blob: 3f4bba092e6346a76dbc29009ed24d50772e40a8 [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 <math.h>
#include "./aom_dsp_rtcd.h"
#include "./av1_rtcd.h"
#include "aom_dsp/aom_dsp_common.h"
#include "aom_mem/aom_mem.h"
#include "aom_ports/mem.h"
#include "aom_ports/system_state.h"
#include "av1/common/common.h"
#include "av1/common/entropy.h"
#include "av1/common/entropymode.h"
#include "av1/common/idct.h"
#include "av1/common/mvref_common.h"
#include "av1/common/pred_common.h"
#include "av1/common/quant_common.h"
#include "av1/common/reconinter.h"
#include "av1/common/reconintra.h"
#include "av1/common/scan.h"
#include "av1/common/seg_common.h"
#include "av1/encoder/aq_variance.h"
#include "av1/encoder/cost.h"
#include "av1/encoder/encodemb.h"
#include "av1/encoder/encodemv.h"
#include "av1/encoder/encoder.h"
#include "av1/encoder/hybrid_fwd_txfm.h"
#include "av1/encoder/mcomp.h"
#include "av1/encoder/quantize.h"
#include "av1/encoder/ratectrl.h"
#include "av1/encoder/rd.h"
#include "av1/encoder/rdopt.h"
#if CONFIG_EXT_REFS
#define LAST_FRAME_MODE_MASK \
((1 << INTRA_FRAME) | (1 << LAST2_FRAME) | (1 << LAST3_FRAME) | \
(1 << GOLDEN_FRAME) | (1 << BWDREF_FRAME) | (1 << ALTREF_FRAME)) // NOLINT
#define LAST2_FRAME_MODE_MASK \
((1 << INTRA_FRAME) | (1 << LAST_FRAME) | (1 << LAST3_FRAME) | \
(1 << GOLDEN_FRAME) | (1 << BWDREF_FRAME) | (1 << ALTREF_FRAME)) // NOLINT
#define LAST3_FRAME_MODE_MASK \
((1 << INTRA_FRAME) | (1 << LAST_FRAME) | (1 << LAST2_FRAME) | \
(1 << GOLDEN_FRAME) | (1 << BWDREF_FRAME) | (1 << ALTREF_FRAME)) // NOLINT
#define GOLDEN_FRAME_MODE_MASK \
((1 << INTRA_FRAME) | (1 << LAST_FRAME) | (1 << LAST2_FRAME) | \
(1 << LAST3_FRAME) | (1 << BWDREF_FRAME) | (1 << ALTREF_FRAME)) // NOLINT
#define BWDREF_FRAME_MODE_MASK \
((1 << INTRA_FRAME) | (1 << LAST_FRAME) | (1 << LAST2_FRAME) | \
(1 << LAST3_FRAME) | (1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME)) // NOLINT
#define ALTREF_FRAME_MODE_MASK \
((1 << INTRA_FRAME) | (1 << LAST_FRAME) | (1 << LAST2_FRAME) | \
(1 << LAST3_FRAME) | (1 << GOLDEN_FRAME) | (1 << BWDREF_FRAME)) // NOLINT
#else
#define LAST_FRAME_MODE_MASK \
((1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME) | (1 << INTRA_FRAME))
#define GOLDEN_FRAME_MODE_MASK \
((1 << LAST_FRAME) | (1 << ALTREF_FRAME) | (1 << INTRA_FRAME))
#define ALTREF_FRAME_MODE_MASK \
((1 << LAST_FRAME) | (1 << GOLDEN_FRAME) | (1 << INTRA_FRAME))
#endif // CONFIG_EXT_REFS
#if CONFIG_EXT_REFS
#define SECOND_REF_FRAME_MASK ((1 << ALTREF_FRAME) | (1 << BWDREF_FRAME) | 0x01)
#else
#define SECOND_REF_FRAME_MASK ((1 << ALTREF_FRAME) | 0x01)
#endif // CONFIG_EXT_REFS
#define MIN_EARLY_TERM_INDEX 3
#define NEW_MV_DISCOUNT_FACTOR 8
const double ext_tx_th = 0.99;
typedef struct {
PREDICTION_MODE mode;
MV_REFERENCE_FRAME ref_frame[2];
} MODE_DEFINITION;
typedef struct { MV_REFERENCE_FRAME ref_frame[2]; } REF_DEFINITION;
struct rdcost_block_args {
MACROBLOCK *x;
ENTROPY_CONTEXT t_above[16];
ENTROPY_CONTEXT t_left[16];
int this_rate;
int64_t this_dist;
int64_t this_sse;
int64_t this_rd;
int64_t best_rd;
int exit_early;
int use_fast_coef_costing;
const SCAN_ORDER *scan_order;
uint8_t skippable;
};
#define LAST_NEW_MV_INDEX 6
static const MODE_DEFINITION av1_mode_order[MAX_MODES] = {
{ NEARESTMV, { LAST_FRAME, NONE } },
#if CONFIG_EXT_REFS
{ NEARESTMV, { LAST2_FRAME, NONE } },
{ NEARESTMV, { LAST3_FRAME, NONE } },
{ NEARESTMV, { BWDREF_FRAME, NONE } },
#endif // CONFIG_EXT_REFS
{ NEARESTMV, { ALTREF_FRAME, NONE } },
{ NEARESTMV, { GOLDEN_FRAME, NONE } },
{ DC_PRED, { INTRA_FRAME, NONE } },
{ NEWMV, { LAST_FRAME, NONE } },
#if CONFIG_EXT_REFS
{ NEWMV, { LAST2_FRAME, NONE } },
{ NEWMV, { LAST3_FRAME, NONE } },
{ NEWMV, { BWDREF_FRAME, NONE } },
#endif // CONFIG_EXT_REFS
{ NEWMV, { ALTREF_FRAME, NONE } },
{ NEWMV, { GOLDEN_FRAME, NONE } },
{ NEARMV, { LAST_FRAME, NONE } },
#if CONFIG_EXT_REFS
{ NEARMV, { LAST2_FRAME, NONE } },
{ NEARMV, { LAST3_FRAME, NONE } },
{ NEARMV, { BWDREF_FRAME, NONE } },
#endif // CONFIG_EXT_REFS
{ NEARMV, { ALTREF_FRAME, NONE } },
{ NEARMV, { GOLDEN_FRAME, NONE } },
{ ZEROMV, { LAST_FRAME, NONE } },
#if CONFIG_EXT_REFS
{ ZEROMV, { LAST2_FRAME, NONE } },
{ ZEROMV, { LAST3_FRAME, NONE } },
{ ZEROMV, { BWDREF_FRAME, NONE } },
#endif // CONFIG_EXT_REFS
{ ZEROMV, { GOLDEN_FRAME, NONE } },
{ ZEROMV, { ALTREF_FRAME, NONE } },
// TODO(zoeliu): May need to reconsider the order on the modes to check
{ NEARESTMV, { LAST_FRAME, ALTREF_FRAME } },
#if CONFIG_EXT_REFS
{ NEARESTMV, { LAST2_FRAME, ALTREF_FRAME } },
{ NEARESTMV, { LAST3_FRAME, ALTREF_FRAME } },
#endif // CONFIG_EXT_REFS
{ NEARESTMV, { GOLDEN_FRAME, ALTREF_FRAME } },
#if CONFIG_EXT_REFS
{ NEARESTMV, { LAST_FRAME, BWDREF_FRAME } },
{ NEARESTMV, { LAST2_FRAME, BWDREF_FRAME } },
{ NEARESTMV, { LAST3_FRAME, BWDREF_FRAME } },
{ NEARESTMV, { GOLDEN_FRAME, BWDREF_FRAME } },
#endif // CONFIG_EXT_REFS
{ TM_PRED, { INTRA_FRAME, NONE } },
{ NEARMV, { LAST_FRAME, ALTREF_FRAME } },
{ NEWMV, { LAST_FRAME, ALTREF_FRAME } },
#if CONFIG_EXT_REFS
{ NEARMV, { LAST2_FRAME, ALTREF_FRAME } },
{ NEWMV, { LAST2_FRAME, ALTREF_FRAME } },
{ NEARMV, { LAST3_FRAME, ALTREF_FRAME } },
{ NEWMV, { LAST3_FRAME, ALTREF_FRAME } },
#endif // CONFIG_EXT_REFS
{ NEARMV, { GOLDEN_FRAME, ALTREF_FRAME } },
{ NEWMV, { GOLDEN_FRAME, ALTREF_FRAME } },
#if CONFIG_EXT_REFS
{ NEARMV, { LAST_FRAME, BWDREF_FRAME } },
{ NEWMV, { LAST_FRAME, BWDREF_FRAME } },
{ NEARMV, { LAST2_FRAME, BWDREF_FRAME } },
{ NEWMV, { LAST2_FRAME, BWDREF_FRAME } },
{ NEARMV, { LAST3_FRAME, BWDREF_FRAME } },
{ NEWMV, { LAST3_FRAME, BWDREF_FRAME } },
{ NEARMV, { GOLDEN_FRAME, BWDREF_FRAME } },
{ NEWMV, { GOLDEN_FRAME, BWDREF_FRAME } },
#endif // CONFIG_EXT_REFS
{ ZEROMV, { LAST_FRAME, ALTREF_FRAME } },
#if CONFIG_EXT_REFS
{ ZEROMV, { LAST2_FRAME, ALTREF_FRAME } },
{ ZEROMV, { LAST3_FRAME, ALTREF_FRAME } },
#endif // CONFIG_EXT_REFS
{ ZEROMV, { GOLDEN_FRAME, ALTREF_FRAME } },
#if CONFIG_EXT_REFS
{ ZEROMV, { LAST_FRAME, BWDREF_FRAME } },
{ ZEROMV, { LAST2_FRAME, BWDREF_FRAME } },
{ ZEROMV, { LAST3_FRAME, BWDREF_FRAME } },
{ ZEROMV, { GOLDEN_FRAME, BWDREF_FRAME } },
#endif // CONFIG_EXT_REFS
{ H_PRED, { INTRA_FRAME, NONE } },
{ V_PRED, { INTRA_FRAME, NONE } },
{ D135_PRED, { INTRA_FRAME, NONE } },
{ D207_PRED, { INTRA_FRAME, NONE } },
{ D153_PRED, { INTRA_FRAME, NONE } },
{ D63_PRED, { INTRA_FRAME, NONE } },
{ D117_PRED, { INTRA_FRAME, NONE } },
{ D45_PRED, { INTRA_FRAME, NONE } },
};
static const REF_DEFINITION av1_ref_order[MAX_REFS] = {
{ { LAST_FRAME, NONE } },
#if CONFIG_EXT_REFS
{ { LAST2_FRAME, NONE } }, { { LAST3_FRAME, NONE } },
#endif // CONFIG_EXT_REFS
{ { GOLDEN_FRAME, NONE } },
#if CONFIG_EXT_REFS
{ { BWDREF_FRAME, NONE } },
#endif // CONFIG_EXT_REFS
{ { ALTREF_FRAME, NONE } }, { { LAST_FRAME, ALTREF_FRAME } },
#if CONFIG_EXT_REFS
{ { LAST2_FRAME, ALTREF_FRAME } }, { { LAST3_FRAME, ALTREF_FRAME } },
#endif // CONFIG_EXT_REFS
{ { GOLDEN_FRAME, ALTREF_FRAME } },
#if CONFIG_EXT_REFS
{ { LAST_FRAME, BWDREF_FRAME } }, { { LAST2_FRAME, BWDREF_FRAME } },
{ { LAST3_FRAME, BWDREF_FRAME } }, { { GOLDEN_FRAME, BWDREF_FRAME } },
#endif // CONFIG_EXT_REFS
{ { INTRA_FRAME, NONE } },
};
static void swap_block_ptr(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx, int m, int n,
int min_plane, int max_plane) {
int i;
for (i = min_plane; i < max_plane; ++i) {
struct macroblock_plane *const p = &x->plane[i];
struct macroblockd_plane *const pd = &x->e_mbd.plane[i];
p->coeff = ctx->coeff_pbuf[i][m];
p->qcoeff = ctx->qcoeff_pbuf[i][m];
pd->dqcoeff = ctx->dqcoeff_pbuf[i][m];
p->eobs = ctx->eobs_pbuf[i][m];
ctx->coeff_pbuf[i][m] = ctx->coeff_pbuf[i][n];
ctx->qcoeff_pbuf[i][m] = ctx->qcoeff_pbuf[i][n];
ctx->dqcoeff_pbuf[i][m] = ctx->dqcoeff_pbuf[i][n];
ctx->eobs_pbuf[i][m] = ctx->eobs_pbuf[i][n];
ctx->coeff_pbuf[i][n] = p->coeff;
ctx->qcoeff_pbuf[i][n] = p->qcoeff;
ctx->dqcoeff_pbuf[i][n] = pd->dqcoeff;
ctx->eobs_pbuf[i][n] = p->eobs;
}
}
static void model_rd_for_sb(const AV1_COMP *const cpi, BLOCK_SIZE bsize,
MACROBLOCK *x, MACROBLOCKD *xd, int *out_rate_sum,
int64_t *out_dist_sum, int *skip_txfm_sb,
int64_t *skip_sse_sb) {
// 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.
int i;
int64_t rate_sum = 0;
int64_t dist_sum = 0;
const int ref = xd->mi[0]->mbmi.ref_frame[0];
unsigned int sse;
unsigned int var = 0;
unsigned int sum_sse = 0;
int64_t total_sse = 0;
int skip_flag = 1;
const int shift = 6;
int rate;
int64_t dist;
const int dequant_shift =
#if CONFIG_AOM_HIGHBITDEPTH
(xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd - 5 :
#endif // CONFIG_AOM_HIGHBITDEPTH
3;
x->pred_sse[ref] = 0;
for (i = 0; i < MAX_MB_PLANE; ++i) {
struct macroblock_plane *const p = &x->plane[i];
struct macroblockd_plane *const pd = &xd->plane[i];
const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
const TX_SIZE max_tx_size = max_txsize_lookup[bs];
const BLOCK_SIZE unit_size = txsize_to_bsize[max_tx_size];
const int64_t dc_thr = p->quant_thred[0] >> shift;
const int64_t ac_thr = p->quant_thred[1] >> shift;
// The low thresholds are used to measure if the prediction errors are
// low enough so that we can skip the mode search.
const int64_t low_dc_thr = AOMMIN(50, dc_thr >> 2);
const int64_t low_ac_thr = AOMMIN(80, ac_thr >> 2);
int bw = 1 << (b_width_log2_lookup[bs] - b_width_log2_lookup[unit_size]);
int bh = 1 << (b_height_log2_lookup[bs] - b_width_log2_lookup[unit_size]);
int idx, idy;
int lw = b_width_log2_lookup[unit_size] + 2;
int lh = b_height_log2_lookup[unit_size] + 2;
sum_sse = 0;
for (idy = 0; idy < bh; ++idy) {
for (idx = 0; idx < bw; ++idx) {
uint8_t *src = p->src.buf + (idy * p->src.stride << lh) + (idx << lw);
uint8_t *dst = pd->dst.buf + (idy * pd->dst.stride << lh) + (idx << lh);
int block_idx = (idy << 1) + idx;
int low_err_skip = 0;
var = cpi->fn_ptr[unit_size].vf(src, p->src.stride, dst, pd->dst.stride,
&sse);
x->bsse[(i << 2) + block_idx] = sse;
sum_sse += sse;
x->skip_txfm[(i << 2) + block_idx] = SKIP_TXFM_NONE;
if (!x->select_tx_size) {
// Check if all ac coefficients can be quantized to zero.
if (var < ac_thr || var == 0) {
x->skip_txfm[(i << 2) + block_idx] = SKIP_TXFM_AC_ONLY;
// Check if dc coefficient can be quantized to zero.
if (sse - var < dc_thr || sse == var) {
x->skip_txfm[(i << 2) + block_idx] = SKIP_TXFM_AC_DC;
if (!sse || (var < low_ac_thr && sse - var < low_dc_thr))
low_err_skip = 1;
}
}
}
if (skip_flag && !low_err_skip) skip_flag = 0;
if (i == 0) x->pred_sse[ref] += sse;
}
}
total_sse += sum_sse;
// Fast approximate the modelling function.
if (cpi->sf.simple_model_rd_from_var) {
int64_t rate;
const int64_t square_error = sum_sse;
int quantizer = (pd->dequant[1] >> dequant_shift);
if (quantizer < 120)
rate = (square_error * (280 - quantizer)) >> (16 - AV1_PROB_COST_SHIFT);
else
rate = 0;
dist = (square_error * quantizer) >> 8;
rate_sum += rate;
dist_sum += dist;
} else {
av1_model_rd_from_var_lapndz(sum_sse, num_pels_log2_lookup[bs],
pd->dequant[1] >> dequant_shift, &rate,
&dist);
rate_sum += rate;
dist_sum += dist;
}
}
*skip_txfm_sb = skip_flag;
*skip_sse_sb = total_sse << 4;
*out_rate_sum = (int)rate_sum;
*out_dist_sum = dist_sum << 4;
}
int64_t av1_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff,
intptr_t block_size, int64_t *ssz) {
int i;
int64_t error = 0, sqcoeff = 0;
for (i = 0; i < block_size; i++) {
const int diff = coeff[i] - dqcoeff[i];
error += diff * diff;
sqcoeff += coeff[i] * coeff[i];
}
*ssz = sqcoeff;
return error;
}
int64_t av1_block_error_fp_c(const int16_t *coeff, const int16_t *dqcoeff,
int block_size) {
int i;
int64_t error = 0;
for (i = 0; i < block_size; i++) {
const int diff = coeff[i] - dqcoeff[i];
error += diff * diff;
}
return error;
}
#if CONFIG_AOM_HIGHBITDEPTH
int64_t av1_highbd_block_error_c(const tran_low_t *coeff,
const tran_low_t *dqcoeff, intptr_t block_size,
int64_t *ssz, int bd) {
int i;
int64_t error = 0, sqcoeff = 0;
int shift = 2 * (bd - 8);
int rounding = shift > 0 ? 1 << (shift - 1) : 0;
for (i = 0; i < block_size; i++) {
const int64_t diff = coeff[i] - dqcoeff[i];
error += diff * diff;
sqcoeff += (int64_t)coeff[i] * (int64_t)coeff[i];
}
assert(error >= 0 && sqcoeff >= 0);
error = (error + rounding) >> shift;
sqcoeff = (sqcoeff + rounding) >> shift;
*ssz = sqcoeff;
return error;
}
#endif // CONFIG_AOM_HIGHBITDEPTH
/* The trailing '0' is a terminator which is used inside cost_coeffs() to
* decide whether to include cost of a trailing EOB node or not (i.e. we
* can skip this if the last coefficient in this transform block, e.g. the
* 16th coefficient in a 4x4 block or the 64th coefficient in a 8x8 block,
* were non-zero). */
static const int16_t band_counts[TX_SIZES][8] = {
{ 1, 2, 3, 4, 3, 16 - 13, 0 },
{ 1, 2, 3, 4, 11, 64 - 21, 0 },
{ 1, 2, 3, 4, 11, 256 - 21, 0 },
{ 1, 2, 3, 4, 11, 1024 - 21, 0 },
};
static int cost_coeffs(MACROBLOCK *x, int plane, int block, ENTROPY_CONTEXT *A,
ENTROPY_CONTEXT *L, TX_SIZE tx_size, const int16_t *scan,
const int16_t *nb, int use_fast_coef_costing) {
MACROBLOCKD *const xd = &x->e_mbd;
MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
const struct macroblock_plane *p = &x->plane[plane];
const struct macroblockd_plane *pd = &xd->plane[plane];
const PLANE_TYPE type = pd->plane_type;
const int16_t *band_count = &band_counts[tx_size][1];
const int eob = p->eobs[block];
const tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
unsigned int(*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
x->token_costs[tx_size][type][is_inter_block(mbmi)];
uint8_t token_cache[32 * 32];
int pt = combine_entropy_contexts(*A, *L);
int c, cost;
#if CONFIG_AOM_HIGHBITDEPTH
const int *cat6_high_cost = av1_get_high_cost_table(xd->bd);
#else
const int *cat6_high_cost = av1_get_high_cost_table(8);
#endif
// Check for consistency of tx_size with mode info
assert(type == PLANE_TYPE_Y ? mbmi->tx_size == tx_size
: get_uv_tx_size(mbmi, pd) == tx_size);
if (eob == 0) {
// single eob token
cost = token_costs[0][0][pt][EOB_TOKEN];
c = 0;
} else {
int band_left = *band_count++;
// dc token
int v = qcoeff[0];
int16_t prev_t;
EXTRABIT e;
av1_get_token_extra(v, &prev_t, &e);
cost =
(*token_costs)[0][pt][prev_t] + av1_get_cost(prev_t, e, cat6_high_cost);
token_cache[0] = av1_pt_energy_class[prev_t];
++token_costs;
// ac tokens
for (c = 1; c < eob; c++) {
const int rc = scan[c];
int16_t t;
v = qcoeff[rc];
av1_get_token_extra(v, &t, &e);
if (use_fast_coef_costing) {
cost += (*token_costs)[!prev_t][!prev_t][t] +
av1_get_cost(t, e, cat6_high_cost);
} else {
pt = get_coef_context(nb, token_cache, c);
cost +=
(*token_costs)[!prev_t][pt][t] + av1_get_cost(t, e, cat6_high_cost);
token_cache[rc] = av1_pt_energy_class[t];
}
prev_t = t;
if (!--band_left) {
band_left = *band_count++;
++token_costs;
}
}
// eob token
if (band_left) {
if (use_fast_coef_costing) {
cost += (*token_costs)[0][!prev_t][EOB_TOKEN];
} else {
pt = get_coef_context(nb, token_cache, c);
cost += (*token_costs)[0][pt][EOB_TOKEN];
}
}
}
// is eob first coefficient;
*A = *L = (c > 0);
return cost;
}
static void dist_block(MACROBLOCK *x, int plane, int block, TX_SIZE tx_size,
int64_t *out_dist, int64_t *out_sse) {
const int ss_txfrm_size = tx_size << 1;
MACROBLOCKD *const xd = &x->e_mbd;
const struct macroblock_plane *const p = &x->plane[plane];
const struct macroblockd_plane *const pd = &xd->plane[plane];
int64_t this_sse;
int shift = tx_size == TX_32X32 ? 0 : 2;
tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
#if CONFIG_AOM_HIGHBITDEPTH
const int bd = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd : 8;
*out_dist = av1_highbd_block_error(coeff, dqcoeff, 16 << ss_txfrm_size,
&this_sse, bd) >>
shift;
#else
*out_dist =
av1_block_error(coeff, dqcoeff, 16 << ss_txfrm_size, &this_sse) >> shift;
#endif // CONFIG_AOM_HIGHBITDEPTH
*out_sse = this_sse >> shift;
}
static int rate_block(int plane, int block, int blk_row, int blk_col,
TX_SIZE tx_size, struct rdcost_block_args *args) {
return cost_coeffs(args->x, plane, block, args->t_above + blk_col,
args->t_left + blk_row, tx_size, args->scan_order->scan,
args->scan_order->neighbors, args->use_fast_coef_costing);
}
static void block_rd_txfm(int plane, int block, int blk_row, int blk_col,
BLOCK_SIZE plane_bsize, TX_SIZE tx_size, void *arg) {
struct rdcost_block_args *args = arg;
MACROBLOCK *const x = args->x;
MACROBLOCKD *const xd = &x->e_mbd;
MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
int64_t rd1, rd2, rd;
int rate;
int64_t dist;
int64_t sse;
if (args->exit_early) return;
if (!is_inter_block(mbmi)) {
struct encode_b_args arg = { x, NULL, &mbmi->skip };
av1_encode_block_intra(plane, block, blk_row, blk_col, plane_bsize, tx_size,
&arg);
dist_block(x, plane, block, tx_size, &dist, &sse);
} else if (max_txsize_lookup[plane_bsize] == tx_size) {
if (x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))] ==
SKIP_TXFM_NONE) {
// full forward transform and quantization
av1_xform_quant(x, plane, block, blk_row, blk_col, plane_bsize, tx_size);
dist_block(x, plane, block, tx_size, &dist, &sse);
} else if (x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))] ==
SKIP_TXFM_AC_ONLY) {
// compute DC coefficient
tran_low_t *const coeff = BLOCK_OFFSET(x->plane[plane].coeff, block);
tran_low_t *const dqcoeff = BLOCK_OFFSET(xd->plane[plane].dqcoeff, block);
av1_xform_quant_dc(x, plane, block, blk_row, blk_col, plane_bsize,
tx_size);
sse = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4;
dist = sse;
if (x->plane[plane].eobs[block]) {
const int64_t orig_sse = (int64_t)coeff[0] * coeff[0];
const int64_t resd_sse = coeff[0] - dqcoeff[0];
int64_t dc_correct = orig_sse - resd_sse * resd_sse;
#if CONFIG_AOM_HIGHBITDEPTH
dc_correct >>= ((xd->bd - 8) * 2);
#endif
if (tx_size != TX_32X32) dc_correct >>= 2;
dist = AOMMAX(0, sse - dc_correct);
}
} else {
// SKIP_TXFM_AC_DC
// skip forward transform
x->plane[plane].eobs[block] = 0;
sse = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4;
dist = sse;
}
} else {
// full forward transform and quantization
av1_xform_quant(x, plane, block, blk_row, blk_col, plane_bsize, tx_size);
dist_block(x, plane, block, tx_size, &dist, &sse);
}
rd = RDCOST(x->rdmult, x->rddiv, 0, dist);
if (args->this_rd + rd > args->best_rd) {
args->exit_early = 1;
return;
}
rate = rate_block(plane, block, blk_row, blk_col, tx_size, args);
rd1 = RDCOST(x->rdmult, x->rddiv, rate, dist);
rd2 = RDCOST(x->rdmult, x->rddiv, 0, sse);
// TODO(jingning): temporarily enabled only for luma component
rd = AOMMIN(rd1, rd2);
if (plane == 0)
x->zcoeff_blk[tx_size][block] =
!x->plane[plane].eobs[block] ||
(rd1 > rd2 && !xd->lossless[mbmi->segment_id]);
args->this_rate += rate;
args->this_dist += dist;
args->this_sse += sse;
args->this_rd += rd;
if (args->this_rd > args->best_rd) {
args->exit_early = 1;
return;
}
args->skippable &= !x->plane[plane].eobs[block];
}
static void txfm_rd_in_plane(MACROBLOCK *x, int *rate, int64_t *distortion,
int *skippable, int64_t *sse, int64_t ref_best_rd,
int plane, BLOCK_SIZE bsize, TX_SIZE tx_size,
int use_fast_coef_casting) {
MACROBLOCKD *const xd = &x->e_mbd;
const struct macroblockd_plane *const pd = &xd->plane[plane];
TX_TYPE tx_type;
struct rdcost_block_args args;
av1_zero(args);
args.x = x;
args.best_rd = ref_best_rd;
args.use_fast_coef_costing = use_fast_coef_casting;
args.skippable = 1;
if (plane == 0) xd->mi[0]->mbmi.tx_size = tx_size;
av1_get_entropy_contexts(bsize, tx_size, pd, args.t_above, args.t_left);
tx_type = get_tx_type(pd->plane_type, xd, 0);
args.scan_order = get_scan(tx_size, tx_type);
av1_foreach_transformed_block_in_plane(xd, bsize, plane, block_rd_txfm,
&args);
if (args.exit_early) {
*rate = INT_MAX;
*distortion = INT64_MAX;
*sse = INT64_MAX;
*skippable = 0;
} else {
*distortion = args.this_dist;
*rate = args.this_rate;
*sse = args.this_sse;
*skippable = args.skippable;
}
}
static void choose_largest_tx_size(const AV1_COMP *const cpi, MACROBLOCK *x,
int *rate, int64_t *distortion, int *skip,
int64_t *sse, int64_t ref_best_rd,
BLOCK_SIZE bs) {
const TX_SIZE max_tx_size = max_txsize_lookup[bs];
const AV1_COMMON *const cm = &cpi->common;
const TX_SIZE largest_tx_size = tx_mode_to_biggest_tx_size[cm->tx_mode];
MACROBLOCKD *const xd = &x->e_mbd;
MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
TX_TYPE tx_type, best_tx_type = DCT_DCT;
int r, s;
int64_t d, psse, this_rd, best_rd = INT64_MAX;
aom_prob skip_prob = av1_get_skip_prob(cm, xd);
int s0 = av1_cost_bit(skip_prob, 0);
int s1 = av1_cost_bit(skip_prob, 1);
const int is_inter = is_inter_block(mbmi);
mbmi->tx_size = AOMMIN(max_tx_size, largest_tx_size);
if (mbmi->tx_size < TX_32X32 && !xd->lossless[mbmi->segment_id]) {
for (tx_type = 0; tx_type < TX_TYPES; ++tx_type) {
mbmi->tx_type = tx_type;
txfm_rd_in_plane(x, &r, &d, &s, &psse, ref_best_rd, 0, bs, mbmi->tx_size,
cpi->sf.use_fast_coef_costing);
if (r == INT_MAX) continue;
if (is_inter)
r += cpi->inter_tx_type_costs[mbmi->tx_size][mbmi->tx_type];
else
r += cpi->intra_tx_type_costs[mbmi->tx_size]
[intra_mode_to_tx_type_context[mbmi->mode]]
[mbmi->tx_type];
if (s)
this_rd = RDCOST(x->rdmult, x->rddiv, s1, psse);
else
this_rd = RDCOST(x->rdmult, x->rddiv, r + s0, d);
if (is_inter && !xd->lossless[mbmi->segment_id] && !s)
this_rd = AOMMIN(this_rd, RDCOST(x->rdmult, x->rddiv, s1, psse));
if (this_rd < ((best_tx_type == DCT_DCT) ? ext_tx_th : 1) * best_rd) {
best_rd = this_rd;
best_tx_type = mbmi->tx_type;
}
}
}
mbmi->tx_type = best_tx_type;
txfm_rd_in_plane(x, rate, distortion, skip, sse, ref_best_rd, 0, bs,
mbmi->tx_size, cpi->sf.use_fast_coef_costing);
if (mbmi->tx_size < TX_32X32 && !xd->lossless[mbmi->segment_id] &&
*rate != INT_MAX) {
if (is_inter)
*rate += cpi->inter_tx_type_costs[mbmi->tx_size][mbmi->tx_type];
else
*rate += cpi->intra_tx_type_costs
[mbmi->tx_size][intra_mode_to_tx_type_context[mbmi->mode]]
[mbmi->tx_type];
}
}
static void choose_smallest_tx_size(const AV1_COMP *const cpi, MACROBLOCK *x,
int *rate, int64_t *distortion, int *skip,
int64_t *sse, int64_t ref_best_rd,
BLOCK_SIZE bs) {
MACROBLOCKD *const xd = &x->e_mbd;
MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
mbmi->tx_size = TX_4X4;
mbmi->tx_type = DCT_DCT;
txfm_rd_in_plane(x, rate, distortion, skip, sse, ref_best_rd, 0, bs,
mbmi->tx_size, cpi->sf.use_fast_coef_costing);
}
static void choose_tx_size_from_rd(const AV1_COMP *const cpi, MACROBLOCK *x,
int *rate, int64_t *distortion, int *skip,
int64_t *psse, int64_t ref_best_rd,
BLOCK_SIZE bs) {
const TX_SIZE max_tx_size = max_txsize_lookup[bs];
const AV1_COMMON *const cm = &cpi->common;
MACROBLOCKD *const xd = &x->e_mbd;
MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
aom_prob skip_prob = av1_get_skip_prob(cm, xd);
int r, s;
int64_t d, sse;
int64_t rd = INT64_MAX;
int n, m;
int s0, s1;
int64_t best_rd = INT64_MAX, last_rd = INT64_MAX;
TX_SIZE best_tx = TX_SIZES;
int start_tx, end_tx;
const int tx_select = cm->tx_mode == TX_MODE_SELECT;
TX_TYPE tx_type, best_tx_type = DCT_DCT;
const int is_inter = is_inter_block(mbmi);
uint8_t zcoeff_blk[TX_SIZES][256];
int num_4x4_blks = 1 << (num_pels_log2_lookup[bs] - 4);
const aom_prob *tx_probs = get_tx_probs2(max_tx_size, xd, &cm->fc->tx_probs);
assert(skip_prob > 0);
s0 = av1_cost_bit(skip_prob, 0);
s1 = av1_cost_bit(skip_prob, 1);
if (tx_select) {
start_tx = max_tx_size;
end_tx = (max_tx_size == TX_32X32) ? TX_8X8 : TX_4X4;
} else {
const TX_SIZE chosen_tx_size =
AOMMIN(max_tx_size, tx_mode_to_biggest_tx_size[cm->tx_mode]);
start_tx = chosen_tx_size;
end_tx = chosen_tx_size;
}
*distortion = INT64_MAX;
*rate = INT_MAX;
*skip = 0;
*psse = INT64_MAX;
for (tx_type = DCT_DCT; tx_type < TX_TYPES; ++tx_type) {
#if CONFIG_REF_MV
if (mbmi->ref_mv_idx > 0 && tx_type != DCT_DCT) continue;
#endif
last_rd = INT64_MAX;
for (n = start_tx; n >= end_tx; --n) {
int r_tx_size = 0;
for (m = 0; m <= n - (n == (int)max_tx_size); ++m) {
if (m == n)
r_tx_size += av1_cost_zero(tx_probs[m]);
else
r_tx_size += av1_cost_one(tx_probs[m]);
}
if (n >= TX_32X32 && tx_type != DCT_DCT) {
continue;
}
mbmi->tx_type = tx_type;
txfm_rd_in_plane(x, &r, &d, &s, &sse, ref_best_rd, 0, bs, n,
cpi->sf.use_fast_coef_costing);
if (n < TX_32X32 && !xd->lossless[xd->mi[0]->mbmi.segment_id] &&
r != INT_MAX) {
if (is_inter)
r += cpi->inter_tx_type_costs[mbmi->tx_size][mbmi->tx_type];
else
r += cpi->intra_tx_type_costs
[mbmi->tx_size][intra_mode_to_tx_type_context[mbmi->mode]]
[mbmi->tx_type];
}
if (r == INT_MAX) continue;
if (s) {
if (is_inter) {
rd = RDCOST(x->rdmult, x->rddiv, s1, sse);
} else {
rd = RDCOST(x->rdmult, x->rddiv, s1 + r_tx_size * tx_select, sse);
}
} else {
rd = RDCOST(x->rdmult, x->rddiv, r + s0 + r_tx_size * tx_select, d);
}
if (tx_select && !(s && is_inter)) r += r_tx_size;
if (is_inter && !xd->lossless[xd->mi[0]->mbmi.segment_id] && !s)
rd = AOMMIN(rd, RDCOST(x->rdmult, x->rddiv, s1, sse));
// Early termination in transform size search.
if (cpi->sf.tx_size_search_breakout &&
(rd == INT64_MAX || (s == 1 && tx_type != DCT_DCT && n < start_tx) ||
(n < (int)max_tx_size && rd > last_rd)))
break;
last_rd = rd;
if (rd <
(is_inter && best_tx_type == DCT_DCT ? ext_tx_th : 1) * best_rd) {
best_tx = n;
best_rd = rd;
*distortion = d;
*rate = r;
*skip = s;
*psse = sse;
best_tx_type = mbmi->tx_type;
memcpy(zcoeff_blk, x->zcoeff_blk[n], num_4x4_blks);
}
}
}
mbmi->tx_size = best_tx;
mbmi->tx_type = best_tx_type;
if (mbmi->tx_size >= TX_32X32) assert(mbmi->tx_type == DCT_DCT);
if (best_tx < TX_SIZES)
memcpy(x->zcoeff_blk[best_tx], zcoeff_blk, num_4x4_blks);
}
static void super_block_yrd(const AV1_COMP *const cpi, MACROBLOCK *x, int *rate,
int64_t *distortion, int *skip, int64_t *psse,
BLOCK_SIZE bs, int64_t ref_best_rd) {
MACROBLOCKD *xd = &x->e_mbd;
int64_t sse;
int64_t *ret_sse = psse ? psse : &sse;
assert(bs == xd->mi[0]->mbmi.sb_type);
if (CONFIG_MISC_FIXES && xd->lossless[0]) {
choose_smallest_tx_size(cpi, x, rate, distortion, skip, ret_sse,
ref_best_rd, bs);
} else if (cpi->sf.tx_size_search_method == USE_LARGESTALL ||
xd->lossless[xd->mi[0]->mbmi.segment_id]) {
choose_largest_tx_size(cpi, x, rate, distortion, skip, ret_sse, ref_best_rd,
bs);
} else {
choose_tx_size_from_rd(cpi, x, rate, distortion, skip, ret_sse, ref_best_rd,
bs);
}
}
static int conditional_skipintra(PREDICTION_MODE mode,
PREDICTION_MODE best_intra_mode) {
if (mode == D117_PRED && best_intra_mode != V_PRED &&
best_intra_mode != D135_PRED)
return 1;
if (mode == D63_PRED && best_intra_mode != V_PRED &&
best_intra_mode != D45_PRED)
return 1;
if (mode == D207_PRED && best_intra_mode != H_PRED &&
best_intra_mode != D45_PRED)
return 1;
if (mode == D153_PRED && best_intra_mode != H_PRED &&
best_intra_mode != D135_PRED)
return 1;
return 0;
}
static int64_t rd_pick_intra4x4block(const AV1_COMP *const cpi, MACROBLOCK *x,
int row, int col,
PREDICTION_MODE *best_mode,
const int *bmode_costs, ENTROPY_CONTEXT *a,
ENTROPY_CONTEXT *l, int *bestrate,
int *bestratey, int64_t *bestdistortion,
BLOCK_SIZE bsize, int64_t rd_thresh) {
PREDICTION_MODE mode;
MACROBLOCKD *const xd = &x->e_mbd;
int64_t best_rd = rd_thresh;
struct macroblock_plane *p = &x->plane[0];
struct macroblockd_plane *pd = &xd->plane[0];
const int src_stride = p->src.stride;
const int dst_stride = pd->dst.stride;
const uint8_t *src_init = &p->src.buf[row * 4 * src_stride + col * 4];
uint8_t *dst_init = &pd->dst.buf[row * 4 * src_stride + col * 4];
ENTROPY_CONTEXT ta[2], tempa[2];
ENTROPY_CONTEXT tl[2], templ[2];
const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
int idx, idy;
uint8_t best_dst[8 * 8];
#if CONFIG_AOM_HIGHBITDEPTH
uint16_t best_dst16[8 * 8];
#endif
memcpy(ta, a, num_4x4_blocks_wide * sizeof(a[0]));
memcpy(tl, l, num_4x4_blocks_high * sizeof(l[0]));
xd->mi[0]->mbmi.tx_size = TX_4X4;
#if CONFIG_AOM_HIGHBITDEPTH
if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
int64_t this_rd;
int ratey = 0;
int64_t distortion = 0;
int rate = bmode_costs[mode];
if (!(cpi->sf.intra_y_mode_mask[TX_4X4] & (1 << mode))) continue;
// Only do the oblique modes if the best so far is
// one of the neighboring directional modes
if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
if (conditional_skipintra(mode, *best_mode)) continue;
}
memcpy(tempa, ta, num_4x4_blocks_wide * sizeof(ta[0]));
memcpy(templ, tl, num_4x4_blocks_high * sizeof(tl[0]));
for (idy = 0; idy < num_4x4_blocks_high; ++idy) {
for (idx = 0; idx < num_4x4_blocks_wide; ++idx) {
const int block = (row + idy) * 2 + (col + idx);
const uint8_t *const src = &src_init[idx * 4 + idy * 4 * src_stride];
uint8_t *const dst = &dst_init[idx * 4 + idy * 4 * dst_stride];
int16_t *const src_diff =
av1_raster_block_offset_int16(BLOCK_8X8, block, p->src_diff);
tran_low_t *const coeff = BLOCK_OFFSET(x->plane[0].coeff, block);
xd->mi[0]->bmi[block].as_mode = mode;
av1_predict_intra_block(xd, 1, 1, TX_4X4, mode, dst, dst_stride, dst,
dst_stride, col + idx, row + idy, 0);
aom_highbd_subtract_block(4, 4, src_diff, 8, src, src_stride, dst,
dst_stride, xd->bd);
if (xd->lossless[xd->mi[0]->mbmi.segment_id]) {
TX_TYPE tx_type = get_tx_type(PLANE_TYPE_Y, xd, block);
const SCAN_ORDER *scan_order = get_scan(TX_4X4, tx_type);
av1_highbd_fwd_txfm_4x4(src_diff, coeff, 8, DCT_DCT, 1);
av1_regular_quantize_b_4x4(x, 0, block, scan_order->scan,
scan_order->iscan);
ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
scan_order->scan, scan_order->neighbors,
cpi->sf.use_fast_coef_costing);
if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
goto next_highbd;
av1_highbd_inv_txfm_add_4x4(BLOCK_OFFSET(pd->dqcoeff, block), dst,
dst_stride, p->eobs[block], xd->bd,
DCT_DCT, 1);
} else {
int64_t unused;
TX_TYPE tx_type = get_tx_type(PLANE_TYPE_Y, xd, block);
const SCAN_ORDER *scan_order = get_scan(TX_4X4, tx_type);
av1_highbd_fwd_txfm_4x4(src_diff, coeff, 8, tx_type, 0);
av1_regular_quantize_b_4x4(x, 0, block, scan_order->scan,
scan_order->iscan);
ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
scan_order->scan, scan_order->neighbors,
cpi->sf.use_fast_coef_costing);
distortion +=
av1_highbd_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, block),
16, &unused, xd->bd) >>
2;
if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
goto next_highbd;
av1_highbd_inv_txfm_add_4x4(BLOCK_OFFSET(pd->dqcoeff, block), dst,
dst_stride, p->eobs[block], xd->bd,
tx_type, 0);
}
}
}
rate += ratey;
this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
if (this_rd < best_rd) {
*bestrate = rate;
*bestratey = ratey;
*bestdistortion = distortion;
best_rd = this_rd;
*best_mode = mode;
memcpy(a, tempa, num_4x4_blocks_wide * sizeof(tempa[0]));
memcpy(l, templ, num_4x4_blocks_high * sizeof(templ[0]));
for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy) {
memcpy(best_dst16 + idy * 8,
CONVERT_TO_SHORTPTR(dst_init + idy * dst_stride),
num_4x4_blocks_wide * 4 * sizeof(uint16_t));
}
}
next_highbd : {}
}
if (best_rd >= rd_thresh) return best_rd;
for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy) {
memcpy(CONVERT_TO_SHORTPTR(dst_init + idy * dst_stride),
best_dst16 + idy * 8, num_4x4_blocks_wide * 4 * sizeof(uint16_t));
}
return best_rd;
}
#endif // CONFIG_AOM_HIGHBITDEPTH
for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
int64_t this_rd;
int ratey = 0;
int64_t distortion = 0;
int rate = bmode_costs[mode];
if (!(cpi->sf.intra_y_mode_mask[TX_4X4] & (1 << mode))) continue;
// Only do the oblique modes if the best so far is
// one of the neighboring directional modes
if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
if (conditional_skipintra(mode, *best_mode)) continue;
}
memcpy(tempa, ta, num_4x4_blocks_wide * sizeof(ta[0]));
memcpy(templ, tl, num_4x4_blocks_high * sizeof(tl[0]));
for (idy = 0; idy < num_4x4_blocks_high; ++idy) {
for (idx = 0; idx < num_4x4_blocks_wide; ++idx) {
const int block = (row + idy) * 2 + (col + idx);
const uint8_t *const src = &src_init[idx * 4 + idy * 4 * src_stride];
uint8_t *const dst = &dst_init[idx * 4 + idy * 4 * dst_stride];
int16_t *const src_diff =
av1_raster_block_offset_int16(BLOCK_8X8, block, p->src_diff);
tran_low_t *const coeff = BLOCK_OFFSET(x->plane[0].coeff, block);
xd->mi[0]->bmi[block].as_mode = mode;
av1_predict_intra_block(xd, 1, 1, TX_4X4, mode, dst, dst_stride, dst,
dst_stride, col + idx, row + idy, 0);
aom_subtract_block(4, 4, src_diff, 8, src, src_stride, dst, dst_stride);
if (xd->lossless[xd->mi[0]->mbmi.segment_id]) {
TX_TYPE tx_type = get_tx_type(PLANE_TYPE_Y, xd, block);
const SCAN_ORDER *scan_order = get_scan(TX_4X4, tx_type);
av1_fwd_txfm_4x4(src_diff, coeff, 8, DCT_DCT, 1);
av1_regular_quantize_b_4x4(x, 0, block, scan_order->scan,
scan_order->iscan);
ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
scan_order->scan, scan_order->neighbors,
cpi->sf.use_fast_coef_costing);
if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
goto next;
av1_inv_txfm_add_4x4(BLOCK_OFFSET(pd->dqcoeff, block), dst,
dst_stride, p->eobs[block], DCT_DCT, 1);
} else {
int64_t unused;
TX_TYPE tx_type = get_tx_type(PLANE_TYPE_Y, xd, block);
const SCAN_ORDER *scan_order = get_scan(TX_4X4, tx_type);
av1_fwd_txfm_4x4(src_diff, coeff, 8, tx_type, 0);
av1_regular_quantize_b_4x4(x, 0, block, scan_order->scan,
scan_order->iscan);
ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
scan_order->scan, scan_order->neighbors,
cpi->sf.use_fast_coef_costing);
distortion += av1_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, block),
16, &unused) >>
2;
if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
goto next;
av1_inv_txfm_add_4x4(BLOCK_OFFSET(pd->dqcoeff, block), dst,
dst_stride, p->eobs[block], tx_type, 0);
}
}
}
rate += ratey;
this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
if (this_rd < best_rd) {
*bestrate = rate;
*bestratey = ratey;
*bestdistortion = distortion;
best_rd = this_rd;
*best_mode = mode;
memcpy(a, tempa, num_4x4_blocks_wide * sizeof(tempa[0]));
memcpy(l, templ, num_4x4_blocks_high * sizeof(templ[0]));
for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
memcpy(best_dst + idy * 8, dst_init + idy * dst_stride,
num_4x4_blocks_wide * 4);
}
next : {}
}
if (best_rd >= rd_thresh) return best_rd;
for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
memcpy(dst_init + idy * dst_stride, best_dst + idy * 8,
num_4x4_blocks_wide * 4);
return best_rd;
}
static int64_t rd_pick_intra_sub_8x8_y_mode(const AV1_COMP *const cpi,
MACROBLOCK *mb, int *rate,
int *rate_y, int64_t *distortion,
int64_t best_rd) {
int i, j;
const MACROBLOCKD *const xd = &mb->e_mbd;
MODE_INFO *const mic = xd->mi[0];
const MODE_INFO *above_mi = xd->above_mi;
const MODE_INFO *left_mi = xd->left_mi;
const BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
int idx, idy;
int cost = 0;
int64_t total_distortion = 0;
int tot_rate_y = 0;
int64_t total_rd = 0;
const int *bmode_costs = cpi->mbmode_cost;
#if CONFIG_EXT_INTRA
mic->mbmi.intra_angle_delta[0] = 0;
#endif // CONFIG_EXT_INTRA
// Pick modes for each sub-block (of size 4x4, 4x8, or 8x4) in an 8x8 block.
for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
PREDICTION_MODE best_mode = DC_PRED;
int r = INT_MAX, ry = INT_MAX;
int64_t d = INT64_MAX, this_rd = INT64_MAX;
i = idy * 2 + idx;
if (cpi->common.frame_type == KEY_FRAME) {
const PREDICTION_MODE A = av1_above_block_mode(mic, above_mi, i);
const PREDICTION_MODE L = av1_left_block_mode(mic, left_mi, i);
bmode_costs = cpi->y_mode_costs[A][L];
}
this_rd = rd_pick_intra4x4block(
cpi, mb, idy, idx, &best_mode, bmode_costs,
xd->plane[0].above_context + idx, xd->plane[0].left_context + idy, &r,
&ry, &d, bsize, best_rd - total_rd);
if (this_rd >= best_rd - total_rd) return INT64_MAX;
total_rd += this_rd;
cost += r;
total_distortion += d;
tot_rate_y += ry;
mic->bmi[i].as_mode = best_mode;
for (j = 1; j < num_4x4_blocks_high; ++j)
mic->bmi[i + j * 2].as_mode = best_mode;
for (j = 1; j < num_4x4_blocks_wide; ++j)
mic->bmi[i + j].as_mode = best_mode;
if (total_rd >= best_rd) return INT64_MAX;
}
}
*rate = cost;
*rate_y = tot_rate_y;
*distortion = total_distortion;
mic->mbmi.mode = mic->bmi[3].as_mode;
return RDCOST(mb->rdmult, mb->rddiv, cost, total_distortion);
}
#if CONFIG_EXT_INTRA
static INLINE int write_uniform_cost(int n, int v) {
const int l = get_unsigned_bits(n), m = (1 << l) - n;
if (l == 0) return 0;
return (v < m) ? ((l - 1) * av1_cost_bit(128, 0))
: (l * av1_cost_bit(128, 0));
}
static int64_t pick_intra_angle_routine_sby(
const AV1_COMP *const cpi, MACROBLOCK *x, int8_t angle_delta,
int max_angle_delta, int *rate, int *rate_tokenonly, int64_t *distortion,
int *skippable, int8_t *best_angle_delta, TX_SIZE *best_tx_size,
TX_TYPE *best_tx_type, BLOCK_SIZE bsize, int mode_cost, int64_t *best_rd,
int64_t best_rd_in) {
int this_rate, this_rate_tokenonly, s;
int64_t this_distortion, this_rd;
MB_MODE_INFO *const mbmi = &x->e_mbd.mi[0]->mbmi;
mbmi->intra_angle_delta[0] = angle_delta;
super_block_yrd(cpi, x, &this_rate_tokenonly, &this_distortion, &s, NULL,
bsize, best_rd_in);
if (this_rate_tokenonly == INT_MAX) return INT64_MAX;
this_rate = this_rate_tokenonly + mode_cost +
write_uniform_cost(2 * max_angle_delta + 1,
mbmi->intra_angle_delta[0] + max_angle_delta);
this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
if (this_rd < *best_rd) {
*best_rd = this_rd;
*best_angle_delta = mbmi->intra_angle_delta[0];
*best_tx_size = mbmi->tx_size;
*best_tx_type = mbmi->tx_type;
*rate = this_rate;
*rate_tokenonly = this_rate_tokenonly;
*distortion = this_distortion;
*skippable = s;
}
return this_rd;
}
static int64_t rd_pick_intra_angle_sby(const AV1_COMP *const cpi, MACROBLOCK *x,
int *rate, int *rate_tokenonly,
int64_t *distortion, int *skippable,
BLOCK_SIZE bsize, int mode_cost,
int64_t best_rd) {
MB_MODE_INFO *const mbmi = &x->e_mbd.mi[0]->mbmi;
const int max_angle_delta =
av1_max_angle_delta_y[max_txsize_lookup[bsize]][mbmi->mode];
int i;
int8_t angle_delta, best_angle_delta = 0;
int64_t this_rd, best_rd_in, rd_cost[2 * (MAX_ANGLE_DELTA + 2)];
TX_SIZE best_tx_size = mbmi->tx_size;
TX_TYPE best_tx_type = mbmi->tx_type;
for (i = 0; i < 2 * (MAX_ANGLE_DELTA + 2); ++i) rd_cost[i] = INT64_MAX;
for (angle_delta = 0; angle_delta <= MAX_ANGLE_DELTA; angle_delta += 2) {
if (angle_delta > max_angle_delta) continue;
for (i = 0; i < 2; ++i) {
best_rd_in = (best_rd == INT64_MAX)
? INT64_MAX
: (best_rd + (best_rd >> ((angle_delta == 0) ? 3 : 5)));
this_rd = pick_intra_angle_routine_sby(
cpi, x, (1 - 2 * i) * angle_delta, max_angle_delta, rate,
rate_tokenonly, distortion, skippable, &best_angle_delta,
&best_tx_size, &best_tx_type, bsize, mode_cost, &best_rd, best_rd_in);
rd_cost[2 * angle_delta + i] = this_rd;
if (angle_delta == 0) {
if (this_rd == INT64_MAX) return best_rd;
rd_cost[1] = this_rd;
break;
}
}
}
assert(best_rd != INT64_MAX);
for (angle_delta = 1; angle_delta <= MAX_ANGLE_DELTA; angle_delta += 2) {
int skip_search;
int64_t rd_thresh;
if (angle_delta > max_angle_delta) continue;
for (i = 0; i < 2; ++i) {
skip_search = 0;
rd_thresh = best_rd + (best_rd >> 5);
if (rd_cost[2 * (angle_delta + 1) + i] > rd_thresh &&
rd_cost[2 * (angle_delta - 1) + i] > rd_thresh)
skip_search = 1;
if (!skip_search) {
this_rd = pick_intra_angle_routine_sby(
cpi, x, (1 - 2 * i) * angle_delta, max_angle_delta, rate,
rate_tokenonly, distortion, skippable, &best_angle_delta,
&best_tx_size, &best_tx_type, bsize, mode_cost, &best_rd, best_rd);
}
}
}
mbmi->tx_size = best_tx_size;
mbmi->intra_angle_delta[0] = best_angle_delta;
mbmi->tx_type = best_tx_type;
return best_rd;
}
// Indices are sign, integer, and fractional part of the gradient value
static const uint8_t gradient_to_angle_bin[2][7][16] = {
{
{
6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 7, 7, 0, 0, 0, 0,
},
{
0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1,
},
{
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
},
{
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
},
{
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
},
{
2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
},
{
2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
},
},
{
{
6, 6, 6, 6, 5, 5, 5, 5, 5, 5, 5, 5, 4, 4, 4, 4,
},
{
4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 3, 3, 3, 3, 3, 3,
},
{
3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3,
},
{
3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3,
},
{
3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3,
},
{
3, 3, 3, 3, 3, 3, 3, 2, 2, 2, 2, 2, 2, 2, 2, 2,
},
{
2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
},
},
};
static const uint8_t mode_to_angle_bin[INTRA_MODES] = {
0, 2, 6, 0, 4, 3, 5, 7, 1, 0,
};
// Use gradient analysis to calculate angle histogram. Prediction modes
// corresponding to angles of small percentage will be marked in the mask.
static void angle_estimation(const uint8_t *src, const int src_stride,
const int rows, const int cols,
uint8_t *directional_mode_skip_mask) {
int i, r, c, index, dx, dy, temp, sn, remd, quot;
const int angle_skip_thresh = 10;
uint64_t hist[DIRECTIONAL_MODES];
uint64_t hist_sum = 0;
memset(hist, 0, DIRECTIONAL_MODES * sizeof(hist[0]));
src += src_stride;
for (r = 1; r < rows; ++r) {
for (c = 1; c < cols; ++c) {
dx = src[c] - src[c - 1];
dy = src[c] - src[c - src_stride];
temp = dx * dx + dy * dy;
if (dy == 0) {
index = 2;
} else {
sn = (dx > 0) ^ (dy > 0);
dx = abs(dx);
dy = abs(dy);
remd = dx % dy;
quot = dx / dy;
remd = remd * 16 / dy;
index = gradient_to_angle_bin[sn][AOMMIN(quot, 6)][AOMMIN(remd, 15)];
}
hist[index] += temp;
}
src += src_stride;
}
for (i = 0; i < DIRECTIONAL_MODES; ++i) hist_sum += hist[i];
for (i = 0; i < INTRA_MODES; ++i) {
if (i != DC_PRED && i != TM_PRED) {
int index = mode_to_angle_bin[i];
uint64_t score = 2 * hist[index];
int weight = 2;
if (index > 0) {
score += hist[index - 1];
weight += 1;
}
if (index < DIRECTIONAL_MODES - 1) {
score += hist[index + 1];
weight += 1;
}
if (score * angle_skip_thresh < hist_sum * weight) {
directional_mode_skip_mask[i] = 1;
}
}
}
}
#if CONFIG_AOM_HIGHBITDEPTH
static void highbd_angle_estimation(const uint8_t *src8, const int src_stride,
const int rows, const int cols,
uint8_t *directional_mode_skip_mask) {
int i, r, c, index, dx, dy, temp, sn, remd, quot;
const int angle_skip_thresh = 10;
uint64_t hist[DIRECTIONAL_MODES];
uint64_t hist_sum = 0;
uint16_t *src = CONVERT_TO_SHORTPTR(src8);
memset(hist, 0, DIRECTIONAL_MODES * sizeof(hist[0]));
src += src_stride;
for (r = 1; r < rows; ++r) {
for (c = 1; c < cols; ++c) {
dx = src[c] - src[c - 1];
dy = src[c] - src[c - src_stride];
temp = dx * dx + dy * dy;
if (dy == 0) {
index = 2;
} else {
sn = (dx > 0) ^ (dy > 0);
dx = abs(dx);
dy = abs(dy);
remd = dx % dy;
quot = dx / dy;
remd = remd * 16 / dy;
index = gradient_to_angle_bin[sn][AOMMIN(quot, 6)][AOMMIN(remd, 15)];
}
hist[index] += temp;
}
src += src_stride;
}
for (i = 0; i < DIRECTIONAL_MODES; ++i) hist_sum += hist[i];
for (i = 0; i < INTRA_MODES; ++i) {
if (i != DC_PRED && i != TM_PRED) {
int index = mode_to_angle_bin[i];
uint64_t score = 2 * hist[index];
int weight = 2;
if (index > 0) {
score += hist[index - 1];
weight += 1;
}
if (index < DIRECTIONAL_MODES - 1) {
score += hist[index + 1];
weight += 1;
}
if (score * angle_skip_thresh < hist_sum * weight)
directional_mode_skip_mask[i] = 1;
}
}
}
#endif // CONFIG_AOM_HIGHBITDEPTH
#endif // CONFIG_EXT_INTRA
// This function is used only for intra_only frames
static int64_t rd_pick_intra_sby_mode(const AV1_COMP *const cpi, MACROBLOCK *x,
int *rate, int *rate_tokenonly,
int64_t *distortion, int *skippable,
BLOCK_SIZE bsize, int64_t best_rd) {
PREDICTION_MODE mode;
PREDICTION_MODE mode_selected = DC_PRED;
MACROBLOCKD *const xd = &x->e_mbd;
MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
int this_rate, this_rate_tokenonly, s;
int64_t this_distortion, this_rd;
TX_SIZE best_tx = TX_4X4;
#if CONFIG_EXT_INTRA
int8_t best_angle_delta = 0;
uint8_t directional_mode_skip_mask[INTRA_MODES];
const int src_stride = x->plane[0].src.stride;
const uint8_t *src = x->plane[0].src.buf;
const int rows = 4 * num_4x4_blocks_high_lookup[bsize];
const int cols = 4 * num_4x4_blocks_wide_lookup[bsize];
const TX_SIZE max_tx_size = max_txsize_lookup[bsize];
#endif // CONFIG_EXT_INTRA
TX_TYPE best_tx_type = DCT_DCT;
const int *bmode_costs;
const MODE_INFO *above_mi = xd->above_mi;
const MODE_INFO *left_mi = xd->left_mi;
const PREDICTION_MODE A = av1_above_block_mode(xd->mi[0], above_mi, 0);
const PREDICTION_MODE L = av1_left_block_mode(xd->mi[0], left_mi, 0);
bmode_costs = cpi->y_mode_costs[A][L];
memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
#if CONFIG_EXT_INTRA
memset(directional_mode_skip_mask, 0,
sizeof(directional_mode_skip_mask[0]) * INTRA_MODES);
#if CONFIG_AOM_HIGHBITDEPTH
if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
highbd_angle_estimation(src, src_stride, rows, cols,
directional_mode_skip_mask);
else
#endif
angle_estimation(src, src_stride, rows, cols, directional_mode_skip_mask);
#endif // CONFIG_EXT_INTRA
/* Y Search for intra prediction mode */
for (mode = DC_PRED; mode <= TM_PRED; mode++) {
mbmi->mode = mode;
#if CONFIG_EXT_INTRA
if (is_directional_mode(mbmi->mode)) {
if (directional_mode_skip_mask[mbmi->mode]) continue;
this_rate_tokenonly = INT_MAX;
this_rd = rd_pick_intra_angle_sby(
cpi, x, &this_rate, &this_rate_tokenonly, &this_distortion, &s, bsize,
bmode_costs[mbmi->mode], best_rd);
} else {
mbmi->intra_angle_delta[0] = 0;
super_block_yrd(cpi, x, &this_rate_tokenonly, &this_distortion, &s, NULL,
bsize, best_rd);
}
#else
super_block_yrd(cpi, x, &this_rate_tokenonly, &this_distortion, &s, NULL,
bsize, best_rd);
#endif // CONFIG_EXT_INTRA
if (this_rate_tokenonly == INT_MAX) continue;
this_rate = this_rate_tokenonly + bmode_costs[mode];
#if CONFIG_EXT_INTRA
if (is_directional_mode(mbmi->mode)) {
const int max_angle_delta =
av1_max_angle_delta_y[max_tx_size][mbmi->mode];
this_rate +=
write_uniform_cost(2 * max_angle_delta + 1,
max_angle_delta + mbmi->intra_angle_delta[0]);
}
#endif // CONFIG_EXT_INTRA
this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
if (this_rd < best_rd) {
mode_selected = mode;
best_rd = this_rd;
best_tx = mbmi->tx_size;
best_tx_type = mbmi->tx_type;
#if CONFIG_EXT_INTRA
best_angle_delta = mbmi->intra_angle_delta[0];
#endif // CONFIG_EXT_INTRA
*rate = this_rate;
*rate_tokenonly = this_rate_tokenonly;
*distortion = this_distortion;
*skippable = s;
}
}
mbmi->mode = mode_selected;
mbmi->tx_size = best_tx;
mbmi->tx_type = best_tx_type;
#if CONFIG_EXT_INTRA
mbmi->intra_angle_delta[0] = best_angle_delta;
#endif // CONFIG_EXT_INTRA
return best_rd;
}
// Return value 0: early termination triggered, no valid rd cost available;
// 1: rd cost values are valid.
static int super_block_uvrd(const AV1_COMP *const cpi, MACROBLOCK *x, int *rate,
int64_t *distortion, int *skippable, int64_t *sse,
BLOCK_SIZE bsize, int64_t ref_best_rd) {
MACROBLOCKD *const xd = &x->e_mbd;
MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
const TX_SIZE uv_tx_size = get_uv_tx_size(mbmi, &xd->plane[1]);
int plane;
int pnrate = 0, pnskip = 1;
int64_t pndist = 0, pnsse = 0;
int is_cost_valid = 1;
if (ref_best_rd < 0) is_cost_valid = 0;
if (is_inter_block(mbmi) && is_cost_valid) {
int plane;
for (plane = 1; plane < MAX_MB_PLANE; ++plane)
av1_subtract_plane(x, bsize, plane);
}
*rate = 0;
*distortion = 0;
*sse = 0;
*skippable = 1;
if (is_cost_valid) {
for (plane = 1; plane < MAX_MB_PLANE; ++plane) {
txfm_rd_in_plane(x, &pnrate, &pndist, &pnskip, &pnsse, ref_best_rd, plane,
bsize, uv_tx_size, cpi->sf.use_fast_coef_costing);
if (pnrate == INT_MAX) {
is_cost_valid = 0;
break;
}
*rate += pnrate;
*distortion += pndist;
*sse += pnsse;
*skippable &= pnskip;
}
}
if (!is_cost_valid) {
// reset cost value
*rate = INT_MAX;
*distortion = INT64_MAX;
*sse = INT64_MAX;
*skippable = 0;
}
return is_cost_valid;
}
#if CONFIG_EXT_INTRA
static int64_t pick_intra_angle_routine_sbuv(
const AV1_COMP *const cpi, MACROBLOCK *x, int *rate, int *rate_tokenonly,
int64_t *distortion, int *skippable, int8_t *best_angle_delta,
BLOCK_SIZE bsize, int rate_overhead, int64_t *best_rd, int64_t best_rd_in) {
MB_MODE_INFO *mbmi = &x->e_mbd.mi[0]->mbmi;
int this_rate_tokenonly, this_rate, s;
int64_t this_distortion, this_sse, this_rd;
if (!super_block_uvrd(cpi, x, &this_rate_tokenonly, &this_distortion, &s,
&this_sse, bsize, best_rd_in))
return INT64_MAX;
this_rate = this_rate_tokenonly + rate_overhead;
this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
if (this_rd < *best_rd) {
*best_rd = this_rd;
*best_angle_delta = mbmi->intra_angle_delta[1];
*rate = this_rate;
*rate_tokenonly = this_rate_tokenonly;
*distortion = this_distortion;
*skippable = s;
}
return this_rd;
}
static int rd_pick_intra_angle_sbuv(const AV1_COMP *cpi, MACROBLOCK *x,
int *rate, int *rate_tokenonly,
int64_t *distortion, int *skippable,
BLOCK_SIZE bsize, int rate_overhead,
int64_t best_rd) {
MB_MODE_INFO *mbmi = &x->e_mbd.mi[0]->mbmi;
int64_t this_rd, best_rd_in, rd_cost[2 * (MAX_ANGLE_DELTA + 2)];
int8_t angle_delta, best_angle_delta = 0;
int i;
*rate_tokenonly = INT_MAX;
*skippable = 0;
*distortion = INT64_MAX;
for (i = 0; i < 2 * (MAX_ANGLE_DELTA + 2); ++i) rd_cost[i] = INT64_MAX;
for (angle_delta = 0; angle_delta <= MAX_ANGLE_DELTA_UV; angle_delta += 2) {
for (i = 0; i < 2; ++i) {
best_rd_in = (best_rd == INT64_MAX)
? INT64_MAX
: (best_rd + (best_rd >> ((angle_delta == 0) ? 3 : 5)));
mbmi->intra_angle_delta[1] = (1 - 2 * i) * angle_delta;
this_rd = pick_intra_angle_routine_sbuv(
cpi, x, rate, rate_tokenonly, distortion, skippable,
&best_angle_delta, bsize, rate_overhead, &best_rd, best_rd_in);
rd_cost[2 * angle_delta + i] = this_rd;
if (angle_delta == 0) {
if (this_rd == INT64_MAX) return 0;
rd_cost[1] = this_rd;
break;
}
}
}
assert(best_rd != INT64_MAX);
for (angle_delta = 1; angle_delta <= MAX_ANGLE_DELTA_UV; angle_delta += 2) {
int skip_search;
int64_t rd_thresh;
for (i = 0; i < 2; ++i) {
skip_search = 0;
rd_thresh = best_rd + (best_rd >> 5);
if (rd_cost[2 * (angle_delta + 1) + i] > rd_thresh &&
rd_cost[2 * (angle_delta - 1) + i] > rd_thresh)
skip_search = 1;
if (!skip_search) {
mbmi->intra_angle_delta[1] = (1 - 2 * i) * angle_delta;
this_rd = pick_intra_angle_routine_sbuv(
cpi, x, rate, rate_tokenonly, distortion, skippable,
&best_angle_delta, bsize, rate_overhead, &best_rd, best_rd);
}
}
}
mbmi->intra_angle_delta[1] = best_angle_delta;
return *rate_tokenonly != INT_MAX;
}
#endif // CONFIG_EXT_INTRA
static int64_t rd_pick_intra_sbuv_mode(const AV1_COMP *const cpi, MACROBLOCK *x,
PICK_MODE_CONTEXT *ctx, int *rate,
int *rate_tokenonly, int64_t *distortion,
int *skippable, BLOCK_SIZE bsize,
TX_SIZE max_tx_size) {
MB_MODE_INFO *const mbmi = &x->e_mbd.mi[0]->mbmi;
PREDICTION_MODE mode;
PREDICTION_MODE mode_selected = DC_PRED;
#if CONFIG_EXT_INTRA
int8_t best_angle_delta = 0;
int rate_overhead;
#endif // CONFIG_EXT_INTRA
int64_t best_rd = INT64_MAX, this_rd;
int this_rate_tokenonly, this_rate, s;
int64_t this_distortion, this_sse;
memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
if (!(cpi->sf.intra_uv_mode_mask[max_tx_size] & (1 << mode))) continue;
mbmi->uv_mode = mode;
#if CONFIG_EXT_INTRA
rate_overhead = cpi->intra_uv_mode_cost[mbmi->mode][mode] +
write_uniform_cost(2 * MAX_ANGLE_DELTA_UV + 1, 0);
if (mbmi->sb_type >= BLOCK_8X8 && is_directional_mode(mbmi->uv_mode)) {
if (!rd_pick_intra_angle_sbuv(cpi, x, &this_rate, &this_rate_tokenonly,
&this_distortion, &s, bsize, rate_overhead,
best_rd))
continue;
rate_overhead =
cpi->intra_uv_mode_cost[mbmi->mode][mode] +
write_uniform_cost(2 * MAX_ANGLE_DELTA_UV + 1,
MAX_ANGLE_DELTA_UV + mbmi->intra_angle_delta[1]);
} else {
mbmi->intra_angle_delta[1] = 0;
if (!super_block_uvrd(cpi, x, &this_rate_tokenonly, &this_distortion, &s,
&this_sse, bsize, best_rd))
continue;
rate_overhead = cpi->intra_uv_mode_cost[mbmi->mode][mode];
}
this_rate = this_rate_tokenonly + rate_overhead;
#else
if (!super_block_uvrd(cpi, x, &this_rate_tokenonly, &this_distortion, &s,
&this_sse, bsize, best_rd))
continue;
this_rate = this_rate_tokenonly + cpi->intra_uv_mode_cost[mbmi->mode][mode];
#endif // CONFIG_EXT_INTRA
this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
if (this_rd < best_rd) {
mode_selected = mode;
#if CONFIG_EXT_INTRA
best_angle_delta = mbmi->intra_angle_delta[1];
#endif // CONFIG_EXT_INTRA
best_rd = this_rd;
*rate = this_rate;
*rate_tokenonly = this_rate_tokenonly;
*distortion = this_distortion;
*skippable = s;
if (!x->select_tx_size) swap_block_ptr(x, ctx, 2, 0, 1, MAX_MB_PLANE);
}
}
mbmi->uv_mode = mode_selected;
#if CONFIG_EXT_INTRA
mbmi->intra_angle_delta[1] = best_angle_delta;
#endif // CONFIG_EXT_INTRA
return best_rd;
}
static int64_t rd_sbuv_dcpred(const AV1_COMP *cpi, MACROBLOCK *x, int *rate,
int *rate_tokenonly, int64_t *distortion,
int *skippable, BLOCK_SIZE bsize) {
int64_t unused;
x->e_mbd.mi[0]->mbmi.uv_mode = DC_PRED;
memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
super_block_uvrd(cpi, x, rate_tokenonly, distortion, skippable, &unused,
bsize, INT64_MAX);
*rate = *rate_tokenonly +
cpi->intra_uv_mode_cost[x->e_mbd.mi[0]->mbmi.mode][DC_PRED];
return RDCOST(x->rdmult, x->rddiv, *rate, *distortion);
}
static void choose_intra_uv_mode(const AV1_COMP *const cpi, MACROBLOCK *const x,
PICK_MODE_CONTEXT *ctx, BLOCK_SIZE bsize,
TX_SIZE max_tx_size, int *rate_uv,
int *rate_uv_tokenonly, int64_t *dist_uv,
int *skip_uv, PREDICTION_MODE *mode_uv) {
// Use an estimated rd for uv_intra based on DC_PRED if the
// appropriate speed flag is set.
rd_pick_intra_sbuv_mode(cpi, x, ctx, rate_uv, rate_uv_tokenonly, dist_uv,
skip_uv, bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize,
max_tx_size);
*mode_uv = x->e_mbd.mi[0]->mbmi.uv_mode;
}
static int cost_mv_ref(const AV1_COMP *const cpi, PREDICTION_MODE mode,
int16_t mode_context) {
#if CONFIG_REF_MV
int mode_cost = 0;
int16_t mode_ctx = mode_context & NEWMV_CTX_MASK;
int16_t is_all_zero_mv = mode_context & (1 << ALL_ZERO_FLAG_OFFSET);
assert(is_inter_mode(mode));
if (mode == NEWMV) {
mode_cost = cpi->newmv_mode_cost[mode_ctx][0];
return mode_cost;
} else {
mode_cost = cpi->newmv_mode_cost[mode_ctx][1];
mode_ctx = (mode_context >> ZEROMV_OFFSET) & ZEROMV_CTX_MASK;
if (is_all_zero_mv) return mode_cost;
if (mode == ZEROMV) {
mode_cost += cpi->zeromv_mode_cost[mode_ctx][0];
return mode_cost;
} else {
mode_cost += cpi->zeromv_mode_cost[mode_ctx][1];
mode_ctx = (mode_context >> REFMV_OFFSET) & REFMV_CTX_MASK;
if (mode_context & (1 << SKIP_NEARESTMV_OFFSET)) mode_ctx = 6;
if (mode_context & (1 << SKIP_NEARMV_OFFSET)) mode_ctx = 7;
if (mode_context & (1 << SKIP_NEARESTMV_SUB8X8_OFFSET)) mode_ctx = 8;
mode_cost += cpi->refmv_mode_cost[mode_ctx][mode != NEARESTMV];
return mode_cost;
}
}
#else
assert(is_inter_mode(mode));
return cpi->inter_mode_cost[mode_context][INTER_OFFSET(mode)];
#endif
}
static int set_and_cost_bmi_mvs(const AV1_COMP *const cpi, MACROBLOCK *x,
MACROBLOCKD *xd, int i, PREDICTION_MODE mode,
int_mv this_mv[2],
int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
int_mv seg_mvs[MAX_REF_FRAMES],
int_mv *best_ref_mv[2], const int *mvjcost,
int *mvcost[2]) {
MODE_INFO *const mic = xd->mi[0];
const MB_MODE_INFO *const mbmi = &mic->mbmi;
const MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
int thismvcost = 0;
int idx, idy;
const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[mbmi->sb_type];
const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[mbmi->sb_type];
const int is_compound = has_second_ref(mbmi);
int16_t mode_ctx = mbmi_ext->mode_context[mbmi->ref_frame[0]];
switch (mode) {
case NEWMV:
this_mv[0].as_int = seg_mvs[mbmi->ref_frame[0]].as_int;
thismvcost += av1_mv_bit_cost(&this_mv[0].as_mv, &best_ref_mv[0]->as_mv,
mvjcost, mvcost, MV_COST_WEIGHT_SUB);
if (is_compound) {
this_mv[1].as_int = seg_mvs[mbmi->ref_frame[1]].as_int;
thismvcost += av1_mv_bit_cost(&this_mv[1].as_mv, &best_ref_mv[1]->as_mv,
mvjcost, mvcost, MV_COST_WEIGHT_SUB);
}
break;
case NEARMV:
case NEARESTMV:
this_mv[0].as_int = frame_mv[mode][mbmi->ref_frame[0]].as_int;
if (is_compound)
this_mv[1].as_int = frame_mv[mode][mbmi->ref_frame[1]].as_int;
break;
case ZEROMV:
this_mv[0].as_int = 0;
if (is_compound) this_mv[1].as_int = 0;
break;
default: break;
}
mic->bmi[i].as_mv[0].as_int = this_mv[0].as_int;
if (is_compound) mic->bmi[i].as_mv[1].as_int = this_mv[1].as_int;
mic->bmi[i].as_mode = mode;
#if CONFIG_REF_MV
if (mode == NEWMV) {
mic->bmi[i].pred_mv[0].as_int =
mbmi_ext->ref_mvs[mbmi->ref_frame[0]][0].as_int;
if (is_compound)
mic->bmi[i].pred_mv[1].as_int =
mbmi_ext->ref_mvs[mbmi->ref_frame[1]][0].as_int;
} else {
mic->bmi[i].pred_mv[0].as_int = this_mv[0].as_int;
if (is_compound) mic->bmi[i].pred_mv[1].as_int = this_mv[1].as_int;
}
#endif
for (idy = 0; idy < num_4x4_blocks_high; ++idy)
for (idx = 0; idx < num_4x4_blocks_wide; ++idx)
memmove(&mic->bmi[i + idy * 2 + idx], &mic->bmi[i], sizeof(mic->bmi[i]));
#if CONFIG_REF_MV
mode_ctx = av1_mode_context_analyzer(mbmi_ext->mode_context, mbmi->ref_frame,
mbmi->sb_type, i);
#endif
return cost_mv_ref(cpi, mode, mode_ctx) + thismvcost;
}
static int64_t encode_inter_mb_segment(const AV1_COMP *const cpi, MACROBLOCK *x,
int64_t best_yrd, int i, int *labelyrate,
int64_t *distortion, int64_t *sse,
ENTROPY_CONTEXT *ta, ENTROPY_CONTEXT *tl,
int ir, int ic, int mi_row, int mi_col) {
int k;
MACROBLOCKD *xd = &x->e_mbd;
struct macroblockd_plane *const pd = &xd->plane[0];
struct macroblock_plane *const p = &x->plane[0];
MODE_INFO *const mi = xd->mi[0];
const BLOCK_SIZE plane_bsize = get_plane_block_size(mi->mbmi.sb_type, pd);
const int width = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
const int height = 4 * num_4x4_blocks_high_lookup[plane_bsize];
int idx, idy;
void (*fwd_txm4x4)(const int16_t *input, tran_low_t *output, int stride);
const uint8_t *const src =
&p->src.buf[av1_raster_block_offset(BLOCK_8X8, i, p->src.stride)];
uint8_t *const dst =
&pd->dst.buf[av1_raster_block_offset(BLOCK_8X8, i, pd->dst.stride)];
int64_t thisdistortion = 0, thissse = 0;
int thisrate = 0;
TX_TYPE tx_type = get_tx_type(PLANE_TYPE_Y, xd, i);
const SCAN_ORDER *scan_order = get_scan(TX_4X4, tx_type);
av1_build_inter_predictor_sub8x8(xd, 0, i, ir, ic, mi_row, mi_col);
#if CONFIG_AOM_HIGHBITDEPTH
if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
fwd_txm4x4 = xd->lossless[mi->mbmi.segment_id] ? av1_highbd_fwht4x4
: aom_highbd_fdct4x4;
} else {
fwd_txm4x4 = xd->lossless[mi->mbmi.segment_id] ? av1_fwht4x4 : aom_fdct4x4;
}
#else
fwd_txm4x4 = xd->lossless[mi->mbmi.segment_id] ? av1_fwht4x4 : aom_fdct4x4;
#endif // CONFIG_AOM_HIGHBITDEPTH
#if CONFIG_AOM_HIGHBITDEPTH
if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
aom_highbd_subtract_block(
height, width, av1_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
8, src, p->src.stride, dst, pd->dst.stride, xd->bd);
} else {
aom_subtract_block(height, width,
av1_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
8, src, p->src.stride, dst, pd->dst.stride);
}
#else
aom_subtract_block(height, width,
av1_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
8, src, p->src.stride, dst, pd->dst.stride);
#endif // CONFIG_AOM_HIGHBITDEPTH
k = i;
for (idy = 0; idy < height / 4; ++idy) {
for (idx = 0; idx < width / 4; ++idx) {
int64_t ssz, rd, rd1, rd2;
tran_low_t *coeff;
k += (idy * 2 + idx);
coeff = BLOCK_OFFSET(p->coeff, k);
fwd_txm4x4(av1_raster_block_offset_int16(BLOCK_8X8, k, p->src_diff),
coeff, 8);
av1_regular_quantize_b_4x4(x, 0, k, scan_order->scan, scan_order->iscan);
#if CONFIG_AOM_HIGHBITDEPTH
if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
thisdistortion += av1_highbd_block_error(
coeff, BLOCK_OFFSET(pd->dqcoeff, k), 16, &ssz, xd->bd);
} else {
thisdistortion +=
av1_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, k), 16, &ssz);
}
#else
thisdistortion +=
av1_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, k), 16, &ssz);
#endif // CONFIG_AOM_HIGHBITDEPTH
thissse += ssz;
thisrate += cost_coeffs(x, 0, k, ta + (k & 1), tl + (k >> 1), TX_4X4,
scan_order->scan, scan_order->neighbors,
cpi->sf.use_fast_coef_costing);
rd1 = RDCOST(x->rdmult, x->rddiv, thisrate, thisdistortion >> 2);
rd2 = RDCOST(x->rdmult, x->rddiv, 0, thissse >> 2);
rd = AOMMIN(rd1, rd2);
if (rd >= best_yrd) return INT64_MAX;
}
}
*distortion = thisdistortion >> 2;
*labelyrate = thisrate;
*sse = thissse >> 2;
return RDCOST(x->rdmult, x->rddiv, *labelyrate, *distortion);
}
typedef struct {
int eobs;
int brate;
int byrate;
int64_t bdist;
int64_t bsse;
int64_t brdcost;
int_mv mvs[2];
ENTROPY_CONTEXT ta[2];
ENTROPY_CONTEXT tl[2];
} SEG_RDSTAT;
typedef struct {
int_mv *ref_mv[2];
int_mv mvp;
int64_t segment_rd;
int r;
int64_t d;
int64_t sse;
int segment_yrate;
PREDICTION_MODE modes[4];
SEG_RDSTAT rdstat[4][INTER_MODES];
int mvthresh;
} BEST_SEG_INFO;
static INLINE int mv_check_bounds(const MACROBLOCK *x, const MV *mv) {
return (mv->row >> 3) < x->mv_row_min || (mv->row >> 3) > x->mv_row_max ||
(mv->col >> 3) < x->mv_col_min || (mv->col >> 3) > x->mv_col_max;
}
static INLINE void mi_buf_shift(MACROBLOCK *x, int i) {
MB_MODE_INFO *const mbmi = &x->e_mbd.mi[0]->mbmi;
struct macroblock_plane *const p = &x->plane[0];
struct macroblockd_plane *const pd = &x->e_mbd.plane[0];
p->src.buf =
&p->src.buf[av1_raster_block_offset(BLOCK_8X8, i, p->src.stride)];
assert(((intptr_t)pd->pre[0].buf & 0x7) == 0);
pd->pre[0].buf =
&pd->pre[0].buf[av1_raster_block_offset(BLOCK_8X8, i, pd->pre[0].stride)];
if (has_second_ref(mbmi))
pd->pre[1].buf =
&pd->pre[1]
.buf[av1_raster_block_offset(BLOCK_8X8, i, pd->pre[1].stride)];
}
static INLINE void mi_buf_restore(MACROBLOCK *x, struct buf_2d orig_src,
struct buf_2d orig_pre[2]) {
MB_MODE_INFO *mbmi = &x->e_mbd.mi[0]->mbmi;
x->plane[0].src = orig_src;
x->e_mbd.plane[0].pre[0] = orig_pre[0];
if (has_second_ref(mbmi)) x->e_mbd.plane[0].pre[1] = orig_pre[1];
}
static INLINE int mv_has_subpel(const MV *mv) {
return (mv->row & 0x0F) || (mv->col & 0x0F);
}
// Check if NEARESTMV/NEARMV/ZEROMV is the cheapest way encode zero motion.
// TODO(aconverse): Find out if this is still productive then clean up or remove
static int check_best_zero_mv(const AV1_COMP *const cpi,
const int16_t mode_context[MAX_REF_FRAMES],
int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
int this_mode,
const MV_REFERENCE_FRAME ref_frames[2],
const BLOCK_SIZE bsize, int block) {
if ((this_mode == NEARMV || this_mode == NEARESTMV || this_mode == ZEROMV) &&
frame_mv[this_mode][ref_frames[0]].as_int == 0 &&
(ref_frames[1] == NONE ||
frame_mv[this_mode][ref_frames[1]].as_int == 0)) {
#if CONFIG_REF_MV
int16_t rfc =
av1_mode_context_analyzer(mode_context, ref_frames, bsize, block);
#else
int16_t rfc = mode_context[ref_frames[0]];
#endif
int c1 = cost_mv_ref(cpi, NEARMV, rfc);
int c2 = cost_mv_ref(cpi, NEARESTMV, rfc);
int c3 = cost_mv_ref(cpi, ZEROMV, rfc);
#if !CONFIG_REF_MV
(void)bsize;
(void)block;
#endif
if (this_mode == NEARMV) {
if (c1 > c3) return 0;
} else if (this_mode == NEARESTMV) {
if (c2 > c3) return 0;
} else {
assert(this_mode == ZEROMV);
if (ref_frames[1] == NONE) {
if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0) ||
(c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0))
return 0;
} else {
if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0 &&
frame_mv[NEARESTMV][ref_frames[1]].as_int == 0) ||
(c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0 &&
frame_mv[NEARMV][ref_frames[1]].as_int == 0))
return 0;
}
}
}
return 1;
}
static void joint_motion_search(const AV1_COMP *cpi, MACROBLOCK *x,
BLOCK_SIZE bsize, int_mv *frame_mv, int mi_row,
int mi_col, int_mv single_newmv[MAX_REF_FRAMES],
int *rate_mv, const int block) {
const AV1_COMMON *const cm = &cpi->common;
const int pw = 4 * num_4x4_blocks_wide_lookup[bsize];
const int ph = 4 * num_4x4_blocks_high_lookup[bsize];
MACROBLOCKD *xd = &x->e_mbd;
MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
const int refs[2] = { mbmi->ref_frame[0],
mbmi->ref_frame[1] < 0 ? 0 : mbmi->ref_frame[1] };
int_mv ref_mv[2];
int ite, ref;
struct scale_factors sf;
// Do joint motion search in compound mode to get more accurate mv.
struct buf_2d backup_yv12[2][MAX_MB_PLANE];
int last_besterr[2] = { INT_MAX, INT_MAX };
const YV12_BUFFER_CONFIG *const scaled_ref_frame[2] = {
av1_get_scaled_ref_frame(cpi, mbmi->ref_frame[0]),
av1_get_scaled_ref_frame(cpi, mbmi->ref_frame[1])
};
// Prediction buffer from second frame.
#if CONFIG_AOM_HIGHBITDEPTH
DECLARE_ALIGNED(16, uint16_t, second_pred_alloc_16[64 * 64]);
uint8_t *second_pred;
#else
DECLARE_ALIGNED(16, uint8_t, second_pred[64 * 64]);
#endif // CONFIG_AOM_HIGHBITDEPTH
for (ref = 0; ref < 2; ++ref) {
ref_mv[ref] = x->mbmi_ext->ref_mvs[refs[ref]][0];
if (scaled_ref_frame[ref]) {
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[ref][i] = xd->plane[i].pre[ref];
av1_setup_pre_planes(xd, ref, scaled_ref_frame[ref], mi_row, mi_col,
NULL);
}
frame_mv[refs[ref]].as_int = single_newmv[refs[ref]].as_int;
}
// Since we have scaled the reference frames to match the size of the current
// frame we must use a unit scaling factor during mode selection.
#if CONFIG_AOM_HIGHBITDEPTH
av1_setup_scale_factors_for_frame(&sf, cm->width, cm->height, cm->width,
cm->height, cm->use_highbitdepth);
#else
av1_setup_scale_factors_for_frame(&sf, cm->width, cm->height, cm->width,
cm->height);
#endif // CONFIG_AOM_HIGHBITDEPTH
// Allow joint search multiple times iteratively for each reference frame
// and break out of the search loop if it couldn't find a better mv.
for (ite = 0; ite < 4; ite++) {
struct buf_2d ref_yv12[2];
int bestsme = INT_MAX;
int sadpb = x->sadperbit16;
MV tmp_mv;
int search_range = 3;
int tmp_col_min = x->mv_col_min;
int tmp_col_max = x->mv_col_max;
int tmp_row_min = x->mv_row_min;
int tmp_row_max = x->mv_row_max;
int id = ite % 2; // Even iterations search in the first reference frame,
// odd iterations search in the second. The predictor
// found for the 'other' reference frame is factored in.
// Initialized here because of compiler problem in Visual Studio.
ref_yv12[0] = xd->plane[0].pre[0];
ref_yv12[1] = xd->plane[0].pre[1];
// Get the prediction block from the 'other' reference frame.
#if CONFIG_AOM_HIGHBITDEPTH
if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
second_pred = CONVERT_TO_BYTEPTR(second_pred_alloc_16);
av1_highbd_build_inter_predictor(
ref_yv12[!id].buf, ref_yv12[!id].stride, second_pred, pw,
&frame_mv[refs[!id]].as_mv, &sf, pw, ph, 0, &mbmi->interp_filter,
MV_PRECISION_Q3, mi_col * MI_SIZE, mi_row * MI_SIZE, xd->bd);
} else {
second_pred = (uint8_t *)second_pred_alloc_16;
av1_build_inter_predictor(
ref_yv12[!id].buf, ref_yv12[!id].stride, second_pred, pw,
&frame_mv[refs[!id]].as_mv, &sf, pw, ph, 0, &mbmi->interp_filter,
MV_PRECISION_Q3, mi_col * MI_SIZE, mi_row * MI_SIZE);
}
#else
av1_build_inter_predictor(ref_yv12[!id].buf, ref_yv12[!id].stride,
second_pred, pw, &frame_mv[refs[!id]].as_mv, &sf