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 
25 
do_16x16_motion_iteration(VP9_COMP * cpi,const MV * ref_mv,MV * dst_mv,int mb_row,int mb_col)26 static unsigned int do_16x16_motion_iteration(VP9_COMP *cpi,
27                                               const MV *ref_mv,
28                                               MV *dst_mv,
29                                               int mb_row,
30                                               int mb_col) {
31   MACROBLOCK *const x = &cpi->td.mb;
32   MACROBLOCKD *const xd = &x->e_mbd;
33   MV_SPEED_FEATURES *const mv_sf = &cpi->sf.mv;
34   const SEARCH_METHODS old_search_method = mv_sf->search_method;
35   const vp9_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[BLOCK_16X16];
36 
37   const int tmp_col_min = x->mv_col_min;
38   const int tmp_col_max = x->mv_col_max;
39   const int tmp_row_min = x->mv_row_min;
40   const int tmp_row_max = x->mv_row_max;
41   MV ref_full;
42   int cost_list[5];
43 
44   // Further step/diamond searches as necessary
45   int step_param = mv_sf->reduce_first_step_size;
46   step_param = VPXMIN(step_param, MAX_MVSEARCH_STEPS - 2);
47 
48   vp9_set_mv_search_range(x, ref_mv);
49 
50   ref_full.col = ref_mv->col >> 3;
51   ref_full.row = ref_mv->row >> 3;
52 
53   mv_sf->search_method = HEX;
54   vp9_full_pixel_search(cpi, x, BLOCK_16X16, &ref_full, step_param,
55                         x->errorperbit, cond_cost_list(cpi, cost_list), ref_mv,
56                         dst_mv, 0, 0);
57   mv_sf->search_method = old_search_method;
58 
59   // Try sub-pixel MC
60   // if (bestsme > error_thresh && bestsme < INT_MAX)
61   {
62     int distortion;
63     unsigned int sse;
64     cpi->find_fractional_mv_step(
65         x, dst_mv, ref_mv, cpi->common.allow_high_precision_mv, x->errorperbit,
66         &v_fn_ptr, 0, mv_sf->subpel_iters_per_step,
67         cond_cost_list(cpi, cost_list),
68         NULL, NULL,
69         &distortion, &sse, NULL, 0, 0);
70   }
71 
72   xd->mi[0]->mbmi.mode = NEWMV;
73   xd->mi[0]->mbmi.mv[0].as_mv = *dst_mv;
74 
75   vp9_build_inter_predictors_sby(xd, mb_row, mb_col, BLOCK_16X16);
76 
77   /* restore UMV window */
78   x->mv_col_min = tmp_col_min;
79   x->mv_col_max = tmp_col_max;
80   x->mv_row_min = tmp_row_min;
81   x->mv_row_max = tmp_row_max;
82 
83   return vpx_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
84                       xd->plane[0].dst.buf, xd->plane[0].dst.stride);
85 }
86 
do_16x16_motion_search(VP9_COMP * cpi,const MV * ref_mv,int_mv * dst_mv,int mb_row,int mb_col)87 static int do_16x16_motion_search(VP9_COMP *cpi, const MV *ref_mv,
88                                   int_mv *dst_mv, int mb_row, int mb_col) {
89   MACROBLOCK *const x = &cpi->td.mb;
90   MACROBLOCKD *const xd = &x->e_mbd;
91   unsigned int err, tmp_err;
92   MV tmp_mv;
93 
94   // Try zero MV first
95   // FIXME should really use something like near/nearest MV and/or MV prediction
96   err = vpx_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
97                      xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride);
98   dst_mv->as_int = 0;
99 
100   // Test last reference frame using the previous best mv as the
101   // starting point (best reference) for the search
102   tmp_err = do_16x16_motion_iteration(cpi, ref_mv, &tmp_mv, mb_row, mb_col);
103   if (tmp_err < err) {
104     err = tmp_err;
105     dst_mv->as_mv = tmp_mv;
106   }
107 
108   // If the current best reference mv is not centered on 0,0 then do a 0,0
109   // based search as well.
110   if (ref_mv->row != 0 || ref_mv->col != 0) {
111     unsigned int tmp_err;
112     MV zero_ref_mv = {0, 0}, tmp_mv;
113 
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_mv = tmp_mv;
118       err = tmp_err;
119     }
120   }
121 
122   return err;
123 }
124 
do_16x16_zerozero_search(VP9_COMP * cpi,int_mv * dst_mv)125 static int do_16x16_zerozero_search(VP9_COMP *cpi, int_mv *dst_mv) {
126   MACROBLOCK *const x = &cpi->td.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 = vpx_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 
135   dst_mv->as_int = 0;
136 
137   return err;
138 }
find_best_16x16_intra(VP9_COMP * cpi,PREDICTION_MODE * pbest_mode)139 static int find_best_16x16_intra(VP9_COMP *cpi, PREDICTION_MODE *pbest_mode) {
140   MACROBLOCK   *const x  = &cpi->td.mb;
141   MACROBLOCKD *const xd = &x->e_mbd;
142   PREDICTION_MODE best_mode = -1, mode;
143   unsigned int best_err = INT_MAX;
144 
145   // calculate SATD for each intra prediction mode;
146   // we're intentionally not doing 4x4, we just want a rough estimate
147   for (mode = DC_PRED; mode <= TM_PRED; mode++) {
148     unsigned int err;
149 
150     xd->mi[0]->mbmi.mode = mode;
151     vp9_predict_intra_block(xd, 2, TX_16X16, mode,
152                             x->plane[0].src.buf, x->plane[0].src.stride,
153                             xd->plane[0].dst.buf, xd->plane[0].dst.stride,
154                             0, 0, 0);
155     err = vpx_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
156                        xd->plane[0].dst.buf, xd->plane[0].dst.stride);
157 
158     // find best
159     if (err < best_err) {
160       best_err  = err;
161       best_mode = mode;
162     }
163   }
164 
165   if (pbest_mode)
166     *pbest_mode = best_mode;
167 
168   return best_err;
169 }
170 
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)171 static void update_mbgraph_mb_stats
172 (
173   VP9_COMP *cpi,
174   MBGRAPH_MB_STATS *stats,
175   YV12_BUFFER_CONFIG *buf,
176   int mb_y_offset,
177   YV12_BUFFER_CONFIG *golden_ref,
178   const MV *prev_golden_ref_mv,
179   YV12_BUFFER_CONFIG *alt_ref,
180   int mb_row,
181   int mb_col
182 ) {
183   MACROBLOCK *const x = &cpi->td.mb;
184   MACROBLOCKD *const xd = &x->e_mbd;
185   int intra_error;
186   VP9_COMMON *cm = &cpi->common;
187 
188   // FIXME in practice we're completely ignoring chroma here
189   x->plane[0].src.buf = buf->y_buffer + mb_y_offset;
190   x->plane[0].src.stride = buf->y_stride;
191 
192   xd->plane[0].dst.buf = get_frame_new_buffer(cm)->y_buffer + mb_y_offset;
193   xd->plane[0].dst.stride = get_frame_new_buffer(cm)->y_stride;
194 
195   // do intra 16x16 prediction
196   intra_error = find_best_16x16_intra(cpi,
197                                       &stats->ref[INTRA_FRAME].m.mode);
198   if (intra_error <= 0)
199     intra_error = 1;
200   stats->ref[INTRA_FRAME].err = intra_error;
201 
202   // Golden frame MV search, if it exists and is different than last frame
203   if (golden_ref) {
204     int g_motion_error;
205     xd->plane[0].pre[0].buf = golden_ref->y_buffer + mb_y_offset;
206     xd->plane[0].pre[0].stride = golden_ref->y_stride;
207     g_motion_error = do_16x16_motion_search(cpi,
208                                             prev_golden_ref_mv,
209                                             &stats->ref[GOLDEN_FRAME].m.mv,
210                                             mb_row, mb_col);
211     stats->ref[GOLDEN_FRAME].err = g_motion_error;
212   } else {
213     stats->ref[GOLDEN_FRAME].err = INT_MAX;
214     stats->ref[GOLDEN_FRAME].m.mv.as_int = 0;
215   }
216 
217   // Do an Alt-ref frame MV search, if it exists and is different than
218   // last/golden frame.
219   if (alt_ref) {
220     int a_motion_error;
221     xd->plane[0].pre[0].buf = alt_ref->y_buffer + mb_y_offset;
222     xd->plane[0].pre[0].stride = alt_ref->y_stride;
223     a_motion_error = do_16x16_zerozero_search(cpi,
224                                               &stats->ref[ALTREF_FRAME].m.mv);
225 
226     stats->ref[ALTREF_FRAME].err = a_motion_error;
227   } else {
228     stats->ref[ALTREF_FRAME].err = INT_MAX;
229     stats->ref[ALTREF_FRAME].m.mv.as_int = 0;
230   }
231 }
232 
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)233 static void update_mbgraph_frame_stats(VP9_COMP *cpi,
234                                        MBGRAPH_FRAME_STATS *stats,
235                                        YV12_BUFFER_CONFIG *buf,
236                                        YV12_BUFFER_CONFIG *golden_ref,
237                                        YV12_BUFFER_CONFIG *alt_ref) {
238   MACROBLOCK *const x = &cpi->td.mb;
239   MACROBLOCKD *const xd = &x->e_mbd;
240   VP9_COMMON *const cm = &cpi->common;
241 
242   int mb_col, mb_row, offset = 0;
243   int mb_y_offset = 0, arf_y_offset = 0, gld_y_offset = 0;
244   MV gld_top_mv = {0, 0};
245   MODE_INFO mi_local;
246 
247   vp9_zero(mi_local);
248   // Set up limit values for motion vectors to prevent them extending outside
249   // the UMV borders.
250   x->mv_row_min     = -BORDER_MV_PIXELS_B16;
251   x->mv_row_max     = (cm->mb_rows - 1) * 8 + BORDER_MV_PIXELS_B16;
252   xd->up_available  = 0;
253   xd->plane[0].dst.stride  = buf->y_stride;
254   xd->plane[0].pre[0].stride  = buf->y_stride;
255   xd->plane[1].dst.stride = buf->uv_stride;
256   xd->mi[0] = &mi_local;
257   mi_local.mbmi.sb_type = BLOCK_16X16;
258   mi_local.mbmi.ref_frame[0] = LAST_FRAME;
259   mi_local.mbmi.ref_frame[1] = NONE;
260 
261   for (mb_row = 0; mb_row < cm->mb_rows; mb_row++) {
262     MV gld_left_mv = gld_top_mv;
263     int mb_y_in_offset  = mb_y_offset;
264     int arf_y_in_offset = arf_y_offset;
265     int gld_y_in_offset = gld_y_offset;
266 
267     // Set up limit values for motion vectors to prevent them extending outside
268     // the UMV borders.
269     x->mv_col_min      = -BORDER_MV_PIXELS_B16;
270     x->mv_col_max      = (cm->mb_cols - 1) * 8 + BORDER_MV_PIXELS_B16;
271     xd->left_available = 0;
272 
273     for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
274       MBGRAPH_MB_STATS *mb_stats = &stats->mb_stats[offset + mb_col];
275 
276       update_mbgraph_mb_stats(cpi, mb_stats, buf, mb_y_in_offset,
277                               golden_ref, &gld_left_mv, alt_ref,
278                               mb_row, mb_col);
279       gld_left_mv = mb_stats->ref[GOLDEN_FRAME].m.mv.as_mv;
280       if (mb_col == 0) {
281         gld_top_mv = gld_left_mv;
282       }
283       xd->left_available = 1;
284       mb_y_in_offset    += 16;
285       gld_y_in_offset   += 16;
286       arf_y_in_offset   += 16;
287       x->mv_col_min     -= 16;
288       x->mv_col_max     -= 16;
289     }
290     xd->up_available = 1;
291     mb_y_offset     += buf->y_stride * 16;
292     gld_y_offset    += golden_ref->y_stride * 16;
293     if (alt_ref)
294       arf_y_offset    += alt_ref->y_stride * 16;
295     x->mv_row_min   -= 16;
296     x->mv_row_max   -= 16;
297     offset          += cm->mb_cols;
298   }
299 }
300 
301 // void separate_arf_mbs_byzz
separate_arf_mbs(VP9_COMP * cpi)302 static void separate_arf_mbs(VP9_COMP *cpi) {
303   VP9_COMMON *const cm = &cpi->common;
304   int mb_col, mb_row, offset, i;
305   int mi_row, mi_col;
306   int ncnt[4] = { 0 };
307   int n_frames = cpi->mbgraph_n_frames;
308 
309   int *arf_not_zz;
310 
311   CHECK_MEM_ERROR(cm, arf_not_zz,
312                   vpx_calloc(cm->mb_rows * cm->mb_cols * sizeof(*arf_not_zz),
313                              1));
314 
315   // We are not interested in results beyond the alt ref itself.
316   if (n_frames > cpi->rc.frames_till_gf_update_due)
317     n_frames = cpi->rc.frames_till_gf_update_due;
318 
319   // defer cost to reference frames
320   for (i = n_frames - 1; i >= 0; i--) {
321     MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
322 
323     for (offset = 0, mb_row = 0; mb_row < cm->mb_rows;
324          offset += cm->mb_cols, mb_row++) {
325       for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
326         MBGRAPH_MB_STATS *mb_stats = &frame_stats->mb_stats[offset + mb_col];
327 
328         int altref_err = mb_stats->ref[ALTREF_FRAME].err;
329         int intra_err  = mb_stats->ref[INTRA_FRAME ].err;
330         int golden_err = mb_stats->ref[GOLDEN_FRAME].err;
331 
332         // Test for altref vs intra and gf and that its mv was 0,0.
333         if (altref_err > 1000 ||
334             altref_err > intra_err ||
335             altref_err > golden_err) {
336           arf_not_zz[offset + mb_col]++;
337         }
338       }
339     }
340   }
341 
342   // arf_not_zz is indexed by MB, but this loop is indexed by MI to avoid out
343   // of bound access in segmentation_map
344   for (mi_row = 0; mi_row < cm->mi_rows; mi_row++) {
345     for (mi_col = 0; mi_col < cm->mi_cols; mi_col++) {
346       // If any of the blocks in the sequence failed then the MB
347       // goes in segment 0
348       if (arf_not_zz[mi_row / 2 * cm->mb_cols + mi_col / 2]) {
349         ncnt[0]++;
350         cpi->segmentation_map[mi_row * cm->mi_cols + mi_col] = 0;
351       } else {
352         cpi->segmentation_map[mi_row * cm->mi_cols + mi_col] = 1;
353         ncnt[1]++;
354       }
355     }
356   }
357 
358   // Only bother with segmentation if over 10% of the MBs in static segment
359   // if ( ncnt[1] && (ncnt[0] / ncnt[1] < 10) )
360   if (1) {
361     // Note % of blocks that are marked as static
362     if (cm->MBs)
363       cpi->static_mb_pct = (ncnt[1] * 100) / (cm->mi_rows * cm->mi_cols);
364 
365     // This error case should not be reachable as this function should
366     // never be called with the common data structure uninitialized.
367     else
368       cpi->static_mb_pct = 0;
369 
370     vp9_enable_segmentation(&cm->seg);
371   } else {
372     cpi->static_mb_pct = 0;
373     vp9_disable_segmentation(&cm->seg);
374   }
375 
376   // Free localy allocated storage
377   vpx_free(arf_not_zz);
378 }
379 
vp9_update_mbgraph_stats(VP9_COMP * cpi)380 void vp9_update_mbgraph_stats(VP9_COMP *cpi) {
381   VP9_COMMON *const cm = &cpi->common;
382   int i, n_frames = vp9_lookahead_depth(cpi->lookahead);
383   YV12_BUFFER_CONFIG *golden_ref = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
384 
385   assert(golden_ref != NULL);
386 
387   // we need to look ahead beyond where the ARF transitions into
388   // being a GF - so exit if we don't look ahead beyond that
389   if (n_frames <= cpi->rc.frames_till_gf_update_due)
390     return;
391 
392   if (n_frames > MAX_LAG_BUFFERS)
393     n_frames = MAX_LAG_BUFFERS;
394 
395   cpi->mbgraph_n_frames = n_frames;
396   for (i = 0; i < n_frames; i++) {
397     MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
398     memset(frame_stats->mb_stats, 0,
399            cm->mb_rows * cm->mb_cols * sizeof(*cpi->mbgraph_stats[i].mb_stats));
400   }
401 
402   // do motion search to find contribution of each reference to data
403   // later on in this GF group
404   // FIXME really, the GF/last MC search should be done forward, and
405   // the ARF MC search backwards, to get optimal results for MV caching
406   for (i = 0; i < n_frames; i++) {
407     MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
408     struct lookahead_entry *q_cur = vp9_lookahead_peek(cpi->lookahead, i);
409 
410     assert(q_cur != NULL);
411 
412     update_mbgraph_frame_stats(cpi, frame_stats, &q_cur->img,
413                                golden_ref, cpi->Source);
414   }
415 
416   vpx_clear_system_state();
417 
418   separate_arf_mbs(cpi);
419 }
420