split-radix FFT

c is 1.9x faster than previous c (on various x86 cpus), sse is 1.6x faster than previous sse.

Originally committed as revision 14698 to svn://svn.ffmpeg.org/ffmpeg/trunk

4 files changed, 523 insertions, 309 deletions

diff --git a/libavcodec/i386/fft_3dn.c b/libavcodec/i386/fft_3dn.c
index 8bd7b89d01..6f2e2e8353 100644
--- a/libavcodec/i386/fft_3dn.c
+++ b/libavcodec/i386/fft_3dn.c
@@ -1,7 +1,6 @@
 /*
  * FFT/MDCT transform with 3DNow! optimizations
- * Copyright (c) 2006 Zuxy MENG Jie, Loren Merritt
- * Based on fft_sse.c copyright (c) 2002 Fabrice Bellard.
+ * Copyright (c) 2008 Loren Merritt
  *
  * This file is part of FFmpeg.
  *
@@ -20,109 +19,5 @@
  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  */
 
-#include "libavutil/x86_cpu.h"
-#include "libavcodec/dsputil.h"
-
-static const int p1m1[2] __attribute__((aligned(8))) =
-    { 0, 1 << 31 };
-
-static const int m1p1[2] __attribute__((aligned(8))) =
-    { 1 << 31, 0 };
-
-void ff_fft_calc_3dn(FFTContext *s, FFTComplex *z)
-{
-    int ln = s->nbits;
-    long j;
-    x86_reg i;
-    long nblocks, nloops;
-    FFTComplex *p, *cptr;
-
-    asm volatile(
-        /* FEMMS is not a must here but recommended by AMD */
-        "femms \n\t"
-        "movq %0, %%mm7 \n\t"
-        ::"m"(*(s->inverse ? m1p1 : p1m1))
-    );
-
-    i = 8 << ln;
-    asm volatile(
-        "1: \n\t"
-        "sub $32, %0 \n\t"
-        "movq    (%0,%1), %%mm0 \n\t"
-        "movq  16(%0,%1), %%mm1 \n\t"
-        "movq   8(%0,%1), %%mm2 \n\t"
-        "movq  24(%0,%1), %%mm3 \n\t"
-        "movq      %%mm0, %%mm4 \n\t"
-        "movq      %%mm1, %%mm5 \n\t"
-        "pfadd     %%mm2, %%mm0 \n\t"
-        "pfadd     %%mm3, %%mm1 \n\t"
-        "pfsub     %%mm2, %%mm4 \n\t"
-        "pfsub     %%mm3, %%mm5 \n\t"
-        "movq      %%mm0, %%mm2 \n\t"
-        "punpckldq %%mm5, %%mm6 \n\t"
-        "punpckhdq %%mm6, %%mm5 \n\t"
-        "movq      %%mm4, %%mm3 \n\t"
-        "pxor      %%mm7, %%mm5 \n\t"
-        "pfadd     %%mm1, %%mm0 \n\t"
-        "pfadd     %%mm5, %%mm4 \n\t"
-        "pfsub     %%mm1, %%mm2 \n\t"
-        "pfsub     %%mm5, %%mm3 \n\t"
-        "movq      %%mm0,   (%0,%1) \n\t"
-        "movq      %%mm4,  8(%0,%1) \n\t"
-        "movq      %%mm2, 16(%0,%1) \n\t"
-        "movq      %%mm3, 24(%0,%1) \n\t"
-        "jg 1b \n\t"
-        :"+r"(i)
-        :"r"(z)
-    );
-    /* pass 2 .. ln-1 */
-
-    nblocks = 1 << (ln-3);
-    nloops = 1 << 2;
-    cptr = s->exptab1;
-    do {
-        p = z;
-        j = nblocks;
-        do {
-            i = nloops*8;
-            asm volatile(
-                "1: \n\t"
-                "sub $16, %0 \n\t"
-                "movq    (%1,%0), %%mm0 \n\t"
-                "movq   8(%1,%0), %%mm1 \n\t"
-                "movq    (%2,%0), %%mm2 \n\t"
-                "movq   8(%2,%0), %%mm3 \n\t"
-                "movq      %%mm2, %%mm4 \n\t"
-                "movq      %%mm3, %%mm5 \n\t"
-                "punpckldq %%mm2, %%mm2 \n\t"
-                "punpckldq %%mm3, %%mm3 \n\t"
-                "punpckhdq %%mm4, %%mm4 \n\t"
-                "punpckhdq %%mm5, %%mm5 \n\t"
-                "pfmul   (%3,%0,2), %%mm2 \n\t" //  cre*re cim*re
-                "pfmul  8(%3,%0,2), %%mm3 \n\t"
-                "pfmul 16(%3,%0,2), %%mm4 \n\t" // -cim*im cre*im
-                "pfmul 24(%3,%0,2), %%mm5 \n\t"
-                "pfadd     %%mm2, %%mm4 \n\t" // cre*re-cim*im cim*re+cre*im
-                "pfadd     %%mm3, %%mm5 \n\t"
-                "movq      %%mm0, %%mm2 \n\t"
-                "movq      %%mm1, %%mm3 \n\t"
-                "pfadd     %%mm4, %%mm0 \n\t"
-                "pfadd     %%mm5, %%mm1 \n\t"
-                "pfsub     %%mm4, %%mm2 \n\t"
-                "pfsub     %%mm5, %%mm3 \n\t"
-                "movq      %%mm0,  (%1,%0) \n\t"
-                "movq      %%mm1, 8(%1,%0) \n\t"
-                "movq      %%mm2,  (%2,%0) \n\t"
-                "movq      %%mm3, 8(%2,%0) \n\t"
-                "jg 1b \n\t"
-                :"+r"(i)
-                :"r"(p), "r"(p + nloops), "r"(cptr)
-            );
-            p += nloops*2;
-        } while (--j);
-        cptr += nloops*2;
-        nblocks >>= 1;
-        nloops <<= 1;
-    } while (nblocks != 0);
-    asm volatile("femms");
-}
+#define EMULATE_3DNOWEXT
+#include "fft_3dn2.c"
diff --git a/libavcodec/i386/fft_3dn2.c b/libavcodec/i386/fft_3dn2.c
index 9068dff24b..635bc76658 100644
--- a/libavcodec/i386/fft_3dn2.c
+++ b/libavcodec/i386/fft_3dn2.c
@@ -23,105 +23,26 @@
 #include "libavutil/x86_cpu.h"
 #include "libavcodec/dsputil.h"
 
-static const int p1m1[2] __attribute__((aligned(8))) =
-    { 0, 1 << 31 };
+#ifdef EMULATE_3DNOWEXT
+#define ff_fft_calc_3dn2 ff_fft_calc_3dn
+#define ff_fft_dispatch_3dn2 ff_fft_dispatch_3dn
+#define ff_fft_dispatch_interleave_3dn2 ff_fft_dispatch_interleave_3dn
+#define ff_imdct_calc_3dn2 ff_imdct_calc_3dn
+#define ff_imdct_half_3dn2 ff_imdct_half_3dn
+#endif
 
-static const int m1p1[2] __attribute__((aligned(8))) =
-    { 1 << 31, 0 };
+void ff_fft_dispatch_3dn2(FFTComplex *z, int nbits);
+void ff_fft_dispatch_interleave_3dn2(FFTComplex *z, int nbits);
 
 void ff_fft_calc_3dn2(FFTContext *s, FFTComplex *z)
 {
-    int ln = s->nbits;
-    long j;
-    x86_reg i;
-    long nblocks, nloops;
-    FFTComplex *p, *cptr;
-
-    asm volatile(
-        /* FEMMS is not a must here but recommended by AMD */
-        "femms \n\t"
-        "movq %0, %%mm7 \n\t"
-        ::"m"(*(s->inverse ? m1p1 : p1m1))
-    );
-
-    i = 8 << ln;
-    asm volatile(
-        "1: \n\t"
-        "sub $32, %0 \n\t"
-        "movq    (%0,%1), %%mm0 \n\t"
-        "movq  16(%0,%1), %%mm1 \n\t"
-        "movq   8(%0,%1), %%mm2 \n\t"
-        "movq  24(%0,%1), %%mm3 \n\t"
-        "movq      %%mm0, %%mm4 \n\t"
-        "movq      %%mm1, %%mm5 \n\t"
-        "pfadd     %%mm2, %%mm0 \n\t"
-        "pfadd     %%mm3, %%mm1 \n\t"
-        "pfsub     %%mm2, %%mm4 \n\t"
-        "pfsub     %%mm3, %%mm5 \n\t"
-        "movq      %%mm0, %%mm2 \n\t"
-        "pswapd    %%mm5, %%mm5 \n\t"
-        "movq      %%mm4, %%mm3 \n\t"
-        "pxor      %%mm7, %%mm5 \n\t"
-        "pfadd     %%mm1, %%mm0 \n\t"
-        "pfadd     %%mm5, %%mm4 \n\t"
-        "pfsub     %%mm1, %%mm2 \n\t"
-        "pfsub     %%mm5, %%mm3 \n\t"
-        "movq      %%mm0,   (%0,%1) \n\t"
-        "movq      %%mm4,  8(%0,%1) \n\t"
-        "movq      %%mm2, 16(%0,%1) \n\t"
-        "movq      %%mm3, 24(%0,%1) \n\t"
-        "jg 1b \n\t"
-        :"+r"(i)
-        :"r"(z)
-    );
-    /* pass 2 .. ln-1 */
-
-    nblocks = 1 << (ln-3);
-    nloops = 1 << 2;
-    cptr = s->exptab1;
-    do {
-        p = z;
-        j = nblocks;
-        do {
-            i = nloops*8;
-            asm volatile(
-                "1: \n\t"
-                "sub $16, %0 \n\t"
-                "movq    (%1,%0), %%mm0 \n\t"
-                "movq   8(%1,%0), %%mm1 \n\t"
-                "movq    (%2,%0), %%mm2 \n\t"
-                "movq   8(%2,%0), %%mm3 \n\t"
-                "movq  (%3,%0,2), %%mm4 \n\t"
-                "movq 8(%3,%0,2), %%mm5 \n\t"
-                "pswapd    %%mm4, %%mm6 \n\t" // no need for cptr[2] & cptr[3]
-                "pswapd    %%mm5, %%mm7 \n\t"
-                "pfmul     %%mm2, %%mm4 \n\t" // cre*re cim*im
-                "pfmul     %%mm3, %%mm5 \n\t"
-                "pfmul     %%mm2, %%mm6 \n\t" // cim*re cre*im
-                "pfmul     %%mm3, %%mm7 \n\t"
-                "pfpnacc   %%mm6, %%mm4 \n\t" // cre*re-cim*im cim*re+cre*im
-                "pfpnacc   %%mm7, %%mm5 \n\t"
-                "movq      %%mm0, %%mm2 \n\t"
-                "movq      %%mm1, %%mm3 \n\t"
-                "pfadd     %%mm4, %%mm0 \n\t"
-                "pfadd     %%mm5, %%mm1 \n\t"
-                "pfsub     %%mm4, %%mm2 \n\t"
-                "pfsub     %%mm5, %%mm3 \n\t"
-                "movq      %%mm0,  (%1,%0) \n\t"
-                "movq      %%mm1, 8(%1,%0) \n\t"
-                "movq      %%mm2,  (%2,%0) \n\t"
-                "movq      %%mm3, 8(%2,%0) \n\t"
-                "jg 1b \n\t"
-                :"+r"(i)
-                :"r"(p), "r"(p + nloops), "r"(cptr)
-            );
-            p += nloops*2;
-        } while (--j);
-        cptr += nloops*2;
-        nblocks >>= 1;
-        nloops <<= 1;
-    } while (nblocks != 0);
+    int n = 1<<s->nbits;
+    int i;
+    ff_fft_dispatch_interleave_3dn2(z, s->nbits);
     asm volatile("femms");
+    if(n <= 8)
+        for(i=0; i<n; i+=2)
+            FFSWAP(FFTSample, z[i].im, z[i+1].re);
 }
 
 static void imdct_3dn2(MDCTContext *s, const FFTSample *input, FFTSample *tmp)
@@ -162,7 +83,7 @@ static void imdct_3dn2(MDCTContext *s, const FFTSample *input, FFTSample *tmp)
         );
     }
 
-    ff_fft_calc(&s->fft, z);
+    ff_fft_calc_3dn2(&s->fft, z);
 
     /* post rotation + reordering */
     for(k = 0; k < n4; k++) {
diff --git a/libavcodec/i386/fft_mmx.asm b/libavcodec/i386/fft_mmx.asm
new file mode 100644
index 0000000000..c0a9bd5b6b
--- /dev/null
+++ b/libavcodec/i386/fft_mmx.asm
@@ -0,0 +1,467 @@
+;******************************************************************************
+;* FFT transform with SSE/3DNow optimizations
+;* Copyright (c) 2008 Loren Merritt
+;*
+;* This file is part of FFmpeg.
+;*
+;* FFmpeg is free software; you can redistribute it and/or
+;* modify it under the terms of the GNU Lesser General Public
+;* License as published by the Free Software Foundation; either
+;* version 2.1 of the License, or (at your option) any later version.
+;*
+;* FFmpeg is distributed in the hope that it will be useful,
+;* but WITHOUT ANY WARRANTY; without even the implied warranty of
+;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+;* Lesser General Public License for more details.
+;*
+;* You should have received a copy of the GNU Lesser General Public
+;* License along with FFmpeg; if not, write to the Free Software
+;* 51, Inc., Foundation Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+;******************************************************************************
+
+; These functions are not individually interchangeable with the C versions.
+; While C takes arrays of FFTComplex, SSE/3DNow leave intermediate results
+; in blocks as conventient to the vector size.
+; i.e. {4x real, 4x imaginary, 4x real, ...} (or 2x respectively)
+
+%include "x86inc.asm"
+
+SECTION_RODATA
+
+%define M_SQRT1_2 0.70710678118654752440
+ps_root2: times 4 dd M_SQRT1_2
+ps_root2mppm: dd -M_SQRT1_2, M_SQRT1_2, M_SQRT1_2, -M_SQRT1_2
+ps_m1p1: dd 1<<31, 0
+
+%assign i 16
+%rep 13
+cextern ff_cos_ %+ i
+%assign i i<<1
+%endrep
+
+%ifdef ARCH_X86_64
+    %define pointer dq
+%else
+    %define pointer dd
+%endif
+
+%macro IF0 1+
+%endmacro
+%macro IF1 1+
+    %1
+%endmacro
+
+section .text align=16
+
+%macro T2_3DN 4 ; z0, z1, mem0, mem1
+    mova     %1, %3
+    mova     %2, %1
+    pfadd    %1, %4
+    pfsub    %2, %4
+%endmacro
+
+%macro T4_3DN 6 ; z0, z1, z2, z3, tmp0, tmp1
+    mova     %5, %3
+    pfsub    %3, %4
+    pfadd    %5, %4 ; {t6,t5}
+    pxor     %3, [ps_m1p1 GLOBAL] ; {t8,t7}
+    mova     %6, %1
+    pswapd   %3, %3
+    pfadd    %1, %5 ; {r0,i0}
+    pfsub    %6, %5 ; {r2,i2}
+    mova     %4, %2
+    pfadd    %2, %3 ; {r1,i1}
+    pfsub    %4, %3 ; {r3,i3}
+    SWAP     %3, %6
+%endmacro
+
+; in:  %1={r0,i0,r1,i1} %2={r2,i2,r3,i3}
+; out: %1={r0,r1,r2,r3} %2={i0,i1,i2,i3}
+%macro T4_SSE 3
+    mova     %3, %1
+    shufps   %1, %2, 0x64 ; {r0,i0,r3,i2}
+    shufps   %3, %2, 0xce ; {r1,i1,r2,i3}
+    mova     %2, %1
+    addps    %1, %3       ; {t1,t2,t6,t5}
+    subps    %2, %3       ; {t3,t4,t8,t7}
+    mova     %3, %1
+    shufps   %1, %2, 0x44 ; {t1,t2,t3,t4}
+    shufps   %3, %2, 0xbe ; {t6,t5,t7,t8}
+    mova     %2, %1
+    addps    %1, %3       ; {r0,i0,r1,i1}
+    subps    %2, %3       ; {r2,i2,r3,i3}
+    mova     %3, %1
+    shufps   %1, %2, 0x88 ; {r0,r1,r2,r3}
+    shufps   %3, %2, 0xdd ; {i0,i1,i2,i3}
+    SWAP     %2, %3
+%endmacro
+
+%macro T8_SSE 6 ; r0,i0,r1,i1,t0,t1
+    mova     %5, %3
+    shufps   %3, %4, 0x44 ; {r4,i4,r6,i6}
+    shufps   %5, %4, 0xee ; {r5,i5,r7,i7}
+    mova     %6, %3
+    subps    %3, %5       ; {r5,i5,r7,i7}
+    addps    %6, %5       ; {t1,t2,t3,t4}
+    mova     %5, %3
+    shufps   %5, %5, 0xb1 ; {i5,r5,i7,r7}
+    mulps    %3, [ps_root2mppm GLOBAL] ; {-r5,i5,r7,-i7}
+    mulps    %5, [ps_root2 GLOBAL]
+    addps    %3, %5       ; {t8,t7,ta,t9}
+    mova     %5, %6
+    shufps   %6, %3, 0x36 ; {t3,t2,t9,t8}
+    shufps   %5, %3, 0x9c ; {t1,t4,t7,ta}
+    mova     %3, %6
+    addps    %6, %5       ; {t1,t2,t9,ta}
+    subps    %3, %5       ; {t6,t5,tc,tb}
+    mova     %5, %6
+    shufps   %6, %3, 0xd8 ; {t1,t9,t5,tb}
+    shufps   %5, %3, 0x8d ; {t2,ta,t6,tc}
+    mova     %3, %1
+    mova     %4, %2
+    addps    %1, %6       ; {r0,r1,r2,r3}
+    addps    %2, %5       ; {i0,i1,i2,i3}
+    subps    %3, %6       ; {r4,r5,r6,r7}
+    subps    %4, %5       ; {i4,i5,i6,i7}
+%endmacro
+
+; scheduled for cpu-bound sizes
+%macro PASS_SMALL 3 ; (to load m4-m7), wre, wim
+IF%1 mova    m4, Z(4)
+IF%1 mova    m5, Z(5)
+    mova     m0, %2 ; wre
+    mova     m2, m4
+    mova     m1, %3 ; wim
+    mova     m3, m5
+    mulps    m2, m0 ; r2*wre
+IF%1 mova    m6, Z(6)
+    mulps    m3, m1 ; i2*wim
+IF%1 mova    m7, Z(7)
+    mulps    m4, m1 ; r2*wim
+    mulps    m5, m0 ; i2*wre
+    addps    m2, m3 ; r2*wre + i2*wim
+    mova     m3, m1
+    mulps    m1, m6 ; r3*wim
+    subps    m5, m4 ; i2*wre - r2*wim
+    mova     m4, m0
+    mulps    m3, m7 ; i3*wim
+    mulps    m4, m6 ; r3*wre
+    mulps    m0, m7 ; i3*wre
+    subps    m4, m3 ; r3*wre - i3*wim
+    mova     m3, Z(0)
+    addps    m0, m1 ; i3*wre + r3*wim
+    mova     m1, m4
+    addps    m4, m2 ; t5
+    subps    m1, m2 ; t3
+    subps    m3, m4 ; r2
+    addps    m4, Z(0) ; r0
+    mova     m6, Z(2)
+    mova   Z(4), m3
+    mova   Z(0), m4
+    mova     m3, m5
+    subps    m5, m0 ; t4
+    mova     m4, m6
+    subps    m6, m5 ; r3
+    addps    m5, m4 ; r1
+    mova   Z(6), m6
+    mova   Z(2), m5
+    mova     m2, Z(3)
+    addps    m3, m0 ; t6
+    subps    m2, m1 ; i3
+    mova     m7, Z(1)
+    addps    m1, Z(3) ; i1
+    mova   Z(7), m2
+    mova   Z(3), m1
+    mova     m4, m7
+    subps    m7, m3 ; i2
+    addps    m3, m4 ; i0
+    mova   Z(5), m7
+    mova   Z(1), m3
+%endmacro
+
+; scheduled to avoid store->load aliasing
+%macro PASS_BIG 1 ; (!interleave)
+    mova     m4, Z(4) ; r2
+    mova     m5, Z(5) ; i2
+    mova     m2, m4
+    mova     m0, [wq] ; wre
+    mova     m3, m5
+    mova     m1, [wq+o1q] ; wim
+    mulps    m2, m0 ; r2*wre
+    mova     m6, Z(6) ; r3
+    mulps    m3, m1 ; i2*wim
+    mova     m7, Z(7) ; i3
+    mulps    m4, m1 ; r2*wim
+    mulps    m5, m0 ; i2*wre
+    addps    m2, m3 ; r2*wre + i2*wim
+    mova     m3, m1
+    mulps    m1, m6 ; r3*wim
+    subps    m5, m4 ; i2*wre - r2*wim
+    mova     m4, m0
+    mulps    m3, m7 ; i3*wim
+    mulps    m4, m6 ; r3*wre
+    mulps    m0, m7 ; i3*wre
+    subps    m4, m3 ; r3*wre - i3*wim
+    mova     m3, Z(0)
+    addps    m0, m1 ; i3*wre + r3*wim
+    mova     m1, m4
+    addps    m4, m2 ; t5
+    subps    m1, m2 ; t3
+    subps    m3, m4 ; r2
+    addps    m4, Z(0) ; r0
+    mova     m6, Z(2)
+    mova   Z(4), m3
+    mova   Z(0), m4
+    mova     m3, m5
+    subps    m5, m0 ; t4
+    mova     m4, m6
+    subps    m6, m5 ; r3
+    addps    m5, m4 ; r1
+IF%1 mova  Z(6), m6
+IF%1 mova  Z(2), m5
+    mova     m2, Z(3)
+    addps    m3, m0 ; t6
+    subps    m2, m1 ; i3
+    mova     m7, Z(1)
+    addps    m1, Z(3) ; i1
+IF%1 mova  Z(7), m2
+IF%1 mova  Z(3), m1
+    mova     m4, m7
+    subps    m7, m3 ; i2
+    addps    m3, m4 ; i0
+IF%1 mova  Z(5), m7
+IF%1 mova  Z(1), m3
+%if %1==0
+    mova     m4, m5 ; r1
+    mova     m0, m6 ; r3
+    unpcklps m5, m1
+    unpckhps m4, m1
+    unpcklps m6, m2
+    unpckhps m0, m2
+    mova     m1, Z(0)
+    mova     m2, Z(4)
+    mova   Z(2), m5
+    mova   Z(3), m4
+    mova   Z(6), m6
+    mova   Z(7), m0
+    mova     m5, m1 ; r0
+    mova     m4, m2 ; r2
+    unpcklps m1, m3
+    unpckhps m5, m3
+    unpcklps m2, m7
+    unpckhps m4, m7
+    mova   Z(0), m1
+    mova   Z(1), m5
+    mova   Z(4), m2
+    mova   Z(5), m4
+%endif
+%endmacro
+
+%macro PUNPCK 3
+    mova      %3, %1
+    punpckldq %1, %2
+    punpckhdq %3, %2
+%endmacro
+
+INIT_XMM
+
+%define Z(x) [r0+mmsize*x]
+
+align 16
+fft4_sse:
+    mova     m0, Z(0)
+    mova     m1, Z(1)
+    T4_SSE   m0, m1, m2
+    mova   Z(0), m0
+    mova   Z(1), m1
+    ret
+
+align 16
+fft8_sse:
+    mova     m0, Z(0)
+    mova     m1, Z(1)
+    T4_SSE   m0, m1, m2
+    mova     m2, Z(2)
+    mova     m3, Z(3)
+    T8_SSE   m0, m1, m2, m3, m4, m5
+    mova   Z(0), m0
+    mova   Z(1), m1
+    mova   Z(2), m2
+    mova   Z(3), m3
+    ret
+
+align 16
+fft16_sse:
+    mova     m0, Z(0)
+    mova     m1, Z(1)
+    T4_SSE   m0, m1, m2
+    mova     m2, Z(2)
+    mova     m3, Z(3)
+    T8_SSE   m0, m1, m2, m3, m4, m5
+    mova     m4, Z(4)
+    mova     m5, Z(5)
+    mova   Z(0), m0
+    mova   Z(1), m1
+    mova   Z(2), m2
+    mova   Z(3), m3
+    T4_SSE   m4, m5, m6
+    mova     m6, Z(6)
+    mova     m7, Z(7)
+    T4_SSE   m6, m7, m0
+    PASS_SMALL 0, [ff_cos_16 GLOBAL], [ff_cos_16+16 GLOBAL]
+    ret
+
+
+INIT_MMX
+
+%macro FFT48_3DN 1
+align 16
+fft4%1:
+    T2_3DN   m0, m1, Z(0), Z(1)
+    mova     m2, Z(2)
+    mova     m3, Z(3)
+    T4_3DN   m0, m1, m2, m3, m4, m5
+    PUNPCK   m0, m1, m4
+    PUNPCK   m2, m3, m5
+    mova   Z(0), m0
+    mova   Z(1), m4
+    mova   Z(2), m2
+    mova   Z(3), m5
+    ret
+
+align 16
+fft8%1:
+    T2_3DN   m0, m1, Z(0), Z(1)
+    mova     m2, Z(2)
+    mova     m3, Z(3)
+    T4_3DN   m0, m1, m2, m3, m4, m5
+    mova   Z(0), m0
+    mova   Z(2), m2
+    T2_3DN   m4, m5, Z(4), Z(5)
+    T2_3DN   m6, m7, Z(6), Z(7)
+    pswapd   m0, m5
+    pswapd   m2, m7
+    pxor     m0, [ps_m1p1 GLOBAL]
+    pxor     m2, [ps_m1p1 GLOBAL]
+    pfsub    m5, m0
+    pfadd    m7, m2
+    pfmul    m5, [ps_root2 GLOBAL]
+    pfmul    m7, [ps_root2 GLOBAL]
+    T4_3DN   m1, m3, m5, m7, m0, m2
+    mova   Z(5), m5
+    mova   Z(7), m7
+    mova     m0, Z(0)
+    mova     m2, Z(2)
+    T4_3DN   m0, m2, m4, m6, m5, m7
+    PUNPCK   m0, m1, m5
+    PUNPCK   m2, m3, m7
+    mova   Z(0), m0
+    mova   Z(1), m5
+    mova   Z(2), m2
+    mova   Z(3), m7
+    PUNPCK   m4, Z(5), m5
+    PUNPCK   m6, Z(7), m7
+    mova   Z(4), m4
+    mova   Z(5), m5
+    mova   Z(6), m6
+    mova   Z(7), m7
+    ret
+%endmacro
+
+FFT48_3DN _3dn2
+
+%macro pswapd 2
+%ifidn %1, %2
+    movd [r0+12], %1
+    punpckhdq %1, [r0+8]
+%else
+    movq  %1, %2
+    psrlq %1, 32
+    punpckldq %1, %2
+%endif
+%endmacro
+
+FFT48_3DN _3dn
+
+
+%define Z(x) [zq + o1q*(x&6)*((x/6)^1) + o3q*(x/6) + mmsize*(x&1)]
+
+%macro DECL_PASS 2+ ; name, payload
+align 16
+%1:
+DEFINE_ARGS z, w, n, o1, o3
+    lea o3q, [nq*3]
+    lea o1q, [nq*8]
+    shl o3q, 4
+.loop:
+    %2
+    add zq, mmsize*2
+    add wq, mmsize
+    sub nd, mmsize/8
+    jg .loop
+    rep ret
+%endmacro
+
+INIT_XMM
+DECL_PASS pass_sse, PASS_BIG 1
+DECL_PASS pass_interleave_sse, PASS_BIG 0
+
+INIT_MMX
+%define mulps pfmul
+%define addps pfadd
+%define subps pfsub
+%define unpcklps punpckldq
+%define unpckhps punpckhdq
+DECL_PASS pass_3dn, PASS_SMALL 1, [wq], [wq+o1q]
+DECL_PASS pass_interleave_3dn, PASS_BIG 0
+%define pass_3dn2 pass_3dn
+%define pass_interleave_3dn2 pass_interleave_3dn
+
+
+%macro DECL_FFT 2-3 ; nbits, cpu, suffix
+%xdefine list_of_fft fft4%2, fft8%2
+%if %1==5
+%xdefine list_of_fft list_of_fft, fft16%2
+%endif
+
+%assign n 1<<%1
+%rep 17-%1
+%assign n2 n/2
+%assign n4 n/4
+%xdefine list_of_fft list_of_fft, fft %+ n %+ %3%2
+
+align 16
+fft %+ n %+ %3%2:
+    call fft %+ n2 %+ %2
+    add r0, n*4 - (n&(-2<<%1))
+    call fft %+ n4 %+ %2
+    add r0, n*2 - (n2&(-2<<%1))
+    call fft %+ n4 %+ %2
+    sub r0, n*6 + (n2&(-2<<%1))
+    lea r1, [ff_cos_ %+ n GLOBAL]
+    mov r2d, n4/2
+    jmp pass%3%2
+
+%assign n n*2
+%endrep
+%undef n
+
+align 8
+dispatch_tab%3%2: pointer list_of_fft
+
+; On x86_32, this function does the register saving and restoring for all of fft.
+; The others pass args in registers and don't spill anything.
+cglobal ff_fft_dispatch%3%2, 2,5,0, z, nbits
+    lea r2, [dispatch_tab%3%2 GLOBAL]
+    mov r2, [r2 + (nbitsq-2)*gprsize]
+    call r2
+    RET
+%endmacro ; DECL_FFT
+
+DECL_FFT 5, _sse
+DECL_FFT 5, _sse, _interleave
+DECL_FFT 4, _3dn
+DECL_FFT 4, _3dn, _interleave
+DECL_FFT 4, _3dn2
+DECL_FFT 4, _3dn2, _interleave
+
diff --git a/libavcodec/i386/fft_sse.c b/libavcodec/i386/fft_sse.c
index 305f44a0ce..77a579011a 100644
--- a/libavcodec/i386/fft_sse.c
+++ b/libavcodec/i386/fft_sse.c
@@ -22,124 +22,55 @@
 #include "libavutil/x86_cpu.h"
 #include "libavcodec/dsputil.h"
 
-static const int p1p1p1m1[4] __attribute__((aligned(16))) =
-    { 0, 0, 0, 1 << 31 };
-
-static const int p1p1m1p1[4] __attribute__((aligned(16))) =
-    { 0, 0, 1 << 31, 0 };
-
-static const int p1p1m1m1[4] __attribute__((aligned(16))) =
-    { 0, 0, 1 << 31, 1 << 31 };
-
 static const int p1m1p1m1[4] __attribute__((aligned(16))) =
     { 0, 1 << 31, 0, 1 << 31 };
 
 static const int m1m1m1m1[4] __attribute__((aligned(16))) =
     { 1 << 31, 1 << 31, 1 << 31, 1 << 31 };
 
-#if 0
-static void print_v4sf(const char *str, __m128 a)
-{
-    float *p = (float *)&a;
-    printf("%s: %f %f %f %f\n",
-           str, p[0], p[1], p[2], p[3]);
-}
-#endif
+void ff_fft_dispatch_sse(FFTComplex *z, int nbits);
+void ff_fft_dispatch_interleave_sse(FFTComplex *z, int nbits);
 
-/* XXX: handle reverse case */
 void ff_fft_calc_sse(FFTContext *s, FFTComplex *z)
 {
-    int ln = s->nbits;
-    x86_reg i;
-    long j;
-    long nblocks, nloops;
-    FFTComplex *p, *cptr;
+    int n = 1 << s->nbits;
 
-    asm volatile(
-        "movaps %0, %%xmm4 \n\t"
-        "movaps %1, %%xmm5 \n\t"
-        ::"m"(*p1p1m1m1),
-          "m"(*(s->inverse ? p1p1m1p1 : p1p1p1m1))
-    );
+    ff_fft_dispatch_interleave_sse(z, s->nbits);
 
-    i = 8 << ln;
-    asm volatile(
-        "1: \n\t"
-        "sub $32, %0 \n\t"
-        /* do the pass 0 butterfly */
-        "movaps   (%0,%1), %%xmm0 \n\t"
-        "movaps    %%xmm0, %%xmm1 \n\t"
-        "shufps     $0x4E, %%xmm0, %%xmm0 \n\t"
-        "xorps     %%xmm4, %%xmm1 \n\t"
-        "addps     %%xmm1, %%xmm0 \n\t"
-        "movaps 16(%0,%1), %%xmm2 \n\t"
-        "movaps    %%xmm2, %%xmm3 \n\t"
-        "shufps     $0x4E, %%xmm2, %%xmm2 \n\t"
-        "xorps     %%xmm4, %%xmm3 \n\t"
-        "addps     %%xmm3, %%xmm2 \n\t"
-        /* multiply third by -i */
-        /* by toggling the sign bit */
-        "shufps     $0xB4, %%xmm2, %%xmm2 \n\t"
-        "xorps     %%xmm5, %%xmm2 \n\t"
-        /* do the pass 1 butterfly */
-        "movaps    %%xmm0, %%xmm1 \n\t"
-        "addps     %%xmm2, %%xmm0 \n\t"
-        "subps     %%xmm2, %%xmm1 \n\t"
-        "movaps    %%xmm0,   (%0,%1) \n\t"
-        "movaps    %%xmm1, 16(%0,%1) \n\t"
-        "jg 1b \n\t"
-        :"+r"(i)
-        :"r"(z)
-    );
-    /* pass 2 .. ln-1 */
+    if(n <= 16) {
+        x86_reg i = -8*n;
+        asm volatile(
+            "1: \n"
+            "movaps     (%0,%1), %%xmm0 \n"
+            "movaps      %%xmm0, %%xmm1 \n"
+            "unpcklps 16(%0,%1), %%xmm0 \n"
+            "unpckhps 16(%0,%1), %%xmm1 \n"
+            "movaps      %%xmm0,   (%0,%1) \n"
+            "movaps      %%xmm1, 16(%0,%1) \n"
+            "add $32, %0 \n"
+            "jl 1b \n"
+            :"+r"(i)
+            :"r"(z+n)
+            :"memory"
+        );
+    }
+}
 
-    nblocks = 1 << (ln-3);
-    nloops = 1 << 2;
-    cptr = s->exptab1;
-    do {
-        p = z;
-        j = nblocks;
-        do {
-            i = nloops*8;
-            asm volatile(
-                "1: \n\t"
-                "sub $32, %0 \n\t"
-                "movaps    (%2,%0), %%xmm1 \n\t"
-                "movaps    (%1,%0), %%xmm0 \n\t"
-                "movaps  16(%2,%0), %%xmm5 \n\t"
-                "movaps  16(%1,%0), %%xmm4 \n\t"
-                "movaps     %%xmm1, %%xmm2 \n\t"
-                "movaps     %%xmm5, %%xmm6 \n\t"
-                "shufps      $0xA0, %%xmm1, %%xmm1 \n\t"
-                "shufps      $0xF5, %%xmm2, %%xmm2 \n\t"
-                "shufps      $0xA0, %%xmm5, %%xmm5 \n\t"
-                "shufps      $0xF5, %%xmm6, %%xmm6 \n\t"
-                "mulps   (%3,%0,2), %%xmm1 \n\t" //  cre*re cim*re
-                "mulps 16(%3,%0,2), %%xmm2 \n\t" // -cim*im cre*im
-                "mulps 32(%3,%0,2), %%xmm5 \n\t" //  cre*re cim*re
-                "mulps 48(%3,%0,2), %%xmm6 \n\t" // -cim*im cre*im
-                "addps      %%xmm2, %%xmm1 \n\t"
-                "addps      %%xmm6, %%xmm5 \n\t"
-                "movaps     %%xmm0, %%xmm3 \n\t"
-                "movaps     %%xmm4, %%xmm7 \n\t"
-                "addps      %%xmm1, %%xmm0 \n\t"
-                "subps      %%xmm1, %%xmm3 \n\t"
-                "addps      %%xmm5, %%xmm4 \n\t"
-                "subps      %%xmm5, %%xmm7 \n\t"
-                "movaps     %%xmm0, (%1,%0) \n\t"
-                "movaps     %%xmm3, (%2,%0) \n\t"
-                "movaps     %%xmm4, 16(%1,%0) \n\t"
-                "movaps     %%xmm7, 16(%2,%0) \n\t"
-                "jg 1b \n\t"
-                :"+r"(i)
-                :"r"(p), "r"(p + nloops), "r"(cptr)
-            );
-            p += nloops*2;
-        } while (--j);
-        cptr += nloops*2;
-        nblocks >>= 1;
-        nloops <<= 1;
-    } while (nblocks != 0);
+void ff_fft_permute_sse(FFTContext *s, FFTComplex *z)
+{
+    int n = 1 << s->nbits;
+    int i;
+    for(i=0; i<n; i+=2) {
+        asm volatile(
+            "movaps %2, %%xmm0 \n"
+            "movlps %%xmm0, %0 \n"
+            "movhps %%xmm0, %1 \n"
+            :"=m"(s->tmp_buf[s->revtab[i]]),
+             "=m"(s->tmp_buf[s->revtab[i+1]])
+            :"m"(z[i])
+        );
+    }
+    memcpy(z, s->tmp_buf, n*sizeof(FFTComplex));
 }
 
 static void imdct_sse(MDCTContext *s, const FFTSample *input, FFTSample *tmp)