2 * Copyright (c) 2010 The VP8 project authors. All Rights Reserved.
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
12 #include "vpx_ports/config.h"
18 #include "entropymode.h"
19 #include "quant_common.h"
20 #include "segmentation_common.h"
21 #include "setupintrarecon.h"
22 #include "encodeintra.h"
23 #include "reconinter.h"
25 #include "pickinter.h"
26 #include "findnearmv.h"
27 #include "reconintra.h"
31 #include "vpx_ports/vpx_timer.h"
33 #if CONFIG_RUNTIME_CPU_DETECT
34 #define RTCD(x) &cpi->common.rtcd.x
35 #define IF_RTCD(x) (x)
38 #define IF_RTCD(x) NULL
40 extern void vp8_stuff_mb(VP8_COMP *cpi, MACROBLOCKD *x, TOKENEXTRA **t) ;
42 extern void vp8cx_initialize_me_consts(VP8_COMP *cpi, int QIndex);
43 extern void vp8_auto_select_speed(VP8_COMP *cpi);
44 extern void vp8cx_init_mbrthread_data(VP8_COMP *cpi,
49 void vp8_build_block_offsets(MACROBLOCK *x);
50 void vp8_setup_block_ptrs(MACROBLOCK *x);
51 int vp8cx_encode_inter_macroblock(VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t, int recon_yoffset, int recon_uvoffset);
52 int vp8cx_encode_intra_macro_block(VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t);
55 unsigned int inter_y_modes[10] = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
56 unsigned int inter_uv_modes[4] = {0, 0, 0, 0};
57 unsigned int inter_b_modes[15] = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
58 unsigned int y_modes[5] = {0, 0, 0, 0, 0};
59 unsigned int uv_modes[4] = {0, 0, 0, 0};
60 unsigned int b_modes[14] = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
63 // The first four entries are dummy values
64 static const int qrounding_factors[129] =
66 56, 56, 56, 56, 56, 56, 56, 56,
67 48, 48, 48, 48, 48, 48, 48, 48,
68 48, 48, 48, 48, 48, 48, 48, 48,
69 48, 48, 48, 48, 48, 48, 48, 48,
70 48, 48, 48, 48, 48, 48, 48, 48,
71 48, 48, 48, 48, 48, 48, 48, 48,
72 48, 48, 48, 48, 48, 48, 48, 48,
73 48, 48, 48, 48, 48, 48, 48, 48,
74 48, 48, 48, 48, 48, 48, 48, 48,
75 48, 48, 48, 48, 48, 48, 48, 48,
76 48, 48, 48, 48, 48, 48, 48, 48,
77 48, 48, 48, 48, 48, 48, 48, 48,
78 48, 48, 48, 48, 48, 48, 48, 48,
79 48, 48, 48, 48, 48, 48, 48, 48,
80 48, 48, 48, 48, 48, 48, 48, 48,
81 48, 48, 48, 48, 48, 48, 48, 48,
85 static const int qzbin_factors[129] =
87 64, 64, 64, 64, 80, 80, 80, 80,
88 80, 80, 80, 80, 80, 80, 80, 80,
89 80, 80, 80, 80, 80, 80, 80, 80,
90 80, 80, 80, 80, 80, 80, 80, 80,
91 80, 80, 80, 80, 80, 80, 80, 80,
92 80, 80, 80, 80, 80, 80, 80, 80,
93 80, 80, 80, 80, 80, 80, 80, 80,
94 80, 80, 80, 80, 80, 80, 80, 80,
95 80, 80, 80, 80, 80, 80, 80, 80,
96 80, 80, 80, 80, 80, 80, 80, 80,
97 80, 80, 80, 80, 80, 80, 80, 80,
98 80, 80, 80, 80, 80, 80, 80, 80,
99 80, 80, 80, 80, 80, 80, 80, 80,
100 80, 80, 80, 80, 80, 80, 80, 80,
101 80, 80, 80, 80, 80, 80, 80, 80,
102 80, 80, 80, 80, 80, 80, 80, 80,
106 void vp8cx_init_quantizer(VP8_COMP *cpi)
113 int zbin_boost[16] = {0, 0, 8, 10, 12, 14, 16, 20, 24, 28, 32, 36, 40, 44, 44, 44};
115 for (Q = 0; Q < QINDEX_RANGE; Q++)
118 quant_val = vp8_dc_quant(Q, cpi->common.y1dc_delta_q);
119 cpi->Y1quant[Q][0][0] = (1 << 16) / quant_val;
120 cpi->Y1zbin[Q][0][0] = ((qzbin_factors[Q] * quant_val) + 64) >> 7;
121 cpi->Y1round[Q][0][0] = (qrounding_factors[Q] * quant_val) >> 7;
122 cpi->common.Y1dequant[Q][0][0] = quant_val;
123 cpi->zrun_zbin_boost_y1[Q][0] = (quant_val * zbin_boost[0]) >> 7;
125 quant_val = vp8_dc2quant(Q, cpi->common.y2dc_delta_q);
126 cpi->Y2quant[Q][0][0] = (1 << 16) / quant_val;
127 cpi->Y2zbin[Q][0][0] = ((qzbin_factors[Q] * quant_val) + 64) >> 7;
128 cpi->Y2round[Q][0][0] = (qrounding_factors[Q] * quant_val) >> 7;
129 cpi->common.Y2dequant[Q][0][0] = quant_val;
130 cpi->zrun_zbin_boost_y2[Q][0] = (quant_val * zbin_boost[0]) >> 7;
132 quant_val = vp8_dc_uv_quant(Q, cpi->common.uvdc_delta_q);
133 cpi->UVquant[Q][0][0] = (1 << 16) / quant_val;
134 cpi->UVzbin[Q][0][0] = ((qzbin_factors[Q] * quant_val) + 64) >> 7;;
135 cpi->UVround[Q][0][0] = (qrounding_factors[Q] * quant_val) >> 7;
136 cpi->common.UVdequant[Q][0][0] = quant_val;
137 cpi->zrun_zbin_boost_uv[Q][0] = (quant_val * zbin_boost[0]) >> 7;
139 // all the ac values = ;
140 for (i = 1; i < 16; i++)
142 int rc = vp8_default_zig_zag1d[i];
146 quant_val = vp8_ac_yquant(Q);
147 cpi->Y1quant[Q][r][c] = (1 << 16) / quant_val;
148 cpi->Y1zbin[Q][r][c] = ((qzbin_factors[Q] * quant_val) + 64) >> 7;
149 cpi->Y1round[Q][r][c] = (qrounding_factors[Q] * quant_val) >> 7;
150 cpi->common.Y1dequant[Q][r][c] = quant_val;
151 cpi->zrun_zbin_boost_y1[Q][i] = (quant_val * zbin_boost[i]) >> 7;
153 quant_val = vp8_ac2quant(Q, cpi->common.y2ac_delta_q);
154 cpi->Y2quant[Q][r][c] = (1 << 16) / quant_val;
155 cpi->Y2zbin[Q][r][c] = ((qzbin_factors[Q] * quant_val) + 64) >> 7;
156 cpi->Y2round[Q][r][c] = (qrounding_factors[Q] * quant_val) >> 7;
157 cpi->common.Y2dequant[Q][r][c] = quant_val;
158 cpi->zrun_zbin_boost_y2[Q][i] = (quant_val * zbin_boost[i]) >> 7;
160 quant_val = vp8_ac_uv_quant(Q, cpi->common.uvac_delta_q);
161 cpi->UVquant[Q][r][c] = (1 << 16) / quant_val;
162 cpi->UVzbin[Q][r][c] = ((qzbin_factors[Q] * quant_val) + 64) >> 7;
163 cpi->UVround[Q][r][c] = (qrounding_factors[Q] * quant_val) >> 7;
164 cpi->common.UVdequant[Q][r][c] = quant_val;
165 cpi->zrun_zbin_boost_uv[Q][i] = (quant_val * zbin_boost[i]) >> 7;
170 void vp8cx_mb_init_quantizer(VP8_COMP *cpi, MACROBLOCK *x)
174 MACROBLOCKD *xd = &x->e_mbd;
175 MB_MODE_INFO *mbmi = &xd->mbmi;
178 // Select the baseline MB Q index.
179 if (xd->segmentation_enabled)
182 if (xd->mb_segement_abs_delta == SEGMENT_ABSDATA)
183 QIndex = xd->segment_feature_data[MB_LVL_ALT_Q][mbmi->segment_id];
188 QIndex = cpi->common.base_qindex + xd->segment_feature_data[MB_LVL_ALT_Q][mbmi->segment_id];
189 QIndex = (QIndex >= 0) ? ((QIndex <= MAXQ) ? QIndex : MAXQ) : 0; // Clamp to valid range
193 QIndex = cpi->common.base_qindex;
196 zbin_extra = (cpi->common.Y1dequant[QIndex][0][1] * (cpi->zbin_over_quant + cpi->zbin_mode_boost)) >> 7;
198 for (i = 0; i < 16; i++)
200 x->block[i].quant = cpi->Y1quant[QIndex];
201 x->block[i].zbin = cpi->Y1zbin[QIndex];
202 x->block[i].round = cpi->Y1round[QIndex];
203 x->e_mbd.block[i].dequant = cpi->common.Y1dequant[QIndex];
204 x->block[i].zrun_zbin_boost = cpi->zrun_zbin_boost_y1[QIndex];
205 x->block[i].zbin_extra = (short)zbin_extra;
209 zbin_extra = (cpi->common.UVdequant[QIndex][0][1] * (cpi->zbin_over_quant + cpi->zbin_mode_boost)) >> 7;
211 for (i = 16; i < 24; i++)
213 x->block[i].quant = cpi->UVquant[QIndex];
214 x->block[i].zbin = cpi->UVzbin[QIndex];
215 x->block[i].round = cpi->UVround[QIndex];
216 x->e_mbd.block[i].dequant = cpi->common.UVdequant[QIndex];
217 x->block[i].zrun_zbin_boost = cpi->zrun_zbin_boost_uv[QIndex];
218 x->block[i].zbin_extra = (short)zbin_extra;
222 zbin_extra = (cpi->common.Y2dequant[QIndex][0][1] * ((cpi->zbin_over_quant / 2) + cpi->zbin_mode_boost)) >> 7;
223 x->block[24].quant = cpi->Y2quant[QIndex];
224 x->block[24].zbin = cpi->Y2zbin[QIndex];
225 x->block[24].round = cpi->Y2round[QIndex];
226 x->e_mbd.block[24].dequant = cpi->common.Y2dequant[QIndex];
227 x->block[24].zrun_zbin_boost = cpi->zrun_zbin_boost_y2[QIndex];
228 x->block[24].zbin_extra = (short)zbin_extra;
231 void vp8cx_frame_init_quantizer(VP8_COMP *cpi)
233 // vp8cx_init_quantizer() is first called in vp8_create_compressor(). A check is added here so that vp8cx_init_quantizer() is only called
234 // when these values are not all zero.
235 if (cpi->common.y1dc_delta_q | cpi->common.y2dc_delta_q | cpi->common.uvdc_delta_q | cpi->common.y2ac_delta_q | cpi->common.uvac_delta_q)
237 vp8cx_init_quantizer(cpi);
240 // MB level quantizer setup
241 vp8cx_mb_init_quantizer(cpi, &cpi->mb);
247 void encode_mb_row(VP8_COMP *cpi,
257 int recon_yoffset, recon_uvoffset;
259 int recon_y_stride = cm->last_frame.y_stride;
260 int recon_uv_stride = cm->last_frame.uv_stride;
261 int seg_map_index = (mb_row * cpi->common.mb_cols);
264 // reset above block coeffs
265 xd->above_context[Y1CONTEXT] = cm->above_context[Y1CONTEXT];
266 xd->above_context[UCONTEXT ] = cm->above_context[UCONTEXT ];
267 xd->above_context[VCONTEXT ] = cm->above_context[VCONTEXT ];
268 xd->above_context[Y2CONTEXT] = cm->above_context[Y2CONTEXT];
270 xd->up_available = (mb_row != 0);
271 recon_yoffset = (mb_row * recon_y_stride * 16);
272 recon_uvoffset = (mb_row * recon_uv_stride * 8);
274 cpi->tplist[mb_row].start = *tp;
275 //printf("Main mb_row = %d\n", mb_row);
277 // for each macroblock col in image
278 for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
280 // Distance of Mb to the various image edges.
281 // These specified to 8th pel as they are always compared to values that are in 1/8th pel units
282 xd->mb_to_left_edge = -((mb_col * 16) << 3);
283 xd->mb_to_right_edge = ((cm->mb_cols - 1 - mb_col) * 16) << 3;
284 xd->mb_to_top_edge = -((mb_row * 16) << 3);
285 xd->mb_to_bottom_edge = ((cm->mb_rows - 1 - mb_row) * 16) << 3;
287 // Set up limit values for motion vectors used to prevent them extending outside the UMV borders
288 x->mv_col_min = -((mb_col * 16) + (VP8BORDERINPIXELS - 16));
289 x->mv_col_max = ((cm->mb_cols - 1 - mb_col) * 16) + (VP8BORDERINPIXELS - 16);
290 x->mv_row_min = -((mb_row * 16) + (VP8BORDERINPIXELS - 16));
291 x->mv_row_max = ((cm->mb_rows - 1 - mb_row) * 16) + (VP8BORDERINPIXELS - 16);
293 xd->dst.y_buffer = cm->new_frame.y_buffer + recon_yoffset;
294 xd->dst.u_buffer = cm->new_frame.u_buffer + recon_uvoffset;
295 xd->dst.v_buffer = cm->new_frame.v_buffer + recon_uvoffset;
296 xd->left_available = (mb_col != 0);
298 // Is segmentation enabled
299 // MB level adjutment to quantizer
300 if (xd->segmentation_enabled)
302 // Code to set segment id in xd->mbmi.segment_id for current MB (with range checking)
303 if (cpi->segmentation_map[seg_map_index+mb_col] <= 3)
304 xd->mbmi.segment_id = cpi->segmentation_map[seg_map_index+mb_col];
306 xd->mbmi.segment_id = 0;
308 vp8cx_mb_init_quantizer(cpi, x);
311 xd->mbmi.segment_id = 0; // Set to Segment 0 by default
313 x->active_ptr = cpi->active_map + seg_map_index + mb_col;
315 if (cm->frame_type == KEY_FRAME)
317 *totalrate += vp8cx_encode_intra_macro_block(cpi, x, tp);
319 y_modes[xd->mbmi.mode] ++;
324 *totalrate += vp8cx_encode_inter_macroblock(cpi, x, tp, recon_yoffset, recon_uvoffset);
327 inter_y_modes[xd->mbmi.mode] ++;
329 if (xd->mbmi.mode == SPLITMV)
333 for (b = 0; b < xd->mbmi.partition_count; b++)
335 inter_b_modes[xd->mbmi.partition_bmi[b].mode] ++;
341 // Count of last ref frame 0,0 useage
342 if ((xd->mbmi.mode == ZEROMV) && (xd->mbmi.ref_frame == LAST_FRAME))
343 cpi->inter_zz_count ++;
345 // Special case code for cyclic refresh
346 // If cyclic update enabled then copy xd->mbmi.segment_id; (which may have been updated based on mode
347 // during vp8cx_encode_inter_macroblock()) back into the global sgmentation map
348 if (cpi->cyclic_refresh_mode_enabled && xd->segmentation_enabled)
350 cpi->segmentation_map[seg_map_index+mb_col] = xd->mbmi.segment_id;
352 // If the block has been refreshed mark it as clean (the magnitude of the -ve influences how long it will be before we consider another refresh):
353 // Else if it was coded (last frame 0,0) and has not already been refreshed then mark it as a candidate for cleanup next time (marked 0)
354 // else mark it as dirty (1).
355 if (xd->mbmi.segment_id)
356 cpi->cyclic_refresh_map[seg_map_index+mb_col] = -1;
357 else if ((xd->mbmi.mode == ZEROMV) && (xd->mbmi.ref_frame == LAST_FRAME))
359 if (cpi->cyclic_refresh_map[seg_map_index+mb_col] == 1)
360 cpi->cyclic_refresh_map[seg_map_index+mb_col] = 0;
363 cpi->cyclic_refresh_map[seg_map_index+mb_col] = 1;
368 cpi->tplist[mb_row].stop = *tp;
370 xd->gf_active_ptr++; // Increment pointer into gf useage flags structure for next mb
372 // store macroblock mode info into context array
373 vpx_memcpy(&xd->mode_info_context->mbmi, &xd->mbmi, sizeof(xd->mbmi));
375 for (i = 0; i < 16; i++)
376 vpx_memcpy(&xd->mode_info_context->bmi[i], &xd->block[i].bmi, sizeof(xd->block[i].bmi));
378 // adjust to the next column of macroblocks
379 x->src.y_buffer += 16;
380 x->src.u_buffer += 8;
381 x->src.v_buffer += 8;
386 // Keep track of segment useage
387 segment_counts[xd->mbmi.segment_id] ++;
390 xd->mode_info_context++;
392 xd->above_context[Y1CONTEXT] += 4;
393 xd->above_context[UCONTEXT ] += 2;
394 xd->above_context[VCONTEXT ] += 2;
395 xd->above_context[Y2CONTEXT] ++;
396 cpi->current_mb_col_main = mb_col;
399 //extend the recon for intra prediction
402 xd->dst.y_buffer + 16,
403 xd->dst.u_buffer + 8,
404 xd->dst.v_buffer + 8);
406 // this is to account for the border
407 xd->mode_info_context++;
414 void vp8_encode_frame(VP8_COMP *cpi)
417 MACROBLOCK *const x = & cpi->mb;
418 VP8_COMMON *const cm = & cpi->common;
419 MACROBLOCKD *const xd = & x->e_mbd;
422 TOKENEXTRA *tp = cpi->tok;
423 int segment_counts[MAX_MB_SEGMENTS];
426 if (cm->frame_type != KEY_FRAME)
428 if (cm->mcomp_filter_type == SIXTAP)
430 xd->subpixel_predict = SUBPIX_INVOKE(&cpi->common.rtcd.subpix, sixtap4x4);
431 xd->subpixel_predict8x4 = SUBPIX_INVOKE(&cpi->common.rtcd.subpix, sixtap8x4);
432 xd->subpixel_predict8x8 = SUBPIX_INVOKE(&cpi->common.rtcd.subpix, sixtap8x8);
433 xd->subpixel_predict16x16 = SUBPIX_INVOKE(&cpi->common.rtcd.subpix, sixtap16x16);
437 xd->subpixel_predict = SUBPIX_INVOKE(&cpi->common.rtcd.subpix, bilinear4x4);
438 xd->subpixel_predict8x4 = SUBPIX_INVOKE(&cpi->common.rtcd.subpix, bilinear8x4);
439 xd->subpixel_predict8x8 = SUBPIX_INVOKE(&cpi->common.rtcd.subpix, bilinear8x8);
440 xd->subpixel_predict16x16 = SUBPIX_INVOKE(&cpi->common.rtcd.subpix, bilinear16x16);
446 // For key frames make sure the intra ref frame probability value
447 // is set to "all intra"
448 //cpi->prob_intra_coded = 255;
452 xd->gf_active_ptr = (signed char *)cm->gf_active_flags; // Point to base of GF active flags data structure
454 x->vector_range = 32;
456 // Count of MBs using the alternate Q if any
459 // Reset frame count of inter 0,0 motion vector useage.
460 cpi->inter_zz_count = 0;
462 vpx_memset(segment_counts, 0, sizeof(segment_counts));
464 cpi->prediction_error = 0;
465 cpi->intra_error = 0;
466 cpi->skip_true_count = 0;
467 cpi->skip_false_count = 0;
471 cpi->frame_distortion = 0;
472 cpi->last_mb_distortion = 0;
477 xd->mode_info = cm->mi - 1;
479 xd->mode_info_context = cm->mi;
480 xd->mode_info_stride = cm->mode_info_stride;
482 xd->frame_type = cm->frame_type;
484 xd->frames_since_golden = cm->frames_since_golden;
485 xd->frames_till_alt_ref_frame = cm->frames_till_alt_ref_frame;
486 vp8_zero(cpi->MVcount);
487 // vp8_zero( Contexts)
488 vp8_zero(cpi->coef_counts);
490 // reset intra mode contexts
491 if (cm->frame_type == KEY_FRAME)
492 vp8_init_mbmode_probs(cm);
495 vp8cx_frame_init_quantizer(cpi);
497 if (cpi->compressor_speed == 2)
499 if (cpi->oxcf.cpu_used < 0)
500 cpi->Speed = -(cpi->oxcf.cpu_used);
502 vp8_auto_select_speed(cpi);
505 vp8_initialize_rd_consts(cpi, vp8_dc_quant(cm->base_qindex, cm->y1dc_delta_q));
506 //vp8_initialize_rd_consts( cpi, vp8_dc_quant(cpi->avg_frame_qindex, cm->y1dc_delta_q) );
507 vp8cx_initialize_me_consts(cpi, cm->base_qindex);
508 //vp8cx_initialize_me_consts( cpi, cpi->avg_frame_qindex);
510 // Copy data over into macro block data sturctures.
512 x->src = * cpi->Source;
513 xd->pre = cm->last_frame;
514 xd->dst = cm->new_frame;
516 // set up frame new frame for intra coded blocks
518 vp8_setup_intra_recon(&cm->new_frame);
520 vp8_build_block_offsets(x);
522 vp8_setup_block_dptrs(&x->e_mbd);
524 vp8_setup_block_ptrs(x);
526 x->rddiv = cpi->RDDIV;
527 x->rdmult = cpi->RDMULT;
530 // Experimental rd code
531 // 2 Pass - Possibly set Rdmult based on last frame distortion + this frame target bits or other metrics
532 // such as cpi->rate_correction_factor that indicate relative complexity.
533 /*if ( cpi->pass == 2 && (cpi->last_frame_distortion > 0) && (cpi->target_bits_per_mb > 0) )
535 //x->rdmult = ((cpi->last_frame_distortion * 256)/cpi->common.MBs)/ cpi->target_bits_per_mb;
536 x->rdmult = (int)(cpi->RDMULT * cpi->rate_correction_factor);
539 x->rdmult = cpi->RDMULT; */
540 //x->rdmult = (int)(cpi->RDMULT * pow( (cpi->rate_correction_factor * 2.0), 0.75 ));
543 xd->mbmi.mode = DC_PRED;
544 xd->mbmi.uv_mode = DC_PRED;
546 xd->left_context = cm->left_context;
548 vp8_zero(cpi->count_mb_ref_frame_usage)
549 vp8_zero(cpi->ymode_count)
550 vp8_zero(cpi->uv_mode_count)
554 // vp8_zero( entropy_stats)
556 ENTROPY_CONTEXT **p = cm->above_context;
557 const size_t L = cm->mb_cols;
559 vp8_zero_array(p [Y1CONTEXT], L * 4)
560 vp8_zero_array(p [ UCONTEXT], L * 2)
561 vp8_zero_array(p [ VCONTEXT], L * 2)
562 vp8_zero_array(p [Y2CONTEXT], L)
567 struct vpx_usec_timer emr_timer;
568 vpx_usec_timer_start(&emr_timer);
570 if (!cpi->b_multi_threaded)
572 // for each macroblock row in image
573 for (mb_row = 0; mb_row < cm->mb_rows; mb_row++)
576 vp8_zero(cm->left_context)
578 encode_mb_row(cpi, cm, mb_row, x, xd, &tp, segment_counts, &totalrate);
580 // adjust to the next row of mbs
581 x->src.y_buffer += 16 * x->src.y_stride - 16 * cm->mb_cols;
582 x->src.u_buffer += 8 * x->src.uv_stride - 8 * cm->mb_cols;
583 x->src.v_buffer += 8 * x->src.uv_stride - 8 * cm->mb_cols;
586 cpi->tok_count = tp - cpi->tok;
591 #if CONFIG_MULTITHREAD
592 vp8cx_init_mbrthread_data(cpi, x, cpi->mb_row_ei, 1, cpi->encoding_thread_count);
594 for (mb_row = 0; mb_row < cm->mb_rows; mb_row += (cpi->encoding_thread_count + 1))
597 cpi->current_mb_col_main = -1;
599 for (i = 0; i < cpi->encoding_thread_count; i++)
601 if ((mb_row + i + 1) >= cm->mb_rows)
604 cpi->mb_row_ei[i].mb_row = mb_row + i + 1;
605 cpi->mb_row_ei[i].tp = cpi->tok + (mb_row + i + 1) * (cm->mb_cols * 16 * 24);
606 cpi->mb_row_ei[i].current_mb_col = -1;
607 //SetEvent(cpi->h_event_mbrencoding[i]);
608 sem_post(&cpi->h_event_mbrencoding[i]);
611 vp8_zero(cm->left_context)
613 tp = cpi->tok + mb_row * (cm->mb_cols * 16 * 24);
615 encode_mb_row(cpi, cm, mb_row, x, xd, &tp, segment_counts, &totalrate);
617 // adjust to the next row of mbs
618 x->src.y_buffer += 16 * x->src.y_stride * (cpi->encoding_thread_count + 1) - 16 * cm->mb_cols;
619 x->src.u_buffer += 8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
620 x->src.v_buffer += 8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
622 xd->mode_info_context += xd->mode_info_stride * cpi->encoding_thread_count;
624 if (mb_row < cm->mb_rows - 1)
625 //WaitForSingleObject(cpi->h_event_main, INFINITE);
626 sem_wait(&cpi->h_event_main);
630 for( ;mb_row<cm->mb_rows; mb_row ++)
632 vp8_zero( cm->left_context)
634 tp = cpi->tok + mb_row * (cm->mb_cols * 16 * 24);
636 encode_mb_row(cpi, cm, mb_row, x, xd, &tp, segment_counts, &totalrate);
637 // adjust to the next row of mbs
638 x->src.y_buffer += 16 * x->src.y_stride - 16 * cm->mb_cols;
639 x->src.u_buffer += 8 * x->src.uv_stride - 8 * cm->mb_cols;
640 x->src.v_buffer += 8 * x->src.uv_stride - 8 * cm->mb_cols;
646 for (mb_row = 0; mb_row < cm->mb_rows; mb_row ++)
648 cpi->tok_count += cpi->tplist[mb_row].stop - cpi->tplist[mb_row].start;
651 if (xd->segmentation_enabled)
656 if (xd->segmentation_enabled)
659 for (i = 0; i < cpi->encoding_thread_count; i++)
661 for (j = 0; j < 4; j++)
662 segment_counts[j] += cpi->mb_row_ei[i].segment_counts[j];
668 for (i = 0; i < cpi->encoding_thread_count; i++)
670 totalrate += cpi->mb_row_ei[i].totalrate;
677 vpx_usec_timer_mark(&emr_timer);
678 cpi->time_encode_mb_row += vpx_usec_timer_elapsed(&emr_timer);
683 // Work out the segment probabilites if segmentation is enabled
684 if (xd->segmentation_enabled)
690 vpx_memset(xd->mb_segment_tree_probs, 255 , sizeof(xd->mb_segment_tree_probs));
692 tot_count = segment_counts[0] + segment_counts[1] + segment_counts[2] + segment_counts[3];
696 xd->mb_segment_tree_probs[0] = ((segment_counts[0] + segment_counts[1]) * 255) / tot_count;
698 tot_count = segment_counts[0] + segment_counts[1];
702 xd->mb_segment_tree_probs[1] = (segment_counts[0] * 255) / tot_count;
705 tot_count = segment_counts[2] + segment_counts[3];
708 xd->mb_segment_tree_probs[2] = (segment_counts[2] * 255) / tot_count;
710 // Zero probabilities not allowed
711 for (i = 0; i < MB_FEATURE_TREE_PROBS; i ++)
713 if (xd->mb_segment_tree_probs[i] == 0)
714 xd->mb_segment_tree_probs[i] = 1;
719 // 256 rate units to the bit
720 cpi->projected_frame_size = totalrate >> 8; // projected_frame_size in units of BYTES
722 // Make a note of the percentage MBs coded Intra.
723 if (cm->frame_type == KEY_FRAME)
725 cpi->this_frame_percent_intra = 100;
731 tot_modes = cpi->count_mb_ref_frame_usage[INTRA_FRAME]
732 + cpi->count_mb_ref_frame_usage[LAST_FRAME]
733 + cpi->count_mb_ref_frame_usage[GOLDEN_FRAME]
734 + cpi->count_mb_ref_frame_usage[ALTREF_FRAME];
737 cpi->this_frame_percent_intra = cpi->count_mb_ref_frame_usage[INTRA_FRAME] * 100 / tot_modes;
744 int flag[2] = {0, 0};
746 for (cnt = 0; cnt < MVPcount; cnt++)
748 if (cm->fc.pre_mvc[0][cnt] != cm->fc.mvc[0][cnt])
751 vpx_memcpy(cm->fc.pre_mvc[0], cm->fc.mvc[0], MVPcount);
756 for (cnt = 0; cnt < MVPcount; cnt++)
758 if (cm->fc.pre_mvc[1][cnt] != cm->fc.mvc[1][cnt])
761 vpx_memcpy(cm->fc.pre_mvc[1], cm->fc.mvc[1], MVPcount);
766 if (flag[0] || flag[1])
767 vp8_build_component_cost_table(cpi->mb.mvcost, cpi->mb.mvsadcost, (const MV_CONTEXT *) cm->fc.mvc, flag);
771 // Adjust the projected reference frame useage probability numbers to reflect
772 // what we have just seen. This may be usefull when we make multiple itterations
773 // of the recode loop rather than continuing to use values from the previous frame.
774 if ((cm->frame_type != KEY_FRAME) && !cm->refresh_alt_ref_frame && !cm->refresh_golden_frame)
776 const int *const rfct = cpi->count_mb_ref_frame_usage;
777 const int rf_intra = rfct[INTRA_FRAME];
778 const int rf_inter = rfct[LAST_FRAME] + rfct[GOLDEN_FRAME] + rfct[ALTREF_FRAME];
780 if ((rf_intra + rf_inter) > 0)
782 cpi->prob_intra_coded = (rf_intra * 255) / (rf_intra + rf_inter);
784 if (cpi->prob_intra_coded < 1)
785 cpi->prob_intra_coded = 1;
787 if ((cm->frames_since_golden > 0) || cpi->source_alt_ref_active)
789 cpi->prob_last_coded = rf_inter ? (rfct[LAST_FRAME] * 255) / rf_inter : 128;
791 if (cpi->prob_last_coded < 1)
792 cpi->prob_last_coded = 1;
794 cpi->prob_gf_coded = (rfct[GOLDEN_FRAME] + rfct[ALTREF_FRAME])
795 ? (rfct[GOLDEN_FRAME] * 255) / (rfct[GOLDEN_FRAME] + rfct[ALTREF_FRAME]) : 128;
797 if (cpi->prob_gf_coded < 1)
798 cpi->prob_gf_coded = 1;
804 // Keep record of the total distortion this time around for future use
805 cpi->last_frame_distortion = cpi->frame_distortion;
809 void vp8_setup_block_ptrs(MACROBLOCK *x)
814 for (r = 0; r < 4; r++)
816 for (c = 0; c < 4; c++)
818 x->block[r*4+c].src_diff = x->src_diff + r * 4 * 16 + c * 4;
822 for (r = 0; r < 2; r++)
824 for (c = 0; c < 2; c++)
826 x->block[16 + r*2+c].src_diff = x->src_diff + 256 + r * 4 * 8 + c * 4;
831 for (r = 0; r < 2; r++)
833 for (c = 0; c < 2; c++)
835 x->block[20 + r*2+c].src_diff = x->src_diff + 320 + r * 4 * 8 + c * 4;
839 x->block[24].src_diff = x->src_diff + 384;
842 for (i = 0; i < 25; i++)
844 x->block[i].coeff = x->coeff + i * 16;
848 void vp8_build_block_offsets(MACROBLOCK *x)
853 vp8_build_block_doffsets(&x->e_mbd);
856 for (br = 0; br < 4; br++)
858 for (bc = 0; bc < 4; bc++)
860 BLOCK *this_block = &x->block[block];
861 this_block->base_src = &x->src.y_buffer;
862 this_block->src_stride = x->src.y_stride;
863 this_block->src = 4 * br * this_block->src_stride + 4 * bc;
869 for (br = 0; br < 2; br++)
871 for (bc = 0; bc < 2; bc++)
873 BLOCK *this_block = &x->block[block];
874 this_block->base_src = &x->src.u_buffer;
875 this_block->src_stride = x->src.uv_stride;
876 this_block->src = 4 * br * this_block->src_stride + 4 * bc;
882 for (br = 0; br < 2; br++)
884 for (bc = 0; bc < 2; bc++)
886 BLOCK *this_block = &x->block[block];
887 this_block->base_src = &x->src.v_buffer;
888 this_block->src_stride = x->src.uv_stride;
889 this_block->src = 4 * br * this_block->src_stride + 4 * bc;
895 static void sum_intra_stats(VP8_COMP *cpi, MACROBLOCK *x)
897 const MACROBLOCKD *xd = & x->e_mbd;
898 const MB_PREDICTION_MODE m = xd->mbmi.mode;
899 const MB_PREDICTION_MODE uvm = xd->mbmi.uv_mode;
902 const int is_key = cpi->common.frame_type == KEY_FRAME;
904 ++ (is_key ? uv_modes : inter_uv_modes)[uvm];
908 unsigned int *const bct = is_key ? b_modes : inter_b_modes;
914 ++ bct[xd->block[b].bmi.mode];
921 ++cpi->ymode_count[m];
922 ++cpi->uv_mode_count[uvm];
925 int vp8cx_encode_intra_macro_block(VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t)
927 int Error4x4, Error16x16, error_uv;
928 B_PREDICTION_MODE intra_bmodes[16];
929 int rate4x4, rate16x16, rateuv;
930 int dist4x4, dist16x16, distuv;
932 int rate4x4_tokenonly = 0;
933 int rate16x16_tokenonly = 0;
934 int rateuv_tokenonly = 0;
937 x->e_mbd.mbmi.ref_frame = INTRA_FRAME;
939 #if !(CONFIG_REALTIME_ONLY)
941 if (cpi->sf.RD || cpi->compressor_speed != 2)
943 Error4x4 = vp8_rd_pick_intra4x4mby_modes(cpi, x, &rate4x4, &rate4x4_tokenonly, &dist4x4);
945 //save the b modes for possible later use
946 for (i = 0; i < 16; i++)
947 intra_bmodes[i] = x->e_mbd.block[i].bmi.mode;
949 Error16x16 = vp8_rd_pick_intra16x16mby_mode(cpi, x, &rate16x16, &rate16x16_tokenonly, &dist16x16);
951 error_uv = vp8_rd_pick_intra_mbuv_mode(cpi, x, &rateuv, &rateuv_tokenonly, &distuv);
953 x->e_mbd.mbmi.mb_skip_coeff = (cpi->common.mb_no_coeff_skip) ? 1 : 0;
955 vp8_encode_intra16x16mbuv(IF_RTCD(&cpi->rtcd), x);
958 if (Error4x4 < Error16x16)
961 x->e_mbd.mbmi.mode = B_PRED;
963 // get back the intra block modes
964 for (i = 0; i < 16; i++)
965 x->e_mbd.block[i].bmi.mode = intra_bmodes[i];
967 vp8_encode_intra4x4mby(IF_RTCD(&cpi->rtcd), x);
968 cpi->prediction_error += Error4x4 ;
970 // Experimental RD code
971 cpi->frame_distortion += dist4x4;
976 vp8_encode_intra16x16mby(IF_RTCD(&cpi->rtcd), x);
980 // Experimental RD code
981 cpi->prediction_error += Error16x16;
982 cpi->frame_distortion += dist16x16;
986 sum_intra_stats(cpi, x);
988 vp8_tokenize_mb(cpi, &x->e_mbd, t);
994 int rate2, distortion2;
995 MB_PREDICTION_MODE mode, best_mode = DC_PRED;
997 Error16x16 = INT_MAX;
999 for (mode = DC_PRED; mode <= TM_PRED; mode ++)
1001 x->e_mbd.mbmi.mode = mode;
1002 vp8_build_intra_predictors_mby_ptr(&x->e_mbd);
1003 distortion2 = VARIANCE_INVOKE(&cpi->rtcd.variance, get16x16prederror)(x->src.y_buffer, x->src.y_stride, x->e_mbd.predictor, 16, 0x7fffffff);
1004 rate2 = x->mbmode_cost[x->e_mbd.frame_type][mode];
1005 this_rd = RD_ESTIMATE(x->rdmult, x->rddiv, rate2, distortion2);
1007 if (Error16x16 > this_rd)
1009 Error16x16 = this_rd;
1014 vp8_pick_intra4x4mby_modes(IF_RTCD(&cpi->rtcd), x, &rate2, &distortion2);
1016 if (distortion2 == INT_MAX)
1019 Error4x4 = RD_ESTIMATE(x->rdmult, x->rddiv, rate2, distortion2);
1021 x->e_mbd.mbmi.mb_skip_coeff = (cpi->common.mb_no_coeff_skip) ? 1 : 0;
1023 if (Error4x4 < Error16x16)
1025 x->e_mbd.mbmi.mode = B_PRED;
1026 vp8_encode_intra4x4mby(IF_RTCD(&cpi->rtcd), x);
1027 cpi->prediction_error += Error4x4;
1031 x->e_mbd.mbmi.mode = best_mode;
1032 vp8_encode_intra16x16mby(IF_RTCD(&cpi->rtcd), x);
1033 cpi->prediction_error += Error16x16;
1036 vp8_pick_intra_mbuv_mode(x);
1037 vp8_encode_intra16x16mbuv(IF_RTCD(&cpi->rtcd), x);
1038 sum_intra_stats(cpi, x);
1039 vp8_tokenize_mb(cpi, &x->e_mbd, t);
1048 extern void vp8_fix_contexts(VP8_COMP *cpi, MACROBLOCKD *x);
1050 int vp8cx_encode_inter_macroblock
1052 VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t,
1053 int recon_yoffset, int recon_uvoffset
1056 MACROBLOCKD *const xd = &x->e_mbd;
1058 int intra_error = 0;
1064 if (xd->segmentation_enabled)
1065 x->encode_breakout = cpi->segment_encode_breakout[xd->mbmi.segment_id];
1067 x->encode_breakout = cpi->oxcf.encode_breakout;
1069 #if !(CONFIG_REALTIME_ONLY)
1073 inter_error = vp8_rd_pick_inter_mode(cpi, x, recon_yoffset, recon_uvoffset, &rate, &distortion, &intra_error);
1077 inter_error = vp8_pick_inter_mode(cpi, x, recon_yoffset, recon_uvoffset, &rate, &distortion, &intra_error);
1080 cpi->prediction_error += inter_error;
1081 cpi->intra_error += intra_error;
1084 // Experimental RD code
1085 cpi->frame_distortion += distortion;
1086 cpi->last_mb_distortion = distortion;
1089 // MB level adjutment to quantizer setup
1090 if (xd->segmentation_enabled || cpi->zbin_mode_boost_enabled)
1092 // If cyclic update enabled
1093 if (cpi->cyclic_refresh_mode_enabled)
1095 // Clear segment_id back to 0 if not coded (last frame 0,0)
1096 if ((xd->mbmi.segment_id == 1) &&
1097 ((xd->mbmi.ref_frame != LAST_FRAME) || (xd->mbmi.mode != ZEROMV)))
1099 xd->mbmi.segment_id = 0;
1103 // Experimental code. Special case for gf and arf zeromv modes. Increase zbin size to supress noise
1104 if (cpi->zbin_mode_boost_enabled)
1106 if ((xd->mbmi.mode == ZEROMV) && (xd->mbmi.ref_frame != LAST_FRAME))
1107 cpi->zbin_mode_boost = GF_ZEROMV_ZBIN_BOOST;
1109 cpi->zbin_mode_boost = 0;
1112 vp8cx_mb_init_quantizer(cpi, x);
1115 cpi->count_mb_ref_frame_usage[xd->mbmi.ref_frame] ++;
1117 if (xd->mbmi.ref_frame == INTRA_FRAME)
1119 x->e_mbd.mbmi.mb_skip_coeff = (cpi->common.mb_no_coeff_skip) ? 1 : 0;
1121 vp8_encode_intra16x16mbuv(IF_RTCD(&cpi->rtcd), x);
1123 if (xd->mbmi.mode == B_PRED)
1125 vp8_encode_intra4x4mby(IF_RTCD(&cpi->rtcd), x);
1129 vp8_encode_intra16x16mby(IF_RTCD(&cpi->rtcd), x);
1132 sum_intra_stats(cpi, x);
1140 vp8_find_near_mvs(xd, xd->mode_info_context,
1141 &nearest, &nearby, &best_ref_mv, mdcounts, xd->mbmi.ref_frame, cpi->common.ref_frame_sign_bias);
1143 vp8_build_uvmvs(xd, cpi->common.full_pixel);
1145 // store motion vectors in our motion vector list
1146 if (xd->mbmi.ref_frame == LAST_FRAME)
1148 // Set up pointers for this macro block into the previous frame recon buffer
1149 xd->pre.y_buffer = cpi->common.last_frame.y_buffer + recon_yoffset;
1150 xd->pre.u_buffer = cpi->common.last_frame.u_buffer + recon_uvoffset;
1151 xd->pre.v_buffer = cpi->common.last_frame.v_buffer + recon_uvoffset;
1153 else if (xd->mbmi.ref_frame == GOLDEN_FRAME)
1155 // Set up pointers for this macro block into the golden frame recon buffer
1156 xd->pre.y_buffer = cpi->common.golden_frame.y_buffer + recon_yoffset;
1157 xd->pre.u_buffer = cpi->common.golden_frame.u_buffer + recon_uvoffset;
1158 xd->pre.v_buffer = cpi->common.golden_frame.v_buffer + recon_uvoffset;
1162 // Set up pointers for this macro block into the alternate reference frame recon buffer
1163 xd->pre.y_buffer = cpi->common.alt_ref_frame.y_buffer + recon_yoffset;
1164 xd->pre.u_buffer = cpi->common.alt_ref_frame.u_buffer + recon_uvoffset;
1165 xd->pre.v_buffer = cpi->common.alt_ref_frame.v_buffer + recon_uvoffset;
1168 if (xd->mbmi.mode == SPLITMV)
1172 for (i = 0; i < 16; i++)
1174 if (xd->block[i].bmi.mode == NEW4X4)
1176 cpi->MVcount[0][mv_max+((xd->block[i].bmi.mv.as_mv.row - best_ref_mv.row) >> 1)]++;
1177 cpi->MVcount[1][mv_max+((xd->block[i].bmi.mv.as_mv.col - best_ref_mv.col) >> 1)]++;
1181 else if (xd->mbmi.mode == NEWMV)
1183 cpi->MVcount[0][mv_max+((xd->block[0].bmi.mv.as_mv.row - best_ref_mv.row) >> 1)]++;
1184 cpi->MVcount[1][mv_max+((xd->block[0].bmi.mv.as_mv.col - best_ref_mv.col) >> 1)]++;
1187 if (!x->skip && !x->e_mbd.mbmi.force_no_skip)
1189 vp8_encode_inter16x16(IF_RTCD(&cpi->rtcd), x);
1191 // Clear mb_skip_coeff if mb_no_coeff_skip is not set
1192 if (!cpi->common.mb_no_coeff_skip)
1193 xd->mbmi.mb_skip_coeff = 0;
1197 vp8_stuff_inter16x16(x);
1201 vp8_tokenize_mb(cpi, xd, t);
1204 if (cpi->common.mb_no_coeff_skip)
1206 if (xd->mbmi.mode != B_PRED && xd->mbmi.mode != SPLITMV)
1207 xd->mbmi.dc_diff = 0;
1209 xd->mbmi.dc_diff = 1;
1211 xd->mbmi.mb_skip_coeff = 1;
1212 cpi->skip_true_count ++;
1213 vp8_fix_contexts(cpi, xd);
1217 vp8_stuff_mb(cpi, xd, t);
1218 xd->mbmi.mb_skip_coeff = 0;
1219 cpi->skip_false_count ++;