First pass at swm_fortran #26
Conversation
Based on swm_c, which is funny because that was based on an old Fortran code that isn't in this library I still need to add write_to_file() and compare those binaries to the C code, but the stdout looks good (diagonals match the C output)
This means dswap() is a copy rather than shuffling pointers around
By default this is not called, but it's useful to have for debugging
|
I ran both the C and Fortran versions for a single time step and compared the differences: I was expecting double-precision round-off differences here, I wonder if some of the C constants are being evaluated in single precision? For Fortran I'm compiling with |
Instead of using pointers and allocate() statements, I explcitly define targets for u, v, and p at 3 different time levels and then point to those targets
Note that swm_fortran_driver is much slower than swm_fortran_driver, presumably because we are passing large arrays to the kernels and then only changing a single value in the arrays
the amrex_driver passes a full 3D array as an argument (inside a loop), while the swm_fortran_driver passes a 2D array (and the loop is inside the kernel).
I think it was fortls giving some lint-like warnings...
|
|
Based on
swm_c, which is funny because that was based on an old Fortran code that isn't in this libraryI still need to add
write_to_file()and compare those binaries to the C code, but the stdout looks good (diagonals match the C output); I'll take this out of draft when it's actually ready for consideration