Interpolate whole array instead of looping through elements [message #94325] |
Tue, 11 April 2017 09:18  |
liam.steele
Messages: 13 Registered: September 2015
|
Junior Member |
|
|
Hi all,
I have a 4D array (say temperatures) ordered as temp[nlat,nlon,nlev,ntime], where the elements are the number of latitudes, longitudes, vertical levels and times. What I want to do is interpolate the lat/lon data to a specified point or points, but for each level and time. So, at the moment my code would look like:
out_vals = fltarr(nlev, ntime)
for i = 0, ntime-1 do begin
for j = 0, nlev-1 do begin
out_vals[j,i] = bilinear(temp[*,*,j,i],ival,jval)
endfor
endfor
where ival and jval are the points I want to interpolate to. However, there are lots of levels and times, and so the loop procedure can take a while.
Is there any method where one or both of the loops can be removed, and the interpolation can be carried out on the entire array (i.e. on the time and level dimensions at the same time)? I've tried to think of a quicker method, but I'm stumped.
Many thanks,
Liam
|
|
|
Re: Interpolate whole array instead of looping through elements [message #94326 is a reply to message #94325] |
Tue, 11 April 2017 12:07   |
Craig Markwardt
Messages: 1869 Registered: November 1996
|
Senior Member |
|
|
On Tuesday, April 11, 2017 at 12:18:30 PM UTC-4, liam....@gmx.co.uk wrote:
...
> out_vals[j,i] = bilinear(temp[*,*,j,i],ival,jval)
...
> Is there any method where one or both of the loops can be removed, and the interpolation can be carried out on the entire array (i.e. on the time and level dimensions at the same time)? I've tried to think of a quicker method, but I'm stumped.
As long as TEMP[*,*,j,i] has a lot of elements, you won't really get much of a speed-up removing the FOR loops. Try timing this loop,
for i = 0, ntime-1 do begin
for j = 0, nlev-1 do begin
DUMMY = 1
endfor
endfor
However long this loop takes to run, is the amount of time you will save by removing the loop in your code.
Craig
|
|
|
Re: Interpolate whole array instead of looping through elements [message #94328 is a reply to message #94325] |
Wed, 12 April 2017 16:50   |
liam.steele
Messages: 13 Registered: September 2015
|
Junior Member |
|
|
thanks for the replies everyone,
Well it is a little bit more complicated than I originally said. I made the original question simpler to avoid confusing things!
Basically, temp is of size [280,280,60,720]. What I am actually doing is getting the average temperatures in a radially symmetric crater. Take the following image as an example: https://s22.postimg.org/6r50fbrld/bilinear.jpg
Imagine the black grid is the lat-lon temperature field at a certain level and time (i.e. temp[*,*,0,0]). Once I have this 2D field, I need to calculate how the average temperature varies from the centre of the crater to the edge. So I define a line (shown in red, with the black dots the locations I want values at), and for each black dot I use the bilinear function to get a value. I then rotate the red line a bit more and do the calculation again, and repeat. On each line there are about 100 points.
Once a full circle of rotations is complete, the average temps from the centre to the edge of the crater are found. But only for one time and one level. At the moment I'm rotating the line by 5 degrees. So each time and each level of data has 36 rotations with each rotation having 100 points on the line to use the bilinear function on. So, it's something like:
for iangle = 0, 35 do begin
for ipoint = 0, 99 do begin
; Find ival and jval of the point we want to interpolate to
ival = ....
jval = ....
for itime = 0, 719 do begin
for ilev = 0, 59 do begin
out_vals[ipoint,ilev,itime] = out_vals[ipoint,ilev,itime] + bilinear(temp[*,*,ilev,itime],ival,jval)/36
endfor
endfor
endfor
endfor
And it goes really rather slowly. Looping through just iangle, ipoint and itime takes 131 seconds (using the TIC,TOC functions). This then needs multiplied by 60 to loop through each atmospheric level, so it takes more than two hours in total. And this is just for one lot of data. At the moment I have 50 or so of these to calculate, so that's almost 5 days of IDL calculation!
I was thinking there was maybe something that could be done where the iangle and ipoint loops still occur (as they have to, in order to find the i and j indices for the bilinear interpolation), but then interpolation could occur for all itime and ilev values at once in some speedy IDL vectorized way (since they are using the same indices). But maybe not! Maybe I need to find something other than IDL that might be quicker. Or just accept it is going to take a while to calculate!
Apologies if none of this makes sense!
Liam
|
|
|
Re: Interpolate whole array instead of looping through elements [message #94332 is a reply to message #94328] |
Wed, 12 April 2017 21:10   |
wlandsman
Messages: 743 Registered: June 2000
|
Senior Member |
|
|
I can't help directly as I get dizzy working in more than 3 dimensions ;-)
But it looks like you can parallelize your code. Just today I used the IDL_IDLBridge for the first time to farm out my code to 16 CPUs, and I was able to speed things up by a factor of 10.
You might be able to use the wrapper to the IDL_IDLBridge developed by Mike Galloy
http://michaelgalloy.com/2015/05/26/a-simple-multicore-libra ry-for-idl.html
though it turns out it wasn't appropriate for my problem. I also found the following Web page useful:
http://daedalus.as.arizona.edu/wiki/ParallelIDL
--Wayne
On Wednesday, April 12, 2017 at 7:50:17 PM UTC-4, liam....@gmx.co.uk wrote:
> thanks for the replies everyone,
>
> Well it is a little bit more complicated than I originally said. I made the original question simpler to avoid confusing things!
>
> Basically, temp is of size [280,280,60,720]. What I am actually doing is getting the average temperatures in a radially symmetric crater. Take the following image as an example: https://s22.postimg.org/6r50fbrld/bilinear.jpg
>
> Imagine the black grid is the lat-lon temperature field at a certain level and time (i.e. temp[*,*,0,0]). Once I have this 2D field, I need to calculate how the average temperature varies from the centre of the crater to the edge. So I define a line (shown in red, with the black dots the locations I want values at), and for each black dot I use the bilinear function to get a value. I then rotate the red line a bit more and do the calculation again, and repeat. On each line there are about 100 points.
>
> Once a full circle of rotations is complete, the average temps from the centre to the edge of the crater are found. But only for one time and one level. At the moment I'm rotating the line by 5 degrees. So each time and each level of data has 36 rotations with each rotation having 100 points on the line to use the bilinear function on. So, it's something like:
>
> for iangle = 0, 35 do begin
> for ipoint = 0, 99 do begin
>
> ; Find ival and jval of the point we want to interpolate to
> ival = ....
> jval = ....
>
> for itime = 0, 719 do begin
> for ilev = 0, 59 do begin
> out_vals[ipoint,ilev,itime] = out_vals[ipoint,ilev,itime] + bilinear(temp[*,*,ilev,itime],ival,jval)/36
> endfor
> endfor
>
> endfor
> endfor
>
> And it goes really rather slowly. Looping through just iangle, ipoint and itime takes 131 seconds (using the TIC,TOC functions). This then needs multiplied by 60 to loop through each atmospheric level, so it takes more than two hours in total. And this is just for one lot of data. At the moment I have 50 or so of these to calculate, so that's almost 5 days of IDL calculation!
>
> I was thinking there was maybe something that could be done where the iangle and ipoint loops still occur (as they have to, in order to find the i and j indices for the bilinear interpolation), but then interpolation could occur for all itime and ilev values at once in some speedy IDL vectorized way (since they are using the same indices). But maybe not! Maybe I need to find something other than IDL that might be quicker. Or just accept it is going to take a while to calculate!
>
> Apologies if none of this makes sense!
>
> Liam
|
|
|
Re: Interpolate whole array instead of looping through elements [message #94335 is a reply to message #94328] |
Thu, 13 April 2017 05:21   |
Markus Schmassmann
Messages: 129 Registered: April 2016
|
Senior Member |
|
|
On 04/13/2017 01:50 AM, liam.steele@gmx.co.uk wrote:
> thanks for the replies everyone,
>
> Well it is a little bit more complicated than I originally said. I
> made the original question simpler to avoid confusing things!
>
> Basically, temp is of size [280,280,60,720]. What I am actually
> doing is getting the average temperatures in a radially symmetric crater. Take
> the following image as an example:
> https://s22.postimg.org/6r50fbrld/bilinear.jpg
>
> Imagine the black grid is the lat-lon temperature field at a certain
> level and time (i.e. temp[*,*,0,0]). Once I have this 2D field, I need
> to calculate how the average temperature varies from the centre of the
> crater to the edge. So I define a line (shown in red, with the black
> dots the locations I want values at), and for each black dot I use the
> bilinear function to get a value. I then rotate the red line a bit more
> and do the calculation again, and repeat. On each line there are about
> 100 points.
>
> Once a full circle of rotations is complete, the average temps from
> the centre to the edge of the crater are found. But only for one time
> and one level. At the moment I'm rotating the line by 5 degrees. So each
> time and each level of data has 36 rotations with each rotation having
> 100 points on the line to use the bilinear function on. So, it's
> something like:
>
> for iangle = 0, 35 do begin
> for ipoint = 0, 99 do begin
>
> ; Find ival and jval of the point we want to interpolate to
> ival = ....
> jval = ....
>
> for itime = 0, 719 do begin
> for ilev = 0, 59 do begin
> out_vals[ipoint,ilev,itime] = out_vals[ipoint,ilev,itime] + $
> bilinear(temp[*,*,ilev,itime],ival,jval)/36
> endfor
> endfor
>
> endfor
> endfor
>
> And it goes really rather slowly. Looping through just iangle,
> ipointand itime takes 131 seconds (using the TIC,TOC functions). This then
> needs multiplied by 60 to loop through each atmospheric level, so it
> takes more than two hours in total. And this is just for one lot of
> data. At the moment I have 50 or so of these to calculate, so that's
> almost 5 days of IDL calculation!
>
> I was thinking there was maybe something that could be done where
> theiangle and ipoint loops still occur (as they have to, in order to find
> the i and j indices for the bilinear interpolation), but then
> interpolation could occur for all itime and ilev values at once in some
> speedy IDL vectorized way (since they are using the same indices). But
> maybe not! Maybe I need to find something other than IDL that might be
> quicker. Or just accept it is going to take a while to calculate!
>
> Apologies if none of this makes sense!
ivals=139.5+1.395*rebin(findgen(1,100),[36,100],/sample) $
*rebin(sin(!pi/18*findgen(36,1)),[36,100],/sample)
jvals=139.5+1.395*rebin(findgen(1,100),[36,100],/sample) $
*rebin(cos(!pi/18*findgen(36,1)),[36,100],/sample)
one simple thing for pure increased speed:
for iangle = 0, 35 do for ipoint = 0, 99 do for itime = 0, 719 do $
for ilev = 0, 59 do out_vals[ipoint,ilev,itime] = $
out_vals[ipoint,ilev,itime] + bilinear(temp[*,*,ilev,itime], $
ivals[iangle,ipoint],jval[iangle,ipoint])/36
but better is to vectorize:
ivals2=rebin(ivals,[36,100,720],/sample)
ivals2=rebin(jvals,[36,100,720],/sample)
tvals2=rebin(findgen(1,1,720),[36,100,720],/sample)
out_vals=fltarr(100,720,60)
for ilev=0,59 do out_vals[0,0,ilev]=mean(interpolate( $
reform(temp[*,*,ilev,*]), ivals2, jvals2, tvals2),dim=1 )
out_vals=transpose(out_vals,[0,2,1])
and if that is not fast enough, do the interpolation manually:
wFF=rebin(floor(ivals)+280l*floor(jvals),[36,100,60,720],/sa mple)+ $
rebin(280l^2*lindgen(1,1,60,720),[36,100,60,720],/sample)
wFC=rebin(floor(ivals)+280l* ceil(jvals),[36,100,60,720],/sample)+ $
rebin(280l^2*lindgen(1,1,60,720),[36,100,60,720],/sample)
wCF=rebin( ceil(ivals)+280l*floor(jvals),[36,100,60,720],/sample)+ $
rebin(280l^2*lindgen(1,1,60,720),[36,100,60,720],/sample)
wCC=rebin( ceil(ivals)+280l* ceil(jvals),[36,100,60,720],/sample)+ $
rebin(280l^2*lindgen(1,1,60,720),[36,100,60,720],/sample)
weightFF=rebin( (1+(ivals mod 1))*(1+(jvals mod 1)), $
[36,100,60,720],/sample)
weightFC=rebin(-(1+(ivals mod 1))* (jvals mod 1) , $
[36,100,60,720],/sample)
weightCF=rebin(- (ivals mod 1) *(1+(jvals mod 1)), $
[36,100,60,720],/sample)
weightCC=rebin( (ivals mod 1) * (jvals mod 1) , $
[36,100,60,720],/sample)
out_vals=mean(weightFF*temp[wFF]+weightFC*temp[wFC]+weightCF *temp[wCF]+weightCC*temp[wCC],dim=1)
of which only the last line has to be repeated for every data set.
I haven't debugged anything, so some corrections might be necessary.
Good luck, Markus
|
|
|
Re: Interpolate whole array instead of looping through elements [message #94336 is a reply to message #94335] |
Thu, 13 April 2017 10:02  |
liam.steele
Messages: 13 Registered: September 2015
|
Junior Member |
|
|
On Thursday, 13 April 2017 07:21:50 UTC-5, Markus Schmassmann wrote:
> On 04/13/2017 01:50 AM, liam.steele@gmx.co.uk wrote:
>> thanks for the replies everyone,
>>
>> Well it is a little bit more complicated than I originally said. I
>> made the original question simpler to avoid confusing things!
>>
>> Basically, temp is of size [280,280,60,720]. What I am actually
>> doing is getting the average temperatures in a radially symmetric crater. Take
>> the following image as an example:
>> https://s22.postimg.org/6r50fbrld/bilinear.jpg
>>
>> Imagine the black grid is the lat-lon temperature field at a certain
>> level and time (i.e. temp[*,*,0,0]). Once I have this 2D field, I need
>> to calculate how the average temperature varies from the centre of the
>> crater to the edge. So I define a line (shown in red, with the black
>> dots the locations I want values at), and for each black dot I use the
>> bilinear function to get a value. I then rotate the red line a bit more
>> and do the calculation again, and repeat. On each line there are about
>> 100 points.
>>
>> Once a full circle of rotations is complete, the average temps from
>> the centre to the edge of the crater are found. But only for one time
>> and one level. At the moment I'm rotating the line by 5 degrees. So each
>> time and each level of data has 36 rotations with each rotation having
>> 100 points on the line to use the bilinear function on. So, it's
>> something like:
>>
>> for iangle = 0, 35 do begin
>> for ipoint = 0, 99 do begin
>>
>> ; Find ival and jval of the point we want to interpolate to
>> ival = ....
>> jval = ....
>>
>> for itime = 0, 719 do begin
>> for ilev = 0, 59 do begin
>> out_vals[ipoint,ilev,itime] = out_vals[ipoint,ilev,itime] + $
>> bilinear(temp[*,*,ilev,itime],ival,jval)/36
>> endfor
>> endfor
>>
>> endfor
>> endfor
>>
>> And it goes really rather slowly. Looping through just iangle,
>> ipointand itime takes 131 seconds (using the TIC,TOC functions). This then
>> needs multiplied by 60 to loop through each atmospheric level, so it
>> takes more than two hours in total. And this is just for one lot of
>> data. At the moment I have 50 or so of these to calculate, so that's
>> almost 5 days of IDL calculation!
>>
>> I was thinking there was maybe something that could be done where
>> theiangle and ipoint loops still occur (as they have to, in order to find
>> the i and j indices for the bilinear interpolation), but then
>> interpolation could occur for all itime and ilev values at once in some
>> speedy IDL vectorized way (since they are using the same indices). But
>> maybe not! Maybe I need to find something other than IDL that might be
>> quicker. Or just accept it is going to take a while to calculate!
>>
>> Apologies if none of this makes sense!
>
> ivals=139.5+1.395*rebin(findgen(1,100),[36,100],/sample) $
> *rebin(sin(!pi/18*findgen(36,1)),[36,100],/sample)
> jvals=139.5+1.395*rebin(findgen(1,100),[36,100],/sample) $
> *rebin(cos(!pi/18*findgen(36,1)),[36,100],/sample)
>
> one simple thing for pure increased speed:
>
> for iangle = 0, 35 do for ipoint = 0, 99 do for itime = 0, 719 do $
> for ilev = 0, 59 do out_vals[ipoint,ilev,itime] = $
> out_vals[ipoint,ilev,itime] + bilinear(temp[*,*,ilev,itime], $
> ivals[iangle,ipoint],jval[iangle,ipoint])/36
>
> but better is to vectorize:
>
> ivals2=rebin(ivals,[36,100,720],/sample)
> ivals2=rebin(jvals,[36,100,720],/sample)
> tvals2=rebin(findgen(1,1,720),[36,100,720],/sample)
> out_vals=fltarr(100,720,60)
> for ilev=0,59 do out_vals[0,0,ilev]=mean(interpolate( $
> reform(temp[*,*,ilev,*]), ivals2, jvals2, tvals2),dim=1 )
> out_vals=transpose(out_vals,[0,2,1])
>
> and if that is not fast enough, do the interpolation manually:
>
> wFF=rebin(floor(ivals)+280l*floor(jvals),[36,100,60,720],/sa mple)+ $
> rebin(280l^2*lindgen(1,1,60,720),[36,100,60,720],/sample)
> wFC=rebin(floor(ivals)+280l* ceil(jvals),[36,100,60,720],/sample)+ $
> rebin(280l^2*lindgen(1,1,60,720),[36,100,60,720],/sample)
> wCF=rebin( ceil(ivals)+280l*floor(jvals),[36,100,60,720],/sample)+ $
> rebin(280l^2*lindgen(1,1,60,720),[36,100,60,720],/sample)
> wCC=rebin( ceil(ivals)+280l* ceil(jvals),[36,100,60,720],/sample)+ $
> rebin(280l^2*lindgen(1,1,60,720),[36,100,60,720],/sample)
> weightFF=rebin( (1+(ivals mod 1))*(1+(jvals mod 1)), $
> [36,100,60,720],/sample)
> weightFC=rebin(-(1+(ivals mod 1))* (jvals mod 1) , $
> [36,100,60,720],/sample)
> weightCF=rebin(- (ivals mod 1) *(1+(jvals mod 1)), $
> [36,100,60,720],/sample)
> weightCC=rebin( (ivals mod 1) * (jvals mod 1) , $
> [36,100,60,720],/sample)
> out_vals=mean(weightFF*temp[wFF]+weightFC*temp[wFC]+weightCF *temp[wCF]+weightCC*temp[wCC],dim=1)
>
> of which only the last line has to be repeated for every data set.
>
> I haven't debugged anything, so some corrections might be necessary.
>
> Good luck, Markus
Awesome! Thanks very much. The vectorized method works super quick. What used to take over two hours using the loops now takes 20 seconds! This has made my life much easier! :-)
Liam
|
|
|