| 1 |
|
;/***************************************************************************** |
| 2 |
|
; * |
| 3 |
|
; * XVID MPEG-4 VIDEO CODEC |
| 4 |
|
; * - MPEG4 Quantization H263 implementation / MMX optimized - |
| 5 |
|
; * |
| 6 |
|
; * Copyright(C) 2001-2003 Peter Ross <pross@xvid.org> |
| 7 |
|
; * 2002-2003 Pascal Massimino <skal@planet-d.net> |
| 8 |
|
; * |
| 9 |
|
; * This program is free software ; you can redistribute it and/or modify |
| 10 |
|
; * it under the terms of the GNU General Public License as published by |
| 11 |
|
; * the Free Software Foundation ; either version 2 of the License, or |
| 12 |
|
; * (at your option) any later version. |
| 13 |
|
; * |
| 14 |
|
; * This program is distributed in the hope that it will be useful, |
| 15 |
|
; * but WITHOUT ANY WARRANTY ; without even the implied warranty of |
| 16 |
|
; * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 17 |
|
; * GNU General Public License for more details. |
| 18 |
|
; * |
| 19 |
|
; * You should have received a copy of the GNU General Public License |
| 20 |
|
; * along with this program ; if not, write to the Free Software |
| 21 |
|
; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
| 22 |
|
; * |
| 23 |
|
; * $Id$ |
| 24 |
|
; * |
| 25 |
|
; ****************************************************************************/ |
| 26 |
|
|
| 27 |
|
; enable dequant saturate [-2048,2047], test purposes only. |
| 28 |
|
%define SATURATE |
| 29 |
|
|
| 30 |
|
; data/text alignment |
| 31 |
|
%define ALIGN 8 |
| 32 |
|
|
| 33 |
|
bits 32 |
| 34 |
|
|
| 35 |
|
%macro cglobal 1 |
| 36 |
|
%ifdef PREFIX |
| 37 |
|
global _%1 |
| 38 |
|
%define %1 _%1 |
| 39 |
|
%else |
| 40 |
|
global %1 |
| 41 |
|
%endif |
| 42 |
|
%endmacro |
| 43 |
|
|
| 44 |
|
;*************************************************************************** |
| 45 |
|
; Local data |
| 46 |
|
;*************************************************************************** |
| 47 |
|
|
| 48 |
|
%ifdef FORMAT_COFF |
| 49 |
|
section .data data |
| 50 |
|
%else |
| 51 |
|
section .data data align=16 |
| 52 |
|
%endif |
| 53 |
|
|
| 54 |
|
align 16 |
| 55 |
|
plus_one: |
| 56 |
|
times 8 dw 1 |
| 57 |
|
|
| 58 |
|
;=========================================================================== |
| 59 |
|
; |
| 60 |
|
; subtract by Q/2 table |
| 61 |
|
; |
| 62 |
|
;=========================================================================== |
| 63 |
|
|
| 64 |
|
align 16 |
| 65 |
|
mmx_sub: |
| 66 |
|
%assign quant 1 |
| 67 |
|
%rep 31 |
| 68 |
|
times 4 dw quant / 2 |
| 69 |
|
%assign quant quant+1 |
| 70 |
|
%endrep |
| 71 |
|
|
| 72 |
|
;=========================================================================== |
| 73 |
|
; |
| 74 |
|
; divide by 2Q table |
| 75 |
|
; |
| 76 |
|
; use a shift of 16 to take full advantage of _pmulhw_ |
| 77 |
|
; for q=1, _pmulhw_ will overflow so it is treated seperately |
| 78 |
|
; (3dnow2 provides _pmulhuw_ which wont cause overflow) |
| 79 |
|
; |
| 80 |
|
;=========================================================================== |
| 81 |
|
|
| 82 |
|
align 16 |
| 83 |
|
mmx_div: |
| 84 |
|
%assign quant 1 |
| 85 |
|
%rep 31 |
| 86 |
|
times 4 dw (1<<16) / (quant*2) + 1 |
| 87 |
|
%assign quant quant+1 |
| 88 |
|
%endrep |
| 89 |
|
|
| 90 |
|
;=========================================================================== |
| 91 |
|
; |
| 92 |
|
; add by (odd(Q) ? Q : Q - 1) table |
| 93 |
|
; |
| 94 |
|
;=========================================================================== |
| 95 |
|
|
| 96 |
|
align 16 |
| 97 |
|
mmx_add: |
| 98 |
|
%assign quant 1 |
| 99 |
|
%rep 31 |
| 100 |
|
%if quant % 2 != 0 |
| 101 |
|
times 4 dw quant |
| 102 |
|
%else |
| 103 |
|
times 4 dw quant - 1 |
| 104 |
|
%endif |
| 105 |
|
%assign quant quant+1 |
| 106 |
|
%endrep |
| 107 |
|
|
| 108 |
|
;=========================================================================== |
| 109 |
|
; |
| 110 |
|
; multiple by 2Q table |
| 111 |
|
; |
| 112 |
|
;=========================================================================== |
| 113 |
|
|
| 114 |
|
align 16 |
| 115 |
|
mmx_mul: |
| 116 |
|
%assign quant 1 |
| 117 |
|
%rep 31 |
| 118 |
|
times 4 dw quant*2 |
| 119 |
|
%assign quant quant+1 |
| 120 |
|
%endrep |
| 121 |
|
|
| 122 |
|
;=========================================================================== |
| 123 |
|
; |
| 124 |
|
; saturation limits |
| 125 |
|
; |
| 126 |
|
;=========================================================================== |
| 127 |
|
|
| 128 |
|
align 16 |
| 129 |
|
sse2_2047: |
| 130 |
|
times 8 dw 2047 |
| 131 |
|
|
| 132 |
|
align 16 |
| 133 |
|
mmx_2047: |
| 134 |
|
times 4 dw 2047 |
| 135 |
|
|
| 136 |
|
align 8 |
| 137 |
|
mmx_32768_minus_2048: |
| 138 |
|
times 4 dw (32768-2048) |
| 139 |
|
|
| 140 |
|
mmx_32767_minus_2047: |
| 141 |
|
times 4 dw (32767-2047) |
| 142 |
|
|
| 143 |
|
|
| 144 |
|
;*************************************************************************** |
| 145 |
|
; Code |
| 146 |
|
;*************************************************************************** |
| 147 |
|
|
| 148 |
|
section .text |
| 149 |
|
|
| 150 |
|
|
| 151 |
|
;=========================================================================== |
| 152 |
|
; |
| 153 |
|
; uint32_t quant_h263_intra_mmx(int16_t * coeff, |
| 154 |
|
; const int16_t const * data, |
| 155 |
|
; const uint32_t quant, |
| 156 |
|
; const uint32_t dcscalar); |
| 157 |
|
; |
| 158 |
|
;=========================================================================== |
| 159 |
|
|
| 160 |
|
align ALIGN |
| 161 |
|
cglobal quant_h263_intra_mmx |
| 162 |
|
quant_h263_intra_mmx: |
| 163 |
|
|
| 164 |
|
push ecx |
| 165 |
|
push esi |
| 166 |
|
push edi |
| 167 |
|
|
| 168 |
|
mov edi, [esp + 12 + 4] ; coeff |
| 169 |
|
mov esi, [esp + 12 + 8] ; data |
| 170 |
|
mov eax, [esp + 12 + 12] ; quant |
| 171 |
|
|
| 172 |
|
xor ecx, ecx |
| 173 |
|
cmp al, 1 |
| 174 |
|
jz .q1loop |
| 175 |
|
|
| 176 |
|
movq mm7, [mmx_div + eax * 8 - 8] |
| 177 |
|
|
| 178 |
|
align ALIGN |
| 179 |
|
.loop |
| 180 |
|
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
| 181 |
|
movq mm3, [esi + 8*ecx + 8] |
| 182 |
|
pxor mm1, mm1 ; mm1 = 0 |
| 183 |
|
pxor mm4, mm4 ; |
| 184 |
|
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
| 185 |
|
pcmpgtw mm4, mm3 ; |
| 186 |
|
pxor mm0, mm1 ; mm0 = |mm0| |
| 187 |
|
pxor mm3, mm4 ; |
| 188 |
|
psubw mm0, mm1 ; displace |
| 189 |
|
psubw mm3, mm4 ; |
| 190 |
|
pmulhw mm0, mm7 ; mm0 = (mm0 / 2Q) >> 16 |
| 191 |
|
pmulhw mm3, mm7 ; |
| 192 |
|
pxor mm0, mm1 ; mm0 *= sign(mm0) |
| 193 |
|
pxor mm3, mm4 ; |
| 194 |
|
psubw mm0, mm1 ; undisplace |
| 195 |
|
psubw mm3, mm4 ; |
| 196 |
|
movq [edi + 8*ecx], mm0 |
| 197 |
|
movq [edi + 8*ecx + 8], mm3 |
| 198 |
|
|
| 199 |
|
add ecx, 2 |
| 200 |
|
cmp ecx, 16 |
| 201 |
|
jnz .loop |
| 202 |
|
|
| 203 |
|
.done |
| 204 |
|
|
| 205 |
|
; caclulate data[0] // (int32_t)dcscalar) |
| 206 |
|
mov ecx, [esp + 12 + 16] ; dcscalar |
| 207 |
|
mov edx, ecx |
| 208 |
|
movsx eax, word [esi] ; data[0] |
| 209 |
|
shr edx, 1 ; edx = dcscalar /2 |
| 210 |
|
cmp eax, 0 |
| 211 |
|
jg .gtzero |
| 212 |
|
|
| 213 |
|
sub eax, edx |
| 214 |
|
jmp short .mul |
| 215 |
|
|
| 216 |
|
.gtzero |
| 217 |
|
add eax, edx |
| 218 |
|
.mul |
| 219 |
|
cdq ; expand eax -> edx:eax |
| 220 |
|
idiv ecx ; eax = edx:eax / dcscalar |
| 221 |
|
mov [edi], ax ; coeff[0] = ax |
| 222 |
|
|
| 223 |
|
xor eax, eax ; return(0); |
| 224 |
|
pop edi |
| 225 |
|
pop esi |
| 226 |
|
pop ecx |
| 227 |
|
|
| 228 |
|
ret |
| 229 |
|
|
| 230 |
|
align ALIGN |
| 231 |
|
.q1loop |
| 232 |
|
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
| 233 |
|
movq mm3, [esi + 8*ecx + 8] |
| 234 |
|
pxor mm1, mm1 ; mm1 = 0 |
| 235 |
|
pxor mm4, mm4 ; |
| 236 |
|
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
| 237 |
|
pcmpgtw mm4, mm3 ; |
| 238 |
|
pxor mm0, mm1 ; mm0 = |mm0| |
| 239 |
|
pxor mm3, mm4 ; |
| 240 |
|
psubw mm0, mm1 ; displace |
| 241 |
|
psubw mm3, mm4 ; |
| 242 |
|
psrlw mm0, 1 ; mm0 >>= 1 (/2) |
| 243 |
|
psrlw mm3, 1 ; |
| 244 |
|
pxor mm0, mm1 ; mm0 *= sign(mm0) |
| 245 |
|
pxor mm3, mm4 |
| 246 |
|
psubw mm0, mm1 ; undisplace |
| 247 |
|
psubw mm3, mm4 ; |
| 248 |
|
movq [edi + 8*ecx], mm0 |
| 249 |
|
movq [edi + 8*ecx + 8], mm3 |
| 250 |
|
|
| 251 |
|
add ecx, 2 |
| 252 |
|
cmp ecx, 16 |
| 253 |
|
jnz .q1loop |
| 254 |
|
jmp short .done |
| 255 |
|
|
| 256 |
|
|
| 257 |
|
|
| 258 |
|
;=========================================================================== |
| 259 |
|
; |
| 260 |
|
; uint32_t quant_h263_intra_sse2(int16_t * coeff, |
| 261 |
|
; const int16_t const * data, |
| 262 |
|
; const uint32_t quant, |
| 263 |
|
; const uint32_t dcscalar); |
| 264 |
|
; |
| 265 |
|
;=========================================================================== |
| 266 |
|
|
| 267 |
|
align ALIGN |
| 268 |
|
cglobal quant_h263_intra_sse2 |
| 269 |
|
quant_h263_intra_sse2: |
| 270 |
|
|
| 271 |
|
push esi |
| 272 |
|
push edi |
| 273 |
|
|
| 274 |
|
mov edi, [esp + 8 + 4] ; coeff |
| 275 |
|
mov esi, [esp + 8 + 8] ; data |
| 276 |
|
mov eax, [esp + 8 + 12] ; quant |
| 277 |
|
|
| 278 |
|
xor ecx, ecx |
| 279 |
|
cmp al, 1 |
| 280 |
|
jz near .qas2_q1loop |
| 281 |
|
|
| 282 |
|
.qas2_not1 |
| 283 |
|
movq mm7, [mmx_div + eax*8 - 8] |
| 284 |
|
movq2dq xmm7, mm7 |
| 285 |
|
movlhps xmm7, xmm7 |
| 286 |
|
|
| 287 |
|
align 16 |
| 288 |
|
.qas2_loop |
| 289 |
|
movdqa xmm0, [esi + ecx*8] ; xmm0 = [1st] |
| 290 |
|
movdqa xmm3, [esi + ecx*8 + 16] ; xmm3 = [2nd] |
| 291 |
|
pxor xmm1, xmm1 |
| 292 |
|
pxor xmm4, xmm4 |
| 293 |
|
pcmpgtw xmm1, xmm0 |
| 294 |
|
pcmpgtw xmm4, xmm3 |
| 295 |
|
pxor xmm0, xmm1 |
| 296 |
|
pxor xmm3, xmm4 |
| 297 |
|
psubw xmm0, xmm1 |
| 298 |
|
psubw xmm3, xmm4 |
| 299 |
|
pmulhw xmm0, xmm7 |
| 300 |
|
pmulhw xmm3, xmm7 |
| 301 |
|
pxor xmm0, xmm1 |
| 302 |
|
pxor xmm3, xmm4 |
| 303 |
|
psubw xmm0, xmm1 |
| 304 |
|
psubw xmm3, xmm4 |
| 305 |
|
movdqa [edi + ecx*8], xmm0 |
| 306 |
|
movdqa [edi + ecx*8 + 16], xmm3 |
| 307 |
|
|
| 308 |
|
add ecx, 4 |
| 309 |
|
cmp ecx, 16 |
| 310 |
|
jnz .qas2_loop |
| 311 |
|
|
| 312 |
|
.qas2_done |
| 313 |
|
mov ecx, [esp + 8 + 16] ; dcscalar |
| 314 |
|
mov edx, ecx |
| 315 |
|
movsx eax, word [esi] |
| 316 |
|
shr edx, 1 |
| 317 |
|
cmp eax, 0 |
| 318 |
|
jg .qas2_gtzero |
| 319 |
|
|
| 320 |
|
sub eax, edx |
| 321 |
|
jmp short .qas2_mul |
| 322 |
|
|
| 323 |
|
.qas2_gtzero |
| 324 |
|
add eax, edx |
| 325 |
|
|
| 326 |
|
.qas2_mul |
| 327 |
|
cdq |
| 328 |
|
idiv ecx |
| 329 |
|
|
| 330 |
|
mov [edi], ax |
| 331 |
|
|
| 332 |
|
xor eax, eax ; return(0); |
| 333 |
|
pop edi |
| 334 |
|
pop esi |
| 335 |
|
|
| 336 |
|
ret |
| 337 |
|
|
| 338 |
|
align 16 |
| 339 |
|
.qas2_q1loop |
| 340 |
|
movdqa xmm0, [esi + ecx*8] ; xmm0 = [1st] |
| 341 |
|
movdqa xmm3, [esi + ecx*8 + 16] ; xmm3 = [2nd] |
| 342 |
|
pxor xmm1, xmm1 |
| 343 |
|
pxor xmm4, xmm4 |
| 344 |
|
pcmpgtw xmm1, xmm0 |
| 345 |
|
pcmpgtw xmm4, xmm3 |
| 346 |
|
pxor xmm0, xmm1 |
| 347 |
|
pxor xmm3, xmm4 |
| 348 |
|
psubw xmm0, xmm1 |
| 349 |
|
psubw xmm3, xmm4 |
| 350 |
|
psrlw xmm0, 1 |
| 351 |
|
psrlw xmm3, 1 |
| 352 |
|
pxor xmm0, xmm1 |
| 353 |
|
pxor xmm3, xmm4 |
| 354 |
|
psubw xmm0, xmm1 |
| 355 |
|
psubw xmm3, xmm4 |
| 356 |
|
movdqa [edi + ecx*8], xmm0 |
| 357 |
|
movdqa [edi + ecx*8 + 16], xmm3 |
| 358 |
|
|
| 359 |
|
add ecx, 4 |
| 360 |
|
cmp ecx, 16 |
| 361 |
|
jnz .qas2_q1loop |
| 362 |
|
jmp near .qas2_done |
| 363 |
|
|
| 364 |
|
|
| 365 |
|
|
| 366 |
|
;=========================================================================== |
| 367 |
|
; |
| 368 |
|
; uint32_t quant_h263_inter_mmx(int16_t * coeff, |
| 369 |
|
; const int16_t const * data, |
| 370 |
|
; const uint32_t quant); |
| 371 |
|
; |
| 372 |
|
;=========================================================================== |
| 373 |
|
|
| 374 |
|
align ALIGN |
| 375 |
|
cglobal quant_h263_inter_mmx |
| 376 |
|
quant_h263_inter_mmx: |
| 377 |
|
|
| 378 |
|
push ecx |
| 379 |
|
push esi |
| 380 |
|
push edi |
| 381 |
|
|
| 382 |
|
mov edi, [esp + 12 + 4] ; coeff |
| 383 |
|
mov esi, [esp + 12 + 8] ; data |
| 384 |
|
mov eax, [esp + 12 + 12] ; quant |
| 385 |
|
|
| 386 |
|
xor ecx, ecx |
| 387 |
|
|
| 388 |
|
pxor mm5, mm5 ; sum |
| 389 |
|
movq mm6, [mmx_sub + eax * 8 - 8]; sub |
| 390 |
|
|
| 391 |
|
cmp al, 1 |
| 392 |
|
jz .q1loop |
| 393 |
|
|
| 394 |
|
movq mm7, [mmx_div + eax * 8 - 8] ; divider |
| 395 |
|
|
| 396 |
|
align ALIGN |
| 397 |
|
.loop |
| 398 |
|
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
| 399 |
|
movq mm3, [esi + 8*ecx + 8] |
| 400 |
|
pxor mm1, mm1 ; mm1 = 0 |
| 401 |
|
pxor mm4, mm4 ; |
| 402 |
|
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
| 403 |
|
pcmpgtw mm4, mm3 ; |
| 404 |
|
pxor mm0, mm1 ; mm0 = |mm0| |
| 405 |
|
pxor mm3, mm4 ; |
| 406 |
|
psubw mm0, mm1 ; displace |
| 407 |
|
psubw mm3, mm4 ; |
| 408 |
|
psubusw mm0, mm6 ; mm0 -= sub (unsigned, dont go < 0) |
| 409 |
|
psubusw mm3, mm6 ; |
| 410 |
|
pmulhw mm0, mm7 ; mm0 = (mm0 / 2Q) >> 16 |
| 411 |
|
pmulhw mm3, mm7 ; |
| 412 |
|
paddw mm5, mm0 ; sum += mm0 |
| 413 |
|
pxor mm0, mm1 ; mm0 *= sign(mm0) |
| 414 |
|
paddw mm5, mm3 ; |
| 415 |
|
pxor mm3, mm4 ; |
| 416 |
|
psubw mm0, mm1 ; undisplace |
| 417 |
|
psubw mm3, mm4 |
| 418 |
|
movq [edi + 8*ecx], mm0 |
| 419 |
|
movq [edi + 8*ecx + 8], mm3 |
| 420 |
|
|
| 421 |
|
add ecx, 2 |
| 422 |
|
cmp ecx, 16 |
| 423 |
|
jnz .loop |
| 424 |
|
|
| 425 |
|
.done |
| 426 |
|
pmaddwd mm5, [plus_one] |
| 427 |
|
movq mm0, mm5 |
| 428 |
|
psrlq mm5, 32 |
| 429 |
|
paddd mm0, mm5 |
| 430 |
|
|
| 431 |
|
movd eax, mm0 ; return sum |
| 432 |
|
pop edi |
| 433 |
|
pop esi |
| 434 |
|
pop ecx |
| 435 |
|
|
| 436 |
|
ret |
| 437 |
|
|
| 438 |
|
align ALIGN |
| 439 |
|
.q1loop |
| 440 |
|
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
| 441 |
|
movq mm3, [esi + 8*ecx+ 8] ; |
| 442 |
|
pxor mm1, mm1 ; mm1 = 0 |
| 443 |
|
pxor mm4, mm4 ; |
| 444 |
|
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
| 445 |
|
pcmpgtw mm4, mm3 ; |
| 446 |
|
pxor mm0, mm1 ; mm0 = |mm0| |
| 447 |
|
pxor mm3, mm4 ; |
| 448 |
|
psubw mm0, mm1 ; displace |
| 449 |
|
psubw mm3, mm4 ; |
| 450 |
|
psubusw mm0, mm6 ; mm0 -= sub (unsigned, dont go < 0) |
| 451 |
|
psubusw mm3, mm6 ; |
| 452 |
|
psrlw mm0, 1 ; mm0 >>= 1 (/2) |
| 453 |
|
psrlw mm3, 1 ; |
| 454 |
|
paddw mm5, mm0 ; sum += mm0 |
| 455 |
|
pxor mm0, mm1 ; mm0 *= sign(mm0) |
| 456 |
|
paddw mm5, mm3 ; |
| 457 |
|
pxor mm3, mm4 ; |
| 458 |
|
psubw mm0, mm1 ; undisplace |
| 459 |
|
psubw mm3, mm4 |
| 460 |
|
movq [edi + 8*ecx], mm0 |
| 461 |
|
movq [edi + 8*ecx + 8], mm3 |
| 462 |
|
|
| 463 |
|
add ecx, 2 |
| 464 |
|
cmp ecx, 16 |
| 465 |
|
jnz .q1loop |
| 466 |
|
|
| 467 |
|
jmp .done |
| 468 |
|
|
| 469 |
|
|
| 470 |
|
|
| 471 |
|
;=========================================================================== |
| 472 |
|
; |
| 473 |
|
; uint32_t quant_h263_inter_sse2(int16_t * coeff, |
| 474 |
|
; const int16_t const * data, |
| 475 |
|
; const uint32_t quant); |
| 476 |
|
; |
| 477 |
|
;=========================================================================== |
| 478 |
|
|
| 479 |
|
align 16 |
| 480 |
|
cglobal quant_h263_inter_sse2 |
| 481 |
|
quant_h263_inter_sse2: |
| 482 |
|
|
| 483 |
|
push esi |
| 484 |
|
push edi |
| 485 |
|
|
| 486 |
|
mov edi, [esp + 8 + 4] ; coeff |
| 487 |
|
mov esi, [esp + 8 + 8] ; data |
| 488 |
|
mov eax, [esp + 8 + 12] ; quant |
| 489 |
|
|
| 490 |
|
xor ecx, ecx |
| 491 |
|
|
| 492 |
|
pxor xmm5, xmm5 ; sum |
| 493 |
|
|
| 494 |
|
movq mm0, [mmx_sub + eax*8 - 8] ; sub |
| 495 |
|
movq2dq xmm6, mm0 ; load into low 8 bytes |
| 496 |
|
movlhps xmm6, xmm6 ; duplicate into high 8 bytes |
| 497 |
|
|
| 498 |
|
cmp al, 1 |
| 499 |
|
jz near .qes2_q1loop |
| 500 |
|
|
| 501 |
|
.qes2_not1 |
| 502 |
|
movq mm0, [mmx_div + eax*8 - 8] ; divider |
| 503 |
|
movq2dq xmm7, mm0 |
| 504 |
|
movlhps xmm7, xmm7 |
| 505 |
|
|
| 506 |
|
align 16 |
| 507 |
|
.qes2_loop |
| 508 |
|
movdqa xmm0, [esi + ecx*8] ; xmm0 = [1st] |
| 509 |
|
movdqa xmm3, [esi + ecx*8 + 16] ; xmm3 = [2nd] |
| 510 |
|
pxor xmm1, xmm1 |
| 511 |
|
pxor xmm4, xmm4 |
| 512 |
|
pcmpgtw xmm1, xmm0 |
| 513 |
|
pcmpgtw xmm4, xmm3 |
| 514 |
|
pxor xmm0, xmm1 |
| 515 |
|
pxor xmm3, xmm4 |
| 516 |
|
psubw xmm0, xmm1 |
| 517 |
|
psubw xmm3, xmm4 |
| 518 |
|
psubusw xmm0, xmm6 |
| 519 |
|
psubusw xmm3, xmm6 |
| 520 |
|
pmulhw xmm0, xmm7 |
| 521 |
|
pmulhw xmm3, xmm7 |
| 522 |
|
paddw xmm5, xmm0 |
| 523 |
|
pxor xmm0, xmm1 |
| 524 |
|
paddw xmm5, xmm3 |
| 525 |
|
pxor xmm3, xmm4 |
| 526 |
|
psubw xmm0, xmm1 |
| 527 |
|
psubw xmm3, xmm4 |
| 528 |
|
movdqa [edi + ecx*8], xmm0 |
| 529 |
|
movdqa [edi + ecx*8 + 16], xmm3 |
| 530 |
|
|
| 531 |
|
add ecx, 4 |
| 532 |
|
cmp ecx, 16 |
| 533 |
|
jnz .qes2_loop |
| 534 |
|
|
| 535 |
|
.qes2_done |
| 536 |
|
movdqu xmm6, [plus_one] |
| 537 |
|
pmaddwd xmm5, xmm6 |
| 538 |
|
movhlps xmm6, xmm5 |
| 539 |
|
paddd xmm5, xmm6 |
| 540 |
|
movdq2q mm0, xmm5 |
| 541 |
|
|
| 542 |
|
movq mm5, mm0 |
| 543 |
|
psrlq mm5, 32 |
| 544 |
|
paddd mm0, mm5 |
| 545 |
|
|
| 546 |
|
movd eax, mm0 ; return sum |
| 547 |
|
|
| 548 |
|
pop edi |
| 549 |
|
pop esi |
| 550 |
|
|
| 551 |
|
ret |
| 552 |
|
|
| 553 |
|
align 16 |
| 554 |
|
.qes2_q1loop |
| 555 |
|
movdqa xmm0, [esi + ecx*8] ; xmm0 = [1st] |
| 556 |
|
movdqa xmm3, [esi + ecx*8 + 16] ; xmm3 = [2nd] |
| 557 |
|
pxor xmm1, xmm1 |
| 558 |
|
pxor xmm4, xmm4 |
| 559 |
|
pcmpgtw xmm1, xmm0 |
| 560 |
|
pcmpgtw xmm4, xmm3 |
| 561 |
|
pxor xmm0, xmm1 |
| 562 |
|
pxor xmm3, xmm4 |
| 563 |
|
psubw xmm0, xmm1 |
| 564 |
|
psubw xmm3, xmm4 |
| 565 |
|
psubusw xmm0, xmm6 |
| 566 |
|
psubusw xmm3, xmm6 |
| 567 |
|
psrlw xmm0, 1 |
| 568 |
|
psrlw xmm3, 1 |
| 569 |
|
paddw xmm5, xmm0 |
| 570 |
|
pxor xmm0, xmm1 |
| 571 |
|
paddw xmm5, xmm3 |
| 572 |
|
pxor xmm3, xmm4 |
| 573 |
|
psubw xmm0, xmm1 |
| 574 |
|
psubw xmm3, xmm4 |
| 575 |
|
movdqa [edi + ecx*8], xmm0 |
| 576 |
|
movdqa [edi + ecx*8 + 16], xmm3 |
| 577 |
|
|
| 578 |
|
add ecx, 4 |
| 579 |
|
cmp ecx, 16 |
| 580 |
|
jnz .qes2_q1loop |
| 581 |
|
jmp .qes2_done |
| 582 |
|
|
| 583 |
|
|
| 584 |
|
;=========================================================================== |
| 585 |
|
; |
| 586 |
|
; uint32_t dequant_h263_intra_mmx(int16_t *data, |
| 587 |
|
; const int16_t const *coeff, |
| 588 |
|
; const uint32_t quant, |
| 589 |
|
; const uint32_t dcscalar); |
| 590 |
|
; |
| 591 |
|
;=========================================================================== |
| 592 |
|
|
| 593 |
|
; note: we only saturate to +2047 *before* restoring the sign. |
| 594 |
|
; Hence, final clamp really is [-2048,2047] |
| 595 |
|
|
| 596 |
|
align ALIGN |
| 597 |
|
cglobal dequant_h263_intra_mmx |
| 598 |
|
dequant_h263_intra_mmx: |
| 599 |
|
|
| 600 |
|
mov edx, [esp+ 4] ; data |
| 601 |
|
mov ecx, [esp+ 8] ; coeff |
| 602 |
|
mov eax, [esp+12] ; quant |
| 603 |
|
movq mm6, [mmx_add + eax*8 - 8] ; quant or quant-1 |
| 604 |
|
movq mm7, [mmx_mul + eax*8 - 8] ; 2*quant |
| 605 |
|
mov eax, -16 |
| 606 |
|
|
| 607 |
|
align ALIGN |
| 608 |
|
.loop |
| 609 |
|
movq mm0, [ecx+8*eax+8*16] ; c = coeff[i] |
| 610 |
|
movq mm3, [ecx+8*eax+8*16 + 8] ; c' = coeff[i+1] |
| 611 |
|
pxor mm1, mm1 |
| 612 |
|
pxor mm4, mm4 |
| 613 |
|
pcmpgtw mm1, mm0 ; sign(c) |
| 614 |
|
pcmpgtw mm4, mm3 ; sign(c') |
| 615 |
|
pxor mm2, mm2 |
| 616 |
|
pxor mm5, mm5 |
| 617 |
|
pcmpeqw mm2, mm0 ; c is zero |
| 618 |
|
pcmpeqw mm5, mm3 ; c' is zero |
| 619 |
|
pandn mm2, mm6 ; offset = isZero ? 0 : quant_add |
| 620 |
|
pandn mm5, mm6 |
| 621 |
|
pxor mm0, mm1 ; negate if negative |
| 622 |
|
pxor mm3, mm4 ; negate if negative |
| 623 |
|
psubw mm0, mm1 |
| 624 |
|
psubw mm3, mm4 |
| 625 |
|
pmullw mm0, mm7 ; *= 2Q |
| 626 |
|
pmullw mm3, mm7 ; *= 2Q |
| 627 |
|
paddw mm0, mm2 ; + offset |
| 628 |
|
paddw mm3, mm5 ; + offset |
| 629 |
|
paddw mm0, mm1 ; negate back |
| 630 |
|
paddw mm3, mm4 ; negate back |
| 631 |
|
|
| 632 |
|
; saturates to +2047 |
| 633 |
|
movq mm2, [mmx_32767_minus_2047] |
| 634 |
|
add eax, 2 |
| 635 |
|
paddsw mm0, mm2 |
| 636 |
|
paddsw mm3, mm2 |
| 637 |
|
psubsw mm0, mm2 |
| 638 |
|
psubsw mm3, mm2 |
| 639 |
|
|
| 640 |
|
pxor mm0, mm1 |
| 641 |
|
pxor mm3, mm4 |
| 642 |
|
movq [edx + 8*eax + 8*16 - 2*8], mm0 |
| 643 |
|
movq [edx + 8*eax + 8*16+8 - 2*8], mm3 |
| 644 |
|
jnz near .loop |
| 645 |
|
|
| 646 |
|
; deal with DC |
| 647 |
|
movd mm0, [ecx] |
| 648 |
|
pmullw mm0, [esp+16] ; dcscalar |
| 649 |
|
movq mm2, [mmx_32767_minus_2047] |
| 650 |
|
paddsw mm0, mm2 |
| 651 |
|
psubsw mm0, mm2 |
| 652 |
|
movq mm3, [mmx_32768_minus_2048] |
| 653 |
|
psubsw mm0, mm3 |
| 654 |
|
paddsw mm0, mm3 |
| 655 |
|
movd eax, mm0 |
| 656 |
|
mov [edx], ax |
| 657 |
|
|
| 658 |
|
xor eax, eax ; return(0); |
| 659 |
|
ret |
| 660 |
|
|
| 661 |
|
;=========================================================================== |
| 662 |
|
; |
| 663 |
|
; uint32_t dequant_h263_intra_xmm(int16_t *data, |
| 664 |
|
; const int16_t const *coeff, |
| 665 |
|
; const uint32_t quant, |
| 666 |
|
; const uint32_t dcscalar); |
| 667 |
|
; |
| 668 |
|
;=========================================================================== |
| 669 |
|
|
| 670 |
|
; this is the same as dequant_inter_mmx, except that we're |
| 671 |
|
; saturating using 'pminsw' (saves 2 cycles/loop => ~5% faster) |
| 672 |
|
|
| 673 |
|
align ALIGN |
| 674 |
|
cglobal dequant_h263_intra_xmm |
| 675 |
|
dequant_h263_intra_xmm: |
| 676 |
|
|
| 677 |
|
mov edx, [esp+ 4] ; data |
| 678 |
|
mov ecx, [esp+ 8] ; coeff |
| 679 |
|
mov eax, [esp+12] ; quant |
| 680 |
|
movq mm6, [mmx_add + eax*8 - 8] ; quant or quant-1 |
| 681 |
|
movq mm7, [mmx_mul + eax*8 - 8] ; 2*quant |
| 682 |
|
mov eax, -16 |
| 683 |
|
|
| 684 |
|
align ALIGN |
| 685 |
|
.loop |
| 686 |
|
movq mm0, [ecx+8*eax+8*16] ; c = coeff[i] |
| 687 |
|
movq mm3, [ecx+8*eax+8*16 + 8] ; c' = coeff[i+1] |
| 688 |
|
pxor mm1, mm1 |
| 689 |
|
pxor mm4, mm4 |
| 690 |
|
pcmpgtw mm1, mm0 ; sign(c) |
| 691 |
|
pcmpgtw mm4, mm3 ; sign(c') |
| 692 |
|
pxor mm2, mm2 |
| 693 |
|
pxor mm5, mm5 |
| 694 |
|
pcmpeqw mm2, mm0 ; c is zero |
| 695 |
|
pcmpeqw mm5, mm3 ; c' is zero |
| 696 |
|
pandn mm2, mm6 ; offset = isZero ? 0 : quant_add |
| 697 |
|
pandn mm5, mm6 |
| 698 |
|
pxor mm0, mm1 ; negate if negative |
| 699 |
|
pxor mm3, mm4 ; negate if negative |
| 700 |
|
psubw mm0, mm1 |
| 701 |
|
psubw mm3, mm4 |
| 702 |
|
pmullw mm0, mm7 ; *= 2Q |
| 703 |
|
pmullw mm3, mm7 ; *= 2Q |
| 704 |
|
paddw mm0, mm2 ; + offset |
| 705 |
|
paddw mm3, mm5 ; + offset |
| 706 |
|
paddw mm0, mm1 ; negate back |
| 707 |
|
paddw mm3, mm4 ; negate back |
| 708 |
|
|
| 709 |
|
; saturates to +2047 |
| 710 |
|
movq mm2, [mmx_2047] |
| 711 |
|
pminsw mm0, mm2 |
| 712 |
|
add eax, 2 |
| 713 |
|
pminsw mm3, mm2 |
| 714 |
|
|
| 715 |
|
pxor mm0, mm1 |
| 716 |
|
pxor mm3, mm4 |
| 717 |
|
movq [edx + 8*eax + 8*16 - 2*8], mm0 |
| 718 |
|
movq [edx + 8*eax + 8*16+8 - 2*8], mm3 |
| 719 |
|
jnz near .loop |
| 720 |
|
|
| 721 |
|
; deal with DC |
| 722 |
|
movd mm0, [ecx] |
| 723 |
|
pmullw mm0, [esp+16] ; dcscalar |
| 724 |
|
movq mm2, [mmx_32767_minus_2047] |
| 725 |
|
paddsw mm0, mm2 |
| 726 |
|
psubsw mm0, mm2 |
| 727 |
|
movq mm2, [mmx_32768_minus_2048] |
| 728 |
|
psubsw mm0, mm2 |
| 729 |
|
paddsw mm0, mm2 |
| 730 |
|
movd eax, mm0 |
| 731 |
|
mov [edx], ax |
| 732 |
|
|
| 733 |
|
xor eax, eax |
| 734 |
|
ret |
| 735 |
|
|
| 736 |
|
|
| 737 |
|
;=========================================================================== |
| 738 |
|
; |
| 739 |
|
; uint32_t dequant_h263_intra_sse2(int16_t *data, |
| 740 |
|
; const int16_t const *coeff, |
| 741 |
|
; const uint32_t quant, |
| 742 |
|
; const uint32_t dcscalar); |
| 743 |
|
; |
| 744 |
|
;=========================================================================== |
| 745 |
|
|
| 746 |
|
align ALIGN |
| 747 |
|
cglobal dequant_h263_intra_sse2 |
| 748 |
|
dequant_h263_intra_sse2: |
| 749 |
|
mov edx, [esp+ 4] ; data |
| 750 |
|
mov ecx, [esp+ 8] ; coeff |
| 751 |
|
mov eax, [esp+12] ; quant |
| 752 |
|
movq mm6, [mmx_add + eax * 8 - 8] |
| 753 |
|
movq mm7, [mmx_mul + eax * 8 - 8] |
| 754 |
|
movq2dq xmm6, mm6 |
| 755 |
|
movq2dq xmm7, mm7 |
| 756 |
|
movlhps xmm6, xmm6 |
| 757 |
|
movlhps xmm7, xmm7 |
| 758 |
|
mov eax, -16 |
| 759 |
|
|
| 760 |
|
align ALIGN |
| 761 |
|
.loop |
| 762 |
|
movdqa xmm0, [ecx + 8*16 + 8*eax] ; c = coeff[i] |
| 763 |
|
movdqa xmm3, [ecx + 8*16 + 8*eax+ 16] |
| 764 |
|
pxor xmm1, xmm1 |
| 765 |
|
pxor xmm4, xmm4 |
| 766 |
|
pcmpgtw xmm1, xmm0 ; sign(c) |
| 767 |
|
pcmpgtw xmm4, xmm3 |
| 768 |
|
pxor xmm2, xmm2 |
| 769 |
|
pxor xmm5, xmm5 |
| 770 |
|
pcmpeqw xmm2, xmm0 ; c is zero |
| 771 |
|
pcmpeqw xmm5, xmm3 |
| 772 |
|
pandn xmm2, xmm6 ; offset = isZero ? 0 : quant_add |
| 773 |
|
pandn xmm5, xmm6 |
| 774 |
|
pxor xmm0, xmm1 ; negate if negative |
| 775 |
|
pxor xmm3, xmm4 |
| 776 |
|
psubw xmm0, xmm1 |
| 777 |
|
psubw xmm3, xmm4 |
| 778 |
|
pmullw xmm0, xmm7 ; *= 2Q |
| 779 |
|
pmullw xmm3, xmm7 |
| 780 |
|
paddw xmm0, xmm2 ; + offset |
| 781 |
|
paddw xmm3, xmm5 |
| 782 |
|
paddw xmm0, xmm1 ; negate back |
| 783 |
|
paddw xmm3, xmm4 |
| 784 |
|
|
| 785 |
|
; saturates to +2047 |
| 786 |
|
movdqa xmm2, [sse2_2047] |
| 787 |
|
pminsw xmm0, xmm2 |
| 788 |
|
add eax, 4 |
| 789 |
|
pminsw xmm3, xmm2 |
| 790 |
|
|
| 791 |
|
pxor xmm0, xmm1 |
| 792 |
|
pxor xmm3, xmm4 |
| 793 |
|
movdqa [edx + 8*16 - 8*4 + 8*eax], xmm0 |
| 794 |
|
movdqa [edx + 8*16 - 8*4 + 8*eax + 16], xmm3 |
| 795 |
|
jnz near .loop |
| 796 |
|
|
| 797 |
|
; deal with DC |
| 798 |
|
movd mm0, [ecx] |
| 799 |
|
pmullw mm0, [esp+16] ; dcscalar |
| 800 |
|
movq mm2, [mmx_32767_minus_2047] |
| 801 |
|
paddsw mm0, mm2 |
| 802 |
|
psubsw mm0, mm2 |
| 803 |
|
movq mm2, [mmx_32768_minus_2048] |
| 804 |
|
psubsw mm0, mm2 |
| 805 |
|
paddsw mm0, mm2 |
| 806 |
|
movd eax, mm0 |
| 807 |
|
mov [edx], ax |
| 808 |
|
|
| 809 |
|
xor eax, eax |
| 810 |
|
ret |
| 811 |
|
|
| 812 |
|
;=========================================================================== |
| 813 |
|
; |
| 814 |
|
; uint32t dequant_h263_inter_mmx(int16_t * data, |
| 815 |
|
; const int16_t * const coeff, |
| 816 |
|
; const uint32_t quant); |
| 817 |
|
; |
| 818 |
|
;=========================================================================== |
| 819 |
|
|
| 820 |
|
align ALIGN |
| 821 |
|
cglobal dequant_h263_inter_mmx |
| 822 |
|
dequant_h263_inter_mmx: |
| 823 |
|
|
| 824 |
|
mov edx, [esp+ 4] ; data |
| 825 |
|
mov ecx, [esp+ 8] ; coeff |
| 826 |
|
mov eax, [esp+12] ; quant |
| 827 |
|
movq mm6, [mmx_add + eax*8 - 8] ; quant or quant-1 |
| 828 |
|
movq mm7, [mmx_mul + eax*8 - 8] ; 2*quant |
| 829 |
|
mov eax, -16 |
| 830 |
|
|
| 831 |
|
align ALIGN |
| 832 |
|
.loop |
| 833 |
|
movq mm0, [ecx+8*eax+8*16] ; c = coeff[i] |
| 834 |
|
movq mm3, [ecx+8*eax+8*16 + 8] ; c' = coeff[i+1] |
| 835 |
|
pxor mm1, mm1 |
| 836 |
|
pxor mm4, mm4 |
| 837 |
|
pcmpgtw mm1, mm0 ; sign(c) |
| 838 |
|
pcmpgtw mm4, mm3 ; sign(c') |
| 839 |
|
pxor mm2, mm2 |
| 840 |
|
pxor mm5, mm5 |
| 841 |
|
pcmpeqw mm2, mm0 ; c is zero |
| 842 |
|
pcmpeqw mm5, mm3 ; c' is zero |
| 843 |
|
pandn mm2, mm6 ; offset = isZero ? 0 : quant_add |
| 844 |
|
pandn mm5, mm6 |
| 845 |
|
pxor mm0, mm1 ; negate if negative |
| 846 |
|
pxor mm3, mm4 ; negate if negative |
| 847 |
|
psubw mm0, mm1 |
| 848 |
|
psubw mm3, mm4 |
| 849 |
|
pmullw mm0, mm7 ; *= 2Q |
| 850 |
|
pmullw mm3, mm7 ; *= 2Q |
| 851 |
|
paddw mm0, mm2 ; + offset |
| 852 |
|
paddw mm3, mm5 ; + offset |
| 853 |
|
paddw mm0, mm1 ; negate back |
| 854 |
|
paddw mm3, mm4 ; negate back |
| 855 |
|
; saturates to +2047 |
| 856 |
|
movq mm2, [mmx_32767_minus_2047] |
| 857 |
|
add eax, 2 |
| 858 |
|
paddsw mm0, mm2 |
| 859 |
|
paddsw mm3, mm2 |
| 860 |
|
psubsw mm0, mm2 |
| 861 |
|
psubsw mm3, mm2 |
| 862 |
|
|
| 863 |
|
pxor mm0, mm1 |
| 864 |
|
pxor mm3, mm4 |
| 865 |
|
movq [edx + 8*eax + 8*16 - 2*8], mm0 |
| 866 |
|
movq [edx + 8*eax + 8*16+8 - 2*8], mm3 |
| 867 |
|
jnz near .loop |
| 868 |
|
|
| 869 |
|
xor eax, eax |
| 870 |
|
ret |
| 871 |
|
|
| 872 |
|
;=========================================================================== |
| 873 |
|
; |
| 874 |
|
; uint32_t dequant_h263_inter_xmm(int16_t * data, |
| 875 |
|
; const int16_t * const coeff, |
| 876 |
|
; const uint32_t quant); |
| 877 |
|
; |
| 878 |
|
;=========================================================================== |
| 879 |
|
|
| 880 |
|
; this is the same as dequant_inter_mmx, |
| 881 |
|
; except that we're saturating using 'pminsw' (saves 2 cycles/loop) |
| 882 |
|
|
| 883 |
|
align ALIGN |
| 884 |
|
cglobal dequant_h263_inter_xmm |
| 885 |
|
dequant_h263_inter_xmm: |
| 886 |
|
|
| 887 |
|
mov edx, [esp+ 4] ; data |
| 888 |
|
mov ecx, [esp+ 8] ; coeff |
| 889 |
|
mov eax, [esp+12] ; quant |
| 890 |
|
movq mm6, [mmx_add + eax*8 - 8] ; quant or quant-1 |
| 891 |
|
movq mm7, [mmx_mul + eax*8 - 8] ; 2*quant |
| 892 |
|
mov eax, -16 |
| 893 |
|
|
| 894 |
|
align ALIGN |
| 895 |
|
.loop |
| 896 |
|
movq mm0, [ecx+8*eax+8*16] ; c = coeff[i] |
| 897 |
|
movq mm3, [ecx+8*eax+8*16 + 8] ; c' = coeff[i+1] |
| 898 |
|
pxor mm1, mm1 |
| 899 |
|
pxor mm4, mm4 |
| 900 |
|
pcmpgtw mm1, mm0 ; sign(c) |
| 901 |
|
pcmpgtw mm4, mm3 ; sign(c') |
| 902 |
|
pxor mm2, mm2 |
| 903 |
|
pxor mm5, mm5 |
| 904 |
|
pcmpeqw mm2, mm0 ; c is zero |
| 905 |
|
pcmpeqw mm5, mm3 ; c' is zero |
| 906 |
|
pandn mm2, mm6 ; offset = isZero ? 0 : quant_add |
| 907 |
|
pandn mm5, mm6 |
| 908 |
|
pxor mm0, mm1 ; negate if negative |
| 909 |
|
pxor mm3, mm4 ; negate if negative |
| 910 |
|
psubw mm0, mm1 |
| 911 |
|
psubw mm3, mm4 |
| 912 |
|
pmullw mm0, mm7 ; *= 2Q |
| 913 |
|
pmullw mm3, mm7 ; *= 2Q |
| 914 |
|
paddw mm0, mm2 ; + offset |
| 915 |
|
paddw mm3, mm5 ; + offset |
| 916 |
|
paddw mm0, mm1 ; start restoring sign |
| 917 |
|
paddw mm3, mm4 ; start restoring sign |
| 918 |
|
; saturates to +2047 |
| 919 |
|
movq mm2, [mmx_2047] |
| 920 |
|
pminsw mm0, mm2 |
| 921 |
|
add eax, 2 |
| 922 |
|
pminsw mm3, mm2 |
| 923 |
|
|
| 924 |
|
pxor mm0, mm1 ; finish restoring sign |
| 925 |
|
pxor mm3, mm4 ; finish restoring sign |
| 926 |
|
movq [edx + 8*eax + 8*16 - 2*8], mm0 |
| 927 |
|
movq [edx + 8*eax + 8*16+8 - 2*8], mm3 |
| 928 |
|
jnz near .loop |
| 929 |
|
|
| 930 |
|
xor eax, eax |
| 931 |
|
ret |
| 932 |
|
|
| 933 |
|
;=========================================================================== |
| 934 |
|
; |
| 935 |
|
; uint32_t dequant_h263_inter_sse2(int16_t * data, |
| 936 |
|
; const int16_t * const coeff, |
| 937 |
|
; const uint32_t quant); |
| 938 |
|
; |
| 939 |
|
;=========================================================================== |
| 940 |
|
|
| 941 |
|
align ALIGN |
| 942 |
|
cglobal dequant_h263_inter_sse2 |
| 943 |
|
dequant_h263_inter_sse2: |
| 944 |
|
mov edx, [esp + 4] ; data |
| 945 |
|
mov ecx, [esp + 8] ; coeff |
| 946 |
|
mov eax, [esp + 12] ; quant |
| 947 |
|
movq mm6, [mmx_add + eax * 8 - 8] |
| 948 |
|
movq mm7, [mmx_mul + eax * 8 - 8] |
| 949 |
|
movq2dq xmm6, mm6 |
| 950 |
|
movq2dq xmm7, mm7 |
| 951 |
|
movlhps xmm6, xmm6 |
| 952 |
|
movlhps xmm7, xmm7 |
| 953 |
|
mov eax, -16 |
| 954 |
|
|
| 955 |
|
align ALIGN |
| 956 |
|
.loop |
| 957 |
|
movdqa xmm0, [ecx + 8*16 + 8*eax] ; c = coeff[i] |
| 958 |
|
movdqa xmm3, [ecx + 8*16 + 8*eax + 16] |
| 959 |
|
|
| 960 |
|
pxor xmm1, xmm1 |
| 961 |
|
pxor xmm4, xmm4 |
| 962 |
|
pcmpgtw xmm1, xmm0 ; sign(c) |
| 963 |
|
pcmpgtw xmm4, xmm3 |
| 964 |
|
pxor xmm2, xmm2 |
| 965 |
|
pxor xmm5, xmm5 |
| 966 |
|
pcmpeqw xmm2, xmm0 ; c is zero |
| 967 |
|
pcmpeqw xmm5, xmm3 |
| 968 |
|
pandn xmm2, xmm6 |
| 969 |
|
pandn xmm5, xmm6 |
| 970 |
|
pxor xmm0, xmm1 ; negate if negative |
| 971 |
|
pxor xmm3, xmm4 |
| 972 |
|
psubw xmm0, xmm1 |
| 973 |
|
psubw xmm3, xmm4 |
| 974 |
|
pmullw xmm0, xmm7 ; *= 2Q |
| 975 |
|
pmullw xmm3, xmm7 |
| 976 |
|
paddw xmm0, xmm2 ; + offset |
| 977 |
|
paddw xmm3, xmm5 |
| 978 |
|
|
| 979 |
|
paddw xmm0, xmm1 ; start restoring sign |
| 980 |
|
paddw xmm3, xmm4 |
| 981 |
|
|
| 982 |
|
; saturates to +2047 |
| 983 |
|
movdqa xmm2, [sse2_2047] |
| 984 |
|
pminsw xmm0, xmm2 |
| 985 |
|
add eax, 4 |
| 986 |
|
pminsw xmm3, xmm2 |
| 987 |
|
|
| 988 |
|
pxor xmm0, xmm1 ; finish restoring sign |
| 989 |
|
pxor xmm3, xmm4 |
| 990 |
|
movdqa [edx + 8*16 - 8*4 + 8*eax], xmm0 |
| 991 |
|
movdqa [edx + 8*16 - 8*4 + 8*eax + 16], xmm3 |
| 992 |
|
jnz near .loop |
| 993 |
|
|
| 994 |
|
xor eax, eax |
| 995 |
|
ret |
Parent Directory
| 
Revision Log
| 
Patch