Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.

git-svn-id: svn://svn.mplayerhq.hu/mplayer/trunk@1246 b3059339-0415-0410-9bf9-f77b7e298cf2

19 files changed, 3210 insertions, 2314 deletions

diff --git a/mp3lib/Makefile b/mp3lib/Makefile
index b82aa6215f..6aa93c4275 100644
--- a/mp3lib/Makefile
+++ b/mp3lib/Makefile
@@ -1,8 +1,10 @@
 
 include config.mak
 
-SRCS = sr1.c d_cpu.s decode_i586.s $(OPTIONAL_SRCS)
-OBJS = sr1.o d_cpu.o decode_i586.o $(OPTIONAL_OBJS)
+SRCS = sr1.c d_cpu.s decode_i586.s dct64_MMX.s decode_MMX.s tabinit_MMX.s\
+dct36_3dnow.s dct64_3dnow.s dct36_k7.s dct64_k7.s
+OBJS = sr1.o d_cpu.o decode_i586.o dct64_MMX.o decode_MMX.o tabinit_MMX.o\
+dct36_3dnow.o dct64_3dnow.o dct36_k7.o dct64_k7.o
 # OBJS = $(SRCS:.c,.s=.o)
 CFLAGS  = $(OPTFLAGS) $(EXTRA_INC)
 
diff --git a/mp3lib/d_cpu.h b/mp3lib/d_cpu.h
index d2c92b9415..3d221f66e4 100644
--- a/mp3lib/d_cpu.h
+++ b/mp3lib/d_cpu.h
@@ -9,9 +9,12 @@
 unsigned int _CpuID;
 unsigned int _i586;
 unsigned int _3dnow;
+unsigned int _isse;
+unsigned int _has_mmx;
 
 extern unsigned long CpuDetect( void );
 extern unsigned long ipentium( void );
+extern unsigned long isse( void );
 extern unsigned long a3dnow( void );
 
 #endif
diff --git a/mp3lib/d_cpu.s b/mp3lib/d_cpu.s
index 0715ccccd1..6df924b241 100644
--- a/mp3lib/d_cpu.s
+++ b/mp3lib/d_cpu.s
@@ -9,6 +9,7 @@
 .globl CpuDetect
 .globl ipentium
 .globl a3dnow
+.globl isse
 
 / ---------------------------------------------------------------------------
 /  in C: unsigned long CpuDetect( void );
@@ -45,7 +46,9 @@ exit_cpudetect:
 
 / ---------------------------------------------------------------------------
 /  in C: unsigled long ipentium( void );
-/   return: 0 if the processor is not P5 or above else above 1.
+/  return: 0 if this processor i386 or i486
+/          1 otherwise
+/          2 if this cpu supports mmx
 / ---------------------------------------------------------------------------
 ipentium:
         pushl  %ebx
@@ -63,10 +66,15 @@ ipentium:
         jz     no_cpuid
         movl   $1,%eax
         cpuid
-        shrl   $8,%eax
-        cmpl   $5,%eax
-        jb     no_cpuid
-        movl   $1,%eax
+	movl   %eax, %ecx
+	xorl   %eax, %eax
+        shrl   $8,%ecx
+        cmpl   $5,%ecx
+        jb     exit
+        incl   %eax
+	test   $0x00800000, %edx
+	jz     exit
+	incl   %eax
         jmp    exit
 no_cpuid:
         xorl   %eax,%eax
@@ -113,3 +121,33 @@ exit2:
         popl   %edx
         popl   %ebx
         ret
+
+/ ---------------------------------------------------------------------------
+/  in C: unsigned long isse( void );
+/  return: 0 if this processor does not support sse
+/          1 otherwise
+/          2 if this cpu supports sse2 extension
+/ ---------------------------------------------------------------------------
+isse:
+        pushl  %ebx
+        pushl  %edx
+        pushl  %ecx
+
+        call   ipentium
+        testl  %eax,%eax
+        jz     exit3
+
+        movl   $1,%eax
+        cpuid
+	xorl   %eax, %eax
+        testl  $0x02000000,%edx
+        jz     exit3
+	incl   %eax
+        testl  $0x04000000,%edx
+        jz     exit3
+        incl   %eax
+exit3:
+        popl   %ecx
+        popl   %edx
+        popl   %ebx
+        ret
diff --git a/mp3lib/dct36.c b/mp3lib/dct36.c
index 04992f09cc..18bb35a5c4 100644
--- a/mp3lib/dct36.c
+++ b/mp3lib/dct36.c
@@ -193,7 +193,7 @@ static void dct36(real *inbuf,real *o1,real *o2,real *wintab,real *tsbuf)
     sum1 = (tmp2b - tmp1b) * tfcos36[(v)]; \
 	MACRO0(v); }
 
-    register const real *c = nCOS9;
+    register const real *c = COS9;
     register real *out2 = o2;
 	register real *w = wintab;
 	register real *out1 = o1;
diff --git a/mp3lib/dct64_3dnow.s b/mp3lib/dct64_3dnow.s
index b7540573a6..dfade383db 100644
--- a/mp3lib/dct64_3dnow.s
+++ b/mp3lib/dct64_3dnow.s
@@ -1,706 +1,932 @@
-///
-/// Replacement of dct64() with AMD's 3DNow! SIMD operations support
-///
-/// Syuuhei Kashiyama <squash@mb.kcom.ne.jp>
-///
-/// The author of this program disclaim whole expressed or implied
-/// warranties with regard to this program, and in no event shall the
-/// author of this program liable to whatever resulted from the use of
-/// this program. Use it at your own risk.
-///
-
-        .globl dct64_3dnow
-        .type    dct64_3dnow,@function
-dct64_3dnow:
-        subl $256,%esp
-        pushl %ebp
-        pushl %edi
-        pushl %esi
-        pushl %ebx
-        leal 16(%esp),%ebx
-        movl 284(%esp),%edi
-        movl 276(%esp),%ebp
-        movl 280(%esp),%edx
-        leal 128(%ebx),%esi
-
-        / femms
-
-        // 1
-        movl pnts,%eax
-        movq 0(%edi),%mm0
-        movq %mm0,%mm1
-        movd 124(%edi),%mm2
-        punpckldq 120(%edi),%mm2
-        movq 0(%eax),%mm3
-        pfadd %mm2,%mm0
-        movq %mm0,0(%ebx)
-        pfsub %mm2,%mm1
-        pfmul %mm3,%mm1
-        movd %mm1,124(%ebx)
-        psrlq $32,%mm1
-        movd %mm1,120(%ebx)
-        movq 8(%edi),%mm4
-        movq %mm4,%mm5
-        movd 116(%edi),%mm6
-        punpckldq 112(%edi),%mm6
-        movq 8(%eax),%mm7
-        pfadd %mm6,%mm4
-        movq %mm4,8(%ebx)
-        pfsub %mm6,%mm5
-        pfmul %mm7,%mm5
-        movd %mm5,116(%ebx)
-        psrlq $32,%mm5
-        movd %mm5,112(%ebx)
-        movq 16(%edi),%mm0
-        movq %mm0,%mm1
-        movd 108(%edi),%mm2
-        punpckldq 104(%edi),%mm2
-        movq 16(%eax),%mm3
-        pfadd %mm2,%mm0
-        movq %mm0,16(%ebx)
-        pfsub %mm2,%mm1
-        pfmul %mm3,%mm1
-        movd %mm1,108(%ebx)
-        psrlq $32,%mm1
-        movd %mm1,104(%ebx)
-        movq 24(%edi),%mm4
-        movq %mm4,%mm5
-        movd 100(%edi),%mm6
-        punpckldq 96(%edi),%mm6
-        movq 24(%eax),%mm7
-        pfadd %mm6,%mm4
-        movq %mm4,24(%ebx)
-        pfsub %mm6,%mm5
-        pfmul %mm7,%mm5
-        movd %mm5,100(%ebx)
-        psrlq $32,%mm5
-        movd %mm5,96(%ebx)
-        movq 32(%edi),%mm0
-        movq %mm0,%mm1
-        movd 92(%edi),%mm2
-        punpckldq 88(%edi),%mm2
-        movq 32(%eax),%mm3
-        pfadd %mm2,%mm0
-        movq %mm0,32(%ebx)
-        pfsub %mm2,%mm1
-        pfmul %mm3,%mm1
-        movd %mm1,92(%ebx)
-        psrlq $32,%mm1
-        movd %mm1,88(%ebx)
-        movq 40(%edi),%mm4
-        movq %mm4,%mm5
-        movd 84(%edi),%mm6
-        punpckldq 80(%edi),%mm6
-        movq 40(%eax),%mm7
-        pfadd %mm6,%mm4
-        movq %mm4,40(%ebx)
-        pfsub %mm6,%mm5
-        pfmul %mm7,%mm5
-        movd %mm5,84(%ebx)
-        psrlq $32,%mm5
-        movd %mm5,80(%ebx)
-        movq 48(%edi),%mm0
-        movq %mm0,%mm1
-        movd 76(%edi),%mm2
-        punpckldq 72(%edi),%mm2
-        movq 48(%eax),%mm3
-        pfadd %mm2,%mm0
-        movq %mm0,48(%ebx)
-        pfsub %mm2,%mm1
-        pfmul %mm3,%mm1
-        movd %mm1,76(%ebx)
-        psrlq $32,%mm1
-        movd %mm1,72(%ebx)
-        movq 56(%edi),%mm4
-        movq %mm4,%mm5
-        movd 68(%edi),%mm6
-        punpckldq 64(%edi),%mm6
-        movq 56(%eax),%mm7
-        pfadd %mm6,%mm4
-        movq %mm4,56(%ebx)
-        pfsub %mm6,%mm5
-        pfmul %mm7,%mm5
-        movd %mm5,68(%ebx)
-        psrlq $32,%mm5
-        movd %mm5,64(%ebx)
-
-        // 2
-        movl pnts+4,%eax
-        / 0, 14
-        movq 0(%ebx),%mm0
-        movq %mm0,%mm1
-        movd 60(%ebx),%mm2
-        punpckldq 56(%ebx),%mm2
-        movq 0(%eax),%mm3
-        pfadd %mm2,%mm0
-        movq %mm0,0(%esi)
-        pfsub %mm2,%mm1
-        pfmul %mm3,%mm1
-        movd %mm1,60(%esi)
-        psrlq $32,%mm1
-        movd %mm1,56(%esi)
-        / 16, 30
-        movq 64(%ebx),%mm0
-        movq %mm0,%mm1
-        movd 124(%ebx),%mm2
-        punpckldq 120(%ebx),%mm2
-        pfadd %mm2,%mm0
-        movq %mm0,64(%esi)
-        pfsubr %mm2,%mm1
-        pfmul %mm3,%mm1
-        movd %mm1,124(%esi)
-        psrlq $32,%mm1
-        movd %mm1,120(%esi)
-        movq 8(%ebx),%mm4
-        / 2, 12
-        movq %mm4,%mm5
-        movd 52(%ebx),%mm6
-        punpckldq 48(%ebx),%mm6
-        movq 8(%eax),%mm7
-        pfadd %mm6,%mm4
-        movq %mm4,8(%esi)
-        pfsub %mm6,%mm5
-        pfmul %mm7,%mm5
-        movd %mm5,52(%esi)
-        psrlq $32,%mm5
-        movd %mm5,48(%esi)
-        movq 72(%ebx),%mm4
-        / 18, 28
-        movq %mm4,%mm5
-        movd 116(%ebx),%mm6
-        punpckldq 112(%ebx),%mm6
-        pfadd %mm6,%mm4
-        movq %mm4,72(%esi)
-        pfsubr %mm6,%mm5
-        pfmul %mm7,%mm5
-        movd %mm5,116(%esi)
-        psrlq $32,%mm5
-        movd %mm5,112(%esi)
-        movq 16(%ebx),%mm0
-        / 4, 10
-        movq %mm0,%mm1
-        movd 44(%ebx),%mm2
-        punpckldq 40(%ebx),%mm2
-        movq 16(%eax),%mm3
-        pfadd %mm2,%mm0
-        movq %mm0,16(%esi)
-        pfsub %mm2,%mm1
-        pfmul %mm3,%mm1
-        movd %mm1,44(%esi)
-        psrlq $32,%mm1
-        movd %mm1,40(%esi)
-        movq 80(%ebx),%mm0
-        / 20, 26
-        movq %mm0,%mm1
-        movd 108(%ebx),%mm2
-        punpckldq 104(%ebx),%mm2
-        pfadd %mm2,%mm0
-        movq %mm0,80(%esi)
-        pfsubr %mm2,%mm1
-        pfmul %mm3,%mm1
-        movd %mm1,108(%esi)
-        psrlq $32,%mm1
-        movd %mm1,104(%esi)
-        movq 24(%ebx),%mm4
-        / 6, 8
-        movq %mm4,%mm5
-        movd 36(%ebx),%mm6
-        punpckldq 32(%ebx),%mm6
-        movq 24(%eax),%mm7
-        pfadd %mm6,%mm4
-        movq %mm4,24(%esi)
-        pfsub %mm6,%mm5
-        pfmul %mm7,%mm5
-        movd %mm5,36(%esi)
-        psrlq $32,%mm5
-        movd %mm5,32(%esi)
-        movq 88(%ebx),%mm4
-        / 22, 24
-        movq %mm4,%mm5
-        movd 100(%ebx),%mm6
-        punpckldq 96(%ebx),%mm6
-        pfadd %mm6,%mm4
-        movq %mm4,88(%esi)
-        pfsubr %mm6,%mm5
-        pfmul %mm7,%mm5
-        movd %mm5,100(%esi)
-        psrlq $32,%mm5
-        movd %mm5,96(%esi)
-
-        // 3
-        movl pnts+8,%eax
-        movq 0(%eax),%mm0
-        movq 8(%eax),%mm1
-        movq 0(%esi),%mm2
-        / 0, 6
-        movq %mm2,%mm3
-        movd 28(%esi),%mm4
-        punpckldq 24(%esi),%mm4
-        pfadd %mm4,%mm2
-        pfsub %mm4,%mm3
-        pfmul %mm0,%mm3
-        movq %mm2,0(%ebx)
-        movd %mm3,28(%ebx)
-        psrlq $32,%mm3
-        movd %mm3,24(%ebx)
-        movq 8(%esi),%mm5
-        / 2, 4
-        movq %mm5,%mm6
-        movd 20(%esi),%mm7
-        punpckldq 16(%esi),%mm7
-        pfadd %mm7,%mm5
-        pfsub %mm7,%mm6
-        pfmul %mm1,%mm6
-        movq %mm5,8(%ebx)
-        movd %mm6,20(%ebx)
-        psrlq $32,%mm6
-        movd %mm6,16(%ebx)
-        movq 32(%esi),%mm2
-        / 8, 14
-        movq %mm2,%mm3
-        movd 60(%esi),%mm4
-        punpckldq 56(%esi),%mm4
-        pfadd %mm4,%mm2
-        pfsubr %mm4,%mm3
-        pfmul %mm0,%mm3
-        movq %mm2,32(%ebx)
-        movd %mm3,60(%ebx)
-        psrlq $32,%mm3
-        movd %mm3,56(%ebx)
-        movq 40(%esi),%mm5
-        / 10, 12
-        movq %mm5,%mm6
-        movd 52(%esi),%mm7
-        punpckldq 48(%esi),%mm7
-        pfadd %mm7,%mm5
-        pfsubr %mm7,%mm6
-        pfmul %mm1,%mm6
-        movq %mm5,40(%ebx)
-        movd %mm6,52(%ebx)
-        psrlq $32,%mm6
-        movd %mm6,48(%ebx)
-        movq 64(%esi),%mm2
-        / 16, 22
-        movq %mm2,%mm3
-        movd 92(%esi),%mm4
-        punpckldq 88(%esi),%mm4
-        pfadd %mm4,%mm2
-        pfsub %mm4,%mm3
-        pfmul %mm0,%mm3
-        movq %mm2,64(%ebx)
-        movd %mm3,92(%ebx)
-        psrlq $32,%mm3
-        movd %mm3,88(%ebx)
-        movq 72(%esi),%mm5
-        / 18, 20
-        movq %mm5,%mm6
-        movd 84(%esi),%mm7
-        punpckldq 80(%esi),%mm7
-        pfadd %mm7,%mm5
-        pfsub %mm7,%mm6
-        pfmul %mm1,%mm6
-        movq %mm5,72(%ebx)
-        movd %mm6,84(%ebx)
-        psrlq $32,%mm6
-        movd %mm6,80(%ebx)
-        movq 96(%esi),%mm2
-        / 24, 30
-        movq %mm2,%mm3
-        movd 124(%esi),%mm4
-        punpckldq 120(%esi),%mm4
-        pfadd %mm4,%mm2
-        pfsubr %mm4,%mm3
-        pfmul %mm0,%mm3
-        movq %mm2,96(%ebx)
-        movd %mm3,124(%ebx)
-        psrlq $32,%mm3
-        movd %mm3,120(%ebx)
-        movq 104(%esi),%mm5
-        / 26, 28
-        movq %mm5,%mm6
-        movd 116(%esi),%mm7
-        punpckldq 112(%esi),%mm7
-        pfadd %mm7,%mm5
-        pfsubr %mm7,%mm6
-        pfmul %mm1,%mm6
-        movq %mm5,104(%ebx)
-        movd %mm6,116(%ebx)
-        psrlq $32,%mm6
-        movd %mm6,112(%ebx)
-
-        // 4
-        movl pnts+12,%eax
-        movq 0(%eax),%mm0
-        movq 0(%ebx),%mm1
-        / 0
-        movq %mm1,%mm2
-        movd 12(%ebx),%mm3
-        punpckldq 8(%ebx),%mm3
-        pfadd %mm3,%mm1
-        pfsub %mm3,%mm2
-        pfmul %mm0,%mm2
-        movq %mm1,0(%esi)
-        movd %mm2,12(%esi)
-        psrlq $32,%mm2
-        movd %mm2,8(%esi)
-        movq 16(%ebx),%mm4
-        / 4
-        movq %mm4,%mm5
-        movd 28(%ebx),%mm6
-        punpckldq 24(%ebx),%mm6
-        pfadd %mm6,%mm4
-        pfsubr %mm6,%mm5
-        pfmul %mm0,%mm5
-        movq %mm4,16(%esi)
-        movd %mm5,28(%esi)
-        psrlq $32,%mm5
-        movd %mm5,24(%esi)
-        movq 32(%ebx),%mm1
-        / 8
-        movq %mm1,%mm2
-        movd 44(%ebx),%mm3
-        punpckldq 40(%ebx),%mm3
-        pfadd %mm3,%mm1
-        pfsub %mm3,%mm2
-        pfmul %mm0,%mm2
-        movq %mm1,32(%esi)
-        movd %mm2,44(%esi)
-        psrlq $32,%mm2
-        movd %mm2,40(%esi)
-        movq 48(%ebx),%mm4
-        / 12
-        movq %mm4,%mm5
-        movd 60(%ebx),%mm6
-        punpckldq 56(%ebx),%mm6
-        pfadd %mm6,%mm4
-        pfsubr %mm6,%mm5
-        pfmul %mm0,%mm5
-        movq %mm4,48(%esi)
-        movd %mm5,60(%esi)
-        psrlq $32,%mm5
-        movd %mm5,56(%esi)
-        movq 64(%ebx),%mm1
-        / 16
-        movq %mm1,%mm2
-        movd 76(%ebx),%mm3
-        punpckldq 72(%ebx),%mm3
-        pfadd %mm3,%mm1
-        pfsub %mm3,%mm2
-        pfmul %mm0,%mm2
-        movq %mm1,64(%esi)
-        movd %mm2,76(%esi)
-        psrlq $32,%mm2
-        movd %mm2,72(%esi)
-        movq 80(%ebx),%mm4
-        / 20
-        movq %mm4,%mm5
-        movd 92(%ebx),%mm6
-        punpckldq 88(%ebx),%mm6
-        pfadd %mm6,%mm4
-        pfsubr %mm6,%mm5
-        pfmul %mm0,%mm5
-        movq %mm4,80(%esi)
-        movd %mm5,92(%esi)
-        psrlq $32,%mm5
-        movd %mm5,88(%esi)
-        movq 96(%ebx),%mm1
-        / 24
-        movq %mm1,%mm2
-        movd 108(%ebx),%mm3
-        punpckldq 104(%ebx),%mm3
-        pfadd %mm3,%mm1
-        pfsub %mm3,%mm2
-        pfmul %mm0,%mm2
-        movq %mm1,96(%esi)
-        movd %mm2,108(%esi)
-        psrlq $32,%mm2
-        movd %mm2,104(%esi)
-        movq 112(%ebx),%mm4
-        / 28
-        movq %mm4,%mm5
-        movd 124(%ebx),%mm6
-        punpckldq 120(%ebx),%mm6
-        pfadd %mm6,%mm4
-        pfsubr %mm6,%mm5
-        pfmul %mm0,%mm5
-        movq %mm4,112(%esi)
-        movd %mm5,124(%esi)
-        psrlq $32,%mm5
-        movd %mm5,120(%esi)
-
-        // 5
-        movl $-1,%eax
-        movd %eax,%mm1
-        movl $1,%eax
-        movd %eax,%mm0
-        / L | H
-        punpckldq %mm1,%mm0
-        pi2fd %mm0,%mm0
-        / 1.0 | -1.0
-        movd %eax,%mm1
-        pi2fd %mm1,%mm1
-        movl pnts+16,%eax
-        movd 0(%eax),%mm2
-        punpckldq %mm2,%mm1
-        / 1.0 | cos0
-        movq 0(%esi),%mm2
-        / 0
-        movq %mm2,%mm3
-        pfmul %mm0,%mm3
-        pfacc %mm3,%mm2
-        pfmul %mm1,%mm2
-        movq %mm2,0(%ebx)
-        movq 8(%esi),%mm4
-        movq %mm4,%mm5
-        pfmul %mm0,%mm5
-        pfacc %mm5,%mm4
-        pfmul %mm0,%mm4
-        pfmul %mm1,%mm4
-        movq %mm4,%mm5
-        psrlq $32,%mm5
-        pfacc %mm5,%mm4
-        movq %mm4,8(%ebx)
-        movq 16(%esi),%mm2
-        / 4
-        movq %mm2,%mm3
-        pfmul %mm0,%mm3
-        pfacc %mm3,%mm2
-        pfmul %mm1,%mm2
-        movq 24(%esi),%mm4
-        movq %mm4,%mm5
-        pfmul %mm0,%mm5
-        pfacc %mm5,%mm4
-        pfmul %mm0,%mm4
-        pfmul %mm1,%mm4
-        movq %mm4,%mm5
-        psrlq $32,%mm5
-        pfacc %mm5,%mm4
-        movq %mm2,%mm3
-        psrlq $32,%mm3
-        pfadd %mm4,%mm2
-        pfadd %mm3,%mm4
-        movq %mm2,16(%ebx)
-        movq %mm4,24(%ebx)
-        movq 32(%esi),%mm2
-        / 8
-        movq %mm2,%mm3
-        pfmul %mm0,%mm3
-        pfacc %mm3,%mm2
-        pfmul %mm1,%mm2
-        movq %mm2,32(%ebx)
-        movq 40(%esi),%mm4
-        movq %mm4,%mm5
-        pfmul %mm0,%mm5
-        pfacc %mm5,%mm4
-        pfmul %mm0,%mm4
-        pfmul %mm1,%mm4
-        movq %mm4,%mm5
-        psrlq $32,%mm5
-        pfacc %mm5,%mm4
-        movq %mm4,40(%ebx)
-        movq 48(%esi),%mm2
-        / 12
-        movq %mm2,%mm3
-        pfmul %mm0,%mm3
-        pfacc %mm3,%mm2
-        pfmul %mm1,%mm2
-        movq 56(%esi),%mm4
-        movq %mm4,%mm5
-        pfmul %mm0,%mm5
-        pfacc %mm5,%mm4
-        pfmul %mm0,%mm4
-        pfmul %mm1,%mm4
-        movq %mm4,%mm5
-        psrlq $32,%mm5
-        pfacc %mm5,%mm4
-        movq %mm2,%mm3
-        psrlq $32,%mm3
-        pfadd %mm4,%mm2
-        pfadd %mm3,%mm4
-        movq %mm2,48(%ebx)
-        movq %mm4,56(%ebx)
-        movq 64(%esi),%mm2
-        / 16
-        movq %mm2,%mm3
-        pfmul %mm0,%mm3
-        pfacc %mm3,%mm2
-        pfmul %mm1,%mm2
-        movq %mm2,64(%ebx)
-        movq 72(%esi),%mm4
-        movq %mm4,%mm5
-        pfmul %mm0,%mm5
-        pfacc %mm5,%mm4
-        pfmul %mm0,%mm4
-        pfmul %mm1,%mm4
-        movq %mm4,%mm5
-        psrlq $32,%mm5
-        pfacc %mm5,%mm4
-        movq %mm4,72(%ebx)
-        movq 80(%esi),%mm2
-        / 20
-        movq %mm2,%mm3
-        pfmul %mm0,%mm3
-        pfacc %mm3,%mm2
-        pfmul %mm1,%mm2
-        movq 88(%esi),%mm4
-        movq %mm4,%mm5
-        pfmul %mm0,%mm5
-        pfacc %mm5,%mm4
-        pfmul %mm0,%mm4
-        pfmul %mm1,%mm4
-        movq %mm4,%mm5
-        psrlq $32,%mm5
-        pfacc %mm5,%mm4
-        movq %mm2,%mm3
-        psrlq $32,%mm3
-        pfadd %mm4,%mm2
-        pfadd %mm3,%mm4
-        movq %mm2,80(%ebx)
-        movq %mm4,88(%ebx)
-        movq 96(%esi),%mm2
-        / 24
-        movq %mm2,%mm3
-        pfmul %mm0,%mm3
-        pfacc %mm3,%mm2
-        pfmul %mm1,%mm2
-        movq %mm2,96(%ebx)
-        movq 104(%esi),%mm4
-        movq %mm4,%mm5
-        pfmul %mm0,%mm5
-        pfacc %mm5,%mm4
-        pfmul %mm0,%mm4
-        pfmul %mm1,%mm4
-        movq %mm4,%mm5
-        psrlq $32,%mm5
-        pfacc %mm5,%mm4
-        movq %mm4,104(%ebx)
-        movq 112(%esi),%mm2
-        / 28
-        movq %mm2,%mm3
-        pfmul %mm0,%mm3
-        pfacc %mm3,%mm2
-        pfmul %mm1,%mm2
-        movq 120(%esi),%mm4
-        movq %mm4,%mm5
-        pfmul %mm0,%mm5
-        pfacc %mm5,%mm4
-        pfmul %mm0,%mm4
-        pfmul %mm1,%mm4
-        movq %mm4,%mm5
-        psrlq $32,%mm5
-        pfacc %mm5,%mm4
-        movq %mm2,%mm3
-        psrlq $32,%mm3
-        pfadd %mm4,%mm2
-        pfadd %mm3,%mm4
-        movq %mm2,112(%ebx)
-        movq %mm4,120(%ebx)
-
-        // Phase6
-        movl 0(%ebx),%eax
-        movl %eax,1024(%ebp)
-        movl 4(%ebx),%eax
-        movl %eax,0(%ebp)
-        movl %eax,0(%edx)
-        movl 8(%ebx),%eax
-        movl %eax,512(%ebp)
-        movl 12(%ebx),%eax
-        movl %eax,512(%edx)
-
-        movl 16(%ebx),%eax
-        movl %eax,768(%ebp)
-        movl 20(%ebx),%eax
-        movl %eax,256(%edx)
-
-        movl 24(%ebx),%eax
-        movl %eax,256(%ebp)
-        movl 28(%ebx),%eax
-        movl %eax,768(%edx)
-
-        movq 32(%ebx),%mm0
-        movq 48(%ebx),%mm1
-        pfadd %mm1,%mm0
-        movd %mm0,896(%ebp)
-        psrlq $32,%mm0
-        movd %mm0,128(%edx)
-        movq 40(%ebx),%mm2
-        pfadd %mm2,%mm1
-        movd %mm1,640(%ebp)
-        psrlq $32,%mm1
-        movd %mm1,384(%edx)
-
-        movq 56(%ebx),%mm3
-        pfadd %mm3,%mm2
-        movd %mm2,384(%ebp)
-        psrlq $32,%mm2
-        movd %mm2,640(%edx)
-
-        movd 36(%ebx),%mm4
-        pfadd %mm4,%mm3
-        movd %mm3,128(%ebp)
-        psrlq $32,%mm3
-        movd %mm3,896(%edx)
-        movq 96(%ebx),%mm0
-        movq 64(%ebx),%mm1
-
-        movq 112(%ebx),%mm2
-        pfadd %mm2,%mm0
-        movq %mm0,%mm3
-        pfadd %mm1,%mm3
-        movd %mm3,960(%ebp)
-        psrlq $32,%mm3
-        movd %mm3,64(%edx)
-        movq 80(%ebx),%mm1
-        pfadd %mm1,%mm0
-        movd %mm0,832(%ebp)
-        psrlq $32,%mm0
-        movd %mm0,192(%edx)
-        movq 104(%ebx),%mm3
-        pfadd %mm3,%mm2
-        movq %mm2,%mm4
-        pfadd %mm1,%mm4
-        movd %mm4,704(%ebp)
-        psrlq $32,%mm4
-        movd %mm4,320(%edx)
-        movq 72(%ebx),%mm1
-        pfadd %mm1,%mm2
-        movd %mm2,576(%ebp)
-        psrlq $32,%mm2
-        movd %mm2,448(%edx)
-
-        movq 120(%ebx),%mm4
-        pfadd %mm4,%mm3
-        movq %mm3,%mm5
-        pfadd %mm1,%mm5
-        movd %mm5,448(%ebp)
-        psrlq $32,%mm5
-        movd %mm5,576(%edx)
-        movq 88(%ebx),%mm1
-        pfadd %mm1,%mm3
-        movd %mm3,320(%ebp)
-        psrlq $32,%mm3
-        movd %mm3,704(%edx)
-
-        movd 100(%ebx),%mm5
-        pfadd %mm5,%mm4
-        movq %mm4,%mm6
-        pfadd %mm1,%mm6
-        movd %mm6,192(%ebp)
-        psrlq $32,%mm6
-        movd %mm6,832(%edx)
-        movd 68(%ebx),%mm1
-        pfadd %mm1,%mm4
-        movd %mm4,64(%ebp)
-        psrlq $32,%mm4
-        movd %mm4,960(%edx)
-
-        / femms
-
-        popl %ebx
-        popl %esi
-        popl %edi
-        popl %ebp
-        addl $256,%esp
-
-        ret
+# This code was taken from http://www.mpg123.org
+# See ChangeLog of mpg123-0.59s-pre.1 for detail
+# Applied to mplayer by Nick Kurshev <nickols_k@mail.ru>
+# Partial 3dnow! optimization by Nick Kurshev
+#
+# TODO: finish 3dnow! optimization at least in scalar mode
+#
+
+.data
+	.align 8
+plus_minus_3dnow: .long 0x00000000, 0x80000000
+costab:
+	.long 1056974725
+	.long 1057056395
+	.long 1057223771
+	.long 1057485416
+	.long 1057855544
+	.long 1058356026
+	.long 1059019886
+	.long 1059897405
+	.long 1061067246
+	.long 1062657950
+	.long 1064892987
+	.long 1066774581
+	.long 1069414683
+	.long 1073984175
+	.long 1079645762
+	.long 1092815430
+	.long 1057005197
+	.long 1057342072
+	.long 1058087743
+	.long 1059427869
+	.long 1061799040
+	.long 1065862217
+	.long 1071413542
+	.long 1084439708
+	.long 1057128951
+	.long 1058664893
+	.long 1063675095
+	.long 1076102863
+	.long 1057655764
+	.long 1067924853
+	.long 1060439283
+
+.text
+
+	.align 16
+
+.globl dct64_MMX_3dnow
+dct64_MMX_3dnow:
+	pushl %ebx
+	pushl %esi
+	pushl %edi
+	subl $256,%esp
+	movl 280(%esp),%eax
+
+	leal 128(%esp),%edx
+	movl 272(%esp),%esi
+	movl 276(%esp),%edi
+	movl $costab,%ebx
+	orl %ecx,%ecx
+	movl %esp,%ecx
+	femms	
+/* Phase 1*/
+	movq	(%eax), %mm0
+	movq	8(%eax), %mm4
+	movq	%mm0, %mm3
+	movq	%mm4, %mm7
+	movq	120(%eax), %mm1
+	movq	112(%eax), %mm5
+	/* n.b.: pswapd*/	
+	movq	%mm1, %mm2
+	movq	%mm5, %mm6
+	psrlq	$32, %mm1
+	psrlq	$32, %mm5
+	punpckldq %mm2, %mm1
+	punpckldq %mm6, %mm5
+	/**/
+	pfadd	%mm1, %mm0
+	pfadd	%mm5, %mm4
+	movq	%mm0, (%edx)
+	movq	%mm4, 8(%edx)
+	pfsub	%mm1, %mm3
+	pfsub	%mm5, %mm7
+	pfmul	(%ebx), %mm3
+	pfmul	8(%ebx), %mm7
+	movd	%mm3, 124(%edx)
+	movd	%mm7, 116(%edx)
+	psrlq	$32, %mm3
+	psrlq	$32, %mm7
+	movd	%mm3, 120(%edx)
+	movd	%mm7, 112(%edx)
+
+	movq	16(%eax), %mm0
+	movq	24(%eax), %mm4
+	movq	%mm0, %mm3
+	movq	%mm4, %mm7
+	movq	104(%eax), %mm1
+	movq	96(%eax), %mm5
+	/* n.b.: pswapd*/	
+	movq	%mm1, %mm2
+	movq	%mm5, %mm6
+	psrlq	$32, %mm1
+	psrlq	$32, %mm5
+	punpckldq %mm2, %mm1
+	punpckldq %mm6, %mm5
+	/**/
+	pfadd	%mm1, %mm0
+	pfadd	%mm5, %mm4
+	movq	%mm0, 16(%edx)
+	movq	%mm4, 24(%edx)
+	pfsub	%mm1, %mm3
+	pfsub	%mm5, %mm7
+	pfmul	16(%ebx), %mm3
+	pfmul	24(%ebx), %mm7
+	movd	%mm3, 108(%edx)
+	movd	%mm7, 100(%edx)
+	psrlq	$32, %mm3
+	psrlq	$32, %mm7
+	movd	%mm3, 104(%edx)
+	movd	%mm7, 96(%edx)
+
+	movq	32(%eax), %mm0
+	movq	40(%eax), %mm4
+	movq	%mm0, %mm3
+	movq	%mm4, %mm7
+	movq	88(%eax), %mm1
+	movq	80(%eax), %mm5
+	/* n.b.: pswapd*/	
+	movq	%mm1, %mm2
+	movq	%mm5, %mm6
+	psrlq	$32, %mm1
+	psrlq	$32, %mm5
+	punpckldq %mm2, %mm1
+	punpckldq %mm6, %mm5
+	/**/
+	pfadd	%mm1, %mm0
+	pfadd	%mm5, %mm4
+	movq	%mm0, 32(%edx)
+	movq	%mm4, 40(%edx)
+	pfsub	%mm1, %mm3
+	pfsub	%mm5, %mm7
+	pfmul	32(%ebx), %mm3
+	pfmul	40(%ebx), %mm7
+	movd	%mm3, 92(%edx)
+	movd	%mm7, 84(%edx)
+	psrlq	$32, %mm3
+	psrlq	$32, %mm7
+	movd	%mm3, 88(%edx)
+	movd	%mm7, 80(%edx)
+
+	movq	48(%eax), %mm0
+	movq	56(%eax), %mm4
+	movq	%mm0, %mm3
+	movq	%mm4, %mm7
+	movq	72(%eax), %mm1
+	movq	64(%eax), %mm5
+	/* n.b.: pswapd*/	
+	movq	%mm1, %mm2
+	movq	%mm5, %mm6
+	psrlq	$32, %mm1
+	psrlq	$32, %mm5
+	punpckldq %mm2, %mm1
+	punpckldq %mm6, %mm5
+	/**/
+	pfadd	%mm1, %mm0
+	pfadd	%mm5, %mm4
+	movq	%mm0, 48(%edx)
+	movq	%mm4, 56(%edx)
+	pfsub	%mm1, %mm3
+	pfsub	%mm5, %mm7
+	pfmul	48(%ebx), %mm3
+	pfmul	56(%ebx), %mm7
+	movd	%mm3, 76(%edx)
+	movd	%mm7, 68(%edx)
+	psrlq	$32, %mm3
+	psrlq	$32, %mm7
+	movd	%mm3, 72(%edx)
+	movd	%mm7, 64(%edx)
+
+/* Phase 2*/
+
+	movq	(%edx), %mm0
+	movq	8(%edx), %mm4
+	movq	%mm0, %mm3
+	movq	%mm4, %mm7
+	movq	56(%edx), %mm1
+	movq	48(%edx), %mm5
+	/* n.b.: pswapd*/	
+	movq	%mm1, %mm2
+	movq	%mm5, %mm6
+	psrlq	$32, %mm1
+	psrlq	$32, %mm5
+	punpckldq %mm2, %mm1
+	punpckldq %mm6, %mm5
+	/**/
+	pfadd	%mm1, %mm0
+	pfadd	%mm5, %mm4
+	movq	%mm0, (%ecx)
+	movq	%mm4, 8(%ecx)
+	pfsub	%mm1, %mm3
+	pfsub	%mm5, %mm7
+	pfmul	64(%ebx), %mm3
+	pfmul	72(%ebx), %mm7
+	movd	%mm3, 60(%ecx)
+	movd	%mm7, 52(%ecx)
+	psrlq	$32, %mm3
+	psrlq	$32, %mm7
+	movd	%mm3, 56(%ecx)
+	movd	%mm7, 48(%ecx)
+	
+	movq	16(%edx), %mm0
+	movq	24(%edx), %mm4
+	movq	%mm0, %mm3
+	movq	%mm4, %mm7
+	movq	40(%edx), %mm1
+	movq	32(%edx), %mm5
+	/* n.b.: pswapd*/	
+	movq	%mm1, %mm2
+	movq	%mm5, %mm6
+	psrlq	$32, %mm1
+	psrlq	$32, %mm5
+	punpckldq %mm2, %mm1
+	punpckldq %mm6, %mm5
+	/**/
+	pfadd	%mm1, %mm0
+	pfadd	%mm5, %mm4
+	movq	%mm0, 16(%ecx)
+	movq	%mm4, 24(%ecx)
+	pfsub	%mm1, %mm3
+	pfsub	%mm5, %mm7
+	pfmul	80(%ebx), %mm3
+	pfmul	88(%ebx), %mm7
+	movd	%mm3, 44(%ecx)