[Numpy-discussion] How to speed up this function?
Tim Hochberg
tim.hochberg at ieee.org
Tue Dec 5 12:16:05 CST 2006
fsenkel at verizon.net wrote:
> Hello,
>
> I'm taking a CFD class, one of the codes I wrote runs very slow. When I look at hotshot is says the function below is the problem. Since this is an explicit step, the for loops are only traversed once, so I think it's caused by memory usage, but I'm not sure if it's the local variables or the loop? I can vectorize the inner loop, would declaring the data structures in the calling routine and passing them in be a better idea than using local storage?
>
> I'm new at python and numpy, I need to look at how to get profiling information for the lines within a function.
>
>
> Thank you,
>
> Frank
>
>
> PS
> I tried to post this via google groups, but it didn't seem to go through, sorry if it ends up as multiple postings
>
>
> def findw(wnext,wprior,phiprior,uprior,vprior):
> #format here is x[i,j] where i's are rows, j's columns, use flipud() to get the
> #print out consistent with the spacial up-down directions
>
> #assign local names that are more
> #inline with the class notation
> w = wprior
> phi = phiprior
> u = uprior
> v = vprior
>
>
> #three of the BC are known so just set them
> #symetry plane
> wnext[0,0:gcols] = 0.0
>
> #upper wall
> wnext[gN,0:gcols] = 2.0/gdy**2 * (phi[gN,0:gcols] - phi[gN-1,0:gcols])
>
> #inlet, off the walls
> wnext[1:grows-1,0] = 0.0
>
>
> upos = where(u>0)
> vpos = where(v>0)
>
> Sx = ones_like(u)
> Sx[upos] = 0.0
>
> Sy = ones_like(v)
> Sy[vpos] = 0.0
>
> uw = u*w
> vw = v*w
>
> #interior nodes
> for j in range(1,gasizej-1):
> for i in range(1,gasizei-1):
>
> wnext[i,j] =( w[i,j] + gnu*gdt/gdx**2 * (w[i,j-1] - 2.0*w[i,j] + w[i,j+1]) +
> gnu*gdt/gdy**2 * (w[i-1,j] - 2.0*w[i,j] + w[i+1,j]) -
> (1.0 - Sx[i,j]) * gdt/gdx * (uw[i,j] - uw[i,j-1]) -
> Sx[i,j] * gdt/gdx * (uw[i,j+1] - uw[i,j]) -
> (1.0 - Sy[i,j]) * gdt/gdy * (vw[i,j] - vw[i-1,j]) -
> Sy[i,j] * gdt/gdy * (vw[i+1,j] - vw[i,j]) )
>
I imagine that this loop is what is killing you. Remove at least the
inner loop, if not both (try removing just the inner loop as well as
both since sometimes it's faster to just remove the inner one due to
memory usage issues..) removing both will look something like:
wnext[1:-1,1:-1] = w[1:-1, 1:-1] + gnu*gdx**2* (w[1:-1, 0:-2] -
2.0*w[1:-1, 1:-1] + w[1:-1,2:] etc, etc
When you're done with that, note also that you have the same array term
present multiple times. You could save more time by collapsing those
terms and using different scalar multipliers. Occasionally that is
numerically unwise, but I doubt it in this case.
There all sorts of other things that you can do such as using inplace
operations, etc. But try vectorizing the loop first.
-tim
> ## print "***wnext****"
> ## print "i: ", i, "j: ", j, "wnext[i,j]: ", wnext[i,j]
>
> #final BC at outlet, off walls
> wnext[1:grows-1,gM] = wnext[1:grows-1,gM-1]
> _______________________________________________
> Numpy-discussion mailing list
> Numpy-discussion at scipy.org
> http://projects.scipy.org/mailman/listinfo/numpy-discussion
>
>
>
More information about the Numpy-discussion
mailing list