diff --git a/kiss_fft.c b/kiss_fft.c
index 4272e65..f9d0ce4 100644
--- a/kiss_fft.c
+++ b/kiss_fft.c
@@ -59,6 +59,8 @@ typedef struct {
 #   define C_FIXDIV(c,div) /* NOOP */
 #endif
 
+#define  C_ADD( res, a,b)\
+    do {    (res).r=(a).r+(b).r;  (res).i=(a).i+(b).i;  }while(0)
 #define  C_SUB( res, a,b)\
     do {    (res).r=(a).r-(b).r;  (res).i=(a).i-(b).i;  }while(0)
 #define C_ADDTO( res , a)\
@@ -162,6 +164,59 @@ void bfly4(
     }while(--m);
 }
 
+void bfly3(
+         kiss_fft_cpx * Fout,
+         int fstride,
+         const kiss_fft_state * st,
+         int m
+         )
+{
+     kiss_fft_cpx *Fout0,*Fout1,*Fout2;
+     kiss_fft_cpx *tw1,*tw2;
+     kiss_fft_cpx * scratch = st->scratch;
+     kiss_fft_cpx * twiddles = st->twiddles;
+     kiss_fft_cpx epi3;
+     epi3 = twiddles[fstride*m];
+
+     Fout0=Fout;
+     Fout1=Fout0+m;
+     Fout2=Fout0+2*m;
+     tw1=tw2=twiddles;
+
+     do{
+
+         C_FIXDIV(*Fout,3); C_FIXDIV(*Fout1,3); C_FIXDIV(*Fout2,3);
+
+         scratch[0] = *Fout0;
+
+         C_MUL(scratch[1],*Fout1 , *tw1);
+         C_MUL(scratch[2],*Fout2 , *tw2);
+         tw1 += fstride;
+         tw2 += fstride*2;
+
+         C_ADD(scratch[4],scratch[1],scratch[2]);
+         C_ADD(*Fout0,scratch[0],scratch[4]);
+
+         scratch[4].r /= -2;
+         scratch[4].i /= -2;
+
+         C_SUB(scratch[5],scratch[1],scratch[2]);
+         scratch[5].r *= epi3.i;
+         scratch[5].i *= epi3.i;
+
+         scratch[3].r = scratch[4].r - scratch[5].i ;
+         scratch[3].i = scratch[4].i + scratch[5].r ;
+
+         C_ADD( *Fout1, scratch[0] , scratch[3] );
+
+         scratch[3].r = scratch[4].r + scratch[5].i;
+         scratch[3].i = scratch[4].i - scratch[5].r;
+         C_ADD( *Fout2, scratch[0] , scratch[3] );
+
+         ++Fout0;++Fout1;++Fout2;
+     }while(--m);
+}
+
 /* perform the butterfly for one stage of a mixed radix FFT */
 void bfly_generic(
         kiss_fft_cpx * Fout,
@@ -221,8 +276,11 @@ void fft_work(
     }
 
     switch (p) {
-        case 4: bfly4(Fout,fstride,st,m); break;
         case 2: bfly2(Fout,fstride,st,m); break;
+#if 1
+        case 3: bfly3(Fout,fstride,st,m); break;
+#endif                
+        case 4: bfly4(Fout,fstride,st,m); break;
         default: bfly_generic(Fout,fstride,st,m,p); break;
     }
 }
diff --git a/test/Makefile b/test/Makefile
index 15bd87a..b07d4ea 100644
--- a/test/Makefile
+++ b/test/Makefile
@@ -1,11 +1,10 @@
 
 ifeq "$(NFFT)" ""
-	NFFT=1024
+ NFFT=1024
 endif
 
-
 ifeq "$(NUMFFTS)" ""
-	NUMFFTS=10000
+ NUMFFTS=10000
 endif
 
 UTILSRC=../kiss_fft.c fftutil.c
@@ -13,7 +12,7 @@ UTILSRC=../kiss_fft.c fftutil.c
 CFLAGS=-Wall -O3 -lm -I.. -ansi -pedantic
 
 ifeq "$(DATATYPE)" ""
-	DATATYPE=float
+  DATATYPE=float
 endif
 UTIL=fftutil_$(DATATYPE)
 
@@ -21,9 +20,9 @@ UTIL=fftutil_$(DATATYPE)
 all: $(UTIL)
 
 ifeq  "$(DATATYPE)" "short"
-	TYPEFLAGS=-DFIXED_POINT -Dkiss_fft_scalar=short
+ TYPEFLAGS=-DFIXED_POINT -Dkiss_fft_scalar=short
 else	
-	TYPEFLAGS=-Dkiss_fft_scalar=$(DATATYPE)
+ TYPEFLAGS=-Dkiss_fft_scalar=$(DATATYPE)
 endif
 
 $(UTIL): $(UTILSRC)
@@ -47,7 +46,7 @@ time: all $(RANDDAT)
 POW2=256 512 1024 2048
 POW3=243 729 2187
 mtime: all
-	@for n in $(POW2) $(POW3) ;do \
+	@for n in $(POW3) ;do \
 		export NFFT=$$n;make time; \
 		done
 
@@ -56,7 +55,7 @@ snr: all
 	@echo
 	@echo "### testing SNR for $(NFFT) point FFTs"
 	@echo "#### $(DATATYPE)"
-	@echo "testkiss($(NFFT),'$(DATATYPE)',1,1/$(NFFT));" | octave -q
+	@echo "testkiss( $(NFFT) , '$(DATATYPE)' );" | octave -q
 
 test: snr time 
 
diff --git a/test/testkiss.m b/test/testkiss.m
index fc86c17..ba04eff 100755
--- a/test/testkiss.m
+++ b/test/testkiss.m
@@ -1,11 +1,11 @@
-function snr= testkiss( nfft , prec ,scale_t2f ,scale_f2t )
-if nargin<1, nfft=1024
-endif
-if nargin<2, prec='float'
-endif
-if nargin<3, scale_t2f=1
-endif
-if nargin<4, scale_f2t=1
+function snr= testkiss( nfft , prec )
+
+if strcmp( prec ,'short')
+    scale_t2f=nfft;
+    scale_f2t=nfft;
+else
+    scale_t2f=1;
+    scale_f2t=1/nfft;
 endif
 
 kfft= sprintf('./fftutil_%s',prec);
diff --git a/tools/Makefile b/tools/Makefile
index 15bd87a..b07d4ea 100644
--- a/tools/Makefile
+++ b/tools/Makefile
@@ -1,11 +1,10 @@
 
 ifeq "$(NFFT)" ""
-	NFFT=1024
+ NFFT=1024
 endif
 
-
 ifeq "$(NUMFFTS)" ""
-	NUMFFTS=10000
+ NUMFFTS=10000
 endif
 
 UTILSRC=../kiss_fft.c fftutil.c
@@ -13,7 +12,7 @@ UTILSRC=../kiss_fft.c fftutil.c
 CFLAGS=-Wall -O3 -lm -I.. -ansi -pedantic
 
 ifeq "$(DATATYPE)" ""
-	DATATYPE=float
+  DATATYPE=float
 endif
 UTIL=fftutil_$(DATATYPE)
 
@@ -21,9 +20,9 @@ UTIL=fftutil_$(DATATYPE)
 all: $(UTIL)
 
 ifeq  "$(DATATYPE)" "short"
-	TYPEFLAGS=-DFIXED_POINT -Dkiss_fft_scalar=short
+ TYPEFLAGS=-DFIXED_POINT -Dkiss_fft_scalar=short
 else	
-	TYPEFLAGS=-Dkiss_fft_scalar=$(DATATYPE)
+ TYPEFLAGS=-Dkiss_fft_scalar=$(DATATYPE)
 endif
 
 $(UTIL): $(UTILSRC)
@@ -47,7 +46,7 @@ time: all $(RANDDAT)
 POW2=256 512 1024 2048
 POW3=243 729 2187
 mtime: all
-	@for n in $(POW2) $(POW3) ;do \
+	@for n in $(POW3) ;do \
 		export NFFT=$$n;make time; \
 		done
 
@@ -56,7 +55,7 @@ snr: all
 	@echo
 	@echo "### testing SNR for $(NFFT) point FFTs"
 	@echo "#### $(DATATYPE)"
-	@echo "testkiss($(NFFT),'$(DATATYPE)',1,1/$(NFFT));" | octave -q
+	@echo "testkiss( $(NFFT) , '$(DATATYPE)' );" | octave -q
 
 test: snr time