1;****************************************************************************** 2;* VP9 loop filter SIMD optimizations 3;* 4;* Copyright (C) 2013-2014 Cl��ment B��sch <u pkh me> 5;* 6;* This file is part of FFmpeg. 7;* 8;* FFmpeg is free software; you can redistribute it and/or 9;* modify it under the terms of the GNU Lesser General Public 10;* License as published by the Free Software Foundation; either 11;* version 2.1 of the License, or (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 GNU 16;* Lesser General Public License for more details. 17;* 18;* You should have received a copy of the GNU Lesser General Public 19;* License 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/x86/x86util.asm" 24 25%if ARCH_X86_64 26 27SECTION_RODATA 28 29cextern pb_3 30cextern pb_80 31 32pb_4: times 16 db 0x04 33pb_10: times 16 db 0x10 34pb_40: times 16 db 0x40 35pb_81: times 16 db 0x81 36pb_f8: times 16 db 0xf8 37pb_fe: times 16 db 0xfe 38 39pw_4: times 8 dw 4 40pw_8: times 8 dw 8 41 42; with mix functions, two 8-bit thresholds are stored in a 16-bit storage, 43; the following mask is used to splat both in the same register 44mask_mix: times 8 db 0 45 times 8 db 1 46 47mask_mix84: times 8 db 0xff 48 times 8 db 0x00 49mask_mix48: times 8 db 0x00 50 times 8 db 0xff 51 52SECTION .text 53 54; %1 = abs(%2-%3) 55%macro ABSSUB 4 ; dst, src1 (RO), src2 (RO), tmp 56 psubusb %1, %3, %2 57 psubusb %4, %2, %3 58 por %1, %4 59%endmacro 60 61; %1 = %1<=%2 62%macro CMP_LTE 3-4 ; src/dst, cmp, tmp, pb_80 63%if %0 == 4 64 pxor %1, %4 65%endif 66 pcmpgtb %3, %2, %1 ; cmp > src? 67 pcmpeqb %1, %2 ; cmp == src? XXX: avoid this with a -1/+1 well placed? 68 por %1, %3 ; cmp >= src? 69%endmacro 70 71; %1 = abs(%2-%3) <= %4 72%macro ABSSUB_CMP 6-7 [pb_80]; dst, src1, src2, cmp, tmp1, tmp2, [pb_80] 73 ABSSUB %1, %2, %3, %6 ; dst = abs(src1-src2) 74 CMP_LTE %1, %4, %6, %7 ; dst <= cmp 75%endmacro 76 77%macro MASK_APPLY 4 ; %1=new_data/dst %2=old_data %3=mask %4=tmp 78 pand %1, %3 ; new &= mask 79 pandn %4, %3, %2 ; tmp = ~mask & old 80 por %1, %4 ; new&mask | old&~mask 81%endmacro 82 83%macro FILTER_SUBx2_ADDx2 8 ; %1=dst %2=h/l %3=cache %4=sub1 %5=sub2 %6=add1 %7=add2 %8=rshift 84 punpck%2bw %3, %4, m0 85 psubw %1, %3 86 punpck%2bw %3, %5, m0 87 psubw %1, %3 88 punpck%2bw %3, %6, m0 89 paddw %1, %3 90 punpck%2bw %3, %7, m0 91 paddw %3, %1 92 psraw %1, %3, %8 93%endmacro 94 95%macro FILTER_INIT 8 ; tmp1, tmp2, cacheL, cacheH, dstp, filterid, mask, source 96 FILTER%6_INIT %1, l, %3 97 FILTER%6_INIT %2, h, %4 98 packuswb %1, %2 99 MASK_APPLY %1, %8, %7, %2 100 mova %5, %1 101%endmacro 102 103%macro FILTER_UPDATE 11-14 ; tmp1, tmp2, cacheL, cacheH, dstp, -, -, +, +, rshift, mask, [source], [preload reg + value] 104%if %0 == 13 ; no source + preload 105 mova %12, %13 106%elif %0 == 14 ; source + preload 107 mova %13, %14 108%endif 109 FILTER_SUBx2_ADDx2 %1, l, %3, %6, %7, %8, %9, %10 110 FILTER_SUBx2_ADDx2 %2, h, %4, %6, %7, %8, %9, %10 111 packuswb %1, %2 112%if %0 == 12 || %0 == 14 113 MASK_APPLY %1, %12, %11, %2 114%else 115 MASK_APPLY %1, %5, %11, %2 116%endif 117 mova %5, %1 118%endmacro 119 120%macro SRSHIFT3B_2X 4 ; reg1, reg2, [pb_10], tmp 121 mova %4, [pb_f8] 122 pand %1, %4 123 pand %2, %4 124 psrlq %1, 3 125 psrlq %2, 3 126 pxor %1, %3 127 pxor %2, %3 128 psubb %1, %3 129 psubb %2, %3 130%endmacro 131 132%macro EXTRACT_POS_NEG 3 ; i8, neg, pos 133 pxor %3, %3 134 pxor %2, %2 135 pcmpgtb %3, %1 ; i8 < 0 mask 136 psubb %2, %1 ; neg values (only the originally - will be kept) 137 pand %2, %3 ; negative values of i8 (but stored as +) 138 pandn %3, %1 ; positive values of i8 139%endmacro 140 141; clip_u8(u8 + i8) 142%macro SIGN_ADD 5 ; dst, u8, i8, tmp1, tmp2 143 EXTRACT_POS_NEG %3, %4, %5 144 psubusb %1, %2, %4 ; sub the negatives 145 paddusb %1, %5 ; add the positives 146%endmacro 147 148; clip_u8(u8 - i8) 149%macro SIGN_SUB 5 ; dst, u8, i8, tmp1, tmp2 150 EXTRACT_POS_NEG %3, %4, %5 151 psubusb %1, %2, %5 ; sub the positives 152 paddusb %1, %4 ; add the negatives 153%endmacro 154 155%macro FILTER6_INIT 3 ; %1=dst %2=h/l %3=cache 156 punpck%2bw %1, m14, m0 ; p3: B->W 157 paddw %3, %1, %1 ; p3*2 158 paddw %3, %1 ; p3*3 159 punpck%2bw %1, m15, m0 ; p2: B->W 160 paddw %3, %1 ; p3*3 + p2 161 paddw %3, %1 ; p3*3 + p2*2 162 punpck%2bw %1, m10, m0 ; p1: B->W 163 paddw %3, %1 ; p3*3 + p2*2 + p1 164 punpck%2bw %1, m11, m0 ; p0: B->W 165 paddw %3, %1 ; p3*3 + p2*2 + p1 + p0 166 punpck%2bw %1, m12, m0 ; q0: B->W 167 paddw %3, %1 ; p3*3 + p2*2 + p1 + p0 + q0 168 paddw %3, [pw_4] ; p3*3 + p2*2 + p1 + p0 + q0 + 4 169 psraw %1, %3, 3 ; (p3*3 + p2*2 + p1 + p0 + q0 + 4) >> 3 170%endmacro 171 172%macro FILTER14_INIT 3 ; %1=dst %2=h/l %3=cache 173 punpck%2bw %1, m2, m0 ; p7: B->W 174 psllw %3, %1, 3 ; p7*8 175 psubw %3, %1 ; p7*7 176 punpck%2bw %1, m3, m0 ; p6: B->W 177 paddw %3, %1 ; p7*7 + p6 178 paddw %3, %1 ; p7*7 + p6*2 179 punpck%2bw %1, m8, m0 ; p5: B->W 180 paddw %3, %1 ; p7*7 + p6*2 + p5 181 punpck%2bw %1, m9, m0 ; p4: B->W 182 paddw %3, %1 ; p7*7 + p6*2 + p5 + p4 183 punpck%2bw %1, m14, m0 ; p3: B->W 184 paddw %3, %1 ; p7*7 + p6*2 + p5 + p4 + p3 185 punpck%2bw %1, m15, m0 ; p2: B->W 186 paddw %3, %1 ; p7*7 + p6*2 + p5 + .. + p2 187 punpck%2bw %1, m10, m0 ; p1: B->W 188 paddw %3, %1 ; p7*7 + p6*2 + p5 + .. + p1 189 punpck%2bw %1, m11, m0 ; p0: B->W 190 paddw %3, %1 ; p7*7 + p6*2 + p5 + .. + p0 191 punpck%2bw %1, m12, m0 ; q0: B->W 192 paddw %3, %1 ; p7*7 + p6*2 + p5 + .. + p0 + q0 193 paddw %3, [pw_8] ; p7*7 + p6*2 + p5 + .. + p0 + q0 + 8 194 psraw %1, %3, 4 ; (p7*7 + p6*2 + p5 + .. + p0 + q0 + 8) >> 4 195%endmacro 196 197%macro TRANSPOSE16x16B 17 198 mova %17, m%16 199 SBUTTERFLY bw, %1, %2, %16 200 SBUTTERFLY bw, %3, %4, %16 201 SBUTTERFLY bw, %5, %6, %16 202 SBUTTERFLY bw, %7, %8, %16 203 SBUTTERFLY bw, %9, %10, %16 204 SBUTTERFLY bw, %11, %12, %16 205 SBUTTERFLY bw, %13, %14, %16 206 mova m%16, %17 207 mova %17, m%14 208 SBUTTERFLY bw, %15, %16, %14 209 SBUTTERFLY wd, %1, %3, %14 210 SBUTTERFLY wd, %2, %4, %14 211 SBUTTERFLY wd, %5, %7, %14 212 SBUTTERFLY wd, %6, %8, %14 213 SBUTTERFLY wd, %9, %11, %14 214 SBUTTERFLY wd, %10, %12, %14 215 SBUTTERFLY wd, %13, %15, %14 216 mova m%14, %17 217 mova %17, m%12 218 SBUTTERFLY wd, %14, %16, %12 219 SBUTTERFLY dq, %1, %5, %12 220 SBUTTERFLY dq, %2, %6, %12 221 SBUTTERFLY dq, %3, %7, %12 222 SBUTTERFLY dq, %4, %8, %12 223 SBUTTERFLY dq, %9, %13, %12 224 SBUTTERFLY dq, %10, %14, %12 225 SBUTTERFLY dq, %11, %15, %12 226 mova m%12, %17 227 mova %17, m%8 228 SBUTTERFLY dq, %12, %16, %8 229 SBUTTERFLY qdq, %1, %9, %8 230 SBUTTERFLY qdq, %2, %10, %8 231 SBUTTERFLY qdq, %3, %11, %8 232 SBUTTERFLY qdq, %4, %12, %8 233 SBUTTERFLY qdq, %5, %13, %8 234 SBUTTERFLY qdq, %6, %14, %8 235 SBUTTERFLY qdq, %7, %15, %8 236 mova m%8, %17 237 mova %17, m%1 238 SBUTTERFLY qdq, %8, %16, %1 239 mova m%1, %17 240 SWAP %2, %9 241 SWAP %3, %5 242 SWAP %4, %13 243 SWAP %6, %11 244 SWAP %8, %15 245 SWAP %12, %14 246%endmacro 247 248; transpose 16 half lines (high part) to 8 full centered lines 249%macro TRANSPOSE16x8B 16 250 punpcklbw m%1, m%2 251 punpcklbw m%3, m%4 252 punpcklbw m%5, m%6 253 punpcklbw m%7, m%8 254 punpcklbw m%9, m%10 255 punpcklbw m%11, m%12 256 punpcklbw m%13, m%14 257 punpcklbw m%15, m%16 258 SBUTTERFLY wd, %1, %3, %2 259 SBUTTERFLY wd, %5, %7, %2 260 SBUTTERFLY wd, %9, %11, %2 261 SBUTTERFLY wd, %13, %15, %2 262 SBUTTERFLY dq, %1, %5, %2 263 SBUTTERFLY dq, %3, %7, %2 264 SBUTTERFLY dq, %9, %13, %2 265 SBUTTERFLY dq, %11, %15, %2 266 SBUTTERFLY qdq, %1, %9, %2 267 SBUTTERFLY qdq, %3, %11, %2 268 SBUTTERFLY qdq, %5, %13, %2 269 SBUTTERFLY qdq, %7, %15, %2 270 SWAP %5, %1 271 SWAP %6, %9 272 SWAP %7, %1 273 SWAP %8, %13 274 SWAP %9, %3 275 SWAP %10, %11 276 SWAP %11, %1 277 SWAP %12, %15 278%endmacro 279 280%macro DEFINE_REAL_P7_TO_Q7 0-1 0 281%define P7 dst1q + 2*mstrideq + %1 282%define P6 dst1q + mstrideq + %1 283%define P5 dst1q + %1 284%define P4 dst1q + strideq + %1 285%define P3 dstq + 4*mstrideq + %1 286%define P2 dstq + mstride3q + %1 287%define P1 dstq + 2*mstrideq + %1 288%define P0 dstq + mstrideq + %1 289%define Q0 dstq + %1 290%define Q1 dstq + strideq + %1 291%define Q2 dstq + 2*strideq + %1 292%define Q3 dstq + stride3q + %1 293%define Q4 dstq + 4*strideq + %1 294%define Q5 dst2q + mstrideq + %1 295%define Q6 dst2q + %1 296%define Q7 dst2q + strideq + %1 297%endmacro 298 299; ..............AB -> AAAAAAAABBBBBBBB 300%macro SPLATB_MIX 1-2 [mask_mix] 301%if cpuflag(ssse3) 302 pshufb %1, %2 303%else 304 punpcklbw %1, %1 305 punpcklqdq %1, %1 306 pshuflw %1, %1, 0 307 pshufhw %1, %1, 0x55 308%endif 309%endmacro 310 311%macro LOOPFILTER 2 ; %1=v/h %2=size1 312 lea mstrideq, [strideq] 313 neg mstrideq 314 315 lea stride3q, [strideq+2*strideq] 316 mov mstride3q, stride3q 317 neg mstride3q 318 319%ifidn %1, h 320%if %2 > 16 321%define movx movh 322 lea dstq, [dstq + 8*strideq - 4] 323%else 324%define movx movu 325 lea dstq, [dstq + 8*strideq - 8] ; go from top center (h pos) to center left (v pos) 326%endif 327%endif 328 329 lea dst1q, [dstq + 2*mstride3q] ; dst1q = &dst[stride * -6] 330 lea dst2q, [dstq + 2* stride3q] ; dst2q = &dst[stride * +6] 331 332 DEFINE_REAL_P7_TO_Q7 333 334%ifidn %1, h 335 movx m0, [P7] 336 movx m1, [P6] 337 movx m2, [P5] 338 movx m3, [P4] 339 movx m4, [P3] 340 movx m5, [P2] 341 movx m6, [P1] 342 movx m7, [P0] 343 movx m8, [Q0] 344 movx m9, [Q1] 345 movx m10, [Q2] 346 movx m11, [Q3] 347 movx m12, [Q4] 348 movx m13, [Q5] 349 movx m14, [Q6] 350 movx m15, [Q7] 351%define P7 rsp + 0 352%define P6 rsp + 16 353%define P5 rsp + 32 354%define P4 rsp + 48 355%define P3 rsp + 64 356%define P2 rsp + 80 357%define P1 rsp + 96 358%define P0 rsp + 112 359%define Q0 rsp + 128 360%define Q1 rsp + 144 361%define Q2 rsp + 160 362%define Q3 rsp + 176 363%define Q4 rsp + 192 364%define Q5 rsp + 208 365%define Q6 rsp + 224 366%define Q7 rsp + 240 367 368%if %2 == 16 369 TRANSPOSE16x16B 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, [rsp] 370 mova [P7], m0 371 mova [P6], m1 372 mova [P5], m2 373 mova [P4], m3 374%else 375 TRANSPOSE16x8B 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 376%endif 377 mova [P3], m4 378 mova [P2], m5 379 mova [P1], m6 380 mova [P0], m7 381 mova [Q0], m8 382 mova [Q1], m9 383 mova [Q2], m10 384 mova [Q3], m11 385%if %2 == 16 386 mova [Q4], m12 387 mova [Q5], m13 388 mova [Q6], m14 389 mova [Q7], m15 390%endif 391%endif 392 393 ; calc fm mask 394%if %2 == 16 395%if cpuflag(ssse3) 396 pxor m0, m0 397%endif 398 SPLATB_REG m2, I, m0 ; I I I I ... 399 SPLATB_REG m3, E, m0 ; E E E E ... 400%else 401%if cpuflag(ssse3) 402 mova m0, [mask_mix] 403%endif 404 movd m2, Id 405 movd m3, Ed 406 SPLATB_MIX m2, m0 407 SPLATB_MIX m3, m0 408%endif 409 mova m0, [pb_80] 410 pxor m2, m0 411 pxor m3, m0 412%ifidn %1, v 413 mova m8, [P3] 414 mova m9, [P2] 415 mova m10, [P1] 416 mova m11, [P0] 417 mova m12, [Q0] 418 mova m13, [Q1] 419 mova m14, [Q2] 420 mova m15, [Q3] 421%else 422 ; In case of horizontal, P3..Q3 are already present in some registers due 423 ; to the previous transpose, so we just swap registers. 424 SWAP 8, 4, 12 425 SWAP 9, 5, 13 426 SWAP 10, 6, 14 427 SWAP 11, 7, 15 428%endif 429 ABSSUB_CMP m5, m8, m9, m2, m6, m7, m0 ; m5 = abs(p3-p2) <= I 430 ABSSUB_CMP m1, m9, m10, m2, m6, m7, m0 ; m1 = abs(p2-p1) <= I 431 pand m5, m1 432 ABSSUB_CMP m1, m10, m11, m2, m6, m7, m0 ; m1 = abs(p1-p0) <= I 433 pand m5, m1 434 ABSSUB_CMP m1, m12, m13, m2, m6, m7, m0 ; m1 = abs(q1-q0) <= I 435 pand m5, m1 436 ABSSUB_CMP m1, m13, m14, m2, m6, m7, m0 ; m1 = abs(q2-q1) <= I 437 pand m5, m1 438 ABSSUB_CMP m1, m14, m15, m2, m6, m7, m0 ; m1 = abs(q3-q2) <= I 439 pand m5, m1 440 ABSSUB m1, m11, m12, m7 ; abs(p0-q0) 441 paddusb m1, m1 ; abs(p0-q0) * 2 442 ABSSUB m2, m10, m13, m7 ; abs(p1-q1) 443 pand m2, [pb_fe] ; drop lsb so shift can work 444 psrlq m2, 1 ; abs(p1-q1)/2 445 paddusb m1, m2 ; abs(p0-q0)*2 + abs(p1-q1)/2 446 pxor m1, m0 447 pcmpgtb m4, m3, m1 ; E > X? 448 pcmpeqb m3, m1 ; E == X? 449 por m3, m4 ; E >= X? 450 pand m3, m5 ; fm final value 451 452 ; (m3: fm, m8..15: p3 p2 p1 p0 q0 q1 q2 q3) 453 ; calc flat8in (if not 44_16) and hev masks 454 mova m6, [pb_81] ; [1 1 1 1 ...] ^ 0x80 455%if %2 != 44 456 ABSSUB_CMP m2, m8, m11, m6, m4, m5 ; abs(p3 - p0) <= 1 457 mova m8, [pb_80] 458 ABSSUB_CMP m1, m9, m11, m6, m4, m5, m8 ; abs(p2 - p0) <= 1 459 pand m2, m1 460 ABSSUB m4, m10, m11, m5 ; abs(p1 - p0) 461%if %2 == 16 462%if cpuflag(ssse3) 463 pxor m0, m0 464%endif 465 SPLATB_REG m7, H, m0 ; H H H H ... 466%else 467 movd m7, Hd 468 SPLATB_MIX m7 469%endif 470 pxor m7, m8 471 pxor m4, m8 472 pcmpgtb m0, m4, m7 ; abs(p1 - p0) > H (1/2 hev condition) 473 CMP_LTE m4, m6, m5 ; abs(p1 - p0) <= 1 474 pand m2, m4 ; (flat8in) 475 ABSSUB m4, m13, m12, m1 ; abs(q1 - q0) 476 pxor m4, m8 477 pcmpgtb m5, m4, m7 ; abs(q1 - q0) > H (2/2 hev condition) 478 por m0, m5 ; hev final value 479 CMP_LTE m4, m6, m5 ; abs(q1 - q0) <= 1 480 pand m2, m4 ; (flat8in) 481 ABSSUB_CMP m1, m14, m12, m6, m4, m5, m8 ; abs(q2 - q0) <= 1 482 pand m2, m1 483 ABSSUB_CMP m1, m15, m12, m6, m4, m5, m8 ; abs(q3 - q0) <= 1 484 pand m2, m1 ; flat8in final value 485%if %2 == 84 || %2 == 48 486 pand m2, [mask_mix%2] 487%endif 488%else 489 mova m6, [pb_80] 490 movd m7, Hd 491 SPLATB_MIX m7 492 pxor m7, m6 493 ABSSUB m4, m10, m11, m1 ; abs(p1 - p0) 494 pxor m4, m6 495 pcmpgtb m0, m4, m7 ; abs(p1 - p0) > H (1/2 hev condition) 496 ABSSUB m4, m13, m12, m1 ; abs(q1 - q0) 497 pxor m4, m6 498 pcmpgtb m5, m4, m7 ; abs(q1 - q0) > H (2/2 hev condition) 499 por m0, m5 ; hev final value 500%endif 501 502%if %2 == 16 503 ; (m0: hev, m2: flat8in, m3: fm, m6: pb_81, m9..15: p2 p1 p0 q0 q1 q2 q3) 504 ; calc flat8out mask 505 mova m8, [P7] 506 mova m9, [P6] 507 ABSSUB_CMP m1, m8, m11, m6, m4, m5 ; abs(p7 - p0) <= 1 508 ABSSUB_CMP m7, m9, m11, m6, m4, m5 ; abs(p6 - p0) <= 1 509 pand m1, m7 510 mova m8, [P5] 511 mova m9, [P4] 512 ABSSUB_CMP m7, m8, m11, m6, m4, m5 ; abs(p5 - p0) <= 1 513 pand m1, m7 514 ABSSUB_CMP m7, m9, m11, m6, m4, m5 ; abs(p4 - p0) <= 1 515 pand m1, m7 516 mova m14, [Q4] 517 mova m15, [Q5] 518 ABSSUB_CMP m7, m14, m12, m6, m4, m5 ; abs(q4 - q0) <= 1 519 pand m1, m7 520 ABSSUB_CMP m7, m15, m12, m6, m4, m5 ; abs(q5 - q0) <= 1 521 pand m1, m7 522 mova m14, [Q6] 523 mova m15, [Q7] 524 ABSSUB_CMP m7, m14, m12, m6, m4, m5 ; abs(q4 - q0) <= 1 525 pand m1, m7 526 ABSSUB_CMP m7, m15, m12, m6, m4, m5 ; abs(q5 - q0) <= 1 527 pand m1, m7 ; flat8out final value 528%endif 529 530 ; if (fm) { 531 ; if (out && in) filter_14() 532 ; else if (in) filter_6() 533 ; else if (hev) filter_2() 534 ; else filter_4() 535 ; } 536 ; 537 ; f14: fm & out & in 538 ; f6: fm & ~f14 & in => fm & ~(out & in) & in => fm & ~out & in 539 ; f2: fm & ~f14 & ~f6 & hev => fm & ~(out & in) & ~(~out & in) & hev => fm & ~in & hev 540 ; f4: fm & ~f14 & ~f6 & ~f2 => fm & ~(out & in) & ~(~out & in) & ~(~in & hev) => fm & ~in & ~hev 541 542 ; (m0: hev, [m1: flat8out], [m2: flat8in], m3: fm, m8..15: p5 p4 p1 p0 q0 q1 q6 q7) 543 ; filter2() 544%if %2 != 44 545 mova m6, [pb_80] ; already in m6 if 44_16 546%endif 547 pxor m15, m12, m6 ; q0 ^ 0x80 548 pxor m14, m11, m6 ; p0 ^ 0x80 549 psubsb m15, m14 ; (signed) q0 - p0 550 pxor m4, m10, m6 ; p1 ^ 0x80 551 pxor m5, m13, m6 ; q1 ^ 0x80 552 psubsb m4, m5 ; (signed) p1 - q1 553 paddsb m4, m15 ; (q0 - p0) + (p1 - q1) 554 paddsb m4, m15 ; 2*(q0 - p0) + (p1 - q1) 555 paddsb m4, m15 ; 3*(q0 - p0) + (p1 - q1) 556 paddsb m6, m4, [pb_4] ; m6: f1 = clip(f + 4, 127) 557 paddsb m4, [pb_3] ; m4: f2 = clip(f + 3, 127) 558 mova m14, [pb_10] ; will be reused in filter4() 559 SRSHIFT3B_2X m6, m4, m14, m7 ; f1 and f2 sign byte shift by 3 560 SIGN_SUB m7, m12, m6, m5, m9 ; m7 = q0 - f1 561 SIGN_ADD m8, m11, m4, m5, m9 ; m8 = p0 + f2 562%if %2 != 44 563 pandn m6, m2, m3 ; ~mask(in) & mask(fm) 564 pand m6, m0 ; (~mask(in) & mask(fm)) & mask(hev) 565%else 566 pand m6, m3, m0 567%endif 568 MASK_APPLY m7, m12, m6, m5 ; m7 = filter2(q0) & mask / we write it in filter4() 569 MASK_APPLY m8, m11, m6, m5 ; m8 = filter2(p0) & mask / we write it in filter4() 570 571 ; (m0: hev, [m1: flat8out], [m2: flat8in], m3: fm, m7..m8: q0' p0', m10..13: p1 p0 q0 q1, m14: pb_10, m15: q0-p0) 572 ; filter4() 573 mova m4, m15 574 paddsb m15, m4 ; 2 * (q0 - p0) 575 paddsb m15, m4 ; 3 * (q0 - p0) 576 paddsb m6, m15, [pb_4] ; m6: f1 = clip(f + 4, 127) 577 paddsb m15, [pb_3] ; m15: f2 = clip(f + 3, 127) 578 SRSHIFT3B_2X m6, m15, m14, m9 ; f1 and f2 sign byte shift by 3 579%if %2 != 44 580%define p0tmp m7 581%define q0tmp m9 582 pandn m5, m2, m3 ; ~mask(in) & mask(fm) 583 pandn m0, m5 ; ~mask(hev) & (~mask(in) & mask(fm)) 584%else 585%define p0tmp m1 586%define q0tmp m2 587 pandn m0, m3 588%endif 589 SIGN_SUB q0tmp, m12, m6, m4, m14 ; q0 - f1 590 MASK_APPLY q0tmp, m7, m0, m5 ; filter4(q0) & mask 591 mova [Q0], q0tmp 592 SIGN_ADD p0tmp, m11, m15, m4, m14 ; p0 + f2 593 MASK_APPLY p0tmp, m8, m0, m5 ; filter4(p0) & mask 594 mova [P0], p0tmp 595 paddb m6, [pb_80] ; 596 pxor m8, m8 ; f=(f1+1)>>1 597 pavgb m6, m8 ; 598 psubb m6, [pb_40] ; 599 SIGN_ADD m7, m10, m6, m8, m9 ; p1 + f 600 SIGN_SUB m4, m13, m6, m8, m9 ; q1 - f 601 MASK_APPLY m7, m10, m0, m14 ; m7 = filter4(p1) 602 MASK_APPLY m4, m13, m0, m14 ; m4 = filter4(q1) 603 mova [P1], m7 604 mova [Q1], m4 605 606 ; ([m1: flat8out], m2: flat8in, m3: fm, m10..13: p1 p0 q0 q1) 607 ; filter6() 608%if %2 != 44 609 pxor m0, m0 610%if %2 > 16 611 pand m3, m2 612%else 613 pand m2, m3 ; mask(fm) & mask(in) 614 pandn m3, m1, m2 ; ~mask(out) & (mask(fm) & mask(in)) 615%endif 616 mova m14, [P3] 617 mova m15, [P2] 618 mova m8, [Q2] 619 mova m9, [Q3] 620 FILTER_INIT m4, m5, m6, m7, [P2], 6, m3, m15 ; [p2] 621 FILTER_UPDATE m6, m7, m4, m5, [P1], m14, m15, m10, m13, 3, m3 ; [p1] -p3 -p2 +p1 +q1 622 FILTER_UPDATE m4, m5, m6, m7, [P0], m14, m10, m11, m8, 3, m3 ; [p0] -p3 -p1 +p0 +q2 623 FILTER_UPDATE m6, m7, m4, m5, [Q0], m14, m11, m12, m9, 3, m3 ; [q0] -p3 -p0 +q0 +q3 624 FILTER_UPDATE m4, m5, m6, m7, [Q1], m15, m12, m13, m9, 3, m3 ; [q1] -p2 -q0 +q1 +q3 625 FILTER_UPDATE m6, m7, m4, m5, [Q2], m10, m13, m8, m9, 3, m3, m8 ; [q2] -p1 -q1 +q2 +q3 626%endif 627 628 ; (m0: 0, [m1: flat8out], m2: fm & flat8in, m8..15: q2 q3 p1 p0 q0 q1 p3 p2) 629 ; filter14() 630 ; 631 ; m2 m3 m8 m9 m14 m15 m10 m11 m12 m13 632 ; 633 ; q2 q3 p3 p2 p1 p0 q0 q1 634 ; p6 -7 p7 p6 p5 p4 . . . . . 635 ; p5 -6 -p7 -p6 +p5 +q1 . . . . 636 ; p4 -5 -p7 -p5 +p4 +q2 . . . q2 637 ; p3 -4 -p7 -p4 +p3 +q3 . . . q3 638 ; p2 -3 -p7 -p3 +p2 +q4 . . . q4 639 ; p1 -2 -p7 -p2 +p1 +q5 . . . q5 640 ; p0 -1 -p7 -p1 +p0 +q6 . . . q6 641 ; q0 +0 -p7 -p0 +q0 +q7 . . . q7 642 ; q1 +1 -p6 -q0 +q1 +q7 q1 . . . 643 ; q2 +2 -p5 -q1 +q2 +q7 . q2 . . 644 ; q3 +3 -p4 -q2 +q3 +q7 . q3 . . 645 ; q4 +4 -p3 -q3 +q4 +q7 . q4 . . 646 ; q5 +5 -p2 -q4 +q5 +q7 . q5 . . 647 ; q6 +6 -p1 -q5 +q6 +q7 . q6 . . 648 649%if %2 == 16 650 pand m1, m2 ; mask(out) & (mask(fm) & mask(in)) 651 mova m2, [P7] 652 mova m3, [P6] 653 mova m8, [P5] 654 mova m9, [P4] 655 FILTER_INIT m4, m5, m6, m7, [P6], 14, m1, m3 656 FILTER_UPDATE m6, m7, m4, m5, [P5], m2, m3, m8, m13, 4, m1, m8 ; [p5] -p7 -p6 +p5 +q1 657 FILTER_UPDATE m4, m5, m6, m7, [P4], m2, m8, m9, m13, 4, m1, m9, m13, [Q2] ; [p4] -p7 -p5 +p4 +q2 658 FILTER_UPDATE m6, m7, m4, m5, [P3], m2, m9, m14, m13, 4, m1, m14, m13, [Q3] ; [p3] -p7 -p4 +p3 +q3 659 FILTER_UPDATE m4, m5, m6, m7, [P2], m2, m14, m15, m13, 4, m1, m13, [Q4] ; [p2] -p7 -p3 +p2 +q4 660 FILTER_UPDATE m6, m7, m4, m5, [P1], m2, m15, m10, m13, 4, m1, m13, [Q5] ; [p1] -p7 -p2 +p1 +q5 661 FILTER_UPDATE m4, m5, m6, m7, [P0], m2, m10, m11, m13, 4, m1, m13, [Q6] ; [p0] -p7 -p1 +p0 +q6 662 FILTER_UPDATE m6, m7, m4, m5, [Q0], m2, m11, m12, m13, 4, m1, m13, [Q7] ; [q0] -p7 -p0 +q0 +q7 663 FILTER_UPDATE m4, m5, m6, m7, [Q1], m3, m12, m2, m13, 4, m1, m2, [Q1] ; [q1] -p6 -q0 +q1 +q7 664 FILTER_UPDATE m6, m7, m4, m5, [Q2], m8, m2, m3, m13, 4, m1, m3, [Q2] ; [q2] -p5 -q1 +q2 +q7 665 FILTER_UPDATE m4, m5, m6, m7, [Q3], m9, m3, m8, m13, 4, m1, m8, m8, [Q3] ; [q3] -p4 -q2 +q3 +q7 666 FILTER_UPDATE m6, m7, m4, m5, [Q4], m14, m8, m9, m13, 4, m1, m9, m9, [Q4] ; [q4] -p3 -q3 +q4 +q7 667 FILTER_UPDATE m4, m5, m6, m7, [Q5], m15, m9, m14, m13, 4, m1, m14, m14, [Q5] ; [q5] -p2 -q4 +q5 +q7 668 FILTER_UPDATE m6, m7, m4, m5, [Q6], m10, m14, m15, m13, 4, m1, m15, m15, [Q6] ; [q6] -p1 -q5 +q6 +q7 669%endif 670 671%ifidn %1, h 672%if %2 == 16 673 mova m0, [P7] 674 mova m1, [P6] 675 mova m2, [P5] 676 mova m3, [P4] 677 mova m4, [P3] 678 mova m5, [P2] 679 mova m6, [P1] 680 mova m7, [P0] 681 mova m8, [Q0] 682 mova m9, [Q1] 683 mova m10, [Q2] 684 mova m11, [Q3] 685 mova m12, [Q4] 686 mova m13, [Q5] 687 mova m14, [Q6] 688 mova m15, [Q7] 689 TRANSPOSE16x16B 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, [rsp] 690 DEFINE_REAL_P7_TO_Q7 691 movu [P7], m0 692 movu [P6], m1 693 movu [P5], m2 694 movu [P4], m3 695 movu [P3], m4 696 movu [P2], m5 697 movu [P1], m6 698 movu [P0], m7 699 movu [Q0], m8 700 movu [Q1], m9 701 movu [Q2], m10 702 movu [Q3], m11 703 movu [Q4], m12 704 movu [Q5], m13 705 movu [Q6], m14 706 movu [Q7], m15 707%elif %2 == 44 708 SWAP 0, 7 ; m0 = p1 709 SWAP 3, 4 ; m3 = q1 710 DEFINE_REAL_P7_TO_Q7 2 711 SBUTTERFLY bw, 0, 1, 8 712 SBUTTERFLY bw, 2, 3, 8 713 SBUTTERFLY wd, 0, 2, 8 714 SBUTTERFLY wd, 1, 3, 8 715 SBUTTERFLY dq, 0, 4, 8 716 SBUTTERFLY dq, 1, 5, 8 717 SBUTTERFLY dq, 2, 6, 8 718 SBUTTERFLY dq, 3, 7, 8 719 movd [P7], m0 720 punpckhqdq m0, m8 721 movd [P6], m0 722 movd [Q0], m1 723 punpckhqdq m1, m9 724 movd [Q1], m1 725 movd [P3], m2 726 punpckhqdq m2, m10 727 movd [P2], m2 728 movd [Q4], m3 729 punpckhqdq m3, m11 730 movd [Q5], m3 731 movd [P5], m4 732 punpckhqdq m4, m12 733 movd [P4], m4 734 movd [Q2], m5 735 punpckhqdq m5, m13 736 movd [Q3], m5 737 movd [P1], m6 738 punpckhqdq m6, m14 739 movd [P0], m6 740 movd [Q6], m7 741 punpckhqdq m7, m8 742 movd [Q7], m7 743%else 744 ; the following code do a transpose of 8 full lines to 16 half 745 ; lines (high part). It is inlined to avoid the need of a staging area 746 mova m0, [P3] 747 mova m1, [P2] 748 mova m2, [P1] 749 mova m3, [P0] 750 mova m4, [Q0] 751 mova m5, [Q1] 752 mova m6, [Q2] 753 mova m7, [Q3] 754 DEFINE_REAL_P7_TO_Q7 755 SBUTTERFLY bw, 0, 1, 8 756 SBUTTERFLY bw, 2, 3, 8 757 SBUTTERFLY bw, 4, 5, 8 758 SBUTTERFLY bw, 6, 7, 8 759 SBUTTERFLY wd, 0, 2, 8 760 SBUTTERFLY wd, 1, 3, 8 761 SBUTTERFLY wd, 4, 6, 8 762 SBUTTERFLY wd, 5, 7, 8 763 SBUTTERFLY dq, 0, 4, 8 764 SBUTTERFLY dq, 1, 5, 8 765 SBUTTERFLY dq, 2, 6, 8 766 SBUTTERFLY dq, 3, 7, 8 767 movh [P7], m0 768 punpckhqdq m0, m8 769 movh [P6], m0 770 movh [Q0], m1 771 punpckhqdq m1, m9 772 movh [Q1], m1 773 movh [P3], m2 774 punpckhqdq m2, m10 775 movh [P2], m2 776 movh [Q4], m3 777 punpckhqdq m3, m11 778 movh [Q5], m3 779 movh [P5], m4 780 punpckhqdq m4, m12 781 movh [P4], m4 782 movh [Q2], m5 783 punpckhqdq m5, m13 784 movh [Q3], m5 785 movh [P1], m6 786 punpckhqdq m6, m14 787 movh [P0], m6 788 movh [Q6], m7 789 punpckhqdq m7, m8 790 movh [Q7], m7 791%endif 792%endif 793 794 RET 795%endmacro 796 797%macro LPF_16_VH 2 798INIT_XMM %2 799cglobal vp9_loop_filter_v_%1_16, 5,10,16, dst, stride, E, I, H, mstride, dst1, dst2, stride3, mstride3 800 LOOPFILTER v, %1 801cglobal vp9_loop_filter_h_%1_16, 5,10,16, 256, dst, stride, E, I, H, mstride, dst1, dst2, stride3, mstride3 802 LOOPFILTER h, %1 803%endmacro 804 805%macro LPF_16_VH_ALL_OPTS 1 806LPF_16_VH %1, sse2 807LPF_16_VH %1, ssse3 808LPF_16_VH %1, avx 809%endmacro 810 811LPF_16_VH_ALL_OPTS 16 812LPF_16_VH_ALL_OPTS 44 813LPF_16_VH_ALL_OPTS 48 814LPF_16_VH_ALL_OPTS 84 815LPF_16_VH_ALL_OPTS 88 816 817%endif ; x86-64 818