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 "./vp9_rtcd.h"
14 #include "./vpx_dsp_rtcd.h"
15
16 #include "vpx_dsp/vpx_dsp_common.h"
17 #include "vpx_mem/vpx_mem.h"
18 #include "vpx_ports/system_state.h"
19 #include "vp9/encoder/vp9_segmentation.h"
20 #include "vp9/encoder/vp9_mcomp.h"
21 #include "vp9/common/vp9_blockd.h"
22 #include "vp9/common/vp9_reconinter.h"
23 #include "vp9/common/vp9_reconintra.h"
24
do_16x16_motion_iteration(VP9_COMP * cpi,const MV * ref_mv,MV * dst_mv,int mb_row,int mb_col)25 static unsigned int do_16x16_motion_iteration(VP9_COMP *cpi, const MV *ref_mv,
26 MV *dst_mv, int mb_row,
27 int mb_col) {
28 MACROBLOCK *const x = &cpi->td.mb;
29 MACROBLOCKD *const xd = &x->e_mbd;
30 MV_SPEED_FEATURES *const mv_sf = &cpi->sf.mv;
31 const SEARCH_METHODS old_search_method = mv_sf->search_method;
32 const vp9_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[BLOCK_16X16];
33 const MvLimits tmp_mv_limits = x->mv_limits;
34 MV ref_full;
35 int cost_list[5];
36
37 // Further step/diamond searches as necessary
38 int step_param = mv_sf->reduce_first_step_size;
39 step_param = VPXMIN(step_param, MAX_MVSEARCH_STEPS - 2);
40
41 vp9_set_mv_search_range(&x->mv_limits, ref_mv);
42
43 ref_full.col = ref_mv->col >> 3;
44 ref_full.row = ref_mv->row >> 3;
45
46 mv_sf->search_method = HEX;
47 vp9_full_pixel_search(cpi, x, BLOCK_16X16, &ref_full, step_param,
48 cpi->sf.mv.search_method, x->errorperbit,
49 cond_cost_list(cpi, cost_list), ref_mv, dst_mv, 0, 0);
50 mv_sf->search_method = old_search_method;
51
52 /* restore UMV window */
53 x->mv_limits = tmp_mv_limits;
54
55 // Try sub-pixel MC
56 // if (bestsme > error_thresh && bestsme < INT_MAX)
57 {
58 uint32_t distortion;
59 uint32_t sse;
60 // TODO(yunqing): may use higher tap interp filter than 2 taps if needed.
61 cpi->find_fractional_mv_step(
62 x, dst_mv, ref_mv, cpi->common.allow_high_precision_mv, x->errorperbit,
63 &v_fn_ptr, 0, mv_sf->subpel_search_level,
64 cond_cost_list(cpi, cost_list), NULL, NULL, &distortion, &sse, NULL, 0,
65 0, USE_2_TAPS);
66 }
67
68 xd->mi[0]->mode = NEWMV;
69 xd->mi[0]->mv[0].as_mv = *dst_mv;
70
71 vp9_build_inter_predictors_sby(xd, mb_row, mb_col, BLOCK_16X16);
72
73 return vpx_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
74 xd->plane[0].dst.buf, xd->plane[0].dst.stride);
75 }
76
do_16x16_motion_search(VP9_COMP * cpi,const MV * ref_mv,int_mv * dst_mv,int mb_row,int mb_col)77 static int do_16x16_motion_search(VP9_COMP *cpi, const MV *ref_mv,
78 int_mv *dst_mv, int mb_row, int mb_col) {
79 MACROBLOCK *const x = &cpi->td.mb;
80 MACROBLOCKD *const xd = &x->e_mbd;
81 unsigned int err, tmp_err;
82 MV tmp_mv;
83
84 // Try zero MV first
85 // FIXME should really use something like near/nearest MV and/or MV prediction
86 err = vpx_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
87 xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride);
88 dst_mv->as_int = 0;
89
90 // Test last reference frame using the previous best mv as the
91 // starting point (best reference) for the search
92 tmp_err = do_16x16_motion_iteration(cpi, ref_mv, &tmp_mv, mb_row, mb_col);
93 if (tmp_err < err) {
94 err = tmp_err;
95 dst_mv->as_mv = tmp_mv;
96 }
97
98 // If the current best reference mv is not centered on 0,0 then do a 0,0
99 // based search as well.
100 if (ref_mv->row != 0 || ref_mv->col != 0) {
101 unsigned int tmp_err;
102 MV zero_ref_mv = { 0, 0 }, tmp_mv;
103
104 tmp_err =
105 do_16x16_motion_iteration(cpi, &zero_ref_mv, &tmp_mv, mb_row, mb_col);
106 if (tmp_err < err) {
107 dst_mv->as_mv = tmp_mv;
108 err = tmp_err;
109 }
110 }
111
112 return err;
113 }
114
do_16x16_zerozero_search(VP9_COMP * cpi,int_mv * dst_mv)115 static int do_16x16_zerozero_search(VP9_COMP *cpi, int_mv *dst_mv) {
116 MACROBLOCK *const x = &cpi->td.mb;
117 MACROBLOCKD *const xd = &x->e_mbd;
118 unsigned int err;
119
120 // Try zero MV first
121 // FIXME should really use something like near/nearest MV and/or MV prediction
122 err = vpx_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
123 xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride);
124
125 dst_mv->as_int = 0;
126
127 return err;
128 }
find_best_16x16_intra(VP9_COMP * cpi,PREDICTION_MODE * pbest_mode)129 static int find_best_16x16_intra(VP9_COMP *cpi, PREDICTION_MODE *pbest_mode) {
130 MACROBLOCK *const x = &cpi->td.mb;
131 MACROBLOCKD *const xd = &x->e_mbd;
132 PREDICTION_MODE best_mode = -1, mode;
133 unsigned int best_err = INT_MAX;
134
135 // calculate SATD for each intra prediction mode;
136 // we're intentionally not doing 4x4, we just want a rough estimate
137 for (mode = DC_PRED; mode <= TM_PRED; mode++) {
138 unsigned int err;
139
140 xd->mi[0]->mode = mode;
141 vp9_predict_intra_block(xd, 2, TX_16X16, mode, x->plane[0].src.buf,
142 x->plane[0].src.stride, xd->plane[0].dst.buf,
143 xd->plane[0].dst.stride, 0, 0, 0);
144 err = vpx_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
145 xd->plane[0].dst.buf, xd->plane[0].dst.stride);
146
147 // find best
148 if (err < best_err) {
149 best_err = err;
150 best_mode = mode;
151 }
152 }
153
154 if (pbest_mode) *pbest_mode = best_mode;
155
156 return best_err;
157 }
158
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)159 static void update_mbgraph_mb_stats(VP9_COMP *cpi, MBGRAPH_MB_STATS *stats,
160 YV12_BUFFER_CONFIG *buf, int mb_y_offset,
161 YV12_BUFFER_CONFIG *golden_ref,
162 const MV *prev_golden_ref_mv,
163 YV12_BUFFER_CONFIG *alt_ref, int mb_row,
164 int mb_col) {
165 MACROBLOCK *const x = &cpi->td.mb;
166 MACROBLOCKD *const xd = &x->e_mbd;
167 int intra_error;
168 VP9_COMMON *cm = &cpi->common;
169
170 // FIXME in practice we're completely ignoring chroma here
171 x->plane[0].src.buf = buf->y_buffer + mb_y_offset;
172 x->plane[0].src.stride = buf->y_stride;
173
174 xd->plane[0].dst.buf = get_frame_new_buffer(cm)->y_buffer + mb_y_offset;
175 xd->plane[0].dst.stride = get_frame_new_buffer(cm)->y_stride;
176
177 // do intra 16x16 prediction
178 intra_error = find_best_16x16_intra(cpi, &stats->ref[INTRA_FRAME].m.mode);
179 if (intra_error <= 0) intra_error = 1;
180 stats->ref[INTRA_FRAME].err = intra_error;
181
182 // Golden frame MV search, if it exists and is different than last frame
183 if (golden_ref) {
184 int g_motion_error;
185 xd->plane[0].pre[0].buf = golden_ref->y_buffer + mb_y_offset;
186 xd->plane[0].pre[0].stride = golden_ref->y_stride;
187 g_motion_error =
188 do_16x16_motion_search(cpi, prev_golden_ref_mv,
189 &stats->ref[GOLDEN_FRAME].m.mv, mb_row, mb_col);
190 stats->ref[GOLDEN_FRAME].err = g_motion_error;
191 } else {
192 stats->ref[GOLDEN_FRAME].err = INT_MAX;
193 stats->ref[GOLDEN_FRAME].m.mv.as_int = 0;
194 }
195
196 // Do an Alt-ref frame MV search, if it exists and is different than
197 // last/golden frame.
198 if (alt_ref) {
199 int a_motion_error;
200 xd->plane[0].pre[0].buf = alt_ref->y_buffer + mb_y_offset;
201 xd->plane[0].pre[0].stride = alt_ref->y_stride;
202 a_motion_error =
203 do_16x16_zerozero_search(cpi, &stats->ref[ALTREF_FRAME].m.mv);
204
205 stats->ref[ALTREF_FRAME].err = a_motion_error;
206 } else {
207 stats->ref[ALTREF_FRAME].err = INT_MAX;
208 stats->ref[ALTREF_FRAME].m.mv.as_int = 0;
209 }
210 }
211
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)212 static void update_mbgraph_frame_stats(VP9_COMP *cpi,
213 MBGRAPH_FRAME_STATS *stats,
214 YV12_BUFFER_CONFIG *buf,
215 YV12_BUFFER_CONFIG *golden_ref,
216 YV12_BUFFER_CONFIG *alt_ref) {
217 MACROBLOCK *const x = &cpi->td.mb;
218 MACROBLOCKD *const xd = &x->e_mbd;
219 VP9_COMMON *const cm = &cpi->common;
220
221 int mb_col, mb_row, offset = 0;
222 int mb_y_offset = 0;
223 MV gld_top_mv = { 0, 0 };
224 MODE_INFO mi_local;
225 MODE_INFO mi_above, mi_left;
226
227 vp9_zero(mi_local);
228 // Set up limit values for motion vectors to prevent them extending outside
229 // the UMV borders.
230 x->mv_limits.row_min = -BORDER_MV_PIXELS_B16;
231 x->mv_limits.row_max = (cm->mb_rows - 1) * 8 + BORDER_MV_PIXELS_B16;
232 // Signal to vp9_predict_intra_block() that above is not available
233 xd->above_mi = NULL;
234
235 xd->plane[0].dst.stride = buf->y_stride;
236 xd->plane[0].pre[0].stride = buf->y_stride;
237 xd->plane[1].dst.stride = buf->uv_stride;
238 xd->mi[0] = &mi_local;
239 mi_local.sb_type = BLOCK_16X16;
240 mi_local.ref_frame[0] = LAST_FRAME;
241 mi_local.ref_frame[1] = NONE;
242
243 for (mb_row = 0; mb_row < cm->mb_rows; mb_row++) {
244 MV gld_left_mv = gld_top_mv;
245 int mb_y_in_offset = mb_y_offset;
246
247 // Set up limit values for motion vectors to prevent them extending outside
248 // the UMV borders.
249 x->mv_limits.col_min = -BORDER_MV_PIXELS_B16;
250 x->mv_limits.col_max = (cm->mb_cols - 1) * 8 + BORDER_MV_PIXELS_B16;
251 // Signal to vp9_predict_intra_block() that left is not available
252 xd->left_mi = NULL;
253
254 for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
255 MBGRAPH_MB_STATS *mb_stats = &stats->mb_stats[offset + mb_col];
256
257 update_mbgraph_mb_stats(cpi, mb_stats, buf, mb_y_in_offset, golden_ref,
258 &gld_left_mv, alt_ref, mb_row, mb_col);
259 gld_left_mv = mb_stats->ref[GOLDEN_FRAME].m.mv.as_mv;
260 if (mb_col == 0) {
261 gld_top_mv = gld_left_mv;
262 }
263 // Signal to vp9_predict_intra_block() that left is available
264 xd->left_mi = &mi_left;
265
266 mb_y_in_offset += 16;
267 x->mv_limits.col_min -= 16;
268 x->mv_limits.col_max -= 16;
269 }
270
271 // Signal to vp9_predict_intra_block() that above is available
272 xd->above_mi = &mi_above;
273
274 mb_y_offset += buf->y_stride * 16;
275 x->mv_limits.row_min -= 16;
276 x->mv_limits.row_max -= 16;
277 offset += cm->mb_cols;
278 }
279 }
280
281 // void separate_arf_mbs_byzz
separate_arf_mbs(VP9_COMP * cpi)282 static void separate_arf_mbs(VP9_COMP *cpi) {
283 VP9_COMMON *const cm = &cpi->common;
284 int mb_col, mb_row, offset, i;
285 int mi_row, mi_col;
286 int ncnt[4] = { 0 };
287 int n_frames = cpi->mbgraph_n_frames;
288
289 int *arf_not_zz;
290
291 CHECK_MEM_ERROR(
292 cm, arf_not_zz,
293 vpx_calloc(cm->mb_rows * cm->mb_cols * sizeof(*arf_not_zz), 1));
294
295 // We are not interested in results beyond the alt ref itself.
296 if (n_frames > cpi->rc.frames_till_gf_update_due)
297 n_frames = cpi->rc.frames_till_gf_update_due;
298
299 // defer cost to reference frames
300 for (i = n_frames - 1; i >= 0; i--) {
301 MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
302
303 for (offset = 0, mb_row = 0; mb_row < cm->mb_rows;
304 offset += cm->mb_cols, mb_row++) {
305 for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
306 MBGRAPH_MB_STATS *mb_stats = &frame_stats->mb_stats[offset + mb_col];
307
308 int altref_err = mb_stats->ref[ALTREF_FRAME].err;
309 int intra_err = mb_stats->ref[INTRA_FRAME].err;
310 int golden_err = mb_stats->ref[GOLDEN_FRAME].err;
311
312 // Test for altref vs intra and gf and that its mv was 0,0.
313 if (altref_err > 1000 || altref_err > intra_err ||
314 altref_err > golden_err) {
315 arf_not_zz[offset + mb_col]++;
316 }
317 }
318 }
319 }
320
321 // arf_not_zz is indexed by MB, but this loop is indexed by MI to avoid out
322 // of bound access in segmentation_map
323 for (mi_row = 0; mi_row < cm->mi_rows; mi_row++) {
324 for (mi_col = 0; mi_col < cm->mi_cols; mi_col++) {
325 // If any of the blocks in the sequence failed then the MB
326 // goes in segment 0
327 if (arf_not_zz[mi_row / 2 * cm->mb_cols + mi_col / 2]) {
328 ncnt[0]++;
329 cpi->segmentation_map[mi_row * cm->mi_cols + mi_col] = 0;
330 } else {
331 cpi->segmentation_map[mi_row * cm->mi_cols + mi_col] = 1;
332 ncnt[1]++;
333 }
334 }
335 }
336
337 // Only bother with segmentation if over 10% of the MBs in static segment
338 // if ( ncnt[1] && (ncnt[0] / ncnt[1] < 10) )
339 if (1) {
340 // Note % of blocks that are marked as static
341 if (cm->MBs)
342 cpi->static_mb_pct = (ncnt[1] * 100) / (cm->mi_rows * cm->mi_cols);
343
344 // This error case should not be reachable as this function should
345 // never be called with the common data structure uninitialized.
346 else
347 cpi->static_mb_pct = 0;
348
349 vp9_enable_segmentation(&cm->seg);
350 } else {
351 cpi->static_mb_pct = 0;
352 vp9_disable_segmentation(&cm->seg);
353 }
354
355 // Free localy allocated storage
356 vpx_free(arf_not_zz);
357 }
358
vp9_update_mbgraph_stats(VP9_COMP * cpi)359 void vp9_update_mbgraph_stats(VP9_COMP *cpi) {
360 VP9_COMMON *const cm = &cpi->common;
361 int i, n_frames = vp9_lookahead_depth(cpi->lookahead);
362 YV12_BUFFER_CONFIG *golden_ref = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
363
364 assert(golden_ref != NULL);
365
366 // we need to look ahead beyond where the ARF transitions into
367 // being a GF - so exit if we don't look ahead beyond that
368 if (n_frames <= cpi->rc.frames_till_gf_update_due) return;
369
370 if (n_frames > MAX_LAG_BUFFERS) n_frames = MAX_LAG_BUFFERS;
371
372 cpi->mbgraph_n_frames = n_frames;
373 for (i = 0; i < n_frames; i++) {
374 MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
375 memset(frame_stats->mb_stats, 0,
376 cm->mb_rows * cm->mb_cols * sizeof(*cpi->mbgraph_stats[i].mb_stats));
377 }
378
379 // do motion search to find contribution of each reference to data
380 // later on in this GF group
381 // FIXME really, the GF/last MC search should be done forward, and
382 // the ARF MC search backwards, to get optimal results for MV caching
383 for (i = 0; i < n_frames; i++) {
384 MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
385 struct lookahead_entry *q_cur = vp9_lookahead_peek(cpi->lookahead, i);
386
387 assert(q_cur != NULL);
388
389 update_mbgraph_frame_stats(cpi, frame_stats, &q_cur->img, golden_ref,
390 cpi->Source);
391 }
392
393 vpx_clear_system_state();
394
395 separate_arf_mbs(cpi);
396 }
397