1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libvpx/vp8/encoder/encodeframe.c Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,1404 @@
1.4 +/*
1.5 + * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
1.6 + *
1.7 + * Use of this source code is governed by a BSD-style license
1.8 + * that can be found in the LICENSE file in the root of the source
1.9 + * tree. An additional intellectual property rights grant can be found
1.10 + * in the file PATENTS. All contributing project authors may
1.11 + * be found in the AUTHORS file in the root of the source tree.
1.12 + */
1.13 +
1.14 +
1.15 +#include "vpx_config.h"
1.16 +#include "vp8_rtcd.h"
1.17 +#include "encodemb.h"
1.18 +#include "encodemv.h"
1.19 +#include "vp8/common/common.h"
1.20 +#include "onyx_int.h"
1.21 +#include "vp8/common/extend.h"
1.22 +#include "vp8/common/entropymode.h"
1.23 +#include "vp8/common/quant_common.h"
1.24 +#include "segmentation.h"
1.25 +#include "vp8/common/setupintrarecon.h"
1.26 +#include "encodeintra.h"
1.27 +#include "vp8/common/reconinter.h"
1.28 +#include "rdopt.h"
1.29 +#include "pickinter.h"
1.30 +#include "vp8/common/findnearmv.h"
1.31 +#include <stdio.h>
1.32 +#include <limits.h>
1.33 +#include "vp8/common/invtrans.h"
1.34 +#include "vpx_ports/vpx_timer.h"
1.35 +#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
1.36 +#include "bitstream.h"
1.37 +#endif
1.38 +#include "encodeframe.h"
1.39 +
1.40 +extern void vp8_stuff_mb(VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t) ;
1.41 +extern void vp8_calc_ref_frame_costs(int *ref_frame_cost,
1.42 + int prob_intra,
1.43 + int prob_last,
1.44 + int prob_garf
1.45 + );
1.46 +extern void vp8_convert_rfct_to_prob(VP8_COMP *const cpi);
1.47 +extern void vp8cx_initialize_me_consts(VP8_COMP *cpi, int QIndex);
1.48 +extern void vp8_auto_select_speed(VP8_COMP *cpi);
1.49 +extern void vp8cx_init_mbrthread_data(VP8_COMP *cpi,
1.50 + MACROBLOCK *x,
1.51 + MB_ROW_COMP *mbr_ei,
1.52 + int count);
1.53 +static void adjust_act_zbin( VP8_COMP *cpi, MACROBLOCK *x );
1.54 +
1.55 +#ifdef MODE_STATS
1.56 +unsigned int inter_y_modes[10] = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
1.57 +unsigned int inter_uv_modes[4] = {0, 0, 0, 0};
1.58 +unsigned int inter_b_modes[15] = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
1.59 +unsigned int y_modes[5] = {0, 0, 0, 0, 0};
1.60 +unsigned int uv_modes[4] = {0, 0, 0, 0};
1.61 +unsigned int b_modes[14] = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
1.62 +#endif
1.63 +
1.64 +
1.65 +/* activity_avg must be positive, or flat regions could get a zero weight
1.66 + * (infinite lambda), which confounds analysis.
1.67 + * This also avoids the need for divide by zero checks in
1.68 + * vp8_activity_masking().
1.69 + */
1.70 +#define VP8_ACTIVITY_AVG_MIN (64)
1.71 +
1.72 +/* This is used as a reference when computing the source variance for the
1.73 + * purposes of activity masking.
1.74 + * Eventually this should be replaced by custom no-reference routines,
1.75 + * which will be faster.
1.76 + */
1.77 +static const unsigned char VP8_VAR_OFFS[16]=
1.78 +{
1.79 + 128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128
1.80 +};
1.81 +
1.82 +
1.83 +/* Original activity measure from Tim T's code. */
1.84 +static unsigned int tt_activity_measure( VP8_COMP *cpi, MACROBLOCK *x )
1.85 +{
1.86 + unsigned int act;
1.87 + unsigned int sse;
1.88 + /* TODO: This could also be done over smaller areas (8x8), but that would
1.89 + * require extensive changes elsewhere, as lambda is assumed to be fixed
1.90 + * over an entire MB in most of the code.
1.91 + * Another option is to compute four 8x8 variances, and pick a single
1.92 + * lambda using a non-linear combination (e.g., the smallest, or second
1.93 + * smallest, etc.).
1.94 + */
1.95 + act = vp8_variance16x16(x->src.y_buffer,
1.96 + x->src.y_stride, VP8_VAR_OFFS, 0, &sse);
1.97 + act = act<<4;
1.98 +
1.99 + /* If the region is flat, lower the activity some more. */
1.100 + if (act < 8<<12)
1.101 + act = act < 5<<12 ? act : 5<<12;
1.102 +
1.103 + return act;
1.104 +}
1.105 +
1.106 +/* Stub for alternative experimental activity measures. */
1.107 +static unsigned int alt_activity_measure( VP8_COMP *cpi,
1.108 + MACROBLOCK *x, int use_dc_pred )
1.109 +{
1.110 + return vp8_encode_intra(cpi,x, use_dc_pred);
1.111 +}
1.112 +
1.113 +
1.114 +/* Measure the activity of the current macroblock
1.115 + * What we measure here is TBD so abstracted to this function
1.116 + */
1.117 +#define ALT_ACT_MEASURE 1
1.118 +static unsigned int mb_activity_measure( VP8_COMP *cpi, MACROBLOCK *x,
1.119 + int mb_row, int mb_col)
1.120 +{
1.121 + unsigned int mb_activity;
1.122 +
1.123 + if ( ALT_ACT_MEASURE )
1.124 + {
1.125 + int use_dc_pred = (mb_col || mb_row) && (!mb_col || !mb_row);
1.126 +
1.127 + /* Or use and alternative. */
1.128 + mb_activity = alt_activity_measure( cpi, x, use_dc_pred );
1.129 + }
1.130 + else
1.131 + {
1.132 + /* Original activity measure from Tim T's code. */
1.133 + mb_activity = tt_activity_measure( cpi, x );
1.134 + }
1.135 +
1.136 + if ( mb_activity < VP8_ACTIVITY_AVG_MIN )
1.137 + mb_activity = VP8_ACTIVITY_AVG_MIN;
1.138 +
1.139 + return mb_activity;
1.140 +}
1.141 +
1.142 +/* Calculate an "average" mb activity value for the frame */
1.143 +#define ACT_MEDIAN 0
1.144 +static void calc_av_activity( VP8_COMP *cpi, int64_t activity_sum )
1.145 +{
1.146 +#if ACT_MEDIAN
1.147 + /* Find median: Simple n^2 algorithm for experimentation */
1.148 + {
1.149 + unsigned int median;
1.150 + unsigned int i,j;
1.151 + unsigned int * sortlist;
1.152 + unsigned int tmp;
1.153 +
1.154 + /* Create a list to sort to */
1.155 + CHECK_MEM_ERROR(sortlist,
1.156 + vpx_calloc(sizeof(unsigned int),
1.157 + cpi->common.MBs));
1.158 +
1.159 + /* Copy map to sort list */
1.160 + vpx_memcpy( sortlist, cpi->mb_activity_map,
1.161 + sizeof(unsigned int) * cpi->common.MBs );
1.162 +
1.163 +
1.164 + /* Ripple each value down to its correct position */
1.165 + for ( i = 1; i < cpi->common.MBs; i ++ )
1.166 + {
1.167 + for ( j = i; j > 0; j -- )
1.168 + {
1.169 + if ( sortlist[j] < sortlist[j-1] )
1.170 + {
1.171 + /* Swap values */
1.172 + tmp = sortlist[j-1];
1.173 + sortlist[j-1] = sortlist[j];
1.174 + sortlist[j] = tmp;
1.175 + }
1.176 + else
1.177 + break;
1.178 + }
1.179 + }
1.180 +
1.181 + /* Even number MBs so estimate median as mean of two either side. */
1.182 + median = ( 1 + sortlist[cpi->common.MBs >> 1] +
1.183 + sortlist[(cpi->common.MBs >> 1) + 1] ) >> 1;
1.184 +
1.185 + cpi->activity_avg = median;
1.186 +
1.187 + vpx_free(sortlist);
1.188 + }
1.189 +#else
1.190 + /* Simple mean for now */
1.191 + cpi->activity_avg = (unsigned int)(activity_sum/cpi->common.MBs);
1.192 +#endif
1.193 +
1.194 + if (cpi->activity_avg < VP8_ACTIVITY_AVG_MIN)
1.195 + cpi->activity_avg = VP8_ACTIVITY_AVG_MIN;
1.196 +
1.197 + /* Experimental code: return fixed value normalized for several clips */
1.198 + if ( ALT_ACT_MEASURE )
1.199 + cpi->activity_avg = 100000;
1.200 +}
1.201 +
1.202 +#define USE_ACT_INDEX 0
1.203 +#define OUTPUT_NORM_ACT_STATS 0
1.204 +
1.205 +#if USE_ACT_INDEX
1.206 +/* Calculate and activity index for each mb */
1.207 +static void calc_activity_index( VP8_COMP *cpi, MACROBLOCK *x )
1.208 +{
1.209 + VP8_COMMON *const cm = & cpi->common;
1.210 + int mb_row, mb_col;
1.211 +
1.212 + int64_t act;
1.213 + int64_t a;
1.214 + int64_t b;
1.215 +
1.216 +#if OUTPUT_NORM_ACT_STATS
1.217 + FILE *f = fopen("norm_act.stt", "a");
1.218 + fprintf(f, "\n%12d\n", cpi->activity_avg );
1.219 +#endif
1.220 +
1.221 + /* Reset pointers to start of activity map */
1.222 + x->mb_activity_ptr = cpi->mb_activity_map;
1.223 +
1.224 + /* Calculate normalized mb activity number. */
1.225 + for (mb_row = 0; mb_row < cm->mb_rows; mb_row++)
1.226 + {
1.227 + /* for each macroblock col in image */
1.228 + for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
1.229 + {
1.230 + /* Read activity from the map */
1.231 + act = *(x->mb_activity_ptr);
1.232 +
1.233 + /* Calculate a normalized activity number */
1.234 + a = act + 4*cpi->activity_avg;
1.235 + b = 4*act + cpi->activity_avg;
1.236 +
1.237 + if ( b >= a )
1.238 + *(x->activity_ptr) = (int)((b + (a>>1))/a) - 1;
1.239 + else
1.240 + *(x->activity_ptr) = 1 - (int)((a + (b>>1))/b);
1.241 +
1.242 +#if OUTPUT_NORM_ACT_STATS
1.243 + fprintf(f, " %6d", *(x->mb_activity_ptr));
1.244 +#endif
1.245 + /* Increment activity map pointers */
1.246 + x->mb_activity_ptr++;
1.247 + }
1.248 +
1.249 +#if OUTPUT_NORM_ACT_STATS
1.250 + fprintf(f, "\n");
1.251 +#endif
1.252 +
1.253 + }
1.254 +
1.255 +#if OUTPUT_NORM_ACT_STATS
1.256 + fclose(f);
1.257 +#endif
1.258 +
1.259 +}
1.260 +#endif
1.261 +
1.262 +/* Loop through all MBs. Note activity of each, average activity and
1.263 + * calculate a normalized activity for each
1.264 + */
1.265 +static void build_activity_map( VP8_COMP *cpi )
1.266 +{
1.267 + MACROBLOCK *const x = & cpi->mb;
1.268 + MACROBLOCKD *xd = &x->e_mbd;
1.269 + VP8_COMMON *const cm = & cpi->common;
1.270 +
1.271 +#if ALT_ACT_MEASURE
1.272 + YV12_BUFFER_CONFIG *new_yv12 = &cm->yv12_fb[cm->new_fb_idx];
1.273 + int recon_yoffset;
1.274 + int recon_y_stride = new_yv12->y_stride;
1.275 +#endif
1.276 +
1.277 + int mb_row, mb_col;
1.278 + unsigned int mb_activity;
1.279 + int64_t activity_sum = 0;
1.280 +
1.281 + /* for each macroblock row in image */
1.282 + for (mb_row = 0; mb_row < cm->mb_rows; mb_row++)
1.283 + {
1.284 +#if ALT_ACT_MEASURE
1.285 + /* reset above block coeffs */
1.286 + xd->up_available = (mb_row != 0);
1.287 + recon_yoffset = (mb_row * recon_y_stride * 16);
1.288 +#endif
1.289 + /* for each macroblock col in image */
1.290 + for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
1.291 + {
1.292 +#if ALT_ACT_MEASURE
1.293 + xd->dst.y_buffer = new_yv12->y_buffer + recon_yoffset;
1.294 + xd->left_available = (mb_col != 0);
1.295 + recon_yoffset += 16;
1.296 +#endif
1.297 + /* Copy current mb to a buffer */
1.298 + vp8_copy_mem16x16(x->src.y_buffer, x->src.y_stride, x->thismb, 16);
1.299 +
1.300 + /* measure activity */
1.301 + mb_activity = mb_activity_measure( cpi, x, mb_row, mb_col );
1.302 +
1.303 + /* Keep frame sum */
1.304 + activity_sum += mb_activity;
1.305 +
1.306 + /* Store MB level activity details. */
1.307 + *x->mb_activity_ptr = mb_activity;
1.308 +
1.309 + /* Increment activity map pointer */
1.310 + x->mb_activity_ptr++;
1.311 +
1.312 + /* adjust to the next column of source macroblocks */
1.313 + x->src.y_buffer += 16;
1.314 + }
1.315 +
1.316 +
1.317 + /* adjust to the next row of mbs */
1.318 + x->src.y_buffer += 16 * x->src.y_stride - 16 * cm->mb_cols;
1.319 +
1.320 +#if ALT_ACT_MEASURE
1.321 + /* extend the recon for intra prediction */
1.322 + vp8_extend_mb_row(new_yv12, xd->dst.y_buffer + 16,
1.323 + xd->dst.u_buffer + 8, xd->dst.v_buffer + 8);
1.324 +#endif
1.325 +
1.326 + }
1.327 +
1.328 + /* Calculate an "average" MB activity */
1.329 + calc_av_activity(cpi, activity_sum);
1.330 +
1.331 +#if USE_ACT_INDEX
1.332 + /* Calculate an activity index number of each mb */
1.333 + calc_activity_index( cpi, x );
1.334 +#endif
1.335 +
1.336 +}
1.337 +
1.338 +/* Macroblock activity masking */
1.339 +void vp8_activity_masking(VP8_COMP *cpi, MACROBLOCK *x)
1.340 +{
1.341 +#if USE_ACT_INDEX
1.342 + x->rdmult += *(x->mb_activity_ptr) * (x->rdmult >> 2);
1.343 + x->errorperbit = x->rdmult * 100 /(110 * x->rddiv);
1.344 + x->errorperbit += (x->errorperbit==0);
1.345 +#else
1.346 + int64_t a;
1.347 + int64_t b;
1.348 + int64_t act = *(x->mb_activity_ptr);
1.349 +
1.350 + /* Apply the masking to the RD multiplier. */
1.351 + a = act + (2*cpi->activity_avg);
1.352 + b = (2*act) + cpi->activity_avg;
1.353 +
1.354 + x->rdmult = (unsigned int)(((int64_t)x->rdmult*b + (a>>1))/a);
1.355 + x->errorperbit = x->rdmult * 100 /(110 * x->rddiv);
1.356 + x->errorperbit += (x->errorperbit==0);
1.357 +#endif
1.358 +
1.359 + /* Activity based Zbin adjustment */
1.360 + adjust_act_zbin(cpi, x);
1.361 +}
1.362 +
1.363 +static
1.364 +void encode_mb_row(VP8_COMP *cpi,
1.365 + VP8_COMMON *cm,
1.366 + int mb_row,
1.367 + MACROBLOCK *x,
1.368 + MACROBLOCKD *xd,
1.369 + TOKENEXTRA **tp,
1.370 + int *segment_counts,
1.371 + int *totalrate)
1.372 +{
1.373 + int recon_yoffset, recon_uvoffset;
1.374 + int mb_col;
1.375 + int ref_fb_idx = cm->lst_fb_idx;
1.376 + int dst_fb_idx = cm->new_fb_idx;
1.377 + int recon_y_stride = cm->yv12_fb[ref_fb_idx].y_stride;
1.378 + int recon_uv_stride = cm->yv12_fb[ref_fb_idx].uv_stride;
1.379 + int map_index = (mb_row * cpi->common.mb_cols);
1.380 +
1.381 +#if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
1.382 + const int num_part = (1 << cm->multi_token_partition);
1.383 + TOKENEXTRA * tp_start = cpi->tok;
1.384 + vp8_writer *w;
1.385 +#endif
1.386 +
1.387 +#if CONFIG_MULTITHREAD
1.388 + const int nsync = cpi->mt_sync_range;
1.389 + const int rightmost_col = cm->mb_cols + nsync;
1.390 + volatile const int *last_row_current_mb_col;
1.391 + volatile int *current_mb_col = &cpi->mt_current_mb_col[mb_row];
1.392 +
1.393 + if ((cpi->b_multi_threaded != 0) && (mb_row != 0))
1.394 + last_row_current_mb_col = &cpi->mt_current_mb_col[mb_row - 1];
1.395 + else
1.396 + last_row_current_mb_col = &rightmost_col;
1.397 +#endif
1.398 +
1.399 +#if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
1.400 + if(num_part > 1)
1.401 + w= &cpi->bc[1 + (mb_row % num_part)];
1.402 + else
1.403 + w = &cpi->bc[1];
1.404 +#endif
1.405 +
1.406 + /* reset above block coeffs */
1.407 + xd->above_context = cm->above_context;
1.408 +
1.409 + xd->up_available = (mb_row != 0);
1.410 + recon_yoffset = (mb_row * recon_y_stride * 16);
1.411 + recon_uvoffset = (mb_row * recon_uv_stride * 8);
1.412 +
1.413 + cpi->tplist[mb_row].start = *tp;
1.414 + /* printf("Main mb_row = %d\n", mb_row); */
1.415 +
1.416 + /* Distance of Mb to the top & bottom edges, specified in 1/8th pel
1.417 + * units as they are always compared to values that are in 1/8th pel
1.418 + */
1.419 + xd->mb_to_top_edge = -((mb_row * 16) << 3);
1.420 + xd->mb_to_bottom_edge = ((cm->mb_rows - 1 - mb_row) * 16) << 3;
1.421 +
1.422 + /* Set up limit values for vertical motion vector components
1.423 + * to prevent them extending beyond the UMV borders
1.424 + */
1.425 + x->mv_row_min = -((mb_row * 16) + (VP8BORDERINPIXELS - 16));
1.426 + x->mv_row_max = ((cm->mb_rows - 1 - mb_row) * 16)
1.427 + + (VP8BORDERINPIXELS - 16);
1.428 +
1.429 + /* Set the mb activity pointer to the start of the row. */
1.430 + x->mb_activity_ptr = &cpi->mb_activity_map[map_index];
1.431 +
1.432 + /* for each macroblock col in image */
1.433 + for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
1.434 + {
1.435 +
1.436 +#if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
1.437 + *tp = cpi->tok;
1.438 +#endif
1.439 + /* Distance of Mb to the left & right edges, specified in
1.440 + * 1/8th pel units as they are always compared to values
1.441 + * that are in 1/8th pel units
1.442 + */
1.443 + xd->mb_to_left_edge = -((mb_col * 16) << 3);
1.444 + xd->mb_to_right_edge = ((cm->mb_cols - 1 - mb_col) * 16) << 3;
1.445 +
1.446 + /* Set up limit values for horizontal motion vector components
1.447 + * to prevent them extending beyond the UMV borders
1.448 + */
1.449 + x->mv_col_min = -((mb_col * 16) + (VP8BORDERINPIXELS - 16));
1.450 + x->mv_col_max = ((cm->mb_cols - 1 - mb_col) * 16)
1.451 + + (VP8BORDERINPIXELS - 16);
1.452 +
1.453 + xd->dst.y_buffer = cm->yv12_fb[dst_fb_idx].y_buffer + recon_yoffset;
1.454 + xd->dst.u_buffer = cm->yv12_fb[dst_fb_idx].u_buffer + recon_uvoffset;
1.455 + xd->dst.v_buffer = cm->yv12_fb[dst_fb_idx].v_buffer + recon_uvoffset;
1.456 + xd->left_available = (mb_col != 0);
1.457 +
1.458 + x->rddiv = cpi->RDDIV;
1.459 + x->rdmult = cpi->RDMULT;
1.460 +
1.461 + /* Copy current mb to a buffer */
1.462 + vp8_copy_mem16x16(x->src.y_buffer, x->src.y_stride, x->thismb, 16);
1.463 +
1.464 +#if CONFIG_MULTITHREAD
1.465 + if (cpi->b_multi_threaded != 0)
1.466 + {
1.467 + *current_mb_col = mb_col - 1; /* set previous MB done */
1.468 +
1.469 + if ((mb_col & (nsync - 1)) == 0)
1.470 + {
1.471 + while (mb_col > (*last_row_current_mb_col - nsync))
1.472 + {
1.473 + x86_pause_hint();
1.474 + thread_sleep(0);
1.475 + }
1.476 + }
1.477 + }
1.478 +#endif
1.479 +
1.480 + if(cpi->oxcf.tuning == VP8_TUNE_SSIM)
1.481 + vp8_activity_masking(cpi, x);
1.482 +
1.483 + /* Is segmentation enabled */
1.484 + /* MB level adjustment to quantizer */
1.485 + if (xd->segmentation_enabled)
1.486 + {
1.487 + /* Code to set segment id in xd->mbmi.segment_id for current MB
1.488 + * (with range checking)
1.489 + */
1.490 + if (cpi->segmentation_map[map_index+mb_col] <= 3)
1.491 + xd->mode_info_context->mbmi.segment_id = cpi->segmentation_map[map_index+mb_col];
1.492 + else
1.493 + xd->mode_info_context->mbmi.segment_id = 0;
1.494 +
1.495 + vp8cx_mb_init_quantizer(cpi, x, 1);
1.496 + }
1.497 + else
1.498 + /* Set to Segment 0 by default */
1.499 + xd->mode_info_context->mbmi.segment_id = 0;
1.500 +
1.501 + x->active_ptr = cpi->active_map + map_index + mb_col;
1.502 +
1.503 + if (cm->frame_type == KEY_FRAME)
1.504 + {
1.505 + *totalrate += vp8cx_encode_intra_macroblock(cpi, x, tp);
1.506 +#ifdef MODE_STATS
1.507 + y_modes[xd->mbmi.mode] ++;
1.508 +#endif
1.509 + }
1.510 + else
1.511 + {
1.512 + *totalrate += vp8cx_encode_inter_macroblock(cpi, x, tp, recon_yoffset, recon_uvoffset, mb_row, mb_col);
1.513 +
1.514 +#ifdef MODE_STATS
1.515 + inter_y_modes[xd->mbmi.mode] ++;
1.516 +
1.517 + if (xd->mbmi.mode == SPLITMV)
1.518 + {
1.519 + int b;
1.520 +
1.521 + for (b = 0; b < xd->mbmi.partition_count; b++)
1.522 + {
1.523 + inter_b_modes[x->partition->bmi[b].mode] ++;
1.524 + }
1.525 + }
1.526 +
1.527 +#endif
1.528 +
1.529 + /* Special case code for cyclic refresh
1.530 + * If cyclic update enabled then copy xd->mbmi.segment_id; (which
1.531 + * may have been updated based on mode during
1.532 + * vp8cx_encode_inter_macroblock()) back into the global
1.533 + * segmentation map
1.534 + */
1.535 + if ((cpi->current_layer == 0) &&
1.536 + (cpi->cyclic_refresh_mode_enabled &&
1.537 + xd->segmentation_enabled))
1.538 + {
1.539 + cpi->segmentation_map[map_index+mb_col] = xd->mode_info_context->mbmi.segment_id;
1.540 +
1.541 + /* If the block has been refreshed mark it as clean (the
1.542 + * magnitude of the -ve influences how long it will be before
1.543 + * we consider another refresh):
1.544 + * Else if it was coded (last frame 0,0) and has not already
1.545 + * been refreshed then mark it as a candidate for cleanup
1.546 + * next time (marked 0) else mark it as dirty (1).
1.547 + */
1.548 + if (xd->mode_info_context->mbmi.segment_id)
1.549 + cpi->cyclic_refresh_map[map_index+mb_col] = -1;
1.550 + else if ((xd->mode_info_context->mbmi.mode == ZEROMV) && (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME))
1.551 + {
1.552 + if (cpi->cyclic_refresh_map[map_index+mb_col] == 1)
1.553 + cpi->cyclic_refresh_map[map_index+mb_col] = 0;
1.554 + }
1.555 + else
1.556 + cpi->cyclic_refresh_map[map_index+mb_col] = 1;
1.557 +
1.558 + }
1.559 + }
1.560 +
1.561 + cpi->tplist[mb_row].stop = *tp;
1.562 +
1.563 +#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
1.564 + /* pack tokens for this MB */
1.565 + {
1.566 + int tok_count = *tp - tp_start;
1.567 + pack_tokens(w, tp_start, tok_count);
1.568 + }
1.569 +#endif
1.570 + /* Increment pointer into gf usage flags structure. */
1.571 + x->gf_active_ptr++;
1.572 +
1.573 + /* Increment the activity mask pointers. */
1.574 + x->mb_activity_ptr++;
1.575 +
1.576 + /* adjust to the next column of macroblocks */
1.577 + x->src.y_buffer += 16;
1.578 + x->src.u_buffer += 8;
1.579 + x->src.v_buffer += 8;
1.580 +
1.581 + recon_yoffset += 16;
1.582 + recon_uvoffset += 8;
1.583 +
1.584 + /* Keep track of segment usage */
1.585 + segment_counts[xd->mode_info_context->mbmi.segment_id] ++;
1.586 +
1.587 + /* skip to next mb */
1.588 + xd->mode_info_context++;
1.589 + x->partition_info++;
1.590 + xd->above_context++;
1.591 + }
1.592 +
1.593 + /* extend the recon for intra prediction */
1.594 + vp8_extend_mb_row( &cm->yv12_fb[dst_fb_idx],
1.595 + xd->dst.y_buffer + 16,
1.596 + xd->dst.u_buffer + 8,
1.597 + xd->dst.v_buffer + 8);
1.598 +
1.599 +#if CONFIG_MULTITHREAD
1.600 + if (cpi->b_multi_threaded != 0)
1.601 + *current_mb_col = rightmost_col;
1.602 +#endif
1.603 +
1.604 + /* this is to account for the border */
1.605 + xd->mode_info_context++;
1.606 + x->partition_info++;
1.607 +}
1.608 +
1.609 +static void init_encode_frame_mb_context(VP8_COMP *cpi)
1.610 +{
1.611 + MACROBLOCK *const x = & cpi->mb;
1.612 + VP8_COMMON *const cm = & cpi->common;
1.613 + MACROBLOCKD *const xd = & x->e_mbd;
1.614 +
1.615 + /* GF active flags data structure */
1.616 + x->gf_active_ptr = (signed char *)cpi->gf_active_flags;
1.617 +
1.618 + /* Activity map pointer */
1.619 + x->mb_activity_ptr = cpi->mb_activity_map;
1.620 +
1.621 + x->act_zbin_adj = 0;
1.622 +
1.623 + x->partition_info = x->pi;
1.624 +
1.625 + xd->mode_info_context = cm->mi;
1.626 + xd->mode_info_stride = cm->mode_info_stride;
1.627 +
1.628 + xd->frame_type = cm->frame_type;
1.629 +
1.630 + /* reset intra mode contexts */
1.631 + if (cm->frame_type == KEY_FRAME)
1.632 + vp8_init_mbmode_probs(cm);
1.633 +
1.634 + /* Copy data over into macro block data structures. */
1.635 + x->src = * cpi->Source;
1.636 + xd->pre = cm->yv12_fb[cm->lst_fb_idx];
1.637 + xd->dst = cm->yv12_fb[cm->new_fb_idx];
1.638 +
1.639 + /* set up frame for intra coded blocks */
1.640 + vp8_setup_intra_recon(&cm->yv12_fb[cm->new_fb_idx]);
1.641 +
1.642 + vp8_build_block_offsets(x);
1.643 +
1.644 + xd->mode_info_context->mbmi.mode = DC_PRED;
1.645 + xd->mode_info_context->mbmi.uv_mode = DC_PRED;
1.646 +
1.647 + xd->left_context = &cm->left_context;
1.648 +
1.649 + x->mvc = cm->fc.mvc;
1.650 +
1.651 + vpx_memset(cm->above_context, 0,
1.652 + sizeof(ENTROPY_CONTEXT_PLANES) * cm->mb_cols);
1.653 +
1.654 + /* Special case treatment when GF and ARF are not sensible options
1.655 + * for reference
1.656 + */
1.657 + if (cpi->ref_frame_flags == VP8_LAST_FRAME)
1.658 + vp8_calc_ref_frame_costs(x->ref_frame_cost,
1.659 + cpi->prob_intra_coded,255,128);
1.660 + else if ((cpi->oxcf.number_of_layers > 1) &&
1.661 + (cpi->ref_frame_flags == VP8_GOLD_FRAME))
1.662 + vp8_calc_ref_frame_costs(x->ref_frame_cost,
1.663 + cpi->prob_intra_coded,1,255);
1.664 + else if ((cpi->oxcf.number_of_layers > 1) &&
1.665 + (cpi->ref_frame_flags == VP8_ALTR_FRAME))
1.666 + vp8_calc_ref_frame_costs(x->ref_frame_cost,
1.667 + cpi->prob_intra_coded,1,1);
1.668 + else
1.669 + vp8_calc_ref_frame_costs(x->ref_frame_cost,
1.670 + cpi->prob_intra_coded,
1.671 + cpi->prob_last_coded,
1.672 + cpi->prob_gf_coded);
1.673 +
1.674 + xd->fullpixel_mask = 0xffffffff;
1.675 + if(cm->full_pixel)
1.676 + xd->fullpixel_mask = 0xfffffff8;
1.677 +
1.678 + vp8_zero(x->coef_counts);
1.679 + vp8_zero(x->ymode_count);
1.680 + vp8_zero(x->uv_mode_count)
1.681 + x->prediction_error = 0;
1.682 + x->intra_error = 0;
1.683 + vp8_zero(x->count_mb_ref_frame_usage);
1.684 +}
1.685 +
1.686 +static void sum_coef_counts(MACROBLOCK *x, MACROBLOCK *x_thread)
1.687 +{
1.688 + int i = 0;
1.689 + do
1.690 + {
1.691 + int j = 0;
1.692 + do
1.693 + {
1.694 + int k = 0;
1.695 + do
1.696 + {
1.697 + /* at every context */
1.698 +
1.699 + /* calc probs and branch cts for this frame only */
1.700 + int t = 0; /* token/prob index */
1.701 +
1.702 + do
1.703 + {
1.704 + x->coef_counts [i][j][k][t] +=
1.705 + x_thread->coef_counts [i][j][k][t];
1.706 + }
1.707 + while (++t < ENTROPY_NODES);
1.708 + }
1.709 + while (++k < PREV_COEF_CONTEXTS);
1.710 + }
1.711 + while (++j < COEF_BANDS);
1.712 + }
1.713 + while (++i < BLOCK_TYPES);
1.714 +}
1.715 +
1.716 +void vp8_encode_frame(VP8_COMP *cpi)
1.717 +{
1.718 + int mb_row;
1.719 + MACROBLOCK *const x = & cpi->mb;
1.720 + VP8_COMMON *const cm = & cpi->common;
1.721 + MACROBLOCKD *const xd = & x->e_mbd;
1.722 + TOKENEXTRA *tp = cpi->tok;
1.723 + int segment_counts[MAX_MB_SEGMENTS];
1.724 + int totalrate;
1.725 +#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
1.726 + BOOL_CODER * bc = &cpi->bc[1]; /* bc[0] is for control partition */
1.727 + const int num_part = (1 << cm->multi_token_partition);
1.728 +#endif
1.729 +
1.730 + vpx_memset(segment_counts, 0, sizeof(segment_counts));
1.731 + totalrate = 0;
1.732 +
1.733 + if (cpi->compressor_speed == 2)
1.734 + {
1.735 + if (cpi->oxcf.cpu_used < 0)
1.736 + cpi->Speed = -(cpi->oxcf.cpu_used);
1.737 + else
1.738 + vp8_auto_select_speed(cpi);
1.739 + }
1.740 +
1.741 + /* Functions setup for all frame types so we can use MC in AltRef */
1.742 + if(!cm->use_bilinear_mc_filter)
1.743 + {
1.744 + xd->subpixel_predict = vp8_sixtap_predict4x4;
1.745 + xd->subpixel_predict8x4 = vp8_sixtap_predict8x4;
1.746 + xd->subpixel_predict8x8 = vp8_sixtap_predict8x8;
1.747 + xd->subpixel_predict16x16 = vp8_sixtap_predict16x16;
1.748 + }
1.749 + else
1.750 + {
1.751 + xd->subpixel_predict = vp8_bilinear_predict4x4;
1.752 + xd->subpixel_predict8x4 = vp8_bilinear_predict8x4;
1.753 + xd->subpixel_predict8x8 = vp8_bilinear_predict8x8;
1.754 + xd->subpixel_predict16x16 = vp8_bilinear_predict16x16;
1.755 + }
1.756 +
1.757 + cpi->mb.skip_true_count = 0;
1.758 + cpi->tok_count = 0;
1.759 +
1.760 +#if 0
1.761 + /* Experimental code */
1.762 + cpi->frame_distortion = 0;
1.763 + cpi->last_mb_distortion = 0;
1.764 +#endif
1.765 +
1.766 + xd->mode_info_context = cm->mi;
1.767 +
1.768 + vp8_zero(cpi->mb.MVcount);
1.769 +
1.770 + vp8cx_frame_init_quantizer(cpi);
1.771 +
1.772 + vp8_initialize_rd_consts(cpi, x,
1.773 + vp8_dc_quant(cm->base_qindex, cm->y1dc_delta_q));
1.774 +
1.775 + vp8cx_initialize_me_consts(cpi, cm->base_qindex);
1.776 +
1.777 + if(cpi->oxcf.tuning == VP8_TUNE_SSIM)
1.778 + {
1.779 + /* Initialize encode frame context. */
1.780 + init_encode_frame_mb_context(cpi);
1.781 +
1.782 + /* Build a frame level activity map */
1.783 + build_activity_map(cpi);
1.784 + }
1.785 +
1.786 + /* re-init encode frame context. */
1.787 + init_encode_frame_mb_context(cpi);
1.788 +
1.789 +#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
1.790 + {
1.791 + int i;
1.792 + for(i = 0; i < num_part; i++)
1.793 + {
1.794 + vp8_start_encode(&bc[i], cpi->partition_d[i + 1],
1.795 + cpi->partition_d_end[i + 1]);
1.796 + bc[i].error = &cm->error;
1.797 + }
1.798 + }
1.799 +
1.800 +#endif
1.801 +
1.802 + {
1.803 + struct vpx_usec_timer emr_timer;
1.804 + vpx_usec_timer_start(&emr_timer);
1.805 +
1.806 +#if CONFIG_MULTITHREAD
1.807 + if (cpi->b_multi_threaded)
1.808 + {
1.809 + int i;
1.810 +
1.811 + vp8cx_init_mbrthread_data(cpi, x, cpi->mb_row_ei,
1.812 + cpi->encoding_thread_count);
1.813 +
1.814 + for (i = 0; i < cm->mb_rows; i++)
1.815 + cpi->mt_current_mb_col[i] = -1;
1.816 +
1.817 + for (i = 0; i < cpi->encoding_thread_count; i++)
1.818 + {
1.819 + sem_post(&cpi->h_event_start_encoding[i]);
1.820 + }
1.821 +
1.822 + for (mb_row = 0; mb_row < cm->mb_rows; mb_row += (cpi->encoding_thread_count + 1))
1.823 + {
1.824 + vp8_zero(cm->left_context)
1.825 +
1.826 +#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
1.827 + tp = cpi->tok;
1.828 +#else
1.829 + tp = cpi->tok + mb_row * (cm->mb_cols * 16 * 24);
1.830 +#endif
1.831 +
1.832 + encode_mb_row(cpi, cm, mb_row, x, xd, &tp, segment_counts, &totalrate);
1.833 +
1.834 + /* adjust to the next row of mbs */
1.835 + x->src.y_buffer += 16 * x->src.y_stride * (cpi->encoding_thread_count + 1) - 16 * cm->mb_cols;
1.836 + x->src.u_buffer += 8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
1.837 + x->src.v_buffer += 8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
1.838 +
1.839 + xd->mode_info_context += xd->mode_info_stride * cpi->encoding_thread_count;
1.840 + x->partition_info += xd->mode_info_stride * cpi->encoding_thread_count;
1.841 + x->gf_active_ptr += cm->mb_cols * cpi->encoding_thread_count;
1.842 +
1.843 + if(mb_row == cm->mb_rows - 1)
1.844 + {
1.845 + sem_post(&cpi->h_event_end_encoding); /* signal frame encoding end */
1.846 + }
1.847 + }
1.848 +
1.849 + sem_wait(&cpi->h_event_end_encoding); /* wait for other threads to finish */
1.850 +
1.851 + for (mb_row = 0; mb_row < cm->mb_rows; mb_row ++)
1.852 + {
1.853 + cpi->tok_count += (unsigned int)
1.854 + (cpi->tplist[mb_row].stop - cpi->tplist[mb_row].start);
1.855 + }
1.856 +
1.857 + if (xd->segmentation_enabled)
1.858 + {
1.859 + int j;
1.860 +
1.861 + if (xd->segmentation_enabled)
1.862 + {
1.863 + for (i = 0; i < cpi->encoding_thread_count; i++)
1.864 + {
1.865 + for (j = 0; j < 4; j++)
1.866 + segment_counts[j] += cpi->mb_row_ei[i].segment_counts[j];
1.867 + }
1.868 + }
1.869 + }
1.870 +
1.871 + for (i = 0; i < cpi->encoding_thread_count; i++)
1.872 + {
1.873 + int mode_count;
1.874 + int c_idx;
1.875 + totalrate += cpi->mb_row_ei[i].totalrate;
1.876 +
1.877 + cpi->mb.skip_true_count += cpi->mb_row_ei[i].mb.skip_true_count;
1.878 +
1.879 + for(mode_count = 0; mode_count < VP8_YMODES; mode_count++)
1.880 + cpi->mb.ymode_count[mode_count] +=
1.881 + cpi->mb_row_ei[i].mb.ymode_count[mode_count];
1.882 +
1.883 + for(mode_count = 0; mode_count < VP8_UV_MODES; mode_count++)
1.884 + cpi->mb.uv_mode_count[mode_count] +=
1.885 + cpi->mb_row_ei[i].mb.uv_mode_count[mode_count];
1.886 +
1.887 + for(c_idx = 0; c_idx < MVvals; c_idx++)
1.888 + {
1.889 + cpi->mb.MVcount[0][c_idx] +=
1.890 + cpi->mb_row_ei[i].mb.MVcount[0][c_idx];
1.891 + cpi->mb.MVcount[1][c_idx] +=
1.892 + cpi->mb_row_ei[i].mb.MVcount[1][c_idx];
1.893 + }
1.894 +
1.895 + cpi->mb.prediction_error +=
1.896 + cpi->mb_row_ei[i].mb.prediction_error;
1.897 + cpi->mb.intra_error += cpi->mb_row_ei[i].mb.intra_error;
1.898 +
1.899 + for(c_idx = 0; c_idx < MAX_REF_FRAMES; c_idx++)
1.900 + cpi->mb.count_mb_ref_frame_usage[c_idx] +=
1.901 + cpi->mb_row_ei[i].mb.count_mb_ref_frame_usage[c_idx];
1.902 +
1.903 + for(c_idx = 0; c_idx < MAX_ERROR_BINS; c_idx++)
1.904 + cpi->mb.error_bins[c_idx] +=
1.905 + cpi->mb_row_ei[i].mb.error_bins[c_idx];
1.906 +
1.907 + /* add up counts for each thread */
1.908 + sum_coef_counts(x, &cpi->mb_row_ei[i].mb);
1.909 + }
1.910 +
1.911 + }
1.912 + else
1.913 +#endif
1.914 + {
1.915 +
1.916 + /* for each macroblock row in image */
1.917 + for (mb_row = 0; mb_row < cm->mb_rows; mb_row++)
1.918 + {
1.919 + vp8_zero(cm->left_context)
1.920 +
1.921 +#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
1.922 + tp = cpi->tok;
1.923 +#endif
1.924 +
1.925 + encode_mb_row(cpi, cm, mb_row, x, xd, &tp, segment_counts, &totalrate);
1.926 +
1.927 + /* adjust to the next row of mbs */
1.928 + x->src.y_buffer += 16 * x->src.y_stride - 16 * cm->mb_cols;
1.929 + x->src.u_buffer += 8 * x->src.uv_stride - 8 * cm->mb_cols;
1.930 + x->src.v_buffer += 8 * x->src.uv_stride - 8 * cm->mb_cols;
1.931 + }
1.932 +
1.933 + cpi->tok_count = (unsigned int)(tp - cpi->tok);
1.934 + }
1.935 +
1.936 +#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
1.937 + {
1.938 + int i;
1.939 + for(i = 0; i < num_part; i++)
1.940 + {
1.941 + vp8_stop_encode(&bc[i]);
1.942 + cpi->partition_sz[i+1] = bc[i].pos;
1.943 + }
1.944 + }
1.945 +#endif
1.946 +
1.947 + vpx_usec_timer_mark(&emr_timer);
1.948 + cpi->time_encode_mb_row += vpx_usec_timer_elapsed(&emr_timer);
1.949 + }
1.950 +
1.951 +
1.952 + // Work out the segment probabilities if segmentation is enabled
1.953 + // and needs to be updated
1.954 + if (xd->segmentation_enabled && xd->update_mb_segmentation_map)
1.955 + {
1.956 + int tot_count;
1.957 + int i;
1.958 +
1.959 + /* Set to defaults */
1.960 + vpx_memset(xd->mb_segment_tree_probs, 255 , sizeof(xd->mb_segment_tree_probs));
1.961 +
1.962 + tot_count = segment_counts[0] + segment_counts[1] + segment_counts[2] + segment_counts[3];
1.963 +
1.964 + if (tot_count)
1.965 + {
1.966 + xd->mb_segment_tree_probs[0] = ((segment_counts[0] + segment_counts[1]) * 255) / tot_count;
1.967 +
1.968 + tot_count = segment_counts[0] + segment_counts[1];
1.969 +
1.970 + if (tot_count > 0)
1.971 + {
1.972 + xd->mb_segment_tree_probs[1] = (segment_counts[0] * 255) / tot_count;
1.973 + }
1.974 +
1.975 + tot_count = segment_counts[2] + segment_counts[3];
1.976 +
1.977 + if (tot_count > 0)
1.978 + xd->mb_segment_tree_probs[2] = (segment_counts[2] * 255) / tot_count;
1.979 +
1.980 + /* Zero probabilities not allowed */
1.981 + for (i = 0; i < MB_FEATURE_TREE_PROBS; i ++)
1.982 + {
1.983 + if (xd->mb_segment_tree_probs[i] == 0)
1.984 + xd->mb_segment_tree_probs[i] = 1;
1.985 + }
1.986 + }
1.987 + }
1.988 +
1.989 + /* projected_frame_size in units of BYTES */
1.990 + cpi->projected_frame_size = totalrate >> 8;
1.991 +
1.992 + /* Make a note of the percentage MBs coded Intra. */
1.993 + if (cm->frame_type == KEY_FRAME)
1.994 + {
1.995 + cpi->this_frame_percent_intra = 100;
1.996 + }
1.997 + else
1.998 + {
1.999 + int tot_modes;
1.1000 +
1.1001 + tot_modes = cpi->mb.count_mb_ref_frame_usage[INTRA_FRAME]
1.1002 + + cpi->mb.count_mb_ref_frame_usage[LAST_FRAME]
1.1003 + + cpi->mb.count_mb_ref_frame_usage[GOLDEN_FRAME]
1.1004 + + cpi->mb.count_mb_ref_frame_usage[ALTREF_FRAME];
1.1005 +
1.1006 + if (tot_modes)
1.1007 + cpi->this_frame_percent_intra =
1.1008 + cpi->mb.count_mb_ref_frame_usage[INTRA_FRAME] * 100 / tot_modes;
1.1009 +
1.1010 + }
1.1011 +
1.1012 +#if ! CONFIG_REALTIME_ONLY
1.1013 + /* Adjust the projected reference frame usage probability numbers to
1.1014 + * reflect what we have just seen. This may be useful when we make
1.1015 + * multiple iterations of the recode loop rather than continuing to use
1.1016 + * values from the previous frame.
1.1017 + */
1.1018 + if ((cm->frame_type != KEY_FRAME) && ((cpi->oxcf.number_of_layers > 1) ||
1.1019 + (!cm->refresh_alt_ref_frame && !cm->refresh_golden_frame)))
1.1020 + {
1.1021 + vp8_convert_rfct_to_prob(cpi);
1.1022 + }
1.1023 +#endif
1.1024 +}
1.1025 +void vp8_setup_block_ptrs(MACROBLOCK *x)
1.1026 +{
1.1027 + int r, c;
1.1028 + int i;
1.1029 +
1.1030 + for (r = 0; r < 4; r++)
1.1031 + {
1.1032 + for (c = 0; c < 4; c++)
1.1033 + {
1.1034 + x->block[r*4+c].src_diff = x->src_diff + r * 4 * 16 + c * 4;
1.1035 + }
1.1036 + }
1.1037 +
1.1038 + for (r = 0; r < 2; r++)
1.1039 + {
1.1040 + for (c = 0; c < 2; c++)
1.1041 + {
1.1042 + x->block[16 + r*2+c].src_diff = x->src_diff + 256 + r * 4 * 8 + c * 4;
1.1043 + }
1.1044 + }
1.1045 +
1.1046 +
1.1047 + for (r = 0; r < 2; r++)
1.1048 + {
1.1049 + for (c = 0; c < 2; c++)
1.1050 + {
1.1051 + x->block[20 + r*2+c].src_diff = x->src_diff + 320 + r * 4 * 8 + c * 4;
1.1052 + }
1.1053 + }
1.1054 +
1.1055 + x->block[24].src_diff = x->src_diff + 384;
1.1056 +
1.1057 +
1.1058 + for (i = 0; i < 25; i++)
1.1059 + {
1.1060 + x->block[i].coeff = x->coeff + i * 16;
1.1061 + }
1.1062 +}
1.1063 +
1.1064 +void vp8_build_block_offsets(MACROBLOCK *x)
1.1065 +{
1.1066 + int block = 0;
1.1067 + int br, bc;
1.1068 +
1.1069 + vp8_build_block_doffsets(&x->e_mbd);
1.1070 +
1.1071 + /* y blocks */
1.1072 + x->thismb_ptr = &x->thismb[0];
1.1073 + for (br = 0; br < 4; br++)
1.1074 + {
1.1075 + for (bc = 0; bc < 4; bc++)
1.1076 + {
1.1077 + BLOCK *this_block = &x->block[block];
1.1078 + this_block->base_src = &x->thismb_ptr;
1.1079 + this_block->src_stride = 16;
1.1080 + this_block->src = 4 * br * 16 + 4 * bc;
1.1081 + ++block;
1.1082 + }
1.1083 + }
1.1084 +
1.1085 + /* u blocks */
1.1086 + for (br = 0; br < 2; br++)
1.1087 + {
1.1088 + for (bc = 0; bc < 2; bc++)
1.1089 + {
1.1090 + BLOCK *this_block = &x->block[block];
1.1091 + this_block->base_src = &x->src.u_buffer;
1.1092 + this_block->src_stride = x->src.uv_stride;
1.1093 + this_block->src = 4 * br * this_block->src_stride + 4 * bc;
1.1094 + ++block;
1.1095 + }
1.1096 + }
1.1097 +
1.1098 + /* v blocks */
1.1099 + for (br = 0; br < 2; br++)
1.1100 + {
1.1101 + for (bc = 0; bc < 2; bc++)
1.1102 + {
1.1103 + BLOCK *this_block = &x->block[block];
1.1104 + this_block->base_src = &x->src.v_buffer;
1.1105 + this_block->src_stride = x->src.uv_stride;
1.1106 + this_block->src = 4 * br * this_block->src_stride + 4 * bc;
1.1107 + ++block;
1.1108 + }
1.1109 + }
1.1110 +}
1.1111 +
1.1112 +static void sum_intra_stats(VP8_COMP *cpi, MACROBLOCK *x)
1.1113 +{
1.1114 + const MACROBLOCKD *xd = & x->e_mbd;
1.1115 + const MB_PREDICTION_MODE m = xd->mode_info_context->mbmi.mode;
1.1116 + const MB_PREDICTION_MODE uvm = xd->mode_info_context->mbmi.uv_mode;
1.1117 +
1.1118 +#ifdef MODE_STATS
1.1119 + const int is_key = cpi->common.frame_type == KEY_FRAME;
1.1120 +
1.1121 + ++ (is_key ? uv_modes : inter_uv_modes)[uvm];
1.1122 +
1.1123 + if (m == B_PRED)
1.1124 + {
1.1125 + unsigned int *const bct = is_key ? b_modes : inter_b_modes;
1.1126 +
1.1127 + int b = 0;
1.1128 +
1.1129 + do
1.1130 + {
1.1131 + ++ bct[xd->block[b].bmi.mode];
1.1132 + }
1.1133 + while (++b < 16);
1.1134 + }
1.1135 +
1.1136 +#endif
1.1137 +
1.1138 + ++x->ymode_count[m];
1.1139 + ++x->uv_mode_count[uvm];
1.1140 +
1.1141 +}
1.1142 +
1.1143 +/* Experimental stub function to create a per MB zbin adjustment based on
1.1144 + * some previously calculated measure of MB activity.
1.1145 + */
1.1146 +static void adjust_act_zbin( VP8_COMP *cpi, MACROBLOCK *x )
1.1147 +{
1.1148 +#if USE_ACT_INDEX
1.1149 + x->act_zbin_adj = *(x->mb_activity_ptr);
1.1150 +#else
1.1151 + int64_t a;
1.1152 + int64_t b;
1.1153 + int64_t act = *(x->mb_activity_ptr);
1.1154 +
1.1155 + /* Apply the masking to the RD multiplier. */
1.1156 + a = act + 4*cpi->activity_avg;
1.1157 + b = 4*act + cpi->activity_avg;
1.1158 +
1.1159 + if ( act > cpi->activity_avg )
1.1160 + x->act_zbin_adj = (int)(((int64_t)b + (a>>1))/a) - 1;
1.1161 + else
1.1162 + x->act_zbin_adj = 1 - (int)(((int64_t)a + (b>>1))/b);
1.1163 +#endif
1.1164 +}
1.1165 +
1.1166 +int vp8cx_encode_intra_macroblock(VP8_COMP *cpi, MACROBLOCK *x,
1.1167 + TOKENEXTRA **t)
1.1168 +{
1.1169 + MACROBLOCKD *xd = &x->e_mbd;
1.1170 + int rate;
1.1171 +
1.1172 + if (cpi->sf.RD && cpi->compressor_speed != 2)
1.1173 + vp8_rd_pick_intra_mode(x, &rate);
1.1174 + else
1.1175 + vp8_pick_intra_mode(x, &rate);
1.1176 +
1.1177 + if(cpi->oxcf.tuning == VP8_TUNE_SSIM)
1.1178 + {
1.1179 + adjust_act_zbin( cpi, x );
1.1180 + vp8_update_zbin_extra(cpi, x);
1.1181 + }
1.1182 +
1.1183 + if (x->e_mbd.mode_info_context->mbmi.mode == B_PRED)
1.1184 + vp8_encode_intra4x4mby(x);
1.1185 + else
1.1186 + vp8_encode_intra16x16mby(x);
1.1187 +
1.1188 + vp8_encode_intra16x16mbuv(x);
1.1189 +
1.1190 + sum_intra_stats(cpi, x);
1.1191 +
1.1192 + vp8_tokenize_mb(cpi, x, t);
1.1193 +
1.1194 + if (xd->mode_info_context->mbmi.mode != B_PRED)
1.1195 + vp8_inverse_transform_mby(xd);
1.1196 +
1.1197 + vp8_dequant_idct_add_uv_block
1.1198 + (xd->qcoeff+16*16, xd->dequant_uv,
1.1199 + xd->dst.u_buffer, xd->dst.v_buffer,
1.1200 + xd->dst.uv_stride, xd->eobs+16);
1.1201 + return rate;
1.1202 +}
1.1203 +#ifdef SPEEDSTATS
1.1204 +extern int cnt_pm;
1.1205 +#endif
1.1206 +
1.1207 +extern void vp8_fix_contexts(MACROBLOCKD *x);
1.1208 +
1.1209 +int vp8cx_encode_inter_macroblock
1.1210 +(
1.1211 + VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t,
1.1212 + int recon_yoffset, int recon_uvoffset,
1.1213 + int mb_row, int mb_col
1.1214 +)
1.1215 +{
1.1216 + MACROBLOCKD *const xd = &x->e_mbd;
1.1217 + int intra_error = 0;
1.1218 + int rate;
1.1219 + int distortion;
1.1220 +
1.1221 + x->skip = 0;
1.1222 +
1.1223 + if (xd->segmentation_enabled)
1.1224 + x->encode_breakout = cpi->segment_encode_breakout[xd->mode_info_context->mbmi.segment_id];
1.1225 + else
1.1226 + x->encode_breakout = cpi->oxcf.encode_breakout;
1.1227 +
1.1228 +#if CONFIG_TEMPORAL_DENOISING
1.1229 + /* Reset the best sse mode/mv for each macroblock. */
1.1230 + x->best_reference_frame = INTRA_FRAME;
1.1231 + x->best_zeromv_reference_frame = INTRA_FRAME;
1.1232 + x->best_sse_inter_mode = 0;
1.1233 + x->best_sse_mv.as_int = 0;
1.1234 + x->need_to_clamp_best_mvs = 0;
1.1235 +#endif
1.1236 +
1.1237 + if (cpi->sf.RD)
1.1238 + {
1.1239 + int zbin_mode_boost_enabled = x->zbin_mode_boost_enabled;
1.1240 +
1.1241 + /* Are we using the fast quantizer for the mode selection? */
1.1242 + if(cpi->sf.use_fastquant_for_pick)
1.1243 + {
1.1244 + x->quantize_b = vp8_fast_quantize_b;
1.1245 + x->quantize_b_pair = vp8_fast_quantize_b_pair;
1.1246 +
1.1247 + /* the fast quantizer does not use zbin_extra, so
1.1248 + * do not recalculate */
1.1249 + x->zbin_mode_boost_enabled = 0;
1.1250 + }
1.1251 + vp8_rd_pick_inter_mode(cpi, x, recon_yoffset, recon_uvoffset, &rate,
1.1252 + &distortion, &intra_error);
1.1253 +
1.1254 + /* switch back to the regular quantizer for the encode */
1.1255 + if (cpi->sf.improved_quant)
1.1256 + {
1.1257 + x->quantize_b = vp8_regular_quantize_b;
1.1258 + x->quantize_b_pair = vp8_regular_quantize_b_pair;
1.1259 + }
1.1260 +
1.1261 + /* restore cpi->zbin_mode_boost_enabled */
1.1262 + x->zbin_mode_boost_enabled = zbin_mode_boost_enabled;
1.1263 +
1.1264 + }
1.1265 + else
1.1266 + {
1.1267 + vp8_pick_inter_mode(cpi, x, recon_yoffset, recon_uvoffset, &rate,
1.1268 + &distortion, &intra_error, mb_row, mb_col);
1.1269 + }
1.1270 +
1.1271 + x->prediction_error += distortion;
1.1272 + x->intra_error += intra_error;
1.1273 +
1.1274 + if(cpi->oxcf.tuning == VP8_TUNE_SSIM)
1.1275 + {
1.1276 + /* Adjust the zbin based on this MB rate. */
1.1277 + adjust_act_zbin( cpi, x );
1.1278 + }
1.1279 +
1.1280 +#if 0
1.1281 + /* Experimental RD code */
1.1282 + cpi->frame_distortion += distortion;
1.1283 + cpi->last_mb_distortion = distortion;
1.1284 +#endif
1.1285 +
1.1286 + /* MB level adjutment to quantizer setup */
1.1287 + if (xd->segmentation_enabled)
1.1288 + {
1.1289 + /* If cyclic update enabled */
1.1290 + if (cpi->current_layer == 0 && cpi->cyclic_refresh_mode_enabled)
1.1291 + {
1.1292 + /* Clear segment_id back to 0 if not coded (last frame 0,0) */
1.1293 + if ((xd->mode_info_context->mbmi.segment_id == 1) &&
1.1294 + ((xd->mode_info_context->mbmi.ref_frame != LAST_FRAME) || (xd->mode_info_context->mbmi.mode != ZEROMV)))
1.1295 + {
1.1296 + xd->mode_info_context->mbmi.segment_id = 0;
1.1297 +
1.1298 + /* segment_id changed, so update */
1.1299 + vp8cx_mb_init_quantizer(cpi, x, 1);
1.1300 + }
1.1301 + }
1.1302 + }
1.1303 +
1.1304 + {
1.1305 + /* Experimental code.
1.1306 + * Special case for gf and arf zeromv modes, for 1 temporal layer.
1.1307 + * Increase zbin size to supress noise.
1.1308 + */
1.1309 + x->zbin_mode_boost = 0;
1.1310 + if (x->zbin_mode_boost_enabled)
1.1311 + {
1.1312 + if ( xd->mode_info_context->mbmi.ref_frame != INTRA_FRAME )
1.1313 + {
1.1314 + if (xd->mode_info_context->mbmi.mode == ZEROMV)
1.1315 + {
1.1316 + if (xd->mode_info_context->mbmi.ref_frame != LAST_FRAME &&
1.1317 + cpi->oxcf.number_of_layers == 1)
1.1318 + x->zbin_mode_boost = GF_ZEROMV_ZBIN_BOOST;
1.1319 + else
1.1320 + x->zbin_mode_boost = LF_ZEROMV_ZBIN_BOOST;
1.1321 + }
1.1322 + else if (xd->mode_info_context->mbmi.mode == SPLITMV)
1.1323 + x->zbin_mode_boost = 0;
1.1324 + else
1.1325 + x->zbin_mode_boost = MV_ZBIN_BOOST;
1.1326 + }
1.1327 + }
1.1328 +
1.1329 + /* The fast quantizer doesn't use zbin_extra, only do so with
1.1330 + * the regular quantizer. */
1.1331 + if (cpi->sf.improved_quant)
1.1332 + vp8_update_zbin_extra(cpi, x);
1.1333 + }
1.1334 +
1.1335 + x->count_mb_ref_frame_usage[xd->mode_info_context->mbmi.ref_frame] ++;
1.1336 +
1.1337 + if (xd->mode_info_context->mbmi.ref_frame == INTRA_FRAME)
1.1338 + {
1.1339 + vp8_encode_intra16x16mbuv(x);
1.1340 +
1.1341 + if (xd->mode_info_context->mbmi.mode == B_PRED)
1.1342 + {
1.1343 + vp8_encode_intra4x4mby(x);
1.1344 + }
1.1345 + else
1.1346 + {
1.1347 + vp8_encode_intra16x16mby(x);
1.1348 + }
1.1349 +
1.1350 + sum_intra_stats(cpi, x);
1.1351 + }
1.1352 + else
1.1353 + {
1.1354 + int ref_fb_idx;
1.1355 +
1.1356 + if (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME)
1.1357 + ref_fb_idx = cpi->common.lst_fb_idx;
1.1358 + else if (xd->mode_info_context->mbmi.ref_frame == GOLDEN_FRAME)
1.1359 + ref_fb_idx = cpi->common.gld_fb_idx;
1.1360 + else
1.1361 + ref_fb_idx = cpi->common.alt_fb_idx;
1.1362 +
1.1363 + xd->pre.y_buffer = cpi->common.yv12_fb[ref_fb_idx].y_buffer + recon_yoffset;
1.1364 + xd->pre.u_buffer = cpi->common.yv12_fb[ref_fb_idx].u_buffer + recon_uvoffset;
1.1365 + xd->pre.v_buffer = cpi->common.yv12_fb[ref_fb_idx].v_buffer + recon_uvoffset;
1.1366 +
1.1367 + if (!x->skip)
1.1368 + {
1.1369 + vp8_encode_inter16x16(x);
1.1370 + }
1.1371 + else
1.1372 + vp8_build_inter16x16_predictors_mb(xd, xd->dst.y_buffer,
1.1373 + xd->dst.u_buffer, xd->dst.v_buffer,
1.1374 + xd->dst.y_stride, xd->dst.uv_stride);
1.1375 +
1.1376 + }
1.1377 +
1.1378 + if (!x->skip)
1.1379 + {
1.1380 + vp8_tokenize_mb(cpi, x, t);
1.1381 +
1.1382 + if (xd->mode_info_context->mbmi.mode != B_PRED)
1.1383 + vp8_inverse_transform_mby(xd);
1.1384 +
1.1385 + vp8_dequant_idct_add_uv_block
1.1386 + (xd->qcoeff+16*16, xd->dequant_uv,
1.1387 + xd->dst.u_buffer, xd->dst.v_buffer,
1.1388 + xd->dst.uv_stride, xd->eobs+16);
1.1389 + }
1.1390 + else
1.1391 + {
1.1392 + /* always set mb_skip_coeff as it is needed by the loopfilter */
1.1393 + xd->mode_info_context->mbmi.mb_skip_coeff = 1;
1.1394 +
1.1395 + if (cpi->common.mb_no_coeff_skip)
1.1396 + {
1.1397 + x->skip_true_count ++;
1.1398 + vp8_fix_contexts(xd);
1.1399 + }
1.1400 + else
1.1401 + {
1.1402 + vp8_stuff_mb(cpi, x, t);
1.1403 + }
1.1404 + }
1.1405 +
1.1406 + return rate;
1.1407 +}
