-
Ronald S. Bultje authored
Use these, instead of the 4/5-dimensional arrays, to hold statistics, counts, accumulations and probabilities for coefficient tokens. This commit also re-allows ENTROPY_STATS to compile. Change-Id: If441ffac936f52a3af91d8f2922ea8a0ceabdaa5
885cf816
/* * 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_ports/config.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_entropymode.h"#include "vp9/common/vp9_quant_common.h"#include "vp9/encoder/vp9_segmentation.h"#include "vp9/common/vp9_setupintrarecon.h"#include "vp9/common/vp9_reconintra4x4.h"#include "vp9/encoder/vp9_encodeintra.h"#include "vp9/common/vp9_reconinter.h"#include "vp9/common/vp9_invtrans.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/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_macroblock(VP9_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t, int recon_yoffset, int recon_uvoffset, int output_enabled, int mb_col, int mb_row);static void encode_superblock(VP9_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t, int recon_yoffset, int recon_uvoffset, int mb_col, int mb_row);static void adjust_act_zbin(VP9_COMP *cpi, MACROBLOCK *x);#ifdef MODE_STATSunsigned int inter_y_modes[MB_MODE_COUNT];unsigned int inter_uv_modes[VP9_UV_MODES];unsigned int inter_b_modes[B_MODE_COUNT];unsigned int y_modes[VP9_YMODES];unsigned int i8x8_modes[VP9_I8X8_MODES];unsigned int uv_modes[VP9_UV_MODES];unsigned int uv_modes_y[VP9_YMODES][VP9_UV_MODES];unsigned int b_modes[B_MODE_COUNT];#endif/* activity_avg must be positive, or flat regions could get a zero weight7172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140
* (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 unsigned char 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 would
* 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->src.y_buffer, x->src.y_stride, VP9_VAR_OFFS, 0,
&sse);
act = 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);
}
// 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
141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210
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(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]) {
// 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 and 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
211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280
// 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
// 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;
// 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->dst.y_buffer = new_yv12->y_buffer + recon_yoffset;
xd->left_available = (mb_col != 0);
recon_yoffset += 16;
281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350
#endif
#if !CONFIG_SUPERBLOCKS
// Copy current mb to a buffer
vp9_copy_mem16x16(x->src.y_buffer, x->src.y_stride, x->thismb, 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->src.y_buffer += 16;
}
// adjust to the next row of mbs
x->src.y_buffer += 16 * x->src.y_stride - 16 * cm->mb_cols;
#if ALT_ACT_MEASURE
// extend the recon for intra prediction
vp9_extend_mb_row(new_yv12, xd->dst.y_buffer + 16,
xd->dst.u_buffer + 8, xd->dst.v_buffer + 8);
#endif
}
// 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);
}
#if CONFIG_NEW_MVREF
static int vp9_cost_mv_ref_id(vp9_prob * ref_id_probs, int mv_ref_id) {
351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420
int cost;
// Encode the index for the MV reference.
switch (mv_ref_id) {
case 0:
cost = vp9_cost_zero(ref_id_probs[0]);
break;
case 1:
cost = vp9_cost_one(ref_id_probs[0]);
cost += vp9_cost_zero(ref_id_probs[1]);
break;
case 2:
cost = vp9_cost_one(ref_id_probs[0]);
cost += vp9_cost_one(ref_id_probs[1]);
cost += vp9_cost_zero(ref_id_probs[2]);
break;
case 3:
cost = vp9_cost_one(ref_id_probs[0]);
cost += vp9_cost_one(ref_id_probs[1]);
cost += vp9_cost_one(ref_id_probs[2]);
break;
// TRAP.. This should not happen
default:
assert(0);
break;
}
return cost;
}
// Estimate the cost of each coding the vector using each reference candidate
static unsigned int pick_best_mv_ref(MACROBLOCK *x,
MV_REFERENCE_FRAME ref_frame,
int_mv target_mv,
int_mv * mv_ref_list,
int_mv * best_ref) {
int i;
int best_index = 0;
int cost, cost2;
int zero_seen = (mv_ref_list[0].as_int) ? FALSE : TRUE;
MACROBLOCKD *xd = &x->e_mbd;
int max_mv = MV_MAX;
cost = vp9_cost_mv_ref_id(xd->mb_mv_ref_id_probs[ref_frame], 0) +
vp9_mv_bit_cost(&target_mv, &mv_ref_list[0], x->nmvjointcost,
x->mvcost, 96, xd->allow_high_precision_mv);
for (i = 1; i < MAX_MV_REF_CANDIDATES; ++i) {
// If we see a 0,0 reference vector for a second time we have reached
// the end of the list of valid candidate vectors.
if (!mv_ref_list[i].as_int) {
if (zero_seen)
break;
else
zero_seen = TRUE;
}
// Check for cases where the reference choice would give rise to an
// uncodable/out of range residual for row or col.
if ((abs(target_mv.as_mv.row - mv_ref_list[i].as_mv.row) > max_mv) ||
(abs(target_mv.as_mv.col - mv_ref_list[i].as_mv.col) > max_mv)) {
continue;
}
cost2 = vp9_cost_mv_ref_id(xd->mb_mv_ref_id_probs[ref_frame], i) +
vp9_mv_bit_cost(&target_mv, &mv_ref_list[i], x->nmvjointcost,
x->mvcost, 96, xd->allow_high_precision_mv);
if (cost2 < cost) {
cost = cost2;
421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490
best_index = i;
}
}
best_ref->as_int = mv_ref_list[best_index].as_int;
return best_index;
}
#endif
static void update_state(VP9_COMP *cpi, MACROBLOCK *x,
PICK_MODE_CONTEXT *ctx) {
int i;
MACROBLOCKD *xd = &x->e_mbd;
MODE_INFO *mi = &ctx->mic;
MB_MODE_INFO * mbmi = &xd->mode_info_context->mbmi;
int mb_mode = mi->mbmi.mode;
int mb_mode_index = ctx->best_mode_index;
#if CONFIG_DEBUG
assert(mb_mode < MB_MODE_COUNT);
assert(mb_mode_index < MAX_MODES);
assert(mi->mbmi.ref_frame < MAX_REF_FRAMES);
#endif
// Restore the coding context of the MB to that that was in place
// when the mode was picked for it
vpx_memcpy(xd->mode_info_context, mi, sizeof(MODE_INFO));
#if CONFIG_SUPERBLOCKS
if (mi->mbmi.encoded_as_sb) {
const int mis = cpi->common.mode_info_stride;
if (xd->mb_to_right_edge >= 0)
vpx_memcpy(xd->mode_info_context + 1, mi, sizeof(MODE_INFO));
if (xd->mb_to_bottom_edge >= 0) {
vpx_memcpy(xd->mode_info_context + mis, mi, sizeof(MODE_INFO));
if (xd->mb_to_right_edge >= 0)
vpx_memcpy(xd->mode_info_context + mis + 1, mi, sizeof(MODE_INFO));
}
#if CONFIG_TX32X32 && CONFIG_SUPERBLOCKS
} else {
ctx->txfm_rd_diff[ALLOW_32X32] = ctx->txfm_rd_diff[ALLOW_16X16];
#endif
}
#endif
if (mb_mode == B_PRED) {
for (i = 0; i < 16; i++) {
xd->block[i].bmi.as_mode = xd->mode_info_context->bmi[i].as_mode;
assert(xd->block[i].bmi.as_mode.first < B_MODE_COUNT);
}
} else if (mb_mode == I8X8_PRED) {
for (i = 0; i < 16; i++) {
xd->block[i].bmi = xd->mode_info_context->bmi[i];
}
} else if (mb_mode == SPLITMV) {
vpx_memcpy(x->partition_info, &ctx->partition_info,
sizeof(PARTITION_INFO));
mbmi->mv[0].as_int = x->partition_info->bmi[15].mv.as_int;
mbmi->mv[1].as_int = x->partition_info->bmi[15].second_mv.as_int;
}
{
int segment_id = mbmi->segment_id;
if (!vp9_segfeature_active(xd, segment_id, SEG_LVL_EOB) ||
vp9_get_segdata(xd, segment_id, SEG_LVL_EOB)) {
for (i = 0; i < NB_TXFM_MODES; i++) {
cpi->rd_tx_select_diff[i] += ctx->txfm_rd_diff[i];
}
}
491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560
}
if (cpi->common.frame_type == KEY_FRAME) {
// Restore the coding modes to that held in the coding context
// if (mb_mode == B_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_I8X8_PRED /*I8X8_PRED*/,
THR_B_PRED /*B_PRED*/,
};
cpi->mode_chosen_counts[kf_mode_index[mb_mode]]++;
#endif
} else {
/*
// Reduce the activation RD thresholds for the best choice mode
if ((cpi->rd_baseline_thresh[mb_mode_index] > 0) &&
(cpi->rd_baseline_thresh[mb_mode_index] < (INT_MAX >> 2)))
{
int best_adjustment = (cpi->rd_thresh_mult[mb_mode_index] >> 2);
cpi->rd_thresh_mult[mb_mode_index] =
(cpi->rd_thresh_mult[mb_mode_index]
>= (MIN_THRESHMULT + best_adjustment)) ?
cpi->rd_thresh_mult[mb_mode_index] - best_adjustment :
MIN_THRESHMULT;
cpi->rd_threshes[mb_mode_index] =
(cpi->rd_baseline_thresh[mb_mode_index] >> 7)
* cpi->rd_thresh_mult[mb_mode_index];
}
*/
// Note how often each mode chosen as best
cpi->mode_chosen_counts[mb_mode_index]++;
if (mbmi->mode == SPLITMV || mbmi->mode == NEWMV) {
int_mv best_mv, best_second_mv;
MV_REFERENCE_FRAME rf = mbmi->ref_frame;
#if CONFIG_NEW_MVREF
unsigned int best_index;
MV_REFERENCE_FRAME sec_ref_frame = mbmi->second_ref_frame;
#endif
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[rf][0].as_int;
best_second_mv.as_int = mbmi->ref_mvs[mbmi->second_ref_frame][0].as_int;
#if CONFIG_NEW_MVREF
best_index = pick_best_mv_ref(x, rf, mbmi->mv[0],
mbmi->ref_mvs[rf], &best_mv);
mbmi->best_index = best_index;
if (mbmi->second_ref_frame > 0) {
unsigned int best_index;
best_index =
pick_best_mv_ref(x, sec_ref_frame, mbmi->mv[1],
561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630
mbmi->ref_mvs[sec_ref_frame],
&best_second_mv);
mbmi->best_second_index = best_index;
}
#endif
}
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 CONFIG_COMP_INTERINTRA_PRED
if (mbmi->mode >= NEARESTMV && mbmi->mode < SPLITMV &&
mbmi->second_ref_frame <= INTRA_FRAME) {
if (mbmi->second_ref_frame == INTRA_FRAME) {
++cpi->interintra_count[1];
++cpi->ymode_count[mbmi->interintra_mode];
#if SEPARATE_INTERINTRA_UV
++cpi->y_uv_mode_count[mbmi->interintra_mode][mbmi->interintra_uv_mode];
#endif
} else {
++cpi->interintra_count[0];
}
}
if (cpi->common.mcomp_filter_type == SWITCHABLE &&
mbmi->mode >= NEARESTMV &&
mbmi->mode <= SPLITMV) {
++cpi->switchable_interp_count
[vp9_get_pred_context(&cpi->common, xd, PRED_SWITCHABLE_INTERP)]
[vp9_switchable_interp_map[mbmi->interp_filter]];
}
#endif
cpi->prediction_error += ctx->distortion;
cpi->intra_error += ctx->intra_error;
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;
}
}
static void pick_mb_modes(VP9_COMP *cpi,
VP9_COMMON *cm,
int mb_row,
int mb_col,
MACROBLOCK *x,
MACROBLOCKD *xd,
TOKENEXTRA **tp,
int *totalrate,
int *totaldist) {
int i;
int map_index;
int recon_yoffset, recon_uvoffset;
int ref_fb_idx = cm->lst_fb_idx;
int dst_fb_idx = cm->new_fb_idx;
int recon_y_stride = cm->yv12_fb[ref_fb_idx].y_stride;
int recon_uv_stride = cm->yv12_fb[ref_fb_idx].uv_stride;
ENTROPY_CONTEXT_PLANES left_context[2];
ENTROPY_CONTEXT_PLANES above_context[2];
ENTROPY_CONTEXT_PLANES *initial_above_context_ptr = cm->above_context
+ mb_col;
// Offsets to move pointers from MB to MB within a SB in raster order
int row_delta[4] = { 0, +1, 0, -1};
int col_delta[4] = { +1, -1, +1, +1};
/* Function should not modify L & A contexts; save and restore on exit */
vpx_memcpy(left_context,
cm->left_context,
sizeof(left_context));
631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700
vpx_memcpy(above_context,
initial_above_context_ptr,
sizeof(above_context));
/* Encode MBs in raster order within the SB */
for (i = 0; i < 4; i++) {
int dy = row_delta[i];
int dx = col_delta[i];
int offset_unextended = dy * cm->mb_cols + dx;
int offset_extended = dy * xd->mode_info_stride + dx;
MB_MODE_INFO * mbmi = &xd->mode_info_context->mbmi;
// TODO Many of the index items here can be computed more efficiently!
if ((mb_row >= cm->mb_rows) || (mb_col >= cm->mb_cols)) {
// MB lies outside frame, move on
mb_row += dy;
mb_col += dx;
// Update pointers
x->src.y_buffer += 16 * (dx + dy * x->src.y_stride);
x->src.u_buffer += 8 * (dx + dy * x->src.uv_stride);
x->src.v_buffer += 8 * (dx + dy * x->src.uv_stride);
x->gf_active_ptr += offset_unextended;
x->partition_info += offset_extended;
xd->mode_info_context += offset_extended;
xd->prev_mode_info_context += offset_extended;
#if CONFIG_DEBUG
assert((xd->prev_mode_info_context - cpi->common.prev_mip) ==
(xd->mode_info_context - cpi->common.mip));
#endif
continue;
}
// Index of the MB in the SB 0..3
xd->mb_index = i;
map_index = (mb_row * cpi->common.mb_cols) + mb_col;
x->mb_activity_ptr = &cpi->mb_activity_map[map_index];
// set above context pointer
xd->above_context = cm->above_context + mb_col;
// Restore the appropriate left context depending on which
// row in the SB the MB is situated
xd->left_context = cm->left_context + (i >> 1);
// Set up distance of MB to edge of frame in 1/8th pel units
xd->mb_to_top_edge = -((mb_row * 16) << 3);
xd->mb_to_left_edge = -((mb_col * 16) << 3);
xd->mb_to_bottom_edge = ((cm->mb_rows - 1 - mb_row) * 16) << 3;
xd->mb_to_right_edge = ((cm->mb_cols - 1 - mb_col) * 16) << 3;
// Set up limit values for MV components to prevent them from
// extending beyond the UMV borders assuming 16x16 block size
x->mv_row_min = -((mb_row * 16) + VP9BORDERINPIXELS - VP9_INTERP_EXTEND);
x->mv_col_min = -((mb_col * 16) + VP9BORDERINPIXELS - VP9_INTERP_EXTEND);
x->mv_row_max = ((cm->mb_rows - mb_row) * 16 +
(VP9BORDERINPIXELS - 16 - VP9_INTERP_EXTEND));
x->mv_col_max = ((cm->mb_cols - mb_col) * 16 +
(VP9BORDERINPIXELS - 16 - VP9_INTERP_EXTEND));
xd->up_available = (mb_row != 0);
xd->left_available = (mb_col != 0);
recon_yoffset = (mb_row * recon_y_stride * 16) + (mb_col * 16);
recon_uvoffset = (mb_row * recon_uv_stride * 8) + (mb_col * 8);
xd->dst.y_buffer = cm->yv12_fb[dst_fb_idx].y_buffer + recon_yoffset;
701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770
xd->dst.u_buffer = cm->yv12_fb[dst_fb_idx].u_buffer + recon_uvoffset;
xd->dst.v_buffer = cm->yv12_fb[dst_fb_idx].v_buffer + recon_uvoffset;
#if !CONFIG_SUPERBLOCKS
// Copy current MB to a work buffer
vp9_copy_mem16x16(x->src.y_buffer, x->src.y_stride, x->thismb, 16);
#endif
x->rddiv = cpi->RDDIV;
x->rdmult = cpi->RDMULT;
if (cpi->oxcf.tuning == VP8_TUNE_SSIM)
vp9_activity_masking(cpi, x);
// Is segmentation enabled
if (xd->segmentation_enabled) {
// Code to set segment id in xd->mbmi.segment_id
if (xd->update_mb_segmentation_map)
mbmi->segment_id = cpi->segmentation_map[map_index];
else
mbmi->segment_id = cm->last_frame_seg_map[map_index];
if (mbmi->segment_id > 3)
mbmi->segment_id = 0;
vp9_mb_init_quantizer(cpi, x);
} else
// Set to Segment 0 by default
mbmi->segment_id = 0;
x->active_ptr = cpi->active_map + map_index;
#if CONFIG_SUPERBLOCKS
xd->mode_info_context->mbmi.encoded_as_sb = 0;
#endif
cpi->update_context = 0; // TODO Do we need this now??
vp9_intra_prediction_down_copy(xd);
#ifdef ENC_DEBUG
enc_debug = (cpi->common.current_video_frame == 46 &&
mb_row == 5 && mb_col == 2);
#endif
// 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) {
int r, d;
#ifdef ENC_DEBUG
if (enc_debug)
printf("intra pick_mb_modes %d %d\n", mb_row, mb_col);
#endif
vp9_rd_pick_intra_mode(cpi, x, &r, &d);
*totalrate += r;
*totaldist += d;
// Dummy encode, do not do the tokenization
encode_macroblock(cpi, x, tp,
recon_yoffset, recon_uvoffset, 0, mb_col, mb_row);
// Note the encoder may have changed the segment_id
// Save the coding context
vpx_memcpy(&x->mb_context[i].mic, xd->mode_info_context,
sizeof(MODE_INFO));
} else {
int seg_id, r, d;
if (xd->segmentation_enabled && cpi->seg0_cnt > 0 &&
!vp9_segfeature_active(xd, 0, SEG_LVL_REF_FRAME) &&
vp9_segfeature_active(xd, 1, SEG_LVL_REF_FRAME) &&
vp9_check_segref(xd, 1, INTRA_FRAME) +
771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840
vp9_check_segref(xd, 1, LAST_FRAME) +
vp9_check_segref(xd, 1, GOLDEN_FRAME) +
vp9_check_segref(xd, 1, ALTREF_FRAME) == 1) {
cpi->seg0_progress = (cpi->seg0_idx << 16) / cpi->seg0_cnt;
} else {
cpi->seg0_progress = (((mb_col & ~1) * 2 + (mb_row & ~1) * cm->mb_cols + i) << 16) / cm->MBs;
}
#ifdef ENC_DEBUG
if (enc_debug)
printf("inter pick_mb_modes %d %d\n", mb_row, mb_col);
#endif
vp9_pick_mode_inter_macroblock(cpi, x, recon_yoffset,
recon_uvoffset, &r, &d);
*totalrate += r;
*totaldist += d;
// Dummy encode, do not do the tokenization
encode_macroblock(cpi, x, tp,
recon_yoffset, recon_uvoffset, 0, mb_col, mb_row);
seg_id = mbmi->segment_id;
if (cpi->mb.e_mbd.segmentation_enabled && seg_id == 0) {
cpi->seg0_idx++;
}
if (!xd->segmentation_enabled ||
!vp9_segfeature_active(xd, seg_id, SEG_LVL_REF_FRAME) ||
vp9_check_segref(xd, seg_id, INTRA_FRAME) +
vp9_check_segref(xd, seg_id, LAST_FRAME) +
vp9_check_segref(xd, seg_id, GOLDEN_FRAME) +
vp9_check_segref(xd, seg_id, ALTREF_FRAME) > 1) {
// Get the prediction context and status
int pred_flag = vp9_get_pred_flag(xd, PRED_REF);
int pred_context = vp9_get_pred_context(cm, xd, PRED_REF);
// Count prediction success
cpi->ref_pred_count[pred_context][pred_flag]++;
}
}
// Next MB
mb_row += dy;
mb_col += dx;
x->src.y_buffer += 16 * (dx + dy * x->src.y_stride);
x->src.u_buffer += 8 * (dx + dy * x->src.uv_stride);
x->src.v_buffer += 8 * (dx + dy * x->src.uv_stride);
x->gf_active_ptr += offset_unextended;
x->partition_info += offset_extended;
xd->mode_info_context += offset_extended;
xd->prev_mode_info_context += offset_extended;
#if CONFIG_DEBUG
assert((xd->prev_mode_info_context - cpi->common.prev_mip) ==
(xd->mode_info_context - cpi->common.mip));
#endif
}
/* Restore L & A coding context to those in place on entry */
vpx_memcpy(cm->left_context,
left_context,
sizeof(left_context));
vpx_memcpy(initial_above_context_ptr,
above_context,
sizeof(above_context));
}
#if CONFIG_SUPERBLOCKS
static void pick_sb_modes (VP9_COMP *cpi,
841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910
VP9_COMMON *cm,
int mb_row,
int mb_col,
MACROBLOCK *x,
MACROBLOCKD *xd,
TOKENEXTRA **tp,
int *totalrate,
int *totaldist)
{
int map_index;
int recon_yoffset, recon_uvoffset;
int ref_fb_idx = cm->lst_fb_idx;
int dst_fb_idx = cm->new_fb_idx;
int recon_y_stride = cm->yv12_fb[ref_fb_idx].y_stride;
int recon_uv_stride = cm->yv12_fb[ref_fb_idx].uv_stride;
ENTROPY_CONTEXT_PLANES left_context[2];
ENTROPY_CONTEXT_PLANES above_context[2];
ENTROPY_CONTEXT_PLANES *initial_above_context_ptr = cm->above_context
+ mb_col;
/* Function should not modify L & A contexts; save and restore on exit */
vpx_memcpy (left_context,
cm->left_context,
sizeof(left_context));
vpx_memcpy (above_context,
initial_above_context_ptr,
sizeof(above_context));
map_index = (mb_row * cpi->common.mb_cols) + mb_col;
x->mb_activity_ptr = &cpi->mb_activity_map[map_index];
/* set above context pointer */
xd->above_context = cm->above_context + mb_col;
/* Restore the appropriate left context depending on which
* row in the SB the MB is situated */
xd->left_context = cm->left_context;
// Set up distance of MB to edge of frame in 1/8th pel units
xd->mb_to_top_edge = -((mb_row * 16) << 3);
xd->mb_to_left_edge = -((mb_col * 16) << 3);
xd->mb_to_bottom_edge = ((cm->mb_rows - 2 - mb_row) * 16) << 3;
xd->mb_to_right_edge = ((cm->mb_cols - 2 - mb_col) * 16) << 3;
/* Set up limit values for MV components to prevent them from
* extending beyond the UMV borders assuming 16x16 block size */
x->mv_row_min = -((mb_row * 16) + VP9BORDERINPIXELS - VP9_INTERP_EXTEND);
x->mv_col_min = -((mb_col * 16) + VP9BORDERINPIXELS - VP9_INTERP_EXTEND);
x->mv_row_max = ((cm->mb_rows - mb_row) * 16 +
(VP9BORDERINPIXELS - 32 - VP9_INTERP_EXTEND));
x->mv_col_max = ((cm->mb_cols - mb_col) * 16 +
(VP9BORDERINPIXELS - 32 - VP9_INTERP_EXTEND));
xd->up_available = (mb_row != 0);
xd->left_available = (mb_col != 0);
recon_yoffset = (mb_row * recon_y_stride * 16) + (mb_col * 16);
recon_uvoffset = (mb_row * recon_uv_stride * 8) + (mb_col * 8);
xd->dst.y_buffer = cm->yv12_fb[dst_fb_idx].y_buffer + recon_yoffset;
xd->dst.u_buffer = cm->yv12_fb[dst_fb_idx].u_buffer + recon_uvoffset;
xd->dst.v_buffer = cm->yv12_fb[dst_fb_idx].v_buffer + recon_uvoffset;
#if 0 // FIXME
/* Copy current MB to a work buffer */
vp9_copy_mem16x16(x->src.y_buffer, x->src.y_stride, x->thismb, 16);
#endif
x->rddiv = cpi->RDDIV;
x->rdmult = cpi->RDMULT;
if(cpi->oxcf.tuning == VP8_TUNE_SSIM)
vp9_activity_masking(cpi, x);
911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980
/* Is segmentation enabled */
if (xd->segmentation_enabled)
{
/* Code to set segment id in xd->mbmi.segment_id */
if (xd->update_mb_segmentation_map)
xd->mode_info_context->mbmi.segment_id =
cpi->segmentation_map[map_index] &&
cpi->segmentation_map[map_index + 1] &&
cpi->segmentation_map[map_index + cm->mb_cols] &&
cpi->segmentation_map[map_index + cm->mb_cols + 1];
else
xd->mode_info_context->mbmi.segment_id =
cm->last_frame_seg_map[map_index] &&
cm->last_frame_seg_map[map_index + 1] &&
cm->last_frame_seg_map[map_index + cm->mb_cols] &&
cm->last_frame_seg_map[map_index + cm->mb_cols + 1];
if (xd->mode_info_context->mbmi.segment_id > 3)
xd->mode_info_context->mbmi.segment_id = 0;
vp9_mb_init_quantizer(cpi, x);
}
else
/* Set to Segment 0 by default */
xd->mode_info_context->mbmi.segment_id = 0;
x->active_ptr = cpi->active_map + map_index;
cpi->update_context = 0; // TODO Do we need this now??
/* 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);
/* Save the coding context */
vpx_memcpy(&x->sb_context[0].mic, xd->mode_info_context,
sizeof(MODE_INFO));
} else {
if (xd->segmentation_enabled && cpi->seg0_cnt > 0 &&
!vp9_segfeature_active(xd, 0, SEG_LVL_REF_FRAME) &&
vp9_segfeature_active(xd, 1, SEG_LVL_REF_FRAME) &&
vp9_check_segref(xd, 1, INTRA_FRAME) +
vp9_check_segref(xd, 1, LAST_FRAME) +
vp9_check_segref(xd, 1, GOLDEN_FRAME) +
vp9_check_segref(xd, 1, ALTREF_FRAME) == 1) {
cpi->seg0_progress = (cpi->seg0_idx << 16) / cpi->seg0_cnt;
} else {
cpi->seg0_progress =
(((mb_col & ~1) * 2 + (mb_row & ~1) * cm->mb_cols) << 16) / cm->MBs;
}
vp9_rd_pick_inter_mode_sb(cpi, x,
recon_yoffset,
recon_uvoffset,
totalrate,
totaldist);
}
/* Restore L & A coding context to those in place on entry */
vpx_memcpy (cm->left_context,
left_context,
sizeof(left_context));
vpx_memcpy (initial_above_context_ptr,
above_context,
sizeof(above_context));
}
#endif
981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050
static void encode_sb(VP9_COMP *cpi,
VP9_COMMON *cm,
int mbrow,
int mbcol,
MACROBLOCK *x,
MACROBLOCKD *xd,
TOKENEXTRA **tp) {
int i;
int map_index;
int mb_row, mb_col;
int recon_yoffset, recon_uvoffset;
int ref_fb_idx = cm->lst_fb_idx;
int dst_fb_idx = cm->new_fb_idx;
int recon_y_stride = cm->yv12_fb[ref_fb_idx].y_stride;
int recon_uv_stride = cm->yv12_fb[ref_fb_idx].uv_stride;
int row_delta[4] = { 0, +1, 0, -1};
int col_delta[4] = { +1, -1, +1, +1};
mb_row = mbrow;
mb_col = mbcol;
/* Encode MBs in raster order within the SB */
for (i = 0; i < 4; i++) {
int dy = row_delta[i];
int dx = col_delta[i];
int offset_extended = dy * xd->mode_info_stride + dx;
int offset_unextended = dy * cm->mb_cols + dx;
MB_MODE_INFO * mbmi = &xd->mode_info_context->mbmi;
if ((mb_row >= cm->mb_rows) || (mb_col >= cm->mb_cols)) {
// MB lies outside frame, move on
mb_row += dy;
mb_col += dx;
x->src.y_buffer += 16 * (dx + dy * x->src.y_stride);
x->src.u_buffer += 8 * (dx + dy * x->src.uv_stride);
x->src.v_buffer += 8 * (dx + dy * x->src.uv_stride);
x->gf_active_ptr += offset_unextended;
x->partition_info += offset_extended;
xd->mode_info_context += offset_extended;
xd->prev_mode_info_context += offset_extended;
#if CONFIG_DEBUG
assert((xd->prev_mode_info_context - cpi->common.prev_mip) ==
(xd->mode_info_context - cpi->common.mip));
#endif
continue;
}
xd->mb_index = i;
// Restore MB state to that when it was picked
#if CONFIG_SUPERBLOCKS
if (xd->mode_info_context->mbmi.encoded_as_sb) {
update_state(cpi, x, &x->sb_context[i]);
cpi->sb_count++;
} else
#endif
update_state(cpi, x, &x->mb_context[i]);
map_index = (mb_row * cpi->common.mb_cols) + mb_col;
x->mb_activity_ptr = &cpi->mb_activity_map[map_index];
// reset above block coeffs
xd->above_context = cm->above_context + mb_col;
xd->left_context = cm->left_context + (i >> 1);
// Set up distance of MB to edge of the frame in 1/8th pel units
1051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120
// Set up limit values for MV components to prevent them from
// extending beyond the UMV borders assuming 32x32 block size
x->mv_row_min = -((mb_row * 16) + VP9BORDERINPIXELS - VP9_INTERP_EXTEND);
x->mv_col_min = -((mb_col * 16) + VP9BORDERINPIXELS - VP9_INTERP_EXTEND);
xd->mb_to_top_edge = -((mb_row * 16) << 3);
xd->mb_to_left_edge = -((mb_col * 16) << 3);
#if CONFIG_SUPERBLOCKS
if (xd->mode_info_context->mbmi.encoded_as_sb) {
x->mv_row_max = ((cm->mb_rows - mb_row) * 16 +
(VP9BORDERINPIXELS - 32 - VP9_INTERP_EXTEND));
x->mv_col_max = ((cm->mb_cols - mb_col) * 16 +
(VP9BORDERINPIXELS - 32 - VP9_INTERP_EXTEND));
xd->mb_to_bottom_edge = ((cm->mb_rows - 2 - mb_row) * 16) << 3;
xd->mb_to_right_edge = ((cm->mb_cols - 2 - mb_col) * 16) << 3;
} else {
#endif
x->mv_row_max = ((cm->mb_rows - mb_row) * 16 +
(VP9BORDERINPIXELS - 16 - VP9_INTERP_EXTEND));
x->mv_col_max = ((cm->mb_cols - mb_col) * 16 +
(VP9BORDERINPIXELS - 16 - VP9_INTERP_EXTEND));
xd->mb_to_bottom_edge = ((cm->mb_rows - 1 - mb_row) * 16) << 3;
xd->mb_to_right_edge = ((cm->mb_cols - 1 - mb_col) * 16) << 3;
#if CONFIG_SUPERBLOCKS
}
#endif
xd->up_available = (mb_row != 0);
xd->left_available = (mb_col != 0);
recon_yoffset = (mb_row * recon_y_stride * 16) + (mb_col * 16);
recon_uvoffset = (mb_row * recon_uv_stride * 8) + (mb_col * 8);
xd->dst.y_buffer = cm->yv12_fb[dst_fb_idx].y_buffer + recon_yoffset;
xd->dst.u_buffer = cm->yv12_fb[dst_fb_idx].u_buffer + recon_uvoffset;
xd->dst.v_buffer = cm->yv12_fb[dst_fb_idx].v_buffer + recon_uvoffset;
#if !CONFIG_SUPERBLOCKS
// Copy current MB to a work buffer
vp9_copy_mem16x16(x->src.y_buffer, x->src.y_stride, x->thismb, 16);
#endif
if (cpi->oxcf.tuning == VP8_TUNE_SSIM)
vp9_activity_masking(cpi, x);
// Is segmentation enabled
if (xd->segmentation_enabled) {
vp9_mb_init_quantizer(cpi, x);
}
x->active_ptr = cpi->active_map + map_index;
cpi->update_context = 0;
#if CONFIG_SUPERBLOCKS
if (!xd->mode_info_context->mbmi.encoded_as_sb)
#endif
vp9_intra_prediction_down_copy(xd);
if (cm->frame_type == KEY_FRAME) {
#if CONFIG_SUPERBLOCKS
if (xd->mode_info_context->mbmi.encoded_as_sb)
encode_superblock(cpi, x, tp, recon_yoffset, recon_uvoffset,
mb_col, mb_row);
else
#endif
encode_macroblock(cpi, x, tp, recon_yoffset, recon_uvoffset, 1,