1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libvpx/vp8/encoder/ethreading.c Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,655 @@
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 +#include "onyx_int.h"
1.15 +#include "vp8/common/threading.h"
1.16 +#include "vp8/common/common.h"
1.17 +#include "vp8/common/extend.h"
1.18 +#include "bitstream.h"
1.19 +#include "encodeframe.h"
1.20 +
1.21 +#if CONFIG_MULTITHREAD
1.22 +
1.23 +extern void vp8cx_mb_init_quantizer(VP8_COMP *cpi, MACROBLOCK *x, int ok_to_skip);
1.24 +
1.25 +extern void vp8_loopfilter_frame(VP8_COMP *cpi, VP8_COMMON *cm);
1.26 +
1.27 +static THREAD_FUNCTION thread_loopfilter(void *p_data)
1.28 +{
1.29 + VP8_COMP *cpi = (VP8_COMP *)(((LPFTHREAD_DATA *)p_data)->ptr1);
1.30 + VP8_COMMON *cm = &cpi->common;
1.31 +
1.32 + while (1)
1.33 + {
1.34 + if (cpi->b_multi_threaded == 0)
1.35 + break;
1.36 +
1.37 + if (sem_wait(&cpi->h_event_start_lpf) == 0)
1.38 + {
1.39 + if (cpi->b_multi_threaded == 0) /* we're shutting down */
1.40 + break;
1.41 +
1.42 + vp8_loopfilter_frame(cpi, cm);
1.43 +
1.44 + sem_post(&cpi->h_event_end_lpf);
1.45 + }
1.46 + }
1.47 +
1.48 + return 0;
1.49 +}
1.50 +
1.51 +static
1.52 +THREAD_FUNCTION thread_encoding_proc(void *p_data)
1.53 +{
1.54 + int ithread = ((ENCODETHREAD_DATA *)p_data)->ithread;
1.55 + VP8_COMP *cpi = (VP8_COMP *)(((ENCODETHREAD_DATA *)p_data)->ptr1);
1.56 + MB_ROW_COMP *mbri = (MB_ROW_COMP *)(((ENCODETHREAD_DATA *)p_data)->ptr2);
1.57 + ENTROPY_CONTEXT_PLANES mb_row_left_context;
1.58 +
1.59 + while (1)
1.60 + {
1.61 + if (cpi->b_multi_threaded == 0)
1.62 + break;
1.63 +
1.64 + if (sem_wait(&cpi->h_event_start_encoding[ithread]) == 0)
1.65 + {
1.66 + const int nsync = cpi->mt_sync_range;
1.67 + VP8_COMMON *cm = &cpi->common;
1.68 + int mb_row;
1.69 + MACROBLOCK *x = &mbri->mb;
1.70 + MACROBLOCKD *xd = &x->e_mbd;
1.71 + TOKENEXTRA *tp ;
1.72 +#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
1.73 + TOKENEXTRA *tp_start = cpi->tok + (1 + ithread) * (16 * 24);
1.74 + const int num_part = (1 << cm->multi_token_partition);
1.75 +#endif
1.76 +
1.77 + int *segment_counts = mbri->segment_counts;
1.78 + int *totalrate = &mbri->totalrate;
1.79 +
1.80 + if (cpi->b_multi_threaded == 0) /* we're shutting down */
1.81 + break;
1.82 +
1.83 + for (mb_row = ithread + 1; mb_row < cm->mb_rows; mb_row += (cpi->encoding_thread_count + 1))
1.84 + {
1.85 +
1.86 + int recon_yoffset, recon_uvoffset;
1.87 + int mb_col;
1.88 + int ref_fb_idx = cm->lst_fb_idx;
1.89 + int dst_fb_idx = cm->new_fb_idx;
1.90 + int recon_y_stride = cm->yv12_fb[ref_fb_idx].y_stride;
1.91 + int recon_uv_stride = cm->yv12_fb[ref_fb_idx].uv_stride;
1.92 + int map_index = (mb_row * cm->mb_cols);
1.93 + volatile const int *last_row_current_mb_col;
1.94 + volatile int *current_mb_col = &cpi->mt_current_mb_col[mb_row];
1.95 +
1.96 +#if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
1.97 + vp8_writer *w = &cpi->bc[1 + (mb_row % num_part)];
1.98 +#else
1.99 + tp = cpi->tok + (mb_row * (cm->mb_cols * 16 * 24));
1.100 + cpi->tplist[mb_row].start = tp;
1.101 +#endif
1.102 +
1.103 + last_row_current_mb_col = &cpi->mt_current_mb_col[mb_row - 1];
1.104 +
1.105 + /* reset above block coeffs */
1.106 + xd->above_context = cm->above_context;
1.107 + xd->left_context = &mb_row_left_context;
1.108 +
1.109 + vp8_zero(mb_row_left_context);
1.110 +
1.111 + xd->up_available = (mb_row != 0);
1.112 + recon_yoffset = (mb_row * recon_y_stride * 16);
1.113 + recon_uvoffset = (mb_row * recon_uv_stride * 8);
1.114 +
1.115 + /* Set the mb activity pointer to the start of the row. */
1.116 + x->mb_activity_ptr = &cpi->mb_activity_map[map_index];
1.117 +
1.118 + /* for each macroblock col in image */
1.119 + for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
1.120 + {
1.121 + *current_mb_col = mb_col - 1;
1.122 +
1.123 + if ((mb_col & (nsync - 1)) == 0)
1.124 + {
1.125 + while (mb_col > (*last_row_current_mb_col - nsync))
1.126 + {
1.127 + x86_pause_hint();
1.128 + thread_sleep(0);
1.129 + }
1.130 + }
1.131 +
1.132 +#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
1.133 + tp = tp_start;
1.134 +#endif
1.135 +
1.136 + /* Distance of Mb to the various image edges.
1.137 + * These specified to 8th pel as they are always compared
1.138 + * to values that are in 1/8th pel units
1.139 + */
1.140 + xd->mb_to_left_edge = -((mb_col * 16) << 3);
1.141 + xd->mb_to_right_edge = ((cm->mb_cols - 1 - mb_col) * 16) << 3;
1.142 + xd->mb_to_top_edge = -((mb_row * 16) << 3);
1.143 + xd->mb_to_bottom_edge = ((cm->mb_rows - 1 - mb_row) * 16) << 3;
1.144 +
1.145 + /* Set up limit values for motion vectors used to prevent
1.146 + * them extending outside the UMV borders
1.147 + */
1.148 + x->mv_col_min = -((mb_col * 16) + (VP8BORDERINPIXELS - 16));
1.149 + x->mv_col_max = ((cm->mb_cols - 1 - mb_col) * 16) + (VP8BORDERINPIXELS - 16);
1.150 + x->mv_row_min = -((mb_row * 16) + (VP8BORDERINPIXELS - 16));
1.151 + x->mv_row_max = ((cm->mb_rows - 1 - mb_row) * 16) + (VP8BORDERINPIXELS - 16);
1.152 +
1.153 + xd->dst.y_buffer = cm->yv12_fb[dst_fb_idx].y_buffer + recon_yoffset;
1.154 + xd->dst.u_buffer = cm->yv12_fb[dst_fb_idx].u_buffer + recon_uvoffset;
1.155 + xd->dst.v_buffer = cm->yv12_fb[dst_fb_idx].v_buffer + recon_uvoffset;
1.156 + xd->left_available = (mb_col != 0);
1.157 +
1.158 + x->rddiv = cpi->RDDIV;
1.159 + x->rdmult = cpi->RDMULT;
1.160 +
1.161 + /* Copy current mb to a buffer */
1.162 + vp8_copy_mem16x16(x->src.y_buffer, x->src.y_stride, x->thismb, 16);
1.163 +
1.164 + if (cpi->oxcf.tuning == VP8_TUNE_SSIM)
1.165 + vp8_activity_masking(cpi, x);
1.166 +
1.167 + /* Is segmentation enabled */
1.168 + /* MB level adjustment to quantizer */
1.169 + if (xd->segmentation_enabled)
1.170 + {
1.171 + /* Code to set segment id in xd->mbmi.segment_id for
1.172 + * current MB (with range checking)
1.173 + */
1.174 + if (cpi->segmentation_map[map_index + mb_col] <= 3)
1.175 + xd->mode_info_context->mbmi.segment_id = cpi->segmentation_map[map_index + mb_col];
1.176 + else
1.177 + xd->mode_info_context->mbmi.segment_id = 0;
1.178 +
1.179 + vp8cx_mb_init_quantizer(cpi, x, 1);
1.180 + }
1.181 + else
1.182 + /* Set to Segment 0 by default */
1.183 + xd->mode_info_context->mbmi.segment_id = 0;
1.184 +
1.185 + x->active_ptr = cpi->active_map + map_index + mb_col;
1.186 +
1.187 + if (cm->frame_type == KEY_FRAME)
1.188 + {
1.189 + *totalrate += vp8cx_encode_intra_macroblock(cpi, x, &tp);
1.190 +#ifdef MODE_STATS
1.191 + y_modes[xd->mbmi.mode] ++;
1.192 +#endif
1.193 + }
1.194 + else
1.195 + {
1.196 + *totalrate += vp8cx_encode_inter_macroblock(cpi, x, &tp, recon_yoffset, recon_uvoffset, mb_row, mb_col);
1.197 +
1.198 +#ifdef MODE_STATS
1.199 + inter_y_modes[xd->mbmi.mode] ++;
1.200 +
1.201 + if (xd->mbmi.mode == SPLITMV)
1.202 + {
1.203 + int b;
1.204 +
1.205 + for (b = 0; b < xd->mbmi.partition_count; b++)
1.206 + {
1.207 + inter_b_modes[x->partition->bmi[b].mode] ++;
1.208 + }
1.209 + }
1.210 +
1.211 +#endif
1.212 +
1.213 + /* Special case code for cyclic refresh
1.214 + * If cyclic update enabled then copy
1.215 + * xd->mbmi.segment_id; (which may have been updated
1.216 + * based on mode during
1.217 + * vp8cx_encode_inter_macroblock()) back into the
1.218 + * global segmentation map
1.219 + */
1.220 + if ((cpi->current_layer == 0) &&
1.221 + (cpi->cyclic_refresh_mode_enabled &&
1.222 + xd->segmentation_enabled))
1.223 + {
1.224 + const MB_MODE_INFO * mbmi = &xd->mode_info_context->mbmi;
1.225 + cpi->segmentation_map[map_index + mb_col] = mbmi->segment_id;
1.226 +
1.227 + /* If the block has been refreshed mark it as clean
1.228 + * (the magnitude of the -ve influences how long it
1.229 + * will be before we consider another refresh):
1.230 + * Else if it was coded (last frame 0,0) and has
1.231 + * not already been refreshed then mark it as a
1.232 + * candidate for cleanup next time (marked 0) else
1.233 + * mark it as dirty (1).
1.234 + */
1.235 + if (mbmi->segment_id)
1.236 + cpi->cyclic_refresh_map[map_index + mb_col] = -1;
1.237 + else if ((mbmi->mode == ZEROMV) && (mbmi->ref_frame == LAST_FRAME))
1.238 + {
1.239 + if (cpi->cyclic_refresh_map[map_index + mb_col] == 1)
1.240 + cpi->cyclic_refresh_map[map_index + mb_col] = 0;
1.241 + }
1.242 + else
1.243 + cpi->cyclic_refresh_map[map_index + mb_col] = 1;
1.244 +
1.245 + }
1.246 + }
1.247 +
1.248 +#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
1.249 + /* pack tokens for this MB */
1.250 + {
1.251 + int tok_count = tp - tp_start;
1.252 + pack_tokens(w, tp_start, tok_count);
1.253 + }
1.254 +#else
1.255 + cpi->tplist[mb_row].stop = tp;
1.256 +#endif
1.257 + /* Increment pointer into gf usage flags structure. */
1.258 + x->gf_active_ptr++;
1.259 +
1.260 + /* Increment the activity mask pointers. */
1.261 + x->mb_activity_ptr++;
1.262 +
1.263 + /* adjust to the next column of macroblocks */
1.264 + x->src.y_buffer += 16;
1.265 + x->src.u_buffer += 8;
1.266 + x->src.v_buffer += 8;
1.267 +
1.268 + recon_yoffset += 16;
1.269 + recon_uvoffset += 8;
1.270 +
1.271 + /* Keep track of segment usage */
1.272 + segment_counts[xd->mode_info_context->mbmi.segment_id]++;
1.273 +
1.274 + /* skip to next mb */
1.275 + xd->mode_info_context++;
1.276 + x->partition_info++;
1.277 + xd->above_context++;
1.278 + }
1.279 +
1.280 + vp8_extend_mb_row( &cm->yv12_fb[dst_fb_idx],
1.281 + xd->dst.y_buffer + 16,
1.282 + xd->dst.u_buffer + 8,
1.283 + xd->dst.v_buffer + 8);
1.284 +
1.285 + *current_mb_col = mb_col + nsync;
1.286 +
1.287 + /* this is to account for the border */
1.288 + xd->mode_info_context++;
1.289 + x->partition_info++;
1.290 +
1.291 + x->src.y_buffer += 16 * x->src.y_stride * (cpi->encoding_thread_count + 1) - 16 * cm->mb_cols;
1.292 + x->src.u_buffer += 8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
1.293 + x->src.v_buffer += 8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
1.294 +
1.295 + xd->mode_info_context += xd->mode_info_stride * cpi->encoding_thread_count;
1.296 + x->partition_info += xd->mode_info_stride * cpi->encoding_thread_count;
1.297 + x->gf_active_ptr += cm->mb_cols * cpi->encoding_thread_count;
1.298 +
1.299 + if (mb_row == cm->mb_rows - 1)
1.300 + {
1.301 + sem_post(&cpi->h_event_end_encoding); /* signal frame encoding end */
1.302 + }
1.303 + }
1.304 + }
1.305 + }
1.306 +
1.307 + /* printf("exit thread %d\n", ithread); */
1.308 + return 0;
1.309 +}
1.310 +
1.311 +static void setup_mbby_copy(MACROBLOCK *mbdst, MACROBLOCK *mbsrc)
1.312 +{
1.313 +
1.314 + MACROBLOCK *x = mbsrc;
1.315 + MACROBLOCK *z = mbdst;
1.316 + int i;
1.317 +
1.318 + z->ss = x->ss;
1.319 + z->ss_count = x->ss_count;
1.320 + z->searches_per_step = x->searches_per_step;
1.321 + z->errorperbit = x->errorperbit;
1.322 +
1.323 + z->sadperbit16 = x->sadperbit16;
1.324 + z->sadperbit4 = x->sadperbit4;
1.325 +
1.326 + /*
1.327 + z->mv_col_min = x->mv_col_min;
1.328 + z->mv_col_max = x->mv_col_max;
1.329 + z->mv_row_min = x->mv_row_min;
1.330 + z->mv_row_max = x->mv_row_max;
1.331 + */
1.332 +
1.333 + z->short_fdct4x4 = x->short_fdct4x4;
1.334 + z->short_fdct8x4 = x->short_fdct8x4;
1.335 + z->short_walsh4x4 = x->short_walsh4x4;
1.336 + z->quantize_b = x->quantize_b;
1.337 + z->quantize_b_pair = x->quantize_b_pair;
1.338 + z->optimize = x->optimize;
1.339 +
1.340 + /*
1.341 + z->mvc = x->mvc;
1.342 + z->src.y_buffer = x->src.y_buffer;
1.343 + z->src.u_buffer = x->src.u_buffer;
1.344 + z->src.v_buffer = x->src.v_buffer;
1.345 + */
1.346 +
1.347 + z->mvcost[0] = x->mvcost[0];
1.348 + z->mvcost[1] = x->mvcost[1];
1.349 + z->mvsadcost[0] = x->mvsadcost[0];
1.350 + z->mvsadcost[1] = x->mvsadcost[1];
1.351 +
1.352 + z->token_costs = x->token_costs;
1.353 + z->inter_bmode_costs = x->inter_bmode_costs;
1.354 + z->mbmode_cost = x->mbmode_cost;
1.355 + z->intra_uv_mode_cost = x->intra_uv_mode_cost;
1.356 + z->bmode_costs = x->bmode_costs;
1.357 +
1.358 + for (i = 0; i < 25; i++)
1.359 + {
1.360 + z->block[i].quant = x->block[i].quant;
1.361 + z->block[i].quant_fast = x->block[i].quant_fast;
1.362 + z->block[i].quant_shift = x->block[i].quant_shift;
1.363 + z->block[i].zbin = x->block[i].zbin;
1.364 + z->block[i].zrun_zbin_boost = x->block[i].zrun_zbin_boost;
1.365 + z->block[i].round = x->block[i].round;
1.366 + z->block[i].src_stride = x->block[i].src_stride;
1.367 + }
1.368 +
1.369 + z->q_index = x->q_index;
1.370 + z->act_zbin_adj = x->act_zbin_adj;
1.371 + z->last_act_zbin_adj = x->last_act_zbin_adj;
1.372 +
1.373 + {
1.374 + MACROBLOCKD *xd = &x->e_mbd;
1.375 + MACROBLOCKD *zd = &z->e_mbd;
1.376 +
1.377 + /*
1.378 + zd->mode_info_context = xd->mode_info_context;
1.379 + zd->mode_info = xd->mode_info;
1.380 +
1.381 + zd->mode_info_stride = xd->mode_info_stride;
1.382 + zd->frame_type = xd->frame_type;
1.383 + zd->up_available = xd->up_available ;
1.384 + zd->left_available = xd->left_available;
1.385 + zd->left_context = xd->left_context;
1.386 + zd->last_frame_dc = xd->last_frame_dc;
1.387 + zd->last_frame_dccons = xd->last_frame_dccons;
1.388 + zd->gold_frame_dc = xd->gold_frame_dc;
1.389 + zd->gold_frame_dccons = xd->gold_frame_dccons;
1.390 + zd->mb_to_left_edge = xd->mb_to_left_edge;
1.391 + zd->mb_to_right_edge = xd->mb_to_right_edge;
1.392 + zd->mb_to_top_edge = xd->mb_to_top_edge ;
1.393 + zd->mb_to_bottom_edge = xd->mb_to_bottom_edge;
1.394 + zd->gf_active_ptr = xd->gf_active_ptr;
1.395 + zd->frames_since_golden = xd->frames_since_golden;
1.396 + zd->frames_till_alt_ref_frame = xd->frames_till_alt_ref_frame;
1.397 + */
1.398 + zd->subpixel_predict = xd->subpixel_predict;
1.399 + zd->subpixel_predict8x4 = xd->subpixel_predict8x4;
1.400 + zd->subpixel_predict8x8 = xd->subpixel_predict8x8;
1.401 + zd->subpixel_predict16x16 = xd->subpixel_predict16x16;
1.402 + zd->segmentation_enabled = xd->segmentation_enabled;
1.403 + zd->mb_segement_abs_delta = xd->mb_segement_abs_delta;
1.404 + vpx_memcpy(zd->segment_feature_data, xd->segment_feature_data,
1.405 + sizeof(xd->segment_feature_data));
1.406 +
1.407 + vpx_memcpy(zd->dequant_y1_dc, xd->dequant_y1_dc,
1.408 + sizeof(xd->dequant_y1_dc));
1.409 + vpx_memcpy(zd->dequant_y1, xd->dequant_y1, sizeof(xd->dequant_y1));
1.410 + vpx_memcpy(zd->dequant_y2, xd->dequant_y2, sizeof(xd->dequant_y2));
1.411 + vpx_memcpy(zd->dequant_uv, xd->dequant_uv, sizeof(xd->dequant_uv));
1.412 +
1.413 +#if 1
1.414 + /*TODO: Remove dequant from BLOCKD. This is a temporary solution until
1.415 + * the quantizer code uses a passed in pointer to the dequant constants.
1.416 + * This will also require modifications to the x86 and neon assembly.
1.417 + * */
1.418 + for (i = 0; i < 16; i++)
1.419 + zd->block[i].dequant = zd->dequant_y1;
1.420 + for (i = 16; i < 24; i++)
1.421 + zd->block[i].dequant = zd->dequant_uv;
1.422 + zd->block[24].dequant = zd->dequant_y2;
1.423 +#endif
1.424 +
1.425 +
1.426 + vpx_memcpy(z->rd_threshes, x->rd_threshes, sizeof(x->rd_threshes));
1.427 + vpx_memcpy(z->rd_thresh_mult, x->rd_thresh_mult,
1.428 + sizeof(x->rd_thresh_mult));
1.429 +
1.430 + z->zbin_over_quant = x->zbin_over_quant;
1.431 + z->zbin_mode_boost_enabled = x->zbin_mode_boost_enabled;
1.432 + z->zbin_mode_boost = x->zbin_mode_boost;
1.433 +
1.434 + vpx_memset(z->error_bins, 0, sizeof(z->error_bins));
1.435 + }
1.436 +}
1.437 +
1.438 +void vp8cx_init_mbrthread_data(VP8_COMP *cpi,
1.439 + MACROBLOCK *x,
1.440 + MB_ROW_COMP *mbr_ei,
1.441 + int count
1.442 + )
1.443 +{
1.444 +
1.445 + VP8_COMMON *const cm = & cpi->common;
1.446 + MACROBLOCKD *const xd = & x->e_mbd;
1.447 + int i;
1.448 +
1.449 + for (i = 0; i < count; i++)
1.450 + {
1.451 + MACROBLOCK *mb = & mbr_ei[i].mb;
1.452 + MACROBLOCKD *mbd = &mb->e_mbd;
1.453 +
1.454 + mbd->subpixel_predict = xd->subpixel_predict;
1.455 + mbd->subpixel_predict8x4 = xd->subpixel_predict8x4;
1.456 + mbd->subpixel_predict8x8 = xd->subpixel_predict8x8;
1.457 + mbd->subpixel_predict16x16 = xd->subpixel_predict16x16;
1.458 + mb->gf_active_ptr = x->gf_active_ptr;
1.459 +
1.460 + vpx_memset(mbr_ei[i].segment_counts, 0, sizeof(mbr_ei[i].segment_counts));
1.461 + mbr_ei[i].totalrate = 0;
1.462 +
1.463 + mb->partition_info = x->pi + x->e_mbd.mode_info_stride * (i + 1);
1.464 +
1.465 + mbd->mode_info_context = cm->mi + x->e_mbd.mode_info_stride * (i + 1);
1.466 + mbd->mode_info_stride = cm->mode_info_stride;
1.467 +
1.468 + mbd->frame_type = cm->frame_type;
1.469 +
1.470 + mb->src = * cpi->Source;
1.471 + mbd->pre = cm->yv12_fb[cm->lst_fb_idx];
1.472 + mbd->dst = cm->yv12_fb[cm->new_fb_idx];
1.473 +
1.474 + mb->src.y_buffer += 16 * x->src.y_stride * (i + 1);
1.475 + mb->src.u_buffer += 8 * x->src.uv_stride * (i + 1);
1.476 + mb->src.v_buffer += 8 * x->src.uv_stride * (i + 1);
1.477 +
1.478 + vp8_build_block_offsets(mb);
1.479 +
1.480 + mbd->left_context = &cm->left_context;
1.481 + mb->mvc = cm->fc.mvc;
1.482 +
1.483 + setup_mbby_copy(&mbr_ei[i].mb, x);
1.484 +
1.485 + mbd->fullpixel_mask = 0xffffffff;
1.486 + if(cm->full_pixel)
1.487 + mbd->fullpixel_mask = 0xfffffff8;
1.488 +
1.489 + vp8_zero(mb->coef_counts);
1.490 + vp8_zero(x->ymode_count);
1.491 + mb->skip_true_count = 0;
1.492 + vp8_zero(mb->MVcount);
1.493 + mb->prediction_error = 0;
1.494 + mb->intra_error = 0;
1.495 + vp8_zero(mb->count_mb_ref_frame_usage);
1.496 + mb->mbs_tested_so_far = 0;
1.497 + }
1.498 +}
1.499 +
1.500 +int vp8cx_create_encoder_threads(VP8_COMP *cpi)
1.501 +{
1.502 + const VP8_COMMON * cm = &cpi->common;
1.503 +
1.504 + cpi->b_multi_threaded = 0;
1.505 + cpi->encoding_thread_count = 0;
1.506 + cpi->b_lpf_running = 0;
1.507 +
1.508 + if (cm->processor_core_count > 1 && cpi->oxcf.multi_threaded > 1)
1.509 + {
1.510 + int ithread;
1.511 + int th_count = cpi->oxcf.multi_threaded - 1;
1.512 + int rc = 0;
1.513 +
1.514 + /* don't allocate more threads than cores available */
1.515 + if (cpi->oxcf.multi_threaded > cm->processor_core_count)
1.516 + th_count = cm->processor_core_count - 1;
1.517 +
1.518 + /* we have th_count + 1 (main) threads processing one row each */
1.519 + /* no point to have more threads than the sync range allows */
1.520 + if(th_count > ((cm->mb_cols / cpi->mt_sync_range) - 1))
1.521 + {
1.522 + th_count = (cm->mb_cols / cpi->mt_sync_range) - 1;
1.523 + }
1.524 +
1.525 + if(th_count == 0)
1.526 + return 0;
1.527 +
1.528 + CHECK_MEM_ERROR(cpi->h_encoding_thread,
1.529 + vpx_malloc(sizeof(pthread_t) * th_count));
1.530 + CHECK_MEM_ERROR(cpi->h_event_start_encoding,
1.531 + vpx_malloc(sizeof(sem_t) * th_count));
1.532 + CHECK_MEM_ERROR(cpi->mb_row_ei,
1.533 + vpx_memalign(32, sizeof(MB_ROW_COMP) * th_count));
1.534 + vpx_memset(cpi->mb_row_ei, 0, sizeof(MB_ROW_COMP) * th_count);
1.535 + CHECK_MEM_ERROR(cpi->en_thread_data,
1.536 + vpx_malloc(sizeof(ENCODETHREAD_DATA) * th_count));
1.537 +
1.538 + sem_init(&cpi->h_event_end_encoding, 0, 0);
1.539 +
1.540 + cpi->b_multi_threaded = 1;
1.541 + cpi->encoding_thread_count = th_count;
1.542 +
1.543 + /*
1.544 + printf("[VP8:] multi_threaded encoding is enabled with %d threads\n\n",
1.545 + (cpi->encoding_thread_count +1));
1.546 + */
1.547 +
1.548 + for (ithread = 0; ithread < th_count; ithread++)
1.549 + {
1.550 + ENCODETHREAD_DATA *ethd = &cpi->en_thread_data[ithread];
1.551 +
1.552 + /* Setup block ptrs and offsets */
1.553 + vp8_setup_block_ptrs(&cpi->mb_row_ei[ithread].mb);
1.554 + vp8_setup_block_dptrs(&cpi->mb_row_ei[ithread].mb.e_mbd);
1.555 +
1.556 + sem_init(&cpi->h_event_start_encoding[ithread], 0, 0);
1.557 +
1.558 + ethd->ithread = ithread;
1.559 + ethd->ptr1 = (void *)cpi;
1.560 + ethd->ptr2 = (void *)&cpi->mb_row_ei[ithread];
1.561 +
1.562 + rc = pthread_create(&cpi->h_encoding_thread[ithread], 0,
1.563 + thread_encoding_proc, ethd);
1.564 + if(rc)
1.565 + break;
1.566 + }
1.567 +
1.568 + if(rc)
1.569 + {
1.570 + /* shutdown other threads */
1.571 + cpi->b_multi_threaded = 0;
1.572 + for(--ithread; ithread >= 0; ithread--)
1.573 + {
1.574 + pthread_join(cpi->h_encoding_thread[ithread], 0);
1.575 + sem_destroy(&cpi->h_event_start_encoding[ithread]);
1.576 + }
1.577 + sem_destroy(&cpi->h_event_end_encoding);
1.578 +
1.579 + /* free thread related resources */
1.580 + vpx_free(cpi->h_event_start_encoding);
1.581 + vpx_free(cpi->h_encoding_thread);
1.582 + vpx_free(cpi->mb_row_ei);
1.583 + vpx_free(cpi->en_thread_data);
1.584 +
1.585 + return -1;
1.586 + }
1.587 +
1.588 +
1.589 + {
1.590 + LPFTHREAD_DATA * lpfthd = &cpi->lpf_thread_data;
1.591 +
1.592 + sem_init(&cpi->h_event_start_lpf, 0, 0);
1.593 + sem_init(&cpi->h_event_end_lpf, 0, 0);
1.594 +
1.595 + lpfthd->ptr1 = (void *)cpi;
1.596 + rc = pthread_create(&cpi->h_filter_thread, 0, thread_loopfilter,
1.597 + lpfthd);
1.598 +
1.599 + if(rc)
1.600 + {
1.601 + /* shutdown other threads */
1.602 + cpi->b_multi_threaded = 0;
1.603 + for(--ithread; ithread >= 0; ithread--)
1.604 + {
1.605 + sem_post(&cpi->h_event_start_encoding[ithread]);
1.606 + pthread_join(cpi->h_encoding_thread[ithread], 0);
1.607 + sem_destroy(&cpi->h_event_start_encoding[ithread]);
1.608 + }
1.609 + sem_destroy(&cpi->h_event_end_encoding);
1.610 + sem_destroy(&cpi->h_event_end_lpf);
1.611 + sem_destroy(&cpi->h_event_start_lpf);
1.612 +
1.613 + /* free thread related resources */
1.614 + vpx_free(cpi->h_event_start_encoding);
1.615 + vpx_free(cpi->h_encoding_thread);
1.616 + vpx_free(cpi->mb_row_ei);
1.617 + vpx_free(cpi->en_thread_data);
1.618 +
1.619 + return -2;
1.620 + }
1.621 + }
1.622 + }
1.623 + return 0;
1.624 +}
1.625 +
1.626 +void vp8cx_remove_encoder_threads(VP8_COMP *cpi)
1.627 +{
1.628 + if (cpi->b_multi_threaded)
1.629 + {
1.630 + /* shutdown other threads */
1.631 + cpi->b_multi_threaded = 0;
1.632 + {
1.633 + int i;
1.634 +
1.635 + for (i = 0; i < cpi->encoding_thread_count; i++)
1.636 + {
1.637 + sem_post(&cpi->h_event_start_encoding[i]);
1.638 + pthread_join(cpi->h_encoding_thread[i], 0);
1.639 +
1.640 + sem_destroy(&cpi->h_event_start_encoding[i]);
1.641 + }
1.642 +
1.643 + sem_post(&cpi->h_event_start_lpf);
1.644 + pthread_join(cpi->h_filter_thread, 0);
1.645 + }
1.646 +
1.647 + sem_destroy(&cpi->h_event_end_encoding);
1.648 + sem_destroy(&cpi->h_event_end_lpf);
1.649 + sem_destroy(&cpi->h_event_start_lpf);
1.650 +
1.651 + /* free thread related resources */
1.652 + vpx_free(cpi->h_event_start_encoding);
1.653 + vpx_free(cpi->h_encoding_thread);
1.654 + vpx_free(cpi->mb_row_ei);
1.655 + vpx_free(cpi->en_thread_data);
1.656 + }
1.657 +}
1.658 +#endif
