Mercurial > libavcodec.hg

comparison libpostproc/postprocess_template.c @ 210:c2b6d68a0671 libavcodec

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
mangle for win32 in postproc
author atmos4
date Sat, 19 Jan 2002 05:14:46 +0000
parents 3ccd74a91074
children f1074f0d4969
comparison
equal deleted inserted replaced
209:c0d8ecae7ac5 210:c2b6d68a0671
58 asm volatile( 58 asm volatile(
59 "leal (%1, %2), %%eax \n\t" 59 "leal (%1, %2), %%eax \n\t"
60 "leal (%%eax, %2, 4), %%ebx \n\t" 60 "leal (%%eax, %2, 4), %%ebx \n\t"
61 // 0 1 2 3 4 5 6 7 8 9 61 // 0 1 2 3 4 5 6 7 8 9
62 // %1 eax eax+%2 eax+2%2 %1+4%2 ebx ebx+%2 ebx+2%2 %1+8%2 ebx+4%2 62 // %1 eax eax+%2 eax+2%2 %1+4%2 ebx ebx+%2 ebx+2%2 %1+8%2 ebx+4%2
63 "movq mmxDCOffset, %%mm7 \n\t" // mm7 = 0x7F 63 "movq "MANGLE(mmxDCOffset)", %%mm7 \n\t" // mm7 = 0x7F
64 "movq mmxDCThreshold, %%mm6 \n\t" // mm6 = 0x7D 64 "movq "MANGLE(mmxDCThreshold)", %%mm6 \n\t" // mm6 = 0x7D
65 "movq (%1), %%mm0 \n\t" 65 "movq (%1), %%mm0 \n\t"
66 "movq (%%eax), %%mm1 \n\t" 66 "movq (%%eax), %%mm1 \n\t"
67 "psubb %%mm1, %%mm0 \n\t" // mm0 = differnece 67 "psubb %%mm1, %%mm0 \n\t" // mm0 = differnece
68 "paddb %%mm7, %%mm0 \n\t" 68 "paddb %%mm7, %%mm0 \n\t"
69 "pcmpgtb %%mm6, %%mm0 \n\t" 69 "pcmpgtb %%mm6, %%mm0 \n\t"
169 "movq %%mm0, %%mm2 \n\t" 169 "movq %%mm0, %%mm2 \n\t"
170 "psubusb %%mm1, %%mm0 \n\t" 170 "psubusb %%mm1, %%mm0 \n\t"
171 "psubusb %%mm2, %%mm1 \n\t" 171 "psubusb %%mm2, %%mm1 \n\t"
172 "por %%mm1, %%mm0 \n\t" // ABS Diff 172 "por %%mm1, %%mm0 \n\t" // ABS Diff
173 173
174 "movq pQPb, %%mm7 \n\t" // QP,..., QP 174 "movq "MANGLE(pQPb)", %%mm7 \n\t" // QP,..., QP
175 "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP 175 "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP
176 "psubusb %%mm7, %%mm0 \n\t" // Diff <= 2QP -> 0 176 "psubusb %%mm7, %%mm0 \n\t" // Diff <= 2QP -> 0
177 "pcmpeqd b00, %%mm0 \n\t" 177 "pcmpeqd "MANGLE(b00)", %%mm0 \n\t"
178 "psrlq $16, %%mm0 \n\t" 178 "psrlq $16, %%mm0 \n\t"
179 "pcmpeqd bFF, %%mm0 \n\t" 179 "pcmpeqd "MANGLE(bFF)", %%mm0 \n\t"
180 // "movd %%mm0, (%1, %2, 4)\n\t" 180 // "movd %%mm0, (%1, %2, 4)\n\t"
181 "movd %%mm0, %0 \n\t" 181 "movd %%mm0, %0 \n\t"
182 : "=r" (isOk) 182 : "=r" (isOk)
183 : "r" (src), "r" (stride) 183 : "r" (src), "r" (stride)
184 ); 184 );
217 static inline void RENAME(doVertLowPass)(uint8_t *src, int stride, int QP) 217 static inline void RENAME(doVertLowPass)(uint8_t *src, int stride, int QP)
218 { 218 {
219 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) 219 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
220 src+= stride*3; 220 src+= stride*3;
221 asm volatile( //"movv %0 %1 %2\n\t" 221 asm volatile( //"movv %0 %1 %2\n\t"
222 "movq pQPb, %%mm0 \n\t" // QP,..., QP 222 "movq "MANGLE(pQPb)", %%mm0 \n\t" // QP,..., QP
223 223
224 "movq (%0), %%mm6 \n\t" 224 "movq (%0), %%mm6 \n\t"
225 "movq (%0, %1), %%mm5 \n\t" 225 "movq (%0, %1), %%mm5 \n\t"
226 "movq %%mm5, %%mm1 \n\t" 226 "movq %%mm5, %%mm1 \n\t"
227 "movq %%mm6, %%mm2 \n\t" 227 "movq %%mm6, %%mm2 \n\t"
228 "psubusb %%mm6, %%mm5 \n\t" 228 "psubusb %%mm6, %%mm5 \n\t"
229 "psubusb %%mm1, %%mm2 \n\t" 229 "psubusb %%mm1, %%mm2 \n\t"
230 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines 230 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines
231 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0 231 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0
232 "pcmpeqb b00, %%mm2 \n\t" // diff <= QP -> FF 232 "pcmpeqb "MANGLE(b00)", %%mm2 \n\t" // diff <= QP -> FF
233 233
234 "pand %%mm2, %%mm6 \n\t" 234 "pand %%mm2, %%mm6 \n\t"
235 "pandn %%mm1, %%mm2 \n\t" 235 "pandn %%mm1, %%mm2 \n\t"
236 "por %%mm2, %%mm6 \n\t"// First Line to Filter 236 "por %%mm2, %%mm6 \n\t"// First Line to Filter
237 237
245 "movq %%mm7, %%mm2 \n\t" 245 "movq %%mm7, %%mm2 \n\t"
246 "psubusb %%mm7, %%mm5 \n\t" 246 "psubusb %%mm7, %%mm5 \n\t"
247 "psubusb %%mm1, %%mm2 \n\t" 247 "psubusb %%mm1, %%mm2 \n\t"
248 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines 248 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines
249 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0 249 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0
250 "pcmpeqb b00, %%mm2 \n\t" // diff <= QP -> FF 250 "pcmpeqb "MANGLE(b00)", %%mm2 \n\t" // diff <= QP -> FF
251 251
252 "pand %%mm2, %%mm7 \n\t" 252 "pand %%mm2, %%mm7 \n\t"
253 "pandn %%mm1, %%mm2 \n\t" 253 "pandn %%mm1, %%mm2 \n\t"
254 "por %%mm2, %%mm7 \n\t" // First Line to Filter 254 "por %%mm2, %%mm7 \n\t" // First Line to Filter
255 255
401 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) 401 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
402 src+= stride*3; 402 src+= stride*3;
403 // FIXME rounding 403 // FIXME rounding
404 asm volatile( 404 asm volatile(
405 "pxor %%mm7, %%mm7 \n\t" // 0 405 "pxor %%mm7, %%mm7 \n\t" // 0
406 "movq b80, %%mm6 \n\t" // MIN_SIGNED_BYTE 406 "movq "MANGLE(b80)", %%mm6 \n\t" // MIN_SIGNED_BYTE
407 "leal (%0, %1), %%eax \n\t" 407 "leal (%0, %1), %%eax \n\t"
408 "leal (%%eax, %1, 4), %%ebx \n\t" 408 "leal (%%eax, %1, 4), %%ebx \n\t"
409 // 0 1 2 3 4 5 6 7 8 9 409 // 0 1 2 3 4 5 6 7 8 9
410 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 410 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1
411 "movq pQPb, %%mm0 \n\t" // QP,..., QP 411 "movq "MANGLE(pQPb)", %%mm0 \n\t" // QP,..., QP
412 "movq %%mm0, %%mm1 \n\t" // QP,..., QP 412 "movq %%mm0, %%mm1 \n\t" // QP,..., QP
413 "paddusb b02, %%mm0 \n\t" 413 "paddusb "MANGLE(b02)", %%mm0 \n\t"
414 "psrlw $2, %%mm0 \n\t" 414 "psrlw $2, %%mm0 \n\t"
415 "pand b3F, %%mm0 \n\t" // QP/4,..., QP/4 415 "pand "MANGLE(b3F)", %%mm0 \n\t" // QP/4,..., QP/4
416 "paddusb %%mm1, %%mm0 \n\t" // QP*1.25 ... 416 "paddusb %%mm1, %%mm0 \n\t" // QP*1.25 ...
417 "movq (%0, %1, 4), %%mm2 \n\t" // line 4 417 "movq (%0, %1, 4), %%mm2 \n\t" // line 4
418 "movq (%%ebx), %%mm3 \n\t" // line 5 418 "movq (%%ebx), %%mm3 \n\t" // line 5
419 "movq %%mm2, %%mm4 \n\t" // line 4 419 "movq %%mm2, %%mm4 \n\t" // line 4
420 "pcmpeqb %%mm5, %%mm5 \n\t" // -1 420 "pcmpeqb %%mm5, %%mm5 \n\t" // -1
439 // "psubb %%mm6, %%mm2 \n\t" 439 // "psubb %%mm6, %%mm2 \n\t"
440 "movq %%mm2, (%%ebx) \n\t" 440 "movq %%mm2, (%%ebx) \n\t"
441 441
442 "paddb %%mm6, %%mm5 \n\t" 442 "paddb %%mm6, %%mm5 \n\t"
443 "psrlw $2, %%mm5 \n\t" 443 "psrlw $2, %%mm5 \n\t"
444 "pand b3F, %%mm5 \n\t" 444 "pand "MANGLE(b3F)", %%mm5 \n\t"
445 "psubb b20, %%mm5 \n\t" // (l5-l4)/8 445 "psubb "MANGLE(b20)", %%mm5 \n\t" // (l5-l4)/8
446 446
447 "movq (%%eax, %1, 2), %%mm2 \n\t" 447 "movq (%%eax, %1, 2), %%mm2 \n\t"
448 "paddb %%mm6, %%mm2 \n\t" // line 3 + 0x80 448 "paddb %%mm6, %%mm2 \n\t" // line 3 + 0x80
449 "paddsb %%mm5, %%mm2 \n\t" 449 "paddsb %%mm5, %%mm2 \n\t"
450 "psubb %%mm6, %%mm2 \n\t" 450 "psubb %%mm6, %%mm2 \n\t"
501 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) 501 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
502 src+= stride*3; 502 src+= stride*3;
503 503
504 asm volatile( 504 asm volatile(
505 "pxor %%mm7, %%mm7 \n\t" // 0 505 "pxor %%mm7, %%mm7 \n\t" // 0
506 // "movq b80, %%mm6 \n\t" // MIN_SIGNED_BYTE 506 // "movq "MANGLE(b80)", %%mm6 \n\t" // MIN_SIGNED_BYTE
507 "leal (%0, %1), %%eax \n\t" 507 "leal (%0, %1), %%eax \n\t"
508 "leal (%%eax, %1, 4), %%ebx \n\t" 508 "leal (%%eax, %1, 4), %%ebx \n\t"
509 // 0 1 2 3 4 5 6 7 8 9 509 // 0 1 2 3 4 5 6 7 8 9
510 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 510 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1
511 "movq (%%eax, %1, 2), %%mm0 \n\t" // line 3 511 "movq (%%eax, %1, 2), %%mm0 \n\t" // line 3
527 "pcmpeqb %%mm7, %%mm2 \n\t" // (l4 - l5) <= 0 ? -1 : 0 527 "pcmpeqb %%mm7, %%mm2 \n\t" // (l4 - l5) <= 0 ? -1 : 0
528 "psubusb %%mm1, %%mm5 \n\t" 528 "psubusb %%mm1, %%mm5 \n\t"
529 "por %%mm5, %%mm4 \n\t" // |l4 - l5| 529 "por %%mm5, %%mm4 \n\t" // |l4 - l5|
530 "psubusb %%mm0, %%mm4 \n\t" //d = MAX(0, |l4-l5| - (|l2-l3| + |l5-l6|)/2) 530 "psubusb %%mm0, %%mm4 \n\t" //d = MAX(0, |l4-l5| - (|l2-l3| + |l5-l6|)/2)
531 "movq %%mm4, %%mm3 \n\t" // d 531 "movq %%mm4, %%mm3 \n\t" // d
532 "psubusb pQPb, %%mm4 \n\t" 532 "psubusb "MANGLE(pQPb)", %%mm4 \n\t"
533 "pcmpeqb %%mm7, %%mm4 \n\t" // d <= QP ? -1 : 0 533 "pcmpeqb %%mm7, %%mm4 \n\t" // d <= QP ? -1 : 0
534 "psubusb b01, %%mm3 \n\t" 534 "psubusb "MANGLE(b01)", %%mm3 \n\t"
535 "pand %%mm4, %%mm3 \n\t" // d <= QP ? d : 0 535 "pand %%mm4, %%mm3 \n\t" // d <= QP ? d : 0
536 536
537 PAVGB(%%mm7, %%mm3) // d/2 537 PAVGB(%%mm7, %%mm3) // d/2
538 "movq %%mm3, %%mm1 \n\t" // d/2 538 "movq %%mm3, %%mm1 \n\t" // d/2
539 PAVGB(%%mm7, %%mm3) // d/4 539 PAVGB(%%mm7, %%mm3) // d/4
738 "por %%mm6, %%mm2 \n\t" // ~|2l4 - 5l5 + 5l6 - 2l7|/8 738 "por %%mm6, %%mm2 \n\t" // ~|2l4 - 5l5 + 5l6 - 2l7|/8
739 // mm0= SIGN(menergy), mm1= |lenergy|/8, mm2= |renergy|/8, mm3= |menergy|/8, mm7=0 739 // mm0= SIGN(menergy), mm1= |lenergy|/8, mm2= |renergy|/8, mm3= |menergy|/8, mm7=0
740 740
741 741
742 PMINUB(%%mm2, %%mm1, %%mm4) // MIN(|lenergy|,|renergy|)/8 742 PMINUB(%%mm2, %%mm1, %%mm4) // MIN(|lenergy|,|renergy|)/8
743 "movq pQPb, %%mm4 \n\t" // QP //FIXME QP+1 ? 743 "movq "MANGLE(pQPb)", %%mm4 \n\t" // QP //FIXME QP+1 ?
744 "paddusb b01, %%mm4 \n\t" 744 "paddusb "MANGLE(b01)", %%mm4 \n\t"
745 "pcmpgtb %%mm3, %%mm4 \n\t" // |menergy|/8 < QP 745 "pcmpgtb %%mm3, %%mm4 \n\t" // |menergy|/8 < QP
746 "psubusb %%mm1, %%mm3 \n\t" // d=|menergy|/8-MIN(|lenergy|,|renergy|)/8 746 "psubusb %%mm1, %%mm3 \n\t" // d=|menergy|/8-MIN(|lenergy|,|renergy|)/8
747 "pand %%mm4, %%mm3 \n\t" 747 "pand %%mm4, %%mm3 \n\t"
748 748
749 "movq %%mm3, %%mm1 \n\t" 749 "movq %%mm3, %%mm1 \n\t"
750 // "psubusb b01, %%mm3 \n\t" 750 // "psubusb "MANGLE(b01)", %%mm3 \n\t"
751 PAVGB(%%mm7, %%mm3) 751 PAVGB(%%mm7, %%mm3)
752 PAVGB(%%mm7, %%mm3) 752 PAVGB(%%mm7, %%mm3)
753 "paddusb %%mm1, %%mm3 \n\t" 753 "paddusb %%mm1, %%mm3 \n\t"
754 // "paddusb b01, %%mm3 \n\t" 754 // "paddusb "MANGLE(b01)", %%mm3 \n\t"
755 755
756 "movq (%%eax, %1, 2), %%mm6 \n\t" //l3 756 "movq (%%eax, %1, 2), %%mm6 \n\t" //l3
757 "movq (%0, %1, 4), %%mm5 \n\t" //l4 757 "movq (%0, %1, 4), %%mm5 \n\t" //l4
758 "movq (%0, %1, 4), %%mm4 \n\t" //l4 758 "movq (%0, %1, 4), %%mm4 \n\t" //l4
759 "psubusb %%mm6, %%mm5 \n\t" 759 "psubusb %%mm6, %%mm5 \n\t"
762 "pcmpeqb %%mm7, %%mm6 \n\t" // SIGN(l3-l4) 762 "pcmpeqb %%mm7, %%mm6 \n\t" // SIGN(l3-l4)
763 "pxor %%mm6, %%mm0 \n\t" 763 "pxor %%mm6, %%mm0 \n\t"
764 "pand %%mm0, %%mm3 \n\t" 764 "pand %%mm0, %%mm3 \n\t"
765 PMINUB(%%mm5, %%mm3, %%mm0) 765 PMINUB(%%mm5, %%mm3, %%mm0)
766 766
767 "psubusb b01, %%mm3 \n\t" 767 "psubusb "MANGLE(b01)", %%mm3 \n\t"
768 PAVGB(%%mm7, %%mm3) 768 PAVGB(%%mm7, %%mm3)
769 769
770 "movq (%%eax, %1, 2), %%mm0 \n\t" 770 "movq (%%eax, %1, 2), %%mm0 \n\t"
771 "movq (%0, %1, 4), %%mm2 \n\t" 771 "movq (%0, %1, 4), %%mm2 \n\t"
772 "pxor %%mm6, %%mm0 \n\t" 772 "pxor %%mm6, %%mm0 \n\t"
794 794
795 "movq (%%eax, %1, 4), %%mm2 \n\t" // l5 795 "movq (%%eax, %1, 4), %%mm2 \n\t" // l5
796 "movq (%%eax, %1), %%mm3 \n\t" // l2 796 "movq (%%eax, %1), %%mm3 \n\t" // l2
797 "pxor %%mm6, %%mm2 \n\t" // -l5-1 797 "pxor %%mm6, %%mm2 \n\t" // -l5-1
798 "movq %%mm2, %%mm5 \n\t" // -l5-1 798 "movq %%mm2, %%mm5 \n\t" // -l5-1
799 "movq b80, %%mm4 \n\t" // 128 799 "movq "MANGLE(b80)", %%mm4 \n\t" // 128
800 "leal (%%eax, %1, 4), %%ebx \n\t" 800 "leal (%%eax, %1, 4), %%ebx \n\t"
801 PAVGB(%%mm3, %%mm2) // (l2-l5+256)/2 801 PAVGB(%%mm3, %%mm2) // (l2-l5+256)/2
802 PAVGB(%%mm0, %%mm4) // ~(l4-l3)/4 + 128 802 PAVGB(%%mm0, %%mm4) // ~(l4-l3)/4 + 128
803 PAVGB(%%mm2, %%mm4) // ~(l2-l5)/4 +(l4-l3)/8 + 128 803 PAVGB(%%mm2, %%mm4) // ~(l2-l5)/4 +(l4-l3)/8 + 128
804 PAVGB(%%mm0, %%mm4) // ~(l2-l5)/8 +5(l4-l3)/16 + 128 804 PAVGB(%%mm0, %%mm4) // ~(l2-l5)/8 +5(l4-l3)/16 + 128
806 806
807 "movq (%%eax), %%mm2 \n\t" // l1 807 "movq (%%eax), %%mm2 \n\t" // l1
808 "pxor %%mm6, %%mm2 \n\t" // -l1-1 808 "pxor %%mm6, %%mm2 \n\t" // -l1-1
809 PAVGB(%%mm3, %%mm2) // (l2-l1+256)/2 809 PAVGB(%%mm3, %%mm2) // (l2-l1+256)/2
810 PAVGB((%0), %%mm1) // (l0-l3+256)/2 810 PAVGB((%0), %%mm1) // (l0-l3+256)/2
811 "movq b80, %%mm3 \n\t" // 128 811 "movq "MANGLE(b80)", %%mm3 \n\t" // 128
812 PAVGB(%%mm2, %%mm3) // ~(l2-l1)/4 + 128 812 PAVGB(%%mm2, %%mm3) // ~(l2-l1)/4 + 128
813 PAVGB(%%mm1, %%mm3) // ~(l0-l3)/4 +(l2-l1)/8 + 128 813 PAVGB(%%mm1, %%mm3) // ~(l0-l3)/4 +(l2-l1)/8 + 128
814 PAVGB(%%mm2, %%mm3) // ~(l0-l3)/8 +5(l2-l1)/16 + 128 814 PAVGB(%%mm2, %%mm3) // ~(l0-l3)/8 +5(l2-l1)/16 + 128
815 // mm0=128-q, mm3=lenergy/16 + 128, mm4= menergy/16 + 128, mm5= -l5-1 815 // mm0=128-q, mm3=lenergy/16 + 128, mm4= menergy/16 + 128, mm5= -l5-1
816 816
817 PAVGB((%%ebx, %1), %%mm5) // (l6-l5+256)/2 817 PAVGB((%%ebx, %1), %%mm5) // (l6-l5+256)/2
818 "movq (%%ebx, %1, 2), %%mm1 \n\t" // l7 818 "movq (%%ebx, %1, 2), %%mm1 \n\t" // l7
819 "pxor %%mm6, %%mm1 \n\t" // -l7-1 819 "pxor %%mm6, %%mm1 \n\t" // -l7-1
820 PAVGB((%0, %1, 4), %%mm1) // (l4-l7+256)/2 820 PAVGB((%0, %1, 4), %%mm1) // (l4-l7+256)/2
821 "movq b80, %%mm2 \n\t" // 128 821 "movq "MANGLE(b80)", %%mm2 \n\t" // 128
822 PAVGB(%%mm5, %%mm2) // ~(l6-l5)/4 + 128 822 PAVGB(%%mm5, %%mm2) // ~(l6-l5)/4 + 128
823 PAVGB(%%mm1, %%mm2) // ~(l4-l7)/4 +(l6-l5)/8 + 128 823 PAVGB(%%mm1, %%mm2) // ~(l4-l7)/4 +(l6-l5)/8 + 128
824 PAVGB(%%mm5, %%mm2) // ~(l4-l7)/8 +5(l6-l5)/16 + 128 824 PAVGB(%%mm5, %%mm2) // ~(l4-l7)/8 +5(l6-l5)/16 + 128
825 // mm0=128-q, mm2=renergy/16 + 128, mm3=lenergy/16 + 128, mm4= menergy/16 + 128 825 // mm0=128-q, mm2=renergy/16 + 128, mm3=lenergy/16 + 128, mm4= menergy/16 + 128
826 826
827 "movq b00, %%mm1 \n\t" // 0 827 "movq "MANGLE(b00)", %%mm1 \n\t" // 0
828 "movq b00, %%mm5 \n\t" // 0 828 "movq "MANGLE(b00)", %%mm5 \n\t" // 0
829 "psubb %%mm2, %%mm1 \n\t" // 128 - renergy/16 829 "psubb %%mm2, %%mm1 \n\t" // 128 - renergy/16
830 "psubb %%mm3, %%mm5 \n\t" // 128 - lenergy/16 830 "psubb %%mm3, %%mm5 \n\t" // 128 - lenergy/16
831 PMAXUB(%%mm1, %%mm2) // 128 + |renergy/16| 831 PMAXUB(%%mm1, %%mm2) // 128 + |renergy/16|
832 PMAXUB(%%mm5, %%mm3) // 128 + |lenergy/16| 832 PMAXUB(%%mm5, %%mm3) // 128 + |lenergy/16|
833 PMINUB(%%mm2, %%mm3, %%mm1) // 128 + MIN(|lenergy|,|renergy|)/16 833 PMINUB(%%mm2, %%mm3, %%mm1) // 128 + MIN(|lenergy|,|renergy|)/16
834 834
835 // mm0=128-q, mm3=128 + MIN(|lenergy|,|renergy|)/16, mm4= menergy/16 + 128 835 // mm0=128-q, mm3=128 + MIN(|lenergy|,|renergy|)/16, mm4= menergy/16 + 128
836 836
837 "movq b00, %%mm7 \n\t" // 0 837 "movq "MANGLE(b00)", %%mm7 \n\t" // 0
838 "movq pQPb, %%mm2 \n\t" // QP 838 "movq "MANGLE(pQPb)", %%mm2 \n\t" // QP
839 PAVGB(%%mm6, %%mm2) // 128 + QP/2 839 PAVGB(%%mm6, %%mm2) // 128 + QP/2
840 "psubb %%mm6, %%mm2 \n\t" 840 "psubb %%mm6, %%mm2 \n\t"
841 841
842 "movq %%mm4, %%mm1 \n\t" 842 "movq %%mm4, %%mm1 \n\t"
843 "pcmpgtb %%mm7, %%mm1 \n\t" // SIGN(menergy) 843 "pcmpgtb %%mm7, %%mm1 \n\t" // SIGN(menergy)
846 "pcmpgtb %%mm4, %%mm2 \n\t" // |menergy|/16 < QP/2 846 "pcmpgtb %%mm4, %%mm2 \n\t" // |menergy|/16 < QP/2
847 "psubusb %%mm3, %%mm4 \n\t" //d=|menergy|/16 - MIN(|lenergy|,|renergy|)/16 847 "psubusb %%mm3, %%mm4 \n\t" //d=|menergy|/16 - MIN(|lenergy|,|renergy|)/16
848 // mm0=128-q, mm1= SIGN(menergy), mm2= |menergy|/16 < QP/2, mm4= d/16 848 // mm0=128-q, mm1= SIGN(menergy), mm2= |menergy|/16 < QP/2, mm4= d/16
849 849
850 "movq %%mm4, %%mm3 \n\t" // d 850 "movq %%mm4, %%mm3 \n\t" // d
851 "psubusb b01, %%mm4 \n\t" 851 "psubusb "MANGLE(b01)", %%mm4 \n\t"
852 PAVGB(%%mm7, %%mm4) // d/32 852 PAVGB(%%mm7, %%mm4) // d/32
853 PAVGB(%%mm7, %%mm4) // (d + 32)/64 853 PAVGB(%%mm7, %%mm4) // (d + 32)/64
854 "paddb %%mm3, %%mm4 \n\t" // 5d/64 854 "paddb %%mm3, %%mm4 \n\t" // 5d/64
855 "pand %%mm2, %%mm4 \n\t" 855 "pand %%mm2, %%mm4 \n\t"
856 856
857 "movq b80, %%mm5 \n\t" // 128 857 "movq "MANGLE(b80)", %%mm5 \n\t" // 128
858 "psubb %%mm0, %%mm5 \n\t" // q 858 "psubb %%mm0, %%mm5 \n\t" // q
859 "paddsb %%mm6, %%mm5 \n\t" // fix bad rounding 859 "paddsb %%mm6, %%mm5 \n\t" // fix bad rounding
860 "pcmpgtb %%mm5, %%mm7 \n\t" // SIGN(q) 860 "pcmpgtb %%mm5, %%mm7 \n\t" // SIGN(q)
861 "pxor %%mm7, %%mm5 \n\t" 861 "pxor %%mm7, %%mm5 \n\t"
862 862
989 989
990 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3 990 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3
991 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3 991 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3
992 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 992 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
993 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 993 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
994 "movq %%mm0, temp0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 994 "movq %%mm0, "MANGLE(temp0)" \n\t" // 2L0 - 5L1 + 5L2 - 2L3
995 "movq %%mm1, temp1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 995 "movq %%mm1, "MANGLE(temp1)" \n\t" // 2H0 - 5H1 + 5H2 - 2H3
996 996
997 "movq (%0, %1, 4), %%mm0 \n\t" 997 "movq (%0, %1, 4), %%mm0 \n\t"
998 "movq %%mm0, %%mm1 \n\t" 998 "movq %%mm0, %%mm1 \n\t"
999 "punpcklbw %%mm7, %%mm0 \n\t" // L4 999 "punpcklbw %%mm7, %%mm0 \n\t" // L4
1000 "punpckhbw %%mm7, %%mm1 \n\t" // H4 1000 "punpckhbw %%mm7, %%mm1 \n\t" // H4
1001 1001
1002 "psubw %%mm0, %%mm2 \n\t" // L3 - L4 1002 "psubw %%mm0, %%mm2 \n\t" // L3 - L4
1003 "psubw %%mm1, %%mm3 \n\t" // H3 - H4 1003 "psubw %%mm1, %%mm3 \n\t" // H3 - H4
1004 "movq %%mm2, temp2 \n\t" // L3 - L4 1004 "movq %%mm2, "MANGLE(temp2)" \n\t" // L3 - L4
1005 "movq %%mm3, temp3 \n\t" // H3 - H4 1005 "movq %%mm3, "MANGLE(temp3)" \n\t" // H3 - H4
1006 "paddw %%mm4, %%mm4 \n\t" // 2L2 1006 "paddw %%mm4, %%mm4 \n\t" // 2L2
1007 "paddw %%mm5, %%mm5 \n\t" // 2H2 1007 "paddw %%mm5, %%mm5 \n\t" // 2H2
1008 "psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4 1008 "psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4
1009 "psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4 1009 "psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4
1010 1010
1047 "paddw %%mm2, %%mm2 \n\t" // 2L7 1047 "paddw %%mm2, %%mm2 \n\t" // 2L7
1048 "paddw %%mm3, %%mm3 \n\t" // 2H7 1048 "paddw %%mm3, %%mm3 \n\t" // 2H7
1049 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7 1049 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7
1050 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7 1050 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7
1051 1051
1052 "movq temp0, %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 1052 "movq "MANGLE(temp0)", %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
1053 "movq temp1, %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 1053 "movq "MANGLE(temp1)", %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
1054 1054
1055 #ifdef HAVE_MMX2 1055 #ifdef HAVE_MMX2
1056 "movq %%mm7, %%mm6 \n\t" // 0 1056 "movq %%mm7, %%mm6 \n\t" // 0
1057 "psubw %%mm0, %%mm6 \n\t" 1057 "psubw %%mm0, %%mm6 \n\t"
1058 "pmaxsw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7| 1058 "pmaxsw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
1136 //FIXME if *5/64 is supposed to be /13 then we should use 5041 instead of 5120 1136 //FIXME if *5/64 is supposed to be /13 then we should use 5041 instead of 5120
1137 "pmulhw %%mm2, %%mm4 \n\t" // hd/13 1137 "pmulhw %%mm2, %%mm4 \n\t" // hd/13
1138 "pmulhw %%mm2, %%mm5 \n\t" // ld/13 1138 "pmulhw %%mm2, %%mm5 \n\t" // ld/13
1139 */ 1139 */
1140 1140
1141 "movq temp2, %%mm0 \n\t" // L3 - L4 1141 "movq "MANGLE(temp2)", %%mm0 \n\t" // L3 - L4
1142 "movq temp3, %%mm1 \n\t" // H3 - H4 1142 "movq "MANGLE(temp3)", %%mm1 \n\t" // H3 - H4
1143 1143
1144 "pxor %%mm2, %%mm2 \n\t" 1144 "pxor %%mm2, %%mm2 \n\t"
1145 "pxor %%mm3, %%mm3 \n\t" 1145 "pxor %%mm3, %%mm3 \n\t"
1146 1146
1147 "pcmpgtw %%mm0, %%mm2 \n\t" // sign (L3-L4) 1147 "pcmpgtw %%mm0, %%mm2 \n\t" // sign (L3-L4)
1233 1233
1234 static inline void RENAME(dering)(uint8_t src[], int stride, int QP) 1234 static inline void RENAME(dering)(uint8_t src[], int stride, int QP)
1235 { 1235 {
1236 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) 1236 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1237 asm volatile( 1237 asm volatile(
1238 "movq pQPb, %%mm0 \n\t" 1238 "movq "MANGLE(pQPb)", %%mm0 \n\t"
1239 "paddusb %%mm0, %%mm0 \n\t" 1239 "paddusb %%mm0, %%mm0 \n\t"
1240 "movq %%mm0, pQPb2 \n\t" 1240 "movq %%mm0, "MANGLE(pQPb2)" \n\t"
1241 1241
1242 "leal (%0, %1), %%eax \n\t" 1242 "leal (%0, %1), %%eax \n\t"
1243 "leal (%%eax, %1, 4), %%ebx \n\t" 1243 "leal (%%eax, %1, 4), %%ebx \n\t"
1244 // 0 1 2 3 4 5 6 7 8 9 1244 // 0 1 2 3 4 5 6 7 8 9
1245 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 1245 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1
1317 "paddb %%mm4, %%mm6 \n\t" 1317 "paddb %%mm4, %%mm6 \n\t"
1318 #endif 1318 #endif
1319 "movq %%mm6, %%mm0 \n\t" // max 1319 "movq %%mm6, %%mm0 \n\t" // max
1320 "psubb %%mm7, %%mm6 \n\t" // max - min 1320 "psubb %%mm7, %%mm6 \n\t" // max - min
1321 "movd %%mm6, %%ecx \n\t" 1321 "movd %%mm6, %%ecx \n\t"
1322 "cmpb deringThreshold, %%cl \n\t" 1322 "cmpb "MANGLE(deringThreshold)", %%cl \n\t"
1323 " jb 1f \n\t" 1323 " jb 1f \n\t"
1324 PAVGB(%%mm0, %%mm7) // a=(max + min)/2 1324 PAVGB(%%mm0, %%mm7) // a=(max + min)/2
1325 "punpcklbw %%mm7, %%mm7 \n\t" 1325 "punpcklbw %%mm7, %%mm7 \n\t"
1326 "punpcklbw %%mm7, %%mm7 \n\t" 1326 "punpcklbw %%mm7, %%mm7 \n\t"
1327 "punpcklbw %%mm7, %%mm7 \n\t" 1327 "punpcklbw %%mm7, %%mm7 \n\t"
1328 "movq %%mm7, temp0 \n\t" 1328 "movq %%mm7, "MANGLE(temp0)" \n\t"
1329 1329
1330 "movq (%0), %%mm0 \n\t" // L10 1330 "movq (%0), %%mm0 \n\t" // L10
1331 "movq %%mm0, %%mm1 \n\t" // L10 1331 "movq %%mm0, %%mm1 \n\t" // L10
1332 "movq %%mm0, %%mm2 \n\t" // L10 1332 "movq %%mm0, %%mm2 \n\t" // L10
1333 "psllq $8, %%mm1 \n\t" 1333 "psllq $8, %%mm1 \n\t"
1342 PAVGB(%%mm2, %%mm1) // (L20 + L00)/2 1342 PAVGB(%%mm2, %%mm1) // (L20 + L00)/2
1343 PAVGB(%%mm0, %%mm1) // (L20 + L00 + 2L10)/4 1343 PAVGB(%%mm0, %%mm1) // (L20 + L00 + 2L10)/4
1344 "psubusb %%mm7, %%mm0 \n\t" 1344 "psubusb %%mm7, %%mm0 \n\t"
1345 "psubusb %%mm7, %%mm2 \n\t" 1345 "psubusb %%mm7, %%mm2 \n\t"
1346 "psubusb %%mm7, %%mm3 \n\t" 1346 "psubusb %%mm7, %%mm3 \n\t"
1347 "pcmpeqb b00, %%mm0 \n\t" // L10 > a ? 0 : -1 1347 "pcmpeqb "MANGLE(b00)", %%mm0 \n\t" // L10 > a ? 0 : -1
1348 "pcmpeqb b00, %%mm2 \n\t" // L20 > a ? 0 : -1 1348 "pcmpeqb "MANGLE(b00)", %%mm2 \n\t" // L20 > a ? 0 : -1
1349 "pcmpeqb b00, %%mm3 \n\t" // L00 > a ? 0 : -1 1349 "pcmpeqb "MANGLE(b00)", %%mm3 \n\t" // L00 > a ? 0 : -1
1350 "paddb %%mm2, %%mm0 \n\t" 1350 "paddb %%mm2, %%mm0 \n\t"
1351 "paddb %%mm3, %%mm0 \n\t" 1351 "paddb %%mm3, %%mm0 \n\t"
1352 1352
1353 "movq (%%eax), %%mm2 \n\t" // L11 1353 "movq (%%eax), %%mm2 \n\t" // L11
1354 "movq %%mm2, %%mm3 \n\t" // L11 1354 "movq %%mm2, %%mm3 \n\t" // L11
1365 PAVGB(%%mm4, %%mm3) // (L21 + L01)/2 1365 PAVGB(%%mm4, %%mm3) // (L21 + L01)/2
1366 PAVGB(%%mm2, %%mm3) // (L21 + L01 + 2L11)/4 1366 PAVGB(%%mm2, %%mm3) // (L21 + L01 + 2L11)/4
1367 "psubusb %%mm7, %%mm2 \n\t" 1367 "psubusb %%mm7, %%mm2 \n\t"
1368 "psubusb %%mm7, %%mm4 \n\t" 1368 "psubusb %%mm7, %%mm4 \n\t"
1369 "psubusb %%mm7, %%mm5 \n\t" 1369 "psubusb %%mm7, %%mm5 \n\t"
1370 "pcmpeqb b00, %%mm2 \n\t" // L11 > a ? 0 : -1 1370 "pcmpeqb "MANGLE(b00)", %%mm2 \n\t" // L11 > a ? 0 : -1
1371 "pcmpeqb b00, %%mm4 \n\t" // L21 > a ? 0 : -1 1371 "pcmpeqb "MANGLE(b00)", %%mm4 \n\t" // L21 > a ? 0 : -1
1372 "pcmpeqb b00, %%mm5 \n\t" // L01 > a ? 0 : -1 1372 "pcmpeqb "MANGLE(b00)", %%mm5 \n\t" // L01 > a ? 0 : -1
1373 "paddb %%mm4, %%mm2 \n\t" 1373 "paddb %%mm4, %%mm2 \n\t"
1374 "paddb %%mm5, %%mm2 \n\t" 1374 "paddb %%mm5, %%mm2 \n\t"
1375 // 0, 2, 3, 1 1375 // 0, 2, 3, 1
1376 #define DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \ 1376 #define DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \
1377 "movq " #src ", " #sx " \n\t" /* src[0] */\ 1377 "movq " #src ", " #sx " \n\t" /* src[0] */\
1387 "por " #t1 ", " #t0 " \n\t" /* src[+1] */\ 1387 "por " #t1 ", " #t0 " \n\t" /* src[+1] */\
1388 "movq " #lx ", " #t1 " \n\t" /* src[-1] */\ 1388 "movq " #lx ", " #t1 " \n\t" /* src[-1] */\
1389 PAVGB(t0, lx) /* (src[-1] + src[+1])/2 */\ 1389 PAVGB(t0, lx) /* (src[-1] + src[+1])/2 */\
1390 PAVGB(sx, lx) /* (src[-1] + 2src[0] + src[+1])/4 */\ 1390 PAVGB(sx, lx) /* (src[-1] + 2src[0] + src[+1])/4 */\
1391 PAVGB(lx, pplx) \ 1391 PAVGB(lx, pplx) \
1392 "movq " #lx ", temp1 \n\t"\ 1392 "movq " #lx ", "MANGLE(temp1)" \n\t"\
1393 "movq temp0, " #lx " \n\t"\ 1393 "movq "MANGLE(temp0)", " #lx " \n\t"\
1394 "psubusb " #lx ", " #t1 " \n\t"\ 1394 "psubusb " #lx ", " #t1 " \n\t"\
1395 "psubusb " #lx ", " #t0 " \n\t"\ 1395 "psubusb " #lx ", " #t0 " \n\t"\
1396 "psubusb " #lx ", " #sx " \n\t"\ 1396 "psubusb " #lx ", " #sx " \n\t"\
1397 "movq b00, " #lx " \n\t"\ 1397 "movq "MANGLE(b00)", " #lx " \n\t"\
1398 "pcmpeqb " #lx ", " #t1 " \n\t" /* src[-1] > a ? 0 : -1*/\ 1398 "pcmpeqb " #lx ", " #t1 " \n\t" /* src[-1] > a ? 0 : -1*/\
1399 "pcmpeqb " #lx ", " #t0 " \n\t" /* src[+1] > a ? 0 : -1*/\ 1399 "pcmpeqb " #lx ", " #t0 " \n\t" /* src[+1] > a ? 0 : -1*/\
1400 "pcmpeqb " #lx ", " #sx " \n\t" /* src[0] > a ? 0 : -1*/\ 1400 "pcmpeqb " #lx ", " #sx " \n\t" /* src[0] > a ? 0 : -1*/\
1401 "paddb " #t1 ", " #t0 " \n\t"\ 1401 "paddb " #t1 ", " #t0 " \n\t"\
1402 "paddb " #t0 ", " #sx " \n\t"\ 1402 "paddb " #t0 ", " #sx " \n\t"\
1403 \ 1403 \
1404 PAVGB(plx, pplx) /* filtered */\ 1404 PAVGB(plx, pplx) /* filtered */\
1405 "movq " #dst ", " #t0 " \n\t" /* dst */\ 1405 "movq " #dst ", " #t0 " \n\t" /* dst */\
1406 "movq " #t0 ", " #t1 " \n\t" /* dst */\ 1406 "movq " #t0 ", " #t1 " \n\t" /* dst */\
1407 "psubusb pQPb2, " #t0 " \n\t"\ 1407 "psubusb "MANGLE(pQPb2)", " #t0 " \n\t"\
1408 "paddusb pQPb2, " #t1 " \n\t"\ 1408 "paddusb "MANGLE(pQPb2)", " #t1 " \n\t"\
1409 PMAXUB(t0, pplx)\ 1409 PMAXUB(t0, pplx)\
1410 PMINUB(t1, pplx, t0)\ 1410 PMINUB(t1, pplx, t0)\
1411 "paddb " #sx ", " #ppsx " \n\t"\ 1411 "paddb " #sx ", " #ppsx " \n\t"\
1412 "paddb " #psx ", " #ppsx " \n\t"\ 1412 "paddb " #psx ", " #ppsx " \n\t"\
1413 "#paddb b02, " #ppsx " \n\t"\ 1413 "#paddb "MANGLE(b02)", " #ppsx " \n\t"\
1414 "pand b08, " #ppsx " \n\t"\ 1414 "pand "MANGLE(b08)", " #ppsx " \n\t"\
1415 "pcmpeqb " #lx ", " #ppsx " \n\t"\ 1415 "pcmpeqb " #lx ", " #ppsx " \n\t"\
1416 "pand " #ppsx ", " #pplx " \n\t"\ 1416 "pand " #ppsx ", " #pplx " \n\t"\
1417 "pandn " #dst ", " #ppsx " \n\t"\ 1417 "pandn " #dst ", " #ppsx " \n\t"\
1418 "por " #pplx ", " #ppsx " \n\t"\ 1418 "por " #pplx ", " #ppsx " \n\t"\
1419 "movq " #ppsx ", " #dst " \n\t"\ 1419 "movq " #ppsx ", " #dst " \n\t"\
1420 "movq temp1, " #lx " \n\t" 1420 "movq "MANGLE(temp1)", " #lx " \n\t"
1421 1421
1422 /* 1422 /*
1423 0000000 1423 0000000
1424 1111111 1424 1111111
1425 1425
2080 "paddw %%mm5, %%mm6 \n\t" 2080 "paddw %%mm5, %%mm6 \n\t"
2081 "paddw %%mm7, %%mm6 \n\t" 2081 "paddw %%mm7, %%mm6 \n\t"
2082 "paddw %%mm6, %%mm0 \n\t" 2082 "paddw %%mm6, %%mm0 \n\t"
2083 #elif defined (FAST_L2_DIFF) 2083 #elif defined (FAST_L2_DIFF)
2084 "pcmpeqb %%mm7, %%mm7 \n\t" 2084 "pcmpeqb %%mm7, %%mm7 \n\t"
2085 "movq b80, %%mm6 \n\t" 2085 "movq "MANGLE(b80)", %%mm6 \n\t"
2086 "pxor %%mm0, %%mm0 \n\t" 2086 "pxor %%mm0, %%mm0 \n\t"
2087 #define L2_DIFF_CORE(a, b)\ 2087 #define L2_DIFF_CORE(a, b)\
2088 "movq " #a ", %%mm5 \n\t"\ 2088 "movq " #a ", %%mm5 \n\t"\
2089 "movq " #b ", %%mm2 \n\t"\ 2089 "movq " #b ", %%mm2 \n\t"\
2090 "pxor %%mm7, %%mm2 \n\t"\ 2090 "pxor %%mm7, %%mm2 \n\t"\
2150 "addl 1024(%%ebx), %%ecx \n\t" 2150 "addl 1024(%%ebx), %%ecx \n\t"
2151 "shrl $3, %%ecx \n\t" 2151 "shrl $3, %%ecx \n\t"
2152 "movl %%ecx, (%%ebx) \n\t" 2152 "movl %%ecx, (%%ebx) \n\t"
2153 "leal (%%eax, %2, 2), %%ebx \n\t" // 5*stride 2153 "leal (%%eax, %2, 2), %%ebx \n\t" // 5*stride
2154 2154
2155 // "movl %3, %%ecx \n\t" 2155 // "movl %3, %%ecx \n\t"
2156 // "movl %%ecx, test \n\t" 2156 // "movl %%ecx, test \n\t"
2157 // "jmp 4f \n\t" 2157 // "jmp 4f \n\t"
2158 "cmpl 4+maxTmpNoise, %%ecx \n\t" 2158 "cmpl 4+"MANGLE(maxTmpNoise)", %%ecx \n\t"
2159 " jb 2f \n\t" 2159 " jb 2f \n\t"
2160 "cmpl 8+maxTmpNoise, %%ecx \n\t" 2160 "cmpl 8+"MANGLE(maxTmpNoise)", %%ecx \n\t"
2161 " jb 1f \n\t" 2161 " jb 1f \n\t"
2162 2162
2163 "leal (%%ebx, %2, 2), %%ecx \n\t" // 7*stride 2163 "leal (%%ebx, %2, 2), %%ecx \n\t" // 7*stride
2164 "movq (%0), %%mm0 \n\t" // L0 2164 "movq (%0), %%mm0 \n\t" // L0
2165 "movq (%0, %2), %%mm1 \n\t" // L1 2165 "movq (%0, %2), %%mm1 \n\t" // L1
2214 "movq %%mm6, (%0, %%eax, 2) \n\t" // L6 2214 "movq %%mm6, (%0, %%eax, 2) \n\t" // L6
2215 "movq %%mm7, (%0, %%ecx) \n\t" // L7 2215 "movq %%mm7, (%0, %%ecx) \n\t" // L7
2216 "jmp 4f \n\t" 2216 "jmp 4f \n\t"
2217 2217
2218 "2: \n\t" 2218 "2: \n\t"
2219 "cmpl maxTmpNoise, %%ecx \n\t" 2219 "cmpl "MANGLE(maxTmpNoise)", %%ecx \n\t"
2220 " jb 3f \n\t" 2220 " jb 3f \n\t"
2221 2221
2222 "leal (%%ebx, %2, 2), %%ecx \n\t" // 7*stride 2222 "leal (%%ebx, %2, 2), %%ecx \n\t" // 7*stride
2223 "movq (%0), %%mm0 \n\t" // L0 2223 "movq (%0), %%mm0 \n\t" // L0
2224 "movq (%0, %2), %%mm1 \n\t" // L1 2224 "movq (%0, %2), %%mm1 \n\t" // L1
2459 { 2459 {
2460 #ifdef HAVE_MMX 2460 #ifdef HAVE_MMX
2461 asm volatile( 2461 asm volatile(
2462 "leal (%0,%2), %%eax \n\t" 2462 "leal (%0,%2), %%eax \n\t"
2463 "leal (%1,%3), %%ebx \n\t" 2463 "leal (%1,%3), %%ebx \n\t"
2464 "movq packedYOffset, %%mm2 \n\t" 2464 "movq "MANGLE(packedYOffset)", %%mm2\n\t"
2465 "movq packedYScale, %%mm3 \n\t" 2465 "movq "MANGLE(packedYScale)", %%mm3\n\t"
2466 "pxor %%mm4, %%mm4 \n\t" 2466 "pxor %%mm4, %%mm4 \n\t"
2467 #ifdef HAVE_MMX2 2467 #ifdef HAVE_MMX2
2468 #define SCALED_CPY(src1, src2, dst1, dst2) \ 2468 #define SCALED_CPY(src1, src2, dst1, dst2) \
2469 "movq " #src1 ", %%mm0 \n\t"\ 2469 "movq " #src1 ", %%mm0 \n\t"\
2470 "movq " #src1 ", %%mm5 \n\t"\ 2470 "movq " #src1 ", %%mm5 \n\t"\
2882 asm volatile( 2882 asm volatile(
2883 "movd %0, %%mm7 \n\t" 2883 "movd %0, %%mm7 \n\t"
2884 "packuswb %%mm7, %%mm7 \n\t" // 0, 0, 0, QP, 0, 0, 0, QP 2884 "packuswb %%mm7, %%mm7 \n\t" // 0, 0, 0, QP, 0, 0, 0, QP
2885 "packuswb %%mm7, %%mm7 \n\t" // 0,QP, 0, QP, 0,QP, 0, QP 2885 "packuswb %%mm7, %%mm7 \n\t" // 0,QP, 0, QP, 0,QP, 0, QP
2886 "packuswb %%mm7, %%mm7 \n\t" // QP,..., QP 2886 "packuswb %%mm7, %%mm7 \n\t" // QP,..., QP
2887 "movq %%mm7, pQPb \n\t" 2887 "movq %%mm7, "MANGLE(pQPb)" \n\t"
2888 : : "r" (QP) 2888 : : "r" (QP)
2889 ); 2889 );
2890 #endif 2890 #endif
2891 2891
2892 #ifdef MORE_TIMING 2892 #ifdef MORE_TIMING