Parent Directory | Revision Log
Revision 1054 - (view) (download)
1 : | edgomez | 1054 | /***************************************************************************** |
2 : | * | ||
3 : | * XVID MPEG-4 VIDEO CODEC | ||
4 : | * - Motion Compensation related code - | ||
5 : | * | ||
6 : | * Copyright(C) 2002 Peter Ross <pross@xvid.org> | ||
7 : | * 2003 Christoph Lampert <gruel@web.de> | ||
8 : | * | ||
9 : | * This program is free software ; you can redistribute it and/or modify | ||
10 : | * it under the terms of the GNU General Public License as published by | ||
11 : | * the Free Software Foundation ; either version 2 of the License, or | ||
12 : | * (at your option) any later version. | ||
13 : | * | ||
14 : | * This program is distributed in the hope that it will be useful, | ||
15 : | * but WITHOUT ANY WARRANTY ; without even the implied warranty of | ||
16 : | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | ||
17 : | * GNU General Public License for more details. | ||
18 : | * | ||
19 : | * You should have received a copy of the GNU General Public License | ||
20 : | * along with this program ; if not, write to the Free Software | ||
21 : | * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | ||
22 : | * | ||
23 : | * $Id: motion_comp.c,v 1.18.2.6 2003-06-09 13:54:31 edgomez Exp $ | ||
24 : | * | ||
25 : | ****************************************************************************/ | ||
26 : | suxen_drol | 118 | |
27 : | edgomez | 851 | #include <stdio.h> |
28 : | |||
29 : | Isibaar | 3 | #include "../encoder.h" |
30 : | #include "../utils/mbfunctions.h" | ||
31 : | #include "../image/interpolate8x8.h" | ||
32 : | edgomez | 851 | #include "../image/reduced.h" |
33 : | Isibaar | 3 | #include "../utils/timer.h" |
34 : | suxen_drol | 118 | #include "motion.h" |
35 : | Isibaar | 3 | |
36 : | edgomez | 851 | #ifndef RSHIFT |
37 : | #define RSHIFT(a,b) ((a) > 0 ? ((a) + (1<<((b)-1)))>>(b) : ((a) + (1<<((b)-1))-1)>>(b)) | ||
38 : | #endif | ||
39 : | |||
40 : | /* assume b>0 */ | ||
41 : | #ifndef RDIV | ||
42 : | #define RDIV(a,b) (((a)>0 ? (a) + ((b)>>1) : (a) - ((b)>>1))/(b)) | ||
43 : | #endif | ||
44 : | |||
45 : | |||
46 : | syskin | 935 | /* This is borrowed from decoder.c */ |
47 : | edgomez | 851 | static __inline int gmc_sanitize(int value, int quarterpel, int fcode) |
48 : | { | ||
49 : | int length = 1 << (fcode+4); | ||
50 : | |||
51 : | edgomez | 1053 | #if 0 |
52 : | if (quarterpel) value *= 2; | ||
53 : | #endif | ||
54 : | edgomez | 851 | |
55 : | syskin | 935 | if (value < -length) |
56 : | edgomez | 851 | return -length; |
57 : | syskin | 935 | else if (value >= length) |
58 : | edgomez | 851 | return length-1; |
59 : | else return value; | ||
60 : | } | ||
61 : | |||
62 : | /* And this is borrowed from bitstream.c until we find a common solution */ | ||
63 : | |||
64 : | static uint32_t __inline | ||
65 : | log2bin(uint32_t value) | ||
66 : | { | ||
67 : | /* Changed by Chenm001 */ | ||
68 : | #if !defined(_MSC_VER) | ||
69 : | int n = 0; | ||
70 : | |||
71 : | while (value) { | ||
72 : | value >>= 1; | ||
73 : | n++; | ||
74 : | } | ||
75 : | return n; | ||
76 : | #else | ||
77 : | __asm { | ||
78 : | syskin | 935 | bsr eax, value |
79 : | edgomez | 851 | inc eax |
80 : | } | ||
81 : | #endif | ||
82 : | } | ||
83 : | |||
84 : | |||
85 : | edgomez | 195 | static __inline void |
86 : | edgomez | 851 | compensate16x16_interpolate(int16_t * const dct_codes, |
87 : | uint8_t * const cur, | ||
88 : | const uint8_t * const ref, | ||
89 : | const uint8_t * const refh, | ||
90 : | const uint8_t * const refv, | ||
91 : | const uint8_t * const refhv, | ||
92 : | uint8_t * const tmp, | ||
93 : | uint32_t x, | ||
94 : | uint32_t y, | ||
95 : | const int32_t dx, | ||
96 : | const int32_t dy, | ||
97 : | const int32_t stride, | ||
98 : | const int quarterpel, | ||
99 : | const int reduced_resolution, | ||
100 : | const int32_t rounding) | ||
101 : | Isibaar | 3 | { |
102 : | edgomez | 851 | const uint8_t * ptr; |
103 : | Isibaar | 3 | |
104 : | edgomez | 851 | if (!reduced_resolution) { |
105 : | Isibaar | 3 | |
106 : | edgomez | 851 | if(quarterpel) { |
107 : | if ((dx&3) | (dy&3)) { | ||
108 : | interpolate16x16_quarterpel(tmp - y * stride - x, | ||
109 : | (uint8_t *) ref, tmp + 32, | ||
110 : | tmp + 64, tmp + 96, x, y, dx, dy, stride, rounding); | ||
111 : | ptr = tmp; | ||
112 : | edgomez | 1053 | } else ptr = ref + (y + dy/4)*stride + x + dx/4; /* fullpixel position */ |
113 : | Isibaar | 3 | |
114 : | edgomez | 851 | } else ptr = get_ref(ref, refh, refv, refhv, x, y, 1, dx, dy, stride); |
115 : | |||
116 : | edgomez | 195 | transfer_8to16sub(dct_codes, cur + y * stride + x, |
117 : | syskin | 935 | ptr, stride); |
118 : | edgomez | 851 | transfer_8to16sub(dct_codes+64, cur + y * stride + x + 8, |
119 : | syskin | 935 | ptr + 8, stride); |
120 : | edgomez | 851 | transfer_8to16sub(dct_codes+128, cur + y * stride + x + 8*stride, |
121 : | syskin | 935 | ptr + 8*stride, stride); |
122 : | edgomez | 851 | transfer_8to16sub(dct_codes+192, cur + y * stride + x + 8*stride+8, |
123 : | syskin | 935 | ptr + 8*stride + 8, stride); |
124 : | Isibaar | 3 | |
125 : | edgomez | 1053 | } else { /* reduced_resolution */ |
126 : | syskin | 935 | |
127 : | edgomez | 851 | x *= 2; y *= 2; |
128 : | |||
129 : | ptr = get_ref(ref, refh, refv, refhv, x, y, 1, dx, dy, stride); | ||
130 : | syskin | 935 | |
131 : | edgomez | 851 | filter_18x18_to_8x8(dct_codes, cur+y*stride + x, stride); |
132 : | filter_diff_18x18_to_8x8(dct_codes, ptr, stride); | ||
133 : | |||
134 : | filter_18x18_to_8x8(dct_codes+64, cur+y*stride + x + 16, stride); | ||
135 : | filter_diff_18x18_to_8x8(dct_codes+64, ptr + 16, stride); | ||
136 : | |||
137 : | filter_18x18_to_8x8(dct_codes+128, cur+(y+16)*stride + x, stride); | ||
138 : | filter_diff_18x18_to_8x8(dct_codes+128, ptr + 16*stride, stride); | ||
139 : | |||
140 : | filter_18x18_to_8x8(dct_codes+192, cur+(y+16)*stride + x + 16, stride); | ||
141 : | filter_diff_18x18_to_8x8(dct_codes+192, ptr + 16*stride + 16, stride); | ||
142 : | |||
143 : | transfer32x32_copy(cur + y*stride + x, ptr, stride); | ||
144 : | edgomez | 195 | } |
145 : | Isibaar | 3 | } |
146 : | |||
147 : | edgomez | 851 | static __inline void |
148 : | compensate8x8_interpolate( int16_t * const dct_codes, | ||
149 : | uint8_t * const cur, | ||
150 : | const uint8_t * const ref, | ||
151 : | const uint8_t * const refh, | ||
152 : | const uint8_t * const refv, | ||
153 : | const uint8_t * const refhv, | ||
154 : | uint8_t * const tmp, | ||
155 : | uint32_t x, | ||
156 : | uint32_t y, | ||
157 : | const int32_t dx, | ||
158 : | const int32_t dy, | ||
159 : | const int32_t stride, | ||
160 : | const int32_t quarterpel, | ||
161 : | const int reduced_resolution, | ||
162 : | const int32_t rounding) | ||
163 : | { | ||
164 : | const uint8_t * ptr; | ||
165 : | Isibaar | 3 | |
166 : | edgomez | 851 | if (!reduced_resolution) { |
167 : | Isibaar | 3 | |
168 : | edgomez | 851 | if(quarterpel) { |
169 : | if ((dx&3) | (dy&3)) { | ||
170 : | interpolate8x8_quarterpel(tmp - y*stride - x, | ||
171 : | (uint8_t *) ref, tmp + 32, | ||
172 : | tmp + 64, tmp + 96, x, y, dx, dy, stride, rounding); | ||
173 : | ptr = tmp; | ||
174 : | edgomez | 1053 | } else ptr = ref + (y + dy/4)*stride + x + dx/4; /* fullpixel position */ |
175 : | edgomez | 851 | } else ptr = get_ref(ref, refh, refv, refhv, x, y, 1, dx, dy, stride); |
176 : | |||
177 : | transfer_8to16sub(dct_codes, cur + y * stride + x, ptr, stride); | ||
178 : | |||
179 : | edgomez | 1053 | } else { /* reduced_resolution */ |
180 : | edgomez | 851 | |
181 : | x *= 2; y *= 2; | ||
182 : | |||
183 : | ptr = get_ref(ref, refh, refv, refhv, x, y, 1, dx, dy, stride); | ||
184 : | |||
185 : | filter_18x18_to_8x8(dct_codes, cur+y*stride + x, stride); | ||
186 : | filter_diff_18x18_to_8x8(dct_codes, ptr, stride); | ||
187 : | syskin | 935 | |
188 : | edgomez | 851 | transfer16x16_copy(cur + y*stride + x, ptr, stride); |
189 : | } | ||
190 : | } | ||
191 : | |||
192 : | /* XXX: slow, inelegant... */ | ||
193 : | static void | ||
194 : | interpolate18x18_switch(uint8_t * const cur, | ||
195 : | const uint8_t * const refn, | ||
196 : | const uint32_t x, | ||
197 : | const uint32_t y, | ||
198 : | const int32_t dx, | ||
199 : | const int dy, | ||
200 : | const int32_t stride, | ||
201 : | const int32_t rounding) | ||
202 : | { | ||
203 : | interpolate8x8_switch(cur, refn, x-1, y-1, dx, dy, stride, rounding); | ||
204 : | interpolate8x8_switch(cur, refn, x+7, y-1, dx, dy, stride, rounding); | ||
205 : | interpolate8x8_switch(cur, refn, x+9, y-1, dx, dy, stride, rounding); | ||
206 : | |||
207 : | interpolate8x8_switch(cur, refn, x-1, y+7, dx, dy, stride, rounding); | ||
208 : | interpolate8x8_switch(cur, refn, x+7, y+7, dx, dy, stride, rounding); | ||
209 : | interpolate8x8_switch(cur, refn, x+9, y+7, dx, dy, stride, rounding); | ||
210 : | |||
211 : | interpolate8x8_switch(cur, refn, x-1, y+9, dx, dy, stride, rounding); | ||
212 : | interpolate8x8_switch(cur, refn, x+7, y+9, dx, dy, stride, rounding); | ||
213 : | interpolate8x8_switch(cur, refn, x+9, y+9, dx, dy, stride, rounding); | ||
214 : | } | ||
215 : | |||
216 : | static void | ||
217 : | CompensateChroma( int dx, int dy, | ||
218 : | const int i, const int j, | ||
219 : | IMAGE * const Cur, | ||
220 : | const IMAGE * const Ref, | ||
221 : | uint8_t * const temp, | ||
222 : | int16_t * const coeff, | ||
223 : | const int32_t stride, | ||
224 : | const int rounding, | ||
225 : | const int rrv) | ||
226 : | { /* uv-block-based compensation */ | ||
227 : | |||
228 : | if (!rrv) { | ||
229 : | transfer_8to16sub(coeff, Cur->u + 8 * j * stride + 8 * i, | ||
230 : | syskin | 935 | interpolate8x8_switch2(temp, Ref->u, 8 * i, 8 * j, |
231 : | edgomez | 851 | dx, dy, stride, rounding), |
232 : | stride); | ||
233 : | transfer_8to16sub(coeff + 64, Cur->v + 8 * j * stride + 8 * i, | ||
234 : | syskin | 935 | interpolate8x8_switch2(temp, Ref->v, 8 * i, 8 * j, |
235 : | edgomez | 851 | dx, dy, stride, rounding), |
236 : | stride); | ||
237 : | } else { | ||
238 : | uint8_t * current, * reference; | ||
239 : | |||
240 : | current = Cur->u + 16*j*stride + 16*i; | ||
241 : | reference = temp - 16*j*stride - 16*i; | ||
242 : | interpolate18x18_switch(reference, Ref->u, 16*i, 16*j, dx, dy, stride, rounding); | ||
243 : | filter_18x18_to_8x8(coeff, current, stride); | ||
244 : | filter_diff_18x18_to_8x8(coeff, temp, stride); | ||
245 : | transfer16x16_copy(current, temp, stride); | ||
246 : | |||
247 : | current = Cur->v + 16*j*stride + 16*i; | ||
248 : | interpolate18x18_switch(reference, Ref->v, 16*i, 16*j, dx, dy, stride, rounding); | ||
249 : | filter_18x18_to_8x8(coeff + 64, current, stride); | ||
250 : | filter_diff_18x18_to_8x8(coeff + 64, temp, stride); | ||
251 : | transfer16x16_copy(current, temp, stride); | ||
252 : | } | ||
253 : | } | ||
254 : | |||
255 : | edgomez | 195 | void |
256 : | MBMotionCompensation(MACROBLOCK * const mb, | ||
257 : | syskin | 935 | const uint32_t i, |
258 : | const uint32_t j, | ||
259 : | const IMAGE * const ref, | ||
260 : | const IMAGE * const refh, | ||
261 : | const IMAGE * const refv, | ||
262 : | const IMAGE * const refhv, | ||
263 : | const IMAGE * const refGMC, | ||
264 : | IMAGE * const cur, | ||
265 : | int16_t * dct_codes, | ||
266 : | const uint32_t width, | ||
267 : | const uint32_t height, | ||
268 : | const uint32_t edged_width, | ||
269 : | const int32_t quarterpel, | ||
270 : | const int reduced_resolution, | ||
271 : | const int32_t rounding) | ||
272 : | Isibaar | 3 | { |
273 : | edgomez | 851 | int32_t dx; |
274 : | int32_t dy; | ||
275 : | Isibaar | 3 | |
276 : | edgomez | 851 | uint8_t * const tmp = refv->u; |
277 : | Isibaar | 3 | |
278 : | edgomez | 851 | if ( (!reduced_resolution) && (mb->mode == MODE_NOT_CODED) ) { /* quick copy for early SKIP */ |
279 : | /* early SKIP is only activated in P-VOPs, not in S-VOPs, so mcsel can never be 1 */ | ||
280 : | Isibaar | 3 | |
281 : | edgomez | 851 | transfer16x16_copy(cur->y + 16 * (i + j * edged_width), |
282 : | syskin | 935 | ref->y + 16 * (i + j * edged_width), |
283 : | edged_width); | ||
284 : | |||
285 : | edgomez | 851 | transfer8x8_copy(cur->u + 8 * (i + j * edged_width/2), |
286 : | ref->u + 8 * (i + j * edged_width/2), | ||
287 : | edged_width / 2); | ||
288 : | transfer8x8_copy(cur->v + 8 * (i + j * edged_width/2), | ||
289 : | ref->v + 8 * (i + j * edged_width/2), | ||
290 : | edged_width / 2); | ||
291 : | return; | ||
292 : | } | ||
293 : | Isibaar | 3 | |
294 : | syskin | 935 | if ((mb->mode == MODE_NOT_CODED || mb->mode == MODE_INTER |
295 : | edgomez | 851 | || mb->mode == MODE_INTER_Q)) { |
296 : | chl | 437 | |
297 : | edgomez | 851 | /* reduced resolution + GMC: not possible */ |
298 : | Isibaar | 3 | |
299 : | edgomez | 851 | if (mb->mcsel) { |
300 : | syskin | 935 | |
301 : | edgomez | 851 | /* call normal routine once, easier than "if (mcsel)"ing all the time */ |
302 : | syskin | 935 | |
303 : | edgomez | 851 | transfer_8to16sub(&dct_codes[0*64], cur->y + 16*j*edged_width + 16*i, |
304 : | syskin | 935 | refGMC->y + 16*j*edged_width + 16*i, edged_width); |
305 : | edgomez | 851 | transfer_8to16sub(&dct_codes[1*64], cur->y + 16*j*edged_width + 16*i+8, |
306 : | syskin | 935 | refGMC->y + 16*j*edged_width + 16*i+8, edged_width); |
307 : | edgomez | 851 | transfer_8to16sub(&dct_codes[2*64], cur->y + (16*j+8)*edged_width + 16*i, |
308 : | syskin | 935 | refGMC->y + (16*j+8)*edged_width + 16*i, edged_width); |
309 : | edgomez | 851 | transfer_8to16sub(&dct_codes[3*64], cur->y + (16*j+8)*edged_width + 16*i+8, |
310 : | syskin | 935 | refGMC->y + (16*j+8)*edged_width + 16*i+8, edged_width); |
311 : | chl | 437 | |
312 : | edgomez | 851 | /* lumi is needed earlier for mode decision, but chroma should be done block-based, but it isn't, yet. */ |
313 : | |||
314 : | transfer_8to16sub(&dct_codes[4 * 64], cur->u + 8 *j*edged_width/2 + 8*i, | ||
315 : | refGMC->u + 8 *j*edged_width/2 + 8*i, edged_width/2); | ||
316 : | |||
317 : | transfer_8to16sub(&dct_codes[5 * 64], cur->v + 8*j* edged_width/2 + 8*i, | ||
318 : | refGMC->v + 8*j* edged_width/2 + 8*i, edged_width/2); | ||
319 : | |||
320 : | return; | ||
321 : | } | ||
322 : | |||
323 : | /* ordinary compensation */ | ||
324 : | syskin | 935 | |
325 : | edgomez | 851 | dx = (quarterpel ? mb->qmvs[0].x : mb->mvs[0].x); |
326 : | dy = (quarterpel ? mb->qmvs[0].y : mb->mvs[0].y); | ||
327 : | |||
328 : | if (reduced_resolution) { | ||
329 : | dx = RRV_MV_SCALEUP(dx); | ||
330 : | dy = RRV_MV_SCALEUP(dy); | ||
331 : | } | ||
332 : | |||
333 : | compensate16x16_interpolate(&dct_codes[0 * 64], cur->y, ref->y, refh->y, | ||
334 : | refv->y, refhv->y, tmp, 16 * i, 16 * j, dx, dy, | ||
335 : | edged_width, quarterpel, reduced_resolution, rounding); | ||
336 : | |||
337 : | syskin | 935 | if (quarterpel) { dx /= 2; dy /= 2; } |
338 : | |||
339 : | edgomez | 851 | dx = (dx >> 1) + roundtab_79[dx & 0x3]; |
340 : | dy = (dy >> 1) + roundtab_79[dy & 0x3]; | ||
341 : | |||
342 : | edgomez | 1053 | } else { /* mode == MODE_INTER4V */ |
343 : | edgomez | 851 | int k, sumx = 0, sumy = 0; |
344 : | const VECTOR * const mvs = (quarterpel ? mb->qmvs : mb->mvs); | ||
345 : | |||
346 : | for (k = 0; k < 4; k++) { | ||
347 : | dx = mvs[k].x; | ||
348 : | dy = mvs[k].y; | ||
349 : | syskin | 935 | sumx += quarterpel ? dx/2 : dx; |
350 : | sumy += quarterpel ? dy/2 : dy; | ||
351 : | edgomez | 851 | |
352 : | if (reduced_resolution){ | ||
353 : | dx = RRV_MV_SCALEUP(dx); | ||
354 : | dy = RRV_MV_SCALEUP(dy); | ||
355 : | } | ||
356 : | |||
357 : | compensate8x8_interpolate(&dct_codes[k * 64], cur->y, ref->y, refh->y, | ||
358 : | refv->y, refhv->y, tmp, 16 * i + 8*(k&1), 16 * j + 8*(k>>1), dx, | ||
359 : | dy, edged_width, quarterpel, reduced_resolution, rounding); | ||
360 : | } | ||
361 : | dx = (sumx >> 3) + roundtab_76[sumx & 0xf]; | ||
362 : | dy = (sumy >> 3) + roundtab_76[sumy & 0xf]; | ||
363 : | } | ||
364 : | |||
365 : | CompensateChroma(dx, dy, i, j, cur, ref, tmp, | ||
366 : | &dct_codes[4 * 64], edged_width / 2, rounding, reduced_resolution); | ||
367 : | } | ||
368 : | |||
369 : | |||
370 : | void | ||
371 : | MBMotionCompensationBVOP(MBParam * pParam, | ||
372 : | MACROBLOCK * const mb, | ||
373 : | const uint32_t i, | ||
374 : | const uint32_t j, | ||
375 : | IMAGE * const cur, | ||
376 : | const IMAGE * const f_ref, | ||
377 : | const IMAGE * const f_refh, | ||
378 : | const IMAGE * const f_refv, | ||
379 : | const IMAGE * const f_refhv, | ||
380 : | const IMAGE * const b_ref, | ||
381 : | const IMAGE * const b_refh, | ||
382 : | const IMAGE * const b_refv, | ||
383 : | const IMAGE * const b_refhv, | ||
384 : | int16_t * dct_codes) | ||
385 : | { | ||
386 : | const uint32_t edged_width = pParam->edged_width; | ||
387 : | int32_t dx, dy, b_dx, b_dy, sumx, sumy, b_sumx, b_sumy; | ||
388 : | int k; | ||
389 : | edgomez | 949 | const int quarterpel = pParam->vol_flags & XVID_VOL_QUARTERPEL; |
390 : | edgomez | 851 | const uint8_t * ptr1, * ptr2; |
391 : | uint8_t * const tmp = f_refv->u; | ||
392 : | const VECTOR * const fmvs = (quarterpel ? mb->qmvs : mb->mvs); | ||
393 : | const VECTOR * const bmvs = (quarterpel ? mb->b_qmvs : mb->b_mvs); | ||
394 : | |||
395 : | switch (mb->mode) { | ||
396 : | case MODE_FORWARD: | ||
397 : | dx = fmvs->x; dy = fmvs->y; | ||
398 : | |||
399 : | compensate16x16_interpolate(&dct_codes[0 * 64], cur->y, f_ref->y, f_refh->y, | ||
400 : | f_refv->y, f_refhv->y, tmp, 16 * i, 16 * j, dx, | ||
401 : | dy, edged_width, quarterpel, 0, 0); | ||
402 : | |||
403 : | if (quarterpel) { dx /= 2; dy /= 2; } | ||
404 : | |||
405 : | CompensateChroma( (dx >> 1) + roundtab_79[dx & 0x3], | ||
406 : | (dy >> 1) + roundtab_79[dy & 0x3], | ||
407 : | i, j, cur, f_ref, tmp, | ||
408 : | &dct_codes[4 * 64], edged_width / 2, 0, 0); | ||
409 : | |||
410 : | return; | ||
411 : | |||
412 : | case MODE_BACKWARD: | ||
413 : | b_dx = bmvs->x; b_dy = bmvs->y; | ||
414 : | |||
415 : | syskin | 935 | compensate16x16_interpolate(&dct_codes[0 * 64], cur->y, b_ref->y, b_refh->y, |
416 : | edgomez | 851 | b_refv->y, b_refhv->y, tmp, 16 * i, 16 * j, b_dx, |
417 : | syskin | 935 | b_dy, edged_width, quarterpel, 0, 0); |
418 : | edgomez | 851 | |
419 : | if (quarterpel) { b_dx /= 2; b_dy /= 2; } | ||
420 : | |||
421 : | CompensateChroma( (b_dx >> 1) + roundtab_79[b_dx & 0x3], | ||
422 : | (b_dy >> 1) + roundtab_79[b_dy & 0x3], | ||
423 : | i, j, cur, b_ref, tmp, | ||
424 : | &dct_codes[4 * 64], edged_width / 2, 0, 0); | ||
425 : | |||
426 : | return; | ||
427 : | |||
428 : | case MODE_INTERPOLATE: /* _could_ use DIRECT, but would be overkill (no 4MV there) */ | ||
429 : | case MODE_DIRECT_NO4V: | ||
430 : | dx = fmvs->x; dy = fmvs->y; | ||
431 : | b_dx = bmvs->x; b_dy = bmvs->y; | ||
432 : | |||
433 : | if (quarterpel) { | ||
434 : | syskin | 935 | |
435 : | edgomez | 851 | if ((dx&3) | (dy&3)) { |
436 : | interpolate16x16_quarterpel(tmp - i * 16 - j * 16 * edged_width, | ||
437 : | (uint8_t *) f_ref->y, tmp + 32, | ||
438 : | tmp + 64, tmp + 96, 16*i, 16*j, dx, dy, edged_width, 0); | ||
439 : | ptr1 = tmp; | ||
440 : | edgomez | 1053 | } else ptr1 = f_ref->y + (16*j + dy/4)*edged_width + 16*i + dx/4; /* fullpixel position */ |
441 : | edgomez | 851 | |
442 : | if ((b_dx&3) | (b_dy&3)) { | ||
443 : | interpolate16x16_quarterpel(tmp - i * 16 - j * 16 * edged_width + 16, | ||
444 : | (uint8_t *) b_ref->y, tmp + 32, | ||
445 : | tmp + 64, tmp + 96, 16*i, 16*j, b_dx, b_dy, edged_width, 0); | ||
446 : | ptr2 = tmp + 16; | ||
447 : | edgomez | 1053 | } else ptr2 = b_ref->y + (16*j + b_dy/4)*edged_width + 16*i + b_dx/4; /* fullpixel position */ |
448 : | edgomez | 851 | |
449 : | b_dx /= 2; | ||
450 : | b_dy /= 2; | ||
451 : | dx /= 2; | ||
452 : | dy /= 2; | ||
453 : | |||
454 : | } else { | ||
455 : | ptr1 = get_ref(f_ref->y, f_refh->y, f_refv->y, f_refhv->y, | ||
456 : | i, j, 16, dx, dy, edged_width); | ||
457 : | |||
458 : | ptr2 = get_ref(b_ref->y, b_refh->y, b_refv->y, b_refhv->y, | ||
459 : | i, j, 16, b_dx, b_dy, edged_width); | ||
460 : | } | ||
461 : | for (k = 0; k < 4; k++) | ||
462 : | transfer_8to16sub2(&dct_codes[k * 64], | ||
463 : | cur->y + (i * 16+(k&1)*8) + (j * 16+((k>>1)*8)) * edged_width, | ||
464 : | ptr1 + (k&1)*8 + (k>>1)*8*edged_width, | ||
465 : | ptr2 + (k&1)*8 + (k>>1)*8*edged_width, edged_width); | ||
466 : | |||
467 : | |||
468 : | dx = (dx >> 1) + roundtab_79[dx & 0x3]; | ||
469 : | dy = (dy >> 1) + roundtab_79[dy & 0x3]; | ||
470 : | |||
471 : | b_dx = (b_dx >> 1) + roundtab_79[b_dx & 0x3]; | ||
472 : | b_dy = (b_dy >> 1) + roundtab_79[b_dy & 0x3]; | ||
473 : | |||
474 : | break; | ||
475 : | syskin | 935 | |
476 : | edgomez | 1053 | default: /* MODE_DIRECT (or MODE_DIRECT_NONE_MV in case of bframes decoding) */ |
477 : | edgomez | 851 | sumx = sumy = b_sumx = b_sumy = 0; |
478 : | |||
479 : | for (k = 0; k < 4; k++) { | ||
480 : | syskin | 935 | |
481 : | edgomez | 851 | dx = fmvs[k].x; dy = fmvs[k].y; |
482 : | b_dx = bmvs[k].x; b_dy = bmvs[k].y; | ||
483 : | |||
484 : | if (quarterpel) { | ||
485 : | sumx += dx/2; sumy += dy/2; | ||
486 : | b_sumx += b_dx/2; b_sumy += b_dy/2; | ||
487 : | |||
488 : | if ((dx&3) | (dy&3)) { | ||
489 : | interpolate8x8_quarterpel(tmp - (i * 16+(k&1)*8) - (j * 16+((k>>1)*8)) * edged_width, | ||
490 : | syskin | 935 | (uint8_t *) f_ref->y, |
491 : | tmp + 32, tmp + 64, tmp + 96, | ||
492 : | edgomez | 851 | 16*i + (k&1)*8, 16*j + (k>>1)*8, dx, dy, edged_width, 0); |
493 : | ptr1 = tmp; | ||
494 : | } else ptr1 = f_ref->y + (16*j + (k>>1)*8 + dy/4)*edged_width + 16*i + (k&1)*8 + dx/4; | ||
495 : | |||
496 : | if ((b_dx&3) | (b_dy&3)) { | ||
497 : | interpolate8x8_quarterpel(tmp - (i * 16+(k&1)*8) - (j * 16+((k>>1)*8)) * edged_width + 16, | ||
498 : | (uint8_t *) b_ref->y, | ||
499 : | syskin | 935 | tmp + 16, tmp + 32, tmp + 48, |
500 : | edgomez | 851 | 16*i + (k&1)*8, 16*j + (k>>1)*8, b_dx, b_dy, edged_width, 0); |
501 : | ptr2 = tmp + 16; | ||
502 : | } else ptr2 = b_ref->y + (16*j + (k>>1)*8 + b_dy/4)*edged_width + 16*i + (k&1)*8 + b_dx/4; | ||
503 : | } else { | ||
504 : | sumx += dx; sumy += dy; | ||
505 : | b_sumx += b_dx; b_sumy += b_dy; | ||
506 : | |||
507 : | syskin | 935 | ptr1 = get_ref(f_ref->y, f_refh->y, f_refv->y, f_refhv->y, |
508 : | edgomez | 851 | 2*i + (k&1), 2*j + (k>>1), 8, dx, dy, edged_width); |
509 : | syskin | 935 | ptr2 = get_ref(b_ref->y, b_refh->y, b_refv->y, b_refhv->y, |
510 : | edgomez | 851 | 2*i + (k&1), 2*j + (k>>1), 8, b_dx, b_dy, edged_width); |
511 : | } | ||
512 : | transfer_8to16sub2(&dct_codes[k * 64], | ||
513 : | cur->y + (i * 16+(k&1)*8) + (j * 16+((k>>1)*8)) * edged_width, | ||
514 : | ptr1, ptr2, edged_width); | ||
515 : | syskin | 935 | |
516 : | edgomez | 851 | } |
517 : | |||
518 : | dx = (sumx >> 3) + roundtab_76[sumx & 0xf]; | ||
519 : | dy = (sumy >> 3) + roundtab_76[sumy & 0xf]; | ||
520 : | b_dx = (b_sumx >> 3) + roundtab_76[b_sumx & 0xf]; | ||
521 : | b_dy = (b_sumy >> 3) + roundtab_76[b_sumy & 0xf]; | ||
522 : | |||
523 : | break; | ||
524 : | } | ||
525 : | |||
526 : | edgomez | 1053 | /* v block-based chroma interpolation for direct and interpolate modes */ |
527 : | edgomez | 851 | transfer_8to16sub2(&dct_codes[4 * 64], |
528 : | cur->u + (j * 8) * edged_width / 2 + (i * 8), | ||
529 : | interpolate8x8_switch2(tmp, b_ref->u, 8 * i, 8 * j, | ||
530 : | b_dx, b_dy, edged_width / 2, 0), | ||
531 : | interpolate8x8_switch2(tmp + 8, f_ref->u, 8 * i, 8 * j, | ||
532 : | dx, dy, edged_width / 2, 0), | ||
533 : | edged_width / 2); | ||
534 : | |||
535 : | transfer_8to16sub2(&dct_codes[5 * 64], | ||
536 : | cur->v + (j * 8) * edged_width / 2 + (i * 8), | ||
537 : | interpolate8x8_switch2(tmp, b_ref->v, 8 * i, 8 * j, | ||
538 : | b_dx, b_dy, edged_width / 2, 0), | ||
539 : | interpolate8x8_switch2(tmp + 8, f_ref->v, 8 * i, 8 * j, | ||
540 : | dx, dy, edged_width / 2, 0), | ||
541 : | edged_width / 2); | ||
542 : | } | ||
543 : | |||
544 : | |||
545 : | |||
546 : | void generate_GMCparameters( const int num_wp, const int res, | ||
547 : | syskin | 935 | const WARPPOINTS *const warp, |
548 : | const int width, const int height, | ||
549 : | GMC_DATA *const gmc) | ||
550 : | edgomez | 851 | { |
551 : | syskin | 935 | const int du0 = warp->duv[0].x; |
552 : | const int dv0 = warp->duv[0].y; | ||
553 : | const int du1 = warp->duv[1].x; | ||
554 : | const int dv1 = warp->duv[1].y; | ||
555 : | const int du2 = warp->duv[2].x; | ||
556 : | const int dv2 = warp->duv[2].y; | ||
557 : | edgomez | 851 | |
558 : | syskin | 935 | gmc->W = width; |
559 : | gmc->H = height; | ||
560 : | edgomez | 851 | |
561 : | edgomez | 1053 | gmc->rho = 4 - log2bin(res-1); /* = {3,2,1,0} for res={2,4,8,16} */ |
562 : | edgomez | 851 | |
563 : | syskin | 935 | gmc->alpha = log2bin(gmc->W-1); |
564 : | gmc->Ws = (1 << gmc->alpha); | ||
565 : | edgomez | 851 | |
566 : | syskin | 935 | gmc->dxF = 16*gmc->Ws + RDIV( 8*gmc->Ws*du1, gmc->W ); |
567 : | gmc->dxG = RDIV( 8*gmc->Ws*dv1, gmc->W ); | ||
568 : | gmc->Fo = (res*du0 + 1) << (gmc->alpha+gmc->rho-1); | ||
569 : | gmc->Go = (res*dv0 + 1) << (gmc->alpha+gmc->rho-1); | ||
570 : | edgomez | 851 | |
571 : | syskin | 935 | if (num_wp==2) { |
572 : | gmc->dyF = -gmc->dxG; | ||
573 : | gmc->dyG = gmc->dxF; | ||
574 : | } else if (num_wp==3) { | ||
575 : | gmc->beta = log2bin(gmc->H-1); | ||
576 : | gmc->Hs = (1 << gmc->beta); | ||
577 : | gmc->dyF = RDIV( 8*gmc->Hs*du2, gmc->H ); | ||
578 : | gmc->dyG = 16*gmc->Hs + RDIV( 8*gmc->Hs*dv2, gmc->H ); | ||
579 : | if (gmc->beta > gmc->alpha) { | ||
580 : | gmc->dxF <<= (gmc->beta - gmc->alpha); | ||
581 : | gmc->dxG <<= (gmc->beta - gmc->alpha); | ||
582 : | gmc->alpha = gmc->beta; | ||
583 : | gmc->Ws = 1<< gmc->beta; | ||
584 : | } else { | ||
585 : | gmc->dyF <<= gmc->alpha - gmc->beta; | ||
586 : | gmc->dyG <<= gmc->alpha - gmc->beta; | ||
587 : | } | ||
588 : | } | ||
589 : | edgomez | 851 | |
590 : | syskin | 935 | gmc->cFo = gmc->dxF + gmc->dyF + (1 << (gmc->alpha+gmc->rho+1)); |
591 : | gmc->cFo += 16*gmc->Ws*(du0-1); | ||
592 : | |||
593 : | gmc->cGo = gmc->dxG + gmc->dyG + (1 << (gmc->alpha+gmc->rho+1)); | ||
594 : | gmc->cGo += 16*gmc->Ws*(dv0-1); | ||
595 : | edgomez | 851 | } |
596 : | |||
597 : | syskin | 935 | void |
598 : | edgomez | 1053 | generate_GMCimage( const GMC_DATA *const gmc_data, /* [input] precalculated data */ |
599 : | const IMAGE *const pRef, /* [input] */ | ||
600 : | syskin | 935 | const int mb_width, |
601 : | edgomez | 851 | const int mb_height, |
602 : | const int stride, | ||
603 : | syskin | 935 | const int stride2, |
604 : | edgomez | 1053 | const int fcode, /* [input] some parameters... */ |
605 : | const int32_t quarterpel, /* [input] for rounding avgMV */ | ||
606 : | const int reduced_resolution, /* [input] ignored */ | ||
607 : | const int32_t rounding, /* [input] for rounding image data */ | ||
608 : | MACROBLOCK *const pMBs, /* [output] average motion vectors */ | ||
609 : | IMAGE *const pGMC) /* [output] full warped image */ | ||
610 : | edgomez | 851 | { |
611 : | |||
612 : | unsigned int mj,mi; | ||
613 : | VECTOR avgMV; | ||
614 : | Isibaar | 3 | |
615 : | syskin | 935 | for (mj = 0; mj < (unsigned int)mb_height; mj++) |
616 : | for (mi = 0; mi < (unsigned int)mb_width; mi++) { | ||
617 : | |||
618 : | avgMV = generate_GMCimageMB(gmc_data, pRef, mi, mj, | ||
619 : | stride, stride2, quarterpel, rounding, pGMC); | ||
620 : | |||
621 : | pMBs[mj*mb_width+mi].amv.x = gmc_sanitize(avgMV.x, quarterpel, fcode); | ||
622 : | pMBs[mj*mb_width+mi].amv.y = gmc_sanitize(avgMV.y, quarterpel, fcode); | ||
623 : | pMBs[mj*mb_width+mi].mcsel = 0; /* until mode decision */ | ||
624 : | edgomez | 851 | } |
625 : | } | ||
626 : | Isibaar | 3 | |
627 : | |||
628 : | |||
629 : | edgomez | 851 | #define MLT(i) (((16-(i))<<16) + (i)) |
630 : | static const uint32_t MTab[16] = { | ||
631 : | MLT( 0), MLT( 1), MLT( 2), MLT( 3), MLT( 4), MLT( 5), MLT( 6), MLT(7), | ||
632 : | MLT( 8), MLT( 9), MLT(10), MLT(11), MLT(12), MLT(13), MLT(14), MLT(15) | ||
633 : | }; | ||
634 : | #undef MLT | ||
635 : | Isibaar | 3 | |
636 : | edgomez | 851 | VECTOR generate_GMCimageMB( const GMC_DATA *const gmc_data, |
637 : | syskin | 935 | const IMAGE *const pRef, |
638 : | const int mi, const int mj, | ||
639 : | const int stride, | ||
640 : | const int stride2, | ||
641 : | const int quarterpel, | ||
642 : | const int rounding, | ||
643 : | IMAGE *const pGMC) | ||
644 : | edgomez | 851 | { |
645 : | syskin | 935 | const int W = gmc_data->W; |
646 : | const int H = gmc_data->H; | ||
647 : | edgomez | 851 | |
648 : | syskin | 935 | const int rho = gmc_data->rho; |
649 : | const int alpha = gmc_data->alpha; | ||
650 : | edgomez | 851 | |
651 : | syskin | 935 | const int rounder = ( 128 - (rounding<<(rho+rho)) ) << 16; |
652 : | edgomez | 851 | |
653 : | syskin | 935 | const int dxF = gmc_data->dxF; |
654 : | const int dyF = gmc_data->dyF; | ||
655 : | const int dxG = gmc_data->dxG; | ||
656 : | const int dyG = gmc_data->dyG; | ||
657 : | edgomez | 851 | |
658 : | syskin | 935 | uint8_t *dstY, *dstU, *dstV; |
659 : | edgomez | 851 | |
660 : | syskin | 935 | int I,J; |
661 : | VECTOR avgMV = {0,0}; | ||
662 : | edgomez | 851 | |
663 : | syskin | 935 | int32_t Fj, Gj; |
664 : | edgomez | 851 | |
665 : | syskin | 935 | dstY = &pGMC->y[(mj*16)*stride+mi*16] + 16; |
666 : | edgomez | 851 | |
667 : | syskin | 935 | Fj = gmc_data->Fo + dyF*mj*16 + dxF*mi*16; |
668 : | Gj = gmc_data->Go + dyG*mj*16 + dxG*mi*16; | ||
669 : | |||
670 : | for (J = 16; J > 0; --J) { | ||
671 : | int32_t Fi, Gi; | ||
672 : | edgomez | 851 | |
673 : | syskin | 935 | Fi = Fj; Fj += dyF; |
674 : | Gi = Gj; Gj += dyG; | ||
675 : | for (I = -16; I < 0; ++I) { | ||
676 : | int32_t F, G; | ||
677 : | uint32_t ri, rj; | ||
678 : | edgomez | 851 | |
679 : | syskin | 935 | F = ( Fi >> (alpha+rho) ) << rho; Fi += dxF; |
680 : | G = ( Gi >> (alpha+rho) ) << rho; Gi += dxG; | ||
681 : | edgomez | 851 | |
682 : | syskin | 935 | avgMV.x += F; |
683 : | avgMV.y += G; | ||
684 : | edgomez | 851 | |
685 : | syskin | 935 | ri = MTab[F&15]; |
686 : | rj = MTab[G&15]; | ||
687 : | edgomez | 851 | |
688 : | syskin | 935 | F >>= 4; |
689 : | G >>= 4; | ||
690 : | edgomez | 851 | |
691 : | syskin | 935 | if (F < -1) F = -1; |
692 : | else if (F > W) F = W; | ||
693 : | if (G< -1) G=-1; | ||
694 : | else if (G>H) G=H; | ||
695 : | edgomez | 851 | |
696 : | edgomez | 1053 | { /* MMX-like bilinear... */ |
697 : | syskin | 935 | const int offset = G*stride + F; |
698 : | uint32_t f0, f1; | ||
699 : | f0 = pRef->y[ offset +0 ]; | ||
700 : | f0 |= pRef->y[ offset +1 ] << 16; | ||
701 : | f1 = pRef->y[ offset+stride +0 ]; | ||
702 : | f1 |= pRef->y[ offset+stride +1 ] << 16; | ||
703 : | f0 = (ri*f0)>>16; | ||
704 : | f1 = (ri*f1) & 0x0fff0000; | ||
705 : | f0 |= f1; | ||
706 : | f0 = ( rj*f0 + rounder ) >> 24; | ||
707 : | edgomez | 851 | |
708 : | syskin | 935 | dstY[I] = (uint8_t)f0; |
709 : | } | ||
710 : | } | ||
711 : | |||
712 : | dstY += stride; | ||
713 : | } | ||
714 : | edgomez | 851 | |
715 : | syskin | 935 | dstU = &pGMC->u[(mj*8)*stride2+mi*8] + 8; |
716 : | dstV = &pGMC->v[(mj*8)*stride2+mi*8] + 8; | ||
717 : | edgomez | 851 | |
718 : | syskin | 935 | Fj = gmc_data->cFo + dyF*4 *mj*8 + dxF*4 *mi*8; |
719 : | Gj = gmc_data->cGo + dyG*4 *mj*8 + dxG*4 *mi*8; | ||
720 : | |||
721 : | for (J = 8; J > 0; --J) { | ||
722 : | int32_t Fi, Gi; | ||
723 : | Fi = Fj; Fj += 4*dyF; | ||
724 : | Gi = Gj; Gj += 4*dyG; | ||
725 : | edgomez | 851 | |
726 : | syskin | 935 | for (I = -8; I < 0; ++I) { |
727 : | int32_t F, G; | ||
728 : | uint32_t ri, rj; | ||
729 : | edgomez | 851 | |
730 : | syskin | 935 | F = ( Fi >> (alpha+rho+2) ) << rho; Fi += 4*dxF; |
731 : | G = ( Gi >> (alpha+rho+2) ) << rho; Gi += 4*dxG; | ||
732 : | edgomez | 851 | |
733 : | syskin | 935 | ri = MTab[F&15]; |
734 : | rj = MTab[G&15]; | ||
735 : | edgomez | 851 | |
736 : | syskin | 935 | F >>= 4; |
737 : | G >>= 4; | ||
738 : | edgomez | 851 | |
739 : | syskin | 935 | if (F < -1) F=-1; |
740 : | else if (F >= W/2) F = W/2; | ||
741 : | if (G < -1) G = -1; | ||
742 : | else if (G >= H/2) G = H/2; | ||
743 : | edgomez | 851 | |
744 : | syskin | 935 | { |
745 : | const int offset = G*stride2 + F; | ||
746 : | uint32_t f0, f1; | ||
747 : | edgomez | 851 | |
748 : | syskin | 935 | f0 = pRef->u[ offset +0 ]; |
749 : | f0 |= pRef->u[ offset +1 ] << 16; | ||
750 : | f1 = pRef->u[ offset+stride2 +0 ]; | ||
751 : | f1 |= pRef->u[ offset+stride2 +1 ] << 16; | ||
752 : | f0 = (ri*f0)>>16; | ||
753 : | f1 = (ri*f1) & 0x0fff0000; | ||
754 : | f0 |= f1; | ||
755 : | f0 = ( rj*f0 + rounder ) >> 24; | ||
756 : | edgomez | 851 | |
757 : | syskin | 935 | dstU[I] = (uint8_t)f0; |
758 : | edgomez | 851 | |
759 : | |||
760 : | syskin | 935 | f0 = pRef->v[ offset +0 ]; |
761 : | f0 |= pRef->v[ offset +1 ] << 16; | ||
762 : | f1 = pRef->v[ offset+stride2 +0 ]; | ||
763 : | f1 |= pRef->v[ offset+stride2 +1 ] << 16; | ||
764 : | f0 = (ri*f0)>>16; | ||
765 : | f1 = (ri*f1) & 0x0fff0000; | ||
766 : | f0 |= f1; | ||
767 : | f0 = ( rj*f0 + rounder ) >> 24; | ||
768 : | edgomez | 851 | |
769 : | syskin | 935 | dstV[I] = (uint8_t)f0; |
770 : | } | ||
771 : | } | ||
772 : | dstU += stride2; | ||
773 : | dstV += stride2; | ||
774 : | } | ||
775 : | edgomez | 851 | |
776 : | |||
777 : | edgomez | 1053 | avgMV.x -= 16*((256*mi+120)<<4); /* 120 = 15*16/2 */ |
778 : | syskin | 935 | avgMV.y -= 16*((256*mj+120)<<4); |
779 : | |||
780 : | avgMV.x = RSHIFT( avgMV.x, (4+7-quarterpel) ); | ||
781 : | avgMV.y = RSHIFT( avgMV.y, (4+7-quarterpel) ); | ||
782 : | |||
783 : | return avgMV; | ||
784 : | edgomez | 851 | } |
785 : | syskin | 935 | |
786 : | |||
787 : | |||
788 : | edgomez | 851 | #ifdef OLD_GRUEL_GMC |
789 : | syskin | 935 | void |
790 : | edgomez | 1053 | generate_GMCparameters( const int num_wp, /* [input]: number of warppoints */ |
791 : | const int res, /* [input]: resolution */ | ||
792 : | const WARPPOINTS *const warp, /* [input]: warp points */ | ||
793 : | edgomez | 851 | const int width, const int height, |
794 : | edgomez | 1053 | GMC_DATA *const gmc) /* [output] precalculated parameters */ |
795 : | edgomez | 851 | { |
796 : | |||
797 : | syskin | 935 | /* We follow mainly two sources: The original standard, which is ugly, and the |
798 : | thesis from Andreas Dehnhardt, which is much nicer. | ||
799 : | edgomez | 851 | |
800 : | syskin | 935 | Notation is: indices are written next to the variable, |
801 : | primes in the standard are denoted by a suffix 'p'. | ||
802 : | types are "c"=constant, "i"=input parameter, "f"=calculated, then fixed, | ||
803 : | "o"=output data, " "=other, "u" = unused, "p"=calc for every pixel | ||
804 : | |||
805 : | type | variable name | ISO name (TeX-style) | value or range | usage | ||
806 : | edgomez | 851 | ------------------------------------------------------------------------------------- |
807 : | c | H | H | [16 , ?] | image width (w/o edges) | ||
808 : | c | W | W | [16 , ?] | image height (w/o edges) | ||
809 : | |||
810 : | c | i0 | i_0 | 0 | ref. point #1, X | ||
811 : | c | j0 | j_0 | 0 | ref. point #1, Y | ||
812 : | c | i1 | i_1 | W | ref. point #2, X | ||
813 : | c | j1 | j_1 | 0 | ref. point #2, Y | ||
814 : | cu | i2 | i_2 | 0 | ref. point #3, X | ||
815 : | syskin | 935 | cu | i2 | j_2 | H | ref. point #3, Y |
816 : | edgomez | 851 | |
817 : | i | du0 | du[0] | [-16863,16863] | warp vector #1, Y | ||
818 : | i | dv0 | dv[0] | [-16863,16863] | warp vector #1, Y | ||
819 : | i | du1 | du[1] | [-16863,16863] | warp vector #2, Y | ||
820 : | i | dv1 | dv[1] | [-16863,16863] | warp vector #2, Y | ||
821 : | syskin | 935 | iu | du2 | du[2] | [-16863,16863] | warp vector #3, Y |
822 : | edgomez | 851 | iu | dv2 | dv[2] | [-16863,16863] | warp vector #3, Y |
823 : | |||
824 : | syskin | 935 | i | s | s | {2,4,8,16} | interpol. resolution |
825 : | f | sigma | - | log2(s) | X / s == X >> sigma | ||
826 : | f | r | r | =16/s | complementary res. | ||
827 : | f | rho | \rho | log2(r) | X / r == X >> rho | ||
828 : | |||
829 : | f | i0s | i'_0 | | | ||
830 : | f | j0s | j'_0 | | | ||
831 : | f | i1s | i'_1 | | | ||
832 : | f | j1s | j'_1 | | | ||
833 : | f | i2s | i'_2 | | | ||
834 : | f | j2s | j'_2 | | | ||
835 : | |||
836 : | f | alpha | \alpha | | 2^{alpha-1} < W <= 2^alpha | ||
837 : | f | beta | \beta | | 2^{beta-1} < H <= 2^beta | ||
838 : | |||
839 : | f | Ws | W' | W = 2^{alpha} | scaled width | ||
840 : | f | Hs | H' | W = 2^{beta} | scaled height | ||
841 : | |||
842 : | f | i1ss | i''_1 | "virtual sprite stuff" | ||
843 : | f | j1ss | j''_1 | "virtual sprite stuff" | ||
844 : | f | i2ss | i''_2 | "virtual sprite stuff" | ||
845 : | f | j2ss | j''_2 | "virtual sprite stuff" | ||
846 : | edgomez | 851 | */ |
847 : | |||
848 : | /* Some calculations are disabled because we only use 2 warppoints at the moment */ | ||
849 : | |||
850 : | int du0 = warp->duv[0].x; | ||
851 : | int dv0 = warp->duv[0].y; | ||
852 : | int du1 = warp->duv[1].x; | ||
853 : | int dv1 = warp->duv[1].y; | ||
854 : | edgomez | 1053 | #if 0 |
855 : | int du2 = warp->duv[2].x; | ||
856 : | int dv2 = warp->duv[2].y; | ||
857 : | #endif | ||
858 : | syskin | 935 | |
859 : | edgomez | 851 | gmc->num_wp = num_wp; |
860 : | |||
861 : | gmc->s = res; /* scaling parameters 2,4,8 or 16 */ | ||
862 : | gmc->sigma = log2bin(res-1); /* log2bin(15)=4, log2bin(16)=5, log2bin(17)=5 */ | ||
863 : | syskin | 935 | gmc->r = 16/res; |
864 : | edgomez | 851 | gmc->rho = 4 - gmc->sigma; /* = log2bin(r-1) */ |
865 : | |||
866 : | gmc->W = width; | ||
867 : | gmc->H = height; /* fixed reference coordinates */ | ||
868 : | |||
869 : | gmc->alpha = log2bin(gmc->W-1); | ||
870 : | syskin | 935 | gmc->Ws= 1<<gmc->alpha; |
871 : | edgomez | 851 | |
872 : | edgomez | 1053 | #if 0 |
873 : | gmc->beta = log2bin(gmc->H-1); | ||
874 : | gmc->Hs= 1<<gmc->beta; | ||
875 : | #endif | ||
876 : | edgomez | 851 | |
877 : | edgomez | 1053 | #if 0 |
878 : | printf("du0=%d dv0=%d du1=%d dv1=%d s=%d sigma=%d W=%d alpha=%d, Ws=%d, rho=%d\n",du0,dv0,du1,dv1,gmc->s,gmc->sigma,gmc->W,gmc->alpha,gmc->Ws,gmc->rho); | ||
879 : | #endif | ||
880 : | edgomez | 851 | |
881 : | edgomez | 1053 | /* |
882 : | * i2s is only needed for num_wp >= 3, etc. | ||
883 : | * the 's' values are in 1/s pel resolution | ||
884 : | */ | ||
885 : | edgomez | 851 | gmc->i0s = res/2 * ( du0 ); |
886 : | gmc->j0s = res/2 * ( dv0 ); | ||
887 : | gmc->i1s = res/2 * (2*width + du1 + du0 ); | ||
888 : | gmc->j1s = res/2 * ( dv1 + dv0 ); | ||
889 : | edgomez | 1053 | #if 0 |
890 : | gmc->i2s = res/2 * ( du2 + du0 ); | ||
891 : | gmc->j2s = res/2 * (2*height + dv2 + dv0 ); | ||
892 : | #endif | ||
893 : | syskin | 935 | |
894 : | edgomez | 851 | /* i2s and i2ss are only needed for num_wp == 3, etc. */ |
895 : | |||
896 : | /* the 'ss' values are in 1/16 pel resolution */ | ||
897 : | syskin | 935 | gmc->i1ss = 16*gmc->Ws + ROUNDED_DIV(((gmc->W-gmc->Ws)*(gmc->r*gmc->i0s) + gmc->Ws*(gmc->r*gmc->i1s - 16*gmc->W)),gmc->W); |
898 : | gmc->j1ss = ROUNDED_DIV( ((gmc->W - gmc->Ws)*(gmc->r*gmc->j0s) + gmc->Ws*gmc->r*gmc->j1s) ,gmc->W ); | ||
899 : | edgomez | 851 | |
900 : | edgomez | 1053 | #if 0 |
901 : | gmc->i2ss = ROUNDED_DIV( ((gmc->H - gmc->Hs)*(gmc->r*gmc->i0s) + gmc->Hs*(gmc->r*gmc->i2s)), gmc->H); | ||
902 : | gmc->j2ss = 16*gmc->Hs + ROUNDED_DIV( ((gmc->H-gmc->Hs)*(gmc->r*gmc->j0s) + gmc->Ws*(gmc->r*gmc->j2s - 16*gmc->H)), gmc->H); | ||
903 : | #endif | ||
904 : | edgomez | 851 | |
905 : | syskin | 935 | return; |
906 : | edgomez | 851 | } |
907 : | |||
908 : | syskin | 935 | void |
909 : | edgomez | 1053 | generate_GMCimage( const GMC_DATA *const gmc_data, /* [input] precalculated data */ |
910 : | const IMAGE *const pRef, /* [input] */ | ||
911 : | syskin | 935 | const int mb_width, |
912 : | edgomez | 851 | const int mb_height, |
913 : | const int stride, | ||
914 : | syskin | 935 | const int stride2, |
915 : | edgomez | 1053 | const int fcode, /* [input] some parameters... */ |
916 : | const int32_t quarterpel, /* [input] for rounding avgMV */ | ||
917 : | const int reduced_resolution, /* [input] ignored */ | ||
918 : | const int32_t rounding, /* [input] for rounding image data */ | ||
919 : | MACROBLOCK *const pMBs, /* [output] average motion vectors */ | ||
920 : | IMAGE *const pGMC) /* [output] full warped image */ | ||
921 : | edgomez | 851 | { |
922 : | |||
923 : | unsigned int mj,mi; | ||
924 : | VECTOR avgMV; | ||
925 : | syskin | 935 | |
926 : | for (mj = 0;mj < mb_height; mj++) | ||
927 : | for (mi = 0;mi < mb_width; mi++) { | ||
928 : | |||
929 : | avgMV = generate_GMCimageMB(gmc_data, pRef, mi, mj, | ||
930 : | edgomez | 851 | stride, stride2, quarterpel, rounding, pGMC); |
931 : | |||
932 : | pMBs[mj*mb_width+mi].amv.x = gmc_sanitize(avgMV.x, quarterpel, fcode); | ||
933 : | pMBs[mj*mb_width+mi].amv.y = gmc_sanitize(avgMV.y, quarterpel, fcode); | ||
934 : | pMBs[mj*mb_width+mi].mcsel = 0; /* until mode decision */ | ||
935 : | Isibaar | 3 | } |
936 : | } | ||
937 : | edgomez | 851 | |
938 : | |||
939 : | VECTOR generate_GMCimageMB( const GMC_DATA *const gmc_data, /* [input] all precalc data */ | ||
940 : | const IMAGE *const pRef, /* [input] */ | ||
941 : | const int mi, const int mj, /* [input] MB position */ | ||
942 : | const int stride, /* [input] Lumi stride */ | ||
943 : | const int stride2, /* [input] chroma stride */ | ||
944 : | const int quarterpel, /* [input] for rounding of avgMV */ | ||
945 : | const int rounding, /* [input] for rounding of imgae data */ | ||
946 : | IMAGE *const pGMC) /* [outut] generate image */ | ||
947 : | syskin | 935 | |
948 : | edgomez | 851 | /* |
949 : | syskin | 935 | type | variable name | ISO name (TeX-style) | value or range | usage |
950 : | edgomez | 851 | ------------------------------------------------------------------------------------- |
951 : | syskin | 935 | p | F | F(i,j) | | pelwise motion vector X in s-th pel |
952 : | p | G | G(i,j) | | pelwise motion vector Y in s-th pel | ||
953 : | p | Fc | F_c(i,j) | | | ||
954 : | p | Gc | G_c(i,j) | | same for chroma | ||
955 : | edgomez | 851 | |
956 : | syskin | 935 | p | Y00 | Y_{00} | [0,255*s*s] | first: 4 neighbouring Y-values |
957 : | p | Y01 | Y_{01} | [0,255] | at fullpel position, around the | ||
958 : | p | Y10 | Y_{10} | [0,255*s] | position where pelweise MV points to | ||
959 : | p | Y11 | Y_{11} | [0,255] | later: bilinear interpol Y-values in Y00 | ||
960 : | edgomez | 851 | |
961 : | syskin | 935 | p | C00 | C_{00} | [0,255*s*s] | same for chroma Cb and Cr |
962 : | p | C01 | C_{01} | [0,255] | | ||
963 : | p | C10 | C_{10} | [0,255*s] | | ||
964 : | p | C11 | C_{11} | [0,255] | | ||
965 : | |||
966 : | edgomez | 851 | */ |
967 : | { | ||
968 : | const int W = gmc_data->W; | ||
969 : | const int H = gmc_data->H; | ||
970 : | |||
971 : | const int s = gmc_data->s; | ||
972 : | const int sigma = gmc_data->sigma; | ||
973 : | syskin | 935 | |
974 : | edgomez | 851 | const int r = gmc_data->r; |
975 : | syskin | 935 | const int rho = gmc_data->rho; |
976 : | |||
977 : | edgomez | 851 | const int i0s = gmc_data->i0s; |
978 : | const int j0s = gmc_data->j0s; | ||
979 : | syskin | 935 | |
980 : | edgomez | 851 | const int i1ss = gmc_data->i1ss; |
981 : | syskin | 935 | const int j1ss = gmc_data->j1ss; |
982 : | edgomez | 1053 | #if 0 |
983 : | const int i2ss = gmc_data->i2ss; | ||
984 : | const int j2ss = gmc_data->j2ss; | ||
985 : | #endif | ||
986 : | edgomez | 851 | |
987 : | syskin | 935 | const int alpha = gmc_data->alpha; |
988 : | const int Ws = gmc_data->Ws; | ||
989 : | edgomez | 851 | |
990 : | edgomez | 1053 | #if 0 |
991 : | const int beta = gmc_data->beta; | ||
992 : | const int Hs = gmc_data->Hs; | ||
993 : | #endif | ||
994 : | edgomez | 851 | |
995 : | int I,J; | ||
996 : | VECTOR avgMV = {0,0}; | ||
997 : | syskin | 935 | |
998 : | edgomez | 851 | for (J=16*mj;J<16*(mj+1);J++) |
999 : | for (I=16*mi;I<16*(mi+1);I++) | ||
1000 : | { | ||
1001 : | int F= i0s + ( ((-r*i0s+i1ss)*I + (r*j0s-j1ss)*J + (1<<(alpha+rho-1))) >> (alpha+rho) ); | ||
1002 : | int G= j0s + ( ((-r*j0s+j1ss)*I + (-r*i0s+i1ss)*J + (1<<(alpha+rho-1))) >> (alpha+rho) ); | ||
1003 : | |||
1004 : | syskin | 935 | /* this naive implementation (with lots of multiplications) isn't slower (rather faster) than |
1005 : | edgomez | 851 | working incremental. Don't ask me why... maybe the whole this is memory bound? */ |
1006 : | |||
1007 : | edgomez | 1053 | const int ri= F & (s-1); /* fractional part of pelwise MV X */ |
1008 : | const int rj= G & (s-1); /* fractional part of pelwise MV Y */ | ||
1009 : | syskin | 935 | |
1010 : | edgomez | 851 | int Y00,Y01,Y10,Y11; |
1011 : | |||
1012 : | /* unclipped values are used for avgMV */ | ||
1013 : | avgMV.x += F-(I<<sigma); /* shift position to 1/s-pel, as the MV is */ | ||
1014 : | avgMV.y += G-(J<<sigma); /* TODO: don't do this (of course) */ | ||
1015 : | |||
1016 : | F >>= sigma; | ||
1017 : | G >>= sigma; | ||
1018 : | |||
1019 : | syskin | 935 | /* clip values to be in range. Since we have edges, clip to 1 less than lower boundary |
1020 : | edgomez | 851 | this way positions F+1/G+1 are still right */ |
1021 : | |||
1022 : | if (F< -1) | ||
1023 : | F=-1; | ||
1024 : | else if (F>W) | ||
1025 : | F=W; /* W or W-1 doesn't matter, so save 1 subtract ;-) */ | ||
1026 : | if (G< -1) | ||
1027 : | G=-1; | ||
1028 : | else if (G>H) | ||
1029 : | G=H; /* dito */ | ||
1030 : | |||
1031 : | edgomez | 1053 | Y00 = pRef->y[ G*stride + F ]; /* Lumi values */ |
1032 : | edgomez | 851 | Y01 = pRef->y[ G*stride + F+1 ]; |
1033 : | Y10 = pRef->y[ G*stride + F+stride ]; | ||
1034 : | Y11 = pRef->y[ G*stride + F+stride+1 ]; | ||
1035 : | syskin | 935 | |
1036 : | edgomez | 851 | /* bilinear interpolation */ |
1037 : | Y00 = ((s-ri)*Y00 + ri*Y01); | ||
1038 : | Y10 = ((s-ri)*Y10 + ri*Y11); | ||
1039 : | syskin | 935 | Y00 = ((s-rj)*Y00 + rj*Y10 + s*s/2 - rounding ) >> (sigma+sigma); |
1040 : | |||
1041 : | edgomez | 851 | pGMC->y[J*stride+I] = (uint8_t)Y00; /* output 1 Y-pixel */ |
1042 : | } | ||
1043 : | |||
1044 : | syskin | 935 | |
1045 : | /* doing chroma _here_ is even more stupid and slow, because won't be used until Compensation and | ||
1046 : | edgomez | 851 | most likely not even then (only if the block really _is_ GMC) |
1047 : | */ | ||
1048 : | |||
1049 : | for (J=8*mj;J<8*(mj+1);J++) /* this plays the role of j_c,i_c in the standard */ | ||
1050 : | for (I=8*mi;I<8*(mi+1);I++) /* For I_c we have to use I_c = 4*i_c+1 ! */ | ||
1051 : | { | ||
1052 : | /* same positions for both chroma components, U=Cb and V=Cr */ | ||
1053 : | syskin | 935 | int Fc=((-r*i0s+i1ss)*(4*I+1) + (r*j0s-j1ss)*(4*J+1) +2*Ws*r*i0s |
1054 : | edgomez | 851 | -16*Ws +(1<<(alpha+rho+1)))>>(alpha+rho+2); |
1055 : | syskin | 935 | int Gc=((-r*j0s+j1ss)*(4*I+1) +(-r*i0s+i1ss)*(4*J+1) +2*Ws*r*j0s |
1056 : | edgomez | 851 | -16*Ws +(1<<(alpha+rho+1))) >>(alpha+rho+2); |
1057 : | syskin | 935 | |
1058 : | edgomez | 1053 | const int ri= Fc & (s-1); /* fractional part of pelwise MV X */ |
1059 : | const int rj= Gc & (s-1); /* fractional part of pelwise MV Y */ | ||
1060 : | edgomez | 851 | |
1061 : | int C00,C01,C10,C11; | ||
1062 : | syskin | 935 | |
1063 : | edgomez | 851 | Fc >>= sigma; |
1064 : | Gc >>= sigma; | ||
1065 : | syskin | 935 | |
1066 : | edgomez | 851 | if (Fc< -1) |
1067 : | Fc=-1; | ||
1068 : | else if (Fc>=W/2) | ||
1069 : | Fc=W/2; /* W or W-1 doesn't matter, so save 1 subtraction ;-) */ | ||
1070 : | if (Gc< -1) | ||
1071 : | Gc=-1; | ||
1072 : | else if (Gc>=H/2) | ||
1073 : | Gc=H/2; /* dito */ | ||
1074 : | |||
1075 : | syskin | 935 | /* now calculate U data */ |
1076 : | edgomez | 1053 | C00 = pRef->u[ Gc*stride2 + Fc ]; /* chroma-value Cb */ |
1077 : | edgomez | 851 | C01 = pRef->u[ Gc*stride2 + Fc+1 ]; |
1078 : | C10 = pRef->u[ (Gc+1)*stride2 + Fc ]; | ||
1079 : | C11 = pRef->u[ (Gc+1)*stride2 + Fc+1 ]; | ||
1080 : | syskin | 935 | |
1081 : | edgomez | 851 | /* bilinear interpolation */ |
1082 : | C00 = ((s-ri)*C00 + ri*C01); | ||
1083 : | C10 = ((s-ri)*C10 + ri*C11); | ||
1084 : | syskin | 935 | C00 = ((s-rj)*C00 + rj*C10 + s*s/2 - rounding ) >> (sigma+sigma); |
1085 : | |||
1086 : | edgomez | 851 | pGMC->u[J*stride2+I] = (uint8_t)C00; /* output 1 U-pixel */ |
1087 : | syskin | 935 | |
1088 : | edgomez | 851 | /* now calculate V data */ |
1089 : | edgomez | 1053 | C00 = pRef->v[ Gc*stride2 + Fc ]; /* chroma-value Cr */ |
1090 : | edgomez | 851 | C01 = pRef->v[ Gc*stride2 + Fc+1 ]; |
1091 : | C10 = pRef->v[ (Gc+1)*stride2 + Fc ]; | ||
1092 : | C11 = pRef->v[ (Gc+1)*stride2 + Fc+1 ]; | ||
1093 : | syskin | 935 | |
1094 : | edgomez | 851 | /* bilinear interpolation */ |
1095 : | C00 = ((s-ri)*C00 + ri*C01); | ||
1096 : | C10 = ((s-ri)*C10 + ri*C11); | ||
1097 : | syskin | 935 | C00 = ((s-rj)*C00 + rj*C10 + s*s/2 - rounding ) >> (sigma+sigma); |
1098 : | |||
1099 : | edgomez | 851 | pGMC->v[J*stride2+I] = (uint8_t)C00; /* output 1 V-pixel */ |
1100 : | } | ||
1101 : | syskin | 935 | |
1102 : | edgomez | 1053 | /* The average vector is rounded from 1/s-pel to 1/2 or 1/4 using the '//' operator */ |
1103 : | edgomez | 851 | |
1104 : | avgMV.x = RSHIFT( avgMV.x, (sigma+7-quarterpel) ); | ||
1105 : | avgMV.y = RSHIFT( avgMV.y, (sigma+7-quarterpel) ); | ||
1106 : | |||
1107 : | /* ^^^^ this is the way MS Reference Software does it */ | ||
1108 : | syskin | 935 | |
1109 : | edgomez | 851 | return avgMV; /* clipping to fcode area is done outside! */ |
1110 : | } | ||
1111 : | |||
1112 : | #endif |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |