Parent Directory
| 
Revision Log
Revision 1.1 - (view) (download)
| 1 : | edgomez | 1.1 | ;/**************************************************************************** |
| 2 : | ; * | ||
| 3 : | ; * XVID MPEG-4 VIDEO CODEC | ||
| 4 : | ; * - 3dne Quantization/Dequantization - | ||
| 5 : | ; * | ||
| 6 : | ; * Copyright (C) 2002-2003 Peter Ross <pross@xvid.org> | ||
| 7 : | ; * 2002 Jaan Kalda | ||
| 8 : | ; * 2004 Andre Werthmann <wertmann@aei.mpg.de> | ||
| 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 : | ; * $Id$ | ||
| 25 : | ; * | ||
| 26 : | ; ***************************************************************************/ | ||
| 27 : | |||
| 28 : | ; _3dne functions are compatible with iSSE, but are optimized specifically | ||
| 29 : | ; for K7 pipelines | ||
| 30 : | |||
| 31 : | %define SATURATE | ||
| 32 : | |||
| 33 : | BITS 64 | ||
| 34 : | |||
| 35 : | %macro cglobal 1 | ||
| 36 : | %ifdef PREFIX | ||
| 37 : | %ifdef MARK_FUNCS | ||
| 38 : | global _%1:function %1.endfunc-%1 | ||
| 39 : | %define %1 _%1:function %1.endfunc-%1 | ||
| 40 : | %else | ||
| 41 : | global _%1 | ||
| 42 : | %define %1 _%1 | ||
| 43 : | %endif | ||
| 44 : | %else | ||
| 45 : | %ifdef MARK_FUNCS | ||
| 46 : | global %1:function %1.endfunc-%1 | ||
| 47 : | %else | ||
| 48 : | global %1 | ||
| 49 : | %endif | ||
| 50 : | %endif | ||
| 51 : | %endmacro | ||
| 52 : | |||
| 53 : | ;============================================================================= | ||
| 54 : | ; Local data | ||
| 55 : | ;============================================================================= | ||
| 56 : | |||
| 57 : | %ifdef FORMAT_COFF | ||
| 58 : | SECTION .rodata | ||
| 59 : | %else | ||
| 60 : | SECTION .rodata align=16 | ||
| 61 : | %endif | ||
| 62 : | |||
| 63 : | ALIGN 8 | ||
| 64 : | mmzero: | ||
| 65 : | dd 0,0 | ||
| 66 : | mmx_one: | ||
| 67 : | times 4 dw 1 | ||
| 68 : | |||
| 69 : | ;----------------------------------------------------------------------------- | ||
| 70 : | ; divide by 2Q table | ||
| 71 : | ;----------------------------------------------------------------------------- | ||
| 72 : | |||
| 73 : | ALIGN 16 | ||
| 74 : | mmx_divs: ;i>2 | ||
| 75 : | %assign i 1 | ||
| 76 : | %rep 31 | ||
| 77 : | times 4 dw ((1 << 15) / i + 1) | ||
| 78 : | %assign i i+1 | ||
| 79 : | %endrep | ||
| 80 : | |||
| 81 : | ALIGN 16 | ||
| 82 : | mmx_div: ;quant>2 | ||
| 83 : | times 4 dw 65535 ; the div by 2 formula will overflow for the case | ||
| 84 : | ; quant=1 but we don't care much because quant=1 | ||
| 85 : | ; is handled by a different piece of code that | ||
| 86 : | ; doesn't use this table. | ||
| 87 : | %assign quant 2 | ||
| 88 : | %rep 31 | ||
| 89 : | times 4 dw ((1 << 16) / quant + 1) | ||
| 90 : | %assign quant quant+1 | ||
| 91 : | %endrep | ||
| 92 : | |||
| 93 : | %macro FIXX 1 | ||
| 94 : | dw (1 << 16) / (%1) + 1 | ||
| 95 : | %endmacro | ||
| 96 : | |||
| 97 : | %define nop4 db 08Dh, 074h, 026h,0 | ||
| 98 : | %define nop3 add esp, byte 0 | ||
| 99 : | %define nop2 mov esp, esp | ||
| 100 : | %define nop7 db 08dh, 02ch, 02dh,0,0,0,0 | ||
| 101 : | %define nop6 add ebp, dword 0 | ||
| 102 : | |||
| 103 : | ;----------------------------------------------------------------------------- | ||
| 104 : | ; quantd table | ||
| 105 : | ;----------------------------------------------------------------------------- | ||
| 106 : | |||
| 107 : | %define VM18P 3 | ||
| 108 : | %define VM18Q 4 | ||
| 109 : | |||
| 110 : | ALIGN 16 | ||
| 111 : | quantd: | ||
| 112 : | %assign i 1 | ||
| 113 : | %rep 31 | ||
| 114 : | times 4 dw (((VM18P*i) + (VM18Q/2)) / VM18Q) | ||
| 115 : | %assign i i+1 | ||
| 116 : | %endrep | ||
| 117 : | |||
| 118 : | ;----------------------------------------------------------------------------- | ||
| 119 : | ; multiple by 2Q table | ||
| 120 : | ;----------------------------------------------------------------------------- | ||
| 121 : | |||
| 122 : | ALIGN 16 | ||
| 123 : | mmx_mul_quant: | ||
| 124 : | %assign i 1 | ||
| 125 : | %rep 31 | ||
| 126 : | times 4 dw i | ||
| 127 : | %assign i i+1 | ||
| 128 : | %endrep | ||
| 129 : | |||
| 130 : | ;----------------------------------------------------------------------------- | ||
| 131 : | ; saturation limits | ||
| 132 : | ;----------------------------------------------------------------------------- | ||
| 133 : | |||
| 134 : | ALIGN 16 | ||
| 135 : | mmx_32767_minus_2047: | ||
| 136 : | times 4 dw (32767-2047) | ||
| 137 : | mmx_32768_minus_2048: | ||
| 138 : | times 4 dw (32768-2048) | ||
| 139 : | mmx_2047: | ||
| 140 : | times 4 dw 2047 | ||
| 141 : | mmx_minus_2048: | ||
| 142 : | times 4 dw (-2048) | ||
| 143 : | zero: | ||
| 144 : | times 4 dw 0 | ||
| 145 : | |||
| 146 : | int_div: | ||
| 147 : | dd 0 | ||
| 148 : | %assign i 1 | ||
| 149 : | %rep 255 | ||
| 150 : | dd (1 << 17) / ( i) + 1 | ||
| 151 : | %assign i i+1 | ||
| 152 : | %endrep | ||
| 153 : | |||
| 154 : | ;============================================================================= | ||
| 155 : | ; Code | ||
| 156 : | ;============================================================================= | ||
| 157 : | |||
| 158 : | SECTION .text align=16 | ||
| 159 : | |||
| 160 : | cglobal quant_mpeg_intra_x86_64 | ||
| 161 : | cglobal quant_mpeg_inter_x86_64 | ||
| 162 : | cglobal dequant_mpeg_intra_x86_64 | ||
| 163 : | cglobal dequant_mpeg_inter_x86_64 | ||
| 164 : | |||
| 165 : | ;----------------------------------------------------------------------------- | ||
| 166 : | ; | ||
| 167 : | ; uint32_t quant_mpeg_intra_x86_64(int16_t * coeff, | ||
| 168 : | ; const int16_t const * data, | ||
| 169 : | ; const uint32_t quant, | ||
| 170 : | ; const uint32_t dcscalar, | ||
| 171 : | ; const uint16_t *mpeg_matrices); | ||
| 172 : | ; Ported from its 32bit xmm cousin | ||
| 173 : | ;----------------------------------------------------------------------------- | ||
| 174 : | |||
| 175 : | ALIGN 16 | ||
| 176 : | quant_mpeg_intra_x86_64: | ||
| 177 : | mov rax, rsi ; data | ||
| 178 : | mov r9, rcx ; save dcscalar | ||
| 179 : | mov rcx, rdx ; quant | ||
| 180 : | mov rdx, rdi ; coeff | ||
| 181 : | |||
| 182 : | push rbx | ||
| 183 : | |||
| 184 : | mov rdi, r8 ; mpeg_quant_matrices | ||
| 185 : | |||
| 186 : | mov rsi, -14 | ||
| 187 : | pxor mm0, mm0 | ||
| 188 : | pxor mm3, mm3 | ||
| 189 : | cmp rcx, byte 1 | ||
| 190 : | je near .q1loop | ||
| 191 : | cmp rcx, byte 19 | ||
| 192 : | jg near .lloop | ||
| 193 : | |||
| 194 : | ALIGN 16 | ||
| 195 : | .loop | ||
| 196 : | movq mm1, [rax + 8*rsi+112] ; mm0 = [1st] | ||
| 197 : | psubw mm0, mm1 ;-mm1 | ||
| 198 : | movq mm4, [rax + 8*rsi + 120] ; | ||
| 199 : | psubw mm3, mm4 ;-mm4 | ||
| 200 : | pmaxsw mm0, mm1 ;|src| | ||
| 201 : | pmaxsw mm3,mm4 | ||
| 202 : | ; nop2 | ||
| 203 : | psraw mm1, 15 ;sign src | ||
| 204 : | psraw mm4, 15 | ||
| 205 : | psllw mm0, 4 ;level << 4 ; | ||
| 206 : | psllw mm3, 4 | ||
| 207 : | paddw mm0, [rdi + 128 + 8*rsi+112] | ||
| 208 : | paddw mm3, [rdi + 128 + 8*rsi+120] | ||
| 209 : | movq mm5, [rdi + 384 + 8*rsi+112] | ||
| 210 : | movq mm7, [rdi + 384 + 8*rsi+120] | ||
| 211 : | pmulhuw mm5, mm0 | ||
| 212 : | pmulhuw mm7, mm3 | ||
| 213 : | ; mov esp, esp | ||
| 214 : | movq mm2, [rdi + 8*rsi+112] | ||
| 215 : | movq mm6, [rdi + 8*rsi+120] | ||
| 216 : | pmullw mm2, mm5 | ||
| 217 : | pmullw mm6, mm7 | ||
| 218 : | psubw mm0, mm2 | ||
| 219 : | psubw mm3, mm6 | ||
| 220 : | ; nop4 | ||
| 221 : | lea r11, [quantd wrt rip] | ||
| 222 : | movq mm2, [r11 + rcx * 8 - 8] | ||
| 223 : | lea r11, [mmx_divs wrt rip] | ||
| 224 : | movq mm6, [r11 + rcx * 8 - 8] | ||
| 225 : | paddw mm5, mm2 | ||
| 226 : | paddw mm7, mm2 | ||
| 227 : | ; mov esp, esp | ||
| 228 : | pmulhuw mm0, [rdi + 256 + 8*rsi+112] | ||
| 229 : | pmulhuw mm3, [rdi + 256 + 8*rsi+120] | ||
| 230 : | paddw mm5, mm0 | ||
| 231 : | paddw mm7, mm3 | ||
| 232 : | pxor mm0, mm0 | ||
| 233 : | pxor mm3, mm3 | ||
| 234 : | pmulhuw mm5, mm6 ; mm0 = (mm0 / 2Q) >> 16 | ||
| 235 : | pmulhuw mm7, mm6 ; (level + quantd) / quant (0<quant<32) | ||
| 236 : | pxor mm5, mm1 ; mm0 *= sign(mm0) | ||
| 237 : | pxor mm7, mm4 ; | ||
| 238 : | psubw mm5, mm1 ; undisplace | ||
| 239 : | psubw mm7, mm4 ; | ||
| 240 : | movq [rdx + 8*rsi+112], mm5 | ||
| 241 : | movq [rdx + 8*rsi +120], mm7 | ||
| 242 : | add rsi, byte 2 | ||
| 243 : | jng near .loop | ||
| 244 : | |||
| 245 : | .done | ||
| 246 : | ; calculate data[0] // (int32_t)dcscalar) | ||
| 247 : | ; mov esi, [esp + 12 + 16] ; dcscalar | ||
| 248 : | mov rsi, r9 ; dcscalar | ||
| 249 : | movsx rcx, word [rax] | ||
| 250 : | mov rdi, rcx | ||
| 251 : | ; mov edx, [esp + 12 + 16] | ||
| 252 : | mov r11, rdx ; save rdx | ||
| 253 : | mov rdx, r9 ; | ||
| 254 : | shr edx, 1 ; ebx = dcscalar /2 | ||
| 255 : | sar edi, 31 ; cdq is vectorpath | ||
| 256 : | xor edx, edi ; ebx = eax V -eax -1 | ||
| 257 : | sub ecx, edi | ||
| 258 : | add ecx, edx | ||
| 259 : | ;; mov rdx, [dword esp + 12 + 4] | ||
| 260 : | mov rdx, r11 ; restore rdx | ||
| 261 : | lea r11, [int_div wrt rip] | ||
| 262 : | mov rsi, [r11+4*rsi] | ||
| 263 : | imul ecx, esi | ||
| 264 : | sar ecx, 17 | ||
| 265 : | lea rbx, [byte rcx + 1] | ||
| 266 : | cmovs rcx, rbx | ||
| 267 : | ; idiv cx ; ecx = edi:ecx / dcscalar | ||
| 268 : | |||
| 269 : | ; mov ebx, [esp] | ||
| 270 : | ; mov edi, [esp+4] | ||
| 271 : | ; mov esi, [esp+8] | ||
| 272 : | ; add esp, byte 12 ; pops... | ||
| 273 : | pop rbx | ||
| 274 : | ; mov [rdx], rcx ; coeff[0] = ax | ||
| 275 : | mov [rdx], cx ; coeff[0] = cx | ||
| 276 : | |||
| 277 : | xor rax, rax | ||
| 278 : | ret | ||
| 279 : | |||
| 280 : | ALIGN 16 | ||
| 281 : | .q1loop | ||
| 282 : | movq mm1, [rax + 8*rsi+112] ; mm0 = [1st] | ||
| 283 : | psubw mm0, mm1 ;-mm1 | ||
| 284 : | movq mm4, [rax + 8*rsi+120] ; | ||
| 285 : | psubw mm3, mm4 ;-mm4 | ||
| 286 : | pmaxsw mm0, mm1 ;|src| | ||
| 287 : | pmaxsw mm3, mm4 | ||
| 288 : | ; nop2 | ||
| 289 : | psraw mm1, 15 ;sign src | ||
| 290 : | psraw mm4, 15 | ||
| 291 : | psllw mm0, 4 ; level << 4 | ||
| 292 : | psllw mm3, 4 | ||
| 293 : | paddw mm0, [rdi + 128 + 8*rsi+112] ;mm0 is to be divided | ||
| 294 : | paddw mm3, [rdi + 128 + 8*rsi+120] ;intra1 contains fix for division by 1 | ||
| 295 : | movq mm5, [rdi + 384 + 8*rsi+112] ;with rounding down | ||
| 296 : | movq mm7, [rdi + 384 + 8*rsi+120] | ||
| 297 : | pmulhuw mm5, mm0 | ||
| 298 : | pmulhuw mm7, mm3 ;mm7: first approx of division | ||
| 299 : | ; mov esp, esp | ||
| 300 : | movq mm2, [rdi + 8*rsi+112] | ||
| 301 : | movq mm6, [rdi + 8*rsi+120] ; divs for q<=16 | ||
| 302 : | pmullw mm2, mm5 ;test value <= original | ||
| 303 : | pmullw mm6, mm7 | ||
| 304 : | psubw mm0, mm2 ;mismatch | ||
| 305 : | psubw mm3, mm6 | ||
| 306 : | ; nop4 | ||
| 307 : | lea r11, [quantd wrt rip] | ||
| 308 : | movq mm2, [r11 + rcx * 8 - 8] | ||
| 309 : | paddw mm5, mm2 ;first approx with quantd | ||
| 310 : | paddw mm7, mm2 | ||
| 311 : | ; mov esp, esp | ||
| 312 : | pmulhuw mm0, [rdi + 256 + 8*rsi+112] ;correction | ||
| 313 : | pmulhuw mm3, [rdi + 256 + 8*rsi+120] | ||
| 314 : | paddw mm5, mm0 ;final result with quantd | ||
| 315 : | paddw mm7, mm3 | ||
| 316 : | pxor mm0, mm0 | ||
| 317 : | pxor mm3, mm3 | ||
| 318 : | ; mov esp, esp | ||
| 319 : | psrlw mm5, 1 ; (level + quantd) /2 (quant = 1) | ||
| 320 : | psrlw mm7, 1 | ||
| 321 : | pxor mm5, mm1 ; mm0 *= sign(mm0) | ||
| 322 : | pxor mm7, mm4 ; | ||
| 323 : | psubw mm5, mm1 ; undisplace | ||
| 324 : | psubw mm7, mm4 ; | ||
| 325 : | movq [rdx + 8*rsi+112], mm5 | ||
| 326 : | movq [rdx + 8*rsi +120], mm7 | ||
| 327 : | add rsi, byte 2 | ||
| 328 : | jng near .q1loop | ||
| 329 : | jmp near .done | ||
| 330 : | |||
| 331 : | ALIGN 8 | ||
| 332 : | .lloop | ||
| 333 : | movq mm1, [rax + 8*rsi+112] ; mm0 = [1st] | ||
| 334 : | psubw mm0, mm1 ;-mm1 | ||
| 335 : | movq mm4, [rax + 8*rsi+120] | ||
| 336 : | psubw mm3, mm4 ;-mm4 | ||
| 337 : | pmaxsw mm0, mm1 ;|src| | ||
| 338 : | pmaxsw mm3, mm4 | ||
| 339 : | ; nop2 | ||
| 340 : | psraw mm1, 15 ;sign src | ||
| 341 : | psraw mm4, 15 | ||
| 342 : | psllw mm0, 4 ; level << 4 | ||
| 343 : | psllw mm3, 4 ; | ||
| 344 : | paddw mm0, [rdi + 128 + 8*rsi+112] ;mm0 is to be divided intra1 contains fix for division by 1 | ||
| 345 : | paddw mm3, [rdi + 128 + 8*rsi+120] | ||
| 346 : | movq mm5, [rdi + 384 + 8*rsi+112] | ||
| 347 : | movq mm7, [rdi + 384 + 8*rsi+120] | ||
| 348 : | pmulhuw mm5, mm0 | ||
| 349 : | pmulhuw mm7, mm3 ;mm7: first approx of division | ||
| 350 : | ; mov esp, esp | ||
| 351 : | movq mm2, [rdi + 8*rsi+112] | ||
| 352 : | movq mm6, [rdi + 8*rsi+120] | ||
| 353 : | pmullw mm2, mm5 ;test value <= original | ||
| 354 : | pmullw mm6, mm7 | ||
| 355 : | psubw mm0, mm2 ;mismatch | ||
| 356 : | psubw mm3, mm6 | ||
| 357 : | ; nop4 | ||
| 358 : | lea r11, [quantd wrt rip] | ||
| 359 : | movq mm2, [r11 + rcx * 8 - 8] | ||
| 360 : | lea r11, [mmx_div wrt rip] | ||
| 361 : | movq mm6, [r11 + rcx * 8 - 8] ; divs for q<=16 | ||
| 362 : | paddw mm5, mm2 ;first approx with quantd | ||
| 363 : | paddw mm7, mm2 | ||
| 364 : | ; mov esp, esp | ||
| 365 : | pmulhuw mm0, [rdi + 256 + 8*rsi+112] ;correction | ||
| 366 : | pmulhuw mm3, [rdi + 256 + 8*rsi+120] | ||
| 367 : | paddw mm5, mm0 ;final result with quantd | ||
| 368 : | paddw mm7, mm3 | ||
| 369 : | pxor mm0, mm0 | ||
| 370 : | pxor mm3, mm3 | ||
| 371 : | ; mov esp, esp | ||
| 372 : | pmulhuw mm5, mm6 ; mm0 = (mm0 / 2Q) >> 16 | ||
| 373 : | pmulhuw mm7, mm6 ; (level + quantd) / quant (0<quant<32) | ||
| 374 : | psrlw mm5, 1 ; (level + quantd) / (2*quant) | ||
| 375 : | psrlw mm7, 1 | ||
| 376 : | pxor mm5, mm1 ; mm0 *= sign(mm0) | ||
| 377 : | pxor mm7, mm4 ; | ||
| 378 : | psubw mm5, mm1 ; undisplace | ||
| 379 : | psubw mm7, mm4 ; | ||
| 380 : | movq [rdx + 8*rsi+112], mm5 | ||
| 381 : | movq [rdx + 8*rsi +120], mm7 | ||
| 382 : | add rsi,byte 2 | ||
| 383 : | jng near .lloop | ||
| 384 : | jmp near .done | ||
| 385 : | .endfunc | ||
| 386 : | |||
| 387 : | ;----------------------------------------------------------------------------- | ||
| 388 : | ; | ||
| 389 : | ; uint32_t quant_mpeg_inter_x86_64(int16_t * coeff, | ||
| 390 : | ; const int16_t const * data, | ||
| 391 : | ; const uint32_t quant, | ||
| 392 : | ; const uint16_t *mpeg_matrices); | ||
| 393 : | ; Ported from its 32bit xmm cousin | ||
| 394 : | ;----------------------------------------------------------------------------- | ||
| 395 : | |||
| 396 : | ALIGN 16 | ||
| 397 : | quant_mpeg_inter_x86_64: | ||
| 398 : | mov rax, rsi ; data | ||
| 399 : | mov r8, rdi ; save coeff | ||
| 400 : | mov rdi, rcx ; mpeg_matrices | ||
| 401 : | mov rcx, rdx ; quant | ||
| 402 : | mov rdx, r8 ; coeff | ||
| 403 : | |||
| 404 : | push rbx | ||
| 405 : | |||
| 406 : | mov rsi, -14 | ||
| 407 : | mov rbx, rsp | ||
| 408 : | sub rsp, byte 24 ; 16 would be enough, but it isn't important | ||
| 409 : | lea rbx, [rsp+8] | ||
| 410 : | and rbx, byte -8 ;ALIGN 8 | ||
| 411 : | pxor mm0, mm0 | ||
| 412 : | pxor mm3, mm3 | ||
| 413 : | movq [byte rbx],mm0 | ||
| 414 : | movq [rbx+8],mm0 | ||
| 415 : | cmp rcx, byte 1 | ||
| 416 : | je near .q1loop | ||
| 417 : | cmp rcx, byte 19 | ||
| 418 : | jg near .lloop | ||
| 419 : | |||
| 420 : | ALIGN 16 | ||
| 421 : | .loop | ||
| 422 : | movq mm1, [rax + 8*rsi+112] ; mm0 = [1st] | ||
| 423 : | psubw mm0, mm1 ;-mm1 | ||
| 424 : | movq mm4, [rax + 8*rsi + 120] ; | ||
| 425 : | psubw mm3, mm4 ;-mm4 | ||
| 426 : | pmaxsw mm0, mm1 ;|src| | ||
| 427 : | pmaxsw mm3, mm4 | ||
| 428 : | ; nop2 | ||
| 429 : | psraw mm1, 15 ;sign src | ||
| 430 : | psraw mm4, 15 | ||
| 431 : | psllw mm0, 4 ; level << 4 | ||
| 432 : | psllw mm3, 4 ; | ||
| 433 : | paddw mm0, [rdi + 640 + 8*rsi+112] | ||
| 434 : | paddw mm3, [rdi + 640 + 8*rsi+120] | ||
| 435 : | movq mm5, [rdi + 896 + 8*rsi+112] | ||
| 436 : | movq mm7, [rdi + 896 + 8*rsi+120] | ||
| 437 : | pmulhuw mm5, mm0 | ||
| 438 : | pmulhuw mm7, mm3 | ||
| 439 : | ; mov esp, esp | ||
| 440 : | movq mm2, [rdi + 512 + 8*rsi+112] | ||
| 441 : | movq mm6, [rdi + 512 + 8*rsi+120] | ||
| 442 : | pmullw mm2, mm5 | ||
| 443 : | pmullw mm6, mm7 | ||
| 444 : | psubw mm0, mm2 | ||
| 445 : | psubw mm3, mm6 | ||
| 446 : | movq mm2, [byte rbx] | ||
| 447 : | lea r11, [mmx_divs wrt rip] | ||
| 448 : | movq mm6, [r11 + rcx * 8 - 8] | ||
| 449 : | pmulhuw mm0, [rdi + 768 + 8*rsi+112] | ||
| 450 : | pmulhuw mm3, [rdi + 768 + 8*rsi+120] | ||
| 451 : | paddw mm2, [rbx+8] ;sum | ||
| 452 : | paddw mm5, mm0 | ||
| 453 : | paddw mm7, mm3 | ||
| 454 : | pxor mm0, mm0 | ||
| 455 : | pxor mm3, mm3 | ||
| 456 : | pmulhuw mm5, mm6 ; mm0 = (mm0 / 2Q) >> 16 | ||
| 457 : | pmulhuw mm7, mm6 ; (level ) / quant (0<quant<32) | ||
| 458 : | add rsi, byte 2 | ||
| 459 : | paddw mm2, mm5 ;sum += x1 | ||
| 460 : | movq [rbx], mm7 ;store x2 | ||
| 461 : | pxor mm5, mm1 ; mm0 *= sign(mm0) | ||
| 462 : | pxor mm7, mm4 ; | ||
| 463 : | psubw mm5, mm1 ; undisplace | ||
| 464 : | psubw mm7, mm4 ; | ||
| 465 : | ; db 0Fh, 7Fh, 54h, 23h, 08 ;movq [ebx+8],mm2 ;store sum | ||
| 466 : | movq [rbx+8], mm2 ;store sum | ||
| 467 : | movq [rdx + 8*rsi+112-16], mm5 | ||
| 468 : | movq [rdx + 8*rsi +120-16], mm7 | ||
| 469 : | jng near .loop | ||
| 470 : | |||
| 471 : | .done | ||
| 472 : | ; calculate data[0] // (int32_t)dcscalar) | ||
| 473 : | paddw mm2, [rbx] | ||
| 474 : | add rsp, byte 24 | ||
| 475 : | pop rbx | ||
| 476 : | pmaddwd mm2, [mmx_one wrt rip] | ||
| 477 : | punpckldq mm0, mm2 ;get low dw to mm0:high | ||
| 478 : | paddd mm0,mm2 | ||
| 479 : | punpckhdq mm0, mm0 ;get result to low | ||
| 480 : | movd rax, mm0 | ||
| 481 : | |||
| 482 : | ret | ||
| 483 : | |||
| 484 : | ALIGN 16 | ||
| 485 : | .q1loop | ||
| 486 : | movq mm1, [rax + 8*rsi+112] ; mm0 = [1st] | ||
| 487 : | psubw mm0, mm1 ;-mm1 | ||
| 488 : | movq mm4, [rax + 8*rsi+120] | ||
| 489 : | psubw mm3, mm4 ;-mm4 | ||
| 490 : | pmaxsw mm0, mm1 ;|src| | ||
| 491 : | pmaxsw mm3, mm4 | ||
| 492 : | ; nop2 | ||
| 493 : | psraw mm1, 15 ; sign src | ||
| 494 : | psraw mm4, 15 | ||
| 495 : | psllw mm0, 4 ; level << 4 | ||
| 496 : | psllw mm3, 4 | ||
| 497 : | paddw mm0, [rdi + 640 + 8*rsi+112] ;mm0 is to be divided | ||
| 498 : | paddw mm3, [rdi + 640 + 8*rsi+120] ; inter1 contains fix for division by 1 | ||
| 499 : | movq mm5, [rdi + 896 + 8*rsi+112] ;with rounding down | ||
| 500 : | movq mm7, [rdi + 896 + 8*rsi+120] | ||
| 501 : | pmulhuw mm5, mm0 | ||
| 502 : | pmulhuw mm7, mm3 ;mm7: first approx of division | ||
| 503 : | ; mov esp, esp | ||
| 504 : | movq mm2, [rdi + 512 + 8*rsi+112] | ||
| 505 : | movq mm6, [rdi + 512 + 8*rsi+120] ; divs for q<=16 | ||
| 506 : | pmullw mm2, mm5 ;test value <= original | ||
| 507 : | pmullw mm6, mm7 | ||
| 508 : | psubw mm0, mm2 ;mismatch | ||
| 509 : | psubw mm3, mm6 | ||
| 510 : | movq mm2, [byte rbx] | ||
| 511 : | pmulhuw mm0, [rdi + 768 + 8*rsi+112] ;correction | ||
| 512 : | pmulhuw mm3, [rdi + 768 + 8*rsi+120] | ||
| 513 : | paddw mm2, [rbx+8] ;sum | ||
| 514 : | paddw mm5, mm0 ;final result | ||
| 515 : | paddw mm7, mm3 | ||
| 516 : | pxor mm0, mm0 | ||
| 517 : | pxor mm3, mm3 | ||
| 518 : | psrlw mm5, 1 ; (level ) /2 (quant = 1) | ||
| 519 : | psrlw mm7, 1 | ||
| 520 : | add rsi, byte 2 | ||
| 521 : | paddw mm2, mm5 ;sum += x1 | ||
| 522 : | movq [rbx], mm7 ;store x2 | ||
| 523 : | pxor mm5, mm1 ; mm0 *= sign(mm0) | ||
| 524 : | pxor mm7, mm4 ; | ||
| 525 : | psubw mm5, mm1 ; undisplace | ||
| 526 : | psubw mm7, mm4 ; | ||
| 527 : | movq [rbx+8], mm2 ;store sum | ||
| 528 : | movq [rdx + 8*rsi+112-16], mm5 | ||
| 529 : | movq [rdx + 8*rsi +120-16], mm7 | ||
| 530 : | jng near .q1loop | ||
| 531 : | jmp near .done | ||
| 532 : | |||
| 533 : | ALIGN 8 | ||
| 534 : | .lloop | ||
| 535 : | movq mm1, [rax + 8*rsi+112] ; mm0 = [1st] | ||
| 536 : | psubw mm0,mm1 ;-mm1 | ||
| 537 : | movq mm4, [rax + 8*rsi+120] | ||
| 538 : | psubw mm3,mm4 ;-mm4 | ||
| 539 : | pmaxsw mm0,mm1 ;|src| | ||
| 540 : | pmaxsw mm3,mm4 | ||
| 541 : | ; nop2 | ||
| 542 : | psraw mm1,15 ;sign src | ||
| 543 : | psraw mm4,15 | ||
| 544 : | psllw mm0, 4 ; level << 4 | ||
| 545 : | psllw mm3, 4 ; | ||
| 546 : | paddw mm0, [rdi + 640 + 8*rsi+112] ;mm0 is to be divided inter1 contains fix for division by 1 | ||
| 547 : | paddw mm3, [rdi + 640 + 8*rsi+120] | ||
| 548 : | movq mm5,[rdi + 896 + 8*rsi+112] | ||
| 549 : | movq mm7,[rdi + 896 + 8*rsi+120] | ||
| 550 : | pmulhuw mm5,mm0 | ||
| 551 : | pmulhuw mm7,mm3 ;mm7: first approx of division | ||
| 552 : | ; mov esp,esp | ||
| 553 : | movq mm2,[rdi + 512 + 8*rsi+112] | ||
| 554 : | movq mm6,[rdi + 512 + 8*rsi+120] | ||
| 555 : | pmullw mm2,mm5 ;test value <= original | ||
| 556 : | pmullw mm6,mm7 | ||
| 557 : | psubw mm0,mm2 ;mismatch | ||
| 558 : | psubw mm3,mm6 | ||
| 559 : | movq mm2,[byte rbx] | ||
| 560 : | lea r11, [mmx_div wrt rip] | ||
| 561 : | movq mm6,[r11 + rcx * 8 - 8] ; divs for q<=16 | ||
| 562 : | pmulhuw mm0,[rdi + 768 + 8*rsi+112] ;correction | ||
| 563 : | pmulhuw mm3,[rdi + 768 + 8*rsi+120] | ||
| 564 : | paddw mm2,[rbx+8] ;sum | ||
| 565 : | paddw mm5,mm0 ;final result | ||
| 566 : | paddw mm7,mm3 | ||
| 567 : | pxor mm0,mm0 | ||
| 568 : | pxor mm3,mm3 | ||
| 569 : | pmulhuw mm5, mm6 ; mm0 = (mm0 / 2Q) >> 16 | ||
| 570 : | pmulhuw mm7, mm6 ; (level ) / quant (0<quant<32) | ||
| 571 : | add rsi,byte 2 | ||
| 572 : | psrlw mm5, 1 ; (level ) / (2*quant) | ||
| 573 : | paddw mm2,mm5 ;sum += x1 | ||
| 574 : | psrlw mm7, 1 | ||
| 575 : | movq [rbx],mm7 ;store x2 | ||
| 576 : | pxor mm5, mm1 ; mm0 *= sign(mm0) | ||
| 577 : | pxor mm7, mm4 ; | ||
| 578 : | psubw mm5, mm1 ; undisplace | ||
| 579 : | psubw mm7, mm4 ; | ||
| 580 : | ; db 0Fh, 7Fh, 54h, 23h, 08 ;movq [ebx+8],mm2 ;store sum | ||
| 581 : | movq [rbx+8], mm2 ;store sum | ||
| 582 : | movq [rdx + 8*rsi+112-16], mm5 | ||
| 583 : | movq [rdx + 8*rsi +120-16], mm7 | ||
| 584 : | jng near .lloop | ||
| 585 : | jmp near .done | ||
| 586 : | .endfunc | ||
| 587 : | |||
| 588 : | ;----------------------------------------------------------------------------- | ||
| 589 : | ; | ||
| 590 : | ; uint32_t dequant_mpeg_intra_x86_64(int16_t *data, | ||
| 591 : | ; const int16_t const *coeff, | ||
| 592 : | ; const uint32_t quant, | ||
| 593 : | ; const uint32_t dcscalar, | ||
| 594 : | ; const uint16_t *mpeg_matrices); | ||
| 595 : | ; Ported from the 32bit 3dne cousin | ||
| 596 : | ;----------------------------------------------------------------------------- | ||
| 597 : | |||
| 598 : | ; Note: in order to saturate 'easily', we pre-shift the quantifier | ||
| 599 : | ; by 4. Then, the high-word of (coeff[]*matrix[i]*quant) are used to | ||
| 600 : | ; build a saturating mask. It is non-zero only when an overflow occured. | ||
| 601 : | ; We thus avoid packing/unpacking toward double-word. | ||
| 602 : | ; Moreover, we perform the mult (matrix[i]*quant) first, instead of, e.g., | ||
| 603 : | ; (coeff[i]*matrix[i]). This is less prone to overflow if coeff[] are not | ||
| 604 : | ; checked. Input ranges are: coeff in [-127,127], inter_matrix in [1..255],a | ||
| 605 : | ; and quant in [1..31]. | ||
| 606 : | ; | ||
| 607 : | |||
| 608 : | %macro DEQUANT4INTRAMMX 1 | ||
| 609 : | movq mm1, [byte rcx+ 16 * %1] ; mm0 = c = coeff[i] | ||
| 610 : | movq mm4, [rcx+ 16 * %1 +8] ; mm3 = c' = coeff[i+1] | ||
| 611 : | psubw mm0, mm1 | ||
| 612 : | psubw mm3, mm4 | ||
| 613 : | pmaxsw mm0, mm1 | ||
| 614 : | pmaxsw mm3, mm4 | ||
| 615 : | psraw mm1, 15 | ||
| 616 : | psraw mm4, 15 | ||
| 617 : | %if %1 | ||
| 618 : | movq mm2, [rsp-16] | ||
| 619 : | movq mm7, [rsp-16] | ||
| 620 : | %endif | ||
| 621 : | pmullw mm2, [rdi + 16 * %1 ] ; matrix[i]*quant | ||
| 622 : | pmullw mm7, [rdi + 16 * %1 +8] ; matrix[i+1]*quant | ||
| 623 : | movq mm5, mm0 | ||
| 624 : | movq mm6, mm3 | ||
| 625 : | pmulhw mm0, mm2 ; high of coeff*(matrix*quant) | ||
| 626 : | pmulhw mm3, mm7 ; high of coeff*(matrix*quant) | ||
| 627 : | pmullw mm2, mm5 ; low of coeff*(matrix*quant) | ||
| 628 : | pmullw mm7, mm6 ; low of coeff*(matrix*quant) | ||
| 629 : | pcmpgtw mm0, [rsp-8] | ||
| 630 : | pcmpgtw mm3, [rsp-8] | ||
| 631 : | paddusw mm2, mm0 | ||
| 632 : | paddusw mm7, mm3 | ||
| 633 : | psrlw mm2, 5 | ||
| 634 : | psrlw mm7, 5 | ||
| 635 : | pxor mm2, mm1 ; start negating back | ||
| 636 : | pxor mm7, mm4 ; start negating back | ||
| 637 : | psubusw mm1, mm0 | ||
| 638 : | psubusw mm4, mm3 | ||
| 639 : | movq mm0, [rsp-8] | ||
| 640 : | movq mm3, [rsp-8] | ||
| 641 : | psubw mm2, mm1 ; finish negating back | ||
| 642 : | psubw mm7, mm4 ; finish negating back | ||
| 643 : | movq [byte rdx + 16 * %1], mm2 ; data[i] | ||
| 644 : | movq [rdx + 16 * %1 +8], mm7 ; data[i+1] | ||
| 645 : | %endmacro | ||
| 646 : | |||
| 647 : | ALIGN 16 | ||
| 648 : | dequant_mpeg_intra_x86_64: | ||
| 649 : | mov rax, rdx ; quant | ||
| 650 : | mov rdx, rdi ; data | ||
| 651 : | mov r9, rcx ; dcscalar | ||
| 652 : | mov rcx, rsi ; coeff | ||
| 653 : | |||
| 654 : | lea r11, [mmx_mul_quant wrt rip] | ||
| 655 : | movq mm7, [r11 + rax*8 - 8] | ||
| 656 : | psllw mm7, 2 ; << 2. See comment. | ||
| 657 : | push rbx | ||
| 658 : | |||
| 659 : | movsx ebx, word [rcx] | ||
| 660 : | pxor mm0, mm0 | ||
| 661 : | pxor mm3, mm3 | ||
| 662 : | |||
| 663 : | movq [rsp-8], mm0 | ||
| 664 : | movq [rsp-16], mm7 | ||
| 665 : | |||
| 666 : | imul ebx, r9d | ||
| 667 : | movq mm2, mm7 | ||
| 668 : | mov rdi, r8 ; mpeg_quant_matrices | ||
| 669 : | ALIGN 4 | ||
| 670 : | |||
| 671 : | DEQUANT4INTRAMMX 0 | ||
| 672 : | |||
| 673 : | mov esi, -2048 | ||
| 674 : | ; nop | ||
| 675 : | cmp ebx, esi | ||
| 676 : | |||
| 677 : | DEQUANT4INTRAMMX 1 | ||
| 678 : | |||
| 679 : | cmovl ebx, esi | ||
| 680 : | neg esi | ||
| 681 : | sub esi, byte 1 ;2047 | ||
| 682 : | |||
| 683 : | DEQUANT4INTRAMMX 2 | ||
| 684 : | |||
| 685 : | cmp ebx, esi | ||
| 686 : | cmovg ebx, esi | ||
| 687 : | |||
| 688 : | DEQUANT4INTRAMMX 3 | ||
| 689 : | |||
| 690 : | mov [byte rdx], bx | ||
| 691 : | |||
| 692 : | DEQUANT4INTRAMMX 4 | ||
| 693 : | DEQUANT4INTRAMMX 5 | ||
| 694 : | DEQUANT4INTRAMMX 6 | ||
| 695 : | DEQUANT4INTRAMMX 7 | ||
| 696 : | |||
| 697 : | pop rbx | ||
| 698 : | |||
| 699 : | xor rax, rax | ||
| 700 : | ret | ||
| 701 : | .endfunc | ||
| 702 : | |||
| 703 : | ;----------------------------------------------------------------------------- | ||
| 704 : | ; | ||
| 705 : | ; uint32_t dequant_mpeg_inter_3dne(int16_t * data, | ||
| 706 : | ; const int16_t * const coeff, | ||
| 707 : | ; const uint32_t quant, | ||
| 708 : | ; const uint16_t *mpeg_matrices); | ||
| 709 : | ; Ported from 32bit 3dne cousin | ||
| 710 : | ;----------------------------------------------------------------------------- | ||
| 711 : | |||
| 712 : | ; Note: We use (2*c + sgn(c) - sgn(-c)) as multiplier | ||
| 713 : | ; so we handle the 3 cases: c<0, c==0, and c>0 in one shot. | ||
| 714 : | ; sgn(x) is the result of 'pcmpgtw 0,x': 0 if x>=0, -1 if x<0. | ||
| 715 : | ; It's mixed with the extraction of the absolute value. | ||
| 716 : | |||
| 717 : | ALIGN 16 | ||
| 718 : | dequant_mpeg_inter_x86_64: | ||
| 719 : | mov rax, rdx ; quant | ||
| 720 : | mov rdx, rdi ; data | ||
| 721 : | mov rdi, rcx ; mpeg_matrices | ||
| 722 : | mov rcx, rsi ; coeff | ||
| 723 : | |||
| 724 : | lea r11, [mmx_mul_quant wrt rip] | ||
| 725 : | movq mm7, [r11 + rax*8 - 8] | ||
| 726 : | mov rax, -14 | ||
| 727 : | paddw mm7, mm7 ; << 1 | ||
| 728 : | pxor mm6, mm6 ; mismatch sum | ||
| 729 : | pxor mm1, mm1 | ||
| 730 : | pxor mm3, mm3 | ||
| 731 : | |||
| 732 : | ALIGN 16 | ||
| 733 : | .loop | ||
| 734 : | movq mm0, [rcx+8*rax + 7*16 ] ; mm0 = coeff[i] | ||
| 735 : | pcmpgtw mm1, mm0 ; mm1 = sgn(c) (preserved) | ||
| 736 : | movq mm2, [rcx+8*rax + 7*16 +8] ; mm2 = coeff[i+1] | ||
| 737 : | pcmpgtw mm3, mm2 ; mm3 = sgn(c') (preserved) | ||
| 738 : | paddsw mm0, mm1 ; c += sgn(c) | ||
| 739 : | paddsw mm2, mm3 ; c += sgn(c') | ||
| 740 : | paddw mm0, mm0 ; c *= 2 | ||
| 741 : | paddw mm2, mm2 ; c'*= 2 | ||
| 742 : | |||
| 743 : | movq mm4, [mmzero wrt rip] | ||
| 744 : | movq mm5, [mmzero wrt rip] | ||
| 745 : | psubw mm4, mm0 ; -c | ||
| 746 : | psubw mm5, mm2 ; -c' | ||
| 747 : | |||
| 748 : | psraw mm4, 16 ; mm4 = sgn(-c) | ||
| 749 : | psraw mm5, 16 ; mm5 = sgn(-c') | ||
| 750 : | psubsw mm0, mm4 ; c -= sgn(-c) | ||
| 751 : | psubsw mm2, mm5 ; c' -= sgn(-c') | ||
| 752 : | pxor mm0, mm1 ; finish changing sign if needed | ||
| 753 : | pxor mm2, mm3 ; finish changing sign if needed | ||
| 754 : | |||
| 755 : | ; we're short on register, here. Poor pairing... | ||
| 756 : | |||
| 757 : | movq mm4, mm7 ; (matrix*quant) | ||
| 758 : | ; nop | ||
| 759 : | pmullw mm4, [rdi + 512 + 8*rax + 7*16] | ||
| 760 : | movq mm5, mm4 | ||
| 761 : | pmulhw mm5, mm0 ; high of c*(matrix*quant) | ||
| 762 : | pmullw mm0, mm4 ; low of c*(matrix*quant) | ||
| 763 : | |||
| 764 : | movq mm4, mm7 ; (matrix*quant) | ||
| 765 : | pmullw mm4, [rdi + 512 + 8*rax + 7*16 + 8] | ||
| 766 : | add rax, byte 2 | ||
| 767 : | |||
| 768 : | pcmpgtw mm5, [mmzero wrt rip] | ||
| 769 : | paddusw mm0, mm5 | ||
| 770 : | psrlw mm0, 5 | ||
| 771 : | pxor mm0, mm1 ; start restoring sign | ||
| 772 : | psubusw mm1, mm5 | ||
| 773 : | |||
| 774 : | movq mm5, mm4 | ||
| 775 : | pmulhw mm5, mm2 ; high of c*(matrix*quant) | ||
| 776 : | pmullw mm2, mm4 ; low of c*(matrix*quant) | ||
| 777 : | psubw mm0, mm1 ; finish restoring sign | ||
| 778 : | |||
| 779 : | pcmpgtw mm5, [mmzero wrt rip] | ||
| 780 : | paddusw mm2, mm5 | ||
| 781 : | psrlw mm2, 5 | ||
| 782 : | pxor mm2, mm3 ; start restoring sign | ||
| 783 : | psubusw mm3, mm5 | ||
| 784 : | psubw mm2, mm3 ; finish restoring sign | ||
| 785 : | movq mm1, [mmzero wrt rip] | ||
| 786 : | movq mm3, [byte mmzero wrt rip] | ||
| 787 : | pxor mm6, mm0 ; mismatch control | ||
| 788 : | movq [rdx + 8*rax + 7*16 -2*8 ], mm0 ; data[i] | ||
| 789 : | pxor mm6, mm2 ; mismatch control | ||
| 790 : | movq [rdx + 8*rax + 7*16 -2*8 +8], mm2 ; data[i+1] | ||
| 791 : | |||
| 792 : | jng .loop | ||
| 793 : | ; nop | ||
| 794 : | |||
| 795 : | ; mismatch control | ||
| 796 : | |||
| 797 : | pshufw mm0, mm6, 01010101b | ||
| 798 : | pshufw mm1, mm6, 10101010b | ||
| 799 : | pshufw mm2, mm6, 11111111b | ||
| 800 : | pxor mm6, mm0 | ||
| 801 : | pxor mm1, mm2 | ||
| 802 : | pxor mm6, mm1 | ||
| 803 : | movd rax, mm6 | ||
| 804 : | and rax, byte 1 | ||
| 805 : | xor rax, byte 1 | ||
| 806 : | xor word [rdx + 2*63], ax | ||
| 807 : | |||
| 808 : | xor rax, rax | ||
| 809 : | ret | ||
| 810 : | .endfunc |
| No admin address has been configured | ViewVC Help |
| Powered by ViewVC 1.0.4 |