Parent Directory
| 
Revision Log
Revision 1104 - (view) (download)
| 1 : | edgomez | 851 | ;/* |
| 2 : | ; * Simple IDCT MMX | ||
| 3 : | ; * | ||
| 4 : | ; * Copyright (c) 2001, 2002 Michael Niedermayer <michaelni@gmx.at> | ||
| 5 : | ; * | ||
| 6 : | ; * This library is free software; you can redistribute it and/or | ||
| 7 : | ; * modify it under the terms of the GNU Lesser General Public | ||
| 8 : | ; * License as published by the Free Software Foundation; either | ||
| 9 : | ; * version 2 of the License, or (at your option) any later version. | ||
| 10 : | ; * | ||
| 11 : | ; * This library is distributed in the hope that it will be useful, | ||
| 12 : | ; * but WITHOUT ANY WARRANTY; without even the implied warranty of | ||
| 13 : | ; * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | ||
| 14 : | ; * Lesser General Public License for more details. | ||
| 15 : | ; * | ||
| 16 : | ; * You should have received a copy of the GNU Lesser General Public | ||
| 17 : | ; * License along with this library; if not, write to the Free Software | ||
| 18 : | ; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | ||
| 19 : | ; * | ||
| 20 : | ; * Ported to nasm by Peter Ross <pross@xvid.org> | ||
| 21 : | ; */ | ||
| 22 : | |||
| 23 : | bits 32 | ||
| 24 : | |||
| 25 : | |||
| 26 : | ;=========================================================================== | ||
| 27 : | ; data | ||
| 28 : | ;=========================================================================== | ||
| 29 : | |||
| 30 : | %ifdef FORMAT_COFF | ||
| 31 : | section .data | ||
| 32 : | align 8 | ||
| 33 : | %else | ||
| 34 : | section .data data align=8 | ||
| 35 : | %endif | ||
| 36 : | |||
| 37 : | wm1010 dw 0, 0xffff, 0, 0xffff | ||
| 38 : | d40000 dd 0x40000, 0 | ||
| 39 : | |||
| 40 : | |||
| 41 : | %define ROW_SHIFT 11 | ||
| 42 : | %define COL_SHIFT 20 | ||
| 43 : | %define C0 23170 ;cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5 = 23170.475006 | ||
| 44 : | %define C1 22725 ;cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5 = 22725.260826 | ||
| 45 : | %define C2 21407 ;cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5 = 21406.727617 | ||
| 46 : | %define C3 19266 ;cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5 = 19265.545870 | ||
| 47 : | %define C4 16383 ;cos(i*M_PI/16)*sqrt(2)*(1<<14) - 0.5 = 16384.000000 | ||
| 48 : | %define C5 12873 ;cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5 = 12872.826198 | ||
| 49 : | %define C6 8867 ;cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5 = 8866.956905 | ||
| 50 : | %define C7 4520 ;cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5 = 4520.335430 | ||
| 51 : | |||
| 52 : | coeffs | ||
| 53 : | dw 1<<(ROW_SHIFT-1), 0, 1<<(ROW_SHIFT-1), 0, ; 0 | ||
| 54 : | dw 1<<(ROW_SHIFT-1), 1, 1<<(ROW_SHIFT-1), 0, ; 8 | ||
| 55 : | |||
| 56 : | dw C4, C4, C4, C4 ; 16 | ||
| 57 : | dw C4, -C4, C4, -C4 ; 24 | ||
| 58 : | |||
| 59 : | dw C2, C6, C2, C6 ; 32 | ||
| 60 : | dw C6, -C2, C6, -C2 ; 40 | ||
| 61 : | |||
| 62 : | dw C1, C3, C1, C3 ; 48 | ||
| 63 : | dw C5, C7, C5, C7 ; 56 | ||
| 64 : | |||
| 65 : | dw C3, -C7, C3, -C7 ; 64 | ||
| 66 : | dw -C1, -C5, -C1, -C5 ; 72 | ||
| 67 : | |||
| 68 : | dw C5, -C1, C5, -C1 ; 80 | ||
| 69 : | dw C7, C3, C7, C3 ; 88 | ||
| 70 : | |||
| 71 : | dw C7, -C5, C7, -C5 ; 96 | ||
| 72 : | dw C3, -C1, C3, -C1 ; 104 | ||
| 73 : | |||
| 74 : | |||
| 75 : | ;=========================================================================== | ||
| 76 : | ; text | ||
| 77 : | ;=========================================================================== | ||
| 78 : | section .text | ||
| 79 : | |||
| 80 : | %macro DC_COND_IDCT 8 | ||
| 81 : | %define src0 %1 | ||
| 82 : | %define src4 %2 | ||
| 83 : | %define src1 %3 | ||
| 84 : | %define src5 %4 | ||
| 85 : | %define dst %5 | ||
| 86 : | %define rounder_op %6 | ||
| 87 : | %define rounder_arg %7 | ||
| 88 : | %define shift %8 | ||
| 89 : | movq mm0,[src0] ; R4 R0 r4 r0 | ||
| 90 : | movq mm1,[src4] ; R6 R2 r6 r2 | ||
| 91 : | movq mm2,[src1] ; R3 R1 r3 r1 | ||
| 92 : | movq mm3,[src5] ; R7 R5 r7 r5 | ||
| 93 : | movq mm4,[wm1010] | ||
| 94 : | pand mm4,mm0 | ||
| 95 : | por mm4,mm1 | ||
| 96 : | por mm4,mm2 | ||
| 97 : | por mm4,mm3 | ||
| 98 : | packssdw mm4,mm4 | ||
| 99 : | movd eax,mm4 | ||
| 100 : | or eax,eax | ||
| 101 : | jz near .skip1 | ||
| 102 : | movq mm4,[coeffs+16] ; C4 C4 C4 C4 | ||
| 103 : | pmaddwd mm4,mm0 ; C4R4+C4R0 C4r4+C4r0 | ||
| 104 : | movq mm5,[coeffs+24] ; -C4 C4 -C4 C4 | ||
| 105 : | pmaddwd mm0,mm5 ; -C4R4+C4R0 -C4r4+C4r0 | ||
| 106 : | movq mm5,[coeffs+32] ; C6 C2 C6 C2 | ||
| 107 : | pmaddwd mm5,mm1 ; C6R6+C2R2 C6r6+C2r2 | ||
| 108 : | movq mm6,[coeffs+40] ; -C2 C6 -C2 C6 | ||
| 109 : | pmaddwd mm1,mm6 ; -C2R6+C6R2 -C2r6+C6r2 | ||
| 110 : | movq mm7,[coeffs+48] ; C3 C1 C3 C1 | ||
| 111 : | pmaddwd mm7,mm2 ; C3R3+C1R1 C3r3+C1r1 | ||
| 112 : | rounder_op mm4, rounder_arg | ||
| 113 : | movq mm6,mm4 ; C4R4+C4R0 C4r4+C4r0 | ||
| 114 : | paddd mm4,mm5 ; A0 a0 | ||
| 115 : | psubd mm6,mm5 ; A3 a3 | ||
| 116 : | movq mm5,[coeffs+56] ; C7 C5 C7 C5 | ||
| 117 : | pmaddwd mm5,mm3 ; C7R7+C5R5 C7r7+C5r5 | ||
| 118 : | rounder_op mm0, rounder_arg | ||
| 119 : | paddd mm1,mm0 ; A1 a1 | ||
| 120 : | paddd mm0,mm0 | ||
| 121 : | psubd mm0,mm1 ; A2 a2 | ||
| 122 : | pmaddwd mm2,[coeffs+64] ; -C7R3+C3R1 -C7r3+C3r1 | ||
| 123 : | paddd mm7,mm5 ; B0 b0 | ||
| 124 : | movq mm5,[coeffs+72] ; -C5 -C1 -C5 -C1 | ||
| 125 : | pmaddwd mm5,mm3 ; -C5R7-C1R5 -C5r7-C1r5 | ||
| 126 : | paddd mm7,mm4 ; A0+B0 a0+b0 | ||
| 127 : | paddd mm4,mm4 ; 2A0 2a0 | ||
| 128 : | psubd mm4,mm7 ; A0-B0 a0-b0 | ||
| 129 : | paddd mm5,mm2 ; B1 b1 | ||
| 130 : | psrad mm7,shift | ||
| 131 : | psrad mm4,shift | ||
| 132 : | movq mm2,mm1 ; A1 a1 | ||
| 133 : | paddd mm1,mm5 ; A1+B1 a1+b1 | ||
| 134 : | psubd mm2,mm5 ; A1-B1 a1-b1 | ||
| 135 : | psrad mm1,shift | ||
| 136 : | psrad mm2,shift | ||
| 137 : | packssdw mm7,mm1 ; A1+B1 a1+b1 A0+B0 a0+b0 | ||
| 138 : | packssdw mm2,mm4 ; A0-B0 a0-b0 A1-B1 a1-b1 | ||
| 139 : | movq [dst],mm7 | ||
| 140 : | movq mm1,[src1] ; R3 R1 r3 r1 | ||
| 141 : | movq mm4,[coeffs+80] ;-C1 C5 -C1 C5 | ||
| 142 : | movq [dst + 24],mm2 | ||
| 143 : | pmaddwd mm4,mm1 ; -C1R3+C5R1 -C1r3+C5r1 | ||
| 144 : | movq mm7,[coeffs+88] ; C3 C7 C3 C7 | ||
| 145 : | pmaddwd mm1,[coeffs+96] ; -C5R3+C7R1 -C5r3+C7r1 | ||
| 146 : | pmaddwd mm7,mm3 ; C3R7+C7R5 C3r7+C7r5 | ||
| 147 : | movq mm2,mm0 ; A2 a2 | ||
| 148 : | pmaddwd mm3,[coeffs+104] ; -C1R7+C3R5 -C1r7+C3r5 | ||
| 149 : | paddd mm4,mm7 ; B2 b2 | ||
| 150 : | paddd mm2,mm4 ; A2+B2 a2+b2 | ||
| 151 : | psubd mm0,mm4 ; a2-B2 a2-b2 | ||
| 152 : | psrad mm2,shift | ||
| 153 : | psrad mm0,shift | ||
| 154 : | movq mm4,mm6 ; A3 a3 | ||
| 155 : | paddd mm3,mm1 ; B3 b3 | ||
| 156 : | paddd mm6,mm3 ; A3+B3 a3+b3 | ||
| 157 : | psubd mm4,mm3 ; a3-B3 a3-b3 | ||
| 158 : | psrad mm6,shift | ||
| 159 : | packssdw mm2,mm6 ; A3+B3 a3+b3 A2+B2 a2+b2 | ||
| 160 : | movq [ dst + 8],mm2 | ||
| 161 : | psrad mm4,shift | ||
| 162 : | packssdw mm4,mm0 ; A2-B2 a2-b2 A3-B3 a3-b3 | ||
| 163 : | movq [ dst + 16],mm4 | ||
| 164 : | jmp short .skip2 | ||
| 165 : | .skip1 | ||
| 166 : | pslld mm0,16 | ||
| 167 : | paddd mm0,[d40000] | ||
| 168 : | psrad mm0,13 | ||
| 169 : | packssdw mm0,mm0 | ||
| 170 : | movq [ dst ],mm0 | ||
| 171 : | movq [ dst + 8],mm0 | ||
| 172 : | movq [ dst + 16],mm0 | ||
| 173 : | movq [ dst + 24],mm0 | ||
| 174 : | .skip2 | ||
| 175 : | %undef src0 | ||
| 176 : | %undef src4 | ||
| 177 : | %undef src1 | ||
| 178 : | %undef src5 | ||
| 179 : | %undef dst | ||
| 180 : | %undef rounder_op | ||
| 181 : | %undef rounder_arg | ||
| 182 : | %undef shift | ||
| 183 : | %endmacro | ||
| 184 : | |||
| 185 : | |||
| 186 : | |||
| 187 : | %macro Z_COND_IDCT 9 | ||
| 188 : | %define src0 %1 | ||
| 189 : | %define src4 %2 | ||
| 190 : | %define src1 %3 | ||
| 191 : | %define src5 %4 | ||
| 192 : | %define dst %5 | ||
| 193 : | %define rounder_op %6 | ||
| 194 : | %define rounder_arg %7 | ||
| 195 : | %define shift %8 | ||
| 196 : | %define bt %9 | ||
| 197 : | movq mm0,[src0] ; R4 R0 r4 r0 | ||
| 198 : | movq mm1,[src4] ; R6 R2 r6 r2 | ||
| 199 : | movq mm2,[src1] ; R3 R1 r3 r1 | ||
| 200 : | movq mm3,[src5] ; R7 R5 r7 r5 | ||
| 201 : | movq mm4,mm0 | ||
| 202 : | por mm4,mm1 | ||
| 203 : | por mm4,mm2 | ||
| 204 : | por mm4,mm3 | ||
| 205 : | packssdw mm4,mm4 | ||
| 206 : | movd eax,mm4 | ||
| 207 : | or eax,eax | ||
| 208 : | jz near bt | ||
| 209 : | movq mm4,[coeffs+16] ; C4 C4 C4 C4 | ||
| 210 : | pmaddwd mm4,mm0 ; C4R4+C4R0 C4r4+C4r0 | ||
| 211 : | movq mm5,[coeffs+24] ; -C4 C4 -C4 C4 | ||
| 212 : | pmaddwd mm0,mm5 ; -C4R4+C4R0 -C4r4+C4r0 | ||
| 213 : | movq mm5,[coeffs+32] ; C6 C2 C6 C2 | ||
| 214 : | pmaddwd mm5,mm1 ; C6R6+C2R2 C6r6+C2r2 | ||
| 215 : | movq mm6,[coeffs+40] ; -C2 C6 -C2 C6 | ||
| 216 : | pmaddwd mm1,mm6 ; -C2R6+C6R2 -C2r6+C6r2 | ||
| 217 : | movq mm7,[coeffs+48] ; C3 C1 C3 C1 | ||
| 218 : | pmaddwd mm7,mm2 ; C3R3+C1R1 C3r3+C1r1 | ||
| 219 : | rounder_op mm4, rounder_arg | ||
| 220 : | movq mm6,mm4 ; C4R4+C4R0 C4r4+C4r0 | ||
| 221 : | paddd mm4,mm5 ; A0 a0 | ||
| 222 : | psubd mm6,mm5 ; A3 a3 | ||
| 223 : | movq mm5,[coeffs+56] ; C7 C5 C7 C5 | ||
| 224 : | pmaddwd mm5,mm3 ; C7R7+C5R5 C7r7+C5r5 | ||
| 225 : | rounder_op mm0, rounder_arg | ||
| 226 : | paddd mm1,mm0 ; A1 a1 | ||
| 227 : | paddd mm0,mm0 | ||
| 228 : | psubd mm0,mm1 ; A2 a2 | ||
| 229 : | pmaddwd mm2,[coeffs+64] ; -C7R3+C3R1 -C7r3+C3r1 | ||
| 230 : | paddd mm7,mm5 ; B0 b0 | ||
| 231 : | movq mm5,[coeffs+72] ; -C5 -C1 -C5 -C1 | ||
| 232 : | pmaddwd mm5,mm3 ; -C5R7-C1R5 -C5r7-C1r5 | ||
| 233 : | paddd mm7,mm4 ; A0+B0 a0+b0 | ||
| 234 : | paddd mm4,mm4 ; 2A0 2a0 | ||
| 235 : | psubd mm4,mm7 ; A0-B0 a0-b0 | ||
| 236 : | paddd mm5,mm2 ; B1 b1 | ||
| 237 : | psrad mm7,shift | ||
| 238 : | psrad mm4,shift | ||
| 239 : | movq mm2,mm1 ; A1 a1 | ||
| 240 : | paddd mm1,mm5 ; A1+B1 a1+b1 | ||
| 241 : | psubd mm2,mm5 ; A1-B1 a1-b1 | ||
| 242 : | psrad mm1,shift | ||
| 243 : | psrad mm2,shift | ||
| 244 : | packssdw mm7,mm1 ; A1+B1 a1+b1 A0+B0 a0+b0 | ||
| 245 : | packssdw mm2,mm4 ; A0-B0 a0-b0 A1-B1 a1-b1 | ||
| 246 : | movq [ dst ],mm7 | ||
| 247 : | movq mm1,[src1] ; R3 R1 r3 r1 | ||
| 248 : | movq mm4,[coeffs+80] ; -C1 C5 -C1 C5 | ||
| 249 : | movq [ dst + 24 ],mm2 | ||
| 250 : | pmaddwd mm4,mm1 ; -C1R3+C5R1 -C1r3+C5r1 | ||
| 251 : | movq mm7,[coeffs+88] ; C3 C7 C3 C7 | ||
| 252 : | pmaddwd mm1,[coeffs+96] ; -C5R3+C7R1 -C5r3+C7r1 | ||
| 253 : | pmaddwd mm7,mm3 ; C3R7+C7R5 C3r7+C7r5 | ||
| 254 : | movq mm2,mm0 ; A2 a2 | ||
| 255 : | pmaddwd mm3,[coeffs+104] ; -C1R7+C3R5 -C1r7+C3r5 | ||
| 256 : | paddd mm4,mm7 ; B2 b2 | ||
| 257 : | paddd mm2,mm4 ; A2+B2 a2+b2 | ||
| 258 : | psubd mm0,mm4 ; a2-B2 a2-b2 | ||
| 259 : | psrad mm2,shift | ||
| 260 : | psrad mm0,shift | ||
| 261 : | movq mm4,mm6 ; A3 a3 | ||
| 262 : | paddd mm3,mm1 ; B3 b3 | ||
| 263 : | paddd mm6,mm3 ; A3+B3 a3+b3 | ||
| 264 : | psubd mm4,mm3 ; a3-B3 a3-b3 | ||
| 265 : | psrad mm6,shift | ||
| 266 : | packssdw mm2,mm6 ; A3+B3 a3+b3 A2+B2 a2+b2 | ||
| 267 : | movq [ dst + 8],mm2 | ||
| 268 : | psrad mm4,shift | ||
| 269 : | packssdw mm4,mm0 ; A2-B2 a2-b2 A3-B3 a3-b3 | ||
| 270 : | movq [dst + 16],mm4 | ||
| 271 : | %undef src0 | ||
| 272 : | %undef src4 | ||
| 273 : | %undef src1 | ||
| 274 : | %undef src5 | ||
| 275 : | %undef dst | ||
| 276 : | %undef rounder_op | ||
| 277 : | %undef rounder_arg | ||
| 278 : | %undef shift | ||
| 279 : | %undef bt | ||
| 280 : | %endmacro | ||
| 281 : | |||
| 282 : | |||
| 283 : | |||
| 284 : | %macro IDCT0 8 | ||
| 285 : | %define src0 %1 | ||
| 286 : | %define src4 %2 | ||
| 287 : | %define src1 %3 | ||
| 288 : | %define src5 %4 | ||
| 289 : | %define dst %5 | ||
| 290 : | %define rounder_op %6 | ||
| 291 : | %define rounder_arg %7 | ||
| 292 : | %define shift %8 | ||
| 293 : | movq mm0,[src0] ; R4 R0 r4 r0 | ||
| 294 : | movq mm1,[src4] ; R6 R2 r6 r2 | ||
| 295 : | movq mm2,[src1] ; R3 R1 r3 r1 | ||
| 296 : | movq mm3,[src5] ; R7 R5 r7 r5 | ||
| 297 : | movq mm4,[coeffs+16] ; C4 C4 C4 C4 | ||
| 298 : | pmaddwd mm4,mm0 ; C4R4+C4R0 C4r4+C4r0 | ||
| 299 : | movq mm5,[coeffs+24] ; -C4 C4 -C4 C4 | ||
| 300 : | pmaddwd mm0,mm5 ; -C4R4+C4R0 -C4r4+C4r0 | ||
| 301 : | movq mm5,[coeffs+32] ; C6 C2 C6 C2 | ||
| 302 : | pmaddwd mm5,mm1 ; C6R6+C2R2 C6r6+C2r2 | ||
| 303 : | movq mm6,[coeffs+40] ; -C2 C6 -C2 C6 | ||
| 304 : | pmaddwd mm1,mm6 ; -C2R6+C6R2 -C2r6+C6r2 | ||
| 305 : | ; rounder_op mm4, rounder_arg | ||
| 306 : | movq mm6,mm4 ; C4R4+C4R0 C4r4+C4r0 | ||
| 307 : | movq mm7,[coeffs+48] ; C3 C1 C3 C1 | ||
| 308 : | ; rounder_op mm0, rounder_arg | ||
| 309 : | pmaddwd mm7,mm2 ; C3R3+C1R1 C3r3+C1r1 | ||
| 310 : | paddd mm4,mm5 ; A0 a0 | ||
| 311 : | psubd mm6,mm5 ; A3 a3 | ||
| 312 : | movq mm5,mm0 ; -C4R4+C4R0 -C4r4+C4r0 | ||
| 313 : | paddd mm0,mm1 ; A1 a1 | ||
| 314 : | psubd mm5,mm1 ; A2 a2 | ||
| 315 : | movq mm1,[coeffs+56] ; C7 C5 C7 C5 | ||
| 316 : | pmaddwd mm1,mm3 ; C7R7+C5R5 C7r7+C5r5 | ||
| 317 : | pmaddwd mm2,[coeffs+64] ; -C7R3+C3R1 -C7r3+C3r1 | ||
| 318 : | paddd mm7,mm1 ; B0 b0 | ||
| 319 : | movq mm1,[coeffs+72] ; -C5 -C1 -C5 -C1 | ||
| 320 : | pmaddwd mm1,mm3 ; -C5R7-C1R5 -C5r7-C1r5 | ||
| 321 : | paddd mm7,mm4 ; A0+B0 a0+b0 | ||
| 322 : | paddd mm4,mm4 ; 2A0 2a0 | ||
| 323 : | psubd mm4,mm7 ; A0-B0 a0-b0 | ||
| 324 : | paddd mm1,mm2 ; B1 b1 | ||
| 325 : | psrad mm7,shift | ||
| 326 : | psrad mm4,shift | ||
| 327 : | movq mm2,mm0 ; A1 a1 | ||
| 328 : | paddd mm0,mm1 ; A1+B1 a1+b1 | ||
| 329 : | psubd mm2,mm1 ; A1-B1 a1-b1 | ||
| 330 : | psrad mm0,shift | ||
| 331 : | psrad mm2,shift | ||
| 332 : | packssdw mm7,mm7 ; A0+B0 a0+b0 | ||
| 333 : | movd [ dst ],mm7 | ||
| 334 : | packssdw mm0,mm0 ; A1+B1 a1+b1 | ||
| 335 : | movd [ dst + 16],mm0 | ||
| 336 : | packssdw mm2,mm2 ; A1-B1 a1-b1 | ||
| 337 : | movd [ dst + 96 ],mm2 | ||
| 338 : | packssdw mm4,mm4 ; A0-B0 a0-b0 | ||
| 339 : | movd [ dst + 112],mm4 | ||
| 340 : | movq mm0,[src1] ; R3 R1 r3 r1 | ||
| 341 : | movq mm4,[coeffs+80] ; -C1 C5 -C1 C5 | ||
| 342 : | pmaddwd mm4,mm0 ; -C1R3+C5R1 -C1r3+C5r1 | ||
| 343 : | movq mm7,[coeffs+88] ; C3 C7 C3 C7 | ||
| 344 : | pmaddwd mm0,[coeffs+96] ; -C5R3+C7R1 -C5r3+C7r1 | ||
| 345 : | pmaddwd mm7,mm3 ; C3R7+C7R5 C3r7+C7r5 | ||
| 346 : | movq mm2,mm5 ; A2 a2 | ||
| 347 : | pmaddwd mm3,[coeffs+104] ; -C1R7+C3R5 -C1r7+C3r5 | ||
| 348 : | paddd mm4,mm7 ; B2 b2 | ||
| 349 : | paddd mm2,mm4 ; A2+B2 a2+b2 | ||
| 350 : | psubd mm5,mm4 ; a2-B2 a2-b2 | ||
| 351 : | psrad mm2,shift | ||
| 352 : | psrad mm5,shift | ||
| 353 : | movq mm4,mm6 ; A3 a3 | ||
| 354 : | paddd mm3,mm0 ; B3 b3 | ||
| 355 : | paddd mm6,mm3 ; A3+B3 a3+b3 | ||
| 356 : | psubd mm4,mm3 ; a3-B3 a3-b3 | ||
| 357 : | psrad mm6,shift | ||
| 358 : | psrad mm4,shift | ||
| 359 : | packssdw mm2,mm2 ; A2+B2 a2+b2 | ||
| 360 : | packssdw mm6,mm6 ; A3+B3 a3+b3 | ||
| 361 : | movd [ dst + 32 ],mm2 | ||
| 362 : | packssdw mm4,mm4 ; A3-B3 a3-b3 | ||
| 363 : | packssdw mm5,mm5 ; A2-B2 a2-b2 | ||
| 364 : | movd [ dst + 48 ],mm6 | ||
| 365 : | movd [ dst + 64 ],mm4 | ||
| 366 : | movd [ dst + 80 ],mm5 | ||
| 367 : | %undef src0 | ||
| 368 : | %undef src4 | ||
| 369 : | %undef src1 | ||
| 370 : | %undef src5 | ||
| 371 : | %undef dst | ||
| 372 : | %undef rounder_op | ||
| 373 : | %undef rounder_arg | ||
| 374 : | %undef shift | ||
| 375 : | %endmacro | ||
| 376 : | |||
| 377 : | |||
| 378 : | |||
| 379 : | %macro IDCT4 8 | ||
| 380 : | %define src0 %1 | ||
| 381 : | %define src4 %2 | ||
| 382 : | %define src1 %3 | ||
| 383 : | %define src5 %4 | ||
| 384 : | %define dst %5 | ||
| 385 : | %define rounder_op %6 | ||
| 386 : | %define rounder_arg %7 | ||
| 387 : | %define shift %8 | ||
| 388 : | movq mm0,[src0] ; R4 R0 r4 r0 | ||
| 389 : | movq mm1,[src4] ; R6 R2 r6 r2 | ||
| 390 : | movq mm3,[src5] ; R7 R5 r7 r5 | ||
| 391 : | movq mm4,[coeffs+16] ; C4 C4 C4 C4 | ||
| 392 : | pmaddwd mm4,mm0 ; C4R4+C4R0 C4r4+C4r0 | ||
| 393 : | movq mm5,[coeffs+24] ; -C4 C4 -C4 C4 | ||
| 394 : | pmaddwd mm0,mm5 ; -C4R4+C4R0 -C4r4+C4r0 | ||
| 395 : | movq mm5,[coeffs+32] ; C6 C2 C6 C2 | ||
| 396 : | pmaddwd mm5,mm1 ; C6R6+C2R2 C6r6+C2r2 | ||
| 397 : | movq mm6,[coeffs+40] ; -C2 C6 -C2 C6 | ||
| 398 : | pmaddwd mm1,mm6 ; -C2R6+C6R2 -C2r6+C6r2 | ||
| 399 : | ; rounder_op mm4, rounder_arg | ||
| 400 : | movq mm6,mm4 ; C4R4+C4R0 C4r4+C4r0 | ||
| 401 : | ; rounder_op mm0, rounder_arg | ||
| 402 : | paddd mm4,mm5 ; A0 a0 | ||
| 403 : | psubd mm6,mm5 ; A3 a3 | ||
| 404 : | movq mm5,mm0 ; -C4R4+C4R0 -C4r4+C4r0 | ||
| 405 : | paddd mm0,mm1 ; A1 a1 | ||
| 406 : | psubd mm5,mm1 ; A2 a2 | ||
| 407 : | movq mm1,[coeffs+56] ; C7 C5 C7 C5 | ||
| 408 : | pmaddwd mm1,mm3 ; C7R7+C5R5 C7r7+C5r5 | ||
| 409 : | movq mm7,[coeffs+72] ; -C5 -C1 -C5 -C1 | ||
| 410 : | pmaddwd mm7,mm3 ; -C5R7-C1R5 -C5r7-C1r5 | ||
| 411 : | paddd mm1,mm4 ; A0+B0 a0+b0 | ||
| 412 : | paddd mm4,mm4 ; 2A0 2a0 | ||
| 413 : | psubd mm4,mm1 ; A0-B0 a0-b0 | ||
| 414 : | psrad mm1,shift | ||
| 415 : | psrad mm4,shift | ||
| 416 : | movq mm2,mm0 ; A1 a1 | ||
| 417 : | paddd mm0,mm7 ; A1+B1 a1+b1 | ||
| 418 : | psubd mm2,mm7 ; A1-B1 a1-b1 | ||
| 419 : | psrad mm0,shift | ||
| 420 : | psrad mm2,shift | ||
| 421 : | packssdw mm1,mm1 ; A0+B0 a0+b0 | ||
| 422 : | movd [ dst ],mm1 | ||
| 423 : | packssdw mm0,mm0 ; A1+B1 a1+b1 | ||
| 424 : | movd [ dst + 16 ],mm0 | ||
| 425 : | packssdw mm2,mm2 ; A1-B1 a1-b1 | ||
| 426 : | movd [ dst + 96 ],mm2 | ||
| 427 : | packssdw mm4,mm4 ; A0-B0 a0-b0 | ||
| 428 : | movd [ dst + 112 ],mm4 | ||
| 429 : | movq mm1,[coeffs+88] ; C3 C7 C3 C7 | ||
| 430 : | pmaddwd mm1,mm3 ; C3R7+C7R5 C3r7+C7r5 | ||
| 431 : | movq mm2,mm5 ; A2 a2 | ||
| 432 : | pmaddwd mm3,[coeffs+104] ; -C1R7+C3R5 -C1r7+C3r5 | ||
| 433 : | paddd mm2,mm1 ; A2+B2 a2+b2 | ||
| 434 : | psubd mm5,mm1 ; a2-B2 a2-b2 | ||
| 435 : | psrad mm2,shift | ||
| 436 : | psrad mm5,shift | ||
| 437 : | movq mm1,mm6 ; A3 a3 | ||
| 438 : | paddd mm6,mm3 ; A3+B3 a3+b3 | ||
| 439 : | psubd mm1,mm3 ; a3-B3 a3-b3 | ||
| 440 : | psrad mm6,shift | ||
| 441 : | psrad mm1,shift | ||
| 442 : | packssdw mm2,mm2 ; A2+B2 a2+b2 | ||
| 443 : | packssdw mm6,mm6 ; A3+B3 a3+b3 | ||
| 444 : | movd [dst + 32],mm2 | ||
| 445 : | packssdw mm1,mm1 ; A3-B3 a3-b3 | ||
| 446 : | packssdw mm5,mm5 ; A2-B2 a2-b2 | ||
| 447 : | movd [dst + 48],mm6 | ||
| 448 : | movd [dst + 64],mm1 | ||
| 449 : | movd [dst + 80],mm5 | ||
| 450 : | %undef src0 | ||
| 451 : | %undef src4 | ||
| 452 : | %undef src1 | ||
| 453 : | %undef src5 | ||
| 454 : | %undef dst | ||
| 455 : | %undef rounder_op | ||
| 456 : | %undef rounder_arg | ||
| 457 : | %undef shift | ||
| 458 : | %endmacro | ||
| 459 : | |||
| 460 : | |||
| 461 : | |||
| 462 : | %macro IDCT6 8 | ||
| 463 : | %define src0 %1 | ||
| 464 : | %define src4 %2 | ||
| 465 : | %define src1 %3 | ||
| 466 : | %define src5 %4 | ||
| 467 : | %define dst %5 | ||
| 468 : | %define rounder_op %6 | ||
| 469 : | %define rounder_arg %7 | ||
| 470 : | %define shift %8 | ||
| 471 : | movq mm0,[src0] ; R4 R0 r4 r0 | ||
| 472 : | movq mm3,[src5] ; R7 R5 r7 r5 | ||
| 473 : | movq mm4,[coeffs+16] ; C4 C4 C4 C4 | ||
| 474 : | pmaddwd mm4,mm0 ; C4R4+C4R0 C4r4+C4r0 | ||
| 475 : | movq mm5,[coeffs+24] ; -C4 C4 -C4 C4 | ||
| 476 : | pmaddwd mm0,mm5 ; -C4R4+C4R0 -C4r4+C4r0 | ||
| 477 : | ; rounder_op mm4, rounder_arg | ||
| 478 : | movq mm6,mm4 ; C4R4+C4R0 C4r4+C4r0 | ||
| 479 : | ; rounder_op mm0, rounder_arg | ||
| 480 : | movq mm5,mm0 ; -C4R4+C4R0 -C4r4+C4r0 | ||
| 481 : | movq mm1,[coeffs+56] ; C7 C5 C7 C5 | ||
| 482 : | pmaddwd mm1,mm3 ; C7R7+C5R5 C7r7+C5r5 | ||
| 483 : | movq mm7,[coeffs+72] ; -C5 -C1 -C5 -C1 | ||
| 484 : | pmaddwd mm7,mm3 ; -C5R7-C1R5 -C5r7-C1r5 | ||
| 485 : | paddd mm1,mm4 ; A0+B0 a0+b0 | ||
| 486 : | paddd mm4,mm4 ; 2A0 2a0 | ||
| 487 : | psubd mm4,mm1 ; A0-B0 a0-b0 | ||
| 488 : | psrad mm1,shift | ||
| 489 : | psrad mm4,shift | ||
| 490 : | movq mm2,mm0 ; A1 a1 | ||
| 491 : | paddd mm0,mm7 ; A1+B1 a1+b1 | ||
| 492 : | psubd mm2,mm7 ; A1-B1 a1-b1 | ||
| 493 : | psrad mm0,shift | ||
| 494 : | psrad mm2,shift | ||
| 495 : | packssdw mm1,mm1 ; A0+B0 a0+b0 | ||
| 496 : | movd [ dst ],mm1 | ||
| 497 : | packssdw mm0,mm0 ; A1+B1 a1+b1 | ||
| 498 : | movd [ dst + 16 ],mm0 | ||
| 499 : | packssdw mm2,mm2 ; A1-B1 a1-b1 | ||
| 500 : | movd [ dst + 96 ],mm2 | ||
| 501 : | packssdw mm4,mm4 ; A0-B0 a0-b0 | ||
| 502 : | movd [ dst + 112 ],mm4 | ||
| 503 : | movq mm1,[coeffs+88] ; C3 C7 C3 C7 | ||
| 504 : | pmaddwd mm1,mm3 ; C3R7+C7R5 C3r7+C7r5 | ||
| 505 : | movq mm2,mm5 ; A2 a2 | ||
| 506 : | pmaddwd mm3,[coeffs+104] ; -C1R7+C3R5 -C1r7+C3r5 | ||
| 507 : | paddd mm2,mm1 ; A2+B2 a2+b2 | ||
| 508 : | psubd mm5,mm1 ; a2-B2 a2-b2 | ||
| 509 : | psrad mm2,shift | ||
| 510 : | psrad mm5,shift | ||
| 511 : | movq mm1,mm6 ; A3 a3 | ||
| 512 : | paddd mm6,mm3 ; A3+B3 a3+b3 | ||
| 513 : | psubd mm1,mm3 ; a3-B3 a3-b3 | ||
| 514 : | psrad mm6,shift | ||
| 515 : | psrad mm1,shift | ||
| 516 : | packssdw mm2,mm2 ; A2+B2 a2+b2 | ||
| 517 : | packssdw mm6,mm6 ; A3+B3 a3+b3 | ||
| 518 : | movd [dst + 32],mm2 | ||
| 519 : | packssdw mm1,mm1 ; A3-B3 a3-b3 | ||
| 520 : | packssdw mm5,mm5 ; A2-B2 a2-b2 | ||
| 521 : | movd [dst + 48],mm6 | ||
| 522 : | movd [dst + 64],mm1 | ||
| 523 : | movd [dst + 80],mm5 | ||
| 524 : | %undef src0 | ||
| 525 : | %undef src4 | ||
| 526 : | %undef src1 | ||
| 527 : | %undef src5 | ||
| 528 : | %undef dst | ||
| 529 : | %undef rounder_op | ||
| 530 : | %undef rounder_arg | ||
| 531 : | %undef shift | ||
| 532 : | %endmacro | ||
| 533 : | |||
| 534 : | |||
| 535 : | |||
| 536 : | |||
| 537 : | %macro IDCT2 8 | ||
| 538 : | %define src0 %1 | ||
| 539 : | %define src4 %2 | ||
| 540 : | %define src1 %3 | ||
| 541 : | %define src5 %4 | ||
| 542 : | %define dst %5 | ||
| 543 : | %define rounder_op %6 | ||
| 544 : | %define rounder_arg %7 | ||
| 545 : | %define shift %8 | ||
| 546 : | movq mm0,[src0] ; R4 R0 r4 r0 | ||
| 547 : | movq mm2,[src1] ; R3 R1 r3 r1 | ||
| 548 : | movq mm3,[src5] ; R7 R5 r7 r5 | ||
| 549 : | movq mm4,[coeffs+16] ; C4 C4 C4 C4 | ||
| 550 : | pmaddwd mm4,mm0 ; C4R4+C4R0 C4r4+C4r0 | ||
| 551 : | movq mm5,[coeffs+24] ; -C4 C4 -C4 C4 | ||
| 552 : | pmaddwd mm0,mm5 ; -C4R4+C4R0 -C4r4+C4r0 | ||
| 553 : | ; rounder_op mm4, rounder_arg | ||
| 554 : | movq mm6,mm4 ; C4R4+C4R0 C4r4+C4r0 | ||
| 555 : | movq mm7,[coeffs+48] ; C3 C1 C3 C1 | ||
| 556 : | ; rounder_op mm0, rounder_arg | ||
| 557 : | pmaddwd mm7,mm2 ; C3R3+C1R1 C3r3+C1r1 | ||
| 558 : | movq mm5,mm0 ; -C4R4+C4R0 -C4r4+C4r0 | ||
| 559 : | movq mm1,[coeffs+56] ; C7 C5 C7 C5 | ||
| 560 : | pmaddwd mm1,mm3 ; C7R7+C5R5 C7r7+C5r5 | ||
| 561 : | pmaddwd mm2,[coeffs+64] ; -C7R3+C3R1 -C7r3+C3r1 | ||
| 562 : | paddd mm7,mm1 ; B0 b0 | ||
| 563 : | movq mm1,[coeffs+72] ; -C5 -C1 -C5 -C1 | ||
| 564 : | pmaddwd mm1,mm3 ; -C5R7-C1R5 -C5r7-C1r5 | ||
| 565 : | paddd mm7,mm4 ; A0+B0 a0+b0 | ||
| 566 : | paddd mm4,mm4 ; 2A0 2a0 | ||
| 567 : | psubd mm4,mm7 ; A0-B0 a0-b0 | ||
| 568 : | paddd mm1,mm2 ; B1 b1 | ||
| 569 : | psrad mm7,shift | ||
| 570 : | psrad mm4,shift | ||
| 571 : | movq mm2,mm0 ; A1 a1 | ||
| 572 : | paddd mm0,mm1 ; A1+B1 a1+b1 | ||
| 573 : | psubd mm2,mm1 ; A1-B1 a1-b1 | ||
| 574 : | psrad mm0,shift | ||
| 575 : | psrad mm2,shift | ||
| 576 : | packssdw mm7,mm7 ; A0+B0 a0+b0 | ||
| 577 : | movd [dst],mm7 | ||
| 578 : | packssdw mm0,mm0 ; A1+B1 a1+b1 | ||
| 579 : | movd [dst + 16],mm0 | ||
| 580 : | packssdw mm2,mm2 ; A1-B1 a1-b1 | ||
| 581 : | movd [dst + 96],mm2 | ||
| 582 : | packssdw mm4,mm4 ; A0-B0 a0-b0 | ||
| 583 : | movd [dst + 112],mm4 | ||
| 584 : | movq mm0,[src1] ; R3 R1 r3 r1 | ||
| 585 : | movq mm4,[coeffs+80] ; -C1 C5 -C1 C5 | ||
| 586 : | pmaddwd mm4,mm0 ; -C1R3+C5R1 -C1r3+C5r1 | ||
| 587 : | movq mm7,[coeffs+88] ; C3 C7 C3 C7 | ||
| 588 : | pmaddwd mm0,[coeffs+96] ; -C5R3+C7R1 -C5r3+C7r1 | ||
| 589 : | pmaddwd mm7,mm3 ; C3R7+C7R5 C3r7+C7r5 | ||
| 590 : | movq mm2,mm5 ; A2 a2 | ||
| 591 : | pmaddwd mm3,[coeffs+104] ; -C1R7+C3R5 -C1r7+C3r5 | ||
| 592 : | paddd mm4,mm7 ; B2 b2 | ||
| 593 : | paddd mm2,mm4 ; A2+B2 a2+b2 | ||
| 594 : | psubd mm5,mm4 ; a2-B2 a2-b2 | ||
| 595 : | psrad mm2,shift | ||
| 596 : | psrad mm5,shift | ||
| 597 : | movq mm4,mm6 ; A3 a3 | ||
| 598 : | paddd mm3,mm0 ; B3 b3 | ||
| 599 : | paddd mm6,mm3 ; A3+B3 a3+b3 | ||
| 600 : | psubd mm4,mm3 ; a3-B3 a3-b3 | ||
| 601 : | psrad mm6,shift | ||
| 602 : | psrad mm4,shift | ||
| 603 : | packssdw mm2,mm2 ; A2+B2 a2+b2 | ||
| 604 : | packssdw mm6,mm6 ; A3+B3 a3+b3 | ||
| 605 : | movd [dst + 32],mm2 | ||
| 606 : | packssdw mm4,mm4 ; A3-B3 a3-b3 | ||
| 607 : | packssdw mm5,mm5 ; A2-B2 a2-b2 | ||
| 608 : | movd [dst + 48],mm6 | ||
| 609 : | movd [dst + 64],mm4 | ||
| 610 : | movd [dst + 80],mm5 | ||
| 611 : | %undef src0 | ||
| 612 : | %undef src4 | ||
| 613 : | %undef src1 | ||
| 614 : | %undef src5 | ||
| 615 : | %undef dst | ||
| 616 : | %undef rounder_op | ||
| 617 : | %undef rounder_arg | ||
| 618 : | %undef shift | ||
| 619 : | %endmacro | ||
| 620 : | |||
| 621 : | |||
| 622 : | |||
| 623 : | %macro IDCT3 8 | ||
| 624 : | %define src0 %1 | ||
| 625 : | %define src4 %2 | ||
| 626 : | %define src1 %3 | ||
| 627 : | %define src5 %4 | ||
| 628 : | %define dst %5 | ||
| 629 : | %define rounder_op %6 | ||
| 630 : | %define rounder_arg %7 | ||
| 631 : | %define shift %8 | ||
| 632 : | movq mm0,[src0] ; R4 R0 r4 r0 | ||
| 633 : | movq mm2,[src1] ; R3 R1 r3 r1 | ||
| 634 : | movq mm4,[coeffs+16] ; C4 C4 C4 C4 | ||
| 635 : | pmaddwd mm4,mm0 ; C4R4+C4R0 C4r4+C4r0 | ||
| 636 : | movq mm5,[coeffs+24] ; -C4 C4 -C4 C4 | ||
| 637 : | pmaddwd mm0,mm5 ; -C4R4+C4R0 -C4r4+C4r0 | ||
| 638 : | ; rounder_op mm4, rounder_arg | ||
| 639 : | movq mm6,mm4 ; C4R4+C4R0 C4r4+C4r0 | ||
| 640 : | movq mm7,[coeffs+48] ; C3 C1 C3 C1 | ||
| 641 : | ; rounder_op mm0, rounder_arg | ||
| 642 : | pmaddwd mm7,mm2 ; C3R3+C1R1 C3r3+C1r1 | ||
| 643 : | movq mm5,mm0 ; -C4R4+C4R0 -C4r4+C4r0 | ||
| 644 : | movq mm3,[coeffs+64] | ||
| 645 : | pmaddwd mm3,mm2 ; -C7R3+C3R1 -C7r3+C3r1 | ||
| 646 : | paddd mm7,mm4 ; A0+B0 a0+b0 | ||
| 647 : | paddd mm4,mm4 ; 2A0 2a0 | ||
| 648 : | psubd mm4,mm7 ; A0-B0 a0-b0 | ||
| 649 : | psrad mm7,shift | ||
| 650 : | psrad mm4,shift | ||
| 651 : | movq mm1,mm0 ; A1 a1 | ||
| 652 : | paddd mm0,mm3 ; A1+B1 a1+b1 | ||
| 653 : | psubd mm1,mm3 ; A1-B1 a1-b1 | ||
| 654 : | psrad mm0,shift | ||
| 655 : | psrad mm1,shift | ||
| 656 : | packssdw mm7,mm7 ; A0+B0 a0+b0 | ||
| 657 : | movd [dst],mm7 | ||
| 658 : | packssdw mm0,mm0 ; A1+B1 a1+b1 | ||
| 659 : | movd [dst + 16],mm0 | ||
| 660 : | packssdw mm1,mm1 ; A1-B1 a1-b1 | ||
| 661 : | movd [dst + 96],mm1 | ||
| 662 : | packssdw mm4,mm4 ; A0-B0 a0-b0 | ||
| 663 : | movd [dst + 112],mm4 | ||
| 664 : | movq mm4,[coeffs+80] ; -C1 C5 -C1 C5 | ||
| 665 : | pmaddwd mm4,mm2 ; -C1R3+C5R1 -C1r3+C5r1 | ||
| 666 : | pmaddwd mm2,[coeffs+96] ; -C5R3+C7R1 -C5r3+C7r1 | ||
| 667 : | movq mm1,mm5 ; A2 a2 | ||
| 668 : | paddd mm1,mm4 ; A2+B2 a2+b2 | ||
| 669 : | psubd mm5,mm4 ; a2-B2 a2-b2 | ||
| 670 : | psrad mm1,shift | ||
| 671 : | psrad mm5,shift | ||
| 672 : | movq mm4,mm6 ; A3 a3 | ||
| 673 : | paddd mm6,mm2 ; A3+B3 a3+b3 | ||
| 674 : | psubd mm4,mm2 ; a3-B3 a3-b3 | ||
| 675 : | psrad mm6,shift | ||
| 676 : | psrad mm4,shift | ||
| 677 : | packssdw mm1,mm1 ; A2+B2 a2+b2 | ||
| 678 : | packssdw mm6,mm6 ; A3+B3 a3+b3 | ||
| 679 : | movd [dst + 32],mm1 | ||
| 680 : | packssdw mm4,mm4 ; A3-B3 a3-b3 | ||
| 681 : | packssdw mm5,mm5 ; A2-B2 a2-b2 | ||
| 682 : | movd [dst + 48],mm6 | ||
| 683 : | movd [dst + 64],mm4 | ||
| 684 : | movd [dst + 80],mm5 | ||
| 685 : | %undef src0 | ||
| 686 : | %undef src4 | ||
| 687 : | %undef src1 | ||
| 688 : | %undef src5 | ||
| 689 : | %undef dst | ||
| 690 : | %undef rounder_op | ||
| 691 : | %undef rounder_arg | ||
| 692 : | %undef shift | ||
| 693 : | %endmacro | ||
| 694 : | |||
| 695 : | |||
| 696 : | |||
| 697 : | %macro IDCT5 8 | ||
| 698 : | %define src0 %1 | ||
| 699 : | %define src4 %2 | ||
| 700 : | %define src1 %3 | ||
| 701 : | %define src5 %4 | ||
| 702 : | %define dst %5 | ||
| 703 : | %define rounder_op %6 | ||
| 704 : | %define rounder_arg %7 | ||
| 705 : | %define shift %8 | ||
| 706 : | movq mm0,[src0] ; R4 R0 r4 r0 | ||
| 707 : | movq mm1,[src4] ; R6 R2 r6 r2 | ||
| 708 : | movq mm4,[coeffs+16] ; C4 C4 C4 C4 | ||
| 709 : | pmaddwd mm4,mm0 ; C4R4+C4R0 C4r4+C4r0 | ||
| 710 : | movq mm5,[coeffs+24] ; -C4 C4 -C4 C4 | ||
| 711 : | pmaddwd mm0,mm5 ; -C4R4+C4R0 -C4r4+C4r0 | ||
| 712 : | movq mm5,[coeffs+32] ; C6 C2 C6 C2 | ||
| 713 : | pmaddwd mm5,mm1 ; C6R6+C2R2 C6r6+C2r2 | ||
| 714 : | movq mm6,[coeffs+40] ; -C2 C6 -C2 C6 | ||
| 715 : | pmaddwd mm1,mm6 ; -C2R6+C6R2 -C2r6+C6r2 | ||
| 716 : | ; rounder_op mm4, rounder_arg | ||
| 717 : | movq mm6,mm4 ; C4R4+C4R0 C4r4+C4r0 | ||
| 718 : | paddd mm4,mm5 ; A0 a0 | ||
| 719 : | ; rounder_op mm0, rounder_arg | ||
| 720 : | psubd mm6,mm5 ; A3 a3 | ||
| 721 : | movq mm5,mm0 ; -C4R4+C4R0 -C4r4+C4r0 | ||
| 722 : | paddd mm0,mm1 ; A1 a1 | ||
| 723 : | psubd mm5,mm1 ; A2 a2 | ||
| 724 : | movq mm2,[src0 + 8] ; R4 R0 r4 r0 | ||
| 725 : | movq mm3,[src4 + 8] ; R6 R2 r6 r2 | ||
| 726 : | movq mm1,[coeffs+16] ; C4 C4 C4 C4 | ||
| 727 : | pmaddwd mm1,mm2 ; C4R4+C4R0 C4r4+C4r0 | ||
| 728 : | movq mm7,[coeffs+24] ; -C4 C4 -C4 C4 | ||
| 729 : | pmaddwd mm2,mm7 ; -C4R4+C4R0 -C4r4+C4r0 | ||
| 730 : | movq mm7,[coeffs+32] ; C6 C2 C6 C2 | ||
| 731 : | pmaddwd mm7,mm3 ; C6R6+C2R2 C6r6+C2r2 | ||
| 732 : | pmaddwd mm3,[coeffs+40] ; -C2R6+C6R2 -C2r6+C6r2 | ||
| 733 : | ; rounder_op mm1, rounder_arg | ||
| 734 : | paddd mm7,mm1 ; A0 a0 | ||
| 735 : | paddd mm1,mm1 ; 2C0 2c0 | ||
| 736 : | ; rounder_op mm2, rounder_arg | ||
| 737 : | psubd mm1,mm7 ; A3 a3 | ||
| 738 : | paddd mm3,mm2 ; A1 a1 | ||
| 739 : | paddd mm2,mm2 ; 2C1 2c1 | ||
| 740 : | psubd mm2,mm3 ; A2 a2 | ||
| 741 : | psrad mm4,shift | ||
| 742 : | psrad mm7,shift | ||
| 743 : | psrad mm3,shift | ||
| 744 : | packssdw mm4,mm7 ; A0 a0 | ||
| 745 : | movq [dst],mm4 | ||
| 746 : | psrad mm0,shift | ||
| 747 : | packssdw mm0,mm3 ; A1 a1 | ||
| 748 : | movq [dst + 16],mm0 | ||
| 749 : | movq [dst + 96],mm0 | ||
| 750 : | movq [dst + 112],mm4 | ||
| 751 : | psrad mm5,shift | ||
| 752 : | psrad mm6,shift | ||
| 753 : | psrad mm2,shift | ||
| 754 : | packssdw mm5,mm2 ; A2-B2 a2-b2 | ||
| 755 : | movq [dst + 32],mm5 | ||
| 756 : | psrad mm1,shift | ||
| 757 : | packssdw mm6,mm1 ; A3+B3 a3+b3 | ||
| 758 : | movq [dst + 48],mm6 | ||
| 759 : | movq [dst + 64],mm6 | ||
| 760 : | movq [dst + 80],mm5 | ||
| 761 : | %undef src0 | ||
| 762 : | %undef src4 | ||
| 763 : | %undef src1 | ||
| 764 : | %undef src5 | ||
| 765 : | %undef dst | ||
| 766 : | %undef rounder_op | ||
| 767 : | %undef rounder_arg | ||
| 768 : | %undef shift | ||
| 769 : | %endmacro | ||
| 770 : | |||
| 771 : | |||
| 772 : | %macro IDCT1 8 | ||
| 773 : | %define src0 %1 | ||
| 774 : | %define src4 %2 | ||
| 775 : | %define src1 %3 | ||
| 776 : | %define src5 %4 | ||
| 777 : | %define dst %5 | ||
| 778 : | %define rounder_op %6 | ||
| 779 : | %define rounder_arg %7 | ||
| 780 : | %define shift %8 | ||
| 781 : | movq mm0,[src0] ; R4 R0 r4 r0 | ||
| 782 : | movq mm1,[src4] ; R6 R2 r6 r2 | ||
| 783 : | movq mm2,[src1] ; R3 R1 r3 r1 | ||
| 784 : | movq mm4,[coeffs+16] ; C4 C4 C4 C4 | ||
| 785 : | pmaddwd mm4,mm0 ; C4R4+C4R0 C4r4+C4r0 | ||
| 786 : | movq mm5,[coeffs+24] ; -C4 C4 -C4 C4 | ||
| 787 : | pmaddwd mm0,mm5 ; -C4R4+C4R0 -C4r4+C4r0 | ||
| 788 : | movq mm5,[coeffs+32] ; C6 C2 C6 C2 | ||
| 789 : | pmaddwd mm5,mm1 ; C6R6+C2R2 C6r6+C2r2 | ||
| 790 : | movq mm6,[coeffs+40] ; -C2 C6 -C2 C6 | ||
| 791 : | pmaddwd mm1,mm6 ; -C2R6+C6R2 -C2r6+C6r2 | ||
| 792 : | ; rounder_op mm4, rounder_arg | ||
| 793 : | movq mm6,mm4 ; C4R4+C4R0 C4r4+C4r0 | ||
| 794 : | movq mm7,[coeffs+48] ; C3 C1 C3 C1 | ||
| 795 : | ; rounder_op mm0, rounder_arg | ||
| 796 : | pmaddwd mm7,mm2 ; C3R3+C1R1 C3r3+C1r1 | ||
| 797 : | paddd mm4,mm5 ; A0 a0 | ||
| 798 : | psubd mm6,mm5 ; A3 a3 | ||
| 799 : | movq mm5,mm0 ; -C4R4+C4R0 -C4r4+C4r0 | ||
| 800 : | paddd mm0,mm1 ; A1 a1 | ||
| 801 : | psubd mm5,mm1 ; A2 a2 | ||
| 802 : | movq mm1,[coeffs+64] | ||
| 803 : | pmaddwd mm1,mm2 ; -C7R3+C3R1 -C7r3+C3r1 | ||
| 804 : | paddd mm7,mm4 ; A0+B0 a0+b0 | ||
| 805 : | paddd mm4,mm4 ; 2A0 2a0 | ||
| 806 : | psubd mm4,mm7 ; A0-B0 a0-b0 | ||
| 807 : | psrad mm7,shift | ||
| 808 : | psrad mm4,shift | ||
| 809 : | movq mm3,mm0 ; A1 a1 | ||
| 810 : | paddd mm0,mm1 ; A1+B1 a1+b1 | ||
| 811 : | psubd mm3,mm1 ; A1-B1 a1-b1 | ||
| 812 : | psrad mm0,shift | ||
| 813 : | psrad mm3,shift | ||
| 814 : | packssdw mm7,mm7 ; A0+B0 a0+b0 | ||
| 815 : | movd [dst],mm7 | ||
| 816 : | packssdw mm0,mm0 ; A1+B1 a1+b1 | ||
| 817 : | movd [dst + 16],mm0 | ||
| 818 : | packssdw mm3,mm3 ; A1-B1 a1-b1 | ||
| 819 : | movd [dst + 96],mm3 | ||
| 820 : | packssdw mm4,mm4 ; A0-B0 a0-b0 | ||
| 821 : | movd [dst + 112],mm4 | ||
| 822 : | movq mm4,[coeffs+80] ; -C1 C5 -C1 C5 | ||
| 823 : | pmaddwd mm4,mm2 ; -C1R3+C5R1 -C1r3+C5r1 | ||
| 824 : | pmaddwd mm2,[coeffs+96] ; -C5R3+C7R1 -C5r3+C7r1 | ||
| 825 : | movq mm3,mm5 ; A2 a2 | ||
| 826 : | paddd mm3,mm4 ; A2+B2 a2+b2 | ||
| 827 : | psubd mm5,mm4 ; a2-B2 a2-b2 | ||
| 828 : | psrad mm3,shift | ||
| 829 : | psrad mm5,shift | ||
| 830 : | movq mm4,mm6 ; A3 a3 | ||
| 831 : | paddd mm6,mm2 ; A3+B3 a3+b3 | ||
| 832 : | psubd mm4,mm2 ; a3-B3 a3-b3 | ||
| 833 : | psrad mm6,shift | ||
| 834 : | packssdw mm3,mm3 ; A2+B2 a2+b2 | ||
| 835 : | movd [dst + 32],mm3 | ||
| 836 : | psrad mm4,shift | ||
| 837 : | packssdw mm6,mm6 ; A3+B3 a3+b3 | ||
| 838 : | movd [dst + 48],mm6 | ||
| 839 : | packssdw mm4,mm4 ; A3-B3 a3-b3 | ||
| 840 : | packssdw mm5,mm5 ; A2-B2 a2-b2 | ||
| 841 : | movd [dst + 64],mm4 | ||
| 842 : | movd [dst + 80],mm5 | ||
| 843 : | %undef src0 | ||
| 844 : | %undef src4 | ||
| 845 : | %undef src1 | ||
| 846 : | %undef src5 | ||
| 847 : | %undef dst | ||
| 848 : | %undef rounder_op | ||
| 849 : | %undef rounder_arg | ||
| 850 : | %undef shift | ||
| 851 : | %endmacro | ||
| 852 : | |||
| 853 : | |||
| 854 : | |||
| 855 : | |||
| 856 : | %macro IDCT7 8 | ||
| 857 : | %define src0 %1 | ||
| 858 : | %define src4 %2 | ||
| 859 : | %define src1 %3 | ||
| 860 : | %define src5 %4 | ||
| 861 : | %define dst %5 | ||
| 862 : | %define rounder_op %6 | ||
| 863 : | %define rounder_arg %7 | ||
| 864 : | %define shift %8 | ||
| 865 : | movq mm0,[src0] ; R4 R0 r4 r0 | ||
| 866 : | movq mm4,[coeffs+16] ; C4 C4 C4 C4 | ||
| 867 : | pmaddwd mm4,mm0 ; C4R4+C4R0 C4r4+C4r0 | ||
| 868 : | movq mm5,[coeffs+24] ; -C4 C4 -C4 C4 | ||
| 869 : | pmaddwd mm0,mm5 ; -C4R4+C4R0 -C4r4+C4r0 | ||
| 870 : | ; rounder_op mm4, rounder_arg | ||
| 871 : | ; rounder_op mm0, rounder_arg | ||
| 872 : | psrad mm4,shift | ||
| 873 : | psrad mm0,shift | ||
| 874 : | movq mm2,[src0 + 8] ; R4 R0 r4 r0 | ||
| 875 : | movq mm1,[coeffs+16] ; C4 C4 C4 C4 | ||
| 876 : | pmaddwd mm1,mm2 ; C4R4+C4R0 C4r4+C4r0 | ||
| 877 : | movq mm7,[coeffs+24] ; -C4 C4 -C4 C4 | ||
| 878 : | pmaddwd mm2,mm7 ; -C4R4+C4R0 -C4r4+C4r0 | ||
| 879 : | movq mm7,[coeffs+32] ; C6 C2 C6 C2 | ||
| 880 : | ; rounder_op mm1, rounder_arg | ||
| 881 : | ; rounder_op mm2, rounder_arg | ||
| 882 : | psrad mm1,shift | ||
| 883 : | packssdw mm4,mm1 ; A0 a0 | ||
| 884 : | movq [dst],mm4 | ||
| 885 : | psrad mm2,shift | ||
| 886 : | packssdw mm0,mm2 ; A1 a1 | ||
| 887 : | movq [dst + 16],mm0 | ||
| 888 : | movq [dst + 96],mm0 | ||
| 889 : | movq [dst + 112],mm4 | ||
| 890 : | movq [dst + 32],mm0 | ||
| 891 : | movq [dst + 48],mm4 | ||
| 892 : | movq [dst + 64],mm4 | ||
| 893 : | movq [dst + 80],mm0 | ||
| 894 : | %undef src0 | ||
| 895 : | %undef src4 | ||
| 896 : | %undef src1 | ||
| 897 : | %undef src5 | ||
| 898 : | %undef dst | ||
| 899 : | %undef rounder_op | ||
| 900 : | %undef rounder_arg | ||
| 901 : | %undef shift | ||
| 902 : | %endmacro | ||
| 903 : | |||
| 904 : | |||
| 905 : | |||
| 906 : | %macro cglobal 1 | ||
| 907 : | %ifdef PREFIX | ||
| 908 : | global _%1 | ||
| 909 : | %define %1 _%1 | ||
| 910 : | %else | ||
| 911 : | global %1 | ||
| 912 : | %endif | ||
| 913 : | %endmacro | ||
| 914 : | |||
| 915 : | |||
| 916 : | Isibaar | 1055 | ; void simple_idct_mmx_P(int16_t * const block); |
| 917 : | ; expects input data to be permutated | ||
| 918 : | ; | ||
| 919 : | edgomez | 851 | align 16 |
| 920 : | Isibaar | 1055 | cglobal simple_idct_mmx_P |
| 921 : | simple_idct_mmx_P | ||
| 922 : | edgomez | 851 | sub esp, 128 |
| 923 : | mov edx, [esp+128+4] | ||
| 924 : | |||
| 925 : | ; src0, src4, src1, src5, dst, rndop, rndarg, shift, bt | ||
| 926 : | |||
| 927 : | DC_COND_IDCT edx+0, edx+8, edx+16, edx+24, esp, paddd, [coeffs+8], 11 | ||
| 928 : | Z_COND_IDCT edx+32, edx+40, edx+48, edx+56, esp+32, paddd, [coeffs], 11, .four | ||
| 929 : | Z_COND_IDCT edx+64, edx+72, edx+80, edx+88, esp+64, paddd, [coeffs], 11, .two | ||
| 930 : | Z_COND_IDCT edx+96, edx+104,edx+112,edx+120,esp+96, paddd, [coeffs], 11, .one | ||
| 931 : | IDCT0 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
| 932 : | IDCT0 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
| 933 : | IDCT0 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
| 934 : | IDCT0 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
| 935 : | jmp .ret | ||
| 936 : | |||
| 937 : | align 16 | ||
| 938 : | .four | ||
| 939 : | Z_COND_IDCT edx+64, edx+72, edx+80, edx+88, esp+64, paddd, [coeffs], 11, .six | ||
| 940 : | Z_COND_IDCT edx+96, edx+104,edx+112,edx+120,esp+96, paddd, [coeffs], 11, .five | ||
| 941 : | IDCT4 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
| 942 : | IDCT4 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
| 943 : | IDCT4 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
| 944 : | IDCT4 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
| 945 : | jmp .ret | ||
| 946 : | |||
| 947 : | align 16 | ||
| 948 : | .six | ||
| 949 : | Z_COND_IDCT edx+96, edx+104,edx+112,edx+120,esp+96, paddd, [coeffs], 11, .seven | ||
| 950 : | IDCT6 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
| 951 : | IDCT6 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
| 952 : | IDCT6 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
| 953 : | IDCT6 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
| 954 : | jmp .ret | ||
| 955 : | |||
| 956 : | align 16 | ||
| 957 : | .two | ||
| 958 : | Z_COND_IDCT edx+96, edx+104,edx+112,edx+120,esp+96, paddd, [coeffs], 11, .three | ||
| 959 : | IDCT2 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
| 960 : | IDCT2 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
| 961 : | IDCT2 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
| 962 : | IDCT2 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
| 963 : | jmp .ret | ||
| 964 : | |||
| 965 : | align 16 | ||
| 966 : | .three | ||
| 967 : | IDCT3 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
| 968 : | IDCT3 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
| 969 : | IDCT3 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
| 970 : | IDCT3 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
| 971 : | jmp .ret | ||
| 972 : | |||
| 973 : | align 16 | ||
| 974 : | .five | ||
| 975 : | IDCT5 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
| 976 : | ; IDCT5 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
| 977 : | IDCT5 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
| 978 : | ; IDCT5 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
| 979 : | jmp .ret | ||
| 980 : | |||
| 981 : | align 16 | ||
| 982 : | .one | ||
| 983 : | IDCT1 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
| 984 : | IDCT1 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
| 985 : | IDCT1 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
| 986 : | IDCT1 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
| 987 : | jmp .ret | ||
| 988 : | |||
| 989 : | align 16 | ||
| 990 : | .seven | ||
| 991 : | IDCT7 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
| 992 : | ; IDCT7 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
| 993 : | IDCT7 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
| 994 : | ; IDCT7 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
| 995 : | |||
| 996 : | .ret | ||
| 997 : | add esp, 128 | ||
| 998 : | ret | ||
| 999 : | |||
| 1000 : | |||
| 1001 : | Isibaar | 1055 | ;------------------ again with permuted parms -------- |
| 1002 : | ; | ||
| 1003 : | ; simple_idct_mmx is the same function as simple_idct_mmx_P above except that on entry it will | ||
| 1004 : | ; do a fast in-line and in-place permutation on the iDCT parm list. This means that same parm list | ||
| 1005 : | ; will also not have to be copied on the way out. - trbarry 6/2003 | ||
| 1006 : | |||
| 1007 : | %macro XLODA 2 | ||
| 1008 : | mov bx, [srcP+2*%2] ; get src contents | ||
| 1009 : | mov ax, [srcP+2*%1] ; get dest contents | ||
| 1010 : | mov [srcP+2*%1], bx ; store new dest val | ||
| 1011 : | %endmacro | ||
| 1012 : | |||
| 1013 : | %macro XCHGA 2 | ||
| 1014 : | mov ax, [srcP+2*%1] ; get dest contents | ||
| 1015 : | mov [srcP+2*%1], bx ; store new dest val | ||
| 1016 : | %endmacro | ||
| 1017 : | |||
| 1018 : | %macro XCHGB 2 | ||
| 1019 : | mov bx, [srcP+2*%1] ; get dest contents | ||
| 1020 : | mov [srcP+2*%1], ax ; store new dest val | ||
| 1021 : | %endmacro | ||
| 1022 : | |||
| 1023 : | %macro XSTRA 2 | ||
| 1024 : | mov [srcP+2*%1], bx ; store dest val | ||
| 1025 : | %endmacro | ||
| 1026 : | |||
| 1027 : | %macro XSTRB 2 | ||
| 1028 : | mov [srcP+2*%1], ax ; store dest val | ||
| 1029 : | %endmacro | ||
| 1030 : | |||
| 1031 : | %macro PERMUTEP 1 | ||
| 1032 : | %define srcP %1 | ||
| 1033 : | push ebx | ||
| 1034 : | |||
| 1035 : | ; XCHGA 0x00, 0x00 ; nothing to do | ||
| 1036 : | |||
| 1037 : | XLODA 0x08, 0x01 | ||
| 1038 : | XCHGB 0x10, 0x08 | ||
| 1039 : | XCHGA 0x20, 0x10 | ||
| 1040 : | XCHGB 0x02, 0x20 | ||
| 1041 : | XCHGA 0x04, 0x02 | ||
| 1042 : | XSTRB 0x01, 0x04 | ||
| 1043 : | |||
| 1044 : | XLODA 0x09, 0x03 | ||
| 1045 : | XCHGB 0x18, 0x09 | ||
| 1046 : | XCHGA 0x12, 0x18 | ||
| 1047 : | XCHGB 0x24, 0x12 | ||
| 1048 : | XSTRA 0x03, 0x24 | ||
| 1049 : | |||
| 1050 : | XLODA 0x0C, 0x05 | ||
| 1051 : | XCHGB 0x11, 0x0C | ||
| 1052 : | XCHGA 0x28, 0x11 | ||
| 1053 : | XCHGB 0x30, 0x28 | ||
| 1054 : | XCHGA 0x22, 0x30 | ||
| 1055 : | XCHGB 0x06, 0x22 | ||
| 1056 : | XSTRA 0x05, 0x06 | ||
| 1057 : | |||
| 1058 : | XLODA 0x0D, 0x07 | ||
| 1059 : | XCHGB 0x1C, 0x0D | ||
| 1060 : | XCHGA 0x13, 0x1C | ||
| 1061 : | XCHGB 0x29, 0x13 | ||
| 1062 : | XCHGA 0x38, 0x29 | ||
| 1063 : | XCHGB 0x32, 0x38 | ||
| 1064 : | XCHGA 0x26, 0x32 | ||
| 1065 : | XSTRB 0x07, 0x26 | ||
| 1066 : | |||
| 1067 : | XLODA 0x14, 0x0A | ||
| 1068 : | XCHGB 0x21, 0x14 | ||
| 1069 : | XSTRA 0x0A, 0x21 | ||
| 1070 : | |||
| 1071 : | XLODA 0x19, 0x0B | ||
| 1072 : | XCHGB 0x1A, 0x19 | ||
| 1073 : | XCHGA 0x16, 0x1A | ||
| 1074 : | XCHGB 0x25, 0x16 | ||
| 1075 : | XCHGA 0x0E, 0x25 | ||
| 1076 : | XCHGB 0x15, 0x0E | ||
| 1077 : | XCHGA 0x2C, 0x15 | ||
| 1078 : | XCHGB 0x31, 0x2C | ||
| 1079 : | XCHGA 0x2A, 0x31 | ||
| 1080 : | XCHGB 0x34, 0x2A | ||
| 1081 : | XCHGA 0x23, 0x34 | ||
| 1082 : | XSTRB 0x0B, 0x23 | ||
| 1083 : | |||
| 1084 : | XLODA 0x1D, 0x0F | ||
| 1085 : | XCHGB 0x1E, 0x1D | ||
| 1086 : | XCHGA 0x17, 0x1E | ||
| 1087 : | XCHGB 0x2D, 0x17 | ||
| 1088 : | XCHGA 0x3C, 0x2D | ||
| 1089 : | XCHGB 0x33, 0x3C | ||
| 1090 : | XCHGA 0x2B, 0x33 | ||
| 1091 : | XCHGB 0x39, 0x2B | ||
| 1092 : | XCHGA 0x3A, 0x39 | ||
| 1093 : | XCHGB 0x36, 0x3A | ||
| 1094 : | XCHGA 0x27, 0x36 | ||
| 1095 : | XSTRB 0x0F, 0x27 | ||
| 1096 : | |||
| 1097 : | ; XCHGA 0x1B, 0x1B | ||
| 1098 : | |||
| 1099 : | ; XCHGA 0x1F, 0x1F | ||
| 1100 : | |||
| 1101 : | XLODA 0x35, 0x2E | ||
| 1102 : | XSTRB 0x2E, 0x35 | ||
| 1103 : | |||
| 1104 : | XLODA 0x3D, 0x2F | ||
| 1105 : | XCHGB 0x3E, 0x3D | ||
| 1106 : | XCHGA 0x37, 0x3E | ||
| 1107 : | XSTRB 0x2F, 0x37 | ||
| 1108 : | |||
| 1109 : | ; XCHGA 0x3B, 0x3B | ||
| 1110 : | |||
| 1111 : | ; XCHGA 0x3F, 0x3F | ||
| 1112 : | pop ebx | ||
| 1113 : | %undef srcP | ||
| 1114 : | %endmacro | ||
| 1115 : | |||
| 1116 : | ; void simple_idct_mmx(int16_t * const block); | ||
| 1117 : | align 16 | ||
| 1118 : | cglobal simple_idct_mmx | ||
| 1119 : | |||
| 1120 : | simple_idct_mmx | ||
| 1121 : | sub esp, 128 | ||
| 1122 : | mov edx, [esp+128+4] | ||
| 1123 : | PERMUTEP edx ; permute parm list in place | ||
| 1124 : | |||
| 1125 : | ; src0, src4, src1, src5, dst, rndop, rndarg, shift, bt | ||
| 1126 : | |||
| 1127 : | DC_COND_IDCT edx+0, edx+8, edx+16, edx+24, esp, paddd, [coeffs+8], 11 | ||
| 1128 : | Z_COND_IDCT edx+32, edx+40, edx+48, edx+56, esp+32, paddd, [coeffs], 11, .fourP | ||
| 1129 : | Z_COND_IDCT edx+64, edx+72, edx+80, edx+88, esp+64, paddd, [coeffs], 11, .twoP | ||
| 1130 : | Z_COND_IDCT edx+96, edx+104,edx+112,edx+120,esp+96, paddd, [coeffs], 11, .oneP | ||
| 1131 : | IDCT0 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
| 1132 : | IDCT0 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
| 1133 : | IDCT0 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
| 1134 : | IDCT0 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
| 1135 : | jmp .retP | ||
| 1136 : | |||
| 1137 : | align 16 | ||
| 1138 : | .fourP | ||
| 1139 : | Z_COND_IDCT edx+64, edx+72, edx+80, edx+88, esp+64, paddd, [coeffs], 11, .sixP | ||
| 1140 : | Z_COND_IDCT edx+96, edx+104,edx+112,edx+120,esp+96, paddd, [coeffs], 11, .fiveP | ||
| 1141 : | IDCT4 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
| 1142 : | IDCT4 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
| 1143 : | IDCT4 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
| 1144 : | IDCT4 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
| 1145 : | jmp .retP | ||
| 1146 : | |||
| 1147 : | align 16 | ||
| 1148 : | .sixP | ||
| 1149 : | Z_COND_IDCT edx+96, edx+104,edx+112,edx+120,esp+96, paddd, [coeffs], 11, .sevenP | ||
| 1150 : | IDCT6 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
| 1151 : | IDCT6 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
| 1152 : | IDCT6 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
| 1153 : | IDCT6 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
| 1154 : | jmp .retP | ||
| 1155 : | |||
| 1156 : | align 16 | ||
| 1157 : | .twoP | ||
| 1158 : | Z_COND_IDCT edx+96, edx+104,edx+112,edx+120,esp+96, paddd, [coeffs], 11, .threeP | ||
| 1159 : | IDCT2 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
| 1160 : | IDCT2 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
| 1161 : | IDCT2 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
| 1162 : | IDCT2 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
| 1163 : | jmp .retP | ||
| 1164 : | |||
| 1165 : | align 16 | ||
| 1166 : | .threeP | ||
| 1167 : | IDCT3 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
| 1168 : | IDCT3 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
| 1169 : | IDCT3 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
| 1170 : | IDCT3 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
| 1171 : | jmp .retP | ||
| 1172 : | |||
| 1173 : | align 16 | ||
| 1174 : | .fiveP | ||
| 1175 : | IDCT5 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
| 1176 : | ; IDCT5 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
| 1177 : | IDCT5 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
| 1178 : | ; IDCT5 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
| 1179 : | jmp .retP | ||
| 1180 : | |||
| 1181 : | align 16 | ||
| 1182 : | .oneP | ||
| 1183 : | IDCT1 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
| 1184 : | IDCT1 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
| 1185 : | IDCT1 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
| 1186 : | IDCT1 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
| 1187 : | jmp .retP | ||
| 1188 : | |||
| 1189 : | align 16 | ||
| 1190 : | .sevenP | ||
| 1191 : | IDCT7 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
| 1192 : | ; IDCT7 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
| 1193 : | IDCT7 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
| 1194 : | ; IDCT7 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
| 1195 : | |||
| 1196 : | .retP | ||
| 1197 : | add esp, 128 | ||
| 1198 : | ret | ||
| 1199 : | |||
| 1200 : |
| No admin address has been configured | ViewVC Help |
| Powered by ViewVC 1.0.4 |