The Tor Browser / file comparison
comparison: media/libvpx/vp9/encoder/vp9_mbgraph.c
media/libvpx/vp9/encoder/vp9_mbgraph.c
- changeset 0
- 6474c204b198
equal
deleted
inserted
replaced
| -1:000000000000 | 0:294b8fe3d4cb |
|---|---|
| 1 /* | |
| 2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved. | |
| 3 * | |
| 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. | |
| 9 */ | |
| 10 | |
| 11 #include <limits.h> | |
| 12 | |
| 13 #include "vpx_mem/vpx_mem.h" | |
| 14 #include "vp9/encoder/vp9_encodeintra.h" | |
| 15 #include "vp9/encoder/vp9_rdopt.h" | |
| 16 #include "vp9/encoder/vp9_segmentation.h" | |
| 17 #include "vp9/encoder/vp9_mcomp.h" | |
| 18 #include "vp9/common/vp9_blockd.h" | |
| 19 #include "vp9/common/vp9_reconinter.h" | |
| 20 #include "vp9/common/vp9_reconintra.h" | |
| 21 #include "vp9/common/vp9_systemdependent.h" | |
| 22 | |
| 23 | |
| 24 | |
| 25 static unsigned int do_16x16_motion_iteration(VP9_COMP *cpi, | |
| 26 int_mv *ref_mv, | |
| 27 int_mv *dst_mv, | |
| 28 int mb_row, | |
| 29 int mb_col) { | |
| 30 MACROBLOCK *const x = &cpi->mb; | |
| 31 MACROBLOCKD *const xd = &x->e_mbd; | |
| 32 vp9_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[BLOCK_16X16]; | |
| 33 unsigned int best_err; | |
| 34 | |
| 35 const int tmp_col_min = x->mv_col_min; | |
| 36 const int tmp_col_max = x->mv_col_max; | |
| 37 const int tmp_row_min = x->mv_row_min; | |
| 38 const int tmp_row_max = x->mv_row_max; | |
| 39 int_mv ref_full; | |
| 40 | |
| 41 // Further step/diamond searches as necessary | |
| 42 int step_param = cpi->sf.reduce_first_step_size + | |
| 43 (cpi->speed < 8 ? (cpi->speed > 5 ? 1 : 0) : 2); | |
| 44 step_param = MIN(step_param, (cpi->sf.max_step_search_steps - 2)); | |
| 45 | |
| 46 vp9_clamp_mv_min_max(x, &ref_mv->as_mv); | |
| 47 | |
| 48 ref_full.as_mv.col = ref_mv->as_mv.col >> 3; | |
| 49 ref_full.as_mv.row = ref_mv->as_mv.row >> 3; | |
| 50 | |
| 51 /*cpi->sf.search_method == HEX*/ | |
| 52 best_err = vp9_hex_search(x, &ref_full.as_mv, step_param, x->errorperbit, | |
| 53 0, &v_fn_ptr, | |
| 54 0, &ref_mv->as_mv, &dst_mv->as_mv); | |
| 55 | |
| 56 // Try sub-pixel MC | |
| 57 // if (bestsme > error_thresh && bestsme < INT_MAX) | |
| 58 { | |
| 59 int distortion; | |
| 60 unsigned int sse; | |
| 61 best_err = cpi->find_fractional_mv_step( | |
| 62 x, | |
| 63 &dst_mv->as_mv, &ref_mv->as_mv, | |
| 64 cpi->common.allow_high_precision_mv, | |
| 65 x->errorperbit, &v_fn_ptr, | |
| 66 0, cpi->sf.subpel_iters_per_step, NULL, NULL, | |
| 67 & distortion, &sse); | |
| 68 } | |
| 69 | |
| 70 vp9_set_mbmode_and_mvs(x, NEWMV, dst_mv); | |
| 71 vp9_build_inter_predictors_sby(xd, mb_row, mb_col, BLOCK_16X16); | |
| 72 best_err = vp9_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride, | |
| 73 xd->plane[0].dst.buf, xd->plane[0].dst.stride, | |
| 74 INT_MAX); | |
| 75 | |
| 76 /* restore UMV window */ | |
| 77 x->mv_col_min = tmp_col_min; | |
| 78 x->mv_col_max = tmp_col_max; | |
| 79 x->mv_row_min = tmp_row_min; | |
| 80 x->mv_row_max = tmp_row_max; | |
| 81 | |
| 82 return best_err; | |
| 83 } | |
| 84 | |
| 85 static int do_16x16_motion_search(VP9_COMP *cpi, int_mv *ref_mv, int_mv *dst_mv, | |
| 86 int mb_row, int mb_col) { | |
| 87 MACROBLOCK *const x = &cpi->mb; | |
| 88 MACROBLOCKD *const xd = &x->e_mbd; | |
| 89 unsigned int err, tmp_err; | |
| 90 int_mv tmp_mv; | |
| 91 | |
| 92 // Try zero MV first | |
| 93 // FIXME should really use something like near/nearest MV and/or MV prediction | |
| 94 err = vp9_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride, | |
| 95 xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride, | |
| 96 INT_MAX); | |
| 97 dst_mv->as_int = 0; | |
| 98 | |
| 99 // Test last reference frame using the previous best mv as the | |
| 100 // starting point (best reference) for the search | |
| 101 tmp_err = do_16x16_motion_iteration(cpi, ref_mv, &tmp_mv, mb_row, mb_col); | |
| 102 if (tmp_err < err) { | |
| 103 err = tmp_err; | |
| 104 dst_mv->as_int = tmp_mv.as_int; | |
| 105 } | |
| 106 | |
| 107 // If the current best reference mv is not centered on 0,0 then do a 0,0 | |
| 108 // based search as well. | |
| 109 if (ref_mv->as_int) { | |
| 110 unsigned int tmp_err; | |
| 111 int_mv zero_ref_mv, tmp_mv; | |
| 112 | |
| 113 zero_ref_mv.as_int = 0; | |
| 114 tmp_err = do_16x16_motion_iteration(cpi, &zero_ref_mv, &tmp_mv, | |
| 115 mb_row, mb_col); | |
| 116 if (tmp_err < err) { | |
| 117 dst_mv->as_int = tmp_mv.as_int; | |
| 118 err = tmp_err; | |
| 119 } | |
| 120 } | |
| 121 | |
| 122 return err; | |
| 123 } | |
| 124 | |
| 125 static int do_16x16_zerozero_search(VP9_COMP *cpi, int_mv *dst_mv) { | |
| 126 MACROBLOCK *const x = &cpi->mb; | |
| 127 MACROBLOCKD *const xd = &x->e_mbd; | |
| 128 unsigned int err; | |
| 129 | |
| 130 // Try zero MV first | |
| 131 // FIXME should really use something like near/nearest MV and/or MV prediction | |
| 132 err = vp9_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride, | |
| 133 xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride, | |
| 134 INT_MAX); | |
| 135 | |
| 136 dst_mv->as_int = 0; | |
| 137 | |
| 138 return err; | |
| 139 } | |
| 140 static int find_best_16x16_intra(VP9_COMP *cpi, | |
| 141 int mb_y_offset, | |
| 142 MB_PREDICTION_MODE *pbest_mode) { | |
| 143 MACROBLOCK *const x = &cpi->mb; | |
| 144 MACROBLOCKD *const xd = &x->e_mbd; | |
| 145 MB_PREDICTION_MODE best_mode = -1, mode; | |
| 146 unsigned int best_err = INT_MAX; | |
| 147 | |
| 148 // calculate SATD for each intra prediction mode; | |
| 149 // we're intentionally not doing 4x4, we just want a rough estimate | |
| 150 for (mode = DC_PRED; mode <= TM_PRED; mode++) { | |
| 151 unsigned int err; | |
| 152 | |
| 153 xd->mi_8x8[0]->mbmi.mode = mode; | |
| 154 vp9_predict_intra_block(xd, 0, 2, TX_16X16, mode, | |
| 155 x->plane[0].src.buf, x->plane[0].src.stride, | |
| 156 xd->plane[0].dst.buf, xd->plane[0].dst.stride); | |
| 157 err = vp9_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride, | |
| 158 xd->plane[0].dst.buf, xd->plane[0].dst.stride, best_err); | |
| 159 | |
| 160 // find best | |
| 161 if (err < best_err) { | |
| 162 best_err = err; | |
| 163 best_mode = mode; | |
| 164 } | |
| 165 } | |
| 166 | |
| 167 if (pbest_mode) | |
| 168 *pbest_mode = best_mode; | |
| 169 | |
| 170 return best_err; | |
| 171 } | |
| 172 | |
| 173 static void update_mbgraph_mb_stats | |
| 174 ( | |
| 175 VP9_COMP *cpi, | |
| 176 MBGRAPH_MB_STATS *stats, | |
| 177 YV12_BUFFER_CONFIG *buf, | |
| 178 int mb_y_offset, | |
| 179 YV12_BUFFER_CONFIG *golden_ref, | |
| 180 int_mv *prev_golden_ref_mv, | |
| 181 int gld_y_offset, | |
| 182 YV12_BUFFER_CONFIG *alt_ref, | |
| 183 int_mv *prev_alt_ref_mv, | |
| 184 int arf_y_offset, | |
| 185 int mb_row, | |
| 186 int mb_col | |
| 187 ) { | |
| 188 MACROBLOCK *const x = &cpi->mb; | |
| 189 MACROBLOCKD *const xd = &x->e_mbd; | |
| 190 int intra_error; | |
| 191 VP9_COMMON *cm = &cpi->common; | |
| 192 | |
| 193 // FIXME in practice we're completely ignoring chroma here | |
| 194 x->plane[0].src.buf = buf->y_buffer + mb_y_offset; | |
| 195 x->plane[0].src.stride = buf->y_stride; | |
| 196 | |
| 197 xd->plane[0].dst.buf = get_frame_new_buffer(cm)->y_buffer + mb_y_offset; | |
| 198 xd->plane[0].dst.stride = get_frame_new_buffer(cm)->y_stride; | |
| 199 | |
| 200 // do intra 16x16 prediction | |
| 201 intra_error = find_best_16x16_intra(cpi, mb_y_offset, | |
| 202 &stats->ref[INTRA_FRAME].m.mode); | |
| 203 if (intra_error <= 0) | |
| 204 intra_error = 1; | |
| 205 stats->ref[INTRA_FRAME].err = intra_error; | |
| 206 | |
| 207 // Golden frame MV search, if it exists and is different than last frame | |
| 208 if (golden_ref) { | |
| 209 int g_motion_error; | |
| 210 xd->plane[0].pre[0].buf = golden_ref->y_buffer + mb_y_offset; | |
| 211 xd->plane[0].pre[0].stride = golden_ref->y_stride; | |
| 212 g_motion_error = do_16x16_motion_search(cpi, | |
| 213 prev_golden_ref_mv, | |
| 214 &stats->ref[GOLDEN_FRAME].m.mv, | |
| 215 mb_row, mb_col); | |
| 216 stats->ref[GOLDEN_FRAME].err = g_motion_error; | |
| 217 } else { | |
| 218 stats->ref[GOLDEN_FRAME].err = INT_MAX; | |
| 219 stats->ref[GOLDEN_FRAME].m.mv.as_int = 0; | |
| 220 } | |
| 221 | |
| 222 // Do an Alt-ref frame MV search, if it exists and is different than | |
| 223 // last/golden frame. | |
| 224 if (alt_ref) { | |
| 225 int a_motion_error; | |
| 226 xd->plane[0].pre[0].buf = alt_ref->y_buffer + mb_y_offset; | |
| 227 xd->plane[0].pre[0].stride = alt_ref->y_stride; | |
| 228 a_motion_error = do_16x16_zerozero_search(cpi, | |
| 229 &stats->ref[ALTREF_FRAME].m.mv); | |
| 230 | |
| 231 stats->ref[ALTREF_FRAME].err = a_motion_error; | |
| 232 } else { | |
| 233 stats->ref[ALTREF_FRAME].err = INT_MAX; | |
| 234 stats->ref[ALTREF_FRAME].m.mv.as_int = 0; | |
| 235 } | |
| 236 } | |
| 237 | |
| 238 static void update_mbgraph_frame_stats(VP9_COMP *cpi, | |
| 239 MBGRAPH_FRAME_STATS *stats, | |
| 240 YV12_BUFFER_CONFIG *buf, | |
| 241 YV12_BUFFER_CONFIG *golden_ref, | |
| 242 YV12_BUFFER_CONFIG *alt_ref) { | |
| 243 MACROBLOCK *const x = &cpi->mb; | |
| 244 MACROBLOCKD *const xd = &x->e_mbd; | |
| 245 VP9_COMMON *const cm = &cpi->common; | |
| 246 | |
| 247 int mb_col, mb_row, offset = 0; | |
| 248 int mb_y_offset = 0, arf_y_offset = 0, gld_y_offset = 0; | |
| 249 int_mv arf_top_mv, gld_top_mv; | |
| 250 MODE_INFO mi_local = { { 0 } }; | |
| 251 | |
| 252 // Set up limit values for motion vectors to prevent them extending outside | |
| 253 // the UMV borders. | |
| 254 arf_top_mv.as_int = 0; | |
| 255 gld_top_mv.as_int = 0; | |
| 256 x->mv_row_min = -BORDER_MV_PIXELS_B16; | |
| 257 x->mv_row_max = (cm->mb_rows - 1) * 8 + BORDER_MV_PIXELS_B16; | |
| 258 xd->up_available = 0; | |
| 259 xd->plane[0].dst.stride = buf->y_stride; | |
| 260 xd->plane[0].pre[0].stride = buf->y_stride; | |
| 261 xd->plane[1].dst.stride = buf->uv_stride; | |
| 262 xd->mi_8x8[0] = &mi_local; | |
| 263 mi_local.mbmi.sb_type = BLOCK_16X16; | |
| 264 mi_local.mbmi.ref_frame[0] = LAST_FRAME; | |
| 265 mi_local.mbmi.ref_frame[1] = NONE; | |
| 266 | |
| 267 for (mb_row = 0; mb_row < cm->mb_rows; mb_row++) { | |
| 268 int_mv arf_left_mv, gld_left_mv; | |
| 269 int mb_y_in_offset = mb_y_offset; | |
| 270 int arf_y_in_offset = arf_y_offset; | |
| 271 int gld_y_in_offset = gld_y_offset; | |
| 272 | |
| 273 // Set up limit values for motion vectors to prevent them extending outside | |
| 274 // the UMV borders. | |
| 275 arf_left_mv.as_int = arf_top_mv.as_int; | |
| 276 gld_left_mv.as_int = gld_top_mv.as_int; | |
| 277 x->mv_col_min = -BORDER_MV_PIXELS_B16; | |
| 278 x->mv_col_max = (cm->mb_cols - 1) * 8 + BORDER_MV_PIXELS_B16; | |
| 279 xd->left_available = 0; | |
| 280 | |
| 281 for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) { | |
| 282 MBGRAPH_MB_STATS *mb_stats = &stats->mb_stats[offset + mb_col]; | |
| 283 | |
| 284 update_mbgraph_mb_stats(cpi, mb_stats, buf, mb_y_in_offset, | |
| 285 golden_ref, &gld_left_mv, gld_y_in_offset, | |
| 286 alt_ref, &arf_left_mv, arf_y_in_offset, | |
| 287 mb_row, mb_col); | |
| 288 arf_left_mv.as_int = mb_stats->ref[ALTREF_FRAME].m.mv.as_int; | |
| 289 gld_left_mv.as_int = mb_stats->ref[GOLDEN_FRAME].m.mv.as_int; | |
| 290 if (mb_col == 0) { | |
| 291 arf_top_mv.as_int = arf_left_mv.as_int; | |
| 292 gld_top_mv.as_int = gld_left_mv.as_int; | |
| 293 } | |
| 294 xd->left_available = 1; | |
| 295 mb_y_in_offset += 16; | |
| 296 gld_y_in_offset += 16; | |
| 297 arf_y_in_offset += 16; | |
| 298 x->mv_col_min -= 16; | |
| 299 x->mv_col_max -= 16; | |
| 300 } | |
| 301 xd->up_available = 1; | |
| 302 mb_y_offset += buf->y_stride * 16; | |
| 303 gld_y_offset += golden_ref->y_stride * 16; | |
| 304 if (alt_ref) | |
| 305 arf_y_offset += alt_ref->y_stride * 16; | |
| 306 x->mv_row_min -= 16; | |
| 307 x->mv_row_max -= 16; | |
| 308 offset += cm->mb_cols; | |
| 309 } | |
| 310 } | |
| 311 | |
| 312 // void separate_arf_mbs_byzz | |
| 313 static void separate_arf_mbs(VP9_COMP *cpi) { | |
| 314 VP9_COMMON *const cm = &cpi->common; | |
| 315 int mb_col, mb_row, offset, i; | |
| 316 int mi_row, mi_col; | |
| 317 int ncnt[4] = { 0 }; | |
| 318 int n_frames = cpi->mbgraph_n_frames; | |
| 319 | |
| 320 int *arf_not_zz; | |
| 321 | |
| 322 CHECK_MEM_ERROR(cm, arf_not_zz, | |
| 323 vpx_calloc(cm->mb_rows * cm->mb_cols * sizeof(*arf_not_zz), | |
| 324 1)); | |
| 325 | |
| 326 // We are not interested in results beyond the alt ref itself. | |
| 327 if (n_frames > cpi->frames_till_gf_update_due) | |
| 328 n_frames = cpi->frames_till_gf_update_due; | |
| 329 | |
| 330 // defer cost to reference frames | |
| 331 for (i = n_frames - 1; i >= 0; i--) { | |
| 332 MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i]; | |
| 333 | |
| 334 for (offset = 0, mb_row = 0; mb_row < cm->mb_rows; | |
| 335 offset += cm->mb_cols, mb_row++) { | |
| 336 for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) { | |
| 337 MBGRAPH_MB_STATS *mb_stats = &frame_stats->mb_stats[offset + mb_col]; | |
| 338 | |
| 339 int altref_err = mb_stats->ref[ALTREF_FRAME].err; | |
| 340 int intra_err = mb_stats->ref[INTRA_FRAME ].err; | |
| 341 int golden_err = mb_stats->ref[GOLDEN_FRAME].err; | |
| 342 | |
| 343 // Test for altref vs intra and gf and that its mv was 0,0. | |
| 344 if (altref_err > 1000 || | |
| 345 altref_err > intra_err || | |
| 346 altref_err > golden_err) { | |
| 347 arf_not_zz[offset + mb_col]++; | |
| 348 } | |
| 349 } | |
| 350 } | |
| 351 } | |
| 352 | |
| 353 // arf_not_zz is indexed by MB, but this loop is indexed by MI to avoid out | |
| 354 // of bound access in segmentation_map | |
| 355 for (mi_row = 0; mi_row < cm->mi_rows; mi_row++) { | |
| 356 for (mi_col = 0; mi_col < cm->mi_cols; mi_col++) { | |
| 357 // If any of the blocks in the sequence failed then the MB | |
| 358 // goes in segment 0 | |
| 359 if (arf_not_zz[mi_row/2*cm->mb_cols + mi_col/2]) { | |
| 360 ncnt[0]++; | |
| 361 cpi->segmentation_map[mi_row * cm->mi_cols + mi_col] = 0; | |
| 362 } else { | |
| 363 cpi->segmentation_map[mi_row * cm->mi_cols + mi_col] = 1; | |
| 364 ncnt[1]++; | |
| 365 } | |
| 366 } | |
| 367 } | |
| 368 | |
| 369 // Only bother with segmentation if over 10% of the MBs in static segment | |
| 370 // if ( ncnt[1] && (ncnt[0] / ncnt[1] < 10) ) | |
| 371 if (1) { | |
| 372 // Note % of blocks that are marked as static | |
| 373 if (cm->MBs) | |
| 374 cpi->static_mb_pct = (ncnt[1] * 100) / (cm->mi_rows * cm->mi_cols); | |
| 375 | |
| 376 // This error case should not be reachable as this function should | |
| 377 // never be called with the common data structure uninitialized. | |
| 378 else | |
| 379 cpi->static_mb_pct = 0; | |
| 380 | |
| 381 cpi->seg0_cnt = ncnt[0]; | |
| 382 vp9_enable_segmentation((VP9_PTR)cpi); | |
| 383 } else { | |
| 384 cpi->static_mb_pct = 0; | |
| 385 vp9_disable_segmentation((VP9_PTR)cpi); | |
| 386 } | |
| 387 | |
| 388 // Free localy allocated storage | |
| 389 vpx_free(arf_not_zz); | |
| 390 } | |
| 391 | |
| 392 void vp9_update_mbgraph_stats(VP9_COMP *cpi) { | |
| 393 VP9_COMMON *const cm = &cpi->common; | |
| 394 int i, n_frames = vp9_lookahead_depth(cpi->lookahead); | |
| 395 YV12_BUFFER_CONFIG *golden_ref = | |
| 396 &cm->yv12_fb[cm->ref_frame_map[cpi->gld_fb_idx]]; | |
| 397 | |
| 398 // we need to look ahead beyond where the ARF transitions into | |
| 399 // being a GF - so exit if we don't look ahead beyond that | |
| 400 if (n_frames <= cpi->frames_till_gf_update_due) | |
| 401 return; | |
| 402 if (n_frames > (int)cpi->frames_till_alt_ref_frame) | |
| 403 n_frames = cpi->frames_till_alt_ref_frame; | |
| 404 if (n_frames > MAX_LAG_BUFFERS) | |
| 405 n_frames = MAX_LAG_BUFFERS; | |
| 406 | |
| 407 cpi->mbgraph_n_frames = n_frames; | |
| 408 for (i = 0; i < n_frames; i++) { | |
| 409 MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i]; | |
| 410 vpx_memset(frame_stats->mb_stats, 0, | |
| 411 cm->mb_rows * cm->mb_cols * | |
| 412 sizeof(*cpi->mbgraph_stats[i].mb_stats)); | |
| 413 } | |
| 414 | |
| 415 // do motion search to find contribution of each reference to data | |
| 416 // later on in this GF group | |
| 417 // FIXME really, the GF/last MC search should be done forward, and | |
| 418 // the ARF MC search backwards, to get optimal results for MV caching | |
| 419 for (i = 0; i < n_frames; i++) { | |
| 420 MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i]; | |
| 421 struct lookahead_entry *q_cur = vp9_lookahead_peek(cpi->lookahead, i); | |
| 422 | |
| 423 assert(q_cur != NULL); | |
| 424 | |
| 425 update_mbgraph_frame_stats(cpi, frame_stats, &q_cur->img, | |
| 426 golden_ref, cpi->Source); | |
| 427 } | |
| 428 | |
| 429 vp9_clear_system_state(); // __asm emms; | |
| 430 | |
| 431 separate_arf_mbs(cpi); | |
| 432 } |
