vp9_encodeframe.c 71.33 KiB
/*
 *  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/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/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_DEBUG
int enc_debug = 0;
#endif
void vp9_select_interp_filter_type(VP9_COMP *cpi);
static 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
7172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140
// 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->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(sortlist, vpx_calloc(sizeof(unsigned int),
141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210
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 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
211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280
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; 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;
281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350
// 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; } // 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; #if CONFIG_DEBUG || CONFIG_INTERNAL_STATS MB_PREDICTION_MODE mb_mode = mi->mbmi.mode; #endif 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); #if CONFIG_DEBUG assert(mb_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); #endif
351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420
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; } } } 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]; } if (mbmi->ref_frame[0] != INTRA_FRAME && mbmi->sb_type < BLOCK_SIZE_SB8X8) { *x->partition_info = ctx->partition_info; mbmi->mv[0].as_int = x->partition_info->bmi[3].mv.as_int; mbmi->mv[1].as_int = x->partition_info->bmi[3].second_mv.as_int; } x->skip = ctx->skip; if (!output_enabled) return; if (!vp9_segfeature_active(xd, 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[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)) ?
421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490
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->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; } if (cpi->common.mcomp_filter_type == SWITCHABLE && is_inter_mode(mbmi->mode)) { ++cpi->common.fc.switchable_interp_count [vp9_get_pred_context(&cpi->common, xd, PRED_SWITCHABLE_INTERP)] [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; } } static unsigned find_seg_id(uint8_t *buf, BLOCK_SIZE_TYPE bsize, int start_y, int height, int start_x, int width) { const int bw = 1 << mi_width_log2(bsize), bh = 1 << mi_height_log2(bsize); const int end_x = MIN(start_x + bw, width); const int end_y = MIN(start_y + bh, height); int x, y; unsigned seg_id = -1; buf += width * start_y; for (y = start_y; y < end_y; y++, buf += width) { for (x = start_x; x < end_x; x++) { seg_id = MIN(seg_id, buf[x]); } } return seg_id; } 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,