1 /* 2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved. 3 * 4 * Use of this source code is governed by a BSD-style license 5 * that can be found in the LICENSE file in the root of the source 6 * tree. An additional intellectual property rights grant can be found 7 * in the file PATENTS. All contributing project authors may 8 * be found in the AUTHORS file in the root of the source tree. 9 */ 10 11 #include "./vpx_config.h" 12 #include "vp9/common/vp9_loopfilter.h" 13 #include "vp9/common/vp9_onyxc_int.h" 14 #include "vp9/common/vp9_reconinter.h" 15 #include "vpx_mem/vpx_mem.h" 16 17 #include "vp9/common/vp9_seg_common.h" 18 19 // This structure holds bit masks for all 8x8 blocks in a 64x64 region. 20 // Each 1 bit represents a position in which we want to apply the loop filter. 21 // Left_ entries refer to whether we apply a filter on the border to the 22 // left of the block. Above_ entries refer to whether or not to apply a 23 // filter on the above border. Int_ entries refer to whether or not to 24 // apply borders on the 4x4 edges within the 8x8 block that each bit 25 // represents. 26 // Since each transform is accompanied by a potentially different type of 27 // loop filter there is a different entry in the array for each transform size. 28 typedef struct { 29 uint64_t left_y[TX_SIZES]; 30 uint64_t above_y[TX_SIZES]; 31 uint64_t int_4x4_y; 32 uint16_t left_uv[TX_SIZES]; 33 uint16_t above_uv[TX_SIZES]; 34 uint16_t int_4x4_uv; 35 uint8_t lfl_y[64]; 36 uint8_t lfl_uv[16]; 37 } LOOP_FILTER_MASK; 38 39 // 64 bit masks for left transform size. Each 1 represents a position where 40 // we should apply a loop filter across the left border of an 8x8 block 41 // boundary. 42 // 43 // In the case of TX_16X16-> ( in low order byte first we end up with 44 // a mask that looks like this 45 // 46 // 10101010 47 // 10101010 48 // 10101010 49 // 10101010 50 // 10101010 51 // 10101010 52 // 10101010 53 // 10101010 54 // 55 // A loopfilter should be applied to every other 8x8 horizontally. 56 static const uint64_t left_64x64_txform_mask[TX_SIZES]= { 57 0xffffffffffffffff, // TX_4X4 58 0xffffffffffffffff, // TX_8x8 59 0x5555555555555555, // TX_16x16 60 0x1111111111111111, // TX_32x32 61 }; 62 63 // 64 bit masks for above transform size. Each 1 represents a position where 64 // we should apply a loop filter across the top border of an 8x8 block 65 // boundary. 66 // 67 // In the case of TX_32x32 -> ( in low order byte first we end up with 68 // a mask that looks like this 69 // 70 // 11111111 71 // 00000000 72 // 00000000 73 // 00000000 74 // 11111111 75 // 00000000 76 // 00000000 77 // 00000000 78 // 79 // A loopfilter should be applied to every other 4 the row vertically. 80 static const uint64_t above_64x64_txform_mask[TX_SIZES]= { 81 0xffffffffffffffff, // TX_4X4 82 0xffffffffffffffff, // TX_8x8 83 0x00ff00ff00ff00ff, // TX_16x16 84 0x000000ff000000ff, // TX_32x32 85 }; 86 87 // 64 bit masks for prediction sizes (left). Each 1 represents a position 88 // where left border of an 8x8 block. These are aligned to the right most 89 // appropriate bit, and then shifted into place. 90 // 91 // In the case of TX_16x32 -> ( low order byte first ) we end up with 92 // a mask that looks like this : 93 // 94 // 10000000 95 // 10000000 96 // 10000000 97 // 10000000 98 // 00000000 99 // 00000000 100 // 00000000 101 // 00000000 102 static const uint64_t left_prediction_mask[BLOCK_SIZES] = { 103 0x0000000000000001, // BLOCK_4X4, 104 0x0000000000000001, // BLOCK_4X8, 105 0x0000000000000001, // BLOCK_8X4, 106 0x0000000000000001, // BLOCK_8X8, 107 0x0000000000000101, // BLOCK_8X16, 108 0x0000000000000001, // BLOCK_16X8, 109 0x0000000000000101, // BLOCK_16X16, 110 0x0000000001010101, // BLOCK_16X32, 111 0x0000000000000101, // BLOCK_32X16, 112 0x0000000001010101, // BLOCK_32X32, 113 0x0101010101010101, // BLOCK_32X64, 114 0x0000000001010101, // BLOCK_64X32, 115 0x0101010101010101, // BLOCK_64X64 116 }; 117 118 // 64 bit mask to shift and set for each prediction size. 119 static const uint64_t above_prediction_mask[BLOCK_SIZES] = { 120 0x0000000000000001, // BLOCK_4X4 121 0x0000000000000001, // BLOCK_4X8 122 0x0000000000000001, // BLOCK_8X4 123 0x0000000000000001, // BLOCK_8X8 124 0x0000000000000001, // BLOCK_8X16, 125 0x0000000000000003, // BLOCK_16X8 126 0x0000000000000003, // BLOCK_16X16 127 0x0000000000000003, // BLOCK_16X32, 128 0x000000000000000f, // BLOCK_32X16, 129 0x000000000000000f, // BLOCK_32X32, 130 0x000000000000000f, // BLOCK_32X64, 131 0x00000000000000ff, // BLOCK_64X32, 132 0x00000000000000ff, // BLOCK_64X64 133 }; 134 // 64 bit mask to shift and set for each prediction size. A bit is set for 135 // each 8x8 block that would be in the left most block of the given block 136 // size in the 64x64 block. 137 static const uint64_t size_mask[BLOCK_SIZES] = { 138 0x0000000000000001, // BLOCK_4X4 139 0x0000000000000001, // BLOCK_4X8 140 0x0000000000000001, // BLOCK_8X4 141 0x0000000000000001, // BLOCK_8X8 142 0x0000000000000101, // BLOCK_8X16, 143 0x0000000000000003, // BLOCK_16X8 144 0x0000000000000303, // BLOCK_16X16 145 0x0000000003030303, // BLOCK_16X32, 146 0x0000000000000f0f, // BLOCK_32X16, 147 0x000000000f0f0f0f, // BLOCK_32X32, 148 0x0f0f0f0f0f0f0f0f, // BLOCK_32X64, 149 0x00000000ffffffff, // BLOCK_64X32, 150 0xffffffffffffffff, // BLOCK_64X64 151 }; 152 153 // These are used for masking the left and above borders. 154 static const uint64_t left_border = 0x1111111111111111; 155 static const uint64_t above_border = 0x000000ff000000ff; 156 157 // 16 bit masks for uv transform sizes. 158 static const uint16_t left_64x64_txform_mask_uv[TX_SIZES]= { 159 0xffff, // TX_4X4 160 0xffff, // TX_8x8 161 0x5555, // TX_16x16 162 0x1111, // TX_32x32 163 }; 164 165 static const uint16_t above_64x64_txform_mask_uv[TX_SIZES]= { 166 0xffff, // TX_4X4 167 0xffff, // TX_8x8 168 0x0f0f, // TX_16x16 169 0x000f, // TX_32x32 170 }; 171 172 // 16 bit left mask to shift and set for each uv prediction size. 173 static const uint16_t left_prediction_mask_uv[BLOCK_SIZES] = { 174 0x0001, // BLOCK_4X4, 175 0x0001, // BLOCK_4X8, 176 0x0001, // BLOCK_8X4, 177 0x0001, // BLOCK_8X8, 178 0x0001, // BLOCK_8X16, 179 0x0001, // BLOCK_16X8, 180 0x0001, // BLOCK_16X16, 181 0x0011, // BLOCK_16X32, 182 0x0001, // BLOCK_32X16, 183 0x0011, // BLOCK_32X32, 184 0x1111, // BLOCK_32X64 185 0x0011, // BLOCK_64X32, 186 0x1111, // BLOCK_64X64 187 }; 188 // 16 bit above mask to shift and set for uv each prediction size. 189 static const uint16_t above_prediction_mask_uv[BLOCK_SIZES] = { 190 0x0001, // BLOCK_4X4 191 0x0001, // BLOCK_4X8 192 0x0001, // BLOCK_8X4 193 0x0001, // BLOCK_8X8 194 0x0001, // BLOCK_8X16, 195 0x0001, // BLOCK_16X8 196 0x0001, // BLOCK_16X16 197 0x0001, // BLOCK_16X32, 198 0x0003, // BLOCK_32X16, 199 0x0003, // BLOCK_32X32, 200 0x0003, // BLOCK_32X64, 201 0x000f, // BLOCK_64X32, 202 0x000f, // BLOCK_64X64 203 }; 204 205 // 64 bit mask to shift and set for each uv prediction size 206 static const uint16_t size_mask_uv[BLOCK_SIZES] = { 207 0x0001, // BLOCK_4X4 208 0x0001, // BLOCK_4X8 209 0x0001, // BLOCK_8X4 210 0x0001, // BLOCK_8X8 211 0x0001, // BLOCK_8X16, 212 0x0001, // BLOCK_16X8 213 0x0001, // BLOCK_16X16 214 0x0011, // BLOCK_16X32, 215 0x0003, // BLOCK_32X16, 216 0x0033, // BLOCK_32X32, 217 0x3333, // BLOCK_32X64, 218 0x00ff, // BLOCK_64X32, 219 0xffff, // BLOCK_64X64 220 }; 221 static const uint16_t left_border_uv = 0x1111; 222 static const uint16_t above_border_uv = 0x000f; 223 224 225 static void lf_init_lut(loop_filter_info_n *lfi) { 226 lfi->mode_lf_lut[DC_PRED] = 0; 227 lfi->mode_lf_lut[D45_PRED] = 0; 228 lfi->mode_lf_lut[D135_PRED] = 0; 229 lfi->mode_lf_lut[D117_PRED] = 0; 230 lfi->mode_lf_lut[D153_PRED] = 0; 231 lfi->mode_lf_lut[D207_PRED] = 0; 232 lfi->mode_lf_lut[D63_PRED] = 0; 233 lfi->mode_lf_lut[V_PRED] = 0; 234 lfi->mode_lf_lut[H_PRED] = 0; 235 lfi->mode_lf_lut[TM_PRED] = 0; 236 lfi->mode_lf_lut[ZEROMV] = 0; 237 lfi->mode_lf_lut[NEARESTMV] = 1; 238 lfi->mode_lf_lut[NEARMV] = 1; 239 lfi->mode_lf_lut[NEWMV] = 1; 240 } 241 242 static void update_sharpness(loop_filter_info_n *lfi, int sharpness_lvl) { 243 int lvl; 244 245 // For each possible value for the loop filter fill out limits 246 for (lvl = 0; lvl <= MAX_LOOP_FILTER; lvl++) { 247 // Set loop filter paramaeters that control sharpness. 248 int block_inside_limit = lvl >> ((sharpness_lvl > 0) + (sharpness_lvl > 4)); 249 250 if (sharpness_lvl > 0) { 251 if (block_inside_limit > (9 - sharpness_lvl)) 252 block_inside_limit = (9 - sharpness_lvl); 253 } 254 255 if (block_inside_limit < 1) 256 block_inside_limit = 1; 257 258 vpx_memset(lfi->lfthr[lvl].lim, block_inside_limit, SIMD_WIDTH); 259 vpx_memset(lfi->lfthr[lvl].mblim, (2 * (lvl + 2) + block_inside_limit), 260 SIMD_WIDTH); 261 } 262 } 263 264 void vp9_loop_filter_init(VP9_COMMON *cm) { 265 loop_filter_info_n *lfi = &cm->lf_info; 266 struct loopfilter *lf = &cm->lf; 267 int lvl; 268 269 // init limits for given sharpness 270 update_sharpness(lfi, lf->sharpness_level); 271 lf->last_sharpness_level = lf->sharpness_level; 272 273 // init LUT for lvl and hev thr picking 274 lf_init_lut(lfi); 275 276 // init hev threshold const vectors 277 for (lvl = 0; lvl <= MAX_LOOP_FILTER; lvl++) 278 vpx_memset(lfi->lfthr[lvl].hev_thr, (lvl >> 4), SIMD_WIDTH); 279 } 280 281 void vp9_loop_filter_frame_init(VP9_COMMON *cm, int default_filt_lvl) { 282 int seg_id; 283 // n_shift is the a multiplier for lf_deltas 284 // the multiplier is 1 for when filter_lvl is between 0 and 31; 285 // 2 when filter_lvl is between 32 and 63 286 const int n_shift = default_filt_lvl >> 5; 287 loop_filter_info_n *const lfi = &cm->lf_info; 288 struct loopfilter *const lf = &cm->lf; 289 struct segmentation *const seg = &cm->seg; 290 291 // update limits if sharpness has changed 292 if (lf->last_sharpness_level != lf->sharpness_level) { 293 update_sharpness(lfi, lf->sharpness_level); 294 lf->last_sharpness_level = lf->sharpness_level; 295 } 296 297 for (seg_id = 0; seg_id < MAX_SEGMENTS; seg_id++) { 298 int lvl_seg = default_filt_lvl, ref, mode, intra_lvl; 299 300 // Set the baseline filter values for each segment 301 if (vp9_segfeature_active(seg, seg_id, SEG_LVL_ALT_LF)) { 302 const int data = vp9_get_segdata(seg, seg_id, SEG_LVL_ALT_LF); 303 lvl_seg = seg->abs_delta == SEGMENT_ABSDATA 304 ? data 305 : clamp(default_filt_lvl + data, 0, MAX_LOOP_FILTER); 306 } 307 308 if (!lf->mode_ref_delta_enabled) { 309 // we could get rid of this if we assume that deltas are set to 310 // zero when not in use; encoder always uses deltas 311 vpx_memset(lfi->lvl[seg_id], lvl_seg, sizeof(lfi->lvl[seg_id])); 312 continue; 313 } 314 315 intra_lvl = lvl_seg + lf->ref_deltas[INTRA_FRAME] * (1 << n_shift); 316 lfi->lvl[seg_id][INTRA_FRAME][0] = clamp(intra_lvl, 0, MAX_LOOP_FILTER); 317 318 for (ref = LAST_FRAME; ref < MAX_REF_FRAMES; ++ref) 319 for (mode = 0; mode < MAX_MODE_LF_DELTAS; ++mode) { 320 const int inter_lvl = lvl_seg + lf->ref_deltas[ref] * (1 << n_shift) 321 + lf->mode_deltas[mode] * (1 << n_shift); 322 lfi->lvl[seg_id][ref][mode] = clamp(inter_lvl, 0, MAX_LOOP_FILTER); 323 } 324 } 325 } 326 327 static uint8_t build_lfi(const loop_filter_info_n *lfi_n, 328 const MB_MODE_INFO *mbmi) { 329 const int seg = mbmi->segment_id; 330 const int ref = mbmi->ref_frame[0]; 331 const int mode = lfi_n->mode_lf_lut[mbmi->mode]; 332 const int filter_level = lfi_n->lvl[seg][ref][mode]; 333 334 return filter_level; 335 } 336 337 static void filter_selectively_vert(uint8_t *s, int pitch, 338 unsigned int mask_16x16, 339 unsigned int mask_8x8, 340 unsigned int mask_4x4, 341 unsigned int mask_4x4_int, 342 const loop_filter_info_n *lfi_n, 343 const uint8_t *lfl) { 344 unsigned int mask; 345 346 for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int; 347 mask; mask >>= 1) { 348 const loop_filter_thresh *lfi = lfi_n->lfthr + *lfl; 349 350 if (mask & 1) { 351 if (mask_16x16 & 1) { 352 vp9_mb_lpf_vertical_edge_w(s, pitch, lfi->mblim, lfi->lim, 353 lfi->hev_thr); 354 assert(!(mask_8x8 & 1)); 355 assert(!(mask_4x4 & 1)); 356 assert(!(mask_4x4_int & 1)); 357 } else if (mask_8x8 & 1) { 358 vp9_mbloop_filter_vertical_edge(s, pitch, lfi->mblim, lfi->lim, 359 lfi->hev_thr, 1); 360 assert(!(mask_16x16 & 1)); 361 assert(!(mask_4x4 & 1)); 362 } else if (mask_4x4 & 1) { 363 vp9_loop_filter_vertical_edge(s, pitch, lfi->mblim, lfi->lim, 364 lfi->hev_thr, 1); 365 assert(!(mask_16x16 & 1)); 366 assert(!(mask_8x8 & 1)); 367 } 368 } 369 if (mask_4x4_int & 1) 370 vp9_loop_filter_vertical_edge(s + 4, pitch, lfi->mblim, lfi->lim, 371 lfi->hev_thr, 1); 372 s += 8; 373 lfl += 1; 374 mask_16x16 >>= 1; 375 mask_8x8 >>= 1; 376 mask_4x4 >>= 1; 377 mask_4x4_int >>= 1; 378 } 379 } 380 381 static void filter_selectively_horiz(uint8_t *s, int pitch, 382 unsigned int mask_16x16, 383 unsigned int mask_8x8, 384 unsigned int mask_4x4, 385 unsigned int mask_4x4_int, 386 const loop_filter_info_n *lfi_n, 387 const uint8_t *lfl) { 388 unsigned int mask; 389 int count; 390 391 for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int; 392 mask; mask >>= count) { 393 const loop_filter_thresh *lfi = lfi_n->lfthr + *lfl; 394 395 count = 1; 396 if (mask & 1) { 397 if (mask_16x16 & 1) { 398 if ((mask_16x16 & 3) == 3) { 399 vp9_mb_lpf_horizontal_edge_w(s, pitch, lfi->mblim, lfi->lim, 400 lfi->hev_thr, 2); 401 count = 2; 402 } else { 403 vp9_mb_lpf_horizontal_edge_w(s, pitch, lfi->mblim, lfi->lim, 404 lfi->hev_thr, 1); 405 } 406 assert(!(mask_8x8 & 1)); 407 assert(!(mask_4x4 & 1)); 408 assert(!(mask_4x4_int & 1)); 409 } else if (mask_8x8 & 1) { 410 if ((mask_8x8 & 3) == 3) { 411 // Next block's thresholds 412 const loop_filter_thresh *lfin = lfi_n->lfthr + *(lfl + 1); 413 414 // TODO(yunqingwang): Combine next 2 calls as 1 wide filtering. 415 vp9_mbloop_filter_horizontal_edge(s, pitch, lfi->mblim, lfi->lim, 416 lfi->hev_thr, 1); 417 vp9_mbloop_filter_horizontal_edge(s + 8, pitch, lfin->mblim, 418 lfin->lim, lfin->hev_thr, 1); 419 420 if ((mask_4x4_int & 3) == 3) { 421 // TODO(yunqingwang): Combine next 2 calls as 1 wide filtering. 422 vp9_loop_filter_horizontal_edge(s + 4 * pitch, pitch, lfi->mblim, 423 lfi->lim, lfi->hev_thr, 1); 424 vp9_loop_filter_horizontal_edge(s + 8 + 4 * pitch, pitch, 425 lfin->mblim, lfin->lim, 426 lfin->hev_thr, 1); 427 } else { 428 if (mask_4x4_int & 1) 429 vp9_loop_filter_horizontal_edge(s + 4 * pitch, pitch, lfi->mblim, 430 lfi->lim, lfi->hev_thr, 1); 431 else if (mask_4x4_int & 2) 432 vp9_loop_filter_horizontal_edge(s + 8 + 4 * pitch, pitch, 433 lfin->mblim, lfin->lim, 434 lfin->hev_thr, 1); 435 } 436 count = 2; 437 } else { 438 vp9_mbloop_filter_horizontal_edge(s, pitch, lfi->mblim, lfi->lim, 439 lfi->hev_thr, 1); 440 441 if (mask_4x4_int & 1) 442 vp9_loop_filter_horizontal_edge(s + 4 * pitch, pitch, lfi->mblim, 443 lfi->lim, lfi->hev_thr, 1); 444 } 445 assert(!(mask_16x16 & 1)); 446 assert(!(mask_4x4 & 1)); 447 } else if (mask_4x4 & 1) { 448 if ((mask_4x4 & 3) == 3) { 449 // Next block's thresholds 450 const loop_filter_thresh *lfin = lfi_n->lfthr + *(lfl + 1); 451 452 // TODO(yunqingwang): Combine next 2 calls as 1 wide filtering. 453 vp9_loop_filter_horizontal_edge(s, pitch, lfi->mblim, lfi->lim, 454 lfi->hev_thr, 1); 455 vp9_loop_filter_horizontal_edge(s + 8, pitch, lfin->mblim, lfin->lim, 456 lfin->hev_thr, 1); 457 458 if ((mask_4x4_int & 3) == 3) { 459 // TODO(yunqingwang): Combine next 2 calls as 1 wide filtering. 460 vp9_loop_filter_horizontal_edge(s + 4 * pitch, pitch, lfi->mblim, 461 lfi->lim, lfi->hev_thr, 1); 462 vp9_loop_filter_horizontal_edge(s + 8 + 4 * pitch, pitch, 463 lfin->mblim, lfin->lim, 464 lfin->hev_thr, 1); 465 } else { 466 if (mask_4x4_int & 1) 467 vp9_loop_filter_horizontal_edge(s + 4 * pitch, pitch, lfi->mblim, 468 lfi->lim, lfi->hev_thr, 1); 469 else if (mask_4x4_int & 2) 470 vp9_loop_filter_horizontal_edge(s + 8 + 4 * pitch, pitch, 471 lfin->mblim, lfin->lim, 472 lfin->hev_thr, 1); 473 } 474 count = 2; 475 } else { 476 vp9_loop_filter_horizontal_edge(s, pitch, lfi->mblim, lfi->lim, 477 lfi->hev_thr, 1); 478 479 if (mask_4x4_int & 1) 480 vp9_loop_filter_horizontal_edge(s + 4 * pitch, pitch, lfi->mblim, 481 lfi->lim, lfi->hev_thr, 1); 482 } 483 assert(!(mask_16x16 & 1)); 484 assert(!(mask_8x8 & 1)); 485 } else if (mask_4x4_int & 1) { 486 vp9_loop_filter_horizontal_edge(s + 4 * pitch, pitch, lfi->mblim, 487 lfi->lim, lfi->hev_thr, 1); 488 } 489 } 490 s += 8 * count; 491 lfl += count; 492 mask_16x16 >>= count; 493 mask_8x8 >>= count; 494 mask_4x4 >>= count; 495 mask_4x4_int >>= count; 496 } 497 } 498 499 // This function ors into the current lfm structure, where to do loop 500 // filters for the specific mi we are looking at. It uses information 501 // including the block_size_type (32x16, 32x32, etc), the transform size, 502 // whether there were any coefficients encoded, and the loop filter strength 503 // block we are currently looking at. Shift is used to position the 504 // 1's we produce. 505 // TODO(JBB) Need another function for different resolution color.. 506 static void build_masks(const loop_filter_info_n *const lfi_n, 507 const MODE_INFO *mi, const int shift_y, 508 const int shift_uv, 509 LOOP_FILTER_MASK *lfm) { 510 const BLOCK_SIZE block_size = mi->mbmi.sb_type; 511 const TX_SIZE tx_size_y = mi->mbmi.tx_size; 512 const TX_SIZE tx_size_uv = get_uv_tx_size(&mi->mbmi); 513 const int skip = mi->mbmi.skip_coeff; 514 const int seg = mi->mbmi.segment_id; 515 const int ref = mi->mbmi.ref_frame[0]; 516 const int mode = lfi_n->mode_lf_lut[mi->mbmi.mode]; 517 const int filter_level = lfi_n->lvl[seg][ref][mode]; 518 uint64_t *left_y = &lfm->left_y[tx_size_y]; 519 uint64_t *above_y = &lfm->above_y[tx_size_y]; 520 uint64_t *int_4x4_y = &lfm->int_4x4_y; 521 uint16_t *left_uv = &lfm->left_uv[tx_size_uv]; 522 uint16_t *above_uv = &lfm->above_uv[tx_size_uv]; 523 uint16_t *int_4x4_uv = &lfm->int_4x4_uv; 524 int i; 525 int w = num_8x8_blocks_wide_lookup[block_size]; 526 int h = num_8x8_blocks_high_lookup[block_size]; 527 528 // If filter level is 0 we don't loop filter. 529 if (!filter_level) { 530 return; 531 } else { 532 int index = shift_y; 533 for (i = 0; i < h; i++) { 534 vpx_memset(&lfm->lfl_y[index], filter_level, w); 535 index += 8; 536 } 537 } 538 539 // These set 1 in the current block size for the block size edges. 540 // For instance if the block size is 32x16, we'll set : 541 // above = 1111 542 // 0000 543 // and 544 // left = 1000 545 // = 1000 546 // NOTE : In this example the low bit is left most ( 1000 ) is stored as 547 // 1, not 8... 548 // 549 // U and v set things on a 16 bit scale. 550 // 551 *above_y |= above_prediction_mask[block_size] << shift_y; 552 *above_uv |= above_prediction_mask_uv[block_size] << shift_uv; 553 *left_y |= left_prediction_mask[block_size] << shift_y; 554 *left_uv |= left_prediction_mask_uv[block_size] << shift_uv; 555 556 // If the block has no coefficients and is not intra we skip applying 557 // the loop filter on block edges. 558 if (skip && ref > INTRA_FRAME) 559 return; 560 561 // Here we are adding a mask for the transform size. The transform 562 // size mask is set to be correct for a 64x64 prediction block size. We 563 // mask to match the size of the block we are working on and then shift it 564 // into place.. 565 *above_y |= (size_mask[block_size] & 566 above_64x64_txform_mask[tx_size_y]) << shift_y; 567 *above_uv |= (size_mask_uv[block_size] & 568 above_64x64_txform_mask_uv[tx_size_uv]) << shift_uv; 569 570 *left_y |= (size_mask[block_size] & 571 left_64x64_txform_mask[tx_size_y]) << shift_y; 572 *left_uv |= (size_mask_uv[block_size] & 573 left_64x64_txform_mask_uv[tx_size_uv]) << shift_uv; 574 575 // Here we are trying to determine what to do with the internal 4x4 block 576 // boundaries. These differ from the 4x4 boundaries on the outside edge of 577 // an 8x8 in that the internal ones can be skipped and don't depend on 578 // the prediction block size. 579 if (tx_size_y == TX_4X4) { 580 *int_4x4_y |= (size_mask[block_size] & 0xffffffffffffffff) << shift_y; 581 } 582 if (tx_size_uv == TX_4X4) { 583 *int_4x4_uv |= (size_mask_uv[block_size] & 0xffff) << shift_uv; 584 } 585 } 586 587 // This function does the same thing as the one above with the exception that 588 // it only affects the y masks. It exists because for blocks < 16x16 in size, 589 // we only update u and v masks on the first block. 590 static void build_y_mask(const loop_filter_info_n *const lfi_n, 591 const MODE_INFO *mi, const int shift_y, 592 LOOP_FILTER_MASK *lfm) { 593 const BLOCK_SIZE block_size = mi->mbmi.sb_type; 594 const TX_SIZE tx_size_y = mi->mbmi.tx_size; 595 const int skip = mi->mbmi.skip_coeff; 596 const int seg = mi->mbmi.segment_id; 597 const int ref = mi->mbmi.ref_frame[0]; 598 const int mode = lfi_n->mode_lf_lut[mi->mbmi.mode]; 599 const int filter_level = lfi_n->lvl[seg][ref][mode]; 600 uint64_t *left_y = &lfm->left_y[tx_size_y]; 601 uint64_t *above_y = &lfm->above_y[tx_size_y]; 602 uint64_t *int_4x4_y = &lfm->int_4x4_y; 603 int i; 604 int w = num_8x8_blocks_wide_lookup[block_size]; 605 int h = num_8x8_blocks_high_lookup[block_size]; 606 607 if (!filter_level) { 608 return; 609 } else { 610 int index = shift_y; 611 for (i = 0; i < h; i++) { 612 vpx_memset(&lfm->lfl_y[index], filter_level, w); 613 index += 8; 614 } 615 } 616 617 *above_y |= above_prediction_mask[block_size] << shift_y; 618 *left_y |= left_prediction_mask[block_size] << shift_y; 619 620 if (skip && ref > INTRA_FRAME) 621 return; 622 623 *above_y |= (size_mask[block_size] & 624 above_64x64_txform_mask[tx_size_y]) << shift_y; 625 626 *left_y |= (size_mask[block_size] & 627 left_64x64_txform_mask[tx_size_y]) << shift_y; 628 629 if (tx_size_y == TX_4X4) { 630 *int_4x4_y |= (size_mask[block_size] & 0xffffffffffffffff) << shift_y; 631 } 632 } 633 634 // This function sets up the bit masks for the entire 64x64 region represented 635 // by mi_row, mi_col. 636 // TODO(JBB): This function only works for yv12. 637 static void setup_mask(VP9_COMMON *const cm, const int mi_row, const int mi_col, 638 MODE_INFO **mi_8x8, const int mode_info_stride, 639 LOOP_FILTER_MASK *lfm) { 640 int idx_32, idx_16, idx_8; 641 const loop_filter_info_n *const lfi_n = &cm->lf_info; 642 MODE_INFO **mip = mi_8x8; 643 MODE_INFO **mip2 = mi_8x8; 644 645 // These are offsets to the next mi in the 64x64 block. It is what gets 646 // added to the mi ptr as we go through each loop. It helps us to avoids 647 // setting up special row and column counters for each index. The last step 648 // brings us out back to the starting position. 649 const int offset_32[] = {4, (mode_info_stride << 2) - 4, 4, 650 -(mode_info_stride << 2) - 4}; 651 const int offset_16[] = {2, (mode_info_stride << 1) - 2, 2, 652 -(mode_info_stride << 1) - 2}; 653 const int offset[] = {1, mode_info_stride - 1, 1, -mode_info_stride - 1}; 654 655 // Following variables represent shifts to position the current block 656 // mask over the appropriate block. A shift of 36 to the left will move 657 // the bits for the final 32 by 32 block in the 64x64 up 4 rows and left 658 // 4 rows to the appropriate spot. 659 const int shift_32_y[] = {0, 4, 32, 36}; 660 const int shift_16_y[] = {0, 2, 16, 18}; 661 const int shift_8_y[] = {0, 1, 8, 9}; 662 const int shift_32_uv[] = {0, 2, 8, 10}; 663 const int shift_16_uv[] = {0, 1, 4, 5}; 664 int i; 665 const int max_rows = (mi_row + MI_BLOCK_SIZE > cm->mi_rows ? 666 cm->mi_rows - mi_row : MI_BLOCK_SIZE); 667 const int max_cols = (mi_col + MI_BLOCK_SIZE > cm->mi_cols ? 668 cm->mi_cols - mi_col : MI_BLOCK_SIZE); 669 670 vp9_zero(*lfm); 671 672 // TODO(jimbankoski): Try moving most of the following code into decode 673 // loop and storing lfm in the mbmi structure so that we don't have to go 674 // through the recursive loop structure multiple times. 675 switch (mip[0]->mbmi.sb_type) { 676 case BLOCK_64X64: 677 build_masks(lfi_n, mip[0] , 0, 0, lfm); 678 break; 679 case BLOCK_64X32: 680 build_masks(lfi_n, mip[0], 0, 0, lfm); 681 mip2 = mip + mode_info_stride * 4; 682 if (4 >= max_rows) 683 break; 684 build_masks(lfi_n, mip2[0], 32, 8, lfm); 685 break; 686 case BLOCK_32X64: 687 build_masks(lfi_n, mip[0], 0, 0, lfm); 688 mip2 = mip + 4; 689 if (4 >= max_cols) 690 break; 691 build_masks(lfi_n, mip2[0], 4, 2, lfm); 692 break; 693 default: 694 for (idx_32 = 0; idx_32 < 4; mip += offset_32[idx_32], ++idx_32) { 695 const int shift_y = shift_32_y[idx_32]; 696 const int shift_uv = shift_32_uv[idx_32]; 697 const int mi_32_col_offset = ((idx_32 & 1) << 2); 698 const int mi_32_row_offset = ((idx_32 >> 1) << 2); 699 if (mi_32_col_offset >= max_cols || mi_32_row_offset >= max_rows) 700 continue; 701 switch (mip[0]->mbmi.sb_type) { 702 case BLOCK_32X32: 703 build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm); 704 break; 705 case BLOCK_32X16: 706 build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm); 707 if (mi_32_row_offset + 2 >= max_rows) 708 continue; 709 mip2 = mip + mode_info_stride * 2; 710 build_masks(lfi_n, mip2[0], shift_y + 16, shift_uv + 4, lfm); 711 break; 712 case BLOCK_16X32: 713 build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm); 714 if (mi_32_col_offset + 2 >= max_cols) 715 continue; 716 mip2 = mip + 2; 717 build_masks(lfi_n, mip2[0], shift_y + 2, shift_uv + 1, lfm); 718 break; 719 default: 720 for (idx_16 = 0; idx_16 < 4; mip += offset_16[idx_16], ++idx_16) { 721 const int shift_y = shift_32_y[idx_32] + shift_16_y[idx_16]; 722 const int shift_uv = shift_32_uv[idx_32] + shift_16_uv[idx_16]; 723 const int mi_16_col_offset = mi_32_col_offset + 724 ((idx_16 & 1) << 1); 725 const int mi_16_row_offset = mi_32_row_offset + 726 ((idx_16 >> 1) << 1); 727 728 if (mi_16_col_offset >= max_cols || mi_16_row_offset >= max_rows) 729 continue; 730 731 switch (mip[0]->mbmi.sb_type) { 732 case BLOCK_16X16: 733 build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm); 734 break; 735 case BLOCK_16X8: 736 build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm); 737 if (mi_16_row_offset + 1 >= max_rows) 738 continue; 739 mip2 = mip + mode_info_stride; 740 build_y_mask(lfi_n, mip2[0], shift_y+8, lfm); 741 break; 742 case BLOCK_8X16: 743 build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm); 744 if (mi_16_col_offset +1 >= max_cols) 745 continue; 746 mip2 = mip + 1; 747 build_y_mask(lfi_n, mip2[0], shift_y+1, lfm); 748 break; 749 default: { 750 const int shift_y = shift_32_y[idx_32] + 751 shift_16_y[idx_16] + 752 shift_8_y[0]; 753 build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm); 754 mip += offset[0]; 755 for (idx_8 = 1; idx_8 < 4; mip += offset[idx_8], ++idx_8) { 756 const int shift_y = shift_32_y[idx_32] + 757 shift_16_y[idx_16] + 758 shift_8_y[idx_8]; 759 const int mi_8_col_offset = mi_16_col_offset + 760 ((idx_8 & 1)); 761 const int mi_8_row_offset = mi_16_row_offset + 762 ((idx_8 >> 1)); 763 764 if (mi_8_col_offset >= max_cols || 765 mi_8_row_offset >= max_rows) 766 continue; 767 build_y_mask(lfi_n, mip[0], shift_y, lfm); 768 } 769 break; 770 } 771 } 772 } 773 break; 774 } 775 } 776 break; 777 } 778 // The largest loopfilter we have is 16x16 so we use the 16x16 mask 779 // for 32x32 transforms also also. 780 lfm->left_y[TX_16X16] |= lfm->left_y[TX_32X32]; 781 lfm->above_y[TX_16X16] |= lfm->above_y[TX_32X32]; 782 lfm->left_uv[TX_16X16] |= lfm->left_uv[TX_32X32]; 783 lfm->above_uv[TX_16X16] |= lfm->above_uv[TX_32X32]; 784 785 // We do at least 8 tap filter on every 32x32 even if the transform size 786 // is 4x4. So if the 4x4 is set on a border pixel add it to the 8x8 and 787 // remove it from the 4x4. 788 lfm->left_y[TX_8X8] |= lfm->left_y[TX_4X4] & left_border; 789 lfm->left_y[TX_4X4] &= ~left_border; 790 lfm->above_y[TX_8X8] |= lfm->above_y[TX_4X4] & above_border; 791 lfm->above_y[TX_4X4] &= ~above_border; 792 lfm->left_uv[TX_8X8] |= lfm->left_uv[TX_4X4] & left_border_uv; 793 lfm->left_uv[TX_4X4] &= ~left_border_uv; 794 lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_4X4] & above_border_uv; 795 lfm->above_uv[TX_4X4] &= ~above_border_uv; 796 797 // We do some special edge handling. 798 if (mi_row + MI_BLOCK_SIZE > cm->mi_rows) { 799 const uint64_t rows = cm->mi_rows - mi_row; 800 801 // Each pixel inside the border gets a 1, 802 const uint64_t mask_y = (((uint64_t) 1 << (rows << 3)) - 1); 803 const uint16_t mask_uv = (((uint16_t) 1 << (((rows + 1) >> 1) << 2)) - 1); 804 805 // Remove values completely outside our border. 806 for (i = 0; i < TX_32X32; i++) { 807 lfm->left_y[i] &= mask_y; 808 lfm->above_y[i] &= mask_y; 809 lfm->left_uv[i] &= mask_uv; 810 lfm->above_uv[i] &= mask_uv; 811 } 812 lfm->int_4x4_y &= mask_y; 813 lfm->int_4x4_uv &= mask_uv; 814 815 // We don't apply a wide loop filter on the last uv block row. If set 816 // apply the shorter one instead. 817 if (rows == 1) { 818 lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_16X16]; 819 lfm->above_uv[TX_16X16] = 0; 820 } 821 if (rows == 5) { 822 lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_16X16] & 0xff00; 823 lfm->above_uv[TX_16X16] &= ~(lfm->above_uv[TX_16X16] & 0xff00); 824 } 825 } 826 827 if (mi_col + MI_BLOCK_SIZE > cm->mi_cols) { 828 const uint64_t columns = cm->mi_cols - mi_col; 829 830 // Each pixel inside the border gets a 1, the multiply copies the border 831 // to where we need it. 832 const uint64_t mask_y = (((1 << columns) - 1)) * 0x0101010101010101; 833 const uint16_t mask_uv = ((1 << ((columns + 1) >> 1)) - 1) * 0x1111; 834 835 // Internal edges are not applied on the last column of the image so 836 // we mask 1 more for the internal edges 837 const uint16_t mask_uv_int = ((1 << (columns >> 1)) - 1) * 0x1111; 838 839 // Remove the bits outside the image edge. 840 for (i = 0; i < TX_32X32; i++) { 841 lfm->left_y[i] &= mask_y; 842 lfm->above_y[i] &= mask_y; 843 lfm->left_uv[i] &= mask_uv; 844 lfm->above_uv[i] &= mask_uv; 845 } 846 lfm->int_4x4_y &= mask_y; 847 lfm->int_4x4_uv &= mask_uv_int; 848 849 // We don't apply a wide loop filter on the last uv column. If set 850 // apply the shorter one instead. 851 if (columns == 1) { 852 lfm->left_uv[TX_8X8] |= lfm->left_uv[TX_16X16]; 853 lfm->left_uv[TX_16X16] = 0; 854 } 855 if (columns == 5) { 856 lfm->left_uv[TX_8X8] |= (lfm->left_uv[TX_16X16] & 0xcccc); 857 lfm->left_uv[TX_16X16] &= ~(lfm->left_uv[TX_16X16] & 0xcccc); 858 } 859 } 860 // We don't a loop filter on the first column in the image. Mask that out. 861 if (mi_col == 0) { 862 for (i = 0; i < TX_32X32; i++) { 863 lfm->left_y[i] &= 0xfefefefefefefefe; 864 lfm->left_uv[i] &= 0xeeee; 865 } 866 } 867 } 868 869 #if CONFIG_NON420 870 static void filter_block_plane_non420(VP9_COMMON *cm, 871 struct macroblockd_plane *plane, 872 MODE_INFO **mi_8x8, 873 int mi_row, int mi_col) { 874 const int ss_x = plane->subsampling_x; 875 const int ss_y = plane->subsampling_y; 876 const int row_step = 1 << ss_x; 877 const int col_step = 1 << ss_y; 878 const int row_step_stride = cm->mode_info_stride * row_step; 879 struct buf_2d *const dst = &plane->dst; 880 uint8_t* const dst0 = dst->buf; 881 unsigned int mask_16x16[MI_BLOCK_SIZE] = {0}; 882 unsigned int mask_8x8[MI_BLOCK_SIZE] = {0}; 883 unsigned int mask_4x4[MI_BLOCK_SIZE] = {0}; 884 unsigned int mask_4x4_int[MI_BLOCK_SIZE] = {0}; 885 uint8_t lfl[MI_BLOCK_SIZE * MI_BLOCK_SIZE]; 886 int r, c; 887 888 for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += row_step) { 889 unsigned int mask_16x16_c = 0; 890 unsigned int mask_8x8_c = 0; 891 unsigned int mask_4x4_c = 0; 892 unsigned int border_mask; 893 894 // Determine the vertical edges that need filtering 895 for (c = 0; c < MI_BLOCK_SIZE && mi_col + c < cm->mi_cols; c += col_step) { 896 const MODE_INFO *mi = mi_8x8[c]; 897 const int skip_this = mi[0].mbmi.skip_coeff 898 && is_inter_block(&mi[0].mbmi); 899 // left edge of current unit is block/partition edge -> no skip 900 const int block_edge_left = b_width_log2(mi[0].mbmi.sb_type) ? 901 !(c & ((1 << (b_width_log2(mi[0].mbmi.sb_type)-1)) - 1)) : 1; 902 const int skip_this_c = skip_this && !block_edge_left; 903 // top edge of current unit is block/partition edge -> no skip 904 const int block_edge_above = b_height_log2(mi[0].mbmi.sb_type) ? 905 !(r & ((1 << (b_height_log2(mi[0].mbmi.sb_type)-1)) - 1)) : 1; 906 const int skip_this_r = skip_this && !block_edge_above; 907 const TX_SIZE tx_size = (plane->plane_type == PLANE_TYPE_UV) 908 ? get_uv_tx_size(&mi[0].mbmi) 909 : mi[0].mbmi.tx_size; 910 const int skip_border_4x4_c = ss_x && mi_col + c == cm->mi_cols - 1; 911 const int skip_border_4x4_r = ss_y && mi_row + r == cm->mi_rows - 1; 912 913 // Filter level can vary per MI 914 if (!(lfl[(r << 3) + (c >> ss_x)] = 915 build_lfi(&cm->lf_info, &mi[0].mbmi))) 916 continue; 917 918 // Build masks based on the transform size of each block 919 if (tx_size == TX_32X32) { 920 if (!skip_this_c && ((c >> ss_x) & 3) == 0) { 921 if (!skip_border_4x4_c) 922 mask_16x16_c |= 1 << (c >> ss_x); 923 else 924 mask_8x8_c |= 1 << (c >> ss_x); 925 } 926 if (!skip_this_r && ((r >> ss_y) & 3) == 0) { 927 if (!skip_border_4x4_r) 928 mask_16x16[r] |= 1 << (c >> ss_x); 929 else 930 mask_8x8[r] |= 1 << (c >> ss_x); 931 } 932 } else if (tx_size == TX_16X16) { 933 if (!skip_this_c && ((c >> ss_x) & 1) == 0) { 934 if (!skip_border_4x4_c) 935 mask_16x16_c |= 1 << (c >> ss_x); 936 else 937 mask_8x8_c |= 1 << (c >> ss_x); 938 } 939 if (!skip_this_r && ((r >> ss_y) & 1) == 0) { 940 if (!skip_border_4x4_r) 941 mask_16x16[r] |= 1 << (c >> ss_x); 942 else 943 mask_8x8[r] |= 1 << (c >> ss_x); 944 } 945 } else { 946 // force 8x8 filtering on 32x32 boundaries 947 if (!skip_this_c) { 948 if (tx_size == TX_8X8 || ((c >> ss_x) & 3) == 0) 949 mask_8x8_c |= 1 << (c >> ss_x); 950 else 951 mask_4x4_c |= 1 << (c >> ss_x); 952 } 953 954 if (!skip_this_r) { 955 if (tx_size == TX_8X8 || ((r >> ss_y) & 3) == 0) 956 mask_8x8[r] |= 1 << (c >> ss_x); 957 else 958 mask_4x4[r] |= 1 << (c >> ss_x); 959 } 960 961 if (!skip_this && tx_size < TX_8X8 && !skip_border_4x4_c) 962 mask_4x4_int[r] |= 1 << (c >> ss_x); 963 } 964 } 965 966 // Disable filtering on the leftmost column 967 border_mask = ~(mi_col == 0); 968 filter_selectively_vert(dst->buf, dst->stride, 969 mask_16x16_c & border_mask, 970 mask_8x8_c & border_mask, 971 mask_4x4_c & border_mask, 972 mask_4x4_int[r], 973 &cm->lf_info, &lfl[r << 3]); 974 dst->buf += 8 * dst->stride; 975 mi_8x8 += row_step_stride; 976 } 977 978 // Now do horizontal pass 979 dst->buf = dst0; 980 for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += row_step) { 981 const int skip_border_4x4_r = ss_y && mi_row + r == cm->mi_rows - 1; 982 const unsigned int mask_4x4_int_r = skip_border_4x4_r ? 0 : mask_4x4_int[r]; 983 984 unsigned int mask_16x16_r; 985 unsigned int mask_8x8_r; 986 unsigned int mask_4x4_r; 987 988 if (mi_row + r == 0) { 989 mask_16x16_r = 0; 990 mask_8x8_r = 0; 991 mask_4x4_r = 0; 992 } else { 993 mask_16x16_r = mask_16x16[r]; 994 mask_8x8_r = mask_8x8[r]; 995 mask_4x4_r = mask_4x4[r]; 996 } 997 998 filter_selectively_horiz(dst->buf, dst->stride, 999 mask_16x16_r, 1000 mask_8x8_r, 1001 mask_4x4_r, 1002 mask_4x4_int_r, 1003 &cm->lf_info, &lfl[r << 3]); 1004 dst->buf += 8 * dst->stride; 1005 } 1006 } 1007 #endif 1008 1009 static void filter_block_plane(VP9_COMMON *const cm, 1010 struct macroblockd_plane *const plane, 1011 int mi_row, 1012 LOOP_FILTER_MASK *lfm) { 1013 struct buf_2d *const dst = &plane->dst; 1014 uint8_t* const dst0 = dst->buf; 1015 unsigned int mask_4x4_int_row[MI_BLOCK_SIZE] = {0}; 1016 int r, c; 1017 1018 if (!plane->plane_type) { 1019 uint64_t mask_16x16 = lfm->left_y[TX_16X16]; 1020 uint64_t mask_8x8 = lfm->left_y[TX_8X8]; 1021 uint64_t mask_4x4 = lfm->left_y[TX_4X4]; 1022 uint64_t mask_4x4_int = lfm->int_4x4_y; 1023 1024 // Vertical pass 1025 for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r++) { 1026 mask_4x4_int_row[r] = mask_4x4_int & 0xff; 1027 1028 // Disable filtering on the leftmost column 1029 filter_selectively_vert(dst->buf, dst->stride, 1030 mask_16x16 & 0xff, 1031 mask_8x8 & 0xff, 1032 mask_4x4 & 0xff, 1033 mask_4x4_int_row[r], 1034 &cm->lf_info, &lfm->lfl_y[r << 3]); 1035 1036 dst->buf += 8 * dst->stride; 1037 mask_16x16 >>= 8; 1038 mask_8x8 >>= 8; 1039 mask_4x4 >>= 8; 1040 mask_4x4_int >>= 8; 1041 } 1042 1043 // Horizontal pass 1044 dst->buf = dst0; 1045 mask_16x16 = lfm->above_y[TX_16X16]; 1046 mask_8x8 = lfm->above_y[TX_8X8]; 1047 mask_4x4 = lfm->above_y[TX_4X4]; 1048 1049 for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r++) { 1050 unsigned int mask_16x16_r; 1051 unsigned int mask_8x8_r; 1052 unsigned int mask_4x4_r; 1053 1054 if (mi_row + r == 0) { 1055 mask_16x16_r = 0; 1056 mask_8x8_r = 0; 1057 mask_4x4_r = 0; 1058 } else { 1059 mask_16x16_r = mask_16x16 & 0xff; 1060 mask_8x8_r = mask_8x8 & 0xff; 1061 mask_4x4_r = mask_4x4 & 0xff; 1062 } 1063 1064 filter_selectively_horiz(dst->buf, dst->stride, 1065 mask_16x16_r, 1066 mask_8x8_r, 1067 mask_4x4_r, 1068 mask_4x4_int_row[r], 1069 &cm->lf_info, &lfm->lfl_y[r << 3]); 1070 1071 dst->buf += 8 * dst->stride; 1072 mask_16x16 >>= 8; 1073 mask_8x8 >>= 8; 1074 mask_4x4 >>= 8; 1075 } 1076 } else { 1077 uint16_t mask_16x16 = lfm->left_uv[TX_16X16]; 1078 uint16_t mask_8x8 = lfm->left_uv[TX_8X8]; 1079 uint16_t mask_4x4 = lfm->left_uv[TX_4X4]; 1080 uint16_t mask_4x4_int = lfm->int_4x4_uv; 1081 1082 // Vertical pass 1083 for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += 2) { 1084 if (plane->plane_type == 1) { 1085 for (c = 0; c < (MI_BLOCK_SIZE >> 1); c++) 1086 lfm->lfl_uv[(r << 1) + c] = lfm->lfl_y[(r << 3) + (c << 1)]; 1087 } 1088 1089 mask_4x4_int_row[r] = mask_4x4_int & 0xf; 1090 // Disable filtering on the leftmost column 1091 filter_selectively_vert(dst->buf, dst->stride, 1092 mask_16x16 & 0xf, 1093 mask_8x8 & 0xf, 1094 mask_4x4 & 0xf, 1095 mask_4x4_int_row[r], 1096 &cm->lf_info, &lfm->lfl_uv[r << 1]); 1097 1098 dst->buf += 8 * dst->stride; 1099 mask_16x16 >>= 4; 1100 mask_8x8 >>= 4; 1101 mask_4x4 >>= 4; 1102 mask_4x4_int >>= 4; 1103 } 1104 1105 // Horizontal pass 1106 dst->buf = dst0; 1107 mask_16x16 = lfm->above_uv[TX_16X16]; 1108 mask_8x8 = lfm->above_uv[TX_8X8]; 1109 mask_4x4 = lfm->above_uv[TX_4X4]; 1110 1111 for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += 2) { 1112 const int skip_border_4x4_r = mi_row + r == cm->mi_rows - 1; 1113 const unsigned int mask_4x4_int_r = skip_border_4x4_r ? 1114 0 : (mask_4x4_int_row[r]); 1115 unsigned int mask_16x16_r; 1116 unsigned int mask_8x8_r; 1117 unsigned int mask_4x4_r; 1118 1119 if (mi_row + r == 0) { 1120 mask_16x16_r = 0; 1121 mask_8x8_r = 0; 1122 mask_4x4_r = 0; 1123 } else { 1124 mask_16x16_r = mask_16x16 & 0xf; 1125 mask_8x8_r = mask_8x8 & 0xf; 1126 mask_4x4_r = mask_4x4 & 0xf; 1127 } 1128 1129 filter_selectively_horiz(dst->buf, dst->stride, 1130 mask_16x16_r, 1131 mask_8x8_r, 1132 mask_4x4_r, 1133 mask_4x4_int_r, 1134 &cm->lf_info, &lfm->lfl_uv[r << 1]); 1135 1136 dst->buf += 8 * dst->stride; 1137 mask_16x16 >>= 4; 1138 mask_8x8 >>= 4; 1139 mask_4x4 >>= 4; 1140 } 1141 } 1142 } 1143 1144 void vp9_loop_filter_rows(const YV12_BUFFER_CONFIG *frame_buffer, 1145 VP9_COMMON *cm, MACROBLOCKD *xd, 1146 int start, int stop, int y_only) { 1147 const int num_planes = y_only ? 1 : MAX_MB_PLANE; 1148 int mi_row, mi_col; 1149 LOOP_FILTER_MASK lfm; 1150 #if CONFIG_NON420 1151 int use_420 = y_only || (xd->plane[1].subsampling_y == 1 && 1152 xd->plane[1].subsampling_x == 1); 1153 #endif 1154 1155 for (mi_row = start; mi_row < stop; mi_row += MI_BLOCK_SIZE) { 1156 MODE_INFO **mi_8x8 = cm->mi_grid_visible + mi_row * cm->mode_info_stride; 1157 1158 for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_BLOCK_SIZE) { 1159 int plane; 1160 1161 setup_dst_planes(xd, frame_buffer, mi_row, mi_col); 1162 1163 // TODO(JBB): Make setup_mask work for non 420. 1164 #if CONFIG_NON420 1165 if (use_420) 1166 #endif 1167 setup_mask(cm, mi_row, mi_col, mi_8x8 + mi_col, cm->mode_info_stride, 1168 &lfm); 1169 1170 for (plane = 0; plane < num_planes; ++plane) { 1171 #if CONFIG_NON420 1172 if (use_420) 1173 #endif 1174 filter_block_plane(cm, &xd->plane[plane], mi_row, &lfm); 1175 #if CONFIG_NON420 1176 else 1177 filter_block_plane_non420(cm, &xd->plane[plane], mi_8x8 + mi_col, 1178 mi_row, mi_col); 1179 #endif 1180 } 1181 } 1182 } 1183 } 1184 1185 void vp9_loop_filter_frame(VP9_COMMON *cm, MACROBLOCKD *xd, 1186 int frame_filter_level, 1187 int y_only, int partial) { 1188 int start_mi_row, end_mi_row, mi_rows_to_filter; 1189 if (!frame_filter_level) return; 1190 start_mi_row = 0; 1191 mi_rows_to_filter = cm->mi_rows; 1192 if (partial && cm->mi_rows > 8) { 1193 start_mi_row = cm->mi_rows >> 1; 1194 start_mi_row &= 0xfffffff8; 1195 mi_rows_to_filter = MAX(cm->mi_rows / 8, 8); 1196 } 1197 end_mi_row = start_mi_row + mi_rows_to_filter; 1198 vp9_loop_filter_frame_init(cm, frame_filter_level); 1199 vp9_loop_filter_rows(cm->frame_to_show, cm, xd, 1200 start_mi_row, end_mi_row, 1201 y_only); 1202 } 1203 1204 int vp9_loop_filter_worker(void *arg1, void *arg2) { 1205 LFWorkerData *const lf_data = (LFWorkerData*)arg1; 1206 (void)arg2; 1207 vp9_loop_filter_rows(lf_data->frame_buffer, lf_data->cm, &lf_data->xd, 1208 lf_data->start, lf_data->stop, lf_data->y_only); 1209 return 1; 1210 } 1211
