diff --git a/tools/fftutil.c b/tools/fftutil.c
index bf050e0..60fd497 100644
--- a/tools/fftutil.c
+++ b/tools/fftutil.c
@@ -45,22 +45,19 @@ void fft_filend(FILE * fin,FILE * fout,int *dims,int ndims,int isinverse)
 {
     void *st;
     kiss_fft_cpx *buf;
-    kiss_fft_cpx *bufout;
     int dimprod=1,i;
     for (i=0;i<ndims;++i) 
         dimprod *= dims[i];
 
     buf = (kiss_fft_cpx *) malloc (sizeof (kiss_fft_cpx) * dimprod);
-    bufout = (kiss_fft_cpx *) malloc (sizeof (kiss_fft_cpx) * dimprod);
     st = kiss_fftnd_alloc (dims, ndims, isinverse, 0, 0);
 
     while (fread (buf, sizeof (kiss_fft_cpx) * dimprod, 1, fin) > 0) {
-        kiss_fftnd (st, buf, bufout);
-        fwrite (bufout, sizeof (kiss_fft_cpx), dimprod, fout);
+        kiss_fftnd (st, buf, buf);
+        fwrite (buf, sizeof (kiss_fft_cpx), dimprod, fout);
     }
     free (st);
     free (buf);
-    free (bufout);
 }
 
 void fft_file_real(FILE * fin,FILE * fout,int nfft,int isinverse)
@@ -99,7 +96,7 @@ int get_dims(char * arg,int * dims)
         if (p0)
             *p0++ = '\0';
         dims[ndims++] = atoi(arg);
-        fprintf(stderr,"dims[%d] = %d\n",ndims-1,dims[ndims-1]);
+        /* fprintf(stderr,"dims[%d] = %d\n",ndims-1,dims[ndims-1]); */
         arg = p0;
     }while (p0);
     return ndims;
diff --git a/tools/kiss_fftnd.c b/tools/kiss_fftnd.c
index 798cd3e..e95290a 100644
--- a/tools/kiss_fftnd.c
+++ b/tools/kiss_fftnd.c
@@ -146,15 +146,19 @@ void kiss_fftnd(void * cfg,const kiss_fft_cpx *fin,kiss_fft_cpx *fout)
     kiss_fft_cpx * bufout;
 
     /*arrange it so the last bufout == fout*/
-    if ( st->ndims & 1 )
+    if ( st->ndims & 1 ) {
         bufout = fout;
-    else
+        if (fin==fout) {
+            memcpy( st->tmpbuf, fin, sizeof(kiss_fft_cpx) * st->dimprod );
+            bufin = st->tmpbuf;
+        }
+    }else
         bufout = st->tmpbuf;
 
     for ( k=0; k < st->ndims; ++k) {
         int curdim = st->dims[k];
         int stride = st->dimprod / curdim;
-        
+
         for ( i=0 ; i<stride ; ++i ) 
             kiss_fft_stride( st->states[k], bufin+i , bufout+i*curdim, stride );