| Re: Convolution of Stick Spectra [message #21721] |
Sat, 09 September 2000 17:44  |
bjackel
Messages: 10
Registered: September 1993
|
Junior Member |
|
|
Hi Todd
It helps a bit to pre-calculate your variance term.
Cuts execution time from 27 to 18 seconds on my PC.
variance_term= ( ( (.12*sqrt(energy/1000))/1.6651)*1000)^2
FOR indx=0L,nstick-1 DO BEGIN
result = result + intensity[indx]* $
exp(-((energy_scale - energy[indx])^2)/
variance_term[indx] )
ENDFOR
Nothing else springs to mind. Your energy dependent
width is a real problem ie. can't use single FFT for
convolution.
If precision isn't a major issue, you could pre-calculate
Gaussians for each energy bin (2048 of them) and just
multiply by the appropriate intensities. Only useful
if you'll be reading in lots of different spectra in
each run.
It'll be interesting to see what others come up with...
Brian Jackel
Todd Clements wrote:
>
> Hi all...
>
> In my on-going effort to speed up the code in our lab, I have another
> 'challenge' for you (I have to put 'challenge' in quotes because it seems
> that no matter what I ask, someone knows the ansewr off (or at least
> nearly off) the top of their head!)
>
> We have a situation where we need to convolute (with energy dependent
> gaussians) a number of stick spectra on a well-defined energy axis. The
> stick spectra are read in from another program as a 2-dimensional array,
> using ddread. The [0,*] elements are the energies of the sticks, and the
> [1,*] values are the intensities. These have no inherent spacing, they are
> just calculated intensities at whatever energy the calculation returns.
|
|
|
|
| Re: Convolution of Stick Spectra [message #21768 is a reply to message #21721] |
Thu, 14 September 2000 00:00  |
mole6e23
Messages: 31
Registered: December 1998
|
Member |
|
|
bjackel@phys.ucalgary.ca wrote:
> It helps a bit to pre-calculate your variance term.
> Cuts execution time from 27 to 18 seconds on my PC.
>
> variance_term= ( ( (.12*sqrt(energy/1000))/1.6651)*1000)^2
> FOR indx=0L,nstick-1 DO BEGIN
> result = result + intensity[indx]* $
> exp(-((energy_scale - energy[indx])^2)/
> variance_term[indx] )
> ENDFOR
>
Actually, I was surprised to find that didn't change the timing any on my
sytem (alpha OSF unix 5.3 Nov 11 1999) to any appreciable degree. With the
code I first posted (attached below as a procedure), it took about 10.5
seconds either way.
IDL> test
time other: 10.706325
time precalc: 10.545039
Jumping up to 20,000 elements, the timing was still pretty much identical
(42.9 vs. 42.7 seconds).
Todd
----
pro test
convoluted = fltarr( 2, 2048 )
convoluted[0,*] = findgen( 2048 ) / 2047. * 4.
;; Let's fake a stick spectrum, we usually have at least this many elements
stick = abs(randomn( systime(1), 2, 5000 ))
stick[0,*] = stick[0,*] * 4.
stick[1,*] = stick[1,*] * 1000.
time = systime(1)
variance = (((.12*sqrt(stick[0,*]/1000))/1.6651)*1000)^2
for i=0L, n_elements( stick ) / 2 -1 do $
convoluted[1,*] = stick[1,i]*exp(-((convoluted[0,*] - stick[0,i])^2)/ $
variance[i]) $
+ convoluted[1,*]
time2 = systime(1)
for i=0L, n_elements( stick ) / 2 -1 do $
convoluted[1,*] = stick[1,i]*exp(-((convoluted[0,*] - stick[0,i])^2)/ $
(((.12*sqrt(stick[0,i]/1000))/1.6651)*1000)^2) $
+ convoluted[1,*]
print, 'time other: ', systime(1) - time2
print, 'time precalc: ', time2 - time
end
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
