Parent Directory | Revision Log
Revision 1.18.2.5 - (view) (download)
1 : | edgomez | 1.18.2.5 | /* |
2 : | * 30.10.2002 corrected qpel chroma rounding | ||
3 : | * 04.10.2002 added qpel support to MBMotionCompensation | ||
4 : | * 01.05.2002 updated MBMotionCompensationBVOP | ||
5 : | * 14.04.2002 bframe compensation | ||
6 : | */ | ||
7 : | edgomez | 1.17 | |
8 : | #include <stdio.h> | ||
9 : | suxen_drol | 1.3 | |
10 : | Isibaar | 1.1 | #include "../encoder.h" |
11 : | #include "../utils/mbfunctions.h" | ||
12 : | #include "../image/interpolate8x8.h" | ||
13 : | edgomez | 1.17 | #include "../image/reduced.h" |
14 : | Isibaar | 1.1 | #include "../utils/timer.h" |
15 : | suxen_drol | 1.3 | #include "motion.h" |
16 : | Isibaar | 1.1 | |
17 : | edgomez | 1.17 | #ifndef RSHIFT |
18 : | #define RSHIFT(a,b) ((a) > 0 ? ((a) + (1<<((b)-1)))>>(b) : ((a) + (1<<((b)-1))-1)>>(b)) | ||
19 : | #endif | ||
20 : | |||
21 : | /* assume b>0 */ | ||
22 : | #ifndef RDIV | ||
23 : | #define RDIV(a,b) (((a)>0 ? (a) + ((b)>>1) : (a) - ((b)>>1))/(b)) | ||
24 : | #endif | ||
25 : | |||
26 : | |||
27 : | syskin | 1.18.2.2 | /* This is borrowed from decoder.c */ |
28 : | edgomez | 1.17 | static __inline int gmc_sanitize(int value, int quarterpel, int fcode) |
29 : | { | ||
30 : | int length = 1 << (fcode+4); | ||
31 : | |||
32 : | edgomez | 1.18.2.5 | #if 0 |
33 : | if (quarterpel) value *= 2; | ||
34 : | #endif | ||
35 : | edgomez | 1.17 | |
36 : | syskin | 1.18.2.2 | if (value < -length) |
37 : | edgomez | 1.17 | return -length; |
38 : | syskin | 1.18.2.2 | else if (value >= length) |
39 : | edgomez | 1.17 | return length-1; |
40 : | else return value; | ||
41 : | } | ||
42 : | |||
43 : | /* And this is borrowed from bitstream.c until we find a common solution */ | ||
44 : | |||
45 : | static uint32_t __inline | ||
46 : | log2bin(uint32_t value) | ||
47 : | { | ||
48 : | /* Changed by Chenm001 */ | ||
49 : | #if !defined(_MSC_VER) | ||
50 : | int n = 0; | ||
51 : | |||
52 : | while (value) { | ||
53 : | value >>= 1; | ||
54 : | n++; | ||
55 : | } | ||
56 : | return n; | ||
57 : | #else | ||
58 : | __asm { | ||
59 : | syskin | 1.18.2.2 | bsr eax, value |
60 : | edgomez | 1.17 | inc eax |
61 : | } | ||
62 : | #endif | ||
63 : | } | ||
64 : | |||
65 : | Isibaar | 1.1 | |
66 : | edgomez | 1.5 | static __inline void |
67 : | edgomez | 1.17 | compensate16x16_interpolate(int16_t * const dct_codes, |
68 : | uint8_t * const cur, | ||
69 : | const uint8_t * const ref, | ||
70 : | const uint8_t * const refh, | ||
71 : | const uint8_t * const refv, | ||
72 : | const uint8_t * const refhv, | ||
73 : | uint8_t * const tmp, | ||
74 : | uint32_t x, | ||
75 : | uint32_t y, | ||
76 : | const int32_t dx, | ||
77 : | const int32_t dy, | ||
78 : | const int32_t stride, | ||
79 : | const int quarterpel, | ||
80 : | const int reduced_resolution, | ||
81 : | const int32_t rounding) | ||
82 : | Isibaar | 1.1 | { |
83 : | edgomez | 1.17 | const uint8_t * ptr; |
84 : | |||
85 : | if (!reduced_resolution) { | ||
86 : | |||
87 : | if(quarterpel) { | ||
88 : | if ((dx&3) | (dy&3)) { | ||
89 : | interpolate16x16_quarterpel(tmp - y * stride - x, | ||
90 : | (uint8_t *) ref, tmp + 32, | ||
91 : | tmp + 64, tmp + 96, x, y, dx, dy, stride, rounding); | ||
92 : | ptr = tmp; | ||
93 : | edgomez | 1.18.2.5 | } else ptr = ref + (y + dy/4)*stride + x + dx/4; /* fullpixel position */ |
94 : | edgomez | 1.17 | |
95 : | } else ptr = get_ref(ref, refh, refv, refhv, x, y, 1, dx, dy, stride); | ||
96 : | Isibaar | 1.1 | |
97 : | edgomez | 1.5 | transfer_8to16sub(dct_codes, cur + y * stride + x, |
98 : | syskin | 1.18.2.2 | ptr, stride); |
99 : | edgomez | 1.17 | transfer_8to16sub(dct_codes+64, cur + y * stride + x + 8, |
100 : | syskin | 1.18.2.2 | ptr + 8, stride); |
101 : | edgomez | 1.17 | transfer_8to16sub(dct_codes+128, cur + y * stride + x + 8*stride, |
102 : | syskin | 1.18.2.2 | ptr + 8*stride, stride); |
103 : | edgomez | 1.17 | transfer_8to16sub(dct_codes+192, cur + y * stride + x + 8*stride+8, |
104 : | syskin | 1.18.2.2 | ptr + 8*stride + 8, stride); |
105 : | edgomez | 1.17 | |
106 : | edgomez | 1.18.2.5 | } else { /* reduced_resolution */ |
107 : | syskin | 1.18.2.2 | |
108 : | edgomez | 1.17 | x *= 2; y *= 2; |
109 : | |||
110 : | ptr = get_ref(ref, refh, refv, refhv, x, y, 1, dx, dy, stride); | ||
111 : | syskin | 1.18.2.2 | |
112 : | edgomez | 1.17 | filter_18x18_to_8x8(dct_codes, cur+y*stride + x, stride); |
113 : | filter_diff_18x18_to_8x8(dct_codes, ptr, stride); | ||
114 : | |||
115 : | filter_18x18_to_8x8(dct_codes+64, cur+y*stride + x + 16, stride); | ||
116 : | filter_diff_18x18_to_8x8(dct_codes+64, ptr + 16, stride); | ||
117 : | |||
118 : | filter_18x18_to_8x8(dct_codes+128, cur+(y+16)*stride + x, stride); | ||
119 : | filter_diff_18x18_to_8x8(dct_codes+128, ptr + 16*stride, stride); | ||
120 : | |||
121 : | filter_18x18_to_8x8(dct_codes+192, cur+(y+16)*stride + x + 16, stride); | ||
122 : | filter_diff_18x18_to_8x8(dct_codes+192, ptr + 16*stride + 16, stride); | ||
123 : | |||
124 : | transfer32x32_copy(cur + y*stride + x, ptr, stride); | ||
125 : | } | ||
126 : | } | ||
127 : | |||
128 : | static __inline void | ||
129 : | compensate8x8_interpolate( int16_t * const dct_codes, | ||
130 : | uint8_t * const cur, | ||
131 : | const uint8_t * const ref, | ||
132 : | const uint8_t * const refh, | ||
133 : | const uint8_t * const refv, | ||
134 : | const uint8_t * const refhv, | ||
135 : | uint8_t * const tmp, | ||
136 : | uint32_t x, | ||
137 : | uint32_t y, | ||
138 : | const int32_t dx, | ||
139 : | const int32_t dy, | ||
140 : | const int32_t stride, | ||
141 : | const int32_t quarterpel, | ||
142 : | const int reduced_resolution, | ||
143 : | const int32_t rounding) | ||
144 : | { | ||
145 : | const uint8_t * ptr; | ||
146 : | |||
147 : | if (!reduced_resolution) { | ||
148 : | |||
149 : | if(quarterpel) { | ||
150 : | if ((dx&3) | (dy&3)) { | ||
151 : | interpolate8x8_quarterpel(tmp - y*stride - x, | ||
152 : | (uint8_t *) ref, tmp + 32, | ||
153 : | tmp + 64, tmp + 96, x, y, dx, dy, stride, rounding); | ||
154 : | ptr = tmp; | ||
155 : | edgomez | 1.18.2.5 | } else ptr = ref + (y + dy/4)*stride + x + dx/4; /* fullpixel position */ |
156 : | edgomez | 1.17 | } else ptr = get_ref(ref, refh, refv, refhv, x, y, 1, dx, dy, stride); |
157 : | |||
158 : | transfer_8to16sub(dct_codes, cur + y * stride + x, ptr, stride); | ||
159 : | |||
160 : | edgomez | 1.18.2.5 | } else { /* reduced_resolution */ |
161 : | Isibaar | 1.1 | |
162 : | edgomez | 1.17 | x *= 2; y *= 2; |
163 : | Isibaar | 1.1 | |
164 : | edgomez | 1.17 | ptr = get_ref(ref, refh, refv, refhv, x, y, 1, dx, dy, stride); |
165 : | Isibaar | 1.1 | |
166 : | edgomez | 1.17 | filter_18x18_to_8x8(dct_codes, cur+y*stride + x, stride); |
167 : | filter_diff_18x18_to_8x8(dct_codes, ptr, stride); | ||
168 : | syskin | 1.18.2.2 | |
169 : | edgomez | 1.17 | transfer16x16_copy(cur + y*stride + x, ptr, stride); |
170 : | edgomez | 1.5 | } |
171 : | Isibaar | 1.1 | } |
172 : | |||
173 : | edgomez | 1.17 | /* XXX: slow, inelegant... */ |
174 : | static void | ||
175 : | interpolate18x18_switch(uint8_t * const cur, | ||
176 : | const uint8_t * const refn, | ||
177 : | const uint32_t x, | ||
178 : | const uint32_t y, | ||
179 : | const int32_t dx, | ||
180 : | const int dy, | ||
181 : | const int32_t stride, | ||
182 : | const int32_t rounding) | ||
183 : | { | ||
184 : | interpolate8x8_switch(cur, refn, x-1, y-1, dx, dy, stride, rounding); | ||
185 : | interpolate8x8_switch(cur, refn, x+7, y-1, dx, dy, stride, rounding); | ||
186 : | interpolate8x8_switch(cur, refn, x+9, y-1, dx, dy, stride, rounding); | ||
187 : | |||
188 : | interpolate8x8_switch(cur, refn, x-1, y+7, dx, dy, stride, rounding); | ||
189 : | interpolate8x8_switch(cur, refn, x+7, y+7, dx, dy, stride, rounding); | ||
190 : | interpolate8x8_switch(cur, refn, x+9, y+7, dx, dy, stride, rounding); | ||
191 : | |||
192 : | interpolate8x8_switch(cur, refn, x-1, y+9, dx, dy, stride, rounding); | ||
193 : | interpolate8x8_switch(cur, refn, x+7, y+9, dx, dy, stride, rounding); | ||
194 : | interpolate8x8_switch(cur, refn, x+9, y+9, dx, dy, stride, rounding); | ||
195 : | } | ||
196 : | |||
197 : | static void | ||
198 : | CompensateChroma( int dx, int dy, | ||
199 : | const int i, const int j, | ||
200 : | IMAGE * const Cur, | ||
201 : | const IMAGE * const Ref, | ||
202 : | uint8_t * const temp, | ||
203 : | int16_t * const coeff, | ||
204 : | const int32_t stride, | ||
205 : | const int rounding, | ||
206 : | const int rrv) | ||
207 : | { /* uv-block-based compensation */ | ||
208 : | |||
209 : | if (!rrv) { | ||
210 : | transfer_8to16sub(coeff, Cur->u + 8 * j * stride + 8 * i, | ||
211 : | syskin | 1.18.2.2 | interpolate8x8_switch2(temp, Ref->u, 8 * i, 8 * j, |
212 : | edgomez | 1.17 | dx, dy, stride, rounding), |
213 : | stride); | ||
214 : | transfer_8to16sub(coeff + 64, Cur->v + 8 * j * stride + 8 * i, | ||
215 : | syskin | 1.18.2.2 | interpolate8x8_switch2(temp, Ref->v, 8 * i, 8 * j, |
216 : | edgomez | 1.17 | dx, dy, stride, rounding), |
217 : | stride); | ||
218 : | } else { | ||
219 : | uint8_t * current, * reference; | ||
220 : | |||
221 : | current = Cur->u + 16*j*stride + 16*i; | ||
222 : | reference = temp - 16*j*stride - 16*i; | ||
223 : | interpolate18x18_switch(reference, Ref->u, 16*i, 16*j, dx, dy, stride, rounding); | ||
224 : | filter_18x18_to_8x8(coeff, current, stride); | ||
225 : | filter_diff_18x18_to_8x8(coeff, temp, stride); | ||
226 : | transfer16x16_copy(current, temp, stride); | ||
227 : | |||
228 : | current = Cur->v + 16*j*stride + 16*i; | ||
229 : | interpolate18x18_switch(reference, Ref->v, 16*i, 16*j, dx, dy, stride, rounding); | ||
230 : | filter_18x18_to_8x8(coeff + 64, current, stride); | ||
231 : | filter_diff_18x18_to_8x8(coeff + 64, temp, stride); | ||
232 : | transfer16x16_copy(current, temp, stride); | ||
233 : | } | ||
234 : | } | ||
235 : | Isibaar | 1.1 | |
236 : | edgomez | 1.5 | void |
237 : | MBMotionCompensation(MACROBLOCK * const mb, | ||
238 : | syskin | 1.18.2.2 | const uint32_t i, |
239 : | const uint32_t j, | ||
240 : | const IMAGE * const ref, | ||
241 : | const IMAGE * const refh, | ||
242 : | const IMAGE * const refv, | ||
243 : | const IMAGE * const refhv, | ||
244 : | const IMAGE * const refGMC, | ||
245 : | IMAGE * const cur, | ||
246 : | int16_t * dct_codes, | ||
247 : | const uint32_t width, | ||
248 : | const uint32_t height, | ||
249 : | const uint32_t edged_width, | ||
250 : | const int32_t quarterpel, | ||
251 : | const int reduced_resolution, | ||
252 : | const int32_t rounding) | ||
253 : | edgomez | 1.17 | { |
254 : | int32_t dx; | ||
255 : | int32_t dy; | ||
256 : | |||
257 : | uint8_t * const tmp = refv->u; | ||
258 : | |||
259 : | if ( (!reduced_resolution) && (mb->mode == MODE_NOT_CODED) ) { /* quick copy for early SKIP */ | ||
260 : | /* early SKIP is only activated in P-VOPs, not in S-VOPs, so mcsel can never be 1 */ | ||
261 : | |||
262 : | transfer16x16_copy(cur->y + 16 * (i + j * edged_width), | ||
263 : | syskin | 1.18.2.2 | ref->y + 16 * (i + j * edged_width), |
264 : | edged_width); | ||
265 : | |||
266 : | edgomez | 1.17 | transfer8x8_copy(cur->u + 8 * (i + j * edged_width/2), |
267 : | ref->u + 8 * (i + j * edged_width/2), | ||
268 : | edged_width / 2); | ||
269 : | transfer8x8_copy(cur->v + 8 * (i + j * edged_width/2), | ||
270 : | ref->v + 8 * (i + j * edged_width/2), | ||
271 : | edged_width / 2); | ||
272 : | return; | ||
273 : | } | ||
274 : | |||
275 : | syskin | 1.18.2.2 | if ((mb->mode == MODE_NOT_CODED || mb->mode == MODE_INTER |
276 : | edgomez | 1.17 | || mb->mode == MODE_INTER_Q)) { |
277 : | |||
278 : | /* reduced resolution + GMC: not possible */ | ||
279 : | |||
280 : | if (mb->mcsel) { | ||
281 : | syskin | 1.18.2.2 | |
282 : | edgomez | 1.17 | /* call normal routine once, easier than "if (mcsel)"ing all the time */ |
283 : | syskin | 1.18.2.2 | |
284 : | edgomez | 1.17 | transfer_8to16sub(&dct_codes[0*64], cur->y + 16*j*edged_width + 16*i, |
285 : | syskin | 1.18.2.2 | refGMC->y + 16*j*edged_width + 16*i, edged_width); |
286 : | edgomez | 1.17 | transfer_8to16sub(&dct_codes[1*64], cur->y + 16*j*edged_width + 16*i+8, |
287 : | syskin | 1.18.2.2 | refGMC->y + 16*j*edged_width + 16*i+8, edged_width); |
288 : | edgomez | 1.17 | transfer_8to16sub(&dct_codes[2*64], cur->y + (16*j+8)*edged_width + 16*i, |
289 : | syskin | 1.18.2.2 | refGMC->y + (16*j+8)*edged_width + 16*i, edged_width); |
290 : | edgomez | 1.17 | transfer_8to16sub(&dct_codes[3*64], cur->y + (16*j+8)*edged_width + 16*i+8, |
291 : | syskin | 1.18.2.2 | refGMC->y + (16*j+8)*edged_width + 16*i+8, edged_width); |
292 : | edgomez | 1.17 | |
293 : | /* lumi is needed earlier for mode decision, but chroma should be done block-based, but it isn't, yet. */ | ||
294 : | |||
295 : | transfer_8to16sub(&dct_codes[4 * 64], cur->u + 8 *j*edged_width/2 + 8*i, | ||
296 : | refGMC->u + 8 *j*edged_width/2 + 8*i, edged_width/2); | ||
297 : | |||
298 : | transfer_8to16sub(&dct_codes[5 * 64], cur->v + 8*j* edged_width/2 + 8*i, | ||
299 : | refGMC->v + 8*j* edged_width/2 + 8*i, edged_width/2); | ||
300 : | |||
301 : | return; | ||
302 : | } | ||
303 : | |||
304 : | /* ordinary compensation */ | ||
305 : | syskin | 1.18.2.2 | |
306 : | edgomez | 1.17 | dx = (quarterpel ? mb->qmvs[0].x : mb->mvs[0].x); |
307 : | dy = (quarterpel ? mb->qmvs[0].y : mb->mvs[0].y); | ||
308 : | |||
309 : | if (reduced_resolution) { | ||
310 : | dx = RRV_MV_SCALEUP(dx); | ||
311 : | dy = RRV_MV_SCALEUP(dy); | ||
312 : | } | ||
313 : | |||
314 : | compensate16x16_interpolate(&dct_codes[0 * 64], cur->y, ref->y, refh->y, | ||
315 : | refv->y, refhv->y, tmp, 16 * i, 16 * j, dx, dy, | ||
316 : | edged_width, quarterpel, reduced_resolution, rounding); | ||
317 : | |||
318 : | syskin | 1.18.2.2 | if (quarterpel) { dx /= 2; dy /= 2; } |
319 : | |||
320 : | edgomez | 1.17 | dx = (dx >> 1) + roundtab_79[dx & 0x3]; |
321 : | dy = (dy >> 1) + roundtab_79[dy & 0x3]; | ||
322 : | |||
323 : | edgomez | 1.18.2.5 | } else { /* mode == MODE_INTER4V */ |
324 : | edgomez | 1.17 | int k, sumx = 0, sumy = 0; |
325 : | const VECTOR * const mvs = (quarterpel ? mb->qmvs : mb->mvs); | ||
326 : | |||
327 : | for (k = 0; k < 4; k++) { | ||
328 : | dx = mvs[k].x; | ||
329 : | dy = mvs[k].y; | ||
330 : | syskin | 1.18.2.2 | sumx += quarterpel ? dx/2 : dx; |
331 : | sumy += quarterpel ? dy/2 : dy; | ||
332 : | edgomez | 1.17 | |
333 : | if (reduced_resolution){ | ||
334 : | dx = RRV_MV_SCALEUP(dx); | ||
335 : | dy = RRV_MV_SCALEUP(dy); | ||
336 : | } | ||
337 : | |||
338 : | compensate8x8_interpolate(&dct_codes[k * 64], cur->y, ref->y, refh->y, | ||
339 : | refv->y, refhv->y, tmp, 16 * i + 8*(k&1), 16 * j + 8*(k>>1), dx, | ||
340 : | dy, edged_width, quarterpel, reduced_resolution, rounding); | ||
341 : | } | ||
342 : | dx = (sumx >> 3) + roundtab_76[sumx & 0xf]; | ||
343 : | dy = (sumy >> 3) + roundtab_76[sumy & 0xf]; | ||
344 : | } | ||
345 : | |||
346 : | CompensateChroma(dx, dy, i, j, cur, ref, tmp, | ||
347 : | &dct_codes[4 * 64], edged_width / 2, rounding, reduced_resolution); | ||
348 : | } | ||
349 : | |||
350 : | |||
351 : | void | ||
352 : | MBMotionCompensationBVOP(MBParam * pParam, | ||
353 : | MACROBLOCK * const mb, | ||
354 : | const uint32_t i, | ||
355 : | const uint32_t j, | ||
356 : | IMAGE * const cur, | ||
357 : | const IMAGE * const f_ref, | ||
358 : | const IMAGE * const f_refh, | ||
359 : | const IMAGE * const f_refv, | ||
360 : | const IMAGE * const f_refhv, | ||
361 : | const IMAGE * const b_ref, | ||
362 : | const IMAGE * const b_refh, | ||
363 : | const IMAGE * const b_refv, | ||
364 : | const IMAGE * const b_refhv, | ||
365 : | int16_t * dct_codes) | ||
366 : | { | ||
367 : | const uint32_t edged_width = pParam->edged_width; | ||
368 : | int32_t dx, dy, b_dx, b_dy, sumx, sumy, b_sumx, b_sumy; | ||
369 : | int k; | ||
370 : | edgomez | 1.18.2.3 | const int quarterpel = pParam->vol_flags & XVID_VOL_QUARTERPEL; |
371 : | edgomez | 1.17 | const uint8_t * ptr1, * ptr2; |
372 : | uint8_t * const tmp = f_refv->u; | ||
373 : | const VECTOR * const fmvs = (quarterpel ? mb->qmvs : mb->mvs); | ||
374 : | const VECTOR * const bmvs = (quarterpel ? mb->b_qmvs : mb->b_mvs); | ||
375 : | |||
376 : | switch (mb->mode) { | ||
377 : | case MODE_FORWARD: | ||
378 : | dx = fmvs->x; dy = fmvs->y; | ||
379 : | |||
380 : | compensate16x16_interpolate(&dct_codes[0 * 64], cur->y, f_ref->y, f_refh->y, | ||
381 : | f_refv->y, f_refhv->y, tmp, 16 * i, 16 * j, dx, | ||
382 : | dy, edged_width, quarterpel, 0, 0); | ||
383 : | |||
384 : | if (quarterpel) { dx /= 2; dy /= 2; } | ||
385 : | |||
386 : | CompensateChroma( (dx >> 1) + roundtab_79[dx & 0x3], | ||
387 : | (dy >> 1) + roundtab_79[dy & 0x3], | ||
388 : | i, j, cur, f_ref, tmp, | ||
389 : | &dct_codes[4 * 64], edged_width / 2, 0, 0); | ||
390 : | |||
391 : | return; | ||
392 : | |||
393 : | case MODE_BACKWARD: | ||
394 : | b_dx = bmvs->x; b_dy = bmvs->y; | ||
395 : | |||
396 : | syskin | 1.18.2.2 | compensate16x16_interpolate(&dct_codes[0 * 64], cur->y, b_ref->y, b_refh->y, |
397 : | edgomez | 1.17 | b_refv->y, b_refhv->y, tmp, 16 * i, 16 * j, b_dx, |
398 : | syskin | 1.18.2.2 | b_dy, edged_width, quarterpel, 0, 0); |
399 : | edgomez | 1.17 | |
400 : | if (quarterpel) { b_dx /= 2; b_dy /= 2; } | ||
401 : | |||
402 : | CompensateChroma( (b_dx >> 1) + roundtab_79[b_dx & 0x3], | ||
403 : | (b_dy >> 1) + roundtab_79[b_dy & 0x3], | ||
404 : | i, j, cur, b_ref, tmp, | ||
405 : | &dct_codes[4 * 64], edged_width / 2, 0, 0); | ||
406 : | |||
407 : | return; | ||
408 : | |||
409 : | case MODE_INTERPOLATE: /* _could_ use DIRECT, but would be overkill (no 4MV there) */ | ||
410 : | case MODE_DIRECT_NO4V: | ||
411 : | dx = fmvs->x; dy = fmvs->y; | ||
412 : | b_dx = bmvs->x; b_dy = bmvs->y; | ||
413 : | |||
414 : | if (quarterpel) { | ||
415 : | syskin | 1.18.2.2 | |
416 : | edgomez | 1.17 | if ((dx&3) | (dy&3)) { |
417 : | interpolate16x16_quarterpel(tmp - i * 16 - j * 16 * edged_width, | ||
418 : | (uint8_t *) f_ref->y, tmp + 32, | ||
419 : | tmp + 64, tmp + 96, 16*i, 16*j, dx, dy, edged_width, 0); | ||
420 : | ptr1 = tmp; | ||
421 : | edgomez | 1.18.2.5 | } else ptr1 = f_ref->y + (16*j + dy/4)*edged_width + 16*i + dx/4; /* fullpixel position */ |
422 : | edgomez | 1.17 | |
423 : | if ((b_dx&3) | (b_dy&3)) { | ||
424 : | interpolate16x16_quarterpel(tmp - i * 16 - j * 16 * edged_width + 16, | ||
425 : | (uint8_t *) b_ref->y, tmp + 32, | ||
426 : | tmp + 64, tmp + 96, 16*i, 16*j, b_dx, b_dy, edged_width, 0); | ||
427 : | ptr2 = tmp + 16; | ||
428 : | edgomez | 1.18.2.5 | } else ptr2 = b_ref->y + (16*j + b_dy/4)*edged_width + 16*i + b_dx/4; /* fullpixel position */ |
429 : | edgomez | 1.17 | |
430 : | b_dx /= 2; | ||
431 : | b_dy /= 2; | ||
432 : | dx /= 2; | ||
433 : | dy /= 2; | ||
434 : | |||
435 : | } else { | ||
436 : | ptr1 = get_ref(f_ref->y, f_refh->y, f_refv->y, f_refhv->y, | ||
437 : | i, j, 16, dx, dy, edged_width); | ||
438 : | |||
439 : | ptr2 = get_ref(b_ref->y, b_refh->y, b_refv->y, b_refhv->y, | ||
440 : | i, j, 16, b_dx, b_dy, edged_width); | ||
441 : | } | ||
442 : | for (k = 0; k < 4; k++) | ||
443 : | transfer_8to16sub2(&dct_codes[k * 64], | ||
444 : | cur->y + (i * 16+(k&1)*8) + (j * 16+((k>>1)*8)) * edged_width, | ||
445 : | ptr1 + (k&1)*8 + (k>>1)*8*edged_width, | ||
446 : | ptr2 + (k&1)*8 + (k>>1)*8*edged_width, edged_width); | ||
447 : | |||
448 : | |||
449 : | dx = (dx >> 1) + roundtab_79[dx & 0x3]; | ||
450 : | dy = (dy >> 1) + roundtab_79[dy & 0x3]; | ||
451 : | |||
452 : | b_dx = (b_dx >> 1) + roundtab_79[b_dx & 0x3]; | ||
453 : | b_dy = (b_dy >> 1) + roundtab_79[b_dy & 0x3]; | ||
454 : | |||
455 : | break; | ||
456 : | syskin | 1.18.2.2 | |
457 : | edgomez | 1.18.2.5 | default: /* MODE_DIRECT (or MODE_DIRECT_NONE_MV in case of bframes decoding) */ |
458 : | edgomez | 1.17 | sumx = sumy = b_sumx = b_sumy = 0; |
459 : | |||
460 : | for (k = 0; k < 4; k++) { | ||
461 : | syskin | 1.18.2.2 | |
462 : | edgomez | 1.17 | dx = fmvs[k].x; dy = fmvs[k].y; |
463 : | b_dx = bmvs[k].x; b_dy = bmvs[k].y; | ||
464 : | |||
465 : | if (quarterpel) { | ||
466 : | sumx += dx/2; sumy += dy/2; | ||
467 : | b_sumx += b_dx/2; b_sumy += b_dy/2; | ||
468 : | |||
469 : | if ((dx&3) | (dy&3)) { | ||
470 : | interpolate8x8_quarterpel(tmp - (i * 16+(k&1)*8) - (j * 16+((k>>1)*8)) * edged_width, | ||
471 : | syskin | 1.18.2.2 | (uint8_t *) f_ref->y, |
472 : | tmp + 32, tmp + 64, tmp + 96, | ||
473 : | edgomez | 1.17 | 16*i + (k&1)*8, 16*j + (k>>1)*8, dx, dy, edged_width, 0); |
474 : | ptr1 = tmp; | ||
475 : | } else ptr1 = f_ref->y + (16*j + (k>>1)*8 + dy/4)*edged_width + 16*i + (k&1)*8 + dx/4; | ||
476 : | |||
477 : | if ((b_dx&3) | (b_dy&3)) { | ||
478 : | interpolate8x8_quarterpel(tmp - (i * 16+(k&1)*8) - (j * 16+((k>>1)*8)) * edged_width + 16, | ||
479 : | (uint8_t *) b_ref->y, | ||
480 : | syskin | 1.18.2.2 | tmp + 16, tmp + 32, tmp + 48, |
481 : | edgomez | 1.17 | 16*i + (k&1)*8, 16*j + (k>>1)*8, b_dx, b_dy, edged_width, 0); |
482 : | ptr2 = tmp + 16; | ||
483 : | } else ptr2 = b_ref->y + (16*j + (k>>1)*8 + b_dy/4)*edged_width + 16*i + (k&1)*8 + b_dx/4; | ||
484 : | } else { | ||
485 : | sumx += dx; sumy += dy; | ||
486 : | b_sumx += b_dx; b_sumy += b_dy; | ||
487 : | |||
488 : | syskin | 1.18.2.2 | ptr1 = get_ref(f_ref->y, f_refh->y, f_refv->y, f_refhv->y, |
489 : | edgomez | 1.17 | 2*i + (k&1), 2*j + (k>>1), 8, dx, dy, edged_width); |
490 : | syskin | 1.18.2.2 | ptr2 = get_ref(b_ref->y, b_refh->y, b_refv->y, b_refhv->y, |
491 : | edgomez | 1.17 | 2*i + (k&1), 2*j + (k>>1), 8, b_dx, b_dy, edged_width); |
492 : | } | ||
493 : | transfer_8to16sub2(&dct_codes[k * 64], | ||
494 : | cur->y + (i * 16+(k&1)*8) + (j * 16+((k>>1)*8)) * edged_width, | ||
495 : | ptr1, ptr2, edged_width); | ||
496 : | syskin | 1.18.2.2 | |
497 : | edgomez | 1.17 | } |
498 : | |||
499 : | dx = (sumx >> 3) + roundtab_76[sumx & 0xf]; | ||
500 : | dy = (sumy >> 3) + roundtab_76[sumy & 0xf]; | ||
501 : | b_dx = (b_sumx >> 3) + roundtab_76[b_sumx & 0xf]; | ||
502 : | b_dy = (b_sumy >> 3) + roundtab_76[b_sumy & 0xf]; | ||
503 : | |||
504 : | break; | ||
505 : | } | ||
506 : | |||
507 : | edgomez | 1.18.2.5 | /* v block-based chroma interpolation for direct and interpolate modes */ |
508 : | edgomez | 1.17 | transfer_8to16sub2(&dct_codes[4 * 64], |
509 : | cur->u + (j * 8) * edged_width / 2 + (i * 8), | ||
510 : | interpolate8x8_switch2(tmp, b_ref->u, 8 * i, 8 * j, | ||
511 : | b_dx, b_dy, edged_width / 2, 0), | ||
512 : | interpolate8x8_switch2(tmp + 8, f_ref->u, 8 * i, 8 * j, | ||
513 : | dx, dy, edged_width / 2, 0), | ||
514 : | edged_width / 2); | ||
515 : | |||
516 : | transfer_8to16sub2(&dct_codes[5 * 64], | ||
517 : | cur->v + (j * 8) * edged_width / 2 + (i * 8), | ||
518 : | interpolate8x8_switch2(tmp, b_ref->v, 8 * i, 8 * j, | ||
519 : | b_dx, b_dy, edged_width / 2, 0), | ||
520 : | interpolate8x8_switch2(tmp + 8, f_ref->v, 8 * i, 8 * j, | ||
521 : | dx, dy, edged_width / 2, 0), | ||
522 : | edged_width / 2); | ||
523 : | } | ||
524 : | |||
525 : | |||
526 : | |||
527 : | void generate_GMCparameters( const int num_wp, const int res, | ||
528 : | syskin | 1.18.2.2 | const WARPPOINTS *const warp, |
529 : | const int width, const int height, | ||
530 : | GMC_DATA *const gmc) | ||
531 : | Isibaar | 1.1 | { |
532 : | syskin | 1.18.2.2 | const int du0 = warp->duv[0].x; |
533 : | const int dv0 = warp->duv[0].y; | ||
534 : | const int du1 = warp->duv[1].x; | ||
535 : | const int dv1 = warp->duv[1].y; | ||
536 : | const int du2 = warp->duv[2].x; | ||
537 : | const int dv2 = warp->duv[2].y; | ||
538 : | edgomez | 1.17 | |
539 : | syskin | 1.18.2.2 | gmc->W = width; |
540 : | gmc->H = height; | ||
541 : | Isibaar | 1.1 | |
542 : | edgomez | 1.18.2.5 | gmc->rho = 4 - log2bin(res-1); /* = {3,2,1,0} for res={2,4,8,16} */ |
543 : | syskin | 1.18.2.2 | |
544 : | gmc->alpha = log2bin(gmc->W-1); | ||
545 : | gmc->Ws = (1 << gmc->alpha); | ||
546 : | |||
547 : | gmc->dxF = 16*gmc->Ws + RDIV( 8*gmc->Ws*du1, gmc->W ); | ||
548 : | gmc->dxG = RDIV( 8*gmc->Ws*dv1, gmc->W ); | ||
549 : | gmc->Fo = (res*du0 + 1) << (gmc->alpha+gmc->rho-1); | ||
550 : | gmc->Go = (res*dv0 + 1) << (gmc->alpha+gmc->rho-1); | ||
551 : | |||
552 : | if (num_wp==2) { | ||
553 : | gmc->dyF = -gmc->dxG; | ||
554 : | gmc->dyG = gmc->dxF; | ||
555 : | } else if (num_wp==3) { | ||
556 : | gmc->beta = log2bin(gmc->H-1); | ||
557 : | gmc->Hs = (1 << gmc->beta); | ||
558 : | gmc->dyF = RDIV( 8*gmc->Hs*du2, gmc->H ); | ||
559 : | gmc->dyG = 16*gmc->Hs + RDIV( 8*gmc->Hs*dv2, gmc->H ); | ||
560 : | if (gmc->beta > gmc->alpha) { | ||
561 : | gmc->dxF <<= (gmc->beta - gmc->alpha); | ||
562 : | gmc->dxG <<= (gmc->beta - gmc->alpha); | ||
563 : | gmc->alpha = gmc->beta; | ||
564 : | gmc->Ws = 1<< gmc->beta; | ||
565 : | } else { | ||
566 : | gmc->dyF <<= gmc->alpha - gmc->beta; | ||
567 : | gmc->dyG <<= gmc->alpha - gmc->beta; | ||
568 : | } | ||
569 : | } | ||
570 : | |||
571 : | gmc->cFo = gmc->dxF + gmc->dyF + (1 << (gmc->alpha+gmc->rho+1)); | ||
572 : | gmc->cFo += 16*gmc->Ws*(du0-1); | ||
573 : | |||
574 : | gmc->cGo = gmc->dxG + gmc->dyG + (1 << (gmc->alpha+gmc->rho+1)); | ||
575 : | gmc->cGo += 16*gmc->Ws*(dv0-1); | ||
576 : | edgomez | 1.17 | } |
577 : | Isibaar | 1.1 | |
578 : | syskin | 1.18.2.2 | void |
579 : | edgomez | 1.18.2.5 | generate_GMCimage( const GMC_DATA *const gmc_data, /* [input] precalculated data */ |
580 : | const IMAGE *const pRef, /* [input] */ | ||
581 : | syskin | 1.18.2.2 | const int mb_width, |
582 : | edgomez | 1.17 | const int mb_height, |
583 : | const int stride, | ||
584 : | syskin | 1.18.2.2 | const int stride2, |
585 : | edgomez | 1.18.2.5 | const int fcode, /* [input] some parameters... */ |
586 : | const int32_t quarterpel, /* [input] for rounding avgMV */ | ||
587 : | const int reduced_resolution, /* [input] ignored */ | ||
588 : | const int32_t rounding, /* [input] for rounding image data */ | ||
589 : | MACROBLOCK *const pMBs, /* [output] average motion vectors */ | ||
590 : | IMAGE *const pGMC) /* [output] full warped image */ | ||
591 : | edgomez | 1.17 | { |
592 : | chl | 1.13 | |
593 : | edgomez | 1.17 | unsigned int mj,mi; |
594 : | VECTOR avgMV; | ||
595 : | Isibaar | 1.1 | |
596 : | syskin | 1.18.2.2 | for (mj = 0; mj < (unsigned int)mb_height; mj++) |
597 : | for (mi = 0; mi < (unsigned int)mb_width; mi++) { | ||
598 : | |||
599 : | avgMV = generate_GMCimageMB(gmc_data, pRef, mi, mj, | ||
600 : | stride, stride2, quarterpel, rounding, pGMC); | ||
601 : | |||
602 : | pMBs[mj*mb_width+mi].amv.x = gmc_sanitize(avgMV.x, quarterpel, fcode); | ||
603 : | pMBs[mj*mb_width+mi].amv.y = gmc_sanitize(avgMV.y, quarterpel, fcode); | ||
604 : | pMBs[mj*mb_width+mi].mcsel = 0; /* until mode decision */ | ||
605 : | suxen_drol | 1.3 | } |
606 : | } | ||
607 : | edgomez | 1.17 | |
608 : | |||
609 : | |||
610 : | #define MLT(i) (((16-(i))<<16) + (i)) | ||
611 : | static const uint32_t MTab[16] = { | ||
612 : | MLT( 0), MLT( 1), MLT( 2), MLT( 3), MLT( 4), MLT( 5), MLT( 6), MLT(7), | ||
613 : | MLT( 8), MLT( 9), MLT(10), MLT(11), MLT(12), MLT(13), MLT(14), MLT(15) | ||
614 : | }; | ||
615 : | #undef MLT | ||
616 : | |||
617 : | VECTOR generate_GMCimageMB( const GMC_DATA *const gmc_data, | ||
618 : | syskin | 1.18.2.2 | const IMAGE *const pRef, |
619 : | const int mi, const int mj, | ||
620 : | const int stride, | ||
621 : | const int stride2, | ||
622 : | const int quarterpel, | ||
623 : | const int rounding, | ||
624 : | IMAGE *const pGMC) | ||
625 : | edgomez | 1.17 | { |
626 : | syskin | 1.18.2.2 | const int W = gmc_data->W; |
627 : | const int H = gmc_data->H; | ||
628 : | |||
629 : | const int rho = gmc_data->rho; | ||
630 : | const int alpha = gmc_data->alpha; | ||
631 : | |||
632 : | const int rounder = ( 128 - (rounding<<(rho+rho)) ) << 16; | ||
633 : | edgomez | 1.17 | |
634 : | syskin | 1.18.2.2 | const int dxF = gmc_data->dxF; |
635 : | const int dyF = gmc_data->dyF; | ||
636 : | const int dxG = gmc_data->dxG; | ||
637 : | const int dyG = gmc_data->dyG; | ||
638 : | edgomez | 1.17 | |
639 : | syskin | 1.18.2.2 | uint8_t *dstY, *dstU, *dstV; |
640 : | edgomez | 1.17 | |
641 : | syskin | 1.18.2.2 | int I,J; |
642 : | VECTOR avgMV = {0,0}; | ||
643 : | edgomez | 1.17 | |
644 : | syskin | 1.18.2.2 | int32_t Fj, Gj; |
645 : | edgomez | 1.17 | |
646 : | syskin | 1.18.2.2 | dstY = &pGMC->y[(mj*16)*stride+mi*16] + 16; |
647 : | |||
648 : | Fj = gmc_data->Fo + dyF*mj*16 + dxF*mi*16; | ||
649 : | Gj = gmc_data->Go + dyG*mj*16 + dxG*mi*16; | ||
650 : | |||
651 : | for (J = 16; J > 0; --J) { | ||
652 : | int32_t Fi, Gi; | ||
653 : | edgomez | 1.17 | |
654 : | syskin | 1.18.2.2 | Fi = Fj; Fj += dyF; |
655 : | Gi = Gj; Gj += dyG; | ||
656 : | for (I = -16; I < 0; ++I) { | ||
657 : | int32_t F, G; | ||
658 : | uint32_t ri, rj; | ||
659 : | |||
660 : | F = ( Fi >> (alpha+rho) ) << rho; Fi += dxF; | ||
661 : | G = ( Gi >> (alpha+rho) ) << rho; Gi += dxG; | ||
662 : | |||
663 : | avgMV.x += F; | ||
664 : | avgMV.y += G; | ||
665 : | |||
666 : | ri = MTab[F&15]; | ||
667 : | rj = MTab[G&15]; | ||
668 : | |||
669 : | F >>= 4; | ||
670 : | G >>= 4; | ||
671 : | |||
672 : | if (F < -1) F = -1; | ||
673 : | else if (F > W) F = W; | ||
674 : | if (G< -1) G=-1; | ||
675 : | else if (G>H) G=H; | ||
676 : | |||
677 : | edgomez | 1.18.2.5 | { /* MMX-like bilinear... */ |
678 : | syskin | 1.18.2.2 | const int offset = G*stride + F; |
679 : | uint32_t f0, f1; | ||
680 : | f0 = pRef->y[ offset +0 ]; | ||
681 : | f0 |= pRef->y[ offset +1 ] << 16; | ||
682 : | f1 = pRef->y[ offset+stride +0 ]; | ||
683 : | f1 |= pRef->y[ offset+stride +1 ] << 16; | ||
684 : | f0 = (ri*f0)>>16; | ||
685 : | f1 = (ri*f1) & 0x0fff0000; | ||
686 : | f0 |= f1; | ||
687 : | f0 = ( rj*f0 + rounder ) >> 24; | ||
688 : | edgomez | 1.17 | |
689 : | syskin | 1.18.2.2 | dstY[I] = (uint8_t)f0; |
690 : | } | ||
691 : | } | ||
692 : | |||
693 : | dstY += stride; | ||
694 : | } | ||
695 : | |||
696 : | dstU = &pGMC->u[(mj*8)*stride2+mi*8] + 8; | ||
697 : | dstV = &pGMC->v[(mj*8)*stride2+mi*8] + 8; | ||
698 : | |||
699 : | Fj = gmc_data->cFo + dyF*4 *mj*8 + dxF*4 *mi*8; | ||
700 : | Gj = gmc_data->cGo + dyG*4 *mj*8 + dxG*4 *mi*8; | ||
701 : | |||
702 : | for (J = 8; J > 0; --J) { | ||
703 : | int32_t Fi, Gi; | ||
704 : | Fi = Fj; Fj += 4*dyF; | ||
705 : | Gi = Gj; Gj += 4*dyG; | ||
706 : | |||
707 : | for (I = -8; I < 0; ++I) { | ||
708 : | int32_t F, G; | ||
709 : | uint32_t ri, rj; | ||
710 : | |||
711 : | F = ( Fi >> (alpha+rho+2) ) << rho; Fi += 4*dxF; | ||
712 : | G = ( Gi >> (alpha+rho+2) ) << rho; Gi += 4*dxG; | ||
713 : | |||
714 : | ri = MTab[F&15]; | ||
715 : | rj = MTab[G&15]; | ||
716 : | |||
717 : | F >>= 4; | ||
718 : | G >>= 4; | ||
719 : | |||
720 : | if (F < -1) F=-1; | ||
721 : | else if (F >= W/2) F = W/2; | ||
722 : | if (G < -1) G = -1; | ||
723 : | else if (G >= H/2) G = H/2; | ||
724 : | |||
725 : | { | ||
726 : | const int offset = G*stride2 + F; | ||
727 : | uint32_t f0, f1; | ||
728 : | |||
729 : | f0 = pRef->u[ offset +0 ]; | ||
730 : | f0 |= pRef->u[ offset +1 ] << 16; | ||
731 : | f1 = pRef->u[ offset+stride2 +0 ]; | ||
732 : | f1 |= pRef->u[ offset+stride2 +1 ] << 16; | ||
733 : | f0 = (ri*f0)>>16; | ||
734 : | f1 = (ri*f1) & 0x0fff0000; | ||
735 : | f0 |= f1; | ||
736 : | f0 = ( rj*f0 + rounder ) >> 24; | ||
737 : | |||
738 : | dstU[I] = (uint8_t)f0; | ||
739 : | |||
740 : | |||
741 : | f0 = pRef->v[ offset +0 ]; | ||
742 : | f0 |= pRef->v[ offset +1 ] << 16; | ||
743 : | f1 = pRef->v[ offset+stride2 +0 ]; | ||
744 : | f1 |= pRef->v[ offset+stride2 +1 ] << 16; | ||
745 : | f0 = (ri*f0)>>16; | ||
746 : | f1 = (ri*f1) & 0x0fff0000; | ||
747 : | f0 |= f1; | ||
748 : | f0 = ( rj*f0 + rounder ) >> 24; | ||
749 : | |||
750 : | dstV[I] = (uint8_t)f0; | ||
751 : | } | ||
752 : | } | ||
753 : | dstU += stride2; | ||
754 : | dstV += stride2; | ||
755 : | } | ||
756 : | |||
757 : | |||
758 : | edgomez | 1.18.2.5 | avgMV.x -= 16*((256*mi+120)<<4); /* 120 = 15*16/2 */ |
759 : | syskin | 1.18.2.2 | avgMV.y -= 16*((256*mj+120)<<4); |
760 : | |||
761 : | avgMV.x = RSHIFT( avgMV.x, (4+7-quarterpel) ); | ||
762 : | avgMV.y = RSHIFT( avgMV.y, (4+7-quarterpel) ); | ||
763 : | |||
764 : | return avgMV; | ||
765 : | edgomez | 1.17 | } |
766 : | syskin | 1.18.2.2 | |
767 : | |||
768 : | |||
769 : | edgomez | 1.17 | #ifdef OLD_GRUEL_GMC |
770 : | syskin | 1.18.2.2 | void |
771 : | edgomez | 1.18.2.5 | generate_GMCparameters( const int num_wp, /* [input]: number of warppoints */ |
772 : | const int res, /* [input]: resolution */ | ||
773 : | const WARPPOINTS *const warp, /* [input]: warp points */ | ||
774 : | edgomez | 1.17 | const int width, const int height, |
775 : | edgomez | 1.18.2.5 | GMC_DATA *const gmc) /* [output] precalculated parameters */ |
776 : | edgomez | 1.17 | { |
777 : | |||
778 : | syskin | 1.18.2.2 | /* We follow mainly two sources: The original standard, which is ugly, and the |
779 : | thesis from Andreas Dehnhardt, which is much nicer. | ||
780 : | |||
781 : | Notation is: indices are written next to the variable, | ||
782 : | primes in the standard are denoted by a suffix 'p'. | ||
783 : | types are "c"=constant, "i"=input parameter, "f"=calculated, then fixed, | ||
784 : | "o"=output data, " "=other, "u" = unused, "p"=calc for every pixel | ||
785 : | edgomez | 1.17 | |
786 : | syskin | 1.18.2.2 | type | variable name | ISO name (TeX-style) | value or range | usage |
787 : | edgomez | 1.17 | ------------------------------------------------------------------------------------- |
788 : | c | H | H | [16 , ?] | image width (w/o edges) | ||
789 : | c | W | W | [16 , ?] | image height (w/o edges) | ||
790 : | |||
791 : | c | i0 | i_0 | 0 | ref. point #1, X | ||
792 : | c | j0 | j_0 | 0 | ref. point #1, Y | ||
793 : | c | i1 | i_1 | W | ref. point #2, X | ||
794 : | c | j1 | j_1 | 0 | ref. point #2, Y | ||
795 : | cu | i2 | i_2 | 0 | ref. point #3, X | ||
796 : | syskin | 1.18.2.2 | cu | i2 | j_2 | H | ref. point #3, Y |
797 : | edgomez | 1.17 | |
798 : | i | du0 | du[0] | [-16863,16863] | warp vector #1, Y | ||
799 : | i | dv0 | dv[0] | [-16863,16863] | warp vector #1, Y | ||
800 : | i | du1 | du[1] | [-16863,16863] | warp vector #2, Y | ||
801 : | i | dv1 | dv[1] | [-16863,16863] | warp vector #2, Y | ||
802 : | syskin | 1.18.2.2 | iu | du2 | du[2] | [-16863,16863] | warp vector #3, Y |
803 : | edgomez | 1.17 | iu | dv2 | dv[2] | [-16863,16863] | warp vector #3, Y |
804 : | syskin | 1.18.2.2 | |
805 : | i | s | s | {2,4,8,16} | interpol. resolution | ||
806 : | f | sigma | - | log2(s) | X / s == X >> sigma | ||
807 : | f | r | r | =16/s | complementary res. | ||
808 : | f | rho | \rho | log2(r) | X / r == X >> rho | ||
809 : | |||
810 : | f | i0s | i'_0 | | | ||
811 : | f | j0s | j'_0 | | | ||
812 : | f | i1s | i'_1 | | | ||
813 : | f | j1s | j'_1 | | | ||
814 : | f | i2s | i'_2 | | | ||
815 : | f | j2s | j'_2 | | | ||
816 : | |||
817 : | f | alpha | \alpha | | 2^{alpha-1} < W <= 2^alpha | ||
818 : | f | beta | \beta | | 2^{beta-1} < H <= 2^beta | ||
819 : | |||
820 : | f | Ws | W' | W = 2^{alpha} | scaled width | ||
821 : | f | Hs | H' | W = 2^{beta} | scaled height | ||
822 : | |||
823 : | f | i1ss | i''_1 | "virtual sprite stuff" | ||
824 : | f | j1ss | j''_1 | "virtual sprite stuff" | ||
825 : | f | i2ss | i''_2 | "virtual sprite stuff" | ||
826 : | f | j2ss | j''_2 | "virtual sprite stuff" | ||
827 : | edgomez | 1.17 | */ |
828 : | |||
829 : | /* Some calculations are disabled because we only use 2 warppoints at the moment */ | ||
830 : | |||
831 : | int du0 = warp->duv[0].x; | ||
832 : | int dv0 = warp->duv[0].y; | ||
833 : | int du1 = warp->duv[1].x; | ||
834 : | int dv1 = warp->duv[1].y; | ||
835 : | edgomez | 1.18.2.5 | #if 0 |
836 : | int du2 = warp->duv[2].x; | ||
837 : | int dv2 = warp->duv[2].y; | ||
838 : | #endif | ||
839 : | syskin | 1.18.2.2 | |
840 : | edgomez | 1.17 | gmc->num_wp = num_wp; |
841 : | |||
842 : | gmc->s = res; /* scaling parameters 2,4,8 or 16 */ | ||
843 : | gmc->sigma = log2bin(res-1); /* log2bin(15)=4, log2bin(16)=5, log2bin(17)=5 */ | ||
844 : | syskin | 1.18.2.2 | gmc->r = 16/res; |
845 : | edgomez | 1.17 | gmc->rho = 4 - gmc->sigma; /* = log2bin(r-1) */ |
846 : | |||
847 : | gmc->W = width; | ||
848 : | gmc->H = height; /* fixed reference coordinates */ | ||
849 : | |||
850 : | gmc->alpha = log2bin(gmc->W-1); | ||
851 : | syskin | 1.18.2.2 | gmc->Ws= 1<<gmc->alpha; |
852 : | edgomez | 1.17 | |
853 : | edgomez | 1.18.2.5 | #if 0 |
854 : | gmc->beta = log2bin(gmc->H-1); | ||
855 : | gmc->Hs= 1<<gmc->beta; | ||
856 : | #endif | ||
857 : | edgomez | 1.17 | |
858 : | edgomez | 1.18.2.5 | #if 0 |
859 : | 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); | ||
860 : | #endif | ||
861 : | edgomez | 1.17 | |
862 : | edgomez | 1.18.2.5 | /* |
863 : | * i2s is only needed for num_wp >= 3, etc. | ||
864 : | * the 's' values are in 1/s pel resolution | ||
865 : | */ | ||
866 : | edgomez | 1.17 | gmc->i0s = res/2 * ( du0 ); |
867 : | gmc->j0s = res/2 * ( dv0 ); | ||
868 : | gmc->i1s = res/2 * (2*width + du1 + du0 ); | ||
869 : | gmc->j1s = res/2 * ( dv1 + dv0 ); | ||
870 : | edgomez | 1.18.2.5 | #if 0 |
871 : | gmc->i2s = res/2 * ( du2 + du0 ); | ||
872 : | gmc->j2s = res/2 * (2*height + dv2 + dv0 ); | ||
873 : | #endif | ||
874 : | syskin | 1.18.2.2 | |
875 : | edgomez | 1.17 | /* i2s and i2ss are only needed for num_wp == 3, etc. */ |
876 : | |||
877 : | /* the 'ss' values are in 1/16 pel resolution */ | ||
878 : | syskin | 1.18.2.2 | 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); |
879 : | gmc->j1ss = ROUNDED_DIV( ((gmc->W - gmc->Ws)*(gmc->r*gmc->j0s) + gmc->Ws*gmc->r*gmc->j1s) ,gmc->W ); | ||
880 : | edgomez | 1.17 | |
881 : | edgomez | 1.18.2.5 | #if 0 |
882 : | gmc->i2ss = ROUNDED_DIV( ((gmc->H - gmc->Hs)*(gmc->r*gmc->i0s) + gmc->Hs*(gmc->r*gmc->i2s)), gmc->H); | ||
883 : | 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); | ||
884 : | #endif | ||
885 : | edgomez | 1.17 | |
886 : | syskin | 1.18.2.2 | return; |
887 : | edgomez | 1.17 | } |
888 : | |||
889 : | syskin | 1.18.2.2 | void |
890 : | edgomez | 1.18.2.5 | generate_GMCimage( const GMC_DATA *const gmc_data, /* [input] precalculated data */ |
891 : | const IMAGE *const pRef, /* [input] */ | ||
892 : | syskin | 1.18.2.2 | const int mb_width, |
893 : | edgomez | 1.17 | const int mb_height, |
894 : | const int stride, | ||
895 : | syskin | 1.18.2.2 | const int stride2, |
896 : | edgomez | 1.18.2.5 | const int fcode, /* [input] some parameters... */ |
897 : | const int32_t quarterpel, /* [input] for rounding avgMV */ | ||
898 : | const int reduced_resolution, /* [input] ignored */ | ||
899 : | const int32_t rounding, /* [input] for rounding image data */ | ||
900 : | MACROBLOCK *const pMBs, /* [output] average motion vectors */ | ||
901 : | IMAGE *const pGMC) /* [output] full warped image */ | ||
902 : | edgomez | 1.17 | { |
903 : | |||
904 : | unsigned int mj,mi; | ||
905 : | VECTOR avgMV; | ||
906 : | syskin | 1.18.2.2 | |
907 : | for (mj = 0;mj < mb_height; mj++) | ||
908 : | for (mi = 0;mi < mb_width; mi++) { | ||
909 : | |||
910 : | avgMV = generate_GMCimageMB(gmc_data, pRef, mi, mj, | ||
911 : | edgomez | 1.17 | stride, stride2, quarterpel, rounding, pGMC); |
912 : | |||
913 : | pMBs[mj*mb_width+mi].amv.x = gmc_sanitize(avgMV.x, quarterpel, fcode); | ||
914 : | pMBs[mj*mb_width+mi].amv.y = gmc_sanitize(avgMV.y, quarterpel, fcode); | ||
915 : | pMBs[mj*mb_width+mi].mcsel = 0; /* until mode decision */ | ||
916 : | } | ||
917 : | } | ||
918 : | |||
919 : | |||
920 : | VECTOR generate_GMCimageMB( const GMC_DATA *const gmc_data, /* [input] all precalc data */ | ||
921 : | const IMAGE *const pRef, /* [input] */ | ||
922 : | const int mi, const int mj, /* [input] MB position */ | ||
923 : | const int stride, /* [input] Lumi stride */ | ||
924 : | const int stride2, /* [input] chroma stride */ | ||
925 : | const int quarterpel, /* [input] for rounding of avgMV */ | ||
926 : | const int rounding, /* [input] for rounding of imgae data */ | ||
927 : | IMAGE *const pGMC) /* [outut] generate image */ | ||
928 : | syskin | 1.18.2.2 | |
929 : | edgomez | 1.17 | /* |
930 : | syskin | 1.18.2.2 | type | variable name | ISO name (TeX-style) | value or range | usage |
931 : | edgomez | 1.17 | ------------------------------------------------------------------------------------- |
932 : | syskin | 1.18.2.2 | p | F | F(i,j) | | pelwise motion vector X in s-th pel |
933 : | p | G | G(i,j) | | pelwise motion vector Y in s-th pel | ||
934 : | p | Fc | F_c(i,j) | | | ||
935 : | p | Gc | G_c(i,j) | | same for chroma | ||
936 : | |||
937 : | p | Y00 | Y_{00} | [0,255*s*s] | first: 4 neighbouring Y-values | ||
938 : | p | Y01 | Y_{01} | [0,255] | at fullpel position, around the | ||
939 : | p | Y10 | Y_{10} | [0,255*s] | position where pelweise MV points to | ||
940 : | p | Y11 | Y_{11} | [0,255] | later: bilinear interpol Y-values in Y00 | ||
941 : | |||
942 : | p | C00 | C_{00} | [0,255*s*s] | same for chroma Cb and Cr | ||
943 : | p | C01 | C_{01} | [0,255] | | ||
944 : | p | C10 | C_{10} | [0,255*s] | | ||
945 : | p | C11 | C_{11} | [0,255] | | ||
946 : | edgomez | 1.17 | |
947 : | */ | ||
948 : | { | ||
949 : | const int W = gmc_data->W; | ||
950 : | const int H = gmc_data->H; | ||
951 : | |||
952 : | const int s = gmc_data->s; | ||
953 : | const int sigma = gmc_data->sigma; | ||
954 : | syskin | 1.18.2.2 | |
955 : | edgomez | 1.17 | const int r = gmc_data->r; |
956 : | syskin | 1.18.2.2 | const int rho = gmc_data->rho; |
957 : | |||
958 : | edgomez | 1.17 | const int i0s = gmc_data->i0s; |
959 : | const int j0s = gmc_data->j0s; | ||
960 : | syskin | 1.18.2.2 | |
961 : | edgomez | 1.17 | const int i1ss = gmc_data->i1ss; |
962 : | syskin | 1.18.2.2 | const int j1ss = gmc_data->j1ss; |
963 : | edgomez | 1.18.2.5 | #if 0 |
964 : | const int i2ss = gmc_data->i2ss; | ||
965 : | const int j2ss = gmc_data->j2ss; | ||
966 : | #endif | ||
967 : | edgomez | 1.17 | |
968 : | syskin | 1.18.2.2 | const int alpha = gmc_data->alpha; |
969 : | const int Ws = gmc_data->Ws; | ||
970 : | edgomez | 1.17 | |
971 : | edgomez | 1.18.2.5 | #if 0 |
972 : | const int beta = gmc_data->beta; | ||
973 : | const int Hs = gmc_data->Hs; | ||
974 : | #endif | ||
975 : | edgomez | 1.17 | |
976 : | int I,J; | ||
977 : | VECTOR avgMV = {0,0}; | ||
978 : | syskin | 1.18.2.2 | |
979 : | edgomez | 1.17 | for (J=16*mj;J<16*(mj+1);J++) |
980 : | for (I=16*mi;I<16*(mi+1);I++) | ||
981 : | { | ||
982 : | int F= i0s + ( ((-r*i0s+i1ss)*I + (r*j0s-j1ss)*J + (1<<(alpha+rho-1))) >> (alpha+rho) ); | ||
983 : | int G= j0s + ( ((-r*j0s+j1ss)*I + (-r*i0s+i1ss)*J + (1<<(alpha+rho-1))) >> (alpha+rho) ); | ||
984 : | |||
985 : | syskin | 1.18.2.2 | /* this naive implementation (with lots of multiplications) isn't slower (rather faster) than |
986 : | edgomez | 1.17 | working incremental. Don't ask me why... maybe the whole this is memory bound? */ |
987 : | |||
988 : | edgomez | 1.18.2.5 | const int ri= F & (s-1); /* fractional part of pelwise MV X */ |
989 : | const int rj= G & (s-1); /* fractional part of pelwise MV Y */ | ||
990 : | syskin | 1.18.2.2 | |
991 : | edgomez | 1.17 | int Y00,Y01,Y10,Y11; |
992 : | |||
993 : | /* unclipped values are used for avgMV */ | ||
994 : | avgMV.x += F-(I<<sigma); /* shift position to 1/s-pel, as the MV is */ | ||
995 : | avgMV.y += G-(J<<sigma); /* TODO: don't do this (of course) */ | ||
996 : | |||
997 : | F >>= sigma; | ||
998 : | G >>= sigma; | ||
999 : | |||
1000 : | syskin | 1.18.2.2 | /* clip values to be in range. Since we have edges, clip to 1 less than lower boundary |
1001 : | edgomez | 1.17 | this way positions F+1/G+1 are still right */ |
1002 : | |||
1003 : | if (F< -1) | ||
1004 : | F=-1; | ||
1005 : | else if (F>W) | ||
1006 : | F=W; /* W or W-1 doesn't matter, so save 1 subtract ;-) */ | ||
1007 : | if (G< -1) | ||
1008 : | G=-1; | ||
1009 : | else if (G>H) | ||
1010 : | G=H; /* dito */ | ||
1011 : | |||
1012 : | edgomez | 1.18.2.5 | Y00 = pRef->y[ G*stride + F ]; /* Lumi values */ |
1013 : | edgomez | 1.17 | Y01 = pRef->y[ G*stride + F+1 ]; |
1014 : | Y10 = pRef->y[ G*stride + F+stride ]; | ||
1015 : | Y11 = pRef->y[ G*stride + F+stride+1 ]; | ||
1016 : | syskin | 1.18.2.2 | |
1017 : | edgomez | 1.17 | /* bilinear interpolation */ |
1018 : | Y00 = ((s-ri)*Y00 + ri*Y01); | ||
1019 : | Y10 = ((s-ri)*Y10 + ri*Y11); | ||
1020 : | syskin | 1.18.2.2 | Y00 = ((s-rj)*Y00 + rj*Y10 + s*s/2 - rounding ) >> (sigma+sigma); |
1021 : | |||
1022 : | edgomez | 1.17 | pGMC->y[J*stride+I] = (uint8_t)Y00; /* output 1 Y-pixel */ |
1023 : | } | ||
1024 : | |||
1025 : | syskin | 1.18.2.2 | |
1026 : | /* doing chroma _here_ is even more stupid and slow, because won't be used until Compensation and | ||
1027 : | edgomez | 1.17 | most likely not even then (only if the block really _is_ GMC) |
1028 : | */ | ||
1029 : | |||
1030 : | for (J=8*mj;J<8*(mj+1);J++) /* this plays the role of j_c,i_c in the standard */ | ||
1031 : | for (I=8*mi;I<8*(mi+1);I++) /* For I_c we have to use I_c = 4*i_c+1 ! */ | ||
1032 : | { | ||
1033 : | /* same positions for both chroma components, U=Cb and V=Cr */ | ||
1034 : | syskin | 1.18.2.2 | int Fc=((-r*i0s+i1ss)*(4*I+1) + (r*j0s-j1ss)*(4*J+1) +2*Ws*r*i0s |
1035 : | edgomez | 1.17 | -16*Ws +(1<<(alpha+rho+1)))>>(alpha+rho+2); |
1036 : | syskin | 1.18.2.2 | int Gc=((-r*j0s+j1ss)*(4*I+1) +(-r*i0s+i1ss)*(4*J+1) +2*Ws*r*j0s |
1037 : | edgomez | 1.17 | -16*Ws +(1<<(alpha+rho+1))) >>(alpha+rho+2); |
1038 : | syskin | 1.18.2.2 | |
1039 : | edgomez | 1.18.2.5 | const int ri= Fc & (s-1); /* fractional part of pelwise MV X */ |
1040 : | const int rj= Gc & (s-1); /* fractional part of pelwise MV Y */ | ||
1041 : | edgomez | 1.17 | |
1042 : | int C00,C01,C10,C11; | ||
1043 : | syskin | 1.18.2.2 | |
1044 : | edgomez | 1.17 | Fc >>= sigma; |
1045 : | Gc >>= sigma; | ||
1046 : | syskin | 1.18.2.2 | |
1047 : | edgomez | 1.17 | if (Fc< -1) |
1048 : | Fc=-1; | ||
1049 : | else if (Fc>=W/2) | ||
1050 : | Fc=W/2; /* W or W-1 doesn't matter, so save 1 subtraction ;-) */ | ||
1051 : | if (Gc< -1) | ||
1052 : | Gc=-1; | ||
1053 : | else if (Gc>=H/2) | ||
1054 : | Gc=H/2; /* dito */ | ||
1055 : | |||
1056 : | syskin | 1.18.2.2 | /* now calculate U data */ |
1057 : | edgomez | 1.18.2.5 | C00 = pRef->u[ Gc*stride2 + Fc ]; /* chroma-value Cb */ |
1058 : | edgomez | 1.17 | C01 = pRef->u[ Gc*stride2 + Fc+1 ]; |
1059 : | C10 = pRef->u[ (Gc+1)*stride2 + Fc ]; | ||
1060 : | C11 = pRef->u[ (Gc+1)*stride2 + Fc+1 ]; | ||
1061 : | syskin | 1.18.2.2 | |
1062 : | edgomez | 1.17 | /* bilinear interpolation */ |
1063 : | C00 = ((s-ri)*C00 + ri*C01); | ||
1064 : | C10 = ((s-ri)*C10 + ri*C11); | ||
1065 : | syskin | 1.18.2.2 | C00 = ((s-rj)*C00 + rj*C10 + s*s/2 - rounding ) >> (sigma+sigma); |
1066 : | |||
1067 : | edgomez | 1.17 | pGMC->u[J*stride2+I] = (uint8_t)C00; /* output 1 U-pixel */ |
1068 : | syskin | 1.18.2.2 | |
1069 : | edgomez | 1.17 | /* now calculate V data */ |
1070 : | edgomez | 1.18.2.5 | C00 = pRef->v[ Gc*stride2 + Fc ]; /* chroma-value Cr */ |
1071 : | edgomez | 1.17 | C01 = pRef->v[ Gc*stride2 + Fc+1 ]; |
1072 : | C10 = pRef->v[ (Gc+1)*stride2 + Fc ]; | ||
1073 : | C11 = pRef->v[ (Gc+1)*stride2 + Fc+1 ]; | ||
1074 : | syskin | 1.18.2.2 | |
1075 : | edgomez | 1.17 | /* bilinear interpolation */ |
1076 : | C00 = ((s-ri)*C00 + ri*C01); | ||
1077 : | C10 = ((s-ri)*C10 + ri*C11); | ||
1078 : | syskin | 1.18.2.2 | C00 = ((s-rj)*C00 + rj*C10 + s*s/2 - rounding ) >> (sigma+sigma); |
1079 : | |||
1080 : | edgomez | 1.17 | pGMC->v[J*stride2+I] = (uint8_t)C00; /* output 1 V-pixel */ |
1081 : | } | ||
1082 : | syskin | 1.18.2.2 | |
1083 : | edgomez | 1.18.2.5 | /* The average vector is rounded from 1/s-pel to 1/2 or 1/4 using the '//' operator */ |
1084 : | edgomez | 1.17 | |
1085 : | avgMV.x = RSHIFT( avgMV.x, (sigma+7-quarterpel) ); | ||
1086 : | avgMV.y = RSHIFT( avgMV.y, (sigma+7-quarterpel) ); | ||
1087 : | |||
1088 : | /* ^^^^ this is the way MS Reference Software does it */ | ||
1089 : | syskin | 1.18.2.2 | |
1090 : | edgomez | 1.17 | return avgMV; /* clipping to fcode area is done outside! */ |
1091 : | } | ||
1092 : | |||
1093 : | #endif |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |