Parent Directory
| 
Revision Log
Revision 1877 - (view) (download)
| 1 : | edgomez | 1382 | ;/***************************************************************************** |
| 2 : | Isibaar | 262 | ; * |
| 3 : | edgomez | 1382 | ; * XVID MPEG-4 VIDEO CODEC |
| 4 : | ; * - mmx 8x8 block-based halfpel interpolation - | ||
| 5 : | Isibaar | 262 | ; * |
| 6 : | edgomez | 1382 | ; * Copyright(C) 2001 Peter Ross <pross@xvid.org> |
| 7 : | Isibaar | 1795 | ; * 2002-2008 Michael Militzer <michael@xvid.org> |
| 8 : | Isibaar | 262 | ; * |
| 9 : | edgomez | 1382 | ; * This program is free software ; you can redistribute it and/or modify |
| 10 : | ; * it under the terms of the GNU General Public License as published by | ||
| 11 : | ; * the Free Software Foundation ; either version 2 of the License, or | ||
| 12 : | ; * (at your option) any later version. | ||
| 13 : | Isibaar | 262 | ; * |
| 14 : | edgomez | 1382 | ; * This program is distributed in the hope that it will be useful, |
| 15 : | ; * but WITHOUT ANY WARRANTY ; without even the implied warranty of | ||
| 16 : | ; * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | ||
| 17 : | ; * GNU General Public License for more details. | ||
| 18 : | Isibaar | 262 | ; * |
| 19 : | edgomez | 1382 | ; * You should have received a copy of the GNU General Public License |
| 20 : | ; * along with this program ; if not, write to the Free Software | ||
| 21 : | ; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | ||
| 22 : | Isibaar | 262 | ; * |
| 23 : | edgomez | 1382 | ; ****************************************************************************/ |
| 24 : | Isibaar | 262 | |
| 25 : | Isibaar | 1795 | %include "nasm.inc" |
| 26 : | edgomez | 851 | |
| 27 : | edgomez | 1382 | ;============================================================================= |
| 28 : | ; Read only data | ||
| 29 : | ;============================================================================= | ||
| 30 : | Isibaar | 262 | |
| 31 : | Isibaar | 1795 | DATA |
| 32 : | Isibaar | 262 | |
| 33 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 34 : | edgomez | 851 | ; (16 - r) rounding table |
| 35 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 36 : | edgomez | 851 | |
| 37 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 38 : | edgomez | 1382 | rounding_lowpass_mmx: |
| 39 : | times 4 dw 16 | ||
| 40 : | times 4 dw 15 | ||
| 41 : | edgomez | 851 | |
| 42 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 43 : | Isibaar | 262 | ; (1 - r) rounding table |
| 44 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 45 : | Isibaar | 262 | |
| 46 : | edgomez | 1382 | rounding1_mmx: |
| 47 : | times 4 dw 1 | ||
| 48 : | times 4 dw 0 | ||
| 49 : | Isibaar | 262 | |
| 50 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 51 : | ; (2 - r) rounding table | ||
| 52 : | ;----------------------------------------------------------------------------- | ||
| 53 : | Isibaar | 262 | |
| 54 : | edgomez | 1382 | rounding2_mmx: |
| 55 : | times 4 dw 2 | ||
| 56 : | times 4 dw 1 | ||
| 57 : | Isibaar | 262 | |
| 58 : | edgomez | 1382 | mmx_one: |
| 59 : | times 8 db 1 | ||
| 60 : | Isibaar | 262 | |
| 61 : | edgomez | 1382 | mmx_two: |
| 62 : | times 8 db 2 | ||
| 63 : | edgomez | 851 | |
| 64 : | edgomez | 1382 | mmx_three: |
| 65 : | times 8 db 3 | ||
| 66 : | edgomez | 851 | |
| 67 : | edgomez | 1382 | mmx_five: |
| 68 : | times 4 dw 5 | ||
| 69 : | edgomez | 851 | |
| 70 : | edgomez | 1382 | mmx_mask: |
| 71 : | times 8 db 254 | ||
| 72 : | edgomez | 851 | |
| 73 : | edgomez | 1382 | mmx_mask2: |
| 74 : | times 8 db 252 | ||
| 75 : | edgomez | 851 | |
| 76 : | edgomez | 1382 | ;============================================================================= |
| 77 : | ; Code | ||
| 78 : | ;============================================================================= | ||
| 79 : | Isibaar | 262 | |
| 80 : | Isibaar | 1844 | TEXT |
| 81 : | edgomez | 1382 | |
| 82 : | cglobal interpolate8x8_halfpel_h_mmx | ||
| 83 : | cglobal interpolate8x8_halfpel_v_mmx | ||
| 84 : | cglobal interpolate8x8_halfpel_hv_mmx | ||
| 85 : | edgomez | 1530 | |
| 86 : | suxen_drol | 1632 | cglobal interpolate8x4_halfpel_h_mmx |
| 87 : | cglobal interpolate8x4_halfpel_v_mmx | ||
| 88 : | cglobal interpolate8x4_halfpel_hv_mmx | ||
| 89 : | |||
| 90 : | edgomez | 1382 | cglobal interpolate8x8_avg4_mmx |
| 91 : | cglobal interpolate8x8_avg2_mmx | ||
| 92 : | edgomez | 1530 | |
| 93 : | edgomez | 1382 | cglobal interpolate8x8_6tap_lowpass_h_mmx |
| 94 : | cglobal interpolate8x8_6tap_lowpass_v_mmx | ||
| 95 : | |||
| 96 : | edgomez | 1530 | cglobal interpolate8x8_halfpel_add_mmx |
| 97 : | cglobal interpolate8x8_halfpel_h_add_mmx | ||
| 98 : | cglobal interpolate8x8_halfpel_v_add_mmx | ||
| 99 : | cglobal interpolate8x8_halfpel_hv_add_mmx | ||
| 100 : | |||
| 101 : | Isibaar | 262 | %macro CALC_AVG 6 |
| 102 : | edgomez | 1382 | punpcklbw %3, %6 |
| 103 : | punpckhbw %4, %6 | ||
| 104 : | Isibaar | 262 | |
| 105 : | edgomez | 1382 | paddusw %1, %3 ; mm01 += mm23 |
| 106 : | paddusw %2, %4 | ||
| 107 : | paddusw %1, %5 ; mm01 += rounding | ||
| 108 : | paddusw %2, %5 | ||
| 109 : | Isibaar | 262 | |
| 110 : | edgomez | 1382 | psrlw %1, 1 ; mm01 >>= 1 |
| 111 : | psrlw %2, 1 | ||
| 112 : | Isibaar | 262 | %endmacro |
| 113 : | |||
| 114 : | |||
| 115 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 116 : | Isibaar | 262 | ; |
| 117 : | ; void interpolate8x8_halfpel_h_mmx(uint8_t * const dst, | ||
| 118 : | edgomez | 1382 | ; const uint8_t * const src, |
| 119 : | ; const uint32_t stride, | ||
| 120 : | ; const uint32_t rounding); | ||
| 121 : | Isibaar | 262 | ; |
| 122 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 123 : | Isibaar | 262 | |
| 124 : | %macro COPY_H_MMX 0 | ||
| 125 : | Isibaar | 1795 | movq mm0, [TMP0] |
| 126 : | movq mm2, [TMP0 + 1] | ||
| 127 : | edgomez | 1382 | movq mm1, mm0 |
| 128 : | movq mm3, mm2 | ||
| 129 : | Isibaar | 262 | |
| 130 : | edgomez | 1382 | punpcklbw mm0, mm6 ; mm01 = [src] |
| 131 : | punpckhbw mm1, mm6 ; mm23 = [src + 1] | ||
| 132 : | Isibaar | 262 | |
| 133 : | edgomez | 1382 | CALC_AVG mm0, mm1, mm2, mm3, mm7, mm6 |
| 134 : | Isibaar | 262 | |
| 135 : | edgomez | 1382 | packuswb mm0, mm1 |
| 136 : | Isibaar | 1795 | movq [_EAX], mm0 ; [dst] = mm01 |
| 137 : | Isibaar | 262 | |
| 138 : | Isibaar | 1795 | add TMP0, TMP1 ; src += stride |
| 139 : | add _EAX, TMP1 ; dst += stride | ||
| 140 : | Isibaar | 262 | %endmacro |
| 141 : | |||
| 142 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 143 : | edgomez | 1382 | interpolate8x8_halfpel_h_mmx: |
| 144 : | Isibaar | 262 | |
| 145 : | Isibaar | 1795 | mov _EAX, prm4 ; rounding |
| 146 : | lea TMP0, [rounding1_mmx] | ||
| 147 : | movq mm7, [TMP0 + _EAX * 8] | ||
| 148 : | Isibaar | 262 | |
| 149 : | Isibaar | 1795 | mov _EAX, prm1 ; dst |
| 150 : | mov TMP0, prm2 ; src | ||
| 151 : | mov TMP1, prm3 ; stride | ||
| 152 : | Isibaar | 262 | |
| 153 : | Isibaar | 1795 | pxor mm6, mm6 ; zero |
| 154 : | Isibaar | 262 | |
| 155 : | edgomez | 1382 | COPY_H_MMX |
| 156 : | COPY_H_MMX | ||
| 157 : | COPY_H_MMX | ||
| 158 : | COPY_H_MMX | ||
| 159 : | COPY_H_MMX | ||
| 160 : | COPY_H_MMX | ||
| 161 : | COPY_H_MMX | ||
| 162 : | COPY_H_MMX | ||
| 163 : | Isibaar | 262 | |
| 164 : | edgomez | 1382 | ret |
| 165 : | Isibaar | 1793 | ENDFUNC |
| 166 : | Isibaar | 262 | |
| 167 : | |||
| 168 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 169 : | Isibaar | 262 | ; |
| 170 : | ; void interpolate8x8_halfpel_v_mmx(uint8_t * const dst, | ||
| 171 : | edgomez | 1382 | ; const uint8_t * const src, |
| 172 : | ; const uint32_t stride, | ||
| 173 : | ; const uint32_t rounding); | ||
| 174 : | Isibaar | 262 | ; |
| 175 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 176 : | Isibaar | 262 | |
| 177 : | %macro COPY_V_MMX 0 | ||
| 178 : | Isibaar | 1795 | movq mm0, [TMP0] |
| 179 : | movq mm2, [TMP0 + TMP1] | ||
| 180 : | edgomez | 1382 | movq mm1, mm0 |
| 181 : | movq mm3, mm2 | ||
| 182 : | Isibaar | 262 | |
| 183 : | edgomez | 1382 | punpcklbw mm0, mm6 ; mm01 = [src] |
| 184 : | punpckhbw mm1, mm6 ; mm23 = [src + 1] | ||
| 185 : | Isibaar | 262 | |
| 186 : | edgomez | 1382 | CALC_AVG mm0, mm1, mm2, mm3, mm7, mm6 |
| 187 : | Isibaar | 262 | |
| 188 : | edgomez | 1382 | packuswb mm0, mm1 |
| 189 : | Isibaar | 1795 | movq [_EAX], mm0 ; [dst] = mm01 |
| 190 : | Isibaar | 262 | |
| 191 : | Isibaar | 1795 | add TMP0, TMP1 ; src += stride |
| 192 : | add _EAX, TMP1 ; dst += stride | ||
| 193 : | Isibaar | 262 | %endmacro |
| 194 : | |||
| 195 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 196 : | edgomez | 1382 | interpolate8x8_halfpel_v_mmx: |
| 197 : | Isibaar | 262 | |
| 198 : | Isibaar | 1795 | mov _EAX, prm4 ; rounding |
| 199 : | lea TMP0, [rounding1_mmx] | ||
| 200 : | movq mm7, [TMP0 + _EAX * 8] | ||
| 201 : | Isibaar | 262 | |
| 202 : | Isibaar | 1795 | mov _EAX, prm1 ; dst |
| 203 : | mov TMP0, prm2 ; src | ||
| 204 : | mov TMP1, prm3 ; stride | ||
| 205 : | Isibaar | 262 | |
| 206 : | Isibaar | 1795 | pxor mm6, mm6 ; zero |
| 207 : | Isibaar | 262 | |
| 208 : | |||
| 209 : | edgomez | 1382 | COPY_V_MMX |
| 210 : | COPY_V_MMX | ||
| 211 : | COPY_V_MMX | ||
| 212 : | COPY_V_MMX | ||
| 213 : | COPY_V_MMX | ||
| 214 : | COPY_V_MMX | ||
| 215 : | COPY_V_MMX | ||
| 216 : | COPY_V_MMX | ||
| 217 : | Isibaar | 262 | |
| 218 : | edgomez | 1382 | ret |
| 219 : | Isibaar | 1793 | ENDFUNC |
| 220 : | Isibaar | 262 | |
| 221 : | edgomez | 1382 | |
| 222 : | ;----------------------------------------------------------------------------- | ||
| 223 : | Isibaar | 262 | ; |
| 224 : | ; void interpolate8x8_halfpel_hv_mmx(uint8_t * const dst, | ||
| 225 : | edgomez | 1382 | ; const uint8_t * const src, |
| 226 : | ; const uint32_t stride, | ||
| 227 : | ; const uint32_t rounding); | ||
| 228 : | Isibaar | 262 | ; |
| 229 : | ; | ||
| 230 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 231 : | Isibaar | 262 | |
| 232 : | %macro COPY_HV_MMX 0 | ||
| 233 : | edgomez | 1382 | ; current row |
| 234 : | Isibaar | 1795 | movq mm0, [TMP0] |
| 235 : | movq mm2, [TMP0 + 1] | ||
| 236 : | Isibaar | 262 | |
| 237 : | edgomez | 1382 | movq mm1, mm0 |
| 238 : | movq mm3, mm2 | ||
| 239 : | Isibaar | 262 | |
| 240 : | edgomez | 1382 | punpcklbw mm0, mm6 ; mm01 = [src] |
| 241 : | punpcklbw mm2, mm6 ; mm23 = [src + 1] | ||
| 242 : | punpckhbw mm1, mm6 | ||
| 243 : | punpckhbw mm3, mm6 | ||
| 244 : | Isibaar | 262 | |
| 245 : | edgomez | 1382 | paddusw mm0, mm2 ; mm01 += mm23 |
| 246 : | paddusw mm1, mm3 | ||
| 247 : | Isibaar | 262 | |
| 248 : | edgomez | 1382 | ; next row |
| 249 : | Isibaar | 1795 | movq mm4, [TMP0 + TMP1] |
| 250 : | movq mm2, [TMP0 + TMP1 + 1] | ||
| 251 : | Isibaar | 262 | |
| 252 : | edgomez | 1382 | movq mm5, mm4 |
| 253 : | movq mm3, mm2 | ||
| 254 : | Isibaar | 262 | |
| 255 : | edgomez | 1382 | punpcklbw mm4, mm6 ; mm45 = [src + stride] |
| 256 : | punpcklbw mm2, mm6 ; mm23 = [src + stride + 1] | ||
| 257 : | punpckhbw mm5, mm6 | ||
| 258 : | punpckhbw mm3, mm6 | ||
| 259 : | Isibaar | 262 | |
| 260 : | edgomez | 1382 | paddusw mm4, mm2 ; mm45 += mm23 |
| 261 : | paddusw mm5, mm3 | ||
| 262 : | Isibaar | 262 | |
| 263 : | edgomez | 1382 | ; add current + next row |
| 264 : | paddusw mm0, mm4 ; mm01 += mm45 | ||
| 265 : | paddusw mm1, mm5 | ||
| 266 : | paddusw mm0, mm7 ; mm01 += rounding2 | ||
| 267 : | paddusw mm1, mm7 | ||
| 268 : | Isibaar | 262 | |
| 269 : | edgomez | 1382 | psrlw mm0, 2 ; mm01 >>= 2 |
| 270 : | psrlw mm1, 2 | ||
| 271 : | Isibaar | 262 | |
| 272 : | edgomez | 1382 | packuswb mm0, mm1 |
| 273 : | Isibaar | 1795 | movq [_EAX], mm0 ; [dst] = mm01 |
| 274 : | Isibaar | 262 | |
| 275 : | Isibaar | 1795 | add TMP0, TMP1 ; src += stride |
| 276 : | add _EAX, TMP1 ; dst += stride | ||
| 277 : | Isibaar | 262 | %endmacro |
| 278 : | |||
| 279 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 280 : | edgomez | 1382 | interpolate8x8_halfpel_hv_mmx: |
| 281 : | Isibaar | 262 | |
| 282 : | Isibaar | 1795 | mov _EAX, prm4 ; rounding |
| 283 : | lea TMP0, [rounding2_mmx] | ||
| 284 : | movq mm7, [TMP0 + _EAX * 8] | ||
| 285 : | Isibaar | 262 | |
| 286 : | Isibaar | 1795 | mov _EAX, prm1 ; dst |
| 287 : | mov TMP0, prm2 ; src | ||
| 288 : | Isibaar | 262 | |
| 289 : | Isibaar | 1795 | pxor mm6, mm6 ; zero |
| 290 : | Isibaar | 262 | |
| 291 : | Isibaar | 1795 | mov TMP1, prm3 ; stride |
| 292 : | Isibaar | 262 | |
| 293 : | edgomez | 1382 | COPY_HV_MMX |
| 294 : | COPY_HV_MMX | ||
| 295 : | COPY_HV_MMX | ||
| 296 : | COPY_HV_MMX | ||
| 297 : | COPY_HV_MMX | ||
| 298 : | COPY_HV_MMX | ||
| 299 : | COPY_HV_MMX | ||
| 300 : | COPY_HV_MMX | ||
| 301 : | edgomez | 851 | |
| 302 : | edgomez | 1382 | ret |
| 303 : | Isibaar | 1793 | ENDFUNC |
| 304 : | edgomez | 1382 | |
| 305 : | ;----------------------------------------------------------------------------- | ||
| 306 : | edgomez | 851 | ; |
| 307 : | suxen_drol | 1632 | ; void interpolate8x4_halfpel_h_mmx(uint8_t * const dst, |
| 308 : | ; const uint8_t * const src, | ||
| 309 : | ; const uint32_t stride, | ||
| 310 : | ; const uint32_t rounding); | ||
| 311 : | ; | ||
| 312 : | ;----------------------------------------------------------------------------- | ||
| 313 : | |||
| 314 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 315 : | suxen_drol | 1632 | interpolate8x4_halfpel_h_mmx: |
| 316 : | |||
| 317 : | Isibaar | 1795 | mov _EAX, prm4 ; rounding |
| 318 : | lea TMP0, [rounding1_mmx] | ||
| 319 : | movq mm7, [TMP0 + _EAX * 8] | ||
| 320 : | suxen_drol | 1632 | |
| 321 : | Isibaar | 1795 | mov _EAX, prm1 ; dst |
| 322 : | mov TMP0, prm2 ; src | ||
| 323 : | mov TMP1, prm3 ; stride | ||
| 324 : | suxen_drol | 1632 | |
| 325 : | Isibaar | 1795 | pxor mm6, mm6 ; zero |
| 326 : | suxen_drol | 1632 | |
| 327 : | COPY_H_MMX | ||
| 328 : | COPY_H_MMX | ||
| 329 : | COPY_H_MMX | ||
| 330 : | COPY_H_MMX | ||
| 331 : | |||
| 332 : | ret | ||
| 333 : | Isibaar | 1793 | ENDFUNC |
| 334 : | suxen_drol | 1632 | |
| 335 : | |||
| 336 : | ;----------------------------------------------------------------------------- | ||
| 337 : | ; | ||
| 338 : | ; void interpolate8x4_halfpel_v_mmx(uint8_t * const dst, | ||
| 339 : | ; const uint8_t * const src, | ||
| 340 : | ; const uint32_t stride, | ||
| 341 : | ; const uint32_t rounding); | ||
| 342 : | ; | ||
| 343 : | ;----------------------------------------------------------------------------- | ||
| 344 : | |||
| 345 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 346 : | suxen_drol | 1632 | interpolate8x4_halfpel_v_mmx: |
| 347 : | |||
| 348 : | Isibaar | 1795 | mov _EAX, prm4 ; rounding |
| 349 : | lea TMP0, [rounding1_mmx] | ||
| 350 : | movq mm7, [TMP0 + _EAX * 8] | ||
| 351 : | suxen_drol | 1632 | |
| 352 : | Isibaar | 1795 | mov _EAX, prm1 ; dst |
| 353 : | mov TMP0, prm2 ; src | ||
| 354 : | mov TMP1, prm3 ; stride | ||
| 355 : | suxen_drol | 1632 | |
| 356 : | Isibaar | 1795 | pxor mm6, mm6 ; zero |
| 357 : | suxen_drol | 1632 | |
| 358 : | |||
| 359 : | COPY_V_MMX | ||
| 360 : | COPY_V_MMX | ||
| 361 : | COPY_V_MMX | ||
| 362 : | COPY_V_MMX | ||
| 363 : | |||
| 364 : | ret | ||
| 365 : | Isibaar | 1793 | ENDFUNC |
| 366 : | suxen_drol | 1632 | |
| 367 : | |||
| 368 : | ;----------------------------------------------------------------------------- | ||
| 369 : | ; | ||
| 370 : | ; void interpolate8x4_halfpel_hv_mmx(uint8_t * const dst, | ||
| 371 : | ; const uint8_t * const src, | ||
| 372 : | ; const uint32_t stride, | ||
| 373 : | ; const uint32_t rounding); | ||
| 374 : | ; | ||
| 375 : | ; | ||
| 376 : | ;----------------------------------------------------------------------------- | ||
| 377 : | |||
| 378 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 379 : | suxen_drol | 1632 | interpolate8x4_halfpel_hv_mmx: |
| 380 : | |||
| 381 : | Isibaar | 1795 | mov _EAX, prm4 ; rounding |
| 382 : | lea TMP0, [rounding2_mmx] | ||
| 383 : | movq mm7, [TMP0 + _EAX * 8] | ||
| 384 : | suxen_drol | 1632 | |
| 385 : | Isibaar | 1795 | mov _EAX, prm1 ; dst |
| 386 : | mov TMP0, prm2 ; src | ||
| 387 : | suxen_drol | 1632 | |
| 388 : | Isibaar | 1795 | pxor mm6, mm6 ; zero |
| 389 : | suxen_drol | 1632 | |
| 390 : | Isibaar | 1795 | mov TMP1, prm3 ; stride |
| 391 : | suxen_drol | 1632 | |
| 392 : | COPY_HV_MMX | ||
| 393 : | COPY_HV_MMX | ||
| 394 : | COPY_HV_MMX | ||
| 395 : | COPY_HV_MMX | ||
| 396 : | |||
| 397 : | ret | ||
| 398 : | Isibaar | 1793 | ENDFUNC |
| 399 : | suxen_drol | 1632 | |
| 400 : | ;----------------------------------------------------------------------------- | ||
| 401 : | ; | ||
| 402 : | edgomez | 851 | ; void interpolate8x8_avg2_mmx(uint8_t const *dst, |
| 403 : | edgomez | 1382 | ; const uint8_t * const src1, |
| 404 : | ; const uint8_t * const src2, | ||
| 405 : | ; const uint32_t stride, | ||
| 406 : | ; const uint32_t rounding, | ||
| 407 : | ; const uint32_t height); | ||
| 408 : | edgomez | 851 | ; |
| 409 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 410 : | edgomez | 851 | |
| 411 : | %macro AVG2_MMX_RND0 0 | ||
| 412 : | Isibaar | 1795 | movq mm0, [_EAX] ; src1 -> mm0 |
| 413 : | movq mm1, [_EBX] ; src2 -> mm1 | ||
| 414 : | edgomez | 851 | |
| 415 : | Isibaar | 1795 | movq mm4, [_EAX+TMP1] |
| 416 : | movq mm5, [_EBX+TMP1] | ||
| 417 : | edgomez | 851 | |
| 418 : | edgomez | 1382 | movq mm2, mm0 ; src1 -> mm2 |
| 419 : | movq mm3, mm1 ; src2 -> mm3 | ||
| 420 : | edgomez | 851 | |
| 421 : | edgomez | 1382 | pand mm2, mm7 ; isolate the lsb |
| 422 : | pand mm3, mm7 ; isolate the lsb | ||
| 423 : | edgomez | 851 | |
| 424 : | edgomez | 1382 | por mm2, mm3 ; ODD(src1) OR ODD(src2) -> mm2 |
| 425 : | edgomez | 851 | |
| 426 : | edgomez | 1382 | movq mm3, mm4 |
| 427 : | movq mm6, mm5 | ||
| 428 : | edgomez | 851 | |
| 429 : | edgomez | 1382 | pand mm3, mm7 |
| 430 : | pand mm6, mm7 | ||
| 431 : | edgomez | 851 | |
| 432 : | edgomez | 1382 | por mm3, mm6 |
| 433 : | edgomez | 851 | |
| 434 : | edgomez | 1382 | pand mm0, [mmx_mask] |
| 435 : | pand mm1, [mmx_mask] | ||
| 436 : | pand mm4, [mmx_mask] | ||
| 437 : | pand mm5, [mmx_mask] | ||
| 438 : | edgomez | 851 | |
| 439 : | edgomez | 1382 | psrlq mm0, 1 ; src1 / 2 |
| 440 : | psrlq mm1, 1 ; src2 / 2 | ||
| 441 : | edgomez | 851 | |
| 442 : | edgomez | 1382 | psrlq mm4, 1 |
| 443 : | psrlq mm5, 1 | ||
| 444 : | edgomez | 851 | |
| 445 : | edgomez | 1382 | paddb mm0, mm1 ; src1/2 + src2/2 -> mm0 |
| 446 : | paddb mm0, mm2 ; correct rounding error | ||
| 447 : | edgomez | 851 | |
| 448 : | edgomez | 1382 | paddb mm4, mm5 |
| 449 : | paddb mm4, mm3 | ||
| 450 : | edgomez | 851 | |
| 451 : | Isibaar | 1795 | lea _EAX, [_EAX+2*TMP1] |
| 452 : | lea _EBX, [_EBX+2*TMP1] | ||
| 453 : | edgomez | 1382 | |
| 454 : | Isibaar | 1795 | movq [TMP0], mm0 ; (src1 + src2 + 1) / 2 -> dst |
| 455 : | movq [TMP0+TMP1], mm4 | ||
| 456 : | edgomez | 851 | %endmacro |
| 457 : | |||
| 458 : | %macro AVG2_MMX_RND1 0 | ||
| 459 : | Isibaar | 1795 | movq mm0, [_EAX] ; src1 -> mm0 |
| 460 : | movq mm1, [_EBX] ; src2 -> mm1 | ||
| 461 : | edgomez | 851 | |
| 462 : | Isibaar | 1795 | movq mm4, [_EAX+TMP1] |
| 463 : | movq mm5, [_EBX+TMP1] | ||
| 464 : | edgomez | 851 | |
| 465 : | edgomez | 1382 | movq mm2, mm0 ; src1 -> mm2 |
| 466 : | movq mm3, mm1 ; src2 -> mm3 | ||
| 467 : | edgomez | 851 | |
| 468 : | edgomez | 1382 | pand mm2, mm7 ; isolate the lsb |
| 469 : | pand mm3, mm7 ; isolate the lsb | ||
| 470 : | edgomez | 851 | |
| 471 : | edgomez | 1382 | pand mm2, mm3 ; ODD(src1) AND ODD(src2) -> mm2 |
| 472 : | edgomez | 851 | |
| 473 : | edgomez | 1382 | movq mm3, mm4 |
| 474 : | movq mm6, mm5 | ||
| 475 : | edgomez | 851 | |
| 476 : | edgomez | 1382 | pand mm3, mm7 |
| 477 : | pand mm6, mm7 | ||
| 478 : | edgomez | 851 | |
| 479 : | edgomez | 1382 | pand mm3, mm6 |
| 480 : | edgomez | 851 | |
| 481 : | edgomez | 1382 | pand mm0, [mmx_mask] |
| 482 : | pand mm1, [mmx_mask] | ||
| 483 : | pand mm4, [mmx_mask] | ||
| 484 : | pand mm5, [mmx_mask] | ||
| 485 : | edgomez | 851 | |
| 486 : | edgomez | 1382 | psrlq mm0, 1 ; src1 / 2 |
| 487 : | psrlq mm1, 1 ; src2 / 2 | ||
| 488 : | edgomez | 851 | |
| 489 : | edgomez | 1382 | psrlq mm4, 1 |
| 490 : | psrlq mm5, 1 | ||
| 491 : | edgomez | 851 | |
| 492 : | edgomez | 1382 | paddb mm0, mm1 ; src1/2 + src2/2 -> mm0 |
| 493 : | paddb mm0, mm2 ; correct rounding error | ||
| 494 : | edgomez | 851 | |
| 495 : | edgomez | 1382 | paddb mm4, mm5 |
| 496 : | paddb mm4, mm3 | ||
| 497 : | |||
| 498 : | Isibaar | 1795 | lea _EAX, [_EAX+2*TMP1] |
| 499 : | lea _EBX, [_EBX+2*TMP1] | ||
| 500 : | edgomez | 1382 | |
| 501 : | Isibaar | 1795 | movq [TMP0], mm0 ; (src1 + src2 + 1) / 2 -> dst |
| 502 : | movq [TMP0+TMP1], mm4 | ||
| 503 : | edgomez | 851 | %endmacro |
| 504 : | |||
| 505 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 506 : | edgomez | 1382 | interpolate8x8_avg2_mmx: |
| 507 : | edgomez | 851 | |
| 508 : | Isibaar | 1803 | mov eax, prm5d ; rounding |
| 509 : | Isibaar | 1795 | test _EAX, _EAX |
| 510 : | edgomez | 851 | |
| 511 : | edgomez | 1382 | jnz near .rounding1 |
| 512 : | edgomez | 851 | |
| 513 : | Isibaar | 1803 | mov eax, prm6d ; height -> _EAX |
| 514 : | Isibaar | 1795 | sub _EAX, 8 |
| 515 : | edgomez | 851 | |
| 516 : | Isibaar | 1795 | mov TMP0, prm1 ; dst -> edi |
| 517 : | mov _EAX, prm2 ; src1 -> esi | ||
| 518 : | mov TMP1, prm4 ; stride -> TMP1 | ||
| 519 : | edgomez | 851 | |
| 520 : | Isibaar | 1795 | push _EBX |
| 521 : | %ifdef ARCH_IS_X86_64 | ||
| 522 : | mov _EBX, prm3 | ||
| 523 : | %else | ||
| 524 : | mov _EBX, [esp + 4 + 12] ; src2 -> eax | ||
| 525 : | %endif | ||
| 526 : | |||
| 527 : | edgomez | 1382 | movq mm7, [mmx_one] |
| 528 : | edgomez | 851 | |
| 529 : | edgomez | 1382 | jz near .start0 |
| 530 : | edgomez | 851 | |
| 531 : | edgomez | 1382 | AVG2_MMX_RND0 |
| 532 : | Isibaar | 1795 | lea TMP0, [TMP0+2*TMP1] |
| 533 : | edgomez | 1382 | |
| 534 : | Isibaar | 1793 | .start0: |
| 535 : | edgomez | 851 | |
| 536 : | edgomez | 1382 | AVG2_MMX_RND0 |
| 537 : | Isibaar | 1795 | lea TMP0, [TMP0+2*TMP1] |
| 538 : | edgomez | 1382 | AVG2_MMX_RND0 |
| 539 : | Isibaar | 1795 | lea TMP0, [TMP0+2*TMP1] |
| 540 : | edgomez | 1382 | AVG2_MMX_RND0 |
| 541 : | Isibaar | 1795 | lea TMP0, [TMP0+2*TMP1] |
| 542 : | edgomez | 1382 | AVG2_MMX_RND0 |
| 543 : | edgomez | 851 | |
| 544 : | Isibaar | 1795 | pop _EBX |
| 545 : | edgomez | 1382 | ret |
| 546 : | |||
| 547 : | Isibaar | 1793 | .rounding1: |
| 548 : | Isibaar | 1803 | mov eax, prm6d ; height -> _EAX |
| 549 : | Isibaar | 1795 | sub _EAX, 8 |
| 550 : | edgomez | 851 | |
| 551 : | Isibaar | 1795 | mov TMP0, prm1 ; dst -> edi |
| 552 : | mov _EAX, prm2 ; src1 -> esi | ||
| 553 : | mov TMP1, prm4 ; stride -> TMP1 | ||
| 554 : | edgomez | 851 | |
| 555 : | Isibaar | 1795 | push _EBX |
| 556 : | %ifdef ARCH_IS_X86_64 | ||
| 557 : | mov _EBX, prm3 | ||
| 558 : | %else | ||
| 559 : | mov _EBX, [esp + 4 + 12] ; src2 -> eax | ||
| 560 : | %endif | ||
| 561 : | |||
| 562 : | edgomez | 1382 | movq mm7, [mmx_one] |
| 563 : | edgomez | 851 | |
| 564 : | edgomez | 1382 | jz near .start1 |
| 565 : | edgomez | 851 | |
| 566 : | edgomez | 1382 | AVG2_MMX_RND1 |
| 567 : | Isibaar | 1795 | lea TMP0, [TMP0+2*TMP1] |
| 568 : | edgomez | 851 | |
| 569 : | Isibaar | 1793 | .start1: |
| 570 : | edgomez | 851 | |
| 571 : | edgomez | 1382 | AVG2_MMX_RND1 |
| 572 : | Isibaar | 1795 | lea TMP0, [TMP0+2*TMP1] |
| 573 : | edgomez | 1382 | AVG2_MMX_RND1 |
| 574 : | Isibaar | 1795 | lea TMP0, [TMP0+2*TMP1] |
| 575 : | edgomez | 1382 | AVG2_MMX_RND1 |
| 576 : | Isibaar | 1795 | lea TMP0, [TMP0+2*TMP1] |
| 577 : | edgomez | 1382 | AVG2_MMX_RND1 |
| 578 : | edgomez | 851 | |
| 579 : | Isibaar | 1795 | pop _EBX |
| 580 : | edgomez | 1382 | ret |
| 581 : | Isibaar | 1793 | ENDFUNC |
| 582 : | edgomez | 851 | |
| 583 : | |||
| 584 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 585 : | edgomez | 851 | ; |
| 586 : | ; void interpolate8x8_avg4_mmx(uint8_t const *dst, | ||
| 587 : | edgomez | 1382 | ; const uint8_t * const src1, |
| 588 : | ; const uint8_t * const src2, | ||
| 589 : | ; const uint8_t * const src3, | ||
| 590 : | ; const uint8_t * const src4, | ||
| 591 : | ; const uint32_t stride, | ||
| 592 : | ; const uint32_t rounding); | ||
| 593 : | edgomez | 851 | ; |
| 594 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 595 : | edgomez | 851 | |
| 596 : | %macro AVG4_MMX_RND0 0 | ||
| 597 : | Isibaar | 1795 | movq mm0, [_EAX] ; src1 -> mm0 |
| 598 : | movq mm1, [_EBX] ; src2 -> mm1 | ||
| 599 : | edgomez | 851 | |
| 600 : | edgomez | 1382 | movq mm2, mm0 |
| 601 : | movq mm3, mm1 | ||
| 602 : | edgomez | 851 | |
| 603 : | edgomez | 1382 | pand mm2, [mmx_three] |
| 604 : | pand mm3, [mmx_three] | ||
| 605 : | edgomez | 851 | |
| 606 : | edgomez | 1382 | pand mm0, [mmx_mask2] |
| 607 : | pand mm1, [mmx_mask2] | ||
| 608 : | edgomez | 851 | |
| 609 : | edgomez | 1382 | psrlq mm0, 2 |
| 610 : | psrlq mm1, 2 | ||
| 611 : | edgomez | 851 | |
| 612 : | Isibaar | 1795 | lea _EAX, [_EAX+TMP1] |
| 613 : | lea _EBX, [_EBX+TMP1] | ||
| 614 : | edgomez | 851 | |
| 615 : | edgomez | 1382 | paddb mm0, mm1 |
| 616 : | paddb mm2, mm3 | ||
| 617 : | edgomez | 851 | |
| 618 : | Isibaar | 1795 | movq mm4, [_ESI] ; src3 -> mm0 |
| 619 : | movq mm5, [_EDI] ; src4 -> mm1 | ||
| 620 : | edgomez | 851 | |
| 621 : | edgomez | 1382 | movq mm1, mm4 |
| 622 : | movq mm3, mm5 | ||
| 623 : | edgomez | 851 | |
| 624 : | edgomez | 1382 | pand mm1, [mmx_three] |
| 625 : | pand mm3, [mmx_three] | ||
| 626 : | edgomez | 851 | |
| 627 : | edgomez | 1382 | pand mm4, [mmx_mask2] |
| 628 : | pand mm5, [mmx_mask2] | ||
| 629 : | edgomez | 851 | |
| 630 : | edgomez | 1382 | psrlq mm4, 2 |
| 631 : | psrlq mm5, 2 | ||
| 632 : | edgomez | 851 | |
| 633 : | edgomez | 1382 | paddb mm4, mm5 |
| 634 : | paddb mm0, mm4 | ||
| 635 : | edgomez | 851 | |
| 636 : | edgomez | 1382 | paddb mm1, mm3 |
| 637 : | paddb mm2, mm1 | ||
| 638 : | |||
| 639 : | paddb mm2, [mmx_two] | ||
| 640 : | pand mm2, [mmx_mask2] | ||
| 641 : | |||
| 642 : | psrlq mm2, 2 | ||
| 643 : | paddb mm0, mm2 | ||
| 644 : | |||
| 645 : | Isibaar | 1795 | lea _ESI, [_ESI+TMP1] |
| 646 : | lea _EDI, [_EDI+TMP1] | ||
| 647 : | edgomez | 1382 | |
| 648 : | Isibaar | 1795 | movq [TMP0], mm0 ; (src1 + src2 + src3 + src4 + 2) / 4 -> dst |
| 649 : | edgomez | 851 | %endmacro |
| 650 : | |||
| 651 : | %macro AVG4_MMX_RND1 0 | ||
| 652 : | Isibaar | 1795 | movq mm0, [_EAX] ; src1 -> mm0 |
| 653 : | movq mm1, [_EBX] ; src2 -> mm1 | ||
| 654 : | edgomez | 851 | |
| 655 : | edgomez | 1382 | movq mm2, mm0 |
| 656 : | movq mm3, mm1 | ||
| 657 : | edgomez | 851 | |
| 658 : | edgomez | 1382 | pand mm2, [mmx_three] |
| 659 : | pand mm3, [mmx_three] | ||
| 660 : | edgomez | 851 | |
| 661 : | edgomez | 1382 | pand mm0, [mmx_mask2] |
| 662 : | pand mm1, [mmx_mask2] | ||
| 663 : | edgomez | 851 | |
| 664 : | edgomez | 1382 | psrlq mm0, 2 |
| 665 : | psrlq mm1, 2 | ||
| 666 : | edgomez | 851 | |
| 667 : | Isibaar | 1795 | lea _EAX,[_EAX+TMP1] |
| 668 : | lea _EBX,[_EBX+TMP1] | ||
| 669 : | edgomez | 851 | |
| 670 : | edgomez | 1382 | paddb mm0, mm1 |
| 671 : | paddb mm2, mm3 | ||
| 672 : | edgomez | 851 | |
| 673 : | Isibaar | 1795 | movq mm4, [_ESI] ; src3 -> mm0 |
| 674 : | movq mm5, [_EDI] ; src4 -> mm1 | ||
| 675 : | edgomez | 851 | |
| 676 : | edgomez | 1382 | movq mm1, mm4 |
| 677 : | movq mm3, mm5 | ||
| 678 : | edgomez | 851 | |
| 679 : | edgomez | 1382 | pand mm1, [mmx_three] |
| 680 : | pand mm3, [mmx_three] | ||
| 681 : | edgomez | 851 | |
| 682 : | edgomez | 1382 | pand mm4, [mmx_mask2] |
| 683 : | pand mm5, [mmx_mask2] | ||
| 684 : | edgomez | 851 | |
| 685 : | edgomez | 1382 | psrlq mm4, 2 |
| 686 : | psrlq mm5, 2 | ||
| 687 : | edgomez | 851 | |
| 688 : | edgomez | 1382 | paddb mm4, mm5 |
| 689 : | paddb mm0, mm4 | ||
| 690 : | edgomez | 851 | |
| 691 : | edgomez | 1382 | paddb mm1, mm3 |
| 692 : | paddb mm2, mm1 | ||
| 693 : | |||
| 694 : | paddb mm2, [mmx_one] | ||
| 695 : | pand mm2, [mmx_mask2] | ||
| 696 : | |||
| 697 : | psrlq mm2, 2 | ||
| 698 : | paddb mm0, mm2 | ||
| 699 : | |||
| 700 : | Isibaar | 1795 | lea _ESI,[_ESI+TMP1] |
| 701 : | lea _EDI,[_EDI+TMP1] | ||
| 702 : | edgomez | 1382 | |
| 703 : | Isibaar | 1795 | movq [TMP0], mm0 ; (src1 + src2 + src3 + src4 + 2) / 4 -> dst |
| 704 : | edgomez | 851 | %endmacro |
| 705 : | |||
| 706 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 707 : | edgomez | 1382 | interpolate8x8_avg4_mmx: |
| 708 : | edgomez | 851 | |
| 709 : | Isibaar | 1803 | mov eax, prm7d ; rounding |
| 710 : | Isibaar | 1795 | test _EAX, _EAX |
| 711 : | edgomez | 851 | |
| 712 : | Isibaar | 1795 | mov TMP0, prm1 ; dst -> edi |
| 713 : | mov _EAX, prm5 ; src4 -> edi | ||
| 714 : | Isibaar | 1803 | mov TMP1d, prm6d ; stride -> TMP1 |
| 715 : | edgomez | 851 | |
| 716 : | |||
| 717 : | Isibaar | 1795 | push _EBX |
| 718 : | push _EDI | ||
| 719 : | push _ESI | ||
| 720 : | edgomez | 851 | |
| 721 : | Isibaar | 1795 | mov _EDI, _EAX |
| 722 : | |||
| 723 : | %ifdef ARCH_IS_X86_64 | ||
| 724 : | mov _EAX, prm2 | ||
| 725 : | mov _EBX, prm3 | ||
| 726 : | mov _ESI, prm4 | ||
| 727 : | %else | ||
| 728 : | mov _EAX, [esp + 12 + 8] ; src1 -> esi | ||
| 729 : | mov _EBX, [esp + 12 + 12] ; src2 -> _EAX | ||
| 730 : | mov _ESI, [esp + 12 + 16] ; src3 -> esi | ||
| 731 : | %endif | ||
| 732 : | |||
| 733 : | edgomez | 1382 | movq mm7, [mmx_one] |
| 734 : | edgomez | 851 | |
| 735 : | edgomez | 1382 | jnz near .rounding1 |
| 736 : | edgomez | 851 | |
| 737 : | edgomez | 1382 | AVG4_MMX_RND0 |
| 738 : | Isibaar | 1795 | lea TMP0, [TMP0+TMP1] |
| 739 : | edgomez | 1382 | AVG4_MMX_RND0 |
| 740 : | Isibaar | 1795 | lea TMP0, [TMP0+TMP1] |
| 741 : | edgomez | 1382 | AVG4_MMX_RND0 |
| 742 : | Isibaar | 1795 | lea TMP0, [TMP0+TMP1] |
| 743 : | edgomez | 1382 | AVG4_MMX_RND0 |
| 744 : | Isibaar | 1795 | lea TMP0, [TMP0+TMP1] |
| 745 : | edgomez | 1382 | AVG4_MMX_RND0 |
| 746 : | Isibaar | 1795 | lea TMP0, [TMP0+TMP1] |
| 747 : | edgomez | 1382 | AVG4_MMX_RND0 |
| 748 : | Isibaar | 1795 | lea TMP0, [TMP0+TMP1] |
| 749 : | edgomez | 1382 | AVG4_MMX_RND0 |
| 750 : | Isibaar | 1795 | lea TMP0, [TMP0+TMP1] |
| 751 : | edgomez | 1382 | AVG4_MMX_RND0 |
| 752 : | |||
| 753 : | Isibaar | 1795 | pop _ESI |
| 754 : | pop _EDI | ||
| 755 : | pop _EBX | ||
| 756 : | edgomez | 1382 | ret |
| 757 : | |||
| 758 : | Isibaar | 1793 | .rounding1: |
| 759 : | edgomez | 1382 | AVG4_MMX_RND1 |
| 760 : | Isibaar | 1795 | lea TMP0, [TMP0+TMP1] |
| 761 : | edgomez | 1382 | AVG4_MMX_RND1 |
| 762 : | Isibaar | 1795 | lea TMP0, [TMP0+TMP1] |
| 763 : | edgomez | 1382 | AVG4_MMX_RND1 |
| 764 : | Isibaar | 1795 | lea TMP0, [TMP0+TMP1] |
| 765 : | edgomez | 1382 | AVG4_MMX_RND1 |
| 766 : | Isibaar | 1795 | lea TMP0, [TMP0+TMP1] |
| 767 : | edgomez | 1382 | AVG4_MMX_RND1 |
| 768 : | Isibaar | 1795 | lea TMP0, [TMP0+TMP1] |
| 769 : | edgomez | 1382 | AVG4_MMX_RND1 |
| 770 : | Isibaar | 1795 | lea TMP0, [TMP0+TMP1] |
| 771 : | edgomez | 1382 | AVG4_MMX_RND1 |
| 772 : | Isibaar | 1795 | lea TMP0, [TMP0+TMP1] |
| 773 : | edgomez | 1382 | AVG4_MMX_RND1 |
| 774 : | edgomez | 851 | |
| 775 : | Isibaar | 1795 | pop _ESI |
| 776 : | pop _EDI | ||
| 777 : | pop _EBX | ||
| 778 : | edgomez | 1382 | ret |
| 779 : | Isibaar | 1793 | ENDFUNC |
| 780 : | edgomez | 851 | |
| 781 : | |||
| 782 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 783 : | edgomez | 851 | ; |
| 784 : | ; void interpolate8x8_6tap_lowpass_h_mmx(uint8_t const *dst, | ||
| 785 : | edgomez | 1382 | ; const uint8_t * const src, |
| 786 : | ; const uint32_t stride, | ||
| 787 : | ; const uint32_t rounding); | ||
| 788 : | edgomez | 851 | ; |
| 789 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 790 : | edgomez | 851 | |
| 791 : | %macro LOWPASS_6TAP_H_MMX 0 | ||
| 792 : | Isibaar | 1795 | movq mm0, [_EAX] |
| 793 : | movq mm2, [_EAX+1] | ||
| 794 : | edgomez | 851 | |
| 795 : | edgomez | 1382 | movq mm1, mm0 |
| 796 : | movq mm3, mm2 | ||
| 797 : | edgomez | 851 | |
| 798 : | edgomez | 1382 | punpcklbw mm0, mm7 |
| 799 : | punpcklbw mm2, mm7 | ||
| 800 : | edgomez | 851 | |
| 801 : | edgomez | 1382 | punpckhbw mm1, mm7 |
| 802 : | punpckhbw mm3, mm7 | ||
| 803 : | edgomez | 851 | |
| 804 : | edgomez | 1382 | paddw mm0, mm2 |
| 805 : | paddw mm1, mm3 | ||
| 806 : | edgomez | 851 | |
| 807 : | edgomez | 1382 | psllw mm0, 2 |
| 808 : | psllw mm1, 2 | ||
| 809 : | edgomez | 851 | |
| 810 : | Isibaar | 1795 | movq mm2, [_EAX-1] |
| 811 : | movq mm4, [_EAX+2] | ||
| 812 : | edgomez | 851 | |
| 813 : | edgomez | 1382 | movq mm3, mm2 |
| 814 : | movq mm5, mm4 | ||
| 815 : | edgomez | 851 | |
| 816 : | edgomez | 1382 | punpcklbw mm2, mm7 |
| 817 : | punpcklbw mm4, mm7 | ||
| 818 : | edgomez | 851 | |
| 819 : | edgomez | 1382 | punpckhbw mm3, mm7 |
| 820 : | punpckhbw mm5, mm7 | ||
| 821 : | edgomez | 851 | |
| 822 : | edgomez | 1382 | paddw mm2, mm4 |
| 823 : | paddw mm3, mm5 | ||
| 824 : | edgomez | 851 | |
| 825 : | edgomez | 1382 | psubsw mm0, mm2 |
| 826 : | psubsw mm1, mm3 | ||
| 827 : | edgomez | 851 | |
| 828 : | edgomez | 1382 | pmullw mm0, [mmx_five] |
| 829 : | pmullw mm1, [mmx_five] | ||
| 830 : | edgomez | 851 | |
| 831 : | Isibaar | 1795 | movq mm2, [_EAX-2] |
| 832 : | movq mm4, [_EAX+3] | ||
| 833 : | edgomez | 851 | |
| 834 : | edgomez | 1382 | movq mm3, mm2 |
| 835 : | movq mm5, mm4 | ||
| 836 : | edgomez | 851 | |
| 837 : | edgomez | 1382 | punpcklbw mm2, mm7 |
| 838 : | punpcklbw mm4, mm7 | ||
| 839 : | edgomez | 851 | |
| 840 : | edgomez | 1382 | punpckhbw mm3, mm7 |
| 841 : | punpckhbw mm5, mm7 | ||
| 842 : | edgomez | 851 | |
| 843 : | edgomez | 1382 | paddw mm2, mm4 |
| 844 : | paddw mm3, mm5 | ||
| 845 : | edgomez | 851 | |
| 846 : | edgomez | 1382 | paddsw mm0, mm2 |
| 847 : | paddsw mm1, mm3 | ||
| 848 : | edgomez | 851 | |
| 849 : | edgomez | 1382 | paddsw mm0, mm6 |
| 850 : | paddsw mm1, mm6 | ||
| 851 : | edgomez | 851 | |
| 852 : | edgomez | 1382 | psraw mm0, 5 |
| 853 : | psraw mm1, 5 | ||
| 854 : | edgomez | 851 | |
| 855 : | Isibaar | 1795 | lea _EAX, [_EAX+TMP1] |
| 856 : | edgomez | 1382 | packuswb mm0, mm1 |
| 857 : | Isibaar | 1795 | movq [TMP0], mm0 |
| 858 : | edgomez | 851 | %endmacro |
| 859 : | |||
| 860 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 861 : | edgomez | 1382 | interpolate8x8_6tap_lowpass_h_mmx: |
| 862 : | edgomez | 851 | |
| 863 : | Isibaar | 1795 | mov _EAX, prm4 ; rounding |
| 864 : | edgomez | 851 | |
| 865 : | Isibaar | 1795 | lea TMP0, [rounding_lowpass_mmx] |
| 866 : | movq mm6, [TMP0 + _EAX * 8] | ||
| 867 : | edgomez | 851 | |
| 868 : | Isibaar | 1795 | mov TMP0, prm1 ; dst -> edi |
| 869 : | mov _EAX, prm2 ; src -> esi | ||
| 870 : | mov TMP1, prm3 ; stride -> edx | ||
| 871 : | edgomez | 851 | |
| 872 : | edgomez | 1382 | pxor mm7, mm7 |
| 873 : | edgomez | 851 | |
| 874 : | edgomez | 1382 | LOWPASS_6TAP_H_MMX |
| 875 : | Isibaar | 1795 | lea TMP0, [TMP0+TMP1] |
| 876 : | edgomez | 1382 | LOWPASS_6TAP_H_MMX |
| 877 : | Isibaar | 1795 | lea TMP0, [TMP0+TMP1] |
| 878 : | edgomez | 1382 | LOWPASS_6TAP_H_MMX |
| 879 : | Isibaar | 1795 | lea TMP0, [TMP0+TMP1] |
| 880 : | edgomez | 1382 | LOWPASS_6TAP_H_MMX |
| 881 : | Isibaar | 1795 | lea TMP0, [TMP0+TMP1] |
| 882 : | edgomez | 1382 | LOWPASS_6TAP_H_MMX |
| 883 : | Isibaar | 1795 | lea TMP0, [TMP0+TMP1] |
| 884 : | edgomez | 1382 | LOWPASS_6TAP_H_MMX |
| 885 : | Isibaar | 1795 | lea TMP0, [TMP0+TMP1] |
| 886 : | edgomez | 1382 | LOWPASS_6TAP_H_MMX |
| 887 : | Isibaar | 1795 | lea TMP0, [TMP0+TMP1] |
| 888 : | edgomez | 1382 | LOWPASS_6TAP_H_MMX |
| 889 : | edgomez | 851 | |
| 890 : | edgomez | 1382 | ret |
| 891 : | Isibaar | 1793 | ENDFUNC |
| 892 : | edgomez | 1382 | |
| 893 : | ;----------------------------------------------------------------------------- | ||
| 894 : | edgomez | 851 | ; |
| 895 : | ; void interpolate8x8_6tap_lowpass_v_mmx(uint8_t const *dst, | ||
| 896 : | edgomez | 1382 | ; const uint8_t * const src, |
| 897 : | ; const uint32_t stride, | ||
| 898 : | ; const uint32_t rounding); | ||
| 899 : | edgomez | 851 | ; |
| 900 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 901 : | edgomez | 851 | |
| 902 : | %macro LOWPASS_6TAP_V_MMX 0 | ||
| 903 : | Isibaar | 1795 | movq mm0, [_EAX] |
| 904 : | movq mm2, [_EAX+TMP1] | ||
| 905 : | edgomez | 851 | |
| 906 : | edgomez | 1382 | movq mm1, mm0 |
| 907 : | movq mm3, mm2 | ||
| 908 : | edgomez | 851 | |
| 909 : | edgomez | 1382 | punpcklbw mm0, mm7 |
| 910 : | punpcklbw mm2, mm7 | ||
| 911 : | edgomez | 851 | |
| 912 : | edgomez | 1382 | punpckhbw mm1, mm7 |
| 913 : | punpckhbw mm3, mm7 | ||
| 914 : | edgomez | 851 | |
| 915 : | edgomez | 1382 | paddw mm0, mm2 |
| 916 : | paddw mm1, mm3 | ||
| 917 : | edgomez | 851 | |
| 918 : | edgomez | 1382 | psllw mm0, 2 |
| 919 : | psllw mm1, 2 | ||
| 920 : | edgomez | 851 | |
| 921 : | Isibaar | 1795 | movq mm4, [_EAX+2*TMP1] |
| 922 : | sub _EAX, _EBX | ||
| 923 : | movq mm2, [_EAX+2*TMP1] | ||
| 924 : | edgomez | 851 | |
| 925 : | edgomez | 1382 | movq mm3, mm2 |
| 926 : | movq mm5, mm4 | ||
| 927 : | edgomez | 851 | |
| 928 : | edgomez | 1382 | punpcklbw mm2, mm7 |
| 929 : | punpcklbw mm4, mm7 | ||
| 930 : | edgomez | 851 | |
| 931 : | edgomez | 1382 | punpckhbw mm3, mm7 |
| 932 : | punpckhbw mm5, mm7 | ||
| 933 : | edgomez | 851 | |
| 934 : | edgomez | 1382 | paddw mm2, mm4 |
| 935 : | paddw mm3, mm5 | ||
| 936 : | edgomez | 851 | |
| 937 : | edgomez | 1382 | psubsw mm0, mm2 |
| 938 : | psubsw mm1, mm3 | ||
| 939 : | edgomez | 851 | |
| 940 : | edgomez | 1382 | pmullw mm0, [mmx_five] |
| 941 : | pmullw mm1, [mmx_five] | ||
| 942 : | edgomez | 851 | |
| 943 : | Isibaar | 1795 | movq mm2, [_EAX+TMP1] |
| 944 : | movq mm4, [_EAX+2*_EBX] | ||
| 945 : | edgomez | 851 | |
| 946 : | edgomez | 1382 | movq mm3, mm2 |
| 947 : | movq mm5, mm4 | ||
| 948 : | edgomez | 851 | |
| 949 : | edgomez | 1382 | punpcklbw mm2, mm7 |
| 950 : | punpcklbw mm4, mm7 | ||
| 951 : | edgomez | 851 | |
| 952 : | edgomez | 1382 | punpckhbw mm3, mm7 |
| 953 : | punpckhbw mm5, mm7 | ||
| 954 : | edgomez | 851 | |
| 955 : | edgomez | 1382 | paddw mm2, mm4 |
| 956 : | paddw mm3, mm5 | ||
| 957 : | edgomez | 851 | |
| 958 : | edgomez | 1382 | paddsw mm0, mm2 |
| 959 : | paddsw mm1, mm3 | ||
| 960 : | edgomez | 851 | |
| 961 : | edgomez | 1382 | paddsw mm0, mm6 |
| 962 : | paddsw mm1, mm6 | ||
| 963 : | edgomez | 851 | |
| 964 : | edgomez | 1382 | psraw mm0, 5 |
| 965 : | psraw mm1, 5 | ||
| 966 : | edgomez | 851 | |
| 967 : | Isibaar | 1795 | lea _EAX, [_EAX+4*TMP1] |
| 968 : | edgomez | 1382 | packuswb mm0, mm1 |
| 969 : | Isibaar | 1795 | movq [TMP0], mm0 |
| 970 : | edgomez | 851 | %endmacro |
| 971 : | |||
| 972 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 973 : | edgomez | 1382 | interpolate8x8_6tap_lowpass_v_mmx: |
| 974 : | edgomez | 851 | |
| 975 : | Isibaar | 1795 | mov _EAX, prm4 ; rounding |
| 976 : | edgomez | 851 | |
| 977 : | Isibaar | 1795 | lea TMP0, [rounding_lowpass_mmx] |
| 978 : | movq mm6, [TMP0 + _EAX * 8] | ||
| 979 : | edgomez | 851 | |
| 980 : | Isibaar | 1795 | mov TMP0, prm1 ; dst -> edi |
| 981 : | mov _EAX, prm2 ; src -> esi | ||
| 982 : | mov TMP1, prm3 ; stride -> edx | ||
| 983 : | edgomez | 851 | |
| 984 : | Isibaar | 1795 | push _EBX |
| 985 : | Isibaar | 1834 | lea _EBX, [TMP1+TMP1*2] |
| 986 : | edgomez | 851 | |
| 987 : | edgomez | 1382 | pxor mm7, mm7 |
| 988 : | edgomez | 851 | |
| 989 : | edgomez | 1382 | LOWPASS_6TAP_V_MMX |
| 990 : | Isibaar | 1795 | lea TMP0, [TMP0+TMP1] |
| 991 : | edgomez | 1382 | LOWPASS_6TAP_V_MMX |
| 992 : | Isibaar | 1795 | lea TMP0, [TMP0+TMP1] |
| 993 : | edgomez | 1382 | LOWPASS_6TAP_V_MMX |
| 994 : | Isibaar | 1795 | lea TMP0, [TMP0+TMP1] |
| 995 : | edgomez | 1382 | LOWPASS_6TAP_V_MMX |
| 996 : | Isibaar | 1795 | lea TMP0, [TMP0+TMP1] |
| 997 : | edgomez | 1382 | LOWPASS_6TAP_V_MMX |
| 998 : | Isibaar | 1795 | lea TMP0, [TMP0+TMP1] |
| 999 : | edgomez | 1382 | LOWPASS_6TAP_V_MMX |
| 1000 : | Isibaar | 1795 | lea TMP0, [TMP0+TMP1] |
| 1001 : | edgomez | 1382 | LOWPASS_6TAP_V_MMX |
| 1002 : | Isibaar | 1795 | lea TMP0, [TMP0+TMP1] |
| 1003 : | edgomez | 1382 | LOWPASS_6TAP_V_MMX |
| 1004 : | edgomez | 851 | |
| 1005 : | Isibaar | 1795 | pop _EBX |
| 1006 : | edgomez | 1382 | ret |
| 1007 : | Isibaar | 1793 | ENDFUNC |
| 1008 : | edgomez | 1530 | |
| 1009 : | ;=========================================================================== | ||
| 1010 : | ; | ||
| 1011 : | ; The next functions combine both source halfpel interpolation step and the | ||
| 1012 : | ; averaging (with rouding) step to avoid wasting memory bandwidth computing | ||
| 1013 : | ; intermediate halfpel images and then averaging them. | ||
| 1014 : | ; | ||
| 1015 : | ;=========================================================================== | ||
| 1016 : | |||
| 1017 : | %macro PROLOG0 0 | ||
| 1018 : | Isibaar | 1795 | mov TMP0, prm1 ; Dst |
| 1019 : | mov _EAX, prm2 ; Src | ||
| 1020 : | mov TMP1, prm3 ; BpS | ||
| 1021 : | edgomez | 1530 | %endmacro |
| 1022 : | |||
| 1023 : | %macro PROLOG 2 ; %1: Rounder, %2 load Dst-Rounder | ||
| 1024 : | pxor mm6, mm6 | ||
| 1025 : | movq mm7, [%1] ; TODO: dangerous! (eax isn't checked) | ||
| 1026 : | %if %2 | ||
| 1027 : | movq mm5, [rounding1_mmx] | ||
| 1028 : | %endif | ||
| 1029 : | |||
| 1030 : | PROLOG0 | ||
| 1031 : | %endmacro | ||
| 1032 : | |||
| 1033 : | ; performs: mm0 == (mm0+mm2) mm1 == (mm1+mm3) | ||
| 1034 : | %macro MIX 0 | ||
| 1035 : | punpcklbw mm0, mm6 | ||
| 1036 : | punpcklbw mm2, mm6 | ||
| 1037 : | punpckhbw mm1, mm6 | ||
| 1038 : | punpckhbw mm3, mm6 | ||
| 1039 : | paddusw mm0, mm2 | ||
| 1040 : | paddusw mm1, mm3 | ||
| 1041 : | %endmacro | ||
| 1042 : | |||
| 1043 : | %macro MIX_DST 0 | ||
| 1044 : | movq mm3, mm2 | ||
| 1045 : | paddusw mm0, mm7 ; rounder | ||
| 1046 : | paddusw mm1, mm7 ; rounder | ||
| 1047 : | punpcklbw mm2, mm6 | ||
| 1048 : | punpckhbw mm3, mm6 | ||
| 1049 : | psrlw mm0, 1 | ||
| 1050 : | psrlw mm1, 1 | ||
| 1051 : | |||
| 1052 : | paddusw mm0, mm2 ; mix Src(mm0/mm1) with Dst(mm2/mm3) | ||
| 1053 : | paddusw mm1, mm3 | ||
| 1054 : | paddusw mm0, mm5 | ||
| 1055 : | paddusw mm1, mm5 | ||
| 1056 : | psrlw mm0, 1 | ||
| 1057 : | psrlw mm1, 1 | ||
| 1058 : | |||
| 1059 : | packuswb mm0, mm1 | ||
| 1060 : | %endmacro | ||
| 1061 : | |||
| 1062 : | %macro MIX2 0 | ||
| 1063 : | punpcklbw mm0, mm6 | ||
| 1064 : | punpcklbw mm2, mm6 | ||
| 1065 : | paddusw mm0, mm2 | ||
| 1066 : | paddusw mm0, mm7 | ||
| 1067 : | punpckhbw mm1, mm6 | ||
| 1068 : | punpckhbw mm3, mm6 | ||
| 1069 : | paddusw mm1, mm7 | ||
| 1070 : | paddusw mm1, mm3 | ||
| 1071 : | psrlw mm0, 1 | ||
| 1072 : | psrlw mm1, 1 | ||
| 1073 : | |||
| 1074 : | packuswb mm0, mm1 | ||
| 1075 : | %endmacro | ||
| 1076 : | |||
| 1077 : | ;=========================================================================== | ||
| 1078 : | ; | ||
| 1079 : | ; void interpolate8x8_halfpel_add_mmx(uint8_t * const dst, | ||
| 1080 : | ; const uint8_t * const src, | ||
| 1081 : | ; const uint32_t stride, | ||
| 1082 : | ; const uint32_t rounding); | ||
| 1083 : | ; | ||
| 1084 : | ; | ||
| 1085 : | ;=========================================================================== | ||
| 1086 : | |||
| 1087 : | %macro ADD_FF_MMX 1 | ||
| 1088 : | Isibaar | 1795 | movq mm0, [_EAX] |
| 1089 : | movq mm2, [TMP0] | ||
| 1090 : | edgomez | 1530 | movq mm1, mm0 |
| 1091 : | movq mm3, mm2 | ||
| 1092 : | %if (%1!=0) | ||
| 1093 : | Isibaar | 1795 | lea _EAX,[_EAX+%1*TMP1] |
| 1094 : | edgomez | 1530 | %endif |
| 1095 : | MIX | ||
| 1096 : | paddusw mm0, mm5 ; rounder | ||
| 1097 : | paddusw mm1, mm5 ; rounder | ||
| 1098 : | psrlw mm0, 1 | ||
| 1099 : | psrlw mm1, 1 | ||
| 1100 : | |||
| 1101 : | packuswb mm0, mm1 | ||
| 1102 : | Isibaar | 1795 | movq [TMP0], mm0 |
| 1103 : | edgomez | 1530 | %if (%1!=0) |
| 1104 : | Isibaar | 1795 | lea TMP0,[TMP0+%1*TMP1] |
| 1105 : | edgomez | 1530 | %endif |
| 1106 : | %endmacro | ||
| 1107 : | |||
| 1108 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 1109 : | edgomez | 1530 | interpolate8x8_halfpel_add_mmx: |
| 1110 : | PROLOG rounding1_mmx, 1 | ||
| 1111 : | ADD_FF_MMX 1 | ||
| 1112 : | ADD_FF_MMX 1 | ||
| 1113 : | ADD_FF_MMX 1 | ||
| 1114 : | ADD_FF_MMX 1 | ||
| 1115 : | ADD_FF_MMX 1 | ||
| 1116 : | ADD_FF_MMX 1 | ||
| 1117 : | ADD_FF_MMX 1 | ||
| 1118 : | ADD_FF_MMX 0 | ||
| 1119 : | ret | ||
| 1120 : | Isibaar | 1793 | ENDFUNC |
| 1121 : | edgomez | 1530 | |
| 1122 : | ;=========================================================================== | ||
| 1123 : | ; | ||
| 1124 : | ; void interpolate8x8_halfpel_h_add_mmx(uint8_t * const dst, | ||
| 1125 : | ; const uint8_t * const src, | ||
| 1126 : | ; const uint32_t stride, | ||
| 1127 : | ; const uint32_t rounding); | ||
| 1128 : | ; | ||
| 1129 : | ; | ||
| 1130 : | ;=========================================================================== | ||
| 1131 : | |||
| 1132 : | %macro ADD_FH_MMX 0 | ||
| 1133 : | Isibaar | 1795 | movq mm0, [_EAX] |
| 1134 : | movq mm2, [_EAX+1] | ||
| 1135 : | edgomez | 1530 | movq mm1, mm0 |
| 1136 : | movq mm3, mm2 | ||
| 1137 : | |||
| 1138 : | Isibaar | 1795 | lea _EAX,[_EAX+TMP1] |
| 1139 : | edgomez | 1530 | |
| 1140 : | MIX | ||
| 1141 : | Isibaar | 1795 | movq mm2, [TMP0] ; prepare mix with Dst[0] |
| 1142 : | edgomez | 1530 | MIX_DST |
| 1143 : | Isibaar | 1795 | movq [TMP0], mm0 |
| 1144 : | edgomez | 1530 | %endmacro |
| 1145 : | |||
| 1146 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 1147 : | edgomez | 1530 | interpolate8x8_halfpel_h_add_mmx: |
| 1148 : | PROLOG rounding1_mmx, 1 | ||
| 1149 : | |||
| 1150 : | ADD_FH_MMX | ||
| 1151 : | Isibaar | 1795 | lea TMP0,[TMP0+TMP1] |
| 1152 : | edgomez | 1530 | ADD_FH_MMX |
| 1153 : | Isibaar | 1795 | lea TMP0,[TMP0+TMP1] |
| 1154 : | edgomez | 1530 | ADD_FH_MMX |
| 1155 : | Isibaar | 1795 | lea TMP0,[TMP0+TMP1] |
| 1156 : | edgomez | 1530 | ADD_FH_MMX |
| 1157 : | Isibaar | 1795 | lea TMP0,[TMP0+TMP1] |
| 1158 : | edgomez | 1530 | ADD_FH_MMX |
| 1159 : | Isibaar | 1795 | lea TMP0,[TMP0+TMP1] |
| 1160 : | edgomez | 1530 | ADD_FH_MMX |
| 1161 : | Isibaar | 1795 | lea TMP0,[TMP0+TMP1] |
| 1162 : | edgomez | 1530 | ADD_FH_MMX |
| 1163 : | Isibaar | 1795 | lea TMP0,[TMP0+TMP1] |
| 1164 : | edgomez | 1530 | ADD_FH_MMX |
| 1165 : | ret | ||
| 1166 : | Isibaar | 1793 | ENDFUNC |
| 1167 : | edgomez | 1530 | |
| 1168 : | ;=========================================================================== | ||
| 1169 : | ; | ||
| 1170 : | ; void interpolate8x8_halfpel_v_add_mmx(uint8_t * const dst, | ||
| 1171 : | ; const uint8_t * const src, | ||
| 1172 : | ; const uint32_t stride, | ||
| 1173 : | ; const uint32_t rounding); | ||
| 1174 : | ; | ||
| 1175 : | ; | ||
| 1176 : | ;=========================================================================== | ||
| 1177 : | |||
| 1178 : | %macro ADD_HF_MMX 0 | ||
| 1179 : | Isibaar | 1795 | movq mm0, [_EAX] |
| 1180 : | movq mm2, [_EAX+TMP1] | ||
| 1181 : | edgomez | 1530 | movq mm1, mm0 |
| 1182 : | movq mm3, mm2 | ||
| 1183 : | |||
| 1184 : | Isibaar | 1795 | lea _EAX,[_EAX+TMP1] |
| 1185 : | edgomez | 1530 | |
| 1186 : | MIX | ||
| 1187 : | Isibaar | 1795 | movq mm2, [TMP0] ; prepare mix with Dst[0] |
| 1188 : | edgomez | 1530 | MIX_DST |
| 1189 : | Isibaar | 1795 | movq [TMP0], mm0 |
| 1190 : | edgomez | 1530 | |
| 1191 : | %endmacro | ||
| 1192 : | |||
| 1193 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 1194 : | edgomez | 1530 | interpolate8x8_halfpel_v_add_mmx: |
| 1195 : | PROLOG rounding1_mmx, 1 | ||
| 1196 : | |||
| 1197 : | ADD_HF_MMX | ||
| 1198 : | Isibaar | 1795 | lea TMP0,[TMP0+TMP1] |
| 1199 : | edgomez | 1530 | ADD_HF_MMX |
| 1200 : | Isibaar | 1795 | lea TMP0,[TMP0+TMP1] |
| 1201 : | edgomez | 1530 | ADD_HF_MMX |
| 1202 : | Isibaar | 1795 | lea TMP0,[TMP0+TMP1] |
| 1203 : | edgomez | 1530 | ADD_HF_MMX |
| 1204 : | Isibaar | 1795 | lea TMP0,[TMP0+TMP1] |
| 1205 : | edgomez | 1530 | ADD_HF_MMX |
| 1206 : | Isibaar | 1795 | lea TMP0,[TMP0+TMP1] |
| 1207 : | edgomez | 1530 | ADD_HF_MMX |
| 1208 : | Isibaar | 1795 | lea TMP0,[TMP0+TMP1] |
| 1209 : | edgomez | 1530 | ADD_HF_MMX |
| 1210 : | Isibaar | 1795 | lea TMP0,[TMP0+TMP1] |
| 1211 : | edgomez | 1530 | ADD_HF_MMX |
| 1212 : | ret | ||
| 1213 : | Isibaar | 1793 | ENDFUNC |
| 1214 : | edgomez | 1530 | |
| 1215 : | ; The trick is to correct the result of 'pavgb' with some combination of the | ||
| 1216 : | ; lsb's of the 4 input values i,j,k,l, and their intermediate 'pavgb' (s and t). | ||
| 1217 : | ; The boolean relations are: | ||
| 1218 : | ; (i+j+k+l+3)/4 = (s+t+1)/2 - (ij&kl)&st | ||
| 1219 : | ; (i+j+k+l+2)/4 = (s+t+1)/2 - (ij|kl)&st | ||
| 1220 : | ; (i+j+k+l+1)/4 = (s+t+1)/2 - (ij&kl)|st | ||
| 1221 : | ; (i+j+k+l+0)/4 = (s+t+1)/2 - (ij|kl)|st | ||
| 1222 : | ; with s=(i+j+1)/2, t=(k+l+1)/2, ij = i^j, kl = k^l, st = s^t. | ||
| 1223 : | |||
| 1224 : | ; Moreover, we process 2 lines at a times, for better overlapping (~15% faster). | ||
| 1225 : | |||
| 1226 : | ;=========================================================================== | ||
| 1227 : | ; | ||
| 1228 : | ; void interpolate8x8_halfpel_hv_add_mmx(uint8_t * const dst, | ||
| 1229 : | ; const uint8_t * const src, | ||
| 1230 : | ; const uint32_t stride, | ||
| 1231 : | ; const uint32_t rounding); | ||
| 1232 : | ; | ||
| 1233 : | ; | ||
| 1234 : | ;=========================================================================== | ||
| 1235 : | |||
| 1236 : | %macro ADD_HH_MMX 0 | ||
| 1237 : | Isibaar | 1795 | lea _EAX,[_EAX+TMP1] |
| 1238 : | edgomez | 1530 | |
| 1239 : | ; transfert prev line to mm0/mm1 | ||
| 1240 : | movq mm0, mm2 | ||
| 1241 : | movq mm1, mm3 | ||
| 1242 : | |||
| 1243 : | ; load new line in mm2/mm3 | ||
| 1244 : | Isibaar | 1795 | movq mm2, [_EAX] |
| 1245 : | movq mm4, [_EAX+1] | ||
| 1246 : | edgomez | 1530 | movq mm3, mm2 |
| 1247 : | movq mm5, mm4 | ||
| 1248 : | |||
| 1249 : | punpcklbw mm2, mm6 | ||
| 1250 : | punpcklbw mm4, mm6 | ||
| 1251 : | paddusw mm2, mm4 | ||
| 1252 : | punpckhbw mm3, mm6 | ||
| 1253 : | punpckhbw mm5, mm6 | ||
| 1254 : | paddusw mm3, mm5 | ||
| 1255 : | |||
| 1256 : | ; mix current line (mm2/mm3) with previous (mm0,mm1); | ||
| 1257 : | ; we'll preserve mm2/mm3 for next line... | ||
| 1258 : | |||
| 1259 : | paddusw mm0, mm2 | ||
| 1260 : | paddusw mm1, mm3 | ||
| 1261 : | |||
| 1262 : | Isibaar | 1795 | movq mm4, [TMP0] ; prepare mix with Dst[0] |
| 1263 : | edgomez | 1530 | movq mm5, mm4 |
| 1264 : | |||
| 1265 : | paddusw mm0, mm7 ; finish mixing current line | ||
| 1266 : | paddusw mm1, mm7 | ||
| 1267 : | |||
| 1268 : | punpcklbw mm4, mm6 | ||
| 1269 : | punpckhbw mm5, mm6 | ||
| 1270 : | |||
| 1271 : | psrlw mm0, 2 | ||
| 1272 : | psrlw mm1, 2 | ||
| 1273 : | |||
| 1274 : | paddusw mm0, mm4 ; mix Src(mm0/mm1) with Dst(mm2/mm3) | ||
| 1275 : | paddusw mm1, mm5 | ||
| 1276 : | |||
| 1277 : | paddusw mm0, [rounding1_mmx] | ||
| 1278 : | paddusw mm1, [rounding1_mmx] | ||
| 1279 : | |||
| 1280 : | psrlw mm0, 1 | ||
| 1281 : | psrlw mm1, 1 | ||
| 1282 : | |||
| 1283 : | packuswb mm0, mm1 | ||
| 1284 : | |||
| 1285 : | Isibaar | 1795 | movq [TMP0], mm0 |
| 1286 : | edgomez | 1530 | %endmacro |
| 1287 : | |||
| 1288 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 1289 : | edgomez | 1530 | interpolate8x8_halfpel_hv_add_mmx: |
| 1290 : | PROLOG rounding2_mmx, 0 ; mm5 is busy. Don't load dst-rounder | ||
| 1291 : | |||
| 1292 : | ; preprocess first line | ||
| 1293 : | Isibaar | 1795 | movq mm0, [_EAX] |
| 1294 : | movq mm2, [_EAX+1] | ||
| 1295 : | edgomez | 1530 | movq mm1, mm0 |
| 1296 : | movq mm3, mm2 | ||
| 1297 : | |||
| 1298 : | punpcklbw mm0, mm6 | ||
| 1299 : | punpcklbw mm2, mm6 | ||
| 1300 : | punpckhbw mm1, mm6 | ||
| 1301 : | punpckhbw mm3, mm6 | ||
| 1302 : | paddusw mm2, mm0 | ||
| 1303 : | paddusw mm3, mm1 | ||
| 1304 : | |||
| 1305 : | ; Input: mm2/mm3 contains the value (Src[0]+Src[1]) of previous line | ||
| 1306 : | |||
| 1307 : | ADD_HH_MMX | ||
| 1308 : | Isibaar | 1795 | lea TMP0,[TMP0+TMP1] |
| 1309 : | edgomez | 1530 | ADD_HH_MMX |
| 1310 : | Isibaar | 1795 | lea TMP0,[TMP0+TMP1] |
| 1311 : | edgomez | 1530 | ADD_HH_MMX |
| 1312 : | Isibaar | 1795 | lea TMP0,[TMP0+TMP1] |
| 1313 : | edgomez | 1530 | ADD_HH_MMX |
| 1314 : | Isibaar | 1795 | lea TMP0,[TMP0+TMP1] |
| 1315 : | edgomez | 1530 | ADD_HH_MMX |
| 1316 : | Isibaar | 1795 | lea TMP0,[TMP0+TMP1] |
| 1317 : | edgomez | 1530 | ADD_HH_MMX |
| 1318 : | Isibaar | 1795 | lea TMP0,[TMP0+TMP1] |
| 1319 : | edgomez | 1530 | ADD_HH_MMX |
| 1320 : | Isibaar | 1795 | lea TMP0,[TMP0+TMP1] |
| 1321 : | edgomez | 1530 | ADD_HH_MMX |
| 1322 : | |||
| 1323 : | ret | ||
| 1324 : | Isibaar | 1793 | ENDFUNC |
| 1325 : | edgomez | 1530 | |
| 1326 : | Isibaar | 1877 | NON_EXEC_STACK |
| No admin address has been configured | ViewVC Help |
| Powered by ViewVC 1.0.4 |