Parent Directory
| 
Revision Log
Revision 1713 - (view) (download)
| 1 : | edgomez | 1382 | ;/************************************************************************** |
| 2 : | ; * | ||
| 3 : | ; * XVID MPEG-4 VIDEO CODEC | ||
| 4 : | ; * - 3dne Quantization/Dequantization - | ||
| 5 : | ; * | ||
| 6 : | ; * Copyright (C) 2002-2003 Peter Ross <pross@xvid.org> | ||
| 7 : | ; * 2002-2003 Michael Militzer <isibaar@xvid.org> | ||
| 8 : | ; * 2002-2003 Pascal Massimino <skal@planet-d.net> | ||
| 9 : | ; * | ||
| 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 : | syskin | 1713 | ; * $Id: quantize_mpeg_mmx.asm,v 1.6 2006-07-10 08:09:59 syskin Exp $ |
| 25 : | edgomez | 1382 | ; * |
| 26 : | ; *************************************************************************/ | ||
| 27 : | |||
| 28 : | %define SATURATE | ||
| 29 : | |||
| 30 : | BITS 32 | ||
| 31 : | |||
| 32 : | %macro cglobal 1 | ||
| 33 : | %ifdef PREFIX | ||
| 34 : | edgomez | 1535 | %ifdef MARK_FUNCS |
| 35 : | edgomez | 1540 | global _%1:function %1.endfunc-%1 |
| 36 : | %define %1 _%1:function %1.endfunc-%1 | ||
| 37 : | edgomez | 1535 | %else |
| 38 : | global _%1 | ||
| 39 : | %define %1 _%1 | ||
| 40 : | %endif | ||
| 41 : | edgomez | 1382 | %else |
| 42 : | edgomez | 1535 | %ifdef MARK_FUNCS |
| 43 : | edgomez | 1540 | global %1:function %1.endfunc-%1 |
| 44 : | edgomez | 1535 | %else |
| 45 : | global %1 | ||
| 46 : | %endif | ||
| 47 : | edgomez | 1382 | %endif |
| 48 : | %endmacro | ||
| 49 : | |||
| 50 : | %macro cextern 1 | ||
| 51 : | %ifdef PREFIX | ||
| 52 : | extern _%1 | ||
| 53 : | %define %1 _%1 | ||
| 54 : | %else | ||
| 55 : | extern %1 | ||
| 56 : | %endif | ||
| 57 : | %endmacro | ||
| 58 : | |||
| 59 : | ;============================================================================= | ||
| 60 : | ; Local data (Read Only) | ||
| 61 : | ;============================================================================= | ||
| 62 : | |||
| 63 : | %ifdef FORMAT_COFF | ||
| 64 : | edgomez | 1519 | SECTION .rodata |
| 65 : | edgomez | 1382 | %else |
| 66 : | edgomez | 1519 | SECTION .rodata align=16 |
| 67 : | edgomez | 1382 | %endif |
| 68 : | |||
| 69 : | mmx_one: | ||
| 70 : | times 4 dw 1 | ||
| 71 : | |||
| 72 : | ;----------------------------------------------------------------------------- | ||
| 73 : | ; divide by 2Q table | ||
| 74 : | ;----------------------------------------------------------------------------- | ||
| 75 : | |||
| 76 : | ALIGN 16 | ||
| 77 : | mmx_div: | ||
| 78 : | times 4 dw 65535 ; the div by 2 formula will overflow for the case | ||
| 79 : | ; quant=1 but we don't care much because quant=1 | ||
| 80 : | ; is handled by a different piece of code that | ||
| 81 : | ; doesn't use this table. | ||
| 82 : | %assign quant 2 | ||
| 83 : | %rep 30 | ||
| 84 : | times 4 dw (1<<17) / (quant*2) + 1 | ||
| 85 : | %assign quant quant+1 | ||
| 86 : | %endrep | ||
| 87 : | |||
| 88 : | %define VM18P 3 | ||
| 89 : | %define VM18Q 4 | ||
| 90 : | |||
| 91 : | |||
| 92 : | ;----------------------------------------------------------------------------- | ||
| 93 : | ; quantd table | ||
| 94 : | ;----------------------------------------------------------------------------- | ||
| 95 : | |||
| 96 : | quantd: | ||
| 97 : | %assign quant 1 | ||
| 98 : | %rep 31 | ||
| 99 : | times 4 dw ((VM18P*quant) + (VM18Q/2)) / VM18Q | ||
| 100 : | %assign quant quant+1 | ||
| 101 : | %endrep | ||
| 102 : | |||
| 103 : | ;----------------------------------------------------------------------------- | ||
| 104 : | ; multiple by 2Q table | ||
| 105 : | ;----------------------------------------------------------------------------- | ||
| 106 : | |||
| 107 : | mmx_mul_quant: | ||
| 108 : | %assign quant 1 | ||
| 109 : | %rep 31 | ||
| 110 : | times 4 dw quant | ||
| 111 : | %assign quant quant+1 | ||
| 112 : | %endrep | ||
| 113 : | |||
| 114 : | ;----------------------------------------------------------------------------- | ||
| 115 : | ; saturation limits | ||
| 116 : | ;----------------------------------------------------------------------------- | ||
| 117 : | |||
| 118 : | ALIGN 16 | ||
| 119 : | |||
| 120 : | mmx_32767_minus_2047: | ||
| 121 : | times 4 dw (32767-2047) | ||
| 122 : | mmx_32768_minus_2048: | ||
| 123 : | times 4 dw (32768-2048) | ||
| 124 : | mmx_2047: | ||
| 125 : | times 4 dw 2047 | ||
| 126 : | mmx_minus_2048: | ||
| 127 : | times 4 dw (-2048) | ||
| 128 : | zero: | ||
| 129 : | times 4 dw 0 | ||
| 130 : | |||
| 131 : | ;============================================================================= | ||
| 132 : | syskin | 1713 | ; rounding |
| 133 : | ;============================================================================= | ||
| 134 : | |||
| 135 : | mmx_rounding: | ||
| 136 : | dw (1<<13) | ||
| 137 : | dw 0 | ||
| 138 : | dw (1<<13) | ||
| 139 : | dw 0 | ||
| 140 : | |||
| 141 : | ;============================================================================= | ||
| 142 : | edgomez | 1382 | ; Code |
| 143 : | ;============================================================================= | ||
| 144 : | |||
| 145 : | SECTION .text | ||
| 146 : | |||
| 147 : | cglobal quant_mpeg_intra_mmx | ||
| 148 : | cglobal quant_mpeg_inter_mmx | ||
| 149 : | cglobal dequant_mpeg_intra_mmx | ||
| 150 : | cglobal dequant_mpeg_inter_mmx | ||
| 151 : | |||
| 152 : | syskin | 1713 | |
| 153 : | %macro QUANT_MMX 1 | ||
| 154 : | movq mm0, [eax + 16*(%1)] ; data | ||
| 155 : | movq mm2, [ecx + 16*(%1) + 128] ; intra_matrix_rec | ||
| 156 : | movq mm4, [eax + 16*(%1) + 8] ; data | ||
| 157 : | movq mm6, [ecx + 16*(%1) + 128 + 8] ; intra_matrix_rec | ||
| 158 : | |||
| 159 : | movq mm1, mm0 | ||
| 160 : | movq mm5, mm4 | ||
| 161 : | |||
| 162 : | pmullw mm0, mm2 ; low results | ||
| 163 : | pmulhw mm1, mm2 ; high results | ||
| 164 : | pmullw mm4, mm6 ; low results | ||
| 165 : | pmulhw mm5, mm6 ; high results | ||
| 166 : | |||
| 167 : | movq mm2, mm0 | ||
| 168 : | movq mm6, mm4 | ||
| 169 : | |||
| 170 : | punpckhwd mm0, mm1 | ||
| 171 : | punpcklwd mm2, mm1 | ||
| 172 : | punpckhwd mm4, mm5 | ||
| 173 : | punpcklwd mm6, mm5 | ||
| 174 : | |||
| 175 : | paddd mm2, mm7 | ||
| 176 : | paddd mm0, mm7 | ||
| 177 : | paddd mm6, mm7 | ||
| 178 : | paddd mm4, mm7 | ||
| 179 : | |||
| 180 : | psrad mm2, 14 | ||
| 181 : | psrad mm0, 14 | ||
| 182 : | psrad mm6, 14 | ||
| 183 : | psrad mm4, 14 | ||
| 184 : | |||
| 185 : | packssdw mm2, mm0 | ||
| 186 : | packssdw mm6, mm4 | ||
| 187 : | |||
| 188 : | movq [edi + 16*(%1)], mm2 | ||
| 189 : | movq [edi + 16*(%1)+8], mm6 | ||
| 190 : | %endmacro | ||
| 191 : | |||
| 192 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 193 : | ; | ||
| 194 : | ; uint32_t quant_mpeg_intra_mmx(int16_t * coeff, | ||
| 195 : | ; const int16_t const * data, | ||
| 196 : | ; const uint32_t quant, | ||
| 197 : | ; const uint32_t dcscalar, | ||
| 198 : | ; const uint16_t *mpeg_matrices); | ||
| 199 : | ; | ||
| 200 : | ;----------------------------------------------------------------------------- | ||
| 201 : | |||
| 202 : | ALIGN 16 | ||
| 203 : | quant_mpeg_intra_mmx: | ||
| 204 : | |||
| 205 : | push edi | ||
| 206 : | syskin | 1713 | movq mm7, [mmx_rounding] |
| 207 : | edgomez | 1382 | |
| 208 : | syskin | 1713 | mov eax, [esp + 4 + 8] ; data |
| 209 : | mov ecx, [esp + 4 + 20] ; mpeg_quant_matrices | ||
| 210 : | mov edi, [esp + 4 + 4] ; coeff | ||
| 211 : | edgomez | 1382 | |
| 212 : | syskin | 1713 | QUANT_MMX(0) |
| 213 : | QUANT_MMX(1) | ||
| 214 : | QUANT_MMX(2) | ||
| 215 : | QUANT_MMX(3) | ||
| 216 : | QUANT_MMX(4) | ||
| 217 : | QUANT_MMX(5) | ||
| 218 : | QUANT_MMX(6) | ||
| 219 : | QUANT_MMX(7) | ||
| 220 : | edgomez | 1382 | |
| 221 : | syskin | 1713 | ; calculate DC |
| 222 : | movsx eax, word [eax] ; data[0] | ||
| 223 : | mov ecx, [esp + 4 + 16] ; dcscalar | ||
| 224 : | edgomez | 1382 | mov edx, ecx |
| 225 : | shr edx, 1 ; edx = dcscalar /2 | ||
| 226 : | syskin | 1713 | mov edi, edx |
| 227 : | neg edi | ||
| 228 : | |||
| 229 : | edgomez | 1382 | cmp eax, 0 |
| 230 : | syskin | 1713 | cmovg edx, edi |
| 231 : | edgomez | 1382 | sub eax, edx |
| 232 : | syskin | 1713 | |
| 233 : | mov edi, [esp + 4 + 4] ; coeff again | ||
| 234 : | |||
| 235 : | edgomez | 1382 | cdq ; expand eax -> edx:eax |
| 236 : | idiv ecx ; eax = edx:eax / dcscalar | ||
| 237 : | |||
| 238 : | mov [edi], ax ; coeff[0] = ax | ||
| 239 : | |||
| 240 : | pop edi | ||
| 241 : | |||
| 242 : | xor eax, eax ; return(0); | ||
| 243 : | ret | ||
| 244 : | |||
| 245 : | |||
| 246 : | |||
| 247 : | ;----------------------------------------------------------------------------- | ||
| 248 : | ; | ||
| 249 : | ; uint32_t quant_mpeg_inter_mmx(int16_t * coeff, | ||
| 250 : | ; const int16_t const * data, | ||
| 251 : | ; const uint32_t quant, | ||
| 252 : | ; const uint16_t *mpeg_matrices); | ||
| 253 : | ; | ||
| 254 : | ;----------------------------------------------------------------------------- | ||
| 255 : | |||
| 256 : | ALIGN 16 | ||
| 257 : | quant_mpeg_inter_mmx: | ||
| 258 : | |||
| 259 : | push ecx | ||
| 260 : | push esi | ||
| 261 : | push edi | ||
| 262 : | push ebx | ||
| 263 : | |||
| 264 : | mov edi, [esp + 16 + 4] ; coeff | ||
| 265 : | mov esi, [esp + 16 + 8] ; data | ||
| 266 : | mov eax, [esp + 16 + 12] ; quant | ||
| 267 : | mov ebx, [esp + 16 + 16] ; mpeg_quant_matrices | ||
| 268 : | |||
| 269 : | xor ecx, ecx | ||
| 270 : | |||
| 271 : | pxor mm5, mm5 ; sum | ||
| 272 : | |||
| 273 : | cmp al, 1 | ||
| 274 : | jz near .q1loop | ||
| 275 : | |||
| 276 : | cmp al, 2 | ||
| 277 : | jz near .q2loop | ||
| 278 : | |||
| 279 : | movq mm7, [mmx_div + eax * 8 - 8] ; divider | ||
| 280 : | |||
| 281 : | ALIGN 16 | ||
| 282 : | .loop | ||
| 283 : | movq mm0, [esi + 8*ecx] ; mm0 = [1st] | ||
| 284 : | movq mm3, [esi + 8*ecx + 8] ; | ||
| 285 : | pxor mm1, mm1 ; mm1 = 0 | ||
| 286 : | pxor mm4, mm4 ; | ||
| 287 : | pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) | ||
| 288 : | pcmpgtw mm4, mm3 ; | ||
| 289 : | pxor mm0, mm1 ; mm0 = |mm0| | ||
| 290 : | pxor mm3, mm4 ; | ||
| 291 : | psubw mm0, mm1 ; displace | ||
| 292 : | psubw mm3, mm4 ; | ||
| 293 : | psllw mm0, 4 | ||
| 294 : | psllw mm3, 4 | ||
| 295 : | movq mm2, [ebx + 512 + 8*ecx] | ||
| 296 : | psrlw mm2, 1 | ||
| 297 : | paddw mm0, mm2 | ||
| 298 : | movq mm2, [ebx + 768 + ecx*8] | ||
| 299 : | pmulhw mm0, mm2 ; (level<<4 + inter_matrix[i]>>1) / inter_matrix[i] | ||
| 300 : | movq mm2, [ebx + 512 + 8*ecx + 8] | ||
| 301 : | psrlw mm2, 1 | ||
| 302 : | paddw mm3, mm2 | ||
| 303 : | movq mm2, [ebx + 768 + ecx*8 + 8] | ||
| 304 : | pmulhw mm3, mm2 | ||
| 305 : | pmulhw mm0, mm7 ; mm0 = (mm0 / 2Q) >> 16 | ||
| 306 : | pmulhw mm3, mm7 ; | ||
| 307 : | psrlw mm0, 1 ; additional shift by 1 => 16 + 1 = 17 | ||
| 308 : | psrlw mm3, 1 | ||
| 309 : | paddw mm5, mm0 ; sum += mm0 | ||
| 310 : | pxor mm0, mm1 ; mm0 *= sign(mm0) | ||
| 311 : | paddw mm5, mm3 ; | ||
| 312 : | pxor mm3, mm4 ; | ||
| 313 : | psubw mm0, mm1 ; undisplace | ||
| 314 : | psubw mm3, mm4 | ||
| 315 : | movq [edi + 8*ecx], mm0 | ||
| 316 : | movq [edi + 8*ecx + 8], mm3 | ||
| 317 : | |||
| 318 : | add ecx, 2 | ||
| 319 : | cmp ecx, 16 | ||
| 320 : | jnz near .loop | ||
| 321 : | |||
| 322 : | .done | ||
| 323 : | pmaddwd mm5, [mmx_one] | ||
| 324 : | movq mm0, mm5 | ||
| 325 : | psrlq mm5, 32 | ||
| 326 : | paddd mm0, mm5 | ||
| 327 : | movd eax, mm0 ; return sum | ||
| 328 : | |||
| 329 : | pop ebx | ||
| 330 : | pop edi | ||
| 331 : | pop esi | ||
| 332 : | pop ecx | ||
| 333 : | |||
| 334 : | ret | ||
| 335 : | |||
| 336 : | ALIGN 16 | ||
| 337 : | .q1loop | ||
| 338 : | movq mm0, [esi + 8*ecx] ; mm0 = [1st] | ||
| 339 : | movq mm3, [esi + 8*ecx+ 8] | ||
| 340 : | pxor mm1, mm1 ; mm1 = 0 | ||
| 341 : | pxor mm4, mm4 ; | ||
| 342 : | pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) | ||
| 343 : | pcmpgtw mm4, mm3 ; | ||
| 344 : | pxor mm0, mm1 ; mm0 = |mm0| | ||
| 345 : | pxor mm3, mm4 ; | ||
| 346 : | psubw mm0, mm1 ; displace | ||
| 347 : | psubw mm3, mm4 ; | ||
| 348 : | psllw mm0, 4 | ||
| 349 : | psllw mm3, 4 | ||
| 350 : | movq mm2, [ebx + 512 + 8*ecx] | ||
| 351 : | psrlw mm2, 1 | ||
| 352 : | paddw mm0, mm2 | ||
| 353 : | movq mm2, [ebx + 768 + ecx*8] | ||
| 354 : | pmulhw mm0, mm2 ; (level<<4 + inter_matrix[i]>>1) / inter_matrix[i] | ||
| 355 : | movq mm2, [ebx + 512 + 8*ecx + 8] | ||
| 356 : | psrlw mm2, 1 | ||
| 357 : | paddw mm3, mm2 | ||
| 358 : | movq mm2, [ebx + 768 + ecx*8 + 8] | ||
| 359 : | pmulhw mm3, mm2 | ||
| 360 : | psrlw mm0, 1 ; mm0 >>= 1 (/2) | ||
| 361 : | psrlw mm3, 1 ; | ||
| 362 : | paddw mm5, mm0 ; sum += mm0 | ||
| 363 : | pxor mm0, mm1 ; mm0 *= sign(mm0) | ||
| 364 : | paddw mm5, mm3 ; | ||
| 365 : | pxor mm3, mm4 ; | ||
| 366 : | psubw mm0, mm1 ; undisplace | ||
| 367 : | psubw mm3, mm4 | ||
| 368 : | movq [edi + 8*ecx], mm0 | ||
| 369 : | movq [edi + 8*ecx + 8], mm3 | ||
| 370 : | |||
| 371 : | add ecx, 2 | ||
| 372 : | cmp ecx, 16 | ||
| 373 : | jnz near .q1loop | ||
| 374 : | |||
| 375 : | jmp .done | ||
| 376 : | |||
| 377 : | ALIGN 16 | ||
| 378 : | .q2loop | ||
| 379 : | movq mm0, [esi + 8*ecx] ; mm0 = [1st] | ||
| 380 : | movq mm3, [esi + 8*ecx+ 8] | ||
| 381 : | pxor mm1, mm1 ; mm1 = 0 | ||
| 382 : | pxor mm4, mm4 ; | ||
| 383 : | pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) | ||
| 384 : | pcmpgtw mm4, mm3 ; | ||
| 385 : | pxor mm0, mm1 ; mm0 = |mm0| | ||
| 386 : | pxor mm3, mm4 ; | ||
| 387 : | psubw mm0, mm1 ; displace | ||
| 388 : | psubw mm3, mm4 ; | ||
| 389 : | psllw mm0, 4 | ||
| 390 : | psllw mm3, 4 | ||
| 391 : | movq mm2, [ebx + 512 + 8*ecx] | ||
| 392 : | psrlw mm2, 1 | ||
| 393 : | paddw mm0, mm2 | ||
| 394 : | movq mm2, [ebx + 768 + ecx*8] | ||
| 395 : | pmulhw mm0, mm2 ; (level<<4 + inter_matrix[i]>>1) / inter_matrix[i] | ||
| 396 : | movq mm2, [ebx + 512 + 8*ecx + 8] | ||
| 397 : | psrlw mm2, 1 | ||
| 398 : | paddw mm3, mm2 | ||
| 399 : | movq mm2, [ebx + 768 + ecx*8 + 8] | ||
| 400 : | pmulhw mm3, mm2 | ||
| 401 : | psrlw mm0, 2 ; mm0 >>= 1 (/2) | ||
| 402 : | psrlw mm3, 2 ; | ||
| 403 : | paddw mm5, mm0 ; sum += mm0 | ||
| 404 : | pxor mm0, mm1 ; mm0 *= sign(mm0) | ||
| 405 : | paddw mm5, mm3 ; | ||
| 406 : | pxor mm3, mm4 ; | ||
| 407 : | psubw mm0, mm1 ; undisplace | ||
| 408 : | psubw mm3, mm4 | ||
| 409 : | movq [edi + 8*ecx], mm0 | ||
| 410 : | movq [edi + 8*ecx + 8], mm3 | ||
| 411 : | |||
| 412 : | add ecx, 2 | ||
| 413 : | cmp ecx, 16 | ||
| 414 : | jnz near .q2loop | ||
| 415 : | |||
| 416 : | jmp .done | ||
| 417 : | edgomez | 1540 | .endfunc |
| 418 : | edgomez | 1382 | |
| 419 : | |||
| 420 : | ;----------------------------------------------------------------------------- | ||
| 421 : | ; | ||
| 422 : | ; uint32_t dequant_mpeg_intra_mmx(int16_t *data, | ||
| 423 : | ; const int16_t const *coeff, | ||
| 424 : | ; const uint32_t quant, | ||
| 425 : | ; const uint32_t dcscalar, | ||
| 426 : | ; const uint16_t *mpeg_matrices); | ||
| 427 : | ; | ||
| 428 : | ;----------------------------------------------------------------------------- | ||
| 429 : | |||
| 430 : | ; Note: in order to saturate 'easily', we pre-shift the quantifier | ||
| 431 : | ; by 4. Then, the high-word of (coeff[]*matrix[i]*quant) are used to | ||
| 432 : | ; build a saturating mask. It is non-zero only when an overflow occured. | ||
| 433 : | ; We thus avoid packing/unpacking toward double-word. | ||
| 434 : | ; Moreover, we perform the mult (matrix[i]*quant) first, instead of, e.g., | ||
| 435 : | ; (coeff[i]*matrix[i]). This is less prone to overflow if coeff[] are not | ||
| 436 : | ; checked. Input ranges are: coeff in [-127,127], inter_matrix in [1..255],a | ||
| 437 : | ; and quant in [1..31]. | ||
| 438 : | ; | ||
| 439 : | ; The original loop is: | ||
| 440 : | ; | ||
| 441 : | %if 0 | ||
| 442 : | movq mm0, [ecx+8*eax + 8*16] ; mm0 = coeff[i] | ||
| 443 : | pxor mm1, mm1 | ||
| 444 : | pcmpgtw mm1, mm0 | ||
| 445 : | pxor mm0, mm1 ; change sign if negative | ||
| 446 : | psubw mm0, mm1 ; -> mm0 = abs(coeff[i]), mm1 = sign of coeff[i] | ||
| 447 : | |||
| 448 : | movq mm2, mm7 ; mm2 = quant | ||
| 449 : | pmullw mm2, [ebx + 8*eax + 8*16 ] ; matrix[i]*quant. | ||
| 450 : | |||
| 451 : | movq mm6, mm2 | ||
| 452 : | pmulhw mm2, mm0 ; high of coeff*(matrix*quant) (should be 0 if no overflow) | ||
| 453 : | pmullw mm0, mm6 ; low of coeff*(matrix*quant) | ||
| 454 : | |||
| 455 : | pxor mm5, mm5 | ||
| 456 : | pcmpgtw mm2, mm5 ; otherflow? | ||
| 457 : | psrlw mm2, 5 ; =0 if no clamp, 2047 otherwise | ||
| 458 : | psrlw mm0, 5 | ||
| 459 : | paddw mm0, mm1 ; start restoring sign | ||
| 460 : | por mm0, mm2 ; saturate to 2047 if needed | ||
| 461 : | pxor mm0, mm1 ; finish negating back | ||
| 462 : | |||
| 463 : | movq [edx + 8*eax + 8*16], mm0 ; data[i] | ||
| 464 : | add eax, 1 | ||
| 465 : | %endif | ||
| 466 : | |||
| 467 : | ;******************************************************************** | ||
| 468 : | |||
| 469 : | ALIGN 16 | ||
| 470 : | dequant_mpeg_intra_mmx: | ||
| 471 : | |||
| 472 : | push ebx | ||
| 473 : | |||
| 474 : | mov edx, [esp + 4 + 4] ; data | ||
| 475 : | mov ecx, [esp + 4 + 8] ; coeff | ||
| 476 : | mov eax, [esp + 4 + 12] ; quant | ||
| 477 : | mov ebx, [esp + 4 + 20] ; mpeg_quant_matrices | ||
| 478 : | |||
| 479 : | movq mm7, [mmx_mul_quant + eax*8 - 8] | ||
| 480 : | mov eax, -16 ; to keep ALIGNed, we regularly process coeff[0] | ||
| 481 : | psllw mm7, 2 ; << 2. See comment. | ||
| 482 : | pxor mm6, mm6 ; this is a NOP | ||
| 483 : | |||
| 484 : | ALIGN 16 | ||
| 485 : | .loop | ||
| 486 : | movq mm0, [ecx+8*eax + 8*16] ; mm0 = c = coeff[i] | ||
| 487 : | movq mm3, [ecx+8*eax + 8*16 +8]; mm3 = c' = coeff[i+1] | ||
| 488 : | pxor mm1, mm1 | ||
| 489 : | pxor mm4, mm4 | ||
| 490 : | pcmpgtw mm1, mm0 ; mm1 = sgn(c) | ||
| 491 : | movq mm2, mm7 ; mm2 = quant | ||
| 492 : | |||
| 493 : | pcmpgtw mm4, mm3 ; mm4 = sgn(c') | ||
| 494 : | pmullw mm2, [ebx + 8*eax + 8*16 ] ; matrix[i]*quant | ||
| 495 : | |||
| 496 : | pxor mm0, mm1 ; negate if negative | ||
| 497 : | pxor mm3, mm4 ; negate if negative | ||
| 498 : | |||
| 499 : | psubw mm0, mm1 | ||
| 500 : | psubw mm3, mm4 | ||
| 501 : | |||
| 502 : | ; we're short on register, here. Poor pairing... | ||
| 503 : | |||
| 504 : | movq mm5, mm2 | ||
| 505 : | pmullw mm2, mm0 ; low of coeff*(matrix*quant) | ||
| 506 : | |||
| 507 : | pmulhw mm0, mm5 ; high of coeff*(matrix*quant) | ||
| 508 : | movq mm5, mm7 ; mm2 = quant | ||
| 509 : | |||
| 510 : | pmullw mm5, [ebx + 8*eax + 8*16 +8] ; matrix[i+1]*quant | ||
| 511 : | |||
| 512 : | movq mm6, mm5 | ||
| 513 : | add eax,2 ; z-flag will be tested later | ||
| 514 : | |||
| 515 : | pmullw mm6, mm3 ; low of coeff*(matrix*quant) | ||
| 516 : | pmulhw mm3, mm5 ; high of coeff*(matrix*quant) | ||
| 517 : | |||
| 518 : | pcmpgtw mm0, [zero] | ||
| 519 : | paddusw mm2, mm0 | ||
| 520 : | psrlw mm2, 5 | ||
| 521 : | |||
| 522 : | pcmpgtw mm3, [zero] | ||
| 523 : | paddusw mm6, mm3 | ||
| 524 : | psrlw mm6, 5 | ||
| 525 : | |||
| 526 : | pxor mm2, mm1 ; start negating back | ||
| 527 : | pxor mm6, mm4 ; start negating back | ||
| 528 : | |||
| 529 : | psubusw mm1, mm0 | ||
| 530 : | psubusw mm4, mm3 | ||
| 531 : | |||
| 532 : | psubw mm2, mm1 ; finish negating back | ||
| 533 : | psubw mm6, mm4 ; finish negating back | ||
| 534 : | |||
| 535 : | movq [edx + 8*eax + 8*16 -2*8 ], mm2 ; data[i] | ||
| 536 : | movq [edx + 8*eax + 8*16 -2*8 +8], mm6 ; data[i+1] | ||
| 537 : | |||
| 538 : | jnz near .loop | ||
| 539 : | |||
| 540 : | ; deal with DC | ||
| 541 : | movd mm0, [ecx] | ||
| 542 : | pmullw mm0, [esp + 4 + 16] ; dcscalar | ||
| 543 : | movq mm2, [mmx_32767_minus_2047] | ||
| 544 : | paddsw mm0, mm2 | ||
| 545 : | psubsw mm0, mm2 | ||
| 546 : | movq mm2, [mmx_32768_minus_2048] | ||
| 547 : | psubsw mm0, mm2 | ||
| 548 : | paddsw mm0, mm2 | ||
| 549 : | movd eax, mm0 | ||
| 550 : | mov [edx], ax | ||
| 551 : | |||
| 552 : | xor eax, eax | ||
| 553 : | |||
| 554 : | pop ebx | ||
| 555 : | |||
| 556 : | ret | ||
| 557 : | edgomez | 1540 | .endfunc |
| 558 : | edgomez | 1382 | |
| 559 : | ;----------------------------------------------------------------------------- | ||
| 560 : | ; | ||
| 561 : | ; uint32_t dequant_mpeg_inter_mmx(int16_t * data, | ||
| 562 : | ; const int16_t * const coeff, | ||
| 563 : | ; const uint32_t quant, | ||
| 564 : | ; const uint16_t *mpeg_matrices); | ||
| 565 : | ; | ||
| 566 : | ;----------------------------------------------------------------------------- | ||
| 567 : | |||
| 568 : | ; Note: We use (2*c + sgn(c) - sgn(-c)) as multiplier | ||
| 569 : | ; so we handle the 3 cases: c<0, c==0, and c>0 in one shot. | ||
| 570 : | ; sgn(x) is the result of 'pcmpgtw 0,x': 0 if x>=0, -1 if x<0. | ||
| 571 : | ; It's mixed with the extraction of the absolute value. | ||
| 572 : | |||
| 573 : | ALIGN 16 | ||
| 574 : | dequant_mpeg_inter_mmx: | ||
| 575 : | |||
| 576 : | push ebx | ||
| 577 : | |||
| 578 : | mov edx, [esp + 4 + 4] ; data | ||
| 579 : | mov ecx, [esp + 4 + 8] ; coeff | ||
| 580 : | mov eax, [esp + 4 + 12] ; quant | ||
| 581 : | mov ebx, [esp + 4 + 16] ; mpeg_quant_matrices | ||
| 582 : | |||
| 583 : | movq mm7, [mmx_mul_quant + eax*8 - 8] | ||
| 584 : | mov eax, -16 | ||
| 585 : | paddw mm7, mm7 ; << 1 | ||
| 586 : | pxor mm6, mm6 ; mismatch sum | ||
| 587 : | |||
| 588 : | ALIGN 16 | ||
| 589 : | .loop | ||
| 590 : | movq mm0, [ecx+8*eax + 8*16 ] ; mm0 = coeff[i] | ||
| 591 : | movq mm2, [ecx+8*eax + 8*16 +8] ; mm2 = coeff[i+1] | ||
| 592 : | add eax, 2 | ||
| 593 : | |||
| 594 : | pxor mm1, mm1 | ||
| 595 : | pxor mm3, mm3 | ||
| 596 : | pcmpgtw mm1, mm0 ; mm1 = sgn(c) (preserved) | ||
| 597 : | pcmpgtw mm3, mm2 ; mm3 = sgn(c') (preserved) | ||
| 598 : | paddsw mm0, mm1 ; c += sgn(c) | ||
| 599 : | paddsw mm2, mm3 ; c += sgn(c') | ||
| 600 : | paddw mm0, mm0 ; c *= 2 | ||
| 601 : | paddw mm2, mm2 ; c'*= 2 | ||
| 602 : | |||
| 603 : | pxor mm4, mm4 | ||
| 604 : | pxor mm5, mm5 | ||
| 605 : | psubw mm4, mm0 ; -c | ||
| 606 : | psubw mm5, mm2 ; -c' | ||
| 607 : | psraw mm4, 16 ; mm4 = sgn(-c) | ||
| 608 : | psraw mm5, 16 ; mm5 = sgn(-c') | ||
| 609 : | psubsw mm0, mm4 ; c -= sgn(-c) | ||
| 610 : | psubsw mm2, mm5 ; c' -= sgn(-c') | ||
| 611 : | pxor mm0, mm1 ; finish changing sign if needed | ||
| 612 : | pxor mm2, mm3 ; finish changing sign if needed | ||
| 613 : | |||
| 614 : | ; we're short on register, here. Poor pairing... | ||
| 615 : | |||
| 616 : | movq mm4, mm7 ; (matrix*quant) | ||
| 617 : | pmullw mm4, [ebx + 512 + 8*eax + 8*16 -2*8] | ||
| 618 : | movq mm5, mm4 | ||
| 619 : | pmulhw mm5, mm0 ; high of c*(matrix*quant) | ||
| 620 : | pmullw mm0, mm4 ; low of c*(matrix*quant) | ||
| 621 : | |||
| 622 : | movq mm4, mm7 ; (matrix*quant) | ||
| 623 : | pmullw mm4, [ebx + 512 + 8*eax + 8*16 -2*8 + 8] | ||
| 624 : | |||
| 625 : | pcmpgtw mm5, [zero] | ||
| 626 : | paddusw mm0, mm5 | ||
| 627 : | psrlw mm0, 5 | ||
| 628 : | pxor mm0, mm1 ; start restoring sign | ||
| 629 : | psubusw mm1, mm5 | ||
| 630 : | |||
| 631 : | movq mm5, mm4 | ||
| 632 : | pmulhw mm5, mm2 ; high of c*(matrix*quant) | ||
| 633 : | pmullw mm2, mm4 ; low of c*(matrix*quant) | ||
| 634 : | psubw mm0, mm1 ; finish restoring sign | ||
| 635 : | |||
| 636 : | pcmpgtw mm5, [zero] | ||
| 637 : | paddusw mm2, mm5 | ||
| 638 : | psrlw mm2, 5 | ||
| 639 : | pxor mm2, mm3 ; start restoring sign | ||
| 640 : | psubusw mm3, mm5 | ||
| 641 : | psubw mm2, mm3 ; finish restoring sign | ||
| 642 : | |||
| 643 : | pxor mm6, mm0 ; mismatch control | ||
| 644 : | movq [edx + 8*eax + 8*16 -2*8 ], mm0 ; data[i] | ||
| 645 : | pxor mm6, mm2 ; mismatch control | ||
| 646 : | movq [edx + 8*eax + 8*16 -2*8 +8], mm2 ; data[i+1] | ||
| 647 : | |||
| 648 : | jnz near .loop | ||
| 649 : | |||
| 650 : | ; mismatch control | ||
| 651 : | |||
| 652 : | movq mm0, mm6 | ||
| 653 : | psrlq mm0, 48 | ||
| 654 : | movq mm1, mm6 | ||
| 655 : | movq mm2, mm6 | ||
| 656 : | psrlq mm1, 32 | ||
| 657 : | pxor mm6, mm0 | ||
| 658 : | psrlq mm2, 16 | ||
| 659 : | pxor mm6, mm1 | ||
| 660 : | pxor mm6, mm2 | ||
| 661 : | movd eax, mm6 | ||
| 662 : | and eax, 1 | ||
| 663 : | xor eax, 1 | ||
| 664 : | xor word [edx + 2*63], ax | ||
| 665 : | |||
| 666 : | xor eax, eax | ||
| 667 : | |||
| 668 : | pop ebx | ||
| 669 : | |||
| 670 : | ret | ||
| 671 : | edgomez | 1540 | .endfunc |
| 672 : |
| No admin address has been configured | ViewVC Help |
| Powered by ViewVC 1.0.4 |