drizzle
drizzle copied to clipboard
Fix internal handling of floats to be 64-bit
Currently we are seeing round-off errors that mean drizzle results can vary depending on platform and whether OpenBLAS or MKL are used on the machine. Internally, some computations are done using 32-bit FLOAT. These should be 64-bit DOUBLE.
Hopefully this will fix the round-off issues we are seeing.
Should come up with a minimally-reproducible example, though this is difficult due to platform dependency.