Mercurial > libavcodec.hg
annotate vp8.c @ 12358:d596749eb0bc libavcodec
VP8: slightly faster DCT coefficient probability update
author | darkshikari |
---|---|
date | Tue, 03 Aug 2010 23:21:47 +0000 |
parents | ba503df47c2f |
children | 18117b5bb7dc |
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 | |
12339
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
91 uint8_t *intra4x4_pred_mode_top; |
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
92 uint8_t intra4x4_pred_mode_left[4]; |
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]; | |
12340
2d15f62f4f8a
VP8: move zeroing of luma DC block into the WHT
darkshikari
parents:
12339
diff
changeset
|
120 DECLARE_ALIGNED(16, DCTELEM, block_dc)[16]; |
12221 | 121 uint8_t intra4x4_pred_mode_mb[16]; |
11921 | 122 |
123 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
|
124 int segment; ///< segment of the current macroblock |
11921 | 125 |
126 int mbskip_enabled; | |
127 int sign_bias[4]; ///< one state [0, 1] per ref frame type | |
12231 | 128 int ref_count[3]; |
11921 | 129 |
130 /** | |
131 * Base parameters for segmentation, i.e. per-macroblock parameters. | |
132 * These must be kept unchanged even if segmentation is not used for | |
133 * a frame, since the values persist between interframes. | |
134 */ | |
135 struct { | |
136 int enabled; | |
137 int absolute_vals; | |
138 int update_map; | |
139 int8_t base_quant[4]; | |
140 int8_t filter_level[4]; ///< base loop filter level | |
141 } segmentation; | |
142 | |
143 /** | |
144 * Macroblocks can have one of 4 different quants in a frame when | |
145 * segmentation is enabled. | |
146 * If segmentation is disabled, only the first segment's values are used. | |
147 */ | |
148 struct { | |
149 // [0] - DC qmul [1] - AC qmul | |
150 int16_t luma_qmul[2]; | |
151 int16_t luma_dc_qmul[2]; ///< luma dc-only block quant | |
152 int16_t chroma_qmul[2]; | |
153 } qmat[4]; | |
154 | |
155 struct { | |
156 int simple; | |
157 int level; | |
158 int sharpness; | |
159 } filter; | |
160 | |
161 struct { | |
162 int enabled; ///< whether each mb can have a different strength based on mode/ref | |
163 | |
164 /** | |
165 * filter strength adjustment for the following macroblock modes: | |
166 * [0] - i4x4 | |
167 * [1] - zero mv | |
168 * [2] - inter modes except for zero or split mv | |
169 * [3] - split mv | |
170 * i16x16 modes never have any adjustment | |
171 */ | |
172 int8_t mode[4]; | |
173 | |
174 /** | |
175 * filter strength adjustment for macroblocks that reference: | |
176 * [0] - intra / VP56_FRAME_CURRENT | |
177 * [1] - VP56_FRAME_PREVIOUS | |
178 * [2] - VP56_FRAME_GOLDEN | |
179 * [3] - altref / VP56_FRAME_GOLDEN2 | |
180 */ | |
181 int8_t ref[4]; | |
182 } lf_delta; | |
183 | |
184 /** | |
185 * These are all of the updatable probabilities for binary decisions. | |
186 * They are only implictly reset on keyframes, making it quite likely | |
187 * for an interframe to desync if a prior frame's header was corrupt | |
188 * or missing outright! | |
189 */ | |
190 struct { | |
191 uint8_t segmentid[3]; | |
12290
2a09b276db12
b0rk3d FATE + black helicopters hissing -> rolling back to r24556 and sleeping
skal
parents:
12289
diff
changeset
|
192 uint8_t mbskip; |
2a09b276db12
b0rk3d FATE + black helicopters hissing -> rolling back to r24556 and sleeping
skal
parents:
12289
diff
changeset
|
193 uint8_t intra; |
2a09b276db12
b0rk3d FATE + black helicopters hissing -> rolling back to r24556 and sleeping
skal
parents:
12289
diff
changeset
|
194 uint8_t last; |
2a09b276db12
b0rk3d FATE + black helicopters hissing -> rolling back to r24556 and sleeping
skal
parents:
12289
diff
changeset
|
195 uint8_t golden; |
11921 | 196 uint8_t pred16x16[4]; |
197 uint8_t pred8x8c[3]; | |
12343
a18ab740d2db
VP8: eliminate a dereference in coefficient decoding
darkshikari
parents:
12342
diff
changeset
|
198 /* Padded to allow overreads */ |
a18ab740d2db
VP8: eliminate a dereference in coefficient decoding
darkshikari
parents:
12342
diff
changeset
|
199 uint8_t token[4][17][3][NUM_DCT_TOKENS-1]; |
11921 | 200 uint8_t mvc[2][19]; |
201 } prob[2]; | |
202 } VP8Context; | |
203 | |
204 static void vp8_decode_flush(AVCodecContext *avctx) | |
205 { | |
206 VP8Context *s = avctx->priv_data; | |
207 int i; | |
208 | |
209 for (i = 0; i < 4; i++) | |
210 if (s->frames[i].data[0]) | |
211 avctx->release_buffer(avctx, &s->frames[i]); | |
212 memset(s->framep, 0, sizeof(s->framep)); | |
213 | |
214 av_freep(&s->macroblocks_base); | |
12271 | 215 av_freep(&s->filter_strength); |
12339
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
216 av_freep(&s->intra4x4_pred_mode_top); |
11921 | 217 av_freep(&s->top_nnz); |
218 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
|
219 av_freep(&s->top_border); |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
220 av_freep(&s->segmentation_map); |
11921 | 221 |
222 s->macroblocks = NULL; | |
223 } | |
224 | |
225 static int update_dimensions(VP8Context *s, int width, int height) | |
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)); |
12339
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
244 s->intra4x4_pred_mode_top = av_mallocz(s->b4_stride*4); |
11921 | 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 |
12339
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
249 if (!s->macroblocks_base || !s->filter_strength || !s->intra4x4_pred_mode_top || |
12223
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 |
255 return 0; | |
256 } | |
257 | |
258 static void parse_segment_info(VP8Context *s) | |
259 { | |
260 VP56RangeCoder *c = &s->c; | |
261 int i; | |
262 | |
263 s->segmentation.update_map = vp8_rac_get(c); | |
264 | |
265 if (vp8_rac_get(c)) { // update segment feature data | |
266 s->segmentation.absolute_vals = vp8_rac_get(c); | |
267 | |
268 for (i = 0; i < 4; i++) | |
269 s->segmentation.base_quant[i] = vp8_rac_get_sint(c, 7); | |
270 | |
271 for (i = 0; i < 4; i++) | |
272 s->segmentation.filter_level[i] = vp8_rac_get_sint(c, 6); | |
273 } | |
274 if (s->segmentation.update_map) | |
275 for (i = 0; i < 3; i++) | |
276 s->prob->segmentid[i] = vp8_rac_get(c) ? vp8_rac_get_uint(c, 8) : 255; | |
277 } | |
278 | |
279 static void update_lf_deltas(VP8Context *s) | |
280 { | |
281 VP56RangeCoder *c = &s->c; | |
282 int i; | |
283 | |
284 for (i = 0; i < 4; i++) | |
285 s->lf_delta.ref[i] = vp8_rac_get_sint(c, 6); | |
286 | |
287 for (i = 0; i < 4; i++) | |
288 s->lf_delta.mode[i] = vp8_rac_get_sint(c, 6); | |
289 } | |
290 | |
291 static int setup_partitions(VP8Context *s, const uint8_t *buf, int buf_size) | |
292 { | |
293 const uint8_t *sizes = buf; | |
294 int i; | |
295 | |
296 s->num_coeff_partitions = 1 << vp8_rac_get_uint(&s->c, 2); | |
297 | |
298 buf += 3*(s->num_coeff_partitions-1); | |
299 buf_size -= 3*(s->num_coeff_partitions-1); | |
300 if (buf_size < 0) | |
301 return -1; | |
302 | |
303 for (i = 0; i < s->num_coeff_partitions-1; i++) { | |
12247
50a96623366b
VP8: use AV_RL24 instead of defining a new RL24.
darkshikari
parents:
12246
diff
changeset
|
304 int size = AV_RL24(sizes + 3*i); |
11921 | 305 if (buf_size - size < 0) |
306 return -1; | |
307 | |
308 vp56_init_range_decoder(&s->coeff_partition[i], buf, size); | |
309 buf += size; | |
310 buf_size -= size; | |
311 } | |
312 vp56_init_range_decoder(&s->coeff_partition[i], buf, buf_size); | |
313 | |
314 return 0; | |
315 } | |
316 | |
317 static void get_quants(VP8Context *s) | |
318 { | |
319 VP56RangeCoder *c = &s->c; | |
320 int i, base_qi; | |
321 | |
322 int yac_qi = vp8_rac_get_uint(c, 7); | |
323 int ydc_delta = vp8_rac_get_sint(c, 4); | |
324 int y2dc_delta = vp8_rac_get_sint(c, 4); | |
325 int y2ac_delta = vp8_rac_get_sint(c, 4); | |
326 int uvdc_delta = vp8_rac_get_sint(c, 4); | |
327 int uvac_delta = vp8_rac_get_sint(c, 4); | |
328 | |
329 for (i = 0; i < 4; i++) { | |
330 if (s->segmentation.enabled) { | |
331 base_qi = s->segmentation.base_quant[i]; | |
332 if (!s->segmentation.absolute_vals) | |
333 base_qi += yac_qi; | |
334 } else | |
335 base_qi = yac_qi; | |
336 | |
337 s->qmat[i].luma_qmul[0] = vp8_dc_qlookup[av_clip(base_qi + ydc_delta , 0, 127)]; | |
338 s->qmat[i].luma_qmul[1] = vp8_ac_qlookup[av_clip(base_qi , 0, 127)]; | |
339 s->qmat[i].luma_dc_qmul[0] = 2 * vp8_dc_qlookup[av_clip(base_qi + y2dc_delta, 0, 127)]; | |
12290
2a09b276db12
b0rk3d FATE + black helicopters hissing -> rolling back to r24556 and sleeping
skal
parents:
12289
diff
changeset
|
340 s->qmat[i].luma_dc_qmul[1] = 155 * vp8_ac_qlookup[av_clip(base_qi + y2ac_delta, 0, 127)] / 100; |
2a09b276db12
b0rk3d FATE + black helicopters hissing -> rolling back to r24556 and sleeping
skal
parents:
12289
diff
changeset
|
341 s->qmat[i].chroma_qmul[0] = vp8_dc_qlookup[av_clip(base_qi + uvdc_delta, 0, 127)]; |
11921 | 342 s->qmat[i].chroma_qmul[1] = vp8_ac_qlookup[av_clip(base_qi + uvac_delta, 0, 127)]; |
12290
2a09b276db12
b0rk3d FATE + black helicopters hissing -> rolling back to r24556 and sleeping
skal
parents:
12289
diff
changeset
|
343 |
2a09b276db12
b0rk3d FATE + black helicopters hissing -> rolling back to r24556 and sleeping
skal
parents:
12289
diff
changeset
|
344 s->qmat[i].luma_dc_qmul[1] = FFMAX(s->qmat[i].luma_dc_qmul[1], 8); |
2a09b276db12
b0rk3d FATE + black helicopters hissing -> rolling back to r24556 and sleeping
skal
parents:
12289
diff
changeset
|
345 s->qmat[i].chroma_qmul[0] = FFMIN(s->qmat[i].chroma_qmul[0], 132); |
11921 | 346 } |
347 } | |
348 | |
349 /** | |
350 * Determine which buffers golden and altref should be updated with after this frame. | |
351 * The spec isn't clear here, so I'm going by my understanding of what libvpx does | |
352 * | |
353 * Intra frames update all 3 references | |
354 * Inter frames update VP56_FRAME_PREVIOUS if the update_last flag is set | |
355 * If the update (golden|altref) flag is set, it's updated with the current frame | |
356 * if update_last is set, and VP56_FRAME_PREVIOUS otherwise. | |
357 * If the flag is not set, the number read means: | |
358 * 0: no update | |
359 * 1: VP56_FRAME_PREVIOUS | |
360 * 2: update golden with altref, or update altref with golden | |
361 */ | |
362 static VP56Frame ref_to_update(VP8Context *s, int update, VP56Frame ref) | |
363 { | |
364 VP56RangeCoder *c = &s->c; | |
365 | |
366 if (update) | |
367 return VP56_FRAME_CURRENT; | |
368 | |
369 switch (vp8_rac_get_uint(c, 2)) { | |
370 case 1: | |
371 return VP56_FRAME_PREVIOUS; | |
372 case 2: | |
373 return (ref == VP56_FRAME_GOLDEN) ? VP56_FRAME_GOLDEN2 : VP56_FRAME_GOLDEN; | |
374 } | |
375 return VP56_FRAME_NONE; | |
376 } | |
377 | |
378 static void update_refs(VP8Context *s) | |
379 { | |
380 VP56RangeCoder *c = &s->c; | |
381 | |
382 int update_golden = vp8_rac_get(c); | |
383 int update_altref = vp8_rac_get(c); | |
384 | |
385 s->update_golden = ref_to_update(s, update_golden, VP56_FRAME_GOLDEN); | |
386 s->update_altref = ref_to_update(s, update_altref, VP56_FRAME_GOLDEN2); | |
387 } | |
388 | |
389 static int decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size) | |
390 { | |
391 VP56RangeCoder *c = &s->c; | |
12343
a18ab740d2db
VP8: eliminate a dereference in coefficient decoding
darkshikari
parents:
12342
diff
changeset
|
392 int header_size, hscale, vscale, i, j, k, l, m, ret; |
11921 | 393 int width = s->avctx->width; |
394 int height = s->avctx->height; | |
395 | |
396 s->keyframe = !(buf[0] & 1); | |
397 s->profile = (buf[0]>>1) & 7; | |
398 s->invisible = !(buf[0] & 0x10); | |
12247
50a96623366b
VP8: use AV_RL24 instead of defining a new RL24.
darkshikari
parents:
12246
diff
changeset
|
399 header_size = AV_RL24(buf) >> 5; |
11921 | 400 buf += 3; |
401 buf_size -= 3; | |
402 | |
11974 | 403 if (s->profile > 3) |
404 av_log(s->avctx, AV_LOG_WARNING, "Unknown profile %d\n", s->profile); | |
405 | |
406 if (!s->profile) | |
407 memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_epel_pixels_tab, sizeof(s->put_pixels_tab)); | |
408 else // profile 1-3 use bilinear, 4+ aren't defined so whatever | |
409 memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_bilinear_pixels_tab, sizeof(s->put_pixels_tab)); | |
11921 | 410 |
411 if (header_size > buf_size - 7*s->keyframe) { | |
412 av_log(s->avctx, AV_LOG_ERROR, "Header size larger than data provided\n"); | |
413 return AVERROR_INVALIDDATA; | |
414 } | |
415 | |
416 if (s->keyframe) { | |
12247
50a96623366b
VP8: use AV_RL24 instead of defining a new RL24.
darkshikari
parents:
12246
diff
changeset
|
417 if (AV_RL24(buf) != 0x2a019d) { |
50a96623366b
VP8: use AV_RL24 instead of defining a new RL24.
darkshikari
parents:
12246
diff
changeset
|
418 av_log(s->avctx, AV_LOG_ERROR, "Invalid start code 0x%x\n", AV_RL24(buf)); |
11921 | 419 return AVERROR_INVALIDDATA; |
420 } | |
421 width = AV_RL16(buf+3) & 0x3fff; | |
422 height = AV_RL16(buf+5) & 0x3fff; | |
423 hscale = buf[4] >> 6; | |
424 vscale = buf[6] >> 6; | |
425 buf += 7; | |
426 buf_size -= 7; | |
427 | |
11970
c7953ee47af4
vp8: warn and request sample if upscaling specified in header
mru
parents:
11950
diff
changeset
|
428 if (hscale || vscale) |
c7953ee47af4
vp8: warn and request sample if upscaling specified in header
mru
parents:
11950
diff
changeset
|
429 av_log_missing_feature(s->avctx, "Upscaling", 1); |
c7953ee47af4
vp8: warn and request sample if upscaling specified in header
mru
parents:
11950
diff
changeset
|
430 |
11921 | 431 s->update_golden = s->update_altref = VP56_FRAME_CURRENT; |
12343
a18ab740d2db
VP8: eliminate a dereference in coefficient decoding
darkshikari
parents:
12342
diff
changeset
|
432 for (i = 0; i < 4; i++) |
a18ab740d2db
VP8: eliminate a dereference in coefficient decoding
darkshikari
parents:
12342
diff
changeset
|
433 for (j = 0; j < 16; j++) |
a18ab740d2db
VP8: eliminate a dereference in coefficient decoding
darkshikari
parents:
12342
diff
changeset
|
434 memcpy(s->prob->token[i][j], vp8_token_default_probs[i][vp8_coeff_band[j]], |
a18ab740d2db
VP8: eliminate a dereference in coefficient decoding
darkshikari
parents:
12342
diff
changeset
|
435 sizeof(s->prob->token[i][j])); |
11921 | 436 memcpy(s->prob->pred16x16, vp8_pred16x16_prob_inter, sizeof(s->prob->pred16x16)); |
437 memcpy(s->prob->pred8x8c , vp8_pred8x8c_prob_inter , sizeof(s->prob->pred8x8c)); | |
438 memcpy(s->prob->mvc , vp8_mv_default_prob , sizeof(s->prob->mvc)); | |
439 memset(&s->segmentation, 0, sizeof(s->segmentation)); | |
440 } | |
441 | |
442 if (!s->macroblocks_base || /* first frame */ | |
443 width != s->avctx->width || height != s->avctx->height) { | |
444 if ((ret = update_dimensions(s, width, height) < 0)) | |
445 return ret; | |
446 } | |
447 | |
448 vp56_init_range_decoder(c, buf, header_size); | |
449 buf += header_size; | |
450 buf_size -= header_size; | |
451 | |
452 if (s->keyframe) { | |
453 if (vp8_rac_get(c)) | |
454 av_log(s->avctx, AV_LOG_WARNING, "Unspecified colorspace\n"); | |
455 vp8_rac_get(c); // whether we can skip clamping in dsp functions | |
456 } | |
457 | |
458 if ((s->segmentation.enabled = vp8_rac_get(c))) | |
459 parse_segment_info(s); | |
460 else | |
461 s->segmentation.update_map = 0; // FIXME: move this to some init function? | |
462 | |
463 s->filter.simple = vp8_rac_get(c); | |
464 s->filter.level = vp8_rac_get_uint(c, 6); | |
465 s->filter.sharpness = vp8_rac_get_uint(c, 3); | |
466 | |
467 if ((s->lf_delta.enabled = vp8_rac_get(c))) | |
468 if (vp8_rac_get(c)) | |
469 update_lf_deltas(s); | |
470 | |
471 if (setup_partitions(s, buf, buf_size)) { | |
472 av_log(s->avctx, AV_LOG_ERROR, "Invalid partitions\n"); | |
473 return AVERROR_INVALIDDATA; | |
474 } | |
475 | |
476 get_quants(s); | |
477 | |
478 if (!s->keyframe) { | |
479 update_refs(s); | |
480 s->sign_bias[VP56_FRAME_GOLDEN] = vp8_rac_get(c); | |
481 s->sign_bias[VP56_FRAME_GOLDEN2 /* altref */] = vp8_rac_get(c); | |
482 } | |
483 | |
484 // if we aren't saving this frame's probabilities for future frames, | |
485 // make a copy of the current probabilities | |
486 if (!(s->update_probabilities = vp8_rac_get(c))) | |
487 s->prob[1] = s->prob[0]; | |
488 | |
489 s->update_last = s->keyframe || vp8_rac_get(c); | |
490 | |
491 for (i = 0; i < 4; i++) | |
492 for (j = 0; j < 8; j++) | |
493 for (k = 0; k < 3; k++) | |
494 for (l = 0; l < NUM_DCT_TOKENS-1; l++) | |
12343
a18ab740d2db
VP8: eliminate a dereference in coefficient decoding
darkshikari
parents:
12342
diff
changeset
|
495 if (vp56_rac_get_prob_branchy(c, vp8_token_update_probs[i][j][k][l])) { |
a18ab740d2db
VP8: eliminate a dereference in coefficient decoding
darkshikari
parents:
12342
diff
changeset
|
496 int prob = vp8_rac_get_uint(c, 8); |
12358
d596749eb0bc
VP8: slightly faster DCT coefficient probability update
darkshikari
parents:
12354
diff
changeset
|
497 for (m = 0; vp8_coeff_band_indexes[j][m] >= 0; m++) |
d596749eb0bc
VP8: slightly faster DCT coefficient probability update
darkshikari
parents:
12354
diff
changeset
|
498 s->prob->token[i][vp8_coeff_band_indexes[j][m]][k][l] = prob; |
12343
a18ab740d2db
VP8: eliminate a dereference in coefficient decoding
darkshikari
parents:
12342
diff
changeset
|
499 } |
11921 | 500 |
501 if ((s->mbskip_enabled = vp8_rac_get(c))) | |
12290
2a09b276db12
b0rk3d FATE + black helicopters hissing -> rolling back to r24556 and sleeping
skal
parents:
12289
diff
changeset
|
502 s->prob->mbskip = vp8_rac_get_uint(c, 8); |
11921 | 503 |
504 if (!s->keyframe) { | |
12290
2a09b276db12
b0rk3d FATE + black helicopters hissing -> rolling back to r24556 and sleeping
skal
parents:
12289
diff
changeset
|
505 s->prob->intra = vp8_rac_get_uint(c, 8); |
2a09b276db12
b0rk3d FATE + black helicopters hissing -> rolling back to r24556 and sleeping
skal
parents:
12289
diff
changeset
|
506 s->prob->last = vp8_rac_get_uint(c, 8); |
2a09b276db12
b0rk3d FATE + black helicopters hissing -> rolling back to r24556 and sleeping
skal
parents:
12289
diff
changeset
|
507 s->prob->golden = vp8_rac_get_uint(c, 8); |
11921 | 508 |
509 if (vp8_rac_get(c)) | |
510 for (i = 0; i < 4; i++) | |
511 s->prob->pred16x16[i] = vp8_rac_get_uint(c, 8); | |
512 if (vp8_rac_get(c)) | |
513 for (i = 0; i < 3; i++) | |
514 s->prob->pred8x8c[i] = vp8_rac_get_uint(c, 8); | |
515 | |
516 // 17.2 MV probability update | |
517 for (i = 0; i < 2; i++) | |
518 for (j = 0; j < 19; j++) | |
12254
17c151e1280a
VP8: Use vp56_rac_get_prob_branchy when the bit is only used by an if()
conrad
parents:
12253
diff
changeset
|
519 if (vp56_rac_get_prob_branchy(c, vp8_mv_update_prob[i][j])) |
11921 | 520 s->prob->mvc[i][j] = vp8_rac_get_nn(c); |
521 } | |
522 | |
523 return 0; | |
524 } | |
525 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
526 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
527 void clamp_mv(VP8Context *s, VP56mv *dst, const VP56mv *src, int mb_x, int mb_y) |
11921 | 528 { |
529 #define MARGIN (16 << 2) | |
530 dst->x = av_clip(src->x, -((mb_x << 6) + MARGIN), | |
531 ((s->mb_width - 1 - mb_x) << 6) + MARGIN); | |
532 dst->y = av_clip(src->y, -((mb_y << 6) + MARGIN), | |
533 ((s->mb_height - 1 - mb_y) << 6) + MARGIN); | |
534 } | |
535 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
536 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
537 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
|
538 VP56mv near[2], VP56mv *best, uint8_t cnt[4]) |
11921 | 539 { |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
540 VP8Macroblock *mb_edge[3] = { mb + 2 /* top */, |
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
541 mb - 1 /* left */, |
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
542 mb + 1 /* top-left */ }; |
11921 | 543 enum { EDGE_TOP, EDGE_LEFT, EDGE_TOPLEFT }; |
544 VP56mv near_mv[4] = {{ 0 }}; | |
545 enum { CNT_ZERO, CNT_NEAREST, CNT_NEAR, CNT_SPLITMV }; | |
12217 | 546 int idx = CNT_ZERO; |
11921 | 547 int best_idx = CNT_ZERO; |
12217 | 548 int cur_sign_bias = s->sign_bias[mb->ref_frame]; |
549 int *sign_bias = s->sign_bias; | |
11921 | 550 |
551 /* Process MB on top, left and top-left */ | |
12217 | 552 #define MV_EDGE_CHECK(n)\ |
553 {\ | |
554 VP8Macroblock *edge = mb_edge[n];\ | |
555 int edge_ref = edge->ref_frame;\ | |
556 if (edge_ref != VP56_FRAME_CURRENT) {\ | |
557 uint32_t mv = AV_RN32A(&edge->mv);\ | |
558 if (mv) {\ | |
559 if (cur_sign_bias != sign_bias[edge_ref]) {\ | |
560 /* SWAR negate of the values in mv. */\ | |
12242 | 561 mv = ~mv;\ |
562 mv = ((mv&0x7fff7fff) + 0x00010001) ^ (mv&0x80008000);\ | |
12217 | 563 }\ |
564 if (!n || mv != AV_RN32A(&near_mv[idx]))\ | |
565 AV_WN32A(&near_mv[++idx], mv);\ | |
566 cnt[idx] += 1 + (n != 2);\ | |
567 } else\ | |
568 cnt[CNT_ZERO] += 1 + (n != 2);\ | |
569 }\ | |
11921 | 570 } |
12217 | 571 MV_EDGE_CHECK(0) |
572 MV_EDGE_CHECK(1) | |
573 MV_EDGE_CHECK(2) | |
11921 | 574 |
12217 | 575 /* If we have three distinct MVs, merge first and last if they're the same */ |
576 if (cnt[CNT_SPLITMV] && AV_RN32A(&near_mv[1+EDGE_TOP]) == AV_RN32A(&near_mv[1+EDGE_TOPLEFT])) | |
11921 | 577 cnt[CNT_NEAREST] += 1; |
578 | |
579 cnt[CNT_SPLITMV] = ((mb_edge[EDGE_LEFT]->mode == VP8_MVMODE_SPLIT) + | |
580 (mb_edge[EDGE_TOP]->mode == VP8_MVMODE_SPLIT)) * 2 + | |
581 (mb_edge[EDGE_TOPLEFT]->mode == VP8_MVMODE_SPLIT); | |
582 | |
583 /* Swap near and nearest if necessary */ | |
584 if (cnt[CNT_NEAR] > cnt[CNT_NEAREST]) { | |
12217 | 585 FFSWAP(uint8_t, cnt[CNT_NEAREST], cnt[CNT_NEAR]); |
586 FFSWAP( VP56mv, near_mv[CNT_NEAREST], near_mv[CNT_NEAR]); | |
11921 | 587 } |
588 | |
589 /* Choose the best mv out of 0,0 and the nearest mv */ | |
590 if (cnt[CNT_NEAREST] >= cnt[CNT_ZERO]) | |
591 best_idx = CNT_NEAREST; | |
592 | |
12246 | 593 mb->mv = near_mv[best_idx]; |
11921 | 594 near[0] = near_mv[CNT_NEAREST]; |
595 near[1] = near_mv[CNT_NEAR]; | |
596 } | |
597 | |
598 /** | |
599 * Motion vector coding, 17.1. | |
600 */ | |
601 static int read_mv_component(VP56RangeCoder *c, const uint8_t *p) | |
602 { | |
12255 | 603 int bit, x = 0; |
11921 | 604 |
12254
17c151e1280a
VP8: Use vp56_rac_get_prob_branchy when the bit is only used by an if()
conrad
parents:
12253
diff
changeset
|
605 if (vp56_rac_get_prob_branchy(c, p[0])) { |
11921 | 606 int i; |
607 | |
608 for (i = 0; i < 3; i++) | |
609 x += vp56_rac_get_prob(c, p[9 + i]) << i; | |
610 for (i = 9; i > 3; i--) | |
611 x += vp56_rac_get_prob(c, p[9 + i]) << i; | |
612 if (!(x & 0xFFF0) || vp56_rac_get_prob(c, p[12])) | |
613 x += 8; | |
12255 | 614 } else { |
615 // small_mvtree | |
616 const uint8_t *ps = p+2; | |
617 bit = vp56_rac_get_prob(c, *ps); | |
618 ps += 1 + 3*bit; | |
619 x += 4*bit; | |
620 bit = vp56_rac_get_prob(c, *ps); | |
621 ps += 1 + bit; | |
622 x += 2*bit; | |
623 x += vp56_rac_get_prob(c, *ps); | |
624 } | |
11921 | 625 |
626 return (x && vp56_rac_get_prob(c, p[1])) ? -x : x; | |
627 } | |
628 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
629 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
630 const uint8_t *get_submv_prob(uint32_t left, uint32_t top) |
11921 | 631 { |
12219 | 632 if (left == top) |
633 return vp8_submv_prob[4-!!left]; | |
634 if (!top) | |
11921 | 635 return vp8_submv_prob[2]; |
12219 | 636 return vp8_submv_prob[1-!!left]; |
11921 | 637 } |
638 | |
639 /** | |
640 * 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
|
641 * @returns the number of motion vectors parsed (2, 4 or 16) |
11921 | 642 */ |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
643 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
644 int decode_splitmvs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb) |
11921 | 645 { |
12352 | 646 int part_idx; |
647 int n, num; | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
648 VP8Macroblock *top_mb = &mb[2]; |
12219 | 649 VP8Macroblock *left_mb = &mb[-1]; |
650 const uint8_t *mbsplits_left = vp8_mbsplits[left_mb->partitioning], | |
651 *mbsplits_top = vp8_mbsplits[top_mb->partitioning], | |
12352 | 652 *mbsplits_cur, *firstidx; |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
653 VP56mv *top_mv = top_mb->bmv; |
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
654 VP56mv *left_mv = left_mb->bmv; |
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
655 VP56mv *cur_mv = mb->bmv; |
11921 | 656 |
12352 | 657 if (vp56_rac_get_prob_branchy(c, vp8_mbsplit_prob[0])) { |
658 if (vp56_rac_get_prob_branchy(c, vp8_mbsplit_prob[1])) { | |
659 part_idx = VP8_SPLITMVMODE_16x8 + vp56_rac_get_prob(c, vp8_mbsplit_prob[2]); | |
660 } else { | |
661 part_idx = VP8_SPLITMVMODE_8x8; | |
662 } | |
663 } else { | |
664 part_idx = VP8_SPLITMVMODE_4x4; | |
665 } | |
666 | |
667 num = vp8_mbsplit_count[part_idx]; | |
668 mbsplits_cur = vp8_mbsplits[part_idx], | |
669 firstidx = vp8_mbfirstidx[part_idx]; | |
670 mb->partitioning = part_idx; | |
671 | |
11921 | 672 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
|
673 int k = firstidx[n]; |
12219 | 674 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
|
675 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
|
676 |
12219 | 677 if (!(k & 3)) |
678 left = AV_RN32A(&left_mv[mbsplits_left[k + 3]]); | |
679 else | |
680 left = AV_RN32A(&cur_mv[mbsplits_cur[k - 1]]); | |
681 if (k <= 3) | |
682 above = AV_RN32A(&top_mv[mbsplits_top[k + 12]]); | |
683 else | |
684 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
|
685 |
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
686 submv_prob = get_submv_prob(left, above); |
11921 | 687 |
12351 | 688 if (vp56_rac_get_prob_branchy(c, submv_prob[0])) { |
689 if (vp56_rac_get_prob_branchy(c, submv_prob[1])) { | |
690 if (vp56_rac_get_prob_branchy(c, submv_prob[2])) { | |
691 mb->bmv[n].y = mb->mv.y + read_mv_component(c, s->prob->mvc[0]); | |
692 mb->bmv[n].x = mb->mv.x + read_mv_component(c, s->prob->mvc[1]); | |
693 } else { | |
694 AV_ZERO32(&mb->bmv[n]); | |
695 } | |
696 } else { | |
697 AV_WN32A(&mb->bmv[n], above); | |
698 } | |
699 } else { | |
12219 | 700 AV_WN32A(&mb->bmv[n], left); |
11921 | 701 } |
11990
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
702 } |
11921 | 703 |
11990
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
704 return num; |
11921 | 705 } |
706 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
707 static av_always_inline |
12339
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
708 void decode_intra4x4_modes(VP8Context *s, VP56RangeCoder *c, |
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
709 int mb_x, int keyframe) |
11921 | 710 { |
12339
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
711 uint8_t *intra4x4 = s->intra4x4_pred_mode_mb; |
12221 | 712 if (keyframe) { |
12339
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
713 int x, y; |
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
714 uint8_t* const top = s->intra4x4_pred_mode_top + 4 * mb_x; |
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
715 uint8_t* const left = s->intra4x4_pred_mode_left; |
12221 | 716 for (y = 0; y < 4; y++) { |
717 for (x = 0; x < 4; x++) { | |
12339
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
718 const uint8_t *ctx; |
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
719 ctx = vp8_pred4x4_prob_intra[top[x]][left[y]]; |
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
720 *intra4x4 = vp8_rac_get_tree(c, vp8_pred4x4_tree, ctx); |
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
721 left[y] = top[x] = *intra4x4; |
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
722 intra4x4++; |
11921 | 723 } |
724 } | |
12221 | 725 } else { |
12339
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
726 int i; |
12221 | 727 for (i = 0; i < 16; i++) |
728 intra4x4[i] = vp8_rac_get_tree(c, vp8_pred4x4_tree, vp8_pred4x4_prob_inter); | |
11921 | 729 } |
730 } | |
731 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
732 static av_always_inline |
12339
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
733 void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, uint8_t *segment) |
11921 | 734 { |
735 VP56RangeCoder *c = &s->c; | |
736 | |
737 if (s->segmentation.update_map) | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
738 *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
|
739 s->segment = *segment; |
11921 | 740 |
12290
2a09b276db12
b0rk3d FATE + black helicopters hissing -> rolling back to r24556 and sleeping
skal
parents:
12289
diff
changeset
|
741 mb->skip = s->mbskip_enabled ? vp56_rac_get_prob(c, s->prob->mbskip) : 0; |
11921 | 742 |
743 if (s->keyframe) { | |
744 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_intra, vp8_pred16x16_prob_intra); | |
745 | |
746 if (mb->mode == MODE_I4x4) { | |
12339
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
747 decode_intra4x4_modes(s, c, mb_x, 1); |
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
748 } else { |
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
749 const uint32_t modes = vp8_pred4x4_mode[mb->mode] * 0x01010101u; |
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
750 AV_WN32A(s->intra4x4_pred_mode_top + 4 * mb_x, modes); |
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
751 AV_WN32A(s->intra4x4_pred_mode_left, modes); |
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
752 } |
11921 | 753 |
754 s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, vp8_pred8x8c_prob_intra); | |
755 mb->ref_frame = VP56_FRAME_CURRENT; | |
12290
2a09b276db12
b0rk3d FATE + black helicopters hissing -> rolling back to r24556 and sleeping
skal
parents:
12289
diff
changeset
|
756 } else if (vp56_rac_get_prob_branchy(c, s->prob->intra)) { |
11921 | 757 VP56mv near[2], best; |
12217 | 758 uint8_t cnt[4] = { 0 }; |
11921 | 759 |
760 // inter MB, 16.2 | |
12290
2a09b276db12
b0rk3d FATE + black helicopters hissing -> rolling back to r24556 and sleeping
skal
parents:
12289
diff
changeset
|
761 if (vp56_rac_get_prob_branchy(c, s->prob->last)) |
2a09b276db12
b0rk3d FATE + black helicopters hissing -> rolling back to r24556 and sleeping
skal
parents:
12289
diff
changeset
|
762 mb->ref_frame = vp56_rac_get_prob(c, s->prob->golden) ? |
11921 | 763 VP56_FRAME_GOLDEN2 /* altref */ : VP56_FRAME_GOLDEN; |
764 else | |
765 mb->ref_frame = VP56_FRAME_PREVIOUS; | |
12231 | 766 s->ref_count[mb->ref_frame-1]++; |
11921 | 767 |
768 // motion vectors, 16.3 | |
769 find_near_mvs(s, mb, mb_x, mb_y, near, &best, cnt); | |
12350 | 770 if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[0]][0])) { |
771 if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[1]][1])) { | |
772 if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[2]][2])) { | |
773 if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[3]][3])) { | |
774 mb->mode = VP8_MVMODE_SPLIT; | |
775 clamp_mv(s, &mb->mv, &mb->mv, mb_x, mb_y); | |
776 mb->mv = mb->bmv[decode_splitmvs(s, c, mb) - 1]; | |
777 } else { | |
778 mb->mode = VP8_MVMODE_NEW; | |
779 clamp_mv(s, &mb->mv, &mb->mv, mb_x, mb_y); | |
780 mb->mv.y += + read_mv_component(c, s->prob->mvc[0]); | |
781 mb->mv.x += + read_mv_component(c, s->prob->mvc[1]); | |
782 } | |
783 } else { | |
784 mb->mode = VP8_MVMODE_NEAR; | |
785 clamp_mv(s, &mb->mv, &near[1], mb_x, mb_y); | |
786 } | |
787 } else { | |
788 mb->mode = VP8_MVMODE_NEAREST; | |
789 clamp_mv(s, &mb->mv, &near[0], mb_x, mb_y); | |
790 } | |
791 } else { | |
792 mb->mode = VP8_MVMODE_ZERO; | |
12245
ca82c3ce90c1
VP8: use AV_ZERO32 instead of AV_WN32A where relevant
darkshikari
parents:
12244
diff
changeset
|
793 AV_ZERO32(&mb->mv); |
11921 | 794 } |
795 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
|
796 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
|
797 mb->bmv[0] = mb->mv; |
11921 | 798 } |
799 } else { | |
800 // intra MB, 16.1 | |
801 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_inter, s->prob->pred16x16); | |
802 | |
12220
0f635b1f7861
Avoid useless fill_rectangle in P-frames in VP8
darkshikari
parents:
12219
diff
changeset
|
803 if (mb->mode == MODE_I4x4) |
12339
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
804 decode_intra4x4_modes(s, c, mb_x, 0); |
11921 | 805 |
806 s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, s->prob->pred8x8c); | |
807 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
|
808 mb->partitioning = VP8_SPLITMVMODE_NONE; |
12245
ca82c3ce90c1
VP8: use AV_ZERO32 instead of AV_WN32A where relevant
darkshikari
parents:
12244
diff
changeset
|
809 AV_ZERO32(&mb->bmv[0]); |
11921 | 810 } |
811 } | |
812 | |
813 /** | |
12115 | 814 * @param c arithmetic bitstream reader context |
815 * @param block destination for block coefficients | |
816 * @param probs probabilities to use when reading trees from the bitstream | |
11921 | 817 * @param i initial coeff index, 0 unless a separate DC block is coded |
818 * @param zero_nhood the initial prediction context for number of surrounding | |
819 * 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
|
820 * @param qmul array holding the dc/ac dequant factor at position 0/1 |
11921 | 821 * @return 0 if no coeffs were decoded |
822 * otherwise, the index of the last coeff decoded plus one | |
823 */ | |
824 static int decode_block_coeffs(VP56RangeCoder *c, DCTELEM block[16], | |
825 uint8_t probs[8][3][NUM_DCT_TOKENS-1], | |
826 int i, int zero_nhood, int16_t qmul[2]) | |
827 { | |
12343
a18ab740d2db
VP8: eliminate a dereference in coefficient decoding
darkshikari
parents:
12342
diff
changeset
|
828 uint8_t *token_prob = probs[i][zero_nhood]; |
12253
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
829 int nonzero = 0; |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
830 int coeff; |
11921 | 831 |
12253
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
832 do { |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
833 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
|
834 return nonzero; |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
835 |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
836 skip_eob: |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
837 if (!vp56_rac_get_prob_branchy(c, token_prob[1])) { // DCT_0 |
12335 | 838 if (++i == 16) |
839 return nonzero; // invalid input; blocks should end with EOB | |
12343
a18ab740d2db
VP8: eliminate a dereference in coefficient decoding
darkshikari
parents:
12342
diff
changeset
|
840 token_prob = probs[i][0]; |
12335 | 841 goto skip_eob; |
11921 | 842 } |
843 | |
12253
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
844 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
|
845 coeff = 1; |
12343
a18ab740d2db
VP8: eliminate a dereference in coefficient decoding
darkshikari
parents:
12342
diff
changeset
|
846 token_prob = probs[i+1][1]; |
12253
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
847 } else { |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
848 if (!vp56_rac_get_prob_branchy(c, token_prob[3])) { // DCT 2,3,4 |
12354 | 849 coeff = vp56_rac_get_prob_branchy(c, token_prob[4]); |
12253
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
850 if (coeff) |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
851 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
|
852 coeff += 2; |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
853 } else { |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
854 // DCT_CAT* |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
855 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
|
856 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
|
857 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
|
858 } else { // DCT_CAT2 |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
859 coeff = 7; |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
860 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
|
861 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
|
862 } |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
863 } else { // DCT_CAT3 and up |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
864 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
|
865 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
|
866 int cat = (a<<1) + b; |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
867 coeff = 3 + (8<<cat); |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
868 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
|
869 } |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
870 } |
12343
a18ab740d2db
VP8: eliminate a dereference in coefficient decoding
darkshikari
parents:
12342
diff
changeset
|
871 token_prob = probs[i+1][2]; |
12253
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
872 } |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
873 |
11921 | 874 // 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
|
875 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
|
876 nonzero = ++i; |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
877 } while (i < 16); |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
878 |
11921 | 879 return nonzero; |
880 } | |
881 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
882 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
883 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
|
884 uint8_t t_nnz[9], uint8_t l_nnz[9]) |
11921 | 885 { |
886 int i, x, y, luma_start = 0, luma_ctx = 3; | |
887 int nnz_pred, nnz, nnz_total = 0; | |
12224
5b7d690b761b
VP8: Don't store segment in macroblock struct anymore.
darkshikari
parents:
12223
diff
changeset
|
888 int segment = s->segment; |
12342 | 889 int block_dc = 0; |
11921 | 890 |
891 if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) { | |
892 nnz_pred = t_nnz[8] + l_nnz[8]; | |
893 | |
894 // decode DC values and do hadamard | |
12340
2d15f62f4f8a
VP8: move zeroing of luma DC block into the WHT
darkshikari
parents:
12339
diff
changeset
|
895 nnz = decode_block_coeffs(c, s->block_dc, s->prob->token[1], 0, nnz_pred, |
11921 | 896 s->qmat[segment].luma_dc_qmul); |
897 l_nnz[8] = t_nnz[8] = !!nnz; | |
12342 | 898 if (nnz) { |
899 nnz_total += nnz; | |
900 block_dc = 1; | |
901 if (nnz == 1) | |
902 s->vp8dsp.vp8_luma_dc_wht_dc(s->block, s->block_dc); | |
903 else | |
904 s->vp8dsp.vp8_luma_dc_wht(s->block, s->block_dc); | |
905 } | |
11921 | 906 luma_start = 1; |
907 luma_ctx = 0; | |
908 } | |
909 | |
910 // luma blocks | |
911 for (y = 0; y < 4; y++) | |
912 for (x = 0; x < 4; x++) { | |
913 nnz_pred = l_nnz[y] + t_nnz[x]; | |
914 nnz = decode_block_coeffs(c, s->block[y][x], s->prob->token[luma_ctx], luma_start, | |
915 nnz_pred, s->qmat[segment].luma_qmul); | |
12342 | 916 // nnz+block_dc may be one more than the actual last index, but we don't care |
917 s->non_zero_count_cache[y][x] = nnz + block_dc; | |
11921 | 918 t_nnz[x] = l_nnz[y] = !!nnz; |
919 nnz_total += nnz; | |
920 } | |
921 | |
922 // chroma blocks | |
923 // TODO: what to do about dimensions? 2nd dim for luma is x, | |
924 // but for chroma it's (y<<1)|x | |
925 for (i = 4; i < 6; i++) | |
926 for (y = 0; y < 2; y++) | |
927 for (x = 0; x < 2; x++) { | |
928 nnz_pred = l_nnz[i+2*y] + t_nnz[i+2*x]; | |
929 nnz = decode_block_coeffs(c, s->block[i][(y<<1)+x], s->prob->token[2], 0, | |
930 nnz_pred, s->qmat[segment].chroma_qmul); | |
931 s->non_zero_count_cache[i][(y<<1)+x] = nnz; | |
932 t_nnz[i+2*x] = l_nnz[i+2*y] = !!nnz; | |
933 nnz_total += nnz; | |
934 } | |
935 | |
936 // if there were no coded coeffs despite the macroblock not being marked skip, | |
937 // we MUST not do the inner loop filter and should not do IDCT | |
938 // Since skip isn't used for bitstream prediction, just manually set it. | |
939 if (!nnz_total) | |
940 mb->skip = 1; | |
941 } | |
942 | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
943 static av_always_inline |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
944 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
|
945 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
|
946 { |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
947 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
|
948 if (!simple) { |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
949 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
|
950 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
|
951 } |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
952 } |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
953 |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
954 static av_always_inline |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
955 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
|
956 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
|
957 int simple, int xchg) |
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 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
|
960 src_y -= linesize; |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
961 src_cb -= uvlinesize; |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
962 src_cr -= uvlinesize; |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
963 |
12202 | 964 #define XCHG(a,b,xchg) do { \ |
965 if (xchg) AV_SWAP64(b,a); \ | |
966 else AV_COPY64(b,a); \ | |
967 } while (0) | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
968 |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
969 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
|
970 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
|
971 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
|
972 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
|
973 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
|
974 |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
975 // 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
|
976 // 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
|
977 if (!simple || !mb_y) { |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
978 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
|
979 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
|
980 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
|
981 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
|
982 } |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
983 } |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
984 |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
985 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
986 int check_intra_pred_mode(int mode, int mb_x, int mb_y) |
11921 | 987 { |
988 if (mode == DC_PRED8x8) { | |
12243
788445bf10c0
VP8: shave a few clocks off check_intra_pred_mode
darkshikari
parents:
12242
diff
changeset
|
989 if (!mb_x) { |
788445bf10c0
VP8: shave a few clocks off check_intra_pred_mode
darkshikari
parents:
12242
diff
changeset
|
990 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
|
991 } else if (!mb_y) { |
12244 | 992 mode = LEFT_DC_PRED8x8; |
12243
788445bf10c0
VP8: shave a few clocks off check_intra_pred_mode
darkshikari
parents:
12242
diff
changeset
|
993 } |
11921 | 994 } |
995 return mode; | |
996 } | |
997 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
998 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
999 void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb, |
12339
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
1000 int mb_x, int mb_y) |
11921 | 1001 { |
1002 int x, y, mode, nnz, tr; | |
1003 | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
1004 // 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
|
1005 // 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
|
1006 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
|
1007 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
|
1008 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
|
1009 s->filter.simple, 1); |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
1010 |
11921 | 1011 if (mb->mode < MODE_I4x4) { |
1012 mode = check_intra_pred_mode(mb->mode, mb_x, mb_y); | |
1013 s->hpc.pred16x16[mode](dst[0], s->linesize); | |
1014 } else { | |
1015 uint8_t *ptr = dst[0]; | |
12339
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
1016 uint8_t *intra4x4 = s->intra4x4_pred_mode_mb; |
11921 | 1017 |
1018 // all blocks on the right edge of the macroblock use bottom edge | |
1019 // the top macroblock for their topright edge | |
1020 uint8_t *tr_right = ptr - s->linesize + 16; | |
1021 | |
1022 // if we're on the right edge of the frame, said edge is extended | |
1023 // from the top macroblock | |
1024 if (mb_x == s->mb_width-1) { | |
1025 tr = tr_right[-1]*0x01010101; | |
1026 tr_right = (uint8_t *)&tr; | |
1027 } | |
1028 | |
12234
bba849c2a113
VP8: avoid a memset for non-i4x4 blocks with no coefficients
darkshikari
parents:
12233
diff
changeset
|
1029 if (mb->skip) |
bba849c2a113
VP8: avoid a memset for non-i4x4 blocks with no coefficients
darkshikari
parents:
12233
diff
changeset
|
1030 AV_ZERO128(s->non_zero_count_cache); |
bba849c2a113
VP8: avoid a memset for non-i4x4 blocks with no coefficients
darkshikari
parents:
12233
diff
changeset
|
1031 |
11921 | 1032 for (y = 0; y < 4; y++) { |
1033 uint8_t *topright = ptr + 4 - s->linesize; | |
1034 for (x = 0; x < 4; x++) { | |
1035 if (x == 3) | |
1036 topright = tr_right; | |
1037 | |
12221 | 1038 s->hpc.pred4x4[intra4x4[x]](ptr+4*x, topright, s->linesize); |
11921 | 1039 |
1040 nnz = s->non_zero_count_cache[y][x]; | |
1041 if (nnz) { | |
1042 if (nnz == 1) | |
1043 s->vp8dsp.vp8_idct_dc_add(ptr+4*x, s->block[y][x], s->linesize); | |
1044 else | |
1045 s->vp8dsp.vp8_idct_add(ptr+4*x, s->block[y][x], s->linesize); | |
1046 } | |
1047 topright += 4; | |
1048 } | |
1049 | |
1050 ptr += 4*s->linesize; | |
12339
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
1051 intra4x4 += 4; |
11921 | 1052 } |
1053 } | |
1054 | |
1055 mode = check_intra_pred_mode(s->chroma_pred_mode, mb_x, mb_y); | |
1056 s->hpc.pred8x8[mode](dst[1], s->uvlinesize); | |
1057 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
|
1058 |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
1059 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
|
1060 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
|
1061 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
|
1062 s->filter.simple, 0); |
11921 | 1063 } |
1064 | |
1065 /** | |
1066 * Generic MC function. | |
1067 * | |
1068 * @param s VP8 decoding context | |
1069 * @param luma 1 for luma (Y) planes, 0 for chroma (Cb/Cr) planes | |
1070 * @param dst target buffer for block data at block position | |
1071 * @param src reference picture buffer at origin (0, 0) | |
1072 * @param mv motion vector (relative to block position) to get pixel data from | |
1073 * @param x_off horizontal position of block from origin (0, 0) | |
1074 * @param y_off vertical position of block from origin (0, 0) | |
1075 * @param block_w width of block (16, 8 or 4) | |
1076 * @param block_h height of block (always same as block_w) | |
1077 * @param width width of src/dst plane data | |
1078 * @param height height of src/dst plane data | |
1079 * @param linesize size of a single line of plane data, including padding | |
12115 | 1080 * @param mc_func motion compensation function pointers (bilinear or sixtap MC) |
11921 | 1081 */ |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1082 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1083 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
|
1084 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
|
1085 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
|
1086 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
|
1087 vp8_mc_func mc_func[3][3]) |
11921 | 1088 { |
12218 | 1089 if (AV_RN32A(mv)) { |
1090 static const uint8_t idx[8] = { 0, 1, 2, 1, 2, 1, 2, 1 }; | |
1091 int mx = (mv->x << luma)&7, mx_idx = idx[mx]; | |
1092 int my = (mv->y << luma)&7, my_idx = idx[my]; | |
11921 | 1093 |
12218 | 1094 x_off += mv->x >> (3 - luma); |
1095 y_off += mv->y >> (3 - luma); | |
11921 | 1096 |
12218 | 1097 // edge emulation |
1098 src += y_off * linesize + x_off; | |
1099 if (x_off < 2 || x_off >= width - block_w - 3 || | |
1100 y_off < 2 || y_off >= height - block_h - 3) { | |
1101 ff_emulated_edge_mc(s->edge_emu_buffer, src - 2 * linesize - 2, linesize, | |
1102 block_w + 5, block_h + 5, | |
1103 x_off - 2, y_off - 2, width, height); | |
1104 src = s->edge_emu_buffer + 2 + linesize * 2; | |
1105 } | |
1106 mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my); | |
1107 } else | |
1108 mc_func[0][0](dst, linesize, src + y_off * linesize + x_off, linesize, block_h, 0, 0); | |
11921 | 1109 } |
1110 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1111 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1112 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
|
1113 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
|
1114 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
|
1115 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
|
1116 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
|
1117 { |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1118 VP56mv uvmv = *mv; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1119 |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1120 /* Y */ |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1121 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
|
1122 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
|
1123 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
|
1124 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
|
1125 |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1126 /* U/V */ |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1127 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
|
1128 uvmv.x &= ~7; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1129 uvmv.y &= ~7; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1130 } |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1131 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
|
1132 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
|
1133 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
|
1134 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
|
1135 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
|
1136 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
|
1137 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
|
1138 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
|
1139 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
|
1140 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
|
1141 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
|
1142 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
|
1143 } |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1144 |
12215 | 1145 /* Fetch pixels for estimated mv 4 macroblocks ahead. |
1146 * 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
|
1147 static av_always_inline void prefetch_motion(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, int mb_xy, int ref) |
12215 | 1148 { |
12237 | 1149 /* Don't prefetch refs that haven't been used very often this frame. */ |
1150 if (s->ref_count[ref-1] > (mb_xy >> 5)) { | |
12231 | 1151 int x_off = mb_x << 4, y_off = mb_y << 4; |
1152 int mx = mb->mv.x + x_off + 8; | |
1153 int my = mb->mv.y + y_off; | |
1154 uint8_t **src= s->framep[ref]->data; | |
1155 int off= mx + (my + (mb_x&3)*4)*s->linesize + 64; | |
1156 s->dsp.prefetch(src[0]+off, s->linesize, 4); | |
1157 off= (mx>>1) + ((my>>1) + (mb_x&7))*s->uvlinesize + 64; | |
1158 s->dsp.prefetch(src[1]+off, src[2]-src[1], 2); | |
1159 } | |
12215 | 1160 } |
1161 | |
11921 | 1162 /** |
1163 * Apply motion vectors to prediction buffer, chapter 18. | |
1164 */ | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1165 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1166 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
|
1167 int mb_x, int mb_y) |
11921 | 1168 { |
1169 int x_off = mb_x << 4, y_off = mb_y << 4; | |
1170 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
|
1171 AVFrame *ref = s->framep[mb->ref_frame]; |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1172 VP56mv *bmv = mb->bmv; |
11921 | 1173 |
1174 if (mb->mode < VP8_MVMODE_SPLIT) { | |
12228
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1175 vp8_mc_part(s, dst, ref, x_off, y_off, |
11989
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1176 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
|
1177 } 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
|
1178 case VP8_SPLITMVMODE_4x4: { |
11921 | 1179 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
|
1180 VP56mv uvmv; |
11921 | 1181 |
1182 /* Y */ | |
1183 for (y = 0; y < 4; y++) { | |
1184 for (x = 0; x < 4; x++) { | |
1185 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
|
1186 ref->data[0], &bmv[4*y + x], |
11921 | 1187 4*x + x_off, 4*y + y_off, 4, 4, |
1188 width, height, s->linesize, | |
11974 | 1189 s->put_pixels_tab[2]); |
11921 | 1190 } |
1191 } | |
1192 | |
1193 /* U/V */ | |
1194 x_off >>= 1; y_off >>= 1; width >>= 1; height >>= 1; | |
1195 for (y = 0; y < 2; y++) { | |
1196 for (x = 0; x < 2; x++) { | |
1197 uvmv.x = mb->bmv[ 2*y * 4 + 2*x ].x + | |
1198 mb->bmv[ 2*y * 4 + 2*x+1].x + | |
1199 mb->bmv[(2*y+1) * 4 + 2*x ].x + | |
1200 mb->bmv[(2*y+1) * 4 + 2*x+1].x; | |
1201 uvmv.y = mb->bmv[ 2*y * 4 + 2*x ].y + | |
1202 mb->bmv[ 2*y * 4 + 2*x+1].y + | |
1203 mb->bmv[(2*y+1) * 4 + 2*x ].y + | |
1204 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
|
1205 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
|
1206 uvmv.y = (uvmv.y + 2 + (uvmv.y >> (INT_BIT-1))) >> 2; |
11921 | 1207 if (s->profile == 3) { |
1208 uvmv.x &= ~7; | |
1209 uvmv.y &= ~7; | |
1210 } | |
1211 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
|
1212 ref->data[1], &uvmv, |
11921 | 1213 4*x + x_off, 4*y + y_off, 4, 4, |
1214 width, height, s->uvlinesize, | |
11974 | 1215 s->put_pixels_tab[2]); |
11921 | 1216 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
|
1217 ref->data[2], &uvmv, |
11921 | 1218 4*x + x_off, 4*y + y_off, 4, 4, |
1219 width, height, s->uvlinesize, | |
11974 | 1220 s->put_pixels_tab[2]); |
11921 | 1221 } |
1222 } | |
11989
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1223 break; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1224 } |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1225 case VP8_SPLITMVMODE_16x8: |
12228
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1226 vp8_mc_part(s, dst, ref, x_off, y_off, |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1227 0, 0, 16, 8, width, height, &bmv[0]); |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1228 vp8_mc_part(s, dst, ref, x_off, y_off, |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1229 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
|
1230 break; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1231 case VP8_SPLITMVMODE_8x16: |
12228
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1232 vp8_mc_part(s, dst, ref, x_off, y_off, |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1233 0, 0, 8, 16, width, height, &bmv[0]); |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1234 vp8_mc_part(s, dst, ref, x_off, y_off, |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1235 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
|
1236 break; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1237 case VP8_SPLITMVMODE_8x8: |
12228
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1238 vp8_mc_part(s, dst, ref, x_off, y_off, |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1239 0, 0, 8, 8, width, height, &bmv[0]); |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1240 vp8_mc_part(s, dst, ref, x_off, y_off, |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1241 8, 0, 8, 8, width, height, &bmv[1]); |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1242 vp8_mc_part(s, dst, ref, x_off, y_off, |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1243 0, 8, 8, 8, width, height, &bmv[2]); |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1244 vp8_mc_part(s, dst, ref, x_off, y_off, |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1245 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
|
1246 break; |
11921 | 1247 } |
1248 } | |
1249 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1250 static av_always_inline void idct_mb(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb) |
11921 | 1251 { |
12240 | 1252 int x, y, ch; |
11921 | 1253 |
12238 | 1254 if (mb->mode != MODE_I4x4) { |
1255 uint8_t *y_dst = dst[0]; | |
11921 | 1256 for (y = 0; y < 4; y++) { |
12240 | 1257 uint32_t nnz4 = AV_RN32A(s->non_zero_count_cache[y]); |
1258 if (nnz4) { | |
1259 if (nnz4&~0x01010101) { | |
12238 | 1260 for (x = 0; x < 4; x++) { |
12240 | 1261 int nnz = s->non_zero_count_cache[y][x]; |
12238 | 1262 if (nnz) { |
1263 if (nnz == 1) | |
1264 s->vp8dsp.vp8_idct_dc_add(y_dst+4*x, s->block[y][x], s->linesize); | |
1265 else | |
1266 s->vp8dsp.vp8_idct_add(y_dst+4*x, s->block[y][x], s->linesize); | |
1267 } | |
1268 } | |
1269 } else { | |
12241
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1270 s->vp8dsp.vp8_idct_dc_add4y(y_dst, s->block[y], s->linesize); |
11921 | 1271 } |
1272 } | |
1273 y_dst += 4*s->linesize; | |
1274 } | |
12238 | 1275 } |
11921 | 1276 |
12238 | 1277 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
|
1278 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
|
1279 if (nnz4) { |
12238 | 1280 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
|
1281 if (nnz4&~0x01010101) { |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1282 for (y = 0; y < 2; y++) { |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1283 for (x = 0; x < 2; x++) { |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1284 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
|
1285 if (nnz) { |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1286 if (nnz == 1) |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1287 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
|
1288 else |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1289 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
|
1290 } |
12238 | 1291 } |
12241
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1292 ch_dst += 4*s->uvlinesize; |
12238 | 1293 } |
12241
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1294 } else { |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1295 s->vp8dsp.vp8_idct_dc_add4uv(ch_dst, s->block[4+ch], s->uvlinesize); |
11921 | 1296 } |
1297 } | |
1298 } | |
1299 } | |
1300 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1301 static av_always_inline void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb, VP8FilterStrength *f ) |
11921 | 1302 { |
1303 int interior_limit, filter_level; | |
1304 | |
1305 if (s->segmentation.enabled) { | |
12224
5b7d690b761b
VP8: Don't store segment in macroblock struct anymore.
darkshikari
parents:
12223
diff
changeset
|
1306 filter_level = s->segmentation.filter_level[s->segment]; |
11921 | 1307 if (!s->segmentation.absolute_vals) |
1308 filter_level += s->filter.level; | |
1309 } else | |
1310 filter_level = s->filter.level; | |
1311 | |
1312 if (s->lf_delta.enabled) { | |
1313 filter_level += s->lf_delta.ref[mb->ref_frame]; | |
1314 | |
1315 if (mb->ref_frame == VP56_FRAME_CURRENT) { | |
1316 if (mb->mode == MODE_I4x4) | |
1317 filter_level += s->lf_delta.mode[0]; | |
1318 } else { | |
1319 if (mb->mode == VP8_MVMODE_ZERO) | |
1320 filter_level += s->lf_delta.mode[1]; | |
1321 else if (mb->mode == VP8_MVMODE_SPLIT) | |
1322 filter_level += s->lf_delta.mode[3]; | |
1323 else | |
1324 filter_level += s->lf_delta.mode[2]; | |
1325 } | |
1326 } | |
1327 filter_level = av_clip(filter_level, 0, 63); | |
1328 | |
1329 interior_limit = filter_level; | |
1330 if (s->filter.sharpness) { | |
1331 interior_limit >>= s->filter.sharpness > 4 ? 2 : 1; | |
1332 interior_limit = FFMIN(interior_limit, 9 - s->filter.sharpness); | |
1333 } | |
1334 interior_limit = FFMAX(interior_limit, 1); | |
1335 | |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1336 f->filter_level = filter_level; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1337 f->inner_limit = interior_limit; |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1338 f->inner_filter = !mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT; |
11921 | 1339 } |
1340 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1341 static av_always_inline void filter_mb(VP8Context *s, uint8_t *dst[3], VP8FilterStrength *f, int mb_x, int mb_y) |
11921 | 1342 { |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1343 int mbedge_lim, bedge_lim, hev_thresh; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1344 int filter_level = f->filter_level; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1345 int inner_limit = f->inner_limit; |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1346 int inner_filter = f->inner_filter; |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1347 int linesize = s->linesize; |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1348 int uvlinesize = s->uvlinesize; |
11921 | 1349 |
1350 if (!filter_level) | |
1351 return; | |
1352 | |
12081
812e23197d64
VP8: Move calculation of outer filter limit out of dsp functions for normal
conrad
parents:
12062
diff
changeset
|
1353 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
|
1354 bedge_lim = 2* filter_level + inner_limit; |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1355 hev_thresh = filter_level >= 15; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1356 |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1357 if (s->keyframe) { |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1358 if (filter_level >= 40) |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1359 hev_thresh = 2; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1360 } else { |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1361 if (filter_level >= 40) |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1362 hev_thresh = 3; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1363 else if (filter_level >= 20) |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1364 hev_thresh = 2; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1365 } |
12081
812e23197d64
VP8: Move calculation of outer filter limit out of dsp functions for normal
conrad
parents:
12062
diff
changeset
|
1366 |
11921 | 1367 if (mb_x) { |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1368 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
|
1369 mbedge_lim, inner_limit, hev_thresh); |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1370 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
|
1371 mbedge_lim, inner_limit, hev_thresh); |
11921 | 1372 } |
1373 | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1374 if (inner_filter) { |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1375 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
|
1376 inner_limit, hev_thresh); |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1377 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
|
1378 inner_limit, hev_thresh); |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1379 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
|
1380 inner_limit, hev_thresh); |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1381 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
|
1382 uvlinesize, bedge_lim, |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1383 inner_limit, hev_thresh); |
11921 | 1384 } |
1385 | |
1386 if (mb_y) { | |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1387 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
|
1388 mbedge_lim, inner_limit, hev_thresh); |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1389 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
|
1390 mbedge_lim, inner_limit, hev_thresh); |
11921 | 1391 } |
1392 | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1393 if (inner_filter) { |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1394 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
|
1395 linesize, bedge_lim, |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1396 inner_limit, hev_thresh); |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1397 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
|
1398 linesize, bedge_lim, |
12194
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1399 inner_limit, hev_thresh); |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1400 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
|
1401 linesize, bedge_lim, |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1402 inner_limit, hev_thresh); |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1403 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
|
1404 dst[2] + 4 * uvlinesize, |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1405 uvlinesize, bedge_lim, |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1406 inner_limit, hev_thresh); |
11921 | 1407 } |
1408 } | |
1409 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1410 static av_always_inline void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8FilterStrength *f, int mb_x, int mb_y) |
11921 | 1411 { |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1412 int mbedge_lim, bedge_lim; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1413 int filter_level = f->filter_level; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1414 int inner_limit = f->inner_limit; |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1415 int inner_filter = f->inner_filter; |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1416 int linesize = s->linesize; |
11921 | 1417 |
1418 if (!filter_level) | |
1419 return; | |
1420 | |
1421 mbedge_lim = 2*(filter_level+2) + inner_limit; | |
1422 bedge_lim = 2* filter_level + inner_limit; | |
1423 | |
1424 if (mb_x) | |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1425 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
|
1426 if (inner_filter) { |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1427 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
|
1428 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
|
1429 s->vp8dsp.vp8_h_loop_filter_simple(dst+12, linesize, bedge_lim); |
11921 | 1430 } |
1431 | |
1432 if (mb_y) | |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1433 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
|
1434 if (inner_filter) { |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1435 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
|
1436 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
|
1437 s->vp8dsp.vp8_v_loop_filter_simple(dst+12*linesize, linesize, bedge_lim); |
11921 | 1438 } |
1439 } | |
1440 | |
1441 static void filter_mb_row(VP8Context *s, int mb_y) | |
1442 { | |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1443 VP8FilterStrength *f = s->filter_strength; |
11921 | 1444 uint8_t *dst[3] = { |
1445 s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize, | |
1446 s->framep[VP56_FRAME_CURRENT]->data[1] + 8*mb_y*s->uvlinesize, | |
1447 s->framep[VP56_FRAME_CURRENT]->data[2] + 8*mb_y*s->uvlinesize | |
1448 }; | |
1449 int mb_x; | |
1450 | |
1451 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
|
1452 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
|
1453 filter_mb(s, dst, f++, mb_x, mb_y); |
11921 | 1454 dst[0] += 16; |
1455 dst[1] += 8; | |
1456 dst[2] += 8; | |
1457 } | |
1458 } | |
1459 | |
1460 static void filter_mb_row_simple(VP8Context *s, int mb_y) | |
1461 { | |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1462 VP8FilterStrength *f = s->filter_strength; |
11921 | 1463 uint8_t *dst = s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize; |
1464 int mb_x; | |
1465 | |
1466 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
|
1467 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
|
1468 filter_mb_simple(s, dst, f++, mb_x, mb_y); |
11921 | 1469 dst += 16; |
1470 } | |
1471 } | |
1472 | |
1473 static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size, | |
1474 AVPacket *avpkt) | |
1475 { | |
1476 VP8Context *s = avctx->priv_data; | |
1477 int ret, mb_x, mb_y, i, y, referenced; | |
1478 enum AVDiscard skip_thresh; | |
12270
161c205dcdd2
Fix r24445: Instead of needlessly initialising a variable, silence the warning.
cehoyos
parents:
12255
diff
changeset
|
1479 AVFrame *av_uninit(curframe); |
11921 | 1480 |
1481 if ((ret = decode_frame_header(s, avpkt->data, avpkt->size)) < 0) | |
1482 return ret; | |
1483 | |
1484 referenced = s->update_last || s->update_golden == VP56_FRAME_CURRENT | |
1485 || s->update_altref == VP56_FRAME_CURRENT; | |
1486 | |
1487 skip_thresh = !referenced ? AVDISCARD_NONREF : | |
1488 !s->keyframe ? AVDISCARD_NONKEY : AVDISCARD_ALL; | |
1489 | |
1490 if (avctx->skip_frame >= skip_thresh) { | |
1491 s->invisible = 1; | |
1492 goto skip_decode; | |
1493 } | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
1494 s->deblock_filter = s->filter.level && avctx->skip_loop_filter < skip_thresh; |
11921 | 1495 |
1496 for (i = 0; i < 4; i++) | |
1497 if (&s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] && | |
1498 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] && | |
1499 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) { | |
1500 curframe = s->framep[VP56_FRAME_CURRENT] = &s->frames[i]; | |
1501 break; | |
1502 } | |
1503 if (curframe->data[0]) | |
1504 avctx->release_buffer(avctx, curframe); | |
1505 | |
1506 curframe->key_frame = s->keyframe; | |
1507 curframe->pict_type = s->keyframe ? FF_I_TYPE : FF_P_TYPE; | |
1508 curframe->reference = referenced ? 3 : 0; | |
1509 if ((ret = avctx->get_buffer(avctx, curframe))) { | |
1510 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed!\n"); | |
1511 return ret; | |
1512 } | |
1513 | |
1514 // Given that arithmetic probabilities are updated every frame, it's quite likely | |
1515 // that the values we have on a random interframe are complete junk if we didn't | |
1516 // start decode on a keyframe. So just don't display anything rather than junk. | |
1517 if (!s->keyframe && (!s->framep[VP56_FRAME_PREVIOUS] || | |
1518 !s->framep[VP56_FRAME_GOLDEN] || | |
1519 !s->framep[VP56_FRAME_GOLDEN2])) { | |
1520 av_log(avctx, AV_LOG_WARNING, "Discarding interframe without a prior keyframe!\n"); | |
1521 return AVERROR_INVALIDDATA; | |
1522 } | |
1523 | |
1524 s->linesize = curframe->linesize[0]; | |
1525 s->uvlinesize = curframe->linesize[1]; | |
1526 | |
1527 if (!s->edge_emu_buffer) | |
1528 s->edge_emu_buffer = av_malloc(21*s->linesize); | |
1529 | |
1530 memset(s->top_nnz, 0, s->mb_width*sizeof(*s->top_nnz)); | |
1531 | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1532 /* 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
|
1533 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
|
1534 |
11921 | 1535 // 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
|
1536 memset(s->top_border, 127, (s->mb_width+1)*sizeof(*s->top_border)); |
12231 | 1537 memset(s->ref_count, 0, sizeof(s->ref_count)); |
12339
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
1538 if (s->keyframe) |
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
1539 memset(s->intra4x4_pred_mode_top, DC_PRED, s->b4_stride*4); |
11921 | 1540 |
1541 for (mb_y = 0; mb_y < s->mb_height; mb_y++) { | |
1542 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
|
1543 VP8Macroblock *mb = s->macroblocks + (s->mb_height - mb_y - 1)*2; |
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1544 uint8_t *segment_map = s->segmentation_map + mb_y*s->mb_stride; |
12237 | 1545 int mb_xy = mb_y * s->mb_stride; |
11921 | 1546 uint8_t *dst[3] = { |
1547 curframe->data[0] + 16*mb_y*s->linesize, | |
1548 curframe->data[1] + 8*mb_y*s->uvlinesize, | |
1549 curframe->data[2] + 8*mb_y*s->uvlinesize | |
1550 }; | |
1551 | |
1552 memset(s->left_nnz, 0, sizeof(s->left_nnz)); | |
12339
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
1553 AV_WN32A(s->intra4x4_pred_mode_left, DC_PRED*0x01010101); |
11921 | 1554 |
1555 // left edge of 129 for intra prediction | |
1556 if (!(avctx->flags & CODEC_FLAG_EMU_EDGE)) | |
1557 for (i = 0; i < 3; i++) | |
1558 for (y = 0; y < 16>>!!i; y++) | |
1559 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
|
1560 if (mb_y) |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
1561 memset(s->top_border, 129, sizeof(*s->top_border)); |
11921 | 1562 |
12237 | 1563 for (mb_x = 0; mb_x < s->mb_width; mb_x++, mb_xy++, mb++) { |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1564 uint8_t *segment_mb = segment_map+mb_x; |
12221 | 1565 |
12215 | 1566 /* Prefetch the current frame, 4 MBs ahead */ |
1567 s->dsp.prefetch(dst[0] + (mb_x&3)*4*s->linesize + 64, s->linesize, 4); | |
1568 s->dsp.prefetch(dst[1] + (mb_x&7)*s->uvlinesize + 64, dst[2] - dst[1], 2); | |
1569 | |
12339
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
1570 decode_mb_mode(s, mb, mb_x, mb_y, segment_mb); |
11921 | 1571 |
12237 | 1572 prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_PREVIOUS); |
12231 | 1573 |
11921 | 1574 if (!mb->skip) |
1575 decode_mb_coeffs(s, c, mb, s->top_nnz[mb_x], s->left_nnz); | |
1576 | |
12225
c3e11b3108d7
Eliminate a pointless memset for intra blocks in P-frames in VP8
darkshikari
parents:
12224
diff
changeset
|
1577 if (mb->mode <= MODE_I4x4) |
12339
57fc7f2d7b28
only store intra prediction modes on the boundary for keyframes, not as a plane.
skal
parents:
12336
diff
changeset
|
1578 intra_predict(s, dst, mb, mb_x, mb_y); |
12225
c3e11b3108d7
Eliminate a pointless memset for intra blocks in P-frames in VP8
darkshikari
parents:
12224
diff
changeset
|
1579 else |
11921 | 1580 inter_predict(s, dst, mb, mb_x, mb_y); |
1581 | |
12237 | 1582 prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN); |
12231 | 1583 |
11921 | 1584 if (!mb->skip) { |
12238 | 1585 idct_mb(s, dst, mb); |
11921 | 1586 } else { |
1587 AV_ZERO64(s->left_nnz); | |
1588 AV_WN64(s->top_nnz[mb_x], 0); // array of 9, so unaligned | |
1589 | |
1590 // Reset DC block predictors if they would exist if the mb had coefficients | |
1591 if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) { | |
1592 s->left_nnz[8] = 0; | |
1593 s->top_nnz[mb_x][8] = 0; | |
1594 } | |
1595 } | |
1596 | |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1597 if (s->deblock_filter) |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1598 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
|
1599 |
12237 | 1600 prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN2); |
12231 | 1601 |
11921 | 1602 dst[0] += 16; |
1603 dst[1] += 8; | |
1604 dst[2] += 8; | |
1605 } | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
1606 if (s->deblock_filter) { |
11921 | 1607 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
|
1608 filter_mb_row_simple(s, mb_y); |
11921 | 1609 else |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
1610 filter_mb_row(s, mb_y); |
11921 | 1611 } |
1612 } | |
1613 | |
1614 skip_decode: | |
1615 // if future frames don't use the updated probabilities, | |
1616 // reset them to the values we saved | |
1617 if (!s->update_probabilities) | |
1618 s->prob[0] = s->prob[1]; | |
1619 | |
1620 // check if golden and altref are swapped | |
1621 if (s->update_altref == VP56_FRAME_GOLDEN && | |
1622 s->update_golden == VP56_FRAME_GOLDEN2) | |
1623 FFSWAP(AVFrame *, s->framep[VP56_FRAME_GOLDEN], s->framep[VP56_FRAME_GOLDEN2]); | |
1624 else { | |
1625 if (s->update_altref != VP56_FRAME_NONE) | |
1626 s->framep[VP56_FRAME_GOLDEN2] = s->framep[s->update_altref]; | |
1627 | |
1628 if (s->update_golden != VP56_FRAME_NONE) | |
1629 s->framep[VP56_FRAME_GOLDEN] = s->framep[s->update_golden]; | |
1630 } | |
1631 | |
1632 if (s->update_last) // move cur->prev | |
1633 s->framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_CURRENT]; | |
1634 | |
1635 // release no longer referenced frames | |
1636 for (i = 0; i < 4; i++) | |
1637 if (s->frames[i].data[0] && | |
1638 &s->frames[i] != s->framep[VP56_FRAME_CURRENT] && | |
1639 &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] && | |
1640 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] && | |
1641 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) | |
1642 avctx->release_buffer(avctx, &s->frames[i]); | |
1643 | |
1644 if (!s->invisible) { | |
1645 *(AVFrame*)data = *s->framep[VP56_FRAME_CURRENT]; | |
1646 *data_size = sizeof(AVFrame); | |
1647 } | |
1648 | |
1649 return avpkt->size; | |
1650 } | |
1651 | |
1652 static av_cold int vp8_decode_init(AVCodecContext *avctx) | |
1653 { | |
1654 VP8Context *s = avctx->priv_data; | |
1655 | |
1656 s->avctx = avctx; | |
1657 avctx->pix_fmt = PIX_FMT_YUV420P; | |
1658 | |
1659 dsputil_init(&s->dsp, avctx); | |
1660 ff_h264_pred_init(&s->hpc, CODEC_ID_VP8); | |
1661 ff_vp8dsp_init(&s->vp8dsp); | |
1662 | |
1663 // intra pred needs edge emulation among other things | |
1664 if (avctx->flags&CODEC_FLAG_EMU_EDGE) { | |
11947 | 1665 av_log(avctx, AV_LOG_ERROR, "Edge emulation not supported\n"); |
11921 | 1666 return AVERROR_PATCHWELCOME; |
1667 } | |
1668 | |
1669 return 0; | |
1670 } | |
1671 | |
1672 static av_cold int vp8_decode_free(AVCodecContext *avctx) | |
1673 { | |
1674 vp8_decode_flush(avctx); | |
1675 return 0; | |
1676 } | |
1677 | |
1678 AVCodec vp8_decoder = { | |
1679 "vp8", | |
1680 AVMEDIA_TYPE_VIDEO, | |
1681 CODEC_ID_VP8, | |
1682 sizeof(VP8Context), | |
1683 vp8_decode_init, | |
1684 NULL, | |
1685 vp8_decode_free, | |
1686 vp8_decode_frame, | |
1687 CODEC_CAP_DR1, | |
1688 .flush = vp8_decode_flush, | |
1689 .long_name = NULL_IF_CONFIG_SMALL("On2 VP8"), | |
1690 }; |