blob: 967e842c71a89d84c5778ebd598c324405668c37 [file] [log] [blame]
/*
* 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.
*/
#include <assert.h>
#include <limits.h>
#include <math.h>
#include <stdio.h>
#include "config/aom_dsp_rtcd.h"
#include "config/av1_rtcd.h"
#include "aom_dsp/aom_dsp_common.h"
#include "aom_dsp/txfm_common.h"
#include "aom_ports/mem.h"
#include "av1/common/blockd.h"
#include "av1/common/mvref_common.h"
#include "av1/common/pred_common.h"
#include "av1/common/reconinter.h"
#include "av1/common/reconintra.h"
#include "av1/encoder/encodemv.h"
#include "av1/encoder/encoder.h"
#include "av1/encoder/intra_mode_search.h"
#include "av1/encoder/model_rd.h"
#include "av1/encoder/motion_search_facade.h"
#include "av1/encoder/nonrd_opt.h"
#include "av1/encoder/rdopt.h"
#include "av1/encoder/reconinter_enc.h"
#include "av1/encoder/var_based_part.h"
#define CALC_BIASED_RDCOST(rdcost) (7 * (rdcost) >> 3)
extern int g_pick_inter_mode_cnt;
/*!\cond */
typedef struct {
uint8_t *data;
int stride;
int in_use;
} PRED_BUFFER;
typedef struct {
PRED_BUFFER *best_pred;
PREDICTION_MODE best_mode;
TX_SIZE best_tx_size;
TX_TYPE tx_type;
MV_REFERENCE_FRAME best_ref_frame;
MV_REFERENCE_FRAME best_second_ref_frame;
uint8_t best_mode_skip_txfm;
uint8_t best_mode_initial_skip_flag;
int_interpfilters best_pred_filter;
MOTION_MODE best_motion_mode;
WarpedMotionParams wm_params;
int num_proj_ref;
uint8_t blk_skip[MAX_MIB_SIZE * MAX_MIB_SIZE / 4];
PALETTE_MODE_INFO pmi;
int64_t best_sse;
} BEST_PICKMODE;
typedef struct {
MV_REFERENCE_FRAME ref_frame;
PREDICTION_MODE pred_mode;
} REF_MODE;
typedef struct {
MV_REFERENCE_FRAME ref_frame[2];
PREDICTION_MODE pred_mode;
} COMP_REF_MODE;
typedef struct {
InterpFilter filter_x;
InterpFilter filter_y;
} INTER_FILTER;
/*!\brief Structure to store parameters and statistics used in non-rd inter mode
* evaluation.
*/
typedef struct {
BEST_PICKMODE best_pickmode;
RD_STATS this_rdc;
RD_STATS best_rdc;
int64_t uv_dist[RTC_INTER_MODES][REF_FRAMES];
struct buf_2d yv12_mb[REF_FRAMES][MAX_MB_PLANE];
unsigned int vars[RTC_INTER_MODES][REF_FRAMES];
unsigned int ref_costs_single[REF_FRAMES];
int_mv frame_mv[MB_MODE_COUNT][REF_FRAMES];
int_mv frame_mv_best[MB_MODE_COUNT][REF_FRAMES];
int single_inter_mode_costs[RTC_INTER_MODES][REF_FRAMES];
int use_ref_frame_mask[REF_FRAMES];
uint8_t mode_checked[MB_MODE_COUNT][REF_FRAMES];
} InterModeSearchStateNonrd;
/*!\endcond */
#define NUM_COMP_INTER_MODES_RT (6)
#define NUM_INTER_MODES 12
// GLOBALMV in the set below is in fact ZEROMV as we don't do global ME in RT
// mode
static const REF_MODE ref_mode_set[NUM_INTER_MODES] = {
{ LAST_FRAME, NEARESTMV }, { LAST_FRAME, NEARMV },
{ LAST_FRAME, GLOBALMV }, { LAST_FRAME, NEWMV },
{ GOLDEN_FRAME, NEARESTMV }, { GOLDEN_FRAME, NEARMV },
{ GOLDEN_FRAME, GLOBALMV }, { GOLDEN_FRAME, NEWMV },
{ ALTREF_FRAME, NEARESTMV }, { ALTREF_FRAME, NEARMV },
{ ALTREF_FRAME, GLOBALMV }, { ALTREF_FRAME, NEWMV },
};
static const COMP_REF_MODE comp_ref_mode_set[NUM_COMP_INTER_MODES_RT] = {
{ { LAST_FRAME, GOLDEN_FRAME }, GLOBAL_GLOBALMV },
{ { LAST_FRAME, GOLDEN_FRAME }, NEAREST_NEARESTMV },
{ { LAST_FRAME, LAST2_FRAME }, GLOBAL_GLOBALMV },
{ { LAST_FRAME, LAST2_FRAME }, NEAREST_NEARESTMV },
{ { LAST_FRAME, ALTREF_FRAME }, GLOBAL_GLOBALMV },
{ { LAST_FRAME, ALTREF_FRAME }, NEAREST_NEARESTMV },
};
static const INTER_FILTER filters_ref_set[9] = {
{ EIGHTTAP_REGULAR, EIGHTTAP_REGULAR }, { EIGHTTAP_SMOOTH, EIGHTTAP_SMOOTH },
{ EIGHTTAP_REGULAR, EIGHTTAP_SMOOTH }, { EIGHTTAP_SMOOTH, EIGHTTAP_REGULAR },
{ MULTITAP_SHARP, MULTITAP_SHARP }, { EIGHTTAP_REGULAR, MULTITAP_SHARP },
{ MULTITAP_SHARP, EIGHTTAP_REGULAR }, { EIGHTTAP_SMOOTH, MULTITAP_SHARP },
{ MULTITAP_SHARP, EIGHTTAP_SMOOTH }
};
enum {
// INTER_ALL = (1 << NEARESTMV) | (1 << NEARMV) | (1 << NEWMV),
INTER_NEAREST = (1 << NEARESTMV),
INTER_NEAREST_NEW = (1 << NEARESTMV) | (1 << NEWMV),
INTER_NEAREST_NEAR = (1 << NEARESTMV) | (1 << NEARMV),
INTER_NEAR_NEW = (1 << NEARMV) | (1 << NEWMV),
};
// The original scan order (default_scan_8x8) is modified according to the extra
// transpose in hadamard c implementation, i.e., aom_hadamard_lp_8x8_c and
// aom_hadamard_8x8_c.
DECLARE_ALIGNED(16, static const int16_t, default_scan_8x8_transpose[64]) = {
0, 8, 1, 2, 9, 16, 24, 17, 10, 3, 4, 11, 18, 25, 32, 40,
33, 26, 19, 12, 5, 6, 13, 20, 27, 34, 41, 48, 56, 49, 42, 35,
28, 21, 14, 7, 15, 22, 29, 36, 43, 50, 57, 58, 51, 44, 37, 30,
23, 31, 38, 45, 52, 59, 60, 53, 46, 39, 47, 54, 61, 62, 55, 63
};
// The original scan order (av1_default_iscan_8x8) is modified to match
// hadamard AVX2 implementation, i.e., aom_hadamard_lp_8x8_avx2 and
// aom_hadamard_8x8_avx2. Since hadamard AVX2 implementation will modify the
// order of coefficients, such that the normal scan order is no longer
// guaranteed to scan low coefficients first, therefore we modify the scan order
// accordingly.
// Note that this one has to be used together with default_scan_8x8_transpose.
DECLARE_ALIGNED(16, static const int16_t,
av1_default_iscan_8x8_transpose[64]) = {
0, 2, 3, 9, 10, 20, 21, 35, 1, 4, 8, 11, 19, 22, 34, 36,
5, 7, 12, 18, 23, 33, 37, 48, 6, 13, 17, 24, 32, 38, 47, 49,
14, 16, 25, 31, 39, 46, 50, 57, 15, 26, 30, 40, 45, 51, 56, 58,
27, 29, 41, 44, 52, 55, 59, 62, 28, 42, 43, 53, 54, 60, 61, 63
};
// The original scan order (default_scan_16x16) is modified according to the
// extra transpose in hadamard c implementation in lp case, i.e.,
// aom_hadamard_lp_16x16_c.
DECLARE_ALIGNED(16, static const int16_t,
default_scan_lp_16x16_transpose[256]) = {
0, 8, 2, 4, 10, 16, 24, 18, 12, 6, 64, 14, 20, 26, 32,
40, 34, 28, 22, 72, 66, 68, 74, 80, 30, 36, 42, 48, 56, 50,
44, 38, 88, 82, 76, 70, 128, 78, 84, 90, 96, 46, 52, 58, 1,
9, 3, 60, 54, 104, 98, 92, 86, 136, 130, 132, 138, 144, 94, 100,
106, 112, 62, 5, 11, 17, 25, 19, 13, 7, 120, 114, 108, 102, 152,
146, 140, 134, 192, 142, 148, 154, 160, 110, 116, 122, 65, 15, 21, 27,
33, 41, 35, 29, 23, 73, 67, 124, 118, 168, 162, 156, 150, 200, 194,
196, 202, 208, 158, 164, 170, 176, 126, 69, 75, 81, 31, 37, 43, 49,
57, 51, 45, 39, 89, 83, 77, 71, 184, 178, 172, 166, 216, 210, 204,
198, 206, 212, 218, 224, 174, 180, 186, 129, 79, 85, 91, 97, 47, 53,
59, 61, 55, 105, 99, 93, 87, 137, 131, 188, 182, 232, 226, 220, 214,
222, 228, 234, 240, 190, 133, 139, 145, 95, 101, 107, 113, 63, 121, 115,
109, 103, 153, 147, 141, 135, 248, 242, 236, 230, 238, 244, 250, 193, 143,
149, 155, 161, 111, 117, 123, 125, 119, 169, 163, 157, 151, 201, 195, 252,
246, 254, 197, 203, 209, 159, 165, 171, 177, 127, 185, 179, 173, 167, 217,
211, 205, 199, 207, 213, 219, 225, 175, 181, 187, 189, 183, 233, 227, 221,
215, 223, 229, 235, 241, 191, 249, 243, 237, 231, 239, 245, 251, 253, 247,
255
};
#if CONFIG_AV1_HIGHBITDEPTH
// The original scan order (default_scan_16x16) is modified according to the
// extra shift in hadamard c implementation in fp case, i.e.,
// aom_hadamard_16x16_c. Note that 16x16 lp and fp hadamard generate different
// outputs, so we handle them separately.
DECLARE_ALIGNED(16, static const int16_t,
default_scan_fp_16x16_transpose[256]) = {
0, 4, 2, 8, 6, 16, 20, 18, 12, 10, 64, 14, 24, 22, 32,
36, 34, 28, 26, 68, 66, 72, 70, 80, 30, 40, 38, 48, 52, 50,
44, 42, 84, 82, 76, 74, 128, 78, 88, 86, 96, 46, 56, 54, 1,
5, 3, 60, 58, 100, 98, 92, 90, 132, 130, 136, 134, 144, 94, 104,
102, 112, 62, 9, 7, 17, 21, 19, 13, 11, 116, 114, 108, 106, 148,
146, 140, 138, 192, 142, 152, 150, 160, 110, 120, 118, 65, 15, 25, 23,
33, 37, 35, 29, 27, 69, 67, 124, 122, 164, 162, 156, 154, 196, 194,
200, 198, 208, 158, 168, 166, 176, 126, 73, 71, 81, 31, 41, 39, 49,
53, 51, 45, 43, 85, 83, 77, 75, 180, 178, 172, 170, 212, 210, 204,
202, 206, 216, 214, 224, 174, 184, 182, 129, 79, 89, 87, 97, 47, 57,
55, 61, 59, 101, 99, 93, 91, 133, 131, 188, 186, 228, 226, 220, 218,
222, 232, 230, 240, 190, 137, 135, 145, 95, 105, 103, 113, 63, 117, 115,
109, 107, 149, 147, 141, 139, 244, 242, 236, 234, 238, 248, 246, 193, 143,
153, 151, 161, 111, 121, 119, 125, 123, 165, 163, 157, 155, 197, 195, 252,
250, 254, 201, 199, 209, 159, 169, 167, 177, 127, 181, 179, 173, 171, 213,
211, 205, 203, 207, 217, 215, 225, 175, 185, 183, 189, 187, 229, 227, 221,
219, 223, 233, 231, 241, 191, 245, 243, 237, 235, 239, 249, 247, 253, 251,
255
};
#endif
// The original scan order (av1_default_iscan_16x16) is modified to match
// hadamard AVX2 implementation, i.e., aom_hadamard_lp_16x16_avx2.
// Since hadamard AVX2 implementation will modify the order of coefficients,
// such that the normal scan order is no longer guaranteed to scan low
// coefficients first, therefore we modify the scan order accordingly. Note that
// this one has to be used together with default_scan_lp_16x16_transpose.
DECLARE_ALIGNED(16, static const int16_t,
av1_default_iscan_lp_16x16_transpose[256]) = {
0, 44, 2, 46, 3, 63, 9, 69, 1, 45, 4, 64, 8, 68, 11,
87, 5, 65, 7, 67, 12, 88, 18, 94, 6, 66, 13, 89, 17, 93,
24, 116, 14, 90, 16, 92, 25, 117, 31, 123, 15, 91, 26, 118, 30,
122, 41, 148, 27, 119, 29, 121, 42, 149, 48, 152, 28, 120, 43, 150,
47, 151, 62, 177, 10, 86, 20, 96, 21, 113, 35, 127, 19, 95, 22,
114, 34, 126, 37, 144, 23, 115, 33, 125, 38, 145, 52, 156, 32, 124,
39, 146, 51, 155, 58, 173, 40, 147, 50, 154, 59, 174, 73, 181, 49,
153, 60, 175, 72, 180, 83, 198, 61, 176, 71, 179, 84, 199, 98, 202,
70, 178, 85, 200, 97, 201, 112, 219, 36, 143, 54, 158, 55, 170, 77,
185, 53, 157, 56, 171, 76, 184, 79, 194, 57, 172, 75, 183, 80, 195,
102, 206, 74, 182, 81, 196, 101, 205, 108, 215, 82, 197, 100, 204, 109,
216, 131, 223, 99, 203, 110, 217, 130, 222, 140, 232, 111, 218, 129, 221,
141, 233, 160, 236, 128, 220, 142, 234, 159, 235, 169, 245, 78, 193, 104,
208, 105, 212, 135, 227, 103, 207, 106, 213, 134, 226, 136, 228, 107, 214,
133, 225, 137, 229, 164, 240, 132, 224, 138, 230, 163, 239, 165, 241, 139,
231, 162, 238, 166, 242, 189, 249, 161, 237, 167, 243, 188, 248, 190, 250,
168, 244, 187, 247, 191, 251, 210, 254, 186, 246, 192, 252, 209, 253, 211,
255
};
#if CONFIG_AV1_HIGHBITDEPTH
// The original scan order (av1_default_iscan_16x16) is modified to match
// hadamard AVX2 implementation, i.e., aom_hadamard_16x16_avx2.
// Since hadamard AVX2 implementation will modify the order of coefficients,
// such that the normal scan order is no longer guaranteed to scan low
// coefficients first, therefore we modify the scan order accordingly. Note that
// this one has to be used together with default_scan_fp_16x16_transpose.
DECLARE_ALIGNED(16, static const int16_t,
av1_default_iscan_fp_16x16_transpose[256]) = {
0, 44, 2, 46, 1, 45, 4, 64, 3, 63, 9, 69, 8, 68, 11,
87, 5, 65, 7, 67, 6, 66, 13, 89, 12, 88, 18, 94, 17, 93,
24, 116, 14, 90, 16, 92, 15, 91, 26, 118, 25, 117, 31, 123, 30,
122, 41, 148, 27, 119, 29, 121, 28, 120, 43, 150, 42, 149, 48, 152,
47, 151, 62, 177, 10, 86, 20, 96, 19, 95, 22, 114, 21, 113, 35,
127, 34, 126, 37, 144, 23, 115, 33, 125, 32, 124, 39, 146, 38, 145,
52, 156, 51, 155, 58, 173, 40, 147, 50, 154, 49, 153, 60, 175, 59,
174, 73, 181, 72, 180, 83, 198, 61, 176, 71, 179, 70, 178, 85, 200,
84, 199, 98, 202, 97, 201, 112, 219, 36, 143, 54, 158, 53, 157, 56,
171, 55, 170, 77, 185, 76, 184, 79, 194, 57, 172, 75, 183, 74, 182,
81, 196, 80, 195, 102, 206, 101, 205, 108, 215, 82, 197, 100, 204, 99,
203, 110, 217, 109, 216, 131, 223, 130, 222, 140, 232, 111, 218, 129, 221,
128, 220, 142, 234, 141, 233, 160, 236, 159, 235, 169, 245, 78, 193, 104,
208, 103, 207, 106, 213, 105, 212, 135, 227, 134, 226, 136, 228, 107, 214,
133, 225, 132, 224, 138, 230, 137, 229, 164, 240, 163, 239, 165, 241, 139,
231, 162, 238, 161, 237, 167, 243, 166, 242, 189, 249, 188, 248, 190, 250,
168, 244, 187, 247, 186, 246, 192, 252, 191, 251, 210, 254, 209, 253, 211,
255
};
#endif
static INLINE int early_term_inter_search_with_sse(int early_term_idx,
BLOCK_SIZE bsize,
int64_t this_sse,
int64_t best_sse,
PREDICTION_MODE this_mode) {
// Aggressiveness to terminate inter mode search early is adjusted based on
// speed and block size.
static const double early_term_thresh[4][4] = { { 0.65, 0.65, 0.65, 0.7 },
{ 0.6, 0.65, 0.85, 0.9 },
{ 0.5, 0.5, 0.55, 0.6 },
{ 0.6, 0.75, 0.85, 0.85 } };
static const double early_term_thresh_newmv_nearestmv[4] = { 0.3, 0.3, 0.3,
0.3 };
const int size_group = size_group_lookup[bsize];
assert(size_group < 4);
assert((early_term_idx > 0) && (early_term_idx < EARLY_TERM_INDICES));
const double threshold =
((early_term_idx == EARLY_TERM_IDX_4) &&
(this_mode == NEWMV || this_mode == NEARESTMV))
? early_term_thresh_newmv_nearestmv[size_group]
: early_term_thresh[early_term_idx - 1][size_group];
// Terminate inter mode search early based on best sse so far.
if ((early_term_idx > 0) && (threshold * this_sse > best_sse)) {
return 1;
}
return 0;
}
static INLINE void init_best_pickmode(BEST_PICKMODE *bp) {
bp->best_sse = INT64_MAX;
bp->best_mode = NEARESTMV;
bp->best_ref_frame = LAST_FRAME;
bp->best_second_ref_frame = NONE_FRAME;
bp->best_tx_size = TX_8X8;
bp->tx_type = DCT_DCT;
bp->best_pred_filter = av1_broadcast_interp_filter(EIGHTTAP_REGULAR);
bp->best_mode_skip_txfm = 0;
bp->best_mode_initial_skip_flag = 0;
bp->best_pred = NULL;
bp->best_motion_mode = SIMPLE_TRANSLATION;
bp->num_proj_ref = 0;
memset(&bp->wm_params, 0, sizeof(bp->wm_params));
memset(&bp->blk_skip, 0, sizeof(bp->blk_skip));
memset(&bp->pmi, 0, sizeof(bp->pmi));
}
static INLINE int subpel_select(AV1_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize,
int_mv *mv, MV ref_mv, FULLPEL_MV start_mv,
bool fullpel_performed_well) {
const int frame_lowmotion = cpi->rc.avg_frame_low_motion;
// Reduce MV precision for higher int MV value & frame-level motion
if (cpi->sf.rt_sf.reduce_mv_pel_precision_highmotion >= 3) {
int mv_thresh = 4;
const int is_low_resoln =
(cpi->common.width * cpi->common.height <= 320 * 240);
mv_thresh = (bsize > BLOCK_32X32) ? 2 : (bsize > BLOCK_16X16) ? 4 : 6;
if (frame_lowmotion > 0 && frame_lowmotion < 40) mv_thresh = 12;
mv_thresh = (is_low_resoln) ? mv_thresh >> 1 : mv_thresh;
if (abs(mv->as_fullmv.row) >= mv_thresh ||
abs(mv->as_fullmv.col) >= mv_thresh)
return HALF_PEL;
} else if (cpi->sf.rt_sf.reduce_mv_pel_precision_highmotion >= 1) {
int mv_thresh;
const int th_vals[2][3] = { { 4, 8, 10 }, { 4, 6, 8 } };
const int th_idx = cpi->sf.rt_sf.reduce_mv_pel_precision_highmotion - 1;
assert(th_idx >= 0 && th_idx < 2);
if (frame_lowmotion > 0 && frame_lowmotion < 40)
mv_thresh = 12;
else
mv_thresh = (bsize >= BLOCK_32X32) ? th_vals[th_idx][0]
: (bsize >= BLOCK_16X16) ? th_vals[th_idx][1]
: th_vals[th_idx][2];
if (abs(mv->as_fullmv.row) >= (mv_thresh << 1) ||
abs(mv->as_fullmv.col) >= (mv_thresh << 1))
return FULL_PEL;
else if (abs(mv->as_fullmv.row) >= mv_thresh ||
abs(mv->as_fullmv.col) >= mv_thresh)
return HALF_PEL;
}
// Reduce MV precision for relatively static (e.g. background), low-complex
// large areas
if (cpi->sf.rt_sf.reduce_mv_pel_precision_lowcomplex >= 2) {
const int qband = x->qindex >> (QINDEX_BITS - 2);
assert(qband < 4);
if (x->content_state_sb.source_sad_nonrd <= kVeryLowSad &&
bsize > BLOCK_16X16 && qband != 0) {
if (x->source_variance < 500)
return FULL_PEL;
else if (x->source_variance < 5000)
return HALF_PEL;
}
} else if (cpi->sf.rt_sf.reduce_mv_pel_precision_lowcomplex >= 1) {
if (fullpel_performed_well && ref_mv.row == 0 && ref_mv.col == 0 &&
start_mv.row == 0 && start_mv.col == 0)
return HALF_PEL;
}
return cpi->sf.mv_sf.subpel_force_stop;
}
static bool use_aggressive_subpel_search_method(
MACROBLOCK *x, bool use_adaptive_subpel_search,
const bool fullpel_performed_well) {
if (!use_adaptive_subpel_search) return false;
const int qband = x->qindex >> (QINDEX_BITS - 2);
assert(qband < 4);
if ((qband > 0) && (fullpel_performed_well ||
(x->content_state_sb.source_sad_nonrd <= kLowSad) ||
(x->source_variance < 100)))
return true;
return false;
}
/*!\brief Runs Motion Estimation for a specific block and specific ref frame.
*
* \ingroup nonrd_mode_search
* \callgraph
* \callergraph
* Finds the best Motion Vector by running Motion Estimation for a specific
* block and a specific reference frame. Exits early if RDCost of Full Pel part
* exceeds best RD Cost fund so far
* \param[in] cpi Top-level encoder structure
* \param[in] x Pointer to structure holding all the
* data for the current macroblock
* \param[in] bsize Current block size
* \param[in] mi_row Row index in 4x4 units
* \param[in] mi_col Column index in 4x4 units
* \param[in] tmp_mv Pointer to best found New MV
* \param[in] rate_mv Pointer to Rate of the best new MV
* \param[in] best_rd_sofar RD Cost of the best mode found so far
* \param[in] use_base_mv Flag, indicating that tmp_mv holds
* specific MV to start the search with
*
* \return Returns 0 if ME was terminated after Full Pel Search because too
* high RD Cost. Otherwise returns 1. Best New MV is placed into \c tmp_mv.
* Rate estimation for this vector is placed to \c rate_mv
*/
static int combined_motion_search(AV1_COMP *cpi, MACROBLOCK *x,
BLOCK_SIZE bsize, int mi_row, int mi_col,
int_mv *tmp_mv, int *rate_mv,
int64_t best_rd_sofar, int use_base_mv) {
MACROBLOCKD *xd = &x->e_mbd;
const AV1_COMMON *cm = &cpi->common;
const int num_planes = av1_num_planes(cm);
const SPEED_FEATURES *sf = &cpi->sf;
MB_MODE_INFO *mi = xd->mi[0];
struct buf_2d backup_yv12[MAX_MB_PLANE] = { { 0, 0, 0, 0, 0 } };
int step_param = (sf->rt_sf.fullpel_search_step_param)
? sf->rt_sf.fullpel_search_step_param
: cpi->mv_search_params.mv_step_param;
FULLPEL_MV start_mv;
const int ref = mi->ref_frame[0];
const MV ref_mv = av1_get_ref_mv(x, mi->ref_mv_idx).as_mv;
MV center_mv;
int dis;
int rv = 0;
int cost_list[5];
int search_subpel = 1;
const YV12_BUFFER_CONFIG *scaled_ref_frame =
av1_get_scaled_ref_frame(cpi, ref);
if (scaled_ref_frame) {
int i;
// Swap out the reference frame for a version that's been scaled to
// match the resolution of the current frame, allowing the existing
// motion search code to be used without additional modifications.
for (i = 0; i < MAX_MB_PLANE; i++) backup_yv12[i] = xd->plane[i].pre[0];
av1_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL,
num_planes);
}
start_mv = get_fullmv_from_mv(&ref_mv);
if (!use_base_mv)
center_mv = ref_mv;
else
center_mv = tmp_mv->as_mv;
const SEARCH_METHODS search_method = sf->mv_sf.search_method;
const search_site_config *src_search_sites =
av1_get_search_site_config(cpi, x, search_method);
FULLPEL_MOTION_SEARCH_PARAMS full_ms_params;
av1_make_default_fullpel_ms_params(&full_ms_params, cpi, x, bsize, &center_mv,
src_search_sites,
/*fine_search_interval=*/0);
const unsigned int full_var_rd = av1_full_pixel_search(
start_mv, &full_ms_params, step_param, cond_cost_list(cpi, cost_list),
&tmp_mv->as_fullmv, NULL);
// calculate the bit cost on motion vector
MV mvp_full = get_mv_from_fullmv(&tmp_mv->as_fullmv);
*rate_mv = av1_mv_bit_cost(&mvp_full, &ref_mv, x->mv_costs->nmv_joint_cost,
x->mv_costs->mv_cost_stack, MV_COST_WEIGHT);
// TODO(kyslov) Account for Rate Mode!
rv = !(RDCOST(x->rdmult, (*rate_mv), 0) > best_rd_sofar);
if (rv && search_subpel) {
SUBPEL_MOTION_SEARCH_PARAMS ms_params;
av1_make_default_subpel_ms_params(&ms_params, cpi, x, bsize, &ref_mv,
cost_list);
const bool fullpel_performed_well =
(bsize == BLOCK_64X64 && full_var_rd * 40 < 62267 * 7) ||
(bsize == BLOCK_32X32 && full_var_rd * 8 < 42380) ||
(bsize == BLOCK_16X16 && full_var_rd * 8 < 10127);
if (sf->rt_sf.reduce_mv_pel_precision_highmotion ||
sf->rt_sf.reduce_mv_pel_precision_lowcomplex)
ms_params.forced_stop = subpel_select(cpi, x, bsize, tmp_mv, ref_mv,
start_mv, fullpel_performed_well);
MV subpel_start_mv = get_mv_from_fullmv(&tmp_mv->as_fullmv);
assert(av1_is_subpelmv_in_range(&ms_params.mv_limits, subpel_start_mv));
// adaptively downgrade subpel search method based on block properties
if (use_aggressive_subpel_search_method(
x, sf->rt_sf.use_adaptive_subpel_search, fullpel_performed_well))
av1_find_best_sub_pixel_tree_pruned_more(xd, cm, &ms_params,
subpel_start_mv, &tmp_mv->as_mv,
&dis, &x->pred_sse[ref], NULL);
else
cpi->mv_search_params.find_fractional_mv_step(
xd, cm, &ms_params, subpel_start_mv, &tmp_mv->as_mv, &dis,
&x->pred_sse[ref], NULL);
*rate_mv =
av1_mv_bit_cost(&tmp_mv->as_mv, &ref_mv, x->mv_costs->nmv_joint_cost,
x->mv_costs->mv_cost_stack, MV_COST_WEIGHT);
}
if (scaled_ref_frame) {
int i;
for (i = 0; i < MAX_MB_PLANE; i++) xd->plane[i].pre[0] = backup_yv12[i];
}
// The final MV can not be equal to the reference MV as this will trigger an
// assert later. This can happen if both NEAREST and NEAR modes were skipped.
rv = (tmp_mv->as_mv.col != ref_mv.col || tmp_mv->as_mv.row != ref_mv.row);
return rv;
}
/*!\brief Searches for the best New Motion Vector.
*
* \ingroup nonrd_mode_search
* \callgraph
* \callergraph
* Finds the best Motion Vector by doing Motion Estimation. Uses reduced
* complexity ME for non-LAST frames or calls \c combined_motion_search
* for LAST reference frame
* \param[in] cpi Top-level encoder structure
* \param[in] x Pointer to structure holding all the
* data for the current macroblock
* \param[in] frame_mv Array that holds MVs for all modes
* and ref frames
* \param[in] ref_frame Reference frame for which to find
* the best New MVs
* \param[in] gf_temporal_ref Flag, indicating temporal reference
* for GOLDEN frame
* \param[in] bsize Current block size
* \param[in] mi_row Row index in 4x4 units
* \param[in] mi_col Column index in 4x4 units
* \param[in] rate_mv Pointer to Rate of the best new MV
* \param[in] best_rdc Pointer to the RD Cost for the best
* mode found so far
*
* \return Returns -1 if the search was not done, otherwise returns 0.
* Best New MV is placed into \c frame_mv array, Rate estimation for this
* vector is placed to \c rate_mv
*/
static int search_new_mv(AV1_COMP *cpi, MACROBLOCK *x,
int_mv frame_mv[][REF_FRAMES],
MV_REFERENCE_FRAME ref_frame, int gf_temporal_ref,
BLOCK_SIZE bsize, int mi_row, int mi_col, int *rate_mv,
RD_STATS *best_rdc) {
MACROBLOCKD *const xd = &x->e_mbd;
MB_MODE_INFO *const mi = xd->mi[0];
AV1_COMMON *cm = &cpi->common;
if (ref_frame > LAST_FRAME && cpi->oxcf.rc_cfg.mode == AOM_CBR &&
gf_temporal_ref) {
int tmp_sad;
int dis;
if (bsize < BLOCK_16X16) return -1;
tmp_sad = av1_int_pro_motion_estimation(
cpi, x, bsize, mi_row, mi_col,
&x->mbmi_ext.ref_mv_stack[ref_frame][0].this_mv.as_mv);
if (tmp_sad > x->pred_mv_sad[LAST_FRAME]) return -1;
frame_mv[NEWMV][ref_frame].as_int = mi->mv[0].as_int;
int_mv best_mv = mi->mv[0];
best_mv.as_mv.row >>= 3;
best_mv.as_mv.col >>= 3;
MV ref_mv = av1_get_ref_mv(x, 0).as_mv;
frame_mv[NEWMV][ref_frame].as_mv.row >>= 3;
frame_mv[NEWMV][ref_frame].as_mv.col >>= 3;
SUBPEL_MOTION_SEARCH_PARAMS ms_params;
av1_make_default_subpel_ms_params(&ms_params, cpi, x, bsize, &ref_mv, NULL);
if (cpi->sf.rt_sf.reduce_mv_pel_precision_highmotion ||
cpi->sf.rt_sf.reduce_mv_pel_precision_lowcomplex) {
FULLPEL_MV start_mv = { .row = 0, .col = 0 };
ms_params.forced_stop =
subpel_select(cpi, x, bsize, &best_mv, ref_mv, start_mv, false);
}
MV start_mv = get_mv_from_fullmv(&best_mv.as_fullmv);
assert(av1_is_subpelmv_in_range(&ms_params.mv_limits, start_mv));
cpi->mv_search_params.find_fractional_mv_step(
xd, cm, &ms_params, start_mv, &best_mv.as_mv, &dis,
&x->pred_sse[ref_frame], NULL);
frame_mv[NEWMV][ref_frame].as_int = best_mv.as_int;
// When NEWMV is same as ref_mv from the drl, it is preferred to code the
// MV as NEARESTMV or NEARMV. In this case, NEWMV needs to be skipped to
// avoid an assert failure at a later stage. The scenario can occur if
// NEARESTMV was not evaluated for ALTREF.
if (frame_mv[NEWMV][ref_frame].as_mv.col == ref_mv.col &&
frame_mv[NEWMV][ref_frame].as_mv.row == ref_mv.row)
return -1;
*rate_mv = av1_mv_bit_cost(&frame_mv[NEWMV][ref_frame].as_mv, &ref_mv,
x->mv_costs->nmv_joint_cost,
x->mv_costs->mv_cost_stack, MV_COST_WEIGHT);
} else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
&frame_mv[NEWMV][ref_frame], rate_mv,
best_rdc->rdcost, 0)) {
return -1;
}
return 0;
}
static void estimate_single_ref_frame_costs(const AV1_COMMON *cm,
const MACROBLOCKD *xd,
const ModeCosts *mode_costs,
int segment_id, BLOCK_SIZE bsize,
unsigned int *ref_costs_single) {
int seg_ref_active =
segfeature_active(&cm->seg, segment_id, SEG_LVL_REF_FRAME);
if (seg_ref_active) {
memset(ref_costs_single, 0, REF_FRAMES * sizeof(*ref_costs_single));
} else {
int intra_inter_ctx = av1_get_intra_inter_context(xd);
ref_costs_single[INTRA_FRAME] =
mode_costs->intra_inter_cost[intra_inter_ctx][0];
unsigned int base_cost = mode_costs->intra_inter_cost[intra_inter_ctx][1];
if (cm->current_frame.reference_mode == REFERENCE_MODE_SELECT &&
is_comp_ref_allowed(bsize)) {
const int comp_ref_type_ctx = av1_get_comp_reference_type_context(xd);
base_cost += mode_costs->comp_ref_type_cost[comp_ref_type_ctx][1];
}
ref_costs_single[LAST_FRAME] = base_cost;
ref_costs_single[GOLDEN_FRAME] = base_cost;
ref_costs_single[ALTREF_FRAME] = base_cost;
// add cost for last, golden, altref
ref_costs_single[LAST_FRAME] += mode_costs->single_ref_cost[0][0][0];
ref_costs_single[GOLDEN_FRAME] += mode_costs->single_ref_cost[0][0][1];
ref_costs_single[GOLDEN_FRAME] += mode_costs->single_ref_cost[0][1][0];
ref_costs_single[ALTREF_FRAME] += mode_costs->single_ref_cost[0][0][1];
ref_costs_single[ALTREF_FRAME] += mode_costs->single_ref_cost[0][2][0];
}
}
static INLINE void set_force_skip_flag(const AV1_COMP *const cpi,
MACROBLOCK *const x, unsigned int sse,
int *force_skip) {
if (x->txfm_search_params.tx_mode_search_type == TX_MODE_SELECT &&
cpi->sf.rt_sf.tx_size_level_based_on_qstep &&
cpi->sf.rt_sf.tx_size_level_based_on_qstep >= 2) {
const int qstep = x->plane[0].dequant_QTX[1] >> (x->e_mbd.bd - 5);
const unsigned int qstep_sq = qstep * qstep;
// If the sse is low for low source variance blocks, mark those as
// transform skip.
// Note: Though qstep_sq is based on ac qstep, the threshold is kept
// low so that reliable early estimate of tx skip can be obtained
// through its comparison with sse.
if (sse < qstep_sq && x->source_variance < qstep_sq &&
x->color_sensitivity[0] == 0 && x->color_sensitivity[1] == 0)
*force_skip = 1;
}
}
#define CAP_TX_SIZE_FOR_BSIZE_GT32(tx_mode_search_type, bsize) \
(((tx_mode_search_type) != ONLY_4X4 && (bsize) > BLOCK_32X32) ? true : false)
#define TX_SIZE_FOR_BSIZE_GT32 (TX_16X16)
static TX_SIZE calculate_tx_size(const AV1_COMP *const cpi, BLOCK_SIZE bsize,
MACROBLOCK *const x, unsigned int var,
unsigned int sse, int *force_skip) {
MACROBLOCKD *const xd = &x->e_mbd;
TX_SIZE tx_size;
const TxfmSearchParams *txfm_params = &x->txfm_search_params;
if (txfm_params->tx_mode_search_type == TX_MODE_SELECT) {
int multiplier = 8;
unsigned int var_thresh = 0;
unsigned int is_high_var = 1;
// Use quantizer based thresholds to determine transform size.
if (cpi->sf.rt_sf.tx_size_level_based_on_qstep) {
const int qband = x->qindex >> (QINDEX_BITS - 2);
const int mult[4] = { 8, 7, 6, 5 };
assert(qband < 4);
multiplier = mult[qband];
const int qstep = x->plane[0].dequant_QTX[1] >> (xd->bd - 5);
const unsigned int qstep_sq = qstep * qstep;
var_thresh = qstep_sq * 2;
if (cpi->sf.rt_sf.tx_size_level_based_on_qstep >= 2) {
// If the sse is low for low source variance blocks, mark those as
// transform skip.
// Note: Though qstep_sq is based on ac qstep, the threshold is kept
// low so that reliable early estimate of tx skip can be obtained
// through its comparison with sse.
if (sse < qstep_sq && x->source_variance < qstep_sq &&
x->color_sensitivity[0] == 0 && x->color_sensitivity[1] == 0)
*force_skip = 1;
// Further lower transform size based on aq mode only if residual
// variance is high.
is_high_var = (var >= var_thresh);
}
}
// Choose larger transform size for blocks where dc component is dominant or
// the ac component is low.
if (sse > ((var * multiplier) >> 2) || (var < var_thresh))
tx_size =
AOMMIN(max_txsize_lookup[bsize],
tx_mode_to_biggest_tx_size[txfm_params->tx_mode_search_type]);
else
tx_size = TX_8X8;
if (cpi->oxcf.q_cfg.aq_mode == CYCLIC_REFRESH_AQ &&
cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id) && is_high_var)
tx_size = TX_8X8;
else if (tx_size > TX_16X16)
tx_size = TX_16X16;
} else {
tx_size =
AOMMIN(max_txsize_lookup[bsize],
tx_mode_to_biggest_tx_size[txfm_params->tx_mode_search_type]);
}
if (CAP_TX_SIZE_FOR_BSIZE_GT32(txfm_params->tx_mode_search_type, bsize))
tx_size = TX_SIZE_FOR_BSIZE_GT32;
return AOMMIN(tx_size, TX_16X16);
}
static const uint8_t b_width_log2_lookup[BLOCK_SIZES] = { 0, 0, 1, 1, 1, 2,
2, 2, 3, 3, 3, 4,
4, 4, 5, 5 };
static const uint8_t b_height_log2_lookup[BLOCK_SIZES] = { 0, 1, 0, 1, 2, 1,
2, 3, 2, 3, 4, 3,
4, 5, 4, 5 };
static void block_variance(const uint8_t *src, int src_stride,
const uint8_t *ref, int ref_stride, int w, int h,
unsigned int *sse, int *sum, int block_size,
uint32_t *sse8x8, int *sum8x8, uint32_t *var8x8) {
int k = 0;
*sse = 0;
*sum = 0;
// This function is called for block sizes >= BLOCK_32x32. As per the design
// the aom_get_var_sse_sum_8x8_quad() processes four 8x8 blocks (in a 8x32)
// per call. Hence the width and height of the block need to be at least 8 and
// 32 samples respectively.
assert(w >= 32);
assert(h >= 8);
for (int i = 0; i < h; i += block_size) {
for (int j = 0; j < w; j += 32) {
aom_get_var_sse_sum_8x8_quad(
src + src_stride * i + j, src_stride, ref + ref_stride * i + j,
ref_stride, &sse8x8[k], &sum8x8[k], sse, sum, &var8x8[k]);
k += 4;
}
}
}
static void block_variance_16x16_dual(const uint8_t *src, int src_stride,
const uint8_t *ref, int ref_stride, int w,
int h, unsigned int *sse, int *sum,
int block_size, uint32_t *sse16x16,
uint32_t *var16x16) {
int k = 0;
*sse = 0;
*sum = 0;
// This function is called for block sizes >= BLOCK_32x32. As per the design
// the aom_get_var_sse_sum_16x16_dual() processes four 16x16 blocks (in a
// 16x32) per call. Hence the width and height of the block need to be at
// least 16 and 32 samples respectively.
assert(w >= 32);
assert(h >= 16);
for (int i = 0; i < h; i += block_size) {
for (int j = 0; j < w; j += 32) {
aom_get_var_sse_sum_16x16_dual(src + src_stride * i + j, src_stride,
ref + ref_stride * i + j, ref_stride,
&sse16x16[k], sse, sum, &var16x16[k]);
k += 2;
}
}
}
static void calculate_variance(int bw, int bh, TX_SIZE tx_size,
unsigned int *sse_i, int *sum_i,
unsigned int *var_o, unsigned int *sse_o,
int *sum_o) {
const BLOCK_SIZE unit_size = txsize_to_bsize[tx_size];
const int nw = 1 << (bw - b_width_log2_lookup[unit_size]);
const int nh = 1 << (bh - b_height_log2_lookup[unit_size]);
int i, j, k = 0;
for (i = 0; i < nh; i += 2) {
for (j = 0; j < nw; j += 2) {
sse_o[k] = sse_i[i * nw + j] + sse_i[i * nw + j + 1] +
sse_i[(i + 1) * nw + j] + sse_i[(i + 1) * nw + j + 1];
sum_o[k] = sum_i[i * nw + j] + sum_i[i * nw + j + 1] +
sum_i[(i + 1) * nw + j] + sum_i[(i + 1) * nw + j + 1];
var_o[k] = sse_o[k] - (uint32_t)(((int64_t)sum_o[k] * sum_o[k]) >>
(b_width_log2_lookup[unit_size] +
b_height_log2_lookup[unit_size] + 6));
k++;
}
}
}
// Adjust the ac_thr according to speed, width, height and normalized sum
static int ac_thr_factor(const int speed, const int width, const int height,
const int norm_sum) {
if (speed >= 8 && norm_sum < 5) {
if (width <= 640 && height <= 480)
return 4;
else
return 2;
}
return 1;
}
// Sets early_term flag based on chroma planes prediction
static INLINE void set_early_term_based_on_uv_plane(
AV1_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize, MACROBLOCKD *xd, int mi_row,
int mi_col, int *early_term, int num_blk, const unsigned int *sse_tx,
const unsigned int *var_tx, int sum, unsigned int var, unsigned int sse) {
AV1_COMMON *const cm = &cpi->common;
struct macroblock_plane *const p = &x->plane[0];
const uint32_t dc_quant = p->dequant_QTX[0];
const uint32_t ac_quant = p->dequant_QTX[1];
const int64_t dc_thr = dc_quant * dc_quant >> 6;
int64_t ac_thr = ac_quant * ac_quant >> 6;
const int bw = b_width_log2_lookup[bsize];
const int bh = b_height_log2_lookup[bsize];
int ac_test = 1;
int dc_test = 1;
const int norm_sum = abs(sum) >> (bw + bh);
#if CONFIG_AV1_TEMPORAL_DENOISING
if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc(cpi) &&
cpi->oxcf.speed > 5)
ac_thr = av1_scale_acskip_thresh(ac_thr, cpi->denoiser.denoising_level,
norm_sum, cpi->svc.temporal_layer_id);
else
ac_thr *= ac_thr_factor(cpi->oxcf.speed, cm->width, cm->height, norm_sum);
#else
ac_thr *= ac_thr_factor(cpi->oxcf.speed, cm->width, cm->height, norm_sum);
#endif
for (int k = 0; k < num_blk; k++) {
// Check if all ac coefficients can be quantized to zero.
if (!(var_tx[k] < ac_thr || var == 0)) {
ac_test = 0;
break;
}
// Check if dc coefficient can be quantized to zero.
if (!(sse_tx[k] - var_tx[k] < dc_thr || sse == var)) {
dc_test = 0;
break;
}
}
// Check if chroma can be skipped based on ac and dc test flags.
if (ac_test && dc_test) {
int skip_uv[2] = { 0 };
unsigned int var_uv[2];
unsigned int sse_uv[2];
// Transform skipping test in UV planes.
for (int i = 1; i <= 2; i++) {
int j = i - 1;
skip_uv[j] = 1;
if (x->color_sensitivity[j]) {
skip_uv[j] = 0;
struct macroblock_plane *const puv = &x->plane[i];
struct macroblockd_plane *const puvd = &xd->plane[i];
const BLOCK_SIZE uv_bsize = get_plane_block_size(
bsize, puvd->subsampling_x, puvd->subsampling_y);
// Adjust these thresholds for UV.
const int64_t uv_dc_thr =
(puv->dequant_QTX[0] * puv->dequant_QTX[0]) >> 3;
const int64_t uv_ac_thr =
(puv->dequant_QTX[1] * puv->dequant_QTX[1]) >> 3;
av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize, i,
i);
var_uv[j] = cpi->ppi->fn_ptr[uv_bsize].vf(puv->src.buf, puv->src.stride,
puvd->dst.buf,
puvd->dst.stride, &sse_uv[j]);
if ((var_uv[j] < uv_ac_thr || var_uv[j] == 0) &&
(sse_uv[j] - var_uv[j] < uv_dc_thr || sse_uv[j] == var_uv[j]))
skip_uv[j] = 1;
else
break;
}
}
if (skip_uv[0] & skip_uv[1]) {
*early_term = 1;
}
}
}
static INLINE void calc_rate_dist_block_param(AV1_COMP *cpi, MACROBLOCK *x,
RD_STATS *rd_stats,
int calculate_rd, int *early_term,
BLOCK_SIZE bsize,
unsigned int sse) {
if (calculate_rd) {
if (!*early_term) {
const int bw = block_size_wide[bsize];
const int bh = block_size_high[bsize];
model_rd_with_curvfit(cpi, x, bsize, AOM_PLANE_Y, rd_stats->sse, bw * bh,
&rd_stats->rate, &rd_stats->dist);
}
if (*early_term) {
rd_stats->rate = 0;
rd_stats->dist = sse << 4;
}
}
}
static void model_skip_for_sb_y_large_64(AV1_COMP *cpi, BLOCK_SIZE bsize,
int mi_row, int mi_col, MACROBLOCK *x,
MACROBLOCKD *xd, RD_STATS *rd_stats,
int *early_term, int calculate_rd,
int64_t best_sse,
unsigned int *var_output,
unsigned int var_prune_threshold) {
// Note our transform coeffs are 8 times an orthogonal transform.
// Hence quantizer step is also 8 times. To get effective quantizer
// we need to divide by 8 before sending to modeling function.
unsigned int sse;
struct macroblock_plane *const p = &x->plane[0];
struct macroblockd_plane *const pd = &xd->plane[0];
int test_skip = 1;
unsigned int var;
int sum;
const int bw = b_width_log2_lookup[bsize];
const int bh = b_height_log2_lookup[bsize];
unsigned int sse16x16[64] = { 0 };
unsigned int var16x16[64] = { 0 };
assert(xd->mi[0]->tx_size == TX_16X16);
assert(bsize > BLOCK_32X32);
// Calculate variance for whole partition, and also save 16x16 blocks'
// variance to be used in following transform skipping test.
block_variance_16x16_dual(p->src.buf, p->src.stride, pd->dst.buf,
pd->dst.stride, 4 << bw, 4 << bh, &sse, &sum, 16,
sse16x16, var16x16);
var = sse - (unsigned int)(((int64_t)sum * sum) >> (bw + bh + 4));
if (var_output) {
*var_output = var;
if (*var_output > var_prune_threshold) {
return;
}
}
rd_stats->sse = sse;
// Skipping test
*early_term = 0;
set_force_skip_flag(cpi, x, sse, early_term);
// The code below for setting skip flag assumes transform size of at least
// 8x8, so force this lower limit on transform.
MB_MODE_INFO *const mi = xd->mi[0];
if (!calculate_rd && cpi->sf.rt_sf.sse_early_term_inter_search &&
early_term_inter_search_with_sse(
cpi->sf.rt_sf.sse_early_term_inter_search, bsize, sse, best_sse,
mi->mode))
test_skip = 0;
if (*early_term) test_skip = 0;
// Evaluate if the partition block is a skippable block in Y plane.
if (test_skip) {
const unsigned int *sse_tx = sse16x16;
const unsigned int *var_tx = var16x16;
const unsigned int num_block = (1 << (bw + bh - 2)) >> 2;
set_early_term_based_on_uv_plane(cpi, x, bsize, xd, mi_row, mi_col,
early_term, num_block, sse_tx, var_tx, sum,
var, sse);
}
calc_rate_dist_block_param(cpi, x, rd_stats, calculate_rd, early_term, bsize,
sse);
}
static void model_skip_for_sb_y_large(AV1_COMP *cpi, BLOCK_SIZE bsize,
int mi_row, int mi_col, MACROBLOCK *x,
MACROBLOCKD *xd, RD_STATS *rd_stats,
int *early_term, int calculate_rd,
int64_t best_sse,
unsigned int *var_output,
unsigned int var_prune_threshold) {
if (x->force_zeromv_skip_for_blk) {
*early_term = 1;
rd_stats->rate = 0;
rd_stats->dist = 0;
rd_stats->sse = 0;
return;
}
// For block sizes greater than 32x32, the transform size is always 16x16.
// This function avoids calling calculate_variance() for tx_size 16x16 cases
// by directly populating variance at tx_size level from
// block_variance_16x16_dual() function.
const TxfmSearchParams *txfm_params = &x->txfm_search_params;
if (CAP_TX_SIZE_FOR_BSIZE_GT32(txfm_params->tx_mode_search_type, bsize)) {
xd->mi[0]->tx_size = TX_SIZE_FOR_BSIZE_GT32;
model_skip_for_sb_y_large_64(cpi, bsize, mi_row, mi_col, x, xd, rd_stats,
early_term, calculate_rd, best_sse, var_output,
var_prune_threshold);
return;
}
// Note our transform coeffs are 8 times an orthogonal transform.
// Hence quantizer step is also 8 times. To get effective quantizer
// we need to divide by 8 before sending to modeling function.
unsigned int sse;
struct macroblock_plane *const p = &x->plane[0];
struct macroblockd_plane *const pd = &xd->plane[0];
int test_skip = 1;
unsigned int var;
int sum;
const int bw = b_width_log2_lookup[bsize];
const int bh = b_height_log2_lookup[bsize];
unsigned int sse8x8[256] = { 0 };
int sum8x8[256] = { 0 };
unsigned int var8x8[256] = { 0 };
TX_SIZE tx_size;
// Calculate variance for whole partition, and also save 8x8 blocks' variance
// to be used in following transform skipping test.
block_variance(p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride,
4 << bw, 4 << bh, &sse, &sum, 8, sse8x8, sum8x8, var8x8);
var = sse - (unsigned int)(((int64_t)sum * sum) >> (bw + bh + 4));
if (var_output) {
*var_output = var;
if (*var_output > var_prune_threshold) {
return;
}
}
rd_stats->sse = sse;
// Skipping test
*early_term = 0;
tx_size = calculate_tx_size(cpi, bsize, x, var, sse, early_term);
assert(tx_size <= TX_16X16);
// The code below for setting skip flag assumes transform size of at least
// 8x8, so force this lower limit on transform.
if (tx_size < TX_8X8) tx_size = TX_8X8;
xd->mi[0]->tx_size = tx_size;
MB_MODE_INFO *const mi = xd->mi[0];
if (!calculate_rd && cpi->sf.rt_sf.sse_early_term_inter_search &&
early_term_inter_search_with_sse(
cpi->sf.rt_sf.sse_early_term_inter_search, bsize, sse, best_sse,
mi->mode))
test_skip = 0;
if (*early_term) test_skip = 0;
// Evaluate if the partition block is a skippable block in Y plane.
if (test_skip) {
unsigned int sse16x16[64] = { 0 };
int sum16x16[64] = { 0 };
unsigned int var16x16[64] = { 0 };
const unsigned int *sse_tx = sse8x8;
const unsigned int *var_tx = var8x8;
unsigned int num_blks = 1 << (bw + bh - 2);
if (tx_size >= TX_16X16) {
calculate_variance(bw, bh, TX_8X8, sse8x8, sum8x8, var16x16, sse16x16,
sum16x16);
sse_tx = sse16x16;
var_tx = var16x16;
num_blks = num_blks >> 2;
}
set_early_term_based_on_uv_plane(cpi, x, bsize, xd, mi_row, mi_col,
early_term, num_blks, sse_tx, var_tx, sum,
var, sse);
}
calc_rate_dist_block_param(cpi, x, rd_stats, calculate_rd, early_term, bsize,
sse);
}
static void model_rd_for_sb_y(const AV1_COMP *const cpi, BLOCK_SIZE bsize,
MACROBLOCK *x, MACROBLOCKD *xd,
RD_STATS *rd_stats, unsigned int *var_out,
int calculate_rd, int *early_term) {
if (x->force_zeromv_skip_for_blk && early_term != NULL) {
*early_term = 1;
rd_stats->rate = 0;
rd_stats->dist = 0;
rd_stats->sse = 0;
}
// Note our transform coeffs are 8 times an orthogonal transform.
// Hence quantizer step is also 8 times. To get effective quantizer
// we need to divide by 8 before sending to modeling function.
const int ref = xd->mi[0]->ref_frame[0];
assert(bsize < BLOCK_SIZES_ALL);
struct macroblock_plane *const p = &x->plane[0];
struct macroblockd_plane *const pd = &xd->plane[0];
unsigned int sse;
int rate;
int64_t dist;
unsigned int var = cpi->ppi->fn_ptr[bsize].vf(
p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride, &sse);
int force_skip = 0;
xd->mi[0]->tx_size = calculate_tx_size(cpi, bsize, x, var, sse, &force_skip);
if (var_out) {
*var_out = var;
}
if (calculate_rd && (!force_skip || ref == INTRA_FRAME)) {
const int bwide = block_size_wide[bsize];
const int bhigh = block_size_high[bsize];
model_rd_with_curvfit(cpi, x, bsize, AOM_PLANE_Y, sse, bwide * bhigh, &rate,
&dist);
} else {
rate = INT_MAX; // this will be overwritten later with block_yrd
dist = INT_MAX;
}
rd_stats->sse = sse;
x->pred_sse[ref] = (unsigned int)AOMMIN(sse, UINT_MAX);
if (force_skip && ref > INTRA_FRAME) {
rate = 0;
dist = (int64_t)sse << 4;
}
assert(rate >= 0);
rd_stats->skip_txfm = (rate == 0);
rate = AOMMIN(rate, INT_MAX);
rd_stats->rate = rate;
rd_stats->dist = dist;
}
static INLINE void aom_process_hadamard_lp_8x16(MACROBLOCK *x,
int max_blocks_high,
int max_blocks_wide,
int num_4x4_w, int step,
int block_step) {
struct macroblock_plane *const p = &x->plane[0];
const int bw = 4 * num_4x4_w;
const int num_4x4 = AOMMIN(num_4x4_w, max_blocks_wide);
int block = 0;
for (int r = 0; r < max_blocks_high; r += block_step) {
for (int c = 0; c < num_4x4; c += 2 * block_step) {
const int16_t *src_diff = &p->src_diff[(r * bw + c) << 2];
int16_t *low_coeff = (int16_t *)p->coeff + BLOCK_OFFSET(block);
aom_hadamard_lp_8x8_dual(src_diff, (ptrdiff_t)bw, low_coeff);
block += 2 * step;
}
}
}
#define DECLARE_BLOCK_YRD_BUFFERS() \
DECLARE_ALIGNED(64, tran_low_t, dqcoeff_buf[16 * 16]); \
DECLARE_ALIGNED(64, tran_low_t, qcoeff_buf[16 * 16]); \
DECLARE_ALIGNED(64, tran_low_t, coeff_buf[16 * 16]); \
uint16_t eob[1];
#define DECLARE_BLOCK_YRD_VARS() \
/* When is_tx_8x8_dual_applicable is true, we compute the txfm for the \
* entire bsize and write macroblock_plane::coeff. So low_coeff is kept \
* as a non-const so we can reassign it to macroblock_plane::coeff. */ \
int16_t *low_coeff = (int16_t *)coeff_buf; \
int16_t *const low_qcoeff = (int16_t *)qcoeff_buf; \
int16_t *const low_dqcoeff = (int16_t *)dqcoeff_buf; \
const SCAN_ORDER *const scan_order = &av1_scan_orders[tx_size][DCT_DCT]; \
const int diff_stride = bw;
#define DECLARE_LOOP_VARS_BLOCK_YRD() \
const int16_t *src_diff = &p->src_diff[(r * diff_stride + c) << 2];
#if CONFIG_AV1_HIGHBITDEPTH
#define DECLARE_BLOCK_YRD_HBD_VARS() \
tran_low_t *const coeff = coeff_buf; \
tran_low_t *const qcoeff = qcoeff_buf; \
tran_low_t *const dqcoeff = dqcoeff_buf;
static AOM_FORCE_INLINE void update_yrd_loop_vars_hbd(
MACROBLOCK *x, int *skippable, const int step, const int ncoeffs,
tran_low_t *const coeff, tran_low_t *const qcoeff,
tran_low_t *const dqcoeff, RD_STATS *this_rdc, int *eob_cost,
const int tx_blk_id) {
const int is_txfm_skip = (ncoeffs == 0);
*skippable &= is_txfm_skip;
x->txfm_search_info.blk_skip[tx_blk_id] = is_txfm_skip;
*eob_cost += get_msb(ncoeffs + 1);
int64_t dummy;
if (ncoeffs == 1)
this_rdc->rate += (int)abs(qcoeff[0]);
else if (ncoeffs > 1)
this_rdc->rate += aom_satd(qcoeff, step << 4);
this_rdc->dist += av1_block_error(coeff, dqcoeff, step << 4, &dummy) >> 2;
}
#endif
static AOM_FORCE_INLINE void update_yrd_loop_vars(
MACROBLOCK *x, int *skippable, const int step, const int ncoeffs,
int16_t *const low_coeff, int16_t *const low_qcoeff,
int16_t *const low_dqcoeff, RD_STATS *this_rdc, int *eob_cost,
const int tx_blk_id) {
const int is_txfm_skip = (ncoeffs == 0);
*skippable &= is_txfm_skip;
x->txfm_search_info.blk_skip[tx_blk_id] = is_txfm_skip;
*eob_cost += get_msb(ncoeffs + 1);
if (ncoeffs == 1)
this_rdc->rate += (int)abs(low_qcoeff[0]);
else if (ncoeffs > 1)
this_rdc->rate += aom_satd_lp(low_qcoeff, step << 4);
this_rdc->dist += av1_block_error_lp(low_coeff, low_dqcoeff, step << 4) >> 2;
}
/*!\brief Calculates RD Cost using Hadamard transform.
*
* \ingroup nonrd_mode_search
* \callgraph
* \callergraph
* Calculates RD Cost using Hadamard transform. For low bit depth this function
* uses low-precision set of functions (16-bit) and 32 bit for high bit depth
* \param[in] x Pointer to structure holding all the data for
the current macroblock
* \param[in] this_rdc Pointer to calculated RD Cost
* \param[in] skippable Pointer to a flag indicating possible tx skip
* \param[in] bsize Current block size
* \param[in] tx_size Transform size
* \param[in] is_inter_mode Flag to indicate inter mode
*
* \remark Nothing is returned. Instead, calculated RD cost is placed to
* \c this_rdc. \c skippable flag is set if there is no non-zero quantized
* coefficients for Hadamard transform
*/
static void block_yrd(MACROBLOCK *x, RD_STATS *this_rdc, int *skippable,
const BLOCK_SIZE bsize, const TX_SIZE tx_size,
const int is_inter_mode) {
MACROBLOCKD *xd = &x->e_mbd;
const struct macroblockd_plane *pd = &xd->plane[0];
struct macroblock_plane *const p = &x->plane[0];
assert(bsize < BLOCK_SIZES_ALL);
const int num_4x4_w = mi_size_wide[bsize];
const int num_4x4_h = mi_size_high[bsize];
const int step = 1 << (tx_size << 1);
const int block_step = (1 << tx_size);
const int row_step = step * num_4x4_w >> tx_size;
int block = 0;
const int max_blocks_wide =
num_4x4_w + (xd->mb_to_right_edge >= 0 ? 0 : xd->mb_to_right_edge >> 5);
const int max_blocks_high =
num_4x4_h + (xd->mb_to_bottom_edge >= 0 ? 0 : xd->mb_to_bottom_edge >> 5);
int eob_cost = 0;
const int bw = 4 * num_4x4_w;
const int bh = 4 * num_4x4_h;
const int use_hbd = is_cur_buf_hbd(xd);
int num_blk_skip_w = num_4x4_w;
int sh_blk_skip = 0;
if (is_inter_mode) {
num_blk_skip_w = num_4x4_w >> 1;
sh_blk_skip = 1;
}
#if CONFIG_AV1_HIGHBITDEPTH
if (use_hbd) {
aom_highbd_subtract_block(bh, bw, p->src_diff, bw, p->src.buf,
p->src.stride, pd->dst.buf, pd->dst.stride);
} else {
aom_subtract_block(bh, bw, p->src_diff, bw, p->src.buf, p->src.stride,
pd->dst.buf, pd->dst.stride);
}
#else
aom_subtract_block(bh, bw, p->src_diff, bw, p->src.buf, p->src.stride,
pd->dst.buf, pd->dst.stride);
#endif
// Keep the intermediate value on the stack here. Writing directly to
// skippable causes speed regression due to load-and-store issues in
// update_yrd_loop_vars.
int temp_skippable = 1;
this_rdc->dist = 0;
this_rdc->rate = 0;
// For block sizes 8x16 or above, Hadamard txfm of two adjacent 8x8 blocks
// can be done per function call. Hence the call of Hadamard txfm is
// abstracted here for the specified cases.
int is_tx_8x8_dual_applicable =
(tx_size == TX_8X8 && block_size_wide[bsize] >= 16 &&
block_size_high[bsize] >= 8);
#if CONFIG_AV1_HIGHBITDEPTH
// As of now, dual implementation of hadamard txfm is available for low
// bitdepth.
if (use_hbd) is_tx_8x8_dual_applicable = 0;
#endif
if (is_tx_8x8_dual_applicable) {
aom_process_hadamard_lp_8x16(x, max_blocks_high, max_blocks_wide, num_4x4_w,
step, block_step);
}
DECLARE_BLOCK_YRD_BUFFERS()
DECLARE_BLOCK_YRD_VARS()
#if CONFIG_AV1_HIGHBITDEPTH
DECLARE_BLOCK_YRD_HBD_VARS()
#else
(void)use_hbd;
#endif
// Keep track of the row and column of the blocks we use so that we know
// if we are in the unrestricted motion border.
for (int r = 0; r < max_blocks_high; r += block_step) {
for (int c = 0, s = 0; c < max_blocks_wide; c += block_step, s += step) {
DECLARE_LOOP_VARS_BLOCK_YRD()
switch (tx_size) {
#if CONFIG_AV1_HIGHBITDEPTH
case TX_16X16:
if (use_hbd) {
aom_hadamard_16x16(src_diff, diff_stride, coeff);
av1_quantize_fp(coeff, 16 * 16, p->zbin_QTX, p->round_fp_QTX,
p->quant_fp_QTX, p->quant_shift_QTX, qcoeff,
dqcoeff, p->dequant_QTX, eob,
// default_scan_fp_16x16_transpose and
// av1_default_iscan_fp_16x16_transpose have to be
// used together.
default_scan_fp_16x16_transpose,
av1_default_iscan_fp_16x16_transpose);
} else {
aom_hadamard_lp_16x16(src_diff, diff_stride, low_coeff);
av1_quantize_lp(low_coeff, 16 * 16, p->round_fp_QTX,
p->quant_fp_QTX, low_qcoeff, low_dqcoeff,
p->dequant_QTX, eob,
// default_scan_lp_16x16_transpose and
// av1_default_iscan_lp_16x16_transpose have to be
// used together.
default_scan_lp_16x16_transpose,
av1_default_iscan_lp_16x16_transpose);
}
break;
case TX_8X8:
if (use_hbd) {
aom_hadamard_8x8(src_diff, diff_stride, coeff);
av1_quantize_fp(
coeff, 8 * 8, p->zbin_QTX, p->round_fp_QTX, p->quant_fp_QTX,
p->quant_shift_QTX, qcoeff, dqcoeff, p->dequant_QTX, eob,
default_scan_8x8_transpose, av1_default_iscan_8x8_transpose);
} else {
if (is_tx_8x8_dual_applicable) {
// The coeffs are pre-computed for the whole block, so re-assign
// low_coeff to the appropriate location.
const int block_offset = BLOCK_OFFSET(block + s);
low_coeff = (int16_t *)p->coeff + block_offset;
} else {
aom_hadamard_lp_8x8(src_diff, diff_stride, low_coeff);
}
av1_quantize_lp(
low_coeff, 8 * 8, p->round_fp_QTX, p->quant_fp_QTX, low_qcoeff,
low_dqcoeff, p->dequant_QTX, eob,
// default_scan_8x8_transpose and
// av1_default_iscan_8x8_transpose have to be used together.
default_scan_8x8_transpose, av1_default_iscan_8x8_transpose);
}
break;
default:
assert(tx_size == TX_4X4);
// In tx_size=4x4 case, aom_fdct4x4 and aom_fdct4x4_lp generate
// normal coefficients order, so we don't need to change the scan
// order here.
if (use_hbd) {
aom_fdct4x4(src_diff, coeff, diff_stride);
av1_quantize_fp(coeff, 4 * 4, p->zbin_QTX, p->round_fp_QTX,
p->quant_fp_QTX, p->quant_shift_QTX, qcoeff,
dqcoeff, p->dequant_QTX, eob, scan_order->scan,
scan_order->iscan);
} else {
aom_fdct4x4_lp(src_diff, low_coeff, diff_stride);
av1_quantize_lp(low_coeff, 4 * 4, p->round_fp_QTX, p->quant_fp_QTX,
low_qcoeff, low_dqcoeff, p->dequant_QTX, eob,
scan_order->scan, scan_order->iscan);
}
break;
#else
case TX_16X16:
aom_hadamard_lp_16x16(src_diff, diff_stride, low_coeff);
av1_quantize_lp(low_coeff, 16 * 16, p->round_fp_QTX, p->quant_fp_QTX,
low_qcoeff, low_dqcoeff, p->dequant_QTX, eob,
default_scan_lp_16x16_transpose,
av1_default_iscan_lp_16x16_transpose);
break;
case TX_8X8:
if (is_tx_8x8_dual_applicable) {
// The coeffs are pre-computed for the whole block, so re-assign
// low_coeff to the appropriate location.
const int block_offset = BLOCK_OFFSET(block + s);
low_coeff = (int16_t *)p->coeff + block_offset;
} else {
aom_hadamard_lp_8x8(src_diff, diff_stride, low_coeff);
}
av1_quantize_lp(low_coeff, 8 * 8, p->round_fp_QTX, p->quant_fp_QTX,
low_qcoeff, low_dqcoeff, p->dequant_QTX, eob,
default_scan_8x8_transpose,
av1_default_iscan_8x8_transpose);
break;
default:
aom_fdct4x4_lp(src_diff, low_coeff, diff_stride);
av1_quantize_lp(low_coeff, 4 * 4, p->round_fp_QTX, p->quant_fp_QTX,
low_qcoeff, low_dqcoeff, p->dequant_QTX, eob,
scan_order->scan, scan_order->iscan);
break;
#endif
}
assert(*eob <= 1024);
#if CONFIG_AV1_HIGHBITDEPTH
if (use_hbd)
update_yrd_loop_vars_hbd(x, &temp_skippable, step, *eob, coeff, qcoeff,
dqcoeff, this_rdc, &eob_cost,
(r * num_blk_skip_w + c) >> sh_blk_skip);
else
#endif
update_yrd_loop_vars(x, &temp_skippable, step, *eob, low_coeff,
low_qcoeff, low_dqcoeff, this_rdc, &eob_cost,
(r * num_blk_skip_w + c) >> sh_blk_skip);
}
block += row_step;
}
this_rdc->skip_txfm = *skippable = temp_skippable;
if (this_rdc->sse < INT64_MAX) {
this_rdc->sse = (this_rdc->sse << 6) >> 2;
if (temp_skippable) {
this_rdc->dist = 0;
this_rdc->dist = this_rdc->sse;
return;
}
}
// If skippable is set, rate gets clobbered later.
this_rdc->rate <<= (2 + AV1_PROB_COST_SHIFT);
this_rdc->rate += (eob_cost << AV1_PROB_COST_SHIFT);
}
// Explicitly enumerate the cases so the compiler can generate SIMD for the
// function. According to the disassembler, gcc generates SSE codes for each of
// the possible block sizes. The hottest case is tx_width 16, which takes up
// about 8% of the self cycle of av1_nonrd_pick_inter_mode_sb. Since
// av1_nonrd_pick_inter_mode_sb takes up about 3% of total encoding time, the
// potential room of improvement for writing AVX2 optimization is only 3% * 8% =
// 0.24% of total encoding time.
static AOM_INLINE void scale_square_buf_vals(int16_t *dst, const int tx_width,
const int16_t *src,
const int src_stride) {
#define DO_SCALING \
do { \
for (int idy = 0; idy < tx_width; ++idy) { \
for (int idx = 0; idx < tx_width; ++idx) { \
dst[idy * tx_width + idx] = src[idy * src_stride + idx] * 8; \
} \
} \
} while (0)
if (tx_width == 4) {
DO_SCALING;
} else if (tx_width == 8) {
DO_SCALING;
} else if (tx_width == 16) {
DO_SCALING;
} else {
assert(0);
}
#undef DO_SCALING
}
/*!\brief Calculates RD Cost when the block uses Identity transform.
* Note that thie function is only for low bit depth encoding, since it
* is called in real-time mode for now, which sets high bit depth to 0:
* -DCONFIG_AV1_HIGHBITDEPTH=0
*
* \ingroup nonrd_mode_search
* \callgraph
* \callergraph
* Calculates RD Cost. For low bit depth this function
* uses low-precision set of functions (16-bit) and 32 bit for high bit depth
* \param[in] x Pointer to structure holding all the data for
the current macroblock
* \param[in] this_rdc Pointer to calculated RD Cost
* \param[in] skippable Pointer to a flag indicating possible tx skip
* \param[in] bsize Current block size
* \param[in] tx_size Transform size
*
* \remark Nothing is returned. Instead, calculated RD cost is placed to
* \c this_rdc. \c skippable flag is set if all coefficients are zero.
*/
static void block_yrd_idtx(MACROBLOCK *x, RD_STATS *this_rdc, int *skippable,
const BLOCK_SIZE bsize, const TX_SIZE tx_size) {
MACROBLOCKD *xd = &x->e_mbd;
const struct macroblockd_plane *pd = &xd->plane[0];
struct macroblock_plane *const p = &x->plane[0];
assert(bsize < BLOCK_SIZES_ALL);
const int num_4x4_w = mi_size_wide[bsize];
const int num_4x4_h = mi_size_high[bsize];
const int step = 1 << (tx_size << 1);
const int block_step = (1 << tx_size);
const int max_blocks_wide =
num_4x4_w + (xd->mb_to_right_edge >= 0 ? 0 : xd->mb_to_right_edge >> 5);
const int max_blocks_high =
num_4x4_h + (xd->mb_to_bottom_edge >= 0 ? 0 : xd->mb_to_bottom_edge >> 5);
int eob_cost = 0;
const int bw = 4 * num_4x4_w;
const int bh = 4 * num_4x4_h;
const int num_blk_skip_w = num_4x4_w >> 1;
const int sh_blk_skip = 1;
// Keep the intermediate value on the stack here. Writing directly to
// skippable causes speed regression due to load-and-store issues in
// update_yrd_loop_vars.
int temp_skippable = 1;
int tx_wd = 0;
switch (tx_size) {
case TX_64X64:
assert(0); // Not implemented
break;
case TX_32X32:
assert(0); // Not used
break;
case TX_16X16: tx_wd = 16; break;
case TX_8X8: tx_wd = 8; break;
default:
assert(tx_size == TX_4X4);
tx_wd = 4;
break;
}
this_rdc->dist = 0;
this_rdc->rate = 0;
aom_subtract_block(bh, bw, p->src_diff, bw, p->src.buf, p->src.stride,
pd->dst.buf, pd->dst.stride);
// Keep track of the row and column of the blocks we use so that we know
// if we are in the unrestricted motion border.
DECLARE_BLOCK_YRD_BUFFERS()
DECLARE_BLOCK_YRD_VARS()
for (int r = 0; r < max_blocks_high; r += block_step) {
for (int c = 0, s = 0; c < max_blocks_wide; c += block_step, s += step) {
DECLARE_LOOP_VARS_BLOCK_YRD()
scale_square_buf_vals(low_coeff, tx_wd, src_diff, diff_stride);
av1_quantize_lp(low_coeff, tx_wd * tx_wd, p->round_fp_QTX,
p->quant_fp_QTX, low_qcoeff, low_dqcoeff, p->dequant_QTX,
eob, scan_order->scan, scan_order->iscan);
assert(*eob <= 1024);
update_yrd_loop_vars(x, &temp_skippable, step, *eob, low_coeff,
low_qcoeff, low_dqcoeff, this_rdc, &eob_cost,
(r * num_blk_skip_w + c) >> sh_blk_skip);
}
}
this_rdc->skip_txfm = *skippable = temp_skippable;
if (this_rdc->sse < INT64_MAX) {
this_rdc->sse = (this_rdc->sse << 6) >> 2;
if (temp_skippable) {
this_rdc->dist = 0;
this_rdc->dist = this_rdc->sse;
return;
}
}
// If skippable is set, rate gets clobbered later.
this_rdc->rate <<= (2 + AV1_PROB_COST_SHIFT);
this_rdc->rate += (eob_cost << AV1_PROB_COST_SHIFT);
}
static INLINE void init_mbmi(MB_MODE_INFO *mbmi, PREDICTION_MODE pred_mode,
MV_REFERENCE_FRAME ref_frame0,
MV_REFERENCE_FRAME ref_frame1,
const AV1_COMMON *cm) {
PALETTE_MODE_INFO *const pmi = &mbmi->palette_mode_info;
mbmi->ref_mv_idx = 0;
mbmi->mode = pred_mode;
mbmi->uv_mode = UV_DC_PRED;
mbmi->ref_frame[0] = ref_frame0;
mbmi->ref_frame[1] = ref_frame1;
pmi->palette_size[0] = 0;
pmi->palette_size[1] = 0;
mbmi->filter_intra_mode_info.use_filter_intra = 0;
mbmi->mv[0].as_int = mbmi->mv[1].as_int = 0;
mbmi->motion_mode = SIMPLE_TRANSLATION;
mbmi->num_proj_ref = 1;
mbmi->interintra_mode = 0;
set_default_interp_filters(mbmi, cm->features.interp_filter);
}
#if CONFIG_INTERNAL_STATS
static void store_coding_context(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx,
int mode_index) {
#else
static void store_coding_context(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
#endif // CONFIG_INTERNAL_STATS
MACROBLOCKD *const xd = &x->e_mbd;
TxfmSearchInfo *txfm_info = &x->txfm_search_info;
// Take a snapshot of the coding context so it can be
// restored if we decide to encode this way
ctx->rd_stats.skip_txfm = txfm_info->skip_txfm;
ctx->skippable = txfm_info->skip_txfm;
#if CONFIG_INTERNAL_STATS
ctx->best_mode_index = mode_index;
#endif // CONFIG_INTERNAL_STATS
ctx->mic = *xd->mi[0];
ctx->skippable = txfm_info->skip_txfm;
av1_copy_mbmi_ext_to_mbmi_ext_frame(&ctx->mbmi_ext_best, &x->mbmi_ext,
av1_ref_frame_type(xd->mi[0]->ref_frame));
}
static int get_pred_buffer(PRED_BUFFER *p, int len) {
for (int i = 0; i < len; i++) {
if (!p[i].in_use) {
p[i].in_use = 1;
return i;
}
}
return -1;
}
static void free_pred_buffer(PRED_BUFFER *p) {
if (p != NULL) p->in_use = 0;
}
static INLINE int get_drl_cost(const PREDICTION_MODE this_mode,
const int ref_mv_idx,
const MB_MODE_INFO_EXT *mbmi_ext,
const int (*const drl_mode_cost0)[2],
int8_t ref_frame_type) {
int cost = 0;
if (this_mode == NEWMV || this_mode == NEW_NEWMV) {
for (int idx = 0; idx < 2; ++idx) {
if (mbmi_ext->ref_mv_count[ref_frame_type] > idx + 1) {
uint8_t drl_ctx = av1_drl_ctx(mbmi_ext->weight[ref_frame_type], idx);
cost += drl_mode_cost0[drl_ctx][ref_mv_idx != idx];
if (ref_mv_idx == idx) return cost;
}
}
return cost;
}
if (have_nearmv_in_inter_mode(this_mode)) {
for (int idx = 1; idx < 3; ++idx) {
if (mbmi_ext->ref_mv_count[ref_frame_type] > idx + 1) {
uint8_t drl_ctx = av1_drl_ctx(mbmi_ext->weight[ref_frame_type], idx);
cost += drl_mode_cost0[drl_ctx][ref_mv_idx != (idx - 1)];
if (ref_mv_idx == (idx - 1)) return cost;
}
}
return cost;
}
return cost;
}
static int cost_mv_ref(const ModeCosts *const mode_costs, PREDICTION_MODE mode,
int16_t mode_context) {
if (is_inter_compound_mode(mode)) {
return mode_costs
->inter_compound_mode_cost[mode_context][INTER_COMPOUND_OFFSET(mode)];
}
int mode_cost = 0;
int16_t mode_ctx = mode_context & NEWMV_CTX_MASK;
assert(is_inter_mode(mode));
if (mode == NEWMV) {
mode_cost = mode_costs->newmv_mode_cost[mode_ctx][0];
return mode_cost;
} else {
mode_cost = mode_costs->newmv_mode_cost[mode_ctx][1];
mode_ctx = (mode_context >> GLOBALMV_OFFSET) & GLOBALMV_CTX_MASK;
if (mode == GLOBALMV) {
mode_cost += mode_costs->zeromv_mode_cost[mode_ctx][0];
return mode_cost;
} else {
mode_cost += mode_costs->zeromv_mode_cost[mode_ctx][1];
mode_ctx = (mode_context >> REFMV_OFFSET) & REFMV_CTX_MASK;
mode_cost += mode_costs->refmv_mode_cost[mode_ctx][mode != NEARESTMV];
return mode_cost;
}
}
}
static void newmv_diff_bias(MACROBLOCKD *xd, PREDICTION_MODE this_mode,
RD_STATS *this_rdc, BLOCK_SIZE bsize, int mv_row,
int mv_col, int speed, uint32_t spatial_variance,
CONTENT_STATE_SB content_state_sb) {
// Bias against MVs associated with NEWMV mode that are very different from
// top/left neighbors.
if (this_mode == NEWMV) {
int al_mv_average_row;
int al_mv_average_col;
int row_diff, col_diff;
int above_mv_valid = 0;
int left_mv_valid = 0;
int above_row = INVALID_MV_ROW_COL, above_col = INVALID_MV_ROW_COL;
int left_row = INVALID_MV_ROW_COL, left_col = INVALID_MV_ROW_COL;
if (bsize >= BLOCK_64X64 && content_state_sb.source_sad_nonrd != kHighSad &&
spatial_variance < 300 &&
(mv_row > 16 || mv_row < -16 || mv_col > 16 || mv_col < -16)) {
this_rdc->rdcost = this_rdc->rdcost << 2;
return;
}
if (xd->above_mbmi) {
above_mv_valid = xd->above_mbmi->mv[0].as_int != INVALID_MV;
above_row = xd->above_mbmi->mv[0].as_mv.row;
above_col = xd->above_mbmi->mv[0].as_mv.col;
}
if (xd->left_mbmi) {
left_mv_valid = xd->left_mbmi->mv[0].as_int != INVALID_MV;
left_row = xd->left_mbmi->mv[0].as_mv.row;
left_col = xd->left_mbmi->mv[0].as_mv.col;
}
if (above_mv_valid && left_mv_valid) {
al_mv_average_row = (above_row + left_row + 1) >> 1;
al_mv_average_col = (above_col + left_col + 1) >> 1;
} else if (above_mv_valid) {
al_mv_average_row = above_row;
al_mv_average_col = above_col;
} else if (left_mv_valid) {
al_mv_average_row = left_row;
al_mv_average_col = left_col;
} else {
al_mv_average_row = al_mv_average_col = 0;
}
row_diff = al_mv_average_row - mv_row;
col_diff = al_mv_average_col - mv_col;
if (row_diff > 80 || row_diff < -80 || col_diff > 80 || col_diff < -80) {
if (bsize >= BLOCK_32X32)
this_rdc->rdcost = this_rdc->rdcost << 1;
else
this_rdc->rdcost = 5 * this_rdc->rdcost >> 2;
}
} else {
// Bias for speed >= 8 for low spatial variance.
if (speed >= 8 && spatial_variance < 150 &&
(mv_row > 64 || mv_row < -64 || mv_col > 64 || mv_col < -64))
this_rdc->rdcost = 5 * this_rdc->rdcost >> 2;
}
}
static int64_t model_rd_for_sb_uv(AV1_COMP *cpi, BLOCK_SIZE plane_bsize,
MACROBLOCK *x, MACROBLOCKD *xd,
RD_STATS *this_rdc, int start_plane,
int stop_plane) {
// Note our transform coeffs are 8 times an orthogonal transform.
// Hence quantizer step is also 8 times. To get effective quantizer
// we need to divide by 8 before sending to modeling function.
unsigned int sse;
int rate;
int64_t dist;
int i;
int64_t tot_sse = 0;
this_rdc->rate = 0;
this_rdc->dist = 0;
this_rdc->skip_txfm = 0;
for (i = start_plane; i <= stop_plane; ++i) {
struct macroblock_plane *const p = &x->plane[i];
struct macroblockd_plane *const pd = &xd->plane[i];
const uint32_t dc_quant = p->dequant_QTX[0];
const uint32_t ac_quant = p->dequant_QTX[1];
const BLOCK_SIZE bs = plane_bsize;
unsigned int var;
if (!x->color_sensitivity[i - 1]) continue;
var = cpi->ppi->fn_ptr[bs].vf(p->src.buf, p->src.stride, pd->dst.buf,
pd->dst.stride, &sse);
assert(sse >= var);
tot_sse += sse;
av1_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bs],
dc_quant >> 3, &rate, &dist);
this_rdc->rate += rate >> 1;
this_rdc->dist += dist << 3;
av1_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bs], ac_quant >> 3,
&rate, &dist);
this_rdc->rate += rate;
this_rdc->dist += dist << 4;
}
if (this_rdc->rate == 0) {
this_rdc->skip_txfm = 1;
}
if (RDCOST(x->rdmult, this_rdc->rate, this_rdc->dist) >=
RDCOST(x->rdmult, 0, tot_sse << 4)) {
this_rdc->rate = 0;
this_rdc->dist = tot_sse << 4;
this_rdc->skip_txfm = 1;
}
return tot_sse;
}
/*!\cond */
struct estimate_block_intra_args {
AV1_COMP *cpi;
MACROBLOCK *x;
PREDICTION_MODE mode;
int skippable;
RD_STATS *rdc;
};
/*!\endcond */
/*!\brief Estimation of RD cost of an intra mode for Non-RD optimized case.
*
* \ingroup nonrd_mode_search
* \callgraph
* \callergraph
* Calculates RD Cost for an intra mode for a single TX block using Hadamard
* transform.
* \param[in] plane Color plane
* \param[in] block Index of a TX block in a prediction block
* \param[in] row Row of a current TX block
* \param[in] col Column of a current TX block
* \param[in] plane_bsize Block size of a current prediction block
* \param[in] tx_size Transform size
* \param[in] arg Pointer to a structure that holds parameters
* for intra mode search
*
* \remark Nothing is returned. Instead, best mode and RD Cost of the best mode
* are set in \c args->rdc and \c args->mode
*/
static void estimate_block_intra(int plane, int block, int row, int col,
BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
void *arg) {
struct estimate_block_intra_args *const args = arg;
AV1_COMP *const cpi = args->cpi;
AV1_COMMON *const cm = &cpi->common;
MACROBLOCK *const x = args->x;
MACROBLOCKD *const xd = &x->e_mbd;
struct macroblock_plane *const p = &x->plane[plane];
struct macroblockd_plane *const pd = &xd->plane[plane];
const BLOCK_SIZE bsize_tx = txsize_to_bsize[tx_size];
uint8_t *const src_buf_base = p->src.buf;
uint8_t *const dst_buf_base = pd->dst.buf;
const int64_t src_stride = p->src.stride;
const int64_t dst_stride = pd->dst.stride;
RD_STATS this_rdc;
(void)block;
(void)plane_bsize;
av1_predict_intra_block_facade(cm, xd, plane, col, row, tx_size);
av1_invalid_rd_stats(&this_rdc);
p->src.buf = &src_buf_base[4 * (row * src_stride + col)];
pd->dst.buf = &dst_buf_base[4 * (row * dst_stride + col)];
if (plane == 0) {
block_yrd(x, &this_rdc, &args->skippable, bsize_tx,
AOMMIN(tx_size, TX_16X16), 0);
} else {
model_rd_for_sb_uv(cpi, bsize_tx, x, xd, &this_rdc, plane, plane);
}
p->src.buf = src_buf_base;
pd->dst.buf = dst_buf_base;
args->rdc->rate += this_rdc.rate;
args->rdc->dist += this_rdc.dist;
}
static INLINE void update_thresh_freq_fact(AV1_COMP *cpi, MACROBLOCK *x,
BLOCK_SIZE bsize,
MV_REFERENCE_FRAME ref_frame,
THR_MODES best_mode_idx,
PREDICTION_MODE mode) {
const THR_MODES thr_mode_idx = mode_idx[ref_frame][mode_offset(mode)];
const BLOCK_SIZE min_size = AOMMAX(bsize - 3, BLOCK_4X4);
const BLOCK_SIZE max_size = AOMMIN(bsize + 6, BLOCK_128X128);
for (BLOCK_SIZE bs = min_size; bs <= max_size; bs += 3) {
int *freq_fact = &x->thresh_freq_fact[bs][thr_mode_idx];
if (thr_mode_idx == best_mode_idx) {
*freq_fact -= (*freq_fact >> 4);
} else {
*freq_fact =
AOMMIN(*freq_fact + RD_THRESH_INC,
cpi->sf.inter_sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT);
}
}
}
#if CONFIG_AV1_TEMPORAL_DENOISING
static void av1_pickmode_ctx_den_update(
AV1_PICKMODE_CTX_DEN *ctx_den, int64_t zero_last_cost_orig,
unsigned int ref_frame_cost[REF_FRAMES],
int_mv frame_mv[MB_MODE_COUNT][REF_FRAMES], int reuse_inter_pred,
BEST_PICKMODE *bp) {
ctx_den->zero_last_cost_orig = zero_last_cost_orig;
ctx_den->ref_frame_cost = ref_frame_cost;
ctx_den->frame_mv = frame_mv;
ctx_den->reuse_inter_pred = reuse_inter_pred;
ctx_den->best_tx_size = bp->best_tx_size;
ctx_den->best_mode = bp->best_mode;
ctx_den->best_ref_frame = bp->best_ref_frame;
ctx_den->best_pred_filter = bp->best_pred_filter;
ctx_den->best_mode_skip_txfm = bp->best_mode_skip_txfm;
}
static void recheck_zeromv_after_denoising(
AV1_COMP *cpi, MB_MODE_INFO *const mi, MACROBLOCK *x, MACROBLOCKD *const xd,
AV1_DENOISER_DECISION decision, AV1_PICKMODE_CTX_DEN *ctx_den,
struct buf_2d yv12_mb[4][MAX_MB_PLANE], RD_STATS *best_rdc,
BEST_PICKMODE *best_pickmode, BLOCK_SIZE bsize, int mi_row, int mi_col) {
// If INTRA or GOLDEN reference was selected, re-evaluate ZEROMV on
// denoised result. Only do this under noise conditions, and if rdcost of
// ZEROMV on original source is not significantly higher than rdcost of best
// mode.
if (cpi->noise_estimate.enabled && cpi->noise_estimate.level > kLow &&
ctx_den->zero_last_cost_orig < (best_rdc->rdcost << 3) &&
((ctx_den->best_ref_frame == INTRA_FRAME && decision >= FILTER_BLOCK) ||
(ctx_den->best_ref_frame == GOLDEN_FRAME &&
cpi->svc.number_spatial_layers == 1 &&
decision == FILTER_ZEROMV_BLOCK))) {
// Check if we should pick ZEROMV on denoised signal.
AV1_COMMON *const cm = &cpi->common;
RD_STATS this_rdc;
const ModeCosts *mode_costs = &x->mode_costs;
TxfmSearchInfo *txfm_info = &x->txfm_search_info;
MB_MODE_INFO_EXT *const mbmi_ext = &x->mbmi_ext;
mi->mode = GLOBALMV;
mi->ref_frame[0] = LAST_FRAME;
mi->ref_frame[1] = NONE_FRAME;
set_ref_ptrs(cm, xd, mi->ref_frame[0], NONE_FRAME);
mi->mv[0].as_int = 0;
mi->interp_filters = av1_broadcast_interp_filter(EIGHTTAP_REGULAR);
xd->plane[0].pre[0] = yv12_mb[LAST_FRAME][0];
av1_enc_build_inter_predictor_y(xd, mi_row, mi_col);
unsigned int var;
model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc, &var, 1, NULL);
const int16_t mode_ctx =
av1_mode_context_analyzer(mbmi_ext->mode_context, mi->ref_frame);
this_rdc.rate += cost_mv_ref(mode_costs, GLOBALMV, mode_ctx);
this_rdc.rate += ctx_den->ref_frame_cost[LAST_FRAME];
this_rdc.rdcost = RDCOST(x->rdmult, this_rdc.rate, this_rdc.dist);
txfm_info->skip_txfm = this_rdc.skip_txfm;
// Don't switch to ZEROMV if the rdcost for ZEROMV on denoised source
// is higher than best_ref mode (on original source).
if (this_rdc.rdcost > best_rdc->rdcost) {
this_rdc = *best_rdc;
mi->mode = best_pickmode->best_mode;
mi->ref_frame[0] = best_pickmode->best_ref_frame;
set_ref_ptrs(cm, xd, mi->ref_frame[0], NONE_FRAME);
mi->interp_filters = best_pickmode->best_pred_filter;
if (best_pickmode->best_ref_frame == INTRA_FRAME) {
mi->mv[0].as_int = INVALID_MV;
} else {
mi->mv[0].as_int = ctx_den
->frame_mv[best_pickmode->best_mode]
[best_pickmode->best_ref_frame]
.as_int;
if (ctx_den->reuse_inter_pred) {
xd->plane[0].pre[0] = yv12_mb[GOLDEN_FRAME][0];
av1_enc_build_inter_predictor_y(xd, mi_row, mi_col);
}
}
mi->tx_size = best_pickmode->best_tx_size;
txfm_info->skip_txfm = best_pickmode->best_mode_skip_txfm;
} else {
ctx_den->best_ref_frame = LAST_FRAME;
*best_rdc = this_rdc;
}
}
}
#endif // CONFIG_AV1_TEMPORAL_DENOISING
#define FILTER_SEARCH_SIZE 2
/*!\brief Searches for the best interpolation filter
*
* \ingroup nonrd_mode_search
* \callgraph
* \callergraph
* Iterates through subset of possible interpolation filters (EIGHTTAP_REGULAR,
* EIGTHTAP_SMOOTH, MULTITAP_SHARP, depending on FILTER_SEARCH_SIZE) and selects
* the one that gives lowest RD cost. RD cost is calculated using curvfit model.
* Support for dual filters (different filters in the x & y directions) is
* allowed if sf.interp_sf.disable_dual_filter = 0.
*
* \param[in] cpi Top-level encoder structure
* \param[in] x Pointer to structure holding all the
* data for the current macroblock
* \param[in] this_rdc Pointer to calculated RD Cost
* \param[in] inter_pred_params_sr Pointer to structure holding parameters of
inter prediction for single reference
* \param[in] mi_row Row index in 4x4 units
* \param[in] mi_col Column index in 4x4 units
* \param[in] tmp_buffer Pointer to a temporary buffer for
* prediction re-use
* \param[in] bsize Current block size
* \param[in] reuse_inter_pred Flag, indicating prediction re-use
* \param[out] this_mode_pred Pointer to store prediction buffer
* for prediction re-use
* \param[out] this_early_term Flag, indicating that transform can be
* skipped
* \param[out] var The residue variance of the current
* predictor.
* \param[in] use_model_yrd_large Flag, indicating special logic to handle
* large blocks
* \param[in] best_sse Best sse so far.
* \param[in] comp_pred Flag, indicating compound mode.
*
* \remark Nothing is returned. Instead, calculated RD cost is placed to
* \c this_rdc and best filter is placed to \c mi->interp_filters. In case
* \c reuse_inter_pred flag is set, this function also outputs
* \c this_mode_pred. Also \c this_early_temp is set if transform can be
* skipped
*/
static void search_filter_ref(AV1_COMP *cpi, MACROBLOCK *x, RD_STATS *this_rdc,
InterPredParams *inter_pred_params_sr, int mi_row,
int mi_col, PRED_BUFFER *tmp_buffer,
BLOCK_SIZE bsize, int reuse_inter_pred,
PRED_BUFFER **this_mode_pred,
int *this_early_term, unsigned int *var,
int use_model_yrd_large, int64_t best_sse,
int comp_pred) {
AV1_COMMON *const cm = &cpi->common;
MACROBLOCKD *const xd = &x->e_mbd;
struct macroblockd_plane *const pd = &xd->plane[0];
MB_MODE_INFO *const mi = xd->mi[0];
const int bw = block_size_wide[bsize];
int dim_factor =
(cpi->sf.interp_sf.disable_dual_filter == 0) ? FILTER_SEARCH_SIZE : 1;
RD_STATS pf_rd_stats[FILTER_SEARCH_SIZE * FILTER_SEARCH_SIZE] = { 0 };
TX_SIZE pf_tx_size[FILTER_SEARCH_SIZE * FILTER_SEARCH_SIZE] = { 0 };
PRED_BUFFER *current_pred = *this_mode_pred;
int best_skip = 0;
int best_early_term = 0;
int64_t best_cost = INT64_MAX;
int best_filter_index = -1;
SubpelParams subpel_params;
// Initialize inter prediction params at mode level for single reference
// mode.
if (!comp_pred)
init_inter_mode_params(&mi->mv[0].as_mv, inter_pred_params_sr,
&subpel_params, xd->block_ref_scale_factors[0],
pd->pre->width, pd->pre->height);
for (int i = 0; i < FILTER_SEARCH_SIZE * FILTER_SEARCH_SIZE; ++i) {
int64_t cost;
if (cpi->sf.interp_sf.disable_dual_filter &&
filters_ref_set[i].filter_x != filters_ref_set[i].filter_y)
continue;
mi->interp_filters.as_filters.x_filter = filters_ref_set[i].filter_x;
mi->interp_filters.as_filters.y_filter = filters_ref_set[i].filter_y;
if (!comp_pred)
av1_enc_build_inter_predictor_y_nonrd(xd, inter_pred_params_sr,
&subpel_params);
else
av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize, 0, 0);
unsigned int curr_var = UINT_MAX;
if (use_model_yrd_large)
model_skip_for_sb_y_large(cpi, bsize, mi_row, mi_col, x, xd,
&pf_rd_stats[i], this_early_term, 1, best_sse,
&curr_var, UINT_MAX);
else
model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rd_stats[i], &curr_var, 1, NULL);
pf_rd_stats[i].rate += av1_get_switchable_rate(
x, xd, cm->features.interp_filter, cm->seq_params->enable_dual_filter);
cost = RDCOST(x->rdmult, pf_rd_stats[i].rate, pf_rd_stats[i].dist);
pf_tx_size[i] = mi->tx_size;
if (cost < best_cost) {
*var = curr_var;
best_filter_index = i;
best_cost = cost;
best_skip = pf_rd_stats[i].skip_txfm;
best_early_term = *this_early_term;
if (reuse_inter_pred) {
if (*this_mode_pred != current_pred) {
free_pred_buffer(*this_mode_pred);
*this_mode_pred = current_pred;
}
current_pred = &tmp_buffer[get_pred_buffer(tmp_buffer, 3)];
pd->dst.buf = current_pred->data;
pd->dst.stride = bw;
}
}
}
assert(best_filter_index >= 0 &&
best_filter_index < dim_factor * FILTER_SEARCH_SIZE);
if (reuse_inter_pred && *this_mode_pred != current_pred)
free_pred_buffer(current_pred);
mi->interp_filters.as_filters.x_filter =
filters_ref_set[best_filter_index].filter_x;
mi->interp_filters.as_filters.y_filter =
filters_ref_set[best_filter_index].filter_y;
mi->tx_size = pf_tx_size[best_filter_index];
this_rdc->rate = pf_rd_stats[best_filter_index].rate;
this_rdc->dist = pf_rd_stats[best_filter_index].dist;
this_rdc->sse = pf_rd_stats[best_filter_index].sse;
this_rdc->skip_txfm = (best_skip || best_early_term);
*this_early_term = best_early_term;
if (reuse_inter_pred) {
pd->dst.buf = (*this_mode_pred)->data;
pd->dst.stride = (*this_mode_pred)->stride;
} else if (best_filter_index < dim_factor * FILTER_SEARCH_SIZE - 1) {
if (!comp_pred)
av1_enc_build_inter_predictor_y_nonrd(xd, inter_pred_params_sr,
&subpel_params);
else
av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize, 0, 0);
}
}
#if !CONFIG_REALTIME_ONLY
#define MOTION_MODE_SEARCH_SIZE 2
static AOM_INLINE int is_warped_mode_allowed(const AV1_COMP *cpi,
MACROBLOCK *const x,
const MB_MODE_INFO *mbmi) {
const FeatureFlags *const features = &cpi->common.features;
const MACROBLOCKD *xd = &x->e_mbd;
if (cpi->sf.inter_sf.extra_prune_warped) return 0;
if (has_second_ref(mbmi)) return 0;
MOTION_MODE last_motion_mode_allowed = SIMPLE_TRANSLATION;
if (features->switchable_motion_mode) {
// Determine which motion modes to search if more than SIMPLE_TRANSLATION
// is allowed.
last_motion_mode_allowed = motion_mode_allowed(
xd->global_motion, xd, mbmi, features->allow_warped_motion);
}
if (last_motion_mode_allowed == WARPED_CAUSAL) {
return 1;
}
return 0;
}
static void calc_num_proj_ref(AV1_COMP *cpi, MACROBLOCK *x, MB_MODE_INFO *mi) {
AV1_COMMON *const cm = &cpi->common;
MACROBLOCKD *const xd = &x->e_mbd;
const FeatureFlags *const features = &cm->features;
mi->num_proj_ref = 1;
WARP_SAMPLE_INFO *const warp_sample_info =
&x->warp_sample_info[mi->ref_frame[0]];
int *pts0 = warp_sample_info->pts;
int *pts_inref0 = warp_sample_info->pts_inref;
MOTION_MODE last_motion_mode_allowed = SIMPLE_TRANSLATION;
if (features->switchable_motion_mode) {
// Determine which motion modes to search if more than SIMPLE_TRANSLATION
// is allowed.
last_motion_mode_allowed = motion_mode_allowed(
xd->global_motion, xd, mi, features->allow_warped_motion);
}
if (last_motion_mode_allowed == WARPED_CAUSAL) {
if (warp_sample_info->num < 0) {
warp_sample_info->num = av1_findSamples(cm, xd, pts0, pts_inref0);
}
mi->num_proj_ref = warp_sample_info->num;
}
}
static void search_motion_mode(AV1_COMP *cpi, MACROBLOCK *x, RD_STATS *this_rdc,
int mi_row, int mi_col, BLOCK_SIZE bsize,
int *this_early_term, int use_model_yrd_large,
int *rate_mv, int64_t best_sse) {
AV1_COMMON *const cm = &cpi->common;
MACROBLOCKD *const xd = &x->e_mbd;
const FeatureFlags *const features = &cm->features;
MB_MODE_INFO *const mi = xd->mi[0];
RD_STATS pf_rd_stats[MOTION_MODE_SEARCH_SIZE] = { 0 };
int best_skip = 0;
int best_early_term = 0;
int64_t best_cost = INT64_MAX;
int best_mode_index = -1;
const int interp_filter = features->interp_filter;
const MOTION_MODE motion_modes[MOTION_MODE_SEARCH_SIZE] = {
SIMPLE_TRANSLATION, WARPED_CAUSAL
};
int mode_search_size = is_warped_mode_allowed(cpi, x, mi) ? 2 : 1;
WARP_SAMPLE_INFO *const warp_sample_info =
&x->warp_sample_info[mi->ref_frame[0]];
int *pts0 = warp_sample_info->pts;
int *pts_inref0 = warp_sample_info->pts_inref;
const int total_samples = mi->num_proj_ref;
if (total_samples == 0) {
// Do not search WARPED_CAUSAL if there are no samples to use to determine
// warped parameters.
mode_search_size = 1;
}
const MB_MODE_INFO base_mbmi = *mi;
MB_MODE_INFO best_mbmi;
for (int i = 0; i < mode_search_size; ++i) {
int64_t cost = INT64_MAX;
MOTION_MODE motion_mode = motion_modes[i];
*mi = base_mbmi;
mi->motion_mode = motion_mode;
if (motion_mode == SIMPLE_TRANSLATION) {
mi->interp_filters = av1_broadcast_interp_filter(EIGHTTAP_REGULAR);
av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize, 0, 0);
if (use_model_yrd_large)
model_skip_for_sb_y_large(cpi, bsize, mi_row, mi_col, x, xd,
&pf_rd_stats[i], this_early_term, 1, best_sse,
NULL, UINT_MAX);
else
model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rd_stats[i], NULL, 1, NULL);
pf_rd_stats[i].rate +=
av1_get_switchable_rate(x, xd, cm->features.interp_filter,
cm->seq_params->enable_dual_filter);
cost = RDCOST(x->rdmult, pf_rd_stats[i].rate, pf_rd_stats[i].dist);
} else if (motion_mode == WARPED_CAUSAL) {
int pts[SAMPLES_ARRAY_SIZE], pts_inref[SAMPLES_ARRAY_SIZE];
const ModeCosts *mode_costs = &x->mode_costs;
mi->wm_params.wmtype = DEFAULT_WMTYPE;
mi->interp_filters =
av1_broadcast_interp_filter(av1_unswitchable_filter(interp_filter));
memcpy(pts, pts0, total_samples * 2 * sizeof(*pts0));
memcpy(pts_inref, pts_inref0, total_samples * 2 * sizeof(*pts_inref0));
// Select the samples according to motion vector difference
if (mi->num_proj_ref > 1) {
mi->num_proj_ref = av1_selectSamples(&mi->mv[0].as_mv, pts, pts_inref,
mi->num_proj_ref, bsize);
}
// Compute the warped motion parameters with a least squares fit
// using the collected samples
if (!av1_find_projection(mi->num_proj_ref, pts, pts_inref, bsize,
mi->mv[0].as_mv.row, mi->mv[0].as_mv.col,
&mi->wm_params, mi_row, mi_col)) {
if (mi->mode == NEWMV) {
const int_mv mv0 = mi->mv[0];
const WarpedMotionParams wm_params0 = mi->wm_params;
const int num_proj_ref0 = mi->num_proj_ref;
const int_mv ref_mv = av1_get_ref_mv(x, 0);
SUBPEL_MOTION_SEARCH_PARAMS ms_params;
av1_make_default_subpel_ms_params(&ms_params, cpi, x, bsize,
&ref_mv.as_mv, NULL);
// Refine MV in a small range.
av1_refine_warped_mv(xd, cm, &ms_params, bsize, pts0, pts_inref0,
total_samples);
if (mi->mv[0].as_int == ref_mv.as_int) {
continue;
}
if (mv0.as_int != mi->mv[0].as_int) {
// Keep the refined MV and WM parameters.
int tmp_rate_mv = av1_mv_bit_cost(
&mi->mv[0].as_mv, &ref_mv.as_mv, x->mv_costs->nmv_joint_cost,
x->mv_costs->mv_cost_stack, MV_COST_WEIGHT);
*rate_mv = tmp_rate_mv;
} else {
// Restore the old MV and WM parameters.
mi->mv[0] = mv0;
mi->wm_params = wm_params0;
mi->num_proj_ref = num_proj_ref0;
}
}
// Build the warped predictor
av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize, 0,
av1_num_planes(cm) - 1);
if (use_model_yrd_large)
model_skip_for_sb_y_large(cpi, bsize, mi_row, mi_col, x, xd,
&pf_rd_stats[i], this_early_term, 1,
best_sse, NULL, UINT_MAX);
else
model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rd_stats[i], NULL, 1, NULL);
pf_rd_stats[i].rate +=
mode_costs->motion_mode_cost[bsize][mi->motion_mode];
cost = RDCOST(x->rdmult, pf_rd_stats[i].rate, pf_rd_stats[i].dist);
} else {
cost = INT64_MAX;
}
}
if (cost < best_cost) {
best_mode_index = i;
best_cost = cost;
best_skip = pf_rd_stats[i].skip_txfm;
best_early_term = *this_early_term;
best_mbmi = *mi;
}
}
assert(best_mode_index >= 0 && best_mode_index < FILTER_SEARCH_SIZE);
*mi = best_mbmi;
this_rdc->rate = pf_rd_stats[best_mode_index].rate;
this_rdc->dist = pf_rd_stats[best_mode_index].dist;
this_rdc->sse = pf_rd_stats[best_mode_index].sse;
this_rdc->skip_txfm = (best_skip || best_early_term);
*this_early_term = best_early_term;
if (best_mode_index < FILTER_SEARCH_SIZE - 1) {
av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize, 0, 0);
}
}
#endif // !CONFIG_REALTIME_ONLY
#define COLLECT_PICK_MODE_STAT 0
#define COLLECT_NON_SQR_STAT 0
#if COLLECT_PICK_MODE_STAT
#include "aom_ports/aom_timer.h"
typedef struct _mode_search_stat {
int32_t num_blocks[BLOCK_SIZES];
int64_t total_block_times[BLOCK_SIZES];
int32_t num_searches[BLOCK_SIZES][MB_MODE_COUNT];
int32_t num_nonskipped_searches[BLOCK_SIZES][MB_MODE_COUNT];
int64_t search_times[BLOCK_SIZES][MB_MODE_COUNT];
int64_t nonskipped_search_times[BLOCK_SIZES][MB_MODE_COUNT];
int64_t ms_time[BLOCK_SIZES][MB_MODE_COUNT];
int64_t ifs_time[BLOCK_SIZES][MB_MODE_COUNT];
int64_t model_rd_time[BLOCK_SIZES][MB_MODE_COUNT];
int64_t txfm_time[BLOCK_SIZES][MB_MODE_COUNT];
struct aom_usec_timer timer1;
struct aom_usec_timer timer2;
struct aom_usec_timer bsize_timer;
} mode_search_stat;
static mode_search_stat ms_stat;
static AOM_INLINE void print_stage_time(const char *stage_name,
int64_t stage_time,
int64_t total_time) {
printf(" %s: %ld (%f%%)\n", stage_name, stage_time,
100 * stage_time / (float)total_time);
}
static void print_time(const mode_search_stat *const ms_stat,
const BLOCK_SIZE bsize, const int mi_rows,
const int mi_cols, const int mi_row, const int mi_col) {
if ((mi_row + mi_size_high[bsize] >= mi_rows) &&
(mi_col + mi_size_wide[bsize] >= mi_cols)) {
int64_t total_time = 0l;
int32_t total_blocks = 0;
for (BLOCK_SIZE bs = 0; bs < BLOCK_SIZES; bs++) {
total_time += ms_stat->total_block_times[bs];
total_blocks += ms_stat->num_blocks[bs];
}
printf("\n");
for (BLOCK_SIZE bs = 0; bs < BLOCK_SIZES; bs++) {
if (ms_stat->num_blocks[bs] == 0) {
continue;
}
if (!COLLECT_NON_SQR_STAT && block_size_wide[bs] != block_size_high[bs]) {
continue;
}
printf("BLOCK_%dX%d Num %d, Time: %ld (%f%%), Avg_time %f:\n",
block_size_wide[bs], block_size_high[bs], ms_stat->num_blocks[bs],
ms_stat->total_block_times[bs],
100 * ms_stat->total_block_times[bs] / (float)total_time,
(float)ms_stat->total_block_times[bs] / ms_stat->num_blocks[bs]);
for (int j = 0; j < MB_MODE_COUNT; j++) {
if (ms_stat->nonskipped_search_times[bs][j] == 0) {
continue;
}
int64_t total_mode_time = ms_stat->nonskipped_search_times[bs][j];
printf(" Mode %d, %d/%d tps %f\n", j,
ms_stat->num_nonskipped_searches[bs][j],
ms_stat->num_searches[bs][j],
ms_stat->num_nonskipped_searches[bs][j] > 0
? (float)ms_stat->nonskipped_search_times[bs][j] /
ms_stat->num_nonskipped_searches[bs][j]
: 0l);
if (j >= INTER_MODE_START) {