From 1922ba0d4e5753d9c7878cc5550cb61e6e55dcf8 Mon Sep 17 00:00:00 2001
From: Mark Borgerding <mark@borgerding.net>
Date: Tue, 14 Nov 2006 19:40:58 +0000
Subject: [PATCH] update benchmark tool to allow multi-dimensional and/or real
 FFTs

---
 test/Makefile       |   2 +-
 test/benchkiss.c    | 127 ++++++++++++++++++++++++++++++--------------
 tools/kiss_fftndr.c |  28 ++++------
 3 files changed, 98 insertions(+), 59 deletions(-)

diff --git a/test/Makefile b/test/Makefile
index ffebccb..da2dd59 100644
--- a/test/Makefile
+++ b/test/Makefile
@@ -48,7 +48,7 @@ endif
 
 FFTWLIBDIR=-L/usr/local/lib/
 
-SRCFILES=../kiss_fft.c ../tools/kiss_fftnd.c ../tools/kiss_fftr.c pstats.c ../tools/kfc.c
+SRCFILES=../kiss_fft.c ../tools/kiss_fftnd.c ../tools/kiss_fftr.c pstats.c ../tools/kfc.c ../tools/kiss_fftndr.c
 
 all: tools $(BENCHKISS) $(SELFTEST) $(BENCHFFTW) $(TESTREAL) $(TESTKFC)
 
diff --git a/test/benchkiss.c b/test/benchkiss.c
index 378eef5..e7b9cfe 100644
--- a/test/benchkiss.c
+++ b/test/benchkiss.c
@@ -3,72 +3,117 @@
 #include <sys/times.h>
 #include <unistd.h>
 #include "kiss_fft.h"
+#include "kiss_fftr.h"
+#include "kiss_fftnd.h"
+#include "kiss_fftndr.h"
 
 #include "pstats.h"
 
+static
+int getdims(int * dims, char * arg)
+{
+    char *s;
+    int ndims=0;
+    while ( (s=strtok( arg , ",") ) ) {
+        dims[ndims++] = atoi(s);
+        //printf("%s=%d\n",s,dims[ndims-1]);
+        arg=NULL;
+    }
+    return ndims;
+}
 
 int main(int argc,char ** argv)
 {
-    int nfft=1024;
+    int k;
+    int nfft[32];
+    int ndims = 1;
     int isinverse=0;
     int numffts=1000,i;
     kiss_fft_cpx * buf;
     kiss_fft_cpx * bufout;
-    kiss_fft_cfg st;
+    int real = 0;
+
+    nfft[0] = 1024;// default
 
     while (1) {
-      int c = getopt (argc, argv, "n:ix:");
-      if (c == -1)
-        break;
-      switch (c) {
-      case 'n':
-        nfft = atoi (optarg);
-        if (nfft != kiss_fft_next_fast_size(nfft) ) {
-            int ng = kiss_fft_next_fast_size(nfft);
-            fprintf(stderr,"warning: %d might be a better choice for speed than %d\n",ng,nfft);
+        int c = getopt (argc, argv, "n:ix:r");
+        if (c == -1)
+            break;
+        switch (c) {
+            case 'r':
+                real = 1;
+                break;
+            case 'n':
+                ndims = getdims(nfft, optarg );
+                if (nfft[0] != kiss_fft_next_fast_size(nfft[0]) ) {
+                    int ng = kiss_fft_next_fast_size(nfft[0]);
+                    fprintf(stderr,"warning: %d might be a better choice for speed than %d\n",ng,nfft[0]);
+                }
+                break;
+            case 'x':
+                numffts = atoi (optarg);
+                break;
+            case 'i':
+                isinverse = 1;
+                break;
         }
-        break;
-      case 'x':
-        numffts = atoi (optarg);
-        break;
-      case 'i':
-        isinverse = 1;
-        break;
-      }
     }
-#ifdef USE_SIMD        
-    buf=(kiss_fft_cpx*)memalign(sizeof(kiss_fft_cpx),sizeof(kiss_fft_cpx) * nfft);
-    bufout=(kiss_fft_cpx*)memalign(sizeof(kiss_fft_cpx),sizeof(kiss_fft_cpx) * nfft);
+    int nbytes = sizeof(kiss_fft_cpx);
+    for (k=0;k<ndims;++k)
+        nbytes *= nfft[k];
 
+#ifdef USE_SIMD        
     numffts /= 4;
+    nbytes /= 4;
     fprintf(stderr,"since SIMD implementation does 4 ffts at a time, numffts is being reduced to %d\n",numffts);
-#else
-    buf=(kiss_fft_cpx*)malloc(sizeof(kiss_fft_cpx) * nfft);
-    bufout=(kiss_fft_cpx*)malloc(sizeof(kiss_fft_cpx) * nfft);
 #endif
 
-    for (i=0;i<nfft;++i ) {
-#ifdef USE_SIMD        
-        buf[i].r = _mm_set_ps1((float)( rand() - RAND_MAX/2));
-        buf[i].i = _mm_set_ps1((float)( rand() - RAND_MAX/2));
-#else        
-        buf[i].r = rand() - RAND_MAX/2;
-        buf[i].i = rand() - RAND_MAX/2;
-#endif        
-    }
+    buf=(kiss_fft_cpx*)KISS_FFT_MALLOC(nbytes);
+    bufout=(kiss_fft_cpx*)KISS_FFT_MALLOC(nbytes);
+    memset(buf,0,nbytes);
 
     pstats_init();
 
-    st = kiss_fft_alloc( nfft ,isinverse ,0,0);
-
-    for (i=0;i<numffts;++i)
-        kiss_fft( st ,buf,bufout );
-
-    free(st);
+    if (ndims==1) {
+        if (real) {
+            kiss_fftr_cfg st = kiss_fftr_alloc( nfft[0] ,isinverse ,0,0);
+            if (isinverse)
+                for (i=0;i<numffts;++i)
+                    kiss_fftri( st ,(kiss_fft_cpx*)buf,(kiss_fft_scalar*)bufout );
+            else
+                for (i=0;i<numffts;++i)
+                    kiss_fftr( st ,(kiss_fft_scalar*)buf,(kiss_fft_cpx*)bufout );
+            free(st);
+        }else{
+            kiss_fft_cfg st = kiss_fft_alloc( nfft[0] ,isinverse ,0,0);
+            for (i=0;i<numffts;++i)
+                kiss_fft( st ,buf,bufout );
+            free(st);
+        }
+    }else{
+        if (real) {
+            kiss_fftndr_cfg st = kiss_fftndr_alloc( nfft,ndims ,isinverse ,0,0);
+            if (isinverse)
+                for (i=0;i<numffts;++i)
+                    kiss_fftndri( st ,(kiss_fft_cpx*)buf,(kiss_fft_scalar*)bufout );
+            else
+                for (i=0;i<numffts;++i)
+                    kiss_fftndr( st ,(kiss_fft_scalar*)buf,(kiss_fft_cpx*)bufout );
+            free(st);
+        }else{
+            kiss_fftnd_cfg st= kiss_fftnd_alloc(nfft,ndims,isinverse ,0,0);
+            for (i=0;i<numffts;++i)
+                kiss_fftnd( st ,buf,bufout );
+            free(st);
+        }
+    }
 
     free(buf); free(bufout);
 
-    fprintf(stderr,"KISS\tnfft=%d\tnumffts=%d\n" ,nfft,numffts);
+    fprintf(stderr,"KISS\tnfft=");
+    for (k=0;k<ndims;++k)
+        fprintf(stderr, "%d,",nfft[k]);
+    fprintf(stderr,"\tnumffts=%d\n" ,numffts);
     pstats_report();
 
     kiss_fft_cleanup();
diff --git a/tools/kiss_fftndr.c b/tools/kiss_fftndr.c
index 10fc643..ab0f13c 100644
--- a/tools/kiss_fftndr.c
+++ b/tools/kiss_fftndr.c
@@ -73,27 +73,24 @@ void kiss_fftndr(kiss_fftndr_cfg st,const kiss_fft_scalar *timedata,kiss_fft_cpx
     int k1,k2;
     int dimReal = st->dimReal;
     int dimOther = st->dimOther;
+    int nrbins = dimReal/2+1;
 
     kiss_fft_cpx * tmp1 = (kiss_fft_cpx*)st->tmpbuf; 
-    kiss_fft_cpx * tmp2;
-    if (dimReal/2+1 > dimOther)
-        tmp2 = tmp1+dimReal/2+1;
-    else
-        tmp2 = tmp1+dimOther;
+    kiss_fft_cpx * tmp2 = tmp1 + MAX(nrbins,dimOther);
 
     // timedata is N0 x N1 x ... x Nk real
 
     // take a real chunk of data, fft it and place the output at correct intervals
     for (k1=0;k1<dimOther;++k1) {
-        kiss_fftr( st->cfg_r, timedata + k1*dimReal , tmp1 ); // tmp1 now holds dimReal/2+1 complex points
-        for (k2=0;k2<dimReal/2+1;++k2)
+        kiss_fftr( st->cfg_r, timedata + k1*dimReal , tmp1 ); // tmp1 now holds nrbins complex points
+        for (k2=0;k2<nrbins;++k2)
            tmp2[ k2*dimOther+k1 ] = tmp1[k2];
     }
 
-    for (k2=0;k2<dimReal/2+1;++k2) {
+    for (k2=0;k2<nrbins;++k2) {
         kiss_fftnd(st->cfg_nd, tmp2+k2*dimOther, tmp1);  // tmp1 now holds dimOther complex points
         for (k1=0;k1<dimOther;++k1) 
-            freqdata[ k1*(dimReal/2+1) + k2] = tmp1[k1];
+            freqdata[ k1*(nrbins) + k2] = tmp1[k1];
     }
 }
 
@@ -102,21 +99,18 @@ void kiss_fftndri(kiss_fftndr_cfg st,const kiss_fft_cpx *freqdata,kiss_fft_scala
     int k1,k2;
     int dimReal = st->dimReal;
     int dimOther = st->dimOther;
+    int nrbins = dimReal/2+1;
     kiss_fft_cpx * tmp1 = (kiss_fft_cpx*)st->tmpbuf; 
-    kiss_fft_cpx * tmp2;
-    if (dimReal/2+1 > dimOther)
-        tmp2 = tmp1+dimReal/2+1;
-    else
-        tmp2 = tmp1+dimOther;
+    kiss_fft_cpx * tmp2 = tmp1 + MAX(nrbins,dimOther);
 
-    for (k2=0;k2<dimReal/2+1;++k2) {
+    for (k2=0;k2<nrbins;++k2) {
         for (k1=0;k1<dimOther;++k1) 
-            tmp1[k1] = freqdata[ k1*(dimReal/2+1) + k2 ];
+            tmp1[k1] = freqdata[ k1*(nrbins) + k2 ];
         kiss_fftnd(st->cfg_nd, tmp1, tmp2+k2*dimOther);
     }
 
     for (k1=0;k1<dimOther;++k1) {
-        for (k2=0;k2<dimReal/2+1;++k2)
+        for (k2=0;k2<nrbins;++k2)
             tmp1[k2] = tmp2[ k2*dimOther+k1 ];
         kiss_fftri( st->cfg_r,tmp1,timedata + k1*dimReal);
     }