Parent Directory
| 
Revision Log
Revision 1.4 - (view) (download)
| 1 : | edgomez | 1.1 | ;/***************************************************************************** |
| 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 : | Isibaar | 1.3 | %define ENDFUNC .endfunc |
| 34 : | edgomez | 1.1 | %else |
| 35 : | global _%1 | ||
| 36 : | %define %1 _%1 | ||
| 37 : | Isibaar | 1.3 | %define ENDFUNC |
| 38 : | edgomez | 1.1 | %endif |
| 39 : | %else | ||
| 40 : | %ifdef MARK_FUNCS | ||
| 41 : | global %1:function %1.endfunc-%1 | ||
| 42 : | Isibaar | 1.3 | %define ENDFUNC .endfunc |
| 43 : | edgomez | 1.1 | %else |
| 44 : | global %1 | ||
| 45 : | Isibaar | 1.3 | %define ENDFUNC |
| 46 : | edgomez | 1.1 | %endif |
| 47 : | %endif | ||
| 48 : | %endmacro | ||
| 49 : | |||
| 50 : | ;============================================================================= | ||
| 51 : | ; Read only data | ||
| 52 : | ;============================================================================= | ||
| 53 : | |||
| 54 : | %ifdef FORMAT_COFF | ||
| 55 : | SECTION .rodata | ||
| 56 : | %else | ||
| 57 : | SECTION .rodata align=16 | ||
| 58 : | %endif | ||
| 59 : | |||
| 60 : | ALIGN 16 | ||
| 61 : | mmx_one: | ||
| 62 : | times 8 db 1 | ||
| 63 : | |||
| 64 : | SECTION .text align=16 | ||
| 65 : | |||
| 66 : | cglobal interpolate8x8_halfpel_h_x86_64 | ||
| 67 : | cglobal interpolate8x8_halfpel_v_x86_64 | ||
| 68 : | cglobal interpolate8x8_halfpel_hv_x86_64 | ||
| 69 : | |||
| 70 : | cglobal interpolate8x8_halfpel_add_x86_64 | ||
| 71 : | cglobal interpolate8x8_halfpel_h_add_x86_64 | ||
| 72 : | cglobal interpolate8x8_halfpel_v_add_x86_64 | ||
| 73 : | cglobal interpolate8x8_halfpel_hv_add_x86_64 | ||
| 74 : | |||
| 75 : | ;=========================================================================== | ||
| 76 : | ; | ||
| 77 : | ; void interpolate8x8_halfpel_h_x86_64(uint8_t * const dst, | ||
| 78 : | ; const uint8_t * const src, | ||
| 79 : | ; const uint32_t stride, | ||
| 80 : | ; const uint32_t rounding); | ||
| 81 : | ; | ||
| 82 : | ;=========================================================================== | ||
| 83 : | |||
| 84 : | %macro COPY_H_SSE_RND0 0 | ||
| 85 : | movq mm0, [rax] | ||
| 86 : | pavgb mm0, [rax+1] | ||
| 87 : | movq mm1, [rax+rdx] | ||
| 88 : | pavgb mm1, [rax+rdx+1] | ||
| 89 : | lea rax,[rax+2*rdx] | ||
| 90 : | movq [rcx],mm0 | ||
| 91 : | movq [rcx+rdx],mm1 | ||
| 92 : | %endmacro | ||
| 93 : | |||
| 94 : | %macro COPY_H_SSE_RND1 0 | ||
| 95 : | movq mm0, [rax] | ||
| 96 : | movq mm1, [rax+rdx] | ||
| 97 : | movq mm4, mm0 | ||
| 98 : | movq mm5, mm1 | ||
| 99 : | movq mm2, [rax+1] | ||
| 100 : | movq mm3, [rax+rdx+1] | ||
| 101 : | pavgb mm0, mm2 | ||
| 102 : | pxor mm2, mm4 | ||
| 103 : | pavgb mm1, mm3 | ||
| 104 : | lea rax, [rax+2*rdx] | ||
| 105 : | pxor mm3, mm5 | ||
| 106 : | pand mm2, mm7 | ||
| 107 : | pand mm3, mm7 | ||
| 108 : | psubb mm0, mm2 | ||
| 109 : | movq [rcx], mm0 | ||
| 110 : | psubb mm1, mm3 | ||
| 111 : | movq [rcx+rdx], mm1 | ||
| 112 : | %endmacro | ||
| 113 : | |||
| 114 : | ALIGN 16 | ||
| 115 : | interpolate8x8_halfpel_h_x86_64: | ||
| 116 : | |||
| 117 : | mov rax, rcx ; rounding | ||
| 118 : | mov rcx, rdi ; Dst | ||
| 119 : | test rax,rax | ||
| 120 : | mov rax, rsi ; src | ||
| 121 : | ; rdx is stride | ||
| 122 : | |||
| 123 : | jnz near .rounding1 | ||
| 124 : | |||
| 125 : | COPY_H_SSE_RND0 | ||
| 126 : | lea rcx,[rcx+2*rdx] | ||
| 127 : | COPY_H_SSE_RND0 | ||
| 128 : | lea rcx,[rcx+2*rdx] | ||
| 129 : | COPY_H_SSE_RND0 | ||
| 130 : | lea rcx,[rcx+2*rdx] | ||
| 131 : | COPY_H_SSE_RND0 | ||
| 132 : | ret | ||
| 133 : | |||
| 134 : | Isibaar | 1.3 | .rounding1: |
| 135 : | edgomez | 1.1 | ; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 |
| 136 : | movq mm7, [mmx_one wrt rip] | ||
| 137 : | COPY_H_SSE_RND1 | ||
| 138 : | lea rcx, [rcx+2*rdx] | ||
| 139 : | COPY_H_SSE_RND1 | ||
| 140 : | lea rcx,[rcx+2*rdx] | ||
| 141 : | COPY_H_SSE_RND1 | ||
| 142 : | lea rcx,[rcx+2*rdx] | ||
| 143 : | COPY_H_SSE_RND1 | ||
| 144 : | ret | ||
| 145 : | Isibaar | 1.3 | ENDFUNC |
| 146 : | edgomez | 1.1 | |
| 147 : | ;=========================================================================== | ||
| 148 : | ; | ||
| 149 : | ; void interpolate8x8_halfpel_v_x86_64(uint8_t * const dst, | ||
| 150 : | ; const uint8_t * const src, | ||
| 151 : | ; const uint32_t stride, | ||
| 152 : | ; const uint32_t rounding); | ||
| 153 : | ; | ||
| 154 : | ;=========================================================================== | ||
| 155 : | |||
| 156 : | %macro COPY_V_SSE_RND0 0 | ||
| 157 : | movq mm0, [rax] | ||
| 158 : | movq mm1, [rax+rdx] | ||
| 159 : | pavgb mm0, mm1 | ||
| 160 : | pavgb mm1, [rax+2*rdx] | ||
| 161 : | lea rax, [rax+2*rdx] | ||
| 162 : | movq [rcx], mm0 | ||
| 163 : | movq [rcx+rdx],mm1 | ||
| 164 : | %endmacro | ||
| 165 : | |||
| 166 : | %macro COPY_V_SSE_RND1 0 | ||
| 167 : | movq mm0, mm2 | ||
| 168 : | movq mm1, [rax] | ||
| 169 : | movq mm2, [rax+rdx] | ||
| 170 : | lea rax,[rax+2*rdx] | ||
| 171 : | movq mm4, mm0 | ||
| 172 : | movq mm5, mm1 | ||
| 173 : | pavgb mm0, mm1 | ||
| 174 : | pxor mm4, mm1 | ||
| 175 : | pavgb mm1, mm2 | ||
| 176 : | pxor mm5, mm2 | ||
| 177 : | pand mm4, mm7 ; lsb's of (i^j)... | ||
| 178 : | pand mm5, mm7 ; lsb's of (i^j)... | ||
| 179 : | psubb mm0, mm4 ; ...are substracted from result of pavgb | ||
| 180 : | movq [rcx], mm0 | ||
| 181 : | psubb mm1, mm5 ; ...are substracted from result of pavgb | ||
| 182 : | movq [rcx+rdx], mm1 | ||
| 183 : | %endmacro | ||
| 184 : | |||
| 185 : | ALIGN 16 | ||
| 186 : | interpolate8x8_halfpel_v_x86_64: | ||
| 187 : | mov rax, rcx ; rounding | ||
| 188 : | mov rcx, rdi ; Dst | ||
| 189 : | test rax,rax | ||
| 190 : | mov rax, rsi ; Src | ||
| 191 : | ; rdx is stride | ||
| 192 : | |||
| 193 : | ; we process 2 line at a time | ||
| 194 : | jnz near .rounding1 | ||
| 195 : | |||
| 196 : | COPY_V_SSE_RND0 | ||
| 197 : | lea rcx, [rcx+2*rdx] | ||
| 198 : | COPY_V_SSE_RND0 | ||
| 199 : | lea rcx, [rcx+2*rdx] | ||
| 200 : | COPY_V_SSE_RND0 | ||
| 201 : | lea rcx, [rcx+2*rdx] | ||
| 202 : | COPY_V_SSE_RND0 | ||
| 203 : | ret | ||
| 204 : | |||
| 205 : | Isibaar | 1.3 | .rounding1: |
| 206 : | edgomez | 1.1 | ; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 |
| 207 : | movq mm7, [mmx_one wrt rip] | ||
| 208 : | movq mm2, [rax] ; loop invariant | ||
| 209 : | add rax, rdx | ||
| 210 : | |||
| 211 : | COPY_V_SSE_RND1 | ||
| 212 : | lea rcx,[rcx+2*rdx] | ||
| 213 : | COPY_V_SSE_RND1 | ||
| 214 : | lea rcx,[rcx+2*rdx] | ||
| 215 : | COPY_V_SSE_RND1 | ||
| 216 : | lea rcx,[rcx+2*rdx] | ||
| 217 : | COPY_V_SSE_RND1 | ||
| 218 : | ret | ||
| 219 : | Isibaar | 1.3 | ENDFUNC |
| 220 : | edgomez | 1.1 | |
| 221 : | ;=========================================================================== | ||
| 222 : | ; | ||
| 223 : | ; void interpolate8x8_halfpel_hv_xmm(uint8_t * const dst, | ||
| 224 : | ; const uint8_t * const src, | ||
| 225 : | ; const uint32_t stride, | ||
| 226 : | ; const uint32_t rounding); | ||
| 227 : | ; | ||
| 228 : | ; | ||
| 229 : | ;=========================================================================== | ||
| 230 : | |||
| 231 : | ; The trick is to correct the result of 'pavgb' with some combination of the | ||
| 232 : | ; lsb's of the 4 input values i,j,k,l, and their intermediate 'pavgb' (s and t). | ||
| 233 : | ; The boolean relations are: | ||
| 234 : | ; (i+j+k+l+3)/4 = (s+t+1)/2 - (ij&kl)&st | ||
| 235 : | ; (i+j+k+l+2)/4 = (s+t+1)/2 - (ij|kl)&st | ||
| 236 : | ; (i+j+k+l+1)/4 = (s+t+1)/2 - (ij&kl)|st | ||
| 237 : | ; (i+j+k+l+0)/4 = (s+t+1)/2 - (ij|kl)|st | ||
| 238 : | ; with s=(i+j+1)/2, t=(k+l+1)/2, ij = i^j, kl = k^l, st = s^t. | ||
| 239 : | |||
| 240 : | ; Moreover, we process 2 lines at a times, for better overlapping (~15% faster). | ||
| 241 : | |||
| 242 : | %macro COPY_HV_SSE_RND0 0 | ||
| 243 : | lea rax, [rax+rdx] | ||
| 244 : | |||
| 245 : | movq mm0, [rax] | ||
| 246 : | movq mm1, [rax+1] | ||
| 247 : | |||
| 248 : | movq mm6, mm0 | ||
| 249 : | pavgb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step | ||
| 250 : | lea rax, [rax+rdx] | ||
| 251 : | pxor mm1, mm6 ; mm1=(j^k). preserved for next step | ||
| 252 : | |||
| 253 : | por mm3, mm1 ; ij |= jk | ||
| 254 : | movq mm6, mm2 | ||
| 255 : | pxor mm6, mm0 ; mm6 = s^t | ||
| 256 : | pand mm3, mm6 ; (ij|jk) &= st | ||
| 257 : | pavgb mm2, mm0 ; mm2 = (s+t+1)/2 | ||
| 258 : | pand mm3, mm7 ; mask lsb | ||
| 259 : | psubb mm2, mm3 ; apply. | ||
| 260 : | |||
| 261 : | movq [rcx], mm2 | ||
| 262 : | |||
| 263 : | movq mm2, [rax] | ||
| 264 : | movq mm3, [rax+1] | ||
| 265 : | movq mm6, mm2 | ||
| 266 : | pavgb mm2, mm3 ; preserved for next iteration | ||
| 267 : | lea rcx,[rcx+rdx] | ||
| 268 : | pxor mm3, mm6 ; preserved for next iteration | ||
| 269 : | |||
| 270 : | por mm1, mm3 | ||
| 271 : | movq mm6, mm0 | ||
| 272 : | pxor mm6, mm2 | ||
| 273 : | pand mm1, mm6 | ||
| 274 : | pavgb mm0, mm2 | ||
| 275 : | |||
| 276 : | pand mm1, mm7 | ||
| 277 : | psubb mm0, mm1 | ||
| 278 : | |||
| 279 : | movq [rcx], mm0 | ||
| 280 : | %endmacro | ||
| 281 : | |||
| 282 : | %macro COPY_HV_SSE_RND1 0 | ||
| 283 : | lea rax, [rax+rdx] | ||
| 284 : | |||
| 285 : | movq mm0, [rax] | ||
| 286 : | movq mm1, [rax+1] | ||
| 287 : | |||
| 288 : | movq mm6, mm0 | ||
| 289 : | pavgb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step | ||
| 290 : | lea rax, [rax+rdx] | ||
| 291 : | pxor mm1, mm6 ; mm1=(j^k). preserved for next step | ||
| 292 : | |||
| 293 : | pand mm3, mm1 | ||
| 294 : | movq mm6, mm2 | ||
| 295 : | pxor mm6, mm0 | ||
| 296 : | por mm3, mm6 | ||
| 297 : | pavgb mm2, mm0 | ||
| 298 : | pand mm3, mm7 | ||
| 299 : | psubb mm2, mm3 | ||
| 300 : | |||
| 301 : | movq [rcx], mm2 | ||
| 302 : | |||
| 303 : | movq mm2, [rax] | ||
| 304 : | movq mm3, [rax+1] | ||
| 305 : | movq mm6, mm2 | ||
| 306 : | pavgb mm2, mm3 ; preserved for next iteration | ||
| 307 : | lea rcx,[rcx+rdx] | ||
| 308 : | pxor mm3, mm6 ; preserved for next iteration | ||
| 309 : | |||
| 310 : | pand mm1, mm3 | ||
| 311 : | movq mm6, mm0 | ||
| 312 : | pxor mm6, mm2 | ||
| 313 : | por mm1, mm6 | ||
| 314 : | pavgb mm0, mm2 | ||
| 315 : | pand mm1, mm7 | ||
| 316 : | psubb mm0, mm1 | ||
| 317 : | |||
| 318 : | movq [rcx], mm0 | ||
| 319 : | %endmacro | ||
| 320 : | |||
| 321 : | ALIGN 16 | ||
| 322 : | interpolate8x8_halfpel_hv_x86_64: | ||
| 323 : | mov rax, rcx ; rounding | ||
| 324 : | mov rcx, rdi ; Dst | ||
| 325 : | test rax, rax | ||
| 326 : | mov rax, rsi ; Src | ||
| 327 : | ; rdx is stride | ||
| 328 : | |||
| 329 : | movq mm7, [mmx_one wrt rip] | ||
| 330 : | |||
| 331 : | ; loop invariants: mm2=(i+j+1)/2 and mm3= i^j | ||
| 332 : | movq mm2, [rax] | ||
| 333 : | movq mm3, [rax+1] | ||
| 334 : | movq mm6, mm2 | ||
| 335 : | pavgb mm2, mm3 | ||
| 336 : | pxor mm3, mm6 ; mm2/mm3 ready | ||
| 337 : | |||
| 338 : | jnz near .rounding1 | ||
| 339 : | |||
| 340 : | COPY_HV_SSE_RND0 | ||
| 341 : | add rcx, rdx | ||
| 342 : | COPY_HV_SSE_RND0 | ||
| 343 : | add rcx, rdx | ||
| 344 : | COPY_HV_SSE_RND0 | ||
| 345 : | add rcx, rdx | ||
| 346 : | COPY_HV_SSE_RND0 | ||
| 347 : | ret | ||
| 348 : | |||
| 349 : | Isibaar | 1.3 | .rounding1: |
| 350 : | edgomez | 1.1 | COPY_HV_SSE_RND1 |
| 351 : | add rcx, rdx | ||
| 352 : | COPY_HV_SSE_RND1 | ||
| 353 : | add rcx, rdx | ||
| 354 : | COPY_HV_SSE_RND1 | ||
| 355 : | add rcx, rdx | ||
| 356 : | COPY_HV_SSE_RND1 | ||
| 357 : | ret | ||
| 358 : | Isibaar | 1.3 | ENDFUNC |
| 359 : | edgomez | 1.1 | |
| 360 : | ;=========================================================================== | ||
| 361 : | ; | ||
| 362 : | ; The next functions combine both source halfpel interpolation step and the | ||
| 363 : | ; averaging (with rouding) step to avoid wasting memory bandwidth computing | ||
| 364 : | ; intermediate halfpel images and then averaging them. | ||
| 365 : | ; | ||
| 366 : | ;=========================================================================== | ||
| 367 : | |||
| 368 : | %macro PROLOG0 0 | ||
| 369 : | ; rcx fourth | ||
| 370 : | ; rdx third | ||
| 371 : | ; r8 fifth | ||
| 372 : | mov r8, rdx ; saves rounding | ||
| 373 : | mov rcx, rdi ; Dst | ||
| 374 : | mov rax, rsi ; Src | ||
| 375 : | ; rdx is stride | ||
| 376 : | %endmacro | ||
| 377 : | %macro PROLOG1 0 | ||
| 378 : | PROLOG0 | ||
| 379 : | test r8, 1; Rounding? | ||
| 380 : | %endmacro | ||
| 381 : | %macro EPILOG 0 | ||
| 382 : | ret | ||
| 383 : | %endmacro | ||
| 384 : | |||
| 385 : | ;=========================================================================== | ||
| 386 : | ; | ||
| 387 : | ; void interpolate8x8_halfpel_add_xmm(uint8_t * const dst, | ||
| 388 : | ; const uint8_t * const src, | ||
| 389 : | ; const uint32_t stride, | ||
| 390 : | ; const uint32_t rounding); | ||
| 391 : | ; | ||
| 392 : | ; | ||
| 393 : | ;=========================================================================== | ||
| 394 : | |||
| 395 : | %macro ADD_FF 2 | ||
| 396 : | movq mm0, [rax+%1] | ||
| 397 : | movq mm1, [rax+%2] | ||
| 398 : | pavgb mm0, [rcx+%1] | ||
| 399 : | pavgb mm1, [rcx+%2] | ||
| 400 : | movq [rcx+%1], mm0 | ||
| 401 : | movq [rcx+%2], mm1 | ||
| 402 : | %endmacro | ||
| 403 : | |||
| 404 : | ALIGN 16 | ||
| 405 : | interpolate8x8_halfpel_add_x86_64: ; 23c | ||
| 406 : | PROLOG1 | ||
| 407 : | ADD_FF 0, rdx | ||
| 408 : | lea rax,[rax+2*rdx] | ||
| 409 : | lea rcx,[rcx+2*rdx] | ||
| 410 : | ADD_FF 0, rdx | ||
| 411 : | lea rax,[rax+2*rdx] | ||
| 412 : | lea rcx,[rcx+2*rdx] | ||
| 413 : | ADD_FF 0, rdx | ||
| 414 : | lea rax,[rax+2*rdx] | ||
| 415 : | lea rcx,[rcx+2*rdx] | ||
| 416 : | ADD_FF 0, rdx | ||
| 417 : | EPILOG | ||
| 418 : | Isibaar | 1.3 | ENDFUNC |
| 419 : | edgomez | 1.1 | |
| 420 : | ;=========================================================================== | ||
| 421 : | ; | ||
| 422 : | ; void interpolate8x8_halfpel_h_add_xmm(uint8_t * const dst, | ||
| 423 : | ; const uint8_t * const src, | ||
| 424 : | ; const uint32_t stride, | ||
| 425 : | ; const uint32_t rounding); | ||
| 426 : | ; | ||
| 427 : | ; | ||
| 428 : | ;=========================================================================== | ||
| 429 : | |||
| 430 : | |||
| 431 : | %macro ADD_FH_RND0 2 | ||
| 432 : | movq mm0, [rax+%1] | ||
| 433 : | movq mm1, [rax+%2] | ||
| 434 : | pavgb mm0, [rax+%1+1] | ||
| 435 : | pavgb mm1, [rax+%2+1] | ||
| 436 : | pavgb mm0, [rcx+%1] | ||
| 437 : | pavgb mm1, [rcx+%2] | ||
| 438 : | movq [rcx+%1],mm0 | ||
| 439 : | movq [rcx+%2],mm1 | ||
| 440 : | %endmacro | ||
| 441 : | |||
| 442 : | %macro ADD_FH_RND1 2 | ||
| 443 : | movq mm0, [rax+%1] | ||
| 444 : | movq mm1, [rax+%2] | ||
| 445 : | movq mm4, mm0 | ||
| 446 : | movq mm5, mm1 | ||
| 447 : | movq mm2, [rax+%1+1] | ||
| 448 : | movq mm3, [rax+%2+1] | ||
| 449 : | pavgb mm0, mm2 | ||
| 450 : | ; lea ?? | ||
| 451 : | pxor mm2, mm4 | ||
| 452 : | pavgb mm1, mm3 | ||
| 453 : | pxor mm3, mm5 | ||
| 454 : | pand mm2, [mmx_one wrt rip] | ||
| 455 : | pand mm3, [mmx_one wrt rip] | ||
| 456 : | psubb mm0, mm2 | ||
| 457 : | psubb mm1, mm3 | ||
| 458 : | pavgb mm0, [rcx+%1] | ||
| 459 : | pavgb mm1, [rcx+%2] | ||
| 460 : | movq [rcx+%1],mm0 | ||
| 461 : | movq [rcx+%2],mm1 | ||
| 462 : | %endmacro | ||
| 463 : | |||
| 464 : | ALIGN 16 | ||
| 465 : | interpolate8x8_halfpel_h_add_x86_64: ; 32c | ||
| 466 : | PROLOG1 | ||
| 467 : | jnz near .Loop1 | ||
| 468 : | ADD_FH_RND0 0, rdx | ||
| 469 : | lea rax,[rax+2*rdx] | ||
| 470 : | lea rcx,[rcx+2*rdx] | ||
| 471 : | ADD_FH_RND0 0, rdx | ||
| 472 : | lea rax,[rax+2*rdx] | ||
| 473 : | lea rcx,[rcx+2*rdx] | ||
| 474 : | ADD_FH_RND0 0, rdx | ||
| 475 : | lea rax,[rax+2*rdx] | ||
| 476 : | lea rcx,[rcx+2*rdx] | ||
| 477 : | ADD_FH_RND0 0, rdx | ||
| 478 : | EPILOG | ||
| 479 : | |||
| 480 : | Isibaar | 1.3 | .Loop1: |
| 481 : | edgomez | 1.1 | ; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 |
| 482 : | ; movq mm7, [mmx_one wrt rip] | ||
| 483 : | ADD_FH_RND1 0, rdx | ||
| 484 : | lea rax,[rax+2*rdx] | ||
| 485 : | lea rcx,[rcx+2*rdx] | ||
| 486 : | ADD_FH_RND1 0, rdx | ||
| 487 : | lea rax,[rax+2*rdx] | ||
| 488 : | lea rcx,[rcx+2*rdx] | ||
| 489 : | ADD_FH_RND1 0, rdx | ||
| 490 : | lea rax,[rax+2*rdx] | ||
| 491 : | lea rcx,[rcx+2*rdx] | ||
| 492 : | ADD_FH_RND1 0, rdx | ||
| 493 : | EPILOG | ||
| 494 : | Isibaar | 1.3 | ENDFUNC |
| 495 : | edgomez | 1.1 | |
| 496 : | |||
| 497 : | ;=========================================================================== | ||
| 498 : | ; | ||
| 499 : | ; void interpolate8x8_halfpel_v_add_x86_64(uint8_t * const dst, | ||
| 500 : | ; const uint8_t * const src, | ||
| 501 : | ; const uint32_t stride, | ||
| 502 : | ; const uint32_t rounding); | ||
| 503 : | ; | ||
| 504 : | ; | ||
| 505 : | ;=========================================================================== | ||
| 506 : | |||
| 507 : | %macro ADD_8_HF_RND0 0 | ||
| 508 : | movq mm0, [rax] | ||
| 509 : | movq mm1, [rax+rdx] | ||
| 510 : | pavgb mm0, mm1 | ||
| 511 : | pavgb mm1, [rax+2*rdx] | ||
| 512 : | lea rax,[rax+2*rdx] | ||
| 513 : | pavgb mm0, [rcx] | ||
| 514 : | pavgb mm1, [rcx+rdx] | ||
| 515 : | movq [rcx],mm0 | ||
| 516 : | movq [rcx+rdx],mm1 | ||
| 517 : | %endmacro | ||
| 518 : | |||
| 519 : | %macro ADD_8_HF_RND1 0 | ||
| 520 : | movq mm1, [rax+rdx] | ||
| 521 : | movq mm2, [rax+2*rdx] | ||
| 522 : | lea rax,[rax+2*rdx] | ||
| 523 : | movq mm4, mm0 | ||
| 524 : | movq mm5, mm1 | ||
| 525 : | pavgb mm0, mm1 | ||
| 526 : | pxor mm4, mm1 | ||
| 527 : | pavgb mm1, mm2 | ||
| 528 : | pxor mm5, mm2 | ||
| 529 : | pand mm4, mm7 ; lsb's of (i^j)... | ||
| 530 : | pand mm5, mm7 ; lsb's of (i^j)... | ||
| 531 : | psubb mm0, mm4 ; ...are substracted from result of pavgb | ||
| 532 : | pavgb mm0, [rcx] | ||
| 533 : | movq [rcx], mm0 | ||
| 534 : | psubb mm1, mm5 ; ...are substracted from result of pavgb | ||
| 535 : | pavgb mm1, [rcx+rdx] | ||
| 536 : | movq [rcx+rdx], mm1 | ||
| 537 : | %endmacro | ||
| 538 : | |||
| 539 : | ALIGN 16 | ||
| 540 : | interpolate8x8_halfpel_v_add_x86_64: | ||
| 541 : | PROLOG1 | ||
| 542 : | |||
| 543 : | jnz near .Loop1 | ||
| 544 : | pxor mm7, mm7 ; this is a NOP | ||
| 545 : | |||
| 546 : | ADD_8_HF_RND0 | ||
| 547 : | lea rcx,[rcx+2*rdx] | ||
| 548 : | ADD_8_HF_RND0 | ||
| 549 : | lea rcx,[rcx+2*rdx] | ||
| 550 : | ADD_8_HF_RND0 | ||
| 551 : | lea rcx,[rcx+2*rdx] | ||
| 552 : | ADD_8_HF_RND0 | ||
| 553 : | EPILOG | ||
| 554 : | |||
| 555 : | Isibaar | 1.3 | .Loop1: |
| 556 : | edgomez | 1.1 | movq mm0, [rax] ; loop invariant |
| 557 : | movq mm7, [mmx_one wrt rip] | ||
| 558 : | |||
| 559 : | ADD_8_HF_RND1 | ||
| 560 : | movq mm0, mm2 | ||
| 561 : | lea rcx,[rcx+2*rdx] | ||
| 562 : | ADD_8_HF_RND1 | ||
| 563 : | movq mm0, mm2 | ||
| 564 : | lea rcx,[rcx+2*rdx] | ||
| 565 : | ADD_8_HF_RND1 | ||
| 566 : | movq mm0, mm2 | ||
| 567 : | lea rcx,[rcx+2*rdx] | ||
| 568 : | ADD_8_HF_RND1 | ||
| 569 : | EPILOG | ||
| 570 : | Isibaar | 1.3 | ENDFUNC |
| 571 : | edgomez | 1.1 | |
| 572 : | ; The trick is to correct the result of 'pavgb' with some combination of the | ||
| 573 : | ; lsb's of the 4 input values i,j,k,l, and their intermediate 'pavgb' (s and t). | ||
| 574 : | ; The boolean relations are: | ||
| 575 : | ; (i+j+k+l+3)/4 = (s+t+1)/2 - (ij&kl)&st | ||
| 576 : | ; (i+j+k+l+2)/4 = (s+t+1)/2 - (ij|kl)&st | ||
| 577 : | ; (i+j+k+l+1)/4 = (s+t+1)/2 - (ij&kl)|st | ||
| 578 : | ; (i+j+k+l+0)/4 = (s+t+1)/2 - (ij|kl)|st | ||
| 579 : | ; with s=(i+j+1)/2, t=(k+l+1)/2, ij = i^j, kl = k^l, st = s^t. | ||
| 580 : | |||
| 581 : | ; Moreover, we process 2 lines at a times, for better overlapping (~15% faster). | ||
| 582 : | |||
| 583 : | ;=========================================================================== | ||
| 584 : | ; | ||
| 585 : | ; void interpolate8x8_halfpel_hv_add_x86_64(uint8_t * const dst, | ||
| 586 : | ; const uint8_t * const src, | ||
| 587 : | ; const uint32_t stride, | ||
| 588 : | ; const uint32_t rounding); | ||
| 589 : | ; | ||
| 590 : | ; | ||
| 591 : | ;=========================================================================== | ||
| 592 : | |||
| 593 : | %macro ADD_HH_RND0 0 | ||
| 594 : | lea rax,[rax+rdx] | ||
| 595 : | |||
| 596 : | movq mm0, [rax] | ||
| 597 : | movq mm1, [rax+1] | ||
| 598 : | |||
| 599 : | movq mm6, mm0 | ||
| 600 : | pavgb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step | ||
| 601 : | lea rax,[rax+rdx] | ||
| 602 : | pxor mm1, mm6 ; mm1=(j^k). preserved for next step | ||
| 603 : | |||
| 604 : | por mm3, mm1 ; ij |= jk | ||
| 605 : | movq mm6, mm2 | ||
| 606 : | pxor mm6, mm0 ; mm6 = s^t | ||
| 607 : | pand mm3, mm6 ; (ij|jk) &= st | ||
| 608 : | pavgb mm2, mm0 ; mm2 = (s+t+1)/2 | ||
| 609 : | pand mm3, mm7 ; mask lsb | ||
| 610 : | psubb mm2, mm3 ; apply. | ||
| 611 : | |||
| 612 : | pavgb mm2, [rcx] | ||
| 613 : | movq [rcx], mm2 | ||
| 614 : | |||
| 615 : | movq mm2, [rax] | ||
| 616 : | movq mm3, [rax+1] | ||
| 617 : | movq mm6, mm2 | ||
| 618 : | pavgb mm2, mm3 ; preserved for next iteration | ||
| 619 : | lea rcx,[rcx+rdx] | ||
| 620 : | pxor mm3, mm6 ; preserved for next iteration | ||
| 621 : | |||
| 622 : | por mm1, mm3 | ||
| 623 : | movq mm6, mm0 | ||
| 624 : | pxor mm6, mm2 | ||
| 625 : | pand mm1, mm6 | ||
| 626 : | pavgb mm0, mm2 | ||
| 627 : | |||
| 628 : | pand mm1, mm7 | ||
| 629 : | psubb mm0, mm1 | ||
| 630 : | |||
| 631 : | pavgb mm0, [rcx] | ||
| 632 : | movq [rcx], mm0 | ||
| 633 : | %endmacro | ||
| 634 : | |||
| 635 : | %macro ADD_HH_RND1 0 | ||
| 636 : | lea rax,[rax+rdx] | ||
| 637 : | |||
| 638 : | movq mm0, [rax] | ||
| 639 : | movq mm1, [rax+1] | ||
| 640 : | |||
| 641 : | movq mm6, mm0 | ||
| 642 : | pavgb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step | ||
| 643 : | lea rax,[rax+rdx] | ||
| 644 : | pxor mm1, mm6 ; mm1=(j^k). preserved for next step | ||
| 645 : | |||
| 646 : | pand mm3, mm1 | ||
| 647 : | movq mm6, mm2 | ||
| 648 : | pxor mm6, mm0 | ||
| 649 : | por mm3, mm6 | ||
| 650 : | pavgb mm2, mm0 | ||
| 651 : | pand mm3, mm7 | ||
| 652 : | psubb mm2, mm3 | ||
| 653 : | |||
| 654 : | pavgb mm2, [rcx] | ||
| 655 : | movq [rcx], mm2 | ||
| 656 : | |||
| 657 : | movq mm2, [rax] | ||
| 658 : | movq mm3, [rax+1] | ||
| 659 : | movq mm6, mm2 | ||
| 660 : | pavgb mm2, mm3 ; preserved for next iteration | ||
| 661 : | lea rcx,[rcx+rdx] | ||
| 662 : | pxor mm3, mm6 ; preserved for next iteration | ||
| 663 : | |||
| 664 : | pand mm1, mm3 | ||
| 665 : | movq mm6, mm0 | ||
| 666 : | pxor mm6, mm2 | ||
| 667 : | por mm1, mm6 | ||
| 668 : | pavgb mm0, mm2 | ||
| 669 : | pand mm1, mm7 | ||
| 670 : | psubb mm0, mm1 | ||
| 671 : | |||
| 672 : | pavgb mm0, [rcx] | ||
| 673 : | movq [rcx], mm0 | ||
| 674 : | %endmacro | ||
| 675 : | |||
| 676 : | ALIGN 16 | ||
| 677 : | interpolate8x8_halfpel_hv_add_x86_64: | ||
| 678 : | PROLOG1 | ||
| 679 : | |||
| 680 : | movq mm7, [mmx_one wrt rip] | ||
| 681 : | |||
| 682 : | ; loop invariants: mm2=(i+j+1)/2 and mm3= i^j | ||
| 683 : | movq mm2, [rax] | ||
| 684 : | movq mm3, [rax+1] | ||
| 685 : | movq mm6, mm2 | ||
| 686 : | pavgb mm2, mm3 | ||
| 687 : | pxor mm3, mm6 ; mm2/mm3 ready | ||
| 688 : | |||
| 689 : | jnz near .Loop1 | ||
| 690 : | |||
| 691 : | ADD_HH_RND0 | ||
| 692 : | add rcx, rdx | ||
| 693 : | ADD_HH_RND0 | ||
| 694 : | add rcx, rdx | ||
| 695 : | ADD_HH_RND0 | ||
| 696 : | add rcx, rdx | ||
| 697 : | ADD_HH_RND0 | ||
| 698 : | EPILOG | ||
| 699 : | |||
| 700 : | Isibaar | 1.3 | .Loop1: |
| 701 : | edgomez | 1.1 | ADD_HH_RND1 |
| 702 : | add rcx, rdx | ||
| 703 : | ADD_HH_RND1 | ||
| 704 : | add rcx, rdx | ||
| 705 : | ADD_HH_RND1 | ||
| 706 : | add rcx, rdx | ||
| 707 : | ADD_HH_RND1 | ||
| 708 : | |||
| 709 : | EPILOG | ||
| 710 : | Isibaar | 1.3 | ENDFUNC |
| 711 : | Isibaar | 1.2 | |
| 712 : | %ifidn __OUTPUT_FORMAT__,elf | ||
| 713 : | section ".note.GNU-stack" noalloc noexec nowrite progbits | ||
| 714 : | %endif | ||
| 715 : |
| No admin address has been configured | ViewVC Help |
| Powered by ViewVC 1.0.4 |