Speeding up interp2

I'm performing 2-D interpolation on a large gridded dataset. The method of call I'm using is

vals = interp2(V,Xq,Yq)

Where V is a large image, but Xq and Yq are fairly sparse in comparison to the image size. Xq and Yq may be invalid, so I do need to preserve the NaN returns when they are.

The code is painfully slow and running the profiler shows that line 105 in interp2 is the culprit

V = V.';

is taking 82% of the execution time! (Not sure why the author chose to transpose V rather than just swap the index arrays, but whatever, that's a Matlab "bug" I guess.)

Is there an alternate way to do this without interp2 so that I can avoid the unnecessary and wasteful transpose? Should I just call makegriddedinterp() directly the way it looks like interp2 does anyway (and swap X and Y in the call)?

[Philip Freeman]

Digging in, I see that interp2's call to makegriddedinterpolant calls the Matlab builtin griddedInterpolant(). I'll try calling that directly rather than interp2.

[Sean de Wolski]

In general using griddedInterpolant will also be faster for subsequent interpolations because you build it once, it does the underlying computations once, and then are reused. This only helps if you need to call multiple times though.

[Philip Freeman]

Thanks, Sean. That is what I ended up doing and the performance is greatly improved.