Parent Directory
| 
Revision Log
Revision 1586 - (view) (download)
| 1 : | edgomez | 1586 | ;/***************************************************************************** |
| 2 : | ; * | ||
| 3 : | ; * XVID MPEG-4 VIDEO CODEC | ||
| 4 : | ; * - mmx 8x8 block-based halfpel interpolation - | ||
| 5 : | ; * | ||
| 6 : | ; * Copyright(C) 2002 Michael Militzer <isibaar@xvid.org> | ||
| 7 : | ; * 2002 Pascal Massimino <skal@planet-d.net> | ||
| 8 : | ; * 2004 Andre Werthmann <wertmann@aei.mpg.de> | ||
| 9 : | ; * | ||
| 10 : | ; * This program is free software ; you can redistribute it and/or modify | ||
| 11 : | ; * it under the terms of the GNU General Public License as published by | ||
| 12 : | ; * the Free Software Foundation ; either version 2 of the License, or | ||
| 13 : | ; * (at your option) any later version. | ||
| 14 : | ; * | ||
| 15 : | ; * This program is distributed in the hope that it will be useful, | ||
| 16 : | ; * but WITHOUT ANY WARRANTY ; without even the implied warranty of | ||
| 17 : | ; * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | ||
| 18 : | ; * GNU General Public License for more details. | ||
| 19 : | ; * | ||
| 20 : | ; * You should have received a copy of the GNU General Public License | ||
| 21 : | ; * along with this program ; if not, write to the Free Software | ||
| 22 : | ; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | ||
| 23 : | ; * | ||
| 24 : | ; ****************************************************************************/ | ||
| 25 : | |||
| 26 : | BITS 64 | ||
| 27 : | |||
| 28 : | %macro cglobal 1 | ||
| 29 : | %ifdef PREFIX | ||
| 30 : | %ifdef MARK_FUNCS | ||
| 31 : | global _%1:function %1.endfunc-%1 | ||
| 32 : | %define %1 _%1:function %1.endfunc-%1 | ||
| 33 : | %else | ||
| 34 : | global _%1 | ||
| 35 : | %define %1 _%1 | ||
| 36 : | %endif | ||
| 37 : | %else | ||
| 38 : | %ifdef MARK_FUNCS | ||
| 39 : | global %1:function %1.endfunc-%1 | ||
| 40 : | %else | ||
| 41 : | global %1 | ||
| 42 : | %endif | ||
| 43 : | %endif | ||
| 44 : | %endmacro | ||
| 45 : | |||
| 46 : | ;============================================================================= | ||
| 47 : | ; Read only data | ||
| 48 : | ;============================================================================= | ||
| 49 : | |||
| 50 : | %ifdef FORMAT_COFF | ||
| 51 : | SECTION .rodata | ||
| 52 : | %else | ||
| 53 : | SECTION .rodata align=16 | ||
| 54 : | %endif | ||
| 55 : | |||
| 56 : | ALIGN 16 | ||
| 57 : | mmx_one: | ||
| 58 : | times 8 db 1 | ||
| 59 : | |||
| 60 : | SECTION .text align=16 | ||
| 61 : | |||
| 62 : | cglobal interpolate8x8_halfpel_h_x86_64 | ||
| 63 : | cglobal interpolate8x8_halfpel_v_x86_64 | ||
| 64 : | cglobal interpolate8x8_halfpel_hv_x86_64 | ||
| 65 : | |||
| 66 : | cglobal interpolate8x8_halfpel_add_x86_64 | ||
| 67 : | cglobal interpolate8x8_halfpel_h_add_x86_64 | ||
| 68 : | cglobal interpolate8x8_halfpel_v_add_x86_64 | ||
| 69 : | cglobal interpolate8x8_halfpel_hv_add_x86_64 | ||
| 70 : | |||
| 71 : | ;=========================================================================== | ||
| 72 : | ; | ||
| 73 : | ; void interpolate8x8_halfpel_h_x86_64(uint8_t * const dst, | ||
| 74 : | ; const uint8_t * const src, | ||
| 75 : | ; const uint32_t stride, | ||
| 76 : | ; const uint32_t rounding); | ||
| 77 : | ; | ||
| 78 : | ;=========================================================================== | ||
| 79 : | |||
| 80 : | %macro COPY_H_SSE_RND0 0 | ||
| 81 : | movq mm0, [rax] | ||
| 82 : | pavgb mm0, [rax+1] | ||
| 83 : | movq mm1, [rax+rdx] | ||
| 84 : | pavgb mm1, [rax+rdx+1] | ||
| 85 : | lea rax,[rax+2*rdx] | ||
| 86 : | movq [rcx],mm0 | ||
| 87 : | movq [rcx+rdx],mm1 | ||
| 88 : | %endmacro | ||
| 89 : | |||
| 90 : | %macro COPY_H_SSE_RND1 0 | ||
| 91 : | movq mm0, [rax] | ||
| 92 : | movq mm1, [rax+rdx] | ||
| 93 : | movq mm4, mm0 | ||
| 94 : | movq mm5, mm1 | ||
| 95 : | movq mm2, [rax+1] | ||
| 96 : | movq mm3, [rax+rdx+1] | ||
| 97 : | pavgb mm0, mm2 | ||
| 98 : | pxor mm2, mm4 | ||
| 99 : | pavgb mm1, mm3 | ||
| 100 : | lea rax, [rax+2*rdx] | ||
| 101 : | pxor mm3, mm5 | ||
| 102 : | pand mm2, mm7 | ||
| 103 : | pand mm3, mm7 | ||
| 104 : | psubb mm0, mm2 | ||
| 105 : | movq [rcx], mm0 | ||
| 106 : | psubb mm1, mm3 | ||
| 107 : | movq [rcx+rdx], mm1 | ||
| 108 : | %endmacro | ||
| 109 : | |||
| 110 : | ALIGN 16 | ||
| 111 : | interpolate8x8_halfpel_h_x86_64: | ||
| 112 : | |||
| 113 : | mov rax, rcx ; rounding | ||
| 114 : | mov rcx, rdi ; Dst | ||
| 115 : | test rax,rax | ||
| 116 : | mov rax, rsi ; src | ||
| 117 : | ; rdx is stride | ||
| 118 : | |||
| 119 : | jnz near .rounding1 | ||
| 120 : | |||
| 121 : | COPY_H_SSE_RND0 | ||
| 122 : | lea rcx,[rcx+2*rdx] | ||
| 123 : | COPY_H_SSE_RND0 | ||
| 124 : | lea rcx,[rcx+2*rdx] | ||
| 125 : | COPY_H_SSE_RND0 | ||
| 126 : | lea rcx,[rcx+2*rdx] | ||
| 127 : | COPY_H_SSE_RND0 | ||
| 128 : | ret | ||
| 129 : | |||
| 130 : | .rounding1 | ||
| 131 : | ; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 | ||
| 132 : | movq mm7, [mmx_one wrt rip] | ||
| 133 : | COPY_H_SSE_RND1 | ||
| 134 : | lea rcx, [rcx+2*rdx] | ||
| 135 : | COPY_H_SSE_RND1 | ||
| 136 : | lea rcx,[rcx+2*rdx] | ||
| 137 : | COPY_H_SSE_RND1 | ||
| 138 : | lea rcx,[rcx+2*rdx] | ||
| 139 : | COPY_H_SSE_RND1 | ||
| 140 : | ret | ||
| 141 : | .endfunc | ||
| 142 : | |||
| 143 : | ;=========================================================================== | ||
| 144 : | ; | ||
| 145 : | ; void interpolate8x8_halfpel_v_x86_64(uint8_t * const dst, | ||
| 146 : | ; const uint8_t * const src, | ||
| 147 : | ; const uint32_t stride, | ||
| 148 : | ; const uint32_t rounding); | ||
| 149 : | ; | ||
| 150 : | ;=========================================================================== | ||
| 151 : | |||
| 152 : | %macro COPY_V_SSE_RND0 0 | ||
| 153 : | movq mm0, [rax] | ||
| 154 : | movq mm1, [rax+rdx] | ||
| 155 : | pavgb mm0, mm1 | ||
| 156 : | pavgb mm1, [rax+2*rdx] | ||
| 157 : | lea rax, [rax+2*rdx] | ||
| 158 : | movq [rcx], mm0 | ||
| 159 : | movq [rcx+rdx],mm1 | ||
| 160 : | %endmacro | ||
| 161 : | |||
| 162 : | %macro COPY_V_SSE_RND1 0 | ||
| 163 : | movq mm0, mm2 | ||
| 164 : | movq mm1, [rax] | ||
| 165 : | movq mm2, [rax+rdx] | ||
| 166 : | lea rax,[rax+2*rdx] | ||
| 167 : | movq mm4, mm0 | ||
| 168 : | movq mm5, mm1 | ||
| 169 : | pavgb mm0, mm1 | ||
| 170 : | pxor mm4, mm1 | ||
| 171 : | pavgb mm1, mm2 | ||
| 172 : | pxor mm5, mm2 | ||
| 173 : | pand mm4, mm7 ; lsb's of (i^j)... | ||
| 174 : | pand mm5, mm7 ; lsb's of (i^j)... | ||
| 175 : | psubb mm0, mm4 ; ...are substracted from result of pavgb | ||
| 176 : | movq [rcx], mm0 | ||
| 177 : | psubb mm1, mm5 ; ...are substracted from result of pavgb | ||
| 178 : | movq [rcx+rdx], mm1 | ||
| 179 : | %endmacro | ||
| 180 : | |||
| 181 : | ALIGN 16 | ||
| 182 : | interpolate8x8_halfpel_v_x86_64: | ||
| 183 : | mov rax, rcx ; rounding | ||
| 184 : | mov rcx, rdi ; Dst | ||
| 185 : | test rax,rax | ||
| 186 : | mov rax, rsi ; Src | ||
| 187 : | ; rdx is stride | ||
| 188 : | |||
| 189 : | ; we process 2 line at a time | ||
| 190 : | jnz near .rounding1 | ||
| 191 : | |||
| 192 : | COPY_V_SSE_RND0 | ||
| 193 : | lea rcx, [rcx+2*rdx] | ||
| 194 : | COPY_V_SSE_RND0 | ||
| 195 : | lea rcx, [rcx+2*rdx] | ||
| 196 : | COPY_V_SSE_RND0 | ||
| 197 : | lea rcx, [rcx+2*rdx] | ||
| 198 : | COPY_V_SSE_RND0 | ||
| 199 : | ret | ||
| 200 : | |||
| 201 : | .rounding1 | ||
| 202 : | ; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 | ||
| 203 : | movq mm7, [mmx_one wrt rip] | ||
| 204 : | movq mm2, [rax] ; loop invariant | ||
| 205 : | add rax, rdx | ||
| 206 : | |||
| 207 : | COPY_V_SSE_RND1 | ||
| 208 : | lea rcx,[rcx+2*rdx] | ||
| 209 : | COPY_V_SSE_RND1 | ||
| 210 : | lea rcx,[rcx+2*rdx] | ||
| 211 : | COPY_V_SSE_RND1 | ||
| 212 : | lea rcx,[rcx+2*rdx] | ||
| 213 : | COPY_V_SSE_RND1 | ||
| 214 : | ret | ||
| 215 : | .endfunc | ||
| 216 : | |||
| 217 : | ;=========================================================================== | ||
| 218 : | ; | ||
| 219 : | ; void interpolate8x8_halfpel_hv_xmm(uint8_t * const dst, | ||
| 220 : | ; const uint8_t * const src, | ||
| 221 : | ; const uint32_t stride, | ||
| 222 : | ; const uint32_t rounding); | ||
| 223 : | ; | ||
| 224 : | ; | ||
| 225 : | ;=========================================================================== | ||
| 226 : | |||
| 227 : | ; The trick is to correct the result of 'pavgb' with some combination of the | ||
| 228 : | ; lsb's of the 4 input values i,j,k,l, and their intermediate 'pavgb' (s and t). | ||
| 229 : | ; The boolean relations are: | ||
| 230 : | ; (i+j+k+l+3)/4 = (s+t+1)/2 - (ij&kl)&st | ||
| 231 : | ; (i+j+k+l+2)/4 = (s+t+1)/2 - (ij|kl)&st | ||
| 232 : | ; (i+j+k+l+1)/4 = (s+t+1)/2 - (ij&kl)|st | ||
| 233 : | ; (i+j+k+l+0)/4 = (s+t+1)/2 - (ij|kl)|st | ||
| 234 : | ; with s=(i+j+1)/2, t=(k+l+1)/2, ij = i^j, kl = k^l, st = s^t. | ||
| 235 : | |||
| 236 : | ; Moreover, we process 2 lines at a times, for better overlapping (~15% faster). | ||
| 237 : | |||
| 238 : | %macro COPY_HV_SSE_RND0 0 | ||
| 239 : | lea rax, [rax+rdx] | ||
| 240 : | |||
| 241 : | movq mm0, [rax] | ||
| 242 : | movq mm1, [rax+1] | ||
| 243 : | |||
| 244 : | movq mm6, mm0 | ||
| 245 : | pavgb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step | ||
| 246 : | lea rax, [rax+rdx] | ||
| 247 : | pxor mm1, mm6 ; mm1=(j^k). preserved for next step | ||
| 248 : | |||
| 249 : | por mm3, mm1 ; ij |= jk | ||
| 250 : | movq mm6, mm2 | ||
| 251 : | pxor mm6, mm0 ; mm6 = s^t | ||
| 252 : | pand mm3, mm6 ; (ij|jk) &= st | ||
| 253 : | pavgb mm2, mm0 ; mm2 = (s+t+1)/2 | ||
| 254 : | pand mm3, mm7 ; mask lsb | ||
| 255 : | psubb mm2, mm3 ; apply. | ||
| 256 : | |||
| 257 : | movq [rcx], mm2 | ||
| 258 : | |||
| 259 : | movq mm2, [rax] | ||
| 260 : | movq mm3, [rax+1] | ||
| 261 : | movq mm6, mm2 | ||
| 262 : | pavgb mm2, mm3 ; preserved for next iteration | ||
| 263 : | lea rcx,[rcx+rdx] | ||
| 264 : | pxor mm3, mm6 ; preserved for next iteration | ||
| 265 : | |||
| 266 : | por mm1, mm3 | ||
| 267 : | movq mm6, mm0 | ||
| 268 : | pxor mm6, mm2 | ||
| 269 : | pand mm1, mm6 | ||
| 270 : | pavgb mm0, mm2 | ||
| 271 : | |||
| 272 : | pand mm1, mm7 | ||
| 273 : | psubb mm0, mm1 | ||
| 274 : | |||
| 275 : | movq [rcx], mm0 | ||
| 276 : | %endmacro | ||
| 277 : | |||
| 278 : | %macro COPY_HV_SSE_RND1 0 | ||
| 279 : | lea rax, [rax+rdx] | ||
| 280 : | |||
| 281 : | movq mm0, [rax] | ||
| 282 : | movq mm1, [rax+1] | ||
| 283 : | |||
| 284 : | movq mm6, mm0 | ||
| 285 : | pavgb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step | ||
| 286 : | lea rax, [rax+rdx] | ||
| 287 : | pxor mm1, mm6 ; mm1=(j^k). preserved for next step | ||
| 288 : | |||
| 289 : | pand mm3, mm1 | ||
| 290 : | movq mm6, mm2 | ||
| 291 : | pxor mm6, mm0 | ||
| 292 : | por mm3, mm6 | ||
| 293 : | pavgb mm2, mm0 | ||
| 294 : | pand mm3, mm7 | ||
| 295 : | psubb mm2, mm3 | ||
| 296 : | |||
| 297 : | movq [rcx], mm2 | ||
| 298 : | |||
| 299 : | movq mm2, [rax] | ||
| 300 : | movq mm3, [rax+1] | ||
| 301 : | movq mm6, mm2 | ||
| 302 : | pavgb mm2, mm3 ; preserved for next iteration | ||
| 303 : | lea rcx,[rcx+rdx] | ||
| 304 : | pxor mm3, mm6 ; preserved for next iteration | ||
| 305 : | |||
| 306 : | pand mm1, mm3 | ||
| 307 : | movq mm6, mm0 | ||
| 308 : | pxor mm6, mm2 | ||
| 309 : | por mm1, mm6 | ||
| 310 : | pavgb mm0, mm2 | ||
| 311 : | pand mm1, mm7 | ||
| 312 : | psubb mm0, mm1 | ||
| 313 : | |||
| 314 : | movq [rcx], mm0 | ||
| 315 : | %endmacro | ||
| 316 : | |||
| 317 : | ALIGN 16 | ||
| 318 : | interpolate8x8_halfpel_hv_x86_64: | ||
| 319 : | mov rax, rcx ; rounding | ||
| 320 : | mov rcx, rdi ; Dst | ||
| 321 : | test rax, rax | ||
| 322 : | mov rax, rsi ; Src | ||
| 323 : | ; rdx is stride | ||
| 324 : | |||
| 325 : | movq mm7, [mmx_one wrt rip] | ||
| 326 : | |||
| 327 : | ; loop invariants: mm2=(i+j+1)/2 and mm3= i^j | ||
| 328 : | movq mm2, [rax] | ||
| 329 : | movq mm3, [rax+1] | ||
| 330 : | movq mm6, mm2 | ||
| 331 : | pavgb mm2, mm3 | ||
| 332 : | pxor mm3, mm6 ; mm2/mm3 ready | ||
| 333 : | |||
| 334 : | jnz near .rounding1 | ||
| 335 : | |||
| 336 : | COPY_HV_SSE_RND0 | ||
| 337 : | add rcx, rdx | ||
| 338 : | COPY_HV_SSE_RND0 | ||
| 339 : | add rcx, rdx | ||
| 340 : | COPY_HV_SSE_RND0 | ||
| 341 : | add rcx, rdx | ||
| 342 : | COPY_HV_SSE_RND0 | ||
| 343 : | ret | ||
| 344 : | |||
| 345 : | .rounding1 | ||
| 346 : | COPY_HV_SSE_RND1 | ||
| 347 : | add rcx, rdx | ||
| 348 : | COPY_HV_SSE_RND1 | ||
| 349 : | add rcx, rdx | ||
| 350 : | COPY_HV_SSE_RND1 | ||
| 351 : | add rcx, rdx | ||
| 352 : | COPY_HV_SSE_RND1 | ||
| 353 : | ret | ||
| 354 : | .endfunc | ||
| 355 : | |||
| 356 : | ;=========================================================================== | ||
| 357 : | ; | ||
| 358 : | ; The next functions combine both source halfpel interpolation step and the | ||
| 359 : | ; averaging (with rouding) step to avoid wasting memory bandwidth computing | ||
| 360 : | ; intermediate halfpel images and then averaging them. | ||
| 361 : | ; | ||
| 362 : | ;=========================================================================== | ||
| 363 : | |||
| 364 : | %macro PROLOG0 0 | ||
| 365 : | ; rcx fourth | ||
| 366 : | ; rdx third | ||
| 367 : | ; r8 fifth | ||
| 368 : | mov r8, rdx ; saves rounding | ||
| 369 : | mov rcx, rdi ; Dst | ||
| 370 : | mov rax, rsi ; Src | ||
| 371 : | ; rdx is stride | ||
| 372 : | %endmacro | ||
| 373 : | %macro PROLOG1 0 | ||
| 374 : | PROLOG0 | ||
| 375 : | test r8, 1; Rounding? | ||
| 376 : | %endmacro | ||
| 377 : | %macro EPILOG 0 | ||
| 378 : | ret | ||
| 379 : | %endmacro | ||
| 380 : | |||
| 381 : | ;=========================================================================== | ||
| 382 : | ; | ||
| 383 : | ; void interpolate8x8_halfpel_add_xmm(uint8_t * const dst, | ||
| 384 : | ; const uint8_t * const src, | ||
| 385 : | ; const uint32_t stride, | ||
| 386 : | ; const uint32_t rounding); | ||
| 387 : | ; | ||
| 388 : | ; | ||
| 389 : | ;=========================================================================== | ||
| 390 : | |||
| 391 : | %macro ADD_FF 2 | ||
| 392 : | movq mm0, [rax+%1] | ||
| 393 : | movq mm1, [rax+%2] | ||
| 394 : | pavgb mm0, [rcx+%1] | ||
| 395 : | pavgb mm1, [rcx+%2] | ||
| 396 : | movq [rcx+%1], mm0 | ||
| 397 : | movq [rcx+%2], mm1 | ||
| 398 : | %endmacro | ||
| 399 : | |||
| 400 : | ALIGN 16 | ||
| 401 : | interpolate8x8_halfpel_add_x86_64: ; 23c | ||
| 402 : | PROLOG1 | ||
| 403 : | ADD_FF 0, rdx | ||
| 404 : | lea rax,[rax+2*rdx] | ||
| 405 : | lea rcx,[rcx+2*rdx] | ||
| 406 : | ADD_FF 0, rdx | ||
| 407 : | lea rax,[rax+2*rdx] | ||
| 408 : | lea rcx,[rcx+2*rdx] | ||
| 409 : | ADD_FF 0, rdx | ||
| 410 : | lea rax,[rax+2*rdx] | ||
| 411 : | lea rcx,[rcx+2*rdx] | ||
| 412 : | ADD_FF 0, rdx | ||
| 413 : | EPILOG | ||
| 414 : | .endfunc | ||
| 415 : | |||
| 416 : | ;=========================================================================== | ||
| 417 : | ; | ||
| 418 : | ; void interpolate8x8_halfpel_h_add_xmm(uint8_t * const dst, | ||
| 419 : | ; const uint8_t * const src, | ||
| 420 : | ; const uint32_t stride, | ||
| 421 : | ; const uint32_t rounding); | ||
| 422 : | ; | ||
| 423 : | ; | ||
| 424 : | ;=========================================================================== | ||
| 425 : | |||
| 426 : | |||
| 427 : | %macro ADD_FH_RND0 2 | ||
| 428 : | movq mm0, [rax+%1] | ||
| 429 : | movq mm1, [rax+%2] | ||
| 430 : | pavgb mm0, [rax+%1+1] | ||
| 431 : | pavgb mm1, [rax+%2+1] | ||
| 432 : | pavgb mm0, [rcx+%1] | ||
| 433 : | pavgb mm1, [rcx+%2] | ||
| 434 : | movq [rcx+%1],mm0 | ||
| 435 : | movq [rcx+%2],mm1 | ||
| 436 : | %endmacro | ||
| 437 : | |||
| 438 : | %macro ADD_FH_RND1 2 | ||
| 439 : | movq mm0, [rax+%1] | ||
| 440 : | movq mm1, [rax+%2] | ||
| 441 : | movq mm4, mm0 | ||
| 442 : | movq mm5, mm1 | ||
| 443 : | movq mm2, [rax+%1+1] | ||
| 444 : | movq mm3, [rax+%2+1] | ||
| 445 : | pavgb mm0, mm2 | ||
| 446 : | ; lea ?? | ||
| 447 : | pxor mm2, mm4 | ||
| 448 : | pavgb mm1, mm3 | ||
| 449 : | pxor mm3, mm5 | ||
| 450 : | pand mm2, [mmx_one wrt rip] | ||
| 451 : | pand mm3, [mmx_one wrt rip] | ||
| 452 : | psubb mm0, mm2 | ||
| 453 : | psubb mm1, mm3 | ||
| 454 : | pavgb mm0, [rcx+%1] | ||
| 455 : | pavgb mm1, [rcx+%2] | ||
| 456 : | movq [rcx+%1],mm0 | ||
| 457 : | movq [rcx+%2],mm1 | ||
| 458 : | %endmacro | ||
| 459 : | |||
| 460 : | ALIGN 16 | ||
| 461 : | interpolate8x8_halfpel_h_add_x86_64: ; 32c | ||
| 462 : | PROLOG1 | ||
| 463 : | jnz near .Loop1 | ||
| 464 : | ADD_FH_RND0 0, rdx | ||
| 465 : | lea rax,[rax+2*rdx] | ||
| 466 : | lea rcx,[rcx+2*rdx] | ||
| 467 : | ADD_FH_RND0 0, rdx | ||
| 468 : | lea rax,[rax+2*rdx] | ||
| 469 : | lea rcx,[rcx+2*rdx] | ||
| 470 : | ADD_FH_RND0 0, rdx | ||
| 471 : | lea rax,[rax+2*rdx] | ||
| 472 : | lea rcx,[rcx+2*rdx] | ||
| 473 : | ADD_FH_RND0 0, rdx | ||
| 474 : | EPILOG | ||
| 475 : | |||
| 476 : | .Loop1 | ||
| 477 : | ; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 | ||
| 478 : | ; movq mm7, [mmx_one wrt rip] | ||
| 479 : | ADD_FH_RND1 0, rdx | ||
| 480 : | lea rax,[rax+2*rdx] | ||
| 481 : | lea rcx,[rcx+2*rdx] | ||
| 482 : | ADD_FH_RND1 0, rdx | ||
| 483 : | lea rax,[rax+2*rdx] | ||
| 484 : | lea rcx,[rcx+2*rdx] | ||
| 485 : | ADD_FH_RND1 0, rdx | ||
| 486 : | lea rax,[rax+2*rdx] | ||
| 487 : | lea rcx,[rcx+2*rdx] | ||
| 488 : | ADD_FH_RND1 0, rdx | ||
| 489 : | EPILOG | ||
| 490 : | .endfunc | ||
| 491 : | |||
| 492 : | |||
| 493 : | ;=========================================================================== | ||
| 494 : | ; | ||
| 495 : | ; void interpolate8x8_halfpel_v_add_x86_64(uint8_t * const dst, | ||
| 496 : | ; const uint8_t * const src, | ||
| 497 : | ; const uint32_t stride, | ||
| 498 : | ; const uint32_t rounding); | ||
| 499 : | ; | ||
| 500 : | ; | ||
| 501 : | ;=========================================================================== | ||
| 502 : | |||
| 503 : | %macro ADD_8_HF_RND0 0 | ||
| 504 : | movq mm0, [rax] | ||
| 505 : | movq mm1, [rax+rdx] | ||
| 506 : | pavgb mm0, mm1 | ||
| 507 : | pavgb mm1, [rax+2*rdx] | ||
| 508 : | lea rax,[rax+2*rdx] | ||
| 509 : | pavgb mm0, [rcx] | ||
| 510 : | pavgb mm1, [rcx+rdx] | ||
| 511 : | movq [rcx],mm0 | ||
| 512 : | movq [rcx+rdx],mm1 | ||
| 513 : | %endmacro | ||
| 514 : | |||
| 515 : | %macro ADD_8_HF_RND1 0 | ||
| 516 : | movq mm1, [rax+rdx] | ||
| 517 : | movq mm2, [rax+2*rdx] | ||
| 518 : | lea rax,[rax+2*rdx] | ||
| 519 : | movq mm4, mm0 | ||
| 520 : | movq mm5, mm1 | ||
| 521 : | pavgb mm0, mm1 | ||
| 522 : | pxor mm4, mm1 | ||
| 523 : | pavgb mm1, mm2 | ||
| 524 : | pxor mm5, mm2 | ||
| 525 : | pand mm4, mm7 ; lsb's of (i^j)... | ||
| 526 : | pand mm5, mm7 ; lsb's of (i^j)... | ||
| 527 : | psubb mm0, mm4 ; ...are substracted from result of pavgb | ||
| 528 : | pavgb mm0, [rcx] | ||
| 529 : | movq [rcx], mm0 | ||
| 530 : | psubb mm1, mm5 ; ...are substracted from result of pavgb | ||
| 531 : | pavgb mm1, [rcx+rdx] | ||
| 532 : | movq [rcx+rdx], mm1 | ||
| 533 : | %endmacro | ||
| 534 : | |||
| 535 : | ALIGN 16 | ||
| 536 : | interpolate8x8_halfpel_v_add_x86_64: | ||
| 537 : | PROLOG1 | ||
| 538 : | |||
| 539 : | jnz near .Loop1 | ||
| 540 : | pxor mm7, mm7 ; this is a NOP | ||
| 541 : | |||
| 542 : | ADD_8_HF_RND0 | ||
| 543 : | lea rcx,[rcx+2*rdx] | ||
| 544 : | ADD_8_HF_RND0 | ||
| 545 : | lea rcx,[rcx+2*rdx] | ||
| 546 : | ADD_8_HF_RND0 | ||
| 547 : | lea rcx,[rcx+2*rdx] | ||
| 548 : | ADD_8_HF_RND0 | ||
| 549 : | EPILOG | ||
| 550 : | |||
| 551 : | .Loop1 | ||
| 552 : | movq mm0, [rax] ; loop invariant | ||
| 553 : | movq mm7, [mmx_one wrt rip] | ||
| 554 : | |||
| 555 : | ADD_8_HF_RND1 | ||
| 556 : | movq mm0, mm2 | ||
| 557 : | lea rcx,[rcx+2*rdx] | ||
| 558 : | ADD_8_HF_RND1 | ||
| 559 : | movq mm0, mm2 | ||
| 560 : | lea rcx,[rcx+2*rdx] | ||
| 561 : | ADD_8_HF_RND1 | ||
| 562 : | movq mm0, mm2 | ||
| 563 : | lea rcx,[rcx+2*rdx] | ||
| 564 : | ADD_8_HF_RND1 | ||
| 565 : | EPILOG | ||
| 566 : | .endfunc | ||
| 567 : | |||
| 568 : | ; The trick is to correct the result of 'pavgb' with some combination of the | ||
| 569 : | ; lsb's of the 4 input values i,j,k,l, and their intermediate 'pavgb' (s and t). | ||
| 570 : | ; The boolean relations are: | ||
| 571 : | ; (i+j+k+l+3)/4 = (s+t+1)/2 - (ij&kl)&st | ||
| 572 : | ; (i+j+k+l+2)/4 = (s+t+1)/2 - (ij|kl)&st | ||
| 573 : | ; (i+j+k+l+1)/4 = (s+t+1)/2 - (ij&kl)|st | ||
| 574 : | ; (i+j+k+l+0)/4 = (s+t+1)/2 - (ij|kl)|st | ||
| 575 : | ; with s=(i+j+1)/2, t=(k+l+1)/2, ij = i^j, kl = k^l, st = s^t. | ||
| 576 : | |||
| 577 : | ; Moreover, we process 2 lines at a times, for better overlapping (~15% faster). | ||
| 578 : | |||
| 579 : | ;=========================================================================== | ||
| 580 : | ; | ||
| 581 : | ; void interpolate8x8_halfpel_hv_add_x86_64(uint8_t * const dst, | ||
| 582 : | ; const uint8_t * const src, | ||
| 583 : | ; const uint32_t stride, | ||
| 584 : | ; const uint32_t rounding); | ||
| 585 : | ; | ||
| 586 : | ; | ||
| 587 : | ;=========================================================================== | ||
| 588 : | |||
| 589 : | %macro ADD_HH_RND0 0 | ||
| 590 : | lea rax,[rax+rdx] | ||
| 591 : | |||
| 592 : | movq mm0, [rax] | ||
| 593 : | movq mm1, [rax+1] | ||
| 594 : | |||
| 595 : | movq mm6, mm0 | ||
| 596 : | pavgb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step | ||
| 597 : | lea rax,[rax+rdx] | ||
| 598 : | pxor mm1, mm6 ; mm1=(j^k). preserved for next step | ||
| 599 : | |||
| 600 : | por mm3, mm1 ; ij |= jk | ||
| 601 : | movq mm6, mm2 | ||
| 602 : | pxor mm6, mm0 ; mm6 = s^t | ||
| 603 : | pand mm3, mm6 ; (ij|jk) &= st | ||
| 604 : | pavgb mm2, mm0 ; mm2 = (s+t+1)/2 | ||
| 605 : | pand mm3, mm7 ; mask lsb | ||
| 606 : | psubb mm2, mm3 ; apply. | ||
| 607 : | |||
| 608 : | pavgb mm2, [rcx] | ||
| 609 : | movq [rcx], mm2 | ||
| 610 : | |||
| 611 : | movq mm2, [rax] | ||
| 612 : | movq mm3, [rax+1] | ||
| 613 : | movq mm6, mm2 | ||
| 614 : | pavgb mm2, mm3 ; preserved for next iteration | ||
| 615 : | lea rcx,[rcx+rdx] | ||
| 616 : | pxor mm3, mm6 ; preserved for next iteration | ||
| 617 : | |||
| 618 : | por mm1, mm3 | ||
| 619 : | movq mm6, mm0 | ||
| 620 : | pxor mm6, mm2 | ||
| 621 : | pand mm1, mm6 | ||
| 622 : | pavgb mm0, mm2 | ||
| 623 : | |||
| 624 : | pand mm1, mm7 | ||
| 625 : | psubb mm0, mm1 | ||
| 626 : | |||
| 627 : | pavgb mm0, [rcx] | ||
| 628 : | movq [rcx], mm0 | ||
| 629 : | %endmacro | ||
| 630 : | |||
| 631 : | %macro ADD_HH_RND1 0 | ||
| 632 : | lea rax,[rax+rdx] | ||
| 633 : | |||
| 634 : | movq mm0, [rax] | ||
| 635 : | movq mm1, [rax+1] | ||
| 636 : | |||
| 637 : | movq mm6, mm0 | ||
| 638 : | pavgb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step | ||
| 639 : | lea rax,[rax+rdx] | ||
| 640 : | pxor mm1, mm6 ; mm1=(j^k). preserved for next step | ||
| 641 : | |||
| 642 : | pand mm3, mm1 | ||
| 643 : | movq mm6, mm2 | ||
| 644 : | pxor mm6, mm0 | ||
| 645 : | por mm3, mm6 | ||
| 646 : | pavgb mm2, mm0 | ||
| 647 : | pand mm3, mm7 | ||
| 648 : | psubb mm2, mm3 | ||
| 649 : | |||
| 650 : | pavgb mm2, [rcx] | ||
| 651 : | movq [rcx], mm2 | ||
| 652 : | |||
| 653 : | movq mm2, [rax] | ||
| 654 : | movq mm3, [rax+1] | ||
| 655 : | movq mm6, mm2 | ||
| 656 : | pavgb mm2, mm3 ; preserved for next iteration | ||
| 657 : | lea rcx,[rcx+rdx] | ||
| 658 : | pxor mm3, mm6 ; preserved for next iteration | ||
| 659 : | |||
| 660 : | pand mm1, mm3 | ||
| 661 : | movq mm6, mm0 | ||
| 662 : | pxor mm6, mm2 | ||
| 663 : | por mm1, mm6 | ||
| 664 : | pavgb mm0, mm2 | ||
| 665 : | pand mm1, mm7 | ||
| 666 : | psubb mm0, mm1 | ||
| 667 : | |||
| 668 : | pavgb mm0, [rcx] | ||
| 669 : | movq [rcx], mm0 | ||
| 670 : | %endmacro | ||
| 671 : | |||
| 672 : | ALIGN 16 | ||
| 673 : | interpolate8x8_halfpel_hv_add_x86_64: | ||
| 674 : | PROLOG1 | ||
| 675 : | |||
| 676 : | movq mm7, [mmx_one wrt rip] | ||
| 677 : | |||
| 678 : | ; loop invariants: mm2=(i+j+1)/2 and mm3= i^j | ||
| 679 : | movq mm2, [rax] | ||
| 680 : | movq mm3, [rax+1] | ||
| 681 : | movq mm6, mm2 | ||
| 682 : | pavgb mm2, mm3 | ||
| 683 : | pxor mm3, mm6 ; mm2/mm3 ready | ||
| 684 : | |||
| 685 : | jnz near .Loop1 | ||
| 686 : | |||
| 687 : | ADD_HH_RND0 | ||
| 688 : | add rcx, rdx | ||
| 689 : | ADD_HH_RND0 | ||
| 690 : | add rcx, rdx | ||
| 691 : | ADD_HH_RND0 | ||
| 692 : | add rcx, rdx | ||
| 693 : | ADD_HH_RND0 | ||
| 694 : | EPILOG | ||
| 695 : | |||
| 696 : | .Loop1 | ||
| 697 : | ADD_HH_RND1 | ||
| 698 : | add rcx, rdx | ||
| 699 : | ADD_HH_RND1 | ||
| 700 : | add rcx, rdx | ||
| 701 : | ADD_HH_RND1 | ||
| 702 : | add rcx, rdx | ||
| 703 : | ADD_HH_RND1 | ||
| 704 : | |||
| 705 : | EPILOG | ||
| 706 : | .endfunc |
| No admin address has been configured | ViewVC Help |
| Powered by ViewVC 1.0.4 |