diff --git a/_kiss_fft_guts.h b/_kiss_fft_guts.h
index df61c2d..b17c159 100644
--- a/_kiss_fft_guts.h
+++ b/_kiss_fft_guts.h
@@ -88,6 +88,7 @@ void kf_work(
         kiss_fft_cpx * Fout,
         const kiss_fft_cpx * f,
         int fstride,
+        int in_skip,
         int * factors,
         const kiss_fft_state * st
         );
diff --git a/kiss_fft.c b/kiss_fft.c
index 121c872..9894e73 100644
--- a/kiss_fft.c
+++ b/kiss_fft.c
@@ -244,6 +244,7 @@ void kf_work(
         kiss_fft_cpx * Fout,
         const kiss_fft_cpx * f,
         int fstride,
+        int in_stride,
         int * factors,
         const kiss_fft_state * st
         )
@@ -256,8 +257,8 @@ void kf_work(
         if (m==1)
             Fout[q] = *f;
         else
-            kf_work( Fout + m*q, f, fstride*p,factors,st);
-        f += fstride;
+            kf_work( Fout + m*q, f, fstride*p, in_stride, factors,st);
+        f += fstride*in_stride;
     }
 
     switch (p) {
@@ -341,7 +342,7 @@ void * kiss_fft_alloc(int nfft,int inverse_fft,void * mem,size_t * lenmem )
     return st;
 }
 
-void kiss_fft(const void * cfg,const kiss_fft_cpx *fin,kiss_fft_cpx *fout)
+void kiss_fft_stride(const void * cfg,const kiss_fft_cpx *fin,kiss_fft_cpx *fout,int in_stride)
 {
     const kiss_fft_state * st = cfg;
     if (st->nfft < 0) {
@@ -350,9 +351,43 @@ void kiss_fft(const void * cfg,const kiss_fft_cpx *fin,kiss_fft_cpx *fout)
     }
 
     if (fin == fout) {
-        memcpy(st->tmpbuf,fin,sizeof(kiss_fft_cpx)*st->nfft);
-        fin = st->tmpbuf;
+        kf_work(st->tmpbuf,fin,1,in_stride, st->factors,st);
+        memcpy(fout,st->tmpbuf,sizeof(kiss_fft_cpx)*st->nfft);
+    }else{
+        kf_work( fout, fin, 1,in_stride, st->factors,st );
+    }
+}
+
+void kiss_fft(const void * cfg,const kiss_fft_cpx *fin,kiss_fft_cpx *fout)
+{
+    kiss_fft_stride(cfg,fin,fout,1);
+}
+
+void test_stride()
+{
+#define SKIP_FACTOR 7
+#define FFT_SIZE 1800
+    void *cfg;
+    kiss_fft_cpx buf1in[FFT_SIZE],buf1out[FFT_SIZE];
+    kiss_fft_cpx buf2in[SKIP_FACTOR*FFT_SIZE],buf2out[FFT_SIZE];
+    int i;
+    memset(buf2in,0,sizeof(buf2in));
+    for (i=0;i<FFT_SIZE;++i) {
+        buf1in[i].r = rand();
+        buf1in[i].i = rand();
+        buf2in[SKIP_FACTOR*i] = buf1in[i];
+    }
+    cfg= kiss_fft_alloc(FFT_SIZE,0,0,0);
+
+    kiss_fft(cfg,buf1in,buf1out);
+    kiss_fft_stride(cfg,buf2in,buf2out,SKIP_FACTOR);
+    if ( 0==memcmp(buf1out,buf2out,sizeof(buf1out) ) )
+        printf("kiss_fft_stride is working for stride = %d\n",SKIP_FACTOR);
+    else{
+        printf("kiss_fft_stride not working for stride =%d\n",SKIP_FACTOR);
+        for (i=0;i<FFT_SIZE;++i) {
+            fprintf(stderr,"good[%d]=",i);pcpx(buf1out+i);
+            fprintf(stderr,"bad [%d]=",i);pcpx(buf2out+i);
+        }
     }
-
-    kf_work( fout, fin, 1, st->factors,st );
 }
diff --git a/kiss_fft.h b/kiss_fft.h
index 266e944..e0444ec 100644
--- a/kiss_fft.h
+++ b/kiss_fft.h
@@ -62,6 +62,7 @@ void* kiss_fft_alloc(int nfft,int inverse_fft,void * mem,size_t * lenmem);
  * */
 void kiss_fft(const void * cfg,const kiss_fft_cpx *fin,kiss_fft_cpx *fout);
 
+void kiss_fft_stride(const void * cfg,const kiss_fft_cpx *fin,kiss_fft_cpx *fout,int fin_stride);
 
 /* If kiss_fft_alloc allocated a buffer, it is one contiguous 
    buffer and can be simply free()d when no longer needed*/
diff --git a/test/Makefile b/test/Makefile
index aaa1869..b429f3d 100644
--- a/test/Makefile
+++ b/test/Makefile
@@ -34,9 +34,10 @@ endif
 
 all: $(BENCHKISS) $(SELFTEST) $(BENCHFFTW) $(TESTREAL) $(FFTUTIL) $(TESTKFC)
 
-#CFLAGS=-Wall -O3 -ansi -pedantic -march=pentiumpro -ffast-math -fomit-frame-pointer
+CFLAGS=-Wall -O3 -pedantic -march=pentiumpro -ffast-math -fomit-frame-pointer 
+#-DUSE_SKIP 
 # If the above flags do not work, try the following
-CFLAGS=-Wall -O3
+#CFLAGS=-Wall -O3
 
 $(FFTUTIL): ../kiss_fft.c fftutil.c kiss_fft2d.c kiss_fftr.c
 	$(CC) -o $@ $(CFLAGS) -I.. $(TYPEFLAGS) -lm $+
diff --git a/test/benchkiss.c b/test/benchkiss.c
index 6222018..7d08b71 100644
--- a/test/benchkiss.c
+++ b/test/benchkiss.c
@@ -1,6 +1,5 @@
 #include <stdio.h>
 #include <stdlib.h>
-#include <unistd.h>
 #include <sys/times.h>
 #include <unistd.h>
 #include "kiss_fft.h"
diff --git a/test/selftest.c b/test/selftest.c
index 513e46e..05369fb 100644
--- a/test/selftest.c
+++ b/test/selftest.c
@@ -26,6 +26,8 @@ double snr_compare( kiss_fft_cpx * test_vec_out,kiss_fft_cpx * testbuf, int n)
     return snr;
 }
 
+void test_stride();
+
 int main() { int exit_code=0;
 
 #define NFFT 10
@@ -130,7 +132,7 @@ int main() { int exit_code=0;
     }
 #undef NFFT    
 
-
+    test_stride();
     return exit_code;
 }
 
diff --git a/tools/Makefile b/tools/Makefile
index aaa1869..b429f3d 100644
--- a/tools/Makefile
+++ b/tools/Makefile
@@ -34,9 +34,10 @@ endif
 
 all: $(BENCHKISS) $(SELFTEST) $(BENCHFFTW) $(TESTREAL) $(FFTUTIL) $(TESTKFC)
 
-#CFLAGS=-Wall -O3 -ansi -pedantic -march=pentiumpro -ffast-math -fomit-frame-pointer
+CFLAGS=-Wall -O3 -pedantic -march=pentiumpro -ffast-math -fomit-frame-pointer 
+#-DUSE_SKIP 
 # If the above flags do not work, try the following
-CFLAGS=-Wall -O3
+#CFLAGS=-Wall -O3
 
 $(FFTUTIL): ../kiss_fft.c fftutil.c kiss_fft2d.c kiss_fftr.c
 	$(CC) -o $@ $(CFLAGS) -I.. $(TYPEFLAGS) -lm $+
diff --git a/tools/fftutil.c b/tools/fftutil.c
index beb1af5..78d42bf 100644
--- a/tools/fftutil.c
+++ b/tools/fftutil.c
@@ -103,6 +103,14 @@ int main(int argc,char ** argv)
             case 'r':nrows = atoi(optarg);break;
             case 'i':isinverse=1;break;
             case 'R':isreal=1;break;
+            case '?':
+                     fprintf(stderr,"usage options:\n"
+                            "\t-n NFFT: fft size\n"
+                            "\t-r nrows: # rows (implies 2d FFT)\n"
+                            "\t-i : inverse\n"
+                            "\t-R : real input samples, not complex\n");
+                     exit (1);
+            default:fprintf(stderr,"bad %c\n",c);break;
         }
     }