blob: b20d5733273bffc90f800d1a683d6bfe1a1b395c [file] [log] [blame]
/*
* Copyright (c) 2010 The WebM project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include <assert.h>
#include <math.h>
#include "./vp9_rtcd.h"
#include "./vpx_dsp_rtcd.h"
#include "vpx_dsp/vpx_dsp_common.h"
#include "vpx_mem/vpx_mem.h"
#include "vpx_ports/mem.h"
#include "vpx_ports/system_state.h"
#include "vp9/common/vp9_common.h"
#include "vp9/common/vp9_entropy.h"
#include "vp9/common/vp9_entropymode.h"
#include "vp9/common/vp9_idct.h"
#include "vp9/common/vp9_mvref_common.h"
#include "vp9/common/vp9_pred_common.h"
#include "vp9/common/vp9_quant_common.h"
#include "vp9/common/vp9_reconinter.h"
#include "vp9/common/vp9_reconintra.h"
#include "vp9/common/vp9_scan.h"
#include "vp9/common/vp9_seg_common.h"
#include "vp9/encoder/vp9_cost.h"
#include "vp9/encoder/vp9_encodemb.h"
#include "vp9/encoder/vp9_encodemv.h"
#include "vp9/encoder/vp9_encoder.h"
#include "vp9/encoder/vp9_mcomp.h"
#include "vp9/encoder/vp9_quantize.h"
#include "vp9/encoder/vp9_ratectrl.h"
#include "vp9/encoder/vp9_rd.h"
#include "vp9/encoder/vp9_rdopt.h"
#include "vp9/encoder/vp9_aq_variance.h"
#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 ALT_REF_MODE_MASK \
((1 << LAST_FRAME) | (1 << GOLDEN_FRAME) | (1 << INTRA_FRAME))
#define SECOND_REF_FRAME_MASK ((1 << ALTREF_FRAME) | 0x01)
#define MIN_EARLY_TERM_INDEX 3
#define NEW_MV_DISCOUNT_FACTOR 8
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 {
const VP9_COMP *cpi;
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 *so;
uint8_t skippable;
};
#define LAST_NEW_MV_INDEX 6
static const MODE_DEFINITION vp9_mode_order[MAX_MODES] = {
{ NEARESTMV, { LAST_FRAME, NONE } },
{ NEARESTMV, { ALTREF_FRAME, NONE } },
{ NEARESTMV, { GOLDEN_FRAME, NONE } },
{ DC_PRED, { INTRA_FRAME, NONE } },
{ NEWMV, { LAST_FRAME, NONE } },
{ NEWMV, { ALTREF_FRAME, NONE } },
{ NEWMV, { GOLDEN_FRAME, NONE } },
{ NEARMV, { LAST_FRAME, NONE } },
{ NEARMV, { ALTREF_FRAME, NONE } },
{ NEARMV, { GOLDEN_FRAME, NONE } },
{ ZEROMV, { LAST_FRAME, NONE } },
{ ZEROMV, { GOLDEN_FRAME, NONE } },
{ ZEROMV, { ALTREF_FRAME, NONE } },
{ NEARESTMV, { LAST_FRAME, ALTREF_FRAME } },
{ NEARESTMV, { GOLDEN_FRAME, ALTREF_FRAME } },
{ TM_PRED, { INTRA_FRAME, NONE } },
{ NEARMV, { LAST_FRAME, ALTREF_FRAME } },
{ NEWMV, { LAST_FRAME, ALTREF_FRAME } },
{ NEARMV, { GOLDEN_FRAME, ALTREF_FRAME } },
{ NEWMV, { GOLDEN_FRAME, ALTREF_FRAME } },
{ ZEROMV, { LAST_FRAME, ALTREF_FRAME } },
{ ZEROMV, { GOLDEN_FRAME, ALTREF_FRAME } },
{ 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 vp9_ref_order[MAX_REFS] = {
{ { LAST_FRAME, NONE } }, { { GOLDEN_FRAME, NONE } },
{ { ALTREF_FRAME, NONE } }, { { LAST_FRAME, ALTREF_FRAME } },
{ { GOLDEN_FRAME, ALTREF_FRAME } }, { { 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(VP9_COMP *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]->ref_frame[0];
unsigned int sse;
unsigned int var = 0;
int64_t total_sse = 0;
int skip_flag = 1;
const int shift = 6;
int64_t dist;
const int dequant_shift =
#if CONFIG_VP9_HIGHBITDEPTH
(xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd - 5 :
#endif // CONFIG_VP9_HIGHBITDEPTH
3;
unsigned int qstep_vec[MAX_MB_PLANE];
unsigned int nlog2_vec[MAX_MB_PLANE];
unsigned int sum_sse_vec[MAX_MB_PLANE];
int any_zero_sum_sse = 0;
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;
unsigned int sum_sse = 0;
// 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 = VPXMIN(50, dc_thr >> 2);
const int64_t low_ac_thr = VPXMIN(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;
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;
sum_sse_vec[i] = sum_sse;
any_zero_sum_sse = any_zero_sum_sse || (sum_sse == 0);
qstep_vec[i] = pd->dequant[1] >> dequant_shift;
nlog2_vec[i] = num_pels_log2_lookup[bs];
}
// Fast approximate the modelling function.
if (cpi->sf.simple_model_rd_from_var) {
for (i = 0; i < MAX_MB_PLANE; ++i) {
int64_t rate;
const int64_t square_error = sum_sse_vec[i];
int quantizer = qstep_vec[i];
if (quantizer < 120)
rate = (square_error * (280 - quantizer)) >> (16 - VP9_PROB_COST_SHIFT);
else
rate = 0;
dist = (square_error * quantizer) >> 8;
rate_sum += rate;
dist_sum += dist;
}
} else {
if (any_zero_sum_sse) {
for (i = 0; i < MAX_MB_PLANE; ++i) {
int rate;
vp9_model_rd_from_var_lapndz(sum_sse_vec[i], nlog2_vec[i], qstep_vec[i],
&rate, &dist);
rate_sum += rate;
dist_sum += dist;
}
} else {
vp9_model_rd_from_var_lapndz_vec(sum_sse_vec, nlog2_vec, qstep_vec,
&rate_sum, &dist_sum);
}
}
*skip_txfm_sb = skip_flag;
*skip_sse_sb = total_sse << 4;
*out_rate_sum = (int)rate_sum;
*out_dist_sum = dist_sum << 4;
}
#if CONFIG_VP9_HIGHBITDEPTH
int64_t vp9_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;
}
static int64_t vp9_highbd_block_error_dispatch(const tran_low_t *coeff,
const tran_low_t *dqcoeff,
intptr_t block_size,
int64_t *ssz, int bd) {
if (bd == 8) {
return vp9_block_error(coeff, dqcoeff, block_size, ssz);
} else {
return vp9_highbd_block_error(coeff, dqcoeff, block_size, ssz, bd);
}
}
#endif // CONFIG_VP9_HIGHBITDEPTH
int64_t vp9_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 vp9_block_error_fp_c(const tran_low_t *coeff, const tran_low_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;
}
/* 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, TX_SIZE tx_size,
int pt, const int16_t *scan, const int16_t *nb,
int use_fast_coef_costing) {
MACROBLOCKD *const xd = &x->e_mbd;
MODE_INFO *mi = xd->mi[0];
const struct macroblock_plane *p = &x->plane[plane];
const PLANE_TYPE type = get_plane_type(plane);
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(mi)];
uint8_t token_cache[32 * 32];
int cost;
#if CONFIG_VP9_HIGHBITDEPTH
const uint16_t *cat6_high_cost = vp9_get_high_cost_table(xd->bd);
#else
const uint16_t *cat6_high_cost = vp9_get_high_cost_table(8);
#endif
// Check for consistency of tx_size with mode info
assert(type == PLANE_TYPE_Y
? mi->tx_size == tx_size
: get_uv_tx_size(mi, &xd->plane[plane]) == tx_size);
if (eob == 0) {
// single eob token
cost = token_costs[0][0][pt][EOB_TOKEN];
} else {
if (use_fast_coef_costing) {
int band_left = *band_count++;
int c;
// dc token
int v = qcoeff[0];
int16_t prev_t;
cost = vp9_get_token_cost(v, &prev_t, cat6_high_cost);
cost += (*token_costs)[0][pt][prev_t];
token_cache[0] = vp9_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];
cost += vp9_get_token_cost(v, &t, cat6_high_cost);
cost += (*token_costs)[!prev_t][!prev_t][t];
prev_t = t;
if (!--band_left) {
band_left = *band_count++;
++token_costs;
}
}
// eob token
if (band_left) cost += (*token_costs)[0][!prev_t][EOB_TOKEN];
} else { // !use_fast_coef_costing
int band_left = *band_count++;
int c;
// dc token
int v = qcoeff[0];
int16_t tok;
unsigned int(*tok_cost_ptr)[COEFF_CONTEXTS][ENTROPY_TOKENS];
cost = vp9_get_token_cost(v, &tok, cat6_high_cost);
cost += (*token_costs)[0][pt][tok];
token_cache[0] = vp9_pt_energy_class[tok];
++token_costs;
tok_cost_ptr = &((*token_costs)[!tok]);
// ac tokens
for (c = 1; c < eob; c++) {
const int rc = scan[c];
v = qcoeff[rc];
cost += vp9_get_token_cost(v, &tok, cat6_high_cost);
pt = get_coef_context(nb, token_cache, c);
cost += (*tok_cost_ptr)[pt][tok];
token_cache[rc] = vp9_pt_energy_class[tok];
if (!--band_left) {
band_left = *band_count++;
++token_costs;
}
tok_cost_ptr = &((*token_costs)[!tok]);
}
// eob token
if (band_left) {
pt = get_coef_context(nb, token_cache, c);
cost += (*token_costs)[0][pt][EOB_TOKEN];
}
}
}
return cost;
}
static INLINE int num_4x4_to_edge(int plane_4x4_dim, int mb_to_edge_dim,
int subsampling_dim, int blk_dim) {
return plane_4x4_dim + (mb_to_edge_dim >> (5 + subsampling_dim)) - blk_dim;
}
// Compute the pixel domain sum square error on all visible 4x4s in the
// transform block.
static unsigned pixel_sse(const VP9_COMP *const cpi, const MACROBLOCKD *xd,
const struct macroblockd_plane *const pd,
const uint8_t *src, const int src_stride,
const uint8_t *dst, const int dst_stride, int blk_row,
int blk_col, const BLOCK_SIZE plane_bsize,
const BLOCK_SIZE tx_bsize) {
unsigned int sse = 0;
const int plane_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
const int plane_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
const int tx_4x4_w = num_4x4_blocks_wide_lookup[tx_bsize];
const int tx_4x4_h = num_4x4_blocks_high_lookup[tx_bsize];
int b4x4s_to_right_edge = num_4x4_to_edge(plane_4x4_w, xd->mb_to_right_edge,
pd->subsampling_x, blk_col);
int b4x4s_to_bottom_edge = num_4x4_to_edge(plane_4x4_h, xd->mb_to_bottom_edge,
pd->subsampling_y, blk_row);
if (tx_bsize == BLOCK_4X4 ||
(b4x4s_to_right_edge >= tx_4x4_w && b4x4s_to_bottom_edge >= tx_4x4_h)) {
cpi->fn_ptr[tx_bsize].vf(src, src_stride, dst, dst_stride, &sse);
} else {
const vpx_variance_fn_t vf_4x4 = cpi->fn_ptr[BLOCK_4X4].vf;
int r, c;
unsigned this_sse = 0;
int max_r = VPXMIN(b4x4s_to_bottom_edge, tx_4x4_h);
int max_c = VPXMIN(b4x4s_to_right_edge, tx_4x4_w);
sse = 0;
// if we are in the unrestricted motion border.
for (r = 0; r < max_r; ++r) {
// Skip visiting the sub blocks that are wholly within the UMV.
for (c = 0; c < max_c; ++c) {
vf_4x4(src + r * src_stride * 4 + c * 4, src_stride,
dst + r * dst_stride * 4 + c * 4, dst_stride, &this_sse);
sse += this_sse;
}
}
}
return sse;
}
// Compute the squares sum squares on all visible 4x4s in the transform block.
static int64_t sum_squares_visible(const MACROBLOCKD *xd,
const struct macroblockd_plane *const pd,
const int16_t *diff, const int diff_stride,
int blk_row, int blk_col,
const BLOCK_SIZE plane_bsize,
const BLOCK_SIZE tx_bsize) {
int64_t sse;
const int plane_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
const int plane_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
const int tx_4x4_w = num_4x4_blocks_wide_lookup[tx_bsize];
const int tx_4x4_h = num_4x4_blocks_high_lookup[tx_bsize];
int b4x4s_to_right_edge = num_4x4_to_edge(plane_4x4_w, xd->mb_to_right_edge,
pd->subsampling_x, blk_col);
int b4x4s_to_bottom_edge = num_4x4_to_edge(plane_4x4_h, xd->mb_to_bottom_edge,
pd->subsampling_y, blk_row);
if (tx_bsize == BLOCK_4X4 ||
(b4x4s_to_right_edge >= tx_4x4_w && b4x4s_to_bottom_edge >= tx_4x4_h)) {
assert(tx_4x4_w == tx_4x4_h);
sse = (int64_t)vpx_sum_squares_2d_i16(diff, diff_stride, tx_4x4_w << 2);
} else {
int r, c;
int max_r = VPXMIN(b4x4s_to_bottom_edge, tx_4x4_h);
int max_c = VPXMIN(b4x4s_to_right_edge, tx_4x4_w);
sse = 0;
// if we are in the unrestricted motion border.
for (r = 0; r < max_r; ++r) {
// Skip visiting the sub blocks that are wholly within the UMV.
for (c = 0; c < max_c; ++c) {
sse += (int64_t)vpx_sum_squares_2d_i16(
diff + r * diff_stride * 4 + c * 4, diff_stride, 4);
}
}
}
return sse;
}
static void dist_block(const VP9_COMP *cpi, MACROBLOCK *x, int plane,
BLOCK_SIZE plane_bsize, int block, int blk_row,
int blk_col, TX_SIZE tx_size, int64_t *out_dist,
int64_t *out_sse) {
MACROBLOCKD *const xd = &x->e_mbd;
const struct macroblock_plane *const p = &x->plane[plane];
const struct macroblockd_plane *const pd = &xd->plane[plane];
const int eob = p->eobs[block];
if (x->block_tx_domain && eob) {
const int ss_txfrm_size = tx_size << 1;
int64_t this_sse;
const int shift = tx_size == TX_32X32 ? 0 : 2;
const tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
const tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
#if CONFIG_VP9_HIGHBITDEPTH
const int bd = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd : 8;
*out_dist = vp9_highbd_block_error_dispatch(
coeff, dqcoeff, 16 << ss_txfrm_size, &this_sse, bd) >>
shift;
#else
*out_dist =
vp9_block_error(coeff, dqcoeff, 16 << ss_txfrm_size, &this_sse) >>
shift;
#endif // CONFIG_VP9_HIGHBITDEPTH
*out_sse = this_sse >> shift;
if (x->skip_encode && !is_inter_block(xd->mi[0])) {
// TODO(jingning): tune the model to better capture the distortion.
const int64_t p =
(pd->dequant[1] * pd->dequant[1] * (1 << ss_txfrm_size)) >>
#if CONFIG_VP9_HIGHBITDEPTH
(shift + 2 + (bd - 8) * 2);
#else
(shift + 2);
#endif // CONFIG_VP9_HIGHBITDEPTH
*out_dist += (p >> 4);
*out_sse += p;
}
} else {
const BLOCK_SIZE tx_bsize = txsize_to_bsize[tx_size];
const int bs = 4 * num_4x4_blocks_wide_lookup[tx_bsize];
const int src_stride = p->src.stride;
const int dst_stride = pd->dst.stride;
const int src_idx = 4 * (blk_row * src_stride + blk_col);
const int dst_idx = 4 * (blk_row * dst_stride + blk_col);
const uint8_t *src = &p->src.buf[src_idx];
const uint8_t *dst = &pd->dst.buf[dst_idx];
const tran_low_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
unsigned int tmp;
tmp = pixel_sse(cpi, xd, pd, src, src_stride, dst, dst_stride, blk_row,
blk_col, plane_bsize, tx_bsize);
*out_sse = (int64_t)tmp * 16;
if (eob) {
#if CONFIG_VP9_HIGHBITDEPTH
DECLARE_ALIGNED(16, uint16_t, recon16[1024]);
uint8_t *recon = (uint8_t *)recon16;
#else
DECLARE_ALIGNED(16, uint8_t, recon[1024]);
#endif // CONFIG_VP9_HIGHBITDEPTH
#if CONFIG_VP9_HIGHBITDEPTH
if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
vpx_highbd_convolve_copy(CONVERT_TO_SHORTPTR(dst), dst_stride, recon16,
32, NULL, 0, 0, 0, 0, bs, bs, xd->bd);
if (xd->lossless) {
vp9_highbd_iwht4x4_add(dqcoeff, recon16, 32, eob, xd->bd);
} else {
switch (tx_size) {
case TX_4X4:
vp9_highbd_idct4x4_add(dqcoeff, recon16, 32, eob, xd->bd);
break;
case TX_8X8:
vp9_highbd_idct8x8_add(dqcoeff, recon16, 32, eob, xd->bd);
break;
case TX_16X16:
vp9_highbd_idct16x16_add(dqcoeff, recon16, 32, eob, xd->bd);
break;
default:
assert(tx_size == TX_32X32);
vp9_highbd_idct32x32_add(dqcoeff, recon16, 32, eob, xd->bd);
break;
}
}
recon = CONVERT_TO_BYTEPTR(recon16);
} else {
#endif // CONFIG_VP9_HIGHBITDEPTH
vpx_convolve_copy(dst, dst_stride, recon, 32, NULL, 0, 0, 0, 0, bs, bs);
switch (tx_size) {
case TX_32X32: vp9_idct32x32_add(dqcoeff, recon, 32, eob); break;
case TX_16X16: vp9_idct16x16_add(dqcoeff, recon, 32, eob); break;
case TX_8X8: vp9_idct8x8_add(dqcoeff, recon, 32, eob); break;
default:
assert(tx_size == TX_4X4);
// this is like vp9_short_idct4x4 but has a special case around
// eob<=1, which is significant (not just an optimization) for
// the lossless case.
x->inv_txfm_add(dqcoeff, recon, 32, eob);
break;
}
#if CONFIG_VP9_HIGHBITDEPTH
}
#endif // CONFIG_VP9_HIGHBITDEPTH
tmp = pixel_sse(cpi, xd, pd, src, src_stride, recon, 32, blk_row, blk_col,
plane_bsize, tx_bsize);
}
*out_dist = (int64_t)tmp * 16;
}
}
static int rate_block(int plane, int block, TX_SIZE tx_size, int coeff_ctx,
struct rdcost_block_args *args) {
return cost_coeffs(args->x, plane, block, tx_size, coeff_ctx, args->so->scan,
args->so->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;
MODE_INFO *const mi = xd->mi[0];
int64_t rd1, rd2, rd;
int rate;
int64_t dist;
int64_t sse;
const int coeff_ctx =
combine_entropy_contexts(args->t_left[blk_row], args->t_above[blk_col]);
if (args->exit_early) return;
if (!is_inter_block(mi)) {
struct encode_b_args intra_arg = { x, x->block_qcoeff_opt, args->t_above,
args->t_left, &mi->skip };
vp9_encode_block_intra(plane, block, blk_row, blk_col, plane_bsize, tx_size,
&intra_arg);
if (x->block_tx_domain) {
dist_block(args->cpi, x, plane, plane_bsize, block, blk_row, blk_col,
tx_size, &dist, &sse);
} else {
const BLOCK_SIZE tx_bsize = txsize_to_bsize[tx_size];
const struct macroblock_plane *const p = &x->plane[plane];
const struct macroblockd_plane *const pd = &xd->plane[plane];
const int src_stride = p->src.stride;
const int dst_stride = pd->dst.stride;
const int diff_stride = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
const uint8_t *src = &p->src.buf[4 * (blk_row * src_stride + blk_col)];
const uint8_t *dst = &pd->dst.buf[4 * (blk_row * dst_stride + blk_col)];
const int16_t *diff = &p->src_diff[4 * (blk_row * diff_stride + blk_col)];
unsigned int tmp;
sse = sum_squares_visible(xd, pd, diff, diff_stride, blk_row, blk_col,
plane_bsize, tx_bsize);
#if CONFIG_VP9_HIGHBITDEPTH
if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) && (xd->bd > 8))
sse = ROUND64_POWER_OF_TWO(sse, (xd->bd - 8) * 2);
#endif // CONFIG_VP9_HIGHBITDEPTH
sse = sse * 16;
tmp = pixel_sse(args->cpi, xd, pd, src, src_stride, dst, dst_stride,
blk_row, blk_col, plane_bsize, tx_bsize);
dist = (int64_t)tmp * 16;
}
} 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
vp9_xform_quant(x, plane, block, blk_row, blk_col, plane_bsize, tx_size);
if (x->block_qcoeff_opt)
vp9_optimize_b(x, plane, block, tx_size, coeff_ctx);
dist_block(args->cpi, x, plane, plane_bsize, block, blk_row, blk_col,
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);
vp9_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_VP9_HIGHBITDEPTH
dc_correct >>= ((xd->bd - 8) * 2);
#endif
if (tx_size != TX_32X32) dc_correct >>= 2;
dist = VPXMAX(0, sse - dc_correct);
}
} else {
// SKIP_TXFM_AC_DC
// skip forward transform. Because this is handled here, the quantization
// does not need to do it.
x->plane[plane].eobs[block] = 0;
sse = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4;
dist = sse;
}
} else {
// full forward transform and quantization
vp9_xform_quant(x, plane, block, blk_row, blk_col, plane_bsize, tx_size);
if (x->block_qcoeff_opt)
vp9_optimize_b(x, plane, block, tx_size, coeff_ctx);
dist_block(args->cpi, x, plane, plane_bsize, block, blk_row, blk_col,
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, tx_size, coeff_ctx, args);
args->t_above[blk_col] = (x->plane[plane].eobs[block] > 0) ? 1 : 0;
args->t_left[blk_row] = (x->plane[plane].eobs[block] > 0) ? 1 : 0;
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 = VPXMIN(rd1, rd2);
if (plane == 0) {
x->zcoeff_blk[tx_size][block] =
!x->plane[plane].eobs[block] || (rd1 > rd2 && !xd->lossless);
x->sum_y_eobs[tx_size] += x->plane[plane].eobs[block];
}
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(const VP9_COMP *cpi, 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_costing) {
MACROBLOCKD *const xd = &x->e_mbd;
const struct macroblockd_plane *const pd = &xd->plane[plane];
struct rdcost_block_args args;
vp9_zero(args);
args.cpi = cpi;
args.x = x;
args.best_rd = ref_best_rd;
args.use_fast_coef_costing = use_fast_coef_costing;
args.skippable = 1;
if (plane == 0) xd->mi[0]->tx_size = tx_size;
vp9_get_entropy_contexts(bsize, tx_size, pd, args.t_above, args.t_left);
args.so = get_scan(xd, tx_size, get_plane_type(plane), 0);
vp9_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(VP9_COMP *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];
VP9_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;
MODE_INFO *const mi = xd->mi[0];
mi->tx_size = VPXMIN(max_tx_size, largest_tx_size);
txfm_rd_in_plane(cpi, x, rate, distortion, skip, sse, ref_best_rd, 0, bs,
mi->tx_size, cpi->sf.use_fast_coef_costing);
}
static void choose_tx_size_from_rd(VP9_COMP *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];
VP9_COMMON *const cm = &cpi->common;
MACROBLOCKD *const xd = &x->e_mbd;
MODE_INFO *const mi = xd->mi[0];
vpx_prob skip_prob = vp9_get_skip_prob(cm, xd);
int r[TX_SIZES][2], s[TX_SIZES];
int64_t d[TX_SIZES], sse[TX_SIZES];
int64_t rd[TX_SIZES][2] = { { INT64_MAX, INT64_MAX },
{ INT64_MAX, INT64_MAX },
{ INT64_MAX, INT64_MAX },
{ INT64_MAX, INT64_MAX } };
int n;
int s0, s1;
int64_t best_rd = ref_best_rd;
TX_SIZE best_tx = max_tx_size;
int start_tx, end_tx;
const int tx_size_ctx = get_tx_size_context(xd);
assert(skip_prob > 0);
s0 = vp9_cost_bit(skip_prob, 0);
s1 = vp9_cost_bit(skip_prob, 1);
if (cm->tx_mode == TX_MODE_SELECT) {
start_tx = max_tx_size;
end_tx = VPXMAX(start_tx - cpi->sf.tx_size_search_depth, 0);
if (bs > BLOCK_32X32) end_tx = VPXMIN(end_tx + 1, start_tx);
} else {
TX_SIZE chosen_tx_size =
VPXMIN(max_tx_size, tx_mode_to_biggest_tx_size[cm->tx_mode]);
start_tx = chosen_tx_size;
end_tx = chosen_tx_size;
}
for (n = start_tx; n >= end_tx; n--) {
const int r_tx_size = cpi->tx_size_cost[max_tx_size - 1][tx_size_ctx][n];
txfm_rd_in_plane(cpi, x, &r[n][0], &d[n], &s[n], &sse[n], best_rd, 0, bs, n,
cpi->sf.use_fast_coef_costing);
r[n][1] = r[n][0];
if (r[n][0] < INT_MAX) {
r[n][1] += r_tx_size;
}
if (d[n] == INT64_MAX || r[n][0] == INT_MAX) {
rd[n][0] = rd[n][1] = INT64_MAX;
} else if (s[n]) {
if (is_inter_block(mi)) {
rd[n][0] = rd[n][1] = RDCOST(x->rdmult, x->rddiv, s1, sse[n]);
r[n][1] -= r_tx_size;
} else {
rd[n][0] = RDCOST(x->rdmult, x->rddiv, s1, sse[n]);
rd[n][1] = RDCOST(x->rdmult, x->rddiv, s1 + r_tx_size, sse[n]);
}
} else {
rd[n][0] = RDCOST(x->rdmult, x->rddiv, r[n][0] + s0, d[n]);
rd[n][1] = RDCOST(x->rdmult, x->rddiv, r[n][1] + s0, d[n]);
}
if (is_inter_block(mi) && !xd->lossless && !s[n] && sse[n] != INT64_MAX) {
rd[n][0] = VPXMIN(rd[n][0], RDCOST(x->rdmult, x->rddiv, s1, sse[n]));
rd[n][1] = VPXMIN(rd[n][1], RDCOST(x->rdmult, x->rddiv, s1, sse[n]));
}
// Early termination in transform size search.
if (cpi->sf.tx_size_search_breakout &&
(rd[n][1] == INT64_MAX ||
(n < (int)max_tx_size && rd[n][1] > rd[n + 1][1]) || s[n] == 1))
break;
if (rd[n][1] < best_rd) {
best_tx = n;
best_rd = rd[n][1];
}
}
mi->tx_size = best_tx;
*distortion = d[mi->tx_size];
*rate = r[mi->tx_size][cm->tx_mode == TX_MODE_SELECT];
*skip = s[mi->tx_size];
*psse = sse[mi->tx_size];
}
static void super_block_yrd(VP9_COMP *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]->sb_type);
if (cpi->sf.tx_size_search_method == USE_LARGESTALL || xd->lossless) {
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(VP9_COMP *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_VP9_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]->tx_size = TX_4X4;
#if CONFIG_VP9_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];
uint16_t *const dst16 = CONVERT_TO_SHORTPTR(dst);
int16_t *const src_diff =
vp9_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;
vp9_predict_intra_block(xd, 1, TX_4X4, mode,
x->skip_encode ? src : dst,
x->skip_encode ? src_stride : dst_stride, dst,
dst_stride, col + idx, row + idy, 0);
vpx_highbd_subtract_block(4, 4, src_diff, 8, src, src_stride, dst,
dst_stride, xd->bd);
if (xd->lossless) {
const scan_order *so = &vp9_default_scan_orders[TX_4X4];
const int coeff_ctx =
combine_entropy_contexts(tempa[idx], templ[idy]);
vp9_highbd_fwht4x4(src_diff, coeff, 8);
vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
ratey += cost_coeffs(x, 0, block, TX_4X4, coeff_ctx, so->scan,
so->neighbors, cpi->sf.use_fast_coef_costing);
tempa[idx] = templ[idy] = (x->plane[0].eobs[block] > 0 ? 1 : 0);
if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
goto next_highbd;
vp9_highbd_iwht4x4_add(BLOCK_OFFSET(pd->dqcoeff, block), dst16,
dst_stride, p->eobs[block], xd->bd);
} else {
int64_t unused;
const TX_TYPE tx_type = get_tx_type_4x4(PLANE_TYPE_Y, xd, block);
const scan_order *so = &vp9_scan_orders[TX_4X4][tx_type];
const int coeff_ctx =
combine_entropy_contexts(tempa[idx], templ[idy]);
if (tx_type == DCT_DCT)
vpx_highbd_fdct4x4(src_diff, coeff, 8);
else
vp9_highbd_fht4x4(src_diff, coeff, 8, tx_type);
vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
ratey += cost_coeffs(x, 0, block, TX_4X4, coeff_ctx, so->scan,
so->neighbors, cpi->sf.use_fast_coef_costing);
distortion += vp9_highbd_block_error_dispatch(
coeff, BLOCK_OFFSET(pd->dqcoeff, block), 16,
&unused, xd->bd) >>
2;
tempa[idx] = templ[idy] = (x->plane[0].eobs[block] > 0 ? 1 : 0);
if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
goto next_highbd;
vp9_highbd_iht4x4_add(tx_type, BLOCK_OFFSET(pd->dqcoeff, block),
dst16, dst_stride, p->eobs[block], xd->bd);
}
}
}
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 || x->skip_encode) 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_VP9_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 =
vp9_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;
vp9_predict_intra_block(xd, 1, TX_4X4, mode, x->skip_encode ? src : dst,
x->skip_encode ? src_stride : dst_stride, dst,
dst_stride, col + idx, row + idy, 0);
vpx_subtract_block(4, 4, src_diff, 8, src, src_stride, dst, dst_stride);
if (xd->lossless) {
const scan_order *so = &vp9_default_scan_orders[TX_4X4];
const int coeff_ctx =
combine_entropy_contexts(tempa[idx], templ[idy]);
vp9_fwht4x4(src_diff, coeff, 8);
vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
ratey += cost_coeffs(x, 0, block, TX_4X4, coeff_ctx, so->scan,
so->neighbors, cpi->sf.use_fast_coef_costing);
tempa[idx] = templ[idy] = (x->plane[0].eobs[block] > 0) ? 1 : 0;
if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
goto next;
vp9_iwht4x4_add(BLOCK_OFFSET(pd->dqcoeff, block), dst, dst_stride,
p->eobs[block]);
} else {
int64_t unused;
const TX_TYPE tx_type = get_tx_type_4x4(PLANE_TYPE_Y, xd, block);
const scan_order *so = &vp9_scan_orders[TX_4X4][tx_type];
const int coeff_ctx =
combine_entropy_contexts(tempa[idx], templ[idy]);
vp9_fht4x4(src_diff, coeff, 8, tx_type);
vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
ratey += cost_coeffs(x, 0, block, TX_4X4, coeff_ctx, so->scan,
so->neighbors, cpi->sf.use_fast_coef_costing);
tempa[idx] = templ[idy] = (x->plane[0].eobs[block] > 0) ? 1 : 0;
distortion += vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, block),
16, &unused) >>
2;
if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
goto next;
vp9_iht4x4_add(tx_type, BLOCK_OFFSET(pd->dqcoeff, block), dst,
dst_stride, p->eobs[block]);
}
}
}
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 || x->skip_encode) 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(VP9_COMP *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]->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;
// 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 = vp9_above_block_mode(mic, above_mi, i);
const PREDICTION_MODE L = vp9_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->mode = mic->bmi[3].as_mode;
return RDCOST(mb->rdmult, mb->rddiv, cost, total_distortion);
}
// This function is used only for intra_only frames
static int64_t rd_pick_intra_sby_mode(VP9_COMP *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;
MODE_INFO *const mic = xd->mi[0];
int this_rate, this_rate_tokenonly, s;
int64_t this_distortion, this_rd;
TX_SIZE best_tx = TX_4X4;
int *bmode_costs;
const MODE_INFO *above_mi = xd->above_mi;
const MODE_INFO *left_mi = xd->left_mi;
const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, 0);
const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, 0);
bmode_costs = cpi->y_mode_costs[A][L];
memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
/* Y Search for intra prediction mode */
for (mode = DC_PRED; mode <= TM_PRED; mode++) {
if (cpi->sf.use_nonrd_pick_mode) {
// These speed features are turned on in hybrid non-RD and RD mode
// for key frame coding in the context of real-time setting.
if (conditional_skipintra(mode, mode_selected)) continue;
if (*skippable) break;
}
mic->mode = mode;
super_block_yrd(cpi, x, &this_rate_tokenonly, &this_distortion, &s, NULL,
bsize, best_rd);
if (this_rate_tokenonly == INT_MAX) continue;
this_rate = this_rate_tokenonly + bmode_costs[mode];
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 = mic->tx_size;
*rate = this_rate;
*rate_tokenonly = this_rate_tokenonly;
*distortion = this_distortion;
*skippable = s;
}
}
mic->mode = mode_selected;
mic->tx_size = best_tx;
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 VP9_COMP *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;
MODE_INFO *const mi = xd->mi[0];
const TX_SIZE uv_tx_size = get_uv_tx_size(mi, &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(mi) && is_cost_valid) {
int plane;
for (plane = 1; plane < MAX_MB_PLANE; ++plane)
vp9_subtract_plane(x, bsize, plane);
}
*rate = 0;
*distortion = 0;
*sse = 0;
*skippable = 1;
for (plane = 1; plane < MAX_MB_PLANE; ++plane) {
txfm_rd_in_plane(cpi, 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;
}
static int64_t rd_pick_intra_sbuv_mode(VP9_COMP *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) {
MACROBLOCKD *xd = &x->e_mbd;
PREDICTION_MODE mode;
PREDICTION_MODE mode_selected = DC_PRED;
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;
#if CONFIG_BETTER_HW_COMPATIBILITY && CONFIG_VP9_HIGHBITDEPTH
if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) &&
(xd->above_mi == NULL || xd->left_mi == NULL) && need_top_left[mode])
continue;
#endif // CONFIG_BETTER_HW_COMPATIBILITY && CONFIG_VP9_HIGHBITDEPTH
xd->mi[0]->uv_mode = mode;
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[cpi->common.frame_type][xd->mi[0]->mode][mode];
this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
if (this_rd < best_rd) {
mode_selected = mode;
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);
}
}
xd->mi[0]->uv_mode = mode_selected;
return best_rd;
}
static int64_t rd_sbuv_dcpred(const VP9_COMP *cpi, MACROBLOCK *x, int *rate,
int *rate_tokenonly, int64_t *distortion,
int *skippable, BLOCK_SIZE bsize) {
const VP9_COMMON *cm = &cpi->common;
int64_t unused;
x->e_mbd.mi[0]->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[cm->frame_type][x->e_mbd.mi[0]->mode][DC_PRED];
return RDCOST(x->rdmult, x->rddiv, *rate, *distortion);
}
static void choose_intra_uv_mode(VP9_COMP *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.
if (cpi->sf.use_uv_intra_rd_estimate) {
rd_sbuv_dcpred(cpi, x, rate_uv, rate_uv_tokenonly, dist_uv, skip_uv,
bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize);
// Else do a proper rd search for each possible transform size that may
// be considered in the main rd loop.
} else {
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]->uv_mode;
}
static int cost_mv_ref(const VP9_COMP *cpi, PREDICTION_MODE mode,
int mode_context) {
assert(is_inter_mode(mode));
return cpi->inter_mode_cost[mode_context][INTER_OFFSET(mode)];
}
static int set_and_cost_bmi_mvs(VP9_COMP *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 mi = xd->mi[0];
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[mi->sb_type];
const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[mi->sb_type];
const int is_compound = has_second_ref(mi);
switch (mode) {
case NEWMV:
this_mv[0].as_int = seg_mvs[mi->ref_frame[0]].as_int;
thismvcost += vp9_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[mi->ref_frame[1]].as_int;
thismvcost += vp9_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][mi->ref_frame[0]].as_int;
if (is_compound)
this_mv[1].as_int = frame_mv[mode][mi->ref_frame[1]].as_int;
break;
default:
assert(mode == ZEROMV);
this_mv[0].as_int = 0;
if (is_compound) this_mv[1].as_int = 0;
break;
}
mi->bmi[i].as_mv[0].as_int = this_mv[0].as_int;
if (is_compound) mi->bmi[i].as_mv[1].as_int = this_mv[1].as_int;
mi->bmi[i].as_mode = mode;
for (idy = 0; idy < num_4x4_blocks_high; ++idy)
for (idx = 0; idx < num_4x4_blocks_wide; ++idx)
memmove(&mi->bmi[i + idy * 2 + idx], &mi->bmi[i], sizeof(mi->bmi[i]));
return cost_mv_ref(cpi, mode, mbmi_ext->mode_context[mi->ref_frame[0]]) +
thismvcost;
}
static int64_t encode_inter_mb_segment(VP9_COMP *cpi, MACROBLOCK *x,
int64_t best_yrd, int i, int *labelyrate,
int64_t *distortion, int64_t *sse,
ENTROPY_CONTEXT *ta, ENTROPY_CONTEXT *tl,
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->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;
const uint8_t *const src =
&p->src.buf[vp9_raster_block_offset(BLOCK_8X8, i, p->src.stride)];
uint8_t *const dst =
&pd->dst.buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->dst.stride)];
int64_t thisdistortion = 0, thissse = 0;
int thisrate = 0, ref;
const scan_order *so = &vp9_default_scan_orders[TX_4X4];
const int is_compound = has_second_ref(mi);
const InterpKernel *kernel = vp9_filter_kernels[mi->interp_filter];
for (ref = 0; ref < 1 + is_compound; ++ref) {
const int bw = b_width_log2_lookup[BLOCK_8X8];
const int h = 4 * (i >> bw);
const int w = 4 * (i & ((1 << bw) - 1));
const struct scale_factors *sf = &xd->block_refs[ref]->sf;
int y_stride = pd->pre[ref].stride;
uint8_t *pre = pd->pre[ref].buf + (h * pd->pre[ref].stride + w);
if (vp9_is_scaled(sf)) {
const int x_start = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x));
const int y_start = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y));
y_stride = xd->block_refs[ref]->buf->y_stride;
pre = xd->block_refs[ref]->buf->y_buffer;
pre += scaled_buffer_offset(x_start + w, y_start + h, y_stride, sf);
}
#if CONFIG_VP9_HIGHBITDEPTH
if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
vp9_highbd_build_inter_predictor(
CONVERT_TO_SHORTPTR(pre), y_stride, CONVERT_TO_SHORTPTR(dst),
pd->dst.stride, &mi->bmi[i].as_mv[ref].as_mv,
&xd->block_refs[ref]->sf, width, height, ref, kernel, MV_PRECISION_Q3,
mi_col * MI_SIZE + 4 * (i % 2), mi_row * MI_SIZE + 4 * (i / 2),
xd->bd);
} else {
vp9_build_inter_predictor(
pre, y_stride, dst, pd->dst.stride, &mi->bmi[i].as_mv[ref].as_mv,
&xd->block_refs[ref]->sf, width, height, ref, kernel, MV_PRECISION_Q3,
mi_col * MI_SIZE + 4 * (i % 2), mi_row * MI_SIZE + 4 * (i / 2));
}
#else
vp9_build_inter_predictor(
pre, y_stride, dst, pd->dst.stride, &mi->bmi[i].as_mv[ref].as_mv,
&xd->block_refs[ref]->sf, width, height, ref, kernel, MV_PRECISION_Q3,
mi_col * MI_SIZE + 4 * (i % 2), mi_row * MI_SIZE + 4 * (i / 2));
#endif // CONFIG_VP9_HIGHBITDEPTH
}
#if CONFIG_VP9_HIGHBITDEPTH
if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
vpx_highbd_subtract_block(
height, width, vp9_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
8, src, p->src.stride, dst, pd->dst.stride, xd->bd);
} else {
vpx_subtract_block(height, width,
vp9_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
8, src, p->src.stride, dst, pd->dst.stride);
}
#else
vpx_subtract_block(height, width,
vp9_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
8, src, p->src.stride, dst, pd->dst.stride);
#endif // CONFIG_VP9_HIGHBITDEPTH
k = i;
for (idy = 0; idy < height / 4; ++idy) {
for (idx = 0; idx < width / 4; ++idx) {
#if CONFIG_VP9_HIGHBITDEPTH
const int bd = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd : 8;
#endif
int64_t ssz, rd, rd1, rd2;
tran_low_t *coeff;
int coeff_ctx;
k += (idy * 2 + idx);
coeff_ctx = combine_entropy_contexts(ta[k & 1], tl[k >> 1]);
coeff = BLOCK_OFFSET(p->coeff, k);
x->fwd_txfm4x4(vp9_raster_block_offset_int16(BLOCK_8X8, k, p->src_diff),
coeff, 8);
vp9_regular_quantize_b_4x4(x, 0, k, so->scan, so->iscan);
#if CONFIG_VP9_HIGHBITDEPTH
thisdistortion += vp9_highbd_block_error_dispatch(
coeff, BLOCK_OFFSET(pd->dqcoeff, k), 16, &ssz, bd);
#else
thisdistortion +=
vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, k), 16, &ssz);
#endif // CONFIG_VP9_HIGHBITDEPTH
thissse += ssz;
thisrate += cost_coeffs(x, 0, k, TX_4X4, coeff_ctx, so->scan,
so->neighbors, cpi->sf.use_fast_coef_costing);
ta[k & 1] = tl[k >> 1] = (x->plane[0].eobs[k] > 0) ? 1 : 0;
rd1 = RDCOST(x->rdmult, x->rddiv, thisrate, thisdistortion >> 2);
rd2 = RDCOST(x->rdmult, x->rddiv, 0, thissse >> 2);
rd = VPXMIN(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 MvLimits *mv_limits, const MV *mv) {
return (mv->row >> 3) < mv_limits->row_min ||
(mv->row >> 3) > mv_limits->row_max ||
(mv->col >> 3) < mv_limits->col_min ||
(mv->col >> 3) > mv_limits->col_max;
}
static INLINE void mi_buf_shift(MACROBLOCK *x, int i) {
MODE_INFO *const mi = x->e_mbd.mi[0];
struct macroblock_plane *const p = &x->plane[0];
struct macroblockd_plane *const pd = &x->e_mbd.plane[0];
p->src.buf =
&p->src.buf[vp9_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[vp9_raster_block_offset(BLOCK_8X8, i, pd->pre[0].stride)];
if (has_second_ref(mi))
pd->pre[1].buf =
&pd->pre[1]
.buf[vp9_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]) {
MODE_INFO *mi = x->e_mbd.mi[0];
x->plane[0].src = orig_src;
x->e_mbd.plane[0].pre[0] = orig_pre[0];
if (has_second_ref(mi)) 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 VP9_COMP *cpi,
const uint8_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]) {
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)) {
int rfc = mode_context[ref_frames[0]];
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 (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(VP9_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 VP9_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;
MODE_INFO *mi = xd->mi[0];
const int refs[2] = { mi->ref_frame[0],
mi->ref_frame[1] < 0 ? 0 : mi->ref_frame[1] };
int_mv ref_mv[2];
int ite, ref;
const InterpKernel *kernel = vp9_filter_kernels[mi->interp_filter];
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];
uint32_t last_besterr[2] = { UINT_MAX, UINT_MAX };
const YV12_BUFFER_CONFIG *const scaled_ref_frame[2] = {
vp9_get_scaled_ref_frame(cpi, mi->ref_frame[0]),
vp9_get_scaled_ref_frame(cpi, mi->ref_frame[1])
};
// Prediction buffer from second frame.
#if CONFIG_VP9_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_VP9_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];
vp9_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_VP9_HIGHBITDEPTH
vp9_setup_scale_factors_for_frame(&sf, cm->width, cm->height, cm->width,
cm->height, cm->use_highbitdepth);
#else
vp9_setup_scale_factors_for_frame(&sf, cm->width, cm->height, cm->width,
cm->height);
#endif // CONFIG_VP9_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];
uint32_t bestsme = UINT_MAX;
int sadpb = x->sadperbit16;
MV tmp_mv;
int search_range = 3;
const MvLimits tmp_mv_limits = x->mv_limits;
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_VP9_HIGHBITDEPTH
if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
second_pred = CONVERT_TO_BYTEPTR(second_pred_alloc_16);
vp9_highbd_build_inter_predictor(
CONVERT_TO_SHORTPTR(ref_yv12[!id].buf), ref_yv12[!id].stride,
second_pred_alloc_16, pw, &frame_mv[refs[!id]].as_mv, &sf, pw, ph, 0,
kernel, MV_PRECISION_Q3, mi_col * MI_SIZE, mi_row * MI_SIZE, xd->bd);
} else {
second_pred = (uint8_t *)second_pred_alloc_16;
vp9_build_inter_predictor(ref_yv12[!id].buf, ref_yv12[!id].stride,
second_pred, pw, &frame_mv[refs[!id]].as_mv,
&sf, pw, ph, 0, kernel, MV_PRECISION_Q3,
mi_col * MI_SIZE, mi_row * MI_SIZE);
}
#else
vp9_build_inter_predictor(ref_yv12[!id].buf, ref_yv12[!id].stride,
second_pred, pw, &frame_mv[refs[!id]].as_mv, &sf,
pw, ph, 0, kernel, MV_PRECISION_Q3,
mi_col * MI_SIZE, mi_row * MI_SIZE);
#endif // CONFIG_VP9_HIGHBITDEPTH
// Do compound motion search on the current reference frame.
if (id) xd->plane[0].pre[0] = ref_yv12[id];
vp9_set_mv_search_range(&x->mv_limits, &ref_mv[id].as_mv);
// Use the mv result from the single mode as mv predictor.
tmp_mv = frame_mv[refs[id]].as_mv;
tmp_mv.col >>= 3;
tmp_mv.row >>= 3;
// Small-range full-pixel motion search.
bestsme = vp9_refining_search_8p_c(x, &tmp_mv, sadpb, search_range,
&cpi->fn_ptr[bsize], &ref_mv[id].as_mv,
second_pred);
if (bestsme < UINT_MAX)
bestsme = vp9_get_mvpred_av_var(x, &tmp_mv, &ref_mv[id].as_mv,
second_pred, &cpi->fn_ptr[bsize], 1);
x->mv_limits = tmp_mv_limits;
if (bestsme < UINT_MAX) {
uint32_t dis; /* TODO: use dis in distortion calculation later. */
uint32_t sse;
bestsme = cpi->find_fractional_mv_step(
x, &tmp_mv, &ref_mv[id].as_mv, cpi->common.allow_high_precision_mv,
x->errorperbit, &cpi->fn_ptr[bsize], 0,
cpi->sf.mv.subpel_iters_per_step, NULL, x->nmvjointcost, x->mvcost,
&dis, &sse, second_pred, pw, ph);
}
// Restore the pointer to the first (possibly scaled) prediction buffer.
if (id) xd->plane[0].pre[0] = ref_yv12[0];
if (bestsme < last_besterr[id]) {
frame_mv[refs[id]].as_mv = tmp_mv;
last_besterr[id] = bestsme;
} else {
break;
}
}
*rate_mv = 0;
for (ref = 0; ref < 2; ++ref) {
if (scaled_ref_frame[ref]) {
// Restore the prediction frame pointers to their unscaled versions.
int i;
for (i = 0; i < MAX_MB_PLANE; i++)
xd->plane[i].pre[ref] = backup_yv12[ref][i];
}
*rate_mv += vp9_mv_bit_cost(&frame_mv[refs[ref]].as_mv,
&x->mbmi_ext->ref_mvs[refs[ref]][0].as_mv,
x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
}
}
static int64_t rd_pick_best_sub8x8_mode(
VP9_COMP *cpi, MACROBLOCK *x, int_mv *best_ref_mv,
int_mv *second_best_ref_mv, int64_t best_rd, int *returntotrate,
int *returnyrate, int64_t *returndistortion, int *skippable, int64_t *psse,
int mvthresh, int_mv seg_mvs[4][MAX_REF_FRAMES], BEST_SEG_INFO *bsi_buf,
int filter_idx, int mi_row, int mi_col) {
int i;
BEST_SEG_INFO *bsi = bsi_buf + filter_idx;
MACROBLOCKD *xd = &x->e_mbd;
MODE_INFO *mi = xd->mi[0];
int mode_idx;
int k, br = 0, idx, idy;
int64_t bd = 0, block_sse = 0;
PREDICTION_MODE this_mode;
VP9_COMMON *cm = &cpi->common;
struct macroblock_plane *const p = &x->plane[0];
struct macroblockd_plane *const pd = &xd->plane[0];
const int label_count = 4;
int64_t this_segment_rd = 0;
int label_mv_thresh;
int segmentyrate = 0;
const BLOCK_SIZE bsize = mi->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];
ENTROPY_CONTEXT t_above[2], t_left[2];
int subpelmv = 1, have_ref = 0;
SPEED_FEATURES *const sf = &cpi->sf;
const int has_second_rf = has_second_ref(mi);
const int inter_mode_mask = sf->inter_mode_mask[bsize];
MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
vp9_zero(*bsi);
bsi->segment_rd = best_rd;
bsi->ref_mv[0] = best_ref_mv;
bsi->ref_mv[1] = second_best_ref_mv;
bsi->mvp.as_int = best_ref_mv->as_int;
bsi->mvthresh = mvthresh;
for (i = 0; i < 4; i++) bsi->modes[i] = ZEROMV;
memcpy(t_above, pd->above_context, sizeof(t_above));
memcpy(t_left, pd->left_context, sizeof(t_left));
// 64 makes this threshold really big effectively
// making it so that we very rarely check mvs on
// segments. setting this to 1 would make mv thresh
// roughly equal to what it is for macroblocks
label_mv_thresh = 1 * bsi->mvthresh / label_count;
// Segmentation method overheads
for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
// TODO(jingning,rbultje): rewrite the rate-distortion optimization
// loop for 4x4/4x8/8x4 block coding. to be replaced with new rd loop
int_mv mode_mv[MB_MODE_COUNT][2];
int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
PREDICTION_MODE mode_selected = ZEROMV;
int64_t best_rd = INT64_MAX;
const int i = idy * 2 + idx;
int ref;
for (ref = 0; ref < 1 + has_second_rf; ++ref) {
const MV_REFERENCE_FRAME frame = mi->ref_frame[ref];
frame_mv[ZEROMV][frame].as_int = 0;
vp9_append_sub8x8_mvs_for_idx(
cm, xd, i, ref, mi_row, mi_col, &frame_mv[NEARESTMV][frame],
&frame_mv[NEARMV][frame], mbmi_ext->mode_context);
}
// search for the best motion vector on this segment
for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
const struct buf_2d orig_src = x->plane[0].src;
struct buf_2d orig_pre[2];
mode_idx = INTER_OFFSET(this_mode);
bsi->rdstat[i][mode_idx].brdcost = INT64_MAX;
if (!(inter_mode_mask & (1 << this_mode))) continue;
if (!check_best_zero_mv(cpi, mbmi_ext->mode_context, frame_mv,
this_mode, mi->ref_frame))
continue;
memcpy(orig_pre, pd->pre, sizeof(orig_pre));
memcpy(bsi->rdstat[i][mode_idx].ta, t_above,
sizeof(bsi->rdstat[i][mode_idx].ta));
memcpy(bsi->rdstat[i][mode_idx].tl, t_left,
sizeof(bsi->rdstat[i][mode_idx].tl));
// motion search for newmv (single predictor case only)
if (!has_second_rf && this_mode == NEWMV &&
seg_mvs[i][mi->ref_frame[0]].as_int == INVALID_MV) {
MV *const new_mv = &mode_mv[NEWMV][0].as_mv;
int step_param = 0;
uint32_t bestsme = UINT_MAX;
int sadpb = x->sadperbit4;
MV mvp_full;
int max_mv;
int cost_list[5];
const MvLimits tmp_mv_limits = x->mv_limits;
/* Is the best so far sufficiently good that we cant justify doing
* and new motion search. */
if (best_rd < label_mv_thresh) break;
if (cpi->oxcf.mode != BEST) {
// use previous block's result as next block's MV predictor.
if (i > 0) {
bsi->mvp.as_int = mi->bmi[i - 1].as_mv[0].as_int;
if (i == 2) bsi->mvp.as_int = mi->bmi[i - 2].as_mv[0].as_int;
}
}
if (i == 0)
max_mv = x->max_mv_context[mi->ref_frame[0]];
else
max_mv =
VPXMAX(abs(bsi->mvp.as_mv.row), abs(bsi->mvp.as_mv.col)) >> 3;
if (sf->mv.auto_mv_step_size && cm->show_frame) {
// Take wtd average of the step_params based on the last frame's
// max mv magnitude and the best ref mvs of the current block for
// the given reference.
step_param =
(vp9_init_search_range(max_mv) + cpi->mv_step_param) / 2;
} else {
step_param = cpi->mv_step_param;
}
mvp_full.row = bsi->mvp.as_mv.row >> 3;
mvp_full.col = bsi->mvp.as_mv.col >> 3;
if (sf->adaptive_motion_search) {
mvp_full.row = x->pred_mv[mi->ref_frame[0]].row >> 3;
mvp_full.col = x->pred_mv[mi->ref_frame[0]].col >> 3;
step_param = VPXMAX(step_param, 8);
}
// adjust src pointer for this block
mi_buf_shift(x, i);
vp9_set_mv_search_range(&x->mv_limits, &bsi->ref_mv[0]->as_mv);
bestsme = vp9_full_pixel_search(
cpi, x, bsize, &mvp_full, step_param, cpi->sf.mv.search_method,
sadpb,
sf->mv.subpel_search_method != SUBPEL_TREE ? cost_list : NULL,
&bsi->ref_mv[0]->as_mv, new_mv, INT_MAX, 1);
x->mv_limits = tmp_mv_limits;
if (bestsme < UINT_MAX) {
uint32_t distortion;
cpi->find_fractional_mv_step(
x, new_mv, &bsi->ref_mv[0]->as_mv, cm->allow_high_precision_mv,
x->errorperbit, &cpi->fn_ptr[bsize], sf->mv.subpel_force_stop,
sf->mv.subpel_iters_per_step, cond_cost_list(cpi, cost_list),
x->nmvjointcost, x->mvcost, &distortion,
&x->pred_sse[mi->ref_frame[0]], NULL, 0, 0);
// save motion search result for use in compound prediction
seg_mvs[i][mi->ref_frame[0]].as_mv = *new_mv;
}
if (sf->adaptive_motion_search)
x->pred_mv[mi->ref_frame[0]] = *new_mv;
// restore src pointers
mi_buf_restore(x, orig_src, orig_pre);
}
if (has_second_rf) {
if (seg_mvs[i][mi->ref_frame[1]].as_int == INVALID_MV ||
seg_mvs[i][mi->ref_frame[0]].as_int == INVALID_MV)
continue;
}
if (has_second_rf && this_mode == NEWMV &&
mi->interp_filter == EIGHTTAP) {
// adjust src pointers
mi_buf_shift(x, i);
if (sf->comp_inter_joint_search_thresh <= bsize) {
int rate_mv;
joint_motion_search(cpi, x, bsize, frame_mv[this_mode], mi_row,
mi_col, seg_mvs[i], &rate_mv);
seg_mvs[i][mi->ref_frame[0]].as_int =
frame_mv[this_mode][mi->ref_frame[0]].as_int;
seg_mvs[i][mi->ref_frame[1]].as_int =
frame_mv[this_mode][mi->ref_frame[1]].as_int;
}
// restore src pointers
mi_buf_restore(x, orig_src, orig_pre);
}
bsi->rdstat[i][mode_idx].brate = set_and_cost_bmi_mvs(
cpi, x, xd, i, this_mode, mode_mv[this_mode], frame_mv, seg_mvs[i],
bsi->ref_mv, x->nmvjointcost, x->mvcost);
for (ref = 0; ref < 1 + has_second_rf; ++ref) {
bsi->rdstat[i][mode_idx].mvs[ref].as_int =
mode_mv[this_mode][ref].as_int;
if (num_4x4_blocks_wide > 1)
bsi->rdstat[i + 1][mode_idx].mvs[ref].as_int =
mode_mv[this_mode][ref].as_int;
if (num_4x4_blocks_high > 1)
bsi->rdstat[i + 2][mode_idx].mvs[ref].as_int =
mode_mv[this_mode][ref].as_int;
}
// Trap vectors that reach beyond the UMV borders
if (mv_check_bounds(&x->mv_limits, &mode_mv[this_mode][0].as_mv) ||
(has_second_rf &&
mv_check_bounds(&x->mv_limits, &mode_mv[this_mode][1].as_mv)))
continue;
if (filter_idx > 0) {
BEST_SEG_INFO *ref_bsi = bsi_buf;
subpelmv = 0;
have_ref = 1;
for (ref = 0; ref < 1 + has_second_rf; ++ref) {
subpelmv |= mv_has_subpel(&mode_mv[this_mode][ref].as_mv);
have_ref &= mode_mv[this_mode][ref].as_int ==
ref_bsi->rdstat[i][mode_idx].mvs[ref].as_int;
}
if (filter_idx > 1 && !subpelmv && !have_ref) {
ref_bsi = bsi_buf + 1;
have_ref = 1;
for (ref = 0; ref < 1 + has_second_rf; ++ref)
have_ref &= mode_mv[this_mode][ref].as_int ==
ref_bsi->rdstat[i][mode_idx].mvs[ref].as_int;
}
if (!subpelmv && have_ref &&
ref_bsi->rdstat[i][mode_idx].brdcost < INT64_MAX) {
memcpy(&bsi->rdstat[i][mode_idx], &ref_bsi->rdstat[i][mode_idx],
sizeof(SEG_RDSTAT));
if (num_4x4_blocks_wide > 1)
bsi->rdstat[i + 1][mode_idx].eobs =
ref_bsi->rdstat[i + 1][mode_idx].eobs;
if (num_4x4_blocks_high > 1)
bsi->rdstat[i + 2][mode_idx].eobs =
ref_bsi->rdstat[i + 2][mode_idx].eobs;
if (bsi->rdstat[i][mode_idx].brdcost < best_rd) {
mode_selected = this_mode;
best_rd = bsi->rdstat[i][mode_idx].brdcost;
}
continue;
}
}
bsi->rdstat[i][mode_idx].brdcost = encode_inter_mb_segment(
cpi, x, bsi->segment_rd - this_segment_rd, i,
&bsi->rdstat[i][mode_idx].byrate, &bsi->rdstat[i][mode_idx].bdist,
&bsi->rdstat[i][mode_idx].bsse, bsi->rdstat[i][mode_idx].ta,
bsi->rdstat[i][mode_idx].tl, mi_row, mi_col);
if (bsi->rdstat[i][mode_idx].brdcost < INT64_MAX) {
bsi->rdstat[i][mode_idx].brdcost +=
RDCOST(x->rdmult, x->rddiv, bsi->rdstat[i][mode_idx].brate, 0);
bsi->rdstat[i][mode_idx].brate += bsi->rdstat[i][mode_idx].byrate;
bsi->rdstat[i][mode_idx].eobs = p->eobs[i];
if (num_4x4_blocks_wide > 1)
bsi->rdstat[i + 1][mode_idx].eobs = p->eobs[i + 1];
if (num_4x4_blocks_high > 1)
bsi->rdstat[i + 2][mode_idx].eobs = p->eobs[i + 2];
}
if (bsi->rdstat[i][mode_idx].brdcost < best_rd) {
mode_selected = this_mode;
best_rd = bsi->rdstat[i][mode_idx].brdcost;
}
} /*for each 4x4 mode*/
if (best_rd == INT64_MAX) {
int iy, midx;
for (iy = i + 1; iy < 4; ++iy)
for (midx = 0; midx < INTER_MODES; ++midx)
bsi->rdstat[iy][midx].brdcost = INT64_MAX;
bsi->segment_rd = INT64_MAX;
return INT64_MAX;
}
mode_idx = INTER_OFFSET(mode_selected);
memcpy(t_above, bsi->rdstat[i][mode_idx].ta, sizeof(t_above));
memcpy(t_left, bsi->rdstat[i][mode_idx].tl, sizeof(t_left));
set_and_cost_bmi_mvs(cpi, x, xd, i, mode_selected, mode_mv[mode_selected],
frame_mv, seg_mvs[i], bsi->ref_mv, x->nmvjointcost,
x->mvcost);
br += bsi->rdstat[i][mode_idx].brate;
bd += bsi->rdstat[i][mode_idx].bdist;
block_sse += bsi->rdstat[i][mode_idx].bsse;
segmentyrate += bsi->rdstat[i][mode_idx].byrate;
this_segment_rd += bsi->rdstat[i][mode_idx].brdcost;
if (this_segment_rd > bsi->segment_rd) {
int iy, midx;
for (iy = i + 1; iy < 4; ++iy)
for (midx = 0; midx < INTER_MODES; ++midx)
bsi->rdstat[iy][midx].brdcost = INT64_MAX;
bsi->segment_rd = INT64_MAX;
return INT64_MAX;
}
}
} /* for each label */
bsi->r = br;
bsi->d = bd;
bsi->segment_yrate = segmentyrate;
bsi->segment_rd = this_segment_rd;
bsi->sse = block_sse;
// update the coding decisions
for (k = 0; k < 4; ++k) bsi->modes[k] = mi->bmi[k].as_mode;
if (bsi->segment_rd > best_rd) return INT64_MAX;
/* set it to the best */
for (i = 0; i < 4; i++) {
mode_idx = INTER_OFF