def _(iet):
# TODO: we need to pick the rank from `comm_shm`, not `comm`,
# so that we have nranks == ngpus (as long as the user has launched
# the right number of MPI processes per node given the available
# number of GPUs per node)
objcomm = None
for i in iet.parameters:
if isinstance(i, MPICommObject):
objcomm = i
break
devicetype = as_list(self.lang[self.platform])
try:
lang_init = [self.lang['init'](devicetype)]
except TypeError:
# Not all target languages need to be explicitly initialized
lang_init = []
deviceid = DeviceID()
if objcomm is not None:
rank = Symbol(name='rank')
rank_decl = LocalExpression(DummyEq(rank, 0))
rank_init = Call('MPI_Comm_rank', [objcomm, Byref(rank)])
ngpus = Symbol(name='ngpus')
call = self.lang['num-devices'](devicetype)
ngpus_init = LocalExpression(DummyEq(ngpus, call))
osdd_then = self.lang['set-device']([deviceid] + devicetype)
osdd_else = self.lang['set-device']([rank % ngpus] +
devicetype)
body = lang_init + [
Conditional(
CondNe(deviceid, -1),
osdd_then,
List(
body=[rank_decl, rank_init, ngpus_init, osdd_else
]),
)
]
header = c.Comment('Begin of %s+MPI setup' % self.lang['name'])
footer = c.Comment('End of %s+MPI setup' % self.lang['name'])
else:
body = lang_init + [
Conditional(
CondNe(deviceid, -1),
self.lang['set-device']([deviceid] + devicetype))
]
header = c.Comment('Begin of %s setup' % self.lang['name'])
footer = c.Comment('End of %s setup' % self.lang['name'])
init = List(header=header, body=body, footer=(footer, c.Line()))
iet = iet._rebuild(body=(init, ) + iet.body)
return iet, {'args': deviceid}
def _alloc_scalar_on_low_lat_mem(self, site, expr, storage):
"""
Allocate a Scalar in the low latency memory.
"""
key = (site, expr.write
) # Ensure a scalar isn't redeclared in the given site
storage.map(key, expr, LocalExpression(**expr.args))
def _alloc_scalar_on_low_lat_mem(self, scope, expr, storage):
"""Allocate a Scalar in the low latency memory."""
handle = storage._low_lat_mem.setdefault(scope, OrderedDict())
obj = expr.write
if obj in handle:
return
handle[obj] = None # Placeholder to avoid reallocation
storage._low_lat_mem[expr] = LocalExpression(**expr.args)