Mercurial > libavcodec.hg
annotate vp8.c @ 12253:112b3a0db187 libavcodec
Decode DCT tokens by branching to a different code path for each branch
on the huffman tree, instead of traversing the tree in a while loop.
Based on the similar optimization in libvpx's detokenize.c
10% faster at normal bitrates, and 30% faster for high-bitrate intra-only
author | conrad |
---|---|
date | Fri, 23 Jul 2010 21:46:17 +0000 |
parents | 35ee666e4496 |
children | 17c151e1280a |
rev | line source |
---|---|
11921 | 1 /** |
2 * VP8 compatible video decoder | |
3 * | |
4 * Copyright (C) 2010 David Conrad | |
5 * Copyright (C) 2010 Ronald S. Bultje | |
12249 | 6 * Copyright (C) 2010 Jason Garrett-Glaser |
11921 | 7 * |
8 * This file is part of FFmpeg. | |
9 * | |
10 * FFmpeg is free software; you can redistribute it and/or | |
11 * modify it under the terms of the GNU Lesser General Public | |
12 * License as published by the Free Software Foundation; either | |
13 * version 2.1 of the License, or (at your option) any later version. | |
14 * | |
15 * FFmpeg is distributed in the hope that it will be useful, | |
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
18 * Lesser General Public License for more details. | |
19 * | |
20 * You should have received a copy of the GNU Lesser General Public | |
21 * License along with FFmpeg; if not, write to the Free Software | |
22 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | |
23 */ | |
24 | |
25 #include "avcodec.h" | |
26 #include "vp56.h" | |
27 #include "vp8data.h" | |
28 #include "vp8dsp.h" | |
29 #include "h264pred.h" | |
30 #include "rectangle.h" | |
31 | |
32 typedef struct { | |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
33 uint8_t filter_level; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
34 uint8_t inner_limit; |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
35 uint8_t inner_filter; |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
36 } VP8FilterStrength; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
37 |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
38 typedef struct { |
11921 | 39 uint8_t skip; |
40 // todo: make it possible to check for at least (i4x4 or split_mv) | |
41 // in one op. are others needed? | |
42 uint8_t mode; | |
43 uint8_t ref_frame; | |
44 uint8_t partitioning; | |
45 VP56mv mv; | |
46 VP56mv bmv[16]; | |
47 } VP8Macroblock; | |
48 | |
49 typedef struct { | |
50 AVCodecContext *avctx; | |
51 DSPContext dsp; | |
52 VP8DSPContext vp8dsp; | |
53 H264PredContext hpc; | |
11974 | 54 vp8_mc_func put_pixels_tab[3][3][3]; |
11921 | 55 AVFrame frames[4]; |
56 AVFrame *framep[4]; | |
57 uint8_t *edge_emu_buffer; | |
58 VP56RangeCoder c; ///< header context, includes mb modes and motion vectors | |
59 int profile; | |
60 | |
61 int mb_width; /* number of horizontal MB */ | |
62 int mb_height; /* number of vertical MB */ | |
63 int linesize; | |
64 int uvlinesize; | |
65 | |
66 int keyframe; | |
67 int invisible; | |
68 int update_last; ///< update VP56_FRAME_PREVIOUS with the current one | |
69 int update_golden; ///< VP56_FRAME_NONE if not updated, or which frame to copy if so | |
70 int update_altref; | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
71 int deblock_filter; |
11921 | 72 |
73 /** | |
74 * If this flag is not set, all the probability updates | |
75 * are discarded after this frame is decoded. | |
76 */ | |
77 int update_probabilities; | |
78 | |
79 /** | |
80 * All coefficients are contained in separate arith coding contexts. | |
81 * There can be 1, 2, 4, or 8 of these after the header context. | |
82 */ | |
83 int num_coeff_partitions; | |
84 VP56RangeCoder coeff_partition[8]; | |
85 | |
86 VP8Macroblock *macroblocks; | |
87 VP8Macroblock *macroblocks_base; | |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
88 VP8FilterStrength *filter_strength; |
11921 | 89 int mb_stride; |
90 | |
91 uint8_t *intra4x4_pred_mode; | |
92 uint8_t *intra4x4_pred_mode_base; | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
93 uint8_t *segmentation_map; |
11921 | 94 int b4_stride; |
95 | |
96 /** | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
97 * 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
|
98 * 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
|
99 */ |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
100 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
|
101 |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
102 /** |
11921 | 103 * For coeff decode, we need to know whether the above block had non-zero |
104 * coefficients. This means for each macroblock, we need data for 4 luma | |
105 * blocks, 2 u blocks, 2 v blocks, and the luma dc block, for a total of 9 | |
106 * per macroblock. We keep the last row in top_nnz. | |
107 */ | |
108 uint8_t (*top_nnz)[9]; | |
109 DECLARE_ALIGNED(8, uint8_t, left_nnz)[9]; | |
110 | |
111 /** | |
112 * This is the index plus one of the last non-zero coeff | |
113 * for each of the blocks in the current macroblock. | |
114 * So, 0 -> no coeffs | |
115 * 1 -> dc-only (special transform) | |
116 * 2+-> full transform | |
117 */ | |
118 DECLARE_ALIGNED(16, uint8_t, non_zero_count_cache)[6][4]; | |
119 DECLARE_ALIGNED(16, DCTELEM, block)[6][4][16]; | |
12221 | 120 uint8_t intra4x4_pred_mode_mb[16]; |
11921 | 121 |
122 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
|
123 int segment; ///< segment of the current macroblock |
11921 | 124 |
125 int mbskip_enabled; | |
126 int sign_bias[4]; ///< one state [0, 1] per ref frame type | |
12231 | 127 int ref_count[3]; |
11921 | 128 |
129 /** | |
130 * Base parameters for segmentation, i.e. per-macroblock parameters. | |
131 * These must be kept unchanged even if segmentation is not used for | |
132 * a frame, since the values persist between interframes. | |
133 */ | |
134 struct { | |
135 int enabled; | |
136 int absolute_vals; | |
137 int update_map; | |
138 int8_t base_quant[4]; | |
139 int8_t filter_level[4]; ///< base loop filter level | |
140 } segmentation; | |
141 | |
142 /** | |
143 * Macroblocks can have one of 4 different quants in a frame when | |
144 * segmentation is enabled. | |
145 * If segmentation is disabled, only the first segment's values are used. | |
146 */ | |
147 struct { | |
148 // [0] - DC qmul [1] - AC qmul | |
149 int16_t luma_qmul[2]; | |
150 int16_t luma_dc_qmul[2]; ///< luma dc-only block quant | |
151 int16_t chroma_qmul[2]; | |
152 } qmat[4]; | |
153 | |
154 struct { | |
155 int simple; | |
156 int level; | |
157 int sharpness; | |
158 } filter; | |
159 | |
160 struct { | |
161 int enabled; ///< whether each mb can have a different strength based on mode/ref | |
162 | |
163 /** | |
164 * filter strength adjustment for the following macroblock modes: | |
165 * [0] - i4x4 | |
166 * [1] - zero mv | |
167 * [2] - inter modes except for zero or split mv | |
168 * [3] - split mv | |
169 * i16x16 modes never have any adjustment | |
170 */ | |
171 int8_t mode[4]; | |
172 | |
173 /** | |
174 * filter strength adjustment for macroblocks that reference: | |
175 * [0] - intra / VP56_FRAME_CURRENT | |
176 * [1] - VP56_FRAME_PREVIOUS | |
177 * [2] - VP56_FRAME_GOLDEN | |
178 * [3] - altref / VP56_FRAME_GOLDEN2 | |
179 */ | |
180 int8_t ref[4]; | |
181 } lf_delta; | |
182 | |
183 /** | |
184 * These are all of the updatable probabilities for binary decisions. | |
185 * They are only implictly reset on keyframes, making it quite likely | |
186 * for an interframe to desync if a prior frame's header was corrupt | |
187 * or missing outright! | |
188 */ | |
189 struct { | |
190 uint8_t segmentid[3]; | |
191 uint8_t mbskip; | |
192 uint8_t intra; | |
193 uint8_t last; | |
194 uint8_t golden; | |
195 uint8_t pred16x16[4]; | |
196 uint8_t pred8x8c[3]; | |
197 uint8_t token[4][8][3][NUM_DCT_TOKENS-1]; | |
198 uint8_t mvc[2][19]; | |
199 } prob[2]; | |
200 } VP8Context; | |
201 | |
202 static void vp8_decode_flush(AVCodecContext *avctx) | |
203 { | |
204 VP8Context *s = avctx->priv_data; | |
205 int i; | |
206 | |
207 for (i = 0; i < 4; i++) | |
208 if (s->frames[i].data[0]) | |
209 avctx->release_buffer(avctx, &s->frames[i]); | |
210 memset(s->framep, 0, sizeof(s->framep)); | |
211 | |
212 av_freep(&s->macroblocks_base); | |
213 av_freep(&s->intra4x4_pred_mode_base); | |
214 av_freep(&s->top_nnz); | |
215 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
|
216 av_freep(&s->top_border); |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
217 av_freep(&s->segmentation_map); |
11921 | 218 |
219 s->macroblocks = NULL; | |
220 s->intra4x4_pred_mode = NULL; | |
221 } | |
222 | |
223 static int update_dimensions(VP8Context *s, int width, int height) | |
224 { | |
225 int i; | |
226 | |
227 if (avcodec_check_dimensions(s->avctx, width, height)) | |
228 return AVERROR_INVALIDDATA; | |
229 | |
230 vp8_decode_flush(s->avctx); | |
231 | |
232 avcodec_set_dimensions(s->avctx, width, height); | |
233 | |
234 s->mb_width = (s->avctx->coded_width +15) / 16; | |
235 s->mb_height = (s->avctx->coded_height+15) / 16; | |
236 | |
237 // we allocate a border around the top/left of intra4x4 modes | |
238 // this is 4 blocks for intra4x4 to keep 4-byte alignment for fill_rectangle | |
239 s->mb_stride = s->mb_width+1; | |
240 s->b4_stride = 4*s->mb_stride; | |
241 | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
242 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
|
243 s->filter_strength = av_mallocz(s->mb_stride*sizeof(*s->filter_strength)); |
11921 | 244 s->intra4x4_pred_mode_base = av_mallocz(s->b4_stride*(4*s->mb_height+1)); |
245 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
|
246 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
|
247 s->segmentation_map = av_mallocz(s->mb_stride*s->mb_height); |
11921 | 248 |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
249 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
|
250 !s->top_nnz || !s->top_border || !s->segmentation_map) |
12169 | 251 return AVERROR(ENOMEM); |
252 | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
253 s->macroblocks = s->macroblocks_base + 1; |
11921 | 254 s->intra4x4_pred_mode = s->intra4x4_pred_mode_base + 4 + s->b4_stride; |
255 | |
256 memset(s->intra4x4_pred_mode_base, DC_PRED, s->b4_stride); | |
257 for (i = 0; i < 4*s->mb_height; i++) | |
258 s->intra4x4_pred_mode[i*s->b4_stride-1] = DC_PRED; | |
259 | |
260 return 0; | |
261 } | |
262 | |
263 static void parse_segment_info(VP8Context *s) | |
264 { | |
265 VP56RangeCoder *c = &s->c; | |
266 int i; | |
267 | |
268 s->segmentation.update_map = vp8_rac_get(c); | |
269 | |
270 if (vp8_rac_get(c)) { // update segment feature data | |
271 s->segmentation.absolute_vals = vp8_rac_get(c); | |
272 | |
273 for (i = 0; i < 4; i++) | |
274 s->segmentation.base_quant[i] = vp8_rac_get_sint(c, 7); | |
275 | |
276 for (i = 0; i < 4; i++) | |
277 s->segmentation.filter_level[i] = vp8_rac_get_sint(c, 6); | |
278 } | |
279 if (s->segmentation.update_map) | |
280 for (i = 0; i < 3; i++) | |
281 s->prob->segmentid[i] = vp8_rac_get(c) ? vp8_rac_get_uint(c, 8) : 255; | |
282 } | |
283 | |
284 static void update_lf_deltas(VP8Context *s) | |
285 { | |
286 VP56RangeCoder *c = &s->c; | |
287 int i; | |
288 | |
289 for (i = 0; i < 4; i++) | |
290 s->lf_delta.ref[i] = vp8_rac_get_sint(c, 6); | |
291 | |
292 for (i = 0; i < 4; i++) | |
293 s->lf_delta.mode[i] = vp8_rac_get_sint(c, 6); | |
294 } | |
295 | |
296 static int setup_partitions(VP8Context *s, const uint8_t *buf, int buf_size) | |
297 { | |
298 const uint8_t *sizes = buf; | |
299 int i; | |
300 | |
301 s->num_coeff_partitions = 1 << vp8_rac_get_uint(&s->c, 2); | |
302 | |
303 buf += 3*(s->num_coeff_partitions-1); | |
304 buf_size -= 3*(s->num_coeff_partitions-1); | |
305 if (buf_size < 0) | |
306 return -1; | |
307 | |
308 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
|
309 int size = AV_RL24(sizes + 3*i); |
11921 | 310 if (buf_size - size < 0) |
311 return -1; | |
312 | |
313 vp56_init_range_decoder(&s->coeff_partition[i], buf, size); | |
314 buf += size; | |
315 buf_size -= size; | |
316 } | |
317 vp56_init_range_decoder(&s->coeff_partition[i], buf, buf_size); | |
318 | |
319 return 0; | |
320 } | |
321 | |
322 static void get_quants(VP8Context *s) | |
323 { | |
324 VP56RangeCoder *c = &s->c; | |
325 int i, base_qi; | |
326 | |
327 int yac_qi = vp8_rac_get_uint(c, 7); | |
328 int ydc_delta = vp8_rac_get_sint(c, 4); | |
329 int y2dc_delta = vp8_rac_get_sint(c, 4); | |
330 int y2ac_delta = vp8_rac_get_sint(c, 4); | |
331 int uvdc_delta = vp8_rac_get_sint(c, 4); | |
332 int uvac_delta = vp8_rac_get_sint(c, 4); | |
333 | |
334 for (i = 0; i < 4; i++) { | |
335 if (s->segmentation.enabled) { | |
336 base_qi = s->segmentation.base_quant[i]; | |
337 if (!s->segmentation.absolute_vals) | |
338 base_qi += yac_qi; | |
339 } else | |
340 base_qi = yac_qi; | |
341 | |
342 s->qmat[i].luma_qmul[0] = vp8_dc_qlookup[av_clip(base_qi + ydc_delta , 0, 127)]; | |
343 s->qmat[i].luma_qmul[1] = vp8_ac_qlookup[av_clip(base_qi , 0, 127)]; | |
344 s->qmat[i].luma_dc_qmul[0] = 2 * vp8_dc_qlookup[av_clip(base_qi + y2dc_delta, 0, 127)]; | |
345 s->qmat[i].luma_dc_qmul[1] = 155 * vp8_ac_qlookup[av_clip(base_qi + y2ac_delta, 0, 127)] / 100; | |
346 s->qmat[i].chroma_qmul[0] = vp8_dc_qlookup[av_clip(base_qi + uvdc_delta, 0, 127)]; | |
347 s->qmat[i].chroma_qmul[1] = vp8_ac_qlookup[av_clip(base_qi + uvac_delta, 0, 127)]; | |
348 | |
349 s->qmat[i].luma_dc_qmul[1] = FFMAX(s->qmat[i].luma_dc_qmul[1], 8); | |
350 s->qmat[i].chroma_qmul[0] = FFMIN(s->qmat[i].chroma_qmul[0], 132); | |
351 } | |
352 } | |
353 | |
354 /** | |
355 * Determine which buffers golden and altref should be updated with after this frame. | |
356 * The spec isn't clear here, so I'm going by my understanding of what libvpx does | |
357 * | |
358 * Intra frames update all 3 references | |
359 * Inter frames update VP56_FRAME_PREVIOUS if the update_last flag is set | |
360 * If the update (golden|altref) flag is set, it's updated with the current frame | |
361 * if update_last is set, and VP56_FRAME_PREVIOUS otherwise. | |
362 * If the flag is not set, the number read means: | |
363 * 0: no update | |
364 * 1: VP56_FRAME_PREVIOUS | |
365 * 2: update golden with altref, or update altref with golden | |
366 */ | |
367 static VP56Frame ref_to_update(VP8Context *s, int update, VP56Frame ref) | |
368 { | |
369 VP56RangeCoder *c = &s->c; | |
370 | |
371 if (update) | |
372 return VP56_FRAME_CURRENT; | |
373 | |
374 switch (vp8_rac_get_uint(c, 2)) { | |
375 case 1: | |
376 return VP56_FRAME_PREVIOUS; | |
377 case 2: | |
378 return (ref == VP56_FRAME_GOLDEN) ? VP56_FRAME_GOLDEN2 : VP56_FRAME_GOLDEN; | |
379 } | |
380 return VP56_FRAME_NONE; | |
381 } | |
382 | |
383 static void update_refs(VP8Context *s) | |
384 { | |
385 VP56RangeCoder *c = &s->c; | |
386 | |
387 int update_golden = vp8_rac_get(c); | |
388 int update_altref = vp8_rac_get(c); | |
389 | |
390 s->update_golden = ref_to_update(s, update_golden, VP56_FRAME_GOLDEN); | |
391 s->update_altref = ref_to_update(s, update_altref, VP56_FRAME_GOLDEN2); | |
392 } | |
393 | |
394 static int decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size) | |
395 { | |
396 VP56RangeCoder *c = &s->c; | |
397 int header_size, hscale, vscale, i, j, k, l, ret; | |
398 int width = s->avctx->width; | |
399 int height = s->avctx->height; | |
400 | |
401 s->keyframe = !(buf[0] & 1); | |
402 s->profile = (buf[0]>>1) & 7; | |
403 s->invisible = !(buf[0] & 0x10); | |
12247
50a96623366b
VP8: use AV_RL24 instead of defining a new RL24.
darkshikari
parents:
12246
diff
changeset
|
404 header_size = AV_RL24(buf) >> 5; |
11921 | 405 buf += 3; |
406 buf_size -= 3; | |
407 | |
11974 | 408 if (s->profile > 3) |
409 av_log(s->avctx, AV_LOG_WARNING, "Unknown profile %d\n", s->profile); | |
410 | |
411 if (!s->profile) | |
412 memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_epel_pixels_tab, sizeof(s->put_pixels_tab)); | |
413 else // profile 1-3 use bilinear, 4+ aren't defined so whatever | |
414 memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_bilinear_pixels_tab, sizeof(s->put_pixels_tab)); | |
11921 | 415 |
416 if (header_size > buf_size - 7*s->keyframe) { | |
417 av_log(s->avctx, AV_LOG_ERROR, "Header size larger than data provided\n"); | |
418 return AVERROR_INVALIDDATA; | |
419 } | |
420 | |
421 if (s->keyframe) { | |
12247
50a96623366b
VP8: use AV_RL24 instead of defining a new RL24.
darkshikari
parents:
12246
diff
changeset
|
422 if (AV_RL24(buf) != 0x2a019d) { |
50a96623366b
VP8: use AV_RL24 instead of defining a new RL24.
darkshikari
parents:
12246
diff
changeset
|
423 av_log(s->avctx, AV_LOG_ERROR, "Invalid start code 0x%x\n", AV_RL24(buf)); |
11921 | 424 return AVERROR_INVALIDDATA; |
425 } | |
426 width = AV_RL16(buf+3) & 0x3fff; | |
427 height = AV_RL16(buf+5) & 0x3fff; | |
428 hscale = buf[4] >> 6; | |
429 vscale = buf[6] >> 6; | |
430 buf += 7; | |
431 buf_size -= 7; | |
432 | |
11970
c7953ee47af4
vp8: warn and request sample if upscaling specified in header
mru
parents:
11950
diff
changeset
|
433 if (hscale || vscale) |
c7953ee47af4
vp8: warn and request sample if upscaling specified in header
mru
parents:
11950
diff
changeset
|
434 av_log_missing_feature(s->avctx, "Upscaling", 1); |
c7953ee47af4
vp8: warn and request sample if upscaling specified in header
mru
parents:
11950
diff
changeset
|
435 |
11921 | 436 s->update_golden = s->update_altref = VP56_FRAME_CURRENT; |
437 memcpy(s->prob->token , vp8_token_default_probs , sizeof(s->prob->token)); | |
438 memcpy(s->prob->pred16x16, vp8_pred16x16_prob_inter, sizeof(s->prob->pred16x16)); | |
439 memcpy(s->prob->pred8x8c , vp8_pred8x8c_prob_inter , sizeof(s->prob->pred8x8c)); | |
440 memcpy(s->prob->mvc , vp8_mv_default_prob , sizeof(s->prob->mvc)); | |
441 memset(&s->segmentation, 0, sizeof(s->segmentation)); | |
442 } | |
443 | |
444 if (!s->macroblocks_base || /* first frame */ | |
445 width != s->avctx->width || height != s->avctx->height) { | |
446 if ((ret = update_dimensions(s, width, height) < 0)) | |
447 return ret; | |
448 } | |
449 | |
450 vp56_init_range_decoder(c, buf, header_size); | |
451 buf += header_size; | |
452 buf_size -= header_size; | |
453 | |
454 if (s->keyframe) { | |
455 if (vp8_rac_get(c)) | |
456 av_log(s->avctx, AV_LOG_WARNING, "Unspecified colorspace\n"); | |
457 vp8_rac_get(c); // whether we can skip clamping in dsp functions | |
458 } | |
459 | |
460 if ((s->segmentation.enabled = vp8_rac_get(c))) | |
461 parse_segment_info(s); | |
462 else | |
463 s->segmentation.update_map = 0; // FIXME: move this to some init function? | |
464 | |
465 s->filter.simple = vp8_rac_get(c); | |
466 s->filter.level = vp8_rac_get_uint(c, 6); | |
467 s->filter.sharpness = vp8_rac_get_uint(c, 3); | |
468 | |
469 if ((s->lf_delta.enabled = vp8_rac_get(c))) | |
470 if (vp8_rac_get(c)) | |
471 update_lf_deltas(s); | |
472 | |
473 if (setup_partitions(s, buf, buf_size)) { | |
474 av_log(s->avctx, AV_LOG_ERROR, "Invalid partitions\n"); | |
475 return AVERROR_INVALIDDATA; | |
476 } | |
477 | |
478 get_quants(s); | |
479 | |
480 if (!s->keyframe) { | |
481 update_refs(s); | |
482 s->sign_bias[VP56_FRAME_GOLDEN] = vp8_rac_get(c); | |
483 s->sign_bias[VP56_FRAME_GOLDEN2 /* altref */] = vp8_rac_get(c); | |
484 } | |
485 | |
486 // if we aren't saving this frame's probabilities for future frames, | |
487 // make a copy of the current probabilities | |
488 if (!(s->update_probabilities = vp8_rac_get(c))) | |
489 s->prob[1] = s->prob[0]; | |
490 | |
491 s->update_last = s->keyframe || vp8_rac_get(c); | |
492 | |
493 for (i = 0; i < 4; i++) | |
494 for (j = 0; j < 8; j++) | |
495 for (k = 0; k < 3; k++) | |
496 for (l = 0; l < NUM_DCT_TOKENS-1; l++) | |
497 if (vp56_rac_get_prob(c, vp8_token_update_probs[i][j][k][l])) | |
498 s->prob->token[i][j][k][l] = vp8_rac_get_uint(c, 8); | |
499 | |
500 if ((s->mbskip_enabled = vp8_rac_get(c))) | |
501 s->prob->mbskip = vp8_rac_get_uint(c, 8); | |
502 | |
503 if (!s->keyframe) { | |
504 s->prob->intra = vp8_rac_get_uint(c, 8); | |
505 s->prob->last = vp8_rac_get_uint(c, 8); | |
506 s->prob->golden = vp8_rac_get_uint(c, 8); | |
507 | |
508 if (vp8_rac_get(c)) | |
509 for (i = 0; i < 4; i++) | |
510 s->prob->pred16x16[i] = vp8_rac_get_uint(c, 8); | |
511 if (vp8_rac_get(c)) | |
512 for (i = 0; i < 3; i++) | |
513 s->prob->pred8x8c[i] = vp8_rac_get_uint(c, 8); | |
514 | |
515 // 17.2 MV probability update | |
516 for (i = 0; i < 2; i++) | |
517 for (j = 0; j < 19; j++) | |
518 if (vp56_rac_get_prob(c, vp8_mv_update_prob[i][j])) | |
519 s->prob->mvc[i][j] = vp8_rac_get_nn(c); | |
520 } | |
521 | |
522 return 0; | |
523 } | |
524 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
525 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
526 void clamp_mv(VP8Context *s, VP56mv *dst, const VP56mv *src, int mb_x, int mb_y) |
11921 | 527 { |
528 #define MARGIN (16 << 2) | |
529 dst->x = av_clip(src->x, -((mb_x << 6) + MARGIN), | |
530 ((s->mb_width - 1 - mb_x) << 6) + MARGIN); | |
531 dst->y = av_clip(src->y, -((mb_y << 6) + MARGIN), | |
532 ((s->mb_height - 1 - mb_y) << 6) + MARGIN); | |
533 } | |
534 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
535 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
536 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
|
537 VP56mv near[2], VP56mv *best, uint8_t cnt[4]) |
11921 | 538 { |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
539 VP8Macroblock *mb_edge[3] = { mb + 2 /* top */, |
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
540 mb - 1 /* left */, |
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
541 mb + 1 /* top-left */ }; |
11921 | 542 enum { EDGE_TOP, EDGE_LEFT, EDGE_TOPLEFT }; |
543 VP56mv near_mv[4] = {{ 0 }}; | |
544 enum { CNT_ZERO, CNT_NEAREST, CNT_NEAR, CNT_SPLITMV }; | |
12217 | 545 int idx = CNT_ZERO; |
11921 | 546 int best_idx = CNT_ZERO; |
12217 | 547 int cur_sign_bias = s->sign_bias[mb->ref_frame]; |
548 int *sign_bias = s->sign_bias; | |
11921 | 549 |
550 /* Process MB on top, left and top-left */ | |
12217 | 551 #define MV_EDGE_CHECK(n)\ |
552 {\ | |
553 VP8Macroblock *edge = mb_edge[n];\ | |
554 int edge_ref = edge->ref_frame;\ | |
555 if (edge_ref != VP56_FRAME_CURRENT) {\ | |
556 uint32_t mv = AV_RN32A(&edge->mv);\ | |
557 if (mv) {\ | |
558 if (cur_sign_bias != sign_bias[edge_ref]) {\ | |
559 /* SWAR negate of the values in mv. */\ | |
12242 | 560 mv = ~mv;\ |
561 mv = ((mv&0x7fff7fff) + 0x00010001) ^ (mv&0x80008000);\ | |
12217 | 562 }\ |
563 if (!n || mv != AV_RN32A(&near_mv[idx]))\ | |
564 AV_WN32A(&near_mv[++idx], mv);\ | |
565 cnt[idx] += 1 + (n != 2);\ | |
566 } else\ | |
567 cnt[CNT_ZERO] += 1 + (n != 2);\ | |
568 }\ | |
11921 | 569 } |
12217 | 570 MV_EDGE_CHECK(0) |
571 MV_EDGE_CHECK(1) | |
572 MV_EDGE_CHECK(2) | |
11921 | 573 |
12217 | 574 /* If we have three distinct MVs, merge first and last if they're the same */ |
575 if (cnt[CNT_SPLITMV] && AV_RN32A(&near_mv[1+EDGE_TOP]) == AV_RN32A(&near_mv[1+EDGE_TOPLEFT])) | |
11921 | 576 cnt[CNT_NEAREST] += 1; |
577 | |
578 cnt[CNT_SPLITMV] = ((mb_edge[EDGE_LEFT]->mode == VP8_MVMODE_SPLIT) + | |
579 (mb_edge[EDGE_TOP]->mode == VP8_MVMODE_SPLIT)) * 2 + | |
580 (mb_edge[EDGE_TOPLEFT]->mode == VP8_MVMODE_SPLIT); | |
581 | |
582 /* Swap near and nearest if necessary */ | |
583 if (cnt[CNT_NEAR] > cnt[CNT_NEAREST]) { | |
12217 | 584 FFSWAP(uint8_t, cnt[CNT_NEAREST], cnt[CNT_NEAR]); |
585 FFSWAP( VP56mv, near_mv[CNT_NEAREST], near_mv[CNT_NEAR]); | |
11921 | 586 } |
587 | |
588 /* Choose the best mv out of 0,0 and the nearest mv */ | |
589 if (cnt[CNT_NEAREST] >= cnt[CNT_ZERO]) | |
590 best_idx = CNT_NEAREST; | |
591 | |
12246 | 592 mb->mv = near_mv[best_idx]; |
11921 | 593 near[0] = near_mv[CNT_NEAREST]; |
594 near[1] = near_mv[CNT_NEAR]; | |
595 } | |
596 | |
597 /** | |
598 * Motion vector coding, 17.1. | |
599 */ | |
600 static int read_mv_component(VP56RangeCoder *c, const uint8_t *p) | |
601 { | |
602 int x = 0; | |
603 | |
604 if (vp56_rac_get_prob(c, p[0])) { | |
605 int i; | |
606 | |
607 for (i = 0; i < 3; i++) | |
608 x += vp56_rac_get_prob(c, p[9 + i]) << i; | |
609 for (i = 9; i > 3; i--) | |
610 x += vp56_rac_get_prob(c, p[9 + i]) << i; | |
611 if (!(x & 0xFFF0) || vp56_rac_get_prob(c, p[12])) | |
612 x += 8; | |
613 } else | |
614 x = vp8_rac_get_tree(c, vp8_small_mvtree, &p[2]); | |
615 | |
616 return (x && vp56_rac_get_prob(c, p[1])) ? -x : x; | |
617 } | |
618 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
619 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
620 const uint8_t *get_submv_prob(uint32_t left, uint32_t top) |
11921 | 621 { |
12219 | 622 if (left == top) |
623 return vp8_submv_prob[4-!!left]; | |
624 if (!top) | |
11921 | 625 return vp8_submv_prob[2]; |
12219 | 626 return vp8_submv_prob[1-!!left]; |
11921 | 627 } |
628 | |
629 /** | |
630 * 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
|
631 * @returns the number of motion vectors parsed (2, 4 or 16) |
11921 | 632 */ |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
633 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
634 int decode_splitmvs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb) |
11921 | 635 { |
636 int part_idx = mb->partitioning = | |
637 vp8_rac_get_tree(c, vp8_mbsplit_tree, vp8_mbsplit_prob); | |
638 int n, num = vp8_mbsplit_count[part_idx]; | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
639 VP8Macroblock *top_mb = &mb[2]; |
12219 | 640 VP8Macroblock *left_mb = &mb[-1]; |
641 const uint8_t *mbsplits_left = vp8_mbsplits[left_mb->partitioning], | |
642 *mbsplits_top = vp8_mbsplits[top_mb->partitioning], | |
643 *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
|
644 *firstidx = vp8_mbfirstidx[part_idx]; |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
645 VP56mv *top_mv = top_mb->bmv; |
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
646 VP56mv *left_mv = left_mb->bmv; |
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
647 VP56mv *cur_mv = mb->bmv; |
11921 | 648 |
649 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
|
650 int k = firstidx[n]; |
12219 | 651 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
|
652 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
|
653 |
12219 | 654 if (!(k & 3)) |
655 left = AV_RN32A(&left_mv[mbsplits_left[k + 3]]); | |
656 else | |
657 left = AV_RN32A(&cur_mv[mbsplits_cur[k - 1]]); | |
658 if (k <= 3) | |
659 above = AV_RN32A(&top_mv[mbsplits_top[k + 12]]); | |
660 else | |
661 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
|
662 |
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
663 submv_prob = get_submv_prob(left, above); |
11921 | 664 |
665 switch (vp8_rac_get_tree(c, vp8_submv_ref_tree, submv_prob)) { | |
666 case VP8_SUBMVMODE_NEW4X4: | |
12246 | 667 mb->bmv[n].y = mb->mv.y + read_mv_component(c, s->prob->mvc[0]); |
668 mb->bmv[n].x = mb->mv.x + read_mv_component(c, s->prob->mvc[1]); | |
11921 | 669 break; |
670 case VP8_SUBMVMODE_ZERO4X4: | |
12245
ca82c3ce90c1
VP8: use AV_ZERO32 instead of AV_WN32A where relevant
darkshikari
parents:
12244
diff
changeset
|
671 AV_ZERO32(&mb->bmv[n]); |
11921 | 672 break; |
673 case VP8_SUBMVMODE_LEFT4X4: | |
12219 | 674 AV_WN32A(&mb->bmv[n], left); |
11921 | 675 break; |
676 case VP8_SUBMVMODE_TOP4X4: | |
12219 | 677 AV_WN32A(&mb->bmv[n], above); |
11921 | 678 break; |
679 } | |
11990
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
680 } |
11921 | 681 |
11990
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
682 return num; |
11921 | 683 } |
684 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
685 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
686 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
|
687 int stride, int keyframe) |
11921 | 688 { |
12221 | 689 int x, y, t, l, i; |
11921 | 690 |
12221 | 691 if (keyframe) { |
692 const uint8_t *ctx; | |
693 for (y = 0; y < 4; y++) { | |
694 for (x = 0; x < 4; x++) { | |
11921 | 695 t = intra4x4[x - stride]; |
696 l = intra4x4[x - 1]; | |
697 ctx = vp8_pred4x4_prob_intra[t][l]; | |
12221 | 698 intra4x4[x] = vp8_rac_get_tree(c, vp8_pred4x4_tree, ctx); |
11921 | 699 } |
12221 | 700 intra4x4 += stride; |
11921 | 701 } |
12221 | 702 } else { |
703 for (i = 0; i < 16; i++) | |
704 intra4x4[i] = vp8_rac_get_tree(c, vp8_pred4x4_tree, vp8_pred4x4_prob_inter); | |
11921 | 705 } |
706 } | |
707 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
708 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
709 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
|
710 uint8_t *intra4x4, uint8_t *segment) |
11921 | 711 { |
712 VP56RangeCoder *c = &s->c; | |
713 | |
714 if (s->segmentation.update_map) | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
715 *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
|
716 s->segment = *segment; |
11921 | 717 |
718 mb->skip = s->mbskip_enabled ? vp56_rac_get_prob(c, s->prob->mbskip) : 0; | |
719 | |
720 if (s->keyframe) { | |
721 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_intra, vp8_pred16x16_prob_intra); | |
722 | |
723 if (mb->mode == MODE_I4x4) { | |
724 decode_intra4x4_modes(c, intra4x4, s->b4_stride, 1); | |
725 } else | |
726 fill_rectangle(intra4x4, 4, 4, s->b4_stride, vp8_pred4x4_mode[mb->mode], 1); | |
727 | |
728 s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, vp8_pred8x8c_prob_intra); | |
729 mb->ref_frame = VP56_FRAME_CURRENT; | |
730 } else if (vp56_rac_get_prob(c, s->prob->intra)) { | |
731 VP56mv near[2], best; | |
12217 | 732 uint8_t cnt[4] = { 0 }; |
11921 | 733 uint8_t p[4]; |
734 | |
735 // inter MB, 16.2 | |
736 if (vp56_rac_get_prob(c, s->prob->last)) | |
737 mb->ref_frame = vp56_rac_get_prob(c, s->prob->golden) ? | |
738 VP56_FRAME_GOLDEN2 /* altref */ : VP56_FRAME_GOLDEN; | |
739 else | |
740 mb->ref_frame = VP56_FRAME_PREVIOUS; | |
12231 | 741 s->ref_count[mb->ref_frame-1]++; |
11921 | 742 |
743 // motion vectors, 16.3 | |
744 find_near_mvs(s, mb, mb_x, mb_y, near, &best, cnt); | |
12217 | 745 p[0] = vp8_mode_contexts[cnt[0]][0]; |
746 p[1] = vp8_mode_contexts[cnt[1]][1]; | |
747 p[2] = vp8_mode_contexts[cnt[2]][2]; | |
748 p[3] = vp8_mode_contexts[cnt[3]][3]; | |
11921 | 749 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_mvinter, p); |
750 switch (mb->mode) { | |
751 case VP8_MVMODE_SPLIT: | |
12246 | 752 clamp_mv(s, &mb->mv, &mb->mv, mb_x, mb_y); |
753 mb->mv = mb->bmv[decode_splitmvs(s, c, mb) - 1]; | |
11921 | 754 break; |
755 case VP8_MVMODE_ZERO: | |
12245
ca82c3ce90c1
VP8: use AV_ZERO32 instead of AV_WN32A where relevant
darkshikari
parents:
12244
diff
changeset
|
756 AV_ZERO32(&mb->mv); |
11921 | 757 break; |
758 case VP8_MVMODE_NEAREST: | |
759 clamp_mv(s, &mb->mv, &near[0], mb_x, mb_y); | |
760 break; | |
761 case VP8_MVMODE_NEAR: | |
762 clamp_mv(s, &mb->mv, &near[1], mb_x, mb_y); | |
763 break; | |
764 case VP8_MVMODE_NEW: | |
12246 | 765 clamp_mv(s, &mb->mv, &mb->mv, mb_x, mb_y); |
766 mb->mv.y += + read_mv_component(c, s->prob->mvc[0]); | |
767 mb->mv.x += + read_mv_component(c, s->prob->mvc[1]); | |
11921 | 768 break; |
769 } | |
770 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
|
771 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
|
772 mb->bmv[0] = mb->mv; |
11921 | 773 } |
774 } else { | |
775 // intra MB, 16.1 | |
776 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_inter, s->prob->pred16x16); | |
777 | |
12220
0f635b1f7861
Avoid useless fill_rectangle in P-frames in VP8
darkshikari
parents:
12219
diff
changeset
|
778 if (mb->mode == MODE_I4x4) |
12221 | 779 decode_intra4x4_modes(c, intra4x4, 4, 0); |
11921 | 780 |
781 s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, s->prob->pred8x8c); | |
782 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
|
783 mb->partitioning = VP8_SPLITMVMODE_NONE; |
12245
ca82c3ce90c1
VP8: use AV_ZERO32 instead of AV_WN32A where relevant
darkshikari
parents:
12244
diff
changeset
|
784 AV_ZERO32(&mb->bmv[0]); |
11921 | 785 } |
786 } | |
787 | |
788 /** | |
12115 | 789 * @param c arithmetic bitstream reader context |
790 * @param block destination for block coefficients | |
791 * @param probs probabilities to use when reading trees from the bitstream | |
11921 | 792 * @param i initial coeff index, 0 unless a separate DC block is coded |
793 * @param zero_nhood the initial prediction context for number of surrounding | |
794 * 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
|
795 * @param qmul array holding the dc/ac dequant factor at position 0/1 |
11921 | 796 * @return 0 if no coeffs were decoded |
797 * otherwise, the index of the last coeff decoded plus one | |
798 */ | |
799 static int decode_block_coeffs(VP56RangeCoder *c, DCTELEM block[16], | |
800 uint8_t probs[8][3][NUM_DCT_TOKENS-1], | |
801 int i, int zero_nhood, int16_t qmul[2]) | |
802 { | |
12253
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
803 uint8_t *token_prob; |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
804 int nonzero = 0; |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
805 int coeff; |
11921 | 806 |
12253
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
807 do { |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
808 token_prob = probs[vp8_coeff_band[i]][zero_nhood]; |
11921 | 809 |
12253
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
810 if (!vp56_rac_get_prob_branchy(c, token_prob[0])) // DCT_EOB |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
811 return nonzero; |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
812 |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
813 skip_eob: |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
814 if (!vp56_rac_get_prob_branchy(c, token_prob[1])) { // DCT_0 |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
815 zero_nhood = 0; |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
816 token_prob = probs[vp8_coeff_band[++i]][0]; |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
817 if (i < 16) |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
818 goto skip_eob; |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
819 return nonzero; // invalid input; blocks should end with EOB |
11921 | 820 } |
821 | |
12253
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
822 if (!vp56_rac_get_prob_branchy(c, token_prob[2])) { // DCT_1 |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
823 coeff = 1; |
11921 | 824 zero_nhood = 1; |
12253
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
825 } else { |
11921 | 826 zero_nhood = 2; |
827 | |
12253
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
828 if (!vp56_rac_get_prob_branchy(c, token_prob[3])) { // DCT 2,3,4 |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
829 coeff = vp56_rac_get_prob(c, token_prob[4]); |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
830 if (coeff) |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
831 coeff += vp56_rac_get_prob(c, token_prob[5]); |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
832 coeff += 2; |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
833 } else { |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
834 // DCT_CAT* |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
835 if (!vp56_rac_get_prob_branchy(c, token_prob[6])) { |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
836 if (!vp56_rac_get_prob_branchy(c, token_prob[7])) { // DCT_CAT1 |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
837 coeff = 5 + vp56_rac_get_prob(c, vp8_dct_cat1_prob[0]); |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
838 } else { // DCT_CAT2 |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
839 coeff = 7; |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
840 coeff += vp56_rac_get_prob(c, vp8_dct_cat2_prob[0]) << 1; |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
841 coeff += vp56_rac_get_prob(c, vp8_dct_cat2_prob[1]); |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
842 } |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
843 } else { // DCT_CAT3 and up |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
844 int a = vp56_rac_get_prob(c, token_prob[8]); |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
845 int b = vp56_rac_get_prob(c, token_prob[9+a]); |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
846 int cat = (a<<1) + b; |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
847 coeff = 3 + (8<<cat); |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
848 coeff += vp8_rac_get_coeff(c, vp8_dct_cat_prob[cat]); |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
849 } |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
850 } |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
851 } |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
852 |
11921 | 853 // todo: full [16] qmat? load into register? |
12253
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
854 block[zigzag_scan[i]] = (vp8_rac_get(c) ? -coeff : coeff) * qmul[!!i]; |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
855 nonzero = ++i; |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
856 } while (i < 16); |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
857 |
11921 | 858 return nonzero; |
859 } | |
860 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
861 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
862 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
|
863 uint8_t t_nnz[9], uint8_t l_nnz[9]) |
11921 | 864 { |
865 LOCAL_ALIGNED_16(DCTELEM, dc,[16]); | |
866 int i, x, y, luma_start = 0, luma_ctx = 3; | |
867 int nnz_pred, nnz, nnz_total = 0; | |
12224
5b7d690b761b
VP8: Don't store segment in macroblock struct anymore.
darkshikari
parents:
12223
diff
changeset
|
868 int segment = s->segment; |
11921 | 869 |
870 if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) { | |
871 AV_ZERO128(dc); | |
872 AV_ZERO128(dc+8); | |
873 nnz_pred = t_nnz[8] + l_nnz[8]; | |
874 | |
875 // decode DC values and do hadamard | |
876 nnz = decode_block_coeffs(c, dc, s->prob->token[1], 0, nnz_pred, | |
877 s->qmat[segment].luma_dc_qmul); | |
878 l_nnz[8] = t_nnz[8] = !!nnz; | |
879 nnz_total += nnz; | |
880 s->vp8dsp.vp8_luma_dc_wht(s->block, dc); | |
881 luma_start = 1; | |
882 luma_ctx = 0; | |
883 } | |
884 | |
885 // luma blocks | |
886 for (y = 0; y < 4; y++) | |
887 for (x = 0; x < 4; x++) { | |
888 nnz_pred = l_nnz[y] + t_nnz[x]; | |
889 nnz = decode_block_coeffs(c, s->block[y][x], s->prob->token[luma_ctx], luma_start, | |
890 nnz_pred, s->qmat[segment].luma_qmul); | |
891 // nnz+luma_start may be one more than the actual last index, but we don't care | |
892 s->non_zero_count_cache[y][x] = nnz + luma_start; | |
893 t_nnz[x] = l_nnz[y] = !!nnz; | |
894 nnz_total += nnz; | |
895 } | |
896 | |
897 // chroma blocks | |
898 // TODO: what to do about dimensions? 2nd dim for luma is x, | |
899 // but for chroma it's (y<<1)|x | |
900 for (i = 4; i < 6; i++) | |
901 for (y = 0; y < 2; y++) | |
902 for (x = 0; x < 2; x++) { | |
903 nnz_pred = l_nnz[i+2*y] + t_nnz[i+2*x]; | |
904 nnz = decode_block_coeffs(c, s->block[i][(y<<1)+x], s->prob->token[2], 0, | |
905 nnz_pred, s->qmat[segment].chroma_qmul); | |
906 s->non_zero_count_cache[i][(y<<1)+x] = nnz; | |
907 t_nnz[i+2*x] = l_nnz[i+2*y] = !!nnz; | |
908 nnz_total += nnz; | |
909 } | |
910 | |
911 // if there were no coded coeffs despite the macroblock not being marked skip, | |
912 // we MUST not do the inner loop filter and should not do IDCT | |
913 // Since skip isn't used for bitstream prediction, just manually set it. | |
914 if (!nnz_total) | |
915 mb->skip = 1; | |
916 } | |
917 | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
918 static av_always_inline |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
919 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
|
920 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
|
921 { |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
922 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
|
923 if (!simple) { |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
924 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
|
925 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
|
926 } |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
927 } |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
928 |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
929 static av_always_inline |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
930 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
|
931 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
|
932 int simple, int xchg) |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
933 { |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
934 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
|
935 src_y -= linesize; |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
936 src_cb -= uvlinesize; |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
937 src_cr -= uvlinesize; |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
938 |
12202 | 939 #define XCHG(a,b,xchg) do { \ |
940 if (xchg) AV_SWAP64(b,a); \ | |
941 else AV_COPY64(b,a); \ | |
942 } while (0) | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
943 |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
944 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
|
945 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
|
946 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
|
947 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
|
948 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
|
949 |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
950 // 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
|
951 // 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
|
952 if (!simple || !mb_y) { |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
953 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
|
954 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
|
955 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
|
956 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
|
957 } |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
958 } |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
959 |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
960 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
961 int check_intra_pred_mode(int mode, int mb_x, int mb_y) |
11921 | 962 { |
963 if (mode == DC_PRED8x8) { | |
12243
788445bf10c0
VP8: shave a few clocks off check_intra_pred_mode
darkshikari
parents:
12242
diff
changeset
|
964 if (!mb_x) { |
788445bf10c0
VP8: shave a few clocks off check_intra_pred_mode
darkshikari
parents:
12242
diff
changeset
|
965 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
|
966 } else if (!mb_y) { |
12244 | 967 mode = LEFT_DC_PRED8x8; |
12243
788445bf10c0
VP8: shave a few clocks off check_intra_pred_mode
darkshikari
parents:
12242
diff
changeset
|
968 } |
11921 | 969 } |
970 return mode; | |
971 } | |
972 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
973 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
974 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
|
975 uint8_t *intra4x4, int mb_x, int mb_y) |
11921 | 976 { |
977 int x, y, mode, nnz, tr; | |
978 | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
979 // 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
|
980 // 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
|
981 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
|
982 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
|
983 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
|
984 s->filter.simple, 1); |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
985 |
11921 | 986 if (mb->mode < MODE_I4x4) { |
987 mode = check_intra_pred_mode(mb->mode, mb_x, mb_y); | |
988 s->hpc.pred16x16[mode](dst[0], s->linesize); | |
989 } else { | |
990 uint8_t *ptr = dst[0]; | |
12221 | 991 int stride = s->keyframe ? s->b4_stride : 4; |
11921 | 992 |
993 // all blocks on the right edge of the macroblock use bottom edge | |
994 // the top macroblock for their topright edge | |
995 uint8_t *tr_right = ptr - s->linesize + 16; | |
996 | |
997 // if we're on the right edge of the frame, said edge is extended | |
998 // from the top macroblock | |
999 if (mb_x == s->mb_width-1) { | |
1000 tr = tr_right[-1]*0x01010101; | |
1001 tr_right = (uint8_t *)&tr; | |
1002 } | |
1003 | |
12234
bba849c2a113
VP8: avoid a memset for non-i4x4 blocks with no coefficients
darkshikari
parents:
12233
diff
changeset
|
1004 if (mb->skip) |
bba849c2a113
VP8: avoid a memset for non-i4x4 blocks with no coefficients
darkshikari
parents:
12233
diff
changeset
|
1005 AV_ZERO128(s->non_zero_count_cache); |
bba849c2a113
VP8: avoid a memset for non-i4x4 blocks with no coefficients
darkshikari
parents:
12233
diff
changeset
|
1006 |
11921 | 1007 for (y = 0; y < 4; y++) { |
1008 uint8_t *topright = ptr + 4 - s->linesize; | |
1009 for (x = 0; x < 4; x++) { | |
1010 if (x == 3) | |
1011 topright = tr_right; | |
1012 | |
12221 | 1013 s->hpc.pred4x4[intra4x4[x]](ptr+4*x, topright, s->linesize); |
11921 | 1014 |
1015 nnz = s->non_zero_count_cache[y][x]; | |
1016 if (nnz) { | |
1017 if (nnz == 1) | |
1018 s->vp8dsp.vp8_idct_dc_add(ptr+4*x, s->block[y][x], s->linesize); | |
1019 else | |
1020 s->vp8dsp.vp8_idct_add(ptr+4*x, s->block[y][x], s->linesize); | |
1021 } | |
1022 topright += 4; | |
1023 } | |
1024 | |
1025 ptr += 4*s->linesize; | |
12221 | 1026 intra4x4 += stride; |
11921 | 1027 } |
1028 } | |
1029 | |
1030 mode = check_intra_pred_mode(s->chroma_pred_mode, mb_x, mb_y); | |
1031 s->hpc.pred8x8[mode](dst[1], s->uvlinesize); | |
1032 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
|
1033 |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
1034 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
|
1035 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
|
1036 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
|
1037 s->filter.simple, 0); |
11921 | 1038 } |
1039 | |
1040 /** | |
1041 * Generic MC function. | |
1042 * | |
1043 * @param s VP8 decoding context | |
1044 * @param luma 1 for luma (Y) planes, 0 for chroma (Cb/Cr) planes | |
1045 * @param dst target buffer for block data at block position | |
1046 * @param src reference picture buffer at origin (0, 0) | |
1047 * @param mv motion vector (relative to block position) to get pixel data from | |
1048 * @param x_off horizontal position of block from origin (0, 0) | |
1049 * @param y_off vertical position of block from origin (0, 0) | |
1050 * @param block_w width of block (16, 8 or 4) | |
1051 * @param block_h height of block (always same as block_w) | |
1052 * @param width width of src/dst plane data | |
1053 * @param height height of src/dst plane data | |
1054 * @param linesize size of a single line of plane data, including padding | |
12115 | 1055 * @param mc_func motion compensation function pointers (bilinear or sixtap MC) |
11921 | 1056 */ |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1057 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1058 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
|
1059 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
|
1060 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
|
1061 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
|
1062 vp8_mc_func mc_func[3][3]) |
11921 | 1063 { |
12218 | 1064 if (AV_RN32A(mv)) { |
1065 static const uint8_t idx[8] = { 0, 1, 2, 1, 2, 1, 2, 1 }; | |
1066 int mx = (mv->x << luma)&7, mx_idx = idx[mx]; | |
1067 int my = (mv->y << luma)&7, my_idx = idx[my]; | |
11921 | 1068 |
12218 | 1069 x_off += mv->x >> (3 - luma); |
1070 y_off += mv->y >> (3 - luma); | |
11921 | 1071 |
12218 | 1072 // edge emulation |
1073 src += y_off * linesize + x_off; | |
1074 if (x_off < 2 || x_off >= width - block_w - 3 || | |
1075 y_off < 2 || y_off >= height - block_h - 3) { | |
1076 ff_emulated_edge_mc(s->edge_emu_buffer, src - 2 * linesize - 2, linesize, | |
1077 block_w + 5, block_h + 5, | |
1078 x_off - 2, y_off - 2, width, height); | |
1079 src = s->edge_emu_buffer + 2 + linesize * 2; | |
1080 } | |
1081 mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my); | |
1082 } else | |
1083 mc_func[0][0](dst, linesize, src + y_off * linesize + x_off, linesize, block_h, 0, 0); | |
11921 | 1084 } |
1085 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1086 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1087 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
|
1088 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
|
1089 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
|
1090 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
|
1091 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
|
1092 { |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1093 VP56mv uvmv = *mv; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1094 |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1095 /* Y */ |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1096 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
|
1097 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
|
1098 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
|
1099 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
|
1100 |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1101 /* U/V */ |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1102 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
|
1103 uvmv.x &= ~7; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1104 uvmv.y &= ~7; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1105 } |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1106 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
|
1107 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
|
1108 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
|
1109 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
|
1110 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
|
1111 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
|
1112 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
|
1113 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
|
1114 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
|
1115 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
|
1116 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
|
1117 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
|
1118 } |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1119 |
12215 | 1120 /* Fetch pixels for estimated mv 4 macroblocks ahead. |
1121 * 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
|
1122 static av_always_inline void prefetch_motion(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, int mb_xy, int ref) |
12215 | 1123 { |
12237 | 1124 /* Don't prefetch refs that haven't been used very often this frame. */ |
1125 if (s->ref_count[ref-1] > (mb_xy >> 5)) { | |
12231 | 1126 int x_off = mb_x << 4, y_off = mb_y << 4; |
1127 int mx = mb->mv.x + x_off + 8; | |
1128 int my = mb->mv.y + y_off; | |
1129 uint8_t **src= s->framep[ref]->data; | |
1130 int off= mx + (my + (mb_x&3)*4)*s->linesize + 64; | |
1131 s->dsp.prefetch(src[0]+off, s->linesize, 4); | |
1132 off= (mx>>1) + ((my>>1) + (mb_x&7))*s->uvlinesize + 64; | |
1133 s->dsp.prefetch(src[1]+off, src[2]-src[1], 2); | |
1134 } | |
12215 | 1135 } |
1136 | |
11921 | 1137 /** |
1138 * Apply motion vectors to prediction buffer, chapter 18. | |
1139 */ | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1140 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1141 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
|
1142 int mb_x, int mb_y) |
11921 | 1143 { |
1144 int x_off = mb_x << 4, y_off = mb_y << 4; | |
1145 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
|
1146 AVFrame *ref = s->framep[mb->ref_frame]; |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1147 VP56mv *bmv = mb->bmv; |
11921 | 1148 |
1149 if (mb->mode < VP8_MVMODE_SPLIT) { | |
12228
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1150 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
|
1151 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
|
1152 } 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
|
1153 case VP8_SPLITMVMODE_4x4: { |
11921 | 1154 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
|
1155 VP56mv uvmv; |
11921 | 1156 |
1157 /* Y */ | |
1158 for (y = 0; y < 4; y++) { | |
1159 for (x = 0; x < 4; x++) { | |
1160 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
|
1161 ref->data[0], &bmv[4*y + x], |
11921 | 1162 4*x + x_off, 4*y + y_off, 4, 4, |
1163 width, height, s->linesize, | |
11974 | 1164 s->put_pixels_tab[2]); |
11921 | 1165 } |
1166 } | |
1167 | |
1168 /* U/V */ | |
1169 x_off >>= 1; y_off >>= 1; width >>= 1; height >>= 1; | |
1170 for (y = 0; y < 2; y++) { | |
1171 for (x = 0; x < 2; x++) { | |
1172 uvmv.x = mb->bmv[ 2*y * 4 + 2*x ].x + | |
1173 mb->bmv[ 2*y * 4 + 2*x+1].x + | |
1174 mb->bmv[(2*y+1) * 4 + 2*x ].x + | |
1175 mb->bmv[(2*y+1) * 4 + 2*x+1].x; | |
1176 uvmv.y = mb->bmv[ 2*y * 4 + 2*x ].y + | |
1177 mb->bmv[ 2*y * 4 + 2*x+1].y + | |
1178 mb->bmv[(2*y+1) * 4 + 2*x ].y + | |
1179 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
|
1180 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
|
1181 uvmv.y = (uvmv.y + 2 + (uvmv.y >> (INT_BIT-1))) >> 2; |
11921 | 1182 if (s->profile == 3) { |
1183 uvmv.x &= ~7; | |
1184 uvmv.y &= ~7; | |
1185 } | |
1186 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
|
1187 ref->data[1], &uvmv, |
11921 | 1188 4*x + x_off, 4*y + y_off, 4, 4, |
1189 width, height, s->uvlinesize, | |
11974 | 1190 s->put_pixels_tab[2]); |
11921 | 1191 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
|
1192 ref->data[2], &uvmv, |
11921 | 1193 4*x + x_off, 4*y + y_off, 4, 4, |
1194 width, height, s->uvlinesize, | |
11974 | 1195 s->put_pixels_tab[2]); |
11921 | 1196 } |
1197 } | |
11989
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1198 break; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1199 } |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1200 case VP8_SPLITMVMODE_16x8: |
12228
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1201 vp8_mc_part(s, dst, ref, x_off, y_off, |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1202 0, 0, 16, 8, width, height, &bmv[0]); |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1203 vp8_mc_part(s, dst, ref, x_off, y_off, |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1204 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
|
1205 break; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1206 case VP8_SPLITMVMODE_8x16: |
12228
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1207 vp8_mc_part(s, dst, ref, x_off, y_off, |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1208 0, 0, 8, 16, width, height, &bmv[0]); |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1209 vp8_mc_part(s, dst, ref, x_off, y_off, |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1210 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
|
1211 break; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1212 case VP8_SPLITMVMODE_8x8: |
12228
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1213 vp8_mc_part(s, dst, ref, x_off, y_off, |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1214 0, 0, 8, 8, width, height, &bmv[0]); |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1215 vp8_mc_part(s, dst, ref, x_off, y_off, |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1216 8, 0, 8, 8, width, height, &bmv[1]); |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1217 vp8_mc_part(s, dst, ref, x_off, y_off, |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1218 0, 8, 8, 8, width, height, &bmv[2]); |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1219 vp8_mc_part(s, dst, ref, x_off, y_off, |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1220 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
|
1221 break; |
11921 | 1222 } |
1223 } | |
1224 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1225 static av_always_inline void idct_mb(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb) |
11921 | 1226 { |
12240 | 1227 int x, y, ch; |
11921 | 1228 |
12238 | 1229 if (mb->mode != MODE_I4x4) { |
1230 uint8_t *y_dst = dst[0]; | |
11921 | 1231 for (y = 0; y < 4; y++) { |
12240 | 1232 uint32_t nnz4 = AV_RN32A(s->non_zero_count_cache[y]); |
1233 if (nnz4) { | |
1234 if (nnz4&~0x01010101) { | |
12238 | 1235 for (x = 0; x < 4; x++) { |
12240 | 1236 int nnz = s->non_zero_count_cache[y][x]; |
12238 | 1237 if (nnz) { |
1238 if (nnz == 1) | |
1239 s->vp8dsp.vp8_idct_dc_add(y_dst+4*x, s->block[y][x], s->linesize); | |
1240 else | |
1241 s->vp8dsp.vp8_idct_add(y_dst+4*x, s->block[y][x], s->linesize); | |
1242 } | |
1243 } | |
1244 } else { | |
12241
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1245 s->vp8dsp.vp8_idct_dc_add4y(y_dst, s->block[y], s->linesize); |
11921 | 1246 } |
1247 } | |
1248 y_dst += 4*s->linesize; | |
1249 } | |
12238 | 1250 } |
11921 | 1251 |
12238 | 1252 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
|
1253 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
|
1254 if (nnz4) { |
12238 | 1255 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
|
1256 if (nnz4&~0x01010101) { |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1257 for (y = 0; y < 2; y++) { |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1258 for (x = 0; x < 2; x++) { |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1259 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
|
1260 if (nnz) { |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1261 if (nnz == 1) |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1262 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
|
1263 else |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1264 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
|
1265 } |
12238 | 1266 } |
12241
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1267 ch_dst += 4*s->uvlinesize; |
12238 | 1268 } |
12241
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1269 } else { |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1270 s->vp8dsp.vp8_idct_dc_add4uv(ch_dst, s->block[4+ch], s->uvlinesize); |
11921 | 1271 } |
1272 } | |
1273 } | |
1274 } | |
1275 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1276 static av_always_inline void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb, VP8FilterStrength *f ) |
11921 | 1277 { |
1278 int interior_limit, filter_level; | |
1279 | |
1280 if (s->segmentation.enabled) { | |
12224
5b7d690b761b
VP8: Don't store segment in macroblock struct anymore.
darkshikari
parents:
12223
diff
changeset
|
1281 filter_level = s->segmentation.filter_level[s->segment]; |
11921 | 1282 if (!s->segmentation.absolute_vals) |
1283 filter_level += s->filter.level; | |
1284 } else | |
1285 filter_level = s->filter.level; | |
1286 | |
1287 if (s->lf_delta.enabled) { | |
1288 filter_level += s->lf_delta.ref[mb->ref_frame]; | |
1289 | |
1290 if (mb->ref_frame == VP56_FRAME_CURRENT) { | |
1291 if (mb->mode == MODE_I4x4) | |
1292 filter_level += s->lf_delta.mode[0]; | |
1293 } else { | |
1294 if (mb->mode == VP8_MVMODE_ZERO) | |
1295 filter_level += s->lf_delta.mode[1]; | |
1296 else if (mb->mode == VP8_MVMODE_SPLIT) | |
1297 filter_level += s->lf_delta.mode[3]; | |
1298 else | |
1299 filter_level += s->lf_delta.mode[2]; | |
1300 } | |
1301 } | |
1302 filter_level = av_clip(filter_level, 0, 63); | |
1303 | |
1304 interior_limit = filter_level; | |
1305 if (s->filter.sharpness) { | |
1306 interior_limit >>= s->filter.sharpness > 4 ? 2 : 1; | |
1307 interior_limit = FFMIN(interior_limit, 9 - s->filter.sharpness); | |
1308 } | |
1309 interior_limit = FFMAX(interior_limit, 1); | |
1310 | |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1311 f->filter_level = filter_level; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1312 f->inner_limit = interior_limit; |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1313 f->inner_filter = !mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT; |
11921 | 1314 } |
1315 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1316 static av_always_inline void filter_mb(VP8Context *s, uint8_t *dst[3], VP8FilterStrength *f, int mb_x, int mb_y) |
11921 | 1317 { |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1318 int mbedge_lim, bedge_lim, hev_thresh; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1319 int filter_level = f->filter_level; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1320 int inner_limit = f->inner_limit; |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1321 int inner_filter = f->inner_filter; |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1322 int linesize = s->linesize; |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1323 int uvlinesize = s->uvlinesize; |
11921 | 1324 |
1325 if (!filter_level) | |
1326 return; | |
1327 | |
12081
812e23197d64
VP8: Move calculation of outer filter limit out of dsp functions for normal
conrad
parents:
12062
diff
changeset
|
1328 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
|
1329 bedge_lim = 2* filter_level + inner_limit; |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1330 hev_thresh = filter_level >= 15; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1331 |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1332 if (s->keyframe) { |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1333 if (filter_level >= 40) |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1334 hev_thresh = 2; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1335 } else { |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1336 if (filter_level >= 40) |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1337 hev_thresh = 3; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1338 else if (filter_level >= 20) |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1339 hev_thresh = 2; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1340 } |
12081
812e23197d64
VP8: Move calculation of outer filter limit out of dsp functions for normal
conrad
parents:
12062
diff
changeset
|
1341 |
11921 | 1342 if (mb_x) { |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1343 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
|
1344 mbedge_lim, inner_limit, hev_thresh); |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1345 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
|
1346 mbedge_lim, inner_limit, hev_thresh); |
11921 | 1347 } |
1348 | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1349 if (inner_filter) { |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1350 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
|
1351 inner_limit, hev_thresh); |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1352 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
|
1353 inner_limit, hev_thresh); |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1354 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
|
1355 inner_limit, hev_thresh); |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1356 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
|
1357 uvlinesize, bedge_lim, |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1358 inner_limit, hev_thresh); |
11921 | 1359 } |
1360 | |
1361 if (mb_y) { | |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1362 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
|
1363 mbedge_lim, inner_limit, hev_thresh); |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1364 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
|
1365 mbedge_lim, inner_limit, hev_thresh); |
11921 | 1366 } |
1367 | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1368 if (inner_filter) { |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1369 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
|
1370 linesize, bedge_lim, |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1371 inner_limit, hev_thresh); |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1372 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
|
1373 linesize, bedge_lim, |
12194
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1374 inner_limit, hev_thresh); |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1375 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
|
1376 linesize, bedge_lim, |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1377 inner_limit, hev_thresh); |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1378 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
|
1379 dst[2] + 4 * uvlinesize, |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1380 uvlinesize, bedge_lim, |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1381 inner_limit, hev_thresh); |
11921 | 1382 } |
1383 } | |
1384 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1385 static av_always_inline void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8FilterStrength *f, int mb_x, int mb_y) |
11921 | 1386 { |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1387 int mbedge_lim, bedge_lim; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1388 int filter_level = f->filter_level; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1389 int inner_limit = f->inner_limit; |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1390 int inner_filter = f->inner_filter; |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1391 int linesize = s->linesize; |
11921 | 1392 |
1393 if (!filter_level) | |
1394 return; | |
1395 | |
1396 mbedge_lim = 2*(filter_level+2) + inner_limit; | |
1397 bedge_lim = 2* filter_level + inner_limit; | |
1398 | |
1399 if (mb_x) | |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1400 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
|
1401 if (inner_filter) { |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1402 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
|
1403 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
|
1404 s->vp8dsp.vp8_h_loop_filter_simple(dst+12, linesize, bedge_lim); |
11921 | 1405 } |
1406 | |
1407 if (mb_y) | |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1408 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
|
1409 if (inner_filter) { |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1410 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
|
1411 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
|
1412 s->vp8dsp.vp8_v_loop_filter_simple(dst+12*linesize, linesize, bedge_lim); |
11921 | 1413 } |
1414 } | |
1415 | |
1416 static void filter_mb_row(VP8Context *s, int mb_y) | |
1417 { | |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1418 VP8FilterStrength *f = s->filter_strength; |
11921 | 1419 uint8_t *dst[3] = { |
1420 s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize, | |
1421 s->framep[VP56_FRAME_CURRENT]->data[1] + 8*mb_y*s->uvlinesize, | |
1422 s->framep[VP56_FRAME_CURRENT]->data[2] + 8*mb_y*s->uvlinesize | |
1423 }; | |
1424 int mb_x; | |
1425 | |
1426 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
|
1427 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
|
1428 filter_mb(s, dst, f++, mb_x, mb_y); |
11921 | 1429 dst[0] += 16; |
1430 dst[1] += 8; | |
1431 dst[2] += 8; | |
1432 } | |
1433 } | |
1434 | |
1435 static void filter_mb_row_simple(VP8Context *s, int mb_y) | |
1436 { | |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1437 VP8FilterStrength *f = s->filter_strength; |
11921 | 1438 uint8_t *dst = s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize; |
1439 int mb_x; | |
1440 | |
1441 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
|
1442 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
|
1443 filter_mb_simple(s, dst, f++, mb_x, mb_y); |
11921 | 1444 dst += 16; |
1445 } | |
1446 } | |
1447 | |
1448 static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size, | |
1449 AVPacket *avpkt) | |
1450 { | |
1451 VP8Context *s = avctx->priv_data; | |
1452 int ret, mb_x, mb_y, i, y, referenced; | |
1453 enum AVDiscard skip_thresh; | |
12232 | 1454 AVFrame *curframe = NULL; |
11921 | 1455 |
1456 if ((ret = decode_frame_header(s, avpkt->data, avpkt->size)) < 0) | |
1457 return ret; | |
1458 | |
1459 referenced = s->update_last || s->update_golden == VP56_FRAME_CURRENT | |
1460 || s->update_altref == VP56_FRAME_CURRENT; | |
1461 | |
1462 skip_thresh = !referenced ? AVDISCARD_NONREF : | |
1463 !s->keyframe ? AVDISCARD_NONKEY : AVDISCARD_ALL; | |
1464 | |
1465 if (avctx->skip_frame >= skip_thresh) { | |
1466 s->invisible = 1; | |
1467 goto skip_decode; | |
1468 } | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
1469 s->deblock_filter = s->filter.level && avctx->skip_loop_filter < skip_thresh; |
11921 | 1470 |
1471 for (i = 0; i < 4; i++) | |
1472 if (&s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] && | |
1473 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] && | |
1474 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) { | |
1475 curframe = s->framep[VP56_FRAME_CURRENT] = &s->frames[i]; | |
1476 break; | |
1477 } | |
1478 if (curframe->data[0]) | |
1479 avctx->release_buffer(avctx, curframe); | |
1480 | |
1481 curframe->key_frame = s->keyframe; | |
1482 curframe->pict_type = s->keyframe ? FF_I_TYPE : FF_P_TYPE; | |
1483 curframe->reference = referenced ? 3 : 0; | |
1484 if ((ret = avctx->get_buffer(avctx, curframe))) { | |
1485 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed!\n"); | |
1486 return ret; | |
1487 } | |
1488 | |
1489 // Given that arithmetic probabilities are updated every frame, it's quite likely | |
1490 // that the values we have on a random interframe are complete junk if we didn't | |
1491 // start decode on a keyframe. So just don't display anything rather than junk. | |
1492 if (!s->keyframe && (!s->framep[VP56_FRAME_PREVIOUS] || | |
1493 !s->framep[VP56_FRAME_GOLDEN] || | |
1494 !s->framep[VP56_FRAME_GOLDEN2])) { | |
1495 av_log(avctx, AV_LOG_WARNING, "Discarding interframe without a prior keyframe!\n"); | |
1496 return AVERROR_INVALIDDATA; | |
1497 } | |
1498 | |
1499 s->linesize = curframe->linesize[0]; | |
1500 s->uvlinesize = curframe->linesize[1]; | |
1501 | |
1502 if (!s->edge_emu_buffer) | |
1503 s->edge_emu_buffer = av_malloc(21*s->linesize); | |
1504 | |
1505 memset(s->top_nnz, 0, s->mb_width*sizeof(*s->top_nnz)); | |
1506 | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1507 /* 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
|
1508 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
|
1509 |
11921 | 1510 // 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
|
1511 memset(s->top_border, 127, (s->mb_width+1)*sizeof(*s->top_border)); |
12231 | 1512 memset(s->ref_count, 0, sizeof(s->ref_count)); |
11921 | 1513 |
1514 for (mb_y = 0; mb_y < s->mb_height; mb_y++) { | |
1515 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
|
1516 VP8Macroblock *mb = s->macroblocks + (s->mb_height - mb_y - 1)*2; |
11921 | 1517 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
|
1518 uint8_t *segment_map = s->segmentation_map + mb_y*s->mb_stride; |
12237 | 1519 int mb_xy = mb_y * s->mb_stride; |
11921 | 1520 uint8_t *dst[3] = { |
1521 curframe->data[0] + 16*mb_y*s->linesize, | |
1522 curframe->data[1] + 8*mb_y*s->uvlinesize, | |
1523 curframe->data[2] + 8*mb_y*s->uvlinesize | |
1524 }; | |
1525 | |
1526 memset(s->left_nnz, 0, sizeof(s->left_nnz)); | |
1527 | |
1528 // left edge of 129 for intra prediction | |
1529 if (!(avctx->flags & CODEC_FLAG_EMU_EDGE)) | |
1530 for (i = 0; i < 3; i++) | |
1531 for (y = 0; y < 16>>!!i; y++) | |
1532 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
|
1533 if (mb_y) |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
1534 memset(s->top_border, 129, sizeof(*s->top_border)); |
11921 | 1535 |
12237 | 1536 for (mb_x = 0; mb_x < s->mb_width; mb_x++, mb_xy++, mb++) { |
12221 | 1537 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
|
1538 uint8_t *segment_mb = segment_map+mb_x; |
12221 | 1539 |
12215 | 1540 /* Prefetch the current frame, 4 MBs ahead */ |
1541 s->dsp.prefetch(dst[0] + (mb_x&3)*4*s->linesize + 64, s->linesize, 4); | |
1542 s->dsp.prefetch(dst[1] + (mb_x&7)*s->uvlinesize + 64, dst[2] - dst[1], 2); | |
1543 | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1544 decode_mb_mode(s, mb, mb_x, mb_y, intra4x4_mb, segment_mb); |
11921 | 1545 |
12237 | 1546 prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_PREVIOUS); |
12231 | 1547 |
11921 | 1548 if (!mb->skip) |
1549 decode_mb_coeffs(s, c, mb, s->top_nnz[mb_x], s->left_nnz); | |
1550 | |
12225
c3e11b3108d7
Eliminate a pointless memset for intra blocks in P-frames in VP8
darkshikari
parents:
12224
diff
changeset
|
1551 if (mb->mode <= MODE_I4x4) |
12221 | 1552 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
|
1553 else |
11921 | 1554 inter_predict(s, dst, mb, mb_x, mb_y); |
1555 | |
12237 | 1556 prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN); |
12231 | 1557 |
11921 | 1558 if (!mb->skip) { |
12238 | 1559 idct_mb(s, dst, mb); |
11921 | 1560 } else { |
1561 AV_ZERO64(s->left_nnz); | |
1562 AV_WN64(s->top_nnz[mb_x], 0); // array of 9, so unaligned | |
1563 | |
1564 // Reset DC block predictors if they would exist if the mb had coefficients | |
1565 if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) { | |
1566 s->left_nnz[8] = 0; | |
1567 s->top_nnz[mb_x][8] = 0; | |
1568 } | |
1569 } | |
1570 | |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1571 if (s->deblock_filter) |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1572 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
|
1573 |
12237 | 1574 prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN2); |
12231 | 1575 |
11921 | 1576 dst[0] += 16; |
1577 dst[1] += 8; | |
1578 dst[2] += 8; | |
1579 } | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
1580 if (s->deblock_filter) { |
11921 | 1581 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
|
1582 filter_mb_row_simple(s, mb_y); |
11921 | 1583 else |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
1584 filter_mb_row(s, mb_y); |
11921 | 1585 } |
1586 } | |
1587 | |
1588 skip_decode: | |
1589 // if future frames don't use the updated probabilities, | |
1590 // reset them to the values we saved | |
1591 if (!s->update_probabilities) | |
1592 s->prob[0] = s->prob[1]; | |
1593 | |
1594 // check if golden and altref are swapped | |
1595 if (s->update_altref == VP56_FRAME_GOLDEN && | |
1596 s->update_golden == VP56_FRAME_GOLDEN2) | |
1597 FFSWAP(AVFrame *, s->framep[VP56_FRAME_GOLDEN], s->framep[VP56_FRAME_GOLDEN2]); | |
1598 else { | |
1599 if (s->update_altref != VP56_FRAME_NONE) | |
1600 s->framep[VP56_FRAME_GOLDEN2] = s->framep[s->update_altref]; | |
1601 | |
1602 if (s->update_golden != VP56_FRAME_NONE) | |
1603 s->framep[VP56_FRAME_GOLDEN] = s->framep[s->update_golden]; | |
1604 } | |
1605 | |
1606 if (s->update_last) // move cur->prev | |
1607 s->framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_CURRENT]; | |
1608 | |
1609 // release no longer referenced frames | |
1610 for (i = 0; i < 4; i++) | |
1611 if (s->frames[i].data[0] && | |
1612 &s->frames[i] != s->framep[VP56_FRAME_CURRENT] && | |
1613 &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] && | |
1614 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] && | |
1615 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) | |
1616 avctx->release_buffer(avctx, &s->frames[i]); | |
1617 | |
1618 if (!s->invisible) { | |
1619 *(AVFrame*)data = *s->framep[VP56_FRAME_CURRENT]; | |
1620 *data_size = sizeof(AVFrame); | |
1621 } | |
1622 | |
1623 return avpkt->size; | |
1624 } | |
1625 | |
1626 static av_cold int vp8_decode_init(AVCodecContext *avctx) | |
1627 { | |
1628 VP8Context *s = avctx->priv_data; | |
1629 | |
1630 s->avctx = avctx; | |
1631 avctx->pix_fmt = PIX_FMT_YUV420P; | |
1632 | |
1633 dsputil_init(&s->dsp, avctx); | |
1634 ff_h264_pred_init(&s->hpc, CODEC_ID_VP8); | |
1635 ff_vp8dsp_init(&s->vp8dsp); | |
1636 | |
1637 // intra pred needs edge emulation among other things | |
1638 if (avctx->flags&CODEC_FLAG_EMU_EDGE) { | |
11947 | 1639 av_log(avctx, AV_LOG_ERROR, "Edge emulation not supported\n"); |
11921 | 1640 return AVERROR_PATCHWELCOME; |
1641 } | |
1642 | |
1643 return 0; | |
1644 } | |
1645 | |
1646 static av_cold int vp8_decode_free(AVCodecContext *avctx) | |
1647 { | |
1648 vp8_decode_flush(avctx); | |
1649 return 0; | |
1650 } | |
1651 | |
1652 AVCodec vp8_decoder = { | |
1653 "vp8", | |
1654 AVMEDIA_TYPE_VIDEO, | |
1655 CODEC_ID_VP8, | |
1656 sizeof(VP8Context), | |
1657 vp8_decode_init, | |
1658 NULL, | |
1659 vp8_decode_free, | |
1660 vp8_decode_frame, | |
1661 CODEC_CAP_DR1, | |
1662 .flush = vp8_decode_flush, | |
1663 .long_name = NULL_IF_CONFIG_SMALL("On2 VP8"), | |
1664 }; |