1 /*
2 * VP9 compatible video decoder
3 *
4 * Copyright (C) 2013 Ronald S. Bultje <rsbultje gmail com>
5 * Copyright (C) 2013 Clément Bœsch <u pkh me>
6 *
7 * This file is part of FFmpeg.
8 *
9 * FFmpeg is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public
11 * License as published by the Free Software Foundation; either
12 * version 2.1 of the License, or (at your option) any later version.
13 *
14 * FFmpeg is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
18 *
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with FFmpeg; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
22 */
23
24 #include "libavutil/avassert.h"
25
26 #include "avcodec.h"
27 #include "internal.h"
28 #include "videodsp.h"
29 #include "vp56.h"
30 #include "vp9.h"
31 #include "vp9data.h"
32 #include "vp9dec.h"
33
setctx_2d(uint8_t * ptr,int w,int h,ptrdiff_t stride,int v)34 static av_always_inline void setctx_2d(uint8_t *ptr, int w, int h,
35 ptrdiff_t stride, int v)
36 {
37 switch (w) {
38 case 1:
39 do {
40 *ptr = v;
41 ptr += stride;
42 } while (--h);
43 break;
44 case 2: {
45 int v16 = v * 0x0101;
46 do {
47 AV_WN16A(ptr, v16);
48 ptr += stride;
49 } while (--h);
50 break;
51 }
52 case 4: {
53 uint32_t v32 = v * 0x01010101;
54 do {
55 AV_WN32A(ptr, v32);
56 ptr += stride;
57 } while (--h);
58 break;
59 }
60 case 8: {
61 #if HAVE_FAST_64BIT
62 uint64_t v64 = v * 0x0101010101010101ULL;
63 do {
64 AV_WN64A(ptr, v64);
65 ptr += stride;
66 } while (--h);
67 #else
68 uint32_t v32 = v * 0x01010101;
69 do {
70 AV_WN32A(ptr, v32);
71 AV_WN32A(ptr + 4, v32);
72 ptr += stride;
73 } while (--h);
74 #endif
75 break;
76 }
77 }
78 }
79
decode_mode(VP9TileData * td)80 static void decode_mode(VP9TileData *td)
81 {
82 static const uint8_t left_ctx[N_BS_SIZES] = {
83 0x0, 0x8, 0x0, 0x8, 0xc, 0x8, 0xc, 0xe, 0xc, 0xe, 0xf, 0xe, 0xf
84 };
85 static const uint8_t above_ctx[N_BS_SIZES] = {
86 0x0, 0x0, 0x8, 0x8, 0x8, 0xc, 0xc, 0xc, 0xe, 0xe, 0xe, 0xf, 0xf
87 };
88 static const uint8_t max_tx_for_bl_bp[N_BS_SIZES] = {
89 TX_32X32, TX_32X32, TX_32X32, TX_32X32, TX_16X16, TX_16X16,
90 TX_16X16, TX_8X8, TX_8X8, TX_8X8, TX_4X4, TX_4X4, TX_4X4
91 };
92 VP9Context *s = td->s;
93 VP9Block *b = td->b;
94 int row = td->row, col = td->col, row7 = td->row7;
95 enum TxfmMode max_tx = max_tx_for_bl_bp[b->bs];
96 int bw4 = ff_vp9_bwh_tab[1][b->bs][0], w4 = FFMIN(s->cols - col, bw4);
97 int bh4 = ff_vp9_bwh_tab[1][b->bs][1], h4 = FFMIN(s->rows - row, bh4), y;
98 int have_a = row > 0, have_l = col > td->tile_col_start;
99 int vref, filter_id;
100
101 if (!s->s.h.segmentation.enabled) {
102 b->seg_id = 0;
103 } else if (s->s.h.keyframe || s->s.h.intraonly) {
104 b->seg_id = !s->s.h.segmentation.update_map ? 0 :
105 vp8_rac_get_tree(td->c, ff_vp9_segmentation_tree, s->s.h.segmentation.prob);
106 } else if (!s->s.h.segmentation.update_map ||
107 (s->s.h.segmentation.temporal &&
108 vp56_rac_get_prob_branchy(td->c,
109 s->s.h.segmentation.pred_prob[s->above_segpred_ctx[col] +
110 td->left_segpred_ctx[row7]]))) {
111 if (!s->s.h.errorres && s->s.frames[REF_FRAME_SEGMAP].segmentation_map) {
112 int pred = 8, x;
113 uint8_t *refsegmap = s->s.frames[REF_FRAME_SEGMAP].segmentation_map;
114
115 if (!s->s.frames[REF_FRAME_SEGMAP].uses_2pass)
116 ff_thread_await_progress(&s->s.frames[REF_FRAME_SEGMAP].tf, row >> 3, 0);
117 for (y = 0; y < h4; y++) {
118 int idx_base = (y + row) * 8 * s->sb_cols + col;
119 for (x = 0; x < w4; x++)
120 pred = FFMIN(pred, refsegmap[idx_base + x]);
121 }
122 av_assert1(pred < 8);
123 b->seg_id = pred;
124 } else {
125 b->seg_id = 0;
126 }
127
128 memset(&s->above_segpred_ctx[col], 1, w4);
129 memset(&td->left_segpred_ctx[row7], 1, h4);
130 } else {
131 b->seg_id = vp8_rac_get_tree(td->c, ff_vp9_segmentation_tree,
132 s->s.h.segmentation.prob);
133
134 memset(&s->above_segpred_ctx[col], 0, w4);
135 memset(&td->left_segpred_ctx[row7], 0, h4);
136 }
137 if (s->s.h.segmentation.enabled &&
138 (s->s.h.segmentation.update_map || s->s.h.keyframe || s->s.h.intraonly)) {
139 setctx_2d(&s->s.frames[CUR_FRAME].segmentation_map[row * 8 * s->sb_cols + col],
140 bw4, bh4, 8 * s->sb_cols, b->seg_id);
141 }
142
143 b->skip = s->s.h.segmentation.enabled &&
144 s->s.h.segmentation.feat[b->seg_id].skip_enabled;
145 if (!b->skip) {
146 int c = td->left_skip_ctx[row7] + s->above_skip_ctx[col];
147 b->skip = vp56_rac_get_prob(td->c, s->prob.p.skip[c]);
148 td->counts.skip[c][b->skip]++;
149 }
150
151 if (s->s.h.keyframe || s->s.h.intraonly) {
152 b->intra = 1;
153 } else if (s->s.h.segmentation.enabled && s->s.h.segmentation.feat[b->seg_id].ref_enabled) {
154 b->intra = !s->s.h.segmentation.feat[b->seg_id].ref_val;
155 } else {
156 int c, bit;
157
158 if (have_a && have_l) {
159 c = s->above_intra_ctx[col] + td->left_intra_ctx[row7];
160 c += (c == 2);
161 } else {
162 c = have_a ? 2 * s->above_intra_ctx[col] :
163 have_l ? 2 * td->left_intra_ctx[row7] : 0;
164 }
165 bit = vp56_rac_get_prob(td->c, s->prob.p.intra[c]);
166 td->counts.intra[c][bit]++;
167 b->intra = !bit;
168 }
169
170 if ((b->intra || !b->skip) && s->s.h.txfmmode == TX_SWITCHABLE) {
171 int c;
172 if (have_a) {
173 if (have_l) {
174 c = (s->above_skip_ctx[col] ? max_tx :
175 s->above_txfm_ctx[col]) +
176 (td->left_skip_ctx[row7] ? max_tx :
177 td->left_txfm_ctx[row7]) > max_tx;
178 } else {
179 c = s->above_skip_ctx[col] ? 1 :
180 (s->above_txfm_ctx[col] * 2 > max_tx);
181 }
182 } else if (have_l) {
183 c = td->left_skip_ctx[row7] ? 1 :
184 (td->left_txfm_ctx[row7] * 2 > max_tx);
185 } else {
186 c = 1;
187 }
188 switch (max_tx) {
189 case TX_32X32:
190 b->tx = vp56_rac_get_prob(td->c, s->prob.p.tx32p[c][0]);
191 if (b->tx) {
192 b->tx += vp56_rac_get_prob(td->c, s->prob.p.tx32p[c][1]);
193 if (b->tx == 2)
194 b->tx += vp56_rac_get_prob(td->c, s->prob.p.tx32p[c][2]);
195 }
196 td->counts.tx32p[c][b->tx]++;
197 break;
198 case TX_16X16:
199 b->tx = vp56_rac_get_prob(td->c, s->prob.p.tx16p[c][0]);
200 if (b->tx)
201 b->tx += vp56_rac_get_prob(td->c, s->prob.p.tx16p[c][1]);
202 td->counts.tx16p[c][b->tx]++;
203 break;
204 case TX_8X8:
205 b->tx = vp56_rac_get_prob(td->c, s->prob.p.tx8p[c]);
206 td->counts.tx8p[c][b->tx]++;
207 break;
208 case TX_4X4:
209 b->tx = TX_4X4;
210 break;
211 }
212 } else {
213 b->tx = FFMIN(max_tx, s->s.h.txfmmode);
214 }
215
216 if (s->s.h.keyframe || s->s.h.intraonly) {
217 uint8_t *a = &s->above_mode_ctx[col * 2];
218 uint8_t *l = &td->left_mode_ctx[(row7) << 1];
219
220 b->comp = 0;
221 if (b->bs > BS_8x8) {
222 // FIXME the memory storage intermediates here aren't really
223 // necessary, they're just there to make the code slightly
224 // simpler for now
225 b->mode[0] =
226 a[0] = vp8_rac_get_tree(td->c, ff_vp9_intramode_tree,
227 ff_vp9_default_kf_ymode_probs[a[0]][l[0]]);
228 if (b->bs != BS_8x4) {
229 b->mode[1] = vp8_rac_get_tree(td->c, ff_vp9_intramode_tree,
230 ff_vp9_default_kf_ymode_probs[a[1]][b->mode[0]]);
231 l[0] =
232 a[1] = b->mode[1];
233 } else {
234 l[0] =
235 a[1] =
236 b->mode[1] = b->mode[0];
237 }
238 if (b->bs != BS_4x8) {
239 b->mode[2] =
240 a[0] = vp8_rac_get_tree(td->c, ff_vp9_intramode_tree,
241 ff_vp9_default_kf_ymode_probs[a[0]][l[1]]);
242 if (b->bs != BS_8x4) {
243 b->mode[3] = vp8_rac_get_tree(td->c, ff_vp9_intramode_tree,
244 ff_vp9_default_kf_ymode_probs[a[1]][b->mode[2]]);
245 l[1] =
246 a[1] = b->mode[3];
247 } else {
248 l[1] =
249 a[1] =
250 b->mode[3] = b->mode[2];
251 }
252 } else {
253 b->mode[2] = b->mode[0];
254 l[1] =
255 a[1] =
256 b->mode[3] = b->mode[1];
257 }
258 } else {
259 b->mode[0] = vp8_rac_get_tree(td->c, ff_vp9_intramode_tree,
260 ff_vp9_default_kf_ymode_probs[*a][*l]);
261 b->mode[3] =
262 b->mode[2] =
263 b->mode[1] = b->mode[0];
264 // FIXME this can probably be optimized
265 memset(a, b->mode[0], ff_vp9_bwh_tab[0][b->bs][0]);
266 memset(l, b->mode[0], ff_vp9_bwh_tab[0][b->bs][1]);
267 }
268 b->uvmode = vp8_rac_get_tree(td->c, ff_vp9_intramode_tree,
269 ff_vp9_default_kf_uvmode_probs[b->mode[3]]);
270 } else if (b->intra) {
271 b->comp = 0;
272 if (b->bs > BS_8x8) {
273 b->mode[0] = vp8_rac_get_tree(td->c, ff_vp9_intramode_tree,
274 s->prob.p.y_mode[0]);
275 td->counts.y_mode[0][b->mode[0]]++;
276 if (b->bs != BS_8x4) {
277 b->mode[1] = vp8_rac_get_tree(td->c, ff_vp9_intramode_tree,
278 s->prob.p.y_mode[0]);
279 td->counts.y_mode[0][b->mode[1]]++;
280 } else {
281 b->mode[1] = b->mode[0];
282 }
283 if (b->bs != BS_4x8) {
284 b->mode[2] = vp8_rac_get_tree(td->c, ff_vp9_intramode_tree,
285 s->prob.p.y_mode[0]);
286 td->counts.y_mode[0][b->mode[2]]++;
287 if (b->bs != BS_8x4) {
288 b->mode[3] = vp8_rac_get_tree(td->c, ff_vp9_intramode_tree,
289 s->prob.p.y_mode[0]);
290 td->counts.y_mode[0][b->mode[3]]++;
291 } else {
292 b->mode[3] = b->mode[2];
293 }
294 } else {
295 b->mode[2] = b->mode[0];
296 b->mode[3] = b->mode[1];
297 }
298 } else {
299 static const uint8_t size_group[10] = {
300 3, 3, 3, 3, 2, 2, 2, 1, 1, 1
301 };
302 int sz = size_group[b->bs];
303
304 b->mode[0] = vp8_rac_get_tree(td->c, ff_vp9_intramode_tree,
305 s->prob.p.y_mode[sz]);
306 b->mode[1] =
307 b->mode[2] =
308 b->mode[3] = b->mode[0];
309 td->counts.y_mode[sz][b->mode[3]]++;
310 }
311 b->uvmode = vp8_rac_get_tree(td->c, ff_vp9_intramode_tree,
312 s->prob.p.uv_mode[b->mode[3]]);
313 td->counts.uv_mode[b->mode[3]][b->uvmode]++;
314 } else {
315 static const uint8_t inter_mode_ctx_lut[14][14] = {
316 { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },
317 { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },
318 { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },
319 { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },
320 { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },
321 { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },
322 { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },
323 { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },
324 { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },
325 { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },
326 { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },
327 { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },
328 { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 1, 1, 0, 3 },
329 { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 3, 3, 3, 4 },
330 };
331
332 if (s->s.h.segmentation.enabled && s->s.h.segmentation.feat[b->seg_id].ref_enabled) {
333 av_assert2(s->s.h.segmentation.feat[b->seg_id].ref_val != 0);
334 b->comp = 0;
335 b->ref[0] = s->s.h.segmentation.feat[b->seg_id].ref_val - 1;
336 } else {
337 // read comp_pred flag
338 if (s->s.h.comppredmode != PRED_SWITCHABLE) {
339 b->comp = s->s.h.comppredmode == PRED_COMPREF;
340 } else {
341 int c;
342
343 // FIXME add intra as ref=0xff (or -1) to make these easier?
344 if (have_a) {
345 if (have_l) {
346 if (s->above_comp_ctx[col] && td->left_comp_ctx[row7]) {
347 c = 4;
348 } else if (s->above_comp_ctx[col]) {
349 c = 2 + (td->left_intra_ctx[row7] ||
350 td->left_ref_ctx[row7] == s->s.h.fixcompref);
351 } else if (td->left_comp_ctx[row7]) {
352 c = 2 + (s->above_intra_ctx[col] ||
353 s->above_ref_ctx[col] == s->s.h.fixcompref);
354 } else {
355 c = (!s->above_intra_ctx[col] &&
356 s->above_ref_ctx[col] == s->s.h.fixcompref) ^
357 (!td->left_intra_ctx[row7] &&
358 td->left_ref_ctx[row & 7] == s->s.h.fixcompref);
359 }
360 } else {
361 c = s->above_comp_ctx[col] ? 3 :
362 (!s->above_intra_ctx[col] && s->above_ref_ctx[col] == s->s.h.fixcompref);
363 }
364 } else if (have_l) {
365 c = td->left_comp_ctx[row7] ? 3 :
366 (!td->left_intra_ctx[row7] && td->left_ref_ctx[row7] == s->s.h.fixcompref);
367 } else {
368 c = 1;
369 }
370 b->comp = vp56_rac_get_prob(td->c, s->prob.p.comp[c]);
371 td->counts.comp[c][b->comp]++;
372 }
373
374 // read actual references
375 // FIXME probably cache a few variables here to prevent repetitive
376 // memory accesses below
377 if (b->comp) { /* two references */
378 int fix_idx = s->s.h.signbias[s->s.h.fixcompref], var_idx = !fix_idx, c, bit;
379
380 b->ref[fix_idx] = s->s.h.fixcompref;
381 // FIXME can this codeblob be replaced by some sort of LUT?
382 if (have_a) {
383 if (have_l) {
384 if (s->above_intra_ctx[col]) {
385 if (td->left_intra_ctx[row7]) {
386 c = 2;
387 } else {
388 c = 1 + 2 * (td->left_ref_ctx[row7] != s->s.h.varcompref[1]);
389 }
390 } else if (td->left_intra_ctx[row7]) {
391 c = 1 + 2 * (s->above_ref_ctx[col] != s->s.h.varcompref[1]);
392 } else {
393 int refl = td->left_ref_ctx[row7], refa = s->above_ref_ctx[col];
394
395 if (refl == refa && refa == s->s.h.varcompref[1]) {
396 c = 0;
397 } else if (!td->left_comp_ctx[row7] && !s->above_comp_ctx[col]) {
398 if ((refa == s->s.h.fixcompref && refl == s->s.h.varcompref[0]) ||
399 (refl == s->s.h.fixcompref && refa == s->s.h.varcompref[0])) {
400 c = 4;
401 } else {
402 c = (refa == refl) ? 3 : 1;
403 }
404 } else if (!td->left_comp_ctx[row7]) {
405 if (refa == s->s.h.varcompref[1] && refl != s->s.h.varcompref[1]) {
406 c = 1;
407 } else {
408 c = (refl == s->s.h.varcompref[1] &&
409 refa != s->s.h.varcompref[1]) ? 2 : 4;
410 }
411 } else if (!s->above_comp_ctx[col]) {
412 if (refl == s->s.h.varcompref[1] && refa != s->s.h.varcompref[1]) {
413 c = 1;
414 } else {
415 c = (refa == s->s.h.varcompref[1] &&
416 refl != s->s.h.varcompref[1]) ? 2 : 4;
417 }
418 } else {
419 c = (refl == refa) ? 4 : 2;
420 }
421 }
422 } else {
423 if (s->above_intra_ctx[col]) {
424 c = 2;
425 } else if (s->above_comp_ctx[col]) {
426 c = 4 * (s->above_ref_ctx[col] != s->s.h.varcompref[1]);
427 } else {
428 c = 3 * (s->above_ref_ctx[col] != s->s.h.varcompref[1]);
429 }
430 }
431 } else if (have_l) {
432 if (td->left_intra_ctx[row7]) {
433 c = 2;
434 } else if (td->left_comp_ctx[row7]) {
435 c = 4 * (td->left_ref_ctx[row7] != s->s.h.varcompref[1]);
436 } else {
437 c = 3 * (td->left_ref_ctx[row7] != s->s.h.varcompref[1]);
438 }
439 } else {
440 c = 2;
441 }
442 bit = vp56_rac_get_prob(td->c, s->prob.p.comp_ref[c]);
443 b->ref[var_idx] = s->s.h.varcompref[bit];
444 td->counts.comp_ref[c][bit]++;
445 } else /* single reference */ {
446 int bit, c;
447
448 if (have_a && !s->above_intra_ctx[col]) {
449 if (have_l && !td->left_intra_ctx[row7]) {
450 if (td->left_comp_ctx[row7]) {
451 if (s->above_comp_ctx[col]) {
452 c = 1 + (!s->s.h.fixcompref || !td->left_ref_ctx[row7] ||
453 !s->above_ref_ctx[col]);
454 } else {
455 c = (3 * !s->above_ref_ctx[col]) +
456 (!s->s.h.fixcompref || !td->left_ref_ctx[row7]);
457 }
458 } else if (s->above_comp_ctx[col]) {
459 c = (3 * !td->left_ref_ctx[row7]) +
460 (!s->s.h.fixcompref || !s->above_ref_ctx[col]);
461 } else {
462 c = 2 * !td->left_ref_ctx[row7] + 2 * !s->above_ref_ctx[col];
463 }
464 } else if (s->above_intra_ctx[col]) {
465 c = 2;
466 } else if (s->above_comp_ctx[col]) {
467 c = 1 + (!s->s.h.fixcompref || !s->above_ref_ctx[col]);
468 } else {
469 c = 4 * (!s->above_ref_ctx[col]);
470 }
471 } else if (have_l && !td->left_intra_ctx[row7]) {
472 if (td->left_intra_ctx[row7]) {
473 c = 2;
474 } else if (td->left_comp_ctx[row7]) {
475 c = 1 + (!s->s.h.fixcompref || !td->left_ref_ctx[row7]);
476 } else {
477 c = 4 * (!td->left_ref_ctx[row7]);
478 }
479 } else {
480 c = 2;
481 }
482 bit = vp56_rac_get_prob(td->c, s->prob.p.single_ref[c][0]);
483 td->counts.single_ref[c][0][bit]++;
484 if (!bit) {
485 b->ref[0] = 0;
486 } else {
487 // FIXME can this codeblob be replaced by some sort of LUT?
488 if (have_a) {
489 if (have_l) {
490 if (td->left_intra_ctx[row7]) {
491 if (s->above_intra_ctx[col]) {
492 c = 2;
493 } else if (s->above_comp_ctx[col]) {
494 c = 1 + 2 * (s->s.h.fixcompref == 1 ||
495 s->above_ref_ctx[col] == 1);
496 } else if (!s->above_ref_ctx[col]) {
497 c = 3;
498 } else {
499 c = 4 * (s->above_ref_ctx[col] == 1);
500 }
501 } else if (s->above_intra_ctx[col]) {
502 if (td->left_intra_ctx[row7]) {
503 c = 2;
504 } else if (td->left_comp_ctx[row7]) {
505 c = 1 + 2 * (s->s.h.fixcompref == 1 ||
506 td->left_ref_ctx[row7] == 1);
507 } else if (!td->left_ref_ctx[row7]) {
508 c = 3;
509 } else {
510 c = 4 * (td->left_ref_ctx[row7] == 1);
511 }
512 } else if (s->above_comp_ctx[col]) {
513 if (td->left_comp_ctx[row7]) {
514 if (td->left_ref_ctx[row7] == s->above_ref_ctx[col]) {
515 c = 3 * (s->s.h.fixcompref == 1 ||
516 td->left_ref_ctx[row7] == 1);
517 } else {
518 c = 2;
519 }
520 } else if (!td->left_ref_ctx[row7]) {
521 c = 1 + 2 * (s->s.h.fixcompref == 1 ||
522 s->above_ref_ctx[col] == 1);
523 } else {
524 c = 3 * (td->left_ref_ctx[row7] == 1) +
525 (s->s.h.fixcompref == 1 || s->above_ref_ctx[col] == 1);
526 }
527 } else if (td->left_comp_ctx[row7]) {
528 if (!s->above_ref_ctx[col]) {
529 c = 1 + 2 * (s->s.h.fixcompref == 1 ||
530 td->left_ref_ctx[row7] == 1);
531 } else {
532 c = 3 * (s->above_ref_ctx[col] == 1) +
533 (s->s.h.fixcompref == 1 || td->left_ref_ctx[row7] == 1);
534 }
535 } else if (!s->above_ref_ctx[col]) {
536 if (!td->left_ref_ctx[row7]) {
537 c = 3;
538 } else {
539 c = 4 * (td->left_ref_ctx[row7] == 1);
540 }
541 } else if (!td->left_ref_ctx[row7]) {
542 c = 4 * (s->above_ref_ctx[col] == 1);
543 } else {
544 c = 2 * (td->left_ref_ctx[row7] == 1) +
545 2 * (s->above_ref_ctx[col] == 1);
546 }
547 } else {
548 if (s->above_intra_ctx[col] ||
549 (!s->above_comp_ctx[col] && !s->above_ref_ctx[col])) {
550 c = 2;
551 } else if (s->above_comp_ctx[col]) {
552 c = 3 * (s->s.h.fixcompref == 1 || s->above_ref_ctx[col] == 1);
553 } else {
554 c = 4 * (s->above_ref_ctx[col] == 1);
555 }
556 }
557 } else if (have_l) {
558 if (td->left_intra_ctx[row7] ||
559 (!td->left_comp_ctx[row7] && !td->left_ref_ctx[row7])) {
560 c = 2;
561 } else if (td->left_comp_ctx[row7]) {
562 c = 3 * (s->s.h.fixcompref == 1 || td->left_ref_ctx[row7] == 1);
563 } else {
564 c = 4 * (td->left_ref_ctx[row7] == 1);
565 }
566 } else {
567 c = 2;
568 }
569 bit = vp56_rac_get_prob(td->c, s->prob.p.single_ref[c][1]);
570 td->counts.single_ref[c][1][bit]++;
571 b->ref[0] = 1 + bit;
572 }
573 }
574 }
575
576 if (b->bs <= BS_8x8) {
577 if (s->s.h.segmentation.enabled && s->s.h.segmentation.feat[b->seg_id].skip_enabled) {
578 b->mode[0] =
579 b->mode[1] =
580 b->mode[2] =
581 b->mode[3] = ZEROMV;
582 } else {
583 static const uint8_t off[10] = {
584 3, 0, 0, 1, 0, 0, 0, 0, 0, 0
585 };
586
587 // FIXME this needs to use the LUT tables from find_ref_mvs
588 // because not all are -1,0/0,-1
589 int c = inter_mode_ctx_lut[s->above_mode_ctx[col + off[b->bs]]]
590 [td->left_mode_ctx[row7 + off[b->bs]]];
591
592 b->mode[0] = vp8_rac_get_tree(td->c, ff_vp9_inter_mode_tree,
593 s->prob.p.mv_mode[c]);
594 b->mode[1] =
595 b->mode[2] =
596 b->mode[3] = b->mode[0];
597 td->counts.mv_mode[c][b->mode[0] - 10]++;
598 }
599 }
600
601 if (s->s.h.filtermode == FILTER_SWITCHABLE) {
602 int c;
603
604 if (have_a && s->above_mode_ctx[col] >= NEARESTMV) {
605 if (have_l && td->left_mode_ctx[row7] >= NEARESTMV) {
606 c = s->above_filter_ctx[col] == td->left_filter_ctx[row7] ?
607 td->left_filter_ctx[row7] : 3;
608 } else {
609 c = s->above_filter_ctx[col];
610 }
611 } else if (have_l && td->left_mode_ctx[row7] >= NEARESTMV) {
612 c = td->left_filter_ctx[row7];
613 } else {
614 c = 3;
615 }
616
617 filter_id = vp8_rac_get_tree(td->c, ff_vp9_filter_tree,
618 s->prob.p.filter[c]);
619 td->counts.filter[c][filter_id]++;
620 b->filter = ff_vp9_filter_lut[filter_id];
621 } else {
622 b->filter = s->s.h.filtermode;
623 }
624
625 if (b->bs > BS_8x8) {
626 int c = inter_mode_ctx_lut[s->above_mode_ctx[col]][td->left_mode_ctx[row7]];
627
628 b->mode[0] = vp8_rac_get_tree(td->c, ff_vp9_inter_mode_tree,
629 s->prob.p.mv_mode[c]);
630 td->counts.mv_mode[c][b->mode[0] - 10]++;
631 ff_vp9_fill_mv(td, b->mv[0], b->mode[0], 0);
632
633 if (b->bs != BS_8x4) {
634 b->mode[1] = vp8_rac_get_tree(td->c, ff_vp9_inter_mode_tree,
635 s->prob.p.mv_mode[c]);
636 td->counts.mv_mode[c][b->mode[1] - 10]++;
637 ff_vp9_fill_mv(td, b->mv[1], b->mode[1], 1);
638 } else {
639 b->mode[1] = b->mode[0];
640 AV_COPY32(&b->mv[1][0], &b->mv[0][0]);
641 AV_COPY32(&b->mv[1][1], &b->mv[0][1]);
642 }
643
644 if (b->bs != BS_4x8) {
645 b->mode[2] = vp8_rac_get_tree(td->c, ff_vp9_inter_mode_tree,
646 s->prob.p.mv_mode[c]);
647 td->counts.mv_mode[c][b->mode[2] - 10]++;
648 ff_vp9_fill_mv(td, b->mv[2], b->mode[2], 2);
649
650 if (b->bs != BS_8x4) {
651 b->mode[3] = vp8_rac_get_tree(td->c, ff_vp9_inter_mode_tree,
652 s->prob.p.mv_mode[c]);
653 td->counts.mv_mode[c][b->mode[3] - 10]++;
654 ff_vp9_fill_mv(td, b->mv[3], b->mode[3], 3);
655 } else {
656 b->mode[3] = b->mode[2];
657 AV_COPY32(&b->mv[3][0], &b->mv[2][0]);
658 AV_COPY32(&b->mv[3][1], &b->mv[2][1]);
659 }
660 } else {
661 b->mode[2] = b->mode[0];
662 AV_COPY32(&b->mv[2][0], &b->mv[0][0]);
663 AV_COPY32(&b->mv[2][1], &b->mv[0][1]);
664 b->mode[3] = b->mode[1];
665 AV_COPY32(&b->mv[3][0], &b->mv[1][0]);
666 AV_COPY32(&b->mv[3][1], &b->mv[1][1]);
667 }
668 } else {
669 ff_vp9_fill_mv(td, b->mv[0], b->mode[0], -1);
670 AV_COPY32(&b->mv[1][0], &b->mv[0][0]);
671 AV_COPY32(&b->mv[2][0], &b->mv[0][0]);
672 AV_COPY32(&b->mv[3][0], &b->mv[0][0]);
673 AV_COPY32(&b->mv[1][1], &b->mv[0][1]);
674 AV_COPY32(&b->mv[2][1], &b->mv[0][1]);
675 AV_COPY32(&b->mv[3][1], &b->mv[0][1]);
676 }
677
678 vref = b->ref[b->comp ? s->s.h.signbias[s->s.h.varcompref[0]] : 0];
679 }
680
681 #if HAVE_FAST_64BIT
682 #define SPLAT_CTX(var, val, n) \
683 switch (n) { \
684 case 1: var = val; break; \
685 case 2: AV_WN16A(&var, val * 0x0101); break; \
686 case 4: AV_WN32A(&var, val * 0x01010101); break; \
687 case 8: AV_WN64A(&var, val * 0x0101010101010101ULL); break; \
688 case 16: { \
689 uint64_t v64 = val * 0x0101010101010101ULL; \
690 AV_WN64A( &var, v64); \
691 AV_WN64A(&((uint8_t *) &var)[8], v64); \
692 break; \
693 } \
694 }
695 #else
696 #define SPLAT_CTX(var, val, n) \
697 switch (n) { \
698 case 1: var = val; break; \
699 case 2: AV_WN16A(&var, val * 0x0101); break; \
700 case 4: AV_WN32A(&var, val * 0x01010101); break; \
701 case 8: { \
702 uint32_t v32 = val * 0x01010101; \
703 AV_WN32A( &var, v32); \
704 AV_WN32A(&((uint8_t *) &var)[4], v32); \
705 break; \
706 } \
707 case 16: { \
708 uint32_t v32 = val * 0x01010101; \
709 AV_WN32A( &var, v32); \
710 AV_WN32A(&((uint8_t *) &var)[4], v32); \
711 AV_WN32A(&((uint8_t *) &var)[8], v32); \
712 AV_WN32A(&((uint8_t *) &var)[12], v32); \
713 break; \
714 } \
715 }
716 #endif
717
718 switch (ff_vp9_bwh_tab[1][b->bs][0]) {
719 #define SET_CTXS(perf, dir, off, n) \
720 do { \
721 SPLAT_CTX(perf->dir##_skip_ctx[off], b->skip, n); \
722 SPLAT_CTX(perf->dir##_txfm_ctx[off], b->tx, n); \
723 SPLAT_CTX(perf->dir##_partition_ctx[off], dir##_ctx[b->bs], n); \
724 if (!s->s.h.keyframe && !s->s.h.intraonly) { \
725 SPLAT_CTX(perf->dir##_intra_ctx[off], b->intra, n); \
726 SPLAT_CTX(perf->dir##_comp_ctx[off], b->comp, n); \
727 SPLAT_CTX(perf->dir##_mode_ctx[off], b->mode[3], n); \
728 if (!b->intra) { \
729 SPLAT_CTX(perf->dir##_ref_ctx[off], vref, n); \
730 if (s->s.h.filtermode == FILTER_SWITCHABLE) { \
731 SPLAT_CTX(perf->dir##_filter_ctx[off], filter_id, n); \
732 } \
733 } \
734 } \
735 } while (0)
736 case 1: SET_CTXS(s, above, col, 1); break;
737 case 2: SET_CTXS(s, above, col, 2); break;
738 case 4: SET_CTXS(s, above, col, 4); break;
739 case 8: SET_CTXS(s, above, col, 8); break;
740 }
741 switch (ff_vp9_bwh_tab[1][b->bs][1]) {
742 case 1: SET_CTXS(td, left, row7, 1); break;
743 case 2: SET_CTXS(td, left, row7, 2); break;
744 case 4: SET_CTXS(td, left, row7, 4); break;
745 case 8: SET_CTXS(td, left, row7, 8); break;
746 }
747 #undef SPLAT_CTX
748 #undef SET_CTXS
749
750 if (!s->s.h.keyframe && !s->s.h.intraonly) {
751 if (b->bs > BS_8x8) {
752 int mv0 = AV_RN32A(&b->mv[3][0]), mv1 = AV_RN32A(&b->mv[3][1]);
753
754 AV_COPY32(&td->left_mv_ctx[row7 * 2 + 0][0], &b->mv[1][0]);
755 AV_COPY32(&td->left_mv_ctx[row7 * 2 + 0][1], &b->mv[1][1]);
756 AV_WN32A(&td->left_mv_ctx[row7 * 2 + 1][0], mv0);
757 AV_WN32A(&td->left_mv_ctx[row7 * 2 + 1][1], mv1);
758 AV_COPY32(&s->above_mv_ctx[col * 2 + 0][0], &b->mv[2][0]);
759 AV_COPY32(&s->above_mv_ctx[col * 2 + 0][1], &b->mv[2][1]);
760 AV_WN32A(&s->above_mv_ctx[col * 2 + 1][0], mv0);
761 AV_WN32A(&s->above_mv_ctx[col * 2 + 1][1], mv1);
762 } else {
763 int n, mv0 = AV_RN32A(&b->mv[3][0]), mv1 = AV_RN32A(&b->mv[3][1]);
764
765 for (n = 0; n < w4 * 2; n++) {
766 AV_WN32A(&s->above_mv_ctx[col * 2 + n][0], mv0);
767 AV_WN32A(&s->above_mv_ctx[col * 2 + n][1], mv1);
768 }
769 for (n = 0; n < h4 * 2; n++) {
770 AV_WN32A(&td->left_mv_ctx[row7 * 2 + n][0], mv0);
771 AV_WN32A(&td->left_mv_ctx[row7 * 2 + n][1], mv1);
772 }
773 }
774 }
775
776 // FIXME kinda ugly
777 for (y = 0; y < h4; y++) {
778 int x, o = (row + y) * s->sb_cols * 8 + col;
779 VP9mvrefPair *mv = &s->s.frames[CUR_FRAME].mv[o];
780
781 if (b->intra) {
782 for (x = 0; x < w4; x++) {
783 mv[x].ref[0] =
784 mv[x].ref[1] = -1;
785 }
786 } else if (b->comp) {
787 for (x = 0; x < w4; x++) {
788 mv[x].ref[0] = b->ref[0];
789 mv[x].ref[1] = b->ref[1];
790 AV_COPY32(&mv[x].mv[0], &b->mv[3][0]);
791 AV_COPY32(&mv[x].mv[1], &b->mv[3][1]);
792 }
793 } else {
794 for (x = 0; x < w4; x++) {
795 mv[x].ref[0] = b->ref[0];
796 mv[x].ref[1] = -1;
797 AV_COPY32(&mv[x].mv[0], &b->mv[3][0]);
798 }
799 }
800 }
801 }
802
803 // FIXME merge cnt/eob arguments?
804 static av_always_inline int
decode_coeffs_b_generic(VP56RangeCoder * c,int16_t * coef,int n_coeffs,int is_tx32x32,int is8bitsperpixel,int bpp,unsigned (* cnt)[6][3],unsigned (* eob)[6][2],uint8_t (* p)[6][11],int nnz,const int16_t * scan,const int16_t (* nb)[2],const int16_t * band_counts,int16_t * qmul)805 decode_coeffs_b_generic(VP56RangeCoder *c, int16_t *coef, int n_coeffs,
806 int is_tx32x32, int is8bitsperpixel, int bpp, unsigned (*cnt)[6][3],
807 unsigned (*eob)[6][2], uint8_t (*p)[6][11],
808 int nnz, const int16_t *scan, const int16_t (*nb)[2],
809 const int16_t *band_counts, int16_t *qmul)
810 {
811 int i = 0, band = 0, band_left = band_counts[band];
812 const uint8_t *tp = p[0][nnz];
813 uint8_t cache[1024];
814
815 do {
816 int val, rc;
817
818 val = vp56_rac_get_prob_branchy(c, tp[0]); // eob
819 eob[band][nnz][val]++;
820 if (!val)
821 break;
822
823 skip_eob:
824 if (!vp56_rac_get_prob_branchy(c, tp[1])) { // zero
825 cnt[band][nnz][0]++;
826 if (!--band_left)
827 band_left = band_counts[++band];
828 cache[scan[i]] = 0;
829 nnz = (1 + cache[nb[i][0]] + cache[nb[i][1]]) >> 1;
830 tp = p[band][nnz];
831 if (++i == n_coeffs)
832 break; //invalid input; blocks should end with EOB
833 goto skip_eob;
834 }
835
836 rc = scan[i];
837 if (!vp56_rac_get_prob_branchy(c, tp[2])) { // one
838 cnt[band][nnz][1]++;
839 val = 1;
840 cache[rc] = 1;
841 } else {
842 cnt[band][nnz][2]++;
843 if (!vp56_rac_get_prob_branchy(c, tp[3])) { // 2, 3, 4
844 if (!vp56_rac_get_prob_branchy(c, tp[4])) {
845 cache[rc] = val = 2;
846 } else {
847 val = 3 + vp56_rac_get_prob(c, tp[5]);
848 cache[rc] = 3;
849 }
850 } else if (!vp56_rac_get_prob_branchy(c, tp[6])) { // cat1/2
851 cache[rc] = 4;
852 if (!vp56_rac_get_prob_branchy(c, tp[7])) {
853 val = vp56_rac_get_prob(c, 159) + 5;
854 } else {
855 val = (vp56_rac_get_prob(c, 165) << 1) + 7;
856 val += vp56_rac_get_prob(c, 145);
857 }
858 } else { // cat 3-6
859 cache[rc] = 5;
860 if (!vp56_rac_get_prob_branchy(c, tp[8])) {
861 if (!vp56_rac_get_prob_branchy(c, tp[9])) {
862 val = 11 + (vp56_rac_get_prob(c, 173) << 2);
863 val += (vp56_rac_get_prob(c, 148) << 1);
864 val += vp56_rac_get_prob(c, 140);
865 } else {
866 val = 19 + (vp56_rac_get_prob(c, 176) << 3);
867 val += (vp56_rac_get_prob(c, 155) << 2);
868 val += (vp56_rac_get_prob(c, 140) << 1);
869 val += vp56_rac_get_prob(c, 135);
870 }
871 } else if (!vp56_rac_get_prob_branchy(c, tp[10])) {
872 val = (vp56_rac_get_prob(c, 180) << 4) + 35;
873 val += (vp56_rac_get_prob(c, 157) << 3);
874 val += (vp56_rac_get_prob(c, 141) << 2);
875 val += (vp56_rac_get_prob(c, 134) << 1);
876 val += vp56_rac_get_prob(c, 130);
877 } else {
878 val = 67;
879 if (!is8bitsperpixel) {
880 if (bpp == 12) {
881 val += vp56_rac_get_prob(c, 255) << 17;
882 val += vp56_rac_get_prob(c, 255) << 16;
883 }
884 val += (vp56_rac_get_prob(c, 255) << 15);
885 val += (vp56_rac_get_prob(c, 255) << 14);
886 }
887 val += (vp56_rac_get_prob(c, 254) << 13);
888 val += (vp56_rac_get_prob(c, 254) << 12);
889 val += (vp56_rac_get_prob(c, 254) << 11);
890 val += (vp56_rac_get_prob(c, 252) << 10);
891 val += (vp56_rac_get_prob(c, 249) << 9);
892 val += (vp56_rac_get_prob(c, 243) << 8);
893 val += (vp56_rac_get_prob(c, 230) << 7);
894 val += (vp56_rac_get_prob(c, 196) << 6);
895 val += (vp56_rac_get_prob(c, 177) << 5);
896 val += (vp56_rac_get_prob(c, 153) << 4);
897 val += (vp56_rac_get_prob(c, 140) << 3);
898 val += (vp56_rac_get_prob(c, 133) << 2);
899 val += (vp56_rac_get_prob(c, 130) << 1);
900 val += vp56_rac_get_prob(c, 129);
901 }
902 }
903 }
904 #define STORE_COEF(c, i, v) do { \
905 if (is8bitsperpixel) { \
906 c[i] = v; \
907 } else { \
908 AV_WN32A(&c[i * 2], v); \
909 } \
910 } while (0)
911 if (!--band_left)
912 band_left = band_counts[++band];
913 if (is_tx32x32)
914 STORE_COEF(coef, rc, (int)((vp8_rac_get(c) ? -val : val) * (unsigned)qmul[!!i]) / 2);
915 else
916 STORE_COEF(coef, rc, (vp8_rac_get(c) ? -val : val) * (unsigned)qmul[!!i]);
917 nnz = (1 + cache[nb[i][0]] + cache[nb[i][1]]) >> 1;
918 tp = p[band][nnz];
919 } while (++i < n_coeffs);
920
921 return i;
922 }
923
decode_coeffs_b_8bpp(VP9TileData * td,int16_t * coef,int n_coeffs,unsigned (* cnt)[6][3],unsigned (* eob)[6][2],uint8_t (* p)[6][11],int nnz,const int16_t * scan,const int16_t (* nb)[2],const int16_t * band_counts,int16_t * qmul)924 static int decode_coeffs_b_8bpp(VP9TileData *td, int16_t *coef, int n_coeffs,
925 unsigned (*cnt)[6][3], unsigned (*eob)[6][2],
926 uint8_t (*p)[6][11], int nnz, const int16_t *scan,
927 const int16_t (*nb)[2], const int16_t *band_counts,
928 int16_t *qmul)
929 {
930 return decode_coeffs_b_generic(td->c, coef, n_coeffs, 0, 1, 8, cnt, eob, p,
931 nnz, scan, nb, band_counts, qmul);
932 }
933
decode_coeffs_b32_8bpp(VP9TileData * td,int16_t * coef,int n_coeffs,unsigned (* cnt)[6][3],unsigned (* eob)[6][2],uint8_t (* p)[6][11],int nnz,const int16_t * scan,const int16_t (* nb)[2],const int16_t * band_counts,int16_t * qmul)934 static int decode_coeffs_b32_8bpp(VP9TileData *td, int16_t *coef, int n_coeffs,
935 unsigned (*cnt)[6][3], unsigned (*eob)[6][2],
936 uint8_t (*p)[6][11], int nnz, const int16_t *scan,
937 const int16_t (*nb)[2], const int16_t *band_counts,
938 int16_t *qmul)
939 {
940 return decode_coeffs_b_generic(td->c, coef, n_coeffs, 1, 1, 8, cnt, eob, p,
941 nnz, scan, nb, band_counts, qmul);
942 }
943
decode_coeffs_b_16bpp(VP9TileData * td,int16_t * coef,int n_coeffs,unsigned (* cnt)[6][3],unsigned (* eob)[6][2],uint8_t (* p)[6][11],int nnz,const int16_t * scan,const int16_t (* nb)[2],const int16_t * band_counts,int16_t * qmul)944 static int decode_coeffs_b_16bpp(VP9TileData *td, int16_t *coef, int n_coeffs,
945 unsigned (*cnt)[6][3], unsigned (*eob)[6][2],
946 uint8_t (*p)[6][11], int nnz, const int16_t *scan,
947 const int16_t (*nb)[2], const int16_t *band_counts,
948 int16_t *qmul)
949 {
950 return decode_coeffs_b_generic(td->c, coef, n_coeffs, 0, 0, td->s->s.h.bpp, cnt, eob, p,
951 nnz, scan, nb, band_counts, qmul);
952 }
953
decode_coeffs_b32_16bpp(VP9TileData * td,int16_t * coef,int n_coeffs,unsigned (* cnt)[6][3],unsigned (* eob)[6][2],uint8_t (* p)[6][11],int nnz,const int16_t * scan,const int16_t (* nb)[2],const int16_t * band_counts,int16_t * qmul)954 static int decode_coeffs_b32_16bpp(VP9TileData *td, int16_t *coef, int n_coeffs,
955 unsigned (*cnt)[6][3], unsigned (*eob)[6][2],
956 uint8_t (*p)[6][11], int nnz, const int16_t *scan,
957 const int16_t (*nb)[2], const int16_t *band_counts,
958 int16_t *qmul)
959 {
960 return decode_coeffs_b_generic(td->c, coef, n_coeffs, 1, 0, td->s->s.h.bpp, cnt, eob, p,
961 nnz, scan, nb, band_counts, qmul);
962 }
963
decode_coeffs(VP9TileData * td,int is8bitsperpixel)964 static av_always_inline int decode_coeffs(VP9TileData *td, int is8bitsperpixel)
965 {
966 VP9Context *s = td->s;
967 VP9Block *b = td->b;
968 int row = td->row, col = td->col;
969 uint8_t (*p)[6][11] = s->prob.coef[b->tx][0 /* y */][!b->intra];
970 unsigned (*c)[6][3] = td->counts.coef[b->tx][0 /* y */][!b->intra];
971 unsigned (*e)[6][2] = td->counts.eob[b->tx][0 /* y */][!b->intra];
972 int w4 = ff_vp9_bwh_tab[1][b->bs][0] << 1, h4 = ff_vp9_bwh_tab[1][b->bs][1] << 1;
973 int end_x = FFMIN(2 * (s->cols - col), w4);
974 int end_y = FFMIN(2 * (s->rows - row), h4);
975 int n, pl, x, y, ret;
976 int16_t (*qmul)[2] = s->s.h.segmentation.feat[b->seg_id].qmul;
977 int tx = 4 * s->s.h.lossless + b->tx;
978 const int16_t * const *yscans = ff_vp9_scans[tx];
979 const int16_t (* const * ynbs)[2] = ff_vp9_scans_nb[tx];
980 const int16_t *uvscan = ff_vp9_scans[b->uvtx][DCT_DCT];
981 const int16_t (*uvnb)[2] = ff_vp9_scans_nb[b->uvtx][DCT_DCT];
982 uint8_t *a = &s->above_y_nnz_ctx[col * 2];
983 uint8_t *l = &td->left_y_nnz_ctx[(row & 7) << 1];
984 static const int16_t band_counts[4][8] = {
985 { 1, 2, 3, 4, 3, 16 - 13 },
986 { 1, 2, 3, 4, 11, 64 - 21 },
987 { 1, 2, 3, 4, 11, 256 - 21 },
988 { 1, 2, 3, 4, 11, 1024 - 21 },
989 };
990 const int16_t *y_band_counts = band_counts[b->tx];
991 const int16_t *uv_band_counts = band_counts[b->uvtx];
992 int bytesperpixel = is8bitsperpixel ? 1 : 2;
993 int total_coeff = 0;
994
995 #define MERGE(la, end, step, rd) \
996 for (n = 0; n < end; n += step) \
997 la[n] = !!rd(&la[n])
998 #define MERGE_CTX(step, rd) \
999 do { \
1000 MERGE(l, end_y, step, rd); \
1001 MERGE(a, end_x, step, rd); \
1002 } while (0)
1003
1004 #define DECODE_Y_COEF_LOOP(step, mode_index, v) \
1005 for (n = 0, y = 0; y < end_y; y += step) { \
1006 for (x = 0; x < end_x; x += step, n += step * step) { \
1007 enum TxfmType txtp = ff_vp9_intra_txfm_type[b->mode[mode_index]]; \
1008 ret = (is8bitsperpixel ? decode_coeffs_b##v##_8bpp : decode_coeffs_b##v##_16bpp) \
1009 (td, td->block + 16 * n * bytesperpixel, 16 * step * step, \
1010 c, e, p, a[x] + l[y], yscans[txtp], \
1011 ynbs[txtp], y_band_counts, qmul[0]); \
1012 a[x] = l[y] = !!ret; \
1013 total_coeff |= !!ret; \
1014 if (step >= 4) { \
1015 AV_WN16A(&td->eob[n], ret); \
1016 } else { \
1017 td->eob[n] = ret; \
1018 } \
1019 } \
1020 }
1021
1022 #define SPLAT(la, end, step, cond) \
1023 if (step == 2) { \
1024 for (n = 1; n < end; n += step) \
1025 la[n] = la[n - 1]; \
1026 } else if (step == 4) { \
1027 if (cond) { \
1028 for (n = 0; n < end; n += step) \
1029 AV_WN32A(&la[n], la[n] * 0x01010101); \
1030 } else { \
1031 for (n = 0; n < end; n += step) \
1032 memset(&la[n + 1], la[n], FFMIN(end - n - 1, 3)); \
1033 } \
1034 } else /* step == 8 */ { \
1035 if (cond) { \
1036 if (HAVE_FAST_64BIT) { \
1037 for (n = 0; n < end; n += step) \
1038 AV_WN64A(&la[n], la[n] * 0x0101010101010101ULL); \
1039 } else { \
1040 for (n = 0; n < end; n += step) { \
1041 uint32_t v32 = la[n] * 0x01010101; \
1042 AV_WN32A(&la[n], v32); \
1043 AV_WN32A(&la[n + 4], v32); \
1044 } \
1045 } \
1046 } else { \
1047 for (n = 0; n < end; n += step) \
1048 memset(&la[n + 1], la[n], FFMIN(end - n - 1, 7)); \
1049 } \
1050 }
1051 #define SPLAT_CTX(step) \
1052 do { \
1053 SPLAT(a, end_x, step, end_x == w4); \
1054 SPLAT(l, end_y, step, end_y == h4); \
1055 } while (0)
1056
1057 /* y tokens */
1058 switch (b->tx) {
1059 case TX_4X4:
1060 DECODE_Y_COEF_LOOP(1, b->bs > BS_8x8 ? n : 0,);
1061 break;
1062 case TX_8X8:
1063 MERGE_CTX(2, AV_RN16A);
1064 DECODE_Y_COEF_LOOP(2, 0,);
1065 SPLAT_CTX(2);
1066 break;
1067 case TX_16X16:
1068 MERGE_CTX(4, AV_RN32A);
1069 DECODE_Y_COEF_LOOP(4, 0,);
1070 SPLAT_CTX(4);
1071 break;
1072 case TX_32X32:
1073 MERGE_CTX(8, AV_RN64A);
1074 DECODE_Y_COEF_LOOP(8, 0, 32);
1075 SPLAT_CTX(8);
1076 break;
1077 }
1078
1079 #define DECODE_UV_COEF_LOOP(step, v) \
1080 for (n = 0, y = 0; y < end_y; y += step) { \
1081 for (x = 0; x < end_x; x += step, n += step * step) { \
1082 ret = (is8bitsperpixel ? decode_coeffs_b##v##_8bpp : decode_coeffs_b##v##_16bpp) \
1083 (td, td->uvblock[pl] + 16 * n * bytesperpixel, \
1084 16 * step * step, c, e, p, a[x] + l[y], \
1085 uvscan, uvnb, uv_band_counts, qmul[1]); \
1086 a[x] = l[y] = !!ret; \
1087 total_coeff |= !!ret; \
1088 if (step >= 4) { \
1089 AV_WN16A(&td->uveob[pl][n], ret); \
1090 } else { \
1091 td->uveob[pl][n] = ret; \
1092 } \
1093 } \
1094 }
1095
1096 p = s->prob.coef[b->uvtx][1 /* uv */][!b->intra];
1097 c = td->counts.coef[b->uvtx][1 /* uv */][!b->intra];
1098 e = td->counts.eob[b->uvtx][1 /* uv */][!b->intra];
1099 w4 >>= s->ss_h;
1100 end_x >>= s->ss_h;
1101 h4 >>= s->ss_v;
1102 end_y >>= s->ss_v;
1103 for (pl = 0; pl < 2; pl++) {
1104 a = &s->above_uv_nnz_ctx[pl][col << !s->ss_h];
1105 l = &td->left_uv_nnz_ctx[pl][(row & 7) << !s->ss_v];
1106 switch (b->uvtx) {
1107 case TX_4X4:
1108 DECODE_UV_COEF_LOOP(1,);
1109 break;
1110 case TX_8X8:
1111 MERGE_CTX(2, AV_RN16A);
1112 DECODE_UV_COEF_LOOP(2,);
1113 SPLAT_CTX(2);
1114 break;
1115 case TX_16X16:
1116 MERGE_CTX(4, AV_RN32A);
1117 DECODE_UV_COEF_LOOP(4,);
1118 SPLAT_CTX(4);
1119 break;
1120 case TX_32X32:
1121 MERGE_CTX(8, AV_RN64A);
1122 DECODE_UV_COEF_LOOP(8, 32);
1123 SPLAT_CTX(8);
1124 break;
1125 }
1126 }
1127
1128 return total_coeff;
1129 }
1130
decode_coeffs_8bpp(VP9TileData * td)1131 static int decode_coeffs_8bpp(VP9TileData *td)
1132 {
1133 return decode_coeffs(td, 1);
1134 }
1135
decode_coeffs_16bpp(VP9TileData * td)1136 static int decode_coeffs_16bpp(VP9TileData *td)
1137 {
1138 return decode_coeffs(td, 0);
1139 }
1140
mask_edges(uint8_t (* mask)[8][4],int ss_h,int ss_v,int row_and_7,int col_and_7,int w,int h,int col_end,int row_end,enum TxfmMode tx,int skip_inter)1141 static av_always_inline void mask_edges(uint8_t (*mask)[8][4], int ss_h, int ss_v,
1142 int row_and_7, int col_and_7,
1143 int w, int h, int col_end, int row_end,
1144 enum TxfmMode tx, int skip_inter)
1145 {
1146 static const unsigned wide_filter_col_mask[2] = { 0x11, 0x01 };
1147 static const unsigned wide_filter_row_mask[2] = { 0x03, 0x07 };
1148
1149 // FIXME I'm pretty sure all loops can be replaced by a single LUT if
1150 // we make VP9Filter.mask uint64_t (i.e. row/col all single variable)
1151 // and make the LUT 5-indexed (bl, bp, is_uv, tx and row/col), and then
1152 // use row_and_7/col_and_7 as shifts (1*col_and_7+8*row_and_7)
1153
1154 // the intended behaviour of the vp9 loopfilter is to work on 8-pixel
1155 // edges. This means that for UV, we work on two subsampled blocks at
1156 // a time, and we only use the topleft block's mode information to set
1157 // things like block strength. Thus, for any block size smaller than
1158 // 16x16, ignore the odd portion of the block.
1159 if (tx == TX_4X4 && (ss_v | ss_h)) {
1160 if (h == ss_v) {
1161 if (row_and_7 & 1)
1162 return;
1163 if (!row_end)
1164 h += 1;
1165 }
1166 if (w == ss_h) {
1167 if (col_and_7 & 1)
1168 return;
1169 if (!col_end)
1170 w += 1;
1171 }
1172 }
1173
1174 if (tx == TX_4X4 && !skip_inter) {
1175 int t = 1 << col_and_7, m_col = (t << w) - t, y;
1176 // on 32-px edges, use the 8-px wide loopfilter; else, use 4-px wide
1177 int m_row_8 = m_col & wide_filter_col_mask[ss_h], m_row_4 = m_col - m_row_8;
1178
1179 for (y = row_and_7; y < h + row_and_7; y++) {
1180 int col_mask_id = 2 - !(y & wide_filter_row_mask[ss_v]);
1181
1182 mask[0][y][1] |= m_row_8;
1183 mask[0][y][2] |= m_row_4;
1184 // for odd lines, if the odd col is not being filtered,
1185 // skip odd row also:
1186 // .---. <-- a
1187 // | |
1188 // |___| <-- b
1189 // ^ ^
1190 // c d
1191 //
1192 // if a/c are even row/col and b/d are odd, and d is skipped,
1193 // e.g. right edge of size-66x66.webm, then skip b also (bug)
1194 if ((ss_h & ss_v) && (col_end & 1) && (y & 1)) {
1195 mask[1][y][col_mask_id] |= (t << (w - 1)) - t;
1196 } else {
1197 mask[1][y][col_mask_id] |= m_col;
1198 }
1199 if (!ss_h)
1200 mask[0][y][3] |= m_col;
1201 if (!ss_v) {
1202 if (ss_h && (col_end & 1))
1203 mask[1][y][3] |= (t << (w - 1)) - t;
1204 else
1205 mask[1][y][3] |= m_col;
1206 }
1207 }
1208 } else {
1209 int y, t = 1 << col_and_7, m_col = (t << w) - t;
1210
1211 if (!skip_inter) {
1212 int mask_id = (tx == TX_8X8);
1213 int l2 = tx + ss_h - 1, step1d;
1214 static const unsigned masks[4] = { 0xff, 0x55, 0x11, 0x01 };
1215 int m_row = m_col & masks[l2];
1216
1217 // at odd UV col/row edges tx16/tx32 loopfilter edges, force
1218 // 8wd loopfilter to prevent going off the visible edge.
1219 if (ss_h && tx > TX_8X8 && (w ^ (w - 1)) == 1) {
1220 int m_row_16 = ((t << (w - 1)) - t) & masks[l2];
1221 int m_row_8 = m_row - m_row_16;
1222
1223 for (y = row_and_7; y < h + row_and_7; y++) {
1224 mask[0][y][0] |= m_row_16;
1225 mask[0][y][1] |= m_row_8;
1226 }
1227 } else {
1228 for (y = row_and_7; y < h + row_and_7; y++)
1229 mask[0][y][mask_id] |= m_row;
1230 }
1231
1232 l2 = tx + ss_v - 1;
1233 step1d = 1 << l2;
1234 if (ss_v && tx > TX_8X8 && (h ^ (h - 1)) == 1) {
1235 for (y = row_and_7; y < h + row_and_7 - 1; y += step1d)
1236 mask[1][y][0] |= m_col;
1237 if (y - row_and_7 == h - 1)
1238 mask[1][y][1] |= m_col;
1239 } else {
1240 for (y = row_and_7; y < h + row_and_7; y += step1d)
1241 mask[1][y][mask_id] |= m_col;
1242 }
1243 } else if (tx != TX_4X4) {
1244 int mask_id;
1245
1246 mask_id = (tx == TX_8X8) || (h == ss_v);
1247 mask[1][row_and_7][mask_id] |= m_col;
1248 mask_id = (tx == TX_8X8) || (w == ss_h);
1249 for (y = row_and_7; y < h + row_and_7; y++)
1250 mask[0][y][mask_id] |= t;
1251 } else {
1252 int t8 = t & wide_filter_col_mask[ss_h], t4 = t - t8;
1253
1254 for (y = row_and_7; y < h + row_and_7; y++) {
1255 mask[0][y][2] |= t4;
1256 mask[0][y][1] |= t8;
1257 }
1258 mask[1][row_and_7][2 - !(row_and_7 & wide_filter_row_mask[ss_v])] |= m_col;
1259 }
1260 }
1261 }
1262
ff_vp9_decode_block(VP9TileData * td,int row,int col,VP9Filter * lflvl,ptrdiff_t yoff,ptrdiff_t uvoff,enum BlockLevel bl,enum BlockPartition bp)1263 void ff_vp9_decode_block(VP9TileData *td, int row, int col,
1264 VP9Filter *lflvl, ptrdiff_t yoff, ptrdiff_t uvoff,
1265 enum BlockLevel bl, enum BlockPartition bp)
1266 {
1267 VP9Context *s = td->s;
1268 VP9Block *b = td->b;
1269 enum BlockSize bs = bl * 3 + bp;
1270 int bytesperpixel = s->bytesperpixel;
1271 int w4 = ff_vp9_bwh_tab[1][bs][0], h4 = ff_vp9_bwh_tab[1][bs][1], lvl;
1272 int emu[2];
1273 AVFrame *f = s->s.frames[CUR_FRAME].tf.f;
1274
1275 td->row = row;
1276 td->row7 = row & 7;
1277 td->col = col;
1278 td->col7 = col & 7;
1279
1280 td->min_mv.x = -(128 + col * 64);
1281 td->min_mv.y = -(128 + row * 64);
1282 td->max_mv.x = 128 + (s->cols - col - w4) * 64;
1283 td->max_mv.y = 128 + (s->rows - row - h4) * 64;
1284
1285 if (s->pass < 2) {
1286 b->bs = bs;
1287 b->bl = bl;
1288 b->bp = bp;
1289 decode_mode(td);
1290 b->uvtx = b->tx - ((s->ss_h && w4 * 2 == (1 << b->tx)) ||
1291 (s->ss_v && h4 * 2 == (1 << b->tx)));
1292
1293 if (td->block_structure) {
1294 td->block_structure[td->nb_block_structure].row = row;
1295 td->block_structure[td->nb_block_structure].col = col;
1296 td->block_structure[td->nb_block_structure].block_size_idx_x = av_log2(w4);
1297 td->block_structure[td->nb_block_structure].block_size_idx_y = av_log2(h4);
1298 td->nb_block_structure++;
1299 }
1300
1301 if (!b->skip) {
1302 int has_coeffs;
1303
1304 if (bytesperpixel == 1) {
1305 has_coeffs = decode_coeffs_8bpp(td);
1306 } else {
1307 has_coeffs = decode_coeffs_16bpp(td);
1308 }
1309 if (!has_coeffs && b->bs <= BS_8x8 && !b->intra) {
1310 b->skip = 1;
1311 memset(&s->above_skip_ctx[col], 1, w4);
1312 memset(&td->left_skip_ctx[td->row7], 1, h4);
1313 }
1314 } else {
1315 int row7 = td->row7;
1316
1317 #define SPLAT_ZERO_CTX(v, n) \
1318 switch (n) { \
1319 case 1: v = 0; break; \
1320 case 2: AV_ZERO16(&v); break; \
1321 case 4: AV_ZERO32(&v); break; \
1322 case 8: AV_ZERO64(&v); break; \
1323 case 16: AV_ZERO128(&v); break; \
1324 }
1325 #define SPLAT_ZERO_YUV(dir, var, off, n, dir2) \
1326 do { \
1327 SPLAT_ZERO_CTX(dir##_y_##var[off * 2], n * 2); \
1328 if (s->ss_##dir2) { \
1329 SPLAT_ZERO_CTX(dir##_uv_##var[0][off], n); \
1330 SPLAT_ZERO_CTX(dir##_uv_##var[1][off], n); \
1331 } else { \
1332 SPLAT_ZERO_CTX(dir##_uv_##var[0][off * 2], n * 2); \
1333 SPLAT_ZERO_CTX(dir##_uv_##var[1][off * 2], n * 2); \
1334 } \
1335 } while (0)
1336
1337 switch (w4) {
1338 case 1: SPLAT_ZERO_YUV(s->above, nnz_ctx, col, 1, h); break;
1339 case 2: SPLAT_ZERO_YUV(s->above, nnz_ctx, col, 2, h); break;
1340 case 4: SPLAT_ZERO_YUV(s->above, nnz_ctx, col, 4, h); break;
1341 case 8: SPLAT_ZERO_YUV(s->above, nnz_ctx, col, 8, h); break;
1342 }
1343 switch (h4) {
1344 case 1: SPLAT_ZERO_YUV(td->left, nnz_ctx, row7, 1, v); break;
1345 case 2: SPLAT_ZERO_YUV(td->left, nnz_ctx, row7, 2, v); break;
1346 case 4: SPLAT_ZERO_YUV(td->left, nnz_ctx, row7, 4, v); break;
1347 case 8: SPLAT_ZERO_YUV(td->left, nnz_ctx, row7, 8, v); break;
1348 }
1349 }
1350
1351 if (s->pass == 1) {
1352 s->td[0].b++;
1353 s->td[0].block += w4 * h4 * 64 * bytesperpixel;
1354 s->td[0].uvblock[0] += w4 * h4 * 64 * bytesperpixel >> (s->ss_h + s->ss_v);
1355 s->td[0].uvblock[1] += w4 * h4 * 64 * bytesperpixel >> (s->ss_h + s->ss_v);
1356 s->td[0].eob += 4 * w4 * h4;
1357 s->td[0].uveob[0] += 4 * w4 * h4 >> (s->ss_h + s->ss_v);
1358 s->td[0].uveob[1] += 4 * w4 * h4 >> (s->ss_h + s->ss_v);
1359
1360 return;
1361 }
1362 }
1363
1364 // emulated overhangs if the stride of the target buffer can't hold. This
1365 // makes it possible to support emu-edge and so on even if we have large block
1366 // overhangs
1367 emu[0] = (col + w4) * 8 * bytesperpixel > f->linesize[0] ||
1368 (row + h4) > s->rows;
1369 emu[1] = ((col + w4) * 8 >> s->ss_h) * bytesperpixel > f->linesize[1] ||
1370 (row + h4) > s->rows;
1371 if (emu[0]) {
1372 td->dst[0] = td->tmp_y;
1373 td->y_stride = 128;
1374 } else {
1375 td->dst[0] = f->data[0] + yoff;
1376 td->y_stride = f->linesize[0];
1377 }
1378 if (emu[1]) {
1379 td->dst[1] = td->tmp_uv[0];
1380 td->dst[2] = td->tmp_uv[1];
1381 td->uv_stride = 128;
1382 } else {
1383 td->dst[1] = f->data[1] + uvoff;
1384 td->dst[2] = f->data[2] + uvoff;
1385 td->uv_stride = f->linesize[1];
1386 }
1387 if (b->intra) {
1388 if (s->s.h.bpp > 8) {
1389 ff_vp9_intra_recon_16bpp(td, yoff, uvoff);
1390 } else {
1391 ff_vp9_intra_recon_8bpp(td, yoff, uvoff);
1392 }
1393 } else {
1394 if (s->s.h.bpp > 8) {
1395 ff_vp9_inter_recon_16bpp(td);
1396 } else {
1397 ff_vp9_inter_recon_8bpp(td);
1398 }
1399 }
1400 if (emu[0]) {
1401 int w = FFMIN(s->cols - col, w4) * 8, h = FFMIN(s->rows - row, h4) * 8, n, o = 0;
1402
1403 for (n = 0; o < w; n++) {
1404 int bw = 64 >> n;
1405
1406 av_assert2(n <= 4);
1407 if (w & bw) {
1408 s->dsp.mc[n][0][0][0][0](f->data[0] + yoff + o * bytesperpixel, f->linesize[0],
1409 td->tmp_y + o * bytesperpixel, 128, h, 0, 0);
1410 o += bw;
1411 }
1412 }
1413 }
1414 if (emu[1]) {
1415 int w = FFMIN(s->cols - col, w4) * 8 >> s->ss_h;
1416 int h = FFMIN(s->rows - row, h4) * 8 >> s->ss_v, n, o = 0;
1417
1418 for (n = s->ss_h; o < w; n++) {
1419 int bw = 64 >> n;
1420
1421 av_assert2(n <= 4);
1422 if (w & bw) {
1423 s->dsp.mc[n][0][0][0][0](f->data[1] + uvoff + o * bytesperpixel, f->linesize[1],
1424 td->tmp_uv[0] + o * bytesperpixel, 128, h, 0, 0);
1425 s->dsp.mc[n][0][0][0][0](f->data[2] + uvoff + o * bytesperpixel, f->linesize[2],
1426 td->tmp_uv[1] + o * bytesperpixel, 128, h, 0, 0);
1427 o += bw;
1428 }
1429 }
1430 }
1431
1432 // pick filter level and find edges to apply filter to
1433 if (s->s.h.filter.level &&
1434 (lvl = s->s.h.segmentation.feat[b->seg_id].lflvl[b->intra ? 0 : b->ref[0] + 1]
1435 [b->mode[3] != ZEROMV]) > 0) {
1436 int x_end = FFMIN(s->cols - col, w4), y_end = FFMIN(s->rows - row, h4);
1437 int skip_inter = !b->intra && b->skip, col7 = td->col7, row7 = td->row7;
1438
1439 setctx_2d(&lflvl->level[row7 * 8 + col7], w4, h4, 8, lvl);
1440 mask_edges(lflvl->mask[0], 0, 0, row7, col7, x_end, y_end, 0, 0, b->tx, skip_inter);
1441 if (s->ss_h || s->ss_v)
1442 mask_edges(lflvl->mask[1], s->ss_h, s->ss_v, row7, col7, x_end, y_end,
1443 s->cols & 1 && col + w4 >= s->cols ? s->cols & 7 : 0,
1444 s->rows & 1 && row + h4 >= s->rows ? s->rows & 7 : 0,
1445 b->uvtx, skip_inter);
1446 }
1447
1448 if (s->pass == 2) {
1449 s->td[0].b++;
1450 s->td[0].block += w4 * h4 * 64 * bytesperpixel;
1451 s->td[0].uvblock[0] += w4 * h4 * 64 * bytesperpixel >> (s->ss_v + s->ss_h);
1452 s->td[0].uvblock[1] += w4 * h4 * 64 * bytesperpixel >> (s->ss_v + s->ss_h);
1453 s->td[0].eob += 4 * w4 * h4;
1454 s->td[0].uveob[0] += 4 * w4 * h4 >> (s->ss_v + s->ss_h);
1455 s->td[0].uveob[1] += 4 * w4 * h4 >> (s->ss_v + s->ss_h);
1456 }
1457 }
1458