AOMedia AV1 Codec
temporal_filter.h
1/*
2 * Copyright (c) 2016, Alliance for Open Media. 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
12#ifndef AOM_AV1_ENCODER_TEMPORAL_FILTER_H_
13#define AOM_AV1_ENCODER_TEMPORAL_FILTER_H_
14
15#include <stdbool.h>
16
17#include "aom_util/aom_pthread.h"
18
19#ifdef __cplusplus
20extern "C" {
21#endif
23struct AV1_COMP;
24struct AV1EncoderConfig;
25struct ThreadData;
26// TODO(wtc): These two variables are only used in avx2, sse2, neon
27// implementations, where the block size is still hard coded to TF_BLOCK_SIZE.
28// This should be fixed to align with the c implementation.
29#define BH 32
30#define BW 32
31
32// Block size used in temporal filtering.
33#define TF_BLOCK_SIZE BLOCK_32X32
34
35// Window size for temporal filtering.
36#define TF_WINDOW_LENGTH 5
37
38// A constant number, sqrt(pi / 2), used for noise estimation.
39static const double SQRT_PI_BY_2 = 1.25331413732;
40
41// Hyper-parameters used to compute filtering weight. These hyper-parameters can
42// be tuned for a better performance.
43// 0. A scale factor used in temporal filtering to raise the filter weight from
44// `double` with range [0, 1] to `int` with range [0, 1000].
45#define TF_WEIGHT_SCALE 1000
46// 1. Weight factor used to balance the weighted-average between window error
47// and block error. The weight is for window error while the weight for block
48// error is always set as 1.
49#define TF_WINDOW_BLOCK_BALANCE_WEIGHT 5
50// 2. Threshold for using q to adjust the filtering weight. Concretely, when
51// using a small q (high bitrate), we would like to reduce the filtering
52// strength such that more detailed information can be preserved. Hence, when
53// q is smaller than this threshold, we will adjust the filtering weight
54// based on the q-value.
55#define TF_Q_DECAY_THRESHOLD 20
56// 3. Normalization factor used to normalize the motion search error. Since the
57// motion search error can be large and uncontrollable, we will simply
58// normalize it before using it to compute the filtering weight.
59#define TF_SEARCH_ERROR_NORM_WEIGHT 20
60// 4. Threshold for using `arnr_strength` to adjust the filtering strength.
61// Concretely, users can use `arnr_strength` arguments to control the
62// strength of temporal filtering. When `arnr_strength` is small enough (
63// i.e., smaller than this threshold), we will adjust the filtering weight
64// based on the strength value.
65#define TF_STRENGTH_THRESHOLD 4
66// 5. Threshold for using motion search distance to adjust the filtering weight.
67// Concretely, larger motion search vector leads to a higher probability of
68// unreliable search. Hence, we would like to reduce the filtering strength
69// when the distance is large enough. Considering that the distance actually
70// relies on the frame size, this threshold is also a resolution-based
71// threshold. Taking 720p videos as an instance, if this field equals to 0.1,
72// then the actual threshold will be 720 * 0.1 = 72. Similarly, the threshold
73// for 360p videos will be 360 * 0.1 = 36.
74#define TF_SEARCH_DISTANCE_THRESHOLD 0.1
75// 6. Threshold to identify if the q is in a relative high range.
76// Above this cutoff q, a stronger filtering is applied.
77// For a high q, the quantization throws away more information, and thus a
78// stronger filtering is less likely to distort the encoded quality, while a
79// stronger filtering could reduce bit rates.
80// Ror a low q, more details are expected to be retained. Filtering is thus
81// more conservative.
82#define TF_QINDEX_CUTOFF 128
83
84#define NOISE_ESTIMATION_EDGE_THRESHOLD 50
85
86// Sum and SSE source vs filtered frame difference returned by
87// temporal filter.
88typedef struct {
89 int64_t sum;
90 int64_t sse;
91} FRAME_DIFF;
92
98typedef struct {
102 YV12_BUFFER_CONFIG *frames[MAX_LAG_BUFFERS];
107
112
124 struct scale_factors sf;
128 double noise_levels[MAX_MB_PLANE];
150
156#define TF_INFO_BUF_COUNT 2
157
161typedef struct TEMPORAL_FILTER_INFO {
172 YV12_BUFFER_CONFIG tf_buf[TF_INFO_BUF_COUNT];
173
184 FRAME_DIFF frame_diff[TF_INFO_BUF_COUNT];
188 int tf_buf_gf_index[TF_INFO_BUF_COUNT];
192 int tf_buf_display_index_offset[TF_INFO_BUF_COUNT];
196 int tf_buf_valid[TF_INFO_BUF_COUNT];
198
204int av1_is_temporal_filter_on(const struct AV1EncoderConfig *oxcf);
205
212bool av1_tf_info_alloc(TEMPORAL_FILTER_INFO *tf_info,
213 const struct AV1_COMP *cpi);
214
218void av1_tf_info_free(TEMPORAL_FILTER_INFO *tf_info);
219
223void av1_tf_info_reset(TEMPORAL_FILTER_INFO *tf_info);
224
230void av1_tf_info_filtering(TEMPORAL_FILTER_INFO *tf_info, struct AV1_COMP *cpi,
231 const GF_GROUP *gf_group);
232
239YV12_BUFFER_CONFIG *av1_tf_info_get_filtered_buf(TEMPORAL_FILTER_INFO *tf_info,
240 int gf_index,
241 FRAME_DIFF *frame_diff);
242
245// Data related to temporal filtering.
246typedef struct {
247 // Source vs filtered frame error.
248 FRAME_DIFF diff;
249 // Pointer to temporary block info used to store state in temporal filtering
250 // process.
251 MB_MODE_INFO *tmp_mbmi;
252 // Pointer to accumulator buffer used in temporal filtering process.
253 uint32_t *accum;
254 // Pointer to count buffer used in temporal filtering process.
255 uint16_t *count;
256 // Pointer to predictor used in temporal filtering process.
257 uint8_t *pred;
258} TemporalFilterData;
259
260// Data related to temporal filter multi-thread synchronization.
261typedef struct {
262#if CONFIG_MULTITHREAD
263 // Mutex lock used for dispatching jobs.
264 pthread_mutex_t *mutex_;
265#endif // CONFIG_MULTITHREAD
266 // Next temporal filter block row to be filtered.
267 int next_tf_row;
268 // Initialized to false, set to true by the worker thread that encounters an
269 // error in order to abort the processing of other worker threads.
270 bool tf_mt_exit;
271} AV1TemporalFilterSync;
272
273// Estimates noise level from a given frame using a single plane (Y, U, or V).
274// This is an adaptation of the mehtod in the following paper:
275// Shen-Chuan Tai, Shih-Ming Yang, "A fast method for image noise
276// estimation using Laplacian operator and adaptive edge detection",
277// Proc. 3rd International Symposium on Communications, Control and
278// Signal Processing, 2008, St Julians, Malta.
279// Inputs:
280// frame: Pointer to the frame to estimate noise level from.
281// noise_level: Pointer to store the estimated noise.
282// plane_from: Index of the starting plane used for noise estimation.
283// Commonly, 0 for Y-plane, 1 for U-plane, and 2 for V-plane.
284// plane_to: Index of the end plane used for noise estimation.
285// bit_depth: Actual bit-depth instead of the encoding bit-depth of the frame.
286// edge_thresh: Edge threshold.
287void av1_estimate_noise_level(const YV12_BUFFER_CONFIG *frame,
288 double *noise_level, int plane_from, int plane_to,
289 int bit_depth, int edge_thresh);
303void av1_tf_do_filtering_row(struct AV1_COMP *cpi, struct ThreadData *td,
304 int mb_row);
305
331 const int filter_frame_lookahead_idx,
332 int gf_frame_index, FRAME_DIFF *frame_diff,
333 YV12_BUFFER_CONFIG *output_frame);
334
350 const FRAME_DIFF *frame_diff, int q_index,
351 aom_bit_depth_t bit_depth);
352
354// Allocates memory for members of TemporalFilterData.
355// Inputs:
356// tf_data: Pointer to the structure containing temporal filter related data.
357// num_pels: Number of pixels in the block across all planes.
358// is_high_bitdepth: Whether the frame is high-bitdepth or not.
359// Returns:
360// True if allocation is successful and false otherwise.
361static inline bool tf_alloc_and_reset_data(TemporalFilterData *tf_data,
362 int num_pels, int is_high_bitdepth) {
363 tf_data->tmp_mbmi = (MB_MODE_INFO *)aom_calloc(1, sizeof(*tf_data->tmp_mbmi));
364 tf_data->accum =
365 (uint32_t *)aom_memalign(16, num_pels * sizeof(*tf_data->accum));
366 tf_data->count =
367 (uint16_t *)aom_memalign(16, num_pels * sizeof(*tf_data->count));
368 if (is_high_bitdepth)
369 tf_data->pred = CONVERT_TO_BYTEPTR(
370 aom_memalign(32, num_pels * 2 * sizeof(*tf_data->pred)));
371 else
372 tf_data->pred =
373 (uint8_t *)aom_memalign(32, num_pels * sizeof(*tf_data->pred));
374 // In case of an allocation failure, other successfully allocated buffers will
375 // be freed by the tf_dealloc_data() call in encoder_destroy().
376 if (!(tf_data->tmp_mbmi && tf_data->accum && tf_data->count && tf_data->pred))
377 return false;
378 memset(&tf_data->diff, 0, sizeof(tf_data->diff));
379 return true;
380}
381
382// Setup macroblockd params for temporal filtering process.
383// Inputs:
384// mbd: Pointer to the block for filtering.
385// tf_data: Pointer to the structure containing temporal filter related data.
386// scale: Scaling factor.
387// Returns:
388// Nothing will be returned. Contents of mbd will be modified.
389static inline void tf_setup_macroblockd(MACROBLOCKD *mbd,
390 TemporalFilterData *tf_data,
391 const struct scale_factors *scale) {
392 mbd->block_ref_scale_factors[0] = scale;
393 mbd->block_ref_scale_factors[1] = scale;
394 mbd->mi = &tf_data->tmp_mbmi;
395 mbd->mi[0]->motion_mode = SIMPLE_TRANSLATION;
396}
397
398// Deallocates the memory allocated for members of TemporalFilterData.
399// Inputs:
400// tf_data: Pointer to the structure containing temporal filter related data.
401// is_high_bitdepth: Whether the frame is high-bitdepth or not.
402// Returns:
403// Nothing will be returned.
404static inline void tf_dealloc_data(TemporalFilterData *tf_data,
405 int is_high_bitdepth) {
406 if (is_high_bitdepth)
407 tf_data->pred = (uint8_t *)CONVERT_TO_SHORTPTR(tf_data->pred);
408 aom_free(tf_data->tmp_mbmi);
409 tf_data->tmp_mbmi = NULL;
410 aom_free(tf_data->accum);
411 tf_data->accum = NULL;
412 aom_free(tf_data->count);
413 tf_data->count = NULL;
414 aom_free(tf_data->pred);
415 tf_data->pred = NULL;
416}
417
418// Saves the state prior to temporal filter process.
419// Inputs:
420// mbd: Pointer to the block for filtering.
421// input_mbmi: Backup block info to save input state.
422// input_buffer: Backup buffer pointer to save input state.
423// num_planes: Number of planes.
424// Returns:
425// Nothing will be returned. Contents of input_mbmi and input_buffer will be
426// modified.
427static inline void tf_save_state(MACROBLOCKD *mbd, MB_MODE_INFO ***input_mbmi,
428 uint8_t **input_buffer, int num_planes) {
429 for (int i = 0; i < num_planes; i++) {
430 input_buffer[i] = mbd->plane[i].pre[0].buf;
431 }
432 *input_mbmi = mbd->mi;
433}
434
435// Restores the initial state after temporal filter process.
436// Inputs:
437// mbd: Pointer to the block for filtering.
438// input_mbmi: Backup block info from where input state is restored.
439// input_buffer: Backup buffer pointer from where input state is restored.
440// num_planes: Number of planes.
441// Returns:
442// Nothing will be returned. Contents of mbd will be modified.
443static inline void tf_restore_state(MACROBLOCKD *mbd, MB_MODE_INFO **input_mbmi,
444 uint8_t **input_buffer, int num_planes) {
445 for (int i = 0; i < num_planes; i++) {
446 mbd->plane[i].pre[0].buf = input_buffer[i];
447 }
448 mbd->mi = input_mbmi;
449}
450
452#ifdef __cplusplus
453} // extern "C"
454#endif
455
456#endif // AOM_AV1_ENCODER_TEMPORAL_FILTER_H_
enum aom_bit_depth aom_bit_depth_t
Bit depth for codecThis enumeration determines the bit depth of the codec.
int av1_check_show_filtered_frame(const YV12_BUFFER_CONFIG *frame, const FRAME_DIFF *frame_diff, int q_index, aom_bit_depth_t bit_depth)
Check whether a filtered frame can be show directly.
void av1_temporal_filter(struct AV1_COMP *cpi, const int filter_frame_lookahead_idx, int gf_frame_index, FRAME_DIFF *frame_diff, YV12_BUFFER_CONFIG *output_frame)
Performs temporal filtering if needed on a source frame. For example to create a filtered alternate r...
void av1_tf_do_filtering_row(struct AV1_COMP *cpi, struct ThreadData *td, int mb_row)
Does temporal filter for a given macroblock row.
Definition temporal_filter.c:860
Main encoder configuration data structure.
Definition encoder.h:923
Top level encoder structure.
Definition encoder.h:2873
Data related to the current GF/ARF group and the individual frames within the group.
Definition firstpass.h:339
Stores the prediction/txfm mode of the current coding block.
Definition blockd.h:222
MOTION_MODE motion_mode
The motion mode used by the inter prediction.
Definition blockd.h:250
Temporal filter info for a gop.
Definition temporal_filter.h:161
int tf_buf_display_index_offset[2]
Definition temporal_filter.h:192
YV12_BUFFER_CONFIG tf_buf[2]
Definition temporal_filter.h:172
int tf_buf_gf_index[2]
Definition temporal_filter.h:188
FRAME_DIFF frame_diff[2]
Definition temporal_filter.h:184
int is_temporal_filter_on
Definition temporal_filter.h:167
int tf_buf_valid[2]
Definition temporal_filter.h:196
YV12_BUFFER_CONFIG tf_buf_second_arf
Definition temporal_filter.h:180
Parameters related to temporal filtering.
Definition temporal_filter.h:98
YV12_BUFFER_CONFIG * output_frame
Definition temporal_filter.h:111
int q_factor
Definition temporal_filter.h:148
int num_pels
Definition temporal_filter.h:132
int num_frames
Definition temporal_filter.h:106
int compute_frame_diff
Definition temporal_filter.h:120
int mb_rows
Definition temporal_filter.h:136
int mb_cols
Definition temporal_filter.h:140
int is_highbitdepth
Definition temporal_filter.h:144
int filter_frame_idx
Definition temporal_filter.h:116
Variables related to current coding block.
Definition blockd.h:570
struct macroblockd_plane plane[3]
Definition blockd.h:606
const struct scale_factors * block_ref_scale_factors[2]
Definition blockd.h:687
MB_MODE_INFO ** mi
Definition blockd.h:617
YV12 frame buffer data structure.
Definition yv12config.h:46