First, make an array of vertices for your data array (say nx × ny × nz):

nverts = nx*ny*nz

lverts = lindgen(nverts)

Then re-arrange these vertices so that they are xi, yi, zi for all i-elements in your data array:

verts = transpose( [ [lverts mod nx], [lverts/nx mod ny], [lverts/nx/ny] ] )

Next, use MATRIX_MULTIPLY to get the transformed vertices:

verts = ( matrix_multiply(tmat,[float(temporary(verts)), replicate(1.0, 1L, nverts)],/atranspose) )[0:2,*]

Now, interpolate the data to the new vertices:

newarray = interpolate(float(array),verts[0,*],verts[1,*],verts[2,*])

And, finally, reform the newarray as nx × ny × nz:

newarray = reform(newarray,nx,ny,nz,/overwrite)

That does it.  I’ve written it as a function here, with an optional origin keyword.

The results are impressive.  I ran all the “demos” and the difference is about 21-24 X!  I can’t wait to try to do some “real” work using this.

The CPU on this machine is a Dual Quad-core Intel Xeon (X5482 @ 3.2GHz), and it has 64Gb RAM, so it’s not too shabby itself ;^)

Here’s the log from running “bench.pro”:

IDL> dlm_load, ‘gpulib’
% Loaded DLM: GPULIB.
0.756607     0.756607      1.19352     0.206724    0.0188206     0.756607
0.756607     0.756607      1.19352     0.206724    0.0188206     0.756607
N iter   =       50
CPU Time =       0.64100003
GPU Time =      0.031000137
Speedup  =        20.677329
IDL> print, !version
{ x86_64 Win32 Windows Microsoft Windows 7.0 Oct 25 2007      64      64}