An error occurred while loading the file. Please try again.
-
Deb Mukherjee authored
Allows for swtiching/setting interpolation filters at the MB level. A frame level flag indicates whether to use a specifc filter for the entire frame or to signal the interpolation filter for each MB. When switchable filters are used, the encoder chooses between 8-tap and 8-tap sharp filters. The code currently has options to explore other variations as well, which will be cleaned up subsequently. One issue with the framework is that encoding is slow. I tried to do some tricks to speed things up but it is still slow. Decoding speed should not be affected since the number of filter taps remain unchanged. With the current version, we are up 0.5% on derf on average but some videos city/mobile improve by close to 4 and 2% respectively. If we did a full-search by turning the SEARCH_BEST_FILTER flag on, the results are somewhat better. The framework can be combined with filtered prediction, and I seek feedback regarding that. Rebased. Change-Id: I8f632cb2c111e76284140a2bd480945d6d42b77a
52597441
/* * 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 "./vpx_config.h"#include "./vp9_rtcd.h"#include "vp9/encoder/vp9_encodeframe.h"#include "vp9/encoder/vp9_encodemb.h"#include "vp9/encoder/vp9_encodemv.h"#include "vp9/common/vp9_common.h"#include "vp9/encoder/vp9_onyx_int.h"#include "vp9/common/vp9_extend.h"#include "vp9/common/vp9_entropy.h"#include "vp9/common/vp9_entropymode.h"#include "vp9/common/vp9_quant_common.h"#include "vp9/encoder/vp9_segmentation.h"#include "vp9/encoder/vp9_encodeintra.h"#include "vp9/common/vp9_reconinter.h"#include "vp9/encoder/vp9_rdopt.h"#include "vp9/common/vp9_findnearmv.h"#include "vp9/common/vp9_reconintra.h"#include "vp9/common/vp9_seg_common.h"#include "vp9/common/vp9_tile_common.h"#include "vp9/encoder/vp9_tokenize.h"#include "./vp9_rtcd.h"#include <stdio.h>#include <math.h>#include <limits.h>#include "vpx_ports/vpx_timer.h"#include "vp9/common/vp9_pred_common.h"#include "vp9/common/vp9_mvref_common.h"#define DBG_PRNT_SEGMAP 0// #define ENC_DEBUG#ifdef ENC_DEBUGint enc_debug = 0;#endifstatic void encode_superblock(VP9_COMP *cpi, TOKENEXTRA **t, int output_enabled, int mi_row, int mi_col, BLOCK_SIZE_TYPE bsize);static void adjust_act_zbin(VP9_COMP *cpi, MACROBLOCK *x);/* activity_avg must be positive, or flat regions could get a zero weight * (infinite lambda), which confounds analysis. * This also avoids the need for divide by zero checks in * vp9_activity_masking(). */#define VP9_ACTIVITY_AVG_MIN (64)/* This is used as a reference when computing the source variance for the * purposes of activity masking. * Eventually this should be replaced by custom no-reference routines, * which will be faster. */static const uint8_t VP9_VAR_OFFS[16] = {128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128};// Original activity measure from Tim T's code.static unsigned int tt_activity_measure(VP9_COMP *cpi, MACROBLOCK *x) { unsigned int act; unsigned int sse; /* TODO: This could also be done over smaller areas (8x8), but that would7172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140
* require extensive changes elsewhere, as lambda is assumed to be fixed
* over an entire MB in most of the code.
* Another option is to compute four 8x8 variances, and pick a single
* lambda using a non-linear combination (e.g., the smallest, or second
* smallest, etc.).
*/
act = vp9_variance16x16(x->plane[0].src.buf, x->plane[0].src.stride,
VP9_VAR_OFFS, 0, &sse);
act <<= 4;
/* If the region is flat, lower the activity some more. */
if (act < 8 << 12)
act = act < 5 << 12 ? act : 5 << 12;
return act;
}
// Stub for alternative experimental activity measures.
static unsigned int alt_activity_measure(VP9_COMP *cpi, MACROBLOCK *x,
int use_dc_pred) {
return vp9_encode_intra(cpi, x, use_dc_pred);
}
DECLARE_ALIGNED(16, static const uint8_t, vp9_64x64_zeros[64*64]) = {0};
// Measure the activity of the current macroblock
// What we measure here is TBD so abstracted to this function
#define ALT_ACT_MEASURE 1
static unsigned int mb_activity_measure(VP9_COMP *cpi, MACROBLOCK *x,
int mb_row, int mb_col) {
unsigned int mb_activity;
if (ALT_ACT_MEASURE) {
int use_dc_pred = (mb_col || mb_row) && (!mb_col || !mb_row);
// Or use and alternative.
mb_activity = alt_activity_measure(cpi, x, use_dc_pred);
} else {
// Original activity measure from Tim T's code.
mb_activity = tt_activity_measure(cpi, x);
}
if (mb_activity < VP9_ACTIVITY_AVG_MIN)
mb_activity = VP9_ACTIVITY_AVG_MIN;
return mb_activity;
}
// Calculate an "average" mb activity value for the frame
#define ACT_MEDIAN 0
static void calc_av_activity(VP9_COMP *cpi, int64_t activity_sum) {
#if ACT_MEDIAN
// Find median: Simple n^2 algorithm for experimentation
{
unsigned int median;
unsigned int i, j;
unsigned int *sortlist;
unsigned int tmp;
// Create a list to sort to
CHECK_MEM_ERROR(&cpi->common, sortlist, vpx_calloc(sizeof(unsigned int),
cpi->common.MBs));
// Copy map to sort list
vpx_memcpy(sortlist, cpi->mb_activity_map,
sizeof(unsigned int) * cpi->common.MBs);
// Ripple each value down to its correct position
for (i = 1; i < cpi->common.MBs; i ++) {
for (j = i; j > 0; j --) {
if (sortlist[j] < sortlist[j - 1]) {
141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210
// Swap values
tmp = sortlist[j - 1];
sortlist[j - 1] = sortlist[j];
sortlist[j] = tmp;
} else
break;
}
}
// Even number MBs so estimate median as mean of two either side.
median = (1 + sortlist[cpi->common.MBs >> 1] +
sortlist[(cpi->common.MBs >> 1) + 1]) >> 1;
cpi->activity_avg = median;
vpx_free(sortlist);
}
#else
// Simple mean for now
cpi->activity_avg = (unsigned int) (activity_sum / cpi->common.MBs);
#endif
if (cpi->activity_avg < VP9_ACTIVITY_AVG_MIN)
cpi->activity_avg = VP9_ACTIVITY_AVG_MIN;
// Experimental code: return fixed value normalized for several clips
if (ALT_ACT_MEASURE)
cpi->activity_avg = 100000;
}
#define USE_ACT_INDEX 0
#define OUTPUT_NORM_ACT_STATS 0
#if USE_ACT_INDEX
// Calculate an activity index for each mb
static void calc_activity_index(VP9_COMP *cpi, MACROBLOCK *x) {
VP9_COMMON *const cm = &cpi->common;
int mb_row, mb_col;
int64_t act;
int64_t a;
int64_t b;
#if OUTPUT_NORM_ACT_STATS
FILE *f = fopen("norm_act.stt", "a");
fprintf(f, "\n%12d\n", cpi->activity_avg);
#endif
// Reset pointers to start of activity map
x->mb_activity_ptr = cpi->mb_activity_map;
// Calculate normalized mb activity number.
for (mb_row = 0; mb_row < cm->mb_rows; mb_row++) {
// for each macroblock col in image
for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
// Read activity from the map
act = *(x->mb_activity_ptr);
// Calculate a normalized activity number
a = act + 4 * cpi->activity_avg;
b = 4 * act + cpi->activity_avg;
if (b >= a)
*(x->activity_ptr) = (int)((b + (a >> 1)) / a) - 1;
else
*(x->activity_ptr) = 1 - (int)((a + (b >> 1)) / b);
#if OUTPUT_NORM_ACT_STATS
fprintf(f, " %6d", *(x->mb_activity_ptr));
#endif
211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280
// Increment activity map pointers
x->mb_activity_ptr++;
}
#if OUTPUT_NORM_ACT_STATS
fprintf(f, "\n");
#endif
}
#if OUTPUT_NORM_ACT_STATS
fclose(f);
#endif
}
#endif
// Loop through all MBs. Note activity of each, average activity and
// calculate a normalized activity for each
static void build_activity_map(VP9_COMP *cpi) {
MACROBLOCK * const x = &cpi->mb;
MACROBLOCKD *xd = &x->e_mbd;
VP9_COMMON * const cm = &cpi->common;
#if ALT_ACT_MEASURE
YV12_BUFFER_CONFIG *new_yv12 = &cm->yv12_fb[cm->new_fb_idx];
int recon_yoffset;
int recon_y_stride = new_yv12->y_stride;
#endif
int mb_row, mb_col;
unsigned int mb_activity;
int64_t activity_sum = 0;
x->mb_activity_ptr = cpi->mb_activity_map;
// for each macroblock row in image
for (mb_row = 0; mb_row < cm->mb_rows; mb_row++) {
#if ALT_ACT_MEASURE
// reset above block coeffs
xd->up_available = (mb_row != 0);
recon_yoffset = (mb_row * recon_y_stride * 16);
#endif
// for each macroblock col in image
for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
#if ALT_ACT_MEASURE
xd->plane[0].dst.buf = new_yv12->y_buffer + recon_yoffset;
xd->left_available = (mb_col != 0);
recon_yoffset += 16;
#endif
// measure activity
mb_activity = mb_activity_measure(cpi, x, mb_row, mb_col);
// Keep frame sum
activity_sum += mb_activity;
// Store MB level activity details.
*x->mb_activity_ptr = mb_activity;
// Increment activity map pointer
x->mb_activity_ptr++;
// adjust to the next column of source macroblocks
x->plane[0].src.buf += 16;
}
// adjust to the next row of mbs
x->plane[0].src.buf += 16 * x->plane[0].src.stride - 16 * cm->mb_cols;
}
281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350
// Calculate an "average" MB activity
calc_av_activity(cpi, activity_sum);
#if USE_ACT_INDEX
// Calculate an activity index number of each mb
calc_activity_index(cpi, x);
#endif
}
// Macroblock activity masking
void vp9_activity_masking(VP9_COMP *cpi, MACROBLOCK *x) {
#if USE_ACT_INDEX
x->rdmult += *(x->mb_activity_ptr) * (x->rdmult >> 2);
x->errorperbit = x->rdmult * 100 / (110 * x->rddiv);
x->errorperbit += (x->errorperbit == 0);
#else
int64_t a;
int64_t b;
int64_t act = *(x->mb_activity_ptr);
// Apply the masking to the RD multiplier.
a = act + (2 * cpi->activity_avg);
b = (2 * act) + cpi->activity_avg;
x->rdmult = (unsigned int) (((int64_t) x->rdmult * b + (a >> 1)) / a);
x->errorperbit = x->rdmult * 100 / (110 * x->rddiv);
x->errorperbit += (x->errorperbit == 0);
#endif
// Activity based Zbin adjustment
adjust_act_zbin(cpi, x);
}
static void update_state(VP9_COMP *cpi, PICK_MODE_CONTEXT *ctx,
BLOCK_SIZE_TYPE bsize, int output_enabled) {
int i, x_idx, y;
MACROBLOCK * const x = &cpi->mb;
MACROBLOCKD * const xd = &x->e_mbd;
MODE_INFO *mi = &ctx->mic;
MB_MODE_INFO * const mbmi = &xd->mode_info_context->mbmi;
int mb_mode_index = ctx->best_mode_index;
const int mis = cpi->common.mode_info_stride;
const int bh = 1 << mi_height_log2(bsize), bw = 1 << mi_width_log2(bsize);
assert(mi->mbmi.mode < MB_MODE_COUNT);
assert(mb_mode_index < MAX_MODES);
assert(mi->mbmi.ref_frame[0] < MAX_REF_FRAMES);
assert(mi->mbmi.ref_frame[1] < MAX_REF_FRAMES);
assert(mi->mbmi.sb_type == bsize);
// Restore the coding context of the MB to that that was in place
// when the mode was picked for it
for (y = 0; y < bh; y++) {
for (x_idx = 0; x_idx < bw; x_idx++) {
if ((xd->mb_to_right_edge >> (3 + LOG2_MI_SIZE)) + bw > x_idx
&& (xd->mb_to_bottom_edge >> (3 + LOG2_MI_SIZE)) + bh > y) {
MODE_INFO *mi_addr = xd->mode_info_context + x_idx + y * mis;
*mi_addr = *mi;
}
}
}
// FIXME(rbultje) I'm pretty sure this should go to the end of this block
// (i.e. after the output_enabled)
if (bsize < BLOCK_SIZE_SB32X32) {
if (bsize < BLOCK_SIZE_MB16X16)
ctx->txfm_rd_diff[ALLOW_16X16] = ctx->txfm_rd_diff[ALLOW_8X8];
ctx->txfm_rd_diff[ALLOW_32X32] = ctx->txfm_rd_diff[ALLOW_16X16];
351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420
}
if (mbmi->ref_frame[0] != INTRA_FRAME && mbmi->sb_type < BLOCK_SIZE_SB8X8) {
*x->partition_info = ctx->partition_info;
mbmi->mv[0].as_int = mi->bmi[3].as_mv[0].as_int;
mbmi->mv[1].as_int = mi->bmi[3].as_mv[1].as_int;
}
x->skip = ctx->skip;
if (!output_enabled)
return;
if (!vp9_segfeature_active(&xd->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
for (i = 0; i < NB_TXFM_MODES; i++) {
cpi->rd_tx_select_diff[i] += ctx->txfm_rd_diff[i];
}
}
if (cpi->common.frame_type == KEY_FRAME) {
// Restore the coding modes to that held in the coding context
// if (mb_mode == I4X4_PRED)
// for (i = 0; i < 16; i++)
// {
// xd->block[i].bmi.as_mode =
// xd->mode_info_context->bmi[i].as_mode;
// assert(xd->mode_info_context->bmi[i].as_mode < MB_MODE_COUNT);
// }
#if CONFIG_INTERNAL_STATS
static const int kf_mode_index[] = {
THR_DC /*DC_PRED*/,
THR_V_PRED /*V_PRED*/,
THR_H_PRED /*H_PRED*/,
THR_D45_PRED /*D45_PRED*/,
THR_D135_PRED /*D135_PRED*/,
THR_D117_PRED /*D117_PRED*/,
THR_D153_PRED /*D153_PRED*/,
THR_D27_PRED /*D27_PRED*/,
THR_D63_PRED /*D63_PRED*/,
THR_TM /*TM_PRED*/,
THR_B_PRED /*I4X4_PRED*/,
};
cpi->mode_chosen_counts[kf_mode_index[mi->mbmi.mode]]++;
#endif
} else {
// Note how often each mode chosen as best
cpi->mode_chosen_counts[mb_mode_index]++;
if (mbmi->ref_frame[0] != INTRA_FRAME
&& (mbmi->sb_type < BLOCK_SIZE_SB8X8 || mbmi->mode == NEWMV)) {
int_mv best_mv, best_second_mv;
const MV_REFERENCE_FRAME rf1 = mbmi->ref_frame[0];
const MV_REFERENCE_FRAME rf2 = mbmi->ref_frame[1];
best_mv.as_int = ctx->best_ref_mv.as_int;
best_second_mv.as_int = ctx->second_best_ref_mv.as_int;
if (mbmi->mode == NEWMV) {
best_mv.as_int = mbmi->ref_mvs[rf1][0].as_int;
best_second_mv.as_int = mbmi->ref_mvs[rf2][0].as_int;
}
mbmi->best_mv.as_int = best_mv.as_int;
mbmi->best_second_mv.as_int = best_second_mv.as_int;
vp9_update_nmv_count(cpi, x, &best_mv, &best_second_mv);
}
if (bsize > BLOCK_SIZE_SB8X8 && mbmi->mode == NEWMV) {
int i, j;
for (j = 0; j < bh; ++j)
for (i = 0; i < bw; ++i)
if ((xd->mb_to_right_edge >> (3 + LOG2_MI_SIZE)) + bw > i
&& (xd->mb_to_bottom_edge >> (3 + LOG2_MI_SIZE)) + bh > j)
xd->mode_info_context[mis * j + i].mbmi = *mbmi;
}
421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490
if (cpi->common.mcomp_filter_type == SWITCHABLE
&& is_inter_mode(mbmi->mode)) {
++cpi->common.fc.switchable_interp_count[
vp9_get_pred_context_switchable_interp(&cpi->common, xd)]
[vp9_switchable_interp_map[mbmi->interp_filter]];
}
cpi->rd_comp_pred_diff[SINGLE_PREDICTION_ONLY] += ctx->single_pred_diff;
cpi->rd_comp_pred_diff[COMP_PREDICTION_ONLY] += ctx->comp_pred_diff;
cpi->rd_comp_pred_diff[HYBRID_PREDICTION] += ctx->hybrid_pred_diff;
for (i = 0; i <= VP9_SWITCHABLE_FILTERS; i++) {
cpi->rd_filter_diff[i] += ctx->best_filter_diff[i];
}
}
}
void vp9_setup_src_planes(MACROBLOCK *x, const YV12_BUFFER_CONFIG *src,
int mb_row, int mb_col) {
uint8_t *buffers[4] = {src->y_buffer, src->u_buffer, src->v_buffer, src
->alpha_buffer};
int strides[4] = {src->y_stride, src->uv_stride, src->uv_stride, src
->alpha_stride};
int i;
for (i = 0; i < MAX_MB_PLANE; i++) {
setup_pred_plane(&x->plane[i].src, buffers[i], strides[i], mb_row, mb_col,
NULL, x->e_mbd.plane[i].subsampling_x,
x->e_mbd.plane[i].subsampling_y);
}
}
static void set_offsets(VP9_COMP *cpi, int mi_row, int mi_col,
BLOCK_SIZE_TYPE bsize) {
MACROBLOCK * const x = &cpi->mb;
VP9_COMMON * const cm = &cpi->common;
MACROBLOCKD * const xd = &x->e_mbd;
MB_MODE_INFO *mbmi;
const int dst_fb_idx = cm->new_fb_idx;
const int idx_str = xd->mode_info_stride * mi_row + mi_col;
const int bw = 1 << mi_width_log2(bsize), bh = 1 << mi_height_log2(bsize);
const int mb_row = mi_row >> 1;
const int mb_col = mi_col >> 1;
const int idx_map = mb_row * cm->mb_cols + mb_col;
int i;
// entropy context structures
for (i = 0; i < MAX_MB_PLANE; i++) {
xd->plane[i].above_context = cm->above_context[i]
+ (mi_col * 2 >> xd->plane[i].subsampling_x);
xd->plane[i].left_context = cm->left_context[i]
+ (((mi_row * 2) & 15) >> xd->plane[i].subsampling_y);
}
// partition contexts
set_partition_seg_context(cm, xd, mi_row, mi_col);
// Activity map pointer
x->mb_activity_ptr = &cpi->mb_activity_map[idx_map];
x->active_ptr = cpi->active_map + idx_map;
/* pointers to mode info contexts */
x->partition_info = x->pi + idx_str;
xd->mode_info_context = cm->mi + idx_str;
mbmi = &xd->mode_info_context->mbmi;
// Special case: if prev_mi is NULL, the previous mode info context
// cannot be used.
xd->prev_mode_info_context = cm->prev_mi ? cm->prev_mi + idx_str : NULL;
491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560
// Set up destination pointers
setup_dst_planes(xd, &cm->yv12_fb[dst_fb_idx], mi_row, mi_col);
/* Set up limit values for MV components to prevent them from
* extending beyond the UMV borders assuming 16x16 block size */
x->mv_row_min = -((mi_row * MI_SIZE)+ VP9BORDERINPIXELS - VP9_INTERP_EXTEND);
x->mv_col_min = -((mi_col * MI_SIZE)+ VP9BORDERINPIXELS - VP9_INTERP_EXTEND);
x->mv_row_max = ((cm->mi_rows - mi_row) * MI_SIZE
+ (VP9BORDERINPIXELS - MI_SIZE * bh - VP9_INTERP_EXTEND));
x->mv_col_max = ((cm->mi_cols - mi_col) * MI_SIZE
+ (VP9BORDERINPIXELS - MI_SIZE * bw - VP9_INTERP_EXTEND));
// Set up distance of MB to edge of frame in 1/8th pel units
assert(!(mi_col & (bw - 1)) && !(mi_row & (bh - 1)));
set_mi_row_col(cm, xd, mi_row, bh, mi_col, bw);
/* set up source buffers */
vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);
/* R/D setup */
x->rddiv = cpi->RDDIV;
x->rdmult = cpi->RDMULT;
/* segment ID */
if (xd->seg.enabled) {
uint8_t *map = xd->seg.update_map ? cpi->segmentation_map
: cm->last_frame_seg_map;
mbmi->segment_id = vp9_get_segment_id(cm, map, bsize, mi_row, mi_col);
vp9_mb_init_quantizer(cpi, x);
if (xd->seg.enabled && cpi->seg0_cnt > 0
&& !vp9_segfeature_active(&xd->seg, 0, SEG_LVL_REF_FRAME)
&& vp9_segfeature_active(&xd->seg, 1, SEG_LVL_REF_FRAME)) {
cpi->seg0_progress = (cpi->seg0_idx << 16) / cpi->seg0_cnt;
} else {
const int y = mb_row & ~3;
const int x = mb_col & ~3;
const int p16 = ((mb_row & 1) << 1) + (mb_col & 1);
const int p32 = ((mb_row & 2) << 2) + ((mb_col & 2) << 1);
const int tile_progress = cm->cur_tile_mi_col_start * cm->mb_rows >> 1;
const int mb_cols = (cm->cur_tile_mi_col_end - cm->cur_tile_mi_col_start)
>> 1;
cpi->seg0_progress = ((y * mb_cols + x * 4 + p32 + p16 + tile_progress)
<< 16) / cm->MBs;
}
} else {
mbmi->segment_id = 0;
}
}
static void pick_sb_modes(VP9_COMP *cpi, int mi_row, int mi_col,
int *totalrate, int64_t *totaldist,
BLOCK_SIZE_TYPE bsize, PICK_MODE_CONTEXT *ctx,
int64_t best_rd) {
VP9_COMMON *const cm = &cpi->common;
MACROBLOCK *const x = &cpi->mb;
MACROBLOCKD *const xd = &x->e_mbd;
x->rd_search = 1;
if (bsize < BLOCK_SIZE_SB8X8)
if (xd->ab_index != 0)
return;
set_offsets(cpi, mi_row, mi_col, bsize);
xd->mode_info_context->mbmi.sb_type = bsize;
if (cpi->oxcf.tuning == VP8_TUNE_SSIM)
vp9_activity_masking(cpi, x);
561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630
// Find best coding mode & reconstruct the MB so it is available
// as a predictor for MBs that follow in the SB
if (cm->frame_type == KEY_FRAME)
vp9_rd_pick_intra_mode_sb(cpi, x, totalrate, totaldist, bsize, ctx,
best_rd);
else
vp9_rd_pick_inter_mode_sb(cpi, x, mi_row, mi_col, totalrate, totaldist,
bsize, ctx, best_rd);
}
static void update_stats(VP9_COMP *cpi, int mi_row, int mi_col) {
VP9_COMMON * const cm = &cpi->common;
MACROBLOCK * const x = &cpi->mb;
MACROBLOCKD * const xd = &x->e_mbd;
MODE_INFO *mi = xd->mode_info_context;
MB_MODE_INFO * const mbmi = &mi->mbmi;
if (cm->frame_type != KEY_FRAME) {
const int seg_ref_active = vp9_segfeature_active(&xd->seg, mbmi->segment_id,
SEG_LVL_REF_FRAME);
if (!seg_ref_active)
cpi->intra_inter_count[vp9_get_pred_context_intra_inter(cm, xd)][mbmi
->ref_frame[0] > INTRA_FRAME]++;
// If the segment reference feature is enabled we have only a single
// reference frame allowed for the segment so exclude it from
// the reference frame counts used to work out probabilities.
if ((mbmi->ref_frame[0] > INTRA_FRAME) && !seg_ref_active) {
if (cm->comp_pred_mode == HYBRID_PREDICTION)
cpi->comp_inter_count[vp9_get_pred_context_comp_inter_inter(cm, xd)]
[mbmi->ref_frame[1] > INTRA_FRAME]++;
if (mbmi->ref_frame[1] > INTRA_FRAME) {
cpi->comp_ref_count[vp9_get_pred_context_comp_ref_p(cm, xd)][mbmi
->ref_frame[0] == GOLDEN_FRAME]++;
} else {
cpi->single_ref_count[vp9_get_pred_context_single_ref_p1(cm, xd)]
[0][mbmi->ref_frame[0] != LAST_FRAME]++;
if (mbmi->ref_frame[0] != LAST_FRAME)
cpi->single_ref_count[vp9_get_pred_context_single_ref_p2(cm, xd)][1]
[mbmi->ref_frame[0] != GOLDEN_FRAME]++;
}
}
// Count of last ref frame 0,0 usage
if ((mbmi->mode == ZEROMV) && (mbmi->ref_frame[0] == LAST_FRAME))
cpi->inter_zz_count++;
}
}
// TODO(jingning): the variables used here are little complicated. need further
// refactoring on organizing the the temporary buffers, when recursive
// partition down to 4x4 block size is enabled.
static PICK_MODE_CONTEXT *get_block_context(MACROBLOCK *x,
BLOCK_SIZE_TYPE bsize) {
MACROBLOCKD * const xd = &x->e_mbd;
switch (bsize) {
case BLOCK_SIZE_SB64X64:
return &x->sb64_context;
case BLOCK_SIZE_SB64X32:
return &x->sb64x32_context[xd->sb_index];
case BLOCK_SIZE_SB32X64:
return &x->sb32x64_context[xd->sb_index];
case BLOCK_SIZE_SB32X32:
return &x->sb32_context[xd->sb_index];
case BLOCK_SIZE_SB32X16:
return &x->sb32x16_context[xd->sb_index][xd->mb_index];
case BLOCK_SIZE_SB16X32:
631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700
return &x->sb16x32_context[xd->sb_index][xd->mb_index];
case BLOCK_SIZE_MB16X16:
return &x->mb_context[xd->sb_index][xd->mb_index];
case BLOCK_SIZE_SB16X8:
return &x->sb16x8_context[xd->sb_index][xd->mb_index][xd->b_index];
case BLOCK_SIZE_SB8X16:
return &x->sb8x16_context[xd->sb_index][xd->mb_index][xd->b_index];
case BLOCK_SIZE_SB8X8:
return &x->sb8x8_context[xd->sb_index][xd->mb_index][xd->b_index];
case BLOCK_SIZE_SB8X4:
return &x->sb8x4_context[xd->sb_index][xd->mb_index][xd->b_index];
case BLOCK_SIZE_SB4X8:
return &x->sb4x8_context[xd->sb_index][xd->mb_index][xd->b_index];
case BLOCK_SIZE_AB4X4:
return &x->ab4x4_context[xd->sb_index][xd->mb_index][xd->b_index];
default:
assert(0);
return NULL ;
}
}
static BLOCK_SIZE_TYPE *get_sb_partitioning(MACROBLOCK *x,
BLOCK_SIZE_TYPE bsize) {
MACROBLOCKD *xd = &x->e_mbd;
switch (bsize) {
case BLOCK_SIZE_SB64X64:
return &x->sb64_partitioning;
case BLOCK_SIZE_SB32X32:
return &x->sb_partitioning[xd->sb_index];
case BLOCK_SIZE_MB16X16:
return &x->mb_partitioning[xd->sb_index][xd->mb_index];
case BLOCK_SIZE_SB8X8:
return &x->b_partitioning[xd->sb_index][xd->mb_index][xd->b_index];
default:
assert(0);
return NULL ;
}
}
static void restore_context(VP9_COMP *cpi, int mi_row, int mi_col,
ENTROPY_CONTEXT a[16 * MAX_MB_PLANE],
ENTROPY_CONTEXT l[16 * MAX_MB_PLANE],
PARTITION_CONTEXT sa[8], PARTITION_CONTEXT sl[8],
BLOCK_SIZE_TYPE bsize) {
VP9_COMMON * const cm = &cpi->common;
MACROBLOCK * const x = &cpi->mb;
MACROBLOCKD * const xd = &x->e_mbd;
int p;
int bwl = b_width_log2(bsize), bw = 1 << bwl;
int bhl = b_height_log2(bsize), bh = 1 << bhl;
int mwl = mi_width_log2(bsize), mw = 1 << mwl;
int mhl = mi_height_log2(bsize), mh = 1 << mhl;
for (p = 0; p < MAX_MB_PLANE; p++) {
vpx_memcpy(
cm->above_context[p] + ((mi_col * 2) >> xd->plane[p].subsampling_x),
a + bw * p, sizeof(ENTROPY_CONTEXT) * bw >> xd->plane[p].subsampling_x);
vpx_memcpy(
cm->left_context[p]
+ ((mi_row & MI_MASK)* 2 >> xd->plane[p].subsampling_y),l + bh * p,
sizeof(ENTROPY_CONTEXT) * bh >> xd->plane[p].subsampling_y);
}
vpx_memcpy(cm->above_seg_context + mi_col, sa,
sizeof(PARTITION_CONTEXT) * mw);
vpx_memcpy(cm->left_seg_context + (mi_row & MI_MASK), sl,
sizeof(PARTITION_CONTEXT) * mh);
}
static void save_context(VP9_COMP *cpi, int mi_row, int mi_col,
ENTROPY_CONTEXT a[16 * MAX_MB_PLANE],
ENTROPY_CONTEXT l[16 * MAX_MB_PLANE],
PARTITION_CONTEXT sa[8], PARTITION_CONTEXT sl[8],
701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770
BLOCK_SIZE_TYPE bsize) {
VP9_COMMON * const cm = &cpi->common;
MACROBLOCK * const x = &cpi->mb;
MACROBLOCKD * const xd = &x->e_mbd;
int p;
int bwl = b_width_log2(bsize), bw = 1 << bwl;
int bhl = b_height_log2(bsize), bh = 1 << bhl;
int mwl = mi_width_log2(bsize), mw = 1 << mwl;
int mhl = mi_height_log2(bsize), mh = 1 << mhl;
// buffer the above/left context information of the block in search.
for (p = 0; p < MAX_MB_PLANE; ++p) {
vpx_memcpy(
a + bw * p,
cm->above_context[p] + (mi_col * 2 >> xd->plane[p].subsampling_x),
sizeof(ENTROPY_CONTEXT) * bw >> xd->plane[p].subsampling_x);
vpx_memcpy(
l + bh * p,
cm->left_context[p]
+ ((mi_row & MI_MASK)* 2 >> xd->plane[p].subsampling_y),sizeof(ENTROPY_CONTEXT) * bh >> xd->plane[p].subsampling_y);
}
vpx_memcpy(sa, cm->above_seg_context + mi_col,
sizeof(PARTITION_CONTEXT) * mw);
vpx_memcpy(sl, cm->left_seg_context + (mi_row & MI_MASK),
sizeof(PARTITION_CONTEXT) * mh)
;}
static void encode_b(VP9_COMP *cpi, TOKENEXTRA **tp, int mi_row, int mi_col,
int output_enabled, BLOCK_SIZE_TYPE bsize, int sub_index) {
VP9_COMMON * const cm = &cpi->common;
MACROBLOCK * const x = &cpi->mb;
MACROBLOCKD * const xd = &x->e_mbd;
if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
return;
if (sub_index != -1)
*(get_sb_index(xd, bsize)) = sub_index;
if (bsize < BLOCK_SIZE_SB8X8)
if (xd->ab_index > 0)
return;
set_offsets(cpi, mi_row, mi_col, bsize);
update_state(cpi, get_block_context(x, bsize), bsize, output_enabled);
encode_superblock(cpi, tp, output_enabled, mi_row, mi_col, bsize);
if (output_enabled) {
update_stats(cpi, mi_row, mi_col);
(*tp)->token = EOSB_TOKEN;
(*tp)++;
}
}
static void encode_sb(VP9_COMP *cpi, TOKENEXTRA **tp, int mi_row, int mi_col,
int output_enabled, BLOCK_SIZE_TYPE bsize) {
VP9_COMMON * const cm = &cpi->common;
MACROBLOCK * const x = &cpi->mb;
MACROBLOCKD * const xd = &x->e_mbd;
BLOCK_SIZE_TYPE c1 = BLOCK_SIZE_SB8X8;
const int bsl = b_width_log2(bsize), bs = (1 << bsl) / 4;
int bwl, bhl;
int UNINITIALIZED_IS_SAFE(pl);
if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
return;
c1 = BLOCK_SIZE_AB4X4;
if (bsize >= BLOCK_SIZE_SB8X8) {
set_partition_seg_context(cm, xd, mi_row, mi_col);
771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840
pl = partition_plane_context(xd, bsize);
c1 = *(get_sb_partitioning(x, bsize));
}
bwl = b_width_log2(c1), bhl = b_height_log2(c1);
if (bsl == bwl && bsl == bhl) {
if (output_enabled && bsize >= BLOCK_SIZE_SB8X8)
cpi->partition_count[pl][PARTITION_NONE]++;
encode_b(cpi, tp, mi_row, mi_col, output_enabled, c1, -1);
} else if (bsl == bhl && bsl > bwl) {
if (output_enabled)
cpi->partition_count[pl][PARTITION_VERT]++;
encode_b(cpi, tp, mi_row, mi_col, output_enabled, c1, 0);
encode_b(cpi, tp, mi_row, mi_col + bs, output_enabled, c1, 1);
} else if (bsl == bwl && bsl > bhl) {
if (output_enabled)
cpi->partition_count[pl][PARTITION_HORZ]++;
encode_b(cpi, tp, mi_row, mi_col, output_enabled, c1, 0);
encode_b(cpi, tp, mi_row + bs, mi_col, output_enabled, c1, 1);
} else {
BLOCK_SIZE_TYPE subsize;
int i;
assert(bwl < bsl && bhl < bsl);
subsize = get_subsize(bsize, PARTITION_SPLIT);
if (output_enabled)
cpi->partition_count[pl][PARTITION_SPLIT]++;
for (i = 0; i < 4; i++) {
const int x_idx = i & 1, y_idx = i >> 1;
*(get_sb_index(xd, subsize)) = i;
encode_sb(cpi, tp, mi_row + y_idx * bs, mi_col + x_idx * bs,
output_enabled, subsize);
}
}
if (bsize >= BLOCK_SIZE_SB8X8
&& (bsize == BLOCK_SIZE_SB8X8 || bsl == bwl || bsl == bhl)) {
set_partition_seg_context(cm, xd, mi_row, mi_col);
update_partition_context(xd, c1, bsize);
}
}
static void set_partitioning(VP9_COMP *cpi, MODE_INFO *m,
BLOCK_SIZE_TYPE bsize) {
VP9_COMMON *const cm = &cpi->common;
const int mis = cm->mode_info_stride;
int block_row, block_col;
for (block_row = 0; block_row < 8; ++block_row) {
for (block_col = 0; block_col < 8; ++block_col) {
m[block_row * mis + block_col].mbmi.sb_type = bsize;
}
}
}
static void copy_partitioning(VP9_COMP *cpi, MODE_INFO *m, MODE_INFO *p) {
VP9_COMMON *const cm = &cpi->common;
const int mis = cm->mode_info_stride;
int block_row, block_col;
for (block_row = 0; block_row < 8; ++block_row) {
for (block_col = 0; block_col < 8; ++block_col) {
m[block_row * mis + block_col].mbmi.sb_type =
p[block_row * mis + block_col].mbmi.sb_type;
}
}
}
static void set_block_size(VP9_COMMON * const cm, MODE_INFO *m,
841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910
BLOCK_SIZE_TYPE bsize, int mis, int mi_row,
int mi_col) {
int row, col;
int bwl = b_width_log2(bsize);
int bhl = b_height_log2(bsize);
int bsl = (bwl > bhl ? bwl : bhl);
int bs = (1 << bsl) / 2; //
MODE_INFO *m2 = m + mi_row * mis + mi_col;
for (row = 0; row < bs; row++) {
for (col = 0; col < bs; col++) {
if (mi_row + row >= cm->mi_rows || mi_col + col >= cm->mi_cols)
continue;
m2[row * mis + col].mbmi.sb_type = bsize;
}
}
}
typedef struct {
int64_t sum_square_error;
int64_t sum_error;
int count;
int variance;
} var;
typedef struct {
var none;
var horz[2];
var vert[2];
} partition_variance;
#define VT(TYPE, BLOCKSIZE) \
typedef struct { \
partition_variance vt; \
BLOCKSIZE split[4]; } TYPE;
VT(v8x8, var)
VT(v16x16, v8x8)
VT(v32x32, v16x16)
VT(v64x64, v32x32)
typedef struct {
partition_variance *vt;
var *split[4];
} vt_node;
typedef enum {
V16X16,
V32X32,
V64X64,
} TREE_LEVEL;
static void tree_to_node(void *data, BLOCK_SIZE_TYPE block_size, vt_node *node) {
int i;
switch (block_size) {
case BLOCK_SIZE_SB64X64: {
v64x64 *vt = (v64x64 *) data;
node->vt = &vt->vt;
for (i = 0; i < 4; i++)
node->split[i] = &vt->split[i].vt.none;
break;
}
case BLOCK_SIZE_SB32X32: {
v32x32 *vt = (v32x32 *) data;
node->vt = &vt->vt;
for (i = 0; i < 4; i++)
node->split[i] = &vt->split[i].vt.none;
break;
}
case BLOCK_SIZE_MB16X16: {
911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980
v16x16 *vt = (v16x16 *) data;
node->vt = &vt->vt;
for (i = 0; i < 4; i++)
node->split[i] = &vt->split[i].vt.none;
break;
}
case BLOCK_SIZE_SB8X8: {
v8x8 *vt = (v8x8 *) data;
node->vt = &vt->vt;
for (i = 0; i < 4; i++)
node->split[i] = &vt->split[i];
break;
}
default:
node->vt = 0;
for (i = 0; i < 4; i++)
node->split[i] = 0;
assert(-1);
}
}
// Set variance values given sum square error, sum error, count.
static void fill_variance(var *v, int64_t s2, int64_t s, int c) {
v->sum_square_error = s2;
v->sum_error = s;
v->count = c;
if (c > 0)
v->variance = 256
* (v->sum_square_error - v->sum_error * v->sum_error / v->count)
/ v->count;
else
v->variance = 0;
}
// Combine 2 variance structures by summing the sum_error, sum_square_error,
// and counts and then calculating the new variance.
void sum_2_variances(var *r, var *a, var*b) {
fill_variance(r, a->sum_square_error + b->sum_square_error,
a->sum_error + b->sum_error, a->count + b->count);
}
static void fill_variance_tree(void *data, BLOCK_SIZE_TYPE block_size) {
vt_node node;
tree_to_node(data, block_size, &node);
sum_2_variances(&node.vt->horz[0], node.split[0], node.split[1]);
sum_2_variances(&node.vt->horz[1], node.split[2], node.split[3]);
sum_2_variances(&node.vt->vert[0], node.split[0], node.split[2]);
sum_2_variances(&node.vt->vert[1], node.split[1], node.split[3]);
sum_2_variances(&node.vt->none, &node.vt->vert[0], &node.vt->vert[1]);
}
#if PERFORM_RANDOM_PARTITIONING
static int set_vt_partitioning(VP9_COMP *cpi, void *data, MODE_INFO *m,
BLOCK_SIZE_TYPE block_size, int mi_row,
int mi_col, int mi_size) {
VP9_COMMON * const cm = &cpi->common;
vt_node vt;
const int mis = cm->mode_info_stride;
int64_t threshold = 4 * cpi->common.base_qindex * cpi->common.base_qindex;
tree_to_node(data, block_size, &vt);
// split none is available only if we have more than half a block size
// in width and height inside the visible image
if (mi_col + mi_size < cm->mi_cols && mi_row + mi_size < cm->mi_rows &&
(rand() & 3) < 1) {
set_block_size(cm, m, block_size, mis, mi_row, mi_col);
return 1;
}
981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050
// vertical split is available on all but the bottom border
if (mi_row + mi_size < cm->mi_rows && vt.vt->vert[0].variance < threshold
&& (rand() & 3) < 1) {
set_block_size(cm, m, get_subsize(block_size, PARTITION_VERT), mis, mi_row,
mi_col);
return 1;
}
// horizontal split is available on all but the right border
if (mi_col + mi_size < cm->mi_cols && vt.vt->horz[0].variance < threshold
&& (rand() & 3) < 1) {
set_block_size(cm, m, get_subsize(block_size, PARTITION_HORZ), mis, mi_row,
mi_col);
return 1;
}
return 0;
}
#else
static int set_vt_partitioning(VP9_COMP *cpi, void *data, MODE_INFO *m,
BLOCK_SIZE_TYPE block_size, int mi_row,
int mi_col, int mi_size) {
VP9_COMMON * const cm = &cpi->common;
vt_node vt;
const int mis = cm->mode_info_stride;
int64_t threshold = 50 * cpi->common.base_qindex;
tree_to_node(data, block_size, &vt);
// split none is available only if we have more than half a block size
// in width and height inside the visible image
if (mi_col + mi_size < cm->mi_cols && mi_row + mi_size < cm->mi_rows
&& vt.vt->none.variance < threshold) {
set_block_size(cm, m, block_size, mis, mi_row, mi_col);
return 1;
}
// vertical split is available on all but the bottom border
if (mi_row + mi_size < cm->mi_rows && vt.vt->vert[0].variance < threshold
&& vt.vt->vert[1].variance < threshold) {
set_block_size(cm, m, get_subsize(block_size, PARTITION_VERT), mis, mi_row,
mi_col);
return 1;
}
// horizontal split is available on all but the right border
if (mi_col + mi_size < cm->mi_cols && vt.vt->horz[0].variance < threshold
&& vt.vt->horz[1].variance < threshold) {
set_block_size(cm, m, get_subsize(block_size, PARTITION_HORZ), mis, mi_row,
mi_col);
return 1;
}
return 0;
}
#endif
static void choose_partitioning(VP9_COMP *cpi, MODE_INFO *m, int mi_row,
int mi_col) {
VP9_COMMON * const cm = &cpi->common;
MACROBLOCK *x = &cpi->mb;
MACROBLOCKD *xd = &cpi->mb.e_mbd;
const int mis = cm->mode_info_stride;
// TODO(JBB): More experimentation or testing of this threshold;
int64_t threshold = 4;
int i, j, k;
v64x64 vt;
unsigned char * s;
1051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120
int sp;
const unsigned char * d;
int dp;
int pixels_wide = 64, pixels_high = 64;
vpx_memset(&vt, 0, sizeof(vt));
set_offsets(cpi, mi_row, mi_col, BLOCK_SIZE_SB64X64);
if (xd->mb_to_right_edge < 0)
pixels_wide += (xd->mb_to_right_edge >> 3);
if (xd->mb_to_bottom_edge < 0)
pixels_high += (xd->mb_to_bottom_edge >> 3);
s = x->plane[0].src.buf;
sp = x->plane[0].src.stride;
// TODO(JBB): Clearly the higher the quantizer the fewer partitions we want
// but this needs more experimentation.
threshold = threshold * cpi->common.base_qindex * cpi->common.base_qindex;
d = vp9_64x64_zeros;
dp = 64;
if (cm->frame_type != KEY_FRAME) {
int_mv nearest_mv, near_mv;
YV12_BUFFER_CONFIG *ref_fb = &cm->yv12_fb[0];
YV12_BUFFER_CONFIG *second_ref_fb = NULL;
setup_pre_planes(xd, 0, ref_fb, mi_row, mi_col,
&xd->scale_factor[0]);
setup_pre_planes(xd, 1, second_ref_fb, mi_row, mi_col,
&xd->scale_factor[1]);
xd->mode_info_context->mbmi.ref_frame[0] = LAST_FRAME;
xd->mode_info_context->mbmi.sb_type = BLOCK_SIZE_SB64X64;
vp9_find_best_ref_mvs(xd, m->mbmi.ref_mvs[m->mbmi.ref_frame[0]],
&nearest_mv, &near_mv);
xd->mode_info_context->mbmi.mv[0] = nearest_mv;
vp9_build_inter_predictors_sby(xd, mi_row, mi_col, BLOCK_SIZE_SB64X64);
d = xd->plane[0].dst.buf;
dp = xd->plane[0].dst.stride;
}
// Fill in the entire tree of 8x8 variances for splits.
for (i = 0; i < 4; i++) {
const int x32_idx = ((i & 1) << 5);
const int y32_idx = ((i >> 1) << 5);
for (j = 0; j < 4; j++) {
const int x16_idx = x32_idx + ((j & 1) << 4);
const int y16_idx = y32_idx + ((j >> 1) << 4);
v16x16 *vst = &vt.split[i].split[j];
for (k = 0; k < 4; k++) {
int x_idx = x16_idx + ((k & 1) << 3);
int y_idx = y16_idx + ((k >> 1) << 3);
unsigned int sse = 0;
int sum = 0;
if (x_idx < pixels_wide && y_idx < pixels_high)
vp9_get_sse_sum_8x8(s + y_idx * sp + x_idx, sp,
d + y_idx * dp + x_idx, dp, &sse, &sum);
fill_variance(&vst->split[k].vt.none, sse, sum, 64);
}
}
}
// Fill the rest of the variance tree by summing the split partition
// values.
for (i = 0; i < 4; i++) {
for (j = 0; j < 4; j++) {
fill_variance_tree(&vt.split[i].split[j], BLOCK_SIZE_MB16X16);
1121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190
}
fill_variance_tree(&vt.split[i], BLOCK_SIZE_SB32X32);
}
fill_variance_tree(&vt, BLOCK_SIZE_SB64X64);
// Now go through the entire structure, splitting every block size until
// we get to one that's got a variance lower than our threshold, or we
// hit 8x8.
if (!set_vt_partitioning(cpi, &vt, m, BLOCK_SIZE_SB64X64, mi_row, mi_col,
4)) {
for (i = 0; i < 4; ++i) {
const int x32_idx = ((i & 1) << 2);
const int y32_idx = ((i >> 1) << 2);
if (!set_vt_partitioning(cpi, &vt.split[i], m, BLOCK_SIZE_SB32X32,
(mi_row + y32_idx), (mi_col + x32_idx), 2)) {
for (j = 0; j < 4; ++j) {
const int x16_idx = ((j & 1) << 1);
const int y16_idx = ((j >> 1) << 1);
if (!set_vt_partitioning(cpi, &vt.split[i].split[j], m,
BLOCK_SIZE_MB16X16,
(mi_row + y32_idx + y16_idx),
(mi_col + x32_idx + x16_idx), 1)) {
for (k = 0; k < 4; ++k) {
const int x8_idx = (k & 1);
const int y8_idx = (k >> 1);
set_block_size(cm, m, BLOCK_SIZE_SB8X8, mis,
(mi_row + y32_idx + y16_idx + y8_idx),
(mi_col + x32_idx + x16_idx + x8_idx));
}
}
}
}
}
}
}
static void rd_use_partition(VP9_COMP *cpi, MODE_INFO *m, TOKENEXTRA **tp,
int mi_row, int mi_col, BLOCK_SIZE_TYPE bsize,
int *rate, int64_t *dist, int do_recon) {
VP9_COMMON * const cm = &cpi->common;
MACROBLOCK * const x = &cpi->mb;
MACROBLOCKD *xd = &cpi->mb.e_mbd;
const int mis = cm->mode_info_stride;
int bwl = b_width_log2(m->mbmi.sb_type);
int bhl = b_height_log2(m->mbmi.sb_type);
int bsl = b_width_log2(bsize);
int bs = (1 << bsl);
int bh = (1 << bhl);
int ms = bs / 2;
int mh = bh / 2;
int bss = (1 << bsl) / 4;
int i, pl;
PARTITION_TYPE partition = PARTITION_NONE;
BLOCK_SIZE_TYPE subsize;
ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
PARTITION_CONTEXT sl[8], sa[8];
int last_part_rate = INT_MAX;
int64_t last_part_dist = INT_MAX;
int split_rate = INT_MAX;
int64_t split_dist = INT_MAX;
int none_rate = INT_MAX;
int64_t none_dist = INT_MAX;
int chosen_rate = INT_MAX;
int64_t chosen_dist = INT_MAX;
BLOCK_SIZE_TYPE sub_subsize = BLOCK_SIZE_AB4X4;
int splits_below = 0;
BLOCK_SIZE_TYPE bs_type = m->mbmi.sb_type;
if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
return;
// parse the partition type
1191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239124012411242124312441245124612471248124912501251125212531254125512561257125812591260
if ((bwl == bsl) && (bhl == bsl))
partition = PARTITION_NONE;
else if ((bwl == bsl) && (bhl < bsl))
partition = PARTITION_HORZ;
else if ((bwl < bsl) && (bhl == bsl))
partition = PARTITION_VERT;
else if ((bwl < bsl) && (bhl < bsl))
partition = PARTITION_SPLIT;
else
assert(0);
subsize = get_subsize(bsize, partition);
if (bsize < BLOCK_SIZE_SB8X8) {
if (xd->ab_index != 0) {
*rate = 0;
*dist = 0;
return;
}
} else {
*(get_sb_partitioning(x, bsize)) = subsize;
}
save_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
if (cpi->sf.adjust_partitioning_from_last_frame) {
// Check if any of the sub blocks are further split.
if (partition == PARTITION_SPLIT && subsize > BLOCK_SIZE_SB8X8) {
sub_subsize = get_subsize(subsize, PARTITION_SPLIT);
splits_below = 1;
for (i = 0; i < 4; i++) {
int jj = i >> 1, ii = i & 0x01;
if (m[jj * bss * mis + ii * bss].mbmi.sb_type >= sub_subsize) {
splits_below = 0;
}
}
}
// If partition is not none try none unless each of the 4 splits are split
// even further..
if (partition != PARTITION_NONE && !splits_below &&
mi_row + (ms >> 1) < cm->mi_rows &&
mi_col + (ms >> 1) < cm->mi_cols) {
*(get_sb_partitioning(x, bsize)) = bsize;
pick_sb_modes(cpi, mi_row, mi_col, &none_rate, &none_dist, bsize,
get_block_context(x, bsize), INT64_MAX);
set_partition_seg_context(cm, xd, mi_row, mi_col);
pl = partition_plane_context(xd, bsize);
none_rate += x->partition_cost[pl][PARTITION_NONE];
restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
m->mbmi.sb_type = bs_type;
*(get_sb_partitioning(x, bsize)) = subsize;
}
}
switch (partition) {
case PARTITION_NONE:
pick_sb_modes(cpi, mi_row, mi_col, &last_part_rate, &last_part_dist,
bsize, get_block_context(x, bsize), INT64_MAX);
set_partition_seg_context(cm, xd, mi_row, mi_col);
pl = partition_plane_context(xd, bsize);
last_part_rate += x->partition_cost[pl][PARTITION_NONE];
break;
case PARTITION_HORZ:
*(get_sb_index(xd, subsize)) = 0;
pick_sb_modes(cpi, mi_row, mi_col, &last_part_rate, &last_part_dist,
subsize, get_block_context(x, subsize), INT64_MAX);
if (bsize >= BLOCK_SIZE_SB8X8 && mi_row + (mh >> 1) < cm->mi_rows) {
int rt = 0;
1261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330
int64_t dt = 0;
update_state(cpi, get_block_context(x, subsize), subsize, 0);
encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize);
*(get_sb_index(xd, subsize)) = 1;
pick_sb_modes(cpi, mi_row + (ms >> 1), mi_col, &rt, &dt, subsize,
get_block_context(x, subsize), INT64_MAX);
last_part_rate += rt;
last_part_dist += dt;
}
set_partition_seg_context(cm, xd, mi_row, mi_col);
pl = partition_plane_context(xd, bsize);
last_part_rate += x->partition_cost[pl][PARTITION_HORZ];
break;
case PARTITION_VERT:
*(get_sb_index(xd, subsize)) = 0;
pick_sb_modes(cpi, mi_row, mi_col, &last_part_rate, &last_part_dist,
subsize, get_block_context(x, subsize), INT64_MAX);
if (bsize >= BLOCK_SIZE_SB8X8 && mi_col + (ms >> 1) < cm->mi_cols) {
int rt = 0;
int64_t dt = 0;
update_state(cpi, get_block_context(x, subsize), subsize, 0);
encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize);
*(get_sb_index(xd, subsize)) = 1;
pick_sb_modes(cpi, mi_row, mi_col + (ms >> 1), &rt, &dt, subsize,
get_block_context(x, subsize), INT64_MAX);
last_part_rate += rt;
last_part_dist += dt;
}
set_partition_seg_context(cm, xd, mi_row, mi_col);
pl = partition_plane_context(xd, bsize);
last_part_rate += x->partition_cost[pl][PARTITION_VERT];
break;
case PARTITION_SPLIT:
// Split partition.
last_part_rate = 0;
last_part_dist = 0;
for (i = 0; i < 4; i++) {
int x_idx = (i & 1) * (ms >> 1);
int y_idx = (i >> 1) * (ms >> 1);
int jj = i >> 1, ii = i & 0x01;
int rt;
int64_t dt;
if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
continue;
*(get_sb_index(xd, subsize)) = i;
rd_use_partition(cpi, m + jj * bss * mis + ii * bss, tp, mi_row + y_idx,
mi_col + x_idx, subsize, &rt, &dt, i != 3);
last_part_rate += rt;
last_part_dist += dt;
}
set_partition_seg_context(cm, xd, mi_row, mi_col);
pl = partition_plane_context(xd, bsize);
last_part_rate += x->partition_cost[pl][PARTITION_SPLIT];
break;
default:
assert(0);
}
if (cpi->sf.adjust_partitioning_from_last_frame
&& partition != PARTITION_SPLIT && bsize > BLOCK_SIZE_SB8X8
&& (mi_row + ms < cm->mi_rows || mi_row + (ms >> 1) == cm->mi_rows)
&& (mi_col + ms < cm->mi_cols || mi_col + (ms >> 1) == cm->mi_cols)) {
BLOCK_SIZE_TYPE split_subsize = get_subsize(bsize, PARTITION_SPLIT);
split_rate = 0;
split_dist = 0;
restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
// Split partition.
1331133213331334133513361337133813391340134113421343134413451346134713481349135013511352135313541355135613571358135913601361136213631364136513661367136813691370137113721373137413751376137713781379138013811382138313841385138613871388138913901391139213931394139513961397139813991400
for (i = 0; i < 4; i++) {
int x_idx = (i & 1) * (bs >> 2);
int y_idx = (i >> 1) * (bs >> 2);
int rt = 0;
int64_t dt = 0;
ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
PARTITION_CONTEXT sl[8], sa[8];
if ((mi_row + y_idx >= cm->mi_rows)
|| (mi_col + x_idx >= cm->mi_cols))
continue;
*(get_sb_index(xd, split_subsize)) = i;
*(get_sb_partitioning(x, bsize)) = split_subsize;
*(get_sb_partitioning(x, split_subsize)) = split_subsize;
save_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
pick_sb_modes(cpi, mi_row + y_idx, mi_col + x_idx, &rt, &dt,
split_subsize, get_block_context(x, split_subsize),
INT64_MAX);
restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
if (rt < INT_MAX && dt < INT_MAX && i != 3)
encode_sb(cpi, tp, mi_row + y_idx, mi_col + x_idx, 0,
split_subsize);
split_rate += rt;
split_dist += dt;
set_partition_seg_context(cm, xd, mi_row + y_idx, mi_col + x_idx);
pl = partition_plane_context(xd, bsize);
split_rate += x->partition_cost[pl][PARTITION_NONE];
}
set_partition_seg_context(cm, xd, mi_row, mi_col);
pl = partition_plane_context(xd, bsize);
split_rate += x->partition_cost[pl][PARTITION_SPLIT];
chosen_rate = split_rate;
chosen_dist = split_dist;
}
// If last_part is better set the partitioning to that...
if (RDCOST(x->rdmult, x->rddiv, last_part_rate, last_part_dist)
< RDCOST(x->rdmult, x->rddiv, chosen_rate, chosen_dist)) {
m->mbmi.sb_type = bsize;
if (bsize >= BLOCK_SIZE_SB8X8)
*(get_sb_partitioning(x, bsize)) = subsize;
chosen_rate = last_part_rate;
chosen_dist = last_part_dist;
}
// If none was better set the partitioning to that...
if (RDCOST(x->rdmult, x->rddiv, chosen_rate, chosen_dist)
> RDCOST(x->rdmult, x->rddiv, none_rate, none_dist)) {
if (bsize >= BLOCK_SIZE_SB8X8)
*(get_sb_partitioning(x, bsize)) = bsize;
chosen_rate = none_rate;
chosen_dist = none_dist;
}
restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
// We must have chosen a partitioning and encoding or we'll fail later on.
// No other opportunities for success.
assert(chosen_rate < INT_MAX && chosen_dist < INT_MAX);
if (do_recon)
encode_sb(cpi, tp, mi_row, mi_col, bsize == BLOCK_SIZE_SB64X64, bsize);
*rate = chosen_rate;
*dist = chosen_dist;
1401140214031404140514061407140814091410141114121413141414151416141714181419142014211422142314241425142614271428142914301431143214331434143514361437143814391440144114421443144414451446144714481449145014511452145314541455145614571458145914601461146214631464146514661467146814691470
}
// TODO(jingning,jimbankoski,rbultje): properly skip partition types that are
// unlikely to be selected depending on previously rate-distortion optimization
// results, for encoding speed-up.
static void rd_pick_partition(VP9_COMP *cpi, TOKENEXTRA **tp, int mi_row,
int mi_col, BLOCK_SIZE_TYPE bsize, int *rate,
int64_t *dist, int do_recon, int64_t best_rd) {
VP9_COMMON * const cm = &cpi->common;
MACROBLOCK * const x = &cpi->mb;
MACROBLOCKD * const xd = &x->e_mbd;
int bsl = b_width_log2(bsize), bs = 1 << bsl;
int ms = bs / 2;
ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
PARTITION_CONTEXT sl[8], sa[8];
TOKENEXTRA *tp_orig = *tp;
int i, pl;
BLOCK_SIZE_TYPE subsize;
int srate = INT_MAX;
int64_t sdist = INT_MAX;
(void) *tp_orig;
if (bsize < BLOCK_SIZE_SB8X8)
if (xd->ab_index != 0) {
*rate = 0;
*dist = 0;
return;
}
assert(mi_height_log2(bsize) == mi_width_log2(bsize));
save_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
// PARTITION_SPLIT
if (!cpi->sf.use_partitions_greater_than
|| (cpi->sf.use_partitions_greater_than
&& bsize > cpi->sf.greater_than_block_size)) {
if (bsize > BLOCK_SIZE_SB8X8) {
int r4 = 0;
int64_t d4 = 0, sum_rd = 0;
subsize = get_subsize(bsize, PARTITION_SPLIT);
for (i = 0; i < 4 && sum_rd < best_rd; ++i) {
int x_idx = (i & 1) * (ms >> 1);
int y_idx = (i >> 1) * (ms >> 1);
int r = 0;
int64_t d = 0;
if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
continue;
*(get_sb_index(xd, subsize)) = i;
rd_pick_partition(cpi, tp, mi_row + y_idx, mi_col + x_idx, subsize, &r,
&d, i != 3, best_rd - sum_rd);
if (r == INT_MAX) {
r4 = INT_MAX;
sum_rd = INT64_MAX;
} else {
r4 += r;
d4 += d;
sum_rd = RDCOST(x->rdmult, x->rddiv, r4, d4);
}
}
set_partition_seg_context(cm, xd, mi_row, mi_col);
pl = partition_plane_context(xd, bsize);
if (r4 != INT_MAX && i == 4) {
r4 += x->partition_cost[pl][PARTITION_SPLIT];
*(get_sb_partitioning(x, bsize)) = subsize;
1471147214731474147514761477147814791480148114821483148414851486148714881489149014911492149314941495149614971498149915001501150215031504150515061507150815091510151115121513151415151516151715181519152015211522152315241525152615271528152915301531153215331534153515361537153815391540
assert(r4 >= 0);
assert(d4 >= 0);
srate = r4;
sdist = d4;
best_rd = MIN(best_rd, RDCOST(x->rdmult, x->rddiv, r4, d4));
}
restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
}
}
x->fast_ms = 0;
x->pred_mv.as_int = 0;
x->subblock_ref = 0;
// Use 4 subblocks' motion estimation results to speed up current
// partition's checking.
if (cpi->sf.using_small_partition_info) {
// Only use 8x8 result for non HD videos.
// int use_8x8 = (MIN(cpi->common.width, cpi->common.height) < 720) ? 1 : 0;
int use_8x8 = 1;
if (cm->frame_type && !cpi->is_src_frame_alt_ref &&
((use_8x8 && bsize == BLOCK_SIZE_MB16X16) ||
bsize == BLOCK_SIZE_SB32X32 || bsize == BLOCK_SIZE_SB64X64)) {
int ref0 = 0, ref1 = 0, ref2 = 0, ref3 = 0;
if (bsize == BLOCK_SIZE_MB16X16) {
ref0 = x->sb8x8_context[xd->sb_index][xd->mb_index][0].mic.mbmi.
ref_frame[0];
ref1 = x->sb8x8_context[xd->sb_index][xd->mb_index][1].mic.mbmi.
ref_frame[0];
ref2 = x->sb8x8_context[xd->sb_index][xd->mb_index][2].mic.mbmi.
ref_frame[0];
ref3 = x->sb8x8_context[xd->sb_index][xd->mb_index][3].mic.mbmi.
ref_frame[0];
} else if (bsize == BLOCK_SIZE_SB32X32) {
ref0 = x->mb_context[xd->sb_index][0].mic.mbmi.ref_frame[0];
ref1 = x->mb_context[xd->sb_index][1].mic.mbmi.ref_frame[0];
ref2 = x->mb_context[xd->sb_index][2].mic.mbmi.ref_frame[0];
ref3 = x->mb_context[xd->sb_index][3].mic.mbmi.ref_frame[0];
} else if (bsize == BLOCK_SIZE_SB64X64) {
ref0 = x->sb32_context[0].mic.mbmi.ref_frame[0];
ref1 = x->sb32_context[1].mic.mbmi.ref_frame[0];
ref2 = x->sb32_context[2].mic.mbmi.ref_frame[0];
ref3 = x->sb32_context[3].mic.mbmi.ref_frame[0];
}
// Currently, only consider 4 inter ref frames.
if (ref0 && ref1 && ref2 && ref3) {
int16_t mvr0 = 0, mvc0 = 0, mvr1 = 0, mvc1 = 0, mvr2 = 0, mvc2 = 0,
mvr3 = 0, mvc3 = 0;
int d01, d23, d02, d13; // motion vector distance between 2 blocks
// Get each subblock's motion vectors.
if (bsize == BLOCK_SIZE_MB16X16) {
mvr0 = x->sb8x8_context[xd->sb_index][xd->mb_index][0].mic.mbmi.mv[0].
as_mv.row;
mvc0 = x->sb8x8_context[xd->sb_index][xd->mb_index][0].mic.mbmi.mv[0].
as_mv.col;
mvr1 = x->sb8x8_context[xd->sb_index][xd->mb_index][1].mic.mbmi.mv[0].
as_mv.row;
mvc1 = x->sb8x8_context[xd->sb_index][xd->mb_index][1].mic.mbmi.mv[0].
as_mv.col;
mvr2 = x->sb8x8_context[xd->sb_index][xd->mb_index][2].mic.mbmi.mv[0].
as_mv.row;
mvc2 = x->sb8x8_context[xd->sb_index][xd->mb_index][2].mic.mbmi.mv[0].
as_mv.col;
mvr3 = x->sb8x8_context[xd->sb_index][xd->mb_index][3].mic.mbmi.mv[0].
as_mv.row;
mvc3 = x->sb8x8_context[xd->sb_index][xd->mb_index][3].mic.mbmi.mv[0].
1541154215431544154515461547154815491550155115521553155415551556155715581559156015611562156315641565156615671568156915701571157215731574157515761577157815791580158115821583158415851586158715881589159015911592159315941595159615971598159916001601160216031604160516061607160816091610
as_mv.col;
} else if (bsize == BLOCK_SIZE_SB32X32) {
mvr0 = x->mb_context[xd->sb_index][0].mic.mbmi.mv[0].as_mv.row;
mvc0 = x->mb_context[xd->sb_index][0].mic.mbmi.mv[0].as_mv.col;
mvr1 = x->mb_context[xd->sb_index][1].mic.mbmi.mv[0].as_mv.row;
mvc1 = x->mb_context[xd->sb_index][1].mic.mbmi.mv[0].as_mv.col;
mvr2 = x->mb_context[xd->sb_index][2].mic.mbmi.mv[0].as_mv.row;
mvc2 = x->mb_context[xd->sb_index][2].mic.mbmi.mv[0].as_mv.col;
mvr3 = x->mb_context[xd->sb_index][3].mic.mbmi.mv[0].as_mv.row;
mvc3 = x->mb_context[xd->sb_index][3].mic.mbmi.mv[0].as_mv.col;
} else if (bsize == BLOCK_SIZE_SB64X64) {
mvr0 = x->sb32_context[0].mic.mbmi.mv[0].as_mv.row;
mvc0 = x->sb32_context[0].mic.mbmi.mv[0].as_mv.col;
mvr1 = x->sb32_context[1].mic.mbmi.mv[0].as_mv.row;
mvc1 = x->sb32_context[1].mic.mbmi.mv[0].as_mv.col;
mvr2 = x->sb32_context[2].mic.mbmi.mv[0].as_mv.row;
mvc2 = x->sb32_context[2].mic.mbmi.mv[0].as_mv.col;
mvr3 = x->sb32_context[3].mic.mbmi.mv[0].as_mv.row;
mvc3 = x->sb32_context[3].mic.mbmi.mv[0].as_mv.col;
}
// Adjust sign if ref is alt_ref
if (cm->ref_frame_sign_bias[ref0]) {
mvr0 *= -1;
mvc0 *= -1;
}
if (cm->ref_frame_sign_bias[ref1]) {
mvr1 *= -1;
mvc1 *= -1;
}
if (cm->ref_frame_sign_bias[ref2]) {
mvr2 *= -1;
mvc2 *= -1;
}
if (cm->ref_frame_sign_bias[ref3]) {
mvr3 *= -1;
mvc3 *= -1;
}
// Calculate mv distances.
d01 = MAX(abs(mvr0 - mvr1), abs(mvc0 - mvc1));
d23 = MAX(abs(mvr2 - mvr3), abs(mvc2 - mvc3));
d02 = MAX(abs(mvr0 - mvr2), abs(mvc0 - mvc2));
d13 = MAX(abs(mvr1 - mvr3), abs(mvc1 - mvc3));
if (d01 < 24 && d23 < 24 && d02 < 24 && d13 < 24) {
// Set fast motion search level.
x->fast_ms = 1;
// Calculate prediction MV
x->pred_mv.as_mv.row = (mvr0 + mvr1 + mvr2 + mvr3) >> 2;
x->pred_mv.as_mv.col = (mvc0 + mvc1 + mvc2 + mvc3) >> 2;
if (ref0 == ref1 && ref1 == ref2 && ref2 == ref3 &&
d01 < 2 && d23 < 2 && d02 < 2 && d13 < 2) {
// Set fast motion search level.
x->fast_ms = 2;
if (!d01 && !d23 && !d02 && !d13) {
x->fast_ms = 3;
x->subblock_ref = ref0;
}
}
}
}
}
}
1611161216131614161516161617161816191620162116221623162416251626162716281629163016311632163316341635163616371638163916401641164216431644164516461647164816491650165116521653165416551656165716581659166016611662166316641665166616671668166916701671167216731674167516761677167816791680
if (!cpi->sf.use_partitions_less_than
|| (cpi->sf.use_partitions_less_than
&& bsize <= cpi->sf.less_than_block_size)) {
int larger_is_better = 0;
// PARTITION_NONE
if ((mi_row + (ms >> 1) < cm->mi_rows) &&
(mi_col + (ms >> 1) < cm->mi_cols)) {
int r;
int64_t d;
pick_sb_modes(cpi, mi_row, mi_col, &r, &d, bsize,
get_block_context(x, bsize), best_rd);
if (r != INT_MAX && bsize >= BLOCK_SIZE_SB8X8) {
set_partition_seg_context(cm, xd, mi_row, mi_col);
pl = partition_plane_context(xd, bsize);
r += x->partition_cost[pl][PARTITION_NONE];
}
if (r != INT_MAX &&
(bsize == BLOCK_SIZE_SB8X8 ||
RDCOST(x->rdmult, x->rddiv, r, d) <
RDCOST(x->rdmult, x->rddiv, srate, sdist))) {
best_rd = MIN(best_rd, RDCOST(x->rdmult, x->rddiv, r, d));
srate = r;
sdist = d;
larger_is_better = 1;
if (bsize >= BLOCK_SIZE_SB8X8)
*(get_sb_partitioning(x, bsize)) = bsize;
}
}
if (bsize == BLOCK_SIZE_SB8X8) {
int r4 = 0;
int64_t d4 = 0, sum_rd = 0;
subsize = get_subsize(bsize, PARTITION_SPLIT);
for (i = 0; i < 4 && sum_rd < best_rd; ++i) {
int x_idx = (i & 1) * (ms >> 1);
int y_idx = (i >> 1) * (ms >> 1);
int r = 0;
int64_t d = 0;
if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
continue;
*(get_sb_index(xd, subsize)) = i;
rd_pick_partition(cpi, tp, mi_row + y_idx, mi_col + x_idx, subsize, &r,
&d, i != 3, best_rd - sum_rd);
if (r == INT_MAX) {
r4 = INT_MAX;
sum_rd = INT64_MAX;
} else {
r4 += r;
d4 += d;
sum_rd = RDCOST(x->rdmult, x->rddiv, r4, d4);
}
}
set_partition_seg_context(cm, xd, mi_row, mi_col);
pl = partition_plane_context(xd, bsize);
if (r4 != INT_MAX && i == 4) {
r4 += x->partition_cost[pl][PARTITION_SPLIT];
if (RDCOST(x->rdmult, x->rddiv, r4, d4) <
RDCOST(x->rdmult, x->rddiv, srate, sdist)) {
srate = r4;
sdist = d4;
larger_is_better = 0;
*(get_sb_partitioning(x, bsize)) = subsize;
best_rd = MIN(best_rd, RDCOST(x->rdmult, x->rddiv, r4, d4));
}