Mercurial > mplayer.hg

comparison liba52/imdct.c @ 18104:7b408d60de9e

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
add support for intel mac. mp3lib is not fixed yet.
author nplourde
date Sat, 15 Apr 2006 20:46:54 +0000
parents 72764c0dad8a
children 4bad7f00556e
comparison
equal deleted inserted replaced
18103:26ea12332e67 18104:7b408d60de9e
29 * michael did port them from libac3 (untested, perhaps totally broken) 29 * michael did port them from libac3 (untested, perhaps totally broken)
30 * AltiVec optimizations from Romain Dolbeau (romain@dolbeau.org) 30 * AltiVec optimizations from Romain Dolbeau (romain@dolbeau.org)
31 */ 31 */
32 32
33 #include "config.h" 33 #include "config.h"
34 #include "asmalign.h"
34 35
35 #include <math.h> 36 #include <math.h>
36 #include <stdio.h> 37 #include <stdio.h>
37 #ifndef M_PI 38 #ifndef M_PI
38 #define M_PI 3.1415926535897932384626433832795029 39 #define M_PI 3.1415926535897932384626433832795029
790 asm volatile( 791 asm volatile(
791 "xor %%"REG_S", %%"REG_S" \n\t" 792 "xor %%"REG_S", %%"REG_S" \n\t"
792 "lea "MANGLE(bit_reverse_512)", %%"REG_a"\n\t" 793 "lea "MANGLE(bit_reverse_512)", %%"REG_a"\n\t"
793 "mov $1008, %%"REG_D" \n\t" 794 "mov $1008, %%"REG_D" \n\t"
794 "push %%"REG_BP" \n\t" //use ebp without telling gcc 795 "push %%"REG_BP" \n\t" //use ebp without telling gcc
795 ".balign 16 \n\t" 796 ASMALIGN16
796 "1: \n\t" 797 "1: \n\t"
797 "movlps (%0, %%"REG_S"), %%xmm0 \n\t" // XXXI 798 "movlps (%0, %%"REG_S"), %%xmm0 \n\t" // XXXI
798 "movhps 8(%0, %%"REG_D"), %%xmm0 \n\t" // RXXI 799 "movhps 8(%0, %%"REG_D"), %%xmm0 \n\t" // RXXI
799 "movlps 8(%0, %%"REG_S"), %%xmm1 \n\t" // XXXi 800 "movlps 8(%0, %%"REG_S"), %%xmm1 \n\t" // XXXi
800 "movhps (%0, %%"REG_D"), %%xmm1 \n\t" // rXXi 801 "movhps (%0, %%"REG_D"), %%xmm1 \n\t" // rXXi
849 // Note w[0][0]={1,0} 850 // Note w[0][0]={1,0}
850 asm volatile( 851 asm volatile(
851 "xorps %%xmm1, %%xmm1 \n\t" 852 "xorps %%xmm1, %%xmm1 \n\t"
852 "xorps %%xmm2, %%xmm2 \n\t" 853 "xorps %%xmm2, %%xmm2 \n\t"
853 "mov %0, %%"REG_S" \n\t" 854 "mov %0, %%"REG_S" \n\t"
854 ".balign 16 \n\t" 855 ASMALIGN16
855 "1: \n\t" 856 "1: \n\t"
856 "movlps (%%"REG_S"), %%xmm0\n\t" //buf[p] 857 "movlps (%%"REG_S"), %%xmm0\n\t" //buf[p]
857 "movlps 8(%%"REG_S"), %%xmm1\n\t" //buf[q] 858 "movlps 8(%%"REG_S"), %%xmm1\n\t" //buf[q]
858 "movhps (%%"REG_S"), %%xmm0\n\t" //buf[p] 859 "movhps (%%"REG_S"), %%xmm0\n\t" //buf[p]
859 "movhps 8(%%"REG_S"), %%xmm2\n\t" //buf[q] 860 "movhps 8(%%"REG_S"), %%xmm2\n\t" //buf[q]
870 /* 2. iteration */ 871 /* 2. iteration */
871 // Note w[1]={{1,0}, {0,-1}} 872 // Note w[1]={{1,0}, {0,-1}}
872 asm volatile( 873 asm volatile(
873 "movaps "MANGLE(ps111_1)", %%xmm7\n\t" // 1,1,1,-1 874 "movaps "MANGLE(ps111_1)", %%xmm7\n\t" // 1,1,1,-1
874 "mov %0, %%"REG_S" \n\t" 875 "mov %0, %%"REG_S" \n\t"
875 ".balign 16 \n\t" 876 ASMALIGN16
876 "1: \n\t" 877 "1: \n\t"
877 "movaps 16(%%"REG_S"), %%xmm2 \n\t" //r2,i2,r3,i3 878 "movaps 16(%%"REG_S"), %%xmm2 \n\t" //r2,i2,r3,i3
878 "shufps $0xB4, %%xmm2, %%xmm2 \n\t" //r2,i2,i3,r3 879 "shufps $0xB4, %%xmm2, %%xmm2 \n\t" //r2,i2,i3,r3
879 "mulps %%xmm7, %%xmm2 \n\t" //r2,i2,i3,-r3 880 "mulps %%xmm7, %%xmm2 \n\t" //r2,i2,i3,-r3
880 "movaps (%%"REG_S"), %%xmm0 \n\t" //r0,i0,r1,i1 881 "movaps (%%"REG_S"), %%xmm0 \n\t" //r0,i0,r1,i1
901 "movaps 48+"MANGLE(sseW2)", %%xmm6\n\t" 902 "movaps 48+"MANGLE(sseW2)", %%xmm6\n\t"
902 "movaps 16+"MANGLE(sseW2)", %%xmm7\n\t" 903 "movaps 16+"MANGLE(sseW2)", %%xmm7\n\t"
903 "xorps %%xmm5, %%xmm5 \n\t" 904 "xorps %%xmm5, %%xmm5 \n\t"
904 "xorps %%xmm2, %%xmm2 \n\t" 905 "xorps %%xmm2, %%xmm2 \n\t"
905 "mov %0, %%"REG_S" \n\t" 906 "mov %0, %%"REG_S" \n\t"
906 ".balign 16 \n\t" 907 ASMALIGN16
907 "1: \n\t" 908 "1: \n\t"
908 "movaps 32(%%"REG_S"), %%xmm2 \n\t" //r4,i4,r5,i5 909 "movaps 32(%%"REG_S"), %%xmm2 \n\t" //r4,i4,r5,i5
909 "movaps 48(%%"REG_S"), %%xmm3 \n\t" //r6,i6,r7,i7 910 "movaps 48(%%"REG_S"), %%xmm3 \n\t" //r6,i6,r7,i7
910 "movaps "MANGLE(sseW2)", %%xmm4 \n\t" //r4,i4,r5,i5 911 "movaps "MANGLE(sseW2)", %%xmm4 \n\t" //r4,i4,r5,i5
911 "movaps 32+"MANGLE(sseW2)", %%xmm5\n\t" //r6,i6,r7,i7 912 "movaps 32+"MANGLE(sseW2)", %%xmm5\n\t" //r6,i6,r7,i7
942 two_m_plus_one = two_m<<1; 943 two_m_plus_one = two_m<<1;
943 two_m_plus_one_shl3 = (two_m_plus_one<<3); 944 two_m_plus_one_shl3 = (two_m_plus_one<<3);
944 buf_offset = buf+128; 945 buf_offset = buf+128;
945 asm volatile( 946 asm volatile(
946 "mov %0, %%"REG_S" \n\t" 947 "mov %0, %%"REG_S" \n\t"
947 ".balign 16 \n\t" 948 ASMALIGN16
948 "1: \n\t" 949 "1: \n\t"
949 "xor %%"REG_D", %%"REG_D" \n\t" // k 950 "xor %%"REG_D", %%"REG_D" \n\t" // k
950 "lea (%%"REG_S", %3), %%"REG_d" \n\t" 951 "lea (%%"REG_S", %3), %%"REG_d" \n\t"
951 "2: \n\t" 952 "2: \n\t"
952 "movaps (%%"REG_d", %%"REG_D"), %%xmm1 \n\t" 953 "movaps (%%"REG_d", %%"REG_D"), %%xmm1 \n\t"
974 } 975 }
975 976
976 /* Post IFFT complex multiply plus IFFT complex conjugate*/ 977 /* Post IFFT complex multiply plus IFFT complex conjugate*/
977 asm volatile( 978 asm volatile(
978 "mov $-1024, %%"REG_S" \n\t" 979 "mov $-1024, %%"REG_S" \n\t"
979 ".balign 16 \n\t" 980 ASMALIGN16
980 "1: \n\t" 981 "1: \n\t"
981 "movaps (%0, %%"REG_S"), %%xmm0 \n\t" 982 "movaps (%0, %%"REG_S"), %%xmm0 \n\t"
982 "movaps (%0, %%"REG_S"), %%xmm1 \n\t" 983 "movaps (%0, %%"REG_S"), %%xmm1 \n\t"
983 "shufps $0xB1, %%xmm0, %%xmm0 \n\t" 984 "shufps $0xB1, %%xmm0, %%xmm0 \n\t"
984 "mulps 1024+"MANGLE(sseSinCos1c)"(%%"REG_S"), %%xmm1\n\t" 985 "mulps 1024+"MANGLE(sseSinCos1c)"(%%"REG_S"), %%xmm1\n\t"
1000 asm volatile( 1001 asm volatile(
1001 "xor %%"REG_D", %%"REG_D" \n\t" // 0 1002 "xor %%"REG_D", %%"REG_D" \n\t" // 0
1002 "xor %%"REG_S", %%"REG_S" \n\t" // 0 1003 "xor %%"REG_S", %%"REG_S" \n\t" // 0
1003 "movss %3, %%xmm2 \n\t" // bias 1004 "movss %3, %%xmm2 \n\t" // bias
1004 "shufps $0x00, %%xmm2, %%xmm2 \n\t" // bias, bias, ... 1005 "shufps $0x00, %%xmm2, %%xmm2 \n\t" // bias, bias, ...
1005 ".balign 16 \n\t" 1006 ASMALIGN16
1006 "1: \n\t" 1007 "1: \n\t"
1007 "movlps (%0, %%"REG_S"), %%xmm0 \n\t" // ? ? A ? 1008 "movlps (%0, %%"REG_S"), %%xmm0 \n\t" // ? ? A ?
1008 "movlps 8(%0, %%"REG_S"), %%xmm1 \n\t" // ? ? C ? 1009 "movlps 8(%0, %%"REG_S"), %%xmm1 \n\t" // ? ? C ?
1009 "movhps -16(%0, %%"REG_D"), %%xmm1 \n\t" // ? D C ? 1010 "movhps -16(%0, %%"REG_D"), %%xmm1 \n\t" // ? D C ?
1010 "movhps -8(%0, %%"REG_D"), %%xmm0 \n\t" // ? B A ? 1011 "movhps -8(%0, %%"REG_D"), %%xmm0 \n\t" // ? B A ?
1027 asm volatile( 1028 asm volatile(
1028 "mov $1024, %%"REG_D" \n\t" // 512 1029 "mov $1024, %%"REG_D" \n\t" // 512
1029 "xor %%"REG_S", %%"REG_S" \n\t" // 0 1030 "xor %%"REG_S", %%"REG_S" \n\t" // 0
1030 "movss %3, %%xmm2 \n\t" // bias 1031 "movss %3, %%xmm2 \n\t" // bias
1031 "shufps $0x00, %%xmm2, %%xmm2 \n\t" // bias, bias, ... 1032 "shufps $0x00, %%xmm2, %%xmm2 \n\t" // bias, bias, ...
1032 ".balign 16 \n\t" 1033 ASMALIGN16
1033 "1: \n\t" 1034 "1: \n\t"
1034 "movlps (%0, %%"REG_S"), %%xmm0 \n\t" // ? ? ? A 1035 "movlps (%0, %%"REG_S"), %%xmm0 \n\t" // ? ? ? A
1035 "movlps 8(%0, %%"REG_S"), %%xmm1 \n\t" // ? ? ? C 1036 "movlps 8(%0, %%"REG_S"), %%xmm1 \n\t" // ? ? ? C
1036 "movhps -16(%0, %%"REG_D"), %%xmm1 \n\t" // D ? ? C 1037 "movhps -16(%0, %%"REG_D"), %%xmm1 \n\t" // D ? ? C
1037 "movhps -8(%0, %%"REG_D"), %%xmm0 \n\t" // B ? ? A 1038 "movhps -8(%0, %%"REG_D"), %%xmm0 \n\t" // B ? ? A
1054 delay_ptr = delay; 1055 delay_ptr = delay;
1055 1056
1056 asm volatile( 1057 asm volatile(
1057 "xor %%"REG_D", %%"REG_D" \n\t" // 0 1058 "xor %%"REG_D", %%"REG_D" \n\t" // 0
1058 "xor %%"REG_S", %%"REG_S" \n\t" // 0 1059 "xor %%"REG_S", %%"REG_S" \n\t" // 0
1059 ".balign 16 \n\t" 1060 ASMALIGN16
1060 "1: \n\t" 1061 "1: \n\t"
1061 "movlps (%0, %%"REG_S"), %%xmm0 \n\t" // ? ? ? A 1062 "movlps (%0, %%"REG_S"), %%xmm0 \n\t" // ? ? ? A
1062 "movlps 8(%0, %%"REG_S"), %%xmm1 \n\t" // ? ? ? C 1063 "movlps 8(%0, %%"REG_S"), %%xmm1 \n\t" // ? ? ? C
1063 "movhps -16(%0, %%"REG_D"), %%xmm1 \n\t" // D ? ? C 1064 "movhps -16(%0, %%"REG_D"), %%xmm1 \n\t" // D ? ? C
1064 "movhps -8(%0, %%"REG_D"), %%xmm0 \n\t" // B ? ? A 1065 "movhps -8(%0, %%"REG_D"), %%xmm0 \n\t" // B ? ? A
1076 // window_ptr-=128; 1077 // window_ptr-=128;
1077 1078
1078 asm volatile( 1079 asm volatile(
1079 "mov $1024, %%"REG_D" \n\t" // 1024 1080 "mov $1024, %%"REG_D" \n\t" // 1024
1080 "xor %%"REG_S", %%"REG_S" \n\t" // 0 1081 "xor %%"REG_S", %%"REG_S" \n\t" // 0
1081 ".balign 16 \n\t" 1082 ASMALIGN16
1082 "1: \n\t" 1083 "1: \n\t"
1083 "movlps (%0, %%"REG_S"), %%xmm0 \n\t" // ? ? A ? 1084 "movlps (%0, %%"REG_S"), %%xmm0 \n\t" // ? ? A ?
1084 "movlps 8(%0, %%"REG_S"), %%xmm1 \n\t" // ? ? C ? 1085 "movlps 8(%0, %%"REG_S"), %%xmm1 \n\t" // ? ? C ?
1085 "movhps -16(%0, %%"REG_D"), %%xmm1 \n\t" // ? D C ? 1086 "movhps -16(%0, %%"REG_D"), %%xmm1 \n\t" // ? D C ?
1086 "movhps -8(%0, %%"REG_D"), %%xmm0 \n\t" // ? B A ? 1087 "movhps -8(%0, %%"REG_D"), %%xmm0 \n\t" // ? B A ?