Now fast active search works, and can be controlled from config file

This commit is contained in:
Yohai Meiron 2020-05-01 20:41:12 -04:00
parent 2f8f8c582c
commit 1a438449a8
4 changed files with 44 additions and 84 deletions

View file

@ -31,7 +31,7 @@ eta = 0.01
# processes on a machine with a single device, set the value to 1 and use the
# mpirun utility (or whatever is used in your job scheduler) to launch as many
# processes as you like.
devices_per_node = 1
#devices_per_node = 1
##########
@ -168,7 +168,7 @@ devices_per_node = 1
# always included.
#pn_usage = {1, 1, 1, 1, 0, 0, 0}
# The speed of light in N-body units [default: 500]
# The speed of light in N-body units
#pn_c = 477.12
# The spin vectors of the two SMBHs. Only define these if the last component of
@ -188,22 +188,26 @@ devices_per_node = 1
dt_scf = 0.015625
# Name of the mask file for GRAPite [default: grapite.mask]
grapite_mask_file_name = grapite.mask
#grapite_mask_file_name = grapite.mask
# Whether to write to disk a list of SCF coefficients at every dt_disk. [default: false]
etics_dump_coeffs = true
#etics_dump_coeffs = true
# Whether to use an alternative procedure for active particle search that is
# available in the GRAPite library. This requires the number of particles in
# each MPI process to be exactly divisible by 32. This can substantially
# accelerate the calculation in some circumstances [default: false]
grapite_active_search = true
#grapite_active_search = true
# If the number of active particles in a particular bunch is bigger than this
# threshold, then the execution is on the GPU, otherwise on the CPU. When the
# active bunch is small, the overhead of calculating the SCF gravity on the GPU
# makes the operation more expensive than if it is done on the CPU. [default: 32]
#grapite_dev_exec_threshold = 512
# TODO
########
# etics dump mode
# threshold for execution on device for grapite
# scaling parameter override