ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.

2.5% faster fft, 0.5% faster vorbis.

Originally committed as revision 6023 to svn://svn.ffmpeg.org/ffmpeg/trunk

4 files changed, 233 insertions, 285 deletions

diff --git a/libavcodec/fft.c b/libavcodec/fft.c
index 6fd87167bd..bc514724c0 100644
--- a/libavcodec/fft.c
+++ b/libavcodec/fft.c
@@ -58,7 +58,7 @@ int ff_fft_init(FFTContext *s, int nbits, int inverse)
     s->exptab1 = NULL;
 
     /* compute constant table for HAVE_SSE version */
-#if (defined(HAVE_MMX) && (defined(HAVE_BUILTIN_VECTOR) || defined(HAVE_MM3DNOW))) \
+#if defined(HAVE_MMX) \
     || (defined(HAVE_ALTIVEC) && !defined(ALTIVEC_USE_REFERENCE_C_CODE))
     {
         int has_vectors = mm_support();
@@ -67,15 +67,12 @@ int ff_fft_init(FFTContext *s, int nbits, int inverse)
 #if defined(HAVE_MMX)
             if (has_vectors & MM_3DNOWEXT)
                 s->imdct_calc = ff_imdct_calc_3dn2;
-#ifdef HAVE_MM3DNOW
             if (has_vectors & MM_3DNOWEXT)
                 /* 3DNowEx for Athlon(XP) */
                 s->fft_calc = ff_fft_calc_3dn2;
             else if (has_vectors & MM_3DNOW)
                 /* 3DNow! for K6-2/3 */
                 s->fft_calc = ff_fft_calc_3dn;
-#endif
-#ifdef HAVE_BUILTIN_VECTOR
             if (has_vectors & MM_SSE2)
                 /* SSE for P4/K8 */
                 s->fft_calc = ff_fft_calc_sse;
@@ -83,7 +80,6 @@ int ff_fft_init(FFTContext *s, int nbits, int inverse)
                      s->fft_calc == ff_fft_calc_c)
                 /* SSE for P3 */
                 s->fft_calc = ff_fft_calc_sse;
-#endif
 #else /* HAVE_MMX */
             if (has_vectors & MM_ALTIVEC)
                 s->fft_calc = ff_fft_calc_altivec;
diff --git a/libavcodec/i386/fft_3dn.c b/libavcodec/i386/fft_3dn.c
index 16595bddee..00869474ad 100644
--- a/libavcodec/i386/fft_3dn.c
+++ b/libavcodec/i386/fft_3dn.c
@@ -1,6 +1,6 @@
 /*
  * FFT/MDCT transform with 3DNow! optimizations
- * Copyright (c) 2006 Zuxy MENG Jie.
+ * Copyright (c) 2006 Zuxy MENG Jie, Loren Merritt
  * Based on fft_sse.c copyright (c) 2002 Fabrice Bellard.
  *
  * This library is free software; you can redistribute it and/or
@@ -18,11 +18,6 @@
  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  */
 #include "../dsputil.h"
-#include <math.h>
-
-#ifdef HAVE_MM3DNOW
-
-#include <mm3dnow.h>
 
 static const int p1m1[2] __attribute__((aligned(8))) =
     { 0, 1 << 31 };
@@ -33,104 +28,96 @@ static const int m1p1[2] __attribute__((aligned(8))) =
 void ff_fft_calc_3dn(FFTContext *s, FFTComplex *z)
 {
     int ln = s->nbits;
-    int j, np, np2;
-    int nblocks, nloops;
-    register FFTComplex *p, *q;
-    FFTComplex *cptr, *cptr1;
-    int k;
-
-    np = 1 << ln;
-    /* FEMMS not a must here but recommended by AMD */
-    _m_femms();
-
-    {
-        __m64 *r, a0, a1, b0, b1, tmp, c;
-
-        r = (__m64 *)&z[0];
-        if (s->inverse)
-            c = *(__m64 *)m1p1;
-        else
-            c = *(__m64 *)p1m1;
-
-        j = (np >> 2);
-        do {
-            /* do the pass 0 butterfly */
-            a0 = _m_pfadd(r[0], r[1]);
-            a1 = _m_pfsub(r[0], r[1]);
-
-            /* do the pass 0 butterfly */
-            b0 = _m_pfadd(r[2], r[3]);
-            b1 = _m_pfsub(r[2], r[3]);
-
-            /* multiply third by -i */
-            tmp = _m_punpckhdq(b1, b1);
-            b1 = _m_punpckldq(b1, b1);
-            b1 = _m_punpckldq(tmp, b1);
-            b1 = _m_pxor(b1, c);
-
-            /* do the pass 1 butterfly */
-            r[0] = _m_pfadd(a0, b0);
-            r[1] = _m_pfadd(a1, b1);
-            r[2] = _m_pfsub(a0, b0);
-            r[3] = _m_pfsub(a1, b1);
-            r += 4;
-        } while (--j != 0);
-    }
+    long i, j;
+    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 = np >> 3;
+    nblocks = 1 << (ln-3);
     nloops = 1 << 2;
-    np2 = np >> 1;
-
-    cptr1 = s->exptab1;
+    cptr = s->exptab1;
     do {
         p = z;
-        q = z + nloops;
         j = nblocks;
         do {
-            cptr = cptr1;
-            k = nloops >> 1;
-            do {
-                __m64 a0, a1, b0, b1, c0, c1, t10, t11, t20, t21;
-
-                a0 = *(__m64 *)&p[0];
-                a1 = *(__m64 *)&p[1];
-                b0 = *(__m64 *)&q[0];
-                b1 = *(__m64 *)&q[1];
-
-                /* complex mul */
-                c0 = *(__m64 *)&cptr[0];
-                c1 = *(__m64 *)&cptr[1];
-                /*  cre*re cim*re */
-                t10 = _m_pfmul(c0, _m_punpckldq(b0, b0));
-                t11 = _m_pfmul(c1, _m_punpckldq(b1, b1));
-                c0 = *(__m64 *)&cptr[2];
-                c1 = *(__m64 *)&cptr[3];
-                /*  -cim*im cre*im */
-                t20 = _m_pfmul(c0, _m_punpckhdq(b0, b0));
-                t21 = _m_pfmul(c1, _m_punpckhdq(b1, b1));
-                b0 = _m_pfadd(t10, t20);
-                b1 = _m_pfadd(t11, t21);
-
-                /* butterfly */
-                *(__m64 *)&p[0] = _m_pfadd(a0, b0);
-                *(__m64 *)&p[1] = _m_pfadd(a1, b1);
-                *(__m64 *)&q[0] = _m_pfsub(a0, b0);
-                *(__m64 *)&q[1] = _m_pfsub(a1, b1);
-
-                p += 2;
-                q += 2;
-                cptr += 4;
-            } while (--k);
-
-            p += nloops;
-            q += nloops;
+            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);
-        cptr1 += nloops * 2;
-        nblocks = nblocks >> 1;
-        nloops = nloops << 1;
+        cptr += nloops*2;
+        nblocks >>= 1;
+        nloops <<= 1;
     } while (nblocks != 0);
-    _m_femms();
+    asm volatile("femms");
 }
-
-#endif
diff --git a/libavcodec/i386/fft_3dn2.c b/libavcodec/i386/fft_3dn2.c
index b03fae62ee..5f066d175a 100644
--- a/libavcodec/i386/fft_3dn2.c
+++ b/libavcodec/i386/fft_3dn2.c
@@ -18,11 +18,6 @@
  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  */
 #include "../dsputil.h"
-#include <math.h>
-
-#ifdef HAVE_MM3DNOW
-
-#include <mm3dnow.h>
 
 static const int p1m1[2] __attribute__((aligned(8))) =
     { 0, 1 << 31 };
@@ -33,108 +28,97 @@ static const int m1p1[2] __attribute__((aligned(8))) =
 void ff_fft_calc_3dn2(FFTContext *s, FFTComplex *z)
 {
     int ln = s->nbits;
-    int j, np, np2;
-    int nblocks, nloops;
-    register FFTComplex *p, *q;
-    FFTComplex *cptr, *cptr1;
-    int k;
-
-    np = 1 << ln;
-    /* FEMMS is not a must here but recommended by AMD */
-    _m_femms();
-
-    {
-        __m64 *r, a0, a1, b0, b1, c;
-
-        r = (__m64 *)&z[0];
-        if (s->inverse)
-            c = *(__m64 *)m1p1;
-        else
-            c = *(__m64 *)p1m1;
-
-        j = (np >> 2);
-        do {
-            /* do the pass 0 butterfly */
-            a0 = _m_pfadd(r[0], r[1]);
-            a1 = _m_pfsub(r[0], r[1]);
-
-            /* do the pass 0 butterfly */
-            b0 = _m_pfadd(r[2], r[3]);
-            b1 = _m_pfsub(r[2], r[3]);
-
-            /* multiply third by -i */
-            b1 = _m_pswapd(b1);
-            b1 = _m_pxor(b1, c);
-
-            r[0] = _m_pfadd(a0, b0);
-            r[1] = _m_pfadd(a1, b1);
-            r[2] = _m_pfsub(a0, b0);
-            r[3] = _m_pfsub(a1, b1);
-            r += 4;
-        } while (--j != 0);
-    }
+    long i, j;
+    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 = np >> 3;
+    nblocks = 1 << (ln-3);
     nloops = 1 << 2;
-    np2 = np >> 1;
-
-    cptr1 = s->exptab1;
+    cptr = s->exptab1;
     do {
         p = z;
-        q = z + nloops;
         j = nblocks;
         do {
-            cptr = cptr1;
-            k = nloops >> 1;
-            do {
-                __m64 a0, a1, b0, b1, c0, c1, t10, t11, t20, t21;
-
-                a0 = *(__m64 *)&p[0];
-                a1 = *(__m64 *)&p[1];
-                b0 = *(__m64 *)&q[0];
-                b1 = *(__m64 *)&q[1];
-
-                /* complex mul */
-                c0 = *(__m64 *)&cptr[0];
-                c1 = *(__m64 *)&cptr[1];
-                /* cre*re cim*im */
-                t10 = _m_pfmul(c0, b0);
-                t11 = _m_pfmul(c1, b1);
-                /* no need to access cptr[2] & cptr[3] */
-                c0 = _m_pswapd(c0);
-                c1 = _m_pswapd(c1);
-                /* cim*re cre*im */
-                t20 = _m_pfmul(c0, b0);
-                t21 = _m_pfmul(c1, b1);
-
-                /* cre*re-cim*im cim*re+cre*im */
-                b0 = _m_pfpnacc(t10, t20);
-                b1 = _m_pfpnacc(t11, t21);
-
-                /* butterfly */
-                *(__m64 *)&p[0] = _m_pfadd(a0, b0);
-                *(__m64 *)&p[1] = _m_pfadd(a1, b1);
-                *(__m64 *)&q[0] = _m_pfsub(a0, b0);
-                *(__m64 *)&q[1] = _m_pfsub(a1, b1);
-
-                p += 2;
-                q += 2;
-                cptr += 4;
-            } while (--k);
-
-            p += nloops;
-            q += nloops;
+            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);
-        cptr1 += nloops * 2;
-        nblocks = nblocks >> 1;
-        nloops = nloops << 1;
+        cptr += nloops*2;
+        nblocks >>= 1;
+        nloops <<= 1;
     } while (nblocks != 0);
-    _m_femms();
+    asm volatile("femms");
 }
 
-#endif
-
 void ff_imdct_calc_3dn2(MDCTContext *s, FFTSample *output,
                         const FFTSample *input, FFTSample *tmp)
 {
@@ -221,3 +205,4 @@ void ff_imdct_calc_3dn2(MDCTContext *s, FFTSample *output,
     }
     asm volatile("femms");
 }
+
diff --git a/libavcodec/i386/fft_sse.c b/libavcodec/i386/fft_sse.c
index 631848265a..cc2ba2a5ac 100644
--- a/libavcodec/i386/fft_sse.c
+++ b/libavcodec/i386/fft_sse.c
@@ -17,11 +17,6 @@
  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  */
 #include "../dsputil.h"
-#include <math.h>
-
-#ifdef HAVE_BUILTIN_VECTOR
-
-#include <xmmintrin.h>
 
 static const int p1p1p1m1[4] __attribute__((aligned(16))) =
     { 0, 0, 0, 1 << 31 };
@@ -45,96 +40,81 @@ static void print_v4sf(const char *str, __m128 a)
 void ff_fft_calc_sse(FFTContext *s, FFTComplex *z)
 {
     int ln = s->nbits;
-    int         j, np, np2;
-    int         nblocks, nloops;
-    register FFTComplex *p, *q;
-    FFTComplex *cptr, *cptr1;
-    int k;
-
-    np = 1 << ln;
-
-    {
-        __m128 *r, a, b, a1, c1, c2;
-
-        r = (__m128 *)&z[0];
-        c1 = *(__m128 *)p1p1m1m1;
-        if (s->inverse)
-            c2 = *(__m128 *)p1p1m1p1;
-        else
-            c2 = *(__m128 *)p1p1p1m1;
-
-        j = (np >> 2);
-        do {
-            a = r[0];
-            b = _mm_shuffle_ps(a, a, _MM_SHUFFLE(1, 0, 3, 2));
-            a = _mm_xor_ps(a, c1);
-            /* do the pass 0 butterfly */
-            a = _mm_add_ps(a, b);
-
-            a1 = r[1];
-            b = _mm_shuffle_ps(a1, a1, _MM_SHUFFLE(1, 0, 3, 2));
-            a1 = _mm_xor_ps(a1, c1);
-            /* do the pass 0 butterfly */
-            b = _mm_add_ps(a1, b);
-
-            /* multiply third by -i */
-            /* by toggling the sign bit */
-            b = _mm_shuffle_ps(b, b, _MM_SHUFFLE(2, 3, 1, 0));
-            b = _mm_xor_ps(b, c2);
-
-            /* do the pass 1 butterfly */
-            r[0] = _mm_add_ps(a, b);
-            r[1] = _mm_sub_ps(a, b);
-            r += 2;
-        } while (--j != 0);
-    }
+    long i, j;
+    long nblocks, nloops;
+    FFTComplex *p, *cptr;
+
+    asm volatile(
+        "movaps %0, %%xmm4 \n\t"
+        "movaps %1, %%xmm5 \n\t"
+        ::"m"(*p1p1m1m1),
+          "m"(*(s->inverse ? p1p1m1p1 : p1p1p1m1))
+    );
+
+    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 */
 
-    nblocks = np >> 3;
+    nblocks = 1 << (ln-3);
     nloops = 1 << 2;
-    np2 = np >> 1;
-
-    cptr1 = s->exptab1;
+    cptr = s->exptab1;
     do {
         p = z;
-        q = z + nloops;
         j = nblocks;
         do {
-            cptr = cptr1;
-            k = nloops >> 1;
-            do {
-                __m128 a, b, c, t1, t2;
-
-                a = *(__m128 *)p;
-                b = *(__m128 *)q;
-
-                /* complex mul */
-                c = *(__m128 *)cptr;
-                /*  cre*re cim*re */
-                t1 = _mm_mul_ps(c,
-                                _mm_shuffle_ps(b, b, _MM_SHUFFLE(2, 2, 0, 0)));
-                c = *(__m128 *)(cptr + 2);
-                /*  -cim*im cre*im */
-                t2 = _mm_mul_ps(c,
-                                _mm_shuffle_ps(b, b, _MM_SHUFFLE(3, 3, 1, 1)));
-                b = _mm_add_ps(t1, t2);
-
-                /* butterfly */
-                *(__m128 *)p = _mm_add_ps(a, b);
-                *(__m128 *)q = _mm_sub_ps(a, b);
-
-                p += 2;
-                q += 2;
-                cptr += 4;
-            } while (--k);
-
-            p += nloops;
-            q += nloops;
+            i = nloops*8;
+            asm volatile(
+                "1: \n\t"
+                "sub $16, %0 \n\t"
+                "movaps    (%2,%0), %%xmm1 \n\t"
+                "movaps    (%1,%0), %%xmm0 \n\t"
+                "movaps     %%xmm1, %%xmm2 \n\t"
+                "shufps      $0xA0, %%xmm1, %%xmm1 \n\t"
+                "shufps      $0xF5, %%xmm2, %%xmm2 \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
+                "addps      %%xmm2, %%xmm1 \n\t"
+                "movaps     %%xmm0, %%xmm3 \n\t"
+                "addps      %%xmm1, %%xmm0 \n\t"
+                "subps      %%xmm1, %%xmm3 \n\t"
+                "movaps     %%xmm0, (%1,%0) \n\t"
+                "movaps     %%xmm3, (%2,%0) \n\t"
+                "jg 1b \n\t"
+                :"+r"(i)
+                :"r"(p), "r"(p + nloops), "r"(cptr)
+            );
+            p += nloops*2;
         } while (--j);
-        cptr1 += nloops * 2;
-        nblocks = nblocks >> 1;
-        nloops = nloops << 1;
+        cptr += nloops*2;
+        nblocks >>= 1;
+        nloops <<= 1;
     } while (nblocks != 0);
 }
 
-#endif