Parent Directory
| 
Revision Log
Revision 1.5 - (view) (download)
| 1 : | edgomez | 1.3 | ;/***************************************************************************** |
| 2 : | edgomez | 1.2 | ; * |
| 3 : | edgomez | 1.3 | ; * XVID MPEG-4 VIDEO CODEC |
| 4 : | ; * - 3dne pipeline optimized 8x8 block-based halfpel interpolation - | ||
| 5 : | edgomez | 1.2 | ; * |
| 6 : | edgomez | 1.3 | ; * Copyright(C) 2002 Jaan Kalda |
| 7 : | edgomez | 1.2 | ; * |
| 8 : | edgomez | 1.3 | ; * This program is free software ; you can redistribute it and/or modify |
| 9 : | ; * it under the terms of the GNU General Public License as published by | ||
| 10 : | ; * the Free Software Foundation ; either version 2 of the License, or | ||
| 11 : | ; * (at your option) any later version. | ||
| 12 : | edgomez | 1.2 | ; * |
| 13 : | edgomez | 1.3 | ; * This program 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 | ||
| 16 : | ; * GNU General Public License for more details. | ||
| 17 : | edgomez | 1.2 | ; * |
| 18 : | edgomez | 1.3 | ; * You should have received a copy of the GNU General Public License |
| 19 : | ; * along with this program ; if not, write to the Free Software | ||
| 20 : | ; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | ||
| 21 : | edgomez | 1.2 | ; * |
| 22 : | edgomez | 1.3 | ; ****************************************************************************/ |
| 23 : | edgomez | 1.2 | |
| 24 : | edgomez | 1.3 | ; these 3dne functions are compatible with iSSE, but are optimized specifically |
| 25 : | ; for K7 pipelines | ||
| 26 : | edgomez | 1.2 | |
| 27 : | edgomez | 1.3 | BITS 32 |
| 28 : | edgomez | 1.2 | |
| 29 : | edgomez | 1.3 | %macro cglobal 1 |
| 30 : | edgomez | 1.2 | %ifdef PREFIX |
| 31 : | edgomez | 1.5 | %ifdef MARK_FUNCS |
| 32 : | global _%1:function | ||
| 33 : | %define %1 _%1:function | ||
| 34 : | %else | ||
| 35 : | global _%1 | ||
| 36 : | %define %1 _%1 | ||
| 37 : | %endif | ||
| 38 : | edgomez | 1.2 | %else |
| 39 : | edgomez | 1.5 | %ifdef MARK_FUNCS |
| 40 : | global %1:function | ||
| 41 : | %else | ||
| 42 : | global %1 | ||
| 43 : | %endif | ||
| 44 : | edgomez | 1.2 | %endif |
| 45 : | %endmacro | ||
| 46 : | edgomez | 1.3 | |
| 47 : | ;============================================================================= | ||
| 48 : | ; Read only data | ||
| 49 : | ;============================================================================= | ||
| 50 : | |||
| 51 : | edgomez | 1.2 | %ifdef FORMAT_COFF |
| 52 : | edgomez | 1.4 | SECTION .rodata |
| 53 : | edgomez | 1.2 | %else |
| 54 : | edgomez | 1.4 | SECTION .rodata align=16 |
| 55 : | edgomez | 1.2 | %endif |
| 56 : | |||
| 57 : | edgomez | 1.3 | ALIGN 16 |
| 58 : | mmx_one: | ||
| 59 : | times 8 db 1 | ||
| 60 : | |||
| 61 : | ALIGN 8 | ||
| 62 : | mm_minusone: | ||
| 63 : | dd -1,-1 | ||
| 64 : | edgomez | 1.2 | |
| 65 : | edgomez | 1.3 | ;============================================================================= |
| 66 : | ; Macros | ||
| 67 : | ;============================================================================= | ||
| 68 : | edgomez | 1.2 | |
| 69 : | edgomez | 1.3 | %macro nop4 0 |
| 70 : | DB 08Dh,074h,026h,0 | ||
| 71 : | %endmacro | ||
| 72 : | |||
| 73 : | ;============================================================================= | ||
| 74 : | ; Macros | ||
| 75 : | ;============================================================================= | ||
| 76 : | edgomez | 1.2 | |
| 77 : | edgomez | 1.3 | SECTION .text |
| 78 : | edgomez | 1.2 | |
| 79 : | cglobal interpolate8x8_halfpel_h_3dne | ||
| 80 : | cglobal interpolate8x8_halfpel_v_3dne | ||
| 81 : | cglobal interpolate8x8_halfpel_hv_3dne | ||
| 82 : | |||
| 83 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
| 84 : | edgomez | 1.2 | ; |
| 85 : | ; void interpolate8x8_halfpel_h_3dne(uint8_t * const dst, | ||
| 86 : | edgomez | 1.3 | ; const uint8_t * const src, |
| 87 : | ; const uint32_t stride, | ||
| 88 : | ; const uint32_t rounding); | ||
| 89 : | edgomez | 1.2 | ; |
| 90 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
| 91 : | edgomez | 1.2 | |
| 92 : | %macro COPY_H_SSE_RND0 1 | ||
| 93 : | %if (%1) | ||
| 94 : | edgomez | 1.3 | movq mm0, [eax] |
| 95 : | edgomez | 1.2 | %else |
| 96 : | edgomez | 1.3 | movq mm0, [dword eax] |
| 97 : | edgomez | 1.2 | %endif |
| 98 : | pavgb mm0, [eax+1] | ||
| 99 : | edgomez | 1.3 | movq mm1, [eax+edx] |
| 100 : | edgomez | 1.2 | pavgb mm1, [eax+edx+1] |
| 101 : | edgomez | 1.3 | lea eax, [eax+2*edx] |
| 102 : | movq [ecx], mm0 | ||
| 103 : | movq [ecx+edx], mm1 | ||
| 104 : | edgomez | 1.2 | %endmacro |
| 105 : | |||
| 106 : | %macro COPY_H_SSE_RND1 0 | ||
| 107 : | movq mm0, [eax] | ||
| 108 : | movq mm1, [eax+edx] | ||
| 109 : | movq mm4, mm0 | ||
| 110 : | movq mm5, mm1 | ||
| 111 : | edgomez | 1.3 | movq mm2, [eax+1] |
| 112 : | edgomez | 1.2 | movq mm3, [eax+edx+1] |
| 113 : | pavgb mm0, mm2 | ||
| 114 : | pxor mm2, mm4 | ||
| 115 : | pavgb mm1, mm3 | ||
| 116 : | edgomez | 1.3 | lea eax, [eax+2*edx] |
| 117 : | edgomez | 1.2 | pxor mm3, mm5 |
| 118 : | pand mm2, mm7 | ||
| 119 : | pand mm3, mm7 | ||
| 120 : | psubb mm0, mm2 | ||
| 121 : | movq [ecx], mm0 | ||
| 122 : | psubb mm1, mm3 | ||
| 123 : | edgomez | 1.3 | movq [ecx+edx], mm1 |
| 124 : | edgomez | 1.2 | %endmacro |
| 125 : | |||
| 126 : | edgomez | 1.3 | ALIGN 16 |
| 127 : | edgomez | 1.2 | interpolate8x8_halfpel_h_3dne: |
| 128 : | |||
| 129 : | mov eax, [esp+ 8] ; Src | ||
| 130 : | mov edx, [esp+12] ; stride | ||
| 131 : | dec dword [esp+16]; rounding | ||
| 132 : | |||
| 133 : | jz .rounding1 | ||
| 134 : | mov ecx, [esp+ 4] ; Dst | ||
| 135 : | |||
| 136 : | COPY_H_SSE_RND0 0 | ||
| 137 : | lea ecx,[ecx+2*edx] | ||
| 138 : | COPY_H_SSE_RND0 1 | ||
| 139 : | lea ecx,[ecx+2*edx] | ||
| 140 : | COPY_H_SSE_RND0 1 | ||
| 141 : | lea ecx,[ecx+2*edx] | ||
| 142 : | COPY_H_SSE_RND0 1 | ||
| 143 : | ret | ||
| 144 : | |||
| 145 : | .rounding1 | ||
| 146 : | edgomez | 1.3 | ; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 |
| 147 : | edgomez | 1.2 | mov ecx, [esp+ 4] ; Dst |
| 148 : | movq mm7, [mmx_one] | ||
| 149 : | COPY_H_SSE_RND1 | ||
| 150 : | lea ecx, [ecx+2*edx] | ||
| 151 : | COPY_H_SSE_RND1 | ||
| 152 : | lea ecx,[ecx+2*edx] | ||
| 153 : | COPY_H_SSE_RND1 | ||
| 154 : | lea ecx,[ecx+2*edx] | ||
| 155 : | COPY_H_SSE_RND1 | ||
| 156 : | ret | ||
| 157 : | |||
| 158 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
| 159 : | edgomez | 1.2 | ; |
| 160 : | ; void interpolate8x8_halfpel_v_3dne(uint8_t * const dst, | ||
| 161 : | edgomez | 1.3 | ; const uint8_t * const src, |
| 162 : | ; const uint32_t stride, | ||
| 163 : | ; const uint32_t rounding); | ||
| 164 : | edgomez | 1.2 | ; |
| 165 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
| 166 : | edgomez | 1.2 | |
| 167 : | edgomez | 1.3 | ALIGN 16 |
| 168 : | edgomez | 1.2 | interpolate8x8_halfpel_v_3dne: |
| 169 : | |||
| 170 : | mov eax, [esp+ 8] ; Src | ||
| 171 : | mov edx, [esp+12] ; stride | ||
| 172 : | dec dword [esp+16]; rounding | ||
| 173 : | |||
| 174 : | ; we process 2 line at a time | ||
| 175 : | |||
| 176 : | jz .rounding1 | ||
| 177 : | pxor mm2,mm2 | ||
| 178 : | edgomez | 1.3 | movq mm0, [eax] |
| 179 : | movq mm1, [eax+edx] | ||
| 180 : | por mm2, [eax+2*edx] | ||
| 181 : | edgomez | 1.2 | mov ecx, [esp+ 4] ; Dst |
| 182 : | edgomez | 1.3 | lea eax, [eax+2*edx] |
| 183 : | pxor mm4, mm4 | ||
| 184 : | edgomez | 1.2 | pavgb mm0, mm1 |
| 185 : | edgomez | 1.3 | pavgb mm1, mm2 |
| 186 : | movq [byte ecx], mm0 | ||
| 187 : | movq [ecx+edx], mm1 | ||
| 188 : | pxor mm6, mm6 | ||
| 189 : | add eax, edx | ||
| 190 : | lea ecx, [ecx+2*edx] | ||
| 191 : | movq mm3, [byte eax] | ||
| 192 : | por mm4, [eax+edx] | ||
| 193 : | lea eax, [eax+2*edx] | ||
| 194 : | edgomez | 1.2 | pavgb mm2, mm3 |
| 195 : | pavgb mm3, mm4 | ||
| 196 : | edgomez | 1.3 | movq [ecx], mm2 |
| 197 : | movq [ecx+edx], mm3 | ||
| 198 : | lea ecx, [byte ecx+2*edx] | ||
| 199 : | movq mm5, [byte eax] | ||
| 200 : | por mm6, [eax+edx] | ||
| 201 : | lea eax, [eax+2*edx] | ||
| 202 : | edgomez | 1.2 | pavgb mm4, mm5 |
| 203 : | pavgb mm5, mm6 | ||
| 204 : | edgomez | 1.3 | movq [ecx], mm4 |
| 205 : | movq [ecx+edx], mm5 | ||
| 206 : | lea ecx, [ecx+2*edx] | ||
| 207 : | movq mm7, [eax] | ||
| 208 : | movq mm0, [eax+edx] | ||
| 209 : | edgomez | 1.2 | pavgb mm6, mm7 |
| 210 : | pavgb mm7, mm0 | ||
| 211 : | edgomez | 1.3 | movq [ecx], mm6 |
| 212 : | movq [ecx+edx], mm7 | ||
| 213 : | edgomez | 1.2 | ret |
| 214 : | |||
| 215 : | edgomez | 1.3 | ALIGN 8 |
| 216 : | edgomez | 1.2 | .rounding1 |
| 217 : | edgomez | 1.3 | pcmpeqb mm0, mm0 |
| 218 : | psubusb mm0, [eax] | ||
| 219 : | add eax, edx | ||
| 220 : | edgomez | 1.2 | mov ecx, [esp+ 4] ; Dst |
| 221 : | push esi | ||
| 222 : | edgomez | 1.3 | pcmpeqb mm1, mm1 |
| 223 : | pcmpeqb mm2, mm2 | ||
| 224 : | mov esi, mm_minusone | ||
| 225 : | psubusb mm1, [byte eax] | ||
| 226 : | psubusb mm2, [eax+edx] | ||
| 227 : | lea eax, [eax+2*edx] | ||
| 228 : | edgomez | 1.2 | movq mm6, [esi] |
| 229 : | movq mm7, [esi] | ||
| 230 : | pavgb mm0, mm1 | ||
| 231 : | pavgb mm1, mm2 | ||
| 232 : | edgomez | 1.3 | psubusb mm6, mm0 |
| 233 : | psubusb mm7, mm1 | ||
| 234 : | edgomez | 1.2 | movq [ecx], mm6 |
| 235 : | movq [ecx+edx], mm7 | ||
| 236 : | edgomez | 1.3 | lea ecx, [ecx+2*edx] |
| 237 : | pcmpeqb mm3, mm3 | ||
| 238 : | pcmpeqb mm4, mm4 | ||
| 239 : | psubusb mm3, [eax] | ||
| 240 : | psubusb mm4, [eax+edx] | ||
| 241 : | lea eax, [eax+2*edx] | ||
| 242 : | edgomez | 1.2 | pavgb mm2, mm3 |
| 243 : | pavgb mm3, mm4 | ||
| 244 : | movq mm0, [esi] | ||
| 245 : | movq mm1, [esi] | ||
| 246 : | edgomez | 1.3 | psubusb mm0, mm2 |
| 247 : | psubusb mm1, mm3 | ||
| 248 : | edgomez | 1.2 | movq [ecx], mm0 |
| 249 : | movq [ecx+edx], mm1 | ||
| 250 : | lea ecx,[ecx+2*edx] | ||
| 251 : | |||
| 252 : | edgomez | 1.3 | pcmpeqb mm5, mm5 |
| 253 : | pcmpeqb mm6, mm6 | ||
| 254 : | psubusb mm5, [eax] | ||
| 255 : | psubusb mm6, [eax+edx] | ||
| 256 : | lea eax, [eax+2*edx] | ||
| 257 : | edgomez | 1.2 | pavgb mm4, mm5 |
| 258 : | pavgb mm5, mm6 | ||
| 259 : | movq mm2, [esi] | ||
| 260 : | movq mm3, [esi] | ||
| 261 : | edgomez | 1.3 | psubusb mm2, mm4 |
| 262 : | psubusb mm3, mm5 | ||
| 263 : | edgomez | 1.2 | movq [ecx], mm2 |
| 264 : | movq [ecx+edx], mm3 | ||
| 265 : | edgomez | 1.3 | lea ecx, [ecx+2*edx] |
| 266 : | pcmpeqb mm7, mm7 | ||
| 267 : | pcmpeqb mm0, mm0 | ||
| 268 : | psubusb mm7, [eax] | ||
| 269 : | psubusb mm0, [eax+edx] | ||
| 270 : | edgomez | 1.2 | pavgb mm6, mm7 |
| 271 : | pavgb mm7, mm0 | ||
| 272 : | movq mm4, [esi] | ||
| 273 : | movq mm5, [esi] | ||
| 274 : | edgomez | 1.3 | psubusb mm4, mm6 |
| 275 : | edgomez | 1.2 | pop esi |
| 276 : | edgomez | 1.3 | psubusb mm5, mm7 |
| 277 : | edgomez | 1.2 | movq [ecx], mm4 |
| 278 : | movq [ecx+edx], mm5 | ||
| 279 : | ret | ||
| 280 : | edgomez | 1.3 | |
| 281 : | ;----------------------------------------------------------------------------- | ||
| 282 : | edgomez | 1.2 | ; |
| 283 : | ; void interpolate8x8_halfpel_hv_3dne(uint8_t * const dst, | ||
| 284 : | edgomez | 1.3 | ; const uint8_t * const src, |
| 285 : | ; const uint32_t stride, | ||
| 286 : | ; const uint32_t rounding); | ||
| 287 : | edgomez | 1.2 | ; |
| 288 : | ; | ||
| 289 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
| 290 : | edgomez | 1.2 | |
| 291 : | ; The trick is to correct the result of 'pavgb' with some combination of the | ||
| 292 : | ; lsb's of the 4 input values i,j,k,l, and their intermediate 'pavgb' (s and t). | ||
| 293 : | ; The boolean relations are: | ||
| 294 : | edgomez | 1.3 | ; (i+j+k+l+3)/4 = (s+t+1)/2 - (ij&kl)&st |
| 295 : | edgomez | 1.2 | ; (i+j+k+l+2)/4 = (s+t+1)/2 - (ij|kl)&st |
| 296 : | ; (i+j+k+l+1)/4 = (s+t+1)/2 - (ij&kl)|st | ||
| 297 : | ; (i+j+k+l+0)/4 = (s+t+1)/2 - (ij|kl)|st | ||
| 298 : | ; with s=(i+j+1)/2, t=(k+l+1)/2, ij = i^j, kl = k^l, st = s^t. | ||
| 299 : | |||
| 300 : | ; Moreover, we process 2 lines at a times, for better overlapping (~15% faster). | ||
| 301 : | |||
| 302 : | %macro COPY_HV_SSE_RND0 0 | ||
| 303 : | |||
| 304 : | edgomez | 1.3 | movq mm0, [eax+edx] |
| 305 : | movq mm1, [eax+edx+1] | ||
| 306 : | edgomez | 1.2 | |
| 307 : | edgomez | 1.3 | movq mm6, mm0 |
| 308 : | pavgb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step | ||
| 309 : | lea eax, [eax+2*edx] | ||
| 310 : | pxor mm1, mm6 ; mm1=(j^k). preserved for next step | ||
| 311 : | edgomez | 1.2 | |
| 312 : | edgomez | 1.3 | por mm3, mm1 ; ij |= jk |
| 313 : | movq mm6, mm2 | ||
| 314 : | pxor mm6, mm0 ; mm6 = s^t | ||
| 315 : | pand mm3, mm6 ; (ij|jk) &= st | ||
| 316 : | pavgb mm2, mm0 ; mm2 = (s+t+1)/2 | ||
| 317 : | movq mm6, [eax] | ||
| 318 : | pand mm3, mm7 ; mask lsb | ||
| 319 : | psubb mm2, mm3 ; apply. | ||
| 320 : | |||
| 321 : | movq [ecx], mm2 | ||
| 322 : | edgomez | 1.2 | |
| 323 : | edgomez | 1.3 | movq mm2, [eax] |
| 324 : | movq mm3, [eax+1] | ||
| 325 : | pavgb mm2, mm3 ; preserved for next iteration | ||
| 326 : | pxor mm3, mm6 ; preserved for next iteration | ||
| 327 : | |||
| 328 : | por mm1, mm3 | ||
| 329 : | movq mm6, mm0 | ||
| 330 : | pxor mm6, mm2 | ||
| 331 : | pand mm1, mm6 | ||
| 332 : | pavgb mm0, mm2 | ||
| 333 : | |||
| 334 : | pand mm1, mm7 | ||
| 335 : | psubb mm0, mm1 | ||
| 336 : | |||
| 337 : | movq [ecx+edx], mm0 | ||
| 338 : | edgomez | 1.2 | %endmacro |
| 339 : | |||
| 340 : | %macro COPY_HV_SSE_RND1 0 | ||
| 341 : | edgomez | 1.3 | movq mm0, [eax+edx] |
| 342 : | movq mm1, [eax+edx+1] | ||
| 343 : | |||
| 344 : | movq mm6, mm0 | ||
| 345 : | pavgb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step | ||
| 346 : | lea eax,[eax+2*edx] | ||
| 347 : | pxor mm1, mm6 ; mm1=(j^k). preserved for next step | ||
| 348 : | edgomez | 1.2 | |
| 349 : | edgomez | 1.3 | pand mm3, mm1 |
| 350 : | movq mm6, mm2 | ||
| 351 : | pxor mm6, mm0 | ||
| 352 : | por mm3, mm6 | ||
| 353 : | pavgb mm2, mm0 | ||
| 354 : | movq mm6, [eax] | ||
| 355 : | pand mm3, mm7 | ||
| 356 : | psubb mm2, mm3 | ||
| 357 : | |||
| 358 : | movq [ecx], mm2 | ||
| 359 : | |||
| 360 : | movq mm2, [eax] | ||
| 361 : | movq mm3, [eax+1] | ||
| 362 : | pavgb mm2, mm3 ; preserved for next iteration | ||
| 363 : | pxor mm3, mm6 ; preserved for next iteration | ||
| 364 : | |||
| 365 : | pand mm1, mm3 | ||
| 366 : | movq mm6, mm0 | ||
| 367 : | pxor mm6, mm2 | ||
| 368 : | por mm1, mm6 | ||
| 369 : | pavgb mm0, mm2 | ||
| 370 : | pand mm1, mm7 | ||
| 371 : | psubb mm0, mm1 | ||
| 372 : | movq [ecx+edx], mm0 | ||
| 373 : | edgomez | 1.2 | %endmacro |
| 374 : | |||
| 375 : | edgomez | 1.3 | ALIGN 16 |
| 376 : | edgomez | 1.2 | interpolate8x8_halfpel_hv_3dne: |
| 377 : | edgomez | 1.3 | mov eax, [esp+ 8] ; Src |
| 378 : | mov edx, [esp+12] ; stride | ||
| 379 : | dec dword [esp+16] ; rounding | ||
| 380 : | edgomez | 1.2 | |
| 381 : | ; loop invariants: mm2=(i+j+1)/2 and mm3= i^j | ||
| 382 : | movq mm2, [eax] | ||
| 383 : | movq mm3, [eax+1] | ||
| 384 : | movq mm6, mm2 | ||
| 385 : | pavgb mm2, mm3 | ||
| 386 : | edgomez | 1.3 | pxor mm3, mm6 ; mm2/mm3 ready |
| 387 : | mov ecx, [esp+ 4] ; Dst | ||
| 388 : | edgomez | 1.2 | movq mm7, [mmx_one] |
| 389 : | |||
| 390 : | jz near .rounding1 | ||
| 391 : | edgomez | 1.3 | lea ebp,[byte ebp] |
| 392 : | edgomez | 1.2 | COPY_HV_SSE_RND0 |
| 393 : | edgomez | 1.3 | lea ecx,[ecx+2*edx] |
| 394 : | edgomez | 1.2 | COPY_HV_SSE_RND0 |
| 395 : | edgomez | 1.3 | lea ecx,[ecx+2*edx] |
| 396 : | edgomez | 1.2 | COPY_HV_SSE_RND0 |
| 397 : | edgomez | 1.3 | lea ecx,[ecx+2*edx] |
| 398 : | edgomez | 1.2 | COPY_HV_SSE_RND0 |
| 399 : | ret | ||
| 400 : | |||
| 401 : | edgomez | 1.3 | ALIGN 16 |
| 402 : | edgomez | 1.2 | .rounding1 |
| 403 : | COPY_HV_SSE_RND1 | ||
| 404 : | edgomez | 1.3 | lea ecx,[ecx+2*edx] |
| 405 : | edgomez | 1.2 | COPY_HV_SSE_RND1 |
| 406 : | edgomez | 1.3 | lea ecx,[ecx+2*edx] |
| 407 : | edgomez | 1.2 | COPY_HV_SSE_RND1 |
| 408 : | edgomez | 1.3 | lea ecx,[ecx+2*edx] |
| 409 : | edgomez | 1.2 | COPY_HV_SSE_RND1 |
| 410 : | edgomez | 1.3 | ret |
| No admin address has been configured | ViewVC Help |
| Powered by ViewVC 1.0.4 |