From 089923378d610492f1bad80a090005ecbd2da8a3 Mon Sep 17 00:00:00 2001
From: Mark Borgerding <mark@borgerding.net>
Date: Sat, 28 Nov 2009 04:14:47 +0000
Subject: [PATCH] working sourecforge bug

---
 Makefile         |  2 +-
 kiss_fft.c       | 14 ++++++++------
 test/Makefile    |  4 +---
 test/test_real.c | 42 ++++++++++++++++++++++--------------------
 4 files changed, 32 insertions(+), 30 deletions(-)

diff --git a/Makefile b/Makefile
index b272b40..48f536e 100644
--- a/Makefile
+++ b/Makefile
@@ -1,4 +1,4 @@
-KFVER=1_2_8
+KFVER=1_2_9
 
 DISTDIR=kiss_fft_v$(KFVER)
 TARBALL=kiss_fft_v$(KFVER).tar.gz
diff --git a/kiss_fft.c b/kiss_fft.c
index 1f8a8a2..e0540a6 100644
--- a/kiss_fft.c
+++ b/kiss_fft.c
@@ -225,29 +225,30 @@ static void kf_bfly_generic(
     kiss_fft_cpx t;
     int Norig = st->nfft;
 
-    CHECKBUF(scratchbuf,nscratchbuf,p);
+    kiss_fft_cpx * scratch = (kiss_fft_cpx*)malloc(sizeof(kiss_fft_cpx)*p);
 
     for ( u=0; u<m; ++u ) {
         k=u;
         for ( q1=0 ; q1<p ; ++q1 ) {
-            scratchbuf[q1] = Fout[ k  ];
-            C_FIXDIV(scratchbuf[q1],p);
+            scratch[q1] = Fout[ k  ];
+            C_FIXDIV(scratch[q1],p);
             k += m;
         }
 
         k=u;
         for ( q1=0 ; q1<p ; ++q1 ) {
             int twidx=0;
-            Fout[ k ] = scratchbuf[0];
+            Fout[ k ] = scratch[0];
             for (q=1;q<p;++q ) {
                 twidx += fstride * k;
                 if (twidx>=Norig) twidx-=Norig;
-                C_MUL(t,scratchbuf[q] , twiddles[twidx] );
+                C_MUL(t,scratch[q] , twiddles[twidx] );
                 C_ADDTO( Fout[ k ] ,t);
             }
             k += m;
         }
     }
+    free(scratch);
 }
 
 static
@@ -268,7 +269,8 @@ void kf_work(
 #ifdef _OPENMP
     // use openmp extensions at the 
     // top-level (not recursive)
-    if (fstride==1 && m != 1) {
+    if (fstride==1 && p<=5)
+    {
         int k;
 
         // execute the p different work units in different threads
diff --git a/test/Makefile b/test/Makefile
index 61dfb60..5456247 100644
--- a/test/Makefile
+++ b/test/Makefile
@@ -3,14 +3,12 @@ WARNINGS=-W -Wall -Wstrict-prototypes -Wmissing-prototypes -Waggregate-return \
     -Wcast-align -Wcast-qual -Wnested-externs -Wshadow -Wbad-function-cast \
     -Wwrite-strings
 
-# for x86 pentium+ machines , these flags work well
-#
 CFLAGS=-O3 -I.. -I../tools $(WARNINGS)
 CFLAGS+=-ffast-math -fomit-frame-pointer 
 CFLAGS+=-march=prescott 
 #CFLAGS+= -mtune=native
 # TIP: try adding -openmp or -fopenmp  to enable OPENMP directives and use of multiple cores
-#CFLAGS+= -fopenmp
+CFLAGS+=-fopenmp
 
 
 ifeq "$(NFFT)" ""
diff --git a/test/test_real.c b/test/test_real.c
index c926d88..36a0b08 100644
--- a/test/test_real.c
+++ b/test/test_real.c
@@ -59,38 +59,40 @@ double snr_compare( kiss_fft_cpx * vec1,kiss_fft_cpx * vec2, int n)
     }
     return snr;
 }
-#define NFFT 8*3*5
 
 #ifndef NUMFFTS
 #define NUMFFTS 10000
 #endif
 
 
-int main(void)
+int main(int argc,char ** argv)
 {
+    int nfft = 8*3*5;
     double ts,tfft,trfft;
     int i;
-    kiss_fft_cpx cin[NFFT];
-    kiss_fft_cpx cout[NFFT];
-    kiss_fft_cpx sout[NFFT];
+    if (argc>1)
+        nfft = atoi(argv[1]);
+    kiss_fft_cpx cin[nfft];
+    kiss_fft_cpx cout[nfft];
+    kiss_fft_cpx sout[nfft];
     kiss_fft_cfg  kiss_fft_state;
     kiss_fftr_cfg  kiss_fftr_state;
 
-    kiss_fft_scalar rin[NFFT+2];
-    kiss_fft_scalar rout[NFFT+2];
+    kiss_fft_scalar rin[nfft+2];
+    kiss_fft_scalar rout[nfft+2];
     kiss_fft_scalar zero;
     memset(&zero,0,sizeof(zero) ); // ugly way of setting short,int,float,double, or __m128 to zero
 
     srand(time(0));
 
-    for (i=0;i<NFFT;++i) {
+    for (i=0;i<nfft;++i) {
         rin[i] = rand_scalar();
         cin[i].r = rin[i];
         cin[i].i = zero;
     }
 
-    kiss_fft_state = kiss_fft_alloc(NFFT,0,0,0);
-    kiss_fftr_state = kiss_fftr_alloc(NFFT,0,0,0);
+    kiss_fft_state = kiss_fft_alloc(nfft,0,0,0);
+    kiss_fftr_state = kiss_fftr_alloc(nfft,0,0,0);
     kiss_fft(kiss_fft_state,cin,cout);
     kiss_fftr(kiss_fftr_state,rin,sout);
     /*
@@ -105,7 +107,7 @@ int main(void)
     */
         
     printf( "nfft=%d, inverse=%d, snr=%g\n",
-            NFFT,0, snr_compare(cout,sout,(NFFT/2)+1) );
+            nfft,0, snr_compare(cout,sout,(nfft/2)+1) );
     ts = cputime();
     for (i=0;i<NUMFFTS;++i) {
         kiss_fft(kiss_fft_state,cin,cout);
@@ -124,12 +126,12 @@ int main(void)
     free(kiss_fft_state);
     free(kiss_fftr_state);
 
-    kiss_fft_state = kiss_fft_alloc(NFFT,1,0,0);
-    kiss_fftr_state = kiss_fftr_alloc(NFFT,1,0,0);
+    kiss_fft_state = kiss_fft_alloc(nfft,1,0,0);
+    kiss_fftr_state = kiss_fftr_alloc(nfft,1,0,0);
 
     memset(cin,0,sizeof(cin));
 #if 1
-    for (i=1;i< NFFT/2;++i) {
+    for (i=1;i< nfft/2;++i) {
         //cin[i].r = (kiss_fft_scalar)(rand()-RAND_MAX/2);
         cin[i].r = rand_scalar();
         cin[i].i = rand_scalar();
@@ -137,13 +139,13 @@ int main(void)
 #else
     cin[0].r = 12000;
     cin[3].r = 12000;
-    cin[NFFT/2].r = 12000;
+    cin[nfft/2].r = 12000;
 #endif
 
     // conjugate symmetry of real signal 
-    for (i=1;i< NFFT/2;++i) {
-        cin[NFFT-i].r = cin[i].r;
-        cin[NFFT-i].i = - cin[i].i;
+    for (i=1;i< nfft/2;++i) {
+        cin[nfft-i].r = cin[i].r;
+        cin[nfft-i].i = - cin[i].i;
     }
 
     kiss_fft(kiss_fft_state,cin,cout);
@@ -156,13 +158,13 @@ int main(void)
     printf(" results from inverse kiss_fftr: %f,%f,%f,%f,%f ... \n"
             ,(float)rout[0] ,(float)rout[1] ,(float)rout[2] ,(float)rout[3] ,(float)rout[4]);
 */
-    for (i=0;i<NFFT;++i) {
+    for (i=0;i<nfft;++i) {
         sout[i].r = rout[i];
         sout[i].i = zero;
     }
 
     printf( "nfft=%d, inverse=%d, snr=%g\n",
-            NFFT,1, snr_compare(cout,sout,NFFT/2) );
+            nfft,1, snr_compare(cout,sout,nfft/2) );
     free(kiss_fft_state);
     free(kiss_fftr_state);