Resample Lanczos-4 problem

[mschuster]

I am seeing an intermittent problem with Resample when using Lanczos-4 interpolation.

Files are in 'Forum Shared Files/mschuster/lanczos-4'. Open target.fit, drag its view selector to create a clone, and then drag the icon to apply the Resample process to the clone.

The results produced most of the time are fine, but occasionally I see some bad pixel values generated (white pixels along the top of the image). The other fit files in the folder show good and bad results.

The problem seems to be specific to Lanczos-4. I have not seen this problem with the other interpolation modes.

01.07.06.793 Starbuck (eng x86-64) on Win7.

[Nocturnal]

I believe the announcement of the Lanczos interpolation included a warning that under some conditions artifacts can be introduced. I haven't compared your images to those warnings, sorry.

[mschuster]

Hi Sander, could be, but when I run the same process multiple times I get differing results, so maybe this is different.

[Nocturnal]

Hi Mike,

I look forward to hearing from the experts.

[Juan Conejero]

Hi Mike,

Bug confirmed. This should not happen with any interpolation algorithm.

This is one of the weirdest things I've seen in a long time. Since the artifacts are variable and the problem is intermittent, It looks like the typical thread synchronization error. However, the bug happens exactly the same way with parallel execution disabled. With only one thread running one would expect the same result on each execution, so something very strange is happening.

Thank you for detecting this problem. I'll investigate and fix it as soon as possible.

[pfile]

some kind of weird floating point bug perhaps? too bad we don't have other architectures besides x86 anymore to get sanity checks on stuff like this.

[mschuster]

Please check Lanczos-3 also. It looks like it suffers the same problem as Lanczos-4. The problem is very rare however, by far most of the time it works fine.