Mercurial > libavcodec.hg
annotate vp8.c @ 12222:7acdbfd2a222 libavcodec
Calculate deblock strength per-MB instead of per-row
Gives better cache locality, since the VP8Macroblock structs are still in cache.
Inspired by the way x264 does it.
author | darkshikari |
---|---|
date | Thu, 22 Jul 2010 07:24:22 +0000 |
parents | 45852dac8338 |
children | 93e27a5401de |
rev | line source |
---|---|
11921 | 1 /** |
2 * VP8 compatible video decoder | |
3 * | |
4 * Copyright (C) 2010 David Conrad | |
5 * Copyright (C) 2010 Ronald S. Bultje | |
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 "avcodec.h" | |
25 #include "vp56.h" | |
26 #include "vp8data.h" | |
27 #include "vp8dsp.h" | |
28 #include "h264pred.h" | |
29 #include "rectangle.h" | |
30 | |
31 typedef struct { | |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
32 uint8_t filter_level; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
33 uint8_t inner_limit; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
34 } VP8FilterStrength; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
35 |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
36 typedef struct { |
11921 | 37 uint8_t segment; |
38 uint8_t skip; | |
39 // todo: make it possible to check for at least (i4x4 or split_mv) | |
40 // in one op. are others needed? | |
41 uint8_t mode; | |
42 uint8_t ref_frame; | |
43 uint8_t partitioning; | |
44 VP56mv mv; | |
45 VP56mv bmv[16]; | |
46 } VP8Macroblock; | |
47 | |
48 typedef struct { | |
49 AVCodecContext *avctx; | |
50 DSPContext dsp; | |
51 VP8DSPContext vp8dsp; | |
52 H264PredContext hpc; | |
11974 | 53 vp8_mc_func put_pixels_tab[3][3][3]; |
11921 | 54 AVFrame frames[4]; |
55 AVFrame *framep[4]; | |
56 uint8_t *edge_emu_buffer; | |
57 VP56RangeCoder c; ///< header context, includes mb modes and motion vectors | |
58 int profile; | |
59 | |
60 int mb_width; /* number of horizontal MB */ | |
61 int mb_height; /* number of vertical MB */ | |
62 int linesize; | |
63 int uvlinesize; | |
64 | |
65 int keyframe; | |
66 int invisible; | |
67 int update_last; ///< update VP56_FRAME_PREVIOUS with the current one | |
68 int update_golden; ///< VP56_FRAME_NONE if not updated, or which frame to copy if so | |
69 int update_altref; | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
70 int deblock_filter; |
11921 | 71 |
72 /** | |
73 * If this flag is not set, all the probability updates | |
74 * are discarded after this frame is decoded. | |
75 */ | |
76 int update_probabilities; | |
77 | |
78 /** | |
79 * All coefficients are contained in separate arith coding contexts. | |
80 * There can be 1, 2, 4, or 8 of these after the header context. | |
81 */ | |
82 int num_coeff_partitions; | |
83 VP56RangeCoder coeff_partition[8]; | |
84 | |
85 VP8Macroblock *macroblocks; | |
86 VP8Macroblock *macroblocks_base; | |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
87 VP8FilterStrength *filter_strength; |
11921 | 88 int mb_stride; |
89 | |
90 uint8_t *intra4x4_pred_mode; | |
91 uint8_t *intra4x4_pred_mode_base; | |
92 int b4_stride; | |
93 | |
94 /** | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
95 * Cache of the top row needed for intra prediction |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
96 * 16 for luma, 8 for each chroma plane |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
97 */ |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
98 uint8_t (*top_border)[16+8+8]; |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
99 |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
100 /** |
11921 | 101 * For coeff decode, we need to know whether the above block had non-zero |
102 * coefficients. This means for each macroblock, we need data for 4 luma | |
103 * blocks, 2 u blocks, 2 v blocks, and the luma dc block, for a total of 9 | |
104 * per macroblock. We keep the last row in top_nnz. | |
105 */ | |
106 uint8_t (*top_nnz)[9]; | |
107 DECLARE_ALIGNED(8, uint8_t, left_nnz)[9]; | |
108 | |
109 /** | |
110 * This is the index plus one of the last non-zero coeff | |
111 * for each of the blocks in the current macroblock. | |
112 * So, 0 -> no coeffs | |
113 * 1 -> dc-only (special transform) | |
114 * 2+-> full transform | |
115 */ | |
116 DECLARE_ALIGNED(16, uint8_t, non_zero_count_cache)[6][4]; | |
117 DECLARE_ALIGNED(16, DCTELEM, block)[6][4][16]; | |
12221 | 118 uint8_t intra4x4_pred_mode_mb[16]; |
11921 | 119 |
120 int chroma_pred_mode; ///< 8x8c pred mode of the current macroblock | |
121 | |
122 int mbskip_enabled; | |
123 int sign_bias[4]; ///< one state [0, 1] per ref frame type | |
124 | |
125 /** | |
126 * Base parameters for segmentation, i.e. per-macroblock parameters. | |
127 * These must be kept unchanged even if segmentation is not used for | |
128 * a frame, since the values persist between interframes. | |
129 */ | |
130 struct { | |
131 int enabled; | |
132 int absolute_vals; | |
133 int update_map; | |
134 int8_t base_quant[4]; | |
135 int8_t filter_level[4]; ///< base loop filter level | |
136 } segmentation; | |
137 | |
138 /** | |
139 * Macroblocks can have one of 4 different quants in a frame when | |
140 * segmentation is enabled. | |
141 * If segmentation is disabled, only the first segment's values are used. | |
142 */ | |
143 struct { | |
144 // [0] - DC qmul [1] - AC qmul | |
145 int16_t luma_qmul[2]; | |
146 int16_t luma_dc_qmul[2]; ///< luma dc-only block quant | |
147 int16_t chroma_qmul[2]; | |
148 } qmat[4]; | |
149 | |
150 struct { | |
151 int simple; | |
152 int level; | |
153 int sharpness; | |
154 } filter; | |
155 | |
156 struct { | |
157 int enabled; ///< whether each mb can have a different strength based on mode/ref | |
158 | |
159 /** | |
160 * filter strength adjustment for the following macroblock modes: | |
161 * [0] - i4x4 | |
162 * [1] - zero mv | |
163 * [2] - inter modes except for zero or split mv | |
164 * [3] - split mv | |
165 * i16x16 modes never have any adjustment | |
166 */ | |
167 int8_t mode[4]; | |
168 | |
169 /** | |
170 * filter strength adjustment for macroblocks that reference: | |
171 * [0] - intra / VP56_FRAME_CURRENT | |
172 * [1] - VP56_FRAME_PREVIOUS | |
173 * [2] - VP56_FRAME_GOLDEN | |
174 * [3] - altref / VP56_FRAME_GOLDEN2 | |
175 */ | |
176 int8_t ref[4]; | |
177 } lf_delta; | |
178 | |
179 /** | |
180 * These are all of the updatable probabilities for binary decisions. | |
181 * They are only implictly reset on keyframes, making it quite likely | |
182 * for an interframe to desync if a prior frame's header was corrupt | |
183 * or missing outright! | |
184 */ | |
185 struct { | |
186 uint8_t segmentid[3]; | |
187 uint8_t mbskip; | |
188 uint8_t intra; | |
189 uint8_t last; | |
190 uint8_t golden; | |
191 uint8_t pred16x16[4]; | |
192 uint8_t pred8x8c[3]; | |
193 uint8_t token[4][8][3][NUM_DCT_TOKENS-1]; | |
194 uint8_t mvc[2][19]; | |
195 } prob[2]; | |
196 } VP8Context; | |
197 | |
198 #define RL24(p) (AV_RL16(p) + ((p)[2] << 16)) | |
199 | |
200 static void vp8_decode_flush(AVCodecContext *avctx) | |
201 { | |
202 VP8Context *s = avctx->priv_data; | |
203 int i; | |
204 | |
205 for (i = 0; i < 4; i++) | |
206 if (s->frames[i].data[0]) | |
207 avctx->release_buffer(avctx, &s->frames[i]); | |
208 memset(s->framep, 0, sizeof(s->framep)); | |
209 | |
210 av_freep(&s->macroblocks_base); | |
211 av_freep(&s->intra4x4_pred_mode_base); | |
212 av_freep(&s->top_nnz); | |
213 av_freep(&s->edge_emu_buffer); | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
214 av_freep(&s->top_border); |
11921 | 215 |
216 s->macroblocks = NULL; | |
217 s->intra4x4_pred_mode = NULL; | |
218 } | |
219 | |
220 static int update_dimensions(VP8Context *s, int width, int height) | |
221 { | |
222 int i; | |
223 | |
224 if (avcodec_check_dimensions(s->avctx, width, height)) | |
225 return AVERROR_INVALIDDATA; | |
226 | |
227 vp8_decode_flush(s->avctx); | |
228 | |
229 avcodec_set_dimensions(s->avctx, width, height); | |
230 | |
231 s->mb_width = (s->avctx->coded_width +15) / 16; | |
232 s->mb_height = (s->avctx->coded_height+15) / 16; | |
233 | |
234 // we allocate a border around the top/left of intra4x4 modes | |
235 // this is 4 blocks for intra4x4 to keep 4-byte alignment for fill_rectangle | |
236 s->mb_stride = s->mb_width+1; | |
237 s->b4_stride = 4*s->mb_stride; | |
238 | |
239 s->macroblocks_base = av_mallocz(s->mb_stride*(s->mb_height+1)*sizeof(*s->macroblocks)); | |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
240 s->filter_strength = av_mallocz(s->mb_stride*sizeof(*s->filter_strength)); |
11921 | 241 s->intra4x4_pred_mode_base = av_mallocz(s->b4_stride*(4*s->mb_height+1)); |
242 s->top_nnz = av_mallocz(s->mb_width*sizeof(*s->top_nnz)); | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
243 s->top_border = av_mallocz((s->mb_width+1)*sizeof(*s->top_border)); |
11921 | 244 |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
245 if (!s->macroblocks_base || !s->filter_strength || !s->intra4x4_pred_mode_base || !s->top_nnz || !s->top_border) |
12169 | 246 return AVERROR(ENOMEM); |
247 | |
11921 | 248 s->macroblocks = s->macroblocks_base + 1 + s->mb_stride; |
249 s->intra4x4_pred_mode = s->intra4x4_pred_mode_base + 4 + s->b4_stride; | |
250 | |
251 memset(s->intra4x4_pred_mode_base, DC_PRED, s->b4_stride); | |
252 for (i = 0; i < 4*s->mb_height; i++) | |
253 s->intra4x4_pred_mode[i*s->b4_stride-1] = DC_PRED; | |
254 | |
255 return 0; | |
256 } | |
257 | |
258 static void parse_segment_info(VP8Context *s) | |
259 { | |
260 VP56RangeCoder *c = &s->c; | |
261 int i; | |
262 | |
263 s->segmentation.update_map = vp8_rac_get(c); | |
264 | |
265 if (vp8_rac_get(c)) { // update segment feature data | |
266 s->segmentation.absolute_vals = vp8_rac_get(c); | |
267 | |
268 for (i = 0; i < 4; i++) | |
269 s->segmentation.base_quant[i] = vp8_rac_get_sint(c, 7); | |
270 | |
271 for (i = 0; i < 4; i++) | |
272 s->segmentation.filter_level[i] = vp8_rac_get_sint(c, 6); | |
273 } | |
274 if (s->segmentation.update_map) | |
275 for (i = 0; i < 3; i++) | |
276 s->prob->segmentid[i] = vp8_rac_get(c) ? vp8_rac_get_uint(c, 8) : 255; | |
277 } | |
278 | |
279 static void update_lf_deltas(VP8Context *s) | |
280 { | |
281 VP56RangeCoder *c = &s->c; | |
282 int i; | |
283 | |
284 for (i = 0; i < 4; i++) | |
285 s->lf_delta.ref[i] = vp8_rac_get_sint(c, 6); | |
286 | |
287 for (i = 0; i < 4; i++) | |
288 s->lf_delta.mode[i] = vp8_rac_get_sint(c, 6); | |
289 } | |
290 | |
291 static int setup_partitions(VP8Context *s, const uint8_t *buf, int buf_size) | |
292 { | |
293 const uint8_t *sizes = buf; | |
294 int i; | |
295 | |
296 s->num_coeff_partitions = 1 << vp8_rac_get_uint(&s->c, 2); | |
297 | |
298 buf += 3*(s->num_coeff_partitions-1); | |
299 buf_size -= 3*(s->num_coeff_partitions-1); | |
300 if (buf_size < 0) | |
301 return -1; | |
302 | |
303 for (i = 0; i < s->num_coeff_partitions-1; i++) { | |
304 int size = RL24(sizes + 3*i); | |
305 if (buf_size - size < 0) | |
306 return -1; | |
307 | |
308 vp56_init_range_decoder(&s->coeff_partition[i], buf, size); | |
309 buf += size; | |
310 buf_size -= size; | |
311 } | |
312 vp56_init_range_decoder(&s->coeff_partition[i], buf, buf_size); | |
313 | |
314 return 0; | |
315 } | |
316 | |
317 static void get_quants(VP8Context *s) | |
318 { | |
319 VP56RangeCoder *c = &s->c; | |
320 int i, base_qi; | |
321 | |
322 int yac_qi = vp8_rac_get_uint(c, 7); | |
323 int ydc_delta = vp8_rac_get_sint(c, 4); | |
324 int y2dc_delta = vp8_rac_get_sint(c, 4); | |
325 int y2ac_delta = vp8_rac_get_sint(c, 4); | |
326 int uvdc_delta = vp8_rac_get_sint(c, 4); | |
327 int uvac_delta = vp8_rac_get_sint(c, 4); | |
328 | |
329 for (i = 0; i < 4; i++) { | |
330 if (s->segmentation.enabled) { | |
331 base_qi = s->segmentation.base_quant[i]; | |
332 if (!s->segmentation.absolute_vals) | |
333 base_qi += yac_qi; | |
334 } else | |
335 base_qi = yac_qi; | |
336 | |
337 s->qmat[i].luma_qmul[0] = vp8_dc_qlookup[av_clip(base_qi + ydc_delta , 0, 127)]; | |
338 s->qmat[i].luma_qmul[1] = vp8_ac_qlookup[av_clip(base_qi , 0, 127)]; | |
339 s->qmat[i].luma_dc_qmul[0] = 2 * vp8_dc_qlookup[av_clip(base_qi + y2dc_delta, 0, 127)]; | |
340 s->qmat[i].luma_dc_qmul[1] = 155 * vp8_ac_qlookup[av_clip(base_qi + y2ac_delta, 0, 127)] / 100; | |
341 s->qmat[i].chroma_qmul[0] = vp8_dc_qlookup[av_clip(base_qi + uvdc_delta, 0, 127)]; | |
342 s->qmat[i].chroma_qmul[1] = vp8_ac_qlookup[av_clip(base_qi + uvac_delta, 0, 127)]; | |
343 | |
344 s->qmat[i].luma_dc_qmul[1] = FFMAX(s->qmat[i].luma_dc_qmul[1], 8); | |
345 s->qmat[i].chroma_qmul[0] = FFMIN(s->qmat[i].chroma_qmul[0], 132); | |
346 } | |
347 } | |
348 | |
349 /** | |
350 * Determine which buffers golden and altref should be updated with after this frame. | |
351 * The spec isn't clear here, so I'm going by my understanding of what libvpx does | |
352 * | |
353 * Intra frames update all 3 references | |
354 * Inter frames update VP56_FRAME_PREVIOUS if the update_last flag is set | |
355 * If the update (golden|altref) flag is set, it's updated with the current frame | |
356 * if update_last is set, and VP56_FRAME_PREVIOUS otherwise. | |
357 * If the flag is not set, the number read means: | |
358 * 0: no update | |
359 * 1: VP56_FRAME_PREVIOUS | |
360 * 2: update golden with altref, or update altref with golden | |
361 */ | |
362 static VP56Frame ref_to_update(VP8Context *s, int update, VP56Frame ref) | |
363 { | |
364 VP56RangeCoder *c = &s->c; | |
365 | |
366 if (update) | |
367 return VP56_FRAME_CURRENT; | |
368 | |
369 switch (vp8_rac_get_uint(c, 2)) { | |
370 case 1: | |
371 return VP56_FRAME_PREVIOUS; | |
372 case 2: | |
373 return (ref == VP56_FRAME_GOLDEN) ? VP56_FRAME_GOLDEN2 : VP56_FRAME_GOLDEN; | |
374 } | |
375 return VP56_FRAME_NONE; | |
376 } | |
377 | |
378 static void update_refs(VP8Context *s) | |
379 { | |
380 VP56RangeCoder *c = &s->c; | |
381 | |
382 int update_golden = vp8_rac_get(c); | |
383 int update_altref = vp8_rac_get(c); | |
384 | |
385 s->update_golden = ref_to_update(s, update_golden, VP56_FRAME_GOLDEN); | |
386 s->update_altref = ref_to_update(s, update_altref, VP56_FRAME_GOLDEN2); | |
387 } | |
388 | |
389 static int decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size) | |
390 { | |
391 VP56RangeCoder *c = &s->c; | |
392 int header_size, hscale, vscale, i, j, k, l, ret; | |
393 int width = s->avctx->width; | |
394 int height = s->avctx->height; | |
395 | |
396 s->keyframe = !(buf[0] & 1); | |
397 s->profile = (buf[0]>>1) & 7; | |
398 s->invisible = !(buf[0] & 0x10); | |
399 header_size = RL24(buf) >> 5; | |
400 buf += 3; | |
401 buf_size -= 3; | |
402 | |
11974 | 403 if (s->profile > 3) |
404 av_log(s->avctx, AV_LOG_WARNING, "Unknown profile %d\n", s->profile); | |
405 | |
406 if (!s->profile) | |
407 memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_epel_pixels_tab, sizeof(s->put_pixels_tab)); | |
408 else // profile 1-3 use bilinear, 4+ aren't defined so whatever | |
409 memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_bilinear_pixels_tab, sizeof(s->put_pixels_tab)); | |
11921 | 410 |
411 if (header_size > buf_size - 7*s->keyframe) { | |
412 av_log(s->avctx, AV_LOG_ERROR, "Header size larger than data provided\n"); | |
413 return AVERROR_INVALIDDATA; | |
414 } | |
415 | |
416 if (s->keyframe) { | |
417 if (RL24(buf) != 0x2a019d) { | |
418 av_log(s->avctx, AV_LOG_ERROR, "Invalid start code 0x%x\n", RL24(buf)); | |
419 return AVERROR_INVALIDDATA; | |
420 } | |
421 width = AV_RL16(buf+3) & 0x3fff; | |
422 height = AV_RL16(buf+5) & 0x3fff; | |
423 hscale = buf[4] >> 6; | |
424 vscale = buf[6] >> 6; | |
425 buf += 7; | |
426 buf_size -= 7; | |
427 | |
11970
c7953ee47af4
vp8: warn and request sample if upscaling specified in header
mru
parents:
11950
diff
changeset
|
428 if (hscale || vscale) |
c7953ee47af4
vp8: warn and request sample if upscaling specified in header
mru
parents:
11950
diff
changeset
|
429 av_log_missing_feature(s->avctx, "Upscaling", 1); |
c7953ee47af4
vp8: warn and request sample if upscaling specified in header
mru
parents:
11950
diff
changeset
|
430 |
11921 | 431 s->update_golden = s->update_altref = VP56_FRAME_CURRENT; |
432 memcpy(s->prob->token , vp8_token_default_probs , sizeof(s->prob->token)); | |
433 memcpy(s->prob->pred16x16, vp8_pred16x16_prob_inter, sizeof(s->prob->pred16x16)); | |
434 memcpy(s->prob->pred8x8c , vp8_pred8x8c_prob_inter , sizeof(s->prob->pred8x8c)); | |
435 memcpy(s->prob->mvc , vp8_mv_default_prob , sizeof(s->prob->mvc)); | |
436 memset(&s->segmentation, 0, sizeof(s->segmentation)); | |
437 } | |
438 | |
439 if (!s->macroblocks_base || /* first frame */ | |
440 width != s->avctx->width || height != s->avctx->height) { | |
441 if ((ret = update_dimensions(s, width, height) < 0)) | |
442 return ret; | |
443 } | |
444 | |
445 vp56_init_range_decoder(c, buf, header_size); | |
446 buf += header_size; | |
447 buf_size -= header_size; | |
448 | |
449 if (s->keyframe) { | |
450 if (vp8_rac_get(c)) | |
451 av_log(s->avctx, AV_LOG_WARNING, "Unspecified colorspace\n"); | |
452 vp8_rac_get(c); // whether we can skip clamping in dsp functions | |
453 } | |
454 | |
455 if ((s->segmentation.enabled = vp8_rac_get(c))) | |
456 parse_segment_info(s); | |
457 else | |
458 s->segmentation.update_map = 0; // FIXME: move this to some init function? | |
459 | |
460 s->filter.simple = vp8_rac_get(c); | |
461 s->filter.level = vp8_rac_get_uint(c, 6); | |
462 s->filter.sharpness = vp8_rac_get_uint(c, 3); | |
463 | |
464 if ((s->lf_delta.enabled = vp8_rac_get(c))) | |
465 if (vp8_rac_get(c)) | |
466 update_lf_deltas(s); | |
467 | |
468 if (setup_partitions(s, buf, buf_size)) { | |
469 av_log(s->avctx, AV_LOG_ERROR, "Invalid partitions\n"); | |
470 return AVERROR_INVALIDDATA; | |
471 } | |
472 | |
473 get_quants(s); | |
474 | |
475 if (!s->keyframe) { | |
476 update_refs(s); | |
477 s->sign_bias[VP56_FRAME_GOLDEN] = vp8_rac_get(c); | |
478 s->sign_bias[VP56_FRAME_GOLDEN2 /* altref */] = vp8_rac_get(c); | |
479 } | |
480 | |
481 // if we aren't saving this frame's probabilities for future frames, | |
482 // make a copy of the current probabilities | |
483 if (!(s->update_probabilities = vp8_rac_get(c))) | |
484 s->prob[1] = s->prob[0]; | |
485 | |
486 s->update_last = s->keyframe || vp8_rac_get(c); | |
487 | |
488 for (i = 0; i < 4; i++) | |
489 for (j = 0; j < 8; j++) | |
490 for (k = 0; k < 3; k++) | |
491 for (l = 0; l < NUM_DCT_TOKENS-1; l++) | |
492 if (vp56_rac_get_prob(c, vp8_token_update_probs[i][j][k][l])) | |
493 s->prob->token[i][j][k][l] = vp8_rac_get_uint(c, 8); | |
494 | |
495 if ((s->mbskip_enabled = vp8_rac_get(c))) | |
496 s->prob->mbskip = vp8_rac_get_uint(c, 8); | |
497 | |
498 if (!s->keyframe) { | |
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); | |
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++) | |
513 if (vp56_rac_get_prob(c, vp8_mv_update_prob[i][j])) | |
514 s->prob->mvc[i][j] = vp8_rac_get_nn(c); | |
515 } | |
516 | |
517 return 0; | |
518 } | |
519 | |
520 static inline void clamp_mv(VP8Context *s, VP56mv *dst, const VP56mv *src, | |
521 int mb_x, int mb_y) | |
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 | |
530 static void find_near_mvs(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, | |
12217 | 531 VP56mv near[2], VP56mv *best, uint8_t cnt[4]) |
11921 | 532 { |
12217 | 533 int mb_stride = s->mb_stride; |
534 VP8Macroblock *mb_edge[3] = { mb - mb_stride /* top */, | |
535 mb - 1 /* left */, | |
536 mb - mb_stride - 1 /* top-left */ }; | |
11921 | 537 enum { EDGE_TOP, EDGE_LEFT, EDGE_TOPLEFT }; |
538 VP56mv near_mv[4] = {{ 0 }}; | |
539 enum { CNT_ZERO, CNT_NEAREST, CNT_NEAR, CNT_SPLITMV }; | |
12217 | 540 int idx = CNT_ZERO; |
11921 | 541 int best_idx = CNT_ZERO; |
12217 | 542 int cur_sign_bias = s->sign_bias[mb->ref_frame]; |
543 int *sign_bias = s->sign_bias; | |
11921 | 544 |
545 /* Process MB on top, left and top-left */ | |
12217 | 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. */\ | |
555 mv = ((mv&0x80008000) + 0x00010001) ^ (mv&0x7fff7fff);\ | |
556 }\ | |
557 if (!n || mv != AV_RN32A(&near_mv[idx]))\ | |
558 AV_WN32A(&near_mv[++idx], mv);\ | |
559 cnt[idx] += 1 + (n != 2);\ | |
560 } else\ | |
561 cnt[CNT_ZERO] += 1 + (n != 2);\ | |
562 }\ | |
11921 | 563 } |
12217 | 564 MV_EDGE_CHECK(0) |
565 MV_EDGE_CHECK(1) | |
566 MV_EDGE_CHECK(2) | |
11921 | 567 |
12217 | 568 /* If we have three distinct MVs, merge first and last if they're the same */ |
569 if (cnt[CNT_SPLITMV] && AV_RN32A(&near_mv[1+EDGE_TOP]) == AV_RN32A(&near_mv[1+EDGE_TOPLEFT])) | |
11921 | 570 cnt[CNT_NEAREST] += 1; |
571 | |
572 cnt[CNT_SPLITMV] = ((mb_edge[EDGE_LEFT]->mode == VP8_MVMODE_SPLIT) + | |
573 (mb_edge[EDGE_TOP]->mode == VP8_MVMODE_SPLIT)) * 2 + | |
574 (mb_edge[EDGE_TOPLEFT]->mode == VP8_MVMODE_SPLIT); | |
575 | |
576 /* Swap near and nearest if necessary */ | |
577 if (cnt[CNT_NEAR] > cnt[CNT_NEAREST]) { | |
12217 | 578 FFSWAP(uint8_t, cnt[CNT_NEAREST], cnt[CNT_NEAR]); |
579 FFSWAP( VP56mv, near_mv[CNT_NEAREST], near_mv[CNT_NEAR]); | |
11921 | 580 } |
581 | |
582 /* Choose the best mv out of 0,0 and the nearest mv */ | |
583 if (cnt[CNT_NEAREST] >= cnt[CNT_ZERO]) | |
584 best_idx = CNT_NEAREST; | |
585 | |
586 clamp_mv(s, best, &near_mv[best_idx], mb_x, mb_y); | |
587 near[0] = near_mv[CNT_NEAREST]; | |
588 near[1] = near_mv[CNT_NEAR]; | |
589 } | |
590 | |
591 /** | |
592 * Motion vector coding, 17.1. | |
593 */ | |
594 static int read_mv_component(VP56RangeCoder *c, const uint8_t *p) | |
595 { | |
596 int x = 0; | |
597 | |
598 if (vp56_rac_get_prob(c, p[0])) { | |
599 int i; | |
600 | |
601 for (i = 0; i < 3; i++) | |
602 x += vp56_rac_get_prob(c, p[9 + i]) << i; | |
603 for (i = 9; i > 3; i--) | |
604 x += vp56_rac_get_prob(c, p[9 + i]) << i; | |
605 if (!(x & 0xFFF0) || vp56_rac_get_prob(c, p[12])) | |
606 x += 8; | |
607 } else | |
608 x = vp8_rac_get_tree(c, vp8_small_mvtree, &p[2]); | |
609 | |
610 return (x && vp56_rac_get_prob(c, p[1])) ? -x : x; | |
611 } | |
612 | |
12219 | 613 static const uint8_t *get_submv_prob(uint32_t left, uint32_t top) |
11921 | 614 { |
12219 | 615 if (left == top) |
616 return vp8_submv_prob[4-!!left]; | |
617 if (!top) | |
11921 | 618 return vp8_submv_prob[2]; |
12219 | 619 return vp8_submv_prob[1-!!left]; |
11921 | 620 } |
621 | |
622 /** | |
623 * Split motion vector prediction, 16.4. | |
11990
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
624 * @returns the number of motion vectors parsed (2, 4 or 16) |
11921 | 625 */ |
11990
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
626 static int decode_splitmvs(VP8Context *s, VP56RangeCoder *c, |
11921 | 627 VP8Macroblock *mb, VP56mv *base_mv) |
628 { | |
629 int part_idx = mb->partitioning = | |
630 vp8_rac_get_tree(c, vp8_mbsplit_tree, vp8_mbsplit_prob); | |
631 int n, num = vp8_mbsplit_count[part_idx]; | |
12219 | 632 VP8Macroblock *top_mb = &mb[-s->mb_stride]; |
633 VP8Macroblock *left_mb = &mb[-1]; | |
634 const uint8_t *mbsplits_left = vp8_mbsplits[left_mb->partitioning], | |
635 *mbsplits_top = vp8_mbsplits[top_mb->partitioning], | |
636 *mbsplits_cur = vp8_mbsplits[part_idx], | |
11990
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
637 *firstidx = vp8_mbfirstidx[part_idx]; |
12219 | 638 VP56mv *top_mv = top_mb->bmv; |
639 VP56mv *left_mv = left_mb->bmv; | |
640 VP56mv *cur_mv = mb->bmv; | |
11921 | 641 |
642 for (n = 0; n < num; n++) { | |
11990
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
643 int k = firstidx[n]; |
12219 | 644 uint32_t left, above; |
11990
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
645 const uint8_t *submv_prob; |
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
646 |
12219 | 647 if (!(k & 3)) |
648 left = AV_RN32A(&left_mv[mbsplits_left[k + 3]]); | |
649 else | |
650 left = AV_RN32A(&cur_mv[mbsplits_cur[k - 1]]); | |
651 if (k <= 3) | |
652 above = AV_RN32A(&top_mv[mbsplits_top[k + 12]]); | |
653 else | |
654 above = AV_RN32A(&cur_mv[mbsplits_cur[k - 4]]); | |
11990
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
655 |
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
656 submv_prob = get_submv_prob(left, above); |
11921 | 657 |
658 switch (vp8_rac_get_tree(c, vp8_submv_ref_tree, submv_prob)) { | |
659 case VP8_SUBMVMODE_NEW4X4: | |
11990
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
660 mb->bmv[n].y = base_mv->y + read_mv_component(c, s->prob->mvc[0]); |
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
661 mb->bmv[n].x = base_mv->x + read_mv_component(c, s->prob->mvc[1]); |
11921 | 662 break; |
663 case VP8_SUBMVMODE_ZERO4X4: | |
12219 | 664 AV_WN32A(&mb->bmv[n], 0); |
11921 | 665 break; |
666 case VP8_SUBMVMODE_LEFT4X4: | |
12219 | 667 AV_WN32A(&mb->bmv[n], left); |
11921 | 668 break; |
669 case VP8_SUBMVMODE_TOP4X4: | |
12219 | 670 AV_WN32A(&mb->bmv[n], above); |
11921 | 671 break; |
672 } | |
11990
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
673 } |
11921 | 674 |
11990
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
675 return num; |
11921 | 676 } |
677 | |
678 static inline void decode_intra4x4_modes(VP56RangeCoder *c, uint8_t *intra4x4, | |
679 int stride, int keyframe) | |
680 { | |
12221 | 681 int x, y, t, l, i; |
11921 | 682 |
12221 | 683 if (keyframe) { |
684 const uint8_t *ctx; | |
685 for (y = 0; y < 4; y++) { | |
686 for (x = 0; x < 4; x++) { | |
11921 | 687 t = intra4x4[x - stride]; |
688 l = intra4x4[x - 1]; | |
689 ctx = vp8_pred4x4_prob_intra[t][l]; | |
12221 | 690 intra4x4[x] = vp8_rac_get_tree(c, vp8_pred4x4_tree, ctx); |
11921 | 691 } |
12221 | 692 intra4x4 += stride; |
11921 | 693 } |
12221 | 694 } else { |
695 for (i = 0; i < 16; i++) | |
696 intra4x4[i] = vp8_rac_get_tree(c, vp8_pred4x4_tree, vp8_pred4x4_prob_inter); | |
11921 | 697 } |
698 } | |
699 | |
700 static void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, | |
701 uint8_t *intra4x4) | |
702 { | |
703 VP56RangeCoder *c = &s->c; | |
704 | |
705 if (s->segmentation.update_map) | |
706 mb->segment = vp8_rac_get_tree(c, vp8_segmentid_tree, s->prob->segmentid); | |
707 | |
708 mb->skip = s->mbskip_enabled ? vp56_rac_get_prob(c, s->prob->mbskip) : 0; | |
709 | |
710 if (s->keyframe) { | |
711 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_intra, vp8_pred16x16_prob_intra); | |
712 | |
713 if (mb->mode == MODE_I4x4) { | |
714 decode_intra4x4_modes(c, intra4x4, s->b4_stride, 1); | |
715 } else | |
716 fill_rectangle(intra4x4, 4, 4, s->b4_stride, vp8_pred4x4_mode[mb->mode], 1); | |
717 | |
718 s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, vp8_pred8x8c_prob_intra); | |
719 mb->ref_frame = VP56_FRAME_CURRENT; | |
720 } else if (vp56_rac_get_prob(c, s->prob->intra)) { | |
721 VP56mv near[2], best; | |
12217 | 722 uint8_t cnt[4] = { 0 }; |
11921 | 723 uint8_t p[4]; |
724 | |
725 // inter MB, 16.2 | |
726 if (vp56_rac_get_prob(c, s->prob->last)) | |
727 mb->ref_frame = vp56_rac_get_prob(c, s->prob->golden) ? | |
728 VP56_FRAME_GOLDEN2 /* altref */ : VP56_FRAME_GOLDEN; | |
729 else | |
730 mb->ref_frame = VP56_FRAME_PREVIOUS; | |
731 | |
732 // motion vectors, 16.3 | |
733 find_near_mvs(s, mb, mb_x, mb_y, near, &best, cnt); | |
12217 | 734 p[0] = vp8_mode_contexts[cnt[0]][0]; |
735 p[1] = vp8_mode_contexts[cnt[1]][1]; | |
736 p[2] = vp8_mode_contexts[cnt[2]][2]; | |
737 p[3] = vp8_mode_contexts[cnt[3]][3]; | |
11921 | 738 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_mvinter, p); |
739 switch (mb->mode) { | |
740 case VP8_MVMODE_SPLIT: | |
11990
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
741 mb->mv = mb->bmv[decode_splitmvs(s, c, mb, &best) - 1]; |
11921 | 742 break; |
743 case VP8_MVMODE_ZERO: | |
744 mb->mv.x = 0; | |
745 mb->mv.y = 0; | |
746 break; | |
747 case VP8_MVMODE_NEAREST: | |
748 clamp_mv(s, &mb->mv, &near[0], mb_x, mb_y); | |
749 break; | |
750 case VP8_MVMODE_NEAR: | |
751 clamp_mv(s, &mb->mv, &near[1], mb_x, mb_y); | |
752 break; | |
753 case VP8_MVMODE_NEW: | |
754 mb->mv.y = best.y + read_mv_component(c, s->prob->mvc[0]); | |
755 mb->mv.x = best.x + read_mv_component(c, s->prob->mvc[1]); | |
756 break; | |
757 } | |
758 if (mb->mode != VP8_MVMODE_SPLIT) { | |
11990
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
759 mb->partitioning = VP8_SPLITMVMODE_NONE; |
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
760 mb->bmv[0] = mb->mv; |
11921 | 761 } |
762 } else { | |
763 // intra MB, 16.1 | |
764 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_inter, s->prob->pred16x16); | |
765 | |
12220
0f635b1f7861
Avoid useless fill_rectangle in P-frames in VP8
darkshikari
parents:
12219
diff
changeset
|
766 if (mb->mode == MODE_I4x4) |
12221 | 767 decode_intra4x4_modes(c, intra4x4, 4, 0); |
11921 | 768 |
769 s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, s->prob->pred8x8c); | |
770 mb->ref_frame = VP56_FRAME_CURRENT; | |
771 } | |
772 } | |
773 | |
774 /** | |
12115 | 775 * @param c arithmetic bitstream reader context |
776 * @param block destination for block coefficients | |
777 * @param probs probabilities to use when reading trees from the bitstream | |
11921 | 778 * @param i initial coeff index, 0 unless a separate DC block is coded |
779 * @param zero_nhood the initial prediction context for number of surrounding | |
780 * all-zero blocks (only left/top, so 0-2) | |
12062
372f7fed2806
Avoid square brackets in Doxygen comments; Doxygen chokes on them.
diego
parents:
11990
diff
changeset
|
781 * @param qmul array holding the dc/ac dequant factor at position 0/1 |
11921 | 782 * @return 0 if no coeffs were decoded |
783 * otherwise, the index of the last coeff decoded plus one | |
784 */ | |
785 static int decode_block_coeffs(VP56RangeCoder *c, DCTELEM block[16], | |
786 uint8_t probs[8][3][NUM_DCT_TOKENS-1], | |
787 int i, int zero_nhood, int16_t qmul[2]) | |
788 { | |
789 int token, nonzero = 0; | |
790 int offset = 0; | |
791 | |
792 for (; i < 16; i++) { | |
793 token = vp8_rac_get_tree_with_offset(c, vp8_coeff_tree, probs[vp8_coeff_band[i]][zero_nhood], offset); | |
794 | |
795 if (token == DCT_EOB) | |
796 break; | |
797 else if (token >= DCT_CAT1) { | |
798 int cat = token-DCT_CAT1; | |
799 token = vp8_rac_get_coeff(c, vp8_dct_cat_prob[cat]); | |
800 token += vp8_dct_cat_offset[cat]; | |
801 } | |
802 | |
803 // after the first token, the non-zero prediction context becomes | |
804 // based on the last decoded coeff | |
805 if (!token) { | |
806 zero_nhood = 0; | |
807 offset = 1; | |
808 continue; | |
809 } else if (token == 1) | |
810 zero_nhood = 1; | |
811 else | |
812 zero_nhood = 2; | |
813 | |
814 // todo: full [16] qmat? load into register? | |
815 block[zigzag_scan[i]] = (vp8_rac_get(c) ? -token : token) * qmul[!!i]; | |
816 nonzero = i+1; | |
817 offset = 0; | |
818 } | |
819 return nonzero; | |
820 } | |
821 | |
822 static void decode_mb_coeffs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb, | |
823 uint8_t t_nnz[9], uint8_t l_nnz[9]) | |
824 { | |
825 LOCAL_ALIGNED_16(DCTELEM, dc,[16]); | |
826 int i, x, y, luma_start = 0, luma_ctx = 3; | |
827 int nnz_pred, nnz, nnz_total = 0; | |
828 int segment = s->segmentation.enabled ? mb->segment : 0; | |
829 | |
830 s->dsp.clear_blocks((DCTELEM *)s->block); | |
831 | |
832 if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) { | |
833 AV_ZERO128(dc); | |
834 AV_ZERO128(dc+8); | |
835 nnz_pred = t_nnz[8] + l_nnz[8]; | |
836 | |
837 // decode DC values and do hadamard | |
838 nnz = decode_block_coeffs(c, dc, s->prob->token[1], 0, nnz_pred, | |
839 s->qmat[segment].luma_dc_qmul); | |
840 l_nnz[8] = t_nnz[8] = !!nnz; | |
841 nnz_total += nnz; | |
842 s->vp8dsp.vp8_luma_dc_wht(s->block, dc); | |
843 luma_start = 1; | |
844 luma_ctx = 0; | |
845 } | |
846 | |
847 // luma blocks | |
848 for (y = 0; y < 4; y++) | |
849 for (x = 0; x < 4; x++) { | |
850 nnz_pred = l_nnz[y] + t_nnz[x]; | |
851 nnz = decode_block_coeffs(c, s->block[y][x], s->prob->token[luma_ctx], luma_start, | |
852 nnz_pred, s->qmat[segment].luma_qmul); | |
853 // nnz+luma_start may be one more than the actual last index, but we don't care | |
854 s->non_zero_count_cache[y][x] = nnz + luma_start; | |
855 t_nnz[x] = l_nnz[y] = !!nnz; | |
856 nnz_total += nnz; | |
857 } | |
858 | |
859 // chroma blocks | |
860 // TODO: what to do about dimensions? 2nd dim for luma is x, | |
861 // but for chroma it's (y<<1)|x | |
862 for (i = 4; i < 6; i++) | |
863 for (y = 0; y < 2; y++) | |
864 for (x = 0; x < 2; x++) { | |
865 nnz_pred = l_nnz[i+2*y] + t_nnz[i+2*x]; | |
866 nnz = decode_block_coeffs(c, s->block[i][(y<<1)+x], s->prob->token[2], 0, | |
867 nnz_pred, s->qmat[segment].chroma_qmul); | |
868 s->non_zero_count_cache[i][(y<<1)+x] = nnz; | |
869 t_nnz[i+2*x] = l_nnz[i+2*y] = !!nnz; | |
870 nnz_total += nnz; | |
871 } | |
872 | |
873 // if there were no coded coeffs despite the macroblock not being marked skip, | |
874 // we MUST not do the inner loop filter and should not do IDCT | |
875 // Since skip isn't used for bitstream prediction, just manually set it. | |
876 if (!nnz_total) | |
877 mb->skip = 1; | |
878 } | |
879 | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
880 static av_always_inline |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
881 void backup_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
882 int linesize, int uvlinesize, int simple) |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
883 { |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
884 AV_COPY128(top_border, src_y + 15*linesize); |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
885 if (!simple) { |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
886 AV_COPY64(top_border+16, src_cb + 7*uvlinesize); |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
887 AV_COPY64(top_border+24, src_cr + 7*uvlinesize); |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
888 } |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
889 } |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
890 |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
891 static av_always_inline |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
892 void xchg_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
893 int linesize, int uvlinesize, int mb_x, int mb_y, int mb_width, |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
894 int simple, int xchg) |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
895 { |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
896 uint8_t *top_border_m1 = top_border-32; // for TL prediction |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
897 src_y -= linesize; |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
898 src_cb -= uvlinesize; |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
899 src_cr -= uvlinesize; |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
900 |
12202 | 901 #define XCHG(a,b,xchg) do { \ |
902 if (xchg) AV_SWAP64(b,a); \ | |
903 else AV_COPY64(b,a); \ | |
904 } while (0) | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
905 |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
906 XCHG(top_border_m1+8, src_y-8, xchg); |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
907 XCHG(top_border, src_y, xchg); |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
908 XCHG(top_border+8, src_y+8, 1); |
12201
c4b53914f286
vp8: add do { } while(0) around XCHG() macro to avoid confusing if/else
mru
parents:
12200
diff
changeset
|
909 if (mb_x < mb_width-1) |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
910 XCHG(top_border+32, src_y+16, 1); |
12201
c4b53914f286
vp8: add do { } while(0) around XCHG() macro to avoid confusing if/else
mru
parents:
12200
diff
changeset
|
911 |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
912 // only copy chroma for normal loop filter |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
913 // or to initialize the top row to 127 |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
914 if (!simple || !mb_y) { |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
915 XCHG(top_border_m1+16, src_cb-8, xchg); |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
916 XCHG(top_border_m1+24, src_cr-8, xchg); |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
917 XCHG(top_border+16, src_cb, 1); |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
918 XCHG(top_border+24, src_cr, 1); |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
919 } |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
920 } |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
921 |
11921 | 922 static int check_intra_pred_mode(int mode, int mb_x, int mb_y) |
923 { | |
924 if (mode == DC_PRED8x8) { | |
925 if (!(mb_x|mb_y)) | |
926 mode = DC_128_PRED8x8; | |
927 else if (!mb_y) | |
928 mode = LEFT_DC_PRED8x8; | |
929 else if (!mb_x) | |
930 mode = TOP_DC_PRED8x8; | |
931 } | |
932 return mode; | |
933 } | |
934 | |
935 static void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb, | |
12221 | 936 uint8_t *intra4x4, int mb_x, int mb_y) |
11921 | 937 { |
938 int x, y, mode, nnz, tr; | |
939 | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
940 // for the first row, we need to run xchg_mb_border to init the top edge to 127 |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
941 // otherwise, skip it if we aren't going to deblock |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
942 if (s->deblock_filter || !mb_y) |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
943 xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2], |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
944 s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width, |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
945 s->filter.simple, 1); |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
946 |
11921 | 947 if (mb->mode < MODE_I4x4) { |
948 mode = check_intra_pred_mode(mb->mode, mb_x, mb_y); | |
949 s->hpc.pred16x16[mode](dst[0], s->linesize); | |
950 } else { | |
951 uint8_t *ptr = dst[0]; | |
12221 | 952 int stride = s->keyframe ? s->b4_stride : 4; |
11921 | 953 |
954 // all blocks on the right edge of the macroblock use bottom edge | |
955 // the top macroblock for their topright edge | |
956 uint8_t *tr_right = ptr - s->linesize + 16; | |
957 | |
958 // if we're on the right edge of the frame, said edge is extended | |
959 // from the top macroblock | |
960 if (mb_x == s->mb_width-1) { | |
961 tr = tr_right[-1]*0x01010101; | |
962 tr_right = (uint8_t *)&tr; | |
963 } | |
964 | |
965 for (y = 0; y < 4; y++) { | |
966 uint8_t *topright = ptr + 4 - s->linesize; | |
967 for (x = 0; x < 4; x++) { | |
968 if (x == 3) | |
969 topright = tr_right; | |
970 | |
12221 | 971 s->hpc.pred4x4[intra4x4[x]](ptr+4*x, topright, s->linesize); |
11921 | 972 |
973 nnz = s->non_zero_count_cache[y][x]; | |
974 if (nnz) { | |
975 if (nnz == 1) | |
976 s->vp8dsp.vp8_idct_dc_add(ptr+4*x, s->block[y][x], s->linesize); | |
977 else | |
978 s->vp8dsp.vp8_idct_add(ptr+4*x, s->block[y][x], s->linesize); | |
979 } | |
980 topright += 4; | |
981 } | |
982 | |
983 ptr += 4*s->linesize; | |
12221 | 984 intra4x4 += stride; |
11921 | 985 } |
986 } | |
987 | |
988 mode = check_intra_pred_mode(s->chroma_pred_mode, mb_x, mb_y); | |
989 s->hpc.pred8x8[mode](dst[1], s->uvlinesize); | |
990 s->hpc.pred8x8[mode](dst[2], s->uvlinesize); | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
991 |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
992 if (s->deblock_filter || !mb_y) |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
993 xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2], |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
994 s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width, |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
995 s->filter.simple, 0); |
11921 | 996 } |
997 | |
998 /** | |
999 * Generic MC function. | |
1000 * | |
1001 * @param s VP8 decoding context | |
1002 * @param luma 1 for luma (Y) planes, 0 for chroma (Cb/Cr) planes | |
1003 * @param dst target buffer for block data at block position | |
1004 * @param src reference picture buffer at origin (0, 0) | |
1005 * @param mv motion vector (relative to block position) to get pixel data from | |
1006 * @param x_off horizontal position of block from origin (0, 0) | |
1007 * @param y_off vertical position of block from origin (0, 0) | |
1008 * @param block_w width of block (16, 8 or 4) | |
1009 * @param block_h height of block (always same as block_w) | |
1010 * @param width width of src/dst plane data | |
1011 * @param height height of src/dst plane data | |
1012 * @param linesize size of a single line of plane data, including padding | |
12115 | 1013 * @param mc_func motion compensation function pointers (bilinear or sixtap MC) |
11921 | 1014 */ |
1015 static inline void vp8_mc(VP8Context *s, int luma, | |
1016 uint8_t *dst, uint8_t *src, const VP56mv *mv, | |
1017 int x_off, int y_off, int block_w, int block_h, | |
1018 int width, int height, int linesize, | |
11950 | 1019 vp8_mc_func mc_func[3][3]) |
11921 | 1020 { |
12218 | 1021 if (AV_RN32A(mv)) { |
1022 static const uint8_t idx[8] = { 0, 1, 2, 1, 2, 1, 2, 1 }; | |
1023 int mx = (mv->x << luma)&7, mx_idx = idx[mx]; | |
1024 int my = (mv->y << luma)&7, my_idx = idx[my]; | |
11921 | 1025 |
12218 | 1026 x_off += mv->x >> (3 - luma); |
1027 y_off += mv->y >> (3 - luma); | |
11921 | 1028 |
12218 | 1029 // edge emulation |
1030 src += y_off * linesize + x_off; | |
1031 if (x_off < 2 || x_off >= width - block_w - 3 || | |
1032 y_off < 2 || y_off >= height - block_h - 3) { | |
1033 ff_emulated_edge_mc(s->edge_emu_buffer, src - 2 * linesize - 2, linesize, | |
1034 block_w + 5, block_h + 5, | |
1035 x_off - 2, y_off - 2, width, height); | |
1036 src = s->edge_emu_buffer + 2 + linesize * 2; | |
1037 } | |
1038 mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my); | |
1039 } else | |
1040 mc_func[0][0](dst, linesize, src + y_off * linesize + x_off, linesize, block_h, 0, 0); | |
11921 | 1041 } |
1042 | |
11989
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1043 static inline void vp8_mc_part(VP8Context *s, uint8_t *dst[3], |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1044 AVFrame *ref_frame, int x_off, int y_off, |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1045 int bx_off, int by_off, |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1046 int block_w, int block_h, |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1047 int width, int height, VP56mv *mv) |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1048 { |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1049 VP56mv uvmv = *mv; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1050 |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1051 /* Y */ |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1052 vp8_mc(s, 1, dst[0] + by_off * s->linesize + bx_off, |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1053 ref_frame->data[0], mv, x_off + bx_off, y_off + by_off, |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1054 block_w, block_h, width, height, s->linesize, |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1055 s->put_pixels_tab[block_w == 8]); |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1056 |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1057 /* U/V */ |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1058 if (s->profile == 3) { |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1059 uvmv.x &= ~7; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1060 uvmv.y &= ~7; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1061 } |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1062 x_off >>= 1; y_off >>= 1; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1063 bx_off >>= 1; by_off >>= 1; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1064 width >>= 1; height >>= 1; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1065 block_w >>= 1; block_h >>= 1; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1066 vp8_mc(s, 0, dst[1] + by_off * s->uvlinesize + bx_off, |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1067 ref_frame->data[1], &uvmv, x_off + bx_off, y_off + by_off, |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1068 block_w, block_h, width, height, s->uvlinesize, |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1069 s->put_pixels_tab[1 + (block_w == 4)]); |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1070 vp8_mc(s, 0, dst[2] + by_off * s->uvlinesize + bx_off, |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1071 ref_frame->data[2], &uvmv, x_off + bx_off, y_off + by_off, |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1072 block_w, block_h, width, height, s->uvlinesize, |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1073 s->put_pixels_tab[1 + (block_w == 4)]); |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1074 } |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1075 |
12215 | 1076 /* Fetch pixels for estimated mv 4 macroblocks ahead. |
1077 * Optimized for 64-byte cache lines. Inspired by ffh264 prefetch_motion. */ | |
1078 static inline void prefetch_motion(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, int x_off, int y_off, int ref) | |
1079 { | |
1080 if (mb->ref_frame != VP56_FRAME_CURRENT) { | |
1081 int mx = mb->mv.x + x_off + 8; | |
1082 int my = mb->mv.y + y_off; | |
1083 uint8_t **src= s->framep[mb->ref_frame]->data; | |
1084 int off= mx + (my + (mb_x&3)*4)*s->linesize + 64; | |
1085 s->dsp.prefetch(src[0]+off, s->linesize, 4); | |
1086 off= (mx>>1) + ((my>>1) + (mb_x&7))*s->uvlinesize + 64; | |
1087 s->dsp.prefetch(src[1]+off, src[2]-src[1], 2); | |
1088 } | |
1089 } | |
1090 | |
11921 | 1091 /** |
1092 * Apply motion vectors to prediction buffer, chapter 18. | |
1093 */ | |
1094 static void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb, | |
1095 int mb_x, int mb_y) | |
1096 { | |
1097 int x_off = mb_x << 4, y_off = mb_y << 4; | |
1098 int width = 16*s->mb_width, height = 16*s->mb_height; | |
1099 | |
12215 | 1100 prefetch_motion(s, mb, mb_x, mb_y, x_off, y_off, VP56_FRAME_PREVIOUS); |
1101 | |
11921 | 1102 if (mb->mode < VP8_MVMODE_SPLIT) { |
11989
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1103 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off, |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1104 0, 0, 16, 16, width, height, &mb->mv); |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1105 } else switch (mb->partitioning) { |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1106 case VP8_SPLITMVMODE_4x4: { |
11921 | 1107 int x, y; |
11989
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1108 VP56mv uvmv; |
11921 | 1109 |
1110 /* Y */ | |
1111 for (y = 0; y < 4; y++) { | |
1112 for (x = 0; x < 4; x++) { | |
1113 vp8_mc(s, 1, dst[0] + 4*y*s->linesize + x*4, | |
1114 s->framep[mb->ref_frame]->data[0], &mb->bmv[4*y + x], | |
1115 4*x + x_off, 4*y + y_off, 4, 4, | |
1116 width, height, s->linesize, | |
11974 | 1117 s->put_pixels_tab[2]); |
11921 | 1118 } |
1119 } | |
1120 | |
1121 /* U/V */ | |
1122 x_off >>= 1; y_off >>= 1; width >>= 1; height >>= 1; | |
1123 for (y = 0; y < 2; y++) { | |
1124 for (x = 0; x < 2; x++) { | |
1125 uvmv.x = mb->bmv[ 2*y * 4 + 2*x ].x + | |
1126 mb->bmv[ 2*y * 4 + 2*x+1].x + | |
1127 mb->bmv[(2*y+1) * 4 + 2*x ].x + | |
1128 mb->bmv[(2*y+1) * 4 + 2*x+1].x; | |
1129 uvmv.y = mb->bmv[ 2*y * 4 + 2*x ].y + | |
1130 mb->bmv[ 2*y * 4 + 2*x+1].y + | |
1131 mb->bmv[(2*y+1) * 4 + 2*x ].y + | |
1132 mb->bmv[(2*y+1) * 4 + 2*x+1].y; | |
11937
bc617cceacb1
avoid conditional and division in chroma MV calculation
stefang
parents:
11921
diff
changeset
|
1133 uvmv.x = (uvmv.x + 2 + (uvmv.x >> (INT_BIT-1))) >> 2; |
bc617cceacb1
avoid conditional and division in chroma MV calculation
stefang
parents:
11921
diff
changeset
|
1134 uvmv.y = (uvmv.y + 2 + (uvmv.y >> (INT_BIT-1))) >> 2; |
11921 | 1135 if (s->profile == 3) { |
1136 uvmv.x &= ~7; | |
1137 uvmv.y &= ~7; | |
1138 } | |
1139 vp8_mc(s, 0, dst[1] + 4*y*s->uvlinesize + x*4, | |
1140 s->framep[mb->ref_frame]->data[1], &uvmv, | |
1141 4*x + x_off, 4*y + y_off, 4, 4, | |
1142 width, height, s->uvlinesize, | |
11974 | 1143 s->put_pixels_tab[2]); |
11921 | 1144 vp8_mc(s, 0, dst[2] + 4*y*s->uvlinesize + x*4, |
1145 s->framep[mb->ref_frame]->data[2], &uvmv, | |
1146 4*x + x_off, 4*y + y_off, 4, 4, | |
1147 width, height, s->uvlinesize, | |
11974 | 1148 s->put_pixels_tab[2]); |
11921 | 1149 } |
1150 } | |
11989
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1151 break; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1152 } |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1153 case VP8_SPLITMVMODE_16x8: |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1154 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off, |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1155 0, 0, 16, 8, width, height, &mb->bmv[0]); |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1156 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off, |
11990
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
1157 0, 8, 16, 8, width, height, &mb->bmv[1]); |
11989
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1158 break; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1159 case VP8_SPLITMVMODE_8x16: |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1160 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off, |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1161 0, 0, 8, 16, width, height, &mb->bmv[0]); |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1162 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off, |
11990
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
1163 8, 0, 8, 16, width, height, &mb->bmv[1]); |
11989
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1164 break; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1165 case VP8_SPLITMVMODE_8x8: |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1166 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off, |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1167 0, 0, 8, 8, width, height, &mb->bmv[0]); |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1168 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off, |
11990
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
1169 8, 0, 8, 8, width, height, &mb->bmv[1]); |
11989
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1170 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off, |
11990
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
1171 0, 8, 8, 8, width, height, &mb->bmv[2]); |
11989
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1172 vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off, |
11990
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
1173 8, 8, 8, 8, width, height, &mb->bmv[3]); |
11989
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1174 break; |
11921 | 1175 } |
12215 | 1176 |
1177 prefetch_motion(s, mb, mb_x, mb_y, x_off, y_off, VP56_FRAME_GOLDEN); | |
11921 | 1178 } |
1179 | |
1180 static void idct_mb(VP8Context *s, uint8_t *y_dst, uint8_t *u_dst, uint8_t *v_dst, | |
1181 VP8Macroblock *mb) | |
1182 { | |
1183 int x, y, nnz; | |
1184 | |
1185 if (mb->mode != MODE_I4x4) | |
1186 for (y = 0; y < 4; y++) { | |
1187 for (x = 0; x < 4; x++) { | |
1188 nnz = s->non_zero_count_cache[y][x]; | |
1189 if (nnz) { | |
1190 if (nnz == 1) | |
1191 s->vp8dsp.vp8_idct_dc_add(y_dst+4*x, s->block[y][x], s->linesize); | |
1192 else | |
1193 s->vp8dsp.vp8_idct_add(y_dst+4*x, s->block[y][x], s->linesize); | |
1194 } | |
1195 } | |
1196 y_dst += 4*s->linesize; | |
1197 } | |
1198 | |
1199 for (y = 0; y < 2; y++) { | |
1200 for (x = 0; x < 2; x++) { | |
1201 nnz = s->non_zero_count_cache[4][(y<<1)+x]; | |
1202 if (nnz) { | |
1203 if (nnz == 1) | |
1204 s->vp8dsp.vp8_idct_dc_add(u_dst+4*x, s->block[4][(y<<1)+x], s->uvlinesize); | |
1205 else | |
1206 s->vp8dsp.vp8_idct_add(u_dst+4*x, s->block[4][(y<<1)+x], s->uvlinesize); | |
1207 } | |
1208 | |
1209 nnz = s->non_zero_count_cache[5][(y<<1)+x]; | |
1210 if (nnz) { | |
1211 if (nnz == 1) | |
1212 s->vp8dsp.vp8_idct_dc_add(v_dst+4*x, s->block[5][(y<<1)+x], s->uvlinesize); | |
1213 else | |
1214 s->vp8dsp.vp8_idct_add(v_dst+4*x, s->block[5][(y<<1)+x], s->uvlinesize); | |
1215 } | |
1216 } | |
1217 u_dst += 4*s->uvlinesize; | |
1218 v_dst += 4*s->uvlinesize; | |
1219 } | |
1220 } | |
1221 | |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1222 static void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb, VP8FilterStrength *f ) |
11921 | 1223 { |
1224 int interior_limit, filter_level; | |
1225 | |
1226 if (s->segmentation.enabled) { | |
1227 filter_level = s->segmentation.filter_level[mb->segment]; | |
1228 if (!s->segmentation.absolute_vals) | |
1229 filter_level += s->filter.level; | |
1230 } else | |
1231 filter_level = s->filter.level; | |
1232 | |
1233 if (s->lf_delta.enabled) { | |
1234 filter_level += s->lf_delta.ref[mb->ref_frame]; | |
1235 | |
1236 if (mb->ref_frame == VP56_FRAME_CURRENT) { | |
1237 if (mb->mode == MODE_I4x4) | |
1238 filter_level += s->lf_delta.mode[0]; | |
1239 } else { | |
1240 if (mb->mode == VP8_MVMODE_ZERO) | |
1241 filter_level += s->lf_delta.mode[1]; | |
1242 else if (mb->mode == VP8_MVMODE_SPLIT) | |
1243 filter_level += s->lf_delta.mode[3]; | |
1244 else | |
1245 filter_level += s->lf_delta.mode[2]; | |
1246 } | |
1247 } | |
1248 filter_level = av_clip(filter_level, 0, 63); | |
1249 | |
1250 interior_limit = filter_level; | |
1251 if (s->filter.sharpness) { | |
1252 interior_limit >>= s->filter.sharpness > 4 ? 2 : 1; | |
1253 interior_limit = FFMIN(interior_limit, 9 - s->filter.sharpness); | |
1254 } | |
1255 interior_limit = FFMAX(interior_limit, 1); | |
1256 | |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1257 f->filter_level = filter_level; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1258 f->inner_limit = interior_limit; |
11921 | 1259 } |
1260 | |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1261 static void filter_mb(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb, VP8FilterStrength *f, int mb_x, int mb_y) |
11921 | 1262 { |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1263 int mbedge_lim, bedge_lim, hev_thresh; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1264 int filter_level = f->filter_level; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1265 int inner_limit = f->inner_limit; |
11921 | 1266 |
1267 if (!filter_level) | |
1268 return; | |
1269 | |
12081
812e23197d64
VP8: Move calculation of outer filter limit out of dsp functions for normal
conrad
parents:
12062
diff
changeset
|
1270 mbedge_lim = 2*(filter_level+2) + inner_limit; |
812e23197d64
VP8: Move calculation of outer filter limit out of dsp functions for normal
conrad
parents:
12062
diff
changeset
|
1271 bedge_lim = 2* filter_level + inner_limit; |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1272 hev_thresh = filter_level >= 15; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1273 |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1274 if (s->keyframe) { |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1275 if (filter_level >= 40) |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1276 hev_thresh = 2; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1277 } else { |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1278 if (filter_level >= 40) |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1279 hev_thresh = 3; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1280 else if (filter_level >= 20) |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1281 hev_thresh = 2; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1282 } |
12081
812e23197d64
VP8: Move calculation of outer filter limit out of dsp functions for normal
conrad
parents:
12062
diff
changeset
|
1283 |
11921 | 1284 if (mb_x) { |
12194
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1285 s->vp8dsp.vp8_h_loop_filter16y(dst[0], s->linesize, |
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1286 mbedge_lim, inner_limit, hev_thresh); |
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1287 s->vp8dsp.vp8_h_loop_filter8uv(dst[1], dst[2], s->uvlinesize, |
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1288 mbedge_lim, inner_limit, hev_thresh); |
11921 | 1289 } |
1290 | |
1291 if (!mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT) { | |
12194
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1292 s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 4, s->linesize, bedge_lim, |
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1293 inner_limit, hev_thresh); |
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1294 s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 8, s->linesize, bedge_lim, |
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1295 inner_limit, hev_thresh); |
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1296 s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+12, s->linesize, bedge_lim, |
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1297 inner_limit, hev_thresh); |
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1298 s->vp8dsp.vp8_h_loop_filter8uv_inner(dst[1] + 4, dst[2] + 4, |
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1299 s->uvlinesize, bedge_lim, |
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1300 inner_limit, hev_thresh); |
11921 | 1301 } |
1302 | |
1303 if (mb_y) { | |
12194
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1304 s->vp8dsp.vp8_v_loop_filter16y(dst[0], s->linesize, |
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1305 mbedge_lim, inner_limit, hev_thresh); |
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1306 s->vp8dsp.vp8_v_loop_filter8uv(dst[1], dst[2], s->uvlinesize, |
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1307 mbedge_lim, inner_limit, hev_thresh); |
11921 | 1308 } |
1309 | |
1310 if (!mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT) { | |
12194
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1311 s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 4*s->linesize, |
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1312 s->linesize, bedge_lim, |
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1313 inner_limit, hev_thresh); |
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1314 s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 8*s->linesize, |
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1315 s->linesize, bedge_lim, |
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1316 inner_limit, hev_thresh); |
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1317 s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+12*s->linesize, |
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1318 s->linesize, bedge_lim, |
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1319 inner_limit, hev_thresh); |
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1320 s->vp8dsp.vp8_v_loop_filter8uv_inner(dst[1] + 4 * s->uvlinesize, |
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1321 dst[2] + 4 * s->uvlinesize, |
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1322 s->uvlinesize, bedge_lim, |
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1323 inner_limit, hev_thresh); |
11921 | 1324 } |
1325 } | |
1326 | |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1327 static void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8Macroblock *mb, VP8FilterStrength *f, int mb_x, int mb_y) |
11921 | 1328 { |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1329 int mbedge_lim, bedge_lim; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1330 int filter_level = f->filter_level; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1331 int inner_limit = f->inner_limit; |
11921 | 1332 |
1333 if (!filter_level) | |
1334 return; | |
1335 | |
1336 mbedge_lim = 2*(filter_level+2) + inner_limit; | |
1337 bedge_lim = 2* filter_level + inner_limit; | |
1338 | |
1339 if (mb_x) | |
1340 s->vp8dsp.vp8_h_loop_filter_simple(dst, s->linesize, mbedge_lim); | |
1341 if (!mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT) { | |
1342 s->vp8dsp.vp8_h_loop_filter_simple(dst+ 4, s->linesize, bedge_lim); | |
1343 s->vp8dsp.vp8_h_loop_filter_simple(dst+ 8, s->linesize, bedge_lim); | |
1344 s->vp8dsp.vp8_h_loop_filter_simple(dst+12, s->linesize, bedge_lim); | |
1345 } | |
1346 | |
1347 if (mb_y) | |
1348 s->vp8dsp.vp8_v_loop_filter_simple(dst, s->linesize, mbedge_lim); | |
1349 if (!mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT) { | |
1350 s->vp8dsp.vp8_v_loop_filter_simple(dst+ 4*s->linesize, s->linesize, bedge_lim); | |
1351 s->vp8dsp.vp8_v_loop_filter_simple(dst+ 8*s->linesize, s->linesize, bedge_lim); | |
1352 s->vp8dsp.vp8_v_loop_filter_simple(dst+12*s->linesize, s->linesize, bedge_lim); | |
1353 } | |
1354 } | |
1355 | |
1356 static void filter_mb_row(VP8Context *s, int mb_y) | |
1357 { | |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1358 VP8FilterStrength *f = s->filter_strength; |
11921 | 1359 VP8Macroblock *mb = s->macroblocks + mb_y*s->mb_stride; |
1360 uint8_t *dst[3] = { | |
1361 s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize, | |
1362 s->framep[VP56_FRAME_CURRENT]->data[1] + 8*mb_y*s->uvlinesize, | |
1363 s->framep[VP56_FRAME_CURRENT]->data[2] + 8*mb_y*s->uvlinesize | |
1364 }; | |
1365 int mb_x; | |
1366 | |
1367 for (mb_x = 0; mb_x < s->mb_width; mb_x++) { | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
1368 backup_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2], s->linesize, s->uvlinesize, 0); |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1369 filter_mb(s, dst, mb++, f++, mb_x, mb_y); |
11921 | 1370 dst[0] += 16; |
1371 dst[1] += 8; | |
1372 dst[2] += 8; | |
1373 } | |
1374 } | |
1375 | |
1376 static void filter_mb_row_simple(VP8Context *s, int mb_y) | |
1377 { | |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1378 VP8FilterStrength *f = s->filter_strength; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1379 VP8Macroblock *mb = s->macroblocks + mb_y*s->mb_stride; |
11921 | 1380 uint8_t *dst = s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize; |
1381 int mb_x; | |
1382 | |
1383 for (mb_x = 0; mb_x < s->mb_width; mb_x++) { | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
1384 backup_mb_border(s->top_border[mb_x+1], dst, NULL, NULL, s->linesize, 0, 1); |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1385 filter_mb_simple(s, dst, mb++, f++, mb_x, mb_y); |
11921 | 1386 dst += 16; |
1387 } | |
1388 } | |
1389 | |
1390 static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size, | |
1391 AVPacket *avpkt) | |
1392 { | |
1393 VP8Context *s = avctx->priv_data; | |
1394 int ret, mb_x, mb_y, i, y, referenced; | |
1395 enum AVDiscard skip_thresh; | |
1396 AVFrame *curframe; | |
1397 | |
1398 if ((ret = decode_frame_header(s, avpkt->data, avpkt->size)) < 0) | |
1399 return ret; | |
1400 | |
1401 referenced = s->update_last || s->update_golden == VP56_FRAME_CURRENT | |
1402 || s->update_altref == VP56_FRAME_CURRENT; | |
1403 | |
1404 skip_thresh = !referenced ? AVDISCARD_NONREF : | |
1405 !s->keyframe ? AVDISCARD_NONKEY : AVDISCARD_ALL; | |
1406 | |
1407 if (avctx->skip_frame >= skip_thresh) { | |
1408 s->invisible = 1; | |
1409 goto skip_decode; | |
1410 } | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
1411 s->deblock_filter = s->filter.level && avctx->skip_loop_filter < skip_thresh; |
11921 | 1412 |
1413 for (i = 0; i < 4; i++) | |
1414 if (&s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] && | |
1415 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] && | |
1416 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) { | |
1417 curframe = s->framep[VP56_FRAME_CURRENT] = &s->frames[i]; | |
1418 break; | |
1419 } | |
1420 if (curframe->data[0]) | |
1421 avctx->release_buffer(avctx, curframe); | |
1422 | |
1423 curframe->key_frame = s->keyframe; | |
1424 curframe->pict_type = s->keyframe ? FF_I_TYPE : FF_P_TYPE; | |
1425 curframe->reference = referenced ? 3 : 0; | |
1426 if ((ret = avctx->get_buffer(avctx, curframe))) { | |
1427 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed!\n"); | |
1428 return ret; | |
1429 } | |
1430 | |
1431 // Given that arithmetic probabilities are updated every frame, it's quite likely | |
1432 // that the values we have on a random interframe are complete junk if we didn't | |
1433 // start decode on a keyframe. So just don't display anything rather than junk. | |
1434 if (!s->keyframe && (!s->framep[VP56_FRAME_PREVIOUS] || | |
1435 !s->framep[VP56_FRAME_GOLDEN] || | |
1436 !s->framep[VP56_FRAME_GOLDEN2])) { | |
1437 av_log(avctx, AV_LOG_WARNING, "Discarding interframe without a prior keyframe!\n"); | |
1438 return AVERROR_INVALIDDATA; | |
1439 } | |
1440 | |
1441 s->linesize = curframe->linesize[0]; | |
1442 s->uvlinesize = curframe->linesize[1]; | |
1443 | |
1444 if (!s->edge_emu_buffer) | |
1445 s->edge_emu_buffer = av_malloc(21*s->linesize); | |
1446 | |
1447 memset(s->top_nnz, 0, s->mb_width*sizeof(*s->top_nnz)); | |
1448 | |
1449 // top edge of 127 for intra prediction | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
1450 memset(s->top_border, 127, (s->mb_width+1)*sizeof(*s->top_border)); |
11921 | 1451 |
1452 for (mb_y = 0; mb_y < s->mb_height; mb_y++) { | |
1453 VP56RangeCoder *c = &s->coeff_partition[mb_y & (s->num_coeff_partitions-1)]; | |
1454 VP8Macroblock *mb = s->macroblocks + mb_y*s->mb_stride; | |
1455 uint8_t *intra4x4 = s->intra4x4_pred_mode + 4*mb_y*s->b4_stride; | |
1456 uint8_t *dst[3] = { | |
1457 curframe->data[0] + 16*mb_y*s->linesize, | |
1458 curframe->data[1] + 8*mb_y*s->uvlinesize, | |
1459 curframe->data[2] + 8*mb_y*s->uvlinesize | |
1460 }; | |
1461 | |
1462 memset(s->left_nnz, 0, sizeof(s->left_nnz)); | |
1463 | |
1464 // left edge of 129 for intra prediction | |
1465 if (!(avctx->flags & CODEC_FLAG_EMU_EDGE)) | |
1466 for (i = 0; i < 3; i++) | |
1467 for (y = 0; y < 16>>!!i; y++) | |
1468 dst[i][y*curframe->linesize[i]-1] = 129; | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
1469 if (mb_y) |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
1470 memset(s->top_border, 129, sizeof(*s->top_border)); |
11921 | 1471 |
1472 for (mb_x = 0; mb_x < s->mb_width; mb_x++) { | |
12221 | 1473 uint8_t *intra4x4_mb = s->keyframe ? intra4x4 + 4*mb_x : s->intra4x4_pred_mode_mb; |
1474 | |
12215 | 1475 /* Prefetch the current frame, 4 MBs ahead */ |
1476 s->dsp.prefetch(dst[0] + (mb_x&3)*4*s->linesize + 64, s->linesize, 4); | |
1477 s->dsp.prefetch(dst[1] + (mb_x&7)*s->uvlinesize + 64, dst[2] - dst[1], 2); | |
1478 | |
12221 | 1479 decode_mb_mode(s, mb, mb_x, mb_y, intra4x4_mb); |
11921 | 1480 |
1481 if (!mb->skip) | |
1482 decode_mb_coeffs(s, c, mb, s->top_nnz[mb_x], s->left_nnz); | |
1483 else { | |
1484 AV_ZERO128(s->non_zero_count_cache); // luma | |
1485 AV_ZERO64(s->non_zero_count_cache[4]); // chroma | |
1486 } | |
1487 | |
1488 if (mb->mode <= MODE_I4x4) { | |
12221 | 1489 intra_predict(s, dst, mb, intra4x4_mb, mb_x, mb_y); |
11921 | 1490 memset(mb->bmv, 0, sizeof(mb->bmv)); |
1491 } else { | |
1492 inter_predict(s, dst, mb, mb_x, mb_y); | |
1493 } | |
1494 | |
1495 if (!mb->skip) { | |
1496 idct_mb(s, dst[0], dst[1], dst[2], mb); | |
1497 } else { | |
1498 AV_ZERO64(s->left_nnz); | |
1499 AV_WN64(s->top_nnz[mb_x], 0); // array of 9, so unaligned | |
1500 | |
1501 // Reset DC block predictors if they would exist if the mb had coefficients | |
1502 if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) { | |
1503 s->left_nnz[8] = 0; | |
1504 s->top_nnz[mb_x][8] = 0; | |
1505 } | |
1506 } | |
1507 | |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1508 if (s->deblock_filter) |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1509 filter_level_for_mb(s, mb, &s->filter_strength[mb_x]); |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1510 |
11921 | 1511 dst[0] += 16; |
1512 dst[1] += 8; | |
1513 dst[2] += 8; | |
1514 mb++; | |
1515 } | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
1516 if (s->deblock_filter) { |
11921 | 1517 if (s->filter.simple) |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
1518 filter_mb_row_simple(s, mb_y); |
11921 | 1519 else |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
1520 filter_mb_row(s, mb_y); |
11921 | 1521 } |
1522 } | |
1523 | |
1524 skip_decode: | |
1525 // if future frames don't use the updated probabilities, | |
1526 // reset them to the values we saved | |
1527 if (!s->update_probabilities) | |
1528 s->prob[0] = s->prob[1]; | |
1529 | |
1530 // check if golden and altref are swapped | |
1531 if (s->update_altref == VP56_FRAME_GOLDEN && | |
1532 s->update_golden == VP56_FRAME_GOLDEN2) | |
1533 FFSWAP(AVFrame *, s->framep[VP56_FRAME_GOLDEN], s->framep[VP56_FRAME_GOLDEN2]); | |
1534 else { | |
1535 if (s->update_altref != VP56_FRAME_NONE) | |
1536 s->framep[VP56_FRAME_GOLDEN2] = s->framep[s->update_altref]; | |
1537 | |
1538 if (s->update_golden != VP56_FRAME_NONE) | |
1539 s->framep[VP56_FRAME_GOLDEN] = s->framep[s->update_golden]; | |
1540 } | |
1541 | |
1542 if (s->update_last) // move cur->prev | |
1543 s->framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_CURRENT]; | |
1544 | |
1545 // release no longer referenced frames | |
1546 for (i = 0; i < 4; i++) | |
1547 if (s->frames[i].data[0] && | |
1548 &s->frames[i] != s->framep[VP56_FRAME_CURRENT] && | |
1549 &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] && | |
1550 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] && | |
1551 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) | |
1552 avctx->release_buffer(avctx, &s->frames[i]); | |
1553 | |
1554 if (!s->invisible) { | |
1555 *(AVFrame*)data = *s->framep[VP56_FRAME_CURRENT]; | |
1556 *data_size = sizeof(AVFrame); | |
1557 } | |
1558 | |
1559 return avpkt->size; | |
1560 } | |
1561 | |
1562 static av_cold int vp8_decode_init(AVCodecContext *avctx) | |
1563 { | |
1564 VP8Context *s = avctx->priv_data; | |
1565 | |
1566 s->avctx = avctx; | |
1567 avctx->pix_fmt = PIX_FMT_YUV420P; | |
1568 | |
1569 dsputil_init(&s->dsp, avctx); | |
1570 ff_h264_pred_init(&s->hpc, CODEC_ID_VP8); | |
1571 ff_vp8dsp_init(&s->vp8dsp); | |
1572 | |
1573 // intra pred needs edge emulation among other things | |
1574 if (avctx->flags&CODEC_FLAG_EMU_EDGE) { | |
11947 | 1575 av_log(avctx, AV_LOG_ERROR, "Edge emulation not supported\n"); |
11921 | 1576 return AVERROR_PATCHWELCOME; |
1577 } | |
1578 | |
1579 return 0; | |
1580 } | |
1581 | |
1582 static av_cold int vp8_decode_free(AVCodecContext *avctx) | |
1583 { | |
1584 vp8_decode_flush(avctx); | |
1585 return 0; | |
1586 } | |
1587 | |
1588 AVCodec vp8_decoder = { | |
1589 "vp8", | |
1590 AVMEDIA_TYPE_VIDEO, | |
1591 CODEC_ID_VP8, | |
1592 sizeof(VP8Context), | |
1593 vp8_decode_init, | |
1594 NULL, | |
1595 vp8_decode_free, | |
1596 vp8_decode_frame, | |
1597 CODEC_CAP_DR1, | |
1598 .flush = vp8_decode_flush, | |
1599 .long_name = NULL_IF_CONFIG_SMALL("On2 VP8"), | |
1600 }; |