| OLD | NEW |
|---|
| (Empty) |
| 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 | |
| 12 %include "vpx_ports/x86_abi_support.asm" | |
| 13 | |
| 14 %define xmm_filter_shift 7 | |
| 15 | |
| 16 ;void vp9_filter_block2d_bil_var_sse2 | |
| 17 ;( | |
| 18 ; unsigned char *ref_ptr, | |
| 19 ; int ref_pixels_per_line, | |
| 20 ; unsigned char *src_ptr, | |
| 21 ; int src_pixels_per_line, | |
| 22 ; unsigned int Height, | |
| 23 ; int xoffset, | |
| 24 ; int yoffset, | |
| 25 ; int *sum, | |
| 26 ; unsigned int *sumsquared;; | |
| 27 ; | |
| 28 ;) | |
| 29 global sym(vp9_filter_block2d_bil_var_sse2) PRIVATE | |
| 30 sym(vp9_filter_block2d_bil_var_sse2): | |
| 31 push rbp | |
| 32 mov rbp, rsp | |
| 33 SHADOW_ARGS_TO_STACK 9 | |
| 34 SAVE_XMM 7 | |
| 35 GET_GOT rbx | |
| 36 push rsi | |
| 37 push rdi | |
| 38 push rbx | |
| 39 ; end prolog | |
| 40 | |
| 41 pxor xmm6, xmm6 ; | |
| 42 pxor xmm7, xmm7 ; | |
| 43 | |
| 44 lea rsi, [GLOBAL(xmm_bi_rd)] ; rounding | |
| 45 movdqa xmm4, XMMWORD PTR [rsi] | |
| 46 | |
| 47 lea rcx, [GLOBAL(bilinear_filters_sse2)] | |
| 48 movsxd rax, dword ptr arg(5) ; xoffset | |
| 49 | |
| 50 cmp rax, 0 ; skip first_pass f
ilter if xoffset=0 | |
| 51 je filter_block2d_bil_var_sse2_sp_only | |
| 52 | |
| 53 shl rax, 5 ; point to filter c
oeff with xoffset | |
| 54 lea rax, [rax + rcx] ; HFilter | |
| 55 | |
| 56 movsxd rdx, dword ptr arg(6) ; yoffset | |
| 57 | |
| 58 cmp rdx, 0 ; skip second_pass
filter if yoffset=0 | |
| 59 je filter_block2d_bil_var_sse2_fp_only | |
| 60 | |
| 61 shl rdx, 5 | |
| 62 lea rdx, [rdx + rcx] ; VFilter | |
| 63 | |
| 64 mov rsi, arg(0) ;ref_ptr | |
| 65 mov rdi, arg(2) ;src_ptr | |
| 66 movsxd rcx, dword ptr arg(4) ;Height | |
| 67 | |
| 68 pxor xmm0, xmm0 ; | |
| 69 movq xmm1, QWORD PTR [rsi] ; | |
| 70 movq xmm3, QWORD PTR [rsi+1] ; | |
| 71 | |
| 72 punpcklbw xmm1, xmm0 ; | |
| 73 pmullw xmm1, [rax] ; | |
| 74 punpcklbw xmm3, xmm0 | |
| 75 pmullw xmm3, [rax+16] ; | |
| 76 | |
| 77 paddw xmm1, xmm3 ; | |
| 78 paddw xmm1, xmm4 ; | |
| 79 psraw xmm1, xmm_filter_shift ; | |
| 80 movdqa xmm5, xmm1 | |
| 81 | |
| 82 movsxd rbx, dword ptr arg(1) ;ref_pixels_per_line | |
| 83 lea rsi, [rsi + rbx] | |
| 84 %if ABI_IS_32BIT=0 | |
| 85 movsxd r9, dword ptr arg(3) ;src_pixels_per_line | |
| 86 %endif | |
| 87 | |
| 88 filter_block2d_bil_var_sse2_loop: | |
| 89 movq xmm1, QWORD PTR [rsi] ; | |
| 90 movq xmm3, QWORD PTR [rsi+1] ; | |
| 91 | |
| 92 punpcklbw xmm1, xmm0 ; | |
| 93 pmullw xmm1, [rax] ; | |
| 94 punpcklbw xmm3, xmm0 ; | |
| 95 pmullw xmm3, [rax+16] ; | |
| 96 | |
| 97 paddw xmm1, xmm3 ; | |
| 98 paddw xmm1, xmm4 ; | |
| 99 psraw xmm1, xmm_filter_shift ; | |
| 100 | |
| 101 movdqa xmm3, xmm5 ; | |
| 102 movdqa xmm5, xmm1 ; | |
| 103 | |
| 104 pmullw xmm3, [rdx] ; | |
| 105 pmullw xmm1, [rdx+16] ; | |
| 106 paddw xmm1, xmm3 ; | |
| 107 paddw xmm1, xmm4 ; | |
| 108 psraw xmm1, xmm_filter_shift ; | |
| 109 | |
| 110 movq xmm3, QWORD PTR [rdi] ; | |
| 111 punpcklbw xmm3, xmm0 ; | |
| 112 | |
| 113 psubw xmm1, xmm3 ; | |
| 114 paddw xmm6, xmm1 ; | |
| 115 | |
| 116 pmaddwd xmm1, xmm1 ; | |
| 117 paddd xmm7, xmm1 ; | |
| 118 | |
| 119 lea rsi, [rsi + rbx] ;ref_pixels_per_lin
e | |
| 120 %if ABI_IS_32BIT | |
| 121 add rdi, dword ptr arg(3) ;src_pixels_per_lin
e | |
| 122 %else | |
| 123 lea rdi, [rdi + r9] | |
| 124 %endif | |
| 125 | |
| 126 sub rcx, 1 ; | |
| 127 jnz filter_block2d_bil_var_sse2_loop ; | |
| 128 | |
| 129 jmp filter_block2d_bil_variance | |
| 130 | |
| 131 filter_block2d_bil_var_sse2_sp_only: | |
| 132 movsxd rdx, dword ptr arg(6) ; yoffset | |
| 133 | |
| 134 cmp rdx, 0 ; skip all if both
xoffset=0 and yoffset=0 | |
| 135 je filter_block2d_bil_var_sse2_full_pixel | |
| 136 | |
| 137 shl rdx, 5 | |
| 138 lea rdx, [rdx + rcx] ; VFilter | |
| 139 | |
| 140 mov rsi, arg(0) ;ref_ptr | |
| 141 mov rdi, arg(2) ;src_ptr | |
| 142 movsxd rcx, dword ptr arg(4) ;Height | |
| 143 movsxd rax, dword ptr arg(1) ;ref_pixels_per_lin
e | |
| 144 | |
| 145 pxor xmm0, xmm0 ; | |
| 146 movq xmm1, QWORD PTR [rsi] ; | |
| 147 punpcklbw xmm1, xmm0 ; | |
| 148 | |
| 149 movsxd rbx, dword ptr arg(3) ;src_pixels_per_lin
e | |
| 150 lea rsi, [rsi + rax] | |
| 151 | |
| 152 filter_block2d_bil_sp_only_loop: | |
| 153 movq xmm3, QWORD PTR [rsi] ; | |
| 154 punpcklbw xmm3, xmm0 ; | |
| 155 movdqa xmm5, xmm3 | |
| 156 | |
| 157 pmullw xmm1, [rdx] ; | |
| 158 pmullw xmm3, [rdx+16] ; | |
| 159 paddw xmm1, xmm3 ; | |
| 160 paddw xmm1, xmm4 ; | |
| 161 psraw xmm1, xmm_filter_shift ; | |
| 162 | |
| 163 movq xmm3, QWORD PTR [rdi] ; | |
| 164 punpcklbw xmm3, xmm0 ; | |
| 165 | |
| 166 psubw xmm1, xmm3 ; | |
| 167 paddw xmm6, xmm1 ; | |
| 168 | |
| 169 pmaddwd xmm1, xmm1 ; | |
| 170 paddd xmm7, xmm1 ; | |
| 171 | |
| 172 movdqa xmm1, xmm5 ; | |
| 173 lea rsi, [rsi + rax] ;ref_pixels_per_lin
e | |
| 174 lea rdi, [rdi + rbx] ;src_pixels_per_lin
e | |
| 175 | |
| 176 sub rcx, 1 ; | |
| 177 jnz filter_block2d_bil_sp_only_loop ; | |
| 178 | |
| 179 jmp filter_block2d_bil_variance | |
| 180 | |
| 181 filter_block2d_bil_var_sse2_full_pixel: | |
| 182 mov rsi, arg(0) ;ref_ptr | |
| 183 mov rdi, arg(2) ;src_ptr | |
| 184 movsxd rcx, dword ptr arg(4) ;Height | |
| 185 movsxd rax, dword ptr arg(1) ;ref_pixels_per_lin
e | |
| 186 movsxd rbx, dword ptr arg(3) ;src_pixels_per_lin
e | |
| 187 pxor xmm0, xmm0 ; | |
| 188 | |
| 189 filter_block2d_bil_full_pixel_loop: | |
| 190 movq xmm1, QWORD PTR [rsi] ; | |
| 191 punpcklbw xmm1, xmm0 ; | |
| 192 | |
| 193 movq xmm2, QWORD PTR [rdi] ; | |
| 194 punpcklbw xmm2, xmm0 ; | |
| 195 | |
| 196 psubw xmm1, xmm2 ; | |
| 197 paddw xmm6, xmm1 ; | |
| 198 | |
| 199 pmaddwd xmm1, xmm1 ; | |
| 200 paddd xmm7, xmm1 ; | |
| 201 | |
| 202 lea rsi, [rsi + rax] ;ref_pixels_per_lin
e | |
| 203 lea rdi, [rdi + rbx] ;src_pixels_per_lin
e | |
| 204 | |
| 205 sub rcx, 1 ; | |
| 206 jnz filter_block2d_bil_full_pixel_loop ; | |
| 207 | |
| 208 jmp filter_block2d_bil_variance | |
| 209 | |
| 210 filter_block2d_bil_var_sse2_fp_only: | |
| 211 mov rsi, arg(0) ;ref_ptr | |
| 212 mov rdi, arg(2) ;src_ptr | |
| 213 movsxd rcx, dword ptr arg(4) ;Height | |
| 214 movsxd rdx, dword ptr arg(1) ;ref_pixels_per_lin
e | |
| 215 | |
| 216 pxor xmm0, xmm0 ; | |
| 217 movsxd rbx, dword ptr arg(3) ;src_pixels_per_lin
e | |
| 218 | |
| 219 filter_block2d_bil_fp_only_loop: | |
| 220 movq xmm1, QWORD PTR [rsi] ; | |
| 221 movq xmm3, QWORD PTR [rsi+1] ; | |
| 222 | |
| 223 punpcklbw xmm1, xmm0 ; | |
| 224 pmullw xmm1, [rax] ; | |
| 225 punpcklbw xmm3, xmm0 ; | |
| 226 pmullw xmm3, [rax+16] ; | |
| 227 | |
| 228 paddw xmm1, xmm3 ; | |
| 229 paddw xmm1, xmm4 ; | |
| 230 psraw xmm1, xmm_filter_shift ; | |
| 231 | |
| 232 movq xmm3, QWORD PTR [rdi] ; | |
| 233 punpcklbw xmm3, xmm0 ; | |
| 234 | |
| 235 psubw xmm1, xmm3 ; | |
| 236 paddw xmm6, xmm1 ; | |
| 237 | |
| 238 pmaddwd xmm1, xmm1 ; | |
| 239 paddd xmm7, xmm1 ; | |
| 240 lea rsi, [rsi + rdx] | |
| 241 lea rdi, [rdi + rbx] ;src_pixels_per_lin
e | |
| 242 | |
| 243 sub rcx, 1 ; | |
| 244 jnz filter_block2d_bil_fp_only_loop ; | |
| 245 | |
| 246 jmp filter_block2d_bil_variance | |
| 247 | |
| 248 filter_block2d_bil_variance: | |
| 249 movdq2q mm6, xmm6 ; | |
| 250 movdq2q mm7, xmm7 ; | |
| 251 | |
| 252 psrldq xmm6, 8 | |
| 253 psrldq xmm7, 8 | |
| 254 | |
| 255 movdq2q mm2, xmm6 | |
| 256 movdq2q mm3, xmm7 | |
| 257 | |
| 258 paddw mm6, mm2 | |
| 259 paddd mm7, mm3 | |
| 260 | |
| 261 pxor mm3, mm3 ; | |
| 262 pxor mm2, mm2 ; | |
| 263 | |
| 264 punpcklwd mm2, mm6 ; | |
| 265 punpckhwd mm3, mm6 ; | |
| 266 | |
| 267 paddd mm2, mm3 ; | |
| 268 movq mm6, mm2 ; | |
| 269 | |
| 270 psrlq mm6, 32 ; | |
| 271 paddd mm2, mm6 ; | |
| 272 | |
| 273 psrad mm2, 16 ; | |
| 274 movq mm4, mm7 ; | |
| 275 | |
| 276 psrlq mm4, 32 ; | |
| 277 paddd mm4, mm7 ; | |
| 278 | |
| 279 mov rsi, arg(7) ; sum | |
| 280 mov rdi, arg(8) ; sumsquared | |
| 281 | |
| 282 movd [rsi], mm2 ; xsum | |
| 283 movd [rdi], mm4 ; xxsum | |
| 284 | |
| 285 ; begin epilog | |
| 286 pop rbx | |
| 287 pop rdi | |
| 288 pop rsi | |
| 289 RESTORE_GOT | |
| 290 RESTORE_XMM | |
| 291 UNSHADOW_ARGS | |
| 292 pop rbp | |
| 293 ret | |
| 294 | |
| 295 | |
| 296 | |
| 297 ;void vp9_half_horiz_vert_variance16x_h_sse2 | |
| 298 ;( | |
| 299 ; unsigned char *ref_ptr, | |
| 300 ; int ref_pixels_per_line, | |
| 301 ; unsigned char *src_ptr, | |
| 302 ; int src_pixels_per_line, | |
| 303 ; unsigned int Height, | |
| 304 ; int *sum, | |
| 305 ; unsigned int *sumsquared | |
| 306 ;) | |
| 307 global sym(vp9_half_horiz_vert_variance16x_h_sse2) PRIVATE | |
| 308 sym(vp9_half_horiz_vert_variance16x_h_sse2): | |
| 309 push rbp | |
| 310 mov rbp, rsp | |
| 311 SHADOW_ARGS_TO_STACK 7 | |
| 312 SAVE_XMM 7 | |
| 313 GET_GOT rbx | |
| 314 push rsi | |
| 315 push rdi | |
| 316 ; end prolog | |
| 317 | |
| 318 pxor xmm6, xmm6 ; error accumulator | |
| 319 pxor xmm7, xmm7 ; sse eaccumulator | |
| 320 mov rsi, arg(0) ;ref_ptr ; | |
| 321 | |
| 322 mov rdi, arg(2) ;src_ptr ; | |
| 323 movsxd rcx, dword ptr arg(4) ;Height ; | |
| 324 movsxd rax, dword ptr arg(1) ;ref_pixels_per_line | |
| 325 movsxd rdx, dword ptr arg(3) ;src_pixels_per_line | |
| 326 | |
| 327 pxor xmm0, xmm0 ; | |
| 328 | |
| 329 movdqu xmm5, XMMWORD PTR [rsi] | |
| 330 movdqu xmm3, XMMWORD PTR [rsi+1] | |
| 331 pavgb xmm5, xmm3 ; xmm5 = avg(xmm1,x
mm3) horizontal line 1 | |
| 332 | |
| 333 lea rsi, [rsi + rax] | |
| 334 | |
| 335 .half_horiz_vert_variance16x_h_1: | |
| 336 movdqu xmm1, XMMWORD PTR [rsi] ; | |
| 337 movdqu xmm2, XMMWORD PTR [rsi+1] ; | |
| 338 pavgb xmm1, xmm2 ; xmm1 = avg(xmm1,x
mm3) horizontal line i+1 | |
| 339 | |
| 340 pavgb xmm5, xmm1 ; xmm = vertical av
erage of the above | |
| 341 | |
| 342 movdqa xmm4, xmm5 | |
| 343 punpcklbw xmm5, xmm0 ; xmm5 = words of a
bove | |
| 344 punpckhbw xmm4, xmm0 | |
| 345 | |
| 346 movq xmm3, QWORD PTR [rdi] ; xmm3 = d0,d1,d2..
d7 | |
| 347 punpcklbw xmm3, xmm0 ; xmm3 = words of a
bove | |
| 348 psubw xmm5, xmm3 ; xmm5 -= xmm3 | |
| 349 | |
| 350 movq xmm3, QWORD PTR [rdi+8] | |
| 351 punpcklbw xmm3, xmm0 | |
| 352 psubw xmm4, xmm3 | |
| 353 | |
| 354 paddw xmm6, xmm5 ; xmm6 += accumulat
ed column differences | |
| 355 paddw xmm6, xmm4 | |
| 356 pmaddwd xmm5, xmm5 ; xmm5 *= xmm5 | |
| 357 pmaddwd xmm4, xmm4 | |
| 358 paddd xmm7, xmm5 ; xmm7 += accumulat
ed square column differences | |
| 359 paddd xmm7, xmm4 | |
| 360 | |
| 361 movdqa xmm5, xmm1 ; save xmm1 for use
on the next row | |
| 362 | |
| 363 lea rsi, [rsi + rax] | |
| 364 lea rdi, [rdi + rdx] | |
| 365 | |
| 366 sub rcx, 1 ; | |
| 367 jnz .half_horiz_vert_variance16x_h_1 ; | |
| 368 | |
| 369 pxor xmm1, xmm1 | |
| 370 pxor xmm5, xmm5 | |
| 371 | |
| 372 punpcklwd xmm0, xmm6 | |
| 373 punpckhwd xmm1, xmm6 | |
| 374 psrad xmm0, 16 | |
| 375 psrad xmm1, 16 | |
| 376 paddd xmm0, xmm1 | |
| 377 movdqa xmm1, xmm0 | |
| 378 | |
| 379 movdqa xmm6, xmm7 | |
| 380 punpckldq xmm6, xmm5 | |
| 381 punpckhdq xmm7, xmm5 | |
| 382 paddd xmm6, xmm7 | |
| 383 | |
| 384 punpckldq xmm0, xmm5 | |
| 385 punpckhdq xmm1, xmm5 | |
| 386 paddd xmm0, xmm1 | |
| 387 | |
| 388 movdqa xmm7, xmm6 | |
| 389 movdqa xmm1, xmm0 | |
| 390 | |
| 391 psrldq xmm7, 8 | |
| 392 psrldq xmm1, 8 | |
| 393 | |
| 394 paddd xmm6, xmm7 | |
| 395 paddd xmm0, xmm1 | |
| 396 | |
| 397 mov rsi, arg(5) ;[Sum] | |
| 398 mov rdi, arg(6) ;[SSE] | |
| 399 | |
| 400 movd [rsi], xmm0 | |
| 401 movd [rdi], xmm6 | |
| 402 | |
| 403 ; begin epilog | |
| 404 pop rdi | |
| 405 pop rsi | |
| 406 RESTORE_GOT | |
| 407 RESTORE_XMM | |
| 408 UNSHADOW_ARGS | |
| 409 pop rbp | |
| 410 ret | |
| 411 | |
| 412 ;void vp9_half_vert_variance16x_h_sse2 | |
| 413 ;( | |
| 414 ; unsigned char *ref_ptr, | |
| 415 ; int ref_pixels_per_line, | |
| 416 ; unsigned char *src_ptr, | |
| 417 ; int src_pixels_per_line, | |
| 418 ; unsigned int Height, | |
| 419 ; int *sum, | |
| 420 ; unsigned int *sumsquared | |
| 421 ;) | |
| 422 global sym(vp9_half_vert_variance16x_h_sse2) PRIVATE | |
| 423 sym(vp9_half_vert_variance16x_h_sse2): | |
| 424 push rbp | |
| 425 mov rbp, rsp | |
| 426 SHADOW_ARGS_TO_STACK 7 | |
| 427 SAVE_XMM 7 | |
| 428 GET_GOT rbx | |
| 429 push rsi | |
| 430 push rdi | |
| 431 ; end prolog | |
| 432 | |
| 433 pxor xmm6, xmm6 ; error accumulator | |
| 434 pxor xmm7, xmm7 ; sse eaccumulator | |
| 435 mov rsi, arg(0) ;ref_ptr | |
| 436 | |
| 437 mov rdi, arg(2) ;src_ptr | |
| 438 movsxd rcx, dword ptr arg(4) ;Height | |
| 439 movsxd rax, dword ptr arg(1) ;ref_pixels_per_line | |
| 440 movsxd rdx, dword ptr arg(3) ;src_pixels_per_line | |
| 441 | |
| 442 movdqu xmm5, XMMWORD PTR [rsi] | |
| 443 lea rsi, [rsi + rax ] | |
| 444 pxor xmm0, xmm0 | |
| 445 | |
| 446 .half_vert_variance16x_h_1: | |
| 447 movdqu xmm3, XMMWORD PTR [rsi] | |
| 448 | |
| 449 pavgb xmm5, xmm3 ; xmm5 = avg(xmm1,x
mm3) | |
| 450 movdqa xmm4, xmm5 | |
| 451 punpcklbw xmm5, xmm0 | |
| 452 punpckhbw xmm4, xmm0 | |
| 453 | |
| 454 movq xmm2, QWORD PTR [rdi] | |
| 455 punpcklbw xmm2, xmm0 | |
| 456 psubw xmm5, xmm2 | |
| 457 movq xmm2, QWORD PTR [rdi+8] | |
| 458 punpcklbw xmm2, xmm0 | |
| 459 psubw xmm4, xmm2 | |
| 460 | |
| 461 paddw xmm6, xmm5 ; xmm6 += accumulat
ed column differences | |
| 462 paddw xmm6, xmm4 | |
| 463 pmaddwd xmm5, xmm5 ; xmm5 *= xmm5 | |
| 464 pmaddwd xmm4, xmm4 | |
| 465 paddd xmm7, xmm5 ; xmm7 += accumulat
ed square column differences | |
| 466 paddd xmm7, xmm4 | |
| 467 | |
| 468 movdqa xmm5, xmm3 | |
| 469 | |
| 470 lea rsi, [rsi + rax] | |
| 471 lea rdi, [rdi + rdx] | |
| 472 | |
| 473 sub rcx, 1 | |
| 474 jnz .half_vert_variance16x_h_1 | |
| 475 | |
| 476 pxor xmm1, xmm1 | |
| 477 pxor xmm5, xmm5 | |
| 478 | |
| 479 punpcklwd xmm0, xmm6 | |
| 480 punpckhwd xmm1, xmm6 | |
| 481 psrad xmm0, 16 | |
| 482 psrad xmm1, 16 | |
| 483 paddd xmm0, xmm1 | |
| 484 movdqa xmm1, xmm0 | |
| 485 | |
| 486 movdqa xmm6, xmm7 | |
| 487 punpckldq xmm6, xmm5 | |
| 488 punpckhdq xmm7, xmm5 | |
| 489 paddd xmm6, xmm7 | |
| 490 | |
| 491 punpckldq xmm0, xmm5 | |
| 492 punpckhdq xmm1, xmm5 | |
| 493 paddd xmm0, xmm1 | |
| 494 | |
| 495 movdqa xmm7, xmm6 | |
| 496 movdqa xmm1, xmm0 | |
| 497 | |
| 498 psrldq xmm7, 8 | |
| 499 psrldq xmm1, 8 | |
| 500 | |
| 501 paddd xmm6, xmm7 | |
| 502 paddd xmm0, xmm1 | |
| 503 | |
| 504 mov rsi, arg(5) ;[Sum] | |
| 505 mov rdi, arg(6) ;[SSE] | |
| 506 | |
| 507 movd [rsi], xmm0 | |
| 508 movd [rdi], xmm6 | |
| 509 | |
| 510 ; begin epilog | |
| 511 pop rdi | |
| 512 pop rsi | |
| 513 RESTORE_GOT | |
| 514 RESTORE_XMM | |
| 515 UNSHADOW_ARGS | |
| 516 pop rbp | |
| 517 ret | |
| 518 | |
| 519 ;void vp9_half_horiz_variance16x_h_sse2 | |
| 520 ;( | |
| 521 ; unsigned char *ref_ptr, | |
| 522 ; int ref_pixels_per_line, | |
| 523 ; unsigned char *src_ptr, | |
| 524 ; int src_pixels_per_line, | |
| 525 ; unsigned int Height, | |
| 526 ; int *sum, | |
| 527 ; unsigned int *sumsquared | |
| 528 ;) | |
| 529 global sym(vp9_half_horiz_variance16x_h_sse2) PRIVATE | |
| 530 sym(vp9_half_horiz_variance16x_h_sse2): | |
| 531 push rbp | |
| 532 mov rbp, rsp | |
| 533 SHADOW_ARGS_TO_STACK 7 | |
| 534 SAVE_XMM 7 | |
| 535 GET_GOT rbx | |
| 536 push rsi | |
| 537 push rdi | |
| 538 ; end prolog | |
| 539 | |
| 540 pxor xmm6, xmm6 ; error accumulator | |
| 541 pxor xmm7, xmm7 ; sse eaccumulator | |
| 542 mov rsi, arg(0) ;ref_ptr ; | |
| 543 | |
| 544 mov rdi, arg(2) ;src_ptr ; | |
| 545 movsxd rcx, dword ptr arg(4) ;Height ; | |
| 546 movsxd rax, dword ptr arg(1) ;ref_pixels_per_line | |
| 547 movsxd rdx, dword ptr arg(3) ;src_pixels_per_line | |
| 548 | |
| 549 pxor xmm0, xmm0 ; | |
| 550 | |
| 551 .half_horiz_variance16x_h_1: | |
| 552 movdqu xmm5, XMMWORD PTR [rsi] ; xmm5 = s0,s1,s2
..s15 | |
| 553 movdqu xmm3, XMMWORD PTR [rsi+1] ; xmm3 = s1,s2,s3
..s16 | |
| 554 | |
| 555 pavgb xmm5, xmm3 ; xmm5 = avg(xmm1,x
mm3) | |
| 556 movdqa xmm1, xmm5 | |
| 557 punpcklbw xmm5, xmm0 ; xmm5 = words of a
bove | |
| 558 punpckhbw xmm1, xmm0 | |
| 559 | |
| 560 movq xmm3, QWORD PTR [rdi] ; xmm3 = d0,d1,d2..
d7 | |
| 561 punpcklbw xmm3, xmm0 ; xmm3 = words of a
bove | |
| 562 movq xmm2, QWORD PTR [rdi+8] | |
| 563 punpcklbw xmm2, xmm0 | |
| 564 | |
| 565 psubw xmm5, xmm3 ; xmm5 -= xmm3 | |
| 566 psubw xmm1, xmm2 | |
| 567 paddw xmm6, xmm5 ; xmm6 += accumulat
ed column differences | |
| 568 paddw xmm6, xmm1 | |
| 569 pmaddwd xmm5, xmm5 ; xmm5 *= xmm5 | |
| 570 pmaddwd xmm1, xmm1 | |
| 571 paddd xmm7, xmm5 ; xmm7 += accumulat
ed square column differences | |
| 572 paddd xmm7, xmm1 | |
| 573 | |
| 574 lea rsi, [rsi + rax] | |
| 575 lea rdi, [rdi + rdx] | |
| 576 | |
| 577 sub rcx, 1 ; | |
| 578 jnz .half_horiz_variance16x_h_1 ; | |
| 579 | |
| 580 pxor xmm1, xmm1 | |
| 581 pxor xmm5, xmm5 | |
| 582 | |
| 583 punpcklwd xmm0, xmm6 | |
| 584 punpckhwd xmm1, xmm6 | |
| 585 psrad xmm0, 16 | |
| 586 psrad xmm1, 16 | |
| 587 paddd xmm0, xmm1 | |
| 588 movdqa xmm1, xmm0 | |
| 589 | |
| 590 movdqa xmm6, xmm7 | |
| 591 punpckldq xmm6, xmm5 | |
| 592 punpckhdq xmm7, xmm5 | |
| 593 paddd xmm6, xmm7 | |
| 594 | |
| 595 punpckldq xmm0, xmm5 | |
| 596 punpckhdq xmm1, xmm5 | |
| 597 paddd xmm0, xmm1 | |
| 598 | |
| 599 movdqa xmm7, xmm6 | |
| 600 movdqa xmm1, xmm0 | |
| 601 | |
| 602 psrldq xmm7, 8 | |
| 603 psrldq xmm1, 8 | |
| 604 | |
| 605 paddd xmm6, xmm7 | |
| 606 paddd xmm0, xmm1 | |
| 607 | |
| 608 mov rsi, arg(5) ;[Sum] | |
| 609 mov rdi, arg(6) ;[SSE] | |
| 610 | |
| 611 movd [rsi], xmm0 | |
| 612 movd [rdi], xmm6 | |
| 613 | |
| 614 ; begin epilog | |
| 615 pop rdi | |
| 616 pop rsi | |
| 617 RESTORE_GOT | |
| 618 RESTORE_XMM | |
| 619 UNSHADOW_ARGS | |
| 620 pop rbp | |
| 621 ret | |
| 622 | |
| 623 SECTION_RODATA | |
| 624 ; short xmm_bi_rd[8] = { 64, 64, 64, 64,64, 64, 64, 64}; | |
| 625 align 16 | |
| 626 xmm_bi_rd: | |
| 627 times 8 dw 64 | |
| 628 align 16 | |
| 629 bilinear_filters_sse2: | |
| 630 dw 128, 128, 128, 128, 128, 128, 128, 128, 0, 0, 0, 0, 0, 0, 0, 0 | |
| 631 dw 120, 120, 120, 120, 120, 120, 120, 120, 8, 8, 8, 8, 8, 8, 8, 8 | |
| 632 dw 112, 112, 112, 112, 112, 112, 112, 112, 16, 16, 16, 16, 16, 16, 16, 16 | |
| 633 dw 104, 104, 104, 104, 104, 104, 104, 104, 24, 24, 24, 24, 24, 24, 24, 24 | |
| 634 dw 96, 96, 96, 96, 96, 96, 96, 96, 32, 32, 32, 32, 32, 32, 32, 32 | |
| 635 dw 88, 88, 88, 88, 88, 88, 88, 88, 40, 40, 40, 40, 40, 40, 40, 40 | |
| 636 dw 80, 80, 80, 80, 80, 80, 80, 80, 48, 48, 48, 48, 48, 48, 48, 48 | |
| 637 dw 72, 72, 72, 72, 72, 72, 72, 72, 56, 56, 56, 56, 56, 56, 56, 56 | |
| 638 dw 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64 | |
| 639 dw 56, 56, 56, 56, 56, 56, 56, 56, 72, 72, 72, 72, 72, 72, 72, 72 | |
| 640 dw 48, 48, 48, 48, 48, 48, 48, 48, 80, 80, 80, 80, 80, 80, 80, 80 | |
| 641 dw 40, 40, 40, 40, 40, 40, 40, 40, 88, 88, 88, 88, 88, 88, 88, 88 | |
| 642 dw 32, 32, 32, 32, 32, 32, 32, 32, 96, 96, 96, 96, 96, 96, 96, 96 | |
| 643 dw 24, 24, 24, 24, 24, 24, 24, 24, 104, 104, 104, 104, 104, 104, 104, 104 | |
| 644 dw 16, 16, 16, 16, 16, 16, 16, 16, 112, 112, 112, 112, 112, 112, 112, 112 | |
| 645 dw 8, 8, 8, 8, 8, 8, 8, 8, 120, 120, 120, 120, 120, 120, 120, 120 | |
| OLD | NEW |
|---|
Chromium Code Reviews
Issue 17009012:
libvpx: Pull from upstream (Closed)
Base URL: http://src.chromium.org/chrome/trunk/deps/third_party/libvpx/