1/* 2 * AltiVec optimizations (C) 2004 Romain Dolbeau <romain@dolbeau.org> 3 * 4 * based on code by Copyright (C) 2001-2003 Michael Niedermayer (michaelni@gmx.at) 5 * 6 * This file is part of FFmpeg. 7 * 8 * FFmpeg is free software; you can redistribute it and/or modify 9 * it under the terms of the GNU General Public License as published by 10 * the Free Software Foundation; either version 2 of the License, or 11 * (at your option) any later version. 12 * 13 * FFmpeg is distributed in the hope that it will be useful, 14 * but WITHOUT ANY WARRANTY; without even the implied warranty of 15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 16 * GNU General Public License for more details. 17 * 18 * You should have received a copy of the GNU General Public License 19 * along with FFmpeg; if not, write to the Free Software 20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA 21 */ 22 23#include "libavutil/avutil.h" 24#include "libavutil/mem_internal.h" 25 26#define ALTIVEC_TRANSPOSE_8x8_SHORT(src_a,src_b,src_c,src_d,src_e,src_f,src_g,src_h) \ 27 do { \ 28 __typeof__(src_a) tempA1, tempB1, tempC1, tempD1; \ 29 __typeof__(src_a) tempE1, tempF1, tempG1, tempH1; \ 30 __typeof__(src_a) tempA2, tempB2, tempC2, tempD2; \ 31 __typeof__(src_a) tempE2, tempF2, tempG2, tempH2; \ 32 tempA1 = vec_mergeh (src_a, src_e); \ 33 tempB1 = vec_mergel (src_a, src_e); \ 34 tempC1 = vec_mergeh (src_b, src_f); \ 35 tempD1 = vec_mergel (src_b, src_f); \ 36 tempE1 = vec_mergeh (src_c, src_g); \ 37 tempF1 = vec_mergel (src_c, src_g); \ 38 tempG1 = vec_mergeh (src_d, src_h); \ 39 tempH1 = vec_mergel (src_d, src_h); \ 40 tempA2 = vec_mergeh (tempA1, tempE1); \ 41 tempB2 = vec_mergel (tempA1, tempE1); \ 42 tempC2 = vec_mergeh (tempB1, tempF1); \ 43 tempD2 = vec_mergel (tempB1, tempF1); \ 44 tempE2 = vec_mergeh (tempC1, tempG1); \ 45 tempF2 = vec_mergel (tempC1, tempG1); \ 46 tempG2 = vec_mergeh (tempD1, tempH1); \ 47 tempH2 = vec_mergel (tempD1, tempH1); \ 48 src_a = vec_mergeh (tempA2, tempE2); \ 49 src_b = vec_mergel (tempA2, tempE2); \ 50 src_c = vec_mergeh (tempB2, tempF2); \ 51 src_d = vec_mergel (tempB2, tempF2); \ 52 src_e = vec_mergeh (tempC2, tempG2); \ 53 src_f = vec_mergel (tempC2, tempG2); \ 54 src_g = vec_mergeh (tempD2, tempH2); \ 55 src_h = vec_mergel (tempD2, tempH2); \ 56 } while (0) 57 58 59static inline int vertClassify_altivec(uint8_t src[], int stride, PPContext *c) { 60 /* 61 this code makes no assumption on src or stride. 62 One could remove the recomputation of the perm 63 vector by assuming (stride % 16) == 0, unfortunately 64 this is not always true. 65 */ 66 short data_0 = ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1; 67 DECLARE_ALIGNED(16, short, data)[8] = 68 { 69 data_0, 70 data_0 * 2 + 1, 71 c->QP * 2, 72 c->QP * 4 73 }; 74 int numEq; 75 uint8_t *src2 = src; 76 vector signed short v_dcOffset; 77 vector signed short v2QP; 78 vector unsigned short v4QP; 79 vector unsigned short v_dcThreshold; 80 const int properStride = (stride % 16); 81 const int srcAlign = ((unsigned long)src2 % 16); 82 const int two_vectors = ((srcAlign > 8) || properStride) ? 1 : 0; 83 const vector signed int zero = vec_splat_s32(0); 84 const vector signed short mask = vec_splat_s16(1); 85 vector signed int v_numEq = vec_splat_s32(0); 86 vector signed short v_data = vec_ld(0, data); 87 vector signed short v_srcAss0, v_srcAss1, v_srcAss2, v_srcAss3, 88 v_srcAss4, v_srcAss5, v_srcAss6, v_srcAss7; 89//FIXME avoid this mess if possible 90 register int j0 = 0, 91 j1 = stride, 92 j2 = 2 * stride, 93 j3 = 3 * stride, 94 j4 = 4 * stride, 95 j5 = 5 * stride, 96 j6 = 6 * stride, 97 j7 = 7 * stride; 98 vector unsigned char v_srcA0, v_srcA1, v_srcA2, v_srcA3, 99 v_srcA4, v_srcA5, v_srcA6, v_srcA7; 100 101 v_dcOffset = vec_splat(v_data, 0); 102 v_dcThreshold = (vector unsigned short)vec_splat(v_data, 1); 103 v2QP = vec_splat(v_data, 2); 104 v4QP = (vector unsigned short)vec_splat(v_data, 3); 105 106 src2 += stride * 4; 107 108#define LOAD_LINE(i) \ 109 { \ 110 vector unsigned char perm##i = vec_lvsl(j##i, src2); \ 111 vector unsigned char v_srcA2##i; \ 112 vector unsigned char v_srcA1##i = vec_ld(j##i, src2); \ 113 if (two_vectors) \ 114 v_srcA2##i = vec_ld(j##i + 16, src2); \ 115 v_srcA##i = \ 116 vec_perm(v_srcA1##i, v_srcA2##i, perm##i); \ 117 v_srcAss##i = \ 118 (vector signed short)vec_mergeh((vector signed char)zero, \ 119 (vector signed char)v_srcA##i); } 120 121#define LOAD_LINE_ALIGNED(i) \ 122 v_srcA##i = vec_ld(j##i, src2); \ 123 v_srcAss##i = \ 124 (vector signed short)vec_mergeh((vector signed char)zero, \ 125 (vector signed char)v_srcA##i) 126 127 /* Special-casing the aligned case is worthwhile, as all calls from 128 * the (transposed) horizontable deblocks will be aligned, in addition 129 * to the naturally aligned vertical deblocks. */ 130 if (properStride && srcAlign) { 131 LOAD_LINE_ALIGNED(0); 132 LOAD_LINE_ALIGNED(1); 133 LOAD_LINE_ALIGNED(2); 134 LOAD_LINE_ALIGNED(3); 135 LOAD_LINE_ALIGNED(4); 136 LOAD_LINE_ALIGNED(5); 137 LOAD_LINE_ALIGNED(6); 138 LOAD_LINE_ALIGNED(7); 139 } else { 140 LOAD_LINE(0); 141 LOAD_LINE(1); 142 LOAD_LINE(2); 143 LOAD_LINE(3); 144 LOAD_LINE(4); 145 LOAD_LINE(5); 146 LOAD_LINE(6); 147 LOAD_LINE(7); 148 } 149#undef LOAD_LINE 150#undef LOAD_LINE_ALIGNED 151 152#define ITER(i, j) \ 153 const vector signed short v_diff##i = \ 154 vec_sub(v_srcAss##i, v_srcAss##j); \ 155 const vector signed short v_sum##i = \ 156 vec_add(v_diff##i, v_dcOffset); \ 157 const vector signed short v_comp##i = \ 158 (vector signed short)vec_cmplt((vector unsigned short)v_sum##i, \ 159 v_dcThreshold); \ 160 const vector signed short v_part##i = vec_and(mask, v_comp##i); 161 162 { 163 ITER(0, 1) 164 ITER(1, 2) 165 ITER(2, 3) 166 ITER(3, 4) 167 ITER(4, 5) 168 ITER(5, 6) 169 ITER(6, 7) 170 171 v_numEq = vec_sum4s(v_part0, v_numEq); 172 v_numEq = vec_sum4s(v_part1, v_numEq); 173 v_numEq = vec_sum4s(v_part2, v_numEq); 174 v_numEq = vec_sum4s(v_part3, v_numEq); 175 v_numEq = vec_sum4s(v_part4, v_numEq); 176 v_numEq = vec_sum4s(v_part5, v_numEq); 177 v_numEq = vec_sum4s(v_part6, v_numEq); 178 } 179 180#undef ITER 181 182 v_numEq = vec_sums(v_numEq, zero); 183 184 v_numEq = vec_splat(v_numEq, 3); 185 vec_ste(v_numEq, 0, &numEq); 186 187 if (numEq > c->ppMode.flatnessThreshold){ 188 const vector unsigned char mmoP1 = (const vector unsigned char) 189 {0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 190 0x00, 0x01, 0x12, 0x13, 0x08, 0x09, 0x1A, 0x1B}; 191 const vector unsigned char mmoP2 = (const vector unsigned char) 192 {0x04, 0x05, 0x16, 0x17, 0x0C, 0x0D, 0x1E, 0x1F, 193 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f}; 194 const vector unsigned char mmoP = (const vector unsigned char) 195 vec_lvsl(8, (unsigned char*)0); 196 197 vector signed short mmoL1 = vec_perm(v_srcAss0, v_srcAss2, mmoP1); 198 vector signed short mmoL2 = vec_perm(v_srcAss4, v_srcAss6, mmoP2); 199 vector signed short mmoL = vec_perm(mmoL1, mmoL2, mmoP); 200 vector signed short mmoR1 = vec_perm(v_srcAss5, v_srcAss7, mmoP1); 201 vector signed short mmoR2 = vec_perm(v_srcAss1, v_srcAss3, mmoP2); 202 vector signed short mmoR = vec_perm(mmoR1, mmoR2, mmoP); 203 vector signed short mmoDiff = vec_sub(mmoL, mmoR); 204 vector unsigned short mmoSum = (vector unsigned short)vec_add(mmoDiff, v2QP); 205 206 if (vec_any_gt(mmoSum, v4QP)) 207 return 0; 208 else 209 return 1; 210 } 211 else return 2; 212} 213 214static inline void doVertLowPass_altivec(uint8_t *src, int stride, PPContext *c) { 215 /* 216 this code makes no assumption on src or stride. 217 One could remove the recomputation of the perm 218 vector by assuming (stride % 16) == 0, unfortunately 219 this is not always true. Quite a lot of load/stores 220 can be removed by assuming proper alignment of 221 src & stride :-( 222 */ 223 uint8_t *src2 = src; 224 const vector signed int zero = vec_splat_s32(0); 225 const int properStride = (stride % 16); 226 const int srcAlign = ((unsigned long)src2 % 16); 227 DECLARE_ALIGNED(16, short, qp)[8] = {c->QP}; 228 vector signed short vqp = vec_ld(0, qp); 229 vector signed short vb0, vb1, vb2, vb3, vb4, vb5, vb6, vb7, vb8, vb9; 230 vector unsigned char vbA0, av_uninit(vbA1), av_uninit(vbA2), av_uninit(vbA3), av_uninit(vbA4), av_uninit(vbA5), av_uninit(vbA6), av_uninit(vbA7), av_uninit(vbA8), vbA9; 231 vector unsigned char vbB0, av_uninit(vbB1), av_uninit(vbB2), av_uninit(vbB3), av_uninit(vbB4), av_uninit(vbB5), av_uninit(vbB6), av_uninit(vbB7), av_uninit(vbB8), vbB9; 232 vector unsigned char vbT0, vbT1, vbT2, vbT3, vbT4, vbT5, vbT6, vbT7, vbT8, vbT9; 233 vector unsigned char perml0, perml1, perml2, perml3, perml4, 234 perml5, perml6, perml7, perml8, perml9; 235 register int j0 = 0, 236 j1 = stride, 237 j2 = 2 * stride, 238 j3 = 3 * stride, 239 j4 = 4 * stride, 240 j5 = 5 * stride, 241 j6 = 6 * stride, 242 j7 = 7 * stride, 243 j8 = 8 * stride, 244 j9 = 9 * stride; 245 246 vqp = vec_splat(vqp, 0); 247 248 src2 += stride*3; 249 250#define LOAD_LINE(i) \ 251 perml##i = vec_lvsl(i * stride, src2); \ 252 vbA##i = vec_ld(i * stride, src2); \ 253 vbB##i = vec_ld(i * stride + 16, src2); \ 254 vbT##i = vec_perm(vbA##i, vbB##i, perml##i); \ 255 vb##i = \ 256 (vector signed short)vec_mergeh((vector unsigned char)zero, \ 257 (vector unsigned char)vbT##i) 258 259#define LOAD_LINE_ALIGNED(i) \ 260 vbT##i = vec_ld(j##i, src2); \ 261 vb##i = \ 262 (vector signed short)vec_mergeh((vector signed char)zero, \ 263 (vector signed char)vbT##i) 264 265 /* Special-casing the aligned case is worthwhile, as all calls from 266 * the (transposed) horizontable deblocks will be aligned, in addition 267 * to the naturally aligned vertical deblocks. */ 268 if (properStride && srcAlign) { 269 LOAD_LINE_ALIGNED(0); 270 LOAD_LINE_ALIGNED(1); 271 LOAD_LINE_ALIGNED(2); 272 LOAD_LINE_ALIGNED(3); 273 LOAD_LINE_ALIGNED(4); 274 LOAD_LINE_ALIGNED(5); 275 LOAD_LINE_ALIGNED(6); 276 LOAD_LINE_ALIGNED(7); 277 LOAD_LINE_ALIGNED(8); 278 LOAD_LINE_ALIGNED(9); 279 } else { 280 LOAD_LINE(0); 281 LOAD_LINE(1); 282 LOAD_LINE(2); 283 LOAD_LINE(3); 284 LOAD_LINE(4); 285 LOAD_LINE(5); 286 LOAD_LINE(6); 287 LOAD_LINE(7); 288 LOAD_LINE(8); 289 LOAD_LINE(9); 290 } 291#undef LOAD_LINE 292#undef LOAD_LINE_ALIGNED 293 { 294 const vector unsigned short v_2 = vec_splat_u16(2); 295 const vector unsigned short v_4 = vec_splat_u16(4); 296 297 const vector signed short v_diff01 = vec_sub(vb0, vb1); 298 const vector unsigned short v_cmp01 = 299 (const vector unsigned short) vec_cmplt(vec_abs(v_diff01), vqp); 300 const vector signed short v_first = vec_sel(vb1, vb0, v_cmp01); 301 const vector signed short v_diff89 = vec_sub(vb8, vb9); 302 const vector unsigned short v_cmp89 = 303 (const vector unsigned short) vec_cmplt(vec_abs(v_diff89), vqp); 304 const vector signed short v_last = vec_sel(vb8, vb9, v_cmp89); 305 306 const vector signed short temp01 = vec_mladd(v_first, (vector signed short)v_4, vb1); 307 const vector signed short temp02 = vec_add(vb2, vb3); 308 const vector signed short temp03 = vec_add(temp01, (vector signed short)v_4); 309 const vector signed short v_sumsB0 = vec_add(temp02, temp03); 310 311 const vector signed short temp11 = vec_sub(v_sumsB0, v_first); 312 const vector signed short v_sumsB1 = vec_add(temp11, vb4); 313 314 const vector signed short temp21 = vec_sub(v_sumsB1, v_first); 315 const vector signed short v_sumsB2 = vec_add(temp21, vb5); 316 317 const vector signed short temp31 = vec_sub(v_sumsB2, v_first); 318 const vector signed short v_sumsB3 = vec_add(temp31, vb6); 319 320 const vector signed short temp41 = vec_sub(v_sumsB3, v_first); 321 const vector signed short v_sumsB4 = vec_add(temp41, vb7); 322 323 const vector signed short temp51 = vec_sub(v_sumsB4, vb1); 324 const vector signed short v_sumsB5 = vec_add(temp51, vb8); 325 326 const vector signed short temp61 = vec_sub(v_sumsB5, vb2); 327 const vector signed short v_sumsB6 = vec_add(temp61, v_last); 328 329 const vector signed short temp71 = vec_sub(v_sumsB6, vb3); 330 const vector signed short v_sumsB7 = vec_add(temp71, v_last); 331 332 const vector signed short temp81 = vec_sub(v_sumsB7, vb4); 333 const vector signed short v_sumsB8 = vec_add(temp81, v_last); 334 335 const vector signed short temp91 = vec_sub(v_sumsB8, vb5); 336 const vector signed short v_sumsB9 = vec_add(temp91, v_last); 337 338 #define COMPUTE_VR(i, j, k) \ 339 const vector signed short temps1##i = \ 340 vec_add(v_sumsB##i, v_sumsB##k); \ 341 const vector signed short temps2##i = \ 342 vec_mladd(vb##j, (vector signed short)v_2, temps1##i); \ 343 const vector signed short vr##j = vec_sra(temps2##i, v_4) 344 345 COMPUTE_VR(0, 1, 2); 346 COMPUTE_VR(1, 2, 3); 347 COMPUTE_VR(2, 3, 4); 348 COMPUTE_VR(3, 4, 5); 349 COMPUTE_VR(4, 5, 6); 350 COMPUTE_VR(5, 6, 7); 351 COMPUTE_VR(6, 7, 8); 352 COMPUTE_VR(7, 8, 9); 353 354 const vector signed char neg1 = vec_splat_s8(-1); 355 const vector unsigned char permHH = (const vector unsigned char){0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 356 0x18, 0x19, 0x1A, 0x1B, 0x1C, 0x1D, 0x1E, 0x1F}; 357 358#define PACK_AND_STORE(i) \ 359{ const vector unsigned char perms##i = \ 360 vec_lvsr(i * stride, src2); \ 361 const vector unsigned char vf##i = \ 362 vec_packsu(vr##i, (vector signed short)zero); \ 363 const vector unsigned char vg##i = \ 364 vec_perm(vf##i, vbT##i, permHH); \ 365 const vector unsigned char mask##i = \ 366 vec_perm((vector unsigned char)zero, (vector unsigned char)neg1, perms##i); \ 367 const vector unsigned char vg2##i = \ 368 vec_perm(vg##i, vg##i, perms##i); \ 369 const vector unsigned char svA##i = \ 370 vec_sel(vbA##i, vg2##i, mask##i); \ 371 const vector unsigned char svB##i = \ 372 vec_sel(vg2##i, vbB##i, mask##i); \ 373 vec_st(svA##i, i * stride, src2); \ 374 vec_st(svB##i, i * stride + 16, src2);} 375 376#define PACK_AND_STORE_ALIGNED(i) \ 377{ const vector unsigned char vf##i = \ 378 vec_packsu(vr##i, (vector signed short)zero); \ 379 const vector unsigned char vg##i = \ 380 vec_perm(vf##i, vbT##i, permHH); \ 381 vec_st(vg##i, i * stride, src2);} 382 383 /* Special-casing the aligned case is worthwhile, as all calls from 384 * the (transposed) horizontable deblocks will be aligned, in addition 385 * to the naturally aligned vertical deblocks. */ 386 if (properStride && srcAlign) { 387 PACK_AND_STORE_ALIGNED(1) 388 PACK_AND_STORE_ALIGNED(2) 389 PACK_AND_STORE_ALIGNED(3) 390 PACK_AND_STORE_ALIGNED(4) 391 PACK_AND_STORE_ALIGNED(5) 392 PACK_AND_STORE_ALIGNED(6) 393 PACK_AND_STORE_ALIGNED(7) 394 PACK_AND_STORE_ALIGNED(8) 395 } else { 396 PACK_AND_STORE(1) 397 PACK_AND_STORE(2) 398 PACK_AND_STORE(3) 399 PACK_AND_STORE(4) 400 PACK_AND_STORE(5) 401 PACK_AND_STORE(6) 402 PACK_AND_STORE(7) 403 PACK_AND_STORE(8) 404 } 405 #undef PACK_AND_STORE 406 #undef PACK_AND_STORE_ALIGNED 407 } 408} 409 410 411 412static inline void doVertDefFilter_altivec(uint8_t src[], int stride, PPContext *c) { 413 /* 414 this code makes no assumption on src or stride. 415 One could remove the recomputation of the perm 416 vector by assuming (stride % 16) == 0, unfortunately 417 this is not always true. Quite a lot of load/stores 418 can be removed by assuming proper alignment of 419 src & stride :-( 420 */ 421 uint8_t *src2 = src + stride*3; 422 const vector signed int zero = vec_splat_s32(0); 423 DECLARE_ALIGNED(16, short, qp)[8] = {8*c->QP}; 424 vector signed short vqp = vec_splat( 425 (vector signed short)vec_ld(0, qp), 0); 426 427#define LOAD_LINE(i) \ 428 const vector unsigned char perm##i = \ 429 vec_lvsl(i * stride, src2); \ 430 const vector unsigned char vbA##i = \ 431 vec_ld(i * stride, src2); \ 432 const vector unsigned char vbB##i = \ 433 vec_ld(i * stride + 16, src2); \ 434 const vector unsigned char vbT##i = \ 435 vec_perm(vbA##i, vbB##i, perm##i); \ 436 const vector signed short vb##i = \ 437 (vector signed short)vec_mergeh((vector unsigned char)zero, \ 438 (vector unsigned char)vbT##i) 439 440 LOAD_LINE(1); 441 LOAD_LINE(2); 442 LOAD_LINE(3); 443 LOAD_LINE(4); 444 LOAD_LINE(5); 445 LOAD_LINE(6); 446 LOAD_LINE(7); 447 LOAD_LINE(8); 448#undef LOAD_LINE 449 450 const vector signed short v_1 = vec_splat_s16(1); 451 const vector signed short v_2 = vec_splat_s16(2); 452 const vector signed short v_5 = vec_splat_s16(5); 453 const vector signed short v_32 = vec_sl(v_1, 454 (vector unsigned short)v_5); 455 /* middle energy */ 456 const vector signed short l3minusl6 = vec_sub(vb3, vb6); 457 const vector signed short l5minusl4 = vec_sub(vb5, vb4); 458 const vector signed short twotimes_l3minusl6 = vec_mladd(v_2, l3minusl6, (vector signed short)zero); 459 const vector signed short mE = vec_mladd(v_5, l5minusl4, twotimes_l3minusl6); 460 const vector signed short absmE = vec_abs(mE); 461 /* left & right energy */ 462 const vector signed short l1minusl4 = vec_sub(vb1, vb4); 463 const vector signed short l3minusl2 = vec_sub(vb3, vb2); 464 const vector signed short l5minusl8 = vec_sub(vb5, vb8); 465 const vector signed short l7minusl6 = vec_sub(vb7, vb6); 466 const vector signed short twotimes_l1minusl4 = vec_mladd(v_2, l1minusl4, (vector signed short)zero); 467 const vector signed short twotimes_l5minusl8 = vec_mladd(v_2, l5minusl8, (vector signed short)zero); 468 const vector signed short lE = vec_mladd(v_5, l3minusl2, twotimes_l1minusl4); 469 const vector signed short rE = vec_mladd(v_5, l7minusl6, twotimes_l5minusl8); 470 /* d */ 471 const vector signed short ddiff = vec_sub(absmE, 472 vec_min(vec_abs(lE), 473 vec_abs(rE))); 474 const vector signed short ddiffclamp = vec_max(ddiff, (vector signed short)zero); 475 const vector signed short dtimes64 = vec_mladd(v_5, ddiffclamp, v_32); 476 const vector signed short d = vec_sra(dtimes64, vec_splat_u16(6)); 477 const vector signed short minusd = vec_sub((vector signed short)zero, d); 478 const vector signed short finald = vec_sel(minusd, 479 d, 480 vec_cmpgt(vec_sub((vector signed short)zero, mE), 481 (vector signed short)zero)); 482 /* q */ 483 const vector signed short qtimes2 = vec_sub(vb4, vb5); 484 /* for a shift right to behave like /2, we need to add one 485 to all negative integer */ 486 const vector signed short rounddown = vec_sel((vector signed short)zero, 487 v_1, 488 vec_cmplt(qtimes2, (vector signed short)zero)); 489 const vector signed short q = vec_sra(vec_add(qtimes2, rounddown), vec_splat_u16(1)); 490 /* clamp */ 491 const vector signed short dclamp_P1 = vec_max((vector signed short)zero, finald); 492 const vector signed short dclamp_P = vec_min(dclamp_P1, q); 493 const vector signed short dclamp_N1 = vec_min((vector signed short)zero, finald); 494 const vector signed short dclamp_N = vec_max(dclamp_N1, q); 495 496 const vector signed short dclampedfinal = vec_sel(dclamp_N, 497 dclamp_P, 498 vec_cmpgt(q, (vector signed short)zero)); 499 const vector signed short dornotd = vec_sel((vector signed short)zero, 500 dclampedfinal, 501 vec_cmplt(absmE, vqp)); 502 /* add/subtract to l4 and l5 */ 503 const vector signed short vb4minusd = vec_sub(vb4, dornotd); 504 const vector signed short vb5plusd = vec_add(vb5, dornotd); 505 /* finally, stores */ 506 const vector unsigned char st4 = vec_packsu(vb4minusd, (vector signed short)zero); 507 const vector unsigned char st5 = vec_packsu(vb5plusd, (vector signed short)zero); 508 509 const vector signed char neg1 = vec_splat_s8(-1); 510 const vector unsigned char permHH = (const vector unsigned char){0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 511 0x18, 0x19, 0x1A, 0x1B, 0x1C, 0x1D, 0x1E, 0x1F}; 512 513#define STORE(i) \ 514{ const vector unsigned char perms##i = \ 515 vec_lvsr(i * stride, src2); \ 516 const vector unsigned char vg##i = \ 517 vec_perm(st##i, vbT##i, permHH); \ 518 const vector unsigned char mask##i = \ 519 vec_perm((vector unsigned char)zero, (vector unsigned char)neg1, perms##i); \ 520 const vector unsigned char vg2##i = \ 521 vec_perm(vg##i, vg##i, perms##i); \ 522 const vector unsigned char svA##i = \ 523 vec_sel(vbA##i, vg2##i, mask##i); \ 524 const vector unsigned char svB##i = \ 525 vec_sel(vg2##i, vbB##i, mask##i); \ 526 vec_st(svA##i, i * stride, src2); \ 527 vec_st(svB##i, i * stride + 16, src2);} 528 529 STORE(4) 530 STORE(5) 531} 532 533static inline void dering_altivec(uint8_t src[], int stride, PPContext *c) { 534 const vector signed int vsint32_8 = vec_splat_s32(8); 535 const vector unsigned int vuint32_4 = vec_splat_u32(4); 536 const vector signed char neg1 = vec_splat_s8(-1); 537 538 const vector unsigned char permA1 = (vector unsigned char) 539 {0x00, 0x01, 0x02, 0x10, 0x11, 0x12, 0x1F, 0x1F, 540 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F}; 541 const vector unsigned char permA2 = (vector unsigned char) 542 {0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x10, 0x11, 543 0x12, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F}; 544 const vector unsigned char permA1inc = (vector unsigned char) 545 {0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x00, 0x00, 546 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00}; 547 const vector unsigned char permA2inc = (vector unsigned char) 548 {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 549 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00}; 550 const vector unsigned char magic = (vector unsigned char) 551 {0x01, 0x02, 0x01, 0x02, 0x04, 0x02, 0x01, 0x02, 552 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00}; 553 const vector unsigned char extractPerm = (vector unsigned char) 554 {0x10, 0x10, 0x10, 0x01, 0x10, 0x10, 0x10, 0x01, 555 0x10, 0x10, 0x10, 0x01, 0x10, 0x10, 0x10, 0x01}; 556 const vector unsigned char extractPermInc = (vector unsigned char) 557 {0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x01, 558 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x01}; 559 const vector unsigned char identity = vec_lvsl(0,(unsigned char *)0); 560 const vector unsigned char tenRight = (vector unsigned char) 561 {0x00, 0x10, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 562 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00}; 563 const vector unsigned char eightLeft = (vector unsigned char) 564 {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 565 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x08}; 566 567 /* 568 this code makes no assumption on src or stride. 569 One could remove the recomputation of the perm 570 vector by assuming (stride % 16) == 0, unfortunately 571 this is not always true. Quite a lot of load/stores 572 can be removed by assuming proper alignment of 573 src & stride :-( 574 */ 575 uint8_t *srcCopy = src; 576 DECLARE_ALIGNED(16, uint8_t, dt)[16] = { deringThreshold }; 577 const vector signed int zero = vec_splat_s32(0); 578 vector unsigned char v_dt = vec_splat(vec_ld(0, dt), 0); 579 580#define LOAD_LINE(i) \ 581 const vector unsigned char perm##i = \ 582 vec_lvsl(i * stride, srcCopy); \ 583 vector unsigned char sA##i = vec_ld(i * stride, srcCopy); \ 584 vector unsigned char sB##i = vec_ld(i * stride + 16, srcCopy); \ 585 vector unsigned char src##i = vec_perm(sA##i, sB##i, perm##i) 586 587 LOAD_LINE(0); 588 LOAD_LINE(1); 589 LOAD_LINE(2); 590 LOAD_LINE(3); 591 LOAD_LINE(4); 592 LOAD_LINE(5); 593 LOAD_LINE(6); 594 LOAD_LINE(7); 595 LOAD_LINE(8); 596 LOAD_LINE(9); 597#undef LOAD_LINE 598 599 vector unsigned char v_avg; 600 DECLARE_ALIGNED(16, signed int, S)[8]; 601 DECLARE_ALIGNED(16, int, tQP2)[4] = { c->QP/2 + 1 }; 602 vector signed int vQP2 = vec_ld(0, tQP2); 603 vQP2 = vec_splat(vQP2, 0); 604 605 { 606 const vector unsigned char trunc_perm = (vector unsigned char) 607 {0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0x08, 608 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17, 0x18}; 609 const vector unsigned char trunc_src12 = vec_perm(src1, src2, trunc_perm); 610 const vector unsigned char trunc_src34 = vec_perm(src3, src4, trunc_perm); 611 const vector unsigned char trunc_src56 = vec_perm(src5, src6, trunc_perm); 612 const vector unsigned char trunc_src78 = vec_perm(src7, src8, trunc_perm); 613 614#define EXTRACT(op) do { \ 615 const vector unsigned char s_1 = vec_##op(trunc_src12, trunc_src34); \ 616 const vector unsigned char s_2 = vec_##op(trunc_src56, trunc_src78); \ 617 const vector unsigned char s_6 = vec_##op(s_1, s_2); \ 618 const vector unsigned char s_8h = vec_mergeh(s_6, s_6); \ 619 const vector unsigned char s_8l = vec_mergel(s_6, s_6); \ 620 const vector unsigned char s_9 = vec_##op(s_8h, s_8l); \ 621 const vector unsigned char s_9h = vec_mergeh(s_9, s_9); \ 622 const vector unsigned char s_9l = vec_mergel(s_9, s_9); \ 623 const vector unsigned char s_10 = vec_##op(s_9h, s_9l); \ 624 const vector unsigned char s_10h = vec_mergeh(s_10, s_10); \ 625 const vector unsigned char s_10l = vec_mergel(s_10, s_10); \ 626 const vector unsigned char s_11 = vec_##op(s_10h, s_10l); \ 627 const vector unsigned char s_11h = vec_mergeh(s_11, s_11); \ 628 const vector unsigned char s_11l = vec_mergel(s_11, s_11); \ 629 v_##op = vec_##op(s_11h, s_11l); \ 630} while (0) 631 632 vector unsigned char v_min; 633 vector unsigned char v_max; 634 EXTRACT(min); 635 EXTRACT(max); 636#undef EXTRACT 637 638 if (vec_all_lt(vec_sub(v_max, v_min), v_dt)) 639 return; 640 641 v_avg = vec_avg(v_min, v_max); 642 } 643 644 { 645 const vector unsigned short mask1 = (vector unsigned short) 646 {0x0001, 0x0002, 0x0004, 0x0008, 647 0x0010, 0x0020, 0x0040, 0x0080}; 648 const vector unsigned short mask2 = (vector unsigned short) 649 {0x0100, 0x0200, 0x0000, 0x0000, 650 0x0000, 0x0000, 0x0000, 0x0000}; 651 652 const vector unsigned int vuint32_16 = vec_sl(vec_splat_u32(1), vec_splat_u32(4)); 653 const vector unsigned int vuint32_1 = vec_splat_u32(1); 654 655 vector signed int sumA2; 656 vector signed int sumB2; 657 vector signed int sum0, sum1, sum2, sum3, sum4; 658 vector signed int sum5, sum6, sum7, sum8, sum9; 659 660#define COMPARE(i) \ 661 do { \ 662 const vector unsigned char cmp = \ 663 (vector unsigned char)vec_cmpgt(src##i, v_avg); \ 664 const vector unsigned short cmpHi = \ 665 (vector unsigned short)vec_mergeh(cmp, cmp); \ 666 const vector unsigned short cmpLi = \ 667 (vector unsigned short)vec_mergel(cmp, cmp); \ 668 const vector signed short cmpHf = \ 669 (vector signed short)vec_and(cmpHi, mask1); \ 670 const vector signed short cmpLf = \ 671 (vector signed short)vec_and(cmpLi, mask2); \ 672 const vector signed int sump = vec_sum4s(cmpHf, zero); \ 673 const vector signed int sumq = vec_sum4s(cmpLf, sump); \ 674 sum##i = vec_sums(sumq, zero); \ 675 } while (0) 676 677 COMPARE(0); 678 COMPARE(1); 679 COMPARE(2); 680 COMPARE(3); 681 COMPARE(4); 682 COMPARE(5); 683 COMPARE(6); 684 COMPARE(7); 685 COMPARE(8); 686 COMPARE(9); 687#undef COMPARE 688 689 { 690 const vector signed int sump02 = vec_mergel(sum0, sum2); 691 const vector signed int sump13 = vec_mergel(sum1, sum3); 692 const vector signed int sumA = vec_mergel(sump02, sump13); 693 694 const vector signed int sump46 = vec_mergel(sum4, sum6); 695 const vector signed int sump57 = vec_mergel(sum5, sum7); 696 const vector signed int sumB = vec_mergel(sump46, sump57); 697 698 const vector signed int sump8A = vec_mergel(sum8, zero); 699 const vector signed int sump9B = vec_mergel(sum9, zero); 700 const vector signed int sumC = vec_mergel(sump8A, sump9B); 701 702 const vector signed int tA = vec_sl(vec_nor(zero, sumA), vuint32_16); 703 const vector signed int tB = vec_sl(vec_nor(zero, sumB), vuint32_16); 704 const vector signed int tC = vec_sl(vec_nor(zero, sumC), vuint32_16); 705 const vector signed int t2A = vec_or(sumA, tA); 706 const vector signed int t2B = vec_or(sumB, tB); 707 const vector signed int t2C = vec_or(sumC, tC); 708 const vector signed int t3A = vec_and(vec_sra(t2A, vuint32_1), 709 vec_sl(t2A, vuint32_1)); 710 const vector signed int t3B = vec_and(vec_sra(t2B, vuint32_1), 711 vec_sl(t2B, vuint32_1)); 712 const vector signed int t3C = vec_and(vec_sra(t2C, vuint32_1), 713 vec_sl(t2C, vuint32_1)); 714 const vector signed int yA = vec_and(t2A, t3A); 715 const vector signed int yB = vec_and(t2B, t3B); 716 const vector signed int yC = vec_and(t2C, t3C); 717 718 const vector unsigned char strangeperm1 = vec_lvsl(4, (unsigned char*)0); 719 const vector unsigned char strangeperm2 = vec_lvsl(8, (unsigned char*)0); 720 const vector signed int sumAd4 = vec_perm(yA, yB, strangeperm1); 721 const vector signed int sumAd8 = vec_perm(yA, yB, strangeperm2); 722 const vector signed int sumBd4 = vec_perm(yB, yC, strangeperm1); 723 const vector signed int sumBd8 = vec_perm(yB, yC, strangeperm2); 724 const vector signed int sumAp = vec_and(yA, 725 vec_and(sumAd4,sumAd8)); 726 const vector signed int sumBp = vec_and(yB, 727 vec_and(sumBd4,sumBd8)); 728 sumA2 = vec_or(sumAp, 729 vec_sra(sumAp, 730 vuint32_16)); 731 sumB2 = vec_or(sumBp, 732 vec_sra(sumBp, 733 vuint32_16)); 734 } 735 vec_st(sumA2, 0, S); 736 vec_st(sumB2, 16, S); 737 } 738 739 /* I'm not sure the following is actually faster 740 than straight, unvectorized C code :-( */ 741 742#define F_INIT() \ 743 vector unsigned char tenRightM = tenRight; \ 744 vector unsigned char permA1M = permA1; \ 745 vector unsigned char permA2M = permA2; \ 746 vector unsigned char extractPermM = extractPerm 747 748#define F2(i, j, k, l) \ 749 if (S[i] & (1 << (l+1))) { \ 750 const vector unsigned char a_A = vec_perm(src##i, src##j, permA1M); \ 751 const vector unsigned char a_B = vec_perm(a_A, src##k, permA2M); \ 752 const vector signed int a_sump = \ 753 (vector signed int)vec_msum(a_B, magic, (vector unsigned int)zero);\ 754 vector signed int F = vec_sr(vec_sums(a_sump, vsint32_8), vuint32_4); \ 755 const vector signed int p = \ 756 (vector signed int)vec_perm(src##j, (vector unsigned char)zero, \ 757 extractPermM); \ 758 const vector signed int sum = vec_add(p, vQP2); \ 759 const vector signed int diff = vec_sub(p, vQP2); \ 760 vector signed int newpm; \ 761 vector unsigned char newpm2, mask; \ 762 F = vec_splat(F, 3); \ 763 if (vec_all_lt(sum, F)) \ 764 newpm = sum; \ 765 else if (vec_all_gt(diff, F)) \ 766 newpm = diff; \ 767 else newpm = F; \ 768 newpm2 = vec_splat((vector unsigned char)newpm, 15); \ 769 mask = vec_add(identity, tenRightM); \ 770 src##j = vec_perm(src##j, newpm2, mask); \ 771 } \ 772 permA1M = vec_add(permA1M, permA1inc); \ 773 permA2M = vec_add(permA2M, permA2inc); \ 774 tenRightM = vec_sro(tenRightM, eightLeft); \ 775 extractPermM = vec_add(extractPermM, extractPermInc) 776 777#define ITER(i, j, k) do { \ 778 F_INIT(); \ 779 F2(i, j, k, 0); \ 780 F2(i, j, k, 1); \ 781 F2(i, j, k, 2); \ 782 F2(i, j, k, 3); \ 783 F2(i, j, k, 4); \ 784 F2(i, j, k, 5); \ 785 F2(i, j, k, 6); \ 786 F2(i, j, k, 7); \ 787} while (0) 788 789 ITER(0, 1, 2); 790 ITER(1, 2, 3); 791 ITER(2, 3, 4); 792 ITER(3, 4, 5); 793 ITER(4, 5, 6); 794 ITER(5, 6, 7); 795 ITER(6, 7, 8); 796 ITER(7, 8, 9); 797 798#define STORE_LINE(i) do { \ 799 const vector unsigned char permST = \ 800 vec_lvsr(i * stride, srcCopy); \ 801 const vector unsigned char maskST = \ 802 vec_perm((vector unsigned char)zero, \ 803 (vector unsigned char)neg1, permST); \ 804 src##i = vec_perm(src##i ,src##i, permST); \ 805 sA##i= vec_sel(sA##i, src##i, maskST); \ 806 sB##i= vec_sel(src##i, sB##i, maskST); \ 807 vec_st(sA##i, i * stride, srcCopy); \ 808 vec_st(sB##i, i * stride + 16, srcCopy); \ 809} while (0) 810 811 STORE_LINE(1); 812 STORE_LINE(2); 813 STORE_LINE(3); 814 STORE_LINE(4); 815 STORE_LINE(5); 816 STORE_LINE(6); 817 STORE_LINE(7); 818 STORE_LINE(8); 819 820#undef STORE_LINE 821#undef ITER 822#undef F2 823} 824 825#define doHorizLowPass_altivec(a...) doHorizLowPass_C(a) 826#define doHorizDefFilter_altivec(a...) doHorizDefFilter_C(a) 827#define do_a_deblock_altivec(a...) do_a_deblock_C(a) 828 829static inline void tempNoiseReducer_altivec(uint8_t *src, int stride, 830 uint8_t *tempBlurred, uint32_t *tempBlurredPast, int *maxNoise) 831{ 832 const vector signed char neg1 = vec_splat_s8(-1); 833 const vector unsigned char permHH = (const vector unsigned char){0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 834 0x18, 0x19, 0x1A, 0x1B, 0x1C, 0x1D, 0x1E, 0x1F}; 835 836 const vector signed int zero = vec_splat_s32(0); 837 const vector signed short vsint16_1 = vec_splat_s16(1); 838 vector signed int v_dp = zero; 839 vector signed int v_sysdp = zero; 840 int d, sysd, i; 841 842#define LOAD_LINE(src, i) \ 843 register int j##src##i = i * stride; \ 844 vector unsigned char perm##src##i = vec_lvsl(j##src##i, src); \ 845 const vector unsigned char v_##src##A1##i = vec_ld(j##src##i, src); \ 846 const vector unsigned char v_##src##A2##i = vec_ld(j##src##i + 16, src); \ 847 const vector unsigned char v_##src##A##i = \ 848 vec_perm(v_##src##A1##i, v_##src##A2##i, perm##src##i); \ 849 vector signed short v_##src##Ass##i = \ 850 (vector signed short)vec_mergeh((vector signed char)zero, \ 851 (vector signed char)v_##src##A##i) 852 853 LOAD_LINE(src, 0); 854 LOAD_LINE(src, 1); 855 LOAD_LINE(src, 2); 856 LOAD_LINE(src, 3); 857 LOAD_LINE(src, 4); 858 LOAD_LINE(src, 5); 859 LOAD_LINE(src, 6); 860 LOAD_LINE(src, 7); 861 862 LOAD_LINE(tempBlurred, 0); 863 LOAD_LINE(tempBlurred, 1); 864 LOAD_LINE(tempBlurred, 2); 865 LOAD_LINE(tempBlurred, 3); 866 LOAD_LINE(tempBlurred, 4); 867 LOAD_LINE(tempBlurred, 5); 868 LOAD_LINE(tempBlurred, 6); 869 LOAD_LINE(tempBlurred, 7); 870#undef LOAD_LINE 871 872#define ACCUMULATE_DIFFS(i) do { \ 873 vector signed short v_d = vec_sub(v_tempBlurredAss##i, \ 874 v_srcAss##i); \ 875 v_dp = vec_msums(v_d, v_d, v_dp); \ 876 v_sysdp = vec_msums(v_d, vsint16_1, v_sysdp); \ 877 } while (0) 878 879 ACCUMULATE_DIFFS(0); 880 ACCUMULATE_DIFFS(1); 881 ACCUMULATE_DIFFS(2); 882 ACCUMULATE_DIFFS(3); 883 ACCUMULATE_DIFFS(4); 884 ACCUMULATE_DIFFS(5); 885 ACCUMULATE_DIFFS(6); 886 ACCUMULATE_DIFFS(7); 887#undef ACCUMULATE_DIFFS 888 889 tempBlurredPast[127]= maxNoise[0]; 890 tempBlurredPast[128]= maxNoise[1]; 891 tempBlurredPast[129]= maxNoise[2]; 892 893 v_dp = vec_sums(v_dp, zero); 894 v_sysdp = vec_sums(v_sysdp, zero); 895 896 v_dp = vec_splat(v_dp, 3); 897 v_sysdp = vec_splat(v_sysdp, 3); 898 899 vec_ste(v_dp, 0, &d); 900 vec_ste(v_sysdp, 0, &sysd); 901 902 i = d; 903 d = (4*d 904 +(*(tempBlurredPast-256)) 905 +(*(tempBlurredPast-1))+ (*(tempBlurredPast+1)) 906 +(*(tempBlurredPast+256)) 907 +4)>>3; 908 909 *tempBlurredPast=i; 910 911 if (d > maxNoise[1]) { 912 if (d < maxNoise[2]) { 913#define OP(i) v_tempBlurredAss##i = vec_avg(v_tempBlurredAss##i, v_srcAss##i); 914 915 OP(0); 916 OP(1); 917 OP(2); 918 OP(3); 919 OP(4); 920 OP(5); 921 OP(6); 922 OP(7); 923#undef OP 924 } else { 925#define OP(i) v_tempBlurredAss##i = v_srcAss##i; 926 927 OP(0); 928 OP(1); 929 OP(2); 930 OP(3); 931 OP(4); 932 OP(5); 933 OP(6); 934 OP(7); 935#undef OP 936 } 937 } else { 938 if (d < maxNoise[0]) { 939 const vector signed short vsint16_7 = vec_splat_s16(7); 940 const vector signed short vsint16_4 = vec_splat_s16(4); 941 const vector unsigned short vuint16_3 = vec_splat_u16(3); 942 943#define OP(i) do { \ 944 const vector signed short v_temp = \ 945 vec_mladd(v_tempBlurredAss##i, vsint16_7, v_srcAss##i); \ 946 const vector signed short v_temp2 = vec_add(v_temp, vsint16_4); \ 947 v_tempBlurredAss##i = vec_sr(v_temp2, vuint16_3); \ 948 } while (0) 949 950 OP(0); 951 OP(1); 952 OP(2); 953 OP(3); 954 OP(4); 955 OP(5); 956 OP(6); 957 OP(7); 958#undef OP 959 } else { 960 const vector signed short vsint16_3 = vec_splat_s16(3); 961 const vector signed short vsint16_2 = vec_splat_s16(2); 962 963#define OP(i) do { \ 964 const vector signed short v_temp = \ 965 vec_mladd(v_tempBlurredAss##i, vsint16_3, v_srcAss##i); \ 966 const vector signed short v_temp2 = vec_add(v_temp, vsint16_2); \ 967 v_tempBlurredAss##i = \ 968 vec_sr(v_temp2, (vector unsigned short)vsint16_2); \ 969 } while (0) 970 971 OP(0); 972 OP(1); 973 OP(2); 974 OP(3); 975 OP(4); 976 OP(5); 977 OP(6); 978 OP(7); 979#undef OP 980 } 981 } 982 983#define PACK_AND_STORE(src, i) do { \ 984 const vector unsigned char perms = vec_lvsr(i * stride, src); \ 985 const vector unsigned char vf = \ 986 vec_packsu(v_tempBlurredAss##1, (vector signed short)zero); \ 987 const vector unsigned char vg = vec_perm(vf, v_##src##A##i, permHH); \ 988 const vector unsigned char mask = \ 989 vec_perm((vector unsigned char)zero, (vector unsigned char)neg1, perms); \ 990 const vector unsigned char vg2 = vec_perm(vg, vg, perms); \ 991 const vector unsigned char svA = vec_sel(v_##src##A1##i, vg2, mask); \ 992 const vector unsigned char svB = vec_sel(vg2, v_##src##A2##i, mask); \ 993 vec_st(svA, i * stride, src); \ 994 vec_st(svB, i * stride + 16, src); \ 995} while (0) 996 997 PACK_AND_STORE(src, 0); 998 PACK_AND_STORE(src, 1); 999 PACK_AND_STORE(src, 2); 1000 PACK_AND_STORE(src, 3); 1001 PACK_AND_STORE(src, 4); 1002 PACK_AND_STORE(src, 5); 1003 PACK_AND_STORE(src, 6); 1004 PACK_AND_STORE(src, 7); 1005 PACK_AND_STORE(tempBlurred, 0); 1006 PACK_AND_STORE(tempBlurred, 1); 1007 PACK_AND_STORE(tempBlurred, 2); 1008 PACK_AND_STORE(tempBlurred, 3); 1009 PACK_AND_STORE(tempBlurred, 4); 1010 PACK_AND_STORE(tempBlurred, 5); 1011 PACK_AND_STORE(tempBlurred, 6); 1012 PACK_AND_STORE(tempBlurred, 7); 1013#undef PACK_AND_STORE 1014} 1015 1016static inline void transpose_16x8_char_toPackedAlign_altivec(unsigned char* dst, unsigned char* src, int stride) { 1017 const vector unsigned char zero = vec_splat_u8(0); 1018 1019#define LOAD_DOUBLE_LINE(i, j) \ 1020 vector unsigned char perm1##i = vec_lvsl(i * stride, src); \ 1021 vector unsigned char perm2##i = vec_lvsl(j * stride, src); \ 1022 vector unsigned char srcA##i = vec_ld(i * stride, src); \ 1023 vector unsigned char srcB##i = vec_ld(i * stride + 16, src); \ 1024 vector unsigned char srcC##i = vec_ld(j * stride, src); \ 1025 vector unsigned char srcD##i = vec_ld(j * stride+ 16, src); \ 1026 vector unsigned char src##i = vec_perm(srcA##i, srcB##i, perm1##i); \ 1027 vector unsigned char src##j = vec_perm(srcC##i, srcD##i, perm2##i) 1028 1029 LOAD_DOUBLE_LINE(0, 1); 1030 LOAD_DOUBLE_LINE(2, 3); 1031 LOAD_DOUBLE_LINE(4, 5); 1032 LOAD_DOUBLE_LINE(6, 7); 1033#undef LOAD_DOUBLE_LINE 1034 1035 vector unsigned char tempA = vec_mergeh(src0, zero); 1036 vector unsigned char tempB = vec_mergel(src0, zero); 1037 vector unsigned char tempC = vec_mergeh(src1, zero); 1038 vector unsigned char tempD = vec_mergel(src1, zero); 1039 vector unsigned char tempE = vec_mergeh(src2, zero); 1040 vector unsigned char tempF = vec_mergel(src2, zero); 1041 vector unsigned char tempG = vec_mergeh(src3, zero); 1042 vector unsigned char tempH = vec_mergel(src3, zero); 1043 vector unsigned char tempI = vec_mergeh(src4, zero); 1044 vector unsigned char tempJ = vec_mergel(src4, zero); 1045 vector unsigned char tempK = vec_mergeh(src5, zero); 1046 vector unsigned char tempL = vec_mergel(src5, zero); 1047 vector unsigned char tempM = vec_mergeh(src6, zero); 1048 vector unsigned char tempN = vec_mergel(src6, zero); 1049 vector unsigned char tempO = vec_mergeh(src7, zero); 1050 vector unsigned char tempP = vec_mergel(src7, zero); 1051 1052 vector unsigned char temp0 = vec_mergeh(tempA, tempI); 1053 vector unsigned char temp1 = vec_mergel(tempA, tempI); 1054 vector unsigned char temp2 = vec_mergeh(tempB, tempJ); 1055 vector unsigned char temp3 = vec_mergel(tempB, tempJ); 1056 vector unsigned char temp4 = vec_mergeh(tempC, tempK); 1057 vector unsigned char temp5 = vec_mergel(tempC, tempK); 1058 vector unsigned char temp6 = vec_mergeh(tempD, tempL); 1059 vector unsigned char temp7 = vec_mergel(tempD, tempL); 1060 vector unsigned char temp8 = vec_mergeh(tempE, tempM); 1061 vector unsigned char temp9 = vec_mergel(tempE, tempM); 1062 vector unsigned char temp10 = vec_mergeh(tempF, tempN); 1063 vector unsigned char temp11 = vec_mergel(tempF, tempN); 1064 vector unsigned char temp12 = vec_mergeh(tempG, tempO); 1065 vector unsigned char temp13 = vec_mergel(tempG, tempO); 1066 vector unsigned char temp14 = vec_mergeh(tempH, tempP); 1067 vector unsigned char temp15 = vec_mergel(tempH, tempP); 1068 1069 tempA = vec_mergeh(temp0, temp8); 1070 tempB = vec_mergel(temp0, temp8); 1071 tempC = vec_mergeh(temp1, temp9); 1072 tempD = vec_mergel(temp1, temp9); 1073 tempE = vec_mergeh(temp2, temp10); 1074 tempF = vec_mergel(temp2, temp10); 1075 tempG = vec_mergeh(temp3, temp11); 1076 tempH = vec_mergel(temp3, temp11); 1077 tempI = vec_mergeh(temp4, temp12); 1078 tempJ = vec_mergel(temp4, temp12); 1079 tempK = vec_mergeh(temp5, temp13); 1080 tempL = vec_mergel(temp5, temp13); 1081 tempM = vec_mergeh(temp6, temp14); 1082 tempN = vec_mergel(temp6, temp14); 1083 tempO = vec_mergeh(temp7, temp15); 1084 tempP = vec_mergel(temp7, temp15); 1085 1086 temp0 = vec_mergeh(tempA, tempI); 1087 temp1 = vec_mergel(tempA, tempI); 1088 temp2 = vec_mergeh(tempB, tempJ); 1089 temp3 = vec_mergel(tempB, tempJ); 1090 temp4 = vec_mergeh(tempC, tempK); 1091 temp5 = vec_mergel(tempC, tempK); 1092 temp6 = vec_mergeh(tempD, tempL); 1093 temp7 = vec_mergel(tempD, tempL); 1094 temp8 = vec_mergeh(tempE, tempM); 1095 temp9 = vec_mergel(tempE, tempM); 1096 temp10 = vec_mergeh(tempF, tempN); 1097 temp11 = vec_mergel(tempF, tempN); 1098 temp12 = vec_mergeh(tempG, tempO); 1099 temp13 = vec_mergel(tempG, tempO); 1100 temp14 = vec_mergeh(tempH, tempP); 1101 temp15 = vec_mergel(tempH, tempP); 1102 1103 vec_st(temp0, 0, dst); 1104 vec_st(temp1, 16, dst); 1105 vec_st(temp2, 32, dst); 1106 vec_st(temp3, 48, dst); 1107 vec_st(temp4, 64, dst); 1108 vec_st(temp5, 80, dst); 1109 vec_st(temp6, 96, dst); 1110 vec_st(temp7, 112, dst); 1111 vec_st(temp8, 128, dst); 1112 vec_st(temp9, 144, dst); 1113 vec_st(temp10, 160, dst); 1114 vec_st(temp11, 176, dst); 1115 vec_st(temp12, 192, dst); 1116 vec_st(temp13, 208, dst); 1117 vec_st(temp14, 224, dst); 1118 vec_st(temp15, 240, dst); 1119} 1120 1121static inline void transpose_8x16_char_fromPackedAlign_altivec(unsigned char* dst, unsigned char* src, int stride) { 1122 const vector unsigned char zero = vec_splat_u8(0); 1123 const vector signed char neg1 = vec_splat_s8(-1); 1124 1125#define LOAD_DOUBLE_LINE(i, j) \ 1126 vector unsigned char src##i = vec_ld(i * 16, src); \ 1127 vector unsigned char src##j = vec_ld(j * 16, src) 1128 1129 LOAD_DOUBLE_LINE(0, 1); 1130 LOAD_DOUBLE_LINE(2, 3); 1131 LOAD_DOUBLE_LINE(4, 5); 1132 LOAD_DOUBLE_LINE(6, 7); 1133 LOAD_DOUBLE_LINE(8, 9); 1134 LOAD_DOUBLE_LINE(10, 11); 1135 LOAD_DOUBLE_LINE(12, 13); 1136 LOAD_DOUBLE_LINE(14, 15); 1137#undef LOAD_DOUBLE_LINE 1138 1139 vector unsigned char tempA = vec_mergeh(src0, src8); 1140 vector unsigned char tempB; 1141 vector unsigned char tempC = vec_mergeh(src1, src9); 1142 vector unsigned char tempD; 1143 vector unsigned char tempE = vec_mergeh(src2, src10); 1144 vector unsigned char tempG = vec_mergeh(src3, src11); 1145 vector unsigned char tempI = vec_mergeh(src4, src12); 1146 vector unsigned char tempJ; 1147 vector unsigned char tempK = vec_mergeh(src5, src13); 1148 vector unsigned char tempL; 1149 vector unsigned char tempM = vec_mergeh(src6, src14); 1150 vector unsigned char tempO = vec_mergeh(src7, src15); 1151 1152 vector unsigned char temp0 = vec_mergeh(tempA, tempI); 1153 vector unsigned char temp1 = vec_mergel(tempA, tempI); 1154 vector unsigned char temp2; 1155 vector unsigned char temp3; 1156 vector unsigned char temp4 = vec_mergeh(tempC, tempK); 1157 vector unsigned char temp5 = vec_mergel(tempC, tempK); 1158 vector unsigned char temp6; 1159 vector unsigned char temp7; 1160 vector unsigned char temp8 = vec_mergeh(tempE, tempM); 1161 vector unsigned char temp9 = vec_mergel(tempE, tempM); 1162 vector unsigned char temp12 = vec_mergeh(tempG, tempO); 1163 vector unsigned char temp13 = vec_mergel(tempG, tempO); 1164 1165 tempA = vec_mergeh(temp0, temp8); 1166 tempB = vec_mergel(temp0, temp8); 1167 tempC = vec_mergeh(temp1, temp9); 1168 tempD = vec_mergel(temp1, temp9); 1169 tempI = vec_mergeh(temp4, temp12); 1170 tempJ = vec_mergel(temp4, temp12); 1171 tempK = vec_mergeh(temp5, temp13); 1172 tempL = vec_mergel(temp5, temp13); 1173 1174 temp0 = vec_mergeh(tempA, tempI); 1175 temp1 = vec_mergel(tempA, tempI); 1176 temp2 = vec_mergeh(tempB, tempJ); 1177 temp3 = vec_mergel(tempB, tempJ); 1178 temp4 = vec_mergeh(tempC, tempK); 1179 temp5 = vec_mergel(tempC, tempK); 1180 temp6 = vec_mergeh(tempD, tempL); 1181 temp7 = vec_mergel(tempD, tempL); 1182 1183 1184#define STORE_DOUBLE_LINE(i, j) do { \ 1185 vector unsigned char dstAi = vec_ld(i * stride, dst); \ 1186 vector unsigned char dstBi = vec_ld(i * stride + 16, dst); \ 1187 vector unsigned char dstAj = vec_ld(j * stride, dst); \ 1188 vector unsigned char dstBj = vec_ld(j * stride+ 16, dst); \ 1189 vector unsigned char aligni = vec_lvsr(i * stride, dst); \ 1190 vector unsigned char alignj = vec_lvsr(j * stride, dst); \ 1191 vector unsigned char maski = \ 1192 vec_perm(zero, (vector unsigned char)neg1, aligni); \ 1193 vector unsigned char maskj = \ 1194 vec_perm(zero, (vector unsigned char)neg1, alignj); \ 1195 vector unsigned char dstRi = vec_perm(temp##i, temp##i, aligni); \ 1196 vector unsigned char dstRj = vec_perm(temp##j, temp##j, alignj); \ 1197 vector unsigned char dstAFi = vec_sel(dstAi, dstRi, maski); \ 1198 vector unsigned char dstBFi = vec_sel(dstRi, dstBi, maski); \ 1199 vector unsigned char dstAFj = vec_sel(dstAj, dstRj, maskj); \ 1200 vector unsigned char dstBFj = vec_sel(dstRj, dstBj, maskj); \ 1201 vec_st(dstAFi, i * stride, dst); \ 1202 vec_st(dstBFi, i * stride + 16, dst); \ 1203 vec_st(dstAFj, j * stride, dst); \ 1204 vec_st(dstBFj, j * stride + 16, dst); \ 1205} while (0) 1206 1207 STORE_DOUBLE_LINE(0,1); 1208 STORE_DOUBLE_LINE(2,3); 1209 STORE_DOUBLE_LINE(4,5); 1210 STORE_DOUBLE_LINE(6,7); 1211} 1212