Implement a SIMD version of emulated_edge_mc() for x86.
[libav.git] / libavcodec / vp8.c
CommitLineData
3b636f21
DC
1/**
2 * VP8 compatible video decoder
3 *
4 * Copyright (C) 2010 David Conrad
5 * Copyright (C) 2010 Ronald S. Bultje
13a1304b 6 * Copyright (C) 2010 Jason Garrett-Glaser
3b636f21
DC
7 *
8 * This file is part of FFmpeg.
9 *
10 * FFmpeg is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public
12 * License as published by the Free Software Foundation; either
13 * version 2.1 of the License, or (at your option) any later version.
14 *
15 * FFmpeg is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
19 *
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with FFmpeg; if not, write to the Free Software
22 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23 */
24
6ce9b431 25#include "libavcore/imgutils.h"
3b636f21
DC
26#include "avcodec.h"
27#include "vp56.h"
28#include "vp8data.h"
29#include "vp8dsp.h"
30#include "h264pred.h"
31#include "rectangle.h"
32
33typedef struct {
968570d6
JGG
34 uint8_t filter_level;
35 uint8_t inner_limit;
c55e0d34 36 uint8_t inner_filter;
968570d6
JGG
37} VP8FilterStrength;
38
39typedef struct {
3b636f21
DC
40 uint8_t skip;
41 // todo: make it possible to check for at least (i4x4 or split_mv)
42 // in one op. are others needed?
43 uint8_t mode;
44 uint8_t ref_frame;
45 uint8_t partitioning;
46 VP56mv mv;
47 VP56mv bmv[16];
48} VP8Macroblock;
49
50typedef struct {
51 AVCodecContext *avctx;
52 DSPContext dsp;
53 VP8DSPContext vp8dsp;
54 H264PredContext hpc;
0ef1dbed 55 vp8_mc_func put_pixels_tab[3][3][3];
3b636f21
DC
56 AVFrame frames[4];
57 AVFrame *framep[4];
58 uint8_t *edge_emu_buffer;
59 VP56RangeCoder c; ///< header context, includes mb modes and motion vectors
60 int profile;
61
62 int mb_width; /* number of horizontal MB */
63 int mb_height; /* number of vertical MB */
64 int linesize;
65 int uvlinesize;
66
67 int keyframe;
68 int invisible;
69 int update_last; ///< update VP56_FRAME_PREVIOUS with the current one
70 int update_golden; ///< VP56_FRAME_NONE if not updated, or which frame to copy if so
71 int update_altref;
9ac831c2 72 int deblock_filter;
3b636f21
DC
73
74 /**
75 * If this flag is not set, all the probability updates
76 * are discarded after this frame is decoded.
77 */
78 int update_probabilities;
79
80 /**
81 * All coefficients are contained in separate arith coding contexts.
82 * There can be 1, 2, 4, or 8 of these after the header context.
83 */
84 int num_coeff_partitions;
85 VP56RangeCoder coeff_partition[8];
86
87 VP8Macroblock *macroblocks;
88 VP8Macroblock *macroblocks_base;
968570d6 89 VP8FilterStrength *filter_strength;
3b636f21 90
d2840fa4
PM
91 uint8_t *intra4x4_pred_mode_top;
92 uint8_t intra4x4_pred_mode_left[4];
c55e0d34 93 uint8_t *segmentation_map;
3b636f21
DC
94
95 /**
9ac831c2
DC
96 * Cache of the top row needed for intra prediction
97 * 16 for luma, 8 for each chroma plane
98 */
99 uint8_t (*top_border)[16+8+8];
100
101 /**
3b636f21
DC
102 * For coeff decode, we need to know whether the above block had non-zero
103 * coefficients. This means for each macroblock, we need data for 4 luma
104 * blocks, 2 u blocks, 2 v blocks, and the luma dc block, for a total of 9
105 * per macroblock. We keep the last row in top_nnz.
106 */
107 uint8_t (*top_nnz)[9];
108 DECLARE_ALIGNED(8, uint8_t, left_nnz)[9];
109
110 /**
111 * This is the index plus one of the last non-zero coeff
112 * for each of the blocks in the current macroblock.
113 * So, 0 -> no coeffs
114 * 1 -> dc-only (special transform)
115 * 2+-> full transform
116 */
117 DECLARE_ALIGNED(16, uint8_t, non_zero_count_cache)[6][4];
118 DECLARE_ALIGNED(16, DCTELEM, block)[6][4][16];
827d43bb 119 DECLARE_ALIGNED(16, DCTELEM, block_dc)[16];
d1c58fce 120 uint8_t intra4x4_pred_mode_mb[16];
3b636f21
DC
121
122 int chroma_pred_mode; ///< 8x8c pred mode of the current macroblock
b9a7186b 123 int segment; ///< segment of the current macroblock
3b636f21
DC
124
125 int mbskip_enabled;
126 int sign_bias[4]; ///< one state [0, 1] per ref frame type
c4211046 127 int ref_count[3];
3b636f21
DC
128
129 /**
130 * Base parameters for segmentation, i.e. per-macroblock parameters.
131 * These must be kept unchanged even if segmentation is not used for
132 * a frame, since the values persist between interframes.
133 */
134 struct {
135 int enabled;
136 int absolute_vals;
137 int update_map;
138 int8_t base_quant[4];
139 int8_t filter_level[4]; ///< base loop filter level
140 } segmentation;
141
142 /**
143 * Macroblocks can have one of 4 different quants in a frame when
144 * segmentation is enabled.
145 * If segmentation is disabled, only the first segment's values are used.
146 */
147 struct {
148 // [0] - DC qmul [1] - AC qmul
149 int16_t luma_qmul[2];
150 int16_t luma_dc_qmul[2]; ///< luma dc-only block quant
151 int16_t chroma_qmul[2];
152 } qmat[4];
153
154 struct {
155 int simple;
156 int level;
157 int sharpness;
158 } filter;
159
160 struct {
161 int enabled; ///< whether each mb can have a different strength based on mode/ref
162
163 /**
164 * filter strength adjustment for the following macroblock modes:
165 * [0] - i4x4
166 * [1] - zero mv
167 * [2] - inter modes except for zero or split mv
168 * [3] - split mv
169 * i16x16 modes never have any adjustment
170 */
171 int8_t mode[4];
172
173 /**
174 * filter strength adjustment for macroblocks that reference:
175 * [0] - intra / VP56_FRAME_CURRENT
176 * [1] - VP56_FRAME_PREVIOUS
177 * [2] - VP56_FRAME_GOLDEN
178 * [3] - altref / VP56_FRAME_GOLDEN2
179 */
180 int8_t ref[4];
181 } lf_delta;
182
183 /**
184 * These are all of the updatable probabilities for binary decisions.
185 * They are only implictly reset on keyframes, making it quite likely
186 * for an interframe to desync if a prior frame's header was corrupt
187 * or missing outright!
188 */
189 struct {
190 uint8_t segmentid[3];
a8ab0ccc
PM
191 uint8_t mbskip;
192 uint8_t intra;
193 uint8_t last;
194 uint8_t golden;
3b636f21
DC
195 uint8_t pred16x16[4];
196 uint8_t pred8x8c[3];
370b622a
JGG
197 /* Padded to allow overreads */
198 uint8_t token[4][17][3][NUM_DCT_TOKENS-1];
3b636f21
DC
199 uint8_t mvc[2][19];
200 } prob[2];
201} VP8Context;
202
3b636f21
DC
203static void vp8_decode_flush(AVCodecContext *avctx)
204{
205 VP8Context *s = avctx->priv_data;
206 int i;
207
208 for (i = 0; i < 4; i++)
209 if (s->frames[i].data[0])
210 avctx->release_buffer(avctx, &s->frames[i]);
211 memset(s->framep, 0, sizeof(s->framep));
212
213 av_freep(&s->macroblocks_base);
fca05ea8 214 av_freep(&s->filter_strength);
d2840fa4 215 av_freep(&s->intra4x4_pred_mode_top);
3b636f21
DC
216 av_freep(&s->top_nnz);
217 av_freep(&s->edge_emu_buffer);
9ac831c2 218 av_freep(&s->top_border);
c55e0d34 219 av_freep(&s->segmentation_map);
3b636f21
DC
220
221 s->macroblocks = NULL;
3b636f21
DC
222}
223
224static int update_dimensions(VP8Context *s, int width, int height)
225{
e16f217c 226 if (av_image_check_size(width, height, 0, s->avctx))
3b636f21
DC
227 return AVERROR_INVALIDDATA;
228
229 vp8_decode_flush(s->avctx);
230
231 avcodec_set_dimensions(s->avctx, width, height);
232
233 s->mb_width = (s->avctx->coded_width +15) / 16;
234 s->mb_height = (s->avctx->coded_height+15) / 16;
235
aa93c52c
PM
236 s->macroblocks_base = av_mallocz((s->mb_width+s->mb_height*2+1)*sizeof(*s->macroblocks));
237 s->filter_strength = av_mallocz(s->mb_width*sizeof(*s->filter_strength));
ccf13f9e 238 s->intra4x4_pred_mode_top = av_mallocz(s->mb_width*4);
3b636f21 239 s->top_nnz = av_mallocz(s->mb_width*sizeof(*s->top_nnz));
9ac831c2 240 s->top_border = av_mallocz((s->mb_width+1)*sizeof(*s->top_border));
aa93c52c 241 s->segmentation_map = av_mallocz(s->mb_width*s->mb_height);
3b636f21 242
d2840fa4 243 if (!s->macroblocks_base || !s->filter_strength || !s->intra4x4_pred_mode_top ||
c55e0d34 244 !s->top_nnz || !s->top_border || !s->segmentation_map)
b6c420ce
DC
245 return AVERROR(ENOMEM);
246
c55e0d34 247 s->macroblocks = s->macroblocks_base + 1;
3b636f21
DC
248
249 return 0;
250}
251
252static void parse_segment_info(VP8Context *s)
253{
254 VP56RangeCoder *c = &s->c;
255 int i;
256
257 s->segmentation.update_map = vp8_rac_get(c);
258
259 if (vp8_rac_get(c)) { // update segment feature data
260 s->segmentation.absolute_vals = vp8_rac_get(c);
261
262 for (i = 0; i < 4; i++)
263 s->segmentation.base_quant[i] = vp8_rac_get_sint(c, 7);
264
265 for (i = 0; i < 4; i++)
266 s->segmentation.filter_level[i] = vp8_rac_get_sint(c, 6);
267 }
268 if (s->segmentation.update_map)
269 for (i = 0; i < 3; i++)
270 s->prob->segmentid[i] = vp8_rac_get(c) ? vp8_rac_get_uint(c, 8) : 255;
271}
272
273static void update_lf_deltas(VP8Context *s)
274{
275 VP56RangeCoder *c = &s->c;
276 int i;
277
278 for (i = 0; i < 4; i++)
279 s->lf_delta.ref[i] = vp8_rac_get_sint(c, 6);
280
281 for (i = 0; i < 4; i++)
282 s->lf_delta.mode[i] = vp8_rac_get_sint(c, 6);
283}
284
285static int setup_partitions(VP8Context *s, const uint8_t *buf, int buf_size)
286{
287 const uint8_t *sizes = buf;
288 int i;
289
290 s->num_coeff_partitions = 1 << vp8_rac_get_uint(&s->c, 2);
291
292 buf += 3*(s->num_coeff_partitions-1);
293 buf_size -= 3*(s->num_coeff_partitions-1);
294 if (buf_size < 0)
295 return -1;
296
297 for (i = 0; i < s->num_coeff_partitions-1; i++) {
06d50ca8 298 int size = AV_RL24(sizes + 3*i);
3b636f21
DC
299 if (buf_size - size < 0)
300 return -1;
301
905ef0d0 302 ff_vp56_init_range_decoder(&s->coeff_partition[i], buf, size);
3b636f21
DC
303 buf += size;
304 buf_size -= size;
305 }
905ef0d0 306 ff_vp56_init_range_decoder(&s->coeff_partition[i], buf, buf_size);
3b636f21
DC
307
308 return 0;
309}
310
311static void get_quants(VP8Context *s)
312{
313 VP56RangeCoder *c = &s->c;
314 int i, base_qi;
315
316 int yac_qi = vp8_rac_get_uint(c, 7);
317 int ydc_delta = vp8_rac_get_sint(c, 4);
318 int y2dc_delta = vp8_rac_get_sint(c, 4);
319 int y2ac_delta = vp8_rac_get_sint(c, 4);
320 int uvdc_delta = vp8_rac_get_sint(c, 4);
321 int uvac_delta = vp8_rac_get_sint(c, 4);
322
323 for (i = 0; i < 4; i++) {
324 if (s->segmentation.enabled) {
325 base_qi = s->segmentation.base_quant[i];
326 if (!s->segmentation.absolute_vals)
327 base_qi += yac_qi;
328 } else
329 base_qi = yac_qi;
330
331 s->qmat[i].luma_qmul[0] = vp8_dc_qlookup[av_clip(base_qi + ydc_delta , 0, 127)];
332 s->qmat[i].luma_qmul[1] = vp8_ac_qlookup[av_clip(base_qi , 0, 127)];
333 s->qmat[i].luma_dc_qmul[0] = 2 * vp8_dc_qlookup[av_clip(base_qi + y2dc_delta, 0, 127)];
a8ab0ccc
PM
334 s->qmat[i].luma_dc_qmul[1] = 155 * vp8_ac_qlookup[av_clip(base_qi + y2ac_delta, 0, 127)] / 100;
335 s->qmat[i].chroma_qmul[0] = vp8_dc_qlookup[av_clip(base_qi + uvdc_delta, 0, 127)];
3b636f21 336 s->qmat[i].chroma_qmul[1] = vp8_ac_qlookup[av_clip(base_qi + uvac_delta, 0, 127)];
a8ab0ccc
PM
337
338 s->qmat[i].luma_dc_qmul[1] = FFMAX(s->qmat[i].luma_dc_qmul[1], 8);
339 s->qmat[i].chroma_qmul[0] = FFMIN(s->qmat[i].chroma_qmul[0], 132);
3b636f21
DC
340 }
341}
342
343/**
344 * Determine which buffers golden and altref should be updated with after this frame.
345 * The spec isn't clear here, so I'm going by my understanding of what libvpx does
346 *
347 * Intra frames update all 3 references
348 * Inter frames update VP56_FRAME_PREVIOUS if the update_last flag is set
349 * If the update (golden|altref) flag is set, it's updated with the current frame
350 * if update_last is set, and VP56_FRAME_PREVIOUS otherwise.
351 * If the flag is not set, the number read means:
352 * 0: no update
353 * 1: VP56_FRAME_PREVIOUS
354 * 2: update golden with altref, or update altref with golden
355 */
356static VP56Frame ref_to_update(VP8Context *s, int update, VP56Frame ref)
357{
358 VP56RangeCoder *c = &s->c;
359
360 if (update)
361 return VP56_FRAME_CURRENT;
362
363 switch (vp8_rac_get_uint(c, 2)) {
364 case 1:
365 return VP56_FRAME_PREVIOUS;
366 case 2:
367 return (ref == VP56_FRAME_GOLDEN) ? VP56_FRAME_GOLDEN2 : VP56_FRAME_GOLDEN;
368 }
369 return VP56_FRAME_NONE;
370}
371
372static void update_refs(VP8Context *s)
373{
374 VP56RangeCoder *c = &s->c;
375
376 int update_golden = vp8_rac_get(c);
377 int update_altref = vp8_rac_get(c);
378
379 s->update_golden = ref_to_update(s, update_golden, VP56_FRAME_GOLDEN);
380 s->update_altref = ref_to_update(s, update_altref, VP56_FRAME_GOLDEN2);
381}
382
383static int decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size)
384{
385 VP56RangeCoder *c = &s->c;
370b622a 386 int header_size, hscale, vscale, i, j, k, l, m, ret;
3b636f21
DC
387 int width = s->avctx->width;
388 int height = s->avctx->height;
389
390 s->keyframe = !(buf[0] & 1);
391 s->profile = (buf[0]>>1) & 7;
392 s->invisible = !(buf[0] & 0x10);
06d50ca8 393 header_size = AV_RL24(buf) >> 5;
3b636f21
DC
394 buf += 3;
395 buf_size -= 3;
396
0ef1dbed
DC
397 if (s->profile > 3)
398 av_log(s->avctx, AV_LOG_WARNING, "Unknown profile %d\n", s->profile);
399
400 if (!s->profile)
401 memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_epel_pixels_tab, sizeof(s->put_pixels_tab));
402 else // profile 1-3 use bilinear, 4+ aren't defined so whatever
403 memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_bilinear_pixels_tab, sizeof(s->put_pixels_tab));
3b636f21
DC
404
405 if (header_size > buf_size - 7*s->keyframe) {
406 av_log(s->avctx, AV_LOG_ERROR, "Header size larger than data provided\n");
407 return AVERROR_INVALIDDATA;
408 }
409
410 if (s->keyframe) {
06d50ca8
JGG
411 if (AV_RL24(buf) != 0x2a019d) {
412 av_log(s->avctx, AV_LOG_ERROR, "Invalid start code 0x%x\n", AV_RL24(buf));
3b636f21
DC
413 return AVERROR_INVALIDDATA;
414 }
415 width = AV_RL16(buf+3) & 0x3fff;
416 height = AV_RL16(buf+5) & 0x3fff;
417 hscale = buf[4] >> 6;
418 vscale = buf[6] >> 6;
419 buf += 7;
420 buf_size -= 7;
421
92a54426
MR
422 if (hscale || vscale)
423 av_log_missing_feature(s->avctx, "Upscaling", 1);
424
3b636f21 425 s->update_golden = s->update_altref = VP56_FRAME_CURRENT;
370b622a
JGG
426 for (i = 0; i < 4; i++)
427 for (j = 0; j < 16; j++)
428 memcpy(s->prob->token[i][j], vp8_token_default_probs[i][vp8_coeff_band[j]],
429 sizeof(s->prob->token[i][j]));
3b636f21
DC
430 memcpy(s->prob->pred16x16, vp8_pred16x16_prob_inter, sizeof(s->prob->pred16x16));
431 memcpy(s->prob->pred8x8c , vp8_pred8x8c_prob_inter , sizeof(s->prob->pred8x8c));
432 memcpy(s->prob->mvc , vp8_mv_default_prob , sizeof(s->prob->mvc));
433 memset(&s->segmentation, 0, sizeof(s->segmentation));
434 }
435
436 if (!s->macroblocks_base || /* first frame */
437 width != s->avctx->width || height != s->avctx->height) {
438 if ((ret = update_dimensions(s, width, height) < 0))
439 return ret;
440 }
441
905ef0d0 442 ff_vp56_init_range_decoder(c, buf, header_size);
3b636f21
DC
443 buf += header_size;
444 buf_size -= header_size;
445
446 if (s->keyframe) {
447 if (vp8_rac_get(c))
448 av_log(s->avctx, AV_LOG_WARNING, "Unspecified colorspace\n");
449 vp8_rac_get(c); // whether we can skip clamping in dsp functions
450 }
451
452 if ((s->segmentation.enabled = vp8_rac_get(c)))
453 parse_segment_info(s);
454 else
455 s->segmentation.update_map = 0; // FIXME: move this to some init function?
456
457 s->filter.simple = vp8_rac_get(c);
458 s->filter.level = vp8_rac_get_uint(c, 6);
459 s->filter.sharpness = vp8_rac_get_uint(c, 3);
460
461 if ((s->lf_delta.enabled = vp8_rac_get(c)))
462 if (vp8_rac_get(c))
463 update_lf_deltas(s);
464
465 if (setup_partitions(s, buf, buf_size)) {
466 av_log(s->avctx, AV_LOG_ERROR, "Invalid partitions\n");
467 return AVERROR_INVALIDDATA;
468 }
469
470 get_quants(s);
471
472 if (!s->keyframe) {
473 update_refs(s);
474 s->sign_bias[VP56_FRAME_GOLDEN] = vp8_rac_get(c);
475 s->sign_bias[VP56_FRAME_GOLDEN2 /* altref */] = vp8_rac_get(c);
476 }
477
478 // if we aren't saving this frame's probabilities for future frames,
479 // make a copy of the current probabilities
480 if (!(s->update_probabilities = vp8_rac_get(c)))
481 s->prob[1] = s->prob[0];
482
483 s->update_last = s->keyframe || vp8_rac_get(c);
484
485 for (i = 0; i < 4; i++)
486 for (j = 0; j < 8; j++)
487 for (k = 0; k < 3; k++)
488 for (l = 0; l < NUM_DCT_TOKENS-1; l++)
370b622a
JGG
489 if (vp56_rac_get_prob_branchy(c, vp8_token_update_probs[i][j][k][l])) {
490 int prob = vp8_rac_get_uint(c, 8);
b0d58795
JGG
491 for (m = 0; vp8_coeff_band_indexes[j][m] >= 0; m++)
492 s->prob->token[i][vp8_coeff_band_indexes[j][m]][k][l] = prob;
370b622a 493 }
3b636f21
DC
494
495 if ((s->mbskip_enabled = vp8_rac_get(c)))
a8ab0ccc 496 s->prob->mbskip = vp8_rac_get_uint(c, 8);
3b636f21
DC
497
498 if (!s->keyframe) {
a8ab0ccc
PM
499 s->prob->intra = vp8_rac_get_uint(c, 8);
500 s->prob->last = vp8_rac_get_uint(c, 8);
501 s->prob->golden = vp8_rac_get_uint(c, 8);
3b636f21
DC
502
503 if (vp8_rac_get(c))
504 for (i = 0; i < 4; i++)
505 s->prob->pred16x16[i] = vp8_rac_get_uint(c, 8);
506 if (vp8_rac_get(c))
507 for (i = 0; i < 3; i++)
508 s->prob->pred8x8c[i] = vp8_rac_get_uint(c, 8);
509
510 // 17.2 MV probability update
511 for (i = 0; i < 2; i++)
512 for (j = 0; j < 19; j++)
7697cdcf 513 if (vp56_rac_get_prob_branchy(c, vp8_mv_update_prob[i][j]))
3b636f21
DC
514 s->prob->mvc[i][j] = vp8_rac_get_nn(c);
515 }
516
517 return 0;
518}
519
414ac27d
JGG
520static av_always_inline
521void clamp_mv(VP8Context *s, VP56mv *dst, const VP56mv *src, int mb_x, int mb_y)
3b636f21
DC
522{
523#define MARGIN (16 << 2)
524 dst->x = av_clip(src->x, -((mb_x << 6) + MARGIN),
525 ((s->mb_width - 1 - mb_x) << 6) + MARGIN);
526 dst->y = av_clip(src->y, -((mb_y << 6) + MARGIN),
527 ((s->mb_height - 1 - mb_y) << 6) + MARGIN);
528}
529
414ac27d 530static av_always_inline
aa93c52c 531void find_near_mvs(VP8Context *s, VP8Macroblock *mb,
414ac27d 532 VP56mv near[2], VP56mv *best, uint8_t cnt[4])
3b636f21 533{
c55e0d34
JGG
534 VP8Macroblock *mb_edge[3] = { mb + 2 /* top */,
535 mb - 1 /* left */,
536 mb + 1 /* top-left */ };
3b636f21
DC
537 enum { EDGE_TOP, EDGE_LEFT, EDGE_TOPLEFT };
538 VP56mv near_mv[4] = {{ 0 }};
539 enum { CNT_ZERO, CNT_NEAREST, CNT_NEAR, CNT_SPLITMV };
702e8d33 540 int idx = CNT_ZERO;
3b636f21 541 int best_idx = CNT_ZERO;
702e8d33
JGG
542 int cur_sign_bias = s->sign_bias[mb->ref_frame];
543 int *sign_bias = s->sign_bias;
3b636f21
DC
544
545 /* Process MB on top, left and top-left */
702e8d33
JGG
546 #define MV_EDGE_CHECK(n)\
547 {\
548 VP8Macroblock *edge = mb_edge[n];\
549 int edge_ref = edge->ref_frame;\
550 if (edge_ref != VP56_FRAME_CURRENT) {\
551 uint32_t mv = AV_RN32A(&edge->mv);\
552 if (mv) {\
553 if (cur_sign_bias != sign_bias[edge_ref]) {\
554 /* SWAR negate of the values in mv. */\
0087aa47
JGG
555 mv = ~mv;\
556 mv = ((mv&0x7fff7fff) + 0x00010001) ^ (mv&0x80008000);\
702e8d33
JGG
557 }\
558 if (!n || mv != AV_RN32A(&near_mv[idx]))\
559 AV_WN32A(&near_mv[++idx], mv);\
560 cnt[idx] += 1 + (n != 2);\
561 } else\
562 cnt[CNT_ZERO] += 1 + (n != 2);\
563 }\
3b636f21 564 }
702e8d33
JGG
565 MV_EDGE_CHECK(0)
566 MV_EDGE_CHECK(1)
567 MV_EDGE_CHECK(2)
3b636f21 568
702e8d33
JGG
569 /* If we have three distinct MVs, merge first and last if they're the same */
570 if (cnt[CNT_SPLITMV] && AV_RN32A(&near_mv[1+EDGE_TOP]) == AV_RN32A(&near_mv[1+EDGE_TOPLEFT]))
3b636f21
DC
571 cnt[CNT_NEAREST] += 1;
572
573 cnt[CNT_SPLITMV] = ((mb_edge[EDGE_LEFT]->mode == VP8_MVMODE_SPLIT) +
574 (mb_edge[EDGE_TOP]->mode == VP8_MVMODE_SPLIT)) * 2 +
575 (mb_edge[EDGE_TOPLEFT]->mode == VP8_MVMODE_SPLIT);
576
577 /* Swap near and nearest if necessary */
578 if (cnt[CNT_NEAR] > cnt[CNT_NEAREST]) {
702e8d33
JGG
579 FFSWAP(uint8_t, cnt[CNT_NEAREST], cnt[CNT_NEAR]);
580 FFSWAP( VP56mv, near_mv[CNT_NEAREST], near_mv[CNT_NEAR]);
3b636f21
DC
581 }
582
583 /* Choose the best mv out of 0,0 and the nearest mv */
584 if (cnt[CNT_NEAREST] >= cnt[CNT_ZERO])
585 best_idx = CNT_NEAREST;
586
9fddd14a 587 mb->mv = near_mv[best_idx];
3b636f21
DC
588 near[0] = near_mv[CNT_NEAREST];
589 near[1] = near_mv[CNT_NEAR];
590}
591
592/**
593 * Motion vector coding, 17.1.
594 */
595static int read_mv_component(VP56RangeCoder *c, const uint8_t *p)
596{
ca18a478 597 int bit, x = 0;
3b636f21 598
7697cdcf 599 if (vp56_rac_get_prob_branchy(c, p[0])) {
3b636f21
DC
600 int i;
601
602 for (i = 0; i < 3; i++)
603 x += vp56_rac_get_prob(c, p[9 + i]) << i;
604 for (i = 9; i > 3; i--)
605 x += vp56_rac_get_prob(c, p[9 + i]) << i;
606 if (!(x & 0xFFF0) || vp56_rac_get_prob(c, p[12]))
607 x += 8;
ca18a478
DC
608 } else {
609 // small_mvtree
610 const uint8_t *ps = p+2;
611 bit = vp56_rac_get_prob(c, *ps);
612 ps += 1 + 3*bit;
613 x += 4*bit;
614 bit = vp56_rac_get_prob(c, *ps);
615 ps += 1 + bit;
616 x += 2*bit;
617 x += vp56_rac_get_prob(c, *ps);
618 }
3b636f21
DC
619
620 return (x && vp56_rac_get_prob(c, p[1])) ? -x : x;
621}
622
414ac27d
JGG
623static av_always_inline
624const uint8_t *get_submv_prob(uint32_t left, uint32_t top)
3b636f21 625{
7bf254c4
JGG
626 if (left == top)
627 return vp8_submv_prob[4-!!left];
628 if (!top)
3b636f21 629 return vp8_submv_prob[2];
7bf254c4 630 return vp8_submv_prob[1-!!left];
3b636f21
DC
631}
632
633/**
634 * Split motion vector prediction, 16.4.
7ed06b2b 635 * @returns the number of motion vectors parsed (2, 4 or 16)
3b636f21 636 */
414ac27d
JGG
637static av_always_inline
638int decode_splitmvs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb)
3b636f21 639{
0908f1b9
JGG
640 int part_idx;
641 int n, num;
c55e0d34 642 VP8Macroblock *top_mb = &mb[2];
7bf254c4
JGG
643 VP8Macroblock *left_mb = &mb[-1];
644 const uint8_t *mbsplits_left = vp8_mbsplits[left_mb->partitioning],
645 *mbsplits_top = vp8_mbsplits[top_mb->partitioning],
0908f1b9 646 *mbsplits_cur, *firstidx;
c55e0d34
JGG
647 VP56mv *top_mv = top_mb->bmv;
648 VP56mv *left_mv = left_mb->bmv;
649 VP56mv *cur_mv = mb->bmv;
3b636f21 650
0908f1b9
JGG
651 if (vp56_rac_get_prob_branchy(c, vp8_mbsplit_prob[0])) {
652 if (vp56_rac_get_prob_branchy(c, vp8_mbsplit_prob[1])) {
653 part_idx = VP8_SPLITMVMODE_16x8 + vp56_rac_get_prob(c, vp8_mbsplit_prob[2]);
654 } else {
655 part_idx = VP8_SPLITMVMODE_8x8;
656 }
657 } else {
658 part_idx = VP8_SPLITMVMODE_4x4;
659 }
660
661 num = vp8_mbsplit_count[part_idx];
662 mbsplits_cur = vp8_mbsplits[part_idx],
663 firstidx = vp8_mbfirstidx[part_idx];
664 mb->partitioning = part_idx;
665
3b636f21 666 for (n = 0; n < num; n++) {
7ed06b2b 667 int k = firstidx[n];
7bf254c4 668 uint32_t left, above;
7ed06b2b
RB
669 const uint8_t *submv_prob;
670
7bf254c4
JGG
671 if (!(k & 3))
672 left = AV_RN32A(&left_mv[mbsplits_left[k + 3]]);
673 else
674 left = AV_RN32A(&cur_mv[mbsplits_cur[k - 1]]);
675 if (k <= 3)
676 above = AV_RN32A(&top_mv[mbsplits_top[k + 12]]);
677 else
678 above = AV_RN32A(&cur_mv[mbsplits_cur[k - 4]]);
7ed06b2b
RB
679
680 submv_prob = get_submv_prob(left, above);
3b636f21 681
c5dec7f1
JGG
682 if (vp56_rac_get_prob_branchy(c, submv_prob[0])) {
683 if (vp56_rac_get_prob_branchy(c, submv_prob[1])) {
684 if (vp56_rac_get_prob_branchy(c, submv_prob[2])) {
685 mb->bmv[n].y = mb->mv.y + read_mv_component(c, s->prob->mvc[0]);
686 mb->bmv[n].x = mb->mv.x + read_mv_component(c, s->prob->mvc[1]);
687 } else {
688 AV_ZERO32(&mb->bmv[n]);
689 }
690 } else {
691 AV_WN32A(&mb->bmv[n], above);
692 }
693 } else {
7bf254c4 694 AV_WN32A(&mb->bmv[n], left);
3b636f21 695 }
3b636f21 696 }
7ed06b2b
RB
697
698 return num;
3b636f21
DC
699}
700
414ac27d 701static av_always_inline
d2840fa4
PM
702void decode_intra4x4_modes(VP8Context *s, VP56RangeCoder *c,
703 int mb_x, int keyframe)
3b636f21 704{
d2840fa4 705 uint8_t *intra4x4 = s->intra4x4_pred_mode_mb;
d1c58fce 706 if (keyframe) {
d2840fa4
PM
707 int x, y;
708 uint8_t* const top = s->intra4x4_pred_mode_top + 4 * mb_x;
709 uint8_t* const left = s->intra4x4_pred_mode_left;
d1c58fce
JGG
710 for (y = 0; y < 4; y++) {
711 for (x = 0; x < 4; x++) {
d2840fa4
PM
712 const uint8_t *ctx;
713 ctx = vp8_pred4x4_prob_intra[top[x]][left[y]];
714 *intra4x4 = vp8_rac_get_tree(c, vp8_pred4x4_tree, ctx);
715 left[y] = top[x] = *intra4x4;
716 intra4x4++;
3b636f21 717 }
3b636f21 718 }
d1c58fce 719 } else {
d2840fa4 720 int i;
d1c58fce
JGG
721 for (i = 0; i < 16; i++)
722 intra4x4[i] = vp8_rac_get_tree(c, vp8_pred4x4_tree, vp8_pred4x4_prob_inter);
3b636f21
DC
723 }
724}
725
414ac27d 726static av_always_inline
d2840fa4 727void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, uint8_t *segment)
3b636f21
DC
728{
729 VP56RangeCoder *c = &s->c;
3b636f21
DC
730
731 if (s->segmentation.update_map)
c55e0d34 732 *segment = vp8_rac_get_tree(c, vp8_segmentid_tree, s->prob->segmentid);
b9a7186b 733 s->segment = *segment;
3b636f21 734
a8ab0ccc 735 mb->skip = s->mbskip_enabled ? vp56_rac_get_prob(c, s->prob->mbskip) : 0;
3b636f21
DC
736
737 if (s->keyframe) {
738 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_intra, vp8_pred16x16_prob_intra);
739
740 if (mb->mode == MODE_I4x4) {
d2840fa4
PM
741 decode_intra4x4_modes(s, c, mb_x, 1);
742 } else {
743 const uint32_t modes = vp8_pred4x4_mode[mb->mode] * 0x01010101u;
744 AV_WN32A(s->intra4x4_pred_mode_top + 4 * mb_x, modes);
745 AV_WN32A(s->intra4x4_pred_mode_left, modes);
746 }
3b636f21
DC
747
748 s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, vp8_pred8x8c_prob_intra);
749 mb->ref_frame = VP56_FRAME_CURRENT;
a8ab0ccc 750 } else if (vp56_rac_get_prob_branchy(c, s->prob->intra)) {
3b636f21 751 VP56mv near[2], best;
702e8d33 752 uint8_t cnt[4] = { 0 };
3b636f21
DC
753
754 // inter MB, 16.2
a8ab0ccc
PM
755 if (vp56_rac_get_prob_branchy(c, s->prob->last))
756 mb->ref_frame = vp56_rac_get_prob(c, s->prob->golden) ?
3b636f21
DC
757 VP56_FRAME_GOLDEN2 /* altref */ : VP56_FRAME_GOLDEN;
758 else
759 mb->ref_frame = VP56_FRAME_PREVIOUS;
c4211046 760 s->ref_count[mb->ref_frame-1]++;
3b636f21
DC
761
762 // motion vectors, 16.3
aa93c52c 763 find_near_mvs(s, mb, near, &best, cnt);
23117d69
JGG
764 if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[0]][0])) {
765 if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[1]][1])) {
766 if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[2]][2])) {
767 if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[3]][3])) {
768 mb->mode = VP8_MVMODE_SPLIT;
769 clamp_mv(s, &mb->mv, &mb->mv, mb_x, mb_y);
770 mb->mv = mb->bmv[decode_splitmvs(s, c, mb) - 1];
771 } else {
772 mb->mode = VP8_MVMODE_NEW;
773 clamp_mv(s, &mb->mv, &mb->mv, mb_x, mb_y);
2b476e02
JGG
774 mb->mv.y += read_mv_component(c, s->prob->mvc[0]);
775 mb->mv.x += read_mv_component(c, s->prob->mvc[1]);
23117d69
JGG
776 }
777 } else {
778 mb->mode = VP8_MVMODE_NEAR;
779 clamp_mv(s, &mb->mv, &near[1], mb_x, mb_y);
780 }
781 } else {
782 mb->mode = VP8_MVMODE_NEAREST;
783 clamp_mv(s, &mb->mv, &near[0], mb_x, mb_y);
784 }
785 } else {
786 mb->mode = VP8_MVMODE_ZERO;
14767f35 787 AV_ZERO32(&mb->mv);
3b636f21
DC
788 }
789 if (mb->mode != VP8_MVMODE_SPLIT) {
7ed06b2b
RB
790 mb->partitioning = VP8_SPLITMVMODE_NONE;
791 mb->bmv[0] = mb->mv;
3b636f21
DC
792 }
793 } else {
794 // intra MB, 16.1
795 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_inter, s->prob->pred16x16);
796
158e062c 797 if (mb->mode == MODE_I4x4)
d2840fa4 798 decode_intra4x4_modes(s, c, mb_x, 0);
3b636f21
DC
799
800 s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, s->prob->pred8x8c);
801 mb->ref_frame = VP56_FRAME_CURRENT;
b946111f 802 mb->partitioning = VP8_SPLITMVMODE_NONE;
14767f35 803 AV_ZERO32(&mb->bmv[0]);
3b636f21
DC
804 }
805}
806
807/**
e394953e
RB
808 * @param c arithmetic bitstream reader context
809 * @param block destination for block coefficients
810 * @param probs probabilities to use when reading trees from the bitstream
3b636f21
DC
811 * @param i initial coeff index, 0 unless a separate DC block is coded
812 * @param zero_nhood the initial prediction context for number of surrounding
813 * all-zero blocks (only left/top, so 0-2)
3fa76268 814 * @param qmul array holding the dc/ac dequant factor at position 0/1
3b636f21
DC
815 * @return 0 if no coeffs were decoded
816 * otherwise, the index of the last coeff decoded plus one
817 */
1e739679
JGG
818static int decode_block_coeffs_internal(VP56RangeCoder *c, DCTELEM block[16],
819 uint8_t probs[8][3][NUM_DCT_TOKENS-1],
820 int i, uint8_t *token_prob, int16_t qmul[2])
3b636f21 821{
afb54a85 822 goto skip_eob;
fe1b5d97 823 do {
1e739679 824 int coeff;
fe1b5d97 825 if (!vp56_rac_get_prob_branchy(c, token_prob[0])) // DCT_EOB
afb54a85 826 return i;
3b636f21 827
fe1b5d97
DC
828skip_eob:
829 if (!vp56_rac_get_prob_branchy(c, token_prob[1])) { // DCT_0
c22b4468 830 if (++i == 16)
afb54a85 831 return i; // invalid input; blocks should end with EOB
370b622a 832 token_prob = probs[i][0];
c22b4468 833 goto skip_eob;
fe1b5d97
DC
834 }
835
836 if (!vp56_rac_get_prob_branchy(c, token_prob[2])) { // DCT_1
837 coeff = 1;
370b622a 838 token_prob = probs[i+1][1];
fe1b5d97 839 } else {
fe1b5d97 840 if (!vp56_rac_get_prob_branchy(c, token_prob[3])) { // DCT 2,3,4
476be414 841 coeff = vp56_rac_get_prob_branchy(c, token_prob[4]);
fe1b5d97
DC
842 if (coeff)
843 coeff += vp56_rac_get_prob(c, token_prob[5]);
844 coeff += 2;
845 } else {
846 // DCT_CAT*
847 if (!vp56_rac_get_prob_branchy(c, token_prob[6])) {
848 if (!vp56_rac_get_prob_branchy(c, token_prob[7])) { // DCT_CAT1
849 coeff = 5 + vp56_rac_get_prob(c, vp8_dct_cat1_prob[0]);
850 } else { // DCT_CAT2
851 coeff = 7;
852 coeff += vp56_rac_get_prob(c, vp8_dct_cat2_prob[0]) << 1;
853 coeff += vp56_rac_get_prob(c, vp8_dct_cat2_prob[1]);
854 }
855 } else { // DCT_CAT3 and up
856 int a = vp56_rac_get_prob(c, token_prob[8]);
857 int b = vp56_rac_get_prob(c, token_prob[9+a]);
858 int cat = (a<<1) + b;
859 coeff = 3 + (8<<cat);
860 coeff += vp8_rac_get_coeff(c, vp8_dct_cat_prob[cat]);
861 }
862 }
370b622a 863 token_prob = probs[i+1][2];
fe1b5d97 864 }
fe1b5d97 865 block[zigzag_scan[i]] = (vp8_rac_get(c) ? -coeff : coeff) * qmul[!!i];
afb54a85 866 } while (++i < 16);
fe1b5d97 867
afb54a85 868 return i;
3b636f21
DC
869}
870
414ac27d 871static av_always_inline
1e739679
JGG
872int decode_block_coeffs(VP56RangeCoder *c, DCTELEM block[16],
873 uint8_t probs[8][3][NUM_DCT_TOKENS-1],
874 int i, int zero_nhood, int16_t qmul[2])
875{
876 uint8_t *token_prob = probs[i][zero_nhood];
877 if (!vp56_rac_get_prob_branchy(c, token_prob[0])) // DCT_EOB
878 return 0;
879 return decode_block_coeffs_internal(c, block, probs, i, token_prob, qmul);
880}
881
882static av_always_inline
414ac27d
JGG
883void decode_mb_coeffs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb,
884 uint8_t t_nnz[9], uint8_t l_nnz[9])
3b636f21 885{
3b636f21
DC
886 int i, x, y, luma_start = 0, luma_ctx = 3;
887 int nnz_pred, nnz, nnz_total = 0;
b9a7186b 888 int segment = s->segment;
f311208c 889 int block_dc = 0;
3b636f21 890
3b636f21 891 if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
3b636f21
DC
892 nnz_pred = t_nnz[8] + l_nnz[8];
893
894 // decode DC values and do hadamard
827d43bb 895 nnz = decode_block_coeffs(c, s->block_dc, s->prob->token[1], 0, nnz_pred,
3b636f21
DC
896 s->qmat[segment].luma_dc_qmul);
897 l_nnz[8] = t_nnz[8] = !!nnz;
f311208c
JGG
898 if (nnz) {
899 nnz_total += nnz;
900 block_dc = 1;
901 if (nnz == 1)
902 s->vp8dsp.vp8_luma_dc_wht_dc(s->block, s->block_dc);
903 else
904 s->vp8dsp.vp8_luma_dc_wht(s->block, s->block_dc);
905 }
3b636f21
DC
906 luma_start = 1;
907 luma_ctx = 0;
908 }
909
910 // luma blocks
911 for (y = 0; y < 4; y++)
912 for (x = 0; x < 4; x++) {
ffbf0794 913 nnz_pred = l_nnz[y] + t_nnz[x];
3b636f21 914 nnz = decode_block_coeffs(c, s->block[y][x], s->prob->token[luma_ctx], luma_start,
ffbf0794 915 nnz_pred, s->qmat[segment].luma_qmul);
f311208c
JGG
916 // nnz+block_dc may be one more than the actual last index, but we don't care
917 s->non_zero_count_cache[y][x] = nnz + block_dc;
3b636f21
DC
918 t_nnz[x] = l_nnz[y] = !!nnz;
919 nnz_total += nnz;
920 }
921
922 // chroma blocks
923 // TODO: what to do about dimensions? 2nd dim for luma is x,
924 // but for chroma it's (y<<1)|x
925 for (i = 4; i < 6; i++)
926 for (y = 0; y < 2; y++)
927 for (x = 0; x < 2; x++) {
928 nnz_pred = l_nnz[i+2*y] + t_nnz[i+2*x];
929 nnz = decode_block_coeffs(c, s->block[i][(y<<1)+x], s->prob->token[2], 0,
930 nnz_pred, s->qmat[segment].chroma_qmul);
931 s->non_zero_count_cache[i][(y<<1)+x] = nnz;
932 t_nnz[i+2*x] = l_nnz[i+2*y] = !!nnz;
933 nnz_total += nnz;
934 }
935
936 // if there were no coded coeffs despite the macroblock not being marked skip,
937 // we MUST not do the inner loop filter and should not do IDCT
938 // Since skip isn't used for bitstream prediction, just manually set it.
939 if (!nnz_total)
940 mb->skip = 1;
941}
942
9ac831c2
DC
943static av_always_inline
944void backup_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,
945 int linesize, int uvlinesize, int simple)
946{
947 AV_COPY128(top_border, src_y + 15*linesize);
948 if (!simple) {
949 AV_COPY64(top_border+16, src_cb + 7*uvlinesize);
950 AV_COPY64(top_border+24, src_cr + 7*uvlinesize);
951 }
952}
953
954static av_always_inline
955void xchg_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,
956 int linesize, int uvlinesize, int mb_x, int mb_y, int mb_width,
957 int simple, int xchg)
958{
959 uint8_t *top_border_m1 = top_border-32; // for TL prediction
960 src_y -= linesize;
961 src_cb -= uvlinesize;
962 src_cr -= uvlinesize;
963
096971e8
MR
964#define XCHG(a,b,xchg) do { \
965 if (xchg) AV_SWAP64(b,a); \
966 else AV_COPY64(b,a); \
967 } while (0)
9ac831c2
DC
968
969 XCHG(top_border_m1+8, src_y-8, xchg);
970 XCHG(top_border, src_y, xchg);
971 XCHG(top_border+8, src_y+8, 1);
070ce7ef 972 if (mb_x < mb_width-1)
9ac831c2 973 XCHG(top_border+32, src_y+16, 1);
070ce7ef 974
9ac831c2
DC
975 // only copy chroma for normal loop filter
976 // or to initialize the top row to 127
977 if (!simple || !mb_y) {
978 XCHG(top_border_m1+16, src_cb-8, xchg);
979 XCHG(top_border_m1+24, src_cr-8, xchg);
980 XCHG(top_border+16, src_cb, 1);
981 XCHG(top_border+24, src_cr, 1);
982 }
983}
984
414ac27d 985static av_always_inline
ee555de7
RB
986int check_dc_pred8x8_mode(int mode, int mb_x, int mb_y)
987{
988 if (!mb_x) {
989 return mb_y ? TOP_DC_PRED8x8 : DC_128_PRED8x8;
990 } else {
991 return mb_y ? mode : LEFT_DC_PRED8x8;
992 }
993}
994
995static av_always_inline
996int check_tm_pred8x8_mode(int mode, int mb_x, int mb_y)
997{
998 if (!mb_x) {
999 return mb_y ? VERT_PRED8x8 : DC_129_PRED8x8;
1000 } else {
1001 return mb_y ? mode : HOR_PRED8x8;
1002 }
1003}
1004
1005static av_always_inline
1006int check_intra_pred8x8_mode(int mode, int mb_x, int mb_y)
3b636f21
DC
1007{
1008 if (mode == DC_PRED8x8) {
ee555de7
RB
1009 return check_dc_pred8x8_mode(mode, mb_x, mb_y);
1010 } else {
1011 return mode;
1012 }
1013}
1014
1015static av_always_inline
1016int check_intra_pred8x8_mode_emuedge(int mode, int mb_x, int mb_y)
1017{
1018 switch (mode) {
1019 case DC_PRED8x8:
1020 return check_dc_pred8x8_mode(mode, mb_x, mb_y);
1021 case VERT_PRED8x8:
1022 return !mb_y ? DC_127_PRED8x8 : mode;
1023 case HOR_PRED8x8:
1024 return !mb_x ? DC_129_PRED8x8 : mode;
1025 case PLANE_PRED8x8 /*TM*/:
1026 return check_tm_pred8x8_mode(mode, mb_x, mb_y);
1027 }
1028 return mode;
1029}
1030
1031static av_always_inline
1032int check_tm_pred4x4_mode(int mode, int mb_x, int mb_y)
1033{
1034 if (!mb_x) {
1035 return mb_y ? VERT_VP8_PRED : DC_129_PRED;
1036 } else {
1037 return mb_y ? mode : HOR_VP8_PRED;
1038 }
1039}
1040
1041static av_always_inline
1042int check_intra_pred4x4_mode_emuedge(int mode, int mb_x, int mb_y, int *copy_buf)
1043{
1044 switch (mode) {
1045 case VERT_PRED:
1046 if (!mb_x && mb_y) {
1047 *copy_buf = 1;
1048 return mode;
1049 }
1050 /* fall-through */
1051 case DIAG_DOWN_LEFT_PRED:
1052 case VERT_LEFT_PRED:
1053 return !mb_y ? DC_127_PRED : mode;
1054 case HOR_PRED:
1055 if (!mb_y) {
1056 *copy_buf = 1;
1057 return mode;
a71abb71 1058 }
ee555de7
RB
1059 /* fall-through */
1060 case HOR_UP_PRED:
1061 return !mb_x ? DC_129_PRED : mode;
1062 case TM_VP8_PRED:
1063 return check_tm_pred4x4_mode(mode, mb_x, mb_y);
1064 case DC_PRED: // 4x4 DC doesn't use the same "H.264-style" exceptions as 16x16/8x8 DC
1065 case DIAG_DOWN_RIGHT_PRED:
1066 case VERT_RIGHT_PRED:
1067 case HOR_DOWN_PRED:
1068 if (!mb_y || !mb_x)
1069 *copy_buf = 1;
1070 return mode;
3b636f21
DC
1071 }
1072 return mode;
1073}
1074
414ac27d
JGG
1075static av_always_inline
1076void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
d2840fa4 1077 int mb_x, int mb_y)
3b636f21 1078{
ee555de7 1079 AVCodecContext *avctx = s->avctx;
3b636f21
DC
1080 int x, y, mode, nnz, tr;
1081
9ac831c2
DC
1082 // for the first row, we need to run xchg_mb_border to init the top edge to 127
1083 // otherwise, skip it if we aren't going to deblock
ee555de7 1084 if (!(avctx->flags & CODEC_FLAG_EMU_EDGE && !mb_y) && (s->deblock_filter || !mb_y))
9ac831c2
DC
1085 xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],
1086 s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,
1087 s->filter.simple, 1);
1088
3b636f21 1089 if (mb->mode < MODE_I4x4) {
ee555de7
RB
1090 if (avctx->flags & CODEC_FLAG_EMU_EDGE) { // tested
1091 mode = check_intra_pred8x8_mode_emuedge(mb->mode, mb_x, mb_y);
1092 } else {
1093 mode = check_intra_pred8x8_mode(mb->mode, mb_x, mb_y);
1094 }
3b636f21
DC
1095 s->hpc.pred16x16[mode](dst[0], s->linesize);
1096 } else {
1097 uint8_t *ptr = dst[0];
d2840fa4 1098 uint8_t *intra4x4 = s->intra4x4_pred_mode_mb;
ee555de7 1099 uint8_t tr_top[4] = { 127, 127, 127, 127 };
3b636f21
DC
1100
1101 // all blocks on the right edge of the macroblock use bottom edge
1102 // the top macroblock for their topright edge
1103 uint8_t *tr_right = ptr - s->linesize + 16;
1104
1105 // if we're on the right edge of the frame, said edge is extended
1106 // from the top macroblock
7148da48
RB
1107 if (!(!mb_y && avctx->flags & CODEC_FLAG_EMU_EDGE) &&
1108 mb_x == s->mb_width-1) {
3b636f21
DC
1109 tr = tr_right[-1]*0x01010101;
1110 tr_right = (uint8_t *)&tr;
1111 }
1112
b74f70d6
JGG
1113 if (mb->skip)
1114 AV_ZERO128(s->non_zero_count_cache);
1115
3b636f21
DC
1116 for (y = 0; y < 4; y++) {
1117 uint8_t *topright = ptr + 4 - s->linesize;
1118 for (x = 0; x < 4; x++) {
ee555de7
RB
1119 int copy = 0, linesize = s->linesize;
1120 uint8_t *dst = ptr+4*x;
1121 DECLARE_ALIGNED(4, uint8_t, copy_dst)[5*8];
1122
1123 if ((y == 0 || x == 3) && mb_y == 0 && avctx->flags & CODEC_FLAG_EMU_EDGE) {
1124 topright = tr_top;
1125 } else if (x == 3)
3b636f21
DC
1126 topright = tr_right;
1127
ee555de7
RB
1128 if (avctx->flags & CODEC_FLAG_EMU_EDGE) { // mb_x+x or mb_y+y is a hack but works
1129 mode = check_intra_pred4x4_mode_emuedge(intra4x4[x], mb_x + x, mb_y + y, &copy);
1130 if (copy) {
1131 dst = copy_dst + 12;
1132 linesize = 8;
1133 if (!(mb_y + y)) {
1134 copy_dst[3] = 127U;
9d4bdcb7 1135 AV_WN32A(copy_dst+4, 127U * 0x01010101U);
ee555de7 1136 } else {
9d4bdcb7 1137 AV_COPY32(copy_dst+4, ptr+4*x-s->linesize);
ee555de7
RB
1138 if (!(mb_x + x)) {
1139 copy_dst[3] = 129U;
1140 } else {
1141 copy_dst[3] = ptr[4*x-s->linesize-1];
1142 }
1143 }
1144 if (!(mb_x + x)) {
1145 copy_dst[11] =
1146 copy_dst[19] =
1147 copy_dst[27] =
1148 copy_dst[35] = 129U;
1149 } else {
1150 copy_dst[11] = ptr[4*x -1];
1151 copy_dst[19] = ptr[4*x+s->linesize -1];
1152 copy_dst[27] = ptr[4*x+s->linesize*2-1];
1153 copy_dst[35] = ptr[4*x+s->linesize*3-1];
1154 }
1155 }
1156 } else {
1157 mode = intra4x4[x];
1158 }
1159 s->hpc.pred4x4[mode](dst, topright, linesize);
1160 if (copy) {
9d4bdcb7
RB
1161 AV_COPY32(ptr+4*x , copy_dst+12);
1162 AV_COPY32(ptr+4*x+s->linesize , copy_dst+20);
1163 AV_COPY32(ptr+4*x+s->linesize*2, copy_dst+28);
1164 AV_COPY32(ptr+4*x+s->linesize*3, copy_dst+36);
ee555de7 1165 }
3b636f21
DC
1166
1167 nnz = s->non_zero_count_cache[y][x];
1168 if (nnz) {
1169 if (nnz == 1)
1170 s->vp8dsp.vp8_idct_dc_add(ptr+4*x, s->block[y][x], s->linesize);
1171 else
1172 s->vp8dsp.vp8_idct_add(ptr+4*x, s->block[y][x], s->linesize);
1173 }
1174 topright += 4;
1175 }
1176
1177 ptr += 4*s->linesize;
d2840fa4 1178 intra4x4 += 4;
3b636f21
DC
1179 }
1180 }
1181
ee555de7
RB
1182 if (avctx->flags & CODEC_FLAG_EMU_EDGE) {
1183 mode = check_intra_pred8x8_mode_emuedge(s->chroma_pred_mode, mb_x, mb_y);
1184 } else {
1185 mode = check_intra_pred8x8_mode(s->chroma_pred_mode, mb_x, mb_y);
1186 }
3b636f21
DC
1187 s->hpc.pred8x8[mode](dst[1], s->uvlinesize);
1188 s->hpc.pred8x8[mode](dst[2], s->uvlinesize);
9ac831c2 1189
ee555de7 1190 if (!(avctx->flags & CODEC_FLAG_EMU_EDGE && !mb_y) && (s->deblock_filter || !mb_y))
9ac831c2
DC
1191 xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],
1192 s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,
1193 s->filter.simple, 0);
3b636f21
DC
1194}
1195
1196/**
1197 * Generic MC function.
1198 *
1199 * @param s VP8 decoding context
1200 * @param luma 1 for luma (Y) planes, 0 for chroma (Cb/Cr) planes
1201 * @param dst target buffer for block data at block position
1202 * @param src reference picture buffer at origin (0, 0)
1203 * @param mv motion vector (relative to block position) to get pixel data from
1204 * @param x_off horizontal position of block from origin (0, 0)
1205 * @param y_off vertical position of block from origin (0, 0)
1206 * @param block_w width of block (16, 8 or 4)
1207 * @param block_h height of block (always same as block_w)
1208 * @param width width of src/dst plane data
1209 * @param height height of src/dst plane data
1210 * @param linesize size of a single line of plane data, including padding
e394953e 1211 * @param mc_func motion compensation function pointers (bilinear or sixtap MC)
3b636f21 1212 */
414ac27d
JGG
1213static av_always_inline
1214void vp8_mc(VP8Context *s, int luma,
1215 uint8_t *dst, uint8_t *src, const VP56mv *mv,
1216 int x_off, int y_off, int block_w, int block_h,
1217 int width, int height, int linesize,
1218 vp8_mc_func mc_func[3][3])
3b636f21 1219{
c0498b30 1220 if (AV_RN32A(mv)) {
44002d83
RB
1221 static const uint8_t idx[3][8] = {
1222 { 0, 1, 2, 1, 2, 1, 2, 1 }, // nr. of left extra pixels,
1223 // also function pointer index
1224 { 0, 3, 5, 3, 5, 3, 5, 3 }, // nr. of extra pixels required
1225 { 0, 2, 3, 2, 3, 2, 3, 2 }, // nr. of right extra pixels
1226 };
1227 int mx = (mv->x << luma)&7, mx_idx = idx[0][mx];
1228 int my = (mv->y << luma)&7, my_idx = idx[0][my];
c0498b30
JGG
1229
1230 x_off += mv->x >> (3 - luma);
1231 y_off += mv->y >> (3 - luma);
1232
1233 // edge emulation
1234 src += y_off * linesize + x_off;
44002d83
RB
1235 if (x_off < mx_idx || x_off >= width - block_w - idx[2][mx] ||
1236 y_off < my_idx || y_off >= height - block_h - idx[2][my]) {
2e279598 1237 s->dsp.emulated_edge_mc(s->edge_emu_buffer, src - my_idx * linesize - mx_idx, linesize,
44002d83
RB
1238 block_w + idx[1][mx], block_h + idx[1][my],
1239 x_off - mx_idx, y_off - my_idx, width, height);
1240 src = s->edge_emu_buffer + mx_idx + linesize * my_idx;
c0498b30
JGG
1241 }
1242 mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my);
1243 } else
1244 mc_func[0][0](dst, linesize, src + y_off * linesize + x_off, linesize, block_h, 0, 0);
3b636f21
DC
1245}
1246
414ac27d
JGG
1247static av_always_inline
1248void vp8_mc_part(VP8Context *s, uint8_t *dst[3],
1249 AVFrame *ref_frame, int x_off, int y_off,
1250 int bx_off, int by_off,
1251 int block_w, int block_h,
1252 int width, int height, VP56mv *mv)
7c4dcf81
RB
1253{
1254 VP56mv uvmv = *mv;
1255
1256 /* Y */
1257 vp8_mc(s, 1, dst[0] + by_off * s->linesize + bx_off,
1258 ref_frame->data[0], mv, x_off + bx_off, y_off + by_off,
1259 block_w, block_h, width, height, s->linesize,
1260 s->put_pixels_tab[block_w == 8]);
1261
1262 /* U/V */
1263 if (s->profile == 3) {
1264 uvmv.x &= ~7;
1265 uvmv.y &= ~7;
1266 }
1267 x_off >>= 1; y_off >>= 1;
1268 bx_off >>= 1; by_off >>= 1;
1269 width >>= 1; height >>= 1;
1270 block_w >>= 1; block_h >>= 1;
1271 vp8_mc(s, 0, dst[1] + by_off * s->uvlinesize + bx_off,
1272 ref_frame->data[1], &uvmv, x_off + bx_off, y_off + by_off,
1273 block_w, block_h, width, height, s->uvlinesize,
1274 s->put_pixels_tab[1 + (block_w == 4)]);
1275 vp8_mc(s, 0, dst[2] + by_off * s->uvlinesize + bx_off,
1276 ref_frame->data[2], &uvmv, x_off + bx_off, y_off + by_off,
1277 block_w, block_h, width, height, s->uvlinesize,
1278 s->put_pixels_tab[1 + (block_w == 4)]);
1279}
1280
d864dee8
JGG
1281/* Fetch pixels for estimated mv 4 macroblocks ahead.
1282 * Optimized for 64-byte cache lines. Inspired by ffh264 prefetch_motion. */
414ac27d 1283static av_always_inline void prefetch_motion(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, int mb_xy, int ref)
d864dee8 1284{
ef38842f
JGG
1285 /* Don't prefetch refs that haven't been used very often this frame. */
1286 if (s->ref_count[ref-1] > (mb_xy >> 5)) {
c4211046 1287 int x_off = mb_x << 4, y_off = mb_y << 4;
7e13022a
JGG
1288 int mx = (mb->mv.x>>2) + x_off + 8;
1289 int my = (mb->mv.y>>2) + y_off;
c4211046
JGG
1290 uint8_t **src= s->framep[ref]->data;
1291 int off= mx + (my + (mb_x&3)*4)*s->linesize + 64;
1292 s->dsp.prefetch(src[0]+off, s->linesize, 4);
1293 off= (mx>>1) + ((my>>1) + (mb_x&7))*s->uvlinesize + 64;
1294 s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
1295 }
d864dee8
JGG
1296}
1297
3b636f21
DC
1298/**
1299 * Apply motion vectors to prediction buffer, chapter 18.
1300 */
414ac27d
JGG
1301static av_always_inline
1302void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
1303 int mb_x, int mb_y)
3b636f21
DC
1304{
1305 int x_off = mb_x << 4, y_off = mb_y << 4;
1306 int width = 16*s->mb_width, height = 16*s->mb_height;
d292c345
JGG
1307 AVFrame *ref = s->framep[mb->ref_frame];
1308 VP56mv *bmv = mb->bmv;
3b636f21 1309
73be29b0
JGG
1310 switch (mb->partitioning) {
1311 case VP8_SPLITMVMODE_NONE:
d292c345 1312 vp8_mc_part(s, dst, ref, x_off, y_off,
7c4dcf81 1313 0, 0, 16, 16, width, height, &mb->mv);
73be29b0 1314 break;
7c4dcf81 1315 case VP8_SPLITMVMODE_4x4: {
3b636f21 1316 int x, y;
7c4dcf81 1317 VP56mv uvmv;
3b636f21
DC
1318
1319 /* Y */
1320 for (y = 0; y < 4; y++) {
1321 for (x = 0; x < 4; x++) {
1322 vp8_mc(s, 1, dst[0] + 4*y*s->linesize + x*4,
d292c345 1323 ref->data[0], &bmv[4*y + x],
3b636f21
DC
1324 4*x + x_off, 4*y + y_off, 4, 4,
1325 width, height, s->linesize,
0ef1dbed 1326 s->put_pixels_tab[2]);
3b636f21
DC
1327 }
1328 }
1329
1330 /* U/V */
1331 x_off >>= 1; y_off >>= 1; width >>= 1; height >>= 1;
1332 for (y = 0; y < 2; y++) {
1333 for (x = 0; x < 2; x++) {
1334 uvmv.x = mb->bmv[ 2*y * 4 + 2*x ].x +
1335 mb->bmv[ 2*y * 4 + 2*x+1].x +
1336 mb->bmv[(2*y+1) * 4 + 2*x ].x +
1337 mb->bmv[(2*y+1) * 4 + 2*x+1].x;
1338 uvmv.y = mb->bmv[ 2*y * 4 + 2*x ].y +
1339 mb->bmv[ 2*y * 4 + 2*x+1].y +
1340 mb->bmv[(2*y+1) * 4 + 2*x ].y +
1341 mb->bmv[(2*y+1) * 4 + 2*x+1].y;
8f910a56
SG
1342 uvmv.x = (uvmv.x + 2 + (uvmv.x >> (INT_BIT-1))) >> 2;
1343 uvmv.y = (uvmv.y + 2 + (uvmv.y >> (INT_BIT-1))) >> 2;
3b636f21
DC
1344 if (s->profile == 3) {
1345 uvmv.x &= ~7;
1346 uvmv.y &= ~7;
1347 }
1348 vp8_mc(s, 0, dst[1] + 4*y*s->uvlinesize + x*4,
d292c345 1349 ref->data[1], &uvmv,
3b636f21
DC
1350 4*x + x_off, 4*y + y_off, 4, 4,
1351 width, height, s->uvlinesize,
0ef1dbed 1352 s->put_pixels_tab[2]);
3b636f21 1353 vp8_mc(s, 0, dst[2] + 4*y*s->uvlinesize + x*4,
d292c345 1354 ref->data[2], &uvmv,
3b636f21
DC
1355 4*x + x_off, 4*y + y_off, 4, 4,
1356 width, height, s->uvlinesize,
0ef1dbed 1357 s->put_pixels_tab[2]);
3b636f21
DC
1358 }
1359 }
7c4dcf81
RB
1360 break;
1361 }
1362 case VP8_SPLITMVMODE_16x8:
d292c345
JGG
1363 vp8_mc_part(s, dst, ref, x_off, y_off,
1364 0, 0, 16, 8, width, height, &bmv[0]);
1365 vp8_mc_part(s, dst, ref, x_off, y_off,
1366 0, 8, 16, 8, width, height, &bmv[1]);
7c4dcf81
RB
1367 break;
1368 case VP8_SPLITMVMODE_8x16:
d292c345
JGG
1369 vp8_mc_part(s, dst, ref, x_off, y_off,
1370 0, 0, 8, 16, width, height, &bmv[0]);
1371 vp8_mc_part(s, dst, ref, x_off, y_off,
1372 8, 0, 8, 16, width, height, &bmv[1]);
7c4dcf81
RB
1373 break;
1374 case VP8_SPLITMVMODE_8x8:
d292c345
JGG
1375 vp8_mc_part(s, dst, ref, x_off, y_off,
1376 0, 0, 8, 8, width, height, &bmv[0]);
1377 vp8_mc_part(s, dst, ref, x_off, y_off,
1378 8, 0, 8, 8, width, height, &bmv[1]);
1379 vp8_mc_part(s, dst, ref, x_off, y_off,
1380 0, 8, 8, 8, width, height, &bmv[2]);
1381 vp8_mc_part(s, dst, ref, x_off, y_off,
1382 8, 8, 8, 8, width, height, &bmv[3]);
7c4dcf81 1383 break;
3b636f21
DC
1384 }
1385}
1386
414ac27d 1387static av_always_inline void idct_mb(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb)
3b636f21 1388{
3df56f41 1389 int x, y, ch;
3b636f21 1390
8a467b2d
JGG
1391 if (mb->mode != MODE_I4x4) {
1392 uint8_t *y_dst = dst[0];
3b636f21 1393 for (y = 0; y < 4; y++) {
3df56f41
JGG
1394 uint32_t nnz4 = AV_RN32A(s->non_zero_count_cache[y]);
1395 if (nnz4) {
1396 if (nnz4&~0x01010101) {
8a467b2d 1397 for (x = 0; x < 4; x++) {
3df56f41 1398 int nnz = s->non_zero_count_cache[y][x];
8a467b2d
JGG
1399 if (nnz) {
1400 if (nnz == 1)
1401 s->vp8dsp.vp8_idct_dc_add(y_dst+4*x, s->block[y][x], s->linesize);
1402 else
1403 s->vp8dsp.vp8_idct_add(y_dst+4*x, s->block[y][x], s->linesize);
1404 }
1405 }
1406 } else {
3ae079a3 1407 s->vp8dsp.vp8_idct_dc_add4y(y_dst, s->block[y], s->linesize);
3b636f21
DC
1408 }
1409 }
1410 y_dst += 4*s->linesize;
1411 }
8a467b2d 1412 }
3b636f21 1413
8a467b2d 1414 for (ch = 0; ch < 2; ch++) {
3ae079a3
JGG
1415 uint32_t nnz4 = AV_RN32A(s->non_zero_count_cache[4+ch]);
1416 if (nnz4) {
8a467b2d 1417 uint8_t *ch_dst = dst[1+ch];
3ae079a3
JGG
1418 if (nnz4&~0x01010101) {
1419 for (y = 0; y < 2; y++) {
1420 for (x = 0; x < 2; x++) {
1421 int nnz = s->non_zero_count_cache[4+ch][(y<<1)+x];
1422 if (nnz) {
1423 if (nnz == 1)
1424 s->vp8dsp.vp8_idct_dc_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize);
1425 else
1426 s->vp8dsp.vp8_idct_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize);
1427 }
8a467b2d 1428 }
3ae079a3 1429 ch_dst += 4*s->uvlinesize;
8a467b2d 1430 }
3ae079a3
JGG
1431 } else {
1432 s->vp8dsp.vp8_idct_dc_add4uv(ch_dst, s->block[4+ch], s->uvlinesize);
3b636f21
DC
1433 }
1434 }
3b636f21
DC
1435 }
1436}
1437
414ac27d 1438static av_always_inline void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb, VP8FilterStrength *f )
3b636f21
DC
1439{
1440 int interior_limit, filter_level;
1441
1442 if (s->segmentation.enabled) {
b9a7186b 1443 filter_level = s->segmentation.filter_level[s->segment];
3b636f21
DC
1444 if (!s->segmentation.absolute_vals)
1445 filter_level += s->filter.level;
1446 } else
1447 filter_level = s->filter.level;
1448
1449 if (s->lf_delta.enabled) {
1450 filter_level += s->lf_delta.ref[mb->ref_frame];
1451
1452 if (mb->ref_frame == VP56_FRAME_CURRENT) {
1453 if (mb->mode == MODE_I4x4)
1454 filter_level += s->lf_delta.mode[0];
1455 } else {
1456 if (mb->mode == VP8_MVMODE_ZERO)
1457 filter_level += s->lf_delta.mode[1];
1458 else if (mb->mode == VP8_MVMODE_SPLIT)
1459 filter_level += s->lf_delta.mode[3];
1460 else
1461 filter_level += s->lf_delta.mode[2];
1462 }
1463 }
1464 filter_level = av_clip(filter_level, 0, 63);
1465
1466 interior_limit = filter_level;
1467 if (s->filter.sharpness) {
1468 interior_limit >>= s->filter.sharpness > 4 ? 2 : 1;
1469 interior_limit = FFMIN(interior_limit, 9 - s->filter.sharpness);
1470 }
1471 interior_limit = FFMAX(interior_limit, 1);
1472
968570d6
JGG
1473 f->filter_level = filter_level;
1474 f->inner_limit = interior_limit;
c55e0d34 1475 f->inner_filter = !mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT;
3b636f21
DC
1476}
1477
414ac27d 1478static av_always_inline void filter_mb(VP8Context *s, uint8_t *dst[3], VP8FilterStrength *f, int mb_x, int mb_y)
3b636f21 1479{
968570d6
JGG
1480 int mbedge_lim, bedge_lim, hev_thresh;
1481 int filter_level = f->filter_level;
1482 int inner_limit = f->inner_limit;
c55e0d34 1483 int inner_filter = f->inner_filter;
145d3186
JGG
1484 int linesize = s->linesize;
1485 int uvlinesize = s->uvlinesize;
3b636f21 1486
3b636f21
DC
1487 if (!filter_level)
1488 return;
1489
5245c04d
DC
1490 mbedge_lim = 2*(filter_level+2) + inner_limit;
1491 bedge_lim = 2* filter_level + inner_limit;
968570d6
JGG
1492 hev_thresh = filter_level >= 15;
1493
1494 if (s->keyframe) {
1495 if (filter_level >= 40)
1496 hev_thresh = 2;
1497 } else {
1498 if (filter_level >= 40)
1499 hev_thresh = 3;
1500 else if (filter_level >= 20)
1501 hev_thresh = 2;
1502 }
5245c04d 1503
3b636f21 1504 if (mb_x) {
145d3186 1505 s->vp8dsp.vp8_h_loop_filter16y(dst[0], linesize,
3facfc99 1506 mbedge_lim, inner_limit, hev_thresh);
145d3186 1507 s->vp8dsp.vp8_h_loop_filter8uv(dst[1], dst[2], uvlinesize,
3facfc99 1508 mbedge_lim, inner_limit, hev_thresh);
3b636f21
DC
1509 }
1510
c55e0d34 1511 if (inner_filter) {
145d3186
JGG
1512 s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 4, linesize, bedge_lim,
1513 inner_limit, hev_thresh);
1514 s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 8, linesize, bedge_lim,
1515 inner_limit, hev_thresh);
1516 s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+12, linesize, bedge_lim,
1517 inner_limit, hev_thresh);
1518 s->vp8dsp.vp8_h_loop_filter8uv_inner(dst[1] + 4, dst[2] + 4,
1519 uvlinesize, bedge_lim,
1520 inner_limit, hev_thresh);
3b636f21
DC
1521 }
1522
1523 if (mb_y) {
145d3186 1524 s->vp8dsp.vp8_v_loop_filter16y(dst[0], linesize,
3facfc99 1525 mbedge_lim, inner_limit, hev_thresh);
145d3186 1526 s->vp8dsp.vp8_v_loop_filter8uv(dst[1], dst[2], uvlinesize,
3facfc99 1527 mbedge_lim, inner_limit, hev_thresh);
3b636f21
DC
1528 }
1529
c55e0d34 1530 if (inner_filter) {
145d3186
JGG
1531 s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 4*linesize,
1532 linesize, bedge_lim,
1533 inner_limit, hev_thresh);
1534 s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 8*linesize,
1535 linesize, bedge_lim,
1536 inner_limit, hev_thresh);
1537 s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+12*linesize,
1538 linesize, bedge_lim,
1539 inner_limit, hev_thresh);
1540 s->vp8dsp.vp8_v_loop_filter8uv_inner(dst[1] + 4 * uvlinesize,
1541 dst[2] + 4 * uvlinesize,
1542 uvlinesize, bedge_lim,
3facfc99 1543 inner_limit, hev_thresh);
3b636f21
DC
1544 }
1545}
1546
414ac27d 1547static av_always_inline void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8FilterStrength *f, int mb_x, int mb_y)
3b636f21 1548{
968570d6
JGG
1549 int mbedge_lim, bedge_lim;
1550 int filter_level = f->filter_level;
1551 int inner_limit = f->inner_limit;
c55e0d34 1552 int inner_filter = f->inner_filter;
145d3186 1553 int linesize = s->linesize;
3b636f21 1554
3b636f21
DC
1555 if (!filter_level)
1556 return;
1557
1558 mbedge_lim = 2*(filter_level+2) + inner_limit;
1559 bedge_lim = 2* filter_level + inner_limit;
1560
1561 if (mb_x)
145d3186 1562 s->vp8dsp.vp8_h_loop_filter_simple(dst, linesize, mbedge_lim);
c55e0d34 1563 if (inner_filter) {
145d3186
JGG
1564 s->vp8dsp.vp8_h_loop_filter_simple(dst+ 4, linesize, bedge_lim);
1565 s->vp8dsp.vp8_h_loop_filter_simple(dst+ 8, linesize, bedge_lim);
1566 s->vp8dsp.vp8_h_loop_filter_simple(dst+12, linesize, bedge_lim);
3b636f21
DC
1567 }
1568
1569 if (mb_y)
145d3186 1570 s->vp8dsp.vp8_v_loop_filter_simple(dst, linesize, mbedge_lim);
c55e0d34 1571 if (inner_filter) {
145d3186
JGG
1572 s->vp8dsp.vp8_v_loop_filter_simple(dst+ 4*linesize, linesize, bedge_lim);
1573 s->vp8dsp.vp8_v_loop_filter_simple(dst+ 8*linesize, linesize, bedge_lim);
1574 s->vp8dsp.vp8_v_loop_filter_simple(dst+12*linesize, linesize, bedge_lim);
3b636f21
DC
1575 }
1576}
1577
1578static void filter_mb_row(VP8Context *s, int mb_y)
1579{
968570d6 1580 VP8FilterStrength *f = s->filter_strength;
3b636f21
DC
1581 uint8_t *dst[3] = {
1582 s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize,
1583 s->framep[VP56_FRAME_CURRENT]->data[1] + 8*mb_y*s->uvlinesize,
1584 s->framep[VP56_FRAME_CURRENT]->data[2] + 8*mb_y*s->uvlinesize
1585 };
1586 int mb_x;
1587
1588 for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
9ac831c2 1589 backup_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2], s->linesize, s->uvlinesize, 0);
c55e0d34 1590 filter_mb(s, dst, f++, mb_x, mb_y);
3b636f21
DC
1591 dst[0] += 16;
1592 dst[1] += 8;
1593 dst[2] += 8;
1594 }
1595}
1596
1597static void filter_mb_row_simple(VP8Context *s, int mb_y)
1598{
968570d6 1599 VP8FilterStrength *f = s->filter_strength;
968570d6 1600 uint8_t *dst = s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize;
3b636f21
DC
1601 int mb_x;
1602
1603 for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
9ac831c2 1604 backup_mb_border(s->top_border[mb_x+1], dst, NULL, NULL, s->linesize, 0, 1);
c55e0d34 1605 filter_mb_simple(s, dst, f++, mb_x, mb_y);
3b636f21
DC
1606 dst += 16;
1607 }
1608}
1609
1610static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
1611 AVPacket *avpkt)
1612{
1613 VP8Context *s = avctx->priv_data;
1614 int ret, mb_x, mb_y, i, y, referenced;
1615 enum AVDiscard skip_thresh;
28e241de 1616 AVFrame *av_uninit(curframe);
3b636f21
DC
1617
1618 if ((ret = decode_frame_header(s, avpkt->data, avpkt->size)) < 0)
1619 return ret;
1620
1621 referenced = s->update_last || s->update_golden == VP56_FRAME_CURRENT
1622 || s->update_altref == VP56_FRAME_CURRENT;
1623
1624 skip_thresh = !referenced ? AVDISCARD_NONREF :
1625 !s->keyframe ? AVDISCARD_NONKEY : AVDISCARD_ALL;
1626
1627 if (avctx->skip_frame >= skip_thresh) {
1628 s->invisible = 1;
1629 goto skip_decode;
1630 }
9ac831c2 1631 s->deblock_filter = s->filter.level && avctx->skip_loop_filter < skip_thresh;
3b636f21
DC
1632
1633 for (i = 0; i < 4; i++)
1634 if (&s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1635 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1636 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) {
1637 curframe = s->framep[VP56_FRAME_CURRENT] = &s->frames[i];
1638 break;
1639 }
1640 if (curframe->data[0])
1641 avctx->release_buffer(avctx, curframe);
1642
1643 curframe->key_frame = s->keyframe;
1644 curframe->pict_type = s->keyframe ? FF_I_TYPE : FF_P_TYPE;
1645 curframe->reference = referenced ? 3 : 0;
1646 if ((ret = avctx->get_buffer(avctx, curframe))) {
1647 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed!\n");
1648 return ret;
1649 }
1650
1651 // Given that arithmetic probabilities are updated every frame, it's quite likely
1652 // that the values we have on a random interframe are complete junk if we didn't
1653 // start decode on a keyframe. So just don't display anything rather than junk.
1654 if (!s->keyframe && (!s->framep[VP56_FRAME_PREVIOUS] ||
1655 !s->framep[VP56_FRAME_GOLDEN] ||
1656 !s->framep[VP56_FRAME_GOLDEN2])) {
1657 av_log(avctx, AV_LOG_WARNING, "Discarding interframe without a prior keyframe!\n");
1658 return AVERROR_INVALIDDATA;
1659 }
1660
1661 s->linesize = curframe->linesize[0];
1662 s->uvlinesize = curframe->linesize[1];
1663
1664 if (!s->edge_emu_buffer)
1665 s->edge_emu_buffer = av_malloc(21*s->linesize);
1666
1667 memset(s->top_nnz, 0, s->mb_width*sizeof(*s->top_nnz));
1668
aa93c52c
PM
1669 /* Zero macroblock structures for top/top-left prediction from outside the frame. */
1670 memset(s->macroblocks + s->mb_height*2 - 1, 0, (s->mb_width+1)*sizeof(*s->macroblocks));
c55e0d34 1671
3b636f21 1672 // top edge of 127 for intra prediction
ee555de7
RB
1673 if (!(avctx->flags & CODEC_FLAG_EMU_EDGE)) {
1674 s->top_border[0][15] = s->top_border[0][23] = 127;
1675 memset(s->top_border[1]-1, 127, s->mb_width*sizeof(*s->top_border)+1);
1676 }
c4211046 1677 memset(s->ref_count, 0, sizeof(s->ref_count));
d2840fa4 1678 if (s->keyframe)
ccf13f9e 1679 memset(s->intra4x4_pred_mode_top, DC_PRED, s->mb_width*4);
3b636f21
DC
1680
1681 for (mb_y = 0; mb_y < s->mb_height; mb_y++) {
1682 VP56RangeCoder *c = &s->coeff_partition[mb_y & (s->num_coeff_partitions-1)];
c55e0d34 1683 VP8Macroblock *mb = s->macroblocks + (s->mb_height - mb_y - 1)*2;
aa93c52c 1684 int mb_xy = mb_y*s->mb_width;
3b636f21
DC
1685 uint8_t *dst[3] = {
1686 curframe->data[0] + 16*mb_y*s->linesize,
1687 curframe->data[1] + 8*mb_y*s->uvlinesize,
1688 curframe->data[2] + 8*mb_y*s->uvlinesize
1689 };
1690
aa93c52c 1691 memset(mb - 1, 0, sizeof(*mb)); // zero left macroblock
3b636f21 1692 memset(s->left_nnz, 0, sizeof(s->left_nnz));
d2840fa4 1693 AV_WN32A(s->intra4x4_pred_mode_left, DC_PRED*0x01010101);
3b636f21
DC
1694
1695 // left edge of 129 for intra prediction
ee555de7 1696 if (!(avctx->flags & CODEC_FLAG_EMU_EDGE)) {
3b636f21
DC
1697 for (i = 0; i < 3; i++)
1698 for (y = 0; y < 16>>!!i; y++)
1699 dst[i][y*curframe->linesize[i]-1] = 129;
ee555de7
RB
1700 if (mb_y == 1) // top left edge is also 129
1701 s->top_border[0][15] = s->top_border[0][23] = s->top_border[0][31] = 129;
1702 }
3b636f21 1703
ef38842f 1704 for (mb_x = 0; mb_x < s->mb_width; mb_x++, mb_xy++, mb++) {
d864dee8
JGG
1705 /* Prefetch the current frame, 4 MBs ahead */
1706 s->dsp.prefetch(dst[0] + (mb_x&3)*4*s->linesize + 64, s->linesize, 4);
1707 s->dsp.prefetch(dst[1] + (mb_x&7)*s->uvlinesize + 64, dst[2] - dst[1], 2);
1708
aa93c52c 1709 decode_mb_mode(s, mb, mb_x, mb_y, s->segmentation_map + mb_xy);
3b636f21 1710
ef38842f 1711 prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_PREVIOUS);
c4211046 1712
3b636f21
DC
1713 if (!mb->skip)
1714 decode_mb_coeffs(s, c, mb, s->top_nnz[mb_x], s->left_nnz);
3b636f21 1715
b946111f 1716 if (mb->mode <= MODE_I4x4)
d2840fa4 1717 intra_predict(s, dst, mb, mb_x, mb_y);
b946111f 1718 else
3b636f21 1719 inter_predict(s, dst, mb, mb_x, mb_y);
3b636f21 1720
ef38842f 1721 prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN);
c4211046 1722
3b636f21 1723 if (!mb->skip) {
8a467b2d 1724 idct_mb(s, dst, mb);
3b636f21
DC
1725 } else {
1726 AV_ZERO64(s->left_nnz);
1727 AV_WN64(s->top_nnz[mb_x], 0); // array of 9, so unaligned
1728
1729 // Reset DC block predictors if they would exist if the mb had coefficients
1730 if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
1731 s->left_nnz[8] = 0;
1732 s->top_nnz[mb_x][8] = 0;
1733 }
1734 }
1735
968570d6
JGG
1736 if (s->deblock_filter)
1737 filter_level_for_mb(s, mb, &s->filter_strength[mb_x]);
1738
ef38842f 1739 prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN2);
c4211046 1740
3b636f21
DC
1741 dst[0] += 16;
1742 dst[1] += 8;
1743 dst[2] += 8;
3b636f21 1744 }
9ac831c2 1745 if (s->deblock_filter) {
3b636f21 1746 if (s->filter.simple)
9ac831c2 1747 filter_mb_row_simple(s, mb_y);
3b636f21 1748 else
9ac831c2 1749 filter_mb_row(s, mb_y);
3b636f21
DC
1750 }
1751 }
3b636f21
DC
1752
1753skip_decode:
1754 // if future frames don't use the updated probabilities,
1755 // reset them to the values we saved
1756 if (!s->update_probabilities)
1757 s->prob[0] = s->prob[1];
1758
1759 // check if golden and altref are swapped
1760 if (s->update_altref == VP56_FRAME_GOLDEN &&
1761 s->update_golden == VP56_FRAME_GOLDEN2)
1762 FFSWAP(AVFrame *, s->framep[VP56_FRAME_GOLDEN], s->framep[VP56_FRAME_GOLDEN2]);
1763 else {
1764 if (s->update_altref != VP56_FRAME_NONE)
1765 s->framep[VP56_FRAME_GOLDEN2] = s->framep[s->update_altref];
1766
1767 if (s->update_golden != VP56_FRAME_NONE)
1768 s->framep[VP56_FRAME_GOLDEN] = s->framep[s->update_golden];
1769 }
1770
1771 if (s->update_last) // move cur->prev
1772 s->framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_CURRENT];
1773
1774 // release no longer referenced frames
1775 for (i = 0; i < 4; i++)
1776 if (s->frames[i].data[0] &&
1777 &s->frames[i] != s->framep[VP56_FRAME_CURRENT] &&
1778 &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&
1779 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&
1780 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2])
1781 avctx->release_buffer(avctx, &s->frames[i]);
1782
1783 if (!s->invisible) {
1784 *(AVFrame*)data = *s->framep[VP56_FRAME_CURRENT];
1785 *data_size = sizeof(AVFrame);
1786 }
1787
1788 return avpkt->size;
1789}
1790
1791static av_cold int vp8_decode_init(AVCodecContext *avctx)
1792{
1793 VP8Context *s = avctx->priv_data;
1794
1795 s->avctx = avctx;
1796 avctx->pix_fmt = PIX_FMT_YUV420P;
1797
1798 dsputil_init(&s->dsp, avctx);
1799 ff_h264_pred_init(&s->hpc, CODEC_ID_VP8);
1800 ff_vp8dsp_init(&s->vp8dsp);
1801
3b636f21
DC
1802 return 0;
1803}
1804
1805static av_cold int vp8_decode_free(AVCodecContext *avctx)
1806{
1807 vp8_decode_flush(avctx);
1808 return 0;
1809}
1810
d36beb3f 1811AVCodec ff_vp8_decoder = {
3b636f21
DC
1812 "vp8",
1813 AVMEDIA_TYPE_VIDEO,
1814 CODEC_ID_VP8,
1815 sizeof(VP8Context),
1816 vp8_decode_init,
1817 NULL,
1818 vp8_decode_free,
1819 vp8_decode_frame,
1820 CODEC_CAP_DR1,
1821 .flush = vp8_decode_flush,
1822 .long_name = NULL_IF_CONFIG_SMALL("On2 VP8"),
1823};