def finish(self):
new_instructions = []
insns_to_insert = dict(
(insn.id, insn) for insn in self.insns_to_insert)
for orig_insn in self.kernel.instructions:
if orig_insn.id in self.insns_to_update:
new_instructions.append(self.insns_to_update[orig_insn.id])
else:
new_instructions.append(orig_insn)
new_instructions.extend(
sorted(insns_to_insert.values(), key=lambda insn: insn.id))
self.updated_iname_to_tag.update(self.kernel.iname_to_tag)
self.updated_temporary_variables.update(
self.kernel.temporary_variables)
new_domains = list(self.kernel.domains)
import islpy as isl
if self.new_subdomain.dim(isl.dim_type.set) > 0:
new_domains.append(self.new_subdomain)
kernel = self.kernel.copy(
domains=new_domains,
instructions=new_instructions,
iname_to_tag=self.updated_iname_to_tag,
temporary_variables=self.updated_temporary_variables,
overridden_get_grid_sizes_for_insn_ids=None)
# Add nosync directives to any saves or reloads that were added with a
# potential dependency chain.
from loopy.kernel.tools import get_subkernels
for subkernel in get_subkernels(kernel):
relevant_insns = self.subkernel_to_newly_added_insn_ids[subkernel]
from itertools import product
for temporary in self.temporary_to_reload_ids:
for source, sink in product(
relevant_insns
& self.temporary_to_reload_ids[temporary],
relevant_insns
& self.temporary_to_save_ids[temporary]):
kernel = lp.add_nosync(kernel, "global", source, sink)
from loopy.kernel.tools import assign_automatic_axes
return assign_automatic_axes(kernel)
def finish(self):
new_instructions = []
insns_to_insert = dict((insn.id, insn) for insn in self.insns_to_insert)
for orig_insn in self.kernel.instructions:
if orig_insn.id in self.insns_to_update:
new_instructions.append(self.insns_to_update[orig_insn.id])
else:
new_instructions.append(orig_insn)
new_instructions.extend(
sorted(insns_to_insert.values(), key=lambda insn: insn.id))
self.updated_iname_to_tags.update(self.kernel.iname_to_tags)
self.updated_temporary_variables.update(self.kernel.temporary_variables)
new_domains = list(self.kernel.domains)
import islpy as isl
if self.new_subdomain.dim(isl.dim_type.set) > 0:
new_domains.append(self.new_subdomain)
kernel = self.kernel.copy(
domains=new_domains,
instructions=new_instructions,
iname_to_tags=self.updated_iname_to_tags,
temporary_variables=self.updated_temporary_variables,
overridden_get_grid_sizes_for_insn_ids=None)
# Add nosync directives to any saves or reloads that were added with a
# potential dependency chain.
from loopy.kernel.tools import get_subkernels
for subkernel in get_subkernels(kernel):
relevant_insns = self.subkernel_to_newly_added_insn_ids[subkernel]
from itertools import product
for temporary in self.temporary_to_reload_ids:
for source, sink in product(
relevant_insns & self.temporary_to_reload_ids[temporary],
relevant_insns & self.temporary_to_save_ids[temporary]):
kernel = lp.add_nosync(kernel, "global", source, sink)
from loopy.kernel.tools import assign_automatic_axes
return assign_automatic_axes(kernel)
def check_that_temporaries_are_defined_in_subkernels_where_used(kernel):
from loopy.kernel.data import AddressSpace
from loopy.kernel.tools import get_subkernels
for subkernel in get_subkernels(kernel):
defined_base_storage = set()
from loopy.schedule.tools import (temporaries_written_in_subkernel,
temporaries_read_in_subkernel)
for temporary in temporaries_written_in_subkernel(kernel, subkernel):
tval = kernel.temporary_variables[temporary]
if tval.base_storage is not None:
defined_base_storage.add(tval.base_storage)
for temporary in (temporaries_read_in_subkernel(kernel, subkernel) -
temporaries_written_in_subkernel(kernel, subkernel)):
tval = kernel.temporary_variables[temporary]
if tval.initializer is not None:
continue
# For aliased temporaries, check if there is an aliased definition.
if tval.base_storage is not None:
if tval.base_storage not in defined_base_storage:
from loopy.diagnostic import MissingDefinitionError
raise MissingDefinitionError(
"temporary variable '%s' gets "
"used in subkernel '%s' and neither it nor its "
"aliases have a definition" % (temporary, subkernel))
continue
if tval.address_space in (AddressSpace.PRIVATE,
AddressSpace.LOCAL):
from loopy.diagnostic import MissingDefinitionError
raise MissingDefinitionError(
"temporary variable '%s' gets used "
"in subkernel '%s' without a definition (maybe you forgot "
"to call loopy.save_and_reload_temporaries?)" %
(temporary, subkernel))
def check_that_temporaries_are_defined_in_subkernels_where_used(kernel):
from loopy.kernel.data import AddressSpace
from loopy.kernel.tools import get_subkernels
for subkernel in get_subkernels(kernel):
defined_base_storage = set()
from loopy.schedule.tools import (
temporaries_written_in_subkernel, temporaries_read_in_subkernel)
for temporary in temporaries_written_in_subkernel(kernel, subkernel):
tval = kernel.temporary_variables[temporary]
if tval.base_storage is not None:
defined_base_storage.add(tval.base_storage)
for temporary in (
temporaries_read_in_subkernel(kernel, subkernel) -
temporaries_written_in_subkernel(kernel, subkernel)):
tval = kernel.temporary_variables[temporary]
if tval.initializer is not None:
continue
# For aliased temporaries, check if there is an aliased definition.
if tval.base_storage is not None:
if tval.base_storage not in defined_base_storage:
from loopy.diagnostic import MissingDefinitionError
raise MissingDefinitionError("temporary variable '%s' gets "
"used in subkernel '%s' and neither it nor its "
"aliases have a definition" % (temporary, subkernel))
continue
if tval.address_space in (AddressSpace.PRIVATE, AddressSpace.LOCAL):
from loopy.diagnostic import MissingDefinitionError
raise MissingDefinitionError("temporary variable '%s' gets used "
"in subkernel '%s' without a definition (maybe you forgot "
"to call loopy.save_and_reload_temporaries?)"
% (temporary, subkernel))