FRAMES { |
| 747 |
if (fi.refresh_frame_flags & (1 << i)) != 0 { |
| 748 |
coded_data.lookahead_rec_buffer.frames[i] = Some(Arc::clone(&rfs)); |
| 749 |
coded_data.lookahead_rec_buffer.deblock[i] = fs.deblock; |
| 750 |
} |
| 751 |
} |
| 752 |
|
| 753 |
return; |
| 754 |
} |
| 755 |
|
| 756 |
// Our lookahead_rec_buffer should be filled with correct original frame |
| 757 |
// data from the previous frames. Copy it into rec_buffer because that's |
| 758 |
// what the MV search uses. During the actual encoding rec_buffer is |
| 759 |
// overwritten with its correct values anyway. |
| 760 |
fi.rec_buffer = coded_data.lookahead_rec_buffer.clone(); |
| 761 |
|
| 762 |
// Estimate lambda with rate-control dry-run |
| 763 |
fi.set_quantizers(&qps); |
| 764 |
|
| 765 |
// TODO: as in the encoding code, key frames will have no references. |
| 766 |
// However, for block importance purposes we want key frames to act as |
| 767 |
// P-frames in this instance. |
| 768 |
// |
| 769 |
// Compute the motion vectors. |
| 770 |
compute_motion_vectors(fi, fs, &self.inter_cfg); |
| 771 |
|
| 772 |
let coded_data = fi.coded_frame_data.as_mut().unwrap(); |
| 773 |
|
| 774 |
#[cfg(feature = "dump_lookahead_data")] |
| 775 |
{ |
| 776 |
use crate::partition::RefType::*; |
| 777 |
let data_location = Self::build_dump_properties(); |
| 778 |
let file_name = format!("{:010}-mvs", fi.input_frameno); |
| 779 |
let second_ref_frame = if !self.inter_cfg.multiref { |
| 780 |
LAST_FRAME // make second_ref_frame match first |
| 781 |
} else if fi.idx_in_group_output == 0 { |
| 782 |
LAST2_FRAME |
| 783 |
} else { |
| 784 |
ALTREF_FRAME |
| 785 |
}; |
| 786 |
|
| 787 |
// Use the default index, it corresponds to the last P-frame or to the |
| 788 |
// backwards lower reference (so the closest previous frame). |
| 789 |
let index = if second_ref_frame.to_index() != 0 { 0 } else { 1 }; |
| 790 |
|
| 791 |
let me_stats = &fs.frame_me_stats.read().expect("poisoned lock")[index]; |
| 792 |
use byteorder::{NativeEndian, WriteBytesExt}; |
| 793 |
// dynamic allocation: debugging only |
| 794 |
let mut buf = vec![]; |
| 795 |
buf.write_u64::<NativeEndian>(me_stats.rows as u64).unwrap(); |
| 796 |
buf.write_u64::<NativeEndian>(me_stats.cols as u64).unwrap(); |
| 797 |
for y in 0..me_stats.rows { |
| 798 |
for x in 0..me_stats.cols { |
| 799 |
let mv = me_stats[y][x].mv; |
| 800 |
buf.write_i16::<NativeEndian>(mv.row).unwrap(); |
| 801 |
buf.write_i16::<NativeEndian>(mv.col).unwrap(); |
| 802 |
} |
| 803 |
} |
| 804 |
::std::fs::write( |
| 805 |
data_location.join(file_name).with_extension("bin"), |
| 806 |
buf, |
| 807 |
) |
| 808 |
.unwrap(); |
| 809 |
} |
| 810 |
|
| 811 |
// Set lookahead_rec_buffer on this FrameInvariants for future |
| 812 |
// FrameInvariants to pick it up. |
| 813 |
let rfs = Arc::new(ReferenceFrame { |
| 814 |
order_hint: fi.order_hint, |
| 815 |
width: fi.width as u32, |
| 816 |
height: fi.height as u32, |
| 817 |
render_width: fi.render_width, |
| 818 |
render_height: fi.render_height, |
| 819 |
// Use the original frame contents. |
| 820 |
frame: fs.input.clone(), |
| 821 |
input_hres: fs.input_hres.clone(), |
| 822 |
input_qres: fs.input_qres.clone(), |
| 823 |
cdfs: fs.cdfs, |
| 824 |
frame_me_stats: fs.frame_me_stats.clone(), |
| 825 |
output_frameno, |
| 826 |
segmentation: fs.segmentation, |
| 827 |
}); |
| 828 |
for i in 0..REF_FRAMES { |
| 829 |
if (fi.refresh_frame_flags & (1 << i)) != 0 { |
| 830 |
coded_data.lookahead_rec_buffer.frames[i] = Some(Arc::clone(&rfs)); |
| 831 |
coded_data.lookahead_rec_buffer.deblock[i] = fs.deblock; |
| 832 |
} |
| 833 |
} |
| 834 |
} |
| 835 |
|
| 836 |
/// Computes lookahead intra cost approximations and fills in |
| 837 |
/// `lookahead_intra_costs` on the `FrameInvariants`. |
| 838 |
fn compute_lookahead_intra_costs(&mut self, output_frameno: u64) { |
| 839 |
let frame_data = self.frame_data.get(&output_frameno).unwrap(); |
| 840 |
let fd = &frame_data.as_ref(); |
| 841 |
|
| 842 |
// We're only interested in valid frames which are not show-existing-frame. |
| 843 |
if fd.map(|fd| fd.fi.is_show_existing_frame()).unwrap_or(true) { |
| 844 |
return; |
| 845 |
} |
| 846 |
|
| 847 |
let fi = &fd.unwrap().fi; |
| 848 |
|
| 849 |
self |
| 850 |
.frame_data |
| 851 |
.get_mut(&output_frameno) |
| 852 |
.unwrap() |
| 853 |
.as_mut() |
| 854 |
.unwrap() |
| 855 |
.fi |
| 856 |
.coded_frame_data |
| 857 |
.as_mut() |
| 858 |
.unwrap() |
| 859 |
.lookahead_intra_costs = self |
| 860 |
.keyframe_detector |
| 861 |
.intra_costs |
| 862 |
.as_mut() |
| 863 |
.and_then(|intra_costs| intra_costs.remove(&(fi.input_frameno as usize))) |
| 864 |
.unwrap_or_else(|| { |
| 865 |
// We use the cached values from scenechange above if available, |
| 866 |
// otherwise we need to calculate them here. |
| 867 |
let frame = self.frame_q[&fi.input_frameno].as_ref().unwrap(); |
| 868 |
let mut temp_plane = frame.planes[0].clone(); |
| 869 |
|
| 870 |
estimate_intra_costs( |
| 871 |
&mut temp_plane, |
| 872 |
&**frame, |
| 873 |
fi.sequence.bit_depth, |
| 874 |
fi.cpu_feature_level, |
| 875 |
) |
| 876 |
}); |
| 877 |
} |
| 878 |
|
| 879 |
#[profiling::function] |
| 880 |
pub fn compute_keyframe_placement( |
| 881 |
lookahead_frames: &[&Arc<Frame<T>>], keyframes_forced: &BTreeSet<u64>, |
| 882 |
keyframe_detector: &mut SceneChangeDetector<T>, |
| 883 |
next_lookahead_frame: &mut u64, keyframes: &mut BTreeSet<u64>, |
| 884 |
) { |
| 885 |
if keyframes_forced.contains(next_lookahead_frame) |
| 886 |
|| keyframe_detector.analyze_next_frame( |
| 887 |
lookahead_frames, |
| 888 |
*next_lookahead_frame as usize, |
| 889 |
*keyframes.iter().last().unwrap() as usize, |
| 890 |
) |
| 891 |
{ |
| 892 |
keyframes.insert(*next_lookahead_frame); |
| 893 |
} |
| 894 |
|
| 895 |
*next_lookahead_frame += 1; |
| 896 |
} |
| 897 |
|
| 898 |
#[profiling::function] |
| 899 |
pub fn compute_frame_invariants(&mut self) { |
| 900 |
while self.set_frame_properties(self.next_lookahead_output_frameno).is_ok() |
| 901 |
{ |
| 902 |
self |
| 903 |
.compute_lookahead_motion_vectors(self.next_lookahead_output_frameno); |
| 904 |
if self.config.temporal_rdo() { |
| 905 |
self.compute_lookahead_intra_costs(self.next_lookahead_output_frameno); |
| 906 |
} |
| 907 |
self.next_lookahead_output_frameno += 1; |
| 908 |
} |
| 909 |
} |
| 910 |
|
| 911 |
#[profiling::function] |
| 912 |
fn update_block_importances( |
| 913 |
fi: &FrameInvariants<T>, me_stats: &crate::me::FrameMEStats, |
| 914 |
frame: &Frame<T>, reference_frame: &Frame<T>, bit_depth: usize, |
| 915 |
bsize: BlockSize, len: usize, |
| 916 |
reference_frame_block_importances: &mut [f32], |
| 917 |
) { |
| 918 |
let coded_data = fi.coded_frame_data.as_ref().unwrap(); |
| 919 |
let plane_org = &frame.planes[0]; |
| 920 |
let plane_ref = &reference_frame.planes[0]; |
| 921 |
let lookahead_intra_costs_lines = |
| 922 |
coded_data.lookahead_intra_costs.chunks_exact(coded_data.w_in_imp_b); |
| 923 |
let block_importances_lines = |
| 924 |
coded_data.block_importances.chunks_exact(coded_data.w_in_imp_b); |
| 925 |
|
| 926 |
lookahead_intra_costs_lines |
| 927 |
.zip(block_importances_lines) |
| 928 |
.zip(me_stats.rows_iter().step_by(2)) |
| 929 |
.enumerate() |
| 930 |
.flat_map( |
| 931 |
|(y, ((lookahead_intra_costs, block_importances), me_stats_line))| { |
| 932 |
lookahead_intra_costs |
| 933 |
.iter() |
| 934 |
.zip(block_importances.iter()) |
| 935 |
.zip(me_stats_line.iter().step_by(2)) |
| 936 |
.enumerate() |
| 937 |
.map(move |(x, ((&intra_cost, &future_importance), &me_stat))| { |
| 938 |
let mv = me_stat.mv; |
| 939 |
|
| 940 |
// Coordinates of the top-left corner of the reference block, in MV |
| 941 |
// units. |
| 942 |
let reference_x = |
| 943 |
x as i64 * IMP_BLOCK_SIZE_IN_MV_UNITS + mv.col as i64; |
| 944 |
let reference_y = |
| 945 |
y as i64 * IMP_BLOCK_SIZE_IN_MV_UNITS + mv.row as i64; |
| 946 |
|
| 947 |
let region_org = plane_org.region(Area::Rect { |
| 948 |
x: (x * IMPORTANCE_BLOCK_SIZE) as isize, |
| 949 |
y: (y * IMPORTANCE_BLOCK_SIZE) as isize, |
| 950 |
width: IMPORTANCE_BLOCK_SIZE, |
| 951 |
height: IMPORTANCE_BLOCK_SIZE, |
| 952 |
}); |
| 953 |
|
| 954 |
let region_ref = plane_ref.region(Area::Rect { |
| 955 |
x: reference_x as isize |
| 956 |
/ IMP_BLOCK_MV_UNITS_PER_PIXEL as isize, |
| 957 |
y: reference_y as isize |
| 958 |
/ IMP_BLOCK_MV_UNITS_PER_PIXEL as isize, |
| 959 |
width: IMPORTANCE_BLOCK_SIZE, |
| 960 |
height: IMPORTANCE_BLOCK_SIZE, |
| 961 |
}); |
| 962 |
|
| 963 |
let inter_cost = get_satd( |
| 964 |
®ion_org, |
| 965 |
®ion_ref, |
| 966 |
bsize.width(), |
| 967 |
bsize.height(), |
| 968 |
bit_depth, |
| 969 |
fi.cpu_feature_level, |
| 970 |
) as f32; |
| 971 |
|
| 972 |
let intra_cost = intra_cost as f32; |
| 973 |
// let intra_cost = lookahead_intra_costs[x] as f32; |
| 974 |
// let future_importance = block_importances[x]; |
| 975 |
|
| 976 |
let propagate_fraction = if intra_cost <= inter_cost { |
| 977 |
0. |
| 978 |
} else { |
| 979 |
1. - inter_cost / intra_cost |
| 980 |
}; |
| 981 |
|
| 982 |
let propagate_amount = (intra_cost + future_importance) |
| 983 |
* propagate_fraction |
| 984 |
/ len as f32; |
| 985 |
(propagate_amount, reference_x, reference_y) |
| 986 |
}) |
| 987 |
}, |
| 988 |
) |
| 989 |
.for_each(|(propagate_amount, reference_x, reference_y)| { |
| 990 |
let mut propagate = |
| 991 |
|block_x_in_mv_units, block_y_in_mv_units, fraction| { |
| 992 |
let x = block_x_in_mv_units / IMP_BLOCK_SIZE_IN_MV_UNITS; |
| 993 |
let y = block_y_in_mv_units / IMP_BLOCK_SIZE_IN_MV_UNITS; |
| 994 |
|
| 995 |
// TODO: propagate partially if the block is partially off-frame |
| 996 |
// (possible on right and bottom edges)? |
| 997 |
if x >= 0 |
| 998 |
&& y >= 0 |
| 999 |
&& (x as usize) < coded_data.w_in_imp_b |
| 1000 |
&& (y as usize) < coded_data.h_in_imp_b |
| 1001 |
{ |
| 1002 |
reference_frame_block_importances |
| 1003 |
[y as usize * coded_data.w_in_imp_b + x as usize] += |
| 1004 |
propagate_amount * fraction; |
| 1005 |
} |
| 1006 |
}; |
| 1007 |
|
| 1008 |
// Coordinates of the top-left corner of the block intersecting the |
| 1009 |
// reference block from the top-left. |
| 1010 |
let top_left_block_x = (reference_x |
| 1011 |
- if reference_x < 0 { IMP_BLOCK_SIZE_IN_MV_UNITS - 1 } else { 0 }) |
| 1012 |
/ IMP_BLOCK_SIZE_IN_MV_UNITS |
| 1013 |
* IMP_BLOCK_SIZE_IN_MV_UNITS; |
| 1014 |
let top_left_block_y = (reference_y |
| 1015 |
- if reference_y < 0 { IMP_BLOCK_SIZE_IN_MV_UNITS - 1 } else { 0 }) |
| 1016 |
/ IMP_BLOCK_SIZE_IN_MV_UNITS |
| 1017 |
* IMP_BLOCK_SIZE_IN_MV_UNITS; |
| 1018 |
|
| 1019 |
debug_assert!(reference_x >= top_left_block_x); |
| 1020 |
debug_assert!(reference_y >= top_left_block_y); |
| 1021 |
|
| 1022 |
let top_right_block_x = top_left_block_x + IMP_BLOCK_SIZE_IN_MV_UNITS; |
| 1023 |
let top_right_block_y = top_left_block_y; |
| 1024 |
let bottom_left_block_x = top_left_block_x; |
| 1025 |
let bottom_left_block_y = |
| 1026 |
top_left_block_y + IMP_BLOCK_SIZE_IN_MV_UNITS; |
| 1027 |
let bottom_right_block_x = top_right_block_x; |
| 1028 |
let bottom_right_block_y = bottom_left_block_y; |
| 1029 |
|
| 1030 |
let top_left_block_fraction = ((top_right_block_x - reference_x) |
| 1031 |
* (bottom_left_block_y - reference_y)) |
| 1032 |
as f32 |
| 1033 |
/ IMP_BLOCK_AREA_IN_MV_UNITS as f32; |
| 1034 |
|
| 1035 |
propagate(top_left_block_x, top_left_block_y, top_left_block_fraction); |
| 1036 |
|
| 1037 |
let top_right_block_fraction = |
| 1038 |
((reference_x + IMP_BLOCK_SIZE_IN_MV_UNITS - top_right_block_x) |
| 1039 |
* (bottom_left_block_y - reference_y)) as f32 |
| 1040 |
/ IMP_BLOCK_AREA_IN_MV_UNITS as f32; |
| 1041 |
|
| 1042 |
propagate( |
| 1043 |
top_right_block_x, |
| 1044 |
top_right_block_y, |
| 1045 |
top_right_block_fraction, |
| 1046 |
); |
| 1047 |
|
| 1048 |
let bottom_left_block_fraction = ((top_right_block_x - reference_x) |
| 1049 |
* (reference_y + IMP_BLOCK_SIZE_IN_MV_UNITS - bottom_left_block_y)) |
| 1050 |
as f32 |
| 1051 |
/ IMP_BLOCK_AREA_IN_MV_UNITS as f32; |
| 1052 |
|
| 1053 |
propagate( |
| 1054 |
bottom_left_block_x, |
| 1055 |
bottom_left_block_y, |
| 1056 |
bottom_left_block_fraction, |
| 1057 |
); |
| 1058 |
|
| 1059 |
let bottom_right_block_fraction = |
| 1060 |
((reference_x + IMP_BLOCK_SIZE_IN_MV_UNITS - top_right_block_x) |
| 1061 |
* (reference_y + IMP_BLOCK_SIZE_IN_MV_UNITS - bottom_left_block_y)) |
| 1062 |
as f32 |
| 1063 |
/ IMP_BLOCK_AREA_IN_MV_UNITS as f32; |
| 1064 |
|
| 1065 |
propagate( |
| 1066 |
bottom_right_block_x, |
| 1067 |
bottom_right_block_y, |
| 1068 |
bottom_right_block_fraction, |
| 1069 |
); |
| 1070 |
}); |
| 1071 |
} |
| 1072 |
|
| 1073 |
/// Computes the block importances for the current output frame. |
| 1074 |
#[profiling::function] |
| 1075 |
fn compute_block_importances(&mut self) { |
| 1076 |
// SEF don't need block importances. |
| 1077 |
if self.frame_data[&self.output_frameno] |
| 1078 |
.as_ref() |
| 1079 |
.unwrap() |
| 1080 |
.fi |
| 1081 |
.is_show_existing_frame() |
| 1082 |
{ |
| 1083 |
return; |
| 1084 |
} |
| 1085 |
|
| 1086 |
// Get a list of output_framenos that we want to propagate through. |
| 1087 |
let output_framenos = self |
| 1088 |
.get_rdo_lookahead_frames() |
| 1089 |
.map(|(&output_frameno, _)| output_frameno) |
| 1090 |
.collect::<Vec<_>>(); |
| 1091 |
|
| 1092 |
// The first one should be the current output frame. |
| 1093 |
assert_eq!(output_framenos[0], self.output_frameno); |
| 1094 |
|
| 1095 |
// First, initialize them all with zeros. |
| 1096 |
for output_frameno in output_framenos.iter() { |
| 1097 |
let fi = &mut self |
| 1098 |
.frame_data |
| 1099 |
.get_mut(output_frameno) |
| 1100 |
.unwrap() |
| 1101 |
.as_mut() |
| 1102 |
.unwrap() |
| 1103 |
.fi; |
| 1104 |
for x in |
| 1105 |
fi.coded_frame_data.as_mut().unwrap().block_importances.iter_mut() |
| 1106 |
{ |
| 1107 |
*x = 0.; |
| 1108 |
} |
| 1109 |
} |
| 1110 |
|
| 1111 |
// Now compute and propagate the block importances from the end. The |
| 1112 |
// current output frame will get its block importances from the future |
| 1113 |
// frames. |
| 1114 |
let bsize = BlockSize::from_width_and_height( |
| 1115 |
IMPORTANCE_BLOCK_SIZE, |
| 1116 |
IMPORTANCE_BLOCK_SIZE, |
| 1117 |
); |
| 1118 |
|
| 1119 |
for &output_frameno in output_framenos.iter().skip(1).rev() { |
| 1120 |
// TODO: see comment above about key frames not having references. |
| 1121 |
if self |
| 1122 |
.frame_data |
| 1123 |
.get(&output_frameno) |
| 1124 |
.unwrap() |
| 1125 |
.as_ref() |
| 1126 |
.unwrap() |
| 1127 |
.fi |
| 1128 |
.frame_type |
| 1129 |
== FrameType::KEY |
| 1130 |
{ |
| 1131 |
continue; |
| 1132 |
} |
| 1133 |
|
| 1134 |
// Remove fi from the map temporarily and put it back in in the end of |
| 1135 |
// the iteration. This is required because we need to mutably borrow |
| 1136 |
// referenced fis from the map, and that wouldn't be possible if this was |
| 1137 |
// an active borrow. |
| 1138 |
// |
| 1139 |
// Performance note: Contrary to intuition, |
| 1140 |
// removing the data and re-inserting it at the end |
| 1141 |
// is more performant because it avoids a very expensive clone. |
| 1142 |
let output_frame_data = |
| 1143 |
self.frame_data.remove(&output_frameno).unwrap().unwrap(); |
| 1144 |
{ |
| 1145 |
let fi = &output_frame_data.fi; |
| 1146 |
let fs = &output_frame_data.fs; |
| 1147 |
|
| 1148 |
let frame = self.frame_q[&fi.input_frameno].as_ref().unwrap(); |
| 1149 |
|
| 1150 |
// There can be at most 3 of these. |
| 1151 |
let mut unique_indices = ArrayVec::<_, 3>::new(); |
| 1152 |
|
| 1153 |
for (mv_index, &rec_index) in fi.ref_frames.iter().enumerate() { |
| 1154 |
if !unique_indices.iter().any(|&(_, r)| r == rec_index) { |
| 1155 |
unique_indices.push((mv_index, rec_index)); |
| 1156 |
} |
| 1157 |
} |
| 1158 |
|
| 1159 |
let bit_depth = self.config.bit_depth; |
| 1160 |
let frame_data = &mut self.frame_data; |
| 1161 |
let len = unique_indices.len(); |
| 1162 |
|
| 1163 |
let lookahead_me_stats = |
| 1164 |
fs.frame_me_stats.read().expect("poisoned lock"); |
| 1165 |
|
| 1166 |
// Compute and propagate the importance, split evenly between the |
| 1167 |
// referenced frames. |
| 1168 |
unique_indices.iter().for_each(|&(mv_index, rec_index)| { |
| 1169 |
// Use rec_buffer here rather than lookahead_rec_buffer because |
| 1170 |
// rec_buffer still contains the reference frames for the current frame |
| 1171 |
// (it's only overwritten when the frame is encoded), while |
| 1172 |
// lookahead_rec_buffer already contains reference frames for the next |
| 1173 |
// frame (for the reference propagation to work correctly). |
| 1174 |
let reference = |
| 1175 |
fi.rec_buffer.frames[rec_index as usize].as_ref().unwrap(); |
| 1176 |
let reference_frame = &reference.frame; |
| 1177 |
let reference_output_frameno = reference.output_frameno; |
| 1178 |
let me_stats = &lookahead_me_stats[mv_index]; |
| 1179 |
|
| 1180 |
// We should never use frame as its own reference. |
| 1181 |
assert_ne!(reference_output_frameno, output_frameno); |
| 1182 |
|
| 1183 |
if let Some(reference_frame_block_importances) = |
| 1184 |
frame_data.get_mut(&reference_output_frameno).map(|data| { |
| 1185 |
&mut data |
| 1186 |
.as_mut() |
| 1187 |
.unwrap() |
| 1188 |
.fi |
| 1189 |
.coded_frame_data |
| 1190 |
.as_mut() |
| 1191 |
.unwrap() |
| 1192 |
.block_importances |
| 1193 |
}) |
| 1194 |
{ |
| 1195 |
Self::update_block_importances( |
| 1196 |
fi, |
| 1197 |
me_stats, |
| 1198 |
frame, |
| 1199 |
reference_frame, |
| 1200 |
bit_depth, |
| 1201 |
bsize, |
| 1202 |
len, |
| 1203 |
reference_frame_block_importances, |
| 1204 |
); |
| 1205 |
} |
| 1206 |
}); |
| 1207 |
} |
| 1208 |
self.frame_data.insert(output_frameno, Some(output_frame_data)); |
| 1209 |
} |
| 1210 |
|
| 1211 |
if !output_framenos.is_empty() { |
| 1212 |
let fi = &mut self |
| 1213 |
.frame_data |
| 1214 |
.get_mut(&output_framenos[0]) |
| 1215 |
.unwrap() |
| 1216 |
.as_mut() |
| 1217 |
.unwrap() |
| 1218 |
.fi; |
| 1219 |
let coded_data = fi.coded_frame_data.as_mut().unwrap(); |
| 1220 |
let block_importances = coded_data.block_importances.iter(); |
| 1221 |
let lookahead_intra_costs = coded_data.lookahead_intra_costs.iter(); |
| 1222 |
let distortion_scales = coded_data.distortion_scales.iter_mut(); |
| 1223 |
for ((&propagate_cost, &intra_cost), distortion_scale) in |
| 1224 |
block_importances.zip(lookahead_intra_costs).zip(distortion_scales) |
| 1225 |
{ |
| 1226 |
*distortion_scale = crate::rdo::distortion_scale_for( |
| 1227 |
propagate_cost as f64, |
| 1228 |
intra_cost as f64, |
| 1229 |
); |
| 1230 |
} |
| 1231 |
#[cfg(feature = "dump_lookahead_data")] |
| 1232 |
{ |
| 1233 |
use byteorder::{NativeEndian, WriteBytesExt}; |
| 1234 |
|
| 1235 |
let coded_data = fi.coded_frame_data.as_ref().unwrap(); |
| 1236 |
|
| 1237 |
let mut buf = vec![]; |
| 1238 |
let data_location = Self::build_dump_properties(); |
| 1239 |
let file_name = format!("{:010}-imps", fi.input_frameno); |
| 1240 |
buf.write_u64::<NativeEndian>(coded_data.h_in_imp_b as u64).unwrap(); |
| 1241 |
buf.write_u64::<NativeEndian>(coded_data.w_in_imp_b as u64).unwrap(); |
| 1242 |
buf.write_u64::<NativeEndian>(fi.get_frame_subtype() as u64).unwrap(); |
| 1243 |
for y in 0..coded_data.h_in_imp_b { |
| 1244 |
for x in 0..coded_data.w_in_imp_b { |
| 1245 |
buf |
| 1246 |
.write_f32::<NativeEndian>(f64::from( |
| 1247 |
coded_data.distortion_scales[y * coded_data.w_in_imp_b + x], |
| 1248 |
) as f32) |
| 1249 |
.unwrap(); |
| 1250 |
} |
| 1251 |
} |
| 1252 |
::std::fs::write( |
| 1253 |
data_location.join(file_name).with_extension("bin"), |
| 1254 |
buf, |
| 1255 |
) |
| 1256 |
.unwrap(); |
| 1257 |
} |
| 1258 |
} |
| 1259 |
} |
| 1260 |
|
| 1261 |
pub(crate) fn encode_packet( |
| 1262 |
&mut self, cur_output_frameno: u64, |
| 1263 |
) -> Result<Packet<T>, EncoderStatus> { |
| 1264 |
if self |
| 1265 |
.frame_data |
| 1266 |
.get(&cur_output_frameno) |
| 1267 |
.unwrap() |
| 1268 |
.as_ref() |
| 1269 |
.unwrap() |
| 1270 |
.fi |
| 1271 |
.is_show_existing_frame() |
| 1272 |
{ |
| 1273 |
if !self.rc_state.ready() { |
| 1274 |
return Err(EncoderStatus::NotReady); |
| 1275 |
} |
| 1276 |
|
| 1277 |
self.encode_show_existing_packet(cur_output_frameno) |
| 1278 |
} else if let Some(Some(_)) = self.frame_q.get( |
| 1279 |
&self |
| 1280 |
.frame_data |
| 1281 |
.get(&cur_output_frameno) |
| 1282 |
.unwrap() |
| 1283 |
.as_ref() |
| 1284 |
.unwrap() |
| 1285 |
.fi |
| 1286 |
.input_frameno, |
| 1287 |
) { |
| 1288 |
if !self.rc_state.ready() { |
| 1289 |
return Err(EncoderStatus::NotReady); |
| 1290 |
} |
| 1291 |
|
| 1292 |
self.encode_normal_packet(cur_output_frameno) |
| 1293 |
} else { |
| 1294 |
Err(EncoderStatus::NeedMoreData) |
| 1295 |
} |
| 1296 |
} |
| 1297 |
|
| 1298 |
#[profiling::function] |
| 1299 |
pub fn encode_show_existing_packet( |
| 1300 |
&mut self, cur_output_frameno: u64, |
| 1301 |
) -> Result<Packet<T>, EncoderStatus> { |
| 1302 |
let frame_data = |
| 1303 |
self.frame_data.get_mut(&cur_output_frameno).unwrap().as_mut().unwrap(); |
| 1304 |
let sef_data = encode_show_existing_frame( |
| 1305 |
&frame_data.fi, |
| 1306 |
&mut frame_data.fs, |
| 1307 |
&self.inter_cfg, |
| 1308 |
); |
| 1309 |
let bits = (sef_data.len() * 8) as i64; |
| 1310 |
self.packet_data.extend(sef_data); |
| 1311 |
self.rc_state.update_state( |
| 1312 |
bits, |
| 1313 |
FRAME_SUBTYPE_SEF, |
| 1314 |
frame_data.fi.show_frame, |
| 1315 |
0, |
| 1316 |
false, |
| 1317 |
false, |
| 1318 |
); |
| 1319 |
let (rec, source) = if frame_data.fi.show_frame { |
| 1320 |
(Some(frame_data.fs.rec.clone()), Some(frame_data.fs.input.clone())) |
| 1321 |
} else { |
| 1322 |
(None, None) |
| 1323 |
}; |
| 1324 |
|
| 1325 |
self.output_frameno += 1; |
| 1326 |
|
| 1327 |
let input_frameno = frame_data.fi.input_frameno; |
| 1328 |
let frame_type = frame_data.fi.frame_type; |
| 1329 |
let qp = frame_data.fi.base_q_idx; |
| 1330 |
let enc_stats = frame_data.fs.enc_stats.clone(); |
| 1331 |
self.finalize_packet(rec, source, input_frameno, frame_type, qp, enc_stats) |
| 1332 |
} |
| 1333 |
|
| 1334 |
#[profiling::function] |
| 1335 |
pub fn encode_normal_packet( |
| 1336 |
&mut self, cur_output_frameno: u64, |
| 1337 |
) -> Result<Packet<T>, EncoderStatus> { |
| 1338 |
let mut frame_data = |
| 1339 |
self.frame_data.remove(&cur_output_frameno).unwrap().unwrap(); |
| 1340 |
|
| 1341 |
let mut log_isqrt_mean_scale = 0i64; |
| 1342 |
|
| 1343 |
if let Some(coded_data) = frame_data.fi.coded_frame_data.as_mut() { |
| 1344 |
if self.config.tune == Tune::Psychovisual { |
| 1345 |
let frame = |
| 1346 |
self.frame_q[&frame_data.fi.input_frameno].as_ref().unwrap(); |
| 1347 |
coded_data.activity_mask = ActivityMask::from_plane(&frame.planes[0]); |
| 1348 |
coded_data.activity_mask.fill_scales( |
| 1349 |
frame_data.fi.sequence.bit_depth, |
| 1350 |
&mut coded_data.activity_scales, |
| 1351 |
); |
| 1352 |
log_isqrt_mean_scale = coded_data.compute_spatiotemporal_scores(); |
| 1353 |
} else { |
| 1354 |
coded_data.activity_mask = ActivityMask::default(); |
| 1355 |
log_isqrt_mean_scale = coded_data.compute_temporal_scores(); |
| 1356 |
} |
| 1357 |
#[cfg(feature = "dump_lookahead_data")] |
| 1358 |
{ |
| 1359 |
use crate::encoder::Scales::*; |
| 1360 |
let input_frameno = frame_data.fi.input_frameno; |
| 1361 |
if self.config.tune == Tune::Psychovisual { |
| 1362 |
coded_data.dump_scales( |
| 1363 |
Self::build_dump_properties(), |
| 1364 |
ActivityScales, |
| 1365 |
input_frameno, |
| 1366 |
); |
| 1367 |
coded_data.dump_scales( |
| 1368 |
Self::build_dump_properties(), |
| 1369 |
SpatiotemporalScales, |
| 1370 |
input_frameno, |
| 1371 |
); |
| 1372 |
} |
| 1373 |
coded_data.dump_scales( |
| 1374 |
Self::build_dump_properties(), |
| 1375 |
DistortionScales, |
| 1376 |
input_frameno, |
| 1377 |
); |
| 1378 |
} |
| 1379 |
} |
| 1380 |
|
| 1381 |
let fti = frame_data.fi.get_frame_subtype(); |
| 1382 |
let qps = self.rc_state.select_qi( |
| 1383 |
self, |
| 1384 |
cur_output_frameno, |
| 1385 |
fti, |
| 1386 |
self.maybe_prev_log_base_q, |
| 1387 |
log_isqrt_mean_scale, |
| 1388 |
); |
| 1389 |
frame_data.fi.set_quantizers(&qps); |
| 1390 |
|
| 1391 |
if self.rc_state.needs_trial_encode(fti) { |
| 1392 |
let mut trial_fs = frame_data.fs.clone(); |
| 1393 |
let data = encode_frame(&frame_data.fi, &mut trial_fs, &self.inter_cfg); |
| 1394 |
self.rc_state.update_state( |
| 1395 |
(data.len() * 8) as i64, |
| 1396 |
fti, |
| 1397 |
frame_data.fi.show_frame, |
| 1398 |
qps.log_target_q, |
| 1399 |
true, |
| 1400 |
false, |
| 1401 |
); |
| 1402 |
let qps = self.rc_state.select_qi( |
| 1403 |
self, |
| 1404 |
cur_output_frameno, |
| 1405 |
fti, |
| 1406 |
self.maybe_prev_log_base_q, |
| 1407 |
log_isqrt_mean_scale, |
| 1408 |
); |
| 1409 |
frame_data.fi.set_quantizers(&qps); |
| 1410 |
} |
| 1411 |
|
| 1412 |
let data = |
| 1413 |
encode_frame(&frame_data.fi, &mut frame_data.fs, &self.inter_cfg); |
| 1414 |
#[cfg(feature = "dump_lookahead_data")] |
| 1415 |
{ |
| 1416 |
let input_frameno = frame_data.fi.input_frameno; |
| 1417 |
let data_location = Self::build_dump_properties(); |
| 1418 |
frame_data.fs.segmentation.dump_threshold(data_location, input_frameno); |
| 1419 |
} |
| 1420 |
let enc_stats = frame_data.fs.enc_stats.clone(); |
| 1421 |
self.maybe_prev_log_base_q = Some(qps.log_base_q); |
| 1422 |
// TODO: Add support for dropping frames. |
| 1423 |
self.rc_state.update_state( |
| 1424 |
(data.len() * 8) as i64, |
| 1425 |
fti, |
| 1426 |
frame_data.fi.show_frame, |
| 1427 |
qps.log_target_q, |
| 1428 |
false, |
| 1429 |
false, |
| 1430 |
); |
| 1431 |
self.packet_data.extend(data); |
| 1432 |
|
| 1433 |
let planes = |
| 1434 |
if frame_data.fi.sequence.chroma_sampling == Cs400 { 1 } else { 3 }; |
| 1435 |
|
| 1436 |
Arc::get_mut(&mut frame_data.fs.rec).unwrap().pad( |
| 1437 |
frame_data.fi.width, |
| 1438 |
frame_data.fi.height, |
| 1439 |
planes, |
| 1440 |
); |
| 1441 |
|
| 1442 |
let (rec, source) = if frame_data.fi.show_frame { |
| 1443 |
(Some(frame_data.fs.rec.clone()), Some(frame_data.fs.input.clone())) |
| 1444 |
} else { |
| 1445 |
(None, None) |
| 1446 |
}; |
| 1447 |
|
| 1448 |
update_rec_buffer(cur_output_frameno, &mut frame_data.fi, &frame_data.fs); |
| 1449 |
|
| 1450 |
// Copy persistent fields into subsequent FrameInvariants. |
| 1451 |
let rec_buffer = frame_data.fi.rec_buffer.clone(); |
| 1452 |
for subsequent_fi in self |
| 1453 |
.frame_data |
| 1454 |
.iter_mut() |
| 1455 |
.skip_while(|(&output_frameno, _)| output_frameno <= cur_output_frameno) |
| 1456 |
// Here we want the next valid non-show-existing-frame inter frame. |
| 1457 |
// |
| 1458 |
// Copying to show-existing-frame frames isn't actually required |
| 1459 |
// for correct encoding, but it's needed for the reconstruction to |
| 1460 |
// work correctly. |
| 1461 |
.filter_map(|(_, frame_data)| frame_data.as_mut().map(|fd| &mut fd.fi)) |
| 1462 |
.take_while(|fi| fi.frame_type != FrameType::KEY) |
| 1463 |
{ |
| 1464 |
subsequent_fi.rec_buffer = rec_buffer.clone(); |
| 1465 |
subsequent_fi.set_ref_frame_sign_bias(); |
| 1466 |
|
| 1467 |
// Stop after the first non-show-existing-frame. |
| 1468 |
if !subsequent_fi.is_show_existing_frame() { |
| 1469 |
break; |
| 1470 |
} |
| 1471 |
} |
| 1472 |
|
| 1473 |
self.frame_data.insert(cur_output_frameno, Some(frame_data)); |
| 1474 |
let frame_data = |
| 1475 |
self.frame_data.get(&cur_output_frameno).unwrap().as_ref().unwrap(); |
| 1476 |
let fi = &frame_data.fi; |
| 1477 |
|
| 1478 |
self.output_frameno += 1; |
| 1479 |
|
| 1480 |
if fi.show_frame { |
| 1481 |
let input_frameno = fi.input_frameno; |
| 1482 |
let frame_type = fi.frame_type; |
| 1483 |
let qp = fi.base_q_idx; |
| 1484 |
self.finalize_packet( |
| 1485 |
rec, |
| 1486 |
source, |
| 1487 |
input_frameno, |
| 1488 |
frame_type, |
| 1489 |
qp, |
| 1490 |
enc_stats, |
| 1491 |
) |
| 1492 |
} else { |
| 1493 |
Err(EncoderStatus::Encoded) |
| 1494 |
} |
| 1495 |
} |
| 1496 |
|
| 1497 |
#[profiling::function] |
| 1498 |
pub fn receive_packet(&mut self) -> Result<Packet<T>, EncoderStatus> { |
| 1499 |
if self.done_processing() { |
| 1500 |
return Err(EncoderStatus::LimitReached); |
| 1501 |
} |
| 1502 |
|
| 1503 |
if self.needs_more_fi_lookahead() { |
| 1504 |
return Err(EncoderStatus::NeedMoreData); |
| 1505 |
} |
| 1506 |
|
| 1507 |
// Find the next output_frameno corresponding to a non-skipped frame. |
| 1508 |
self.output_frameno = self |
| 1509 |
.frame_data |
| 1510 |
.iter() |
| 1511 |
.skip_while(|(&output_frameno, _)| output_frameno < self.output_frameno) |
| 1512 |
.find(|(_, data)| data.is_some()) |
| 1513 |
.map(|(&output_frameno, _)| output_frameno) |
| 1514 |
.ok_or(EncoderStatus::NeedMoreData)?; // TODO: doesn't play well with the below check? |
| 1515 |
|
| 1516 |
let input_frameno = |
| 1517 |
self.frame_data[&self.output_frameno].as_ref().unwrap().fi.input_frameno; |
| 1518 |
if !self.needs_more_frames(input_frameno) { |
| 1519 |
return Err(EncoderStatus::LimitReached); |
| 1520 |
} |
| 1521 |
|
| 1522 |
if self.config.temporal_rdo() { |
| 1523 |
// Compute the block importances for the current output frame. |
| 1524 |
self.compute_block_importances(); |
| 1525 |
} |
| 1526 |
|
| 1527 |
let cur_output_frameno = self.output_frameno; |
| 1528 |
|
| 1529 |
let mut ret = self.encode_packet(cur_output_frameno); |
| 1530 |
|
| 1531 |
if let Ok(ref mut pkt) = ret { |
| 1532 |
self.garbage_collect(pkt.input_frameno); |
| 1533 |
pkt.opaque = self.opaque_q.remove(&pkt.input_frameno); |
| 1534 |
} |
| 1535 |
|
| 1536 |
ret |
| 1537 |
} |
| 1538 |
|
| 1539 |
fn finalize_packet( |
| 1540 |
&mut self, rec: Option<Arc<Frame<T>>>, source: Option<Arc<Frame<T>>>, |
| 1541 |
input_frameno: u64, frame_type: FrameType, qp: u8, |
| 1542 |
enc_stats: EncoderStats, |
| 1543 |
) -> Result<Packet<T>, EncoderStatus> { |
| 1544 |
let data = self.packet_data.clone(); |
| 1545 |
self.packet_data.clear(); |
| 1546 |
if write_temporal_delimiter(&mut self.packet_data).is_err() { |
| 1547 |
return Err(EncoderStatus::Failure); |
| 1548 |
} |
| 1549 |
|
| 1550 |
self.frames_processed += 1; |
| 1551 |
Ok(Packet { |
| 1552 |
data, |
| 1553 |
rec, |
| 1554 |
source, |
| 1555 |
input_frameno, |
| 1556 |
frame_type, |
| 1557 |
qp, |
| 1558 |
enc_stats, |
| 1559 |
opaque: None, |
| 1560 |
}) |
| 1561 |
} |
| 1562 |
|
| 1563 |
#[profiling::function] |
| 1564 |
fn garbage_collect(&mut self, cur_input_frameno: u64) { |
| 1565 |
if cur_input_frameno == 0 { |
| 1566 |
return; |
| 1567 |
} |
| 1568 |
let frame_q_start = self.frame_q.keys().next().cloned().unwrap_or(0); |
| 1569 |
for i in frame_q_start..cur_input_frameno { |
| 1570 |
self.frame_q.remove(&i); |
| 1571 |
} |
| 1572 |
|
| 1573 |
if self.output_frameno < 2 { |
| 1574 |
return; |
| 1575 |
} |
| 1576 |
let fi_start = self.frame_data.keys().next().cloned().unwrap_or(0); |
| 1577 |
for i in fi_start..(self.output_frameno - 1) { |
| 1578 |
self.frame_data.remove(&i); |
| 1579 |
self.gop_output_frameno_start.remove(&i); |
| 1580 |
self.gop_input_frameno_start.remove(&i); |
| 1581 |
} |
| 1582 |
} |
| 1583 |
|
| 1584 |
/// Counts the number of output frames of each subtype in the next |
| 1585 |
/// `reservoir_frame_delay` temporal units (needed for rate control). |
| 1586 |
/// Returns the number of output frames (excluding SEF frames) and output TUs |
| 1587 |
/// until the last keyframe in the next `reservoir_frame_delay` temporal units, |
| 1588 |
/// or the end of the interval, whichever comes first. |
| 1589 |
/// The former is needed because it indicates the number of rate estimates we |
| 1590 |
/// will make. |
| 1591 |
/// The latter is needed because it indicates the number of times new bitrate |
| 1592 |
/// is added to the buffer. |
| 1593 |
pub(crate) fn guess_frame_subtypes( |
| 1594 |
&self, nframes: &mut [i32; FRAME_NSUBTYPES + 1], |
| 1595 |
reservoir_frame_delay: i32, |
| 1596 |
) -> (i32, i32) { |
| 1597 |
for fti in 0..=FRAME_NSUBTYPES { |
| 1598 |
nframes[fti] = 0; |
| 1599 |
} |
| 1600 |
|
| 1601 |
// Two-pass calls this function before receive_packet(), and in particular |
| 1602 |
// before the very first send_frame(), when the following maps are empty. |
| 1603 |
// In this case, return 0 as the default value. |
| 1604 |
let mut prev_keyframe_input_frameno = *self |
| 1605 |
.gop_input_frameno_start |
| 1606 |
.get(&self.output_frameno) |
| 1607 |
.unwrap_or_else(|| { |
| 1608 |
assert!(self.output_frameno == 0); |
| 1609 |
&0 |
| 1610 |
}); |
| 1611 |
let mut prev_keyframe_output_frameno = *self |
| 1612 |
.gop_output_frameno_start |
| 1613 |
.get(&self.output_frameno) |
| 1614 |
.unwrap_or_else(|| { |
| 1615 |
assert!(self.output_frameno == 0); |
| 1616 |
&0 |
| 1617 |
}); |
| 1618 |
|
| 1619 |
let mut prev_keyframe_ntus = 0; |
| 1620 |
// Does not include SEF frames. |
| 1621 |
let mut prev_keyframe_nframes = 0; |
| 1622 |
let mut acc: [i32; FRAME_NSUBTYPES + 1] = [0; FRAME_NSUBTYPES + 1]; |
| 1623 |
// Updates the frame counts with the accumulated values when we hit a |
| 1624 |
// keyframe. |
| 1625 |
fn collect_counts( |
| 1626 |
nframes: &mut [i32; FRAME_NSUBTYPES + 1], |
| 1627 |
acc: &mut [i32; FRAME_NSUBTYPES + 1], |
| 1628 |
) { |
| 1629 |
for fti in 0..=FRAME_NSUBTYPES { |
| 1630 |
nframes[fti] += acc[fti]; |
| 1631 |
acc[fti] = 0; |
| 1632 |
} |
| 1633 |
acc[FRAME_SUBTYPE_I] += 1; |
| 1634 |
} |
| 1635 |
let mut output_frameno = self.output_frameno; |
| 1636 |
let mut ntus = 0; |
| 1637 |
// Does not include SEF frames. |
| 1638 |
let mut nframes_total = 0; |
| 1639 |
while ntus < reservoir_frame_delay { |
| 1640 |
let output_frameno_in_gop = |
| 1641 |
output_frameno - prev_keyframe_output_frameno; |
| 1642 |
let is_kf = |
| 1643 |
if let Some(Some(frame_data)) = self.frame_data.get(&output_frameno) { |
| 1644 |
if frame_data.fi.frame_type == FrameType::KEY { |
| 1645 |
prev_keyframe_input_frameno = frame_data.fi.input_frameno; |
| 1646 |
// We do not currently use forward keyframes, so they should always |
| 1647 |
// end the current TU (thus we always increment ntus below). |
| 1648 |
debug_assert!(frame_data.fi.show_frame); |
| 1649 |
true |
| 1650 |
} else { |
| 1651 |
false |
| 1652 |
} |
| 1653 |
} else { |
| 1654 |
// It is possible to be invoked for the first time from twopass_out() |
| 1655 |
// before receive_packet() is called, in which case frame_invariants |
| 1656 |
// will not be populated. |
| 1657 |
// Force the first frame in each GOP to be a keyframe in that case. |
| 1658 |
output_frameno_in_gop == 0 |
| 1659 |
}; |
| 1660 |
if is_kf { |
| 1661 |
collect_counts(nframes, &mut acc); |
| 1662 |
prev_keyframe_output_frameno = output_frameno; |
| 1663 |
prev_keyframe_ntus = ntus; |
| 1664 |
prev_keyframe_nframes = nframes_total; |
| 1665 |
output_frameno += 1; |
| 1666 |
ntus += 1; |
| 1667 |
nframes_total += 1; |
| 1668 |
continue; |
| 1669 |
} |
| 1670 |
let idx_in_group_output = |
| 1671 |
self.inter_cfg.get_idx_in_group_output(output_frameno_in_gop); |
| 1672 |
let input_frameno = prev_keyframe_input_frameno |
| 1673 |
+ self |
| 1674 |
.inter_cfg |
| 1675 |
.get_order_hint(output_frameno_in_gop, idx_in_group_output) |
| 1676 |
as u64; |
| 1677 |
// For rate control purposes, ignore any limit on frame count that has |
| 1678 |
// been set. |
| 1679 |
// We pretend that we will keep encoding frames forever to prevent the |
| 1680 |
// control loop from driving us into the rails as we come up against a |
| 1681 |
// hard stop (with no more chance to correct outstanding errors). |
| 1682 |
let next_keyframe_input_frameno = |
| 1683 |
self.next_keyframe_input_frameno(prev_keyframe_input_frameno, true); |
| 1684 |
// If we are re-ordering, we may skip some output frames in the final |
| 1685 |
// re-order group of the GOP. |
| 1686 |
if input_frameno >= next_keyframe_input_frameno { |
| 1687 |
// If we have encoded enough whole groups to reach the next keyframe, |
| 1688 |
// then start the next keyframe gop. |
| 1689 |
if 1 |
| 1690 |
+ (output_frameno - prev_keyframe_output_frameno) |
| 1691 |
/ self.inter_cfg.group_output_len |
| 1692 |
* self.inter_cfg.group_input_len |
| 1693 |
>= next_keyframe_input_frameno - prev_keyframe_input_frameno |
| 1694 |
{ |
| 1695 |
collect_counts(nframes, &mut acc); |
| 1696 |
prev_keyframe_input_frameno = input_frameno; |
| 1697 |
prev_keyframe_output_frameno = output_frameno; |
| 1698 |
prev_keyframe_ntus = ntus; |
| 1699 |
prev_keyframe_nframes = nframes_total; |
| 1700 |
// We do not currently use forward keyframes, so they should always |
| 1701 |
// end the current TU. |
| 1702 |
output_frameno += 1; |
| 1703 |
ntus += 1; |
| 1704 |
} |
| 1705 |
output_frameno += 1; |
| 1706 |
continue; |
| 1707 |
} |
| 1708 |
if self.inter_cfg.get_show_existing_frame(idx_in_group_output) { |
| 1709 |
acc[FRAME_SUBTYPE_SEF] += 1; |
| 1710 |
} else { |
| 1711 |
// TODO: Implement golden P-frames. |
| 1712 |
let fti = FRAME_SUBTYPE_P |
| 1713 |
+ (self.inter_cfg.get_level(idx_in_group_output) as usize); |
| 1714 |
acc[fti] += 1; |
| 1715 |
nframes_total += 1; |
| 1716 |
} |
| 1717 |
if self.inter_cfg.get_show_frame(idx_in_group_output) { |
| 1718 |
ntus += 1; |
| 1719 |
} |
| 1720 |
output_frameno += 1; |
| 1721 |
} |
| 1722 |
if prev_keyframe_output_frameno <= self.output_frameno { |
| 1723 |
// If there were no keyframes at all, or only the first frame was a |
| 1724 |
// keyframe, the accumulators never flushed and still contain counts for |
| 1725 |
// the entire buffer. |
| 1726 |
// In both cases, we return these counts. |
| 1727 |
collect_counts(nframes, &mut acc); |
| 1728 |
(nframes_total, ntus) |
| 1729 |
} else { |
| 1730 |
// Otherwise, we discard what remains in the accumulators as they contain |
| 1731 |
// the counts from and past the last keyframe. |
| 1732 |
(prev_keyframe_nframes, prev_keyframe_ntus) |
| 1733 |
} |
| 1734 |
} |
| 1735 |
} |