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_segmentation.h"
15 #include "vp9/encoder/vp9_mcomp.h"
16 #include "vp9/common/vp9_blockd.h"
17 #include "vp9/common/vp9_reconinter.h"
18 #include "vp9/common/vp9_reconintra.h"
19 #include "vp9/common/vp9_systemdependent.h"
20
21
do_16x16_motion_iteration(VP9_COMP * cpi,const MV * ref_mv,MV * dst_mv,int mb_row,int mb_col)22 static unsigned int do_16x16_motion_iteration(VP9_COMP *cpi,
23 const MV *ref_mv,
24 MV *dst_mv,
25 int mb_row,
26 int mb_col) {
27 MACROBLOCK *const x = &cpi->mb;
28 MACROBLOCKD *const xd = &x->e_mbd;
29 const MV_SPEED_FEATURES *const mv_sf = &cpi->sf.mv;
30 const vp9_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[BLOCK_16X16];
31
32 const int tmp_col_min = x->mv_col_min;
33 const int tmp_col_max = x->mv_col_max;
34 const int tmp_row_min = x->mv_row_min;
35 const int tmp_row_max = x->mv_row_max;
36 MV ref_full;
37 int sad_list[5];
38
39 // Further step/diamond searches as necessary
40 int step_param = mv_sf->reduce_first_step_size;
41 step_param = MIN(step_param, MAX_MVSEARCH_STEPS - 2);
42
43 vp9_set_mv_search_range(x, ref_mv);
44
45 ref_full.col = ref_mv->col >> 3;
46 ref_full.row = ref_mv->row >> 3;
47
48 /*cpi->sf.search_method == HEX*/
49 vp9_hex_search(x, &ref_full, step_param, x->errorperbit, 0,
50 cond_sad_list(cpi, sad_list),
51 &v_fn_ptr, 0, ref_mv, dst_mv);
52
53 // Try sub-pixel MC
54 // if (bestsme > error_thresh && bestsme < INT_MAX)
55 {
56 int distortion;
57 unsigned int sse;
58 cpi->find_fractional_mv_step(
59 x, dst_mv, ref_mv, cpi->common.allow_high_precision_mv, x->errorperbit,
60 &v_fn_ptr, 0, mv_sf->subpel_iters_per_step,
61 cond_sad_list(cpi, sad_list),
62 NULL, NULL,
63 &distortion, &sse, NULL, 0, 0);
64 }
65
66 xd->mi[0].src_mi->mbmi.mode = NEWMV;
67 xd->mi[0].src_mi->mbmi.mv[0].as_mv = *dst_mv;
68
69 vp9_build_inter_predictors_sby(xd, mb_row, mb_col, BLOCK_16X16);
70
71 /* restore UMV window */
72 x->mv_col_min = tmp_col_min;
73 x->mv_col_max = tmp_col_max;
74 x->mv_row_min = tmp_row_min;
75 x->mv_row_max = tmp_row_max;
76
77 return vp9_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
78 xd->plane[0].dst.buf, xd->plane[0].dst.stride);
79 }
80
do_16x16_motion_search(VP9_COMP * cpi,const MV * ref_mv,int_mv * dst_mv,int mb_row,int mb_col)81 static int do_16x16_motion_search(VP9_COMP *cpi, const MV *ref_mv,
82 int_mv *dst_mv, int mb_row, int mb_col) {
83 MACROBLOCK *const x = &cpi->mb;
84 MACROBLOCKD *const xd = &x->e_mbd;
85 unsigned int err, tmp_err;
86 MV tmp_mv;
87
88 // Try zero MV first
89 // FIXME should really use something like near/nearest MV and/or MV prediction
90 err = vp9_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
91 xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride);
92 dst_mv->as_int = 0;
93
94 // Test last reference frame using the previous best mv as the
95 // starting point (best reference) for the search
96 tmp_err = do_16x16_motion_iteration(cpi, ref_mv, &tmp_mv, mb_row, mb_col);
97 if (tmp_err < err) {
98 err = tmp_err;
99 dst_mv->as_mv = tmp_mv;
100 }
101
102 // If the current best reference mv is not centered on 0,0 then do a 0,0
103 // based search as well.
104 if (ref_mv->row != 0 || ref_mv->col != 0) {
105 unsigned int tmp_err;
106 MV zero_ref_mv = {0, 0}, tmp_mv;
107
108 tmp_err = do_16x16_motion_iteration(cpi, &zero_ref_mv, &tmp_mv,
109 mb_row, mb_col);
110 if (tmp_err < err) {
111 dst_mv->as_mv = tmp_mv;
112 err = tmp_err;
113 }
114 }
115
116 return err;
117 }
118
do_16x16_zerozero_search(VP9_COMP * cpi,int_mv * dst_mv)119 static int do_16x16_zerozero_search(VP9_COMP *cpi, int_mv *dst_mv) {
120 MACROBLOCK *const x = &cpi->mb;
121 MACROBLOCKD *const xd = &x->e_mbd;
122 unsigned int err;
123
124 // Try zero MV first
125 // FIXME should really use something like near/nearest MV and/or MV prediction
126 err = vp9_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
127 xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride);
128
129 dst_mv->as_int = 0;
130
131 return err;
132 }
find_best_16x16_intra(VP9_COMP * cpi,PREDICTION_MODE * pbest_mode)133 static int find_best_16x16_intra(VP9_COMP *cpi, PREDICTION_MODE *pbest_mode) {
134 MACROBLOCK *const x = &cpi->mb;
135 MACROBLOCKD *const xd = &x->e_mbd;
136 PREDICTION_MODE best_mode = -1, mode;
137 unsigned int best_err = INT_MAX;
138
139 // calculate SATD for each intra prediction mode;
140 // we're intentionally not doing 4x4, we just want a rough estimate
141 for (mode = DC_PRED; mode <= TM_PRED; mode++) {
142 unsigned int err;
143
144 xd->mi[0].src_mi->mbmi.mode = mode;
145 vp9_predict_intra_block(xd, 0, 2, TX_16X16, mode,
146 x->plane[0].src.buf, x->plane[0].src.stride,
147 xd->plane[0].dst.buf, xd->plane[0].dst.stride,
148 0, 0, 0);
149 err = vp9_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
150 xd->plane[0].dst.buf, xd->plane[0].dst.stride);
151
152 // find best
153 if (err < best_err) {
154 best_err = err;
155 best_mode = mode;
156 }
157 }
158
159 if (pbest_mode)
160 *pbest_mode = best_mode;
161
162 return best_err;
163 }
164
update_mbgraph_mb_stats(VP9_COMP * cpi,MBGRAPH_MB_STATS * stats,YV12_BUFFER_CONFIG * buf,int mb_y_offset,YV12_BUFFER_CONFIG * golden_ref,const MV * prev_golden_ref_mv,YV12_BUFFER_CONFIG * alt_ref,int mb_row,int mb_col)165 static void update_mbgraph_mb_stats
166 (
167 VP9_COMP *cpi,
168 MBGRAPH_MB_STATS *stats,
169 YV12_BUFFER_CONFIG *buf,
170 int mb_y_offset,
171 YV12_BUFFER_CONFIG *golden_ref,
172 const MV *prev_golden_ref_mv,
173 YV12_BUFFER_CONFIG *alt_ref,
174 int mb_row,
175 int mb_col
176 ) {
177 MACROBLOCK *const x = &cpi->mb;
178 MACROBLOCKD *const xd = &x->e_mbd;
179 int intra_error;
180 VP9_COMMON *cm = &cpi->common;
181
182 // FIXME in practice we're completely ignoring chroma here
183 x->plane[0].src.buf = buf->y_buffer + mb_y_offset;
184 x->plane[0].src.stride = buf->y_stride;
185
186 xd->plane[0].dst.buf = get_frame_new_buffer(cm)->y_buffer + mb_y_offset;
187 xd->plane[0].dst.stride = get_frame_new_buffer(cm)->y_stride;
188
189 // do intra 16x16 prediction
190 intra_error = find_best_16x16_intra(cpi,
191 &stats->ref[INTRA_FRAME].m.mode);
192 if (intra_error <= 0)
193 intra_error = 1;
194 stats->ref[INTRA_FRAME].err = intra_error;
195
196 // Golden frame MV search, if it exists and is different than last frame
197 if (golden_ref) {
198 int g_motion_error;
199 xd->plane[0].pre[0].buf = golden_ref->y_buffer + mb_y_offset;
200 xd->plane[0].pre[0].stride = golden_ref->y_stride;
201 g_motion_error = do_16x16_motion_search(cpi,
202 prev_golden_ref_mv,
203 &stats->ref[GOLDEN_FRAME].m.mv,
204 mb_row, mb_col);
205 stats->ref[GOLDEN_FRAME].err = g_motion_error;
206 } else {
207 stats->ref[GOLDEN_FRAME].err = INT_MAX;
208 stats->ref[GOLDEN_FRAME].m.mv.as_int = 0;
209 }
210
211 // Do an Alt-ref frame MV search, if it exists and is different than
212 // last/golden frame.
213 if (alt_ref) {
214 int a_motion_error;
215 xd->plane[0].pre[0].buf = alt_ref->y_buffer + mb_y_offset;
216 xd->plane[0].pre[0].stride = alt_ref->y_stride;
217 a_motion_error = do_16x16_zerozero_search(cpi,
218 &stats->ref[ALTREF_FRAME].m.mv);
219
220 stats->ref[ALTREF_FRAME].err = a_motion_error;
221 } else {
222 stats->ref[ALTREF_FRAME].err = INT_MAX;
223 stats->ref[ALTREF_FRAME].m.mv.as_int = 0;
224 }
225 }
226
update_mbgraph_frame_stats(VP9_COMP * cpi,MBGRAPH_FRAME_STATS * stats,YV12_BUFFER_CONFIG * buf,YV12_BUFFER_CONFIG * golden_ref,YV12_BUFFER_CONFIG * alt_ref)227 static void update_mbgraph_frame_stats(VP9_COMP *cpi,
228 MBGRAPH_FRAME_STATS *stats,
229 YV12_BUFFER_CONFIG *buf,
230 YV12_BUFFER_CONFIG *golden_ref,
231 YV12_BUFFER_CONFIG *alt_ref) {
232 MACROBLOCK *const x = &cpi->mb;
233 MACROBLOCKD *const xd = &x->e_mbd;
234 VP9_COMMON *const cm = &cpi->common;
235
236 int mb_col, mb_row, offset = 0;
237 int mb_y_offset = 0, arf_y_offset = 0, gld_y_offset = 0;
238 MV gld_top_mv = {0, 0};
239 MODE_INFO mi_local;
240
241 vp9_zero(mi_local);
242 // Set up limit values for motion vectors to prevent them extending outside
243 // the UMV borders.
244 x->mv_row_min = -BORDER_MV_PIXELS_B16;
245 x->mv_row_max = (cm->mb_rows - 1) * 8 + BORDER_MV_PIXELS_B16;
246 xd->up_available = 0;
247 xd->plane[0].dst.stride = buf->y_stride;
248 xd->plane[0].pre[0].stride = buf->y_stride;
249 xd->plane[1].dst.stride = buf->uv_stride;
250 xd->mi[0].src_mi = &mi_local;
251 mi_local.mbmi.sb_type = BLOCK_16X16;
252 mi_local.mbmi.ref_frame[0] = LAST_FRAME;
253 mi_local.mbmi.ref_frame[1] = NONE;
254
255 for (mb_row = 0; mb_row < cm->mb_rows; mb_row++) {
256 MV gld_left_mv = gld_top_mv;
257 int mb_y_in_offset = mb_y_offset;
258 int arf_y_in_offset = arf_y_offset;
259 int gld_y_in_offset = gld_y_offset;
260
261 // Set up limit values for motion vectors to prevent them extending outside
262 // the UMV borders.
263 x->mv_col_min = -BORDER_MV_PIXELS_B16;
264 x->mv_col_max = (cm->mb_cols - 1) * 8 + BORDER_MV_PIXELS_B16;
265 xd->left_available = 0;
266
267 for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
268 MBGRAPH_MB_STATS *mb_stats = &stats->mb_stats[offset + mb_col];
269
270 update_mbgraph_mb_stats(cpi, mb_stats, buf, mb_y_in_offset,
271 golden_ref, &gld_left_mv, alt_ref,
272 mb_row, mb_col);
273 gld_left_mv = mb_stats->ref[GOLDEN_FRAME].m.mv.as_mv;
274 if (mb_col == 0) {
275 gld_top_mv = gld_left_mv;
276 }
277 xd->left_available = 1;
278 mb_y_in_offset += 16;
279 gld_y_in_offset += 16;
280 arf_y_in_offset += 16;
281 x->mv_col_min -= 16;
282 x->mv_col_max -= 16;
283 }
284 xd->up_available = 1;
285 mb_y_offset += buf->y_stride * 16;
286 gld_y_offset += golden_ref->y_stride * 16;
287 if (alt_ref)
288 arf_y_offset += alt_ref->y_stride * 16;
289 x->mv_row_min -= 16;
290 x->mv_row_max -= 16;
291 offset += cm->mb_cols;
292 }
293 }
294
295 // void separate_arf_mbs_byzz
separate_arf_mbs(VP9_COMP * cpi)296 static void separate_arf_mbs(VP9_COMP *cpi) {
297 VP9_COMMON *const cm = &cpi->common;
298 int mb_col, mb_row, offset, i;
299 int mi_row, mi_col;
300 int ncnt[4] = { 0 };
301 int n_frames = cpi->mbgraph_n_frames;
302
303 int *arf_not_zz;
304
305 CHECK_MEM_ERROR(cm, arf_not_zz,
306 vpx_calloc(cm->mb_rows * cm->mb_cols * sizeof(*arf_not_zz),
307 1));
308
309 // We are not interested in results beyond the alt ref itself.
310 if (n_frames > cpi->rc.frames_till_gf_update_due)
311 n_frames = cpi->rc.frames_till_gf_update_due;
312
313 // defer cost to reference frames
314 for (i = n_frames - 1; i >= 0; i--) {
315 MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
316
317 for (offset = 0, mb_row = 0; mb_row < cm->mb_rows;
318 offset += cm->mb_cols, mb_row++) {
319 for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
320 MBGRAPH_MB_STATS *mb_stats = &frame_stats->mb_stats[offset + mb_col];
321
322 int altref_err = mb_stats->ref[ALTREF_FRAME].err;
323 int intra_err = mb_stats->ref[INTRA_FRAME ].err;
324 int golden_err = mb_stats->ref[GOLDEN_FRAME].err;
325
326 // Test for altref vs intra and gf and that its mv was 0,0.
327 if (altref_err > 1000 ||
328 altref_err > intra_err ||
329 altref_err > golden_err) {
330 arf_not_zz[offset + mb_col]++;
331 }
332 }
333 }
334 }
335
336 // arf_not_zz is indexed by MB, but this loop is indexed by MI to avoid out
337 // of bound access in segmentation_map
338 for (mi_row = 0; mi_row < cm->mi_rows; mi_row++) {
339 for (mi_col = 0; mi_col < cm->mi_cols; mi_col++) {
340 // If any of the blocks in the sequence failed then the MB
341 // goes in segment 0
342 if (arf_not_zz[mi_row / 2 * cm->mb_cols + mi_col / 2]) {
343 ncnt[0]++;
344 cpi->segmentation_map[mi_row * cm->mi_cols + mi_col] = 0;
345 } else {
346 cpi->segmentation_map[mi_row * cm->mi_cols + mi_col] = 1;
347 ncnt[1]++;
348 }
349 }
350 }
351
352 // Only bother with segmentation if over 10% of the MBs in static segment
353 // if ( ncnt[1] && (ncnt[0] / ncnt[1] < 10) )
354 if (1) {
355 // Note % of blocks that are marked as static
356 if (cm->MBs)
357 cpi->static_mb_pct = (ncnt[1] * 100) / (cm->mi_rows * cm->mi_cols);
358
359 // This error case should not be reachable as this function should
360 // never be called with the common data structure uninitialized.
361 else
362 cpi->static_mb_pct = 0;
363
364 vp9_enable_segmentation(&cm->seg);
365 } else {
366 cpi->static_mb_pct = 0;
367 vp9_disable_segmentation(&cm->seg);
368 }
369
370 // Free localy allocated storage
371 vpx_free(arf_not_zz);
372 }
373
vp9_update_mbgraph_stats(VP9_COMP * cpi)374 void vp9_update_mbgraph_stats(VP9_COMP *cpi) {
375 VP9_COMMON *const cm = &cpi->common;
376 int i, n_frames = vp9_lookahead_depth(cpi->lookahead);
377 YV12_BUFFER_CONFIG *golden_ref = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
378
379 // we need to look ahead beyond where the ARF transitions into
380 // being a GF - so exit if we don't look ahead beyond that
381 if (n_frames <= cpi->rc.frames_till_gf_update_due)
382 return;
383
384 if (n_frames > MAX_LAG_BUFFERS)
385 n_frames = MAX_LAG_BUFFERS;
386
387 cpi->mbgraph_n_frames = n_frames;
388 for (i = 0; i < n_frames; i++) {
389 MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
390 vpx_memset(frame_stats->mb_stats, 0,
391 cm->mb_rows * cm->mb_cols *
392 sizeof(*cpi->mbgraph_stats[i].mb_stats));
393 }
394
395 // do motion search to find contribution of each reference to data
396 // later on in this GF group
397 // FIXME really, the GF/last MC search should be done forward, and
398 // the ARF MC search backwards, to get optimal results for MV caching
399 for (i = 0; i < n_frames; i++) {
400 MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
401 struct lookahead_entry *q_cur = vp9_lookahead_peek(cpi->lookahead, i);
402
403 assert(q_cur != NULL);
404
405 update_mbgraph_frame_stats(cpi, frame_stats, &q_cur->img,
406 golden_ref, cpi->Source);
407 }
408
409 vp9_clear_system_state();
410
411 separate_arf_mbs(cpi);
412 }
413