Mercurial > libavcodec.hg
annotate vp8.c @ 12289:bc679bd1c66d libavcodec
perform the clipping on luma_dc_qmul[1] and chroma_qmul[0] earlier
author | skal |
---|---|
date | Tue, 27 Jul 2010 22:23:50 +0000 |
parents | 6bdbab1d9ed4 |
children | 2a09b276db12 |
rev | line source |
---|---|
11921 | 1 /** |
2 * VP8 compatible video decoder | |
3 * | |
4 * Copyright (C) 2010 David Conrad | |
5 * Copyright (C) 2010 Ronald S. Bultje | |
12249 | 6 * Copyright (C) 2010 Jason Garrett-Glaser |
11921 | 7 * |
8 * This file is part of FFmpeg. | |
9 * | |
10 * FFmpeg is free software; you can redistribute it and/or | |
11 * modify it under the terms of the GNU Lesser General Public | |
12 * License as published by the Free Software Foundation; either | |
13 * version 2.1 of the License, or (at your option) any later version. | |
14 * | |
15 * FFmpeg is distributed in the hope that it will be useful, | |
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
18 * Lesser General Public License for more details. | |
19 * | |
20 * You should have received a copy of the GNU Lesser General Public | |
21 * License along with FFmpeg; if not, write to the Free Software | |
22 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | |
23 */ | |
24 | |
25 #include "avcodec.h" | |
26 #include "vp56.h" | |
27 #include "vp8data.h" | |
28 #include "vp8dsp.h" | |
29 #include "h264pred.h" | |
30 #include "rectangle.h" | |
31 | |
32 typedef struct { | |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
33 uint8_t filter_level; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
34 uint8_t inner_limit; |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
35 uint8_t inner_filter; |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
36 } VP8FilterStrength; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
37 |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
38 typedef struct { |
11921 | 39 uint8_t skip; |
40 // todo: make it possible to check for at least (i4x4 or split_mv) | |
41 // in one op. are others needed? | |
42 uint8_t mode; | |
43 uint8_t ref_frame; | |
44 uint8_t partitioning; | |
45 VP56mv mv; | |
46 VP56mv bmv[16]; | |
47 } VP8Macroblock; | |
48 | |
49 typedef struct { | |
50 AVCodecContext *avctx; | |
51 DSPContext dsp; | |
52 VP8DSPContext vp8dsp; | |
53 H264PredContext hpc; | |
11974 | 54 vp8_mc_func put_pixels_tab[3][3][3]; |
11921 | 55 AVFrame frames[4]; |
56 AVFrame *framep[4]; | |
57 uint8_t *edge_emu_buffer; | |
58 VP56RangeCoder c; ///< header context, includes mb modes and motion vectors | |
59 int profile; | |
60 | |
61 int mb_width; /* number of horizontal MB */ | |
62 int mb_height; /* number of vertical MB */ | |
63 int linesize; | |
64 int uvlinesize; | |
65 | |
66 int keyframe; | |
67 int invisible; | |
68 int update_last; ///< update VP56_FRAME_PREVIOUS with the current one | |
69 int update_golden; ///< VP56_FRAME_NONE if not updated, or which frame to copy if so | |
70 int update_altref; | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
71 int deblock_filter; |
11921 | 72 |
73 /** | |
74 * If this flag is not set, all the probability updates | |
75 * are discarded after this frame is decoded. | |
76 */ | |
77 int update_probabilities; | |
78 | |
79 /** | |
80 * All coefficients are contained in separate arith coding contexts. | |
81 * There can be 1, 2, 4, or 8 of these after the header context. | |
82 */ | |
83 int num_coeff_partitions; | |
84 VP56RangeCoder coeff_partition[8]; | |
85 | |
86 VP8Macroblock *macroblocks; | |
87 VP8Macroblock *macroblocks_base; | |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
88 VP8FilterStrength *filter_strength; |
11921 | 89 int mb_stride; |
90 | |
91 uint8_t *intra4x4_pred_mode; | |
92 uint8_t *intra4x4_pred_mode_base; | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
93 uint8_t *segmentation_map; |
11921 | 94 int b4_stride; |
95 | |
96 /** | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
97 * Cache of the top row needed for intra prediction |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
98 * 16 for luma, 8 for each chroma plane |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
99 */ |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
100 uint8_t (*top_border)[16+8+8]; |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
101 |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
102 /** |
11921 | 103 * For coeff decode, we need to know whether the above block had non-zero |
104 * coefficients. This means for each macroblock, we need data for 4 luma | |
105 * blocks, 2 u blocks, 2 v blocks, and the luma dc block, for a total of 9 | |
106 * per macroblock. We keep the last row in top_nnz. | |
107 */ | |
108 uint8_t (*top_nnz)[9]; | |
109 DECLARE_ALIGNED(8, uint8_t, left_nnz)[9]; | |
110 | |
111 /** | |
112 * This is the index plus one of the last non-zero coeff | |
113 * for each of the blocks in the current macroblock. | |
114 * So, 0 -> no coeffs | |
115 * 1 -> dc-only (special transform) | |
116 * 2+-> full transform | |
117 */ | |
118 DECLARE_ALIGNED(16, uint8_t, non_zero_count_cache)[6][4]; | |
119 DECLARE_ALIGNED(16, DCTELEM, block)[6][4][16]; | |
12221 | 120 uint8_t intra4x4_pred_mode_mb[16]; |
11921 | 121 |
122 int chroma_pred_mode; ///< 8x8c pred mode of the current macroblock | |
12224
5b7d690b761b
VP8: Don't store segment in macroblock struct anymore.
darkshikari
parents:
12223
diff
changeset
|
123 int segment; ///< segment of the current macroblock |
11921 | 124 |
125 int mbskip_enabled; | |
12288
6bdbab1d9ed4
save some copies by moving some fields out of proba[2]
skal
parents:
12271
diff
changeset
|
126 uint8_t mbskip_proba; |
6bdbab1d9ed4
save some copies by moving some fields out of proba[2]
skal
parents:
12271
diff
changeset
|
127 uint8_t intra_proba; |
6bdbab1d9ed4
save some copies by moving some fields out of proba[2]
skal
parents:
12271
diff
changeset
|
128 uint8_t last_proba; |
6bdbab1d9ed4
save some copies by moving some fields out of proba[2]
skal
parents:
12271
diff
changeset
|
129 uint8_t golden_proba; |
11921 | 130 int sign_bias[4]; ///< one state [0, 1] per ref frame type |
12231 | 131 int ref_count[3]; |
11921 | 132 |
133 /** | |
134 * Base parameters for segmentation, i.e. per-macroblock parameters. | |
135 * These must be kept unchanged even if segmentation is not used for | |
136 * a frame, since the values persist between interframes. | |
137 */ | |
138 struct { | |
139 int enabled; | |
140 int absolute_vals; | |
141 int update_map; | |
142 int8_t base_quant[4]; | |
143 int8_t filter_level[4]; ///< base loop filter level | |
144 } segmentation; | |
145 | |
146 /** | |
147 * Macroblocks can have one of 4 different quants in a frame when | |
148 * segmentation is enabled. | |
149 * If segmentation is disabled, only the first segment's values are used. | |
150 */ | |
151 struct { | |
152 // [0] - DC qmul [1] - AC qmul | |
153 int16_t luma_qmul[2]; | |
154 int16_t luma_dc_qmul[2]; ///< luma dc-only block quant | |
155 int16_t chroma_qmul[2]; | |
156 } qmat[4]; | |
157 | |
158 struct { | |
159 int simple; | |
160 int level; | |
161 int sharpness; | |
162 } filter; | |
163 | |
164 struct { | |
165 int enabled; ///< whether each mb can have a different strength based on mode/ref | |
166 | |
167 /** | |
168 * filter strength adjustment for the following macroblock modes: | |
169 * [0] - i4x4 | |
170 * [1] - zero mv | |
171 * [2] - inter modes except for zero or split mv | |
172 * [3] - split mv | |
173 * i16x16 modes never have any adjustment | |
174 */ | |
175 int8_t mode[4]; | |
176 | |
177 /** | |
178 * filter strength adjustment for macroblocks that reference: | |
179 * [0] - intra / VP56_FRAME_CURRENT | |
180 * [1] - VP56_FRAME_PREVIOUS | |
181 * [2] - VP56_FRAME_GOLDEN | |
182 * [3] - altref / VP56_FRAME_GOLDEN2 | |
183 */ | |
184 int8_t ref[4]; | |
185 } lf_delta; | |
186 | |
187 /** | |
188 * These are all of the updatable probabilities for binary decisions. | |
189 * They are only implictly reset on keyframes, making it quite likely | |
190 * for an interframe to desync if a prior frame's header was corrupt | |
191 * or missing outright! | |
192 */ | |
193 struct { | |
194 uint8_t segmentid[3]; | |
195 uint8_t pred16x16[4]; | |
196 uint8_t pred8x8c[3]; | |
197 uint8_t token[4][8][3][NUM_DCT_TOKENS-1]; | |
198 uint8_t mvc[2][19]; | |
199 } prob[2]; | |
200 } VP8Context; | |
201 | |
202 static void vp8_decode_flush(AVCodecContext *avctx) | |
203 { | |
204 VP8Context *s = avctx->priv_data; | |
205 int i; | |
206 | |
207 for (i = 0; i < 4; i++) | |
208 if (s->frames[i].data[0]) | |
209 avctx->release_buffer(avctx, &s->frames[i]); | |
210 memset(s->framep, 0, sizeof(s->framep)); | |
211 | |
212 av_freep(&s->macroblocks_base); | |
12271 | 213 av_freep(&s->filter_strength); |
11921 | 214 av_freep(&s->intra4x4_pred_mode_base); |
215 av_freep(&s->top_nnz); | |
216 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
|
217 av_freep(&s->top_border); |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
218 av_freep(&s->segmentation_map); |
11921 | 219 |
220 s->macroblocks = NULL; | |
221 s->intra4x4_pred_mode = NULL; | |
222 } | |
223 | |
224 static int update_dimensions(VP8Context *s, int width, int height) | |
225 { | |
226 int i; | |
227 | |
228 if (avcodec_check_dimensions(s->avctx, width, height)) | |
229 return AVERROR_INVALIDDATA; | |
230 | |
231 vp8_decode_flush(s->avctx); | |
232 | |
233 avcodec_set_dimensions(s->avctx, width, height); | |
234 | |
235 s->mb_width = (s->avctx->coded_width +15) / 16; | |
236 s->mb_height = (s->avctx->coded_height+15) / 16; | |
237 | |
238 // we allocate a border around the top/left of intra4x4 modes | |
239 // this is 4 blocks for intra4x4 to keep 4-byte alignment for fill_rectangle | |
240 s->mb_stride = s->mb_width+1; | |
241 s->b4_stride = 4*s->mb_stride; | |
242 | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
243 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
|
244 s->filter_strength = av_mallocz(s->mb_stride*sizeof(*s->filter_strength)); |
11921 | 245 s->intra4x4_pred_mode_base = av_mallocz(s->b4_stride*(4*s->mb_height+1)); |
246 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
|
247 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
|
248 s->segmentation_map = av_mallocz(s->mb_stride*s->mb_height); |
11921 | 249 |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
250 if (!s->macroblocks_base || !s->filter_strength || !s->intra4x4_pred_mode_base || |
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
251 !s->top_nnz || !s->top_border || !s->segmentation_map) |
12169 | 252 return AVERROR(ENOMEM); |
253 | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
254 s->macroblocks = s->macroblocks_base + 1; |
11921 | 255 s->intra4x4_pred_mode = s->intra4x4_pred_mode_base + 4 + s->b4_stride; |
256 | |
257 memset(s->intra4x4_pred_mode_base, DC_PRED, s->b4_stride); | |
258 for (i = 0; i < 4*s->mb_height; i++) | |
259 s->intra4x4_pred_mode[i*s->b4_stride-1] = DC_PRED; | |
260 | |
261 return 0; | |
262 } | |
263 | |
264 static void parse_segment_info(VP8Context *s) | |
265 { | |
266 VP56RangeCoder *c = &s->c; | |
267 int i; | |
268 | |
269 s->segmentation.update_map = vp8_rac_get(c); | |
270 | |
271 if (vp8_rac_get(c)) { // update segment feature data | |
272 s->segmentation.absolute_vals = vp8_rac_get(c); | |
273 | |
274 for (i = 0; i < 4; i++) | |
275 s->segmentation.base_quant[i] = vp8_rac_get_sint(c, 7); | |
276 | |
277 for (i = 0; i < 4; i++) | |
278 s->segmentation.filter_level[i] = vp8_rac_get_sint(c, 6); | |
279 } | |
280 if (s->segmentation.update_map) | |
281 for (i = 0; i < 3; i++) | |
282 s->prob->segmentid[i] = vp8_rac_get(c) ? vp8_rac_get_uint(c, 8) : 255; | |
283 } | |
284 | |
285 static void update_lf_deltas(VP8Context *s) | |
286 { | |
287 VP56RangeCoder *c = &s->c; | |
288 int i; | |
289 | |
290 for (i = 0; i < 4; i++) | |
291 s->lf_delta.ref[i] = vp8_rac_get_sint(c, 6); | |
292 | |
293 for (i = 0; i < 4; i++) | |
294 s->lf_delta.mode[i] = vp8_rac_get_sint(c, 6); | |
295 } | |
296 | |
297 static int setup_partitions(VP8Context *s, const uint8_t *buf, int buf_size) | |
298 { | |
299 const uint8_t *sizes = buf; | |
300 int i; | |
301 | |
302 s->num_coeff_partitions = 1 << vp8_rac_get_uint(&s->c, 2); | |
303 | |
304 buf += 3*(s->num_coeff_partitions-1); | |
305 buf_size -= 3*(s->num_coeff_partitions-1); | |
306 if (buf_size < 0) | |
307 return -1; | |
308 | |
309 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
|
310 int size = AV_RL24(sizes + 3*i); |
11921 | 311 if (buf_size - size < 0) |
312 return -1; | |
313 | |
314 vp56_init_range_decoder(&s->coeff_partition[i], buf, size); | |
315 buf += size; | |
316 buf_size -= size; | |
317 } | |
318 vp56_init_range_decoder(&s->coeff_partition[i], buf, buf_size); | |
319 | |
320 return 0; | |
321 } | |
322 | |
323 static void get_quants(VP8Context *s) | |
324 { | |
325 VP56RangeCoder *c = &s->c; | |
326 int i, base_qi; | |
327 | |
328 int yac_qi = vp8_rac_get_uint(c, 7); | |
329 int ydc_delta = vp8_rac_get_sint(c, 4); | |
330 int y2dc_delta = vp8_rac_get_sint(c, 4); | |
331 int y2ac_delta = vp8_rac_get_sint(c, 4); | |
332 int uvdc_delta = vp8_rac_get_sint(c, 4); | |
333 int uvac_delta = vp8_rac_get_sint(c, 4); | |
334 | |
335 for (i = 0; i < 4; i++) { | |
336 if (s->segmentation.enabled) { | |
337 base_qi = s->segmentation.base_quant[i]; | |
338 if (!s->segmentation.absolute_vals) | |
339 base_qi += yac_qi; | |
340 } else | |
341 base_qi = yac_qi; | |
342 | |
343 s->qmat[i].luma_qmul[0] = vp8_dc_qlookup[av_clip(base_qi + ydc_delta , 0, 127)]; | |
344 s->qmat[i].luma_qmul[1] = vp8_ac_qlookup[av_clip(base_qi , 0, 127)]; | |
345 s->qmat[i].luma_dc_qmul[0] = 2 * vp8_dc_qlookup[av_clip(base_qi + y2dc_delta, 0, 127)]; | |
12289
bc679bd1c66d
perform the clipping on luma_dc_qmul[1] and chroma_qmul[0] earlier
skal
parents:
12288
diff
changeset
|
346 s->qmat[i].luma_dc_qmul[1] = 155 * vp8_ac_qlookup[av_clip(base_qi + y2ac_delta, 2, 127)] / 100; |
bc679bd1c66d
perform the clipping on luma_dc_qmul[1] and chroma_qmul[0] earlier
skal
parents:
12288
diff
changeset
|
347 s->qmat[i].chroma_qmul[0] = vp8_dc_qlookup[av_clip(base_qi + uvdc_delta, 0, 117)]; |
11921 | 348 s->qmat[i].chroma_qmul[1] = vp8_ac_qlookup[av_clip(base_qi + uvac_delta, 0, 127)]; |
349 } | |
350 } | |
351 | |
352 /** | |
353 * Determine which buffers golden and altref should be updated with after this frame. | |
354 * The spec isn't clear here, so I'm going by my understanding of what libvpx does | |
355 * | |
356 * Intra frames update all 3 references | |
357 * Inter frames update VP56_FRAME_PREVIOUS if the update_last flag is set | |
358 * If the update (golden|altref) flag is set, it's updated with the current frame | |
359 * if update_last is set, and VP56_FRAME_PREVIOUS otherwise. | |
360 * If the flag is not set, the number read means: | |
361 * 0: no update | |
362 * 1: VP56_FRAME_PREVIOUS | |
363 * 2: update golden with altref, or update altref with golden | |
364 */ | |
365 static VP56Frame ref_to_update(VP8Context *s, int update, VP56Frame ref) | |
366 { | |
367 VP56RangeCoder *c = &s->c; | |
368 | |
369 if (update) | |
370 return VP56_FRAME_CURRENT; | |
371 | |
372 switch (vp8_rac_get_uint(c, 2)) { | |
373 case 1: | |
374 return VP56_FRAME_PREVIOUS; | |
375 case 2: | |
376 return (ref == VP56_FRAME_GOLDEN) ? VP56_FRAME_GOLDEN2 : VP56_FRAME_GOLDEN; | |
377 } | |
378 return VP56_FRAME_NONE; | |
379 } | |
380 | |
381 static void update_refs(VP8Context *s) | |
382 { | |
383 VP56RangeCoder *c = &s->c; | |
384 | |
385 int update_golden = vp8_rac_get(c); | |
386 int update_altref = vp8_rac_get(c); | |
387 | |
388 s->update_golden = ref_to_update(s, update_golden, VP56_FRAME_GOLDEN); | |
389 s->update_altref = ref_to_update(s, update_altref, VP56_FRAME_GOLDEN2); | |
390 } | |
391 | |
392 static int decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size) | |
393 { | |
394 VP56RangeCoder *c = &s->c; | |
395 int header_size, hscale, vscale, i, j, k, l, ret; | |
396 int width = s->avctx->width; | |
397 int height = s->avctx->height; | |
398 | |
399 s->keyframe = !(buf[0] & 1); | |
400 s->profile = (buf[0]>>1) & 7; | |
401 s->invisible = !(buf[0] & 0x10); | |
12247
50a96623366b
VP8: use AV_RL24 instead of defining a new RL24.
darkshikari
parents:
12246
diff
changeset
|
402 header_size = AV_RL24(buf) >> 5; |
11921 | 403 buf += 3; |
404 buf_size -= 3; | |
405 | |
11974 | 406 if (s->profile > 3) |
407 av_log(s->avctx, AV_LOG_WARNING, "Unknown profile %d\n", s->profile); | |
408 | |
409 if (!s->profile) | |
410 memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_epel_pixels_tab, sizeof(s->put_pixels_tab)); | |
411 else // profile 1-3 use bilinear, 4+ aren't defined so whatever | |
412 memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_bilinear_pixels_tab, sizeof(s->put_pixels_tab)); | |
11921 | 413 |
414 if (header_size > buf_size - 7*s->keyframe) { | |
415 av_log(s->avctx, AV_LOG_ERROR, "Header size larger than data provided\n"); | |
416 return AVERROR_INVALIDDATA; | |
417 } | |
418 | |
419 if (s->keyframe) { | |
12247
50a96623366b
VP8: use AV_RL24 instead of defining a new RL24.
darkshikari
parents:
12246
diff
changeset
|
420 if (AV_RL24(buf) != 0x2a019d) { |
50a96623366b
VP8: use AV_RL24 instead of defining a new RL24.
darkshikari
parents:
12246
diff
changeset
|
421 av_log(s->avctx, AV_LOG_ERROR, "Invalid start code 0x%x\n", AV_RL24(buf)); |
11921 | 422 return AVERROR_INVALIDDATA; |
423 } | |
424 width = AV_RL16(buf+3) & 0x3fff; | |
425 height = AV_RL16(buf+5) & 0x3fff; | |
426 hscale = buf[4] >> 6; | |
427 vscale = buf[6] >> 6; | |
428 buf += 7; | |
429 buf_size -= 7; | |
430 | |
11970
c7953ee47af4
vp8: warn and request sample if upscaling specified in header
mru
parents:
11950
diff
changeset
|
431 if (hscale || vscale) |
c7953ee47af4
vp8: warn and request sample if upscaling specified in header
mru
parents:
11950
diff
changeset
|
432 av_log_missing_feature(s->avctx, "Upscaling", 1); |
c7953ee47af4
vp8: warn and request sample if upscaling specified in header
mru
parents:
11950
diff
changeset
|
433 |
11921 | 434 s->update_golden = s->update_altref = VP56_FRAME_CURRENT; |
435 memcpy(s->prob->token , vp8_token_default_probs , sizeof(s->prob->token)); | |
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++) | |
12254
17c151e1280a
VP8: Use vp56_rac_get_prob_branchy when the bit is only used by an if()
conrad
parents:
12253
diff
changeset
|
495 if (vp56_rac_get_prob_branchy(c, vp8_token_update_probs[i][j][k][l])) |
11921 | 496 s->prob->token[i][j][k][l] = vp8_rac_get_uint(c, 8); |
497 | |
498 if ((s->mbskip_enabled = vp8_rac_get(c))) | |
12288
6bdbab1d9ed4
save some copies by moving some fields out of proba[2]
skal
parents:
12271
diff
changeset
|
499 s->mbskip_proba = vp8_rac_get_uint(c, 8); |
11921 | 500 |
501 if (!s->keyframe) { | |
12288
6bdbab1d9ed4
save some copies by moving some fields out of proba[2]
skal
parents:
12271
diff
changeset
|
502 s->intra_proba = vp8_rac_get_uint(c, 8); |
6bdbab1d9ed4
save some copies by moving some fields out of proba[2]
skal
parents:
12271
diff
changeset
|
503 s->last_proba = vp8_rac_get_uint(c, 8); |
6bdbab1d9ed4
save some copies by moving some fields out of proba[2]
skal
parents:
12271
diff
changeset
|
504 s->golden_proba = vp8_rac_get_uint(c, 8); |
11921 | 505 |
506 if (vp8_rac_get(c)) | |
507 for (i = 0; i < 4; i++) | |
508 s->prob->pred16x16[i] = vp8_rac_get_uint(c, 8); | |
509 if (vp8_rac_get(c)) | |
510 for (i = 0; i < 3; i++) | |
511 s->prob->pred8x8c[i] = vp8_rac_get_uint(c, 8); | |
512 | |
513 // 17.2 MV probability update | |
514 for (i = 0; i < 2; i++) | |
515 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
|
516 if (vp56_rac_get_prob_branchy(c, vp8_mv_update_prob[i][j])) |
11921 | 517 s->prob->mvc[i][j] = vp8_rac_get_nn(c); |
518 } | |
519 | |
520 return 0; | |
521 } | |
522 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
523 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
524 void clamp_mv(VP8Context *s, VP56mv *dst, const VP56mv *src, int mb_x, int mb_y) |
11921 | 525 { |
526 #define MARGIN (16 << 2) | |
527 dst->x = av_clip(src->x, -((mb_x << 6) + MARGIN), | |
528 ((s->mb_width - 1 - mb_x) << 6) + MARGIN); | |
529 dst->y = av_clip(src->y, -((mb_y << 6) + MARGIN), | |
530 ((s->mb_height - 1 - mb_y) << 6) + MARGIN); | |
531 } | |
532 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
533 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
534 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
|
535 VP56mv near[2], VP56mv *best, uint8_t cnt[4]) |
11921 | 536 { |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
537 VP8Macroblock *mb_edge[3] = { mb + 2 /* top */, |
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
538 mb - 1 /* left */, |
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
539 mb + 1 /* top-left */ }; |
11921 | 540 enum { EDGE_TOP, EDGE_LEFT, EDGE_TOPLEFT }; |
541 VP56mv near_mv[4] = {{ 0 }}; | |
542 enum { CNT_ZERO, CNT_NEAREST, CNT_NEAR, CNT_SPLITMV }; | |
12217 | 543 int idx = CNT_ZERO; |
11921 | 544 int best_idx = CNT_ZERO; |
12217 | 545 int cur_sign_bias = s->sign_bias[mb->ref_frame]; |
546 int *sign_bias = s->sign_bias; | |
11921 | 547 |
548 /* Process MB on top, left and top-left */ | |
12217 | 549 #define MV_EDGE_CHECK(n)\ |
550 {\ | |
551 VP8Macroblock *edge = mb_edge[n];\ | |
552 int edge_ref = edge->ref_frame;\ | |
553 if (edge_ref != VP56_FRAME_CURRENT) {\ | |
554 uint32_t mv = AV_RN32A(&edge->mv);\ | |
555 if (mv) {\ | |
556 if (cur_sign_bias != sign_bias[edge_ref]) {\ | |
557 /* SWAR negate of the values in mv. */\ | |
12242 | 558 mv = ~mv;\ |
559 mv = ((mv&0x7fff7fff) + 0x00010001) ^ (mv&0x80008000);\ | |
12217 | 560 }\ |
561 if (!n || mv != AV_RN32A(&near_mv[idx]))\ | |
562 AV_WN32A(&near_mv[++idx], mv);\ | |
563 cnt[idx] += 1 + (n != 2);\ | |
564 } else\ | |
565 cnt[CNT_ZERO] += 1 + (n != 2);\ | |
566 }\ | |
11921 | 567 } |
12217 | 568 MV_EDGE_CHECK(0) |
569 MV_EDGE_CHECK(1) | |
570 MV_EDGE_CHECK(2) | |
11921 | 571 |
12217 | 572 /* If we have three distinct MVs, merge first and last if they're the same */ |
573 if (cnt[CNT_SPLITMV] && AV_RN32A(&near_mv[1+EDGE_TOP]) == AV_RN32A(&near_mv[1+EDGE_TOPLEFT])) | |
11921 | 574 cnt[CNT_NEAREST] += 1; |
575 | |
576 cnt[CNT_SPLITMV] = ((mb_edge[EDGE_LEFT]->mode == VP8_MVMODE_SPLIT) + | |
577 (mb_edge[EDGE_TOP]->mode == VP8_MVMODE_SPLIT)) * 2 + | |
578 (mb_edge[EDGE_TOPLEFT]->mode == VP8_MVMODE_SPLIT); | |
579 | |
580 /* Swap near and nearest if necessary */ | |
581 if (cnt[CNT_NEAR] > cnt[CNT_NEAREST]) { | |
12217 | 582 FFSWAP(uint8_t, cnt[CNT_NEAREST], cnt[CNT_NEAR]); |
583 FFSWAP( VP56mv, near_mv[CNT_NEAREST], near_mv[CNT_NEAR]); | |
11921 | 584 } |
585 | |
586 /* Choose the best mv out of 0,0 and the nearest mv */ | |
587 if (cnt[CNT_NEAREST] >= cnt[CNT_ZERO]) | |
588 best_idx = CNT_NEAREST; | |
589 | |
12246 | 590 mb->mv = near_mv[best_idx]; |
11921 | 591 near[0] = near_mv[CNT_NEAREST]; |
592 near[1] = near_mv[CNT_NEAR]; | |
593 } | |
594 | |
595 /** | |
596 * Motion vector coding, 17.1. | |
597 */ | |
598 static int read_mv_component(VP56RangeCoder *c, const uint8_t *p) | |
599 { | |
12255 | 600 int bit, x = 0; |
11921 | 601 |
12254
17c151e1280a
VP8: Use vp56_rac_get_prob_branchy when the bit is only used by an if()
conrad
parents:
12253
diff
changeset
|
602 if (vp56_rac_get_prob_branchy(c, p[0])) { |
11921 | 603 int i; |
604 | |
605 for (i = 0; i < 3; i++) | |
606 x += vp56_rac_get_prob(c, p[9 + i]) << i; | |
607 for (i = 9; i > 3; i--) | |
608 x += vp56_rac_get_prob(c, p[9 + i]) << i; | |
609 if (!(x & 0xFFF0) || vp56_rac_get_prob(c, p[12])) | |
610 x += 8; | |
12255 | 611 } else { |
612 // small_mvtree | |
613 const uint8_t *ps = p+2; | |
614 bit = vp56_rac_get_prob(c, *ps); | |
615 ps += 1 + 3*bit; | |
616 x += 4*bit; | |
617 bit = vp56_rac_get_prob(c, *ps); | |
618 ps += 1 + bit; | |
619 x += 2*bit; | |
620 x += vp56_rac_get_prob(c, *ps); | |
621 } | |
11921 | 622 |
623 return (x && vp56_rac_get_prob(c, p[1])) ? -x : x; | |
624 } | |
625 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
626 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
627 const uint8_t *get_submv_prob(uint32_t left, uint32_t top) |
11921 | 628 { |
12219 | 629 if (left == top) |
630 return vp8_submv_prob[4-!!left]; | |
631 if (!top) | |
11921 | 632 return vp8_submv_prob[2]; |
12219 | 633 return vp8_submv_prob[1-!!left]; |
11921 | 634 } |
635 | |
636 /** | |
637 * 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
|
638 * @returns the number of motion vectors parsed (2, 4 or 16) |
11921 | 639 */ |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
640 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
641 int decode_splitmvs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb) |
11921 | 642 { |
643 int part_idx = mb->partitioning = | |
644 vp8_rac_get_tree(c, vp8_mbsplit_tree, vp8_mbsplit_prob); | |
645 int n, num = vp8_mbsplit_count[part_idx]; | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
646 VP8Macroblock *top_mb = &mb[2]; |
12219 | 647 VP8Macroblock *left_mb = &mb[-1]; |
648 const uint8_t *mbsplits_left = vp8_mbsplits[left_mb->partitioning], | |
649 *mbsplits_top = vp8_mbsplits[top_mb->partitioning], | |
650 *mbsplits_cur = vp8_mbsplits[part_idx], | |
11990
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
651 *firstidx = vp8_mbfirstidx[part_idx]; |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
652 VP56mv *top_mv = top_mb->bmv; |
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
653 VP56mv *left_mv = left_mb->bmv; |
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
654 VP56mv *cur_mv = mb->bmv; |
11921 | 655 |
656 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
|
657 int k = firstidx[n]; |
12219 | 658 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
|
659 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
|
660 |
12219 | 661 if (!(k & 3)) |
662 left = AV_RN32A(&left_mv[mbsplits_left[k + 3]]); | |
663 else | |
664 left = AV_RN32A(&cur_mv[mbsplits_cur[k - 1]]); | |
665 if (k <= 3) | |
666 above = AV_RN32A(&top_mv[mbsplits_top[k + 12]]); | |
667 else | |
668 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
|
669 |
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
670 submv_prob = get_submv_prob(left, above); |
11921 | 671 |
672 switch (vp8_rac_get_tree(c, vp8_submv_ref_tree, submv_prob)) { | |
673 case VP8_SUBMVMODE_NEW4X4: | |
12246 | 674 mb->bmv[n].y = mb->mv.y + read_mv_component(c, s->prob->mvc[0]); |
675 mb->bmv[n].x = mb->mv.x + read_mv_component(c, s->prob->mvc[1]); | |
11921 | 676 break; |
677 case VP8_SUBMVMODE_ZERO4X4: | |
12245
ca82c3ce90c1
VP8: use AV_ZERO32 instead of AV_WN32A where relevant
darkshikari
parents:
12244
diff
changeset
|
678 AV_ZERO32(&mb->bmv[n]); |
11921 | 679 break; |
680 case VP8_SUBMVMODE_LEFT4X4: | |
12219 | 681 AV_WN32A(&mb->bmv[n], left); |
11921 | 682 break; |
683 case VP8_SUBMVMODE_TOP4X4: | |
12219 | 684 AV_WN32A(&mb->bmv[n], above); |
11921 | 685 break; |
686 } | |
11990
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
687 } |
11921 | 688 |
11990
3c51d7ac41c9
Simplify MV parsing, removes laying out 2 or 4 (16x8/8x8/8x16) MVs over all
rbultje
parents:
11989
diff
changeset
|
689 return num; |
11921 | 690 } |
691 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
692 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
693 void decode_intra4x4_modes(VP56RangeCoder *c, uint8_t *intra4x4, |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
694 int stride, int keyframe) |
11921 | 695 { |
12221 | 696 int x, y, t, l, i; |
11921 | 697 |
12221 | 698 if (keyframe) { |
699 const uint8_t *ctx; | |
700 for (y = 0; y < 4; y++) { | |
701 for (x = 0; x < 4; x++) { | |
11921 | 702 t = intra4x4[x - stride]; |
703 l = intra4x4[x - 1]; | |
704 ctx = vp8_pred4x4_prob_intra[t][l]; | |
12221 | 705 intra4x4[x] = vp8_rac_get_tree(c, vp8_pred4x4_tree, ctx); |
11921 | 706 } |
12221 | 707 intra4x4 += stride; |
11921 | 708 } |
12221 | 709 } else { |
710 for (i = 0; i < 16; i++) | |
711 intra4x4[i] = vp8_rac_get_tree(c, vp8_pred4x4_tree, vp8_pred4x4_prob_inter); | |
11921 | 712 } |
713 } | |
714 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
715 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
716 void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
717 uint8_t *intra4x4, uint8_t *segment) |
11921 | 718 { |
719 VP56RangeCoder *c = &s->c; | |
720 | |
721 if (s->segmentation.update_map) | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
722 *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
|
723 s->segment = *segment; |
11921 | 724 |
12288
6bdbab1d9ed4
save some copies by moving some fields out of proba[2]
skal
parents:
12271
diff
changeset
|
725 mb->skip = s->mbskip_enabled ? vp56_rac_get_prob(c, s->mbskip_proba) : 0; |
11921 | 726 |
727 if (s->keyframe) { | |
728 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_intra, vp8_pred16x16_prob_intra); | |
729 | |
730 if (mb->mode == MODE_I4x4) { | |
731 decode_intra4x4_modes(c, intra4x4, s->b4_stride, 1); | |
732 } else | |
733 fill_rectangle(intra4x4, 4, 4, s->b4_stride, vp8_pred4x4_mode[mb->mode], 1); | |
734 | |
735 s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, vp8_pred8x8c_prob_intra); | |
736 mb->ref_frame = VP56_FRAME_CURRENT; | |
12288
6bdbab1d9ed4
save some copies by moving some fields out of proba[2]
skal
parents:
12271
diff
changeset
|
737 } else if (vp56_rac_get_prob_branchy(c, s->intra_proba)) { |
11921 | 738 VP56mv near[2], best; |
12217 | 739 uint8_t cnt[4] = { 0 }; |
11921 | 740 uint8_t p[4]; |
741 | |
742 // inter MB, 16.2 | |
12288
6bdbab1d9ed4
save some copies by moving some fields out of proba[2]
skal
parents:
12271
diff
changeset
|
743 if (vp56_rac_get_prob_branchy(c, s->last_proba)) |
6bdbab1d9ed4
save some copies by moving some fields out of proba[2]
skal
parents:
12271
diff
changeset
|
744 mb->ref_frame = vp56_rac_get_prob(c, s->golden_proba) ? |
11921 | 745 VP56_FRAME_GOLDEN2 /* altref */ : VP56_FRAME_GOLDEN; |
746 else | |
747 mb->ref_frame = VP56_FRAME_PREVIOUS; | |
12231 | 748 s->ref_count[mb->ref_frame-1]++; |
11921 | 749 |
750 // motion vectors, 16.3 | |
751 find_near_mvs(s, mb, mb_x, mb_y, near, &best, cnt); | |
12217 | 752 p[0] = vp8_mode_contexts[cnt[0]][0]; |
753 p[1] = vp8_mode_contexts[cnt[1]][1]; | |
754 p[2] = vp8_mode_contexts[cnt[2]][2]; | |
755 p[3] = vp8_mode_contexts[cnt[3]][3]; | |
11921 | 756 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_mvinter, p); |
757 switch (mb->mode) { | |
758 case VP8_MVMODE_SPLIT: | |
12246 | 759 clamp_mv(s, &mb->mv, &mb->mv, mb_x, mb_y); |
760 mb->mv = mb->bmv[decode_splitmvs(s, c, mb) - 1]; | |
11921 | 761 break; |
762 case VP8_MVMODE_ZERO: | |
12245
ca82c3ce90c1
VP8: use AV_ZERO32 instead of AV_WN32A where relevant
darkshikari
parents:
12244
diff
changeset
|
763 AV_ZERO32(&mb->mv); |
11921 | 764 break; |
765 case VP8_MVMODE_NEAREST: | |
766 clamp_mv(s, &mb->mv, &near[0], mb_x, mb_y); | |
767 break; | |
768 case VP8_MVMODE_NEAR: | |
769 clamp_mv(s, &mb->mv, &near[1], mb_x, mb_y); | |
770 break; | |
771 case VP8_MVMODE_NEW: | |
12246 | 772 clamp_mv(s, &mb->mv, &mb->mv, mb_x, mb_y); |
773 mb->mv.y += + read_mv_component(c, s->prob->mvc[0]); | |
774 mb->mv.x += + read_mv_component(c, s->prob->mvc[1]); | |
11921 | 775 break; |
776 } | |
777 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
|
778 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
|
779 mb->bmv[0] = mb->mv; |
11921 | 780 } |
781 } else { | |
782 // intra MB, 16.1 | |
783 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_inter, s->prob->pred16x16); | |
784 | |
12220
0f635b1f7861
Avoid useless fill_rectangle in P-frames in VP8
darkshikari
parents:
12219
diff
changeset
|
785 if (mb->mode == MODE_I4x4) |
12221 | 786 decode_intra4x4_modes(c, intra4x4, 4, 0); |
11921 | 787 |
788 s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, s->prob->pred8x8c); | |
789 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
|
790 mb->partitioning = VP8_SPLITMVMODE_NONE; |
12245
ca82c3ce90c1
VP8: use AV_ZERO32 instead of AV_WN32A where relevant
darkshikari
parents:
12244
diff
changeset
|
791 AV_ZERO32(&mb->bmv[0]); |
11921 | 792 } |
793 } | |
794 | |
795 /** | |
12115 | 796 * @param c arithmetic bitstream reader context |
797 * @param block destination for block coefficients | |
798 * @param probs probabilities to use when reading trees from the bitstream | |
11921 | 799 * @param i initial coeff index, 0 unless a separate DC block is coded |
800 * @param zero_nhood the initial prediction context for number of surrounding | |
801 * 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
|
802 * @param qmul array holding the dc/ac dequant factor at position 0/1 |
11921 | 803 * @return 0 if no coeffs were decoded |
804 * otherwise, the index of the last coeff decoded plus one | |
805 */ | |
806 static int decode_block_coeffs(VP56RangeCoder *c, DCTELEM block[16], | |
807 uint8_t probs[8][3][NUM_DCT_TOKENS-1], | |
808 int i, int zero_nhood, int16_t qmul[2]) | |
809 { | |
12253
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
810 uint8_t *token_prob; |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
811 int nonzero = 0; |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
812 int coeff; |
11921 | 813 |
12253
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
814 do { |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
815 token_prob = probs[vp8_coeff_band[i]][zero_nhood]; |
11921 | 816 |
12253
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
817 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
|
818 return nonzero; |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
819 |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
820 skip_eob: |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
821 if (!vp56_rac_get_prob_branchy(c, token_prob[1])) { // DCT_0 |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
822 zero_nhood = 0; |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
823 token_prob = probs[vp8_coeff_band[++i]][0]; |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
824 if (i < 16) |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
825 goto skip_eob; |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
826 return nonzero; // invalid input; blocks should end with EOB |
11921 | 827 } |
828 | |
12253
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
829 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
|
830 coeff = 1; |
11921 | 831 zero_nhood = 1; |
12253
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
832 } else { |
11921 | 833 zero_nhood = 2; |
834 | |
12253
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
835 if (!vp56_rac_get_prob_branchy(c, token_prob[3])) { // DCT 2,3,4 |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
836 coeff = vp56_rac_get_prob(c, token_prob[4]); |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
837 if (coeff) |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
838 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
|
839 coeff += 2; |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
840 } else { |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
841 // DCT_CAT* |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
842 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
|
843 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
|
844 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
|
845 } else { // DCT_CAT2 |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
846 coeff = 7; |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
847 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
|
848 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
|
849 } |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
850 } else { // DCT_CAT3 and up |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
851 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
|
852 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
|
853 int cat = (a<<1) + b; |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
854 coeff = 3 + (8<<cat); |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
855 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
|
856 } |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
857 } |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
858 } |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
859 |
11921 | 860 // 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
|
861 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
|
862 nonzero = ++i; |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
863 } while (i < 16); |
112b3a0db187
Decode DCT tokens by branching to a different code path for each branch
conrad
parents:
12249
diff
changeset
|
864 |
11921 | 865 return nonzero; |
866 } | |
867 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
868 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
869 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
|
870 uint8_t t_nnz[9], uint8_t l_nnz[9]) |
11921 | 871 { |
872 LOCAL_ALIGNED_16(DCTELEM, dc,[16]); | |
873 int i, x, y, luma_start = 0, luma_ctx = 3; | |
874 int nnz_pred, nnz, nnz_total = 0; | |
12224
5b7d690b761b
VP8: Don't store segment in macroblock struct anymore.
darkshikari
parents:
12223
diff
changeset
|
875 int segment = s->segment; |
11921 | 876 |
877 if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) { | |
878 AV_ZERO128(dc); | |
879 AV_ZERO128(dc+8); | |
880 nnz_pred = t_nnz[8] + l_nnz[8]; | |
881 | |
882 // decode DC values and do hadamard | |
883 nnz = decode_block_coeffs(c, dc, s->prob->token[1], 0, nnz_pred, | |
884 s->qmat[segment].luma_dc_qmul); | |
885 l_nnz[8] = t_nnz[8] = !!nnz; | |
886 nnz_total += nnz; | |
887 s->vp8dsp.vp8_luma_dc_wht(s->block, dc); | |
888 luma_start = 1; | |
889 luma_ctx = 0; | |
890 } | |
891 | |
892 // luma blocks | |
893 for (y = 0; y < 4; y++) | |
894 for (x = 0; x < 4; x++) { | |
895 nnz_pred = l_nnz[y] + t_nnz[x]; | |
896 nnz = decode_block_coeffs(c, s->block[y][x], s->prob->token[luma_ctx], luma_start, | |
897 nnz_pred, s->qmat[segment].luma_qmul); | |
898 // nnz+luma_start may be one more than the actual last index, but we don't care | |
899 s->non_zero_count_cache[y][x] = nnz + luma_start; | |
900 t_nnz[x] = l_nnz[y] = !!nnz; | |
901 nnz_total += nnz; | |
902 } | |
903 | |
904 // chroma blocks | |
905 // TODO: what to do about dimensions? 2nd dim for luma is x, | |
906 // but for chroma it's (y<<1)|x | |
907 for (i = 4; i < 6; i++) | |
908 for (y = 0; y < 2; y++) | |
909 for (x = 0; x < 2; x++) { | |
910 nnz_pred = l_nnz[i+2*y] + t_nnz[i+2*x]; | |
911 nnz = decode_block_coeffs(c, s->block[i][(y<<1)+x], s->prob->token[2], 0, | |
912 nnz_pred, s->qmat[segment].chroma_qmul); | |
913 s->non_zero_count_cache[i][(y<<1)+x] = nnz; | |
914 t_nnz[i+2*x] = l_nnz[i+2*y] = !!nnz; | |
915 nnz_total += nnz; | |
916 } | |
917 | |
918 // if there were no coded coeffs despite the macroblock not being marked skip, | |
919 // we MUST not do the inner loop filter and should not do IDCT | |
920 // Since skip isn't used for bitstream prediction, just manually set it. | |
921 if (!nnz_total) | |
922 mb->skip = 1; | |
923 } | |
924 | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
925 static av_always_inline |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
926 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
|
927 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
|
928 { |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
929 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
|
930 if (!simple) { |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
931 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
|
932 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
|
933 } |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
934 } |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
935 |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
936 static av_always_inline |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
937 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
|
938 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
|
939 int simple, int xchg) |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
940 { |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
941 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
|
942 src_y -= linesize; |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
943 src_cb -= uvlinesize; |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
944 src_cr -= uvlinesize; |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
945 |
12202 | 946 #define XCHG(a,b,xchg) do { \ |
947 if (xchg) AV_SWAP64(b,a); \ | |
948 else AV_COPY64(b,a); \ | |
949 } while (0) | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
950 |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
951 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
|
952 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
|
953 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
|
954 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
|
955 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
|
956 |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
957 // 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
|
958 // 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
|
959 if (!simple || !mb_y) { |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
960 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
|
961 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
|
962 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
|
963 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
|
964 } |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
965 } |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
966 |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
967 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
968 int check_intra_pred_mode(int mode, int mb_x, int mb_y) |
11921 | 969 { |
970 if (mode == DC_PRED8x8) { | |
12243
788445bf10c0
VP8: shave a few clocks off check_intra_pred_mode
darkshikari
parents:
12242
diff
changeset
|
971 if (!mb_x) { |
788445bf10c0
VP8: shave a few clocks off check_intra_pred_mode
darkshikari
parents:
12242
diff
changeset
|
972 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
|
973 } else if (!mb_y) { |
12244 | 974 mode = LEFT_DC_PRED8x8; |
12243
788445bf10c0
VP8: shave a few clocks off check_intra_pred_mode
darkshikari
parents:
12242
diff
changeset
|
975 } |
11921 | 976 } |
977 return mode; | |
978 } | |
979 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
980 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
981 void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb, |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
982 uint8_t *intra4x4, int mb_x, int mb_y) |
11921 | 983 { |
984 int x, y, mode, nnz, tr; | |
985 | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
986 // 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
|
987 // 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
|
988 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
|
989 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
|
990 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
|
991 s->filter.simple, 1); |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
992 |
11921 | 993 if (mb->mode < MODE_I4x4) { |
994 mode = check_intra_pred_mode(mb->mode, mb_x, mb_y); | |
995 s->hpc.pred16x16[mode](dst[0], s->linesize); | |
996 } else { | |
997 uint8_t *ptr = dst[0]; | |
12221 | 998 int stride = s->keyframe ? s->b4_stride : 4; |
11921 | 999 |
1000 // all blocks on the right edge of the macroblock use bottom edge | |
1001 // the top macroblock for their topright edge | |
1002 uint8_t *tr_right = ptr - s->linesize + 16; | |
1003 | |
1004 // if we're on the right edge of the frame, said edge is extended | |
1005 // from the top macroblock | |
1006 if (mb_x == s->mb_width-1) { | |
1007 tr = tr_right[-1]*0x01010101; | |
1008 tr_right = (uint8_t *)&tr; | |
1009 } | |
1010 | |
12234
bba849c2a113
VP8: avoid a memset for non-i4x4 blocks with no coefficients
darkshikari
parents:
12233
diff
changeset
|
1011 if (mb->skip) |
bba849c2a113
VP8: avoid a memset for non-i4x4 blocks with no coefficients
darkshikari
parents:
12233
diff
changeset
|
1012 AV_ZERO128(s->non_zero_count_cache); |
bba849c2a113
VP8: avoid a memset for non-i4x4 blocks with no coefficients
darkshikari
parents:
12233
diff
changeset
|
1013 |
11921 | 1014 for (y = 0; y < 4; y++) { |
1015 uint8_t *topright = ptr + 4 - s->linesize; | |
1016 for (x = 0; x < 4; x++) { | |
1017 if (x == 3) | |
1018 topright = tr_right; | |
1019 | |
12221 | 1020 s->hpc.pred4x4[intra4x4[x]](ptr+4*x, topright, s->linesize); |
11921 | 1021 |
1022 nnz = s->non_zero_count_cache[y][x]; | |
1023 if (nnz) { | |
1024 if (nnz == 1) | |
1025 s->vp8dsp.vp8_idct_dc_add(ptr+4*x, s->block[y][x], s->linesize); | |
1026 else | |
1027 s->vp8dsp.vp8_idct_add(ptr+4*x, s->block[y][x], s->linesize); | |
1028 } | |
1029 topright += 4; | |
1030 } | |
1031 | |
1032 ptr += 4*s->linesize; | |
12221 | 1033 intra4x4 += stride; |
11921 | 1034 } |
1035 } | |
1036 | |
1037 mode = check_intra_pred_mode(s->chroma_pred_mode, mb_x, mb_y); | |
1038 s->hpc.pred8x8[mode](dst[1], s->uvlinesize); | |
1039 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
|
1040 |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
1041 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
|
1042 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
|
1043 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
|
1044 s->filter.simple, 0); |
11921 | 1045 } |
1046 | |
1047 /** | |
1048 * Generic MC function. | |
1049 * | |
1050 * @param s VP8 decoding context | |
1051 * @param luma 1 for luma (Y) planes, 0 for chroma (Cb/Cr) planes | |
1052 * @param dst target buffer for block data at block position | |
1053 * @param src reference picture buffer at origin (0, 0) | |
1054 * @param mv motion vector (relative to block position) to get pixel data from | |
1055 * @param x_off horizontal position of block from origin (0, 0) | |
1056 * @param y_off vertical position of block from origin (0, 0) | |
1057 * @param block_w width of block (16, 8 or 4) | |
1058 * @param block_h height of block (always same as block_w) | |
1059 * @param width width of src/dst plane data | |
1060 * @param height height of src/dst plane data | |
1061 * @param linesize size of a single line of plane data, including padding | |
12115 | 1062 * @param mc_func motion compensation function pointers (bilinear or sixtap MC) |
11921 | 1063 */ |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1064 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1065 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
|
1066 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
|
1067 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
|
1068 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
|
1069 vp8_mc_func mc_func[3][3]) |
11921 | 1070 { |
12218 | 1071 if (AV_RN32A(mv)) { |
1072 static const uint8_t idx[8] = { 0, 1, 2, 1, 2, 1, 2, 1 }; | |
1073 int mx = (mv->x << luma)&7, mx_idx = idx[mx]; | |
1074 int my = (mv->y << luma)&7, my_idx = idx[my]; | |
11921 | 1075 |
12218 | 1076 x_off += mv->x >> (3 - luma); |
1077 y_off += mv->y >> (3 - luma); | |
11921 | 1078 |
12218 | 1079 // edge emulation |
1080 src += y_off * linesize + x_off; | |
1081 if (x_off < 2 || x_off >= width - block_w - 3 || | |
1082 y_off < 2 || y_off >= height - block_h - 3) { | |
1083 ff_emulated_edge_mc(s->edge_emu_buffer, src - 2 * linesize - 2, linesize, | |
1084 block_w + 5, block_h + 5, | |
1085 x_off - 2, y_off - 2, width, height); | |
1086 src = s->edge_emu_buffer + 2 + linesize * 2; | |
1087 } | |
1088 mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my); | |
1089 } else | |
1090 mc_func[0][0](dst, linesize, src + y_off * linesize + x_off, linesize, block_h, 0, 0); | |
11921 | 1091 } |
1092 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1093 static av_always_inline |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1094 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
|
1095 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
|
1096 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
|
1097 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
|
1098 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
|
1099 { |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1100 VP56mv uvmv = *mv; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1101 |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1102 /* Y */ |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1103 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
|
1104 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
|
1105 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
|
1106 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
|
1107 |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1108 /* U/V */ |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1109 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
|
1110 uvmv.x &= ~7; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1111 uvmv.y &= ~7; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1112 } |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1113 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
|
1114 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
|
1115 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
|
1116 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
|
1117 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
|
1118 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
|
1119 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
|
1120 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
|
1121 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
|
1122 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
|
1123 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
|
1124 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
|
1125 } |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1126 |
12215 | 1127 /* Fetch pixels for estimated mv 4 macroblocks ahead. |
1128 * 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
|
1129 static av_always_inline void prefetch_motion(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, int mb_xy, int ref) |
12215 | 1130 { |
12237 | 1131 /* Don't prefetch refs that haven't been used very often this frame. */ |
1132 if (s->ref_count[ref-1] > (mb_xy >> 5)) { | |
12231 | 1133 int x_off = mb_x << 4, y_off = mb_y << 4; |
1134 int mx = mb->mv.x + x_off + 8; | |
1135 int my = mb->mv.y + y_off; | |
1136 uint8_t **src= s->framep[ref]->data; | |
1137 int off= mx + (my + (mb_x&3)*4)*s->linesize + 64; | |
1138 s->dsp.prefetch(src[0]+off, s->linesize, 4); | |
1139 off= (mx>>1) + ((my>>1) + (mb_x&7))*s->uvlinesize + 64; | |
1140 s->dsp.prefetch(src[1]+off, src[2]-src[1], 2); | |
1141 } | |
12215 | 1142 } |
1143 | |
11921 | 1144 /** |
1145 * Apply motion vectors to prediction buffer, chapter 18. | |
1146 */ | |
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 |
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1148 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
|
1149 int mb_x, int mb_y) |
11921 | 1150 { |
1151 int x_off = mb_x << 4, y_off = mb_y << 4; | |
1152 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
|
1153 AVFrame *ref = s->framep[mb->ref_frame]; |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1154 VP56mv *bmv = mb->bmv; |
11921 | 1155 |
1156 if (mb->mode < VP8_MVMODE_SPLIT) { | |
12228
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1157 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
|
1158 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
|
1159 } 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
|
1160 case VP8_SPLITMVMODE_4x4: { |
11921 | 1161 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
|
1162 VP56mv uvmv; |
11921 | 1163 |
1164 /* Y */ | |
1165 for (y = 0; y < 4; y++) { | |
1166 for (x = 0; x < 4; x++) { | |
1167 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
|
1168 ref->data[0], &bmv[4*y + x], |
11921 | 1169 4*x + x_off, 4*y + y_off, 4, 4, |
1170 width, height, s->linesize, | |
11974 | 1171 s->put_pixels_tab[2]); |
11921 | 1172 } |
1173 } | |
1174 | |
1175 /* U/V */ | |
1176 x_off >>= 1; y_off >>= 1; width >>= 1; height >>= 1; | |
1177 for (y = 0; y < 2; y++) { | |
1178 for (x = 0; x < 2; x++) { | |
1179 uvmv.x = mb->bmv[ 2*y * 4 + 2*x ].x + | |
1180 mb->bmv[ 2*y * 4 + 2*x+1].x + | |
1181 mb->bmv[(2*y+1) * 4 + 2*x ].x + | |
1182 mb->bmv[(2*y+1) * 4 + 2*x+1].x; | |
1183 uvmv.y = mb->bmv[ 2*y * 4 + 2*x ].y + | |
1184 mb->bmv[ 2*y * 4 + 2*x+1].y + | |
1185 mb->bmv[(2*y+1) * 4 + 2*x ].y + | |
1186 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
|
1187 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
|
1188 uvmv.y = (uvmv.y + 2 + (uvmv.y >> (INT_BIT-1))) >> 2; |
11921 | 1189 if (s->profile == 3) { |
1190 uvmv.x &= ~7; | |
1191 uvmv.y &= ~7; | |
1192 } | |
1193 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
|
1194 ref->data[1], &uvmv, |
11921 | 1195 4*x + x_off, 4*y + y_off, 4, 4, |
1196 width, height, s->uvlinesize, | |
11974 | 1197 s->put_pixels_tab[2]); |
11921 | 1198 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
|
1199 ref->data[2], &uvmv, |
11921 | 1200 4*x + x_off, 4*y + y_off, 4, 4, |
1201 width, height, s->uvlinesize, | |
11974 | 1202 s->put_pixels_tab[2]); |
11921 | 1203 } |
1204 } | |
11989
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1205 break; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1206 } |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1207 case VP8_SPLITMVMODE_16x8: |
12228
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1208 vp8_mc_part(s, dst, ref, x_off, y_off, |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1209 0, 0, 16, 8, width, height, &bmv[0]); |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1210 vp8_mc_part(s, dst, ref, x_off, y_off, |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1211 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
|
1212 break; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1213 case VP8_SPLITMVMODE_8x16: |
12228
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1214 vp8_mc_part(s, dst, ref, x_off, y_off, |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1215 0, 0, 8, 16, width, height, &bmv[0]); |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1216 vp8_mc_part(s, dst, ref, x_off, y_off, |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1217 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
|
1218 break; |
176c5deb6756
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
rbultje
parents:
11974
diff
changeset
|
1219 case VP8_SPLITMVMODE_8x8: |
12228
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1220 vp8_mc_part(s, dst, ref, x_off, y_off, |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1221 0, 0, 8, 8, width, height, &bmv[0]); |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1222 vp8_mc_part(s, dst, ref, x_off, y_off, |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1223 8, 0, 8, 8, width, height, &bmv[1]); |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1224 vp8_mc_part(s, dst, ref, x_off, y_off, |
9c63566f623f
Eliminate some repeated dereferences in VP8 inter_predict
darkshikari
parents:
12225
diff
changeset
|
1225 0, 8, 8, 8, width, height, &bmv[2]); |
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 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
|
1228 break; |
11921 | 1229 } |
1230 } | |
1231 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1232 static av_always_inline void idct_mb(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb) |
11921 | 1233 { |
12240 | 1234 int x, y, ch; |
11921 | 1235 |
12238 | 1236 if (mb->mode != MODE_I4x4) { |
1237 uint8_t *y_dst = dst[0]; | |
11921 | 1238 for (y = 0; y < 4; y++) { |
12240 | 1239 uint32_t nnz4 = AV_RN32A(s->non_zero_count_cache[y]); |
1240 if (nnz4) { | |
1241 if (nnz4&~0x01010101) { | |
12238 | 1242 for (x = 0; x < 4; x++) { |
12240 | 1243 int nnz = s->non_zero_count_cache[y][x]; |
12238 | 1244 if (nnz) { |
1245 if (nnz == 1) | |
1246 s->vp8dsp.vp8_idct_dc_add(y_dst+4*x, s->block[y][x], s->linesize); | |
1247 else | |
1248 s->vp8dsp.vp8_idct_add(y_dst+4*x, s->block[y][x], s->linesize); | |
1249 } | |
1250 } | |
1251 } else { | |
12241
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1252 s->vp8dsp.vp8_idct_dc_add4y(y_dst, s->block[y], s->linesize); |
11921 | 1253 } |
1254 } | |
1255 y_dst += 4*s->linesize; | |
1256 } | |
12238 | 1257 } |
11921 | 1258 |
12238 | 1259 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
|
1260 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
|
1261 if (nnz4) { |
12238 | 1262 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
|
1263 if (nnz4&~0x01010101) { |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1264 for (y = 0; y < 2; y++) { |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1265 for (x = 0; x < 2; x++) { |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1266 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
|
1267 if (nnz) { |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1268 if (nnz == 1) |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1269 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
|
1270 else |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1271 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
|
1272 } |
12238 | 1273 } |
12241
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1274 ch_dst += 4*s->uvlinesize; |
12238 | 1275 } |
12241
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1276 } else { |
c7f6ddcc5c01
VP8: optimize DC-only chroma case in the same way as luma.
darkshikari
parents:
12240
diff
changeset
|
1277 s->vp8dsp.vp8_idct_dc_add4uv(ch_dst, s->block[4+ch], s->uvlinesize); |
11921 | 1278 } |
1279 } | |
1280 } | |
1281 } | |
1282 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1283 static av_always_inline void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb, VP8FilterStrength *f ) |
11921 | 1284 { |
1285 int interior_limit, filter_level; | |
1286 | |
1287 if (s->segmentation.enabled) { | |
12224
5b7d690b761b
VP8: Don't store segment in macroblock struct anymore.
darkshikari
parents:
12223
diff
changeset
|
1288 filter_level = s->segmentation.filter_level[s->segment]; |
11921 | 1289 if (!s->segmentation.absolute_vals) |
1290 filter_level += s->filter.level; | |
1291 } else | |
1292 filter_level = s->filter.level; | |
1293 | |
1294 if (s->lf_delta.enabled) { | |
1295 filter_level += s->lf_delta.ref[mb->ref_frame]; | |
1296 | |
1297 if (mb->ref_frame == VP56_FRAME_CURRENT) { | |
1298 if (mb->mode == MODE_I4x4) | |
1299 filter_level += s->lf_delta.mode[0]; | |
1300 } else { | |
1301 if (mb->mode == VP8_MVMODE_ZERO) | |
1302 filter_level += s->lf_delta.mode[1]; | |
1303 else if (mb->mode == VP8_MVMODE_SPLIT) | |
1304 filter_level += s->lf_delta.mode[3]; | |
1305 else | |
1306 filter_level += s->lf_delta.mode[2]; | |
1307 } | |
1308 } | |
1309 filter_level = av_clip(filter_level, 0, 63); | |
1310 | |
1311 interior_limit = filter_level; | |
1312 if (s->filter.sharpness) { | |
1313 interior_limit >>= s->filter.sharpness > 4 ? 2 : 1; | |
1314 interior_limit = FFMIN(interior_limit, 9 - s->filter.sharpness); | |
1315 } | |
1316 interior_limit = FFMAX(interior_limit, 1); | |
1317 | |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1318 f->filter_level = filter_level; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1319 f->inner_limit = interior_limit; |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1320 f->inner_filter = !mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT; |
11921 | 1321 } |
1322 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1323 static av_always_inline void filter_mb(VP8Context *s, uint8_t *dst[3], VP8FilterStrength *f, int mb_x, int mb_y) |
11921 | 1324 { |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1325 int mbedge_lim, bedge_lim, hev_thresh; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1326 int filter_level = f->filter_level; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1327 int inner_limit = f->inner_limit; |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1328 int inner_filter = f->inner_filter; |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1329 int linesize = s->linesize; |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1330 int uvlinesize = s->uvlinesize; |
11921 | 1331 |
1332 if (!filter_level) | |
1333 return; | |
1334 | |
12081
812e23197d64
VP8: Move calculation of outer filter limit out of dsp functions for normal
conrad
parents:
12062
diff
changeset
|
1335 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
|
1336 bedge_lim = 2* filter_level + inner_limit; |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1337 hev_thresh = filter_level >= 15; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1338 |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1339 if (s->keyframe) { |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1340 if (filter_level >= 40) |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1341 hev_thresh = 2; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1342 } else { |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1343 if (filter_level >= 40) |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1344 hev_thresh = 3; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1345 else if (filter_level >= 20) |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1346 hev_thresh = 2; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1347 } |
12081
812e23197d64
VP8: Move calculation of outer filter limit out of dsp functions for normal
conrad
parents:
12062
diff
changeset
|
1348 |
11921 | 1349 if (mb_x) { |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1350 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
|
1351 mbedge_lim, inner_limit, hev_thresh); |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1352 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
|
1353 mbedge_lim, inner_limit, hev_thresh); |
11921 | 1354 } |
1355 | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1356 if (inner_filter) { |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1357 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
|
1358 inner_limit, hev_thresh); |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1359 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
|
1360 inner_limit, hev_thresh); |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1361 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
|
1362 inner_limit, hev_thresh); |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1363 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
|
1364 uvlinesize, bedge_lim, |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1365 inner_limit, hev_thresh); |
11921 | 1366 } |
1367 | |
1368 if (mb_y) { | |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1369 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
|
1370 mbedge_lim, inner_limit, hev_thresh); |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1371 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
|
1372 mbedge_lim, inner_limit, hev_thresh); |
11921 | 1373 } |
1374 | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1375 if (inner_filter) { |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1376 s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 4*linesize, |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1377 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_v_loop_filter16y_inner(dst[0]+ 8*linesize, |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1380 linesize, bedge_lim, |
12194
80b142c2e9f7
Change function prototypes for width=8 inner and mbedge loopfilter functions
rbultje
parents:
12170
diff
changeset
|
1381 inner_limit, hev_thresh); |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1382 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
|
1383 linesize, bedge_lim, |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1384 inner_limit, hev_thresh); |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1385 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
|
1386 dst[2] + 4 * uvlinesize, |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1387 uvlinesize, bedge_lim, |
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1388 inner_limit, hev_thresh); |
11921 | 1389 } |
1390 } | |
1391 | |
12248
121272849def
VP8: always_inline some things to force gcc to do the right thing
darkshikari
parents:
12247
diff
changeset
|
1392 static av_always_inline void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8FilterStrength *f, int mb_x, int mb_y) |
11921 | 1393 { |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1394 int mbedge_lim, bedge_lim; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1395 int filter_level = f->filter_level; |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1396 int inner_limit = f->inner_limit; |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1397 int inner_filter = f->inner_filter; |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1398 int linesize = s->linesize; |
11921 | 1399 |
1400 if (!filter_level) | |
1401 return; | |
1402 | |
1403 mbedge_lim = 2*(filter_level+2) + inner_limit; | |
1404 bedge_lim = 2* filter_level + inner_limit; | |
1405 | |
1406 if (mb_x) | |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1407 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
|
1408 if (inner_filter) { |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1409 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
|
1410 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
|
1411 s->vp8dsp.vp8_h_loop_filter_simple(dst+12, linesize, bedge_lim); |
11921 | 1412 } |
1413 | |
1414 if (mb_y) | |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1415 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
|
1416 if (inner_filter) { |
12233
10b02cbc3cc2
Get rid of more unnecessary dereferences in VP8 deblocking
darkshikari
parents:
12232
diff
changeset
|
1417 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
|
1418 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
|
1419 s->vp8dsp.vp8_v_loop_filter_simple(dst+12*linesize, linesize, bedge_lim); |
11921 | 1420 } |
1421 } | |
1422 | |
1423 static void filter_mb_row(VP8Context *s, int mb_y) | |
1424 { | |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1425 VP8FilterStrength *f = s->filter_strength; |
11921 | 1426 uint8_t *dst[3] = { |
1427 s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize, | |
1428 s->framep[VP56_FRAME_CURRENT]->data[1] + 8*mb_y*s->uvlinesize, | |
1429 s->framep[VP56_FRAME_CURRENT]->data[2] + 8*mb_y*s->uvlinesize | |
1430 }; | |
1431 int mb_x; | |
1432 | |
1433 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
|
1434 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
|
1435 filter_mb(s, dst, f++, mb_x, mb_y); |
11921 | 1436 dst[0] += 16; |
1437 dst[1] += 8; | |
1438 dst[2] += 8; | |
1439 } | |
1440 } | |
1441 | |
1442 static void filter_mb_row_simple(VP8Context *s, int mb_y) | |
1443 { | |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1444 VP8FilterStrength *f = s->filter_strength; |
11921 | 1445 uint8_t *dst = s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize; |
1446 int mb_x; | |
1447 | |
1448 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
|
1449 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
|
1450 filter_mb_simple(s, dst, f++, mb_x, mb_y); |
11921 | 1451 dst += 16; |
1452 } | |
1453 } | |
1454 | |
1455 static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size, | |
1456 AVPacket *avpkt) | |
1457 { | |
1458 VP8Context *s = avctx->priv_data; | |
1459 int ret, mb_x, mb_y, i, y, referenced; | |
1460 enum AVDiscard skip_thresh; | |
12270
161c205dcdd2
Fix r24445: Instead of needlessly initialising a variable, silence the warning.
cehoyos
parents:
12255
diff
changeset
|
1461 AVFrame *av_uninit(curframe); |
11921 | 1462 |
1463 if ((ret = decode_frame_header(s, avpkt->data, avpkt->size)) < 0) | |
1464 return ret; | |
1465 | |
1466 referenced = s->update_last || s->update_golden == VP56_FRAME_CURRENT | |
1467 || s->update_altref == VP56_FRAME_CURRENT; | |
1468 | |
1469 skip_thresh = !referenced ? AVDISCARD_NONREF : | |
1470 !s->keyframe ? AVDISCARD_NONKEY : AVDISCARD_ALL; | |
1471 | |
1472 if (avctx->skip_frame >= skip_thresh) { | |
1473 s->invisible = 1; | |
1474 goto skip_decode; | |
1475 } | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
1476 s->deblock_filter = s->filter.level && avctx->skip_loop_filter < skip_thresh; |
11921 | 1477 |
1478 for (i = 0; i < 4; i++) | |
1479 if (&s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] && | |
1480 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] && | |
1481 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) { | |
1482 curframe = s->framep[VP56_FRAME_CURRENT] = &s->frames[i]; | |
1483 break; | |
1484 } | |
1485 if (curframe->data[0]) | |
1486 avctx->release_buffer(avctx, curframe); | |
1487 | |
1488 curframe->key_frame = s->keyframe; | |
1489 curframe->pict_type = s->keyframe ? FF_I_TYPE : FF_P_TYPE; | |
1490 curframe->reference = referenced ? 3 : 0; | |
1491 if ((ret = avctx->get_buffer(avctx, curframe))) { | |
1492 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed!\n"); | |
1493 return ret; | |
1494 } | |
1495 | |
1496 // Given that arithmetic probabilities are updated every frame, it's quite likely | |
1497 // that the values we have on a random interframe are complete junk if we didn't | |
1498 // start decode on a keyframe. So just don't display anything rather than junk. | |
1499 if (!s->keyframe && (!s->framep[VP56_FRAME_PREVIOUS] || | |
1500 !s->framep[VP56_FRAME_GOLDEN] || | |
1501 !s->framep[VP56_FRAME_GOLDEN2])) { | |
1502 av_log(avctx, AV_LOG_WARNING, "Discarding interframe without a prior keyframe!\n"); | |
1503 return AVERROR_INVALIDDATA; | |
1504 } | |
1505 | |
1506 s->linesize = curframe->linesize[0]; | |
1507 s->uvlinesize = curframe->linesize[1]; | |
1508 | |
1509 if (!s->edge_emu_buffer) | |
1510 s->edge_emu_buffer = av_malloc(21*s->linesize); | |
1511 | |
1512 memset(s->top_nnz, 0, s->mb_width*sizeof(*s->top_nnz)); | |
1513 | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1514 /* 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
|
1515 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
|
1516 |
11921 | 1517 // 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
|
1518 memset(s->top_border, 127, (s->mb_width+1)*sizeof(*s->top_border)); |
12231 | 1519 memset(s->ref_count, 0, sizeof(s->ref_count)); |
11921 | 1520 |
1521 for (mb_y = 0; mb_y < s->mb_height; mb_y++) { | |
1522 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
|
1523 VP8Macroblock *mb = s->macroblocks + (s->mb_height - mb_y - 1)*2; |
11921 | 1524 uint8_t *intra4x4 = s->intra4x4_pred_mode + 4*mb_y*s->b4_stride; |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1525 uint8_t *segment_map = s->segmentation_map + mb_y*s->mb_stride; |
12237 | 1526 int mb_xy = mb_y * s->mb_stride; |
11921 | 1527 uint8_t *dst[3] = { |
1528 curframe->data[0] + 16*mb_y*s->linesize, | |
1529 curframe->data[1] + 8*mb_y*s->uvlinesize, | |
1530 curframe->data[2] + 8*mb_y*s->uvlinesize | |
1531 }; | |
1532 | |
1533 memset(s->left_nnz, 0, sizeof(s->left_nnz)); | |
1534 | |
1535 // left edge of 129 for intra prediction | |
1536 if (!(avctx->flags & CODEC_FLAG_EMU_EDGE)) | |
1537 for (i = 0; i < 3; i++) | |
1538 for (y = 0; y < 16>>!!i; y++) | |
1539 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
|
1540 if (mb_y) |
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
1541 memset(s->top_border, 129, sizeof(*s->top_border)); |
11921 | 1542 |
12237 | 1543 for (mb_x = 0; mb_x < s->mb_width; mb_x++, mb_xy++, mb++) { |
12221 | 1544 uint8_t *intra4x4_mb = s->keyframe ? intra4x4 + 4*mb_x : s->intra4x4_pred_mode_mb; |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1545 uint8_t *segment_mb = segment_map+mb_x; |
12221 | 1546 |
12215 | 1547 /* Prefetch the current frame, 4 MBs ahead */ |
1548 s->dsp.prefetch(dst[0] + (mb_x&3)*4*s->linesize + 64, s->linesize, 4); | |
1549 s->dsp.prefetch(dst[1] + (mb_x&7)*s->uvlinesize + 64, dst[2] - dst[1], 2); | |
1550 | |
12223
93e27a5401de
Convert VP8 macroblock structures to a ring buffer.
darkshikari
parents:
12222
diff
changeset
|
1551 decode_mb_mode(s, mb, mb_x, mb_y, intra4x4_mb, segment_mb); |
11921 | 1552 |
12237 | 1553 prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_PREVIOUS); |
12231 | 1554 |
11921 | 1555 if (!mb->skip) |
1556 decode_mb_coeffs(s, c, mb, s->top_nnz[mb_x], s->left_nnz); | |
1557 | |
12225
c3e11b3108d7
Eliminate a pointless memset for intra blocks in P-frames in VP8
darkshikari
parents:
12224
diff
changeset
|
1558 if (mb->mode <= MODE_I4x4) |
12221 | 1559 intra_predict(s, dst, mb, intra4x4_mb, mb_x, mb_y); |
12225
c3e11b3108d7
Eliminate a pointless memset for intra blocks in P-frames in VP8
darkshikari
parents:
12224
diff
changeset
|
1560 else |
11921 | 1561 inter_predict(s, dst, mb, mb_x, mb_y); |
1562 | |
12237 | 1563 prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN); |
12231 | 1564 |
11921 | 1565 if (!mb->skip) { |
12238 | 1566 idct_mb(s, dst, mb); |
11921 | 1567 } else { |
1568 AV_ZERO64(s->left_nnz); | |
1569 AV_WN64(s->top_nnz[mb_x], 0); // array of 9, so unaligned | |
1570 | |
1571 // Reset DC block predictors if they would exist if the mb had coefficients | |
1572 if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) { | |
1573 s->left_nnz[8] = 0; | |
1574 s->top_nnz[mb_x][8] = 0; | |
1575 } | |
1576 } | |
1577 | |
12222
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1578 if (s->deblock_filter) |
7acdbfd2a222
Calculate deblock strength per-MB instead of per-row
darkshikari
parents:
12221
diff
changeset
|
1579 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
|
1580 |
12237 | 1581 prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN2); |
12231 | 1582 |
11921 | 1583 dst[0] += 16; |
1584 dst[1] += 8; | |
1585 dst[2] += 8; | |
1586 } | |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
1587 if (s->deblock_filter) { |
11921 | 1588 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
|
1589 filter_mb_row_simple(s, mb_y); |
11921 | 1590 else |
12170
6f0db2eeaf70
vp8: Save mb border needed for intra prediction so that loop filter can run
conrad
parents:
12169
diff
changeset
|
1591 filter_mb_row(s, mb_y); |
11921 | 1592 } |
1593 } | |
1594 | |
1595 skip_decode: | |
1596 // if future frames don't use the updated probabilities, | |
1597 // reset them to the values we saved | |
1598 if (!s->update_probabilities) | |
1599 s->prob[0] = s->prob[1]; | |
1600 | |
1601 // check if golden and altref are swapped | |
1602 if (s->update_altref == VP56_FRAME_GOLDEN && | |
1603 s->update_golden == VP56_FRAME_GOLDEN2) | |
1604 FFSWAP(AVFrame *, s->framep[VP56_FRAME_GOLDEN], s->framep[VP56_FRAME_GOLDEN2]); | |
1605 else { | |
1606 if (s->update_altref != VP56_FRAME_NONE) | |
1607 s->framep[VP56_FRAME_GOLDEN2] = s->framep[s->update_altref]; | |
1608 | |
1609 if (s->update_golden != VP56_FRAME_NONE) | |
1610 s->framep[VP56_FRAME_GOLDEN] = s->framep[s->update_golden]; | |
1611 } | |
1612 | |
1613 if (s->update_last) // move cur->prev | |
1614 s->framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_CURRENT]; | |
1615 | |
1616 // release no longer referenced frames | |
1617 for (i = 0; i < 4; i++) | |
1618 if (s->frames[i].data[0] && | |
1619 &s->frames[i] != s->framep[VP56_FRAME_CURRENT] && | |
1620 &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] && | |
1621 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] && | |
1622 &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) | |
1623 avctx->release_buffer(avctx, &s->frames[i]); | |
1624 | |
1625 if (!s->invisible) { | |
1626 *(AVFrame*)data = *s->framep[VP56_FRAME_CURRENT]; | |
1627 *data_size = sizeof(AVFrame); | |
1628 } | |
1629 | |
1630 return avpkt->size; | |
1631 } | |
1632 | |
1633 static av_cold int vp8_decode_init(AVCodecContext *avctx) | |
1634 { | |
1635 VP8Context *s = avctx->priv_data; | |
1636 | |
1637 s->avctx = avctx; | |
1638 avctx->pix_fmt = PIX_FMT_YUV420P; | |
1639 | |
1640 dsputil_init(&s->dsp, avctx); | |
1641 ff_h264_pred_init(&s->hpc, CODEC_ID_VP8); | |
1642 ff_vp8dsp_init(&s->vp8dsp); | |
1643 | |
1644 // intra pred needs edge emulation among other things | |
1645 if (avctx->flags&CODEC_FLAG_EMU_EDGE) { | |
11947 | 1646 av_log(avctx, AV_LOG_ERROR, "Edge emulation not supported\n"); |
11921 | 1647 return AVERROR_PATCHWELCOME; |
1648 } | |
1649 | |
1650 return 0; | |
1651 } | |
1652 | |
1653 static av_cold int vp8_decode_free(AVCodecContext *avctx) | |
1654 { | |
1655 vp8_decode_flush(avctx); | |
1656 return 0; | |
1657 } | |
1658 | |
1659 AVCodec vp8_decoder = { | |
1660 "vp8", | |
1661 AVMEDIA_TYPE_VIDEO, | |
1662 CODEC_ID_VP8, | |
1663 sizeof(VP8Context), | |
1664 vp8_decode_init, | |
1665 NULL, | |
1666 vp8_decode_free, | |
1667 vp8_decode_frame, | |
1668 CODEC_CAP_DR1, | |
1669 .flush = vp8_decode_flush, | |
1670 .long_name = NULL_IF_CONFIG_SMALL("On2 VP8"), | |
1671 }; |