Parent Directory
| 
Revision Log
Revision 1004 - (view) (download)
| 1 : | edgomez | 851 | // 30.10.2002 corrected qpel chroma rounding |
| 2 : | // 04.10.2002 added qpel support to MBMotionCompensation | ||
| 3 : | // 01.05.2002 updated MBMotionCompensationBVOP | ||
| 4 : | // 14.04.2002 bframe compensation | ||
| 5 : | suxen_drol | 118 | |
| 6 : | edgomez | 851 | #include <stdio.h> |
| 7 : | |||
| 8 : | Isibaar | 3 | #include "../encoder.h" |
| 9 : | #include "../utils/mbfunctions.h" | ||
| 10 : | #include "../image/interpolate8x8.h" | ||
| 11 : | edgomez | 851 | #include "../image/reduced.h" |
| 12 : | Isibaar | 3 | #include "../utils/timer.h" |
| 13 : | Isibaar | 1004 | #include "../image/qpel.h" |
| 14 : | suxen_drol | 118 | #include "motion.h" |
| 15 : | Isibaar | 3 | |
| 16 : | edgomez | 851 | #ifndef ABS |
| 17 : | Isibaar | 3 | #define ABS(X) (((X)>0)?(X):-(X)) |
| 18 : | edgomez | 851 | #endif |
| 19 : | #ifndef SIGN | ||
| 20 : | Isibaar | 3 | #define SIGN(X) (((X)>0)?1:-1) |
| 21 : | edgomez | 851 | #endif |
| 22 : | Isibaar | 3 | |
| 23 : | edgomez | 851 | #ifndef RSHIFT |
| 24 : | #define RSHIFT(a,b) ((a) > 0 ? ((a) + (1<<((b)-1)))>>(b) : ((a) + (1<<((b)-1))-1)>>(b)) | ||
| 25 : | #endif | ||
| 26 : | |||
| 27 : | /* assume b>0 */ | ||
| 28 : | #ifndef RDIV | ||
| 29 : | #define RDIV(a,b) (((a)>0 ? (a) + ((b)>>1) : (a) - ((b)>>1))/(b)) | ||
| 30 : | #endif | ||
| 31 : | |||
| 32 : | |||
| 33 : | syskin | 935 | /* This is borrowed from decoder.c */ |
| 34 : | edgomez | 851 | static __inline int gmc_sanitize(int value, int quarterpel, int fcode) |
| 35 : | { | ||
| 36 : | int length = 1 << (fcode+4); | ||
| 37 : | |||
| 38 : | // if (quarterpel) value *= 2; | ||
| 39 : | |||
| 40 : | syskin | 935 | if (value < -length) |
| 41 : | edgomez | 851 | return -length; |
| 42 : | syskin | 935 | else if (value >= length) |
| 43 : | edgomez | 851 | return length-1; |
| 44 : | else return value; | ||
| 45 : | } | ||
| 46 : | |||
| 47 : | /* And this is borrowed from bitstream.c until we find a common solution */ | ||
| 48 : | |||
| 49 : | static uint32_t __inline | ||
| 50 : | log2bin(uint32_t value) | ||
| 51 : | { | ||
| 52 : | /* Changed by Chenm001 */ | ||
| 53 : | #if !defined(_MSC_VER) | ||
| 54 : | int n = 0; | ||
| 55 : | |||
| 56 : | while (value) { | ||
| 57 : | value >>= 1; | ||
| 58 : | n++; | ||
| 59 : | } | ||
| 60 : | return n; | ||
| 61 : | #else | ||
| 62 : | __asm { | ||
| 63 : | syskin | 935 | bsr eax, value |
| 64 : | edgomez | 851 | inc eax |
| 65 : | } | ||
| 66 : | #endif | ||
| 67 : | } | ||
| 68 : | |||
| 69 : | |||
| 70 : | edgomez | 195 | static __inline void |
| 71 : | edgomez | 851 | compensate16x16_interpolate(int16_t * const dct_codes, |
| 72 : | uint8_t * const cur, | ||
| 73 : | const uint8_t * const ref, | ||
| 74 : | const uint8_t * const refh, | ||
| 75 : | const uint8_t * const refv, | ||
| 76 : | const uint8_t * const refhv, | ||
| 77 : | uint8_t * const tmp, | ||
| 78 : | uint32_t x, | ||
| 79 : | uint32_t y, | ||
| 80 : | const int32_t dx, | ||
| 81 : | const int32_t dy, | ||
| 82 : | const int32_t stride, | ||
| 83 : | const int quarterpel, | ||
| 84 : | const int reduced_resolution, | ||
| 85 : | const int32_t rounding) | ||
| 86 : | Isibaar | 3 | { |
| 87 : | edgomez | 851 | const uint8_t * ptr; |
| 88 : | Isibaar | 3 | |
| 89 : | edgomez | 851 | if (!reduced_resolution) { |
| 90 : | Isibaar | 3 | |
| 91 : | edgomez | 851 | if(quarterpel) { |
| 92 : | if ((dx&3) | (dy&3)) { | ||
| 93 : | Isibaar | 1004 | new_interpolate16x16_quarterpel(tmp - y * stride - x, |
| 94 : | edgomez | 851 | (uint8_t *) ref, tmp + 32, |
| 95 : | tmp + 64, tmp + 96, x, y, dx, dy, stride, rounding); | ||
| 96 : | ptr = tmp; | ||
| 97 : | } else ptr = ref + (y + dy/4)*stride + x + dx/4; // fullpixel position | ||
| 98 : | Isibaar | 3 | |
| 99 : | edgomez | 851 | } else ptr = get_ref(ref, refh, refv, refhv, x, y, 1, dx, dy, stride); |
| 100 : | |||
| 101 : | edgomez | 195 | transfer_8to16sub(dct_codes, cur + y * stride + x, |
| 102 : | syskin | 935 | ptr, stride); |
| 103 : | edgomez | 851 | transfer_8to16sub(dct_codes+64, cur + y * stride + x + 8, |
| 104 : | syskin | 935 | ptr + 8, stride); |
| 105 : | edgomez | 851 | transfer_8to16sub(dct_codes+128, cur + y * stride + x + 8*stride, |
| 106 : | syskin | 935 | ptr + 8*stride, stride); |
| 107 : | edgomez | 851 | transfer_8to16sub(dct_codes+192, cur + y * stride + x + 8*stride+8, |
| 108 : | syskin | 935 | ptr + 8*stride + 8, stride); |
| 109 : | Isibaar | 3 | |
| 110 : | edgomez | 851 | } else { //reduced_resolution |
| 111 : | syskin | 935 | |
| 112 : | edgomez | 851 | x *= 2; y *= 2; |
| 113 : | |||
| 114 : | ptr = get_ref(ref, refh, refv, refhv, x, y, 1, dx, dy, stride); | ||
| 115 : | syskin | 935 | |
| 116 : | edgomez | 851 | filter_18x18_to_8x8(dct_codes, cur+y*stride + x, stride); |
| 117 : | filter_diff_18x18_to_8x8(dct_codes, ptr, stride); | ||
| 118 : | |||
| 119 : | filter_18x18_to_8x8(dct_codes+64, cur+y*stride + x + 16, stride); | ||
| 120 : | filter_diff_18x18_to_8x8(dct_codes+64, ptr + 16, stride); | ||
| 121 : | |||
| 122 : | filter_18x18_to_8x8(dct_codes+128, cur+(y+16)*stride + x, stride); | ||
| 123 : | filter_diff_18x18_to_8x8(dct_codes+128, ptr + 16*stride, stride); | ||
| 124 : | |||
| 125 : | filter_18x18_to_8x8(dct_codes+192, cur+(y+16)*stride + x + 16, stride); | ||
| 126 : | filter_diff_18x18_to_8x8(dct_codes+192, ptr + 16*stride + 16, stride); | ||
| 127 : | |||
| 128 : | transfer32x32_copy(cur + y*stride + x, ptr, stride); | ||
| 129 : | edgomez | 195 | } |
| 130 : | Isibaar | 3 | } |
| 131 : | |||
| 132 : | edgomez | 851 | static __inline void |
| 133 : | compensate8x8_interpolate( int16_t * const dct_codes, | ||
| 134 : | uint8_t * const cur, | ||
| 135 : | const uint8_t * const ref, | ||
| 136 : | const uint8_t * const refh, | ||
| 137 : | const uint8_t * const refv, | ||
| 138 : | const uint8_t * const refhv, | ||
| 139 : | uint8_t * const tmp, | ||
| 140 : | uint32_t x, | ||
| 141 : | uint32_t y, | ||
| 142 : | const int32_t dx, | ||
| 143 : | const int32_t dy, | ||
| 144 : | const int32_t stride, | ||
| 145 : | const int32_t quarterpel, | ||
| 146 : | const int reduced_resolution, | ||
| 147 : | const int32_t rounding) | ||
| 148 : | { | ||
| 149 : | const uint8_t * ptr; | ||
| 150 : | Isibaar | 3 | |
| 151 : | edgomez | 851 | if (!reduced_resolution) { |
| 152 : | Isibaar | 3 | |
| 153 : | edgomez | 851 | if(quarterpel) { |
| 154 : | if ((dx&3) | (dy&3)) { | ||
| 155 : | Isibaar | 1004 | new_interpolate8x8_quarterpel(tmp - y*stride - x, |
| 156 : | edgomez | 851 | (uint8_t *) ref, tmp + 32, |
| 157 : | tmp + 64, tmp + 96, x, y, dx, dy, stride, rounding); | ||
| 158 : | ptr = tmp; | ||
| 159 : | } else ptr = ref + (y + dy/4)*stride + x + dx/4; // fullpixel position | ||
| 160 : | } else ptr = get_ref(ref, refh, refv, refhv, x, y, 1, dx, dy, stride); | ||
| 161 : | |||
| 162 : | transfer_8to16sub(dct_codes, cur + y * stride + x, ptr, stride); | ||
| 163 : | |||
| 164 : | } else { //reduced_resolution | ||
| 165 : | |||
| 166 : | x *= 2; y *= 2; | ||
| 167 : | |||
| 168 : | ptr = get_ref(ref, refh, refv, refhv, x, y, 1, dx, dy, stride); | ||
| 169 : | |||
| 170 : | filter_18x18_to_8x8(dct_codes, cur+y*stride + x, stride); | ||
| 171 : | filter_diff_18x18_to_8x8(dct_codes, ptr, stride); | ||
| 172 : | syskin | 935 | |
| 173 : | edgomez | 851 | transfer16x16_copy(cur + y*stride + x, ptr, stride); |
| 174 : | } | ||
| 175 : | } | ||
| 176 : | |||
| 177 : | /* XXX: slow, inelegant... */ | ||
| 178 : | static void | ||
| 179 : | interpolate18x18_switch(uint8_t * const cur, | ||
| 180 : | const uint8_t * const refn, | ||
| 181 : | const uint32_t x, | ||
| 182 : | const uint32_t y, | ||
| 183 : | const int32_t dx, | ||
| 184 : | const int dy, | ||
| 185 : | const int32_t stride, | ||
| 186 : | const int32_t rounding) | ||
| 187 : | { | ||
| 188 : | interpolate8x8_switch(cur, refn, x-1, y-1, dx, dy, stride, rounding); | ||
| 189 : | interpolate8x8_switch(cur, refn, x+7, y-1, dx, dy, stride, rounding); | ||
| 190 : | interpolate8x8_switch(cur, refn, x+9, y-1, dx, dy, stride, rounding); | ||
| 191 : | |||
| 192 : | interpolate8x8_switch(cur, refn, x-1, y+7, dx, dy, stride, rounding); | ||
| 193 : | interpolate8x8_switch(cur, refn, x+7, y+7, dx, dy, stride, rounding); | ||
| 194 : | interpolate8x8_switch(cur, refn, x+9, y+7, dx, dy, stride, rounding); | ||
| 195 : | |||
| 196 : | interpolate8x8_switch(cur, refn, x-1, y+9, dx, dy, stride, rounding); | ||
| 197 : | interpolate8x8_switch(cur, refn, x+7, y+9, dx, dy, stride, rounding); | ||
| 198 : | interpolate8x8_switch(cur, refn, x+9, y+9, dx, dy, stride, rounding); | ||
| 199 : | } | ||
| 200 : | |||
| 201 : | static void | ||
| 202 : | CompensateChroma( int dx, int dy, | ||
| 203 : | const int i, const int j, | ||
| 204 : | IMAGE * const Cur, | ||
| 205 : | const IMAGE * const Ref, | ||
| 206 : | uint8_t * const temp, | ||
| 207 : | int16_t * const coeff, | ||
| 208 : | const int32_t stride, | ||
| 209 : | const int rounding, | ||
| 210 : | const int rrv) | ||
| 211 : | { /* uv-block-based compensation */ | ||
| 212 : | |||
| 213 : | if (!rrv) { | ||
| 214 : | transfer_8to16sub(coeff, Cur->u + 8 * j * stride + 8 * i, | ||
| 215 : | syskin | 935 | interpolate8x8_switch2(temp, Ref->u, 8 * i, 8 * j, |
| 216 : | edgomez | 851 | dx, dy, stride, rounding), |
| 217 : | stride); | ||
| 218 : | transfer_8to16sub(coeff + 64, Cur->v + 8 * j * stride + 8 * i, | ||
| 219 : | syskin | 935 | interpolate8x8_switch2(temp, Ref->v, 8 * i, 8 * j, |
| 220 : | edgomez | 851 | dx, dy, stride, rounding), |
| 221 : | stride); | ||
| 222 : | } else { | ||
| 223 : | uint8_t * current, * reference; | ||
| 224 : | |||
| 225 : | current = Cur->u + 16*j*stride + 16*i; | ||
| 226 : | reference = temp - 16*j*stride - 16*i; | ||
| 227 : | interpolate18x18_switch(reference, Ref->u, 16*i, 16*j, dx, dy, stride, rounding); | ||
| 228 : | filter_18x18_to_8x8(coeff, current, stride); | ||
| 229 : | filter_diff_18x18_to_8x8(coeff, temp, stride); | ||
| 230 : | transfer16x16_copy(current, temp, stride); | ||
| 231 : | |||
| 232 : | current = Cur->v + 16*j*stride + 16*i; | ||
| 233 : | interpolate18x18_switch(reference, Ref->v, 16*i, 16*j, dx, dy, stride, rounding); | ||
| 234 : | filter_18x18_to_8x8(coeff + 64, current, stride); | ||
| 235 : | filter_diff_18x18_to_8x8(coeff + 64, temp, stride); | ||
| 236 : | transfer16x16_copy(current, temp, stride); | ||
| 237 : | } | ||
| 238 : | } | ||
| 239 : | |||
| 240 : | edgomez | 195 | void |
| 241 : | MBMotionCompensation(MACROBLOCK * const mb, | ||
| 242 : | syskin | 935 | const uint32_t i, |
| 243 : | const uint32_t j, | ||
| 244 : | const IMAGE * const ref, | ||
| 245 : | const IMAGE * const refh, | ||
| 246 : | const IMAGE * const refv, | ||
| 247 : | const IMAGE * const refhv, | ||
| 248 : | const IMAGE * const refGMC, | ||
| 249 : | IMAGE * const cur, | ||
| 250 : | int16_t * dct_codes, | ||
| 251 : | const uint32_t width, | ||
| 252 : | const uint32_t height, | ||
| 253 : | const uint32_t edged_width, | ||
| 254 : | const int32_t quarterpel, | ||
| 255 : | const int reduced_resolution, | ||
| 256 : | const int32_t rounding) | ||
| 257 : | Isibaar | 3 | { |
| 258 : | edgomez | 851 | int32_t dx; |
| 259 : | int32_t dy; | ||
| 260 : | Isibaar | 3 | |
| 261 : | edgomez | 851 | uint8_t * const tmp = refv->u; |
| 262 : | Isibaar | 3 | |
| 263 : | edgomez | 851 | if ( (!reduced_resolution) && (mb->mode == MODE_NOT_CODED) ) { /* quick copy for early SKIP */ |
| 264 : | /* early SKIP is only activated in P-VOPs, not in S-VOPs, so mcsel can never be 1 */ | ||
| 265 : | Isibaar | 3 | |
| 266 : | edgomez | 851 | transfer16x16_copy(cur->y + 16 * (i + j * edged_width), |
| 267 : | syskin | 935 | ref->y + 16 * (i + j * edged_width), |
| 268 : | edged_width); | ||
| 269 : | |||
| 270 : | edgomez | 851 | transfer8x8_copy(cur->u + 8 * (i + j * edged_width/2), |
| 271 : | ref->u + 8 * (i + j * edged_width/2), | ||
| 272 : | edged_width / 2); | ||
| 273 : | transfer8x8_copy(cur->v + 8 * (i + j * edged_width/2), | ||
| 274 : | ref->v + 8 * (i + j * edged_width/2), | ||
| 275 : | edged_width / 2); | ||
| 276 : | return; | ||
| 277 : | } | ||
| 278 : | Isibaar | 3 | |
| 279 : | syskin | 935 | if ((mb->mode == MODE_NOT_CODED || mb->mode == MODE_INTER |
| 280 : | edgomez | 851 | || mb->mode == MODE_INTER_Q)) { |
| 281 : | chl | 437 | |
| 282 : | edgomez | 851 | /* reduced resolution + GMC: not possible */ |
| 283 : | Isibaar | 3 | |
| 284 : | edgomez | 851 | if (mb->mcsel) { |
| 285 : | syskin | 935 | |
| 286 : | edgomez | 851 | /* call normal routine once, easier than "if (mcsel)"ing all the time */ |
| 287 : | syskin | 935 | |
| 288 : | edgomez | 851 | transfer_8to16sub(&dct_codes[0*64], cur->y + 16*j*edged_width + 16*i, |
| 289 : | syskin | 935 | refGMC->y + 16*j*edged_width + 16*i, edged_width); |
| 290 : | edgomez | 851 | transfer_8to16sub(&dct_codes[1*64], cur->y + 16*j*edged_width + 16*i+8, |
| 291 : | syskin | 935 | refGMC->y + 16*j*edged_width + 16*i+8, edged_width); |
| 292 : | edgomez | 851 | transfer_8to16sub(&dct_codes[2*64], cur->y + (16*j+8)*edged_width + 16*i, |
| 293 : | syskin | 935 | refGMC->y + (16*j+8)*edged_width + 16*i, edged_width); |
| 294 : | edgomez | 851 | transfer_8to16sub(&dct_codes[3*64], cur->y + (16*j+8)*edged_width + 16*i+8, |
| 295 : | syskin | 935 | refGMC->y + (16*j+8)*edged_width + 16*i+8, edged_width); |
| 296 : | chl | 437 | |
| 297 : | edgomez | 851 | /* lumi is needed earlier for mode decision, but chroma should be done block-based, but it isn't, yet. */ |
| 298 : | |||
| 299 : | transfer_8to16sub(&dct_codes[4 * 64], cur->u + 8 *j*edged_width/2 + 8*i, | ||
| 300 : | refGMC->u + 8 *j*edged_width/2 + 8*i, edged_width/2); | ||
| 301 : | |||
| 302 : | transfer_8to16sub(&dct_codes[5 * 64], cur->v + 8*j* edged_width/2 + 8*i, | ||
| 303 : | refGMC->v + 8*j* edged_width/2 + 8*i, edged_width/2); | ||
| 304 : | |||
| 305 : | return; | ||
| 306 : | } | ||
| 307 : | |||
| 308 : | /* ordinary compensation */ | ||
| 309 : | syskin | 935 | |
| 310 : | edgomez | 851 | dx = (quarterpel ? mb->qmvs[0].x : mb->mvs[0].x); |
| 311 : | dy = (quarterpel ? mb->qmvs[0].y : mb->mvs[0].y); | ||
| 312 : | |||
| 313 : | if (reduced_resolution) { | ||
| 314 : | dx = RRV_MV_SCALEUP(dx); | ||
| 315 : | dy = RRV_MV_SCALEUP(dy); | ||
| 316 : | } | ||
| 317 : | |||
| 318 : | compensate16x16_interpolate(&dct_codes[0 * 64], cur->y, ref->y, refh->y, | ||
| 319 : | refv->y, refhv->y, tmp, 16 * i, 16 * j, dx, dy, | ||
| 320 : | edged_width, quarterpel, reduced_resolution, rounding); | ||
| 321 : | |||
| 322 : | syskin | 935 | if (quarterpel) { dx /= 2; dy /= 2; } |
| 323 : | |||
| 324 : | edgomez | 851 | dx = (dx >> 1) + roundtab_79[dx & 0x3]; |
| 325 : | dy = (dy >> 1) + roundtab_79[dy & 0x3]; | ||
| 326 : | |||
| 327 : | } else { // mode == MODE_INTER4V | ||
| 328 : | int k, sumx = 0, sumy = 0; | ||
| 329 : | const VECTOR * const mvs = (quarterpel ? mb->qmvs : mb->mvs); | ||
| 330 : | |||
| 331 : | for (k = 0; k < 4; k++) { | ||
| 332 : | dx = mvs[k].x; | ||
| 333 : | dy = mvs[k].y; | ||
| 334 : | syskin | 935 | sumx += quarterpel ? dx/2 : dx; |
| 335 : | sumy += quarterpel ? dy/2 : dy; | ||
| 336 : | edgomez | 851 | |
| 337 : | if (reduced_resolution){ | ||
| 338 : | dx = RRV_MV_SCALEUP(dx); | ||
| 339 : | dy = RRV_MV_SCALEUP(dy); | ||
| 340 : | } | ||
| 341 : | |||
| 342 : | compensate8x8_interpolate(&dct_codes[k * 64], cur->y, ref->y, refh->y, | ||
| 343 : | refv->y, refhv->y, tmp, 16 * i + 8*(k&1), 16 * j + 8*(k>>1), dx, | ||
| 344 : | dy, edged_width, quarterpel, reduced_resolution, rounding); | ||
| 345 : | } | ||
| 346 : | dx = (sumx >> 3) + roundtab_76[sumx & 0xf]; | ||
| 347 : | dy = (sumy >> 3) + roundtab_76[sumy & 0xf]; | ||
| 348 : | } | ||
| 349 : | |||
| 350 : | CompensateChroma(dx, dy, i, j, cur, ref, tmp, | ||
| 351 : | &dct_codes[4 * 64], edged_width / 2, rounding, reduced_resolution); | ||
| 352 : | } | ||
| 353 : | |||
| 354 : | |||
| 355 : | void | ||
| 356 : | MBMotionCompensationBVOP(MBParam * pParam, | ||
| 357 : | MACROBLOCK * const mb, | ||
| 358 : | const uint32_t i, | ||
| 359 : | const uint32_t j, | ||
| 360 : | IMAGE * const cur, | ||
| 361 : | const IMAGE * const f_ref, | ||
| 362 : | const IMAGE * const f_refh, | ||
| 363 : | const IMAGE * const f_refv, | ||
| 364 : | const IMAGE * const f_refhv, | ||
| 365 : | const IMAGE * const b_ref, | ||
| 366 : | const IMAGE * const b_refh, | ||
| 367 : | const IMAGE * const b_refv, | ||
| 368 : | const IMAGE * const b_refhv, | ||
| 369 : | int16_t * dct_codes) | ||
| 370 : | { | ||
| 371 : | const uint32_t edged_width = pParam->edged_width; | ||
| 372 : | int32_t dx, dy, b_dx, b_dy, sumx, sumy, b_sumx, b_sumy; | ||
| 373 : | int k; | ||
| 374 : | const int quarterpel = pParam->m_quarterpel; | ||
| 375 : | const uint8_t * ptr1, * ptr2; | ||
| 376 : | uint8_t * const tmp = f_refv->u; | ||
| 377 : | const VECTOR * const fmvs = (quarterpel ? mb->qmvs : mb->mvs); | ||
| 378 : | const VECTOR * const bmvs = (quarterpel ? mb->b_qmvs : mb->b_mvs); | ||
| 379 : | |||
| 380 : | switch (mb->mode) { | ||
| 381 : | case MODE_FORWARD: | ||
| 382 : | dx = fmvs->x; dy = fmvs->y; | ||
| 383 : | |||
| 384 : | compensate16x16_interpolate(&dct_codes[0 * 64], cur->y, f_ref->y, f_refh->y, | ||
| 385 : | f_refv->y, f_refhv->y, tmp, 16 * i, 16 * j, dx, | ||
| 386 : | dy, edged_width, quarterpel, 0, 0); | ||
| 387 : | |||
| 388 : | if (quarterpel) { dx /= 2; dy /= 2; } | ||
| 389 : | |||
| 390 : | CompensateChroma( (dx >> 1) + roundtab_79[dx & 0x3], | ||
| 391 : | (dy >> 1) + roundtab_79[dy & 0x3], | ||
| 392 : | i, j, cur, f_ref, tmp, | ||
| 393 : | &dct_codes[4 * 64], edged_width / 2, 0, 0); | ||
| 394 : | |||
| 395 : | return; | ||
| 396 : | |||
| 397 : | case MODE_BACKWARD: | ||
| 398 : | b_dx = bmvs->x; b_dy = bmvs->y; | ||
| 399 : | |||
| 400 : | syskin | 935 | compensate16x16_interpolate(&dct_codes[0 * 64], cur->y, b_ref->y, b_refh->y, |
| 401 : | edgomez | 851 | b_refv->y, b_refhv->y, tmp, 16 * i, 16 * j, b_dx, |
| 402 : | syskin | 935 | b_dy, edged_width, quarterpel, 0, 0); |
| 403 : | edgomez | 851 | |
| 404 : | if (quarterpel) { b_dx /= 2; b_dy /= 2; } | ||
| 405 : | |||
| 406 : | CompensateChroma( (b_dx >> 1) + roundtab_79[b_dx & 0x3], | ||
| 407 : | (b_dy >> 1) + roundtab_79[b_dy & 0x3], | ||
| 408 : | i, j, cur, b_ref, tmp, | ||
| 409 : | &dct_codes[4 * 64], edged_width / 2, 0, 0); | ||
| 410 : | |||
| 411 : | return; | ||
| 412 : | |||
| 413 : | case MODE_INTERPOLATE: /* _could_ use DIRECT, but would be overkill (no 4MV there) */ | ||
| 414 : | case MODE_DIRECT_NO4V: | ||
| 415 : | dx = fmvs->x; dy = fmvs->y; | ||
| 416 : | b_dx = bmvs->x; b_dy = bmvs->y; | ||
| 417 : | |||
| 418 : | if (quarterpel) { | ||
| 419 : | syskin | 935 | |
| 420 : | edgomez | 851 | if ((dx&3) | (dy&3)) { |
| 421 : | Isibaar | 1004 | new_interpolate16x16_quarterpel(tmp - i * 16 - j * 16 * edged_width, |
| 422 : | edgomez | 851 | (uint8_t *) f_ref->y, tmp + 32, |
| 423 : | tmp + 64, tmp + 96, 16*i, 16*j, dx, dy, edged_width, 0); | ||
| 424 : | ptr1 = tmp; | ||
| 425 : | } else ptr1 = f_ref->y + (16*j + dy/4)*edged_width + 16*i + dx/4; // fullpixel position | ||
| 426 : | |||
| 427 : | if ((b_dx&3) | (b_dy&3)) { | ||
| 428 : | Isibaar | 1004 | new_interpolate16x16_quarterpel(tmp - i * 16 - j * 16 * edged_width + 16, |
| 429 : | edgomez | 851 | (uint8_t *) b_ref->y, tmp + 32, |
| 430 : | tmp + 64, tmp + 96, 16*i, 16*j, b_dx, b_dy, edged_width, 0); | ||
| 431 : | ptr2 = tmp + 16; | ||
| 432 : | } else ptr2 = b_ref->y + (16*j + b_dy/4)*edged_width + 16*i + b_dx/4; // fullpixel position | ||
| 433 : | |||
| 434 : | b_dx /= 2; | ||
| 435 : | b_dy /= 2; | ||
| 436 : | dx /= 2; | ||
| 437 : | dy /= 2; | ||
| 438 : | |||
| 439 : | } else { | ||
| 440 : | ptr1 = get_ref(f_ref->y, f_refh->y, f_refv->y, f_refhv->y, | ||
| 441 : | i, j, 16, dx, dy, edged_width); | ||
| 442 : | |||
| 443 : | ptr2 = get_ref(b_ref->y, b_refh->y, b_refv->y, b_refhv->y, | ||
| 444 : | i, j, 16, b_dx, b_dy, edged_width); | ||
| 445 : | } | ||
| 446 : | for (k = 0; k < 4; k++) | ||
| 447 : | transfer_8to16sub2(&dct_codes[k * 64], | ||
| 448 : | cur->y + (i * 16+(k&1)*8) + (j * 16+((k>>1)*8)) * edged_width, | ||
| 449 : | ptr1 + (k&1)*8 + (k>>1)*8*edged_width, | ||
| 450 : | ptr2 + (k&1)*8 + (k>>1)*8*edged_width, edged_width); | ||
| 451 : | |||
| 452 : | |||
| 453 : | dx = (dx >> 1) + roundtab_79[dx & 0x3]; | ||
| 454 : | dy = (dy >> 1) + roundtab_79[dy & 0x3]; | ||
| 455 : | |||
| 456 : | b_dx = (b_dx >> 1) + roundtab_79[b_dx & 0x3]; | ||
| 457 : | b_dy = (b_dy >> 1) + roundtab_79[b_dy & 0x3]; | ||
| 458 : | |||
| 459 : | break; | ||
| 460 : | syskin | 935 | |
| 461 : | default: // MODE_DIRECT (or MODE_DIRECT_NONE_MV in case of bframes decoding) | ||
| 462 : | edgomez | 851 | sumx = sumy = b_sumx = b_sumy = 0; |
| 463 : | |||
| 464 : | for (k = 0; k < 4; k++) { | ||
| 465 : | syskin | 935 | |
| 466 : | edgomez | 851 | dx = fmvs[k].x; dy = fmvs[k].y; |
| 467 : | b_dx = bmvs[k].x; b_dy = bmvs[k].y; | ||
| 468 : | |||
| 469 : | if (quarterpel) { | ||
| 470 : | sumx += dx/2; sumy += dy/2; | ||
| 471 : | b_sumx += b_dx/2; b_sumy += b_dy/2; | ||
| 472 : | |||
| 473 : | if ((dx&3) | (dy&3)) { | ||
| 474 : | Isibaar | 1004 | new_interpolate8x8_quarterpel(tmp - (i * 16+(k&1)*8) - (j * 16+((k>>1)*8)) * edged_width, |
| 475 : | syskin | 935 | (uint8_t *) f_ref->y, |
| 476 : | tmp + 32, tmp + 64, tmp + 96, | ||
| 477 : | edgomez | 851 | 16*i + (k&1)*8, 16*j + (k>>1)*8, dx, dy, edged_width, 0); |
| 478 : | ptr1 = tmp; | ||
| 479 : | } else ptr1 = f_ref->y + (16*j + (k>>1)*8 + dy/4)*edged_width + 16*i + (k&1)*8 + dx/4; | ||
| 480 : | |||
| 481 : | if ((b_dx&3) | (b_dy&3)) { | ||
| 482 : | Isibaar | 1004 | new_interpolate8x8_quarterpel(tmp - (i * 16+(k&1)*8) - (j * 16+((k>>1)*8)) * edged_width + 16, |
| 483 : | edgomez | 851 | (uint8_t *) b_ref->y, |
| 484 : | syskin | 935 | tmp + 16, tmp + 32, tmp + 48, |
| 485 : | edgomez | 851 | 16*i + (k&1)*8, 16*j + (k>>1)*8, b_dx, b_dy, edged_width, 0); |
| 486 : | ptr2 = tmp + 16; | ||
| 487 : | } else ptr2 = b_ref->y + (16*j + (k>>1)*8 + b_dy/4)*edged_width + 16*i + (k&1)*8 + b_dx/4; | ||
| 488 : | } else { | ||
| 489 : | sumx += dx; sumy += dy; | ||
| 490 : | b_sumx += b_dx; b_sumy += b_dy; | ||
| 491 : | |||
| 492 : | syskin | 935 | ptr1 = get_ref(f_ref->y, f_refh->y, f_refv->y, f_refhv->y, |
| 493 : | edgomez | 851 | 2*i + (k&1), 2*j + (k>>1), 8, dx, dy, edged_width); |
| 494 : | syskin | 935 | ptr2 = get_ref(b_ref->y, b_refh->y, b_refv->y, b_refhv->y, |
| 495 : | edgomez | 851 | 2*i + (k&1), 2*j + (k>>1), 8, b_dx, b_dy, edged_width); |
| 496 : | } | ||
| 497 : | transfer_8to16sub2(&dct_codes[k * 64], | ||
| 498 : | cur->y + (i * 16+(k&1)*8) + (j * 16+((k>>1)*8)) * edged_width, | ||
| 499 : | ptr1, ptr2, edged_width); | ||
| 500 : | syskin | 935 | |
| 501 : | edgomez | 851 | } |
| 502 : | |||
| 503 : | dx = (sumx >> 3) + roundtab_76[sumx & 0xf]; | ||
| 504 : | dy = (sumy >> 3) + roundtab_76[sumy & 0xf]; | ||
| 505 : | b_dx = (b_sumx >> 3) + roundtab_76[b_sumx & 0xf]; | ||
| 506 : | b_dy = (b_sumy >> 3) + roundtab_76[b_sumy & 0xf]; | ||
| 507 : | |||
| 508 : | break; | ||
| 509 : | } | ||
| 510 : | |||
| 511 : | // uv block-based chroma interpolation for direct and interpolate modes | ||
| 512 : | transfer_8to16sub2(&dct_codes[4 * 64], | ||
| 513 : | cur->u + (j * 8) * edged_width / 2 + (i * 8), | ||
| 514 : | interpolate8x8_switch2(tmp, b_ref->u, 8 * i, 8 * j, | ||
| 515 : | b_dx, b_dy, edged_width / 2, 0), | ||
| 516 : | interpolate8x8_switch2(tmp + 8, f_ref->u, 8 * i, 8 * j, | ||
| 517 : | dx, dy, edged_width / 2, 0), | ||
| 518 : | edged_width / 2); | ||
| 519 : | |||
| 520 : | transfer_8to16sub2(&dct_codes[5 * 64], | ||
| 521 : | cur->v + (j * 8) * edged_width / 2 + (i * 8), | ||
| 522 : | interpolate8x8_switch2(tmp, b_ref->v, 8 * i, 8 * j, | ||
| 523 : | b_dx, b_dy, edged_width / 2, 0), | ||
| 524 : | interpolate8x8_switch2(tmp + 8, f_ref->v, 8 * i, 8 * j, | ||
| 525 : | dx, dy, edged_width / 2, 0), | ||
| 526 : | edged_width / 2); | ||
| 527 : | } | ||
| 528 : | |||
| 529 : | |||
| 530 : | |||
| 531 : | void generate_GMCparameters( const int num_wp, const int res, | ||
| 532 : | syskin | 935 | const WARPPOINTS *const warp, |
| 533 : | const int width, const int height, | ||
| 534 : | GMC_DATA *const gmc) | ||
| 535 : | edgomez | 851 | { |
| 536 : | syskin | 935 | const int du0 = warp->duv[0].x; |
| 537 : | const int dv0 = warp->duv[0].y; | ||
| 538 : | const int du1 = warp->duv[1].x; | ||
| 539 : | const int dv1 = warp->duv[1].y; | ||
| 540 : | const int du2 = warp->duv[2].x; | ||
| 541 : | const int dv2 = warp->duv[2].y; | ||
| 542 : | edgomez | 851 | |
| 543 : | syskin | 935 | gmc->W = width; |
| 544 : | gmc->H = height; | ||
| 545 : | edgomez | 851 | |
| 546 : | syskin | 935 | gmc->rho = 4 - log2bin(res-1); // = {3,2,1,0} for res={2,4,8,16} |
| 547 : | edgomez | 851 | |
| 548 : | syskin | 935 | gmc->alpha = log2bin(gmc->W-1); |
| 549 : | gmc->Ws = (1 << gmc->alpha); | ||
| 550 : | edgomez | 851 | |
| 551 : | syskin | 935 | gmc->dxF = 16*gmc->Ws + RDIV( 8*gmc->Ws*du1, gmc->W ); |
| 552 : | gmc->dxG = RDIV( 8*gmc->Ws*dv1, gmc->W ); | ||
| 553 : | gmc->Fo = (res*du0 + 1) << (gmc->alpha+gmc->rho-1); | ||
| 554 : | gmc->Go = (res*dv0 + 1) << (gmc->alpha+gmc->rho-1); | ||
| 555 : | edgomez | 851 | |
| 556 : | syskin | 935 | if (num_wp==2) { |
| 557 : | gmc->dyF = -gmc->dxG; | ||
| 558 : | gmc->dyG = gmc->dxF; | ||
| 559 : | } else if (num_wp==3) { | ||
| 560 : | gmc->beta = log2bin(gmc->H-1); | ||
| 561 : | gmc->Hs = (1 << gmc->beta); | ||
| 562 : | gmc->dyF = RDIV( 8*gmc->Hs*du2, gmc->H ); | ||
| 563 : | gmc->dyG = 16*gmc->Hs + RDIV( 8*gmc->Hs*dv2, gmc->H ); | ||
| 564 : | if (gmc->beta > gmc->alpha) { | ||
| 565 : | gmc->dxF <<= (gmc->beta - gmc->alpha); | ||
| 566 : | gmc->dxG <<= (gmc->beta - gmc->alpha); | ||
| 567 : | gmc->alpha = gmc->beta; | ||
| 568 : | gmc->Ws = 1<< gmc->beta; | ||
| 569 : | } else { | ||
| 570 : | gmc->dyF <<= gmc->alpha - gmc->beta; | ||
| 571 : | gmc->dyG <<= gmc->alpha - gmc->beta; | ||
| 572 : | } | ||
| 573 : | } | ||
| 574 : | edgomez | 851 | |
| 575 : | syskin | 935 | gmc->cFo = gmc->dxF + gmc->dyF + (1 << (gmc->alpha+gmc->rho+1)); |
| 576 : | gmc->cFo += 16*gmc->Ws*(du0-1); | ||
| 577 : | |||
| 578 : | gmc->cGo = gmc->dxG + gmc->dyG + (1 << (gmc->alpha+gmc->rho+1)); | ||
| 579 : | gmc->cGo += 16*gmc->Ws*(dv0-1); | ||
| 580 : | edgomez | 851 | } |
| 581 : | |||
| 582 : | syskin | 935 | void |
| 583 : | generate_GMCimage( const GMC_DATA *const gmc_data, // [input] precalculated data | ||
| 584 : | const IMAGE *const pRef, // [input] | ||
| 585 : | const int mb_width, | ||
| 586 : | edgomez | 851 | const int mb_height, |
| 587 : | const int stride, | ||
| 588 : | syskin | 935 | const int stride2, |
| 589 : | const int fcode, // [input] some parameters... | ||
| 590 : | const int32_t quarterpel, // [input] for rounding avgMV | ||
| 591 : | const int reduced_resolution, // [input] ignored | ||
| 592 : | edgomez | 851 | const int32_t rounding, // [input] for rounding image data |
| 593 : | syskin | 935 | MACROBLOCK *const pMBs, // [output] average motion vectors |
| 594 : | IMAGE *const pGMC) // [output] full warped image | ||
| 595 : | edgomez | 851 | { |
| 596 : | |||
| 597 : | unsigned int mj,mi; | ||
| 598 : | VECTOR avgMV; | ||
| 599 : | Isibaar | 3 | |
| 600 : | syskin | 935 | for (mj = 0; mj < (unsigned int)mb_height; mj++) |
| 601 : | for (mi = 0; mi < (unsigned int)mb_width; mi++) { | ||
| 602 : | |||
| 603 : | avgMV = generate_GMCimageMB(gmc_data, pRef, mi, mj, | ||
| 604 : | stride, stride2, quarterpel, rounding, pGMC); | ||
| 605 : | |||
| 606 : | pMBs[mj*mb_width+mi].amv.x = gmc_sanitize(avgMV.x, quarterpel, fcode); | ||
| 607 : | pMBs[mj*mb_width+mi].amv.y = gmc_sanitize(avgMV.y, quarterpel, fcode); | ||
| 608 : | pMBs[mj*mb_width+mi].mcsel = 0; /* until mode decision */ | ||
| 609 : | edgomez | 851 | } |
| 610 : | } | ||
| 611 : | Isibaar | 3 | |
| 612 : | |||
| 613 : | |||
| 614 : | edgomez | 851 | #define MLT(i) (((16-(i))<<16) + (i)) |
| 615 : | static const uint32_t MTab[16] = { | ||
| 616 : | MLT( 0), MLT( 1), MLT( 2), MLT( 3), MLT( 4), MLT( 5), MLT( 6), MLT(7), | ||
| 617 : | MLT( 8), MLT( 9), MLT(10), MLT(11), MLT(12), MLT(13), MLT(14), MLT(15) | ||
| 618 : | }; | ||
| 619 : | #undef MLT | ||
| 620 : | Isibaar | 3 | |
| 621 : | edgomez | 851 | VECTOR generate_GMCimageMB( const GMC_DATA *const gmc_data, |
| 622 : | syskin | 935 | const IMAGE *const pRef, |
| 623 : | const int mi, const int mj, | ||
| 624 : | const int stride, | ||
| 625 : | const int stride2, | ||
| 626 : | const int quarterpel, | ||
| 627 : | const int rounding, | ||
| 628 : | IMAGE *const pGMC) | ||
| 629 : | edgomez | 851 | { |
| 630 : | syskin | 935 | const int W = gmc_data->W; |
| 631 : | const int H = gmc_data->H; | ||
| 632 : | edgomez | 851 | |
| 633 : | syskin | 935 | const int rho = gmc_data->rho; |
| 634 : | const int alpha = gmc_data->alpha; | ||
| 635 : | edgomez | 851 | |
| 636 : | syskin | 935 | const int rounder = ( 128 - (rounding<<(rho+rho)) ) << 16; |
| 637 : | edgomez | 851 | |
| 638 : | syskin | 935 | const int dxF = gmc_data->dxF; |
| 639 : | const int dyF = gmc_data->dyF; | ||
| 640 : | const int dxG = gmc_data->dxG; | ||
| 641 : | const int dyG = gmc_data->dyG; | ||
| 642 : | edgomez | 851 | |
| 643 : | syskin | 935 | uint8_t *dstY, *dstU, *dstV; |
| 644 : | edgomez | 851 | |
| 645 : | syskin | 935 | int I,J; |
| 646 : | VECTOR avgMV = {0,0}; | ||
| 647 : | edgomez | 851 | |
| 648 : | syskin | 935 | int32_t Fj, Gj; |
| 649 : | edgomez | 851 | |
| 650 : | syskin | 935 | dstY = &pGMC->y[(mj*16)*stride+mi*16] + 16; |
| 651 : | edgomez | 851 | |
| 652 : | syskin | 935 | Fj = gmc_data->Fo + dyF*mj*16 + dxF*mi*16; |
| 653 : | Gj = gmc_data->Go + dyG*mj*16 + dxG*mi*16; | ||
| 654 : | |||
| 655 : | for (J = 16; J > 0; --J) { | ||
| 656 : | int32_t Fi, Gi; | ||
| 657 : | edgomez | 851 | |
| 658 : | syskin | 935 | Fi = Fj; Fj += dyF; |
| 659 : | Gi = Gj; Gj += dyG; | ||
| 660 : | for (I = -16; I < 0; ++I) { | ||
| 661 : | int32_t F, G; | ||
| 662 : | uint32_t ri, rj; | ||
| 663 : | edgomez | 851 | |
| 664 : | syskin | 935 | F = ( Fi >> (alpha+rho) ) << rho; Fi += dxF; |
| 665 : | G = ( Gi >> (alpha+rho) ) << rho; Gi += dxG; | ||
| 666 : | edgomez | 851 | |
| 667 : | syskin | 935 | avgMV.x += F; |
| 668 : | avgMV.y += G; | ||
| 669 : | edgomez | 851 | |
| 670 : | syskin | 935 | ri = MTab[F&15]; |
| 671 : | rj = MTab[G&15]; | ||
| 672 : | edgomez | 851 | |
| 673 : | syskin | 935 | F >>= 4; |
| 674 : | G >>= 4; | ||
| 675 : | edgomez | 851 | |
| 676 : | syskin | 935 | if (F < -1) F = -1; |
| 677 : | else if (F > W) F = W; | ||
| 678 : | if (G< -1) G=-1; | ||
| 679 : | else if (G>H) G=H; | ||
| 680 : | edgomez | 851 | |
| 681 : | syskin | 935 | { // MMX-like bilinear... |
| 682 : | const int offset = G*stride + F; | ||
| 683 : | uint32_t f0, f1; | ||
| 684 : | f0 = pRef->y[ offset +0 ]; | ||
| 685 : | f0 |= pRef->y[ offset +1 ] << 16; | ||
| 686 : | f1 = pRef->y[ offset+stride +0 ]; | ||
| 687 : | f1 |= pRef->y[ offset+stride +1 ] << 16; | ||
| 688 : | f0 = (ri*f0)>>16; | ||
| 689 : | f1 = (ri*f1) & 0x0fff0000; | ||
| 690 : | f0 |= f1; | ||
| 691 : | f0 = ( rj*f0 + rounder ) >> 24; | ||
| 692 : | edgomez | 851 | |
| 693 : | syskin | 935 | dstY[I] = (uint8_t)f0; |
| 694 : | } | ||
| 695 : | } | ||
| 696 : | |||
| 697 : | dstY += stride; | ||
| 698 : | } | ||
| 699 : | edgomez | 851 | |
| 700 : | syskin | 935 | dstU = &pGMC->u[(mj*8)*stride2+mi*8] + 8; |
| 701 : | dstV = &pGMC->v[(mj*8)*stride2+mi*8] + 8; | ||
| 702 : | edgomez | 851 | |
| 703 : | syskin | 935 | Fj = gmc_data->cFo + dyF*4 *mj*8 + dxF*4 *mi*8; |
| 704 : | Gj = gmc_data->cGo + dyG*4 *mj*8 + dxG*4 *mi*8; | ||
| 705 : | |||
| 706 : | for (J = 8; J > 0; --J) { | ||
| 707 : | int32_t Fi, Gi; | ||
| 708 : | Fi = Fj; Fj += 4*dyF; | ||
| 709 : | Gi = Gj; Gj += 4*dyG; | ||
| 710 : | edgomez | 851 | |
| 711 : | syskin | 935 | for (I = -8; I < 0; ++I) { |
| 712 : | int32_t F, G; | ||
| 713 : | uint32_t ri, rj; | ||
| 714 : | edgomez | 851 | |
| 715 : | syskin | 935 | F = ( Fi >> (alpha+rho+2) ) << rho; Fi += 4*dxF; |
| 716 : | G = ( Gi >> (alpha+rho+2) ) << rho; Gi += 4*dxG; | ||
| 717 : | edgomez | 851 | |
| 718 : | syskin | 935 | ri = MTab[F&15]; |
| 719 : | rj = MTab[G&15]; | ||
| 720 : | edgomez | 851 | |
| 721 : | syskin | 935 | F >>= 4; |
| 722 : | G >>= 4; | ||
| 723 : | edgomez | 851 | |
| 724 : | syskin | 935 | if (F < -1) F=-1; |
| 725 : | else if (F >= W/2) F = W/2; | ||
| 726 : | if (G < -1) G = -1; | ||
| 727 : | else if (G >= H/2) G = H/2; | ||
| 728 : | edgomez | 851 | |
| 729 : | syskin | 935 | { |
| 730 : | const int offset = G*stride2 + F; | ||
| 731 : | uint32_t f0, f1; | ||
| 732 : | edgomez | 851 | |
| 733 : | syskin | 935 | f0 = pRef->u[ offset +0 ]; |
| 734 : | f0 |= pRef->u[ offset +1 ] << 16; | ||
| 735 : | f1 = pRef->u[ offset+stride2 +0 ]; | ||
| 736 : | f1 |= pRef->u[ offset+stride2 +1 ] << 16; | ||
| 737 : | f0 = (ri*f0)>>16; | ||
| 738 : | f1 = (ri*f1) & 0x0fff0000; | ||
| 739 : | f0 |= f1; | ||
| 740 : | f0 = ( rj*f0 + rounder ) >> 24; | ||
| 741 : | edgomez | 851 | |
| 742 : | syskin | 935 | dstU[I] = (uint8_t)f0; |
| 743 : | edgomez | 851 | |
| 744 : | |||
| 745 : | syskin | 935 | f0 = pRef->v[ offset +0 ]; |
| 746 : | f0 |= pRef->v[ offset +1 ] << 16; | ||
| 747 : | f1 = pRef->v[ offset+stride2 +0 ]; | ||
| 748 : | f1 |= pRef->v[ offset+stride2 +1 ] << 16; | ||
| 749 : | f0 = (ri*f0)>>16; | ||
| 750 : | f1 = (ri*f1) & 0x0fff0000; | ||
| 751 : | f0 |= f1; | ||
| 752 : | f0 = ( rj*f0 + rounder ) >> 24; | ||
| 753 : | edgomez | 851 | |
| 754 : | syskin | 935 | dstV[I] = (uint8_t)f0; |
| 755 : | } | ||
| 756 : | } | ||
| 757 : | dstU += stride2; | ||
| 758 : | dstV += stride2; | ||
| 759 : | } | ||
| 760 : | edgomez | 851 | |
| 761 : | |||
| 762 : | syskin | 935 | avgMV.x -= 16*((256*mi+120)<<4); // 120 = 15*16/2 |
| 763 : | avgMV.y -= 16*((256*mj+120)<<4); | ||
| 764 : | |||
| 765 : | avgMV.x = RSHIFT( avgMV.x, (4+7-quarterpel) ); | ||
| 766 : | avgMV.y = RSHIFT( avgMV.y, (4+7-quarterpel) ); | ||
| 767 : | |||
| 768 : | return avgMV; | ||
| 769 : | edgomez | 851 | } |
| 770 : | syskin | 935 | |
| 771 : | |||
| 772 : | |||
| 773 : | edgomez | 851 | #ifdef OLD_GRUEL_GMC |
| 774 : | syskin | 935 | void |
| 775 : | edgomez | 851 | generate_GMCparameters( const int num_wp, // [input]: number of warppoints |
| 776 : | syskin | 935 | const int res, // [input]: resolution |
| 777 : | const WARPPOINTS *const warp, // [input]: warp points | ||
| 778 : | edgomez | 851 | const int width, const int height, |
| 779 : | syskin | 935 | GMC_DATA *const gmc) // [output] precalculated parameters |
| 780 : | edgomez | 851 | { |
| 781 : | |||
| 782 : | syskin | 935 | /* We follow mainly two sources: The original standard, which is ugly, and the |
| 783 : | thesis from Andreas Dehnhardt, which is much nicer. | ||
| 784 : | edgomez | 851 | |
| 785 : | syskin | 935 | Notation is: indices are written next to the variable, |
| 786 : | primes in the standard are denoted by a suffix 'p'. | ||
| 787 : | types are "c"=constant, "i"=input parameter, "f"=calculated, then fixed, | ||
| 788 : | "o"=output data, " "=other, "u" = unused, "p"=calc for every pixel | ||
| 789 : | |||
| 790 : | type | variable name | ISO name (TeX-style) | value or range | usage | ||
| 791 : | edgomez | 851 | ------------------------------------------------------------------------------------- |
| 792 : | c | H | H | [16 , ?] | image width (w/o edges) | ||
| 793 : | c | W | W | [16 , ?] | image height (w/o edges) | ||
| 794 : | |||
| 795 : | c | i0 | i_0 | 0 | ref. point #1, X | ||
| 796 : | c | j0 | j_0 | 0 | ref. point #1, Y | ||
| 797 : | c | i1 | i_1 | W | ref. point #2, X | ||
| 798 : | c | j1 | j_1 | 0 | ref. point #2, Y | ||
| 799 : | cu | i2 | i_2 | 0 | ref. point #3, X | ||
| 800 : | syskin | 935 | cu | i2 | j_2 | H | ref. point #3, Y |
| 801 : | edgomez | 851 | |
| 802 : | i | du0 | du[0] | [-16863,16863] | warp vector #1, Y | ||
| 803 : | i | dv0 | dv[0] | [-16863,16863] | warp vector #1, Y | ||
| 804 : | i | du1 | du[1] | [-16863,16863] | warp vector #2, Y | ||
| 805 : | i | dv1 | dv[1] | [-16863,16863] | warp vector #2, Y | ||
| 806 : | syskin | 935 | iu | du2 | du[2] | [-16863,16863] | warp vector #3, Y |
| 807 : | edgomez | 851 | iu | dv2 | dv[2] | [-16863,16863] | warp vector #3, Y |
| 808 : | |||
| 809 : | syskin | 935 | i | s | s | {2,4,8,16} | interpol. resolution |
| 810 : | f | sigma | - | log2(s) | X / s == X >> sigma | ||
| 811 : | f | r | r | =16/s | complementary res. | ||
| 812 : | f | rho | \rho | log2(r) | X / r == X >> rho | ||
| 813 : | |||
| 814 : | f | i0s | i'_0 | | | ||
| 815 : | f | j0s | j'_0 | | | ||
| 816 : | f | i1s | i'_1 | | | ||
| 817 : | f | j1s | j'_1 | | | ||
| 818 : | f | i2s | i'_2 | | | ||
| 819 : | f | j2s | j'_2 | | | ||
| 820 : | |||
| 821 : | f | alpha | \alpha | | 2^{alpha-1} < W <= 2^alpha | ||
| 822 : | f | beta | \beta | | 2^{beta-1} < H <= 2^beta | ||
| 823 : | |||
| 824 : | f | Ws | W' | W = 2^{alpha} | scaled width | ||
| 825 : | f | Hs | H' | W = 2^{beta} | scaled height | ||
| 826 : | |||
| 827 : | f | i1ss | i''_1 | "virtual sprite stuff" | ||
| 828 : | f | j1ss | j''_1 | "virtual sprite stuff" | ||
| 829 : | f | i2ss | i''_2 | "virtual sprite stuff" | ||
| 830 : | f | j2ss | j''_2 | "virtual sprite stuff" | ||
| 831 : | edgomez | 851 | */ |
| 832 : | |||
| 833 : | /* Some calculations are disabled because we only use 2 warppoints at the moment */ | ||
| 834 : | |||
| 835 : | int du0 = warp->duv[0].x; | ||
| 836 : | int dv0 = warp->duv[0].y; | ||
| 837 : | int du1 = warp->duv[1].x; | ||
| 838 : | int dv1 = warp->duv[1].y; | ||
| 839 : | // int du2 = warp->duv[2].x; | ||
| 840 : | // int dv2 = warp->duv[2].y; | ||
| 841 : | syskin | 935 | |
| 842 : | edgomez | 851 | gmc->num_wp = num_wp; |
| 843 : | |||
| 844 : | gmc->s = res; /* scaling parameters 2,4,8 or 16 */ | ||
| 845 : | gmc->sigma = log2bin(res-1); /* log2bin(15)=4, log2bin(16)=5, log2bin(17)=5 */ | ||
| 846 : | syskin | 935 | gmc->r = 16/res; |
| 847 : | edgomez | 851 | gmc->rho = 4 - gmc->sigma; /* = log2bin(r-1) */ |
| 848 : | |||
| 849 : | gmc->W = width; | ||
| 850 : | gmc->H = height; /* fixed reference coordinates */ | ||
| 851 : | |||
| 852 : | gmc->alpha = log2bin(gmc->W-1); | ||
| 853 : | syskin | 935 | gmc->Ws= 1<<gmc->alpha; |
| 854 : | edgomez | 851 | |
| 855 : | // gmc->beta = log2bin(gmc->H-1); | ||
| 856 : | syskin | 935 | // gmc->Hs= 1<<gmc->beta; |
| 857 : | edgomez | 851 | |
| 858 : | // 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); | ||
| 859 : | |||
| 860 : | /* i2s is only needed for num_wp >= 3, etc. */ | ||
| 861 : | /* the 's' values are in 1/s pel resolution */ | ||
| 862 : | gmc->i0s = res/2 * ( du0 ); | ||
| 863 : | gmc->j0s = res/2 * ( dv0 ); | ||
| 864 : | gmc->i1s = res/2 * (2*width + du1 + du0 ); | ||
| 865 : | gmc->j1s = res/2 * ( dv1 + dv0 ); | ||
| 866 : | // gmc->i2s = res/2 * ( du2 + du0 ); | ||
| 867 : | // gmc->j2s = res/2 * (2*height + dv2 + dv0 ); | ||
| 868 : | syskin | 935 | |
| 869 : | edgomez | 851 | /* i2s and i2ss are only needed for num_wp == 3, etc. */ |
| 870 : | |||
| 871 : | /* the 'ss' values are in 1/16 pel resolution */ | ||
| 872 : | 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); |
| 873 : | gmc->j1ss = ROUNDED_DIV( ((gmc->W - gmc->Ws)*(gmc->r*gmc->j0s) + gmc->Ws*gmc->r*gmc->j1s) ,gmc->W ); | ||
| 874 : | edgomez | 851 | |
| 875 : | syskin | 935 | // gmc->i2ss = ROUNDED_DIV( ((gmc->H - gmc->Hs)*(gmc->r*gmc->i0s) + gmc->Hs*(gmc->r*gmc->i2s)), gmc->H); |
| 876 : | edgomez | 851 | // 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); |
| 877 : | |||
| 878 : | syskin | 935 | return; |
| 879 : | edgomez | 851 | } |
| 880 : | |||
| 881 : | syskin | 935 | void |
| 882 : | edgomez | 851 | generate_GMCimage( const GMC_DATA *const gmc_data, // [input] precalculated data |
| 883 : | syskin | 935 | const IMAGE *const pRef, // [input] |
| 884 : | const int mb_width, | ||
| 885 : | edgomez | 851 | const int mb_height, |
| 886 : | const int stride, | ||
| 887 : | syskin | 935 | const int stride2, |
| 888 : | edgomez | 851 | const int fcode, // [input] some parameters... |
| 889 : | const int32_t quarterpel, // [input] for rounding avgMV | ||
| 890 : | const int reduced_resolution, // [input] ignored | ||
| 891 : | const int32_t rounding, // [input] for rounding image data | ||
| 892 : | MACROBLOCK *const pMBs, // [output] average motion vectors | ||
| 893 : | IMAGE *const pGMC) // [output] full warped image | ||
| 894 : | { | ||
| 895 : | |||
| 896 : | unsigned int mj,mi; | ||
| 897 : | VECTOR avgMV; | ||
| 898 : | syskin | 935 | |
| 899 : | for (mj = 0;mj < mb_height; mj++) | ||
| 900 : | for (mi = 0;mi < mb_width; mi++) { | ||
| 901 : | |||
| 902 : | avgMV = generate_GMCimageMB(gmc_data, pRef, mi, mj, | ||
| 903 : | edgomez | 851 | stride, stride2, quarterpel, rounding, pGMC); |
| 904 : | |||
| 905 : | pMBs[mj*mb_width+mi].amv.x = gmc_sanitize(avgMV.x, quarterpel, fcode); | ||
| 906 : | pMBs[mj*mb_width+mi].amv.y = gmc_sanitize(avgMV.y, quarterpel, fcode); | ||
| 907 : | pMBs[mj*mb_width+mi].mcsel = 0; /* until mode decision */ | ||
| 908 : | Isibaar | 3 | } |
| 909 : | } | ||
| 910 : | edgomez | 851 | |
| 911 : | |||
| 912 : | VECTOR generate_GMCimageMB( const GMC_DATA *const gmc_data, /* [input] all precalc data */ | ||
| 913 : | const IMAGE *const pRef, /* [input] */ | ||
| 914 : | const int mi, const int mj, /* [input] MB position */ | ||
| 915 : | const int stride, /* [input] Lumi stride */ | ||
| 916 : | const int stride2, /* [input] chroma stride */ | ||
| 917 : | const int quarterpel, /* [input] for rounding of avgMV */ | ||
| 918 : | const int rounding, /* [input] for rounding of imgae data */ | ||
| 919 : | IMAGE *const pGMC) /* [outut] generate image */ | ||
| 920 : | syskin | 935 | |
| 921 : | edgomez | 851 | /* |
| 922 : | syskin | 935 | type | variable name | ISO name (TeX-style) | value or range | usage |
| 923 : | edgomez | 851 | ------------------------------------------------------------------------------------- |
| 924 : | syskin | 935 | p | F | F(i,j) | | pelwise motion vector X in s-th pel |
| 925 : | p | G | G(i,j) | | pelwise motion vector Y in s-th pel | ||
| 926 : | p | Fc | F_c(i,j) | | | ||
| 927 : | p | Gc | G_c(i,j) | | same for chroma | ||
| 928 : | edgomez | 851 | |
| 929 : | syskin | 935 | p | Y00 | Y_{00} | [0,255*s*s] | first: 4 neighbouring Y-values |
| 930 : | p | Y01 | Y_{01} | [0,255] | at fullpel position, around the | ||
| 931 : | p | Y10 | Y_{10} | [0,255*s] | position where pelweise MV points to | ||
| 932 : | p | Y11 | Y_{11} | [0,255] | later: bilinear interpol Y-values in Y00 | ||
| 933 : | edgomez | 851 | |
| 934 : | syskin | 935 | p | C00 | C_{00} | [0,255*s*s] | same for chroma Cb and Cr |
| 935 : | p | C01 | C_{01} | [0,255] | | ||
| 936 : | p | C10 | C_{10} | [0,255*s] | | ||
| 937 : | p | C11 | C_{11} | [0,255] | | ||
| 938 : | |||
| 939 : | edgomez | 851 | */ |
| 940 : | { | ||
| 941 : | const int W = gmc_data->W; | ||
| 942 : | const int H = gmc_data->H; | ||
| 943 : | |||
| 944 : | const int s = gmc_data->s; | ||
| 945 : | const int sigma = gmc_data->sigma; | ||
| 946 : | syskin | 935 | |
| 947 : | edgomez | 851 | const int r = gmc_data->r; |
| 948 : | syskin | 935 | const int rho = gmc_data->rho; |
| 949 : | |||
| 950 : | edgomez | 851 | const int i0s = gmc_data->i0s; |
| 951 : | const int j0s = gmc_data->j0s; | ||
| 952 : | syskin | 935 | |
| 953 : | edgomez | 851 | const int i1ss = gmc_data->i1ss; |
| 954 : | syskin | 935 | const int j1ss = gmc_data->j1ss; |
| 955 : | // const int i2ss = gmc_data->i2ss; | ||
| 956 : | // const int j2ss = gmc_data->j2ss; | ||
| 957 : | edgomez | 851 | |
| 958 : | syskin | 935 | const int alpha = gmc_data->alpha; |
| 959 : | const int Ws = gmc_data->Ws; | ||
| 960 : | edgomez | 851 | |
| 961 : | // const int beta = gmc_data->beta; | ||
| 962 : | syskin | 935 | // const int Hs = gmc_data->Hs; |
| 963 : | edgomez | 851 | |
| 964 : | int I,J; | ||
| 965 : | VECTOR avgMV = {0,0}; | ||
| 966 : | syskin | 935 | |
| 967 : | edgomez | 851 | for (J=16*mj;J<16*(mj+1);J++) |
| 968 : | for (I=16*mi;I<16*(mi+1);I++) | ||
| 969 : | { | ||
| 970 : | int F= i0s + ( ((-r*i0s+i1ss)*I + (r*j0s-j1ss)*J + (1<<(alpha+rho-1))) >> (alpha+rho) ); | ||
| 971 : | int G= j0s + ( ((-r*j0s+j1ss)*I + (-r*i0s+i1ss)*J + (1<<(alpha+rho-1))) >> (alpha+rho) ); | ||
| 972 : | |||
| 973 : | syskin | 935 | /* this naive implementation (with lots of multiplications) isn't slower (rather faster) than |
| 974 : | edgomez | 851 | working incremental. Don't ask me why... maybe the whole this is memory bound? */ |
| 975 : | |||
| 976 : | const int ri= F & (s-1); // fractional part of pelwise MV X | ||
| 977 : | const int rj= G & (s-1); // fractional part of pelwise MV Y | ||
| 978 : | syskin | 935 | |
| 979 : | edgomez | 851 | int Y00,Y01,Y10,Y11; |
| 980 : | |||
| 981 : | /* unclipped values are used for avgMV */ | ||
| 982 : | avgMV.x += F-(I<<sigma); /* shift position to 1/s-pel, as the MV is */ | ||
| 983 : | avgMV.y += G-(J<<sigma); /* TODO: don't do this (of course) */ | ||
| 984 : | |||
| 985 : | F >>= sigma; | ||
| 986 : | G >>= sigma; | ||
| 987 : | |||
| 988 : | syskin | 935 | /* clip values to be in range. Since we have edges, clip to 1 less than lower boundary |
| 989 : | edgomez | 851 | this way positions F+1/G+1 are still right */ |
| 990 : | |||
| 991 : | if (F< -1) | ||
| 992 : | F=-1; | ||
| 993 : | else if (F>W) | ||
| 994 : | F=W; /* W or W-1 doesn't matter, so save 1 subtract ;-) */ | ||
| 995 : | if (G< -1) | ||
| 996 : | G=-1; | ||
| 997 : | else if (G>H) | ||
| 998 : | G=H; /* dito */ | ||
| 999 : | |||
| 1000 : | Y00 = pRef->y[ G*stride + F ]; // Lumi values | ||
| 1001 : | Y01 = pRef->y[ G*stride + F+1 ]; | ||
| 1002 : | Y10 = pRef->y[ G*stride + F+stride ]; | ||
| 1003 : | Y11 = pRef->y[ G*stride + F+stride+1 ]; | ||
| 1004 : | syskin | 935 | |
| 1005 : | edgomez | 851 | /* bilinear interpolation */ |
| 1006 : | Y00 = ((s-ri)*Y00 + ri*Y01); | ||
| 1007 : | Y10 = ((s-ri)*Y10 + ri*Y11); | ||
| 1008 : | syskin | 935 | Y00 = ((s-rj)*Y00 + rj*Y10 + s*s/2 - rounding ) >> (sigma+sigma); |
| 1009 : | |||
| 1010 : | edgomez | 851 | pGMC->y[J*stride+I] = (uint8_t)Y00; /* output 1 Y-pixel */ |
| 1011 : | } | ||
| 1012 : | |||
| 1013 : | syskin | 935 | |
| 1014 : | /* doing chroma _here_ is even more stupid and slow, because won't be used until Compensation and | ||
| 1015 : | edgomez | 851 | most likely not even then (only if the block really _is_ GMC) |
| 1016 : | */ | ||
| 1017 : | |||
| 1018 : | for (J=8*mj;J<8*(mj+1);J++) /* this plays the role of j_c,i_c in the standard */ | ||
| 1019 : | for (I=8*mi;I<8*(mi+1);I++) /* For I_c we have to use I_c = 4*i_c+1 ! */ | ||
| 1020 : | { | ||
| 1021 : | /* same positions for both chroma components, U=Cb and V=Cr */ | ||
| 1022 : | syskin | 935 | int Fc=((-r*i0s+i1ss)*(4*I+1) + (r*j0s-j1ss)*(4*J+1) +2*Ws*r*i0s |
| 1023 : | edgomez | 851 | -16*Ws +(1<<(alpha+rho+1)))>>(alpha+rho+2); |
| 1024 : | syskin | 935 | int Gc=((-r*j0s+j1ss)*(4*I+1) +(-r*i0s+i1ss)*(4*J+1) +2*Ws*r*j0s |
| 1025 : | edgomez | 851 | -16*Ws +(1<<(alpha+rho+1))) >>(alpha+rho+2); |
| 1026 : | syskin | 935 | |
| 1027 : | edgomez | 851 | const int ri= Fc & (s-1); // fractional part of pelwise MV X |
| 1028 : | const int rj= Gc & (s-1); // fractional part of pelwise MV Y | ||
| 1029 : | |||
| 1030 : | int C00,C01,C10,C11; | ||
| 1031 : | syskin | 935 | |
| 1032 : | edgomez | 851 | Fc >>= sigma; |
| 1033 : | Gc >>= sigma; | ||
| 1034 : | syskin | 935 | |
| 1035 : | edgomez | 851 | if (Fc< -1) |
| 1036 : | Fc=-1; | ||
| 1037 : | else if (Fc>=W/2) | ||
| 1038 : | Fc=W/2; /* W or W-1 doesn't matter, so save 1 subtraction ;-) */ | ||
| 1039 : | if (Gc< -1) | ||
| 1040 : | Gc=-1; | ||
| 1041 : | else if (Gc>=H/2) | ||
| 1042 : | Gc=H/2; /* dito */ | ||
| 1043 : | |||
| 1044 : | syskin | 935 | /* now calculate U data */ |
| 1045 : | edgomez | 851 | C00 = pRef->u[ Gc*stride2 + Fc ]; // chroma-value Cb |
| 1046 : | C01 = pRef->u[ Gc*stride2 + Fc+1 ]; | ||
| 1047 : | C10 = pRef->u[ (Gc+1)*stride2 + Fc ]; | ||
| 1048 : | C11 = pRef->u[ (Gc+1)*stride2 + Fc+1 ]; | ||
| 1049 : | syskin | 935 | |
| 1050 : | edgomez | 851 | /* bilinear interpolation */ |
| 1051 : | C00 = ((s-ri)*C00 + ri*C01); | ||
| 1052 : | C10 = ((s-ri)*C10 + ri*C11); | ||
| 1053 : | syskin | 935 | C00 = ((s-rj)*C00 + rj*C10 + s*s/2 - rounding ) >> (sigma+sigma); |
| 1054 : | |||
| 1055 : | edgomez | 851 | pGMC->u[J*stride2+I] = (uint8_t)C00; /* output 1 U-pixel */ |
| 1056 : | syskin | 935 | |
| 1057 : | edgomez | 851 | /* now calculate V data */ |
| 1058 : | C00 = pRef->v[ Gc*stride2 + Fc ]; // chroma-value Cr | ||
| 1059 : | C01 = pRef->v[ Gc*stride2 + Fc+1 ]; | ||
| 1060 : | C10 = pRef->v[ (Gc+1)*stride2 + Fc ]; | ||
| 1061 : | C11 = pRef->v[ (Gc+1)*stride2 + Fc+1 ]; | ||
| 1062 : | syskin | 935 | |
| 1063 : | edgomez | 851 | /* bilinear interpolation */ |
| 1064 : | C00 = ((s-ri)*C00 + ri*C01); | ||
| 1065 : | C10 = ((s-ri)*C10 + ri*C11); | ||
| 1066 : | syskin | 935 | C00 = ((s-rj)*C00 + rj*C10 + s*s/2 - rounding ) >> (sigma+sigma); |
| 1067 : | |||
| 1068 : | edgomez | 851 | pGMC->v[J*stride2+I] = (uint8_t)C00; /* output 1 V-pixel */ |
| 1069 : | } | ||
| 1070 : | syskin | 935 | |
| 1071 : | edgomez | 851 | /* The average vector is rounded from 1/s-pel to 1/2 or 1/4 using the '//' operator*/ |
| 1072 : | |||
| 1073 : | avgMV.x = RSHIFT( avgMV.x, (sigma+7-quarterpel) ); | ||
| 1074 : | avgMV.y = RSHIFT( avgMV.y, (sigma+7-quarterpel) ); | ||
| 1075 : | |||
| 1076 : | /* ^^^^ this is the way MS Reference Software does it */ | ||
| 1077 : | syskin | 935 | |
| 1078 : | edgomez | 851 | return avgMV; /* clipping to fcode area is done outside! */ |
| 1079 : | } | ||
| 1080 : | |||
| 1081 : | #endif |
| No admin address has been configured | ViewVC Help |
| Powered by ViewVC 1.0.4 |