| 1 | /***************************************************************************** |
| 2 | * Copyright (C) 2013 x265 project |
| 3 | * |
| 4 | * Authors: Steve Borho <steve@borho.org> |
| 5 | * |
| 6 | * This program is free software; you can redistribute it and/or modify |
| 7 | * it under the terms of the GNU General Public License as published by |
| 8 | * the Free Software Foundation; either version 2 of the License, or |
| 9 | * (at your option) any later version. |
| 10 | * |
| 11 | * This program is distributed in the hope that it will be useful, |
| 12 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 13 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 14 | * GNU General Public License for more details. |
| 15 | * |
| 16 | * You should have received a copy of the GNU General Public License |
| 17 | * along with this program; if not, write to the Free Software |
| 18 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111, USA. |
| 19 | * |
| 20 | * This program is also available under a commercial proprietary license. |
| 21 | * For more information, contact us at license @ x265.com. |
| 22 | *****************************************************************************/ |
| 23 | |
| 24 | #include "common.h" |
| 25 | #include "primitives.h" |
| 26 | #include "lowres.h" |
| 27 | #include "motion.h" |
| 28 | #include "x265.h" |
| 29 | |
| 30 | #if _MSC_VER |
| 31 | #pragma warning(disable: 4127) // conditional expression is constant (macros use this construct) |
| 32 | #endif |
| 33 | |
| 34 | using namespace x265; |
| 35 | |
| 36 | namespace { |
| 37 | |
| 38 | struct SubpelWorkload |
| 39 | { |
| 40 | int hpel_iters; |
| 41 | int hpel_dirs; |
| 42 | int qpel_iters; |
| 43 | int qpel_dirs; |
| 44 | bool hpel_satd; |
| 45 | }; |
| 46 | |
| 47 | const SubpelWorkload workload[X265_MAX_SUBPEL_LEVEL + 1] = |
| 48 | { |
| 49 | { 1, 4, 0, 4, false }, // 4 SAD HPEL only |
| 50 | { 1, 4, 1, 4, false }, // 4 SAD HPEL + 4 SATD QPEL |
| 51 | { 1, 4, 1, 4, true }, // 4 SATD HPEL + 4 SATD QPEL |
| 52 | { 2, 4, 1, 4, true }, // 2x4 SATD HPEL + 4 SATD QPEL |
| 53 | { 2, 4, 2, 4, true }, // 2x4 SATD HPEL + 2x4 SATD QPEL |
| 54 | { 1, 8, 1, 8, true }, // 8 SATD HPEL + 8 SATD QPEL (default) |
| 55 | { 2, 8, 1, 8, true }, // 2x8 SATD HPEL + 8 SATD QPEL |
| 56 | { 2, 8, 2, 8, true }, // 2x8 SATD HPEL + 2x8 SATD QPEL |
| 57 | }; |
| 58 | |
| 59 | int sizeScale[NUM_LUMA_PARTITIONS]; |
| 60 | #define SAD_THRESH(v) (bcost < (((v >> 4) * sizeScale[partEnum]))) |
| 61 | |
| 62 | void initScales(void) |
| 63 | { |
| 64 | #define SETUP_SCALE(W, H) \ |
| 65 | sizeScale[LUMA_ ## W ## x ## H] = (H * H) >> 4; |
| 66 | SETUP_SCALE(4, 4); |
| 67 | SETUP_SCALE(8, 8); |
| 68 | SETUP_SCALE(8, 4); |
| 69 | SETUP_SCALE(4, 8); |
| 70 | SETUP_SCALE(16, 16); |
| 71 | SETUP_SCALE(16, 8); |
| 72 | SETUP_SCALE(8, 16); |
| 73 | SETUP_SCALE(16, 12); |
| 74 | SETUP_SCALE(12, 16); |
| 75 | SETUP_SCALE(4, 16); |
| 76 | SETUP_SCALE(16, 4); |
| 77 | SETUP_SCALE(32, 32); |
| 78 | SETUP_SCALE(32, 16); |
| 79 | SETUP_SCALE(16, 32); |
| 80 | SETUP_SCALE(32, 24); |
| 81 | SETUP_SCALE(24, 32); |
| 82 | SETUP_SCALE(32, 8); |
| 83 | SETUP_SCALE(8, 32); |
| 84 | SETUP_SCALE(64, 64); |
| 85 | SETUP_SCALE(64, 32); |
| 86 | SETUP_SCALE(32, 64); |
| 87 | SETUP_SCALE(64, 48); |
| 88 | SETUP_SCALE(48, 64); |
| 89 | SETUP_SCALE(64, 16); |
| 90 | SETUP_SCALE(16, 64); |
| 91 | #undef SETUP_SCALE |
| 92 | } |
| 93 | |
| 94 | /* radius 2 hexagon. repeated entries are to avoid having to compute mod6 every time. */ |
| 95 | const MV hex2[8] = { MV(-1, -2), MV(-2, 0), MV(-1, 2), MV(1, 2), MV(2, 0), MV(1, -2), MV(-1, -2), MV(-2, 0) }; |
| 96 | const uint8_t mod6m1[8] = { 5, 0, 1, 2, 3, 4, 5, 0 }; /* (x-1)%6 */ |
| 97 | const MV square1[9] = { MV(0, 0), MV(0, -1), MV(0, 1), MV(-1, 0), MV(1, 0), MV(-1, -1), MV(-1, 1), MV(1, -1), MV(1, 1) }; |
| 98 | const MV hex4[16] = |
| 99 | { |
| 100 | MV(0, -4), MV(0, 4), MV(-2, -3), MV(2, -3), |
| 101 | MV(-4, -2), MV(4, -2), MV(-4, -1), MV(4, -1), |
| 102 | MV(-4, 0), MV(4, 0), MV(-4, 1), MV(4, 1), |
| 103 | MV(-4, 2), MV(4, 2), MV(-2, 3), MV(2, 3), |
| 104 | }; |
| 105 | const MV offsets[] = |
| 106 | { |
| 107 | MV(-1, 0), MV(0, -1), |
| 108 | MV(-1, -1), MV(1, -1), |
| 109 | MV(-1, 0), MV(1, 0), |
| 110 | MV(-1, 1), MV(-1, -1), |
| 111 | MV(1, -1), MV(1, 1), |
| 112 | MV(-1, 0), MV(0, 1), |
| 113 | MV(-1, 1), MV(1, 1), |
| 114 | MV(1, 0), MV(0, 1), |
| 115 | }; // offsets for Two Point Search |
| 116 | |
| 117 | /* sum of absolute differences between MV candidates, used for adaptive ME range */ |
| 118 | inline int predictorDifference(const MV *mvc, intptr_t numCandidates) |
| 119 | { |
| 120 | int sum = 0; |
| 121 | |
| 122 | for (int i = 0; i < numCandidates - 1; i++) |
| 123 | { |
| 124 | sum += abs(mvc[i].x - mvc[i + 1].x) |
| 125 | + abs(mvc[i].y - mvc[i + 1].y); |
| 126 | } |
| 127 | |
| 128 | return sum; |
| 129 | } |
| 130 | |
| 131 | } |
| 132 | |
| 133 | MotionEstimate::MotionEstimate() |
| 134 | { |
| 135 | ctuAddr = -1; |
| 136 | absPartIdx = -1; |
| 137 | searchMethod = X265_HEX_SEARCH; |
| 138 | subpelRefine = 2; |
| 139 | bChromaSATD = false; |
| 140 | chromaSatd = NULL; |
| 141 | } |
| 142 | |
| 143 | void MotionEstimate::init(int method, int refine, int csp) |
| 144 | { |
| 145 | if (!sizeScale[0]) |
| 146 | initScales(); |
| 147 | |
| 148 | searchMethod = method; |
| 149 | subpelRefine = refine; |
| 150 | fencPUYuv.create(FENC_STRIDE, csp); |
| 151 | } |
| 152 | |
| 153 | MotionEstimate::~MotionEstimate() |
| 154 | { |
| 155 | fencPUYuv.destroy(); |
| 156 | } |
| 157 | |
| 158 | /* Called by lookahead, luma only, no use of PicYuv */ |
| 159 | void MotionEstimate::setSourcePU(pixel *fencY, intptr_t stride, intptr_t offset, int pwidth, int pheight) |
| 160 | { |
| 161 | partEnum = partitionFromSizes(pwidth, pheight); |
| 162 | X265_CHECK(LUMA_4x4 != partEnum, "4x4 inter partition detected!\n"); |
| 163 | sad = primitives.sad[partEnum]; |
| 164 | satd = primitives.satd[partEnum]; |
| 165 | sad_x3 = primitives.sad_x3[partEnum]; |
| 166 | sad_x4 = primitives.sad_x4[partEnum]; |
| 167 | |
| 168 | blockwidth = pwidth; |
| 169 | blockOffset = offset; |
| 170 | absPartIdx = ctuAddr = -1; |
| 171 | |
| 172 | /* copy PU block into cache */ |
| 173 | primitives.luma_copy_pp[partEnum](fencPUYuv.m_buf[0], FENC_STRIDE, fencY + offset, stride); |
| 174 | X265_CHECK(!bChromaSATD, "chroma distortion measurements impossible in this code path\n"); |
| 175 | } |
| 176 | |
| 177 | /* Called by Search::predInterSearch() or --pme equivalent, chroma residual might be considered */ |
| 178 | void MotionEstimate::setSourcePU(const Yuv& srcFencYuv, int _ctuAddr, int cuPartIdx, int puPartIdx, int pwidth, int pheight) |
| 179 | { |
| 180 | partEnum = partitionFromSizes(pwidth, pheight); |
| 181 | X265_CHECK(LUMA_4x4 != partEnum, "4x4 inter partition detected!\n"); |
| 182 | sad = primitives.sad[partEnum]; |
| 183 | satd = primitives.satd[partEnum]; |
| 184 | sad_x3 = primitives.sad_x3[partEnum]; |
| 185 | sad_x4 = primitives.sad_x4[partEnum]; |
| 186 | chromaSatd = primitives.chroma[fencPUYuv.m_csp].satd[partEnum]; |
| 187 | |
| 188 | /* Enable chroma residual cost if subpelRefine level is greater than 2 and chroma block size |
| 189 | * is an even multiple of 4x4 pixels (indicated by non-null chromaSatd pointer) */ |
| 190 | bChromaSATD = subpelRefine > 2 && chromaSatd; |
| 191 | X265_CHECK(!(bChromaSATD && !workload[subpelRefine].hpel_satd), "Chroma SATD cannot be used with SAD hpel\n"); |
| 192 | |
| 193 | ctuAddr = _ctuAddr; |
| 194 | absPartIdx = cuPartIdx + puPartIdx; |
| 195 | blockwidth = pwidth; |
| 196 | blockOffset = 0; |
| 197 | |
| 198 | /* copy PU from CU Yuv */ |
| 199 | fencPUYuv.copyPUFromYuv(srcFencYuv, puPartIdx, partEnum, bChromaSATD); |
| 200 | } |
| 201 | |
| 202 | #define COST_MV_PT_DIST(mx, my, point, dist) \ |
| 203 | do \ |
| 204 | { \ |
| 205 | MV tmv(mx, my); \ |
| 206 | int cost = sad(fenc, FENC_STRIDE, fref + mx + my * stride, stride); \ |
| 207 | cost += mvcost(tmv << 2); \ |
| 208 | if (cost < bcost) { \ |
| 209 | bcost = cost; \ |
| 210 | bmv = tmv; \ |
| 211 | bPointNr = point; \ |
| 212 | bDistance = dist; \ |
| 213 | } \ |
| 214 | } while (0) |
| 215 | |
| 216 | #define COST_MV(mx, my) \ |
| 217 | do \ |
| 218 | { \ |
| 219 | int cost = sad(fenc, FENC_STRIDE, fref + (mx) + (my) * stride, stride); \ |
| 220 | cost += mvcost(MV(mx, my) << 2); \ |
| 221 | COPY2_IF_LT(bcost, cost, bmv, MV(mx, my)); \ |
| 222 | } while (0) |
| 223 | |
| 224 | #define COST_MV_X3_DIR(m0x, m0y, m1x, m1y, m2x, m2y, costs) \ |
| 225 | { \ |
| 226 | pixel *pix_base = fref + bmv.x + bmv.y * stride; \ |
| 227 | sad_x3(fenc, \ |
| 228 | pix_base + (m0x) + (m0y) * stride, \ |
| 229 | pix_base + (m1x) + (m1y) * stride, \ |
| 230 | pix_base + (m2x) + (m2y) * stride, \ |
| 231 | stride, costs); \ |
| 232 | (costs)[0] += mvcost((bmv + MV(m0x, m0y)) << 2); \ |
| 233 | (costs)[1] += mvcost((bmv + MV(m1x, m1y)) << 2); \ |
| 234 | (costs)[2] += mvcost((bmv + MV(m2x, m2y)) << 2); \ |
| 235 | } |
| 236 | |
| 237 | #define COST_MV_PT_DIST_X4(m0x, m0y, p0, d0, m1x, m1y, p1, d1, m2x, m2y, p2, d2, m3x, m3y, p3, d3) \ |
| 238 | { \ |
| 239 | sad_x4(fenc, \ |
| 240 | fref + (m0x) + (m0y) * stride, \ |
| 241 | fref + (m1x) + (m1y) * stride, \ |
| 242 | fref + (m2x) + (m2y) * stride, \ |
| 243 | fref + (m3x) + (m3y) * stride, \ |
| 244 | stride, costs); \ |
| 245 | costs[0] += mvcost(MV(m0x, m0y) << 2); \ |
| 246 | costs[1] += mvcost(MV(m1x, m1y) << 2); \ |
| 247 | costs[2] += mvcost(MV(m2x, m2y) << 2); \ |
| 248 | costs[3] += mvcost(MV(m3x, m3y) << 2); \ |
| 249 | COPY4_IF_LT(bcost, costs[0], bmv, MV(m0x, m0y), bPointNr, p0, bDistance, d0); \ |
| 250 | COPY4_IF_LT(bcost, costs[1], bmv, MV(m1x, m1y), bPointNr, p1, bDistance, d1); \ |
| 251 | COPY4_IF_LT(bcost, costs[2], bmv, MV(m2x, m2y), bPointNr, p2, bDistance, d2); \ |
| 252 | COPY4_IF_LT(bcost, costs[3], bmv, MV(m3x, m3y), bPointNr, p3, bDistance, d3); \ |
| 253 | } |
| 254 | |
| 255 | #define COST_MV_X4(m0x, m0y, m1x, m1y, m2x, m2y, m3x, m3y) \ |
| 256 | { \ |
| 257 | pixel *pix_base = fref + omv.x + omv.y * stride; \ |
| 258 | sad_x4(fenc, \ |
| 259 | pix_base + (m0x) + (m0y) * stride, \ |
| 260 | pix_base + (m1x) + (m1y) * stride, \ |
| 261 | pix_base + (m2x) + (m2y) * stride, \ |
| 262 | pix_base + (m3x) + (m3y) * stride, \ |
| 263 | stride, costs); \ |
| 264 | costs[0] += mvcost((omv + MV(m0x, m0y)) << 2); \ |
| 265 | costs[1] += mvcost((omv + MV(m1x, m1y)) << 2); \ |
| 266 | costs[2] += mvcost((omv + MV(m2x, m2y)) << 2); \ |
| 267 | costs[3] += mvcost((omv + MV(m3x, m3y)) << 2); \ |
| 268 | COPY2_IF_LT(bcost, costs[0], bmv, omv + MV(m0x, m0y)); \ |
| 269 | COPY2_IF_LT(bcost, costs[1], bmv, omv + MV(m1x, m1y)); \ |
| 270 | COPY2_IF_LT(bcost, costs[2], bmv, omv + MV(m2x, m2y)); \ |
| 271 | COPY2_IF_LT(bcost, costs[3], bmv, omv + MV(m3x, m3y)); \ |
| 272 | } |
| 273 | |
| 274 | #define COST_MV_X4_DIR(m0x, m0y, m1x, m1y, m2x, m2y, m3x, m3y, costs) \ |
| 275 | { \ |
| 276 | pixel *pix_base = fref + bmv.x + bmv.y * stride; \ |
| 277 | sad_x4(fenc, \ |
| 278 | pix_base + (m0x) + (m0y) * stride, \ |
| 279 | pix_base + (m1x) + (m1y) * stride, \ |
| 280 | pix_base + (m2x) + (m2y) * stride, \ |
| 281 | pix_base + (m3x) + (m3y) * stride, \ |
| 282 | stride, costs); \ |
| 283 | (costs)[0] += mvcost((bmv + MV(m0x, m0y)) << 2); \ |
| 284 | (costs)[1] += mvcost((bmv + MV(m1x, m1y)) << 2); \ |
| 285 | (costs)[2] += mvcost((bmv + MV(m2x, m2y)) << 2); \ |
| 286 | (costs)[3] += mvcost((bmv + MV(m3x, m3y)) << 2); \ |
| 287 | } |
| 288 | |
| 289 | #define DIA1_ITER(mx, my) \ |
| 290 | { \ |
| 291 | omv.x = mx; omv.y = my; \ |
| 292 | COST_MV_X4(0, -1, 0, 1, -1, 0, 1, 0); \ |
| 293 | } |
| 294 | |
| 295 | #define CROSS(start, x_max, y_max) \ |
| 296 | { \ |
| 297 | int16_t i = start; \ |
| 298 | if ((x_max) <= X265_MIN(mvmax.x - omv.x, omv.x - mvmin.x)) \ |
| 299 | for (; i < (x_max) - 2; i += 4) { \ |
| 300 | COST_MV_X4(i, 0, -i, 0, i + 2, 0, -i - 2, 0); } \ |
| 301 | for (; i < (x_max); i += 2) \ |
| 302 | { \ |
| 303 | if (omv.x + i <= mvmax.x) \ |
| 304 | COST_MV(omv.x + i, omv.y); \ |
| 305 | if (omv.x - i >= mvmin.x) \ |
| 306 | COST_MV(omv.x - i, omv.y); \ |
| 307 | } \ |
| 308 | i = start; \ |
| 309 | if ((y_max) <= X265_MIN(mvmax.y - omv.y, omv.y - mvmin.y)) \ |
| 310 | for (; i < (y_max) - 2; i += 4) { \ |
| 311 | COST_MV_X4(0, i, 0, -i, 0, i + 2, 0, -i - 2); } \ |
| 312 | for (; i < (y_max); i += 2) \ |
| 313 | { \ |
| 314 | if (omv.y + i <= mvmax.y) \ |
| 315 | COST_MV(omv.x, omv.y + i); \ |
| 316 | if (omv.y - i >= mvmin.y) \ |
| 317 | COST_MV(omv.x, omv.y - i); \ |
| 318 | } \ |
| 319 | } |
| 320 | |
| 321 | void MotionEstimate::StarPatternSearch(ReferencePlanes *ref, |
| 322 | const MV & mvmin, |
| 323 | const MV & mvmax, |
| 324 | MV & bmv, |
| 325 | int & bcost, |
| 326 | int & bPointNr, |
| 327 | int & bDistance, |
| 328 | int earlyExitIters, |
| 329 | int merange) |
| 330 | { |
| 331 | ALIGN_VAR_16(int, costs[16]); |
| 332 | pixel* fenc = fencPUYuv.m_buf[0]; |
| 333 | pixel* fref = ref->fpelPlane[0] + blockOffset; |
| 334 | intptr_t stride = ref->lumaStride; |
| 335 | |
| 336 | MV omv = bmv; |
| 337 | int saved = bcost; |
| 338 | int rounds = 0; |
| 339 | |
| 340 | { |
| 341 | int16_t dist = 1; |
| 342 | |
| 343 | /* bPointNr |
| 344 | 2 |
| 345 | 4 * 5 |
| 346 | 7 |
| 347 | */ |
| 348 | const int16_t top = omv.y - dist; |
| 349 | const int16_t bottom = omv.y + dist; |
| 350 | const int16_t left = omv.x - dist; |
| 351 | const int16_t right = omv.x + dist; |
| 352 | |
| 353 | if (top >= mvmin.y && left >= mvmin.x && right <= mvmax.x && bottom <= mvmax.y) |
| 354 | { |
| 355 | COST_MV_PT_DIST_X4(omv.x, top, 2, dist, |
| 356 | left, omv.y, 4, dist, |
| 357 | right, omv.y, 5, dist, |
| 358 | omv.x, bottom, 7, dist); |
| 359 | } |
| 360 | else |
| 361 | { |
| 362 | if (top >= mvmin.y) // check top |
| 363 | { |
| 364 | COST_MV_PT_DIST(omv.x, top, 2, dist); |
| 365 | } |
| 366 | if (left >= mvmin.x) // check middle left |
| 367 | { |
| 368 | COST_MV_PT_DIST(left, omv.y, 4, dist); |
| 369 | } |
| 370 | if (right <= mvmax.x) // check middle right |
| 371 | { |
| 372 | COST_MV_PT_DIST(right, omv.y, 5, dist); |
| 373 | } |
| 374 | if (bottom <= mvmax.y) // check bottom |
| 375 | { |
| 376 | COST_MV_PT_DIST(omv.x, bottom, 7, dist); |
| 377 | } |
| 378 | } |
| 379 | if (bcost < saved) |
| 380 | rounds = 0; |
| 381 | else if (++rounds >= earlyExitIters) |
| 382 | return; |
| 383 | } |
| 384 | |
| 385 | for (int16_t dist = 2; dist <= 8; dist <<= 1) |
| 386 | { |
| 387 | /* bPointNr |
| 388 | 2 |
| 389 | 1 3 |
| 390 | 4 * 5 |
| 391 | 6 8 |
| 392 | 7 |
| 393 | Points 2, 4, 5, 7 are dist |
| 394 | Points 1, 3, 6, 8 are dist>>1 |
| 395 | */ |
| 396 | const int16_t top = omv.y - dist; |
| 397 | const int16_t bottom = omv.y + dist; |
| 398 | const int16_t left = omv.x - dist; |
| 399 | const int16_t right = omv.x + dist; |
| 400 | const int16_t top2 = omv.y - (dist >> 1); |
| 401 | const int16_t bottom2 = omv.y + (dist >> 1); |
| 402 | const int16_t left2 = omv.x - (dist >> 1); |
| 403 | const int16_t right2 = omv.x + (dist >> 1); |
| 404 | saved = bcost; |
| 405 | |
| 406 | if (top >= mvmin.y && left >= mvmin.x && |
| 407 | right <= mvmax.x && bottom <= mvmax.y) // check border |
| 408 | { |
| 409 | COST_MV_PT_DIST_X4(omv.x, top, 2, dist, |
| 410 | left2, top2, 1, dist >> 1, |
| 411 | right2, top2, 3, dist >> 1, |
| 412 | left, omv.y, 4, dist); |
| 413 | COST_MV_PT_DIST_X4(right, omv.y, 5, dist, |
| 414 | left2, bottom2, 6, dist >> 1, |
| 415 | right2, bottom2, 8, dist >> 1, |
| 416 | omv.x, bottom, 7, dist); |
| 417 | } |
| 418 | else // check border for each mv |
| 419 | { |
| 420 | if (top >= mvmin.y) // check top |
| 421 | { |
| 422 | COST_MV_PT_DIST(omv.x, top, 2, dist); |
| 423 | } |
| 424 | if (top2 >= mvmin.y) // check half top |
| 425 | { |
| 426 | if (left2 >= mvmin.x) // check half left |
| 427 | { |
| 428 | COST_MV_PT_DIST(left2, top2, 1, (dist >> 1)); |
| 429 | } |
| 430 | if (right2 <= mvmax.x) // check half right |
| 431 | { |
| 432 | COST_MV_PT_DIST(right2, top2, 3, (dist >> 1)); |
| 433 | } |
| 434 | } |
| 435 | if (left >= mvmin.x) // check left |
| 436 | { |
| 437 | COST_MV_PT_DIST(left, omv.y, 4, dist); |
| 438 | } |
| 439 | if (right <= mvmax.x) // check right |
| 440 | { |
| 441 | COST_MV_PT_DIST(right, omv.y, 5, dist); |
| 442 | } |
| 443 | if (bottom2 <= mvmax.y) // check half bottom |
| 444 | { |
| 445 | if (left2 >= mvmin.x) // check half left |
| 446 | { |
| 447 | COST_MV_PT_DIST(left2, bottom2, 6, (dist >> 1)); |
| 448 | } |
| 449 | if (right2 <= mvmax.x) // check half right |
| 450 | { |
| 451 | COST_MV_PT_DIST(right2, bottom2, 8, (dist >> 1)); |
| 452 | } |
| 453 | } |
| 454 | if (bottom <= mvmax.y) // check bottom |
| 455 | { |
| 456 | COST_MV_PT_DIST(omv.x, bottom, 7, dist); |
| 457 | } |
| 458 | } |
| 459 | |
| 460 | if (bcost < saved) |
| 461 | rounds = 0; |
| 462 | else if (++rounds >= earlyExitIters) |
| 463 | return; |
| 464 | } |
| 465 | |
| 466 | for (int16_t dist = 16; dist <= (int16_t)merange; dist <<= 1) |
| 467 | { |
| 468 | const int16_t top = omv.y - dist; |
| 469 | const int16_t bottom = omv.y + dist; |
| 470 | const int16_t left = omv.x - dist; |
| 471 | const int16_t right = omv.x + dist; |
| 472 | |
| 473 | saved = bcost; |
| 474 | if (top >= mvmin.y && left >= mvmin.x && |
| 475 | right <= mvmax.x && bottom <= mvmax.y) // check border |
| 476 | { |
| 477 | /* index |
| 478 | 0 |
| 479 | 3 |
| 480 | 2 |
| 481 | 1 |
| 482 | 0 3 2 1 * 1 2 3 0 |
| 483 | 1 |
| 484 | 2 |
| 485 | 3 |
| 486 | 0 |
| 487 | */ |
| 488 | |
| 489 | COST_MV_PT_DIST_X4(omv.x, top, 0, dist, |
| 490 | left, omv.y, 0, dist, |
| 491 | right, omv.y, 0, dist, |
| 492 | omv.x, bottom, 0, dist); |
| 493 | |
| 494 | for (int16_t index = 1; index < 4; index++) |
| 495 | { |
| 496 | int16_t posYT = top + ((dist >> 2) * index); |
| 497 | int16_t posYB = bottom - ((dist >> 2) * index); |
| 498 | int16_t posXL = omv.x - ((dist >> 2) * index); |
| 499 | int16_t posXR = omv.x + ((dist >> 2) * index); |
| 500 | |
| 501 | COST_MV_PT_DIST_X4(posXL, posYT, 0, dist, |
| 502 | posXR, posYT, 0, dist, |
| 503 | posXL, posYB, 0, dist, |
| 504 | posXR, posYB, 0, dist); |
| 505 | } |
| 506 | } |
| 507 | else // check border for each mv |
| 508 | { |
| 509 | if (top >= mvmin.y) // check top |
| 510 | { |
| 511 | COST_MV_PT_DIST(omv.x, top, 0, dist); |
| 512 | } |
| 513 | if (left >= mvmin.x) // check left |
| 514 | { |
| 515 | COST_MV_PT_DIST(left, omv.y, 0, dist); |
| 516 | } |
| 517 | if (right <= mvmax.x) // check right |
| 518 | { |
| 519 | COST_MV_PT_DIST(right, omv.y, 0, dist); |
| 520 | } |
| 521 | if (bottom <= mvmax.y) // check bottom |
| 522 | { |
| 523 | COST_MV_PT_DIST(omv.x, bottom, 0, dist); |
| 524 | } |
| 525 | for (int16_t index = 1; index < 4; index++) |
| 526 | { |
| 527 | int16_t posYT = top + ((dist >> 2) * index); |
| 528 | int16_t posYB = bottom - ((dist >> 2) * index); |
| 529 | int16_t posXL = omv.x - ((dist >> 2) * index); |
| 530 | int16_t posXR = omv.x + ((dist >> 2) * index); |
| 531 | |
| 532 | if (posYT >= mvmin.y) // check top |
| 533 | { |
| 534 | if (posXL >= mvmin.x) // check left |
| 535 | { |
| 536 | COST_MV_PT_DIST(posXL, posYT, 0, dist); |
| 537 | } |
| 538 | if (posXR <= mvmax.x) // check right |
| 539 | { |
| 540 | COST_MV_PT_DIST(posXR, posYT, 0, dist); |
| 541 | } |
| 542 | } |
| 543 | if (posYB <= mvmax.y) // check bottom |
| 544 | { |
| 545 | if (posXL >= mvmin.x) // check left |
| 546 | { |
| 547 | COST_MV_PT_DIST(posXL, posYB, 0, dist); |
| 548 | } |
| 549 | if (posXR <= mvmax.x) // check right |
| 550 | { |
| 551 | COST_MV_PT_DIST(posXR, posYB, 0, dist); |
| 552 | } |
| 553 | } |
| 554 | } |
| 555 | } |
| 556 | |
| 557 | if (bcost < saved) |
| 558 | rounds = 0; |
| 559 | else if (++rounds >= earlyExitIters) |
| 560 | return; |
| 561 | } |
| 562 | } |
| 563 | |
| 564 | int MotionEstimate::motionEstimate(ReferencePlanes *ref, |
| 565 | const MV & mvmin, |
| 566 | const MV & mvmax, |
| 567 | const MV & qmvp, |
| 568 | int numCandidates, |
| 569 | const MV * mvc, |
| 570 | int merange, |
| 571 | MV & outQMv) |
| 572 | { |
| 573 | ALIGN_VAR_16(int, costs[16]); |
| 574 | if (ctuAddr >= 0) |
| 575 | blockOffset = ref->reconPic->getLumaAddr(ctuAddr, absPartIdx) - ref->reconPic->getLumaAddr(0); |
| 576 | intptr_t stride = ref->lumaStride; |
| 577 | pixel* fenc = fencPUYuv.m_buf[0]; |
| 578 | pixel* fref = ref->fpelPlane[0] + blockOffset; |
| 579 | |
| 580 | setMVP(qmvp); |
| 581 | |
| 582 | MV qmvmin = mvmin.toQPel(); |
| 583 | MV qmvmax = mvmax.toQPel(); |
| 584 | |
| 585 | /* The term cost used here means satd/sad values for that particular search. |
| 586 | * The costs used in ME integer search only includes the SAD cost of motion |
| 587 | * residual and sqrtLambda times MVD bits. The subpel refine steps use SATD |
| 588 | * cost of residual and sqrtLambda * MVD bits. Mode decision will be based |
| 589 | * on video distortion cost (SSE/PSNR) plus lambda times all signaling bits |
| 590 | * (mode + MVD bits). */ |
| 591 | |
| 592 | // measure SAD cost at clipped QPEL MVP |
| 593 | MV pmv = qmvp.clipped(qmvmin, qmvmax); |
| 594 | MV bestpre = pmv; |
| 595 | int bprecost; |
| 596 | |
| 597 | if (ref->isLowres) |
| 598 | bprecost = ref->lowresQPelCost(fenc, blockOffset, pmv, sad); |
| 599 | else |
| 600 | bprecost = subpelCompare(ref, pmv, sad); |
| 601 | |
| 602 | /* re-measure full pel rounded MVP with SAD as search start point */ |
| 603 | MV bmv = pmv.roundToFPel(); |
| 604 | int bcost = bprecost; |
| 605 | if (pmv.isSubpel()) |
| 606 | bcost = sad(fenc, FENC_STRIDE, fref + bmv.x + bmv.y * stride, stride) + mvcost(bmv << 2); |
| 607 | |
| 608 | // measure SAD cost at MV(0) if MVP is not zero |
| 609 | if (pmv.notZero()) |
| 610 | { |
| 611 | int cost = sad(fenc, FENC_STRIDE, fref, stride) + mvcost(MV(0, 0)); |
| 612 | if (cost < bcost) |
| 613 | { |
| 614 | bcost = cost; |
| 615 | bmv = 0; |
| 616 | } |
| 617 | } |
| 618 | |
| 619 | // measure SAD cost at each QPEL motion vector candidate |
| 620 | if (ref->isLowres) |
| 621 | { |
| 622 | for (int i = 0; i < numCandidates; i++) |
| 623 | { |
| 624 | MV m = mvc[i].clipped(qmvmin, qmvmax); |
| 625 | if (m.notZero() && m != pmv && m != bestpre) // check already measured |
| 626 | { |
| 627 | int cost = ref->lowresQPelCost(fenc, blockOffset, m, sad) + mvcost(m); |
| 628 | if (cost < bprecost) |
| 629 | { |
| 630 | bprecost = cost; |
| 631 | bestpre = m; |
| 632 | } |
| 633 | } |
| 634 | } |
| 635 | } |
| 636 | else |
| 637 | { |
| 638 | for (int i = 0; i < numCandidates; i++) |
| 639 | { |
| 640 | MV m = mvc[i].clipped(qmvmin, qmvmax); |
| 641 | if (m.notZero() && m != pmv && m != bestpre) // check already measured |
| 642 | { |
| 643 | int cost = subpelCompare(ref, m, sad) + mvcost(m); |
| 644 | if (cost < bprecost) |
| 645 | { |
| 646 | bprecost = cost; |
| 647 | bestpre = m; |
| 648 | } |
| 649 | } |
| 650 | } |
| 651 | } |
| 652 | |
| 653 | pmv = pmv.roundToFPel(); |
| 654 | MV omv = bmv; // current search origin or starting point |
| 655 | |
| 656 | switch (searchMethod) |
| 657 | { |
| 658 | case X265_DIA_SEARCH: |
| 659 | { |
| 660 | /* diamond search, radius 1 */ |
| 661 | bcost <<= 4; |
| 662 | int i = merange; |
| 663 | do |
| 664 | { |
| 665 | COST_MV_X4_DIR(0, -1, 0, 1, -1, 0, 1, 0, costs); |
| 666 | COPY1_IF_LT(bcost, (costs[0] << 4) + 1); |
| 667 | COPY1_IF_LT(bcost, (costs[1] << 4) + 3); |
| 668 | COPY1_IF_LT(bcost, (costs[2] << 4) + 4); |
| 669 | COPY1_IF_LT(bcost, (costs[3] << 4) + 12); |
| 670 | if (!(bcost & 15)) |
| 671 | break; |
| 672 | bmv.x -= (bcost << 28) >> 30; |
| 673 | bmv.y -= (bcost << 30) >> 30; |
| 674 | bcost &= ~15; |
| 675 | } |
| 676 | while (--i && bmv.checkRange(mvmin, mvmax)); |
| 677 | bcost >>= 4; |
| 678 | break; |
| 679 | } |
| 680 | |
| 681 | case X265_HEX_SEARCH: |
| 682 | { |
| 683 | me_hex2: |
| 684 | /* hexagon search, radius 2 */ |
| 685 | #if 0 |
| 686 | for (int i = 0; i < merange / 2; i++) |
| 687 | { |
| 688 | omv = bmv; |
| 689 | COST_MV(omv.x - 2, omv.y); |
| 690 | COST_MV(omv.x - 1, omv.y + 2); |
| 691 | COST_MV(omv.x + 1, omv.y + 2); |
| 692 | COST_MV(omv.x + 2, omv.y); |
| 693 | COST_MV(omv.x + 1, omv.y - 2); |
| 694 | COST_MV(omv.x - 1, omv.y - 2); |
| 695 | if (omv == bmv) |
| 696 | break; |
| 697 | if (!bmv.checkRange(mvmin, mvmax)) |
| 698 | break; |
| 699 | } |
| 700 | |
| 701 | #else // if 0 |
| 702 | /* equivalent to the above, but eliminates duplicate candidates */ |
| 703 | COST_MV_X3_DIR(-2, 0, -1, 2, 1, 2, costs); |
| 704 | bcost <<= 3; |
| 705 | COPY1_IF_LT(bcost, (costs[0] << 3) + 2); |
| 706 | COPY1_IF_LT(bcost, (costs[1] << 3) + 3); |
| 707 | COPY1_IF_LT(bcost, (costs[2] << 3) + 4); |
| 708 | COST_MV_X3_DIR(2, 0, 1, -2, -1, -2, costs); |
| 709 | COPY1_IF_LT(bcost, (costs[0] << 3) + 5); |
| 710 | COPY1_IF_LT(bcost, (costs[1] << 3) + 6); |
| 711 | COPY1_IF_LT(bcost, (costs[2] << 3) + 7); |
| 712 | |
| 713 | if (bcost & 7) |
| 714 | { |
| 715 | int dir = (bcost & 7) - 2; |
| 716 | bmv += hex2[dir + 1]; |
| 717 | |
| 718 | /* half hexagon, not overlapping the previous iteration */ |
| 719 | for (int i = (merange >> 1) - 1; i > 0 && bmv.checkRange(mvmin, mvmax); i--) |
| 720 | { |
| 721 | COST_MV_X3_DIR(hex2[dir + 0].x, hex2[dir + 0].y, |
| 722 | hex2[dir + 1].x, hex2[dir + 1].y, |
| 723 | hex2[dir + 2].x, hex2[dir + 2].y, |
| 724 | costs); |
| 725 | bcost &= ~7; |
| 726 | COPY1_IF_LT(bcost, (costs[0] << 3) + 1); |
| 727 | COPY1_IF_LT(bcost, (costs[1] << 3) + 2); |
| 728 | COPY1_IF_LT(bcost, (costs[2] << 3) + 3); |
| 729 | if (!(bcost & 7)) |
| 730 | break; |
| 731 | dir += (bcost & 7) - 2; |
| 732 | dir = mod6m1[dir + 1]; |
| 733 | bmv += hex2[dir + 1]; |
| 734 | } |
| 735 | } |
| 736 | bcost >>= 3; |
| 737 | #endif // if 0 |
| 738 | |
| 739 | /* square refine */ |
| 740 | int dir = 0; |
| 741 | COST_MV_X4_DIR(0, -1, 0, 1, -1, 0, 1, 0, costs); |
| 742 | COPY2_IF_LT(bcost, costs[0], dir, 1); |
| 743 | COPY2_IF_LT(bcost, costs[1], dir, 2); |
| 744 | COPY2_IF_LT(bcost, costs[2], dir, 3); |
| 745 | COPY2_IF_LT(bcost, costs[3], dir, 4); |
| 746 | COST_MV_X4_DIR(-1, -1, -1, 1, 1, -1, 1, 1, costs); |
| 747 | COPY2_IF_LT(bcost, costs[0], dir, 5); |
| 748 | COPY2_IF_LT(bcost, costs[1], dir, 6); |
| 749 | COPY2_IF_LT(bcost, costs[2], dir, 7); |
| 750 | COPY2_IF_LT(bcost, costs[3], dir, 8); |
| 751 | bmv += square1[dir]; |
| 752 | break; |
| 753 | } |
| 754 | |
| 755 | case X265_UMH_SEARCH: |
| 756 | { |
| 757 | int ucost1, ucost2; |
| 758 | int16_t cross_start = 1; |
| 759 | |
| 760 | /* refine predictors */ |
| 761 | omv = bmv; |
| 762 | ucost1 = bcost; |
| 763 | DIA1_ITER(pmv.x, pmv.y); |
| 764 | if (pmv.notZero()) |
| 765 | DIA1_ITER(0, 0); |
| 766 | |
| 767 | ucost2 = bcost; |
| 768 | if (bmv.notZero() && bmv != pmv) |
| 769 | DIA1_ITER(bmv.x, bmv.y); |
| 770 | if (bcost == ucost2) |
| 771 | cross_start = 3; |
| 772 | |
| 773 | /* Early Termination */ |
| 774 | omv = bmv; |
| 775 | if (bcost == ucost2 && SAD_THRESH(2000)) |
| 776 | { |
| 777 | COST_MV_X4(0, -2, -1, -1, 1, -1, -2, 0); |
| 778 | COST_MV_X4(2, 0, -1, 1, 1, 1, 0, 2); |
| 779 | if (bcost == ucost1 && SAD_THRESH(500)) |
| 780 | break; |
| 781 | if (bcost == ucost2) |
| 782 | { |
| 783 | int16_t range = (int16_t)(merange >> 1) | 1; |
| 784 | CROSS(3, range, range); |
| 785 | COST_MV_X4(-1, -2, 1, -2, -2, -1, 2, -1); |
| 786 | COST_MV_X4(-2, 1, 2, 1, -1, 2, 1, 2); |
| 787 | if (bcost == ucost2) |
| 788 | break; |
| 789 | cross_start = range + 2; |
| 790 | } |
| 791 | } |
| 792 | |
| 793 | // TODO: Need to study x264's logic for building mvc list to understand why they |
| 794 | // have special cases here for 16x16, and whether they apply to HEVC CTU |
| 795 | |
| 796 | // adaptive search range based on mvc variability |
| 797 | if (numCandidates) |
| 798 | { |
| 799 | /* range multipliers based on casual inspection of some statistics of |
| 800 | * average distance between current predictor and final mv found by ESA. |
| 801 | * these have not been tuned much by actual encoding. */ |
| 802 | static const uint8_t range_mul[4][4] = |
| 803 | { |
| 804 | { 3, 3, 4, 4 }, |
| 805 | { 3, 4, 4, 4 }, |
| 806 | { 4, 4, 4, 5 }, |
| 807 | { 4, 4, 5, 6 }, |
| 808 | }; |
| 809 | |
| 810 | int mvd; |
| 811 | int sad_ctx, mvd_ctx; |
| 812 | int denom = 1; |
| 813 | |
| 814 | if (numCandidates == 1) |
| 815 | { |
| 816 | if (LUMA_64x64 == partEnum) |
| 817 | /* mvc is probably the same as mvp, so the difference isn't meaningful. |
| 818 | * but prediction usually isn't too bad, so just use medium range */ |
| 819 | mvd = 25; |
| 820 | else |
| 821 | mvd = abs(qmvp.x - mvc[0].x) + abs(qmvp.y - mvc[0].y); |
| 822 | } |
| 823 | else |
| 824 | { |
| 825 | /* calculate the degree of agreement between predictors. */ |
| 826 | |
| 827 | /* in 64x64, mvc includes all the neighbors used to make mvp, |
| 828 | * so don't count mvp separately. */ |
| 829 | |
| 830 | denom = numCandidates - 1; |
| 831 | mvd = 0; |
| 832 | if (partEnum != LUMA_64x64) |
| 833 | { |
| 834 | mvd = abs(qmvp.x - mvc[0].x) + abs(qmvp.y - mvc[0].y); |
| 835 | denom++; |
| 836 | } |
| 837 | mvd += predictorDifference(mvc, numCandidates); |
| 838 | } |
| 839 | |
| 840 | sad_ctx = SAD_THRESH(1000) ? 0 |
| 841 | : SAD_THRESH(2000) ? 1 |
| 842 | : SAD_THRESH(4000) ? 2 : 3; |
| 843 | mvd_ctx = mvd < 10 * denom ? 0 |
| 844 | : mvd < 20 * denom ? 1 |
| 845 | : mvd < 40 * denom ? 2 : 3; |
| 846 | |
| 847 | merange = (merange * range_mul[mvd_ctx][sad_ctx]) >> 2; |
| 848 | } |
| 849 | |
| 850 | /* FIXME if the above DIA2/OCT2/CROSS found a new mv, it has not updated omx/omy. |
| 851 | * we are still centered on the same place as the DIA2. is this desirable? */ |
| 852 | CROSS(cross_start, merange, merange >> 1); |
| 853 | COST_MV_X4(-2, -2, -2, 2, 2, -2, 2, 2); |
| 854 | |
| 855 | /* hexagon grid */ |
| 856 | omv = bmv; |
| 857 | const uint16_t *p_cost_omvx = m_cost_mvx + omv.x * 4; |
| 858 | const uint16_t *p_cost_omvy = m_cost_mvy + omv.y * 4; |
| 859 | uint16_t i = 1; |
| 860 | do |
| 861 | { |
| 862 | if (4 * i > X265_MIN4(mvmax.x - omv.x, omv.x - mvmin.x, |
| 863 | mvmax.y - omv.y, omv.y - mvmin.y)) |
| 864 | { |
| 865 | for (int j = 0; j < 16; j++) |
| 866 | { |
| 867 | MV mv = omv + (hex4[j] * i); |
| 868 | if (mv.checkRange(mvmin, mvmax)) |
| 869 | COST_MV(mv.x, mv.y); |
| 870 | } |
| 871 | } |
| 872 | else |
| 873 | { |
| 874 | int16_t dir = 0; |
| 875 | pixel *fref_base = fref + omv.x + (omv.y - 4 * i) * stride; |
| 876 | size_t dy = (size_t)i * stride; |
| 877 | #define SADS(k, x0, y0, x1, y1, x2, y2, x3, y3) \ |
| 878 | sad_x4(fenc, \ |
| 879 | fref_base x0 * i + (y0 - 2 * k + 4) * dy, \ |
| 880 | fref_base x1 * i + (y1 - 2 * k + 4) * dy, \ |
| 881 | fref_base x2 * i + (y2 - 2 * k + 4) * dy, \ |
| 882 | fref_base x3 * i + (y3 - 2 * k + 4) * dy, \ |
| 883 | stride, costs + 4 * k); \ |
| 884 | fref_base += 2 * dy; |
| 885 | #define ADD_MVCOST(k, x, y) costs[k] += p_cost_omvx[x * 4 * i] + p_cost_omvy[y * 4 * i] |
| 886 | #define MIN_MV(k, x, y) COPY2_IF_LT(bcost, costs[k], dir, x * 16 + (y & 15)) |
| 887 | |
| 888 | SADS(0, +0, -4, +0, +4, -2, -3, +2, -3); |
| 889 | SADS(1, -4, -2, +4, -2, -4, -1, +4, -1); |
| 890 | SADS(2, -4, +0, +4, +0, -4, +1, +4, +1); |
| 891 | SADS(3, -4, +2, +4, +2, -2, +3, +2, +3); |
| 892 | ADD_MVCOST(0, 0, -4); |
| 893 | ADD_MVCOST(1, 0, 4); |
| 894 | ADD_MVCOST(2, -2, -3); |
| 895 | ADD_MVCOST(3, 2, -3); |
| 896 | ADD_MVCOST(4, -4, -2); |
| 897 | ADD_MVCOST(5, 4, -2); |
| 898 | ADD_MVCOST(6, -4, -1); |
| 899 | ADD_MVCOST(7, 4, -1); |
| 900 | ADD_MVCOST(8, -4, 0); |
| 901 | ADD_MVCOST(9, 4, 0); |
| 902 | ADD_MVCOST(10, -4, 1); |
| 903 | ADD_MVCOST(11, 4, 1); |
| 904 | ADD_MVCOST(12, -4, 2); |
| 905 | ADD_MVCOST(13, 4, 2); |
| 906 | ADD_MVCOST(14, -2, 3); |
| 907 | ADD_MVCOST(15, 2, 3); |
| 908 | MIN_MV(0, 0, -4); |
| 909 | MIN_MV(1, 0, 4); |
| 910 | MIN_MV(2, -2, -3); |
| 911 | MIN_MV(3, 2, -3); |
| 912 | MIN_MV(4, -4, -2); |
| 913 | MIN_MV(5, 4, -2); |
| 914 | MIN_MV(6, -4, -1); |
| 915 | MIN_MV(7, 4, -1); |
| 916 | MIN_MV(8, -4, 0); |
| 917 | MIN_MV(9, 4, 0); |
| 918 | MIN_MV(10, -4, 1); |
| 919 | MIN_MV(11, 4, 1); |
| 920 | MIN_MV(12, -4, 2); |
| 921 | MIN_MV(13, 4, 2); |
| 922 | MIN_MV(14, -2, 3); |
| 923 | MIN_MV(15, 2, 3); |
| 924 | #undef SADS |
| 925 | #undef ADD_MVCOST |
| 926 | #undef MIN_MV |
| 927 | if (dir) |
| 928 | { |
| 929 | bmv.x = omv.x + i * (dir >> 4); |
| 930 | bmv.y = omv.y + i * ((dir << 28) >> 28); |
| 931 | } |
| 932 | } |
| 933 | } |
| 934 | while (++i <= merange >> 2); |
| 935 | if (bmv.checkRange(mvmin, mvmax)) |
| 936 | goto me_hex2; |
| 937 | break; |
| 938 | } |
| 939 | |
| 940 | case X265_STAR_SEARCH: // Adapted from HM ME |
| 941 | { |
| 942 | int bPointNr = 0; |
| 943 | int bDistance = 0; |
| 944 | |
| 945 | const int EarlyExitIters = 3; |
| 946 | StarPatternSearch(ref, mvmin, mvmax, bmv, bcost, bPointNr, bDistance, EarlyExitIters, merange); |
| 947 | if (bDistance == 1) |
| 948 | { |
| 949 | // if best distance was only 1, check two missing points. If no new point is found, stop |
| 950 | if (bPointNr) |
| 951 | { |
| 952 | /* For a given direction 1 to 8, check nearest two outer X pixels |
| 953 | X X |
| 954 | X 1 2 3 X |
| 955 | 4 * 5 |
| 956 | X 6 7 8 X |
| 957 | X X |
| 958 | */ |
| 959 | int saved = bcost; |
| 960 | const MV mv1 = bmv + offsets[(bPointNr - 1) * 2]; |
| 961 | const MV mv2 = bmv + offsets[(bPointNr - 1) * 2 + 1]; |
| 962 | if (mv1.checkRange(mvmin, mvmax)) |
| 963 | { |
| 964 | COST_MV(mv1.x, mv1.y); |
| 965 | } |
| 966 | if (mv2.checkRange(mvmin, mvmax)) |
| 967 | { |
| 968 | COST_MV(mv2.x, mv2.y); |
| 969 | } |
| 970 | if (bcost == saved) |
| 971 | break; |
| 972 | } |
| 973 | else |
| 974 | break; |
| 975 | } |
| 976 | |
| 977 | const int RasterDistance = 5; |
| 978 | if (bDistance > RasterDistance) |
| 979 | { |
| 980 | // raster search refinement if original search distance was too big |
| 981 | MV tmv; |
| 982 | for (tmv.y = mvmin.y; tmv.y <= mvmax.y; tmv.y += RasterDistance) |
| 983 | { |
| 984 | for (tmv.x = mvmin.x; tmv.x <= mvmax.x; tmv.x += RasterDistance) |
| 985 | { |
| 986 | if (tmv.x + (RasterDistance * 3) <= mvmax.x) |
| 987 | { |
| 988 | pixel *pix_base = fref + tmv.y * stride + tmv.x; |
| 989 | sad_x4(fenc, |
| 990 | pix_base, |
| 991 | pix_base + RasterDistance, |
| 992 | pix_base + RasterDistance * 2, |
| 993 | pix_base + RasterDistance * 3, |
| 994 | stride, costs); |
| 995 | costs[0] += mvcost(tmv << 2); |
| 996 | COPY2_IF_LT(bcost, costs[0], bmv, tmv); |
| 997 | tmv.x += RasterDistance; |
| 998 | costs[1] += mvcost(tmv << 2); |
| 999 | COPY2_IF_LT(bcost, costs[1], bmv, tmv); |
| 1000 | tmv.x += RasterDistance; |
| 1001 | costs[2] += mvcost(tmv << 2); |
| 1002 | COPY2_IF_LT(bcost, costs[2], bmv, tmv); |
| 1003 | tmv.x += RasterDistance; |
| 1004 | costs[3] += mvcost(tmv << 3); |
| 1005 | COPY2_IF_LT(bcost, costs[3], bmv, tmv); |
| 1006 | } |
| 1007 | else |
| 1008 | COST_MV(tmv.x, tmv.y); |
| 1009 | } |
| 1010 | } |
| 1011 | } |
| 1012 | |
| 1013 | while (bDistance > 0) |
| 1014 | { |
| 1015 | // center a new search around current best |
| 1016 | bDistance = 0; |
| 1017 | bPointNr = 0; |
| 1018 | const int MaxIters = 32; |
| 1019 | StarPatternSearch(ref, mvmin, mvmax, bmv, bcost, bPointNr, bDistance, MaxIters, merange); |
| 1020 | |
| 1021 | if (bDistance == 1) |
| 1022 | { |
| 1023 | if (!bPointNr) |
| 1024 | break; |
| 1025 | |
| 1026 | /* For a given direction 1 to 8, check nearest 2 outer X pixels |
| 1027 | X X |
| 1028 | X 1 2 3 X |
| 1029 | 4 * 5 |
| 1030 | X 6 7 8 X |
| 1031 | X X |
| 1032 | */ |
| 1033 | const MV mv1 = bmv + offsets[(bPointNr - 1) * 2]; |
| 1034 | const MV mv2 = bmv + offsets[(bPointNr - 1) * 2 + 1]; |
| 1035 | if (mv1.checkRange(mvmin, mvmax)) |
| 1036 | { |
| 1037 | COST_MV(mv1.x, mv1.y); |
| 1038 | } |
| 1039 | if (mv2.checkRange(mvmin, mvmax)) |
| 1040 | { |
| 1041 | COST_MV(mv2.x, mv2.y); |
| 1042 | } |
| 1043 | break; |
| 1044 | } |
| 1045 | } |
| 1046 | |
| 1047 | break; |
| 1048 | } |
| 1049 | |
| 1050 | case X265_FULL_SEARCH: |
| 1051 | { |
| 1052 | // dead slow exhaustive search, but at least it uses sad_x4() |
| 1053 | MV tmv; |
| 1054 | for (tmv.y = mvmin.y; tmv.y <= mvmax.y; tmv.y++) |
| 1055 | { |
| 1056 | for (tmv.x = mvmin.x; tmv.x <= mvmax.x; tmv.x++) |
| 1057 | { |
| 1058 | if (tmv.x + 3 <= mvmax.x) |
| 1059 | { |
| 1060 | pixel *pix_base = fref + tmv.y * stride + tmv.x; |
| 1061 | sad_x4(fenc, |
| 1062 | pix_base, |
| 1063 | pix_base + 1, |
| 1064 | pix_base + 2, |
| 1065 | pix_base + 3, |
| 1066 | stride, costs); |
| 1067 | costs[0] += mvcost(tmv << 2); |
| 1068 | COPY2_IF_LT(bcost, costs[0], bmv, tmv); |
| 1069 | tmv.x++; |
| 1070 | costs[1] += mvcost(tmv << 2); |
| 1071 | COPY2_IF_LT(bcost, costs[1], bmv, tmv); |
| 1072 | tmv.x++; |
| 1073 | costs[2] += mvcost(tmv << 2); |
| 1074 | COPY2_IF_LT(bcost, costs[2], bmv, tmv); |
| 1075 | tmv.x++; |
| 1076 | costs[3] += mvcost(tmv << 2); |
| 1077 | COPY2_IF_LT(bcost, costs[3], bmv, tmv); |
| 1078 | } |
| 1079 | else |
| 1080 | COST_MV(tmv.x, tmv.y); |
| 1081 | } |
| 1082 | } |
| 1083 | |
| 1084 | break; |
| 1085 | } |
| 1086 | |
| 1087 | default: |
| 1088 | X265_CHECK(0, "invalid motion estimate mode\n"); |
| 1089 | break; |
| 1090 | } |
| 1091 | |
| 1092 | if (bprecost < bcost) |
| 1093 | { |
| 1094 | bmv = bestpre; |
| 1095 | bcost = bprecost; |
| 1096 | } |
| 1097 | else |
| 1098 | bmv = bmv.toQPel(); // promote search bmv to qpel |
| 1099 | |
| 1100 | const SubpelWorkload& wl = workload[this->subpelRefine]; |
| 1101 | |
| 1102 | if (!bcost) |
| 1103 | { |
| 1104 | /* if there was zero residual at the clipped MVP, we can skip subpel |
| 1105 | * refine, but we do need to include the mvcost in the returned cost */ |
| 1106 | bcost = mvcost(bmv); |
| 1107 | } |
| 1108 | else if (ref->isLowres) |
| 1109 | { |
| 1110 | int bdir = 0; |
| 1111 | for (int i = 1; i <= wl.hpel_dirs; i++) |
| 1112 | { |
| 1113 | MV qmv = bmv + square1[i] * 2; |
| 1114 | int cost = ref->lowresQPelCost(fenc, blockOffset, qmv, sad) + mvcost(qmv); |
| 1115 | COPY2_IF_LT(bcost, cost, bdir, i); |
| 1116 | } |
| 1117 | |
| 1118 | bmv += square1[bdir] * 2; |
| 1119 | bcost = ref->lowresQPelCost(fenc, blockOffset, bmv, satd) + mvcost(bmv); |
| 1120 | |
| 1121 | bdir = 0; |
| 1122 | for (int i = 1; i <= wl.qpel_dirs; i++) |
| 1123 | { |
| 1124 | MV qmv = bmv + square1[i]; |
| 1125 | int cost = ref->lowresQPelCost(fenc, blockOffset, qmv, satd) + mvcost(qmv); |
| 1126 | COPY2_IF_LT(bcost, cost, bdir, i); |
| 1127 | } |
| 1128 | |
| 1129 | bmv += square1[bdir]; |
| 1130 | } |
| 1131 | else |
| 1132 | { |
| 1133 | pixelcmp_t hpelcomp; |
| 1134 | |
| 1135 | if (wl.hpel_satd) |
| 1136 | { |
| 1137 | bcost = subpelCompare(ref, bmv, satd) + mvcost(bmv); |
| 1138 | hpelcomp = satd; |
| 1139 | } |
| 1140 | else |
| 1141 | hpelcomp = sad; |
| 1142 | |
| 1143 | for (int iter = 0; iter < wl.hpel_iters; iter++) |
| 1144 | { |
| 1145 | int bdir = 0; |
| 1146 | for (int i = 1; i <= wl.hpel_dirs; i++) |
| 1147 | { |
| 1148 | MV qmv = bmv + square1[i] * 2; |
| 1149 | int cost = subpelCompare(ref, qmv, hpelcomp) + mvcost(qmv); |
| 1150 | COPY2_IF_LT(bcost, cost, bdir, i); |
| 1151 | } |
| 1152 | |
| 1153 | if (bdir) |
| 1154 | bmv += square1[bdir] * 2; |
| 1155 | else |
| 1156 | break; |
| 1157 | } |
| 1158 | |
| 1159 | /* if HPEL search used SAD, remeasure with SATD before QPEL */ |
| 1160 | if (!wl.hpel_satd) |
| 1161 | bcost = subpelCompare(ref, bmv, satd) + mvcost(bmv); |
| 1162 | |
| 1163 | for (int iter = 0; iter < wl.qpel_iters; iter++) |
| 1164 | { |
| 1165 | int bdir = 0; |
| 1166 | for (int i = 1; i <= wl.qpel_dirs; i++) |
| 1167 | { |
| 1168 | MV qmv = bmv + square1[i]; |
| 1169 | int cost = subpelCompare(ref, qmv, satd) + mvcost(qmv); |
| 1170 | COPY2_IF_LT(bcost, cost, bdir, i); |
| 1171 | } |
| 1172 | |
| 1173 | if (bdir) |
| 1174 | bmv += square1[bdir]; |
| 1175 | else |
| 1176 | break; |
| 1177 | } |
| 1178 | } |
| 1179 | |
| 1180 | x265_emms(); |
| 1181 | outQMv = bmv; |
| 1182 | return bcost; |
| 1183 | } |
| 1184 | |
| 1185 | int MotionEstimate::subpelCompare(ReferencePlanes *ref, const MV& qmv, pixelcmp_t cmp) |
| 1186 | { |
| 1187 | intptr_t refStride = ref->lumaStride; |
| 1188 | pixel *fref = ref->fpelPlane[0] + blockOffset + (qmv.x >> 2) + (qmv.y >> 2) * refStride; |
| 1189 | int xFrac = qmv.x & 0x3; |
| 1190 | int yFrac = qmv.y & 0x3; |
| 1191 | int cost; |
| 1192 | intptr_t lclStride = fencPUYuv.m_size; |
| 1193 | X265_CHECK(lclStride == FENC_STRIDE, "fenc buffer is assumed to have FENC_STRIDE by sad_x3 and sad_x4\n"); |
| 1194 | |
| 1195 | if (!(yFrac | xFrac)) |
| 1196 | cost = cmp(fencPUYuv.m_buf[0], lclStride, fref, refStride); |
| 1197 | else |
| 1198 | { |
| 1199 | /* we are taking a short-cut here if the reference is weighted. To be |
| 1200 | * accurate we should be interpolating unweighted pixels and weighting |
| 1201 | * the final 16bit values prior to rounding and down shifting. Instead we |
| 1202 | * are simply interpolating the weighted full-pel pixels. Not 100% |
| 1203 | * accurate but good enough for fast qpel ME */ |
| 1204 | ALIGN_VAR_32(pixel, subpelbuf[64 * 64]); |
| 1205 | if (!yFrac) |
| 1206 | primitives.luma_hpp[partEnum](fref, refStride, subpelbuf, lclStride, xFrac); |
| 1207 | else if (!xFrac) |
| 1208 | primitives.luma_vpp[partEnum](fref, refStride, subpelbuf, lclStride, yFrac); |
| 1209 | else |
| 1210 | { |
| 1211 | ALIGN_VAR_32(int16_t, immed[64 * (64 + NTAPS_LUMA)]); |
| 1212 | |
| 1213 | int filterSize = NTAPS_LUMA; |
| 1214 | int halfFilterSize = filterSize >> 1; |
| 1215 | primitives.luma_hps[partEnum](fref, refStride, immed, blockwidth, xFrac, 1); |
| 1216 | primitives.luma_vsp[partEnum](immed + (halfFilterSize - 1) * blockwidth, blockwidth, subpelbuf, lclStride, yFrac); |
| 1217 | } |
| 1218 | cost = cmp(fencPUYuv.m_buf[0], lclStride, subpelbuf, lclStride); |
| 1219 | } |
| 1220 | |
| 1221 | if (bChromaSATD) |
| 1222 | { |
| 1223 | int csp = fencPUYuv.m_csp; |
| 1224 | int hshift = fencPUYuv.m_hChromaShift; |
| 1225 | int vshift = fencPUYuv.m_vChromaShift; |
| 1226 | int shiftHor = (2 + hshift); |
| 1227 | int shiftVer = (2 + vshift); |
| 1228 | lclStride = fencPUYuv.m_csize; |
| 1229 | |
| 1230 | intptr_t refStrideC = ref->reconPic->m_strideC; |
| 1231 | intptr_t refOffset = (qmv.x >> shiftHor) + (qmv.y >> shiftVer) * refStrideC; |
| 1232 | |
| 1233 | const pixel* refCb = ref->getCbAddr(ctuAddr, absPartIdx) + refOffset; |
| 1234 | const pixel* refCr = ref->getCrAddr(ctuAddr, absPartIdx) + refOffset; |
| 1235 | |
| 1236 | xFrac = qmv.x & ((1 << shiftHor) - 1); |
| 1237 | yFrac = qmv.y & ((1 << shiftVer) - 1); |
| 1238 | |
| 1239 | if (!(yFrac | xFrac)) |
| 1240 | { |
| 1241 | cost += chromaSatd(fencPUYuv.m_buf[1], lclStride, refCb, refStrideC); |
| 1242 | cost += chromaSatd(fencPUYuv.m_buf[2], lclStride, refCr, refStrideC); |
| 1243 | } |
| 1244 | else |
| 1245 | { |
| 1246 | ALIGN_VAR_32(pixel, subpelbuf[64 * 64]); |
| 1247 | if (!yFrac) |
| 1248 | { |
| 1249 | primitives.chroma[csp].filter_hpp[partEnum](refCb, refStrideC, subpelbuf, lclStride, xFrac << (1 - hshift)); |
| 1250 | cost += chromaSatd(fencPUYuv.m_buf[1], lclStride, subpelbuf, lclStride); |
| 1251 | |
| 1252 | primitives.chroma[csp].filter_hpp[partEnum](refCr, refStrideC, subpelbuf, lclStride, xFrac << (1 - hshift)); |
| 1253 | cost += chromaSatd(fencPUYuv.m_buf[2], lclStride, subpelbuf, lclStride); |
| 1254 | } |
| 1255 | else if (!xFrac) |
| 1256 | { |
| 1257 | primitives.chroma[csp].filter_vpp[partEnum](refCb, refStrideC, subpelbuf, lclStride, yFrac << (1 - vshift)); |
| 1258 | cost += chromaSatd(fencPUYuv.m_buf[1], lclStride, subpelbuf, lclStride); |
| 1259 | |
| 1260 | primitives.chroma[csp].filter_vpp[partEnum](refCr, refStrideC, subpelbuf, lclStride, yFrac << (1 - vshift)); |
| 1261 | cost += chromaSatd(fencPUYuv.m_buf[2], lclStride, subpelbuf, lclStride); |
| 1262 | } |
| 1263 | else |
| 1264 | { |
| 1265 | ALIGN_VAR_32(int16_t, immed[64 * (64 + NTAPS_CHROMA)]); |
| 1266 | |
| 1267 | int extStride = blockwidth >> hshift; |
| 1268 | int filterSize = NTAPS_CHROMA; |
| 1269 | int halfFilterSize = (filterSize >> 1); |
| 1270 | |
| 1271 | primitives.chroma[csp].filter_hps[partEnum](refCb, refStrideC, immed, extStride, xFrac << (1 - hshift), 1); |
| 1272 | primitives.chroma[csp].filter_vsp[partEnum](immed + (halfFilterSize - 1) * extStride, extStride, subpelbuf, lclStride, yFrac << (1 - vshift)); |
| 1273 | cost += chromaSatd(fencPUYuv.m_buf[1], lclStride, subpelbuf, lclStride); |
| 1274 | |
| 1275 | primitives.chroma[csp].filter_hps[partEnum](refCr, refStrideC, immed, extStride, xFrac << (1 - hshift), 1); |
| 1276 | primitives.chroma[csp].filter_vsp[partEnum](immed + (halfFilterSize - 1) * extStride, extStride, subpelbuf, lclStride, yFrac << (1 - vshift)); |
| 1277 | cost += chromaSatd(fencPUYuv.m_buf[2], lclStride, subpelbuf, lclStride); |
| 1278 | } |
| 1279 | } |
| 1280 | } |
| 1281 | |
| 1282 | return cost; |
| 1283 | } |