From ae305ca4009575b4cf427dc3903a482de5e99591 Mon Sep 17 00:00:00 2001
From: Mark Borgerding <mark@borgerding.net>
Date: Fri, 7 Nov 2003 03:31:59 +0000
Subject: [PATCH] slightly faster

---
 kiss_fft.c | 22 ++++++++++++----------
 1 file changed, 12 insertions(+), 10 deletions(-)

diff --git a/kiss_fft.c b/kiss_fft.c
index 8df9c52..f58f88a 100644
--- a/kiss_fft.c
+++ b/kiss_fft.c
@@ -287,12 +287,12 @@ void bfly_generic(
         int p
         )
 {
-    int u,q,d,fsm,halfp,mp;
+    int u,q,d,fsm,halfp,mp,uf;
     kiss_fft_cpx * scratch = st->scratch;
     kiss_fft_cpx * scratch2 = scratch + p;
     kiss_fft_cpx * tw = st->twiddles;
     kiss_fft_cpx tlo,t3,t4;
-    kiss_fft_cpx *Foutlo,*Fouthi;
+    kiss_fft_cpx *Foutlo,*Fouthi,*tw2;
 
     fsm = fstride*m;
     halfp=p/2;
@@ -313,31 +313,33 @@ void bfly_generic(
     for ( u=0; u<m; ++u ) {
         scratch[0] = Fout[0];
         scratch2 = scratch + p;
+        uf=u*fstride;
 
+        /* d==0 */
+        Foutlo=Fout;
         for ( q=1 ; q<p ; ++q ) {
-            C_MUL( scratch[q] , Fout[  m*q ] ,  tw[u*fstride*q] );
-            Fout[  m*q ] = scratch[0];
+            Foutlo += m;
+            C_MUL( scratch[q] , *Foutlo ,  tw[uf*q] );
+            *Foutlo = scratch[0];
         }
         
         for ( q=1; q<p ; ++q ) {
-            
-
             C_ADDTO(*Fout , scratch[q] );
             Foutlo=Fout;
             Fouthi=Fout + mp;
-            
 
             for ( d=1; d<=halfp;++d) {
-                Foutlo += m;
-                Fouthi -= m;
-
                 tlo = *scratch2++;
 
+                Foutlo += m;
+
                 t3.r = scratch[q].r * tlo.r;
                 t3.i = scratch[q].r * tlo.i;
                 t4.r = scratch[q].i * tlo.i;
                 t4.i = - scratch[q].i * tlo.r;
 
+                Fouthi -= m;
+
                 Foutlo->r += t3.r - t4.r;
                 Fouthi->r += t3.r + t4.r;
                 Foutlo->i += t3.i - t4.i;