def _make_partree(self, candidates, nthreads=None):
"""
Parallelize the `candidates` Iterations attaching suitable OpenMP pragmas
for parallelism. In particular:
* All parallel Iterations not *writing* to a host Function, that
is a Function `f` such that ``is_on_device(f) == False`, are offloaded
to the device.
* The remaining ones, that is those writing to a host Function,
are parallelized on the host.
"""
assert candidates
root = candidates[0]
if is_on_device(root, self.gpu_fit, only_writes=True):
# The typical case: all written Functions are device Functions, that is
# they're mapped in the device memory. Then we offload `root` to the device
# Get the collapsable Iterations
collapsable = self._find_collapsable(root, candidates)
ncollapse = 1 + len(collapsable)
body = self._Iteration(gpu_fit=self.gpu_fit, ncollapse=ncollapse, **root.args)
partree = ParallelTree([], body, nthreads=nthreads)
collapsed = [partree] + collapsable
return root, partree, collapsed
elif not self.par_disabled:
# Resort to host parallelism
return super()._make_partree(candidates, nthreads)
else:
return root, None, None
def _make_partree(self, candidates, nthreads=None):
"""
Parallelize the `candidates` Iterations attaching suitable OpenMP pragmas
for GPU offloading.
"""
assert candidates
root = candidates[0]
# Get the collapsable Iterations
collapsable = self._find_collapsable(root, candidates)
ncollapse = 1 + len(collapsable)
# Prepare to build a ParallelTree
# Create a ParallelTree
body = self._Iteration(ncollapse=ncollapse, **root.args)
partree = ParallelTree([], body, nthreads=nthreads)
collapsed = [partree] + collapsable
return root, partree, collapsed
def _make_partree(self, candidates, nthreads=None):
"""
Parallelize the `candidates` Iterations attaching suitable OpenMP pragmas
for GPU offloading.
"""
assert candidates
root = candidates[0]
# Get the collapsable Iterations
collapsable = self._find_collapsable(root, candidates)
ncollapse = 1 + len(collapsable)
# Prepare to build a ParallelTree
omp_pragma = self.lang['par-for-teams'](ncollapse)
# Create a ParallelTree
body = root._rebuild(pragmas=root.pragmas + (omp_pragma, ),
properties=root.properties +
(COLLAPSED(ncollapse), ))
partree = ParallelTree([], body, nthreads=nthreads)
collapsed = [partree] + collapsable
return root, partree, collapsed