Yaowu Xu | c27fc14 | 2016-08-22 16:08:15 -0700 | [diff] [blame] | 1 | ; |
Yaowu Xu | 9c01aa1 | 2016-09-01 14:32:49 -0700 | [diff] [blame] | 2 | ; Copyright (c) 2016, Alliance for Open Media. All rights reserved |
Yaowu Xu | c27fc14 | 2016-08-22 16:08:15 -0700 | [diff] [blame] | 3 | ; |
Yaowu Xu | 9c01aa1 | 2016-09-01 14:32:49 -0700 | [diff] [blame] | 4 | ; This source code is subject to the terms of the BSD 2 Clause License and |
| 5 | ; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License |
| 6 | ; was not distributed with this source code in the LICENSE file, you can |
| 7 | ; obtain it at www.aomedia.org/license/software. If the Alliance for Open |
| 8 | ; Media Patent License 1.0 was not distributed with this source code in the |
| 9 | ; PATENTS file, you can obtain it at www.aomedia.org/license/patent. |
| 10 | ; |
| 11 | |
Yaowu Xu | c27fc14 | 2016-08-22 16:08:15 -0700 | [diff] [blame] | 12 | ; |
| 13 | |
| 14 | %include "aom_ports/x86_abi_support.asm" |
| 15 | |
| 16 | %macro GET_PARAM_4 0 |
| 17 | mov rdx, arg(5) ;filter ptr |
| 18 | mov rsi, arg(0) ;src_ptr |
| 19 | mov rdi, arg(2) ;output_ptr |
| 20 | mov rcx, 0x0400040 |
| 21 | |
| 22 | movdqa xmm3, [rdx] ;load filters |
| 23 | pshuflw xmm4, xmm3, 11111111b ;k3 |
| 24 | psrldq xmm3, 8 |
| 25 | pshuflw xmm3, xmm3, 0b ;k4 |
| 26 | punpcklqdq xmm4, xmm3 ;k3k4 |
| 27 | |
| 28 | movq xmm3, rcx ;rounding |
| 29 | pshufd xmm3, xmm3, 0 |
| 30 | |
| 31 | pxor xmm2, xmm2 |
| 32 | |
| 33 | movsxd rax, DWORD PTR arg(1) ;pixels_per_line |
| 34 | movsxd rdx, DWORD PTR arg(3) ;out_pitch |
| 35 | movsxd rcx, DWORD PTR arg(4) ;output_height |
| 36 | %endm |
| 37 | |
| 38 | %macro APPLY_FILTER_4 1 |
| 39 | |
| 40 | punpckldq xmm0, xmm1 ;two row in one register |
| 41 | punpcklbw xmm0, xmm2 ;unpack to word |
| 42 | pmullw xmm0, xmm4 ;multiply the filter factors |
| 43 | |
| 44 | movdqa xmm1, xmm0 |
| 45 | psrldq xmm1, 8 |
| 46 | paddsw xmm0, xmm1 |
| 47 | |
| 48 | paddsw xmm0, xmm3 ;rounding |
| 49 | psraw xmm0, 7 ;shift |
| 50 | packuswb xmm0, xmm0 ;pack to byte |
| 51 | |
| 52 | %if %1 |
| 53 | movd xmm1, [rdi] |
| 54 | pavgb xmm0, xmm1 |
| 55 | %endif |
| 56 | |
| 57 | movd [rdi], xmm0 |
| 58 | lea rsi, [rsi + rax] |
| 59 | lea rdi, [rdi + rdx] |
| 60 | dec rcx |
| 61 | %endm |
| 62 | |
| 63 | %macro GET_PARAM 0 |
| 64 | mov rdx, arg(5) ;filter ptr |
| 65 | mov rsi, arg(0) ;src_ptr |
| 66 | mov rdi, arg(2) ;output_ptr |
| 67 | mov rcx, 0x0400040 |
| 68 | |
| 69 | movdqa xmm7, [rdx] ;load filters |
| 70 | |
| 71 | pshuflw xmm6, xmm7, 11111111b ;k3 |
| 72 | pshufhw xmm7, xmm7, 0b ;k4 |
| 73 | punpcklwd xmm6, xmm6 |
| 74 | punpckhwd xmm7, xmm7 |
| 75 | |
| 76 | movq xmm4, rcx ;rounding |
| 77 | pshufd xmm4, xmm4, 0 |
| 78 | |
| 79 | pxor xmm5, xmm5 |
| 80 | |
| 81 | movsxd rax, DWORD PTR arg(1) ;pixels_per_line |
| 82 | movsxd rdx, DWORD PTR arg(3) ;out_pitch |
| 83 | movsxd rcx, DWORD PTR arg(4) ;output_height |
| 84 | %endm |
| 85 | |
| 86 | %macro APPLY_FILTER_8 1 |
| 87 | punpcklbw xmm0, xmm5 |
| 88 | punpcklbw xmm1, xmm5 |
| 89 | |
| 90 | pmullw xmm0, xmm6 |
| 91 | pmullw xmm1, xmm7 |
| 92 | paddsw xmm0, xmm1 |
| 93 | paddsw xmm0, xmm4 ;rounding |
| 94 | psraw xmm0, 7 ;shift |
| 95 | packuswb xmm0, xmm0 ;pack back to byte |
| 96 | %if %1 |
| 97 | movq xmm1, [rdi] |
| 98 | pavgb xmm0, xmm1 |
| 99 | %endif |
| 100 | movq [rdi], xmm0 ;store the result |
| 101 | |
| 102 | lea rsi, [rsi + rax] |
| 103 | lea rdi, [rdi + rdx] |
| 104 | dec rcx |
| 105 | %endm |
| 106 | |
| 107 | %macro APPLY_FILTER_16 1 |
| 108 | punpcklbw xmm0, xmm5 |
| 109 | punpcklbw xmm1, xmm5 |
| 110 | punpckhbw xmm2, xmm5 |
| 111 | punpckhbw xmm3, xmm5 |
| 112 | |
| 113 | pmullw xmm0, xmm6 |
| 114 | pmullw xmm1, xmm7 |
| 115 | pmullw xmm2, xmm6 |
| 116 | pmullw xmm3, xmm7 |
| 117 | |
| 118 | paddsw xmm0, xmm1 |
| 119 | paddsw xmm2, xmm3 |
| 120 | |
| 121 | paddsw xmm0, xmm4 ;rounding |
| 122 | paddsw xmm2, xmm4 |
| 123 | psraw xmm0, 7 ;shift |
| 124 | psraw xmm2, 7 |
| 125 | packuswb xmm0, xmm2 ;pack back to byte |
| 126 | %if %1 |
| 127 | movdqu xmm1, [rdi] |
| 128 | pavgb xmm0, xmm1 |
| 129 | %endif |
| 130 | movdqu [rdi], xmm0 ;store the result |
| 131 | |
| 132 | lea rsi, [rsi + rax] |
| 133 | lea rdi, [rdi + rdx] |
| 134 | dec rcx |
| 135 | %endm |
| 136 | |
Yaowu Xu | f883b42 | 2016-08-30 14:01:10 -0700 | [diff] [blame] | 137 | global sym(aom_filter_block1d4_v2_sse2) PRIVATE |
| 138 | sym(aom_filter_block1d4_v2_sse2): |
Yaowu Xu | c27fc14 | 2016-08-22 16:08:15 -0700 | [diff] [blame] | 139 | push rbp |
| 140 | mov rbp, rsp |
| 141 | SHADOW_ARGS_TO_STACK 6 |
| 142 | push rsi |
| 143 | push rdi |
| 144 | ; end prolog |
| 145 | |
| 146 | GET_PARAM_4 |
| 147 | .loop: |
| 148 | movd xmm0, [rsi] ;load src |
| 149 | movd xmm1, [rsi + rax] |
| 150 | |
| 151 | APPLY_FILTER_4 0 |
| 152 | jnz .loop |
| 153 | |
| 154 | ; begin epilog |
| 155 | pop rdi |
| 156 | pop rsi |
| 157 | UNSHADOW_ARGS |
| 158 | pop rbp |
| 159 | ret |
| 160 | |
Yaowu Xu | f883b42 | 2016-08-30 14:01:10 -0700 | [diff] [blame] | 161 | global sym(aom_filter_block1d8_v2_sse2) PRIVATE |
| 162 | sym(aom_filter_block1d8_v2_sse2): |
Yaowu Xu | c27fc14 | 2016-08-22 16:08:15 -0700 | [diff] [blame] | 163 | push rbp |
| 164 | mov rbp, rsp |
| 165 | SHADOW_ARGS_TO_STACK 6 |
| 166 | SAVE_XMM 7 |
| 167 | push rsi |
| 168 | push rdi |
| 169 | ; end prolog |
| 170 | |
| 171 | GET_PARAM |
| 172 | .loop: |
| 173 | movq xmm0, [rsi] ;0 |
| 174 | movq xmm1, [rsi + rax] ;1 |
| 175 | |
| 176 | APPLY_FILTER_8 0 |
| 177 | jnz .loop |
| 178 | |
| 179 | ; begin epilog |
| 180 | pop rdi |
| 181 | pop rsi |
| 182 | RESTORE_XMM |
| 183 | UNSHADOW_ARGS |
| 184 | pop rbp |
| 185 | ret |
| 186 | |
Yaowu Xu | f883b42 | 2016-08-30 14:01:10 -0700 | [diff] [blame] | 187 | global sym(aom_filter_block1d16_v2_sse2) PRIVATE |
| 188 | sym(aom_filter_block1d16_v2_sse2): |
Yaowu Xu | c27fc14 | 2016-08-22 16:08:15 -0700 | [diff] [blame] | 189 | push rbp |
| 190 | mov rbp, rsp |
| 191 | SHADOW_ARGS_TO_STACK 6 |
| 192 | SAVE_XMM 7 |
| 193 | push rsi |
| 194 | push rdi |
| 195 | ; end prolog |
| 196 | |
| 197 | GET_PARAM |
| 198 | .loop: |
| 199 | movdqu xmm0, [rsi] ;0 |
| 200 | movdqu xmm1, [rsi + rax] ;1 |
| 201 | movdqa xmm2, xmm0 |
| 202 | movdqa xmm3, xmm1 |
| 203 | |
| 204 | APPLY_FILTER_16 0 |
| 205 | jnz .loop |
| 206 | |
| 207 | ; begin epilog |
| 208 | pop rdi |
| 209 | pop rsi |
| 210 | RESTORE_XMM |
| 211 | UNSHADOW_ARGS |
| 212 | pop rbp |
| 213 | ret |
| 214 | |
Yaowu Xu | f883b42 | 2016-08-30 14:01:10 -0700 | [diff] [blame] | 215 | global sym(aom_filter_block1d4_v2_avg_sse2) PRIVATE |
| 216 | sym(aom_filter_block1d4_v2_avg_sse2): |
Yaowu Xu | c27fc14 | 2016-08-22 16:08:15 -0700 | [diff] [blame] | 217 | push rbp |
| 218 | mov rbp, rsp |
| 219 | SHADOW_ARGS_TO_STACK 6 |
| 220 | push rsi |
| 221 | push rdi |
| 222 | ; end prolog |
| 223 | |
| 224 | GET_PARAM_4 |
| 225 | .loop: |
| 226 | movd xmm0, [rsi] ;load src |
| 227 | movd xmm1, [rsi + rax] |
| 228 | |
| 229 | APPLY_FILTER_4 1 |
| 230 | jnz .loop |
| 231 | |
| 232 | ; begin epilog |
| 233 | pop rdi |
| 234 | pop rsi |
| 235 | UNSHADOW_ARGS |
| 236 | pop rbp |
| 237 | ret |
| 238 | |
Yaowu Xu | f883b42 | 2016-08-30 14:01:10 -0700 | [diff] [blame] | 239 | global sym(aom_filter_block1d8_v2_avg_sse2) PRIVATE |
| 240 | sym(aom_filter_block1d8_v2_avg_sse2): |
Yaowu Xu | c27fc14 | 2016-08-22 16:08:15 -0700 | [diff] [blame] | 241 | push rbp |
| 242 | mov rbp, rsp |
| 243 | SHADOW_ARGS_TO_STACK 6 |
| 244 | SAVE_XMM 7 |
| 245 | push rsi |
| 246 | push rdi |
| 247 | ; end prolog |
| 248 | |
| 249 | GET_PARAM |
| 250 | .loop: |
| 251 | movq xmm0, [rsi] ;0 |
| 252 | movq xmm1, [rsi + rax] ;1 |
| 253 | |
| 254 | APPLY_FILTER_8 1 |
| 255 | jnz .loop |
| 256 | |
| 257 | ; begin epilog |
| 258 | pop rdi |
| 259 | pop rsi |
| 260 | RESTORE_XMM |
| 261 | UNSHADOW_ARGS |
| 262 | pop rbp |
| 263 | ret |
| 264 | |
Yaowu Xu | f883b42 | 2016-08-30 14:01:10 -0700 | [diff] [blame] | 265 | global sym(aom_filter_block1d16_v2_avg_sse2) PRIVATE |
| 266 | sym(aom_filter_block1d16_v2_avg_sse2): |
Yaowu Xu | c27fc14 | 2016-08-22 16:08:15 -0700 | [diff] [blame] | 267 | push rbp |
| 268 | mov rbp, rsp |
| 269 | SHADOW_ARGS_TO_STACK 6 |
| 270 | SAVE_XMM 7 |
| 271 | push rsi |
| 272 | push rdi |
| 273 | ; end prolog |
| 274 | |
| 275 | GET_PARAM |
| 276 | .loop: |
| 277 | movdqu xmm0, [rsi] ;0 |
| 278 | movdqu xmm1, [rsi + rax] ;1 |
| 279 | movdqa xmm2, xmm0 |
| 280 | movdqa xmm3, xmm1 |
| 281 | |
| 282 | APPLY_FILTER_16 1 |
| 283 | jnz .loop |
| 284 | |
| 285 | ; begin epilog |
| 286 | pop rdi |
| 287 | pop rsi |
| 288 | RESTORE_XMM |
| 289 | UNSHADOW_ARGS |
| 290 | pop rbp |
| 291 | ret |
| 292 | |
Yaowu Xu | f883b42 | 2016-08-30 14:01:10 -0700 | [diff] [blame] | 293 | global sym(aom_filter_block1d4_h2_sse2) PRIVATE |
| 294 | sym(aom_filter_block1d4_h2_sse2): |
Yaowu Xu | c27fc14 | 2016-08-22 16:08:15 -0700 | [diff] [blame] | 295 | push rbp |
| 296 | mov rbp, rsp |
| 297 | SHADOW_ARGS_TO_STACK 6 |
| 298 | push rsi |
| 299 | push rdi |
| 300 | ; end prolog |
| 301 | |
| 302 | GET_PARAM_4 |
| 303 | .loop: |
| 304 | movdqu xmm0, [rsi] ;load src |
| 305 | movdqa xmm1, xmm0 |
| 306 | psrldq xmm1, 1 |
| 307 | |
| 308 | APPLY_FILTER_4 0 |
| 309 | jnz .loop |
| 310 | |
| 311 | ; begin epilog |
| 312 | pop rdi |
| 313 | pop rsi |
| 314 | UNSHADOW_ARGS |
| 315 | pop rbp |
| 316 | ret |
| 317 | |
Yaowu Xu | f883b42 | 2016-08-30 14:01:10 -0700 | [diff] [blame] | 318 | global sym(aom_filter_block1d8_h2_sse2) PRIVATE |
| 319 | sym(aom_filter_block1d8_h2_sse2): |
Yaowu Xu | c27fc14 | 2016-08-22 16:08:15 -0700 | [diff] [blame] | 320 | push rbp |
| 321 | mov rbp, rsp |
| 322 | SHADOW_ARGS_TO_STACK 6 |
| 323 | SAVE_XMM 7 |
| 324 | push rsi |
| 325 | push rdi |
| 326 | ; end prolog |
| 327 | |
| 328 | GET_PARAM |
| 329 | .loop: |
| 330 | movdqu xmm0, [rsi] ;load src |
| 331 | movdqa xmm1, xmm0 |
| 332 | psrldq xmm1, 1 |
| 333 | |
| 334 | APPLY_FILTER_8 0 |
| 335 | jnz .loop |
| 336 | |
| 337 | ; begin epilog |
| 338 | pop rdi |
| 339 | pop rsi |
| 340 | RESTORE_XMM |
| 341 | UNSHADOW_ARGS |
| 342 | pop rbp |
| 343 | ret |
| 344 | |
Yaowu Xu | f883b42 | 2016-08-30 14:01:10 -0700 | [diff] [blame] | 345 | global sym(aom_filter_block1d16_h2_sse2) PRIVATE |
| 346 | sym(aom_filter_block1d16_h2_sse2): |
Yaowu Xu | c27fc14 | 2016-08-22 16:08:15 -0700 | [diff] [blame] | 347 | push rbp |
| 348 | mov rbp, rsp |
| 349 | SHADOW_ARGS_TO_STACK 6 |
| 350 | SAVE_XMM 7 |
| 351 | push rsi |
| 352 | push rdi |
| 353 | ; end prolog |
| 354 | |
| 355 | GET_PARAM |
| 356 | .loop: |
| 357 | movdqu xmm0, [rsi] ;load src |
| 358 | movdqu xmm1, [rsi + 1] |
| 359 | movdqa xmm2, xmm0 |
| 360 | movdqa xmm3, xmm1 |
| 361 | |
| 362 | APPLY_FILTER_16 0 |
| 363 | jnz .loop |
| 364 | |
| 365 | ; begin epilog |
| 366 | pop rdi |
| 367 | pop rsi |
| 368 | RESTORE_XMM |
| 369 | UNSHADOW_ARGS |
| 370 | pop rbp |
| 371 | ret |
| 372 | |
Yaowu Xu | f883b42 | 2016-08-30 14:01:10 -0700 | [diff] [blame] | 373 | global sym(aom_filter_block1d4_h2_avg_sse2) PRIVATE |
| 374 | sym(aom_filter_block1d4_h2_avg_sse2): |
Yaowu Xu | c27fc14 | 2016-08-22 16:08:15 -0700 | [diff] [blame] | 375 | push rbp |
| 376 | mov rbp, rsp |
| 377 | SHADOW_ARGS_TO_STACK 6 |
| 378 | push rsi |
| 379 | push rdi |
| 380 | ; end prolog |
| 381 | |
| 382 | GET_PARAM_4 |
| 383 | .loop: |
| 384 | movdqu xmm0, [rsi] ;load src |
| 385 | movdqa xmm1, xmm0 |
| 386 | psrldq xmm1, 1 |
| 387 | |
| 388 | APPLY_FILTER_4 1 |
| 389 | jnz .loop |
| 390 | |
| 391 | ; begin epilog |
| 392 | pop rdi |
| 393 | pop rsi |
| 394 | UNSHADOW_ARGS |
| 395 | pop rbp |
| 396 | ret |
| 397 | |
Yaowu Xu | f883b42 | 2016-08-30 14:01:10 -0700 | [diff] [blame] | 398 | global sym(aom_filter_block1d8_h2_avg_sse2) PRIVATE |
| 399 | sym(aom_filter_block1d8_h2_avg_sse2): |
Yaowu Xu | c27fc14 | 2016-08-22 16:08:15 -0700 | [diff] [blame] | 400 | push rbp |
| 401 | mov rbp, rsp |
| 402 | SHADOW_ARGS_TO_STACK 6 |
| 403 | SAVE_XMM 7 |
| 404 | push rsi |
| 405 | push rdi |
| 406 | ; end prolog |
| 407 | |
| 408 | GET_PARAM |
| 409 | .loop: |
| 410 | movdqu xmm0, [rsi] ;load src |
| 411 | movdqa xmm1, xmm0 |
| 412 | psrldq xmm1, 1 |
| 413 | |
| 414 | APPLY_FILTER_8 1 |
| 415 | jnz .loop |
| 416 | |
| 417 | ; begin epilog |
| 418 | pop rdi |
| 419 | pop rsi |
| 420 | RESTORE_XMM |
| 421 | UNSHADOW_ARGS |
| 422 | pop rbp |
| 423 | ret |
| 424 | |
Yaowu Xu | f883b42 | 2016-08-30 14:01:10 -0700 | [diff] [blame] | 425 | global sym(aom_filter_block1d16_h2_avg_sse2) PRIVATE |
| 426 | sym(aom_filter_block1d16_h2_avg_sse2): |
Yaowu Xu | c27fc14 | 2016-08-22 16:08:15 -0700 | [diff] [blame] | 427 | push rbp |
| 428 | mov rbp, rsp |
| 429 | SHADOW_ARGS_TO_STACK 6 |
| 430 | SAVE_XMM 7 |
| 431 | push rsi |
| 432 | push rdi |
| 433 | ; end prolog |
| 434 | |
| 435 | GET_PARAM |
| 436 | .loop: |
| 437 | movdqu xmm0, [rsi] ;load src |
| 438 | movdqu xmm1, [rsi + 1] |
| 439 | movdqa xmm2, xmm0 |
| 440 | movdqa xmm3, xmm1 |
| 441 | |
| 442 | APPLY_FILTER_16 1 |
| 443 | jnz .loop |
| 444 | |
| 445 | ; begin epilog |
| 446 | pop rdi |
| 447 | pop rsi |
| 448 | RESTORE_XMM |
| 449 | UNSHADOW_ARGS |
| 450 | pop rbp |
| 451 | ret |