Mercurial > mplayer.hg

comparison liba52/imdct.c @ 4247:2dbd637ffe05

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
mangle for win32 in liba52 (includes dummy mangle.h pointing to the one in main)
author atmos4
date Sat, 19 Jan 2002 05:12:34 +0000
parents 0cc94b1eec0f
children d3aedd7db02c
comparison
equal deleted inserted replaced
4246:3f677202418b 4247:2dbd637ffe05
35 #include <inttypes.h> 35 #include <inttypes.h>
36 36
37 #include "a52.h" 37 #include "a52.h"
38 #include "a52_internal.h" 38 #include "a52_internal.h"
39 #include "mm_accel.h" 39 #include "mm_accel.h"
40 #include "mangle.h"
40 41
41 #ifdef RUNTIME_CPUDETECT 42 #ifdef RUNTIME_CPUDETECT
42 #undef HAVE_3DNOWEX 43 #undef HAVE_3DNOWEX
43 #endif 44 #endif
44 45
658 659
659 /* Pre IFFT complex multiply plus IFFT cmplx conjugate */ 660 /* Pre IFFT complex multiply plus IFFT cmplx conjugate */
660 /* Bit reversed shuffling */ 661 /* Bit reversed shuffling */
661 asm volatile( 662 asm volatile(
662 "xorl %%esi, %%esi \n\t" 663 "xorl %%esi, %%esi \n\t"
663 "leal bit_reverse_512, %%eax \n\t" 664 "leal "MANGLE(bit_reverse_512)", %%eax \n\t"
664 "movl $1008, %%edi \n\t" 665 "movl $1008, %%edi \n\t"
665 "pushl %%ebp \n\t" //use ebp without telling gcc 666 "pushl %%ebp \n\t" //use ebp without telling gcc
666 ".balign 16 \n\t" 667 ".balign 16 \n\t"
667 "1: \n\t" 668 "1: \n\t"
668 "movlps (%0, %%esi), %%xmm0 \n\t" // XXXI 669 "movlps (%0, %%esi), %%xmm0 \n\t" // XXXI
669 "movhps 8(%0, %%edi), %%xmm0 \n\t" // RXXI 670 "movhps 8(%0, %%edi), %%xmm0 \n\t" // RXXI
670 "movlps 8(%0, %%esi), %%xmm1 \n\t" // XXXi 671 "movlps 8(%0, %%esi), %%xmm1 \n\t" // XXXi
671 "movhps (%0, %%edi), %%xmm1 \n\t" // rXXi 672 "movhps (%0, %%edi), %%xmm1 \n\t" // rXXi
672 "shufps $0x33, %%xmm1, %%xmm0 \n\t" // irIR 673 "shufps $0x33, %%xmm1, %%xmm0 \n\t" // irIR
673 "movaps sseSinCos1c(%%esi), %%xmm2 \n\t" 674 "movaps "MANGLE(sseSinCos1c)"(%%esi), %%xmm2\n\t"
674 "mulps %%xmm0, %%xmm2 \n\t" 675 "mulps %%xmm0, %%xmm2 \n\t"
675 "shufps $0xB1, %%xmm0, %%xmm0 \n\t" // riRI 676 "shufps $0xB1, %%xmm0, %%xmm0 \n\t" // riRI
676 "mulps sseSinCos1d(%%esi), %%xmm0 \n\t" 677 "mulps "MANGLE(sseSinCos1d)"(%%esi), %%xmm0\n\t"
677 "subps %%xmm0, %%xmm2 \n\t" 678 "subps %%xmm0, %%xmm2 \n\t"
678 "movzbl (%%eax), %%edx \n\t" 679 "movzbl (%%eax), %%edx \n\t"
679 "movzbl 1(%%eax), %%ebp \n\t" 680 "movzbl 1(%%eax), %%ebp \n\t"
680 "movlps %%xmm2, (%1, %%edx,8) \n\t" 681 "movlps %%xmm2, (%1, %%edx,8) \n\t"
681 "movhps %%xmm2, (%1, %%ebp,8) \n\t" 682 "movhps %%xmm2, (%1, %%ebp,8) \n\t"
739 ); 740 );
740 741
741 /* 2. iteration */ 742 /* 2. iteration */
742 // Note w[1]={{1,0}, {0,-1}} 743 // Note w[1]={{1,0}, {0,-1}}
743 asm volatile( 744 asm volatile(
744 "movaps ps111_1, %%xmm7 \n\t" // 1,1,1,-1 745 "movaps "MANGLE(ps111_1)", %%xmm7\n\t" // 1,1,1,-1
745 "movl %0, %%esi \n\t" 746 "movl %0, %%esi \n\t"
746 ".balign 16 \n\t" 747 ".balign 16 \n\t"
747 "1: \n\t" 748 "1: \n\t"
748 "movaps 16(%%esi), %%xmm2 \n\t" //r2,i2,r3,i3 749 "movaps 16(%%esi), %%xmm2 \n\t" //r2,i2,r3,i3
749 "shufps $0xB4, %%xmm2, %%xmm2 \n\t" //r2,i2,i3,r3 750 "shufps $0xB4, %%xmm2, %%xmm2 \n\t" //r2,i2,i3,r3
767 Note sseW2+16={0,0,sqrt(2),-sqrt(2)) 768 Note sseW2+16={0,0,sqrt(2),-sqrt(2))
768 Note sseW2+32={0,0,-sqrt(2),-sqrt(2)) 769 Note sseW2+32={0,0,-sqrt(2),-sqrt(2))
769 Note sseW2+48={1,-1,sqrt(2),-sqrt(2)) 770 Note sseW2+48={1,-1,sqrt(2),-sqrt(2))
770 */ 771 */
771 asm volatile( 772 asm volatile(
772 "movaps 48+sseW2, %%xmm6 \n\t" 773 "movaps 48+"MANGLE(sseW2)", %%xmm6\n\t"
773 "movaps 16+sseW2, %%xmm7 \n\t" 774 "movaps 16+"MANGLE(sseW2)", %%xmm7\n\t"
774 "xorps %%xmm5, %%xmm5 \n\t" 775 "xorps %%xmm5, %%xmm5 \n\t"
775 "xorps %%xmm2, %%xmm2 \n\t" 776 "xorps %%xmm2, %%xmm2 \n\t"
776 "movl %0, %%esi \n\t" 777 "movl %0, %%esi \n\t"
777 ".balign 16 \n\t" 778 ".balign 16 \n\t"
778 "1: \n\t" 779 "1: \n\t"
779 "movaps 32(%%esi), %%xmm2 \n\t" //r4,i4,r5,i5 780 "movaps 32(%%esi), %%xmm2 \n\t" //r4,i4,r5,i5
780 "movaps 48(%%esi), %%xmm3 \n\t" //r6,i6,r7,i7 781 "movaps 48(%%esi), %%xmm3 \n\t" //r6,i6,r7,i7
781 "movaps sseW2, %%xmm4 \n\t" //r4,i4,r5,i5 782 "movaps "MANGLE(sseW2)", %%xmm4 \n\t" //r4,i4,r5,i5
782 "movaps 32+sseW2, %%xmm5 \n\t" //r6,i6,r7,i7 783 "movaps 32+"MANGLE(sseW2)", %%xmm5\n\t" //r6,i6,r7,i7
783 "mulps %%xmm2, %%xmm4 \n\t" 784 "mulps %%xmm2, %%xmm4 \n\t"
784 "mulps %%xmm3, %%xmm5 \n\t" 785 "mulps %%xmm3, %%xmm5 \n\t"
785 "shufps $0xB1, %%xmm2, %%xmm2 \n\t" //i4,r4,i5,r5 786 "shufps $0xB1, %%xmm2, %%xmm2 \n\t" //i4,r4,i5,r5
786 "shufps $0xB1, %%xmm3, %%xmm3 \n\t" //i6,r6,i7,r7 787 "shufps $0xB1, %%xmm3, %%xmm3 \n\t" //i6,r6,i7,r7
787 "mulps %%xmm6, %%xmm3 \n\t" 788 "mulps %%xmm6, %%xmm3 \n\t"
842 ); 843 );
843 } 844 }
844 845
845 /* Post IFFT complex multiply plus IFFT complex conjugate*/ 846 /* Post IFFT complex multiply plus IFFT complex conjugate*/
846 asm volatile( 847 asm volatile(
847 "movl $-1024, %%esi \n\t" 848 "movl $-1024, %%esi \n\t"
848 ".balign 16 \n\t" 849 ".balign 16 \n\t"
849 "1: \n\t" 850 "1: \n\t"
850 "movaps (%0, %%esi), %%xmm0 \n\t" 851 "movaps (%0, %%esi), %%xmm0 \n\t"
851 "movaps (%0, %%esi), %%xmm1 \n\t" 852 "movaps (%0, %%esi), %%xmm1 \n\t"
852 "shufps $0xB1, %%xmm0, %%xmm0 \n\t" 853 "shufps $0xB1, %%xmm0, %%xmm0 \n\t"
853 "mulps 1024+sseSinCos1c(%%esi), %%xmm1 \n\t" 854 "mulps 1024+"MANGLE(sseSinCos1c)"(%%esi), %%xmm1\n\t"
854 "mulps 1024+sseSinCos1d(%%esi), %%xmm0 \n\t" 855 "mulps 1024+"MANGLE(sseSinCos1d)"(%%esi), %%xmm0\n\t"
855 "addps %%xmm1, %%xmm0 \n\t" 856 "addps %%xmm1, %%xmm0 \n\t"
856 "movaps %%xmm0, (%0, %%esi) \n\t" 857 "movaps %%xmm0, (%0, %%esi) \n\t"
857 "addl $16, %%esi \n\t" 858 "addl $16, %%esi \n\t"
858 " jnz 1b \n\t" 859 " jnz 1b \n\t"
859 :: "r" (buf+128) 860 :: "r" (buf+128)
876 "movlps (%0, %%esi), %%xmm0 \n\t" // ? ? A ? 877 "movlps (%0, %%esi), %%xmm0 \n\t" // ? ? A ?
877 "movlps 8(%0, %%esi), %%xmm1 \n\t" // ? ? C ? 878 "movlps 8(%0, %%esi), %%xmm1 \n\t" // ? ? C ?
878 "movhps -16(%0, %%edi), %%xmm1 \n\t" // ? D C ? 879 "movhps -16(%0, %%edi), %%xmm1 \n\t" // ? D C ?
879 "movhps -8(%0, %%edi), %%xmm0 \n\t" // ? B A ? 880 "movhps -8(%0, %%edi), %%xmm0 \n\t" // ? B A ?
880 "shufps $0x99, %%xmm1, %%xmm0 \n\t" // D C B A 881 "shufps $0x99, %%xmm1, %%xmm0 \n\t" // D C B A
881 "mulps sseWindow(%%esi), %%xmm0 \n\t" 882 "mulps "MANGLE(sseWindow)"(%%esi), %%xmm0\n\t"
882 "addps (%2, %%esi), %%xmm0 \n\t" 883 "addps (%2, %%esi), %%xmm0 \n\t"
883 "addps %%xmm2, %%xmm0 \n\t" 884 "addps %%xmm2, %%xmm0 \n\t"
884 "movaps %%xmm0, (%1, %%esi) \n\t" 885 "movaps %%xmm0, (%1, %%esi) \n\t"
885 "addl $16, %%esi \n\t" 886 "addl $16, %%esi \n\t"
886 "subl $16, %%edi \n\t" 887 "subl $16, %%edi \n\t"
903 "movlps (%0, %%esi), %%xmm0 \n\t" // ? ? ? A 904 "movlps (%0, %%esi), %%xmm0 \n\t" // ? ? ? A
904 "movlps 8(%0, %%esi), %%xmm1 \n\t" // ? ? ? C 905 "movlps 8(%0, %%esi), %%xmm1 \n\t" // ? ? ? C
905 "movhps -16(%0, %%edi), %%xmm1 \n\t" // D ? ? C 906 "movhps -16(%0, %%edi), %%xmm1 \n\t" // D ? ? C
906 "movhps -8(%0, %%edi), %%xmm0 \n\t" // B ? ? A 907 "movhps -8(%0, %%edi), %%xmm0 \n\t" // B ? ? A
907 "shufps $0xCC, %%xmm1, %%xmm0 \n\t" // D C B A 908 "shufps $0xCC, %%xmm1, %%xmm0 \n\t" // D C B A
908 "mulps 512+sseWindow(%%esi), %%xmm0 \n\t" 909 "mulps 512+"MANGLE(sseWindow)"(%%esi), %%xmm0\n\t"
909 "addps (%2, %%esi), %%xmm0 \n\t" 910 "addps (%2, %%esi), %%xmm0 \n\t"
910 "addps %%xmm2, %%xmm0 \n\t" 911 "addps %%xmm2, %%xmm0 \n\t"
911 "movaps %%xmm0, (%1, %%esi) \n\t" 912 "movaps %%xmm0, (%1, %%esi) \n\t"
912 "addl $16, %%esi \n\t" 913 "addl $16, %%esi \n\t"
913 "subl $16, %%edi \n\t" 914 "subl $16, %%edi \n\t"
930 "movlps (%0, %%esi), %%xmm0 \n\t" // ? ? ? A 931 "movlps (%0, %%esi), %%xmm0 \n\t" // ? ? ? A
931 "movlps 8(%0, %%esi), %%xmm1 \n\t" // ? ? ? C 932 "movlps 8(%0, %%esi), %%xmm1 \n\t" // ? ? ? C
932 "movhps -16(%0, %%edi), %%xmm1 \n\t" // D ? ? C 933 "movhps -16(%0, %%edi), %%xmm1 \n\t" // D ? ? C
933 "movhps -8(%0, %%edi), %%xmm0 \n\t" // B ? ? A 934 "movhps -8(%0, %%edi), %%xmm0 \n\t" // B ? ? A
934 "shufps $0xCC, %%xmm1, %%xmm0 \n\t" // D C B A 935 "shufps $0xCC, %%xmm1, %%xmm0 \n\t" // D C B A
935 "mulps 1024+sseWindow(%%esi), %%xmm0 \n\t" 936 "mulps 1024+"MANGLE(sseWindow)"(%%esi), %%xmm0\n\t"
936 "movaps %%xmm0, (%1, %%esi) \n\t" 937 "movaps %%xmm0, (%1, %%esi) \n\t"
937 "addl $16, %%esi \n\t" 938 "addl $16, %%esi \n\t"
938 "subl $16, %%edi \n\t" 939 "subl $16, %%edi \n\t"
939 "cmpl $512, %%esi \n\t" 940 "cmpl $512, %%esi \n\t"
940 " jb 1b \n\t" 941 " jb 1b \n\t"
952 "movlps (%0, %%esi), %%xmm0 \n\t" // ? ? A ? 953 "movlps (%0, %%esi), %%xmm0 \n\t" // ? ? A ?
953 "movlps 8(%0, %%esi), %%xmm1 \n\t" // ? ? C ? 954 "movlps 8(%0, %%esi), %%xmm1 \n\t" // ? ? C ?
954 "movhps -16(%0, %%edi), %%xmm1 \n\t" // ? D C ? 955 "movhps -16(%0, %%edi), %%xmm1 \n\t" // ? D C ?
955 "movhps -8(%0, %%edi), %%xmm0 \n\t" // ? B A ? 956 "movhps -8(%0, %%edi), %%xmm0 \n\t" // ? B A ?
956 "shufps $0x99, %%xmm1, %%xmm0 \n\t" // D C B A 957 "shufps $0x99, %%xmm1, %%xmm0 \n\t" // D C B A
957 "mulps 1536+sseWindow(%%esi), %%xmm0 \n\t" 958 "mulps 1536+"MANGLE(sseWindow)"(%%esi), %%xmm0\n\t"
958 "movaps %%xmm0, (%1, %%esi) \n\t" 959 "movaps %%xmm0, (%1, %%esi) \n\t"
959 "addl $16, %%esi \n\t" 960 "addl $16, %%esi \n\t"
960 "subl $16, %%edi \n\t" 961 "subl $16, %%edi \n\t"
961 "cmpl $512, %%esi \n\t" 962 "cmpl $512, %%esi \n\t"
962 " jb 1b \n\t" 963 " jb 1b \n\t"