Parent Directory
| 
Revision Log
Revision 651 -
(view)
(download)
Original Path: trunk/xvidcore/src/image/x86_asm/interpolate8x8_xmm.asm
| 1 : | chl | 434 | ;/***************************************************************************** |
| 2 : | Isibaar | 262 | ; * |
| 3 : | chl | 434 | ; * XVID MPEG-4 VIDEO CODEC |
| 4 : | ; * xmm 8x8 block-based halfpel interpolation | ||
| 5 : | Isibaar | 262 | ; * |
| 6 : | chl | 434 | ; * Copyright(C) 2002 Michael Militzer <michael@xvid.org> |
| 7 : | ; * Copyright(C) 2002 -Skal- | ||
| 8 : | Isibaar | 262 | ; * |
| 9 : | edgomez | 651 | ; * This file is part of XviD, a free MPEG-4 video encoder/decoder |
| 10 : | Isibaar | 262 | ; * |
| 11 : | edgomez | 651 | ; * XviD is free software; you can redistribute it and/or modify it |
| 12 : | ; * under the terms of the GNU General Public License as published by | ||
| 13 : | chl | 434 | ; * the Free Software Foundation; either version 2 of the License, or |
| 14 : | ; * (at your option) any later version. | ||
| 15 : | Isibaar | 262 | ; * |
| 16 : | chl | 434 | ; * This program is distributed in the hope that it will be useful, |
| 17 : | ; * but WITHOUT ANY WARRANTY; without even the implied warranty of | ||
| 18 : | ; * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | ||
| 19 : | ; * GNU General Public License for more details. | ||
| 20 : | Isibaar | 262 | ; * |
| 21 : | chl | 434 | ; * You should have received a copy of the GNU General Public License |
| 22 : | ; * along with this program; if not, write to the Free Software | ||
| 23 : | ; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | ||
| 24 : | Isibaar | 262 | ; * |
| 25 : | edgomez | 651 | ; * Under section 8 of the GNU General Public License, the copyright |
| 26 : | ; * holders of XVID explicitly forbid distribution in the following | ||
| 27 : | ; * countries: | ||
| 28 : | ; * | ||
| 29 : | ; * - Japan | ||
| 30 : | ; * - United States of America | ||
| 31 : | ; * | ||
| 32 : | ; * Linking XviD statically or dynamically with other modules is making a | ||
| 33 : | ; * combined work based on XviD. Thus, the terms and conditions of the | ||
| 34 : | ; * GNU General Public License cover the whole combination. | ||
| 35 : | ; * | ||
| 36 : | ; * As a special exception, the copyright holders of XviD give you | ||
| 37 : | ; * permission to link XviD with independent modules that communicate with | ||
| 38 : | ; * XviD solely through the VFW1.1 and DShow interfaces, regardless of the | ||
| 39 : | ; * license terms of these independent modules, and to copy and distribute | ||
| 40 : | ; * the resulting combined work under terms of your choice, provided that | ||
| 41 : | ; * every copy of the combined work is accompanied by a complete copy of | ||
| 42 : | ; * the source code of XviD (the version of XviD used to produce the | ||
| 43 : | ; * combined work), being distributed under the terms of the GNU General | ||
| 44 : | ; * Public License plus this exception. An independent module is a module | ||
| 45 : | ; * which is not derived from or based on XviD. | ||
| 46 : | ; * | ||
| 47 : | ; * Note that people who make modified versions of XviD are not obligated | ||
| 48 : | ; * to grant this special exception for their modified versions; it is | ||
| 49 : | ; * their choice whether to do so. The GNU General Public License gives | ||
| 50 : | ; * permission to release a modified version without this exception; this | ||
| 51 : | ; * exception also makes it possible to release a modified version which | ||
| 52 : | ; * carries forward this exception. | ||
| 53 : | ; * | ||
| 54 : | ; * $Id: interpolate8x8_xmm.asm,v 1.3 2002-11-17 00:20:30 edgomez Exp $ | ||
| 55 : | ; * | ||
| 56 : | chl | 434 | ; ****************************************************************************/ |
| 57 : | Isibaar | 262 | |
| 58 : | bits 32 | ||
| 59 : | |||
| 60 : | %macro cglobal 1 | ||
| 61 : | %ifdef PREFIX | ||
| 62 : | global _%1 | ||
| 63 : | %define %1 _%1 | ||
| 64 : | %else | ||
| 65 : | global %1 | ||
| 66 : | %endif | ||
| 67 : | %endmacro | ||
| 68 : | |||
| 69 : | section .data | ||
| 70 : | |||
| 71 : | |||
| 72 : | align 16 | ||
| 73 : | |||
| 74 : | mmx_one | ||
| 75 : | times 8 db 1 | ||
| 76 : | |||
| 77 : | section .text | ||
| 78 : | |||
| 79 : | cglobal interpolate8x8_halfpel_h_xmm | ||
| 80 : | cglobal interpolate8x8_halfpel_v_xmm | ||
| 81 : | cglobal interpolate8x8_halfpel_hv_xmm | ||
| 82 : | |||
| 83 : | ;=========================================================================== | ||
| 84 : | ; | ||
| 85 : | ; void interpolate8x8_halfpel_h_xmm(uint8_t * const dst, | ||
| 86 : | ; const uint8_t * const src, | ||
| 87 : | ; const uint32_t stride, | ||
| 88 : | ; const uint32_t rounding); | ||
| 89 : | ; | ||
| 90 : | ;=========================================================================== | ||
| 91 : | |||
| 92 : | %macro COPY_H_SSE_RND0 0 | ||
| 93 : | movq mm0, [eax] | ||
| 94 : | pavgb mm0, [eax+1] | ||
| 95 : | movq mm1, [eax+edx] | ||
| 96 : | pavgb mm1, [eax+edx+1] | ||
| 97 : | lea eax,[eax+2*edx] | ||
| 98 : | movq [ecx],mm0 | ||
| 99 : | movq [ecx+edx],mm1 | ||
| 100 : | %endmacro | ||
| 101 : | |||
| 102 : | %macro COPY_H_SSE_RND1 0 | ||
| 103 : | movq mm0, [eax] | ||
| 104 : | movq mm1, [eax+edx] | ||
| 105 : | movq mm4, mm0 | ||
| 106 : | movq mm5, mm1 | ||
| 107 : | movq mm2, [eax+1] | ||
| 108 : | movq mm3, [eax+edx+1] | ||
| 109 : | pavgb mm0, mm2 | ||
| 110 : | pxor mm2, mm4 | ||
| 111 : | pavgb mm1, mm3 | ||
| 112 : | lea eax,[eax+2*edx] | ||
| 113 : | pxor mm3, mm5 | ||
| 114 : | pand mm2, mm7 | ||
| 115 : | pand mm3, mm7 | ||
| 116 : | psubb mm0, mm2 | ||
| 117 : | movq [ecx], mm0 | ||
| 118 : | psubb mm1, mm3 | ||
| 119 : | movq [ecx+edx], mm1 | ||
| 120 : | %endmacro | ||
| 121 : | |||
| 122 : | align 16 | ||
| 123 : | interpolate8x8_halfpel_h_xmm: | ||
| 124 : | |||
| 125 : | mov eax, [esp+16]; rounding | ||
| 126 : | mov ecx, [esp+ 4] ; Dst | ||
| 127 : | test eax,eax | ||
| 128 : | mov eax, [esp+ 8] ; Src | ||
| 129 : | mov edx, [esp+12] ; stride | ||
| 130 : | |||
| 131 : | jnz near .rounding1 | ||
| 132 : | |||
| 133 : | COPY_H_SSE_RND0 | ||
| 134 : | lea ecx,[ecx+2*edx] | ||
| 135 : | COPY_H_SSE_RND0 | ||
| 136 : | lea ecx,[ecx+2*edx] | ||
| 137 : | COPY_H_SSE_RND0 | ||
| 138 : | lea ecx,[ecx+2*edx] | ||
| 139 : | COPY_H_SSE_RND0 | ||
| 140 : | ret | ||
| 141 : | |||
| 142 : | .rounding1 | ||
| 143 : | ; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 | ||
| 144 : | movq mm7, [mmx_one] | ||
| 145 : | COPY_H_SSE_RND1 | ||
| 146 : | lea ecx, [ecx+2*edx] | ||
| 147 : | COPY_H_SSE_RND1 | ||
| 148 : | lea ecx,[ecx+2*edx] | ||
| 149 : | COPY_H_SSE_RND1 | ||
| 150 : | lea ecx,[ecx+2*edx] | ||
| 151 : | COPY_H_SSE_RND1 | ||
| 152 : | ret | ||
| 153 : | |||
| 154 : | ;=========================================================================== | ||
| 155 : | ; | ||
| 156 : | ; void interpolate8x8_halfpel_v_xmm(uint8_t * const dst, | ||
| 157 : | ; const uint8_t * const src, | ||
| 158 : | ; const uint32_t stride, | ||
| 159 : | ; const uint32_t rounding); | ||
| 160 : | ; | ||
| 161 : | ;=========================================================================== | ||
| 162 : | |||
| 163 : | %macro COPY_V_SSE_RND0 0 | ||
| 164 : | movq mm0, [eax] | ||
| 165 : | movq mm1, [eax+edx] | ||
| 166 : | pavgb mm0, mm1 | ||
| 167 : | pavgb mm1, [eax+2*edx] | ||
| 168 : | lea eax,[eax+2*edx] | ||
| 169 : | movq [ecx],mm0 | ||
| 170 : | movq [ecx+edx],mm1 | ||
| 171 : | %endmacro | ||
| 172 : | |||
| 173 : | %macro COPY_V_SSE_RND1 0 | ||
| 174 : | movq mm0, mm2 | ||
| 175 : | movq mm1, [eax] | ||
| 176 : | movq mm2, [eax+edx] | ||
| 177 : | lea eax,[eax+2*edx] | ||
| 178 : | movq mm4, mm0 | ||
| 179 : | movq mm5, mm1 | ||
| 180 : | pavgb mm0, mm1 | ||
| 181 : | pxor mm4, mm1 | ||
| 182 : | pavgb mm1, mm2 | ||
| 183 : | pxor mm5, mm2 | ||
| 184 : | pand mm4, mm7 ; lsb's of (i^j)... | ||
| 185 : | pand mm5, mm7 ; lsb's of (i^j)... | ||
| 186 : | psubb mm0, mm4 ; ...are substracted from result of pavgb | ||
| 187 : | movq [ecx], mm0 | ||
| 188 : | psubb mm1, mm5 ; ...are substracted from result of pavgb | ||
| 189 : | movq [ecx+edx], mm1 | ||
| 190 : | %endmacro | ||
| 191 : | |||
| 192 : | align 16 | ||
| 193 : | interpolate8x8_halfpel_v_xmm: | ||
| 194 : | |||
| 195 : | mov eax, [esp+16]; rounding | ||
| 196 : | mov ecx, [esp+ 4] ; Dst | ||
| 197 : | test eax,eax | ||
| 198 : | mov eax, [esp+ 8] ; Src | ||
| 199 : | mov edx, [esp+12] ; stride | ||
| 200 : | |||
| 201 : | ; we process 2 line at a time | ||
| 202 : | |||
| 203 : | jnz near .rounding1 | ||
| 204 : | |||
| 205 : | COPY_V_SSE_RND0 | ||
| 206 : | lea ecx, [ecx+2*edx] | ||
| 207 : | COPY_V_SSE_RND0 | ||
| 208 : | lea ecx, [ecx+2*edx] | ||
| 209 : | COPY_V_SSE_RND0 | ||
| 210 : | lea ecx, [ecx+2*edx] | ||
| 211 : | COPY_V_SSE_RND0 | ||
| 212 : | ret | ||
| 213 : | |||
| 214 : | .rounding1 | ||
| 215 : | ; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 | ||
| 216 : | movq mm7, [mmx_one] | ||
| 217 : | movq mm2, [eax] ; loop invariant | ||
| 218 : | add eax, edx | ||
| 219 : | |||
| 220 : | COPY_V_SSE_RND1 | ||
| 221 : | lea ecx,[ecx+2*edx] | ||
| 222 : | COPY_V_SSE_RND1 | ||
| 223 : | lea ecx,[ecx+2*edx] | ||
| 224 : | COPY_V_SSE_RND1 | ||
| 225 : | lea ecx,[ecx+2*edx] | ||
| 226 : | COPY_V_SSE_RND1 | ||
| 227 : | ret | ||
| 228 : | |||
| 229 : | ;=========================================================================== | ||
| 230 : | ; | ||
| 231 : | ; void interpolate8x8_halfpel_hv_xmm(uint8_t * const dst, | ||
| 232 : | ; const uint8_t * const src, | ||
| 233 : | ; const uint32_t stride, | ||
| 234 : | ; const uint32_t rounding); | ||
| 235 : | ; | ||
| 236 : | ; | ||
| 237 : | ;=========================================================================== | ||
| 238 : | |||
| 239 : | ; The trick is to correct the result of 'pavgb' with some combination of the | ||
| 240 : | ; lsb's of the 4 input values i,j,k,l, and their intermediate 'pavgb' (s and t). | ||
| 241 : | ; The boolean relations are: | ||
| 242 : | ; (i+j+k+l+3)/4 = (s+t+1)/2 - (ij&kl)&st | ||
| 243 : | ; (i+j+k+l+2)/4 = (s+t+1)/2 - (ij|kl)&st | ||
| 244 : | ; (i+j+k+l+1)/4 = (s+t+1)/2 - (ij&kl)|st | ||
| 245 : | ; (i+j+k+l+0)/4 = (s+t+1)/2 - (ij|kl)|st | ||
| 246 : | ; with s=(i+j+1)/2, t=(k+l+1)/2, ij = i^j, kl = k^l, st = s^t. | ||
| 247 : | |||
| 248 : | ; Moreover, we process 2 lines at a times, for better overlapping (~15% faster). | ||
| 249 : | |||
| 250 : | %macro COPY_HV_SSE_RND0 0 | ||
| 251 : | lea eax,[eax+edx] | ||
| 252 : | |||
| 253 : | movq mm0, [eax] | ||
| 254 : | movq mm1, [eax+1] | ||
| 255 : | |||
| 256 : | movq mm6, mm0 | ||
| 257 : | pavgb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step | ||
| 258 : | lea eax,[eax+edx] | ||
| 259 : | pxor mm1, mm6 ; mm1=(j^k). preserved for next step | ||
| 260 : | |||
| 261 : | por mm3, mm1 ; ij |= jk | ||
| 262 : | movq mm6, mm2 | ||
| 263 : | pxor mm6, mm0 ; mm6 = s^t | ||
| 264 : | pand mm3, mm6 ; (ij|jk) &= st | ||
| 265 : | pavgb mm2, mm0 ; mm2 = (s+t+1)/2 | ||
| 266 : | pand mm3, mm7 ; mask lsb | ||
| 267 : | psubb mm2, mm3 ; apply. | ||
| 268 : | |||
| 269 : | movq [ecx], mm2 | ||
| 270 : | |||
| 271 : | movq mm2, [eax] | ||
| 272 : | movq mm3, [eax+1] | ||
| 273 : | movq mm6, mm2 | ||
| 274 : | pavgb mm2, mm3 ; preserved for next iteration | ||
| 275 : | lea ecx,[ecx+edx] | ||
| 276 : | pxor mm3, mm6 ; preserved for next iteration | ||
| 277 : | |||
| 278 : | por mm1, mm3 | ||
| 279 : | movq mm6, mm0 | ||
| 280 : | pxor mm6, mm2 | ||
| 281 : | pand mm1, mm6 | ||
| 282 : | pavgb mm0, mm2 | ||
| 283 : | |||
| 284 : | pand mm1, mm7 | ||
| 285 : | psubb mm0, mm1 | ||
| 286 : | |||
| 287 : | movq [ecx], mm0 | ||
| 288 : | %endmacro | ||
| 289 : | |||
| 290 : | %macro COPY_HV_SSE_RND1 0 | ||
| 291 : | lea eax,[eax+edx] | ||
| 292 : | |||
| 293 : | movq mm0, [eax] | ||
| 294 : | movq mm1, [eax+1] | ||
| 295 : | |||
| 296 : | movq mm6, mm0 | ||
| 297 : | pavgb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step | ||
| 298 : | lea eax,[eax+edx] | ||
| 299 : | pxor mm1, mm6 ; mm1=(j^k). preserved for next step | ||
| 300 : | |||
| 301 : | pand mm3, mm1 | ||
| 302 : | movq mm6, mm2 | ||
| 303 : | pxor mm6, mm0 | ||
| 304 : | por mm3, mm6 | ||
| 305 : | pavgb mm2, mm0 | ||
| 306 : | pand mm3, mm7 | ||
| 307 : | psubb mm2, mm3 | ||
| 308 : | |||
| 309 : | movq [ecx], mm2 | ||
| 310 : | |||
| 311 : | movq mm2, [eax] | ||
| 312 : | movq mm3, [eax+1] | ||
| 313 : | movq mm6, mm2 | ||
| 314 : | pavgb mm2, mm3 ; preserved for next iteration | ||
| 315 : | lea ecx,[ecx+edx] | ||
| 316 : | pxor mm3, mm6 ; preserved for next iteration | ||
| 317 : | |||
| 318 : | pand mm1, mm3 | ||
| 319 : | movq mm6, mm0 | ||
| 320 : | pxor mm6, mm2 | ||
| 321 : | por mm1, mm6 | ||
| 322 : | pavgb mm0, mm2 | ||
| 323 : | pand mm1, mm7 | ||
| 324 : | psubb mm0, mm1 | ||
| 325 : | |||
| 326 : | movq [ecx], mm0 | ||
| 327 : | %endmacro | ||
| 328 : | |||
| 329 : | align 16 | ||
| 330 : | interpolate8x8_halfpel_hv_xmm: | ||
| 331 : | mov eax, [esp+16] ; rounding | ||
| 332 : | mov ecx, [esp+ 4] ; Dst | ||
| 333 : | test eax,eax | ||
| 334 : | mov eax, [esp+ 8] ; Src | ||
| 335 : | mov edx, [esp+12] ; stride | ||
| 336 : | |||
| 337 : | movq mm7, [mmx_one] | ||
| 338 : | |||
| 339 : | ; loop invariants: mm2=(i+j+1)/2 and mm3= i^j | ||
| 340 : | movq mm2, [eax] | ||
| 341 : | movq mm3, [eax+1] | ||
| 342 : | movq mm6, mm2 | ||
| 343 : | pavgb mm2, mm3 | ||
| 344 : | pxor mm3, mm6 ; mm2/mm3 ready | ||
| 345 : | |||
| 346 : | jnz near .rounding1 | ||
| 347 : | |||
| 348 : | COPY_HV_SSE_RND0 | ||
| 349 : | add ecx, edx | ||
| 350 : | COPY_HV_SSE_RND0 | ||
| 351 : | add ecx, edx | ||
| 352 : | COPY_HV_SSE_RND0 | ||
| 353 : | add ecx, edx | ||
| 354 : | COPY_HV_SSE_RND0 | ||
| 355 : | ret | ||
| 356 : | |||
| 357 : | .rounding1 | ||
| 358 : | COPY_HV_SSE_RND1 | ||
| 359 : | add ecx, edx | ||
| 360 : | COPY_HV_SSE_RND1 | ||
| 361 : | add ecx, edx | ||
| 362 : | COPY_HV_SSE_RND1 | ||
| 363 : | add ecx, edx | ||
| 364 : | COPY_HV_SSE_RND1 | ||
| 365 : | ret |
| No admin address has been configured | ViewVC Help |
| Powered by ViewVC 1.0.4 |