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_rdopt.h"
15 #include "vp9/encoder/vp9_segmentation.h"
16 #include "vp9/encoder/vp9_mcomp.h"
17 #include "vp9/common/vp9_blockd.h"
18 #include "vp9/common/vp9_reconinter.h"
19 #include "vp9/common/vp9_reconintra.h"
20 #include "vp9/common/vp9_systemdependent.h"
21
22
23
do_16x16_motion_iteration(VP9_COMP * cpi,const MV * ref_mv,MV * dst_mv,int mb_row,int mb_col)24 static unsigned int do_16x16_motion_iteration(VP9_COMP *cpi,
25 const MV *ref_mv,
26 MV *dst_mv,
27 int mb_row,
28 int mb_col) {
29 MACROBLOCK *const x = &cpi->mb;
30 MACROBLOCKD *const xd = &x->e_mbd;
31 vp9_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[BLOCK_16X16];
32
33 const int tmp_col_min = x->mv_col_min;
34 const int tmp_col_max = x->mv_col_max;
35 const int tmp_row_min = x->mv_row_min;
36 const int tmp_row_max = x->mv_row_max;
37 MV ref_full;
38
39 // Further step/diamond searches as necessary
40 int step_param = cpi->sf.reduce_first_step_size +
41 (cpi->speed < 8 ? (cpi->speed > 5 ? 1 : 0) : 2);
42 step_param = MIN(step_param, (cpi->sf.max_step_search_steps - 2));
43
44 vp9_set_mv_search_range(x, ref_mv);
45
46 ref_full.col = ref_mv->col >> 3;
47 ref_full.row = ref_mv->row >> 3;
48
49 /*cpi->sf.search_method == HEX*/
50 vp9_hex_search(x, &ref_full, step_param, x->errorperbit, 0, &v_fn_ptr, 0,
51 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, cpi->sf.subpel_iters_per_step, NULL, NULL, &distortion,
61 &sse);
62 }
63
64 xd->mi[0]->mbmi.mode = NEWMV;
65 xd->mi[0]->mbmi.mv[0].as_mv = *dst_mv;
66
67 vp9_build_inter_predictors_sby(xd, mb_row, mb_col, BLOCK_16X16);
68
69 /* restore UMV window */
70 x->mv_col_min = tmp_col_min;
71 x->mv_col_max = tmp_col_max;
72 x->mv_row_min = tmp_row_min;
73 x->mv_row_max = tmp_row_max;
74
75 return vp9_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
76 xd->plane[0].dst.buf, xd->plane[0].dst.stride,
77 INT_MAX);
78 }
79
do_16x16_motion_search(VP9_COMP * cpi,const int_mv * ref_mv,int_mv * dst_mv,int mb_row,int mb_col)80 static int do_16x16_motion_search(VP9_COMP *cpi, const int_mv *ref_mv,
81 int_mv *dst_mv, int mb_row, int mb_col) {
82 MACROBLOCK *const x = &cpi->mb;
83 MACROBLOCKD *const xd = &x->e_mbd;
84 unsigned int err, tmp_err;
85 int_mv tmp_mv;
86
87 // Try zero MV first
88 // FIXME should really use something like near/nearest MV and/or MV prediction
89 err = vp9_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
90 xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride,
91 INT_MAX);
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->as_mv, &tmp_mv.as_mv,
97 mb_row, mb_col);
98 if (tmp_err < err) {
99 err = tmp_err;
100 dst_mv->as_int = tmp_mv.as_int;
101 }
102
103 // If the current best reference mv is not centered on 0,0 then do a 0,0
104 // based search as well.
105 if (ref_mv->as_int) {
106 unsigned int tmp_err;
107 int_mv zero_ref_mv, tmp_mv;
108
109 zero_ref_mv.as_int = 0;
110 tmp_err = do_16x16_motion_iteration(cpi, &zero_ref_mv.as_mv, &tmp_mv.as_mv,
111 mb_row, mb_col);
112 if (tmp_err < err) {
113 dst_mv->as_int = tmp_mv.as_int;
114 err = tmp_err;
115 }
116 }
117
118 return err;
119 }
120
do_16x16_zerozero_search(VP9_COMP * cpi,int_mv * dst_mv)121 static int do_16x16_zerozero_search(VP9_COMP *cpi, int_mv *dst_mv) {
122 MACROBLOCK *const x = &cpi->mb;
123 MACROBLOCKD *const xd = &x->e_mbd;
124 unsigned int err;
125
126 // Try zero MV first
127 // FIXME should really use something like near/nearest MV and/or MV prediction
128 err = vp9_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
129 xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride,
130 INT_MAX);
131
132 dst_mv->as_int = 0;
133
134 return err;
135 }
find_best_16x16_intra(VP9_COMP * cpi,MB_PREDICTION_MODE * pbest_mode)136 static int find_best_16x16_intra(VP9_COMP *cpi,
137 MB_PREDICTION_MODE *pbest_mode) {
138 MACROBLOCK *const x = &cpi->mb;
139 MACROBLOCKD *const xd = &x->e_mbd;
140 MB_PREDICTION_MODE best_mode = -1, mode;
141 unsigned int best_err = INT_MAX;
142
143 // calculate SATD for each intra prediction mode;
144 // we're intentionally not doing 4x4, we just want a rough estimate
145 for (mode = DC_PRED; mode <= TM_PRED; mode++) {
146 unsigned int err;
147
148 xd->mi[0]->mbmi.mode = mode;
149 vp9_predict_intra_block(xd, 0, 2, TX_16X16, mode,
150 x->plane[0].src.buf, x->plane[0].src.stride,
151 xd->plane[0].dst.buf, xd->plane[0].dst.stride,
152 0, 0, 0);
153 err = vp9_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
154 xd->plane[0].dst.buf, xd->plane[0].dst.stride, best_err);
155
156 // find best
157 if (err < best_err) {
158 best_err = err;
159 best_mode = mode;
160 }
161 }
162
163 if (pbest_mode)
164 *pbest_mode = best_mode;
165
166 return best_err;
167 }
168
update_mbgraph_mb_stats(VP9_COMP * cpi,MBGRAPH_MB_STATS * stats,YV12_BUFFER_CONFIG * buf,int mb_y_offset,YV12_BUFFER_CONFIG * golden_ref,int_mv * prev_golden_ref_mv,YV12_BUFFER_CONFIG * alt_ref,int mb_row,int mb_col)169 static void update_mbgraph_mb_stats
170 (
171 VP9_COMP *cpi,
172 MBGRAPH_MB_STATS *stats,
173 YV12_BUFFER_CONFIG *buf,
174 int mb_y_offset,
175 YV12_BUFFER_CONFIG *golden_ref,
176 int_mv *prev_golden_ref_mv,
177 YV12_BUFFER_CONFIG *alt_ref,
178 int mb_row,
179 int mb_col
180 ) {
181 MACROBLOCK *const x = &cpi->mb;
182 MACROBLOCKD *const xd = &x->e_mbd;
183 int intra_error;
184 VP9_COMMON *cm = &cpi->common;
185
186 // FIXME in practice we're completely ignoring chroma here
187 x->plane[0].src.buf = buf->y_buffer + mb_y_offset;
188 x->plane[0].src.stride = buf->y_stride;
189
190 xd->plane[0].dst.buf = get_frame_new_buffer(cm)->y_buffer + mb_y_offset;
191 xd->plane[0].dst.stride = get_frame_new_buffer(cm)->y_stride;
192
193 // do intra 16x16 prediction
194 intra_error = find_best_16x16_intra(cpi,
195 &stats->ref[INTRA_FRAME].m.mode);
196 if (intra_error <= 0)
197 intra_error = 1;
198 stats->ref[INTRA_FRAME].err = intra_error;
199
200 // Golden frame MV search, if it exists and is different than last frame
201 if (golden_ref) {
202 int g_motion_error;
203 xd->plane[0].pre[0].buf = golden_ref->y_buffer + mb_y_offset;
204 xd->plane[0].pre[0].stride = golden_ref->y_stride;
205 g_motion_error = do_16x16_motion_search(cpi,
206 prev_golden_ref_mv,
207 &stats->ref[GOLDEN_FRAME].m.mv,
208 mb_row, mb_col);
209 stats->ref[GOLDEN_FRAME].err = g_motion_error;
210 } else {
211 stats->ref[GOLDEN_FRAME].err = INT_MAX;
212 stats->ref[GOLDEN_FRAME].m.mv.as_int = 0;
213 }
214
215 // Do an Alt-ref frame MV search, if it exists and is different than
216 // last/golden frame.
217 if (alt_ref) {
218 int a_motion_error;
219 xd->plane[0].pre[0].buf = alt_ref->y_buffer + mb_y_offset;
220 xd->plane[0].pre[0].stride = alt_ref->y_stride;
221 a_motion_error = do_16x16_zerozero_search(cpi,
222 &stats->ref[ALTREF_FRAME].m.mv);
223
224 stats->ref[ALTREF_FRAME].err = a_motion_error;
225 } else {
226 stats->ref[ALTREF_FRAME].err = INT_MAX;
227 stats->ref[ALTREF_FRAME].m.mv.as_int = 0;
228 }
229 }
230
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)231 static void update_mbgraph_frame_stats(VP9_COMP *cpi,
232 MBGRAPH_FRAME_STATS *stats,
233 YV12_BUFFER_CONFIG *buf,
234 YV12_BUFFER_CONFIG *golden_ref,
235 YV12_BUFFER_CONFIG *alt_ref) {
236 MACROBLOCK *const x = &cpi->mb;
237 MACROBLOCKD *const xd = &x->e_mbd;
238 VP9_COMMON *const cm = &cpi->common;
239
240 int mb_col, mb_row, offset = 0;
241 int mb_y_offset = 0, arf_y_offset = 0, gld_y_offset = 0;
242 int_mv arf_top_mv, gld_top_mv;
243 MODE_INFO mi_local = { { 0 } };
244
245 // Set up limit values for motion vectors to prevent them extending outside
246 // the UMV borders.
247 arf_top_mv.as_int = 0;
248 gld_top_mv.as_int = 0;
249 x->mv_row_min = -BORDER_MV_PIXELS_B16;
250 x->mv_row_max = (cm->mb_rows - 1) * 8 + BORDER_MV_PIXELS_B16;
251 xd->up_available = 0;
252 xd->plane[0].dst.stride = buf->y_stride;
253 xd->plane[0].pre[0].stride = buf->y_stride;
254 xd->plane[1].dst.stride = buf->uv_stride;
255 xd->mi[0] = &mi_local;
256 mi_local.mbmi.sb_type = BLOCK_16X16;
257 mi_local.mbmi.ref_frame[0] = LAST_FRAME;
258 mi_local.mbmi.ref_frame[1] = NONE;
259
260 for (mb_row = 0; mb_row < cm->mb_rows; mb_row++) {
261 int_mv arf_left_mv, gld_left_mv;
262 int mb_y_in_offset = mb_y_offset;
263 int arf_y_in_offset = arf_y_offset;
264 int gld_y_in_offset = gld_y_offset;
265
266 // Set up limit values for motion vectors to prevent them extending outside
267 // the UMV borders.
268 arf_left_mv.as_int = arf_top_mv.as_int;
269 gld_left_mv.as_int = gld_top_mv.as_int;
270 x->mv_col_min = -BORDER_MV_PIXELS_B16;
271 x->mv_col_max = (cm->mb_cols - 1) * 8 + BORDER_MV_PIXELS_B16;
272 xd->left_available = 0;
273
274 for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
275 MBGRAPH_MB_STATS *mb_stats = &stats->mb_stats[offset + mb_col];
276
277 update_mbgraph_mb_stats(cpi, mb_stats, buf, mb_y_in_offset,
278 golden_ref, &gld_left_mv, alt_ref,
279 mb_row, mb_col);
280 arf_left_mv.as_int = mb_stats->ref[ALTREF_FRAME].m.mv.as_int;
281 gld_left_mv.as_int = mb_stats->ref[GOLDEN_FRAME].m.mv.as_int;
282 if (mb_col == 0) {
283 arf_top_mv.as_int = arf_left_mv.as_int;
284 gld_top_mv.as_int = gld_left_mv.as_int;
285 }
286 xd->left_available = 1;
287 mb_y_in_offset += 16;
288 gld_y_in_offset += 16;
289 arf_y_in_offset += 16;
290 x->mv_col_min -= 16;
291 x->mv_col_max -= 16;
292 }
293 xd->up_available = 1;
294 mb_y_offset += buf->y_stride * 16;
295 gld_y_offset += golden_ref->y_stride * 16;
296 if (alt_ref)
297 arf_y_offset += alt_ref->y_stride * 16;
298 x->mv_row_min -= 16;
299 x->mv_row_max -= 16;
300 offset += cm->mb_cols;
301 }
302 }
303
304 // void separate_arf_mbs_byzz
separate_arf_mbs(VP9_COMP * cpi)305 static void separate_arf_mbs(VP9_COMP *cpi) {
306 VP9_COMMON *const cm = &cpi->common;
307 int mb_col, mb_row, offset, i;
308 int mi_row, mi_col;
309 int ncnt[4] = { 0 };
310 int n_frames = cpi->mbgraph_n_frames;
311
312 int *arf_not_zz;
313
314 CHECK_MEM_ERROR(cm, arf_not_zz,
315 vpx_calloc(cm->mb_rows * cm->mb_cols * sizeof(*arf_not_zz),
316 1));
317
318 // We are not interested in results beyond the alt ref itself.
319 if (n_frames > cpi->rc.frames_till_gf_update_due)
320 n_frames = cpi->rc.frames_till_gf_update_due;
321
322 // defer cost to reference frames
323 for (i = n_frames - 1; i >= 0; i--) {
324 MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
325
326 for (offset = 0, mb_row = 0; mb_row < cm->mb_rows;
327 offset += cm->mb_cols, mb_row++) {
328 for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
329 MBGRAPH_MB_STATS *mb_stats = &frame_stats->mb_stats[offset + mb_col];
330
331 int altref_err = mb_stats->ref[ALTREF_FRAME].err;
332 int intra_err = mb_stats->ref[INTRA_FRAME ].err;
333 int golden_err = mb_stats->ref[GOLDEN_FRAME].err;
334
335 // Test for altref vs intra and gf and that its mv was 0,0.
336 if (altref_err > 1000 ||
337 altref_err > intra_err ||
338 altref_err > golden_err) {
339 arf_not_zz[offset + mb_col]++;
340 }
341 }
342 }
343 }
344
345 // arf_not_zz is indexed by MB, but this loop is indexed by MI to avoid out
346 // of bound access in segmentation_map
347 for (mi_row = 0; mi_row < cm->mi_rows; mi_row++) {
348 for (mi_col = 0; mi_col < cm->mi_cols; mi_col++) {
349 // If any of the blocks in the sequence failed then the MB
350 // goes in segment 0
351 if (arf_not_zz[mi_row / 2 * cm->mb_cols + mi_col / 2]) {
352 ncnt[0]++;
353 cpi->segmentation_map[mi_row * cm->mi_cols + mi_col] = 0;
354 } else {
355 cpi->segmentation_map[mi_row * cm->mi_cols + mi_col] = 1;
356 ncnt[1]++;
357 }
358 }
359 }
360
361 // Only bother with segmentation if over 10% of the MBs in static segment
362 // if ( ncnt[1] && (ncnt[0] / ncnt[1] < 10) )
363 if (1) {
364 // Note % of blocks that are marked as static
365 if (cm->MBs)
366 cpi->static_mb_pct = (ncnt[1] * 100) / (cm->mi_rows * cm->mi_cols);
367
368 // This error case should not be reachable as this function should
369 // never be called with the common data structure uninitialized.
370 else
371 cpi->static_mb_pct = 0;
372
373 cpi->seg0_cnt = ncnt[0];
374 vp9_enable_segmentation(&cm->seg);
375 } else {
376 cpi->static_mb_pct = 0;
377 vp9_disable_segmentation(&cm->seg);
378 }
379
380 // Free localy allocated storage
381 vpx_free(arf_not_zz);
382 }
383
vp9_update_mbgraph_stats(VP9_COMP * cpi)384 void vp9_update_mbgraph_stats(VP9_COMP *cpi) {
385 VP9_COMMON *const cm = &cpi->common;
386 int i, n_frames = vp9_lookahead_depth(cpi->lookahead);
387 YV12_BUFFER_CONFIG *golden_ref = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
388
389 // we need to look ahead beyond where the ARF transitions into
390 // being a GF - so exit if we don't look ahead beyond that
391 if (n_frames <= cpi->rc.frames_till_gf_update_due)
392 return;
393
394 if (n_frames > MAX_LAG_BUFFERS)
395 n_frames = MAX_LAG_BUFFERS;
396
397 cpi->mbgraph_n_frames = n_frames;
398 for (i = 0; i < n_frames; i++) {
399 MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
400 vpx_memset(frame_stats->mb_stats, 0,
401 cm->mb_rows * cm->mb_cols *
402 sizeof(*cpi->mbgraph_stats[i].mb_stats));
403 }
404
405 // do motion search to find contribution of each reference to data
406 // later on in this GF group
407 // FIXME really, the GF/last MC search should be done forward, and
408 // the ARF MC search backwards, to get optimal results for MV caching
409 for (i = 0; i < n_frames; i++) {
410 MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
411 struct lookahead_entry *q_cur = vp9_lookahead_peek(cpi->lookahead, i);
412
413 assert(q_cur != NULL);
414
415 update_mbgraph_frame_stats(cpi, frame_stats, &q_cur->img,
416 golden_ref, cpi->Source);
417 }
418
419 vp9_clear_system_state();
420
421 separate_arf_mbs(cpi);
422 }
423