Parent Directory
| 
Revision Log
Revision 1192 - (view) (download)
| 1 : | edgomez | 1176 | ;/************************************************************************** |
| 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 : | edgomez | 1192 | ; * $Id: quantize_mpeg_mmx.asm,v 1.1.2.3 2003-10-28 22:23:03 edgomez Exp $ |
| 25 : | edgomez | 1176 | ; * |
| 26 : | ; *************************************************************************/ | ||
| 27 : | edgomez | 1174 | |
| 28 : | %define SATURATE | ||
| 29 : | |||
| 30 : | edgomez | 1192 | BITS 32 |
| 31 : | edgomez | 1174 | |
| 32 : | edgomez | 1192 | %macro cglobal 1 |
| 33 : | edgomez | 1174 | %ifdef PREFIX |
| 34 : | edgomez | 1192 | global _%1 |
| 35 : | edgomez | 1174 | %define %1 _%1 |
| 36 : | %else | ||
| 37 : | global %1 | ||
| 38 : | %endif | ||
| 39 : | %endmacro | ||
| 40 : | |||
| 41 : | edgomez | 1192 | %macro cextern 1 |
| 42 : | edgomez | 1174 | %ifdef PREFIX |
| 43 : | edgomez | 1192 | extern _%1 |
| 44 : | edgomez | 1174 | %define %1 _%1 |
| 45 : | %else | ||
| 46 : | extern %1 | ||
| 47 : | %endif | ||
| 48 : | %endmacro | ||
| 49 : | |||
| 50 : | edgomez | 1192 | ;============================================================================= |
| 51 : | ; Local data (Read Only) | ||
| 52 : | ;============================================================================= | ||
| 53 : | edgomez | 1174 | |
| 54 : | edgomez | 1192 | SECTION .rodata |
| 55 : | edgomez | 1176 | |
| 56 : | mmx_one: | ||
| 57 : | times 4 dw 1 | ||
| 58 : | |||
| 59 : | edgomez | 1192 | ;----------------------------------------------------------------------------- |
| 60 : | ; divide by 2Q table | ||
| 61 : | ;----------------------------------------------------------------------------- | ||
| 62 : | edgomez | 1174 | |
| 63 : | edgomez | 1192 | ALIGN 16 |
| 64 : | edgomez | 1176 | mmx_div: |
| 65 : | edgomez | 1192 | times 4 dw 65535 ; the div by 2 formula will overflow for the case |
| 66 : | ; quant=1 but we don't care much because quant=1 | ||
| 67 : | ; is handled by a different piece of code that | ||
| 68 : | ; doesn't use this table. | ||
| 69 : | %assign quant 2 | ||
| 70 : | %rep 30 | ||
| 71 : | edgomez | 1176 | times 4 dw (1<<17) / (quant*2) + 1 |
| 72 : | %assign quant quant+1 | ||
| 73 : | %endrep | ||
| 74 : | edgomez | 1174 | |
| 75 : | edgomez | 1192 | ;----------------------------------------------------------------------------- |
| 76 : | ; intra matrix | ||
| 77 : | ;----------------------------------------------------------------------------- | ||
| 78 : | edgomez | 1174 | |
| 79 : | cextern intra_matrix | ||
| 80 : | cextern intra_matrix_fix | ||
| 81 : | |||
| 82 : | edgomez | 1192 | ;----------------------------------------------------------------------------- |
| 83 : | edgomez | 1174 | ; inter matrix |
| 84 : | edgomez | 1192 | ;----------------------------------------------------------------------------- |
| 85 : | edgomez | 1174 | |
| 86 : | cextern inter_matrix | ||
| 87 : | cextern inter_matrix_fix | ||
| 88 : | |||
| 89 : | |||
| 90 : | %define VM18P 3 | ||
| 91 : | %define VM18Q 4 | ||
| 92 : | |||
| 93 : | |||
| 94 : | edgomez | 1192 | ;----------------------------------------------------------------------------- |
| 95 : | ; quantd table | ||
| 96 : | ;----------------------------------------------------------------------------- | ||
| 97 : | edgomez | 1174 | |
| 98 : | edgomez | 1176 | quantd: |
| 99 : | %assign quant 1 | ||
| 100 : | %rep 31 | ||
| 101 : | times 4 dw ((VM18P*quant) + (VM18Q/2)) / VM18Q | ||
| 102 : | %assign quant quant+1 | ||
| 103 : | %endrep | ||
| 104 : | edgomez | 1174 | |
| 105 : | edgomez | 1192 | ;----------------------------------------------------------------------------- |
| 106 : | edgomez | 1174 | ; multiple by 2Q table |
| 107 : | edgomez | 1192 | ;----------------------------------------------------------------------------- |
| 108 : | edgomez | 1174 | |
| 109 : | edgomez | 1176 | mmx_mul_quant: |
| 110 : | %assign quant 1 | ||
| 111 : | %rep 31 | ||
| 112 : | times 4 dw quant | ||
| 113 : | %assign quant quant+1 | ||
| 114 : | %endrep | ||
| 115 : | edgomez | 1174 | |
| 116 : | edgomez | 1192 | ;----------------------------------------------------------------------------- |
| 117 : | ; saturation limits | ||
| 118 : | ;----------------------------------------------------------------------------- | ||
| 119 : | edgomez | 1174 | |
| 120 : | edgomez | 1192 | ALIGN 16 |
| 121 : | edgomez | 1174 | |
| 122 : | edgomez | 1176 | mmx_32767_minus_2047: |
| 123 : | times 4 dw (32767-2047) | ||
| 124 : | mmx_32768_minus_2048: | ||
| 125 : | times 4 dw (32768-2048) | ||
| 126 : | mmx_2047: | ||
| 127 : | times 4 dw 2047 | ||
| 128 : | mmx_minus_2048: | ||
| 129 : | times 4 dw (-2048) | ||
| 130 : | zero: | ||
| 131 : | times 4 dw 0 | ||
| 132 : | edgomez | 1174 | |
| 133 : | edgomez | 1192 | ;============================================================================= |
| 134 : | ; Code | ||
| 135 : | ;============================================================================= | ||
| 136 : | edgomez | 1176 | |
| 137 : | edgomez | 1192 | SECTION .text |
| 138 : | edgomez | 1174 | |
| 139 : | edgomez | 1192 | cglobal quant_mpeg_intra_mmx |
| 140 : | cglobal quant_mpeg_inter_mmx | ||
| 141 : | cglobal dequant_mpeg_intra_mmx | ||
| 142 : | cglobal dequant_mpeg_inter_mmx | ||
| 143 : | |||
| 144 : | ;----------------------------------------------------------------------------- | ||
| 145 : | edgomez | 1174 | ; |
| 146 : | edgomez | 1192 | ; uint32_t quant_mpeg_intra_mmx(int16_t * coeff, |
| 147 : | edgomez | 1176 | ; const int16_t const * data, |
| 148 : | ; const uint32_t quant, | ||
| 149 : | ; const uint32_t dcscalar); | ||
| 150 : | edgomez | 1174 | ; |
| 151 : | edgomez | 1192 | ;----------------------------------------------------------------------------- |
| 152 : | edgomez | 1174 | |
| 153 : | edgomez | 1192 | ALIGN 16 |
| 154 : | edgomez | 1174 | quant_mpeg_intra_mmx: |
| 155 : | |||
| 156 : | edgomez | 1192 | push ecx |
| 157 : | push esi | ||
| 158 : | push edi | ||
| 159 : | edgomez | 1174 | |
| 160 : | edgomez | 1192 | mov edi, [esp + 12 + 4] ; coeff |
| 161 : | mov esi, [esp + 12 + 8] ; data | ||
| 162 : | mov eax, [esp + 12 + 12] ; quant | ||
| 163 : | edgomez | 1174 | |
| 164 : | edgomez | 1192 | movq mm5, [quantd + eax * 8 - 8] ; quantd -> mm5 |
| 165 : | edgomez | 1174 | |
| 166 : | edgomez | 1192 | xor ecx, ecx |
| 167 : | cmp al, 1 | ||
| 168 : | jz near .q1loop | ||
| 169 : | edgomez | 1174 | |
| 170 : | edgomez | 1192 | cmp al, 2 |
| 171 : | jz near .q2loop | ||
| 172 : | edgomez | 1174 | |
| 173 : | edgomez | 1192 | movq mm7, [mmx_div + eax * 8 - 8] ; multipliers[quant] -> mm7 |
| 174 : | |||
| 175 : | ALIGN 16 | ||
| 176 : | edgomez | 1174 | .loop |
| 177 : | edgomez | 1192 | movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
| 178 : | movq mm3, [esi + 8*ecx + 8] ; | ||
| 179 : | pxor mm1, mm1 ; mm1 = 0 | ||
| 180 : | pxor mm4, mm4 | ||
| 181 : | pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) | ||
| 182 : | pcmpgtw mm4, mm3 | ||
| 183 : | pxor mm0, mm1 ; mm0 = |mm0| | ||
| 184 : | pxor mm3, mm4 ; | ||
| 185 : | psubw mm0, mm1 ; displace | ||
| 186 : | psubw mm3, mm4 ; | ||
| 187 : | psllw mm0, 4 ; level << 4 | ||
| 188 : | psllw mm3, 4 | ||
| 189 : | movq mm2, [intra_matrix + 8*ecx] | ||
| 190 : | psrlw mm2, 1 ; intra_matrix[i]>>1 | ||
| 191 : | paddw mm0, mm2 | ||
| 192 : | movq mm2, [intra_matrix_fix + ecx*8] | ||
| 193 : | pmulhw mm0, mm2 ; (level<<4 + intra_matrix[i]>>1) / intra_matrix[i] | ||
| 194 : | movq mm2, [intra_matrix + 8*ecx + 8] | ||
| 195 : | psrlw mm2, 1 | ||
| 196 : | paddw mm3, mm2 | ||
| 197 : | movq mm2, [intra_matrix_fix + ecx*8 + 8] | ||
| 198 : | pmulhw mm3, mm2 | ||
| 199 : | paddw mm0, mm5 ; + quantd | ||
| 200 : | paddw mm3, mm5 | ||
| 201 : | pmulhw mm0, mm7 ; mm0 = (mm0 / 2Q) >> 16 | ||
| 202 : | pmulhw mm3, mm7 ; | ||
| 203 : | psrlw mm0, 1 ; additional shift by 1 => 16 + 1 = 17 | ||
| 204 : | psrlw mm3, 1 | ||
| 205 : | pxor mm0, mm1 ; mm0 *= sign(mm0) | ||
| 206 : | pxor mm3, mm4 ; | ||
| 207 : | psubw mm0, mm1 ; undisplace | ||
| 208 : | psubw mm3, mm4 ; | ||
| 209 : | edgomez | 1174 | |
| 210 : | edgomez | 1192 | movq [edi + 8*ecx], mm0 |
| 211 : | movq [edi + 8*ecx + 8], mm3 | ||
| 212 : | edgomez | 1174 | |
| 213 : | edgomez | 1192 | add ecx,2 |
| 214 : | cmp ecx,16 | ||
| 215 : | jnz near .loop | ||
| 216 : | edgomez | 1174 | |
| 217 : | edgomez | 1192 | .done |
| 218 : | ; caclulate data[0] // (int32_t)dcscalar) | ||
| 219 : | mov ecx, [esp + 12 + 16] ; dcscalar | ||
| 220 : | mov edx, ecx | ||
| 221 : | movsx eax, word [esi] ; data[0] | ||
| 222 : | shr edx, 1 ; edx = dcscalar /2 | ||
| 223 : | cmp eax, 0 | ||
| 224 : | jg .gtzero | ||
| 225 : | edgomez | 1174 | |
| 226 : | edgomez | 1192 | sub eax, edx |
| 227 : | jmp short .mul | ||
| 228 : | edgomez | 1174 | .gtzero |
| 229 : | edgomez | 1192 | add eax, edx |
| 230 : | edgomez | 1174 | .mul |
| 231 : | edgomez | 1192 | cdq ; expand eax -> edx:eax |
| 232 : | idiv ecx ; eax = edx:eax / dcscalar | ||
| 233 : | edgomez | 1174 | |
| 234 : | edgomez | 1192 | mov [edi], ax ; coeff[0] = ax |
| 235 : | edgomez | 1174 | |
| 236 : | edgomez | 1192 | pop edi |
| 237 : | pop esi | ||
| 238 : | pop ecx | ||
| 239 : | edgomez | 1174 | |
| 240 : | edgomez | 1192 | xor eax, eax ; return(0); |
| 241 : | ret | ||
| 242 : | |||
| 243 : | ALIGN 16 | ||
| 244 : | edgomez | 1174 | .q1loop |
| 245 : | edgomez | 1192 | movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
| 246 : | movq mm3, [esi + 8*ecx + 8] ; | ||
| 247 : | pxor mm1, mm1 ; mm1 = 0 | ||
| 248 : | pxor mm4, mm4 ; | ||
| 249 : | pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) | ||
| 250 : | pcmpgtw mm4, mm3 ; | ||
| 251 : | pxor mm0, mm1 ; mm0 = |mm0| | ||
| 252 : | pxor mm3, mm4 ; | ||
| 253 : | psubw mm0, mm1 ; displace | ||
| 254 : | psubw mm3, mm4 ; | ||
| 255 : | psllw mm0, 4 | ||
| 256 : | psllw mm3, 4 | ||
| 257 : | movq mm2, [intra_matrix + 8*ecx] | ||
| 258 : | psrlw mm2, 1 | ||
| 259 : | paddw mm0, mm2 | ||
| 260 : | movq mm2, [intra_matrix_fix + ecx*8] | ||
| 261 : | pmulhw mm0, mm2 ; (level<<4 + intra_matrix[i]>>1) / intra_matrix[i] | ||
| 262 : | movq mm2, [intra_matrix + 8*ecx + 8] | ||
| 263 : | psrlw mm2, 1 | ||
| 264 : | paddw mm3, mm2 | ||
| 265 : | movq mm2, [intra_matrix_fix + ecx*8 + 8] | ||
| 266 : | pmulhw mm3, mm2 | ||
| 267 : | paddw mm0, mm5 | ||
| 268 : | paddw mm3, mm5 | ||
| 269 : | psrlw mm0, 1 ; mm0 >>= 1 (/2) | ||
| 270 : | psrlw mm3, 1 ; | ||
| 271 : | pxor mm0, mm1 ; mm0 *= sign(mm0) | ||
| 272 : | pxor mm3, mm4 ; | ||
| 273 : | psubw mm0, mm1 ; undisplace | ||
| 274 : | psubw mm3, mm4 ; | ||
| 275 : | movq [edi + 8*ecx], mm0 | ||
| 276 : | movq [edi + 8*ecx + 8], mm3 | ||
| 277 : | edgomez | 1174 | |
| 278 : | edgomez | 1192 | add ecx, 2 |
| 279 : | cmp ecx, 16 | ||
| 280 : | jnz near .q1loop | ||
| 281 : | jmp near .done | ||
| 282 : | edgomez | 1174 | |
| 283 : | |||
| 284 : | edgomez | 1192 | ALIGN 16 |
| 285 : | edgomez | 1174 | .q2loop |
| 286 : | edgomez | 1192 | movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
| 287 : | movq mm3, [esi + 8*ecx + 8] ; | ||
| 288 : | pxor mm1, mm1 ; mm1 = 0 | ||
| 289 : | pxor mm4, mm4 ; | ||
| 290 : | pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) | ||
| 291 : | pcmpgtw mm4, mm3 ; | ||
| 292 : | pxor mm0, mm1 ; mm0 = |mm0| | ||
| 293 : | pxor mm3, mm4 ; | ||
| 294 : | psubw mm0, mm1 ; displace | ||
| 295 : | psubw mm3, mm4 ; | ||
| 296 : | psllw mm0, 4 | ||
| 297 : | psllw mm3, 4 | ||
| 298 : | movq mm2, [intra_matrix + 8*ecx] | ||
| 299 : | psrlw mm2, 1 | ||
| 300 : | paddw mm0, mm2 | ||
| 301 : | movq mm2, [intra_matrix_fix + ecx*8] | ||
| 302 : | pmulhw mm0, mm2 ; (level<<4 + intra_matrix[i]>>1) / intra_matrix[i] | ||
| 303 : | movq mm2, [intra_matrix + 8*ecx + 8] | ||
| 304 : | psrlw mm2, 1 | ||
| 305 : | paddw mm3, mm2 | ||
| 306 : | movq mm2, [intra_matrix_fix + ecx*8 + 8] | ||
| 307 : | pmulhw mm3, mm2 | ||
| 308 : | paddw mm0, mm5 | ||
| 309 : | paddw mm3, mm5 | ||
| 310 : | psrlw mm0, 2 ; mm0 >>= 1 (/4) | ||
| 311 : | psrlw mm3, 2 ; | ||
| 312 : | pxor mm0, mm1 ; mm0 *= sign(mm0) | ||
| 313 : | pxor mm3, mm4 ; | ||
| 314 : | psubw mm0, mm1 ; undisplace | ||
| 315 : | psubw mm3, mm4 ; | ||
| 316 : | movq [edi + 8*ecx], mm0 | ||
| 317 : | movq [edi + 8*ecx + 8], mm3 | ||
| 318 : | edgomez | 1174 | |
| 319 : | edgomez | 1192 | add ecx,2 |
| 320 : | cmp ecx,16 | ||
| 321 : | jnz near .q2loop | ||
| 322 : | jmp near .done | ||
| 323 : | edgomez | 1174 | |
| 324 : | |||
| 325 : | edgomez | 1192 | ;----------------------------------------------------------------------------- |
| 326 : | edgomez | 1174 | ; |
| 327 : | edgomez | 1176 | ; uint32_t quant_mpeg_inter_mmx(int16_t * coeff, |
| 328 : | ; const int16_t const * data, | ||
| 329 : | ; const uint32_t quant); | ||
| 330 : | edgomez | 1174 | ; |
| 331 : | edgomez | 1192 | ;----------------------------------------------------------------------------- |
| 332 : | edgomez | 1174 | |
| 333 : | edgomez | 1192 | ALIGN 16 |
| 334 : | edgomez | 1174 | quant_mpeg_inter_mmx: |
| 335 : | |||
| 336 : | edgomez | 1192 | push ecx |
| 337 : | push esi | ||
| 338 : | push edi | ||
| 339 : | edgomez | 1174 | |
| 340 : | edgomez | 1192 | mov edi, [esp + 12 + 4] ; coeff |
| 341 : | mov esi, [esp + 12 + 8] ; data | ||
| 342 : | mov eax, [esp + 12 + 12] ; quant | ||
| 343 : | edgomez | 1174 | |
| 344 : | edgomez | 1192 | xor ecx, ecx |
| 345 : | edgomez | 1174 | |
| 346 : | edgomez | 1192 | pxor mm5, mm5 ; sum |
| 347 : | edgomez | 1174 | |
| 348 : | edgomez | 1192 | cmp al, 1 |
| 349 : | jz near .q1loop | ||
| 350 : | edgomez | 1174 | |
| 351 : | edgomez | 1192 | cmp al, 2 |
| 352 : | jz near .q2loop | ||
| 353 : | edgomez | 1174 | |
| 354 : | edgomez | 1192 | movq mm7, [mmx_div + eax * 8 - 8] ; divider |
| 355 : | edgomez | 1174 | |
| 356 : | edgomez | 1192 | ALIGN 16 |
| 357 : | edgomez | 1174 | .loop |
| 358 : | edgomez | 1192 | movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
| 359 : | movq mm3, [esi + 8*ecx + 8] ; | ||
| 360 : | pxor mm1, mm1 ; mm1 = 0 | ||
| 361 : | pxor mm4, mm4 ; | ||
| 362 : | pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) | ||
| 363 : | pcmpgtw mm4, mm3 ; | ||
| 364 : | pxor mm0, mm1 ; mm0 = |mm0| | ||
| 365 : | pxor mm3, mm4 ; | ||
| 366 : | psubw mm0, mm1 ; displace | ||
| 367 : | psubw mm3, mm4 ; | ||
| 368 : | psllw mm0, 4 | ||
| 369 : | psllw mm3, 4 | ||
| 370 : | movq mm2, [inter_matrix + 8*ecx] | ||
| 371 : | psrlw mm2, 1 | ||
| 372 : | paddw mm0, mm2 | ||
| 373 : | movq mm2, [inter_matrix_fix + ecx*8] | ||
| 374 : | pmulhw mm0, mm2 ; (level<<4 + inter_matrix[i]>>1) / inter_matrix[i] | ||
| 375 : | movq mm2, [inter_matrix + 8*ecx + 8] | ||
| 376 : | psrlw mm2, 1 | ||
| 377 : | paddw mm3, mm2 | ||
| 378 : | movq mm2, [inter_matrix_fix + ecx*8 + 8] | ||
| 379 : | pmulhw mm3, mm2 | ||
| 380 : | pmulhw mm0, mm7 ; mm0 = (mm0 / 2Q) >> 16 | ||
| 381 : | pmulhw mm3, mm7 ; | ||
| 382 : | psrlw mm0, 1 ; additional shift by 1 => 16 + 1 = 17 | ||
| 383 : | psrlw mm3, 1 | ||
| 384 : | paddw mm5, mm0 ; sum += mm0 | ||
| 385 : | pxor mm0, mm1 ; mm0 *= sign(mm0) | ||
| 386 : | paddw mm5, mm3 ; | ||
| 387 : | pxor mm3, mm4 ; | ||
| 388 : | psubw mm0, mm1 ; undisplace | ||
| 389 : | psubw mm3, mm4 | ||
| 390 : | movq [edi + 8*ecx], mm0 | ||
| 391 : | movq [edi + 8*ecx + 8], mm3 | ||
| 392 : | edgomez | 1174 | |
| 393 : | edgomez | 1192 | add ecx, 2 |
| 394 : | cmp ecx, 16 | ||
| 395 : | jnz near .loop | ||
| 396 : | edgomez | 1174 | |
| 397 : | .done | ||
| 398 : | edgomez | 1192 | pmaddwd mm5, [mmx_one] |
| 399 : | movq mm0, mm5 | ||
| 400 : | psrlq mm5, 32 | ||
| 401 : | paddd mm0, mm5 | ||
| 402 : | movd eax, mm0 ; return sum | ||
| 403 : | edgomez | 1174 | |
| 404 : | edgomez | 1192 | pop edi |
| 405 : | pop esi | ||
| 406 : | pop ecx | ||
| 407 : | edgomez | 1174 | |
| 408 : | edgomez | 1192 | ret |
| 409 : | edgomez | 1174 | |
| 410 : | edgomez | 1192 | ALIGN 16 |
| 411 : | edgomez | 1174 | .q1loop |
| 412 : | edgomez | 1192 | movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
| 413 : | movq mm3, [esi + 8*ecx+ 8] | ||
| 414 : | pxor mm1, mm1 ; mm1 = 0 | ||
| 415 : | pxor mm4, mm4 ; | ||
| 416 : | pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) | ||
| 417 : | pcmpgtw mm4, mm3 ; | ||
| 418 : | pxor mm0, mm1 ; mm0 = |mm0| | ||
| 419 : | pxor mm3, mm4 ; | ||
| 420 : | psubw mm0, mm1 ; displace | ||
| 421 : | psubw mm3, mm4 ; | ||
| 422 : | psllw mm0, 4 | ||
| 423 : | psllw mm3, 4 | ||
| 424 : | movq mm2, [inter_matrix + 8*ecx] | ||
| 425 : | psrlw mm2, 1 | ||
| 426 : | paddw mm0, mm2 | ||
| 427 : | movq mm2, [inter_matrix_fix + ecx*8] | ||
| 428 : | pmulhw mm0, mm2 ; (level<<4 + inter_matrix[i]>>1) / inter_matrix[i] | ||
| 429 : | movq mm2, [inter_matrix + 8*ecx + 8] | ||
| 430 : | psrlw mm2, 1 | ||
| 431 : | paddw mm3, mm2 | ||
| 432 : | movq mm2, [inter_matrix_fix + ecx*8 + 8] | ||
| 433 : | pmulhw mm3, mm2 | ||
| 434 : | psrlw mm0, 1 ; mm0 >>= 1 (/2) | ||
| 435 : | psrlw mm3, 1 ; | ||
| 436 : | paddw mm5, mm0 ; sum += mm0 | ||
| 437 : | pxor mm0, mm1 ; mm0 *= sign(mm0) | ||
| 438 : | paddw mm5, mm3 ; | ||
| 439 : | pxor mm3, mm4 ; | ||
| 440 : | psubw mm0, mm1 ; undisplace | ||
| 441 : | psubw mm3, mm4 | ||
| 442 : | movq [edi + 8*ecx], mm0 | ||
| 443 : | movq [edi + 8*ecx + 8], mm3 | ||
| 444 : | edgomez | 1174 | |
| 445 : | edgomez | 1192 | add ecx, 2 |
| 446 : | cmp ecx, 16 | ||
| 447 : | jnz near .q1loop | ||
| 448 : | edgomez | 1174 | |
| 449 : | edgomez | 1192 | jmp .done |
| 450 : | edgomez | 1174 | |
| 451 : | |||
| 452 : | edgomez | 1192 | ALIGN 16 |
| 453 : | edgomez | 1174 | .q2loop |
| 454 : | edgomez | 1192 | movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
| 455 : | movq mm3, [esi + 8*ecx+ 8] | ||
| 456 : | pxor mm1, mm1 ; mm1 = 0 | ||
| 457 : | pxor mm4, mm4 ; | ||
| 458 : | pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) | ||
| 459 : | pcmpgtw mm4, mm3 ; | ||
| 460 : | pxor mm0, mm1 ; mm0 = |mm0| | ||
| 461 : | pxor mm3, mm4 ; | ||
| 462 : | psubw mm0, mm1 ; displace | ||
| 463 : | psubw mm3, mm4 ; | ||
| 464 : | psllw mm0, 4 | ||
| 465 : | psllw mm3, 4 | ||
| 466 : | movq mm2, [inter_matrix + 8*ecx] | ||
| 467 : | psrlw mm2, 1 | ||
| 468 : | paddw mm0, mm2 | ||
| 469 : | movq mm2, [inter_matrix_fix + ecx*8] | ||
| 470 : | pmulhw mm0, mm2 ; (level<<4 + inter_matrix[i]>>1) / inter_matrix[i] | ||
| 471 : | movq mm2, [inter_matrix + 8*ecx + 8] | ||
| 472 : | psrlw mm2, 1 | ||
| 473 : | paddw mm3, mm2 | ||
| 474 : | movq mm2, [inter_matrix_fix + ecx*8 + 8] | ||
| 475 : | pmulhw mm3, mm2 | ||
| 476 : | psrlw mm0, 2 ; mm0 >>= 1 (/2) | ||
| 477 : | psrlw mm3, 2 ; | ||
| 478 : | paddw mm5, mm0 ; sum += mm0 | ||
| 479 : | pxor mm0, mm1 ; mm0 *= sign(mm0) | ||
| 480 : | paddw mm5, mm3 ; | ||
| 481 : | pxor mm3, mm4 ; | ||
| 482 : | psubw mm0, mm1 ; undisplace | ||
| 483 : | psubw mm3, mm4 | ||
| 484 : | movq [edi + 8*ecx], mm0 | ||
| 485 : | movq [edi + 8*ecx + 8], mm3 | ||
| 486 : | edgomez | 1174 | |
| 487 : | edgomez | 1192 | add ecx, 2 |
| 488 : | cmp ecx, 16 | ||
| 489 : | jnz near .q2loop | ||
| 490 : | edgomez | 1174 | |
| 491 : | edgomez | 1192 | jmp .done |
| 492 : | edgomez | 1174 | |
| 493 : | |||
| 494 : | edgomez | 1192 | ;----------------------------------------------------------------------------- |
| 495 : | edgomez | 1174 | ; |
| 496 : | edgomez | 1176 | ; uint32_t dequant_mpeg_intra_mmx(int16_t *data, |
| 497 : | ; const int16_t const *coeff, | ||
| 498 : | ; const uint32_t quant, | ||
| 499 : | ; const uint32_t dcscalar); | ||
| 500 : | edgomez | 1174 | ; |
| 501 : | edgomez | 1192 | ;----------------------------------------------------------------------------- |
| 502 : | edgomez | 1174 | |
| 503 : | ; Note: in order to saturate 'easily', we pre-shift the quantifier | ||
| 504 : | ; by 4. Then, the high-word of (coeff[]*matrix[i]*quant) are used to | ||
| 505 : | ; build a saturating mask. It is non-zero only when an overflow occured. | ||
| 506 : | ; We thus avoid packing/unpacking toward double-word. | ||
| 507 : | ; Moreover, we perform the mult (matrix[i]*quant) first, instead of, e.g., | ||
| 508 : | ; (coeff[i]*matrix[i]). This is less prone to overflow if coeff[] are not | ||
| 509 : | ; checked. Input ranges are: coeff in [-127,127], inter_matrix in [1..255],a | ||
| 510 : | edgomez | 1192 | ; and quant in [1..31]. |
| 511 : | edgomez | 1174 | ; |
| 512 : | ; The original loop is: | ||
| 513 : | ; | ||
| 514 : | %if 0 | ||
| 515 : | edgomez | 1192 | movq mm0, [ecx+8*eax + 8*16] ; mm0 = coeff[i] |
| 516 : | pxor mm1, mm1 | ||
| 517 : | pcmpgtw mm1, mm0 | ||
| 518 : | pxor mm0, mm1 ; change sign if negative | ||
| 519 : | psubw mm0, mm1 ; -> mm0 = abs(coeff[i]), mm1 = sign of coeff[i] | ||
| 520 : | edgomez | 1174 | |
| 521 : | edgomez | 1192 | movq mm2, mm7 ; mm2 = quant |
| 522 : | pmullw mm2, [intra_matrix + 8*eax + 8*16 ] ; matrix[i]*quant. | ||
| 523 : | edgomez | 1174 | |
| 524 : | edgomez | 1192 | movq mm6, mm2 |
| 525 : | pmulhw mm2, mm0 ; high of coeff*(matrix*quant) (should be 0 if no overflow) | ||
| 526 : | pmullw mm0, mm6 ; low of coeff*(matrix*quant) | ||
| 527 : | edgomez | 1174 | |
| 528 : | edgomez | 1192 | pxor mm5, mm5 |
| 529 : | pcmpgtw mm2, mm5 ; otherflow? | ||
| 530 : | psrlw mm2, 5 ; =0 if no clamp, 2047 otherwise | ||
| 531 : | psrlw mm0, 5 | ||
| 532 : | paddw mm0, mm1 ; start restoring sign | ||
| 533 : | por mm0, mm2 ; saturate to 2047 if needed | ||
| 534 : | pxor mm0, mm1 ; finish negating back | ||
| 535 : | edgomez | 1174 | |
| 536 : | edgomez | 1192 | movq [edx + 8*eax + 8*16], mm0 ; data[i] |
| 537 : | add eax, 1 | ||
| 538 : | edgomez | 1174 | %endif |
| 539 : | |||
| 540 : | ;******************************************************************** | ||
| 541 : | |||
| 542 : | edgomez | 1192 | ALIGN 16 |
| 543 : | edgomez | 1174 | dequant_mpeg_intra_mmx: |
| 544 : | |||
| 545 : | edgomez | 1192 | mov edx, [esp+4] ; data |
| 546 : | mov ecx, [esp+8] ; coeff | ||
| 547 : | mov eax, [esp+12] ; quant | ||
| 548 : | edgomez | 1174 | |
| 549 : | edgomez | 1192 | movq mm7, [mmx_mul_quant + eax*8 - 8] |
| 550 : | mov eax, -16 ; to keep ALIGNed, we regularly process coeff[0] | ||
| 551 : | psllw mm7, 2 ; << 2. See comment. | ||
| 552 : | pxor mm6, mm6 ; this is a NOP | ||
| 553 : | edgomez | 1174 | |
| 554 : | edgomez | 1192 | ALIGN 16 |
| 555 : | edgomez | 1174 | .loop |
| 556 : | edgomez | 1192 | movq mm0, [ecx+8*eax + 8*16] ; mm0 = c = coeff[i] |
| 557 : | movq mm3, [ecx+8*eax + 8*16 +8]; mm3 = c' = coeff[i+1] | ||
| 558 : | pxor mm1, mm1 | ||
| 559 : | pxor mm4, mm4 | ||
| 560 : | pcmpgtw mm1, mm0 ; mm1 = sgn(c) | ||
| 561 : | movq mm2, mm7 ; mm2 = quant | ||
| 562 : | edgomez | 1174 | |
| 563 : | edgomez | 1192 | pcmpgtw mm4, mm3 ; mm4 = sgn(c') |
| 564 : | pmullw mm2, [intra_matrix + 8*eax + 8*16 ] ; matrix[i]*quant | ||
| 565 : | edgomez | 1174 | |
| 566 : | edgomez | 1192 | pxor mm0, mm1 ; negate if negative |
| 567 : | pxor mm3, mm4 ; negate if negative | ||
| 568 : | edgomez | 1174 | |
| 569 : | edgomez | 1192 | psubw mm0, mm1 |
| 570 : | psubw mm3, mm4 | ||
| 571 : | edgomez | 1174 | |
| 572 : | edgomez | 1192 | ; we're short on register, here. Poor pairing... |
| 573 : | edgomez | 1174 | |
| 574 : | edgomez | 1192 | movq mm5, mm2 |
| 575 : | pmullw mm2, mm0 ; low of coeff*(matrix*quant) | ||
| 576 : | edgomez | 1174 | |
| 577 : | edgomez | 1192 | pmulhw mm0, mm5 ; high of coeff*(matrix*quant) |
| 578 : | movq mm5, mm7 ; mm2 = quant | ||
| 579 : | edgomez | 1174 | |
| 580 : | edgomez | 1192 | pmullw mm5, [intra_matrix + 8*eax + 8*16 +8] ; matrix[i+1]*quant |
| 581 : | edgomez | 1174 | |
| 582 : | edgomez | 1192 | movq mm6, mm5 |
| 583 : | add eax,2 ; z-flag will be tested later | ||
| 584 : | edgomez | 1174 | |
| 585 : | edgomez | 1192 | pmullw mm6, mm3 ; low of coeff*(matrix*quant) |
| 586 : | pmulhw mm3, mm5 ; high of coeff*(matrix*quant) | ||
| 587 : | edgomez | 1174 | |
| 588 : | edgomez | 1192 | pcmpgtw mm0, [zero] |
| 589 : | paddusw mm2, mm0 | ||
| 590 : | psrlw mm2, 5 | ||
| 591 : | edgomez | 1174 | |
| 592 : | edgomez | 1192 | pcmpgtw mm3, [zero] |
| 593 : | paddusw mm6, mm3 | ||
| 594 : | psrlw mm6, 5 | ||
| 595 : | edgomez | 1174 | |
| 596 : | edgomez | 1192 | pxor mm2, mm1 ; start negating back |
| 597 : | pxor mm6, mm4 ; start negating back | ||
| 598 : | edgomez | 1174 | |
| 599 : | edgomez | 1192 | psubusw mm1, mm0 |
| 600 : | psubusw mm4, mm3 | ||
| 601 : | edgomez | 1174 | |
| 602 : | edgomez | 1192 | psubw mm2, mm1 ; finish negating back |
| 603 : | psubw mm6, mm4 ; finish negating back | ||
| 604 : | edgomez | 1174 | |
| 605 : | edgomez | 1192 | movq [edx + 8*eax + 8*16 -2*8 ], mm2 ; data[i] |
| 606 : | movq [edx + 8*eax + 8*16 -2*8 +8], mm6 ; data[i+1] | ||
| 607 : | edgomez | 1174 | |
| 608 : | edgomez | 1192 | jnz near .loop |
| 609 : | edgomez | 1174 | |
| 610 : | edgomez | 1192 | ; deal with DC |
| 611 : | movd mm0, [ecx] | ||
| 612 : | pmullw mm0, [esp+16] ; dcscalar | ||
| 613 : | movq mm2, [mmx_32767_minus_2047] | ||
| 614 : | paddsw mm0, mm2 | ||
| 615 : | psubsw mm0, mm2 | ||
| 616 : | movq mm2, [mmx_32768_minus_2048] | ||
| 617 : | psubsw mm0, mm2 | ||
| 618 : | paddsw mm0, mm2 | ||
| 619 : | movd eax, mm0 | ||
| 620 : | mov [edx], ax | ||
| 621 : | |||
| 622 : | xor eax, eax | ||
| 623 : | ret | ||
| 624 : | |||
| 625 : | ;----------------------------------------------------------------------------- | ||
| 626 : | edgomez | 1174 | ; |
| 627 : | edgomez | 1176 | ; uint32_t dequant_mpeg_inter_mmx(int16_t * data, |
| 628 : | ; const int16_t * const coeff, | ||
| 629 : | ; const uint32_t quant); | ||
| 630 : | edgomez | 1174 | ; |
| 631 : | edgomez | 1192 | ;----------------------------------------------------------------------------- |
| 632 : | edgomez | 1174 | |
| 633 : | ; Note: We use (2*c + sgn(c) - sgn(-c)) as multiplier | ||
| 634 : | ; so we handle the 3 cases: c<0, c==0, and c>0 in one shot. | ||
| 635 : | ; sgn(x) is the result of 'pcmpgtw 0,x': 0 if x>=0, -1 if x<0. | ||
| 636 : | ; It's mixed with the extraction of the absolute value. | ||
| 637 : | |||
| 638 : | edgomez | 1192 | ALIGN 16 |
| 639 : | edgomez | 1174 | dequant_mpeg_inter_mmx: |
| 640 : | |||
| 641 : | edgomez | 1192 | mov edx, [esp+ 4] ; data |
| 642 : | mov ecx, [esp+ 8] ; coeff | ||
| 643 : | mov eax, [esp+12] ; quant | ||
| 644 : | movq mm7, [mmx_mul_quant + eax*8 - 8] | ||
| 645 : | mov eax, -16 | ||
| 646 : | paddw mm7, mm7 ; << 1 | ||
| 647 : | pxor mm6, mm6 ; mismatch sum | ||
| 648 : | edgomez | 1174 | |
| 649 : | edgomez | 1192 | ALIGN 16 |
| 650 : | edgomez | 1174 | .loop |
| 651 : | edgomez | 1192 | movq mm0, [ecx+8*eax + 8*16 ] ; mm0 = coeff[i] |
| 652 : | movq mm2, [ecx+8*eax + 8*16 +8] ; mm2 = coeff[i+1] | ||
| 653 : | add eax, 2 | ||
| 654 : | edgomez | 1174 | |
| 655 : | edgomez | 1192 | pxor mm1, mm1 |
| 656 : | pxor mm3, mm3 | ||
| 657 : | pcmpgtw mm1, mm0 ; mm1 = sgn(c) (preserved) | ||
| 658 : | pcmpgtw mm3, mm2 ; mm3 = sgn(c') (preserved) | ||
| 659 : | paddsw mm0, mm1 ; c += sgn(c) | ||
| 660 : | paddsw mm2, mm3 ; c += sgn(c') | ||
| 661 : | paddw mm0, mm0 ; c *= 2 | ||
| 662 : | paddw mm2, mm2 ; c'*= 2 | ||
| 663 : | edgomez | 1174 | |
| 664 : | edgomez | 1192 | pxor mm4, mm4 |
| 665 : | pxor mm5, mm5 | ||
| 666 : | psubw mm4, mm0 ; -c | ||
| 667 : | psubw mm5, mm2 ; -c' | ||
| 668 : | psraw mm4, 16 ; mm4 = sgn(-c) | ||
| 669 : | psraw mm5, 16 ; mm5 = sgn(-c') | ||
| 670 : | psubsw mm0, mm4 ; c -= sgn(-c) | ||
| 671 : | psubsw mm2, mm5 ; c' -= sgn(-c') | ||
| 672 : | pxor mm0, mm1 ; finish changing sign if needed | ||
| 673 : | pxor mm2, mm3 ; finish changing sign if needed | ||
| 674 : | edgomez | 1174 | |
| 675 : | ; we're short on register, here. Poor pairing... | ||
| 676 : | |||
| 677 : | edgomez | 1192 | movq mm4, mm7 ; (matrix*quant) |
| 678 : | pmullw mm4, [inter_matrix + 8*eax + 8*16 -2*8] | ||
| 679 : | movq mm5, mm4 | ||
| 680 : | pmulhw mm5, mm0 ; high of c*(matrix*quant) | ||
| 681 : | pmullw mm0, mm4 ; low of c*(matrix*quant) | ||
| 682 : | edgomez | 1174 | |
| 683 : | edgomez | 1192 | movq mm4, mm7 ; (matrix*quant) |
| 684 : | pmullw mm4, [inter_matrix + 8*eax + 8*16 -2*8 + 8] | ||
| 685 : | edgomez | 1174 | |
| 686 : | edgomez | 1192 | pcmpgtw mm5, [zero] |
| 687 : | paddusw mm0, mm5 | ||
| 688 : | psrlw mm0, 5 | ||
| 689 : | pxor mm0, mm1 ; start restoring sign | ||
| 690 : | psubusw mm1, mm5 | ||
| 691 : | edgomez | 1174 | |
| 692 : | edgomez | 1192 | movq mm5, mm4 |
| 693 : | pmulhw mm5, mm2 ; high of c*(matrix*quant) | ||
| 694 : | pmullw mm2, mm4 ; low of c*(matrix*quant) | ||
| 695 : | psubw mm0, mm1 ; finish restoring sign | ||
| 696 : | edgomez | 1174 | |
| 697 : | edgomez | 1192 | pcmpgtw mm5, [zero] |
| 698 : | paddusw mm2, mm5 | ||
| 699 : | psrlw mm2, 5 | ||
| 700 : | pxor mm2, mm3 ; start restoring sign | ||
| 701 : | psubusw mm3, mm5 | ||
| 702 : | psubw mm2, mm3 ; finish restoring sign | ||
| 703 : | edgomez | 1174 | |
| 704 : | edgomez | 1192 | pxor mm6, mm0 ; mismatch control |
| 705 : | movq [edx + 8*eax + 8*16 -2*8 ], mm0 ; data[i] | ||
| 706 : | pxor mm6, mm2 ; mismatch control | ||
| 707 : | movq [edx + 8*eax + 8*16 -2*8 +8], mm2 ; data[i+1] | ||
| 708 : | edgomez | 1174 | |
| 709 : | edgomez | 1192 | jnz near .loop |
| 710 : | edgomez | 1174 | |
| 711 : | ; mismatch control | ||
| 712 : | |||
| 713 : | edgomez | 1192 | movq mm0, mm6 |
| 714 : | psrlq mm0, 48 | ||
| 715 : | movq mm1, mm6 | ||
| 716 : | movq mm2, mm6 | ||
| 717 : | psrlq mm1, 32 | ||
| 718 : | pxor mm6, mm0 | ||
| 719 : | psrlq mm2, 16 | ||
| 720 : | pxor mm6, mm1 | ||
| 721 : | pxor mm6, mm2 | ||
| 722 : | movd eax, mm6 | ||
| 723 : | and eax, 1 | ||
| 724 : | xor eax, 1 | ||
| 725 : | xor word [edx + 2*63], ax | ||
| 726 : | edgomez | 1174 | |
| 727 : | edgomez | 1192 | xor eax, eax |
| 728 : | ret |
| No admin address has been configured | ViewVC Help |
| Powered by ViewVC 1.0.4 |