| Different FFT times for same array size ? [message #45285] |
Tue, 23 August 2005 22:55 |
jcami
Messages: 4
Registered: May 2005
|
Junior Member |
|
|
Hi,
First of all :
IDL> print, !version
{ x86 linux unix linux 6.2 Jun 20 2005 32 64}
I am using the blk_con program that comes with the IDL distribution.
Basically, this uses fft's to do a convolution. I have a large number
of arrays that all have the same number of elements, and I convolve
them all with the same kernel using the blk_con program. I would think
that the execution time should then be roughly the same, but apparently
it is not. To get some more information on the timing, I added a short
timing command in the while loop so that that part now reads :
while loop le k-1L do begin
input = [input[blen-p+1:blen-1], signal[lower:upper]]
tt=systime(1)
output = fft((z_filter * fft(input, -1) ),1) * blen
print, "blk_con ", loop, k-1L, systime(1)-tt
;Next output segment.
result[ip] = double ?
DOUBLE(output[p-1:blen-1]):FLOAT(output[p-1:blen-1])
ip = ip + blen - p + 1L
lower = upper + 1L
upper = lower + blen - p
loop = loop + 1L
endwhile
Since the sizes of the chunks that go into the fft are always the same,
and always a power of 2, the execution time should always be (roughly)
the same, right ??
Now here's an example of what the print statement says for one of my
arrays :
blk_con 2 150 0.034678936
blk_con 3 150 0.036569834
blk_con 4 150 0.033971071
blk_con 5 150 0.036265135
blk_con 6 150 0.039370060
blk_con 7 150 0.049736023
blk_con 8 150 0.039547920
blk_con 9 150 0.054105043
blk_con 10 150 0.036910057
blk_con 11 150 0.046661854
blk_con 12 150 0.13956809
blk_con 13 150 0.046625137
blk_con 14 150 0.058930874
blk_con 15 150 0.036964893
blk_con 16 150 0.045490980
blk_con 17 150 0.063371181
blk_con 18 150 0.037928820
blk_con 19 150 0.048100948
blk_con 20 150 0.032199144
blk_con 21 150 0.042580128
blk_con 22 150 0.034416914
blk_con 23 150 0.036769152
blk_con 24 150 0.033674002
blk_con 25 150 0.045473099
blk_con 26 150 0.035262823
blk_con 27 150 0.039651871
blk_con 28 150 0.037656784
blk_con 29 150 0.040482998
blk_con 30 150 2.7050159
blk_con 31 150 2.8096600
blk_con 32 150 2.8986051
blk_con 33 150 2.7503240
blk_con 34 150 2.8176029
...
blk_con 73 150 2.6031480
blk_con 74 150 0.032071114
blk_con 75 150 0.029819965
blk_con 76 150 2.5321939
...
blk_con 150 150 2.5533669
So the first 28 segments go fast, and from then on things slow down by
a factor of 50 (well, except those 2 funny ones) !!
Am I missing something here ???
Note : I get the same sort of troubles if I use the astrolib's convolve
function, which does pretty much the same.
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
