The branch https://bitbucket.org/eschnett/carpet/branch/ianhinder/fasterp_opt#diff contains an optimisation to the CarpetInterp2 interpolation routine which improved the performance of Llama interpatch interpolation in my test by a factor of 5. I did this a while ago, and haven't looked at it recently. Before merging, the following should be done: 1. Check that is applies cleanly to the current version of Carpet 2. Decide whether the vectorisation pragmas need to be protected by Cactus preprocessor guards Or any other changes which people think might be necessary. This should wait until after the upcoming (May 2015) release of the ET.