temporal_filter.h

/*
 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
 *
 * This source code is subject to the terms of the BSD 2 Clause License and
 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
 * was not distributed with this source code in the LICENSE file, you can
 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
 * Media Patent License 1.0 was not distributed with this source code in the
 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
 */
 
#ifndef AOM_AV1_ENCODER_TEMPORAL_FILTER_H_
#define AOM_AV1_ENCODER_TEMPORAL_FILTER_H_
 
#include <stdbool.h>
 
#ifdef __cplusplus
extern "C" {
#endif
 
struct AV1_COMP;
struct AV1EncoderConfig;
struct ThreadData;
// TODO(any): These two variables are only used in avx2, sse2, sse4
// implementations, where the block size is still hard coded. This should be
// fixed to align with the c implementation.
#define BH 32
#define BW 32
 
// Block size used in temporal filtering.
#define TF_BLOCK_SIZE BLOCK_32X32
 
// Window size for temporal filtering.
#define TF_WINDOW_LENGTH 5
 
// Hyper-parameters used to compute filtering weight. These hyper-parameters can
// be tuned for a better performance.
// 0. A scale factor used in temporal filtering to raise the filter weight from
//    `double` with range [0, 1] to `int` with range [0, 1000].
#define TF_WEIGHT_SCALE 1000
// 1. Weight factor used to balance the weighted-average between window error
//    and block error. The weight is for window error while the weight for block
//    error is always set as 1.
#define TF_WINDOW_BLOCK_BALANCE_WEIGHT 5
// 2. Threshold for using q to adjust the filtering weight. Concretely, when
//    using a small q (high bitrate), we would like to reduce the filtering
//    strength such that more detailed information can be preserved. Hence, when
//    q is smaller than this threshold, we will adjust the filtering weight
//    based on the q-value.
#define TF_Q_DECAY_THRESHOLD 20
// 3. Normalization factor used to normalize the motion search error. Since the
//    motion search error can be large and uncontrollable, we will simply
//    normalize it before using it to compute the filtering weight.
#define TF_SEARCH_ERROR_NORM_WEIGHT 20
// 4. Threshold for using `arnr_strength` to adjust the filtering strength.
//    Concretely, users can use `arnr_strength` arguments to control the
//    strength of temporal filtering. When `arnr_strength` is small enough (
//    i.e., smaller than this threshold), we will adjust the filtering weight
//    based on the strength value.
#define TF_STRENGTH_THRESHOLD 4
// 5. Threshold for using motion search distance to adjust the filtering weight.
//    Concretely, larger motion search vector leads to a higher probability of
//    unreliable search. Hence, we would like to reduce the filtering strength
//    when the distance is large enough. Considering that the distance actually
//    relies on the frame size, this threshold is also a resolution-based
//    threshold. Taking 720p videos as an instance, if this field equals to 0.1,
//    then the actual threshold will be 720 * 0.1 = 72. Similarly, the threshold
//    for 360p videos will be 360 * 0.1 = 36.
#define TF_SEARCH_DISTANCE_THRESHOLD 0.1
// 6. Threshold to identify if the q is in a relative high range.
//    Above this cutoff q, a stronger filtering is applied.
//    For a high q, the quantization throws away more information, and thus a
//    stronger filtering is less likely to distort the encoded quality, while a
//    stronger filtering could reduce bit rates.
//    Ror a low q, more details are expected to be retained. Filtering is thus
//    more conservative.
#define TF_QINDEX_CUTOFF 128
 
#define NOISE_ESTIMATION_EDGE_THRESHOLD 50
 
// Sum and SSE source vs filtered frame difference returned by
// temporal filter.
typedef struct {
  int64_t sum;
  int64_t sse;
} FRAME_DIFF;
 
typedef struct {
  YV12_BUFFER_CONFIG *frames[MAX_LAG_BUFFERS];
  int num_frames;
 
  YV12_BUFFER_CONFIG *output_frame;
 
  int filter_frame_idx;
  int compute_frame_diff;
  struct scale_factors sf;
  double noise_levels[MAX_MB_PLANE];
  int num_pels;
  int mb_rows;
  int mb_cols;
  int is_highbitdepth;
  int q_factor;
} TemporalFilterCtx;
 
#define TF_INFO_BUF_COUNT 2
 
typedef struct TEMPORAL_FILTER_INFO {
  int is_temporal_filter_on;
  YV12_BUFFER_CONFIG tf_buf[TF_INFO_BUF_COUNT];
 
  YV12_BUFFER_CONFIG tf_buf_second_arf;
  FRAME_DIFF frame_diff[TF_INFO_BUF_COUNT];
  int tf_buf_gf_index[TF_INFO_BUF_COUNT];
  int tf_buf_display_index_offset[TF_INFO_BUF_COUNT];
  int tf_buf_valid[TF_INFO_BUF_COUNT];
} TEMPORAL_FILTER_INFO;
 
int av1_is_temporal_filter_on(const struct AV1EncoderConfig *oxcf);
 
void av1_tf_info_alloc(TEMPORAL_FILTER_INFO *tf_info, struct AV1_COMP *cpi);
 
void av1_tf_info_free(TEMPORAL_FILTER_INFO *tf_info);
 
void av1_tf_info_reset(TEMPORAL_FILTER_INFO *tf_info);
 
void av1_tf_info_filtering(TEMPORAL_FILTER_INFO *tf_info, struct AV1_COMP *cpi,
                           const GF_GROUP *gf_group);
 
YV12_BUFFER_CONFIG *av1_tf_info_get_filtered_buf(TEMPORAL_FILTER_INFO *tf_info,
                                                 int gf_index,
                                                 FRAME_DIFF *frame_diff);
 
// Data related to temporal filtering.
typedef struct {
  // Source vs filtered frame error.
  FRAME_DIFF diff;
  // Pointer to temporary block info used to store state in temporal filtering
  // process.
  MB_MODE_INFO *tmp_mbmi;
  // Pointer to accumulator buffer used in temporal filtering process.
  uint32_t *accum;
  // Pointer to count buffer used in temporal filtering process.
  uint16_t *count;
  // Pointer to predictor used in temporal filtering process.
  uint8_t *pred;
} TemporalFilterData;
 
// Data related to temporal filter multi-thread synchronization.
typedef struct {
#if CONFIG_MULTITHREAD
  // Mutex lock used for dispatching jobs.
  pthread_mutex_t *mutex_;
#endif  // CONFIG_MULTITHREAD
  // Next temporal filter block row to be filtered.
  int next_tf_row;
} AV1TemporalFilterSync;
 
// Estimates noise level from a given frame using a single plane (Y, U, or V).
// This is an adaptation of the mehtod in the following paper:
// Shen-Chuan Tai, Shih-Ming Yang, "A fast method for image noise
// estimation using Laplacian operator and adaptive edge detection",
// Proc. 3rd International Symposium on Communications, Control and
// Signal Processing, 2008, St Julians, Malta.
// Inputs:
//   frame: Pointer to the frame to estimate noise level from.
//   plane: Index of the plane used for noise estimation. Commonly, 0 for
//          Y-plane, 1 for U-plane, and 2 for V-plane.
//   bit_depth: Actual bit-depth instead of the encoding bit-depth of the frame.
// Returns:
//   The estimated noise, or -1.0 if there are too few smooth pixels.
double av1_estimate_noise_from_single_plane(const YV12_BUFFER_CONFIG *frame,
                                            const int plane,
                                            const int bit_depth,
                                            const int edge_thresh);
void av1_tf_do_filtering_row(struct AV1_COMP *cpi, struct ThreadData *td,
                             int mb_row);
 
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);
 
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);
 
// Helper function to get `q` used for encoding.
int av1_get_q(const struct AV1_COMP *cpi);
 
// Allocates memory for members of TemporalFilterData.
// Inputs:
//   tf_data: Pointer to the structure containing temporal filter related data.
//   num_pels: Number of pixels in the block across all planes.
//   is_high_bitdepth: Whether the frame is high-bitdepth or not.
// Returns:
//   Nothing will be returned. But the contents of tf_data will be modified.
static AOM_INLINE bool tf_alloc_and_reset_data(TemporalFilterData *tf_data,
                                               int num_pels,
                                               int is_high_bitdepth) {
  tf_data->tmp_mbmi = (MB_MODE_INFO *)malloc(sizeof(*tf_data->tmp_mbmi));
  memset(tf_data->tmp_mbmi, 0, sizeof(*tf_data->tmp_mbmi));
  tf_data->accum =
      (uint32_t *)aom_memalign(16, num_pels * sizeof(*tf_data->accum));
  tf_data->count =
      (uint16_t *)aom_memalign(16, num_pels * sizeof(*tf_data->count));
  memset(&tf_data->diff, 0, sizeof(tf_data->diff));
  if (is_high_bitdepth)
    tf_data->pred = CONVERT_TO_BYTEPTR(
        aom_memalign(32, num_pels * 2 * sizeof(*tf_data->pred)));
  else
    tf_data->pred =
        (uint8_t *)aom_memalign(32, num_pels * sizeof(*tf_data->pred));
  if (!(tf_data->accum && tf_data->count && tf_data->pred)) {
    aom_free(tf_data->accum);
    aom_free(tf_data->count);
    aom_free(tf_data->pred);
    return false;
  }
  return true;
}
 
// Setup macroblockd params for temporal filtering process.
// Inputs:
//   mbd: Pointer to the block for filtering.
//   tf_data: Pointer to the structure containing temporal filter related data.
//   scale: Scaling factor.
// Returns:
//   Nothing will be returned. Contents of mbd will be modified.
static AOM_INLINE void tf_setup_macroblockd(MACROBLOCKD *mbd,
                                            TemporalFilterData *tf_data,
                                            const struct scale_factors *scale) {
  mbd->block_ref_scale_factors[0] = scale;
  mbd->block_ref_scale_factors[1] = scale;
  mbd->mi = &tf_data->tmp_mbmi;
  mbd->mi[0]->motion_mode = SIMPLE_TRANSLATION;
}
 
// Deallocates the memory allocated for members of TemporalFilterData.
// Inputs:
//   tf_data: Pointer to the structure containing temporal filter related data.
//   is_high_bitdepth: Whether the frame is high-bitdepth or not.
// Returns:
//   Nothing will be returned.
static AOM_INLINE void tf_dealloc_data(TemporalFilterData *tf_data,
                                       int is_high_bitdepth) {
  if (is_high_bitdepth)
    tf_data->pred = (uint8_t *)CONVERT_TO_SHORTPTR(tf_data->pred);
  free(tf_data->tmp_mbmi);
  aom_free(tf_data->accum);
  aom_free(tf_data->count);
  aom_free(tf_data->pred);
}
 
// Saves the state prior to temporal filter process.
// Inputs:
//   mbd: Pointer to the block for filtering.
//   input_mbmi: Backup block info to save input state.
//   input_buffer: Backup buffer pointer to save input state.
//   num_planes: Number of planes.
// Returns:
//   Nothing will be returned. Contents of input_mbmi and input_buffer will be
//   modified.
static INLINE void tf_save_state(MACROBLOCKD *mbd, MB_MODE_INFO ***input_mbmi,
                                 uint8_t **input_buffer, int num_planes) {
  for (int i = 0; i < num_planes; i++) {
    input_buffer[i] = mbd->plane[i].pre[0].buf;
  }
  *input_mbmi = mbd->mi;
}
 
// Restores the initial state after temporal filter process.
// Inputs:
//   mbd: Pointer to the block for filtering.
//   input_mbmi: Backup block info from where input state is restored.
//   input_buffer: Backup buffer pointer from where input state is restored.
//   num_planes: Number of planes.
// Returns:
//   Nothing will be returned. Contents of mbd will be modified.
static INLINE void tf_restore_state(MACROBLOCKD *mbd, MB_MODE_INFO **input_mbmi,
                                    uint8_t **input_buffer, int num_planes) {
  for (int i = 0; i < num_planes; i++) {
    mbd->plane[i].pre[0].buf = input_buffer[i];
  }
  mbd->mi = input_mbmi;
}
 
#ifdef __cplusplus
}  // extern "C"
#endif
 
#endif  // AOM_AV1_ENCODER_TEMPORAL_FILTER_H_