Parent Directory
| 
Revision Log
Revision 1089 - (view) (download)
| 1 : | edgomez | 851 | ;/************************************************************************** |
| 2 : | ; * | ||
| 3 : | edgomez | 1089 | ; * XVID MPEG-4 VIDEO CODEC |
| 4 : | ; * - mmx quantization/dequantization - | ||
| 5 : | edgomez | 851 | ; * |
| 6 : | edgomez | 1089 | ; * Copyright(C) 2001-2003 XviD Team <xvid-devel@xvid.org> |
| 7 : | edgomez | 851 | ; * |
| 8 : | edgomez | 1089 | ; * This program is free software ; you can redistribute it and/or modify |
| 9 : | ; * it under the terms of the GNU General Public License as published by | ||
| 10 : | ; * the Free Software Foundation ; either version 2 of the License, or | ||
| 11 : | ; * (at your option) any later version. | ||
| 12 : | edgomez | 851 | ; * |
| 13 : | edgomez | 1089 | ; * This program is distributed in the hope that it will be useful, |
| 14 : | ; * but WITHOUT ANY WARRANTY ; without even the implied warranty of | ||
| 15 : | ; * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | ||
| 16 : | ; * GNU General Public License for more details. | ||
| 17 : | edgomez | 851 | ; * |
| 18 : | edgomez | 1089 | ; * You should have received a copy of the GNU General Public License |
| 19 : | ; * along with this program ; if not, write to the Free Software | ||
| 20 : | ; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | ||
| 21 : | ; * | ||
| 22 : | ; * $Id: quantize_3dne.asm,v 1.2.2.1 2003-07-16 22:59:20 edgomez Exp $ | ||
| 23 : | ; * | ||
| 24 : | edgomez | 851 | ; *************************************************************************/ |
| 25 : | ; these 3dne functions are compatible with iSSE, but are optimized specifically for | ||
| 26 : | ; K7 pipelines | ||
| 27 : | ; | ||
| 28 : | ;------------------------------------------------------------------------------ | ||
| 29 : | ; 09.12.2002 Athlon optimizations contributed by Jaan Kalda | ||
| 30 : | ;------------------------------------------------------------------------------ | ||
| 31 : | |||
| 32 : | ; enable dequant saturate [-2048,2047], test purposes only. | ||
| 33 : | %define SATURATE | ||
| 34 : | |||
| 35 : | ; data/text alignment | ||
| 36 : | %define ALIGN 16 | ||
| 37 : | |||
| 38 : | bits 32 | ||
| 39 : | |||
| 40 : | %ifdef FORMAT_COFF | ||
| 41 : | section .data data | ||
| 42 : | %else | ||
| 43 : | section .data data align=16 | ||
| 44 : | %endif | ||
| 45 : | |||
| 46 : | %macro cglobal 1 | ||
| 47 : | %ifdef PREFIX | ||
| 48 : | global _%1 | ||
| 49 : | %define %1 _%1 | ||
| 50 : | %else | ||
| 51 : | global %1 | ||
| 52 : | %endif | ||
| 53 : | %endmacro | ||
| 54 : | align 4 | ||
| 55 : | int_div | ||
| 56 : | dd 0 | ||
| 57 : | %assign i 1 | ||
| 58 : | %rep 255 | ||
| 59 : | edgomez | 1089 | dd (1 << 16) / (i) + 1 |
| 60 : | edgomez | 851 | %assign i i+1 |
| 61 : | %endrep | ||
| 62 : | |||
| 63 : | align 16 | ||
| 64 : | |||
| 65 : | plus_one times 8 dw 1 | ||
| 66 : | |||
| 67 : | ;=========================================================================== | ||
| 68 : | ; | ||
| 69 : | ; subtract by Q/2 table | ||
| 70 : | ; | ||
| 71 : | ;=========================================================================== | ||
| 72 : | |||
| 73 : | %macro MMX_SUB 1 | ||
| 74 : | times 4 dw %1 / 2 | ||
| 75 : | %endmacro | ||
| 76 : | |||
| 77 : | |||
| 78 : | align 16 | ||
| 79 : | mmx_sub | ||
| 80 : | %assign i 1 | ||
| 81 : | %rep 31 | ||
| 82 : | times 4 dw i / 2 | ||
| 83 : | %assign i i+1 | ||
| 84 : | %endrep | ||
| 85 : | |||
| 86 : | |||
| 87 : | ;=========================================================================== | ||
| 88 : | ; | ||
| 89 : | ; divide by 2Q table | ||
| 90 : | ; | ||
| 91 : | ; use a shift of 16 to take full advantage of _pmulhw_ | ||
| 92 : | ; for q=1, _pmulhw_ will overflow so it is treated seperately | ||
| 93 : | ; (3dnow2 provides _pmulhuw_ which wont cause overflow) | ||
| 94 : | ; | ||
| 95 : | ;=========================================================================== | ||
| 96 : | |||
| 97 : | align 16 | ||
| 98 : | mmx_div | ||
| 99 : | |||
| 100 : | %assign i 1 | ||
| 101 : | %rep 31 | ||
| 102 : | times 4 dw (1 << 16) / (i * 2) + 1 | ||
| 103 : | %assign i i+1 | ||
| 104 : | %endrep | ||
| 105 : | |||
| 106 : | ;=========================================================================== | ||
| 107 : | ; | ||
| 108 : | ; add by (odd(Q) ? Q : Q - 1) table | ||
| 109 : | ; | ||
| 110 : | ;=========================================================================== | ||
| 111 : | |||
| 112 : | %macro MMX_ADD 1 | ||
| 113 : | %if %1 % 2 != 0 | ||
| 114 : | times 4 dw %1 | ||
| 115 : | %else | ||
| 116 : | times 4 dw %1 - 1 | ||
| 117 : | %endif | ||
| 118 : | %endmacro | ||
| 119 : | |||
| 120 : | align 16 | ||
| 121 : | mmx_add | ||
| 122 : | |||
| 123 : | %assign i 1 | ||
| 124 : | %rep 31 | ||
| 125 : | MMX_ADD i | ||
| 126 : | %assign i i+1 | ||
| 127 : | %endrep | ||
| 128 : | |||
| 129 : | ;=========================================================================== | ||
| 130 : | ; | ||
| 131 : | ; multiple by 2Q table | ||
| 132 : | ; | ||
| 133 : | ;=========================================================================== | ||
| 134 : | |||
| 135 : | %macro MMX_MUL 1 | ||
| 136 : | times 4 dw %1 * 2 | ||
| 137 : | %endmacro | ||
| 138 : | |||
| 139 : | align 16 | ||
| 140 : | mmx_mul | ||
| 141 : | |||
| 142 : | %assign i 1 | ||
| 143 : | %rep 31 | ||
| 144 : | times 4 dw i * 2 | ||
| 145 : | %assign i i+1 | ||
| 146 : | %endrep | ||
| 147 : | |||
| 148 : | ;=========================================================================== | ||
| 149 : | ; | ||
| 150 : | ; saturation limits | ||
| 151 : | ; | ||
| 152 : | ;=========================================================================== | ||
| 153 : | |||
| 154 : | align 8 | ||
| 155 : | mmx_32768_minus_2048 times 4 dw (32768-2048) | ||
| 156 : | mmx_32767_minus_2047 times 4 dw (32767-2047) | ||
| 157 : | |||
| 158 : | align 16 | ||
| 159 : | mmx_2047 times 4 dw 2047 | ||
| 160 : | |||
| 161 : | align 8 | ||
| 162 : | mmzero dd 0, 0 | ||
| 163 : | int2047 dd 2047 | ||
| 164 : | int_2048 dd -2048 | ||
| 165 : | |||
| 166 : | section .text | ||
| 167 : | |||
| 168 : | |||
| 169 : | ;=========================================================================== | ||
| 170 : | ; | ||
| 171 : | ; void quant_intra_3dne(int16_t * coeff, | ||
| 172 : | ; const int16_t const * data, | ||
| 173 : | ; const uint32_t quant, | ||
| 174 : | ; const uint32_t dcscalar); | ||
| 175 : | ; | ||
| 176 : | ;=========================================================================== | ||
| 177 : | ;This is Athlon-optimized code (ca 70 clk per call) | ||
| 178 : | ;Optimized by Jaan, 30 Nov 2002 | ||
| 179 : | |||
| 180 : | edgomez | 1089 | %macro quant_intra1 1 |
| 181 : | |||
| 182 : | psubw mm1, mm0 ;A3 | ||
| 183 : | psubw mm3, mm2 ;B3 | ||
| 184 : | edgomez | 851 | %if (%1) |
| 185 : | edgomez | 1089 | psubw mm5, mm4 ;C8 |
| 186 : | psubw mm7, mm6 ;D8 | ||
| 187 : | edgomez | 851 | %endif |
| 188 : | |||
| 189 : | align 8 | ||
| 190 : | edgomez | 1089 | db 0Fh, 6Fh, 64h, 21h, (%1 * 32 +16) ;movq mm4, [ecx + %1 * 32 +16+32] ;C1 |
| 191 : | pmaxsw mm1, mm0 ;A4 | ||
| 192 : | db 0Fh, 6Fh, 74h, 21h, (%1 * 32 +24) ;movq mm6, [ecx + %1 * 32 +24+32] ;D1 | ||
| 193 : | pmaxsw mm3, mm2 ;B4 | ||
| 194 : | edgomez | 851 | |
| 195 : | |||
| 196 : | edgomez | 1089 | psraw mm0, 15 ;A5 |
| 197 : | psraw mm2, 15 ;B5 | ||
| 198 : | edgomez | 851 | %if (%1) |
| 199 : | edgomez | 1089 | movq [edx + %1 * 32 + 16-32], mm5 ;C9 |
| 200 : | movq [edx + %1 * 32 + 24-32], mm7 ;D9 | ||
| 201 : | edgomez | 851 | %endif |
| 202 : | |||
| 203 : | edgomez | 1089 | psrlw mm1, 1 ;A6 |
| 204 : | psrlw mm3, 1 ;B6 | ||
| 205 : | movq mm5, [ebx] ;C2 | ||
| 206 : | movq mm7, [ebx] ;D2 | ||
| 207 : | edgomez | 851 | |
| 208 : | edgomez | 1089 | pxor mm1, mm0 ;A7 |
| 209 : | pxor mm3, mm2 ;B7 | ||
| 210 : | edgomez | 851 | |
| 211 : | edgomez | 1089 | psubw mm5, mm4 ;C3 |
| 212 : | psubw mm7, mm6 ;D3 | ||
| 213 : | psubw mm1, mm0 ;A8 | ||
| 214 : | psubw mm3, mm2 ;B8 | ||
| 215 : | edgomez | 851 | |
| 216 : | %if (%1 == 0) | ||
| 217 : | edgomez | 1089 | push ebp |
| 218 : | movq mm0, [ecx + %1 * 32 +32] | ||
| 219 : | edgomez | 851 | %elif (%1 < 3) |
| 220 : | edgomez | 1089 | db 0Fh, 6Fh, 44h, 21h, (%1 * 32 +32) ;movq mm0, [ecx + %1 * 32 +32] ;A1 |
| 221 : | edgomez | 851 | %endif |
| 222 : | edgomez | 1089 | pmaxsw mm5, mm4 ;C4 |
| 223 : | edgomez | 851 | %if (%1 < 3) |
| 224 : | edgomez | 1089 | db 0Fh, 6Fh, 54h, 21h, ( %1 * 32 +8+32) ;movq mm2, [ecx + %1 * 32 +8+32] ;B1 |
| 225 : | edgomez | 851 | %else |
| 226 : | edgomez | 1089 | cmp esp, esp |
| 227 : | edgomez | 851 | %endif |
| 228 : | edgomez | 1089 | pmaxsw mm7, mm6 ;D4 |
| 229 : | edgomez | 851 | |
| 230 : | edgomez | 1089 | psraw mm4, 15 ;C5 |
| 231 : | psraw mm6, 15 ;D5 | ||
| 232 : | movq [byte edx + %1 * 32], mm1 ;A9 | ||
| 233 : | movq [edx + %1 * 32+8], mm3 ;B9 | ||
| 234 : | edgomez | 851 | |
| 235 : | |||
| 236 : | edgomez | 1089 | psrlw mm5, 1 ;C6 |
| 237 : | psrlw mm7, 1 ;D6 | ||
| 238 : | edgomez | 851 | %if (%1 < 3) |
| 239 : | edgomez | 1089 | movq mm1, [ebx] ;A2 |
| 240 : | movq mm3, [ebx] ;B2 | ||
| 241 : | edgomez | 851 | %endif |
| 242 : | %if (%1 == 3) | ||
| 243 : | edgomez | 1089 | imul eax, [int_div+4*edi] |
| 244 : | edgomez | 851 | %endif |
| 245 : | edgomez | 1089 | pxor mm5, mm4 ;C7 |
| 246 : | pxor mm7, mm6 ;D7 | ||
| 247 : | edgomez | 851 | %endm |
| 248 : | |||
| 249 : | |||
| 250 : | edgomez | 1089 | %macro quant_intra 1 |
| 251 : | ; Rules for athlon: | ||
| 252 : | ; 1) schedule latencies | ||
| 253 : | ; 2) add/mul and load/store in 2:1 proportion | ||
| 254 : | ; 3) avoid spliting >3byte instructions over 8byte boundaries | ||
| 255 : | |||
| 256 : | psubw mm1, mm0 ;A3 | ||
| 257 : | psubw mm3, mm2 ;B3 | ||
| 258 : | edgomez | 851 | %if (%1) |
| 259 : | edgomez | 1089 | psubw mm5, mm4 ;C8 |
| 260 : | psubw mm7, mm6 ;D8 | ||
| 261 : | edgomez | 851 | %endif |
| 262 : | |||
| 263 : | align 8 | ||
| 264 : | edgomez | 1089 | db 0Fh, 6Fh, 64h, 21h, (%1 * 32 +16) ;movq mm4, [ecx + %1 * 32 +16+32] ;C1 |
| 265 : | pmaxsw mm1, mm0 ;A4 | ||
| 266 : | db 0Fh, 6Fh, 74h, 21h, (%1 * 32 +24) ;movq mm6, [ecx + %1 * 32 +24+32] ;D1 | ||
| 267 : | pmaxsw mm3, mm2 ;B4 | ||
| 268 : | edgomez | 851 | |
| 269 : | |||
| 270 : | edgomez | 1089 | psraw mm0, 15 ;A5 |
| 271 : | psraw mm2, 15 ;B5 | ||
| 272 : | edgomez | 851 | %if (%1) |
| 273 : | edgomez | 1089 | movq [edx + %1 * 32 + 16-32], mm5 ;C9 |
| 274 : | movq [edx + %1 * 32 + 24-32], mm7 ;D9 | ||
| 275 : | edgomez | 851 | %endif |
| 276 : | |||
| 277 : | edgomez | 1089 | pmulhw mm1, [esi] ;A6 |
| 278 : | pmulhw mm3, [esi] ;B6 | ||
| 279 : | movq mm5, [ebx] ;C2 | ||
| 280 : | movq mm7, [ebx] ;D2 | ||
| 281 : | edgomez | 851 | |
| 282 : | edgomez | 1089 | nop |
| 283 : | nop | ||
| 284 : | pxor mm1, mm0 ;A7 | ||
| 285 : | pxor mm3, mm2 ;B7 | ||
| 286 : | edgomez | 851 | |
| 287 : | edgomez | 1089 | psubw mm5, mm4 ;C3 |
| 288 : | psubw mm7, mm6 ;D3 | ||
| 289 : | psubw mm1, mm0 ;A8 | ||
| 290 : | psubw mm3, mm2 ;B8 | ||
| 291 : | edgomez | 851 | |
| 292 : | |||
| 293 : | %if (%1 < 3) | ||
| 294 : | edgomez | 1089 | db 0Fh, 6Fh, 44h, 21h, (%1 * 32 +32) ;movq mm0, [ecx + %1 * 32 +32] ;A1 |
| 295 : | edgomez | 851 | %endif |
| 296 : | edgomez | 1089 | pmaxsw mm5, mm4 ;C4 |
| 297 : | edgomez | 851 | %if (%1 < 3) |
| 298 : | edgomez | 1089 | db 0Fh, 6Fh, 54h, 21h, ( %1 * 32 +8+32) ;movq mm2, [ecx + %1 * 32 +8+32] ;B1 |
| 299 : | edgomez | 851 | %else |
| 300 : | edgomez | 1089 | cmp esp, esp |
| 301 : | edgomez | 851 | %endif |
| 302 : | edgomez | 1089 | pmaxsw mm7,mm6 ;D4 |
| 303 : | edgomez | 851 | |
| 304 : | edgomez | 1089 | psraw mm4, 15 ;C5 |
| 305 : | psraw mm6, 15 ;D5 | ||
| 306 : | movq [byte edx + %1 * 32], mm1 ;A9 | ||
| 307 : | movq [edx + %1 * 32+8], mm3 ;B9 | ||
| 308 : | edgomez | 851 | |
| 309 : | |||
| 310 : | edgomez | 1089 | pmulhw mm5, [esi] ;C6 |
| 311 : | pmulhw mm7, [esi] ;D6 | ||
| 312 : | edgomez | 851 | %if (%1 < 3) |
| 313 : | edgomez | 1089 | movq mm1, [ebx] ;A2 |
| 314 : | movq mm3, [ebx] ;B2 | ||
| 315 : | edgomez | 851 | %endif |
| 316 : | %if (%1 == 0) | ||
| 317 : | edgomez | 1089 | push ebp |
| 318 : | edgomez | 851 | %elif (%1 < 3) |
| 319 : | edgomez | 1089 | nop |
| 320 : | edgomez | 851 | %endif |
| 321 : | edgomez | 1089 | nop |
| 322 : | edgomez | 851 | %if (%1 == 3) |
| 323 : | edgomez | 1089 | imul eax, [int_div+4*edi] |
| 324 : | edgomez | 851 | %endif |
| 325 : | edgomez | 1089 | pxor mm5, mm4 ;C7 |
| 326 : | pxor mm7, mm6 ;D7 | ||
| 327 : | edgomez | 851 | %endmacro |
| 328 : | |||
| 329 : | |||
| 330 : | align ALIGN | ||
| 331 : | cglobal quant_intra_3dne | ||
| 332 : | quant_intra_3dne: | ||
| 333 : | |||
| 334 : | edgomez | 1089 | mov eax, [esp + 12] ; quant |
| 335 : | mov ecx, [esp + 8] ; data | ||
| 336 : | mov edx, [esp + 4] ; coeff | ||
| 337 : | cmp al, 1 | ||
| 338 : | pxor mm1, mm1 | ||
| 339 : | pxor mm3, mm3 | ||
| 340 : | movq mm0, [ecx] ; mm0 = [1st] | ||
| 341 : | movq mm2, [ecx + 8] | ||
| 342 : | push esi | ||
| 343 : | lea esi, [mmx_div + eax*8 - 8] | ||
| 344 : | edgomez | 851 | |
| 345 : | edgomez | 1089 | push ebx |
| 346 : | mov ebx, mmzero | ||
| 347 : | push edi | ||
| 348 : | jz near .q1loop | ||
| 349 : | |||
| 350 : | edgomez | 851 | quant_intra 0 |
| 351 : | edgomez | 1089 | mov ebp, [esp + 16 + 16] ; dcscalar |
| 352 : | ; NB -- there are 3 pushes in the function preambule and one more | ||
| 353 : | ; in "quant_intra 0", thus an added offset of 16 bytes | ||
| 354 : | movsx eax, word [byte ecx] ; DC | ||
| 355 : | |||
| 356 : | edgomez | 851 | quant_intra 1 |
| 357 : | edgomez | 1089 | mov edi, eax |
| 358 : | sar edi, 31 ; sign(DC) | ||
| 359 : | shr ebp, byte 1 ; ebp = dcscalar/2 | ||
| 360 : | |||
| 361 : | edgomez | 851 | quant_intra 2 |
| 362 : | edgomez | 1089 | sub eax, edi ; DC (+1) |
| 363 : | xor ebp, edi ; sign(DC) dcscalar /2 (-1) | ||
| 364 : | mov edi, [esp + 16 + 16] ; dscalar | ||
| 365 : | lea eax, [byte eax + ebp] ; DC + sign(DC) dcscalar/2 | ||
| 366 : | mov ebp, [byte esp] | ||
| 367 : | |||
| 368 : | edgomez | 851 | quant_intra 3 |
| 369 : | edgomez | 1089 | psubw mm5, mm4 ;C8 |
| 370 : | mov esi, [esp + 12] ; pop back the register value | ||
| 371 : | mov edi, [esp + 4] ; pop back the register value | ||
| 372 : | sar eax, 16 | ||
| 373 : | lea ebx, [byte eax + 1] ; workaround for eax < 0 | ||
| 374 : | cmovs eax, ebx ; conditionnaly move the corrected value | ||
| 375 : | mov [edx], ax ; coeff[0] = ax | ||
| 376 : | mov ebx, [esp + 8] ; pop back the register value | ||
| 377 : | add esp, byte 16 ; "quant_intra 0" pushed ebp, but we don't restore that one, just correct the stack offset by 16 | ||
| 378 : | psubw mm7, mm6 ;D8 | ||
| 379 : | movq [edx + 3 * 32 + 16], mm5 ;C9 | ||
| 380 : | movq [edx + 3 * 32 + 24], mm7 ;D9 | ||
| 381 : | |||
| 382 : | ret | ||
| 383 : | |||
| 384 : | align 16 | ||
| 385 : | |||
| 386 : | edgomez | 851 | .q1loop |
| 387 : | quant_intra1 0 | ||
| 388 : | edgomez | 1089 | mov ebp, [esp + 16 + 16] ; dcscalar |
| 389 : | movsx eax, word [byte ecx] ; DC | ||
| 390 : | |||
| 391 : | edgomez | 851 | quant_intra1 1 |
| 392 : | edgomez | 1089 | mov edi, eax |
| 393 : | sar edi, 31 ; sign(DC) | ||
| 394 : | shr ebp, byte 1 ; ebp = dcscalar /2 | ||
| 395 : | |||
| 396 : | edgomez | 851 | quant_intra1 2 |
| 397 : | edgomez | 1089 | sub eax, edi ; DC (+1) |
| 398 : | xor ebp, edi ; sign(DC) dcscalar /2 (-1) | ||
| 399 : | mov edi, [esp + 16 + 16] ; dcscalar | ||
| 400 : | lea eax, [byte eax + ebp] ; DC + sign(DC) dcscalar /2 | ||
| 401 : | mov ebp, [byte esp] | ||
| 402 : | |||
| 403 : | edgomez | 851 | quant_intra1 3 |
| 404 : | edgomez | 1089 | psubw mm5, mm4 ;C8 |
| 405 : | mov esi, [dword esp + 12] ; pop back the register value | ||
| 406 : | mov edi, [esp + 4] ; pop back the register value | ||
| 407 : | sar eax, 16 | ||
| 408 : | lea ebx, [byte eax + 1] ; workaround for eax < 0 | ||
| 409 : | cmovs eax, ebx ; conditionnaly move the corrected value | ||
| 410 : | mov [edx], ax ; coeff[0] = ax | ||
| 411 : | mov ebx, [esp + 8] ; pop back the register value | ||
| 412 : | add esp, byte 16 ; "quant_intra 0" pushed ebp, but we don't restore that one, just correct the stack offset by 16 | ||
| 413 : | psubw mm7, mm6 ;D8 | ||
| 414 : | movq [edx + 3 * 32 + 16], mm5 ;C9 | ||
| 415 : | movq [edx + 3 * 32 + 24], mm7 ;D9 | ||
| 416 : | edgomez | 851 | |
| 417 : | edgomez | 1089 | ret |
| 418 : | edgomez | 851 | |
| 419 : | |||
| 420 : | |||
| 421 : | edgomez | 1089 | |
| 422 : | edgomez | 851 | ;=========================================================================== |
| 423 : | ; | ||
| 424 : | ; uint32_t quant_inter_3dne(int16_t * coeff, | ||
| 425 : | ; const int16_t const * data, | ||
| 426 : | ; const uint32_t quant); | ||
| 427 : | ; | ||
| 428 : | ;=========================================================================== | ||
| 429 : | ;This is Athlon-optimized code (ca 90 clk per call) | ||
| 430 : | ;Optimized by Jaan, 30 Nov 2002 | ||
| 431 : | |||
| 432 : | |||
| 433 : | %macro quantinter 1 | ||
| 434 : | edgomez | 1089 | movq mm1, [eax] ;A2 |
| 435 : | psraw mm3, 15 ;B6 | ||
| 436 : | edgomez | 851 | %if (%1) |
| 437 : | edgomez | 1089 | psubw mm2, mm6 ;C10 |
| 438 : | edgomez | 851 | %endif |
| 439 : | edgomez | 1089 | psubw mm1, mm0 ;A3 |
| 440 : | pmulhw mm4, mm7 ;B7 | ||
| 441 : | movq mm6, [ecx + %1*24+16] ;C1 | ||
| 442 : | pmaxsw mm1, mm0 ;A4 | ||
| 443 : | paddw mm5, mm4 ;B8 | ||
| 444 : | edgomez | 851 | %if (%1) |
| 445 : | edgomez | 1089 | movq [edx + %1*24+16-24], mm2 ;C11 |
| 446 : | edgomez | 851 | %endif |
| 447 : | edgomez | 1089 | psubusw mm1, [ebx] ;A5 mm0 -= sub (unsigned, dont go < 0) |
| 448 : | pxor mm4, mm3 ;B9 | ||
| 449 : | movq mm2, [eax] ;C2 | ||
| 450 : | psraw mm0, 15 ;A6 | ||
| 451 : | psubw mm4, mm3 ;B10 | ||
| 452 : | psubw mm2, mm6 ;C3 | ||
| 453 : | pmulhw mm1, mm7 ;A7 mm0 = (mm0 / 2Q) >> 24 | ||
| 454 : | movq mm3, [ecx + %1*24+8] ;B1 | ||
| 455 : | pmaxsw mm2, mm6 ;C4 | ||
| 456 : | paddw mm5, mm1 ;A8 sum += mm0 | ||
| 457 : | edgomez | 851 | %if (%1) |
| 458 : | edgomez | 1089 | movq [edx + %1*24+8-24], mm4 ;B11 |
| 459 : | edgomez | 851 | %else |
| 460 : | edgomez | 1089 | movq [edx + 120], mm4 ;B11 |
| 461 : | edgomez | 851 | %endif |
| 462 : | edgomez | 1089 | psubusw mm2, [ebx] ;C5 |
| 463 : | pxor mm1, mm0 ;A9 mm0 *= sign(mm0) | ||
| 464 : | movq mm4, [eax] ;B2 | ||
| 465 : | psraw mm6, 15 ;C6 | ||
| 466 : | psubw mm1, mm0 ;A10 undisplace | ||
| 467 : | psubw mm4, mm3 ;B3 | ||
| 468 : | pmulhw mm2, mm7 ;C7 | ||
| 469 : | movq mm0, [ecx + %1*24+24] ;A1 mm0 = [1st] | ||
| 470 : | pmaxsw mm4, mm3 ;B4 | ||
| 471 : | paddw mm5, mm2 ;C8 | ||
| 472 : | movq [byte edx + %1*24], mm1 ;A11 | ||
| 473 : | psubusw mm4, [ebx] ;B5 | ||
| 474 : | pxor mm2, mm6 ;C9 | ||
| 475 : | edgomez | 851 | %endmacro |
| 476 : | |||
| 477 : | %macro quantinter1 1 | ||
| 478 : | edgomez | 1089 | movq mm0, [byte ecx + %1*16] ;mm0 = [1st] |
| 479 : | movq mm3, [ecx + %1*16+8] ; | ||
| 480 : | movq mm1, [eax] | ||
| 481 : | movq mm4, [eax] | ||
| 482 : | psubw mm1, mm0 | ||
| 483 : | psubw mm4, mm3 | ||
| 484 : | pmaxsw mm1, mm0 | ||
| 485 : | pmaxsw mm4, mm3 | ||
| 486 : | psubusw mm1, mm6 ; mm0 -= sub (unsigned, dont go < 0) | ||
| 487 : | psubusw mm4, mm6 ; | ||
| 488 : | psraw mm0, 15 | ||
| 489 : | psraw mm3, 15 | ||
| 490 : | psrlw mm1, 1 ; mm0 = (mm0 / 2Q) >> 16 | ||
| 491 : | psrlw mm4, 1 ; | ||
| 492 : | paddw mm5, mm1 ; sum += mm0 | ||
| 493 : | pxor mm1, mm0 ; mm0 *= sign(mm0) | ||
| 494 : | paddw mm5, mm4 | ||
| 495 : | pxor mm4, mm3 ; | ||
| 496 : | psubw mm1, mm0 ; undisplace | ||
| 497 : | psubw mm4, mm3 | ||
| 498 : | cmp esp, esp | ||
| 499 : | movq [byte edx + %1*16], mm1 | ||
| 500 : | movq [edx + %1*16+8], mm4 | ||
| 501 : | edgomez | 851 | %endmacro |
| 502 : | |||
| 503 : | align ALIGN | ||
| 504 : | cglobal quant_inter_3dne | ||
| 505 : | edgomez | 1089 | quant_inter_3dne |
| 506 : | mov edx, [esp + 4] ; coeff | ||
| 507 : | mov ecx, [esp + 8] ; data | ||
| 508 : | mov eax, [esp + 12] ; quant | ||
| 509 : | push ebx | ||
| 510 : | edgomez | 851 | |
| 511 : | edgomez | 1089 | pxor mm5, mm5 ; sum |
| 512 : | nop | ||
| 513 : | lea ebx,[mmx_sub + eax * 8 - 8] ; sub | ||
| 514 : | movq mm7, [mmx_div + eax * 8 - 8] ; divider | ||
| 515 : | edgomez | 851 | |
| 516 : | edgomez | 1089 | cmp al, 1 |
| 517 : | lea eax, [mmzero] | ||
| 518 : | jz near .q1loop | ||
| 519 : | cmp esp, esp | ||
| 520 : | edgomez | 851 | align 8 |
| 521 : | edgomez | 1089 | movq mm3, [ecx + 120] ;B1 |
| 522 : | pxor mm4, mm4 ;B2 | ||
| 523 : | psubw mm4, mm3 ;B3 | ||
| 524 : | movq mm0, [ecx] ;A1 mm0 = [1st] | ||
| 525 : | pmaxsw mm4, mm3 ;B4 | ||
| 526 : | psubusw mm4, [ebx] ;B5 | ||
| 527 : | edgomez | 851 | |
| 528 : | edgomez | 1089 | quantinter 0 |
| 529 : | quantinter 1 | ||
| 530 : | quantinter 2 | ||
| 531 : | quantinter 3 | ||
| 532 : | quantinter 4 | ||
| 533 : | edgomez | 851 | |
| 534 : | edgomez | 1089 | psraw mm3, 15 ;B6 |
| 535 : | psubw mm2, mm6 ;C10 | ||
| 536 : | pmulhw mm4, mm7 ;B7 | ||
| 537 : | paddw mm5, mm4 ;B8 | ||
| 538 : | pxor mm4, mm3 ;B9 | ||
| 539 : | psubw mm4, mm3 ;B10 | ||
| 540 : | movq [edx + 4*24+16], mm2 ;C11 | ||
| 541 : | pop ebx | ||
| 542 : | movq [edx + 4*24+8], mm4 ;B11 | ||
| 543 : | pmaddwd mm5, [plus_one] | ||
| 544 : | movq mm0, mm5 | ||
| 545 : | punpckhdq mm5, mm5 | ||
| 546 : | paddd mm0, mm5 | ||
| 547 : | movd eax, mm0 ; return sum | ||
| 548 : | |||
| 549 : | ret | ||
| 550 : | |||
| 551 : | edgomez | 851 | align ALIGN |
| 552 : | .q1loop | ||
| 553 : | edgomez | 1089 | movq mm6, [byte ebx] |
| 554 : | edgomez | 851 | |
| 555 : | edgomez | 1089 | quantinter1 0 |
| 556 : | quantinter1 1 | ||
| 557 : | quantinter1 2 | ||
| 558 : | quantinter1 3 | ||
| 559 : | quantinter1 4 | ||
| 560 : | quantinter1 5 | ||
| 561 : | quantinter1 6 | ||
| 562 : | quantinter1 7 | ||
| 563 : | |||
| 564 : | pmaddwd mm5, [plus_one] | ||
| 565 : | movq mm0, mm5 | ||
| 566 : | psrlq mm5, 32 | ||
| 567 : | paddd mm0, mm5 | ||
| 568 : | movd eax, mm0 ; return sum | ||
| 569 : | edgomez | 851 | |
| 570 : | edgomez | 1089 | pop ebx |
| 571 : | edgomez | 851 | |
| 572 : | edgomez | 1089 | ret |
| 573 : | edgomez | 851 | |
| 574 : | ;=========================================================================== | ||
| 575 : | ; | ||
| 576 : | ; void dequant_intra_3dne(int16_t *data, | ||
| 577 : | ; const int16_t const *coeff, | ||
| 578 : | ; const uint32_t quant, | ||
| 579 : | ; const uint32_t dcscalar); | ||
| 580 : | ; | ||
| 581 : | ;=========================================================================== | ||
| 582 : | |||
| 583 : | ; this is the same as dequant_inter_3dne, except that we're | ||
| 584 : | ; saturating using 'pminsw' (saves 2 cycles/loop => ~5% faster) | ||
| 585 : | |||
| 586 : | ;This is Athlon-optimized code (ca 106 clk per call) | ||
| 587 : | |||
| 588 : | %macro dequant 1 | ||
| 589 : | edgomez | 1089 | movq mm1, [ecx+%1*24] ; c = coeff[i] ;A2 |
| 590 : | psubw mm0, mm1 ;-c ;A3 (1st dep) | ||
| 591 : | edgomez | 851 | %if (%1) |
| 592 : | edgomez | 1089 | paddw mm4, mm6 ;C11 mm6 free (4th+) |
| 593 : | edgomez | 851 | %endif |
| 594 : | edgomez | 1089 | pmaxsw mm0, mm1 ;|c| ;A4 (2nd) |
| 595 : | edgomez | 851 | %if (%1) |
| 596 : | edgomez | 1089 | mov ebp, ebp |
| 597 : | pminsw mm4, [ebx] ;C12 saturates to +2047 (5th+) later | ||
| 598 : | edgomez | 851 | %endif |
| 599 : | edgomez | 1089 | movq mm6, [esi] ;0 ;A5 mm6 in use |
| 600 : | pandn mm7, [eax] ;B9 offset = isZero ? 0 : quant_add (2nd) | ||
| 601 : | edgomez | 851 | %if (%1) |
| 602 : | edgomez | 1089 | pxor mm5, mm4 ;C13 (6th+) 1later |
| 603 : | edgomez | 851 | %endif |
| 604 : | edgomez | 1089 | movq mm4, [esi] ;C1 ;0 |
| 605 : | mov esp, esp | ||
| 606 : | pcmpeqw mm6, [ecx+%1*24] ;A6 (c ==0) ? -1 : 0 (1st) | ||
| 607 : | edgomez | 851 | align 4 |
| 608 : | edgomez | 1089 | psraw mm1, 15 ; sign(c) ;A7 (2nd) |
| 609 : | edgomez | 851 | %if (%1) |
| 610 : | edgomez | 1089 | movq [edx+%1*24+16-24], mm5 ; C14 (7th) 2later |
| 611 : | edgomez | 851 | %endif |
| 612 : | edgomez | 1089 | paddw mm7, mm3 ;B10 offset +negate back (3rd) |
| 613 : | pmullw mm0, [edi] ;*= 2Q ;A8 (3rd+) | ||
| 614 : | paddw mm2, mm7 ;B11 mm7 free (4th+) | ||
| 615 : | lea ebp, [byte ebp] | ||
| 616 : | movq mm5, [ecx+%1*24+16] ;C2 ; c = coeff[i] | ||
| 617 : | psubw mm4, mm5 ;-c ;C3 (1st dep) | ||
| 618 : | pandn mm6, [eax] ;A9 offset = isZero ? 0 : quant_add (2nd) | ||
| 619 : | pminsw mm2, [ebx] ;B12 saturates to +2047 (5th+) | ||
| 620 : | pxor mm3, mm2 ;B13 (6th+) | ||
| 621 : | movq mm2, [byte esi] ;B1 ;0 | ||
| 622 : | edgomez | 851 | %if (%1) |
| 623 : | edgomez | 1089 | movq [edx+%1*24+8-24], mm3 ;B14 (7th) |
| 624 : | edgomez | 851 | %else |
| 625 : | edgomez | 1089 | movq [edx+120], mm3 |
| 626 : | edgomez | 851 | %endif |
| 627 : | edgomez | 1089 | pmaxsw mm4, mm5 ;|c| ;C4 (2nd) |
| 628 : | paddw mm6, mm1 ;A10 offset +negate back (3rd) | ||
| 629 : | movq mm3, [ecx+%1*24 + 8] ;B2 ; c = coeff[i] | ||
| 630 : | psubw mm2, mm3 ;-c ;B3 (1st dep) | ||
| 631 : | paddw mm0, mm6 ;A11 mm6 free (4th+) | ||
| 632 : | movq mm6, [byte esi] ;0 ;C5 mm6 in use | ||
| 633 : | pcmpeqw mm6, [ecx+%1*24+16] ;C6 (c ==0) ? -1 : 0 (1st) | ||
| 634 : | pminsw mm0, [ebx] ;A12 saturates to +2047 (5th+) | ||
| 635 : | pmaxsw mm2, mm3 ;|c| ;B4 (2nd) | ||
| 636 : | pxor mm1, mm0 ;A13 (6th+) | ||
| 637 : | pmullw mm4, [edi] ;*= 2Q ;C8 (3rd+) | ||
| 638 : | psraw mm5, 15 ; sign(c) ;C7 (2nd) | ||
| 639 : | movq mm7, [byte esi] ;0 ;B5 mm7 in use | ||
| 640 : | pcmpeqw mm7, [ecx+%1*24 + 8] ;B6 (c ==0) ? -1 : 0 (1st) | ||
| 641 : | edgomez | 851 | %if (%1 < 4) |
| 642 : | edgomez | 1089 | movq mm0, [byte esi] ;A1 ;0 |
| 643 : | edgomez | 851 | %endif |
| 644 : | edgomez | 1089 | pandn mm6, [byte eax] ;C9 offset = isZero ? 0 : quant_add (2nd) |
| 645 : | psraw mm3, 15 ;sign(c) ;B7 (2nd) | ||
| 646 : | movq [byte edx+%1*24], mm1 ;A14 (7th) | ||
| 647 : | paddw mm6, mm5 ;C10 offset +negate back (3rd) | ||
| 648 : | pmullw mm2, [edi] ;*= 2Q ;B8 (3rd+) | ||
| 649 : | mov esp, esp | ||
| 650 : | edgomez | 851 | %endmacro |
| 651 : | |||
| 652 : | |||
| 653 : | align ALIGN | ||
| 654 : | cglobal dequant_intra_3dne | ||
| 655 : | dequant_intra_3dne: | ||
| 656 : | edgomez | 1089 | mov ecx, [esp+ 8] ; coeff |
| 657 : | mov eax, [esp+12] ; quant | ||
| 658 : | pxor mm0, mm0 | ||
| 659 : | pxor mm2, mm2 | ||
| 660 : | push edi | ||
| 661 : | push ebx | ||
| 662 : | lea edi, [mmx_mul + eax*8 - 8] ; 2*quant | ||
| 663 : | push ebp | ||
| 664 : | mov ebx, mmx_2047 | ||
| 665 : | movsx ebp, word [ecx] | ||
| 666 : | lea eax, [mmx_add + eax*8 - 8] ; quant or quant-1 | ||
| 667 : | push esi | ||
| 668 : | mov esi, mmzero | ||
| 669 : | pxor mm7, mm7 | ||
| 670 : | movq mm3, [ecx+120] ;B2 ; c = coeff[i] | ||
| 671 : | pcmpeqw mm7, [ecx+120] ;B6 (c ==0) ? -1 : 0 (1st) | ||
| 672 : | edgomez | 851 | |
| 673 : | edgomez | 1089 | imul ebp, [esp+16+16] ; dcscalar |
| 674 : | psubw mm2, mm3 ;-c ;B3 (1st dep) | ||
| 675 : | pmaxsw mm2, mm3 ;|c| ;B4 (2nd) | ||
| 676 : | pmullw mm2, [edi] ;*= 2Q ;B8 (3rd+) | ||
| 677 : | psraw mm3, 15 ; sign(c) ;B7 (2nd) | ||
| 678 : | mov edx, [esp+ 4+16] ; data | ||
| 679 : | |||
| 680 : | edgomez | 851 | align 8 |
| 681 : | edgomez | 1089 | dequant 0 |
| 682 : | edgomez | 851 | |
| 683 : | edgomez | 1089 | cmp ebp, -2048 |
| 684 : | mov esp, esp | ||
| 685 : | edgomez | 851 | |
| 686 : | edgomez | 1089 | dequant 1 |
| 687 : | |||
| 688 : | cmovl ebp, [int_2048] | ||
| 689 : | nop | ||
| 690 : | |||
| 691 : | dequant 2 | ||
| 692 : | |||
| 693 : | cmp ebp, 2047 | ||
| 694 : | mov esp, esp | ||
| 695 : | |||
| 696 : | dequant 3 | ||
| 697 : | |||
| 698 : | cmovg ebp, [int2047] | ||
| 699 : | nop | ||
| 700 : | |||
| 701 : | dequant 4 | ||
| 702 : | |||
| 703 : | paddw mm4, mm6 ;C11 mm6 free (4th+) | ||
| 704 : | pminsw mm4, [ebx] ;C12 saturates to +2047 (5th+) | ||
| 705 : | pandn mm7, [eax] ;B9 offset = isZero ? 0 : quant_add (2nd) | ||
| 706 : | mov eax, ebp | ||
| 707 : | mov esi, [esp] | ||
| 708 : | mov ebp, [esp+4] | ||
| 709 : | pxor mm5, mm4 ;C13 (6th+) | ||
| 710 : | paddw mm7, mm3 ;B10 offset +negate back (3rd) | ||
| 711 : | movq [edx+4*24+16], mm5 ;C14 (7th) | ||
| 712 : | paddw mm2, mm7 ;B11 mm7 free (4th+) | ||
| 713 : | pminsw mm2, [ebx] ;B12 saturates to +2047 (5th+) | ||
| 714 : | mov ebx, [esp+8] | ||
| 715 : | mov edi, [esp+12] | ||
| 716 : | add esp, byte 16 | ||
| 717 : | pxor mm3, mm2 ;B13 (6th+) | ||
| 718 : | movq [edx+4*24+8], mm3 ;B14 (7th) | ||
| 719 : | mov [edx], ax | ||
| 720 : | ret | ||
| 721 : | |||
| 722 : | edgomez | 851 | ;=========================================================================== |
| 723 : | ; | ||
| 724 : | ; void dequant_inter_3dne(int16_t * data, | ||
| 725 : | ; const int16_t * const coeff, | ||
| 726 : | ; const uint32_t quant); | ||
| 727 : | ; | ||
| 728 : | ;=========================================================================== | ||
| 729 : | |||
| 730 : | edgomez | 1089 | ; this is the same as dequant_inter_3dne, |
| 731 : | ; except that we're saturating using 'pminsw' (saves 2 cycles/loop) | ||
| 732 : | edgomez | 851 | ;This is Athlon-optimized code (ca 100 clk per call) |
| 733 : | ;Optimized by Jaan, 30 Nov 2002 | ||
| 734 : | |||
| 735 : | align ALIGN | ||
| 736 : | cglobal dequant_inter_3dne | ||
| 737 : | dequant_inter_3dne: | ||
| 738 : | edgomez | 1089 | mov ecx, [esp+ 8] ; coeff |
| 739 : | mov eax, [esp+12] ; quant | ||
| 740 : | pxor mm0, mm0 | ||
| 741 : | pxor mm2, mm2 | ||
| 742 : | push edi | ||
| 743 : | push ebx | ||
| 744 : | push esi | ||
| 745 : | lea edi, [mmx_mul + eax*8 - 8] ; 2*quant | ||
| 746 : | mov ebx, mmx_2047 | ||
| 747 : | pxor mm7, mm7 | ||
| 748 : | movq mm3, [ecx+120] ;B2 ; c = coeff[i] | ||
| 749 : | pcmpeqw mm7, [ecx+120] ;B6 (c ==0) ? -1 : 0 (1st) | ||
| 750 : | lea eax, [mmx_add + eax*8 - 8] ; quant or quant-1 | ||
| 751 : | psubw mm2, mm3 ;-c ;B3 (1st dep) | ||
| 752 : | mov esi, mmzero | ||
| 753 : | pmaxsw mm2, mm3 ;|c| ;B4 (2nd) | ||
| 754 : | pmullw mm2, [edi] ;*= 2Q ;B8 (3rd+) | ||
| 755 : | psraw mm3, 15 ; sign(c) ;B7 (2nd) | ||
| 756 : | mov edx, [dword esp+ 4+12] ; data | ||
| 757 : | edgomez | 851 | |
| 758 : | align 8 | ||
| 759 : | |||
| 760 : | edgomez | 1089 | dequant 0 |
| 761 : | dequant 1 | ||
| 762 : | dequant 2 | ||
| 763 : | dequant 3 | ||
| 764 : | dequant 4 | ||
| 765 : | |||
| 766 : | paddw mm4, mm6 ;C11 mm6 free (4th+) | ||
| 767 : | pminsw mm4, [ebx] ;C12 saturates to +2047 (5th+) | ||
| 768 : | pandn mm7, [eax] ;B9 offset = isZero ? 0 : quant_add (2nd) | ||
| 769 : | mov esi, [esp] | ||
| 770 : | pxor mm5, mm4 ;C13 (6th+) | ||
| 771 : | paddw mm7, mm3 ;B10 offset +negate back (3rd) | ||
| 772 : | movq [edx+4*24+16], mm5 ;C14 (7th) | ||
| 773 : | paddw mm2, mm7 ;B11 mm7 free (4th+) | ||
| 774 : | pminsw mm2, [ebx] ;B12 saturates to +2047 (5th+) | ||
| 775 : | mov ebx, [esp+4] | ||
| 776 : | mov edi, [esp+8] | ||
| 777 : | add esp, byte 12 | ||
| 778 : | pxor mm3, mm2 ;B13 (6th+) | ||
| 779 : | movq [edx+4*24+8], mm3 ;B14 (7th) | ||
| 780 : | |||
| 781 : | ret |
| No admin address has been configured | ViewVC Help |
| Powered by ViewVC 1.0.4 |