New:
---
GPU porting of pres, swap_timelevel. Adjustments of openACC directives.
Further porting of poisfft, which now runs completely on GPU without any
host/device data transfer for serial an parallel runs (but parallel runs
require data transfer before and after the MPI transpositions).
GPU-porting of tridiagonal solver:
tridiagonal routines split into extermal subroutines (instead using CONTAINS),
no distinction between parallel/non-parallel in poisfft and tridia any more,
tridia routines moved to end of file because of probable bug in PGI compiler
(otherwise "invalid device function" is indicated during runtime).
(cuda_fft_interfaces, fft_xy, flow_statistics, init_3d_model, palm, poisfft, pres, prognostic_equations, swap_timelevel, time_integration, transpose)
output of accelerator board information. (header)
optimization of tridia routines: constant elements and coefficients of tri are
stored in seperate arrays ddzuw and tric, last dimension of tri reduced from 5 to 2,
(init_grid, init_3d_model, modules, palm, poisfft)
poisfft_init is now called internally from poisfft,
(Makefile, Makefile_check, init_pegrid, poisfft, poisfft_hybrid)
CPU-time per grid point and timestep is output to CPU_MEASURES file
(cpu_statistics, modules, time_integration)
Changed:
resorting from/to array work changed, work now has 4 dimensions instead of 1 (transpose)
array diss allocated only if required (init_3d_model)
pressure boundary condition "Neumann+inhomo" removed from the code
(check_parameters, header, poisfft, poisfft_hybrid, pres)
Errors:
bugfix: dependency added for cuda_fft_interfaces (Makefile)
bugfix: CUDA fft plans adjusted for domain decomposition (before they always
used total domain) (fft_xy)
