Parent Directory
| 
Revision Log
Revision 1839 - (view) (download)
| 1 : | edgomez | 1382 | ;/***************************************************************************** |
| 2 : | ; * | ||
| 3 : | ; * XVID MPEG-4 VIDEO CODEC | ||
| 4 : | ; * - MPEG4 Quantization H263 implementation / MMX optimized - | ||
| 5 : | ; * | ||
| 6 : | ; * Copyright(C) 2001-2003 Peter Ross <pross@xvid.org> | ||
| 7 : | ; * 2002-2003 Pascal Massimino <skal@planet-d.net> | ||
| 8 : | edgomez | 1424 | ; * 2004 Jean-Marc Bastide <jmtest@voila.fr> |
| 9 : | edgomez | 1382 | ; * |
| 10 : | ; * This program is free software ; you can redistribute it and/or modify | ||
| 11 : | ; * it under the terms of the GNU General Public License as published by | ||
| 12 : | ; * the Free Software Foundation ; either version 2 of the License, or | ||
| 13 : | ; * (at your option) any later version. | ||
| 14 : | ; * | ||
| 15 : | ; * This program is distributed in the hope that it will be useful, | ||
| 16 : | ; * but WITHOUT ANY WARRANTY ; without even the implied warranty of | ||
| 17 : | ; * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | ||
| 18 : | ; * GNU General Public License for more details. | ||
| 19 : | ; * | ||
| 20 : | ; * You should have received a copy of the GNU General Public License | ||
| 21 : | ; * along with this program ; if not, write to the Free Software | ||
| 22 : | ; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | ||
| 23 : | ; * | ||
| 24 : | Isibaar | 1839 | ; * $Id: quantize_h263_mmx.asm,v 1.13 2008-12-02 13:44:55 Isibaar Exp $ |
| 25 : | edgomez | 1382 | ; * |
| 26 : | ; ****************************************************************************/ | ||
| 27 : | |||
| 28 : | ; enable dequant saturate [-2048,2047], test purposes only. | ||
| 29 : | %define SATURATE | ||
| 30 : | |||
| 31 : | Isibaar | 1795 | %include "nasm.inc" |
| 32 : | edgomez | 1382 | |
| 33 : | ;============================================================================= | ||
| 34 : | ; Read only Local data | ||
| 35 : | ;============================================================================= | ||
| 36 : | |||
| 37 : | Isibaar | 1795 | DATA |
| 38 : | edgomez | 1382 | |
| 39 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 40 : | edgomez | 1382 | plus_one: |
| 41 : | times 8 dw 1 | ||
| 42 : | |||
| 43 : | ;----------------------------------------------------------------------------- | ||
| 44 : | ; | ||
| 45 : | edgomez | 1424 | ; quant table |
| 46 : | ; | ||
| 47 : | ;----------------------------------------------------------------------------- | ||
| 48 : | |||
| 49 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 50 : | edgomez | 1424 | mmx_quant: |
| 51 : | %assign quant 0 | ||
| 52 : | %rep 32 | ||
| 53 : | times 4 dw quant | ||
| 54 : | %assign quant quant+1 | ||
| 55 : | %endrep | ||
| 56 : | |||
| 57 : | ;----------------------------------------------------------------------------- | ||
| 58 : | ; | ||
| 59 : | edgomez | 1382 | ; subtract by Q/2 table |
| 60 : | ; | ||
| 61 : | ;----------------------------------------------------------------------------- | ||
| 62 : | |||
| 63 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 64 : | edgomez | 1382 | mmx_sub: |
| 65 : | %assign quant 1 | ||
| 66 : | %rep 31 | ||
| 67 : | times 4 dw quant / 2 | ||
| 68 : | %assign quant quant+1 | ||
| 69 : | %endrep | ||
| 70 : | |||
| 71 : | ;----------------------------------------------------------------------------- | ||
| 72 : | ; | ||
| 73 : | ; divide by 2Q table | ||
| 74 : | ; | ||
| 75 : | ; use a shift of 16 to take full advantage of _pmulhw_ | ||
| 76 : | ; for q=1, _pmulhw_ will overflow so it is treated seperately | ||
| 77 : | ; (3dnow2 provides _pmulhuw_ which wont cause overflow) | ||
| 78 : | ; | ||
| 79 : | ;----------------------------------------------------------------------------- | ||
| 80 : | |||
| 81 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 82 : | edgomez | 1382 | mmx_div: |
| 83 : | %assign quant 1 | ||
| 84 : | %rep 31 | ||
| 85 : | times 4 dw (1<<16) / (quant*2) + 1 | ||
| 86 : | %assign quant quant+1 | ||
| 87 : | %endrep | ||
| 88 : | |||
| 89 : | ;============================================================================= | ||
| 90 : | ; Code | ||
| 91 : | ;============================================================================= | ||
| 92 : | |||
| 93 : | Isibaar | 1795 | SECTION .rotext align=SECTION_ALIGN |
| 94 : | edgomez | 1382 | |
| 95 : | cglobal quant_h263_intra_mmx | ||
| 96 : | cglobal quant_h263_intra_sse2 | ||
| 97 : | cglobal quant_h263_inter_mmx | ||
| 98 : | cglobal quant_h263_inter_sse2 | ||
| 99 : | cglobal dequant_h263_intra_mmx | ||
| 100 : | cglobal dequant_h263_intra_xmm | ||
| 101 : | cglobal dequant_h263_intra_sse2 | ||
| 102 : | cglobal dequant_h263_inter_mmx | ||
| 103 : | cglobal dequant_h263_inter_xmm | ||
| 104 : | cglobal dequant_h263_inter_sse2 | ||
| 105 : | |||
| 106 : | ;----------------------------------------------------------------------------- | ||
| 107 : | ; | ||
| 108 : | ; uint32_t quant_h263_intra_mmx(int16_t * coeff, | ||
| 109 : | ; const int16_t const * data, | ||
| 110 : | ; const uint32_t quant, | ||
| 111 : | ; const uint32_t dcscalar, | ||
| 112 : | ; const uint16_t *mpeg_matrices); | ||
| 113 : | ; | ||
| 114 : | ;----------------------------------------------------------------------------- | ||
| 115 : | |||
| 116 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 117 : | edgomez | 1382 | quant_h263_intra_mmx: |
| 118 : | |||
| 119 : | Isibaar | 1795 | mov _EAX, prm2 ; data |
| 120 : | mov TMP0, prm4 ; dcscalar | ||
| 121 : | movsx _EAX, word [_EAX] ; data[0] | ||
| 122 : | |||
| 123 : | sar TMP0, 1 ; dcscalar /2 | ||
| 124 : | mov TMP1, _EAX | ||
| 125 : | sar TMP1, 31 ; sgn(data[0]) | ||
| 126 : | xor TMP0,TMP1 ; *sgn(data[0]) | ||
| 127 : | sub _EAX,TMP1 | ||
| 128 : | add _EAX,TMP0 ; + (dcscalar/2)*sgn(data[0]) | ||
| 129 : | edgomez | 1382 | |
| 130 : | Isibaar | 1795 | mov TMP0, prm3 ; quant |
| 131 : | Isibaar | 1803 | lea TMP1, [mmx_div] |
| 132 : | movq mm7, [TMP1+TMP0 * 8 - 8] | ||
| 133 : | %ifdef ARCH_IS_X86_64 | ||
| 134 : | %ifdef WINDOWS | ||
| 135 : | mov TMP1, prm2 | ||
| 136 : | %endif | ||
| 137 : | %endif | ||
| 138 : | edgomez | 1424 | cdq |
| 139 : | Isibaar | 1795 | idiv prm4d ; dcscalar |
| 140 : | Isibaar | 1803 | %ifdef ARCH_IS_X86_64 |
| 141 : | %ifdef WINDOWS | ||
| 142 : | mov prm2, TMP1 | ||
| 143 : | %endif | ||
| 144 : | %endif | ||
| 145 : | Isibaar | 1795 | cmp TMP0, 1 |
| 146 : | mov TMP1, prm1 ; coeff | ||
| 147 : | edgomez | 1424 | je .low |
| 148 : | |||
| 149 : | Isibaar | 1795 | mov TMP0, prm2 ; data |
| 150 : | push _EAX ; DC | ||
| 151 : | mov _EAX, TMP0 | ||
| 152 : | edgomez | 1382 | |
| 153 : | Isibaar | 1795 | mov TMP0,4 |
| 154 : | |||
| 155 : | Isibaar | 1793 | .loop: |
| 156 : | Isibaar | 1795 | movq mm0, [_EAX] ; data |
| 157 : | edgomez | 1424 | pxor mm4,mm4 |
| 158 : | Isibaar | 1795 | movq mm1, [_EAX + 8] |
| 159 : | edgomez | 1424 | pcmpgtw mm4,mm0 ; (data<0) |
| 160 : | pxor mm5,mm5 | ||
| 161 : | pmulhw mm0,mm7 ; /(2*quant) | ||
| 162 : | pcmpgtw mm5,mm1 | ||
| 163 : | Isibaar | 1795 | movq mm2, [_EAX+16] |
| 164 : | edgomez | 1424 | psubw mm0,mm4 ; +(data<0) |
| 165 : | pmulhw mm1,mm7 | ||
| 166 : | pxor mm4,mm4 | ||
| 167 : | Isibaar | 1795 | movq mm3,[_EAX+24] |
| 168 : | edgomez | 1424 | pcmpgtw mm4,mm2 |
| 169 : | psubw mm1,mm5 | ||
| 170 : | pmulhw mm2,mm7 | ||
| 171 : | pxor mm5,mm5 | ||
| 172 : | pcmpgtw mm5,mm3 | ||
| 173 : | pmulhw mm3,mm7 | ||
| 174 : | psubw mm2,mm4 | ||
| 175 : | psubw mm3,mm5 | ||
| 176 : | Isibaar | 1795 | movq [TMP1], mm0 |
| 177 : | lea _EAX, [_EAX+32] | ||
| 178 : | movq [TMP1 + 8], mm1 | ||
| 179 : | movq [TMP1 + 16], mm2 | ||
| 180 : | movq [TMP1 + 24], mm3 | ||
| 181 : | edgomez | 1424 | |
| 182 : | Isibaar | 1795 | dec TMP0 |
| 183 : | lea TMP1, [TMP1+32] | ||
| 184 : | edgomez | 1424 | jne .loop |
| 185 : | jmp .end | ||
| 186 : | |||
| 187 : | Isibaar | 1793 | .low: |
| 188 : | Isibaar | 1795 | movd mm7,TMP0d |
| 189 : | |||
| 190 : | mov TMP0, prm2 | ||
| 191 : | push _EAX | ||
| 192 : | mov _EAX, TMP0 | ||
| 193 : | |||
| 194 : | mov TMP0,4 | ||
| 195 : | Isibaar | 1793 | .loop_low: |
| 196 : | Isibaar | 1795 | movq mm0, [_EAX] |
| 197 : | edgomez | 1424 | pxor mm4,mm4 |
| 198 : | Isibaar | 1795 | movq mm1, [_EAX + 8] |
| 199 : | edgomez | 1424 | pcmpgtw mm4,mm0 |
| 200 : | pxor mm5,mm5 | ||
| 201 : | psubw mm0,mm4 | ||
| 202 : | pcmpgtw mm5,mm1 | ||
| 203 : | psraw mm0,mm7 | ||
| 204 : | psubw mm1,mm5 | ||
| 205 : | Isibaar | 1795 | movq mm2,[_EAX+16] |
| 206 : | edgomez | 1424 | pxor mm4,mm4 |
| 207 : | psraw mm1,mm7 | ||
| 208 : | pcmpgtw mm4,mm2 | ||
| 209 : | pxor mm5,mm5 | ||
| 210 : | psubw mm2,mm4 | ||
| 211 : | Isibaar | 1795 | movq mm3,[_EAX+24] |
| 212 : | edgomez | 1424 | pcmpgtw mm5,mm3 |
| 213 : | psraw mm2,mm7 | ||
| 214 : | psubw mm3,mm5 | ||
| 215 : | Isibaar | 1795 | movq [TMP1], mm0 |
| 216 : | edgomez | 1424 | psraw mm3,mm7 |
| 217 : | Isibaar | 1795 | movq [TMP1 + 8], mm1 |
| 218 : | movq [TMP1+16],mm2 | ||
| 219 : | lea _EAX, [_EAX+32] | ||
| 220 : | movq [TMP1+24],mm3 | ||
| 221 : | edgomez | 1424 | |
| 222 : | Isibaar | 1795 | dec TMP0 |
| 223 : | lea TMP1, [TMP1+32] | ||
| 224 : | edgomez | 1424 | jne .loop_low |
| 225 : | |||
| 226 : | Isibaar | 1793 | .end: |
| 227 : | edgomez | 1382 | |
| 228 : | Isibaar | 1795 | pop _EAX |
| 229 : | |||
| 230 : | mov TMP1, prm1 ; coeff | ||
| 231 : | mov [TMP1],ax | ||
| 232 : | xor _EAX,_EAX ; return 0 | ||
| 233 : | |||
| 234 : | edgomez | 1382 | ret |
| 235 : | Isibaar | 1793 | ENDFUNC |
| 236 : | edgomez | 1424 | |
| 237 : | edgomez | 1382 | |
| 238 : | ;----------------------------------------------------------------------------- | ||
| 239 : | ; | ||
| 240 : | ; uint32_t quant_h263_intra_sse2(int16_t * coeff, | ||
| 241 : | ; const int16_t const * data, | ||
| 242 : | ; const uint32_t quant, | ||
| 243 : | ; const uint32_t dcscalar, | ||
| 244 : | ; const uint16_t *mpeg_matrices); | ||
| 245 : | ; | ||
| 246 : | ;----------------------------------------------------------------------------- | ||
| 247 : | |||
| 248 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 249 : | edgomez | 1382 | quant_h263_intra_sse2: |
| 250 : | Isibaar | 1839 | PUSH_XMM6_XMM7 |
| 251 : | Isibaar | 1795 | mov _EAX, prm2 ; data |
| 252 : | edgomez | 1424 | |
| 253 : | Isibaar | 1795 | movsx _EAX, word [_EAX] ; data[0] |
| 254 : | edgomez | 1424 | |
| 255 : | Isibaar | 1795 | mov TMP0,prm4 ; dcscalar |
| 256 : | mov TMP1,_EAX | ||
| 257 : | sar TMP0,1 | ||
| 258 : | add _EAX,TMP0 | ||
| 259 : | sub TMP1,TMP0 | ||
| 260 : | cmovl _EAX,TMP1 ; +/- dcscalar/2 | ||
| 261 : | mov TMP0, prm3 ; quant | ||
| 262 : | Isibaar | 1803 | lea TMP1, [mmx_div] |
| 263 : | movq xmm7, [TMP1+TMP0 * 8 - 8] | ||
| 264 : | |||
| 265 : | %ifdef ARCH_IS_X86_64 | ||
| 266 : | %ifdef WINDOWS | ||
| 267 : | mov TMP1, prm2 | ||
| 268 : | %endif | ||
| 269 : | %endif | ||
| 270 : | edgomez | 1424 | cdq |
| 271 : | Isibaar | 1795 | idiv prm4d ; dcscalar |
| 272 : | Isibaar | 1803 | %ifdef ARCH_IS_X86_64 |
| 273 : | %ifdef WINDOWS | ||
| 274 : | mov prm2, TMP1 | ||
| 275 : | %endif | ||
| 276 : | %endif | ||
| 277 : | Isibaar | 1795 | cmp TMP0, 1 |
| 278 : | mov TMP1, prm1 ; coeff | ||
| 279 : | je near .low | ||
| 280 : | edgomez | 1424 | |
| 281 : | Isibaar | 1795 | mov TMP0, prm2 |
| 282 : | push _EAX ; DC | ||
| 283 : | mov _EAX, TMP0 | ||
| 284 : | |||
| 285 : | mov TMP0,2 | ||
| 286 : | edgomez | 1424 | movlhps xmm7,xmm7 |
| 287 : | edgomez | 1382 | |
| 288 : | Isibaar | 1793 | .loop: |
| 289 : | Isibaar | 1795 | movdqa xmm0, [_EAX] |
| 290 : | edgomez | 1424 | pxor xmm4,xmm4 |
| 291 : | Isibaar | 1795 | movdqa xmm1, [_EAX + 16] |
| 292 : | edgomez | 1424 | pcmpgtw xmm4,xmm0 |
| 293 : | pxor xmm5,xmm5 | ||
| 294 : | pmulhw xmm0,xmm7 | ||
| 295 : | pcmpgtw xmm5,xmm1 | ||
| 296 : | Isibaar | 1795 | movdqa xmm2, [_EAX+32] |
| 297 : | edgomez | 1424 | psubw xmm0,xmm4 |
| 298 : | pmulhw xmm1,xmm7 | ||
| 299 : | pxor xmm4,xmm4 | ||
| 300 : | Isibaar | 1795 | movdqa xmm3,[_EAX+48] |
| 301 : | edgomez | 1424 | pcmpgtw xmm4,xmm2 |
| 302 : | psubw xmm1,xmm5 | ||
| 303 : | pmulhw xmm2,xmm7 | ||
| 304 : | pxor xmm5,xmm5 | ||
| 305 : | pcmpgtw xmm5,xmm3 | ||
| 306 : | pmulhw xmm3,xmm7 | ||
| 307 : | psubw xmm2,xmm4 | ||
| 308 : | psubw xmm3,xmm5 | ||
| 309 : | Isibaar | 1795 | movdqa [TMP1], xmm0 |
| 310 : | lea _EAX, [_EAX+64] | ||
| 311 : | movdqa [TMP1 + 16], xmm1 | ||
| 312 : | movdqa [TMP1 + 32], xmm2 | ||
| 313 : | movdqa [TMP1 + 48], xmm3 | ||
| 314 : | edgomez | 1424 | |
| 315 : | Isibaar | 1795 | dec TMP0 |
| 316 : | lea TMP1, [TMP1+64] | ||
| 317 : | edgomez | 1424 | jne .loop |
| 318 : | jmp .end | ||
| 319 : | |||
| 320 : | Isibaar | 1793 | .low: |
| 321 : | Isibaar | 1795 | movd xmm7,TMP0d |
| 322 : | |||
| 323 : | mov TMP0, prm2 | ||
| 324 : | push _EAX ; DC | ||
| 325 : | mov _EAX, TMP0 | ||
| 326 : | |||
| 327 : | mov TMP0,2 | ||
| 328 : | Isibaar | 1793 | .loop_low: |
| 329 : | Isibaar | 1795 | movdqa xmm0, [_EAX] |
| 330 : | edgomez | 1424 | pxor xmm4,xmm4 |
| 331 : | Isibaar | 1795 | movdqa xmm1, [_EAX + 16] |
| 332 : | edgomez | 1424 | pcmpgtw xmm4,xmm0 |
| 333 : | pxor xmm5,xmm5 | ||
| 334 : | psubw xmm0,xmm4 | ||
| 335 : | pcmpgtw xmm5,xmm1 | ||
| 336 : | psraw xmm0,xmm7 | ||
| 337 : | psubw xmm1,xmm5 | ||
| 338 : | Isibaar | 1795 | movdqa xmm2,[_EAX+32] |
| 339 : | edgomez | 1424 | pxor xmm4,xmm4 |
| 340 : | psraw xmm1,xmm7 | ||
| 341 : | pcmpgtw xmm4,xmm2 | ||
| 342 : | pxor xmm5,xmm5 | ||
| 343 : | psubw xmm2,xmm4 | ||
| 344 : | Isibaar | 1795 | movdqa xmm3,[_EAX+48] |
| 345 : | edgomez | 1424 | pcmpgtw xmm5,xmm3 |
| 346 : | psraw xmm2,xmm7 | ||
| 347 : | psubw xmm3,xmm5 | ||
| 348 : | Isibaar | 1795 | movdqa [TMP1], xmm0 |
| 349 : | edgomez | 1424 | psraw xmm3,xmm7 |
| 350 : | Isibaar | 1795 | movdqa [TMP1+16], xmm1 |
| 351 : | movdqa [TMP1+32],xmm2 | ||
| 352 : | lea _EAX, [_EAX+64] | ||
| 353 : | movdqa [TMP1+48],xmm3 | ||
| 354 : | edgomez | 1424 | |
| 355 : | Isibaar | 1795 | dec TMP0 |
| 356 : | lea TMP1, [TMP1+64] | ||
| 357 : | edgomez | 1424 | jne .loop_low |
| 358 : | |||
| 359 : | Isibaar | 1793 | .end: |
| 360 : | edgomez | 1382 | |
| 361 : | Isibaar | 1795 | pop _EAX |
| 362 : | |||
| 363 : | mov TMP1, prm1 ; coeff | ||
| 364 : | mov [TMP1],ax | ||
| 365 : | xor _EAX,_EAX ; return 0 | ||
| 366 : | Isibaar | 1839 | POP_XMM6_XMM7 |
| 367 : | edgomez | 1382 | ret |
| 368 : | Isibaar | 1793 | ENDFUNC |
| 369 : | edgomez | 1424 | |
| 370 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 371 : | ; | ||
| 372 : | ; uint32_t quant_h263_inter_mmx(int16_t * coeff, | ||
| 373 : | ; const int16_t const * data, | ||
| 374 : | ; const uint32_t quant, | ||
| 375 : | ; const uint16_t *mpeg_matrices); | ||
| 376 : | ; | ||
| 377 : | ;----------------------------------------------------------------------------- | ||
| 378 : | edgomez | 1424 | |
| 379 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 380 : | edgomez | 1382 | quant_h263_inter_mmx: |
| 381 : | |||
| 382 : | Isibaar | 1795 | mov TMP1, prm1 ; coeff |
| 383 : | mov _EAX, prm3 ; quant | ||
| 384 : | edgomez | 1382 | |
| 385 : | pxor mm5, mm5 ; sum | ||
| 386 : | Isibaar | 1795 | lea TMP0, [mmx_sub] |
| 387 : | movq mm6, [TMP0 + _EAX * 8 - 8] ; sub | ||
| 388 : | edgomez | 1382 | |
| 389 : | cmp al, 1 | ||
| 390 : | Isibaar | 1795 | jz near .q1routine |
| 391 : | edgomez | 1382 | |
| 392 : | Isibaar | 1795 | lea TMP0, [mmx_div] |
| 393 : | movq mm7, [TMP0 + _EAX * 8 - 8] ; divider | ||
| 394 : | edgomez | 1382 | |
| 395 : | Isibaar | 1795 | xor TMP0, TMP0 |
| 396 : | mov _EAX, prm2 ; data | ||
| 397 : | |||
| 398 : | ALIGN SECTION_ALIGN | ||
| 399 : | Isibaar | 1793 | .loop: |
| 400 : | Isibaar | 1795 | movq mm0, [_EAX + 8*TMP0] ; mm0 = [1st] |
| 401 : | movq mm3, [_EAX + 8*TMP0 + 8] | ||
| 402 : | edgomez | 1382 | pxor mm1, mm1 ; mm1 = 0 |
| 403 : | pxor mm4, mm4 ; | ||
| 404 : | pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) | ||
| 405 : | pcmpgtw mm4, mm3 ; | ||
| 406 : | pxor mm0, mm1 ; mm0 = |mm0| | ||
| 407 : | pxor mm3, mm4 ; | ||
| 408 : | psubw mm0, mm1 ; displace | ||
| 409 : | psubw mm3, mm4 ; | ||
| 410 : | psubusw mm0, mm6 ; mm0 -= sub (unsigned, dont go < 0) | ||
| 411 : | psubusw mm3, mm6 ; | ||
| 412 : | pmulhw mm0, mm7 ; mm0 = (mm0 / 2Q) >> 16 | ||
| 413 : | pmulhw mm3, mm7 ; | ||
| 414 : | paddw mm5, mm0 ; sum += mm0 | ||
| 415 : | pxor mm0, mm1 ; mm0 *= sign(mm0) | ||
| 416 : | paddw mm5, mm3 ; | ||
| 417 : | pxor mm3, mm4 ; | ||
| 418 : | psubw mm0, mm1 ; undisplace | ||
| 419 : | psubw mm3, mm4 | ||
| 420 : | Isibaar | 1795 | movq [TMP1 + 8*TMP0], mm0 |
| 421 : | movq [TMP1 + 8*TMP0 + 8], mm3 | ||
| 422 : | edgomez | 1382 | |
| 423 : | Isibaar | 1795 | add TMP0, 2 |
| 424 : | cmp TMP0, 16 | ||
| 425 : | edgomez | 1382 | jnz .loop |
| 426 : | |||
| 427 : | Isibaar | 1793 | .done: |
| 428 : | edgomez | 1382 | pmaddwd mm5, [plus_one] |
| 429 : | movq mm0, mm5 | ||
| 430 : | psrlq mm5, 32 | ||
| 431 : | paddd mm0, mm5 | ||
| 432 : | |||
| 433 : | movd eax, mm0 ; return sum | ||
| 434 : | |||
| 435 : | ret | ||
| 436 : | |||
| 437 : | Isibaar | 1795 | .q1routine: |
| 438 : | xor TMP0, TMP0 | ||
| 439 : | mov _EAX, prm2 ; data | ||
| 440 : | |||
| 441 : | ALIGN SECTION_ALIGN | ||
| 442 : | Isibaar | 1793 | .q1loop: |
| 443 : | Isibaar | 1795 | movq mm0, [_EAX + 8*TMP0] ; mm0 = [1st] |
| 444 : | movq mm3, [_EAX + 8*TMP0+ 8] ; | ||
| 445 : | edgomez | 1382 | pxor mm1, mm1 ; mm1 = 0 |
| 446 : | pxor mm4, mm4 ; | ||
| 447 : | pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) | ||
| 448 : | pcmpgtw mm4, mm3 ; | ||
| 449 : | pxor mm0, mm1 ; mm0 = |mm0| | ||
| 450 : | pxor mm3, mm4 ; | ||
| 451 : | psubw mm0, mm1 ; displace | ||
| 452 : | psubw mm3, mm4 ; | ||
| 453 : | psubusw mm0, mm6 ; mm0 -= sub (unsigned, dont go < 0) | ||
| 454 : | psubusw mm3, mm6 ; | ||
| 455 : | psrlw mm0, 1 ; mm0 >>= 1 (/2) | ||
| 456 : | psrlw mm3, 1 ; | ||
| 457 : | paddw mm5, mm0 ; sum += mm0 | ||
| 458 : | pxor mm0, mm1 ; mm0 *= sign(mm0) | ||
| 459 : | paddw mm5, mm3 ; | ||
| 460 : | pxor mm3, mm4 ; | ||
| 461 : | psubw mm0, mm1 ; undisplace | ||
| 462 : | psubw mm3, mm4 | ||
| 463 : | Isibaar | 1795 | movq [TMP1 + 8*TMP0], mm0 |
| 464 : | movq [TMP1 + 8*TMP0 + 8], mm3 | ||
| 465 : | edgomez | 1382 | |
| 466 : | Isibaar | 1795 | add TMP0, 2 |
| 467 : | cmp TMP0, 16 | ||
| 468 : | edgomez | 1382 | jnz .q1loop |
| 469 : | |||
| 470 : | jmp .done | ||
| 471 : | Isibaar | 1793 | ENDFUNC |
| 472 : | edgomez | 1382 | |
| 473 : | |||
| 474 : | |||
| 475 : | ;----------------------------------------------------------------------------- | ||
| 476 : | ; | ||
| 477 : | ; uint32_t quant_h263_inter_sse2(int16_t * coeff, | ||
| 478 : | ; const int16_t const * data, | ||
| 479 : | ; const uint32_t quant, | ||
| 480 : | ; const uint16_t *mpeg_matrices); | ||
| 481 : | ; | ||
| 482 : | ;----------------------------------------------------------------------------- | ||
| 483 : | |||
| 484 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 485 : | edgomez | 1382 | quant_h263_inter_sse2: |
| 486 : | Isibaar | 1833 | PUSH_XMM6_XMM7 |
| 487 : | |||
| 488 : | Isibaar | 1795 | mov TMP1, prm1 ; coeff |
| 489 : | mov _EAX, prm3 ; quant | ||
| 490 : | edgomez | 1382 | |
| 491 : | pxor xmm5, xmm5 ; sum | ||
| 492 : | |||
| 493 : | Isibaar | 1795 | lea TMP0, [mmx_sub] |
| 494 : | Isibaar | 1839 | movq xmm6, [TMP0 + _EAX*8 - 8] ; sub |
| 495 : | edgomez | 1382 | movlhps xmm6, xmm6 ; duplicate into high 8 bytes |
| 496 : | |||
| 497 : | cmp al, 1 | ||
| 498 : | Isibaar | 1795 | jz near .qes2_q1_routine |
| 499 : | edgomez | 1382 | |
| 500 : | Isibaar | 1793 | .qes2_not1: |
| 501 : | Isibaar | 1795 | lea TMP0, [mmx_div] |
| 502 : | Isibaar | 1839 | movq xmm7, [TMP0 + _EAX*8 - 8] ; divider |
| 503 : | Isibaar | 1795 | |
| 504 : | xor TMP0, TMP0 | ||
| 505 : | mov _EAX, prm2 ; data | ||
| 506 : | |||
| 507 : | edgomez | 1382 | movlhps xmm7, xmm7 |
| 508 : | |||
| 509 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 510 : | Isibaar | 1793 | .qes2_loop: |
| 511 : | Isibaar | 1795 | movdqa xmm0, [_EAX + TMP0*8] ; xmm0 = [1st] |
| 512 : | movdqa xmm3, [_EAX + TMP0*8 + 16] ; xmm3 = [2nd] | ||
| 513 : | edgomez | 1382 | pxor xmm1, xmm1 |
| 514 : | pxor xmm4, xmm4 | ||
| 515 : | pcmpgtw xmm1, xmm0 | ||
| 516 : | pcmpgtw xmm4, xmm3 | ||
| 517 : | pxor xmm0, xmm1 | ||
| 518 : | pxor xmm3, xmm4 | ||
| 519 : | psubw xmm0, xmm1 | ||
| 520 : | psubw xmm3, xmm4 | ||
| 521 : | psubusw xmm0, xmm6 | ||
| 522 : | psubusw xmm3, xmm6 | ||
| 523 : | pmulhw xmm0, xmm7 | ||
| 524 : | pmulhw xmm3, xmm7 | ||
| 525 : | paddw xmm5, xmm0 | ||
| 526 : | pxor xmm0, xmm1 | ||
| 527 : | paddw xmm5, xmm3 | ||
| 528 : | pxor xmm3, xmm4 | ||
| 529 : | psubw xmm0, xmm1 | ||
| 530 : | psubw xmm3, xmm4 | ||
| 531 : | Isibaar | 1795 | movdqa [TMP1 + TMP0*8], xmm0 |
| 532 : | movdqa [TMP1 + TMP0*8 + 16], xmm3 | ||
| 533 : | edgomez | 1382 | |
| 534 : | Isibaar | 1795 | add TMP0, 4 |
| 535 : | cmp TMP0, 16 | ||
| 536 : | edgomez | 1382 | jnz .qes2_loop |
| 537 : | |||
| 538 : | Isibaar | 1793 | .qes2_done: |
| 539 : | Isibaar | 1839 | movdqa xmm6, [plus_one] |
| 540 : | edgomez | 1382 | pmaddwd xmm5, xmm6 |
| 541 : | movhlps xmm6, xmm5 | ||
| 542 : | paddd xmm5, xmm6 | ||
| 543 : | movdq2q mm0, xmm5 | ||
| 544 : | |||
| 545 : | movq mm5, mm0 | ||
| 546 : | psrlq mm5, 32 | ||
| 547 : | paddd mm0, mm5 | ||
| 548 : | |||
| 549 : | movd eax, mm0 ; return sum | ||
| 550 : | |||
| 551 : | Isibaar | 1833 | POP_XMM6_XMM7 |
| 552 : | edgomez | 1382 | ret |
| 553 : | |||
| 554 : | Isibaar | 1795 | .qes2_q1_routine: |
| 555 : | xor TMP0, TMP0 | ||
| 556 : | mov _EAX, prm2 ; data | ||
| 557 : | |||
| 558 : | ALIGN SECTION_ALIGN | ||
| 559 : | Isibaar | 1793 | .qes2_q1loop: |
| 560 : | Isibaar | 1795 | movdqa xmm0, [_EAX + TMP0*8] ; xmm0 = [1st] |
| 561 : | movdqa xmm3, [_EAX + TMP0*8 + 16] ; xmm3 = [2nd] | ||
| 562 : | edgomez | 1382 | pxor xmm1, xmm1 |
| 563 : | pxor xmm4, xmm4 | ||
| 564 : | pcmpgtw xmm1, xmm0 | ||
| 565 : | pcmpgtw xmm4, xmm3 | ||
| 566 : | pxor xmm0, xmm1 | ||
| 567 : | pxor xmm3, xmm4 | ||
| 568 : | psubw xmm0, xmm1 | ||
| 569 : | psubw xmm3, xmm4 | ||
| 570 : | psubusw xmm0, xmm6 | ||
| 571 : | psubusw xmm3, xmm6 | ||
| 572 : | psrlw xmm0, 1 | ||
| 573 : | psrlw xmm3, 1 | ||
| 574 : | paddw xmm5, xmm0 | ||
| 575 : | pxor xmm0, xmm1 | ||
| 576 : | paddw xmm5, xmm3 | ||
| 577 : | pxor xmm3, xmm4 | ||
| 578 : | psubw xmm0, xmm1 | ||
| 579 : | psubw xmm3, xmm4 | ||
| 580 : | Isibaar | 1795 | movdqa [TMP1 + TMP0*8], xmm0 |
| 581 : | movdqa [TMP1 + TMP0*8 + 16], xmm3 | ||
| 582 : | edgomez | 1382 | |
| 583 : | Isibaar | 1795 | add TMP0, 4 |
| 584 : | cmp TMP0, 16 | ||
| 585 : | edgomez | 1382 | jnz .qes2_q1loop |
| 586 : | jmp .qes2_done | ||
| 587 : | Isibaar | 1793 | ENDFUNC |
| 588 : | edgomez | 1382 | |
| 589 : | |||
| 590 : | ;----------------------------------------------------------------------------- | ||
| 591 : | ; | ||
| 592 : | ; uint32_t dequant_h263_intra_mmx(int16_t *data, | ||
| 593 : | ; const int16_t const *coeff, | ||
| 594 : | ; const uint32_t quant, | ||
| 595 : | ; const uint32_t dcscalar, | ||
| 596 : | ; const uint16_t *mpeg_matrices); | ||
| 597 : | ; | ||
| 598 : | ;----------------------------------------------------------------------------- | ||
| 599 : | |||
| 600 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 601 : | edgomez | 1382 | dequant_h263_intra_mmx: |
| 602 : | |||
| 603 : | Isibaar | 1795 | mov TMP0, prm3 ; quant |
| 604 : | mov _EAX, prm2 ; coeff | ||
| 605 : | edgomez | 1424 | pcmpeqw mm0,mm0 |
| 606 : | Isibaar | 1795 | lea TMP1, [mmx_quant] |
| 607 : | movq mm6, [TMP1 + TMP0*8] ; quant | ||
| 608 : | shl TMP0,31 ; quant & 1 ? 0 : - 1 | ||
| 609 : | edgomez | 1424 | movq mm7,mm6 |
| 610 : | movq mm5,mm0 | ||
| 611 : | Isibaar | 1795 | movd mm1,TMP0d |
| 612 : | mov TMP1, prm1 ; data | ||
| 613 : | edgomez | 1424 | psllw mm0,mm1 |
| 614 : | Isibaar | 1795 | paddw mm7,mm7 ; 2*quant |
| 615 : | paddw mm6,mm0 ; quant-1 | ||
| 616 : | edgomez | 1424 | psllw mm5,12 |
| 617 : | Isibaar | 1795 | mov TMP0,8 |
| 618 : | edgomez | 1424 | psrlw mm5,1 |
| 619 : | edgomez | 1382 | |
| 620 : | edgomez | 1424 | .loop: |
| 621 : | Isibaar | 1795 | movq mm0,[_EAX] |
| 622 : | edgomez | 1424 | pxor mm2,mm2 |
| 623 : | pxor mm4,mm4 | ||
| 624 : | pcmpgtw mm2,mm0 | ||
| 625 : | pcmpeqw mm4,mm0 | ||
| 626 : | pmullw mm0,mm7 ; * 2 * quant | ||
| 627 : | Isibaar | 1795 | movq mm1,[_EAX+8] |
| 628 : | edgomez | 1424 | psubw mm0,mm2 |
| 629 : | pxor mm2,mm6 | ||
| 630 : | pxor mm3,mm3 | ||
| 631 : | pandn mm4,mm2 | ||
| 632 : | pxor mm2,mm2 | ||
| 633 : | pcmpgtw mm3,mm1 | ||
| 634 : | pcmpeqw mm2,mm1 | ||
| 635 : | pmullw mm1,mm7 | ||
| 636 : | paddw mm0,mm4 | ||
| 637 : | psubw mm1,mm3 | ||
| 638 : | pxor mm3,mm6 | ||
| 639 : | pandn mm2,mm3 | ||
| 640 : | paddsw mm0, mm5 ; saturate | ||
| 641 : | paddw mm1,mm2 | ||
| 642 : | |||
| 643 : | paddsw mm1, mm5 | ||
| 644 : | psubsw mm0, mm5 | ||
| 645 : | psubsw mm1, mm5 | ||
| 646 : | psubsw mm0, mm5 | ||
| 647 : | psubsw mm1, mm5 | ||
| 648 : | paddsw mm0, mm5 | ||
| 649 : | paddsw mm1, mm5 | ||
| 650 : | |||
| 651 : | Isibaar | 1795 | movq [TMP1],mm0 |
| 652 : | lea _EAX,[_EAX+16] | ||
| 653 : | movq [TMP1+8],mm1 | ||
| 654 : | edgomez | 1424 | |
| 655 : | Isibaar | 1795 | dec TMP0 |
| 656 : | lea TMP1,[TMP1+16] | ||
| 657 : | edgomez | 1424 | jne .loop |
| 658 : | |||
| 659 : | ; deal with DC | ||
| 660 : | Isibaar | 1795 | mov _EAX, prm2 ; coeff |
| 661 : | movd mm1,prm4d ; dcscalar | ||
| 662 : | movd mm0,[_EAX] ; coeff[0] | ||
| 663 : | edgomez | 1432 | pmullw mm0,mm1 ; * dcscalar |
| 664 : | Isibaar | 1795 | mov TMP1, prm1 ; data |
| 665 : | edgomez | 1432 | paddsw mm0, mm5 ; saturate + |
| 666 : | psubsw mm0, mm5 | ||
| 667 : | psubsw mm0, mm5 ; saturate - | ||
| 668 : | paddsw mm0, mm5 | ||
| 669 : | movd eax,mm0 | ||
| 670 : | Isibaar | 1795 | mov [TMP1], ax |
| 671 : | edgomez | 1382 | |
| 672 : | Isibaar | 1795 | xor _EAX, _EAX ; return 0 |
| 673 : | edgomez | 1382 | ret |
| 674 : | Isibaar | 1793 | ENDFUNC |
| 675 : | edgomez | 1382 | |
| 676 : | ;----------------------------------------------------------------------------- | ||
| 677 : | ; | ||
| 678 : | ; uint32_t dequant_h263_intra_xmm(int16_t *data, | ||
| 679 : | ; const int16_t const *coeff, | ||
| 680 : | ; const uint32_t quant, | ||
| 681 : | ; const uint32_t dcscalar, | ||
| 682 : | ; const uint16_t *mpeg_matrices); | ||
| 683 : | ; | ||
| 684 : | ;----------------------------------------------------------------------------- | ||
| 685 : | |||
| 686 : | edgomez | 1424 | |
| 687 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 688 : | edgomez | 1382 | dequant_h263_intra_xmm: |
| 689 : | |||
| 690 : | Isibaar | 1795 | mov TMP0, prm3 ; quant |
| 691 : | mov _EAX, prm2 ; coeff | ||
| 692 : | edgomez | 1424 | |
| 693 : | Isibaar | 1795 | movd mm6,TMP0d ; quant |
| 694 : | edgomez | 1424 | pcmpeqw mm0,mm0 |
| 695 : | Isibaar | 1795 | pshufw mm6,mm6,0 ; all quant |
| 696 : | shl TMP0,31 | ||
| 697 : | edgomez | 1424 | movq mm5,mm0 |
| 698 : | movq mm7,mm6 | ||
| 699 : | Isibaar | 1795 | movd mm1,TMP0d |
| 700 : | mov TMP1, prm1 ; data | ||
| 701 : | psllw mm0,mm1 ; quant & 1 ? 0 : - 1 | ||
| 702 : | edgomez | 1424 | movq mm4,mm5 |
| 703 : | Isibaar | 1795 | paddw mm7,mm7 ; quant*2 |
| 704 : | paddw mm6,mm0 ; quant-1 | ||
| 705 : | psrlw mm4,5 ; mm4=2047 | ||
| 706 : | mov TMP0,8 | ||
| 707 : | pxor mm5,mm4 ; mm5=-2048 | ||
| 708 : | edgomez | 1424 | |
| 709 : | .loop: | ||
| 710 : | Isibaar | 1795 | movq mm0,[_EAX] |
| 711 : | edgomez | 1424 | pxor mm2,mm2 |
| 712 : | pxor mm3,mm3 | ||
| 713 : | edgomez | 1382 | |
| 714 : | edgomez | 1424 | pcmpgtw mm2,mm0 |
| 715 : | pcmpeqw mm3,mm0 ; if coeff==0... | ||
| 716 : | pmullw mm0,mm7 ; * 2 * quant | ||
| 717 : | Isibaar | 1795 | movq mm1,[_EAX+8] |
| 718 : | edgomez | 1424 | |
| 719 : | psubw mm0,mm2 | ||
| 720 : | pxor mm2,mm6 | ||
| 721 : | pandn mm3,mm2 ; ...then data=0 | ||
| 722 : | pxor mm2,mm2 | ||
| 723 : | paddw mm0,mm3 | ||
| 724 : | pxor mm3,mm3 | ||
| 725 : | pcmpeqw mm2,mm1 | ||
| 726 : | pcmpgtw mm3,mm1 | ||
| 727 : | pmullw mm1,mm7 | ||
| 728 : | |||
| 729 : | pminsw mm0,mm4 | ||
| 730 : | psubw mm1,mm3 | ||
| 731 : | pxor mm3,mm6 | ||
| 732 : | pandn mm2,mm3 | ||
| 733 : | paddw mm1,mm2 | ||
| 734 : | |||
| 735 : | pmaxsw mm0,mm5 | ||
| 736 : | pminsw mm1,mm4 | ||
| 737 : | Isibaar | 1795 | movq [TMP1],mm0 |
| 738 : | edgomez | 1424 | pmaxsw mm1,mm5 |
| 739 : | Isibaar | 1795 | lea _EAX,[_EAX+16] |
| 740 : | movq [TMP1+8],mm1 | ||
| 741 : | edgomez | 1424 | |
| 742 : | Isibaar | 1795 | dec TMP0 |
| 743 : | lea TMP1,[TMP1+16] | ||
| 744 : | edgomez | 1424 | jne .loop |
| 745 : | |||
| 746 : | ; deal with DC | ||
| 747 : | Isibaar | 1795 | mov _EAX, prm2 ; coeff |
| 748 : | movd mm1,prm4d ; dcscalar | ||
| 749 : | movd mm0, [_EAX] | ||
| 750 : | edgomez | 1432 | pmullw mm0, mm1 |
| 751 : | Isibaar | 1795 | mov TMP1, prm1 ; data |
| 752 : | edgomez | 1424 | pminsw mm0,mm4 |
| 753 : | pmaxsw mm0,mm5 | ||
| 754 : | edgomez | 1382 | movd eax, mm0 |
| 755 : | Isibaar | 1795 | mov [TMP1], ax |
| 756 : | edgomez | 1382 | |
| 757 : | Isibaar | 1795 | xor _EAX, _EAX ; return 0 |
| 758 : | edgomez | 1382 | ret |
| 759 : | Isibaar | 1793 | ENDFUNC |
| 760 : | edgomez | 1382 | |
| 761 : | |||
| 762 : | ;----------------------------------------------------------------------------- | ||
| 763 : | ; | ||
| 764 : | ; uint32_t dequant_h263_intra_sse2(int16_t *data, | ||
| 765 : | ; const int16_t const *coeff, | ||
| 766 : | ; const uint32_t quant, | ||
| 767 : | ; const uint32_t dcscalar, | ||
| 768 : | ; const uint16_t *mpeg_matrices); | ||
| 769 : | ; | ||
| 770 : | ;----------------------------------------------------------------------------- | ||
| 771 : | |||
| 772 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 773 : | edgomez | 1382 | dequant_h263_intra_sse2: |
| 774 : | Isibaar | 1833 | PUSH_XMM6_XMM7 |
| 775 : | |||
| 776 : | Isibaar | 1795 | mov TMP0, prm3 ; quant |
| 777 : | mov _EAX, prm2 ; coeff | ||
| 778 : | edgomez | 1424 | |
| 779 : | Isibaar | 1795 | movd xmm6,TMP0d ; quant |
| 780 : | edgomez | 1432 | |
| 781 : | Isibaar | 1795 | shl TMP0,31 |
| 782 : | edgomez | 1424 | pshuflw xmm6,xmm6,0 |
| 783 : | pcmpeqw xmm0,xmm0 | ||
| 784 : | movlhps xmm6,xmm6 ; all quant | ||
| 785 : | Isibaar | 1795 | movd xmm1,TMP0d |
| 786 : | edgomez | 1424 | movdqa xmm5,xmm0 |
| 787 : | movdqa xmm7,xmm6 | ||
| 788 : | Isibaar | 1795 | mov TMP1, prm1 ; data |
| 789 : | edgomez | 1424 | paddw xmm7,xmm7 ; quant *2 |
| 790 : | psllw xmm0,xmm1 ; quant & 1 ? 0 : - 1 | ||
| 791 : | movdqa xmm4,xmm5 | ||
| 792 : | paddw xmm6,xmm0 ; quant-1 | ||
| 793 : | psrlw xmm4,5 ; 2047 | ||
| 794 : | Isibaar | 1795 | mov TMP0,4 |
| 795 : | edgomez | 1424 | pxor xmm5,xmm4 ; mm5=-2048 |
| 796 : | |||
| 797 : | .loop: | ||
| 798 : | Isibaar | 1795 | movdqa xmm0,[_EAX] |
| 799 : | edgomez | 1424 | pxor xmm2,xmm2 |
| 800 : | pxor xmm3,xmm3 | ||
| 801 : | edgomez | 1382 | |
| 802 : | edgomez | 1424 | pcmpgtw xmm2,xmm0 |
| 803 : | pcmpeqw xmm3,xmm0 | ||
| 804 : | pmullw xmm0,xmm7 ; * 2 * quant | ||
| 805 : | Isibaar | 1795 | movdqa xmm1,[_EAX+16] |
| 806 : | edgomez | 1424 | |
| 807 : | psubw xmm0,xmm2 | ||
| 808 : | pxor xmm2,xmm6 | ||
| 809 : | pandn xmm3,xmm2 | ||
| 810 : | pxor xmm2,xmm2 | ||
| 811 : | paddw xmm0,xmm3 | ||
| 812 : | pxor xmm3,xmm3 | ||
| 813 : | pcmpeqw xmm2,xmm1 | ||
| 814 : | pcmpgtw xmm3,xmm1 | ||
| 815 : | pmullw xmm1,xmm7 | ||
| 816 : | |||
| 817 : | pminsw xmm0,xmm4 | ||
| 818 : | psubw xmm1,xmm3 | ||
| 819 : | pxor xmm3,xmm6 | ||
| 820 : | pandn xmm2,xmm3 | ||
| 821 : | paddw xmm1,xmm2 | ||
| 822 : | |||
| 823 : | pmaxsw xmm0,xmm5 | ||
| 824 : | pminsw xmm1,xmm4 | ||
| 825 : | Isibaar | 1795 | movdqa [TMP1],xmm0 |
| 826 : | edgomez | 1424 | pmaxsw xmm1,xmm5 |
| 827 : | Isibaar | 1795 | lea _EAX,[_EAX+32] |
| 828 : | movdqa [TMP1+16],xmm1 | ||
| 829 : | edgomez | 1424 | |
| 830 : | Isibaar | 1795 | dec TMP0 |
| 831 : | lea TMP1,[TMP1+32] | ||
| 832 : | edgomez | 1424 | jne .loop |
| 833 : | |||
| 834 : | ; deal with DC | ||
| 835 : | edgomez | 1382 | |
| 836 : | Isibaar | 1795 | mov _EAX, prm2 ; coeff |
| 837 : | movsx _EAX,word [_EAX] | ||
| 838 : | imul prm4d ; dcscalar | ||
| 839 : | mov TMP1, prm1 ; data | ||
| 840 : | edgomez | 1424 | movd xmm0,eax |
| 841 : | pminsw xmm0,xmm4 | ||
| 842 : | pmaxsw xmm0,xmm5 | ||
| 843 : | movd eax,xmm0 | ||
| 844 : | |||
| 845 : | Isibaar | 1795 | mov [TMP1], ax |
| 846 : | edgomez | 1382 | |
| 847 : | Isibaar | 1795 | xor _EAX, _EAX ; return 0 |
| 848 : | Isibaar | 1833 | |
| 849 : | POP_XMM6_XMM7 | ||
| 850 : | edgomez | 1382 | ret |
| 851 : | Isibaar | 1793 | ENDFUNC |
| 852 : | edgomez | 1382 | |
| 853 : | ;----------------------------------------------------------------------------- | ||
| 854 : | ; | ||
| 855 : | ; uint32t dequant_h263_inter_mmx(int16_t * data, | ||
| 856 : | ; const int16_t * const coeff, | ||
| 857 : | ; const uint32_t quant, | ||
| 858 : | ; const uint16_t *mpeg_matrices); | ||
| 859 : | ; | ||
| 860 : | ;----------------------------------------------------------------------------- | ||
| 861 : | |||
| 862 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 863 : | edgomez | 1382 | dequant_h263_inter_mmx: |
| 864 : | |||
| 865 : | Isibaar | 1795 | mov TMP0, prm3 ; quant |
| 866 : | mov _EAX, prm2 ; coeff | ||
| 867 : | edgomez | 1424 | pcmpeqw mm0,mm0 |
| 868 : | Isibaar | 1795 | lea TMP1, [mmx_quant] |
| 869 : | movq mm6, [TMP1 + TMP0*8] ; quant | ||
| 870 : | shl TMP0,31 ; odd/even | ||
| 871 : | edgomez | 1424 | movq mm7,mm6 |
| 872 : | Isibaar | 1795 | movd mm1,TMP0d |
| 873 : | mov TMP1, prm1 ; data | ||
| 874 : | edgomez | 1424 | movq mm5,mm0 |
| 875 : | Isibaar | 1795 | psllw mm0,mm1 ; quant & 1 ? 0 : - 1 |
| 876 : | paddw mm7,mm7 ; quant*2 | ||
| 877 : | paddw mm6,mm0 ; quant & 1 ? quant : quant - 1 | ||
| 878 : | edgomez | 1424 | psllw mm5,12 |
| 879 : | Isibaar | 1795 | mov TMP0,8 |
| 880 : | psrlw mm5,1 ; 32767-2047 (32768-2048) | ||
| 881 : | edgomez | 1382 | |
| 882 : | edgomez | 1424 | .loop: |
| 883 : | Isibaar | 1795 | movq mm0,[_EAX] |
| 884 : | edgomez | 1424 | pxor mm4,mm4 |
| 885 : | pxor mm2,mm2 | ||
| 886 : | pcmpeqw mm4,mm0 ; if coeff==0... | ||
| 887 : | pcmpgtw mm2,mm0 | ||
| 888 : | pmullw mm0,mm7 ; * 2 * quant | ||
| 889 : | pxor mm3,mm3 | ||
| 890 : | psubw mm0,mm2 | ||
| 891 : | Isibaar | 1795 | movq mm1,[_EAX+8] |
| 892 : | edgomez | 1424 | pxor mm2,mm6 |
| 893 : | pcmpgtw mm3,mm1 | ||
| 894 : | pandn mm4,mm2 ; ... then data==0 | ||
| 895 : | pmullw mm1,mm7 | ||
| 896 : | pxor mm2,mm2 | ||
| 897 : | pcmpeqw mm2,mm1 | ||
| 898 : | psubw mm1,mm3 | ||
| 899 : | pxor mm3,mm6 | ||
| 900 : | pandn mm2,mm3 | ||
| 901 : | paddw mm0,mm4 | ||
| 902 : | paddw mm1,mm2 | ||
| 903 : | |||
| 904 : | paddsw mm0, mm5 ; saturate | ||
| 905 : | paddsw mm1, mm5 | ||
| 906 : | psubsw mm0, mm5 | ||
| 907 : | psubsw mm1, mm5 | ||
| 908 : | psubsw mm0, mm5 | ||
| 909 : | psubsw mm1, mm5 | ||
| 910 : | paddsw mm0, mm5 | ||
| 911 : | paddsw mm1, mm5 | ||
| 912 : | |||
| 913 : | Isibaar | 1795 | movq [TMP1],mm0 |
| 914 : | lea _EAX,[_EAX+16] | ||
| 915 : | movq [TMP1+8],mm1 | ||
| 916 : | edgomez | 1424 | |
| 917 : | Isibaar | 1795 | dec TMP0 |
| 918 : | lea TMP1,[TMP1+16] | ||
| 919 : | edgomez | 1424 | jne .loop |
| 920 : | |||
| 921 : | Isibaar | 1795 | xor _EAX, _EAX ; return 0 |
| 922 : | edgomez | 1424 | ret |
| 923 : | Isibaar | 1793 | ENDFUNC |
| 924 : | edgomez | 1382 | |
| 925 : | |||
| 926 : | ;----------------------------------------------------------------------------- | ||
| 927 : | ; | ||
| 928 : | ; uint32_t dequant_h263_inter_xmm(int16_t * data, | ||
| 929 : | ; const int16_t * const coeff, | ||
| 930 : | ; const uint32_t quant, | ||
| 931 : | ; const uint16_t *mpeg_matrices); | ||
| 932 : | ; | ||
| 933 : | ;----------------------------------------------------------------------------- | ||
| 934 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 935 : | edgomez | 1382 | dequant_h263_inter_xmm: |
| 936 : | |||
| 937 : | Isibaar | 1795 | mov TMP0, prm3 ; quant |
| 938 : | mov _EAX, prm2 ; coeff | ||
| 939 : | edgomez | 1424 | pcmpeqw mm0,mm0 |
| 940 : | Isibaar | 1795 | lea TMP1, [mmx_quant] |
| 941 : | movq mm6, [TMP1 + TMP0*8] ; quant | ||
| 942 : | shl TMP0,31 | ||
| 943 : | edgomez | 1424 | movq mm5,mm0 |
| 944 : | Isibaar | 1795 | movd mm1,TMP0d |
| 945 : | edgomez | 1424 | movq mm7,mm6 |
| 946 : | psllw mm0,mm1 | ||
| 947 : | Isibaar | 1795 | mov TMP1, prm1 ; data |
| 948 : | edgomez | 1424 | movq mm4,mm5 |
| 949 : | paddw mm7,mm7 | ||
| 950 : | paddw mm6,mm0 ; quant-1 | ||
| 951 : | edgomez | 1382 | |
| 952 : | edgomez | 1424 | psrlw mm4,5 |
| 953 : | Isibaar | 1795 | mov TMP0,8 |
| 954 : | edgomez | 1424 | pxor mm5,mm4 ; mm5=-2048 |
| 955 : | |||
| 956 : | .loop: | ||
| 957 : | Isibaar | 1795 | movq mm0,[_EAX] |
| 958 : | edgomez | 1424 | pxor mm3,mm3 |
| 959 : | pxor mm2,mm2 | ||
| 960 : | pcmpeqw mm3,mm0 | ||
| 961 : | pcmpgtw mm2,mm0 | ||
| 962 : | pmullw mm0,mm7 ; * 2 * quant | ||
| 963 : | pandn mm3,mm6 | ||
| 964 : | Isibaar | 1795 | movq mm1,[_EAX+8] |
| 965 : | edgomez | 1424 | psubw mm0,mm2 |
| 966 : | pxor mm2,mm3 | ||
| 967 : | pxor mm3,mm3 | ||
| 968 : | paddw mm0,mm2 | ||
| 969 : | pxor mm2,mm2 | ||
| 970 : | pcmpgtw mm3,mm1 | ||
| 971 : | pcmpeqw mm2,mm1 | ||
| 972 : | pmullw mm1,mm7 | ||
| 973 : | pandn mm2,mm6 | ||
| 974 : | psubw mm1,mm3 | ||
| 975 : | pxor mm3,mm2 | ||
| 976 : | paddw mm1,mm3 | ||
| 977 : | |||
| 978 : | pminsw mm0,mm4 | ||
| 979 : | pminsw mm1,mm4 | ||
| 980 : | pmaxsw mm0,mm5 | ||
| 981 : | pmaxsw mm1,mm5 | ||
| 982 : | |||
| 983 : | Isibaar | 1795 | movq [TMP1],mm0 |
| 984 : | lea _EAX,[_EAX+16] | ||
| 985 : | movq [TMP1+8],mm1 | ||
| 986 : | edgomez | 1424 | |
| 987 : | Isibaar | 1795 | dec TMP0 |
| 988 : | lea TMP1,[TMP1+16] | ||
| 989 : | edgomez | 1424 | jne .loop |
| 990 : | edgomez | 1382 | |
| 991 : | Isibaar | 1795 | xor _EAX, _EAX ; return 0 |
| 992 : | edgomez | 1382 | ret |
| 993 : | Isibaar | 1793 | ENDFUNC |
| 994 : | edgomez | 1382 | |
| 995 : | edgomez | 1424 | |
| 996 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 997 : | ; | ||
| 998 : | ; uint32_t dequant_h263_inter_sse2(int16_t * data, | ||
| 999 : | ; const int16_t * const coeff, | ||
| 1000 : | ; const uint32_t quant, | ||
| 1001 : | ; const uint16_t *mpeg_matrices); | ||
| 1002 : | ; | ||
| 1003 : | ;----------------------------------------------------------------------------- | ||
| 1004 : | |||
| 1005 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 1006 : | edgomez | 1382 | dequant_h263_inter_sse2: |
| 1007 : | Isibaar | 1833 | PUSH_XMM6_XMM7 |
| 1008 : | |||
| 1009 : | Isibaar | 1795 | mov TMP0, prm3 ; quant |
| 1010 : | mov _EAX, prm2 ; coeff | ||
| 1011 : | edgomez | 1382 | |
| 1012 : | Isibaar | 1795 | lea TMP1, [mmx_quant] |
| 1013 : | movq xmm6, [TMP1 + TMP0*8] ; quant | ||
| 1014 : | inc TMP0 | ||
| 1015 : | edgomez | 1424 | pcmpeqw xmm5,xmm5 |
| 1016 : | Isibaar | 1795 | and TMP0,1 |
| 1017 : | edgomez | 1424 | movlhps xmm6,xmm6 |
| 1018 : | Isibaar | 1795 | movd xmm0,TMP0d |
| 1019 : | edgomez | 1424 | movdqa xmm7,xmm6 |
| 1020 : | pshuflw xmm0,xmm0,0 | ||
| 1021 : | movdqa xmm4,xmm5 | ||
| 1022 : | Isibaar | 1795 | mov TMP1, prm1 ; data |
| 1023 : | edgomez | 1424 | movlhps xmm0,xmm0 |
| 1024 : | paddw xmm7,xmm7 | ||
| 1025 : | psubw xmm6,xmm0 | ||
| 1026 : | psrlw xmm4,5 ; 2047 | ||
| 1027 : | Isibaar | 1795 | mov TMP0,4 |
| 1028 : | edgomez | 1424 | pxor xmm5,xmm4 ; mm5=-2048 |
| 1029 : | |||
| 1030 : | .loop: | ||
| 1031 : | Isibaar | 1795 | movdqa xmm0,[_EAX] |
| 1032 : | edgomez | 1424 | pxor xmm3,xmm3 |
| 1033 : | pxor xmm2,xmm2 | ||
| 1034 : | pcmpeqw xmm3,xmm0 | ||
| 1035 : | pcmpgtw xmm2,xmm0 | ||
| 1036 : | pmullw xmm0,xmm7 ; * 2 * quant | ||
| 1037 : | pandn xmm3,xmm6 | ||
| 1038 : | Isibaar | 1795 | movdqa xmm1,[_EAX+16] |
| 1039 : | edgomez | 1424 | psubw xmm0,xmm2 |
| 1040 : | pxor xmm2,xmm3 | ||
| 1041 : | pxor xmm3,xmm3 | ||
| 1042 : | paddw xmm0,xmm2 | ||
| 1043 : | pxor xmm2,xmm2 | ||
| 1044 : | pcmpgtw xmm3,xmm1 | ||
| 1045 : | pcmpeqw xmm2,xmm1 | ||
| 1046 : | pmullw xmm1,xmm7 | ||
| 1047 : | pandn xmm2,xmm6 | ||
| 1048 : | psubw xmm1,xmm3 | ||
| 1049 : | pxor xmm3,xmm2 | ||
| 1050 : | paddw xmm1,xmm3 | ||
| 1051 : | |||
| 1052 : | pminsw xmm0,xmm4 | ||
| 1053 : | pminsw xmm1,xmm4 | ||
| 1054 : | pmaxsw xmm0,xmm5 | ||
| 1055 : | pmaxsw xmm1,xmm5 | ||
| 1056 : | |||
| 1057 : | Isibaar | 1795 | movdqa [TMP1],xmm0 |
| 1058 : | lea _EAX,[_EAX+32] | ||
| 1059 : | movdqa [TMP1+16],xmm1 | ||
| 1060 : | edgomez | 1424 | |
| 1061 : | Isibaar | 1795 | dec TMP0 |
| 1062 : | lea TMP1,[TMP1+32] | ||
| 1063 : | edgomez | 1424 | jne .loop |
| 1064 : | edgomez | 1382 | |
| 1065 : | Isibaar | 1795 | xor _EAX, _EAX ; return 0 |
| 1066 : | Isibaar | 1833 | |
| 1067 : | POP_XMM6_XMM7 | ||
| 1068 : | edgomez | 1424 | ret |
| 1069 : | Isibaar | 1793 | ENDFUNC |
| 1070 : | edgomez | 1382 | |
| 1071 : | Isibaar | 1790 | |
| 1072 : | %ifidn __OUTPUT_FORMAT__,elf | ||
| 1073 : | section ".note.GNU-stack" noalloc noexec nowrite progbits | ||
| 1074 : | %endif | ||
| 1075 : |
| No admin address has been configured | ViewVC Help |
| Powered by ViewVC 1.0.4 |