Mercurial > libavcodec.hg
annotate vp8.c @ 12248:121272849def libavcodec
VP8: always_inline some things to force gcc to do the right thing
Mostly seems to help in the MC code, which gets a hundred cycles faster.
author | darkshikari |
---|---|
date | Fri, 23 Jul 2010 21:36:21 +0000 |
parents | 50a96623366b |
children | 35ee666e4496 |
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; |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
34 uint8_t inner_filter; |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
35 } VP8FilterStrength; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
36 |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
37 typedef struct { |
11921 | 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; | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
92 uint8_t *segmentation_map; |
11921 | 93 int b4_stride; |
94 | |
95 /** | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
96 * 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
|
97 * 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
|
98 */ |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
99 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
|
100 |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
101 /** |
11921 | 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]; | |
12221 | 119 uint8_t intra4x4_pred_mode_mb[16]; |
11921 | 120 |
121 int chroma_pred_mode; ///< 8x8c pred mode of the current macroblock | |
12224
5b7d690b761b
VP8: Don't store segment in macroblock struct anymore.
darkshikari
parents:
12223
diff
changeset
|
122 int segment; ///< segment of the current macroblock |
11921 | 123 |
124 int mbskip_enabled; | |
125 int sign_bias[4]; ///< one state [0, 1] per ref frame type | |
12231 | 126 int ref_count[3]; |
11921 | 127 |
128 /** | |
129 * Base parameters for segmentation, i.e. per-macroblock parameters. | |
130 * These must be kept unchanged even if segmentation is not used for | |
131 * a frame, since the values persist between interframes. | |
132 */ | |
133 struct { | |
134 int enabled; | |
135 int absolute_vals; | |
136 int update_map; | |
137 int8_t base_quant[4]; | |
138 int8_t filter_level[4]; ///< base loop filter level | |
139 } segmentation; | |
140 | |
141 /** | |
142 * Macroblocks can have one of 4 different quants in a frame when | |
143 * segmentation is enabled. | |
144 * If segmentation is disabled, only the first segment's values are used. | |
145 */ | |
146 struct { | |
147 // [0] - DC qmul [1] - AC qmul | |
148 int16_t luma_qmul[2]; | |
149 int16_t luma_dc_qmul[2]; ///< luma dc-only block quant | |
150 int16_t chroma_qmul[2]; | |
151 } qmat[4]; | |
152 | |
153 struct { | |
154 int simple; | |
155 int level; | |
156 int sharpness; | |
157 } filter; | |
158 | |
159 struct { | |
160 int enabled; ///< whether each mb can have a different strength based on mode/ref | |
161 | |
162 /** | |
163 * filter strength adjustment for the following macroblock modes: | |
164 * [0] - i4x4 | |
165 * [1] - zero mv | |
166 * [2] - inter modes except for zero or split mv | |
167 * [3] - split mv | |
168 * i16x16 modes never have any adjustment | |
169 */ | |
170 int8_t mode[4]; | |
171 | |
172 /** | |
173 * filter strength adjustment for macroblocks that reference: | |
174 * [0] - intra / VP56_FRAME_CURRENT | |
175 * [1] - VP56_FRAME_PREVIOUS | |
176 * [2] - VP56_FRAME_GOLDEN | |
177 * [3] - altref / VP56_FRAME_GOLDEN2 | |
178 */ | |
179 int8_t ref[4]; | |
180 } lf_delta; | |
181 | |
182 /** | |
183 * These are all of the updatable probabilities for binary decisions. | |
184 * They are only implictly reset on keyframes, making it quite likely | |
185 * for an interframe to desync if a prior frame's header was corrupt | |
186 * or missing outright! | |
187 */ | |
188 struct { | |
189 uint8_t segmentid[3]; | |
190 uint8_t mbskip; | |
191 uint8_t intra; | |
192 uint8_t last; | |
193 uint8_t golden; | |
194 uint8_t pred16x16[4]; | |
195 uint8_t pred8x8c[3]; | |
196 uint8_t token[4][8][3][NUM_DCT_TOKENS-1]; | |
197 uint8_t mvc[2][19]; | |
198 } prob[2]; | |
199 } VP8Context; | |
200 | |
201 static void vp8_decode_flush(AVCodecContext *avctx) | |
202 { | |
203 VP8Context *s = avctx->priv_data; | |
204 int i; | |
205 | |
206 for (i = 0; i < 4; i++) | |
207 if (s->frames[i].data[0]) | |
208 avctx->release_buffer(avctx, &s->frames[i]); | |
209 memset(s->framep, 0, sizeof(s->framep)); | |
210 | |
211 av_freep(&s->macroblocks_base); | |
212 av_freep(&s->intra4x4_pred_mode_base); | |
213 av_freep(&s->top_nnz); | |
214 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
|
215 av_freep(&s->top_border); |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
216 av_freep(&s->segmentation_map); |
11921 | 217 |
218 s->macroblocks = NULL; | |
219 s->intra4x4_pred_mode = NULL; | |
220 } | |
221 | |
222 static int update_dimensions(VP8Context *s, int width, int height) | |
223 { | |
224 int i; | |
225 | |
226 if (avcodec_check_dimensions(s->avctx, width, height)) | |
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 | |
236 // we allocate a border around the top/left of intra4x4 modes | |
237 // this is 4 blocks for intra4x4 to keep 4-byte alignment for fill_rectangle | |
238 s->mb_stride = s->mb_width+1; | |
239 s->b4_stride = 4*s->mb_stride; | |
240 | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
241 s->macroblocks_base = av_mallocz((s->mb_stride+s->mb_height*2+2)*sizeof(*s->macroblocks)); |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
242 s->filter_strength = av_mallocz(s->mb_stride*sizeof(*s->filter_strength)); |
11921 | 243 s->intra4x4_pred_mode_base = av_mallocz(s->b4_stride*(4*s->mb_height+1)); |
244 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
|
245 s->top_border = av_mallocz((s->mb_width+1)*sizeof(*s->top_border)); |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
246 s->segmentation_map = av_mallocz(s->mb_stride*s->mb_height); |
11921 | 247 |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
248 if (!s->macroblocks_base || !s->filter_strength || !s->intra4x4_pred_mode_base || |
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
249 !s->top_nnz || !s->top_border || !s->segmentation_map) |
12169 | 250 return AVERROR(ENOMEM); |
251 | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
252 s->macroblocks = s->macroblocks_base + 1; |
11921 | 253 s->intra4x4_pred_mode = s->intra4x4_pred_mode_base + 4 + s->b4_stride; |
254 | |
255 memset(s->intra4x4_pred_mode_base, DC_PRED, s->b4_stride); | |
256 for (i = 0; i < 4*s->mb_height; i++) | |
257 s->intra4x4_pred_mode[i*s->b4_stride-1] = DC_PRED; | |
258 | |
259 return 0; | |
260 } | |
261 | |
262 static void parse_segment_info(VP8Context *s) | |
263 { | |
264 VP56RangeCoder *c = &s->c; | |
265 int i; | |
266 | |
267 s->segmentation.update_map = vp8_rac_get(c); | |
268 | |
269 if (vp8_rac_get(c)) { // update segment feature data | |
270 s->segmentation.absolute_vals = vp8_rac_get(c); | |
271 | |
272 for (i = 0; i < 4; i++) | |
273 s->segmentation.base_quant[i] = vp8_rac_get_sint(c, 7); | |
274 | |
275 for (i = 0; i < 4; i++) | |
276 s->segmentation.filter_level[i] = vp8_rac_get_sint(c, 6); | |
277 } | |
278 if (s->segmentation.update_map) | |
279 for (i = 0; i < 3; i++) | |
280 s->prob->segmentid[i] = vp8_rac_get(c) ? vp8_rac_get_uint(c, 8) : 255; | |
281 } | |
282 | |
283 static void update_lf_deltas(VP8Context *s) | |
284 { | |
285 VP56RangeCoder *c = &s->c; | |
286 int i; | |
287 | |
288 for (i = 0; i < 4; i++) | |
289 s->lf_delta.ref[i] = vp8_rac_get_sint(c, 6); | |
290 | |
291 for (i = 0; i < 4; i++) | |
292 s->lf_delta.mode[i] = vp8_rac_get_sint(c, 6); | |
293 } | |
294 | |
295 static int setup_partitions(VP8Context *s, const uint8_t *buf, int buf_size) | |
296 { | |
297 const uint8_t *sizes = buf; | |
298 int i; | |
299 | |
300 s->num_coeff_partitions = 1 << vp8_rac_get_uint(&s->c, 2); | |
301 | |
302 buf += 3*(s->num_coeff_partitions-1); | |
303 buf_size -= 3*(s->num_coeff_partitions-1); | |
304 if (buf_size < 0) | |
305 return -1; | |
306 | |
307 for (i = 0; i < s->num_coeff_partitions-1; i++) { | |
12247
50a96623366b
VP8: use AV_RL24 instead of defining a new RL24.
darkshikari
parents:
12246
diff
changeset
|
308 int size = AV_RL24(sizes + 3*i); |
11921 | 309 if (buf_size - size < 0) |
310 return -1; | |
311 | |
312 vp56_init_range_decoder(&s->coeff_partition[i], buf, size); | |
313 buf += size; | |
314 buf_size -= size; | |
315 } | |
316 vp56_init_range_decoder(&s->coeff_partition[i], buf, buf_size); | |
317 | |
318 return 0; | |
319 } | |
320 | |
321 static void get_quants(VP8Context *s) | |
322 { | |
323 VP56RangeCoder *c = &s->c; | |
324 int i, base_qi; | |
325 | |
326 int yac_qi = vp8_rac_get_uint(c, 7); | |
327 int ydc_delta = vp8_rac_get_sint(c, 4); | |
328 int y2dc_delta = vp8_rac_get_sint(c, 4); | |
329 int y2ac_delta = vp8_rac_get_sint(c, 4); | |
330 int uvdc_delta = vp8_rac_get_sint(c, 4); | |
331 int uvac_delta = vp8_rac_get_sint(c, 4); | |
332 | |
333 for (i = 0; i < 4; i++) { | |
334 if (s->segmentation.enabled) { | |
335 base_qi = s->segmentation.base_quant[i]; | |
336 if (!s->segmentation.absolute_vals) | |
337 base_qi += yac_qi; | |
338 } else | |
339 base_qi = yac_qi; | |
340 | |
341 s->qmat[i].luma_qmul[0] = vp8_dc_qlookup[av_clip(base_qi + ydc_delta , 0, 127)]; | |
342 s->qmat[i].luma_qmul[1] = vp8_ac_qlookup[av_clip(base_qi , 0, 127)]; | |
343 s->qmat[i].luma_dc_qmul[0] = 2 * vp8_dc_qlookup[av_clip(base_qi + y2dc_delta, 0, 127)]; | |
344 s->qmat[i].luma_dc_qmul[1] = 155 * vp8_ac_qlookup[av_clip(base_qi + y2ac_delta, 0, 127)] / 100; | |
345 s->qmat[i].chroma_qmul[0] = vp8_dc_qlookup[av_clip(base_qi + uvdc_delta, 0, 127)]; | |
346 s->qmat[i].chroma_qmul[1] = vp8_ac_qlookup[av_clip(base_qi + uvac_delta, 0, 127)]; | |
347 | |
348 s->qmat[i].luma_dc_qmul[1] = FFMAX(s->qmat[i].luma_dc_qmul[1], 8); | |
349 s->qmat[i].chroma_qmul[0] = FFMIN(s->qmat[i].chroma_qmul[0], 132); | |
350 } | |
351 } | |
352 | |
353 /** | |
354 * Determine which buffers golden and altref should be updated with after this frame. | |
355 * The spec isn't clear here, so I'm going by my understanding of what libvpx does | |
356 * | |
357 * Intra frames update all 3 references | |
358 * Inter frames update VP56_FRAME_PREVIOUS if the update_last flag is set | |
359 * If the update (golden|altref) flag is set, it's updated with the current frame | |
360 * if update_last is set, and VP56_FRAME_PREVIOUS otherwise. | |
361 * If the flag is not set, the number read means: | |
362 * 0: no update | |
363 * 1: VP56_FRAME_PREVIOUS | |
364 * 2: update golden with altref, or update altref with golden | |
365 */ | |
366 static VP56Frame ref_to_update(VP8Context *s, int update, VP56Frame ref) | |
367 { | |
368 VP56RangeCoder *c = &s->c; | |
369 | |
370 if (update) | |
371 return VP56_FRAME_CURRENT; | |
372 | |
373 switch (vp8_rac_get_uint(c, 2)) { | |
374 case 1: | |
375 return VP56_FRAME_PREVIOUS; | |
376 case 2: | |
377 return (ref == VP56_FRAME_GOLDEN) ? VP56_FRAME_GOLDEN2 : VP56_FRAME_GOLDEN; | |
378 } | |
379 return VP56_FRAME_NONE; | |
380 } | |
381 | |
382 static void update_refs(VP8Context *s) | |
383 { | |
384 VP56RangeCoder *c = &s->c; | |
385 | |
386 int update_golden = vp8_rac_get(c); | |
387 int update_altref = vp8_rac_get(c); | |
388 | |
389 s->update_golden = ref_to_update(s, update_golden, VP56_FRAME_GOLDEN); | |
390 s->update_altref = ref_to_update(s, update_altref, VP56_FRAME_GOLDEN2); | |
391 } | |
392 | |
393 static int decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size) | |
394 { | |
395 VP56RangeCoder *c = &s->c; | |
396 int header_size, hscale, vscale, i, j, k, l, ret; | |
397 int width = s->avctx->width; | |
398 int height = s->avctx->height; | |
399 | |
400 s->keyframe = !(buf[0] & 1); | |
401 s->profile = (buf[0]>>1) & 7; | |
402 s->invisible = !(buf[0] & 0x10); | |
12247
50a96623366b
VP8: use AV_RL24 instead of defining a new RL24.
darkshikari
parents:
12246
diff
changeset
|
403 header_size = AV_RL24(buf) >> 5; |
11921 | 404 buf += 3; |
405 buf_size -= 3; | |
406 | |
11974 | 407 if (s->profile > 3) |
408 av_log(s->avctx, AV_LOG_WARNING, "Unknown profile %d\n", s->profile); | |
409 | |
410 if (!s->profile) | |
411 memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_epel_pixels_tab, sizeof(s->put_pixels_tab)); | |
412 else // profile 1-3 use bilinear, 4+ aren't defined so whatever | |
413 memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_bilinear_pixels_tab, sizeof(s->put_pixels_tab)); | |
11921 | 414 |
415 if (header_size > buf_size - 7*s->keyframe) { | |
416 av_log(s->avctx, AV_LOG_ERROR, "Header size larger than data provided\n"); | |
417 return AVERROR_INVALIDDATA; | |
418 } | |
419 | |
420 if (s->keyframe) { | |
12247
50a96623366b
VP8: use AV_RL24 instead of defining a new RL24.
darkshikari
parents:
12246
diff
changeset
|
421 if (AV_RL24(buf) != 0x2a019d) { |
50a96623366b
VP8: use AV_RL24 instead of defining a new RL24.
darkshikari
parents:
12246
diff
changeset
|
422 av_log(s->avctx, AV_LOG_ERROR, "Invalid start code 0x%x\n", AV_RL24(buf)); |
11921 | 423 return AVERROR_INVALIDDATA; |
424 } | |
425 width = AV_RL16(buf+3) & 0x3fff; | |
426 height = AV_RL16(buf+5) & 0x3fff; | |
427 hscale = buf[4] >> 6; | |
428 vscale = buf[6] >> 6; | |
429 buf += 7; | |
430 buf_size -= 7; | |
431 | |
11970
c7953ee47af4
vp8: warn and request sample if upscaling specified in header
mru
parents:
11950
diff
changeset
|
432 if (hscale || vscale) |
c7953ee47af4
vp8: warn and request sample if upscaling specified in header
mru
parents:
11950
diff
changeset
|
433 av_log_missing_feature(s->avctx, "Upscaling", 1); |
c7953ee47af4
vp8: warn and request sample if upscaling specified in header
mru
parents:
11950
diff
changeset
|
434 |
11921 | 435 s->update_golden = s->update_altref = VP56_FRAME_CURRENT; |
436 memcpy(s->prob->token , vp8_token_default_probs , sizeof(s->prob->token)); | |
437 memcpy(s->prob->pred16x16, vp8_pred16x16_prob_inter, sizeof(s->prob->pred16x16)); | |
438 memcpy(s->prob->pred8x8c , vp8_pred8x8c_prob_inter , sizeof(s->prob->pred8x8c)); | |
439 memcpy(s->prob->mvc , vp8_mv_default_prob , sizeof(s->prob->mvc)); | |
440 memset(&s->segmentation, 0, sizeof(s->segmentation)); | |
441 } | |
442 | |
443 if (!s->macroblocks_base || /* first frame */ | |
444 width != s->avctx->width || height != s->avctx->height) { | |
445 if ((ret = update_dimensions(s, width, height) < 0)) | |
446 return ret; | |
447 } | |
448 | |
449 vp56_init_range_decoder(c, buf, header_size); | |
450 buf += header_size; | |
451 buf_size -= header_size; | |
452 | |
453 if (s->keyframe) { | |
454 if (vp8_rac_get(c)) | |
455 av_log(s->avctx, AV_LOG_WARNING, "Unspecified colorspace\n"); | |
456 vp8_rac_get(c); // whether we can skip clamping in dsp functions | |
457 } | |
458 | |
459 if ((s->segmentation.enabled = vp8_rac_get(c))) | |
460 parse_segment_info(s); | |
461 else | |
462 s->segmentation.update_map = 0; // FIXME: move this to some init function? | |
463 | |
464 s->filter.simple = vp8_rac_get(c); | |
465 s->filter.level = vp8_rac_get_uint(c, 6); | |
466 s->filter.sharpness = vp8_rac_get_uint(c, 3); | |
467 | |
468 if ((s->lf_delta.enabled = vp8_rac_get(c))) | |
469 if (vp8_rac_get(c)) | |
470 update_lf_deltas(s); | |
471 | |
472 if (setup_partitions(s, buf, buf_size)) { | |
473 av_log(s->avctx, AV_LOG_ERROR, "Invalid partitions\n"); | |
474 return AVERROR_INVALIDDATA; | |
475 } | |
476 | |
477 get_quants(s); | |
478 | |
479 if (!s->keyframe) { | |
480 update_refs(s); | |
481 s->sign_bias[VP56_FRAME_GOLDEN] = vp8_rac_get(c); | |
482 s->sign_bias[VP56_FRAME_GOLDEN2 /* altref */] = vp8_rac_get(c); | |
483 } | |
484 | |
485 // if we aren't saving this frame's probabilities for future frames, | |
486 // make a copy of the current probabilities | |
487 if (!(s->update_probabilities = vp8_rac_get(c))) | |
488 s->prob[1] = s->prob[0]; | |
489 | |
490 s->update_last = s->keyframe || vp8_rac_get(c); | |
491 | |
492 for (i = 0; i < 4; i++) | |
493 for (j = 0; j < 8; j++) | |
494 for (k = 0; k < 3; k++) | |
495 for (l = 0; l < NUM_DCT_TOKENS-1; l++) | |
496 if (vp56_rac_get_prob(c, vp8_token_update_probs[i][j][k][l])) | |
497 s->prob->token[i][j][k][l] = vp8_rac_get_uint(c, 8); | |
498 | |
499 if ((s->mbskip_enabled = vp8_rac_get(c))) | |
500 s->prob->mbskip = vp8_rac_get_uint(c, 8); | |
501 | |
502 if (!s->keyframe) { | |
503 s->prob->intra = vp8_rac_get_uint(c, 8); | |
504 s->prob->last = vp8_rac_get_uint(c, 8); | |
505 s->prob->golden = vp8_rac_get_uint(c, 8); | |
506 | |
507 if (vp8_rac_get(c)) | |
508 for (i = 0; i < 4; i++) | |
509 s->prob->pred16x16[i] = vp8_rac_get_uint(c, 8); | |
510 if (vp8_rac_get(c)) | |
511 for (i = 0; i < 3; i++) | |
512 s->prob->pred8x8c[i] = vp8_rac_get_uint(c, 8); | |
513 | |
514 // 17.2 MV probability update | |
515 for (i = 0; i < 2; i++) | |
516 for (j = 0; j < 19; j++) | |
517 if (vp56_rac_get_prob(c, vp8_mv_update_prob[i][j])) | |
518 s->prob->mvc[i][j] = vp8_rac_get_nn(c); | |
519 } | |
520 | |
521 return 0; | |
522 } | |
523 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
524 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
525 void clamp_mv(VP8Context *s, VP56mv *dst, const VP56mv *src, int mb_x, int mb_y) |
11921 | 526 { |
527 #define MARGIN (16 << 2) | |
528 dst->x = av_clip(src->x, -((mb_x << 6) + MARGIN), | |
529 ((s->mb_width - 1 - mb_x) << 6) + MARGIN); | |
530 dst->y = av_clip(src->y, -((mb_y << 6) + MARGIN), | |
531 ((s->mb_height - 1 - mb_y) << 6) + MARGIN); | |
532 } | |
533 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
534 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
535 void find_near_mvs(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
536 VP56mv near[2], VP56mv *best, uint8_t cnt[4]) |
11921 | 537 { |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
538 VP8Macroblock *mb_edge[3] = { mb + 2 /* top */, |
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
539 mb - 1 /* left */, |
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
540 mb + 1 /* top-left */ }; |
11921 | 541 enum { EDGE_TOP, EDGE_LEFT, EDGE_TOPLEFT }; |
542 VP56mv near_mv[4] = {{ 0 }}; | |
543 enum { CNT_ZERO, CNT_NEAREST, CNT_NEAR, CNT_SPLITMV }; | |
12217 | 544 int idx = CNT_ZERO; |
11921 | 545 int best_idx = CNT_ZERO; |
12217 | 546 int cur_sign_bias = s->sign_bias[mb->ref_frame]; |
547 int *sign_bias = s->sign_bias; | |
11921 | 548 |
549 /* Process MB on top, left and top-left */ | |
12217 | 550 #define MV_EDGE_CHECK(n)\ |
551 {\ | |
552 VP8Macroblock *edge = mb_edge[n];\ | |
553 int edge_ref = edge->ref_frame;\ | |
554 if (edge_ref != VP56_FRAME_CURRENT) {\ | |
555 uint32_t mv = AV_RN32A(&edge->mv);\ | |
556 if (mv) {\ | |
557 if (cur_sign_bias != sign_bias[edge_ref]) {\ | |
558 /* SWAR negate of the values in mv. */\ | |
12242 | 559 mv = ~mv;\ |
560 mv = ((mv&0x7fff7fff) + 0x00010001) ^ (mv&0x80008000);\ | |
12217 | 561 }\ |
562 if (!n || mv != AV_RN32A(&near_mv[idx]))\ | |
563 AV_WN32A(&near_mv[++idx], mv);\ | |
564 cnt[idx] += 1 + (n != 2);\ | |
565 } else\ | |
566 cnt[CNT_ZERO] += 1 + (n != 2);\ | |
567 }\ | |
11921 | 568 } |
12217 | 569 MV_EDGE_CHECK(0) |
570 MV_EDGE_CHECK(1) | |
571 MV_EDGE_CHECK(2) | |
11921 | 572 |
12217 | 573 /* If we have three distinct MVs, merge first and last if they're the same */ |
574 if (cnt[CNT_SPLITMV] && AV_RN32A(&near_mv[1+EDGE_TOP]) == AV_RN32A(&near_mv[1+EDGE_TOPLEFT])) | |
11921 | 575 cnt[CNT_NEAREST] += 1; |
576 | |
577 cnt[CNT_SPLITMV] = ((mb_edge[EDGE_LEFT]->mode == VP8_MVMODE_SPLIT) + | |
578 (mb_edge[EDGE_TOP]->mode == VP8_MVMODE_SPLIT)) * 2 + | |
579 (mb_edge[EDGE_TOPLEFT]->mode == VP8_MVMODE_SPLIT); | |
580 | |
581 /* Swap near and nearest if necessary */ | |
582 if (cnt[CNT_NEAR] > cnt[CNT_NEAREST]) { | |
12217 | 583 FFSWAP(uint8_t, cnt[CNT_NEAREST], cnt[CNT_NEAR]); |
584 FFSWAP( VP56mv, near_mv[CNT_NEAREST], near_mv[CNT_NEAR]); | |
11921 | 585 } |
586 | |
587 /* Choose the best mv out of 0,0 and the nearest mv */ | |
588 if (cnt[CNT_NEAREST] >= cnt[CNT_ZERO]) | |
589 best_idx = CNT_NEAREST; | |
590 | |
12246 | 591 mb->mv = near_mv[best_idx]; |
11921 | 592 near[0] = near_mv[CNT_NEAREST]; |
593 near[1] = near_mv[CNT_NEAR]; | |
594 } | |
595 | |
596 /** | |
597 * Motion vector coding, 17.1. | |
598 */ | |
599 static int read_mv_component(VP56RangeCoder *c, const uint8_t *p) | |
600 { | |
601 int x = 0; | |
602 | |
603 if (vp56_rac_get_prob(c, p[0])) { | |
604 int i; | |
605 | |
606 for (i = 0; i < 3; i++) | |
607 x += vp56_rac_get_prob(c, p[9 + i]) << i; | |
608 for (i = 9; i > 3; i--) | |
609 x += vp56_rac_get_prob(c, p[9 + i]) << i; | |
610 if (!(x & 0xFFF0) || vp56_rac_get_prob(c, p[12])) | |
611 x += 8; | |
612 } else | |
613 x = vp8_rac_get_tree(c, vp8_small_mvtree, &p[2]); | |
614 | |
615 return (x && vp56_rac_get_prob(c, p[1])) ? -x : x; | |
616 } | |
617 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
618 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
619 const uint8_t *get_submv_prob(uint32_t left, uint32_t top) |
11921 | 620 { |
12219 | 621 if (left == top) |
622 return vp8_submv_prob[4-!!left]; | |
623 if (!top) | |
11921 | 624 return vp8_submv_prob[2]; |
12219 | 625 return vp8_submv_prob[1-!!left]; |
11921 | 626 } |
627 | |
628 /** | |
629 * 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
|
630 * @returns the number of motion vectors parsed (2, 4 or 16) |
11921 | 631 */ |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
632 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
633 int decode_splitmvs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb) |
11921 | 634 { |
635 int part_idx = mb->partitioning = | |
636 vp8_rac_get_tree(c, vp8_mbsplit_tree, vp8_mbsplit_prob); | |
637 int n, num = vp8_mbsplit_count[part_idx]; | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
638 VP8Macroblock *top_mb = &mb[2]; |
12219 | 639 VP8Macroblock *left_mb = &mb[-1]; |
640 const uint8_t *mbsplits_left = vp8_mbsplits[left_mb->partitioning], | |
641 *mbsplits_top = vp8_mbsplits[top_mb->partitioning], | |
642 *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
|
643 *firstidx = vp8_mbfirstidx[part_idx]; |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
644 VP56mv *top_mv = top_mb->bmv; |
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
645 VP56mv *left_mv = left_mb->bmv; |
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
646 VP56mv *cur_mv = mb->bmv; |
11921 | 647 |
648 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
|
649 int k = firstidx[n]; |
12219 | 650 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
|
651 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
|
652 |
12219 | 653 if (!(k & 3)) |
654 left = AV_RN32A(&left_mv[mbsplits_left[k + 3]]); | |
655 else | |
656 left = AV_RN32A(&cur_mv[mbsplits_cur[k - 1]]); | |
657 if (k <= 3) | |
658 above = AV_RN32A(&top_mv[mbsplits_top[k + 12]]); | |
659 else | |
660 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
|
661 |
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
662 submv_prob = get_submv_prob(left, above); |
11921 | 663 |
664 switch (vp8_rac_get_tree(c, vp8_submv_ref_tree, submv_prob)) { | |
665 case VP8_SUBMVMODE_NEW4X4: | |
12246 | 666 mb->bmv[n].y = mb->mv.y + read_mv_component(c, s->prob->mvc[0]); |
667 mb->bmv[n].x = mb->mv.x + read_mv_component(c, s->prob->mvc[1]); | |
11921 | 668 break; |
669 case VP8_SUBMVMODE_ZERO4X4: | |
12245
ca82c3ce90c1
VP8: use AV_ZERO32 instead of AV_WN32A where relevant
darkshikari
parents:
12244
diff
changeset
|
670 AV_ZERO32(&mb->bmv[n]); |
11921 | 671 break; |
672 case VP8_SUBMVMODE_LEFT4X4: | |
12219 | 673 AV_WN32A(&mb->bmv[n], left); |
11921 | 674 break; |
675 case VP8_SUBMVMODE_TOP4X4: | |
12219 | 676 AV_WN32A(&mb->bmv[n], above); |
11921 | 677 break; |
678 } | |
11990
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
679 } |
11921 | 680 |
11990
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
681 return num; |
11921 | 682 } |
683 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
684 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
685 void decode_intra4x4_modes(VP56RangeCoder *c, uint8_t *intra4x4, |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
686 int stride, int keyframe) |
11921 | 687 { |
12221 | 688 int x, y, t, l, i; |
11921 | 689 |
12221 | 690 if (keyframe) { |
691 const uint8_t *ctx; | |
692 for (y = 0; y < 4; y++) { | |
693 for (x = 0; x < 4; x++) { | |
11921 | 694 t = intra4x4[x - stride]; |
695 l = intra4x4[x - 1]; | |
696 ctx = vp8_pred4x4_prob_intra[t][l]; | |
12221 | 697 intra4x4[x] = vp8_rac_get_tree(c, vp8_pred4x4_tree, ctx); |
11921 | 698 } |
12221 | 699 intra4x4 += stride; |
11921 | 700 } |
12221 | 701 } else { |
702 for (i = 0; i < 16; i++) | |
703 intra4x4[i] = vp8_rac_get_tree(c, vp8_pred4x4_tree, vp8_pred4x4_prob_inter); | |
11921 | 704 } |
705 } | |
706 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
707 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
708 void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
709 uint8_t *intra4x4, uint8_t *segment) |
11921 | 710 { |
711 VP56RangeCoder *c = &s->c; | |
712 | |
713 if (s->segmentation.update_map) | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
714 *segment = vp8_rac_get_tree(c, vp8_segmentid_tree, s->prob->segmentid); |
12224
5b7d690b761b
VP8: Don't store segment in macroblock struct anymore.
darkshikari
parents:
12223
diff
changeset
|
715 s->segment = *segment; |
11921 | 716 |
717 mb->skip = s->mbskip_enabled ? vp56_rac_get_prob(c, s->prob->mbskip) : 0; | |
718 | |
719 if (s->keyframe) { | |
720 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_intra, vp8_pred16x16_prob_intra); | |
721 | |
722 if (mb->mode == MODE_I4x4) { | |
723 decode_intra4x4_modes(c, intra4x4, s->b4_stride, 1); | |
724 } else | |
725 fill_rectangle(intra4x4, 4, 4, s->b4_stride, vp8_pred4x4_mode[mb->mode], 1); | |
726 | |
727 s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, vp8_pred8x8c_prob_intra); | |
728 mb->ref_frame = VP56_FRAME_CURRENT; | |
729 } else if (vp56_rac_get_prob(c, s->prob->intra)) { | |
730 VP56mv near[2], best; | |
12217 | 731 uint8_t cnt[4] = { 0 }; |
11921 | 732 uint8_t p[4]; |
733 | |
734 // inter MB, 16.2 | |
735 if (vp56_rac_get_prob(c, s->prob->last)) | |
736 mb->ref_frame = vp56_rac_get_prob(c, s->prob->golden) ? | |
737 VP56_FRAME_GOLDEN2 /* altref */ : VP56_FRAME_GOLDEN; | |
738 else | |
739 mb->ref_frame = VP56_FRAME_PREVIOUS; | |
12231 | 740 s->ref_count[mb->ref_frame-1]++; |
11921 | 741 |
742 // motion vectors, 16.3 | |
743 find_near_mvs(s, mb, mb_x, mb_y, near, &best, cnt); | |
12217 | 744 p[0] = vp8_mode_contexts[cnt[0]][0]; |
745 p[1] = vp8_mode_contexts[cnt[1]][1]; | |
746 p[2] = vp8_mode_contexts[cnt[2]][2]; | |
747 p[3] = vp8_mode_contexts[cnt[3]][3]; | |
11921 | 748 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_mvinter, p); |
749 switch (mb->mode) { | |
750 case VP8_MVMODE_SPLIT: | |
12246 | 751 clamp_mv(s, &mb->mv, &mb->mv, mb_x, mb_y); |
752 mb->mv = mb->bmv[decode_splitmvs(s, c, mb) - 1]; | |
11921 | 753 break; |
754 case VP8_MVMODE_ZERO: | |
12245
ca82c3ce90c1
VP8: use AV_ZERO32 instead of AV_WN32A where relevant
darkshikari
parents:
12244
diff
changeset
|
755 AV_ZERO32(&mb->mv); |
11921 | 756 break; |
757 case VP8_MVMODE_NEAREST: | |
758 clamp_mv(s, &mb->mv, &near[0], mb_x, mb_y); | |
759 break; | |
760 case VP8_MVMODE_NEAR: | |
761 clamp_mv(s, &mb->mv, &near[1], mb_x, mb_y); | |
762 break; | |
763 case VP8_MVMODE_NEW: | |
12246 | 764 clamp_mv(s, &mb->mv, &mb->mv, mb_x, mb_y); |
765 mb->mv.y += + read_mv_component(c, s->prob->mvc[0]); | |
766 mb->mv.x += + read_mv_component(c, s->prob->mvc[1]); | |
11921 | 767 break; |
768 } | |
769 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
|
770 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
|
771 mb->bmv[0] = mb->mv; |
11921 | 772 } |
773 } else { | |
774 // intra MB, 16.1 | |
775 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_inter, s->prob->pred16x16); | |
776 | |
12220
0f635b1f7861
Avoid useless fill_rectangle in P-frames in VP8
darkshikari
parents:
12219
diff
changeset
|
777 if (mb->mode == MODE_I4x4) |
12221 | 778 decode_intra4x4_modes(c, intra4x4, 4, 0); |
11921 | 779 |
780 s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, s->prob->pred8x8c); | |
781 mb->ref_frame = VP56_FRAME_CURRENT; | |
12225
c3e11b3108d7
Eliminate a pointless memset for intra blocks in P-frames in VP8
darkshikari
parents:
12224
diff
changeset
|
782 mb->partitioning = VP8_SPLITMVMODE_NONE; |
12245
ca82c3ce90c1
VP8: use AV_ZERO32 instead of AV_WN32A where relevant
darkshikari
parents:
12244
diff
changeset
|
783 AV_ZERO32(&mb->bmv[0]); |
11921 | 784 } |
785 } | |
786 | |
787 /** | |
12115 | 788 * @param c arithmetic bitstream reader context |
789 * @param block destination for block coefficients | |
790 * @param probs probabilities to use when reading trees from the bitstream | |
11921 | 791 * @param i initial coeff index, 0 unless a separate DC block is coded |
792 * @param zero_nhood the initial prediction context for number of surrounding | |
793 * 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
|
794 * @param qmul array holding the dc/ac dequant factor at position 0/1 |
11921 | 795 * @return 0 if no coeffs were decoded |
796 * otherwise, the index of the last coeff decoded plus one | |
797 */ | |
798 static int decode_block_coeffs(VP56RangeCoder *c, DCTELEM block[16], | |
799 uint8_t probs[8][3][NUM_DCT_TOKENS-1], | |
800 int i, int zero_nhood, int16_t qmul[2]) | |
801 { | |
802 int token, nonzero = 0; | |
803 int offset = 0; | |
804 | |
805 for (; i < 16; i++) { | |
806 token = vp8_rac_get_tree_with_offset(c, vp8_coeff_tree, probs[vp8_coeff_band[i]][zero_nhood], offset); | |
807 | |
808 if (token == DCT_EOB) | |
809 break; | |
810 else if (token >= DCT_CAT1) { | |
811 int cat = token-DCT_CAT1; | |
812 token = vp8_rac_get_coeff(c, vp8_dct_cat_prob[cat]); | |
12229
cf675a4d5195
Eliminate a LUT in escape decoding in VP8 decode_block_coeffs
darkshikari
parents:
12228
diff
changeset
|
813 token += 3 + (2<<cat); |
11921 | 814 } |
815 | |
816 // after the first token, the non-zero prediction context becomes | |
817 // based on the last decoded coeff | |
818 if (!token) { | |
819 zero_nhood = 0; | |
820 offset = 1; | |
821 continue; | |
822 } else if (token == 1) | |
823 zero_nhood = 1; | |
824 else | |
825 zero_nhood = 2; | |
826 | |
827 // todo: full [16] qmat? load into register? | |
828 block[zigzag_scan[i]] = (vp8_rac_get(c) ? -token : token) * qmul[!!i]; | |
829 nonzero = i+1; | |
830 offset = 0; | |
831 } | |
832 return nonzero; | |
833 } | |
834 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
835 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
836 void decode_mb_coeffs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb, |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
837 uint8_t t_nnz[9], uint8_t l_nnz[9]) |
11921 | 838 { |
839 LOCAL_ALIGNED_16(DCTELEM, dc,[16]); | |
840 int i, x, y, luma_start = 0, luma_ctx = 3; | |
841 int nnz_pred, nnz, nnz_total = 0; | |
12224
5b7d690b761b
VP8: Don't store segment in macroblock struct anymore.
darkshikari
parents:
12223
diff
changeset
|
842 int segment = s->segment; |
11921 | 843 |
844 if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) { | |
845 AV_ZERO128(dc); | |
846 AV_ZERO128(dc+8); | |
847 nnz_pred = t_nnz[8] + l_nnz[8]; | |
848 | |
849 // decode DC values and do hadamard | |
850 nnz = decode_block_coeffs(c, dc, s->prob->token[1], 0, nnz_pred, | |
851 s->qmat[segment].luma_dc_qmul); | |
852 l_nnz[8] = t_nnz[8] = !!nnz; | |
853 nnz_total += nnz; | |
854 s->vp8dsp.vp8_luma_dc_wht(s->block, dc); | |
855 luma_start = 1; | |
856 luma_ctx = 0; | |
857 } | |
858 | |
859 // luma blocks | |
860 for (y = 0; y < 4; y++) | |
861 for (x = 0; x < 4; x++) { | |
862 nnz_pred = l_nnz[y] + t_nnz[x]; | |
863 nnz = decode_block_coeffs(c, s->block[y][x], s->prob->token[luma_ctx], luma_start, | |
864 nnz_pred, s->qmat[segment].luma_qmul); | |
865 // nnz+luma_start may be one more than the actual last index, but we don't care | |
866 s->non_zero_count_cache[y][x] = nnz + luma_start; | |
867 t_nnz[x] = l_nnz[y] = !!nnz; | |
868 nnz_total += nnz; | |
869 } | |
870 | |
871 // chroma blocks | |
872 // TODO: what to do about dimensions? 2nd dim for luma is x, | |
873 // but for chroma it's (y<<1)|x | |
874 for (i = 4; i < 6; i++) | |
875 for (y = 0; y < 2; y++) | |
876 for (x = 0; x < 2; x++) { | |
877 nnz_pred = l_nnz[i+2*y] + t_nnz[i+2*x]; | |
878 nnz = decode_block_coeffs(c, s->block[i][(y<<1)+x], s->prob->token[2], 0, | |
879 nnz_pred, s->qmat[segment].chroma_qmul); | |
880 s->non_zero_count_cache[i][(y<<1)+x] = nnz; | |
881 t_nnz[i+2*x] = l_nnz[i+2*y] = !!nnz; | |
882 nnz_total += nnz; | |
883 } | |
884 | |
885 // if there were no coded coeffs despite the macroblock not being marked skip, | |
886 // we MUST not do the inner loop filter and should not do IDCT | |
887 // Since skip isn't used for bitstream prediction, just manually set it. | |
888 if (!nnz_total) | |
889 mb->skip = 1; | |
890 } | |
891 | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
892 static av_always_inline |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
893 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
|
894 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
|
895 { |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
896 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
|
897 if (!simple) { |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
898 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
|
899 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
|
900 } |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
901 } |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
902 |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
903 static av_always_inline |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
904 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
|
905 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
|
906 int simple, int xchg) |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
907 { |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
908 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
|
909 src_y -= linesize; |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
910 src_cb -= uvlinesize; |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
911 src_cr -= uvlinesize; |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
912 |
12202 | 913 #define XCHG(a,b,xchg) do { \ |
914 if (xchg) AV_SWAP64(b,a); \ | |
915 else AV_COPY64(b,a); \ | |
916 } while (0) | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
917 |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
918 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
|
919 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
|
920 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
|
921 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
|
922 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
|
923 |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
924 // 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
|
925 // 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
|
926 if (!simple || !mb_y) { |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
927 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
|
928 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
|
929 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
|
930 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
|
931 } |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
932 } |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
933 |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
934 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
935 int check_intra_pred_mode(int mode, int mb_x, int mb_y) |
11921 | 936 { |
937 if (mode == DC_PRED8x8) { | |
12243
788445bf10c0
VP8: shave a few clocks off check_intra_pred_mode
darkshikari
parents:
12242
diff
changeset
|
938 if (!mb_x) { |
788445bf10c0
VP8: shave a few clocks off check_intra_pred_mode
darkshikari
parents:
12242
diff
changeset
|
939 mode = mb_y ? TOP_DC_PRED8x8 : DC_128_PRED8x8; |
788445bf10c0
VP8: shave a few clocks off check_intra_pred_mode
darkshikari
parents:
12242
diff
changeset
|
940 } else if (!mb_y) { |
12244 | 941 mode = LEFT_DC_PRED8x8; |
12243
788445bf10c0
VP8: shave a few clocks off check_intra_pred_mode
darkshikari
parents:
12242
diff
changeset
|
942 } |
11921 | 943 } |
944 return mode; | |
945 } | |
946 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
947 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
948 void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb, |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
949 uint8_t *intra4x4, int mb_x, int mb_y) |
11921 | 950 { |
951 int x, y, mode, nnz, tr; | |
952 | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
953 // 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
|
954 // 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
|
955 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
|
956 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
|
957 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
|
958 s->filter.simple, 1); |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
959 |
11921 | 960 if (mb->mode < MODE_I4x4) { |
961 mode = check_intra_pred_mode(mb->mode, mb_x, mb_y); | |
962 s->hpc.pred16x16[mode](dst[0], s->linesize); | |
963 } else { | |
964 uint8_t *ptr = dst[0]; | |
12221 | 965 int stride = s->keyframe ? s->b4_stride : 4; |
11921 | 966 |
967 // all blocks on the right edge of the macroblock use bottom edge | |
968 // the top macroblock for their topright edge | |
969 uint8_t *tr_right = ptr - s->linesize + 16; | |
970 | |
971 // if we're on the right edge of the frame, said edge is extended | |
972 // from the top macroblock | |
973 if (mb_x == s->mb_width-1) { | |
974 tr = tr_right[-1]*0x01010101; | |
975 tr_right = (uint8_t *)&tr; | |
976 } | |
977 | |
12234
bba849c2a113
VP8: avoid a memset for non-i4x4 blocks with no coefficients
darkshikari
parents:
12233
diff
changeset
|
978 if (mb->skip) |
bba849c2a113
VP8: avoid a memset for non-i4x4 blocks with no coefficients
darkshikari
parents:
12233
diff
changeset
|
979 AV_ZERO128(s->non_zero_count_cache); |
bba849c2a113
VP8: avoid a memset for non-i4x4 blocks with no coefficients
darkshikari
parents:
12233
diff
changeset
|
980 |
11921 | 981 for (y = 0; y < 4; y++) { |
982 uint8_t *topright = ptr + 4 - s->linesize; | |
983 for (x = 0; x < 4; x++) { | |
984 if (x == 3) | |
985 topright = tr_right; | |
986 | |
12221 | 987 s->hpc.pred4x4[intra4x4[x]](ptr+4*x, topright, s->linesize); |
11921 | 988 |
989 nnz = s->non_zero_count_cache[y][x]; | |
990 if (nnz) { | |
991 if (nnz == 1) | |
992 s->vp8dsp.vp8_idct_dc_add(ptr+4*x, s->block[y][x], s->linesize); | |
993 else | |
994 s->vp8dsp.vp8_idct_add(ptr+4*x, s->block[y][x], s->linesize); | |
995 } | |
996 topright += 4; | |
997 } | |
998 | |
999 ptr += 4*s->linesize; | |
12221 | 1000 intra4x4 += stride; |
11921 | 1001 } |
1002 } | |
1003 | |
1004 mode = check_intra_pred_mode(s->chroma_pred_mode, mb_x, mb_y); | |
1005 s->hpc.pred8x8[mode](dst[1], s->uvlinesize); | |
1006 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
|
1007 |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
1008 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
|
1009 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
|
1010 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
|
1011 s->filter.simple, 0); |
11921 | 1012 } |
1013 | |
1014 /** | |
1015 * Generic MC function. | |
1016 * | |
1017 * @param s VP8 decoding context | |
1018 * @param luma 1 for luma (Y) planes, 0 for chroma (Cb/Cr) planes | |
1019 * @param dst target buffer for block data at block position | |
1020 * @param src reference picture buffer at origin (0, 0) | |
1021 * @param mv motion vector (relative to block position) to get pixel data from | |
1022 * @param x_off horizontal position of block from origin (0, 0) | |
1023 * @param y_off vertical position of block from origin (0, 0) | |
1024 * @param block_w width of block (16, 8 or 4) | |
1025 * @param block_h height of block (always same as block_w) | |
1026 * @param width width of src/dst plane data | |
1027 * @param height height of src/dst plane data | |
1028 * @param linesize size of a single line of plane data, including padding | |
12115 | 1029 * @param mc_func motion compensation function pointers (bilinear or sixtap MC) |
11921 | 1030 */ |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1031 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1032 void vp8_mc(VP8Context *s, int luma, |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1033 uint8_t *dst, uint8_t *src, const VP56mv *mv, |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1034 int x_off, int y_off, int block_w, int block_h, |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1035 int width, int height, int linesize, |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1036 vp8_mc_func mc_func[3][3]) |
11921 | 1037 { |
12218 | 1038 if (AV_RN32A(mv)) { |
1039 static const uint8_t idx[8] = { 0, 1, 2, 1, 2, 1, 2, 1 }; | |
1040 int mx = (mv->x << luma)&7, mx_idx = idx[mx]; | |
1041 int my = (mv->y << luma)&7, my_idx = idx[my]; | |
11921 | 1042 |
12218 | 1043 x_off += mv->x >> (3 - luma); |
1044 y_off += mv->y >> (3 - luma); | |
11921 | 1045 |
12218 | 1046 // edge emulation |
1047 src += y_off * linesize + x_off; | |
1048 if (x_off < 2 || x_off >= width - block_w - 3 || | |
1049 y_off < 2 || y_off >= height - block_h - 3) { | |
1050 ff_emulated_edge_mc(s->edge_emu_buffer, src - 2 * linesize - 2, linesize, | |
1051 block_w + 5, block_h + 5, | |
1052 x_off - 2, y_off - 2, width, height); | |
1053 src = s->edge_emu_buffer + 2 + linesize * 2; | |
1054 } | |
1055 mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my); | |
1056 } else | |
1057 mc_func[0][0](dst, linesize, src + y_off * linesize + x_off, linesize, block_h, 0, 0); | |
11921 | 1058 } |
1059 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1060 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1061 void vp8_mc_part(VP8Context *s, uint8_t *dst[3], |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1062 AVFrame *ref_frame, int x_off, int y_off, |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1063 int bx_off, int by_off, |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1064 int block_w, int block_h, |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1065 int width, int height, VP56mv *mv) |
11989
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1066 { |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1067 VP56mv uvmv = *mv; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1068 |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1069 /* Y */ |
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, 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
|
1071 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
|
1072 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
|
1073 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
|
1074 |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1075 /* U/V */ |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1076 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
|
1077 uvmv.x &= ~7; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1078 uvmv.y &= ~7; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1079 } |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1080 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
|
1081 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
|
1082 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
|
1083 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
|
1084 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
|
1085 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
|
1086 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
|
1087 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
|
1088 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
|
1089 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
|
1090 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
|
1091 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
|
1092 } |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1093 |
12215 | 1094 /* Fetch pixels for estimated mv 4 macroblocks ahead. |
1095 * Optimized for 64-byte cache lines. Inspired by ffh264 prefetch_motion. */ | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1096 static av_always_inline void prefetch_motion(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, int mb_xy, int ref) |
12215 | 1097 { |
12237 | 1098 /* Don't prefetch refs that haven't been used very often this frame. */ |
1099 if (s->ref_count[ref-1] > (mb_xy >> 5)) { | |
12231 | 1100 int x_off = mb_x << 4, y_off = mb_y << 4; |
1101 int mx = mb->mv.x + x_off + 8; | |
1102 int my = mb->mv.y + y_off; | |
1103 uint8_t **src= s->framep[ref]->data; | |
1104 int off= mx + (my + (mb_x&3)*4)*s->linesize + 64; | |
1105 s->dsp.prefetch(src[0]+off, s->linesize, 4); | |
1106 off= (mx>>1) + ((my>>1) + (mb_x&7))*s->uvlinesize + 64; | |
1107 s->dsp.prefetch(src[1]+off, src[2]-src[1], 2); | |
1108 } | |
12215 | 1109 } |
1110 | |
11921 | 1111 /** |
1112 * Apply motion vectors to prediction buffer, chapter 18. | |
1113 */ | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1114 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1115 void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb, |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1116 int mb_x, int mb_y) |
11921 | 1117 { |
1118 int x_off = mb_x << 4, y_off = mb_y << 4; | |
1119 int width = 16*s->mb_width, height = 16*s->mb_height; | |
12228
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1120 AVFrame *ref = s->framep[mb->ref_frame]; |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1121 VP56mv *bmv = mb->bmv; |
11921 | 1122 |
1123 if (mb->mode < VP8_MVMODE_SPLIT) { | |
12228
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1124 vp8_mc_part(s, dst, ref, x_off, y_off, |
11989
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1125 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
|
1126 } 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
|
1127 case VP8_SPLITMVMODE_4x4: { |
11921 | 1128 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
|
1129 VP56mv uvmv; |
11921 | 1130 |
1131 /* Y */ | |
1132 for (y = 0; y < 4; y++) { | |
1133 for (x = 0; x < 4; x++) { | |
1134 vp8_mc(s, 1, dst[0] + 4*y*s->linesize + x*4, | |
12228
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1135 ref->data[0], &bmv[4*y + x], |
11921 | 1136 4*x + x_off, 4*y + y_off, 4, 4, |
1137 width, height, s->linesize, | |
11974 | 1138 s->put_pixels_tab[2]); |
11921 | 1139 } |
1140 } | |
1141 | |
1142 /* U/V */ | |
1143 x_off >>= 1; y_off >>= 1; width >>= 1; height >>= 1; | |
1144 for (y = 0; y < 2; y++) { | |
1145 for (x = 0; x < 2; x++) { | |
1146 uvmv.x = mb->bmv[ 2*y * 4 + 2*x ].x + | |
1147 mb->bmv[ 2*y * 4 + 2*x+1].x + | |
1148 mb->bmv[(2*y+1) * 4 + 2*x ].x + | |
1149 mb->bmv[(2*y+1) * 4 + 2*x+1].x; | |
1150 uvmv.y = mb->bmv[ 2*y * 4 + 2*x ].y + | |
1151 mb->bmv[ 2*y * 4 + 2*x+1].y + | |
1152 mb->bmv[(2*y+1) * 4 + 2*x ].y + | |
1153 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
|
1154 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
|
1155 uvmv.y = (uvmv.y + 2 + (uvmv.y >> (INT_BIT-1))) >> 2; |
11921 | 1156 if (s->profile == 3) { |
1157 uvmv.x &= ~7; | |
1158 uvmv.y &= ~7; | |
1159 } | |
1160 vp8_mc(s, 0, dst[1] + 4*y*s->uvlinesize + x*4, | |
12228
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1161 ref->data[1], &uvmv, |
11921 | 1162 4*x + x_off, 4*y + y_off, 4, 4, |
1163 width, height, s->uvlinesize, | |
11974 | 1164 s->put_pixels_tab[2]); |
11921 | 1165 vp8_mc(s, 0, dst[2] + 4*y*s->uvlinesize + x*4, |
12228
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1166 ref->data[2], &uvmv, |
11921 | 1167 4*x + x_off, 4*y + y_off, 4, 4, |
1168 width, height, s->uvlinesize, | |
11974 | 1169 s->put_pixels_tab[2]); |
11921 | 1170 } |
1171 } | |
11989
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1172 break; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1173 } |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1174 case VP8_SPLITMVMODE_16x8: |
12228
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1175 vp8_mc_part(s, dst, ref, x_off, y_off, |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1176 0, 0, 16, 8, width, height, &bmv[0]); |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1177 vp8_mc_part(s, dst, ref, x_off, y_off, |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1178 0, 8, 16, 8, width, height, &bmv[1]); |
11989
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1179 break; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1180 case VP8_SPLITMVMODE_8x16: |
12228
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1181 vp8_mc_part(s, dst, ref, x_off, y_off, |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1182 0, 0, 8, 16, width, height, &bmv[0]); |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1183 vp8_mc_part(s, dst, ref, x_off, y_off, |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1184 8, 0, 8, 16, width, height, &bmv[1]); |
11989
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1185 break; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1186 case VP8_SPLITMVMODE_8x8: |
12228
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1187 vp8_mc_part(s, dst, ref, x_off, y_off, |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1188 0, 0, 8, 8, width, height, &bmv[0]); |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1189 vp8_mc_part(s, dst, ref, x_off, y_off, |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1190 8, 0, 8, 8, width, height, &bmv[1]); |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1191 vp8_mc_part(s, dst, ref, x_off, y_off, |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1192 0, 8, 8, 8, width, height, &bmv[2]); |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1193 vp8_mc_part(s, dst, ref, x_off, y_off, |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1194 8, 8, 8, 8, width, height, &bmv[3]); |
11989
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1195 break; |
11921 | 1196 } |
1197 } | |
1198 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1199 static av_always_inline void idct_mb(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb) |
11921 | 1200 { |
12240 | 1201 int x, y, ch; |
11921 | 1202 |
12238 | 1203 if (mb->mode != MODE_I4x4) { |
1204 uint8_t *y_dst = dst[0]; | |
11921 | 1205 for (y = 0; y < 4; y++) { |
12240 | 1206 uint32_t nnz4 = AV_RN32A(s->non_zero_count_cache[y]); |
1207 if (nnz4) { | |
1208 if (nnz4&~0x01010101) { | |
12238 | 1209 for (x = 0; x < 4; x++) { |
12240 | 1210 int nnz = s->non_zero_count_cache[y][x]; |
12238 | 1211 if (nnz) { |
1212 if (nnz == 1) | |
1213 s->vp8dsp.vp8_idct_dc_add(y_dst+4*x, s->block[y][x], s->linesize); | |
1214 else | |
1215 s->vp8dsp.vp8_idct_add(y_dst+4*x, s->block[y][x], s->linesize); | |
1216 } | |
1217 } | |
1218 } else { | |
12241
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1219 s->vp8dsp.vp8_idct_dc_add4y(y_dst, s->block[y], s->linesize); |
11921 | 1220 } |
1221 } | |
1222 y_dst += 4*s->linesize; | |
1223 } | |
12238 | 1224 } |
11921 | 1225 |
12238 | 1226 for (ch = 0; ch < 2; ch++) { |
12241
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1227 uint32_t nnz4 = AV_RN32A(s->non_zero_count_cache[4+ch]); |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1228 if (nnz4) { |
12238 | 1229 uint8_t *ch_dst = dst[1+ch]; |
12241
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1230 if (nnz4&~0x01010101) { |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1231 for (y = 0; y < 2; y++) { |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1232 for (x = 0; x < 2; x++) { |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1233 int nnz = s->non_zero_count_cache[4+ch][(y<<1)+x]; |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1234 if (nnz) { |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1235 if (nnz == 1) |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1236 s->vp8dsp.vp8_idct_dc_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize); |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1237 else |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1238 s->vp8dsp.vp8_idct_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize); |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1239 } |
12238 | 1240 } |
12241
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1241 ch_dst += 4*s->uvlinesize; |
12238 | 1242 } |
12241
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1243 } else { |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1244 s->vp8dsp.vp8_idct_dc_add4uv(ch_dst, s->block[4+ch], s->uvlinesize); |
11921 | 1245 } |
1246 } | |
1247 } | |
1248 } | |
1249 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1250 static av_always_inline void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb, VP8FilterStrength *f ) |
11921 | 1251 { |
1252 int interior_limit, filter_level; | |
1253 | |
1254 if (s->segmentation.enabled) { | |
12224
5b7d690b761b
VP8: Don't store segment in macroblock struct anymore.
darkshikari
parents:
12223
diff
changeset
|
1255 filter_level = s->segmentation.filter_level[s->segment]; |
11921 | 1256 if (!s->segmentation.absolute_vals) |
1257 filter_level += s->filter.level; | |
1258 } else | |
1259 filter_level = s->filter.level; | |
1260 | |
1261 if (s->lf_delta.enabled) { | |
1262 filter_level += s->lf_delta.ref[mb->ref_frame]; | |
1263 | |
1264 if (mb->ref_frame == VP56_FRAME_CURRENT) { | |
1265 if (mb->mode == MODE_I4x4) | |
1266 filter_level += s->lf_delta.mode[0]; | |
1267 } else { | |
1268 if (mb->mode == VP8_MVMODE_ZERO) | |
1269 filter_level += s->lf_delta.mode[1]; | |
1270 else if (mb->mode == VP8_MVMODE_SPLIT) | |
1271 filter_level += s->lf_delta.mode[3]; | |
1272 else | |
1273 filter_level += s->lf_delta.mode[2]; | |
1274 } | |
1275 } | |
1276 filter_level = av_clip(filter_level, 0, 63); | |
1277 | |
1278 interior_limit = filter_level; | |
1279 if (s->filter.sharpness) { | |
1280 interior_limit >>= s->filter.sharpness > 4 ? 2 : 1; | |
1281 interior_limit = FFMIN(interior_limit, 9 - s->filter.sharpness); | |
1282 } | |
1283 interior_limit = FFMAX(interior_limit, 1); | |
1284 | |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1285 f->filter_level = filter_level; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1286 f->inner_limit = interior_limit; |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1287 f->inner_filter = !mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT; |
11921 | 1288 } |
1289 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1290 static av_always_inline void filter_mb(VP8Context *s, uint8_t *dst[3], VP8FilterStrength *f, int mb_x, int mb_y) |
11921 | 1291 { |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1292 int mbedge_lim, bedge_lim, hev_thresh; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1293 int filter_level = f->filter_level; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1294 int inner_limit = f->inner_limit; |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1295 int inner_filter = f->inner_filter; |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1296 int linesize = s->linesize; |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1297 int uvlinesize = s->uvlinesize; |
11921 | 1298 |
1299 if (!filter_level) | |
1300 return; | |
1301 | |
12081
812e23197d64
VP8: Move calculation of outer filter limit out of dsp functions for normal
conrad
parents:
12062
diff
changeset
|
1302 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
|
1303 bedge_lim = 2* filter_level + inner_limit; |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1304 hev_thresh = filter_level >= 15; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1305 |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1306 if (s->keyframe) { |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1307 if (filter_level >= 40) |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1308 hev_thresh = 2; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1309 } else { |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1310 if (filter_level >= 40) |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1311 hev_thresh = 3; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1312 else if (filter_level >= 20) |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1313 hev_thresh = 2; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1314 } |
12081
812e23197d64
VP8: Move calculation of outer filter limit out of dsp functions for normal
conrad
parents:
12062
diff
changeset
|
1315 |
11921 | 1316 if (mb_x) { |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1317 s->vp8dsp.vp8_h_loop_filter16y(dst[0], linesize, |
12194
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1318 mbedge_lim, inner_limit, hev_thresh); |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1319 s->vp8dsp.vp8_h_loop_filter8uv(dst[1], dst[2], uvlinesize, |
12194
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1320 mbedge_lim, inner_limit, hev_thresh); |
11921 | 1321 } |
1322 | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1323 if (inner_filter) { |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1324 s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 4, linesize, bedge_lim, |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1325 inner_limit, hev_thresh); |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1326 s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 8, linesize, bedge_lim, |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1327 inner_limit, hev_thresh); |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1328 s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+12, linesize, bedge_lim, |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1329 inner_limit, hev_thresh); |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1330 s->vp8dsp.vp8_h_loop_filter8uv_inner(dst[1] + 4, dst[2] + 4, |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1331 uvlinesize, bedge_lim, |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1332 inner_limit, hev_thresh); |
11921 | 1333 } |
1334 | |
1335 if (mb_y) { | |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1336 s->vp8dsp.vp8_v_loop_filter16y(dst[0], linesize, |
12194
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1337 mbedge_lim, inner_limit, hev_thresh); |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1338 s->vp8dsp.vp8_v_loop_filter8uv(dst[1], dst[2], uvlinesize, |
12194
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1339 mbedge_lim, inner_limit, hev_thresh); |
11921 | 1340 } |
1341 | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1342 if (inner_filter) { |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1343 s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 4*linesize, |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1344 linesize, bedge_lim, |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1345 inner_limit, hev_thresh); |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1346 s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 8*linesize, |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1347 linesize, bedge_lim, |
12194
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1348 inner_limit, hev_thresh); |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1349 s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+12*linesize, |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1350 linesize, bedge_lim, |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1351 inner_limit, hev_thresh); |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1352 s->vp8dsp.vp8_v_loop_filter8uv_inner(dst[1] + 4 * uvlinesize, |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1353 dst[2] + 4 * uvlinesize, |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1354 uvlinesize, bedge_lim, |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1355 inner_limit, hev_thresh); |
11921 | 1356 } |
1357 } | |
1358 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1359 static av_always_inline void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8FilterStrength *f, int mb_x, int mb_y) |
11921 | 1360 { |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1361 int mbedge_lim, bedge_lim; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1362 int filter_level = f->filter_level; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1363 int inner_limit = f->inner_limit; |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1364 int inner_filter = f->inner_filter; |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1365 int linesize = s->linesize; |
11921 | 1366 |
1367 if (!filter_level) | |
1368 return; | |
1369 | |
1370 mbedge_lim = 2*(filter_level+2) + inner_limit; | |
1371 bedge_lim = 2* filter_level + inner_limit; | |
1372 | |
1373 if (mb_x) | |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1374 s->vp8dsp.vp8_h_loop_filter_simple(dst, linesize, mbedge_lim); |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1375 if (inner_filter) { |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1376 s->vp8dsp.vp8_h_loop_filter_simple(dst+ 4, linesize, bedge_lim); |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1377 s->vp8dsp.vp8_h_loop_filter_simple(dst+ 8, linesize, bedge_lim); |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1378 s->vp8dsp.vp8_h_loop_filter_simple(dst+12, linesize, bedge_lim); |
11921 | 1379 } |
1380 | |
1381 if (mb_y) | |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1382 s->vp8dsp.vp8_v_loop_filter_simple(dst, linesize, mbedge_lim); |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1383 if (inner_filter) { |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1384 s->vp8dsp.vp8_v_loop_filter_simple(dst+ 4*linesize, linesize, bedge_lim); |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1385 s->vp8dsp.vp8_v_loop_filter_simple(dst+ 8*linesize, linesize, bedge_lim); |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1386 s->vp8dsp.vp8_v_loop_filter_simple(dst+12*linesize, linesize, bedge_lim); |
11921 | 1387 } |
1388 } | |
1389 | |
1390 static void filter_mb_row(VP8Context *s, int mb_y) | |
1391 { | |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1392 VP8FilterStrength *f = s->filter_strength; |
11921 | 1393 uint8_t *dst[3] = { |
1394 s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize, | |
1395 s->framep[VP56_FRAME_CURRENT]->data[1] + 8*mb_y*s->uvlinesize, | |
1396 s->framep[VP56_FRAME_CURRENT]->data[2] + 8*mb_y*s->uvlinesize | |
1397 }; | |
1398 int mb_x; | |
1399 | |
1400 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
|
1401 backup_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2], s->linesize, s->uvlinesize, 0); |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1402 filter_mb(s, dst, f++, mb_x, mb_y); |
11921 | 1403 dst[0] += 16; |
1404 dst[1] += 8; | |
1405 dst[2] += 8; | |
1406 } | |
1407 } | |
1408 | |
1409 static void filter_mb_row_simple(VP8Context *s, int mb_y) | |
1410 { | |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1411 VP8FilterStrength *f = s->filter_strength; |
11921 | 1412 uint8_t *dst = s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize; |
1413 int mb_x; | |
1414 | |
1415 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
|
1416 backup_mb_border(s->top_border[mb_x+1], dst, NULL, NULL, s->linesize, 0, 1); |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1417 filter_mb_simple(s, dst, f++, mb_x, mb_y); |
11921 | 1418 dst += 16; |
1419 } | |
1420 } | |
1421 | |
1422 static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size, | |
1423 AVPacket *avpkt) | |
1424 { | |
1425 VP8Context *s = avctx->priv_data; | |
1426 int ret, mb_x, mb_y, i, y, referenced; | |
1427 enum AVDiscard skip_thresh; | |
12232 | 1428 AVFrame *curframe = NULL; |
11921 | 1429 |
1430 if ((ret = decode_frame_header(s, avpkt->data, avpkt->size)) < 0) | |
1431 return ret; | |
1432 | |
1433 referenced = s->update_last || s->update_golden == VP56_FRAME_CURRENT | |
1434 || s->update_altref == VP56_FRAME_CURRENT; | |
1435 | |
1436 skip_thresh = !referenced ? AVDISCARD_NONREF : | |
1437 !s->keyframe ? AVDISCARD_NONKEY : AVDISCARD_ALL; | |
1438 | |
1439 if (avctx->skip_frame >= skip_thresh) { | |
1440 s->invisible = 1; | |
1441 goto skip_decode; | |
1442 } | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
1443 s->deblock_filter = s->filter.level && avctx->skip_loop_filter < skip_thresh; |
11921 | 1444 |
1445 for (i = 0; i < 4; i++) | |
1446 if (&s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] && | |
1447 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] && | |
1448 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) { | |
1449 curframe = s->framep[VP56_FRAME_CURRENT] = &s->frames[i]; | |
1450 break; | |
1451 } | |
1452 if (curframe->data[0]) | |
1453 avctx->release_buffer(avctx, curframe); | |
1454 | |
1455 curframe->key_frame = s->keyframe; | |
1456 curframe->pict_type = s->keyframe ? FF_I_TYPE : FF_P_TYPE; | |
1457 curframe->reference = referenced ? 3 : 0; | |
1458 if ((ret = avctx->get_buffer(avctx, curframe))) { | |
1459 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed!\n"); | |
1460 return ret; | |
1461 } | |
1462 | |
1463 // Given that arithmetic probabilities are updated every frame, it's quite likely | |
1464 // that the values we have on a random interframe are complete junk if we didn't | |
1465 // start decode on a keyframe. So just don't display anything rather than junk. | |
1466 if (!s->keyframe && (!s->framep[VP56_FRAME_PREVIOUS] || | |
1467 !s->framep[VP56_FRAME_GOLDEN] || | |
1468 !s->framep[VP56_FRAME_GOLDEN2])) { | |
1469 av_log(avctx, AV_LOG_WARNING, "Discarding interframe without a prior keyframe!\n"); | |
1470 return AVERROR_INVALIDDATA; | |
1471 } | |
1472 | |
1473 s->linesize = curframe->linesize[0]; | |
1474 s->uvlinesize = curframe->linesize[1]; | |
1475 | |
1476 if (!s->edge_emu_buffer) | |
1477 s->edge_emu_buffer = av_malloc(21*s->linesize); | |
1478 | |
1479 memset(s->top_nnz, 0, s->mb_width*sizeof(*s->top_nnz)); | |
1480 | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1481 /* Zero macroblock structures for top/left prediction from outside the frame. */ |
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1482 memset(s->macroblocks, 0, (s->mb_width + s->mb_height*2)*sizeof(*s->macroblocks)); |
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1483 |
11921 | 1484 // 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
|
1485 memset(s->top_border, 127, (s->mb_width+1)*sizeof(*s->top_border)); |
12231 | 1486 memset(s->ref_count, 0, sizeof(s->ref_count)); |
11921 | 1487 |
1488 for (mb_y = 0; mb_y < s->mb_height; mb_y++) { | |
1489 VP56RangeCoder *c = &s->coeff_partition[mb_y & (s->num_coeff_partitions-1)]; | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1490 VP8Macroblock *mb = s->macroblocks + (s->mb_height - mb_y - 1)*2; |
11921 | 1491 uint8_t *intra4x4 = s->intra4x4_pred_mode + 4*mb_y*s->b4_stride; |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1492 uint8_t *segment_map = s->segmentation_map + mb_y*s->mb_stride; |
12237 | 1493 int mb_xy = mb_y * s->mb_stride; |
11921 | 1494 uint8_t *dst[3] = { |
1495 curframe->data[0] + 16*mb_y*s->linesize, | |
1496 curframe->data[1] + 8*mb_y*s->uvlinesize, | |
1497 curframe->data[2] + 8*mb_y*s->uvlinesize | |
1498 }; | |
1499 | |
1500 memset(s->left_nnz, 0, sizeof(s->left_nnz)); | |
1501 | |
1502 // left edge of 129 for intra prediction | |
1503 if (!(avctx->flags & CODEC_FLAG_EMU_EDGE)) | |
1504 for (i = 0; i < 3; i++) | |
1505 for (y = 0; y < 16>>!!i; y++) | |
1506 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
|
1507 if (mb_y) |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
1508 memset(s->top_border, 129, sizeof(*s->top_border)); |
11921 | 1509 |
12237 | 1510 for (mb_x = 0; mb_x < s->mb_width; mb_x++, mb_xy++, mb++) { |
12221 | 1511 uint8_t *intra4x4_mb = s->keyframe ? intra4x4 + 4*mb_x : s->intra4x4_pred_mode_mb; |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1512 uint8_t *segment_mb = segment_map+mb_x; |
12221 | 1513 |
12215 | 1514 /* Prefetch the current frame, 4 MBs ahead */ |
1515 s->dsp.prefetch(dst[0] + (mb_x&3)*4*s->linesize + 64, s->linesize, 4); | |
1516 s->dsp.prefetch(dst[1] + (mb_x&7)*s->uvlinesize + 64, dst[2] - dst[1], 2); | |
1517 | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1518 decode_mb_mode(s, mb, mb_x, mb_y, intra4x4_mb, segment_mb); |
11921 | 1519 |
12237 | 1520 prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_PREVIOUS); |
12231 | 1521 |
11921 | 1522 if (!mb->skip) |
1523 decode_mb_coeffs(s, c, mb, s->top_nnz[mb_x], s->left_nnz); | |
1524 | |
12225
c3e11b3108d7
Eliminate a pointless memset for intra blocks in P-frames in VP8
darkshikari
parents:
12224
diff
changeset
|
1525 if (mb->mode <= MODE_I4x4) |
12221 | 1526 intra_predict(s, dst, mb, intra4x4_mb, mb_x, mb_y); |
12225
c3e11b3108d7
Eliminate a pointless memset for intra blocks in P-frames in VP8
darkshikari
parents:
12224
diff
changeset
|
1527 else |
11921 | 1528 inter_predict(s, dst, mb, mb_x, mb_y); |
1529 | |
12237 | 1530 prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN); |
12231 | 1531 |
11921 | 1532 if (!mb->skip) { |
12238 | 1533 idct_mb(s, dst, mb); |
11921 | 1534 } else { |
1535 AV_ZERO64(s->left_nnz); | |
1536 AV_WN64(s->top_nnz[mb_x], 0); // array of 9, so unaligned | |
1537 | |
1538 // Reset DC block predictors if they would exist if the mb had coefficients | |
1539 if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) { | |
1540 s->left_nnz[8] = 0; | |
1541 s->top_nnz[mb_x][8] = 0; | |
1542 } | |
1543 } | |
1544 | |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1545 if (s->deblock_filter) |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1546 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
|
1547 |
12237 | 1548 prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN2); |
12231 | 1549 |
11921 | 1550 dst[0] += 16; |
1551 dst[1] += 8; | |
1552 dst[2] += 8; | |
1553 } | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
1554 if (s->deblock_filter) { |
11921 | 1555 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
|
1556 filter_mb_row_simple(s, mb_y); |
11921 | 1557 else |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
1558 filter_mb_row(s, mb_y); |
11921 | 1559 } |
1560 } | |
1561 | |
1562 skip_decode: | |
1563 // if future frames don't use the updated probabilities, | |
1564 // reset them to the values we saved | |
1565 if (!s->update_probabilities) | |
1566 s->prob[0] = s->prob[1]; | |
1567 | |
1568 // check if golden and altref are swapped | |
1569 if (s->update_altref == VP56_FRAME_GOLDEN && | |
1570 s->update_golden == VP56_FRAME_GOLDEN2) | |
1571 FFSWAP(AVFrame *, s->framep[VP56_FRAME_GOLDEN], s->framep[VP56_FRAME_GOLDEN2]); | |
1572 else { | |
1573 if (s->update_altref != VP56_FRAME_NONE) | |
1574 s->framep[VP56_FRAME_GOLDEN2] = s->framep[s->update_altref]; | |
1575 | |
1576 if (s->update_golden != VP56_FRAME_NONE) | |
1577 s->framep[VP56_FRAME_GOLDEN] = s->framep[s->update_golden]; | |
1578 } | |
1579 | |
1580 if (s->update_last) // move cur->prev | |
1581 s->framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_CURRENT]; | |
1582 | |
1583 // release no longer referenced frames | |
1584 for (i = 0; i < 4; i++) | |
1585 if (s->frames[i].data[0] && | |
1586 &s->frames[i] != s->framep[VP56_FRAME_CURRENT] && | |
1587 &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] && | |
1588 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] && | |
1589 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) | |
1590 avctx->release_buffer(avctx, &s->frames[i]); | |
1591 | |
1592 if (!s->invisible) { | |
1593 *(AVFrame*)data = *s->framep[VP56_FRAME_CURRENT]; | |
1594 *data_size = sizeof(AVFrame); | |
1595 } | |
1596 | |
1597 return avpkt->size; | |
1598 } | |
1599 | |
1600 static av_cold int vp8_decode_init(AVCodecContext *avctx) | |
1601 { | |
1602 VP8Context *s = avctx->priv_data; | |
1603 | |
1604 s->avctx = avctx; | |
1605 avctx->pix_fmt = PIX_FMT_YUV420P; | |
1606 | |
1607 dsputil_init(&s->dsp, avctx); | |
1608 ff_h264_pred_init(&s->hpc, CODEC_ID_VP8); | |
1609 ff_vp8dsp_init(&s->vp8dsp); | |
1610 | |
1611 // intra pred needs edge emulation among other things | |
1612 if (avctx->flags&CODEC_FLAG_EMU_EDGE) { | |
11947 | 1613 av_log(avctx, AV_LOG_ERROR, "Edge emulation not supported\n"); |
11921 | 1614 return AVERROR_PATCHWELCOME; |
1615 } | |
1616 | |
1617 return 0; | |
1618 } | |
1619 | |
1620 static av_cold int vp8_decode_free(AVCodecContext *avctx) | |
1621 { | |
1622 vp8_decode_flush(avctx); | |
1623 return 0; | |
1624 } | |
1625 | |
1626 AVCodec vp8_decoder = { | |
1627 "vp8", | |
1628 AVMEDIA_TYPE_VIDEO, | |
1629 CODEC_ID_VP8, | |
1630 sizeof(VP8Context), | |
1631 vp8_decode_init, | |
1632 NULL, | |
1633 vp8_decode_free, | |
1634 vp8_decode_frame, | |
1635 CODEC_CAP_DR1, | |
1636 .flush = vp8_decode_flush, | |
1637 .long_name = NULL_IF_CONFIG_SMALL("On2 VP8"), | |
1638 }; |