radix 5 works, but is 6x slower than fftw

This commit is contained in:
Mark Borgerding 2003-11-03 03:03:16 +00:00
parent 85764e6437
commit 3c0c0431e2
5 changed files with 51 additions and 152 deletions

View File

@ -278,10 +278,11 @@ void bfly5(
Fout4->i = scratch[0].i + scratch[5].i + scratch[6].i + scratch[7].i + scratch[8].i;
++Fout0;++Fout1;++Fout2;++Fout3;++Fout4;
tw1+=m;
tw2+=2*m;
tw3+=3*m;
tw4+=4*m;
tw1+=fstride;
tw2+=2*fstride;
tw3+=3*fstride;
tw4+=4*fstride;
}
}
@ -347,9 +348,7 @@ void fft_work(
case 2: bfly2(Fout,fstride,st,m); break;
case 3: bfly3(Fout,fstride,st,m); break;
case 4: bfly4(Fout,fstride,st,m); break;
#if 1
case 5: bfly5(Fout,fstride,st,m); break;
#endif
default: bfly_generic(Fout,fstride,st,m,p); break;
}
}

View File

@ -1,23 +1,17 @@
ifeq "$(NFFT)" ""
NFFT=1024
NFFT=1800
endif
ifeq "$(NUMFFTS)" ""
NUMFFTS=10000
endif
UTILSRC=../kiss_fft.c fftutil.c
CFLAGS=-Wall -O3 -lm -I.. -ansi -pedantic
ifeq "$(DATATYPE)" ""
DATATYPE=float
DATATYPE=float
endif
UTIL=fftutil_$(DATATYPE)
BENCH=bm_$(DATATYPE)
all: $(UTIL)
all: $(UTIL) $(BENCH)
ifeq "$(DATATYPE)" "short"
TYPEFLAGS=-DFIXED_POINT -Dkiss_fft_scalar=short
@ -25,31 +19,29 @@ else
TYPEFLAGS=-Dkiss_fft_scalar=$(DATATYPE)
endif
$(UTIL): $(UTILSRC)
gcc -o $@ $(CFLAGS) $(TYPEFLAGS) $(UTILSRC)
CFLAGS=-Wall -O3 -ansi -pedantic
$(UTIL): ../kiss_fft.c fftutil.c
gcc -o $@ $(CFLAGS) -I.. $(TYPEFLAGS) ../kiss_fft.c fftutil.c -lm
RANDDAT=rand_$(DATATYPE)_$(NFFT).dat
$(RANDDAT):
./rand_fft_data.py -n 1 -N $(NFFT) -t $(DATATYPE) > $(RANDDAT)
$(BENCH): benchkiss.c ../kiss_fft.c
gcc -o $@ $(CFLAGS) -I.. benchkiss.c $(TYPEFLAGS) ../kiss_fft.c -lm
time: all $(RANDDAT)
@echo
@echo -n "#### timing $(NUMFFTS) x $(NFFT) point FFTs. "; factor $(NFFT)
@if [ -x ~/fftw/st ] && [ $(DATATYPE) == "double" ]; then \
echo "#### FFTW FFT $(DATATYPE)"; \
time ~/fftw/st -x $(NUMFFTS) -n $(NFFT) < $(RANDDAT) > /dev/null;\
fi
@echo "#### KISS FFT $(DATATYPE)"
@time ./$(UTIL) -x $(NUMFFTS) -n $(NFFT) < $(RANDDAT) > /dev/null
@rm $(RANDDAT)
fftw: bm_fftw
./bm_fftw -x $(NUMFFTS) -n $(NFFT)
bm_fftw: benchfftw.c
gcc -o $@ $(CFLAGS) benchfftw.c -lm -lfftw3 -L /usr/local/lib/
time: all
./$(BENCH) -x $(NUMFFTS) -n $(NFFT)
POW2=256 512 1024 2048
POW3=243 729 2187
POW5=25 125 625
POW5=25 125 625
mtime: all
@for n in $(POW5) ;do \
export NFFT=$$n;make time; \
done
@for n in $(POW2) $(POW3) $(POW5) ;do \
./$(BENCH) -x $(NUMFFTS) -n $$n;\
done
snr: all
@which octave
@ -61,5 +53,4 @@ snr: all
test: snr time
clean:
rm -f *.dat *~ fftutil_*
rm -f *~ fftutil_* bm_*

View File

@ -1,26 +0,0 @@
#!/usr/local/bin/python2.3
import sys
import random
import struct
def main():
from getopt import getopt
opts,args = getopt(sys.argv[1:],'n:N:t:l:h:')
opts=dict(opts)
nbufs = int( opts.get('-n','10000') )
nfft = int( opts.get('-N','1024') )
type = opts.get('-t','f')
lo = float(opts.get('-l','-32768') )
hi = float(opts.get('-h','32767') )
format = {'float':'f','short':'h','double':'d'}[type]
nums = [ random.uniform(lo,hi) for i in range(nfft*2) ]
buf = struct.pack( '%d%s' % ( len( nums ) , format ) , *nums )
for i in range(nbufs):
sys.stdout.write( buf )
if __name__ == '__main__':
main()

View File

@ -1,56 +0,0 @@
#include <stdio.h>
#include <stdlib.h>
#include <fftw3.h>
#include <getopt.h>
int main(int argc,char ** argv)
{
int nfft=1024;
int isinverse=0;
FILE *fin=stdin;
FILE *fout=stdout;
int times=1,i;
fftw_complex * in=NULL;
fftw_complex * out=NULL;
fftw_plan p;
while (1) {
int c = getopt (argc, argv, "n:ix:");
if (c == -1)
break;
switch (c) {
case 'n':
nfft = atoi (optarg);
break;
case 'x':
times = atoi (optarg);
break;
case 'i':
isinverse = 1;
break;
}
}
//fprintf(stderr,"sizeof(fftw_complex) = %d \n" , sizeof(fftw_complex) );
fprintf(stderr,"sizeof(fftw_complex[0]) = %d \n" , sizeof((*in)[0]) );
in=fftw_malloc(sizeof(fftw_complex) * nfft);
out=fftw_malloc(sizeof(fftw_complex) * nfft);
if ( isinverse )
p = fftw_plan_dft_1d(nfft, in, out, FFTW_BACKWARD, FFTW_ESTIMATE);
else
p = fftw_plan_dft_1d(nfft, in, out, FFTW_FORWARD, FFTW_ESTIMATE);
while ( fread( in , sizeof(fftw_complex) , nfft , fin ) > 0 ) {
for (i=0;i<times;++i)
fftw_execute(p);
fwrite( out , sizeof(fftw_complex) , nfft , fout );
}
fftw_destroy_plan(p);
fftw_free(in); fftw_free(out);
return 0;
}

View File

@ -1,23 +1,17 @@
ifeq "$(NFFT)" ""
NFFT=1024
NFFT=1800
endif
ifeq "$(NUMFFTS)" ""
NUMFFTS=10000
endif
UTILSRC=../kiss_fft.c fftutil.c
CFLAGS=-Wall -O3 -lm -I.. -ansi -pedantic
ifeq "$(DATATYPE)" ""
DATATYPE=float
DATATYPE=float
endif
UTIL=fftutil_$(DATATYPE)
BENCH=bm_$(DATATYPE)
all: $(UTIL)
all: $(UTIL) $(BENCH)
ifeq "$(DATATYPE)" "short"
TYPEFLAGS=-DFIXED_POINT -Dkiss_fft_scalar=short
@ -25,31 +19,29 @@ else
TYPEFLAGS=-Dkiss_fft_scalar=$(DATATYPE)
endif
$(UTIL): $(UTILSRC)
gcc -o $@ $(CFLAGS) $(TYPEFLAGS) $(UTILSRC)
CFLAGS=-Wall -O3 -ansi -pedantic
$(UTIL): ../kiss_fft.c fftutil.c
gcc -o $@ $(CFLAGS) -I.. $(TYPEFLAGS) ../kiss_fft.c fftutil.c -lm
RANDDAT=rand_$(DATATYPE)_$(NFFT).dat
$(RANDDAT):
./rand_fft_data.py -n 1 -N $(NFFT) -t $(DATATYPE) > $(RANDDAT)
$(BENCH): benchkiss.c ../kiss_fft.c
gcc -o $@ $(CFLAGS) -I.. benchkiss.c $(TYPEFLAGS) ../kiss_fft.c -lm
time: all $(RANDDAT)
@echo
@echo -n "#### timing $(NUMFFTS) x $(NFFT) point FFTs. "; factor $(NFFT)
@if [ -x ~/fftw/st ] && [ $(DATATYPE) == "double" ]; then \
echo "#### FFTW FFT $(DATATYPE)"; \
time ~/fftw/st -x $(NUMFFTS) -n $(NFFT) < $(RANDDAT) > /dev/null;\
fi
@echo "#### KISS FFT $(DATATYPE)"
@time ./$(UTIL) -x $(NUMFFTS) -n $(NFFT) < $(RANDDAT) > /dev/null
@rm $(RANDDAT)
fftw: bm_fftw
./bm_fftw -x $(NUMFFTS) -n $(NFFT)
bm_fftw: benchfftw.c
gcc -o $@ $(CFLAGS) benchfftw.c -lm -lfftw3 -L /usr/local/lib/
time: all
./$(BENCH) -x $(NUMFFTS) -n $(NFFT)
POW2=256 512 1024 2048
POW3=243 729 2187
POW5=25 125 625
POW5=25 125 625
mtime: all
@for n in $(POW5) ;do \
export NFFT=$$n;make time; \
done
@for n in $(POW2) $(POW3) $(POW5) ;do \
./$(BENCH) -x $(NUMFFTS) -n $$n;\
done
snr: all
@which octave
@ -61,5 +53,4 @@ snr: all
test: snr time
clean:
rm -f *.dat *~ fftutil_*
rm -f *~ fftutil_* bm_*