diff --git a/loopy/check.py b/loopy/check.py
index 9a9ff1fd575883ad4ee212f9e17f7c0197781201..bebd86fffe00d9374da4352206952f1e439ef883 100644
--- a/loopy/check.py
+++ b/loopy/check.py
@@ -168,7 +168,7 @@ def _is_racing_iname_tag(tv, tag):
def check_for_write_races(kernel):
- from loopy.kernel.data import ConcurrentTag, check_iname_tags
+ from loopy.kernel.data import ConcurrentTag, get_iname_tags
for insn in kernel.instructions:
for assignee_name, assignee_indices in zip(
@@ -186,8 +186,9 @@ def check_for_write_races(kernel):
# will cause write races.
raceable_parallel_insn_inames = set(
- iname for iname in kernel.insn_inames(insn)
- if check_iname_tags(kernel.iname_to_tags[iname], ConcurrentTag))
+ iname for iname in kernel.insn_inames(insn)
+ if get_iname_tags(kernel.iname_to_tags[iname],
+ ConcurrentTag))
elif assignee_name in kernel.temporary_variables:
temp_var = kernel.temporary_variables[assignee_name]
@@ -213,7 +214,7 @@ def check_for_write_races(kernel):
def check_for_orphaned_user_hardware_axes(kernel):
- from loopy.kernel.data import LocalIndexTag, check_iname_tags
+ from loopy.kernel.data import LocalIndexTag
for axis in kernel.local_sizes:
found = False
for tags in six.itervalues(kernel.iname_to_tags):
@@ -230,12 +231,12 @@ def check_for_orphaned_user_hardware_axes(kernel):
def check_for_data_dependent_parallel_bounds(kernel):
- from loopy.kernel.data import ConcurrentTag, check_iname_tags
+ from loopy.kernel.data import ConcurrentTag, get_iname_tags
for i, dom in enumerate(kernel.domains):
dom_inames = set(dom.get_var_names(dim_type.set))
par_inames = set(iname for iname in dom_inames
- if check_iname_tags(kernel.iname_to_tags[iname], ConcurrentTag))
+ if get_iname_tags(kernel.iname_to_tags[iname], ConcurrentTag))
if not par_inames:
continue
@@ -651,7 +652,7 @@ def _check_for_unused_hw_axes_in_kernel_chunk(kernel, sched_index=None):
# alternative: just disregard length-1 dimensions?
from loopy.kernel.data import (LocalIndexTag, AutoLocalIndexTagBase,
- GroupIndexTag, check_iname_tags, get_iname_tags)
+ GroupIndexTag, get_iname_tags)
while i < loop_end_i:
sched_item = kernel.schedule[i]
@@ -671,19 +672,13 @@ def _check_for_unused_hw_axes_in_kernel_chunk(kernel, sched_index=None):
for iname in kernel.insn_inames(insn):
tags = kernel.iname_to_tags[iname]
- if check_iname_tags(tags, LocalIndexTag):
- tags = get_iname_tags(tags, LocalIndexTag)
- if len(tags) > 1:
- raise LoopyError("Can only have one LocalIndexTag")
- tag, = tags
+ if get_iname_tags(tags, LocalIndexTag):
+ tag, = get_iname_tags(tags, LocalIndexTag, 1)
local_axes_used.add(tag.axis)
- elif check_iname_tags(tags, GroupIndexTag):
- tags = get_iname_tags(tags, GroupIndexTag)
- if len(tags) > 1:
- raise LoopyError("Can only have one GroupIndexTag")
- tag, = tags
+ elif get_iname_tags(tags, GroupIndexTag):
+ tag, = get_iname_tags(tags, GroupIndexTag, 1)
group_axes_used.add(tag.axis)
- elif check_iname_tags(tags, AutoLocalIndexTagBase):
+ elif get_iname_tags(tags, AutoLocalIndexTagBase):
raise LoopyError("auto local tag encountered")
if group_axes != group_axes_used:
@@ -896,8 +891,6 @@ def check_implemented_domains(kernel, implemented_domains, code=None):
from islpy import align_two
- from loopy.kernel.data import check_iname_tags
-
last_idomains = None
last_insn_inames = None
@@ -930,11 +923,11 @@ def check_implemented_domains(kernel, implemented_domains, code=None):
.project_out_except(insn_inames, [dim_type.set]))
from loopy.kernel.instruction import BarrierInstruction
- from loopy.kernel.data import LocalIndexTag
+ from loopy.kernel.data import LocalIndexTag, get_iname_tags
if isinstance(insn, BarrierInstruction):
# project out local-id-mapped inames, solves #94 on gitlab
non_lid_inames = frozenset(iname for iname in insn_inames
- if not check_iname_tags(kernel.iname_to_tags[iname], LocalIndexTag))
+ if not get_iname_tags(kernel.iname_to_tags[iname], LocalIndexTag))
insn_impl_domain = insn_impl_domain.project_out_except(
non_lid_inames, [dim_type.set])
diff --git a/loopy/codegen/bounds.py b/loopy/codegen/bounds.py
index 886f305a455f66b943660ff653c40ae632360243..240df24e5d2a6dc7ea620a6f2f290937b9119384 100644
--- a/loopy/codegen/bounds.py
+++ b/loopy/codegen/bounds.py
@@ -59,7 +59,7 @@ def get_usable_inames_for_conditional(kernel, sched_index):
from loopy.schedule import (
find_active_inames_at, get_insn_ids_for_block_at, has_barrier_within)
from loopy.kernel.data import (ConcurrentTag, LocalIndexTagBase,
- IlpBaseTag, check_iname_tags)
+ IlpBaseTag, get_iname_tags)
result = find_active_inames_at(kernel, sched_index)
crosses_barrier = has_barrier_within(kernel, sched_index)
@@ -98,9 +98,9 @@ def get_usable_inames_for_conditional(kernel, sched_index):
# at the innermost level of nesting.
if (
- check_iname_tags(tags, ConcurrentTag)
- and not (check_iname_tags(tags, LocalIndexTagBase)
- and crosses_barrier) and not check_iname_tags(tags, IlpBaseTag)
+ get_iname_tags(tags, ConcurrentTag)
+ and not (get_iname_tags(tags, LocalIndexTagBase)
+ and crosses_barrier) and not get_iname_tags(tags, IlpBaseTag)
):
result.add(iname)
diff --git a/loopy/codegen/control.py b/loopy/codegen/control.py
index 22f18760882ae984de56ffd050ee21573b46fbad..41b04e172a2afeef612d31bf2fff323d096945ec 100644
--- a/loopy/codegen/control.py
+++ b/loopy/codegen/control.py
@@ -41,7 +41,7 @@ def get_admissible_conditional_inames_for(codegen_state, sched_index):
kernel = codegen_state.kernel
from loopy.kernel.data import (LocalIndexTag, HardwareConcurrentTag,
- check_iname_tags)
+ get_iname_tags)
from loopy.schedule import find_active_inames_at, has_barrier_within
result = find_active_inames_at(kernel, sched_index)
@@ -49,9 +49,9 @@ def get_admissible_conditional_inames_for(codegen_state, sched_index):
has_barrier = has_barrier_within(kernel, sched_index)
for iname, tags in six.iteritems(kernel.iname_to_tags):
- if (check_iname_tags(tags, HardwareConcurrentTag)
+ if (get_iname_tags(tags, HardwareConcurrentTag)
and codegen_state.is_generating_device_code):
- if not has_barrier or not check_iname_tags(tags, LocalIndexTag):
+ if not has_barrier or not get_iname_tags(tags, LocalIndexTag):
result.add(iname)
return frozenset(result)
@@ -136,13 +136,14 @@ def generate_code_for_sched_index(codegen_state, sched_index):
generate_vectorize_loop,
generate_sequential_loop_dim_code)
- from loopy.kernel.data import (UnrolledIlpTag, UnrollTag, ForceSequentialTag,
- LoopedIlpTag, VectorizeTag, InOrderSequentialSequentialTag, check_iname_tags)
- if check_iname_tags(tags, (UnrollTag, UnrolledIlpTag)):
+ from loopy.kernel.data import (UnrolledIlpTag, UnrollTag,
+ ForceSequentialTag, LoopedIlpTag, VectorizeTag,
+ InOrderSequentialSequentialTag, get_iname_tags)
+ if get_iname_tags(tags, (UnrollTag, UnrolledIlpTag)):
func = generate_unroll_loop
- elif check_iname_tags(tags, VectorizeTag):
+ elif get_iname_tags(tags, VectorizeTag):
func = generate_vectorize_loop
- elif len(tags) == 0 or check_iname_tags(tags, (LoopedIlpTag,
+ elif len(tags) == 0 or get_iname_tags(tags, (LoopedIlpTag,
ForceSequentialTag, InOrderSequentialSequentialTag)):
func = generate_sequential_loop_dim_code
else:
diff --git a/loopy/codegen/loop.py b/loopy/codegen/loop.py
index 01f8a82554595d4eda9dfa899fa26dc262294259..0efa96f967bc597d65ed3fd57c18301d67d6990b 100644
--- a/loopy/codegen/loop.py
+++ b/loopy/codegen/loop.py
@@ -231,7 +231,7 @@ def set_up_hw_parallel_loops(codegen_state, schedule_index, next_func,
kernel = codegen_state.kernel
from loopy.kernel.data import (UniqueTag, HardwareConcurrentTag,
- LocalIndexTag, GroupIndexTag, check_iname_tags)
+ LocalIndexTag, GroupIndexTag, get_iname_tags)
from loopy.schedule import get_insn_ids_for_block_at
insn_ids_for_block = get_insn_ids_for_block_at(kernel.schedule, schedule_index)
@@ -242,7 +242,7 @@ def set_up_hw_parallel_loops(codegen_state, schedule_index, next_func,
all_inames_by_insns |= kernel.insn_inames(insn_id)
hw_inames_left = [iname for iname in all_inames_by_insns
- if check_iname_tags(kernel.iname_to_tags[iname],
+ if get_iname_tags(kernel.iname_to_tags[iname],
HardwareConcurrentTag)]
if not hw_inames_left:
@@ -258,12 +258,8 @@ def set_up_hw_parallel_loops(codegen_state, schedule_index, next_func,
from loopy.symbolic import GroupHardwareAxisIndex, LocalHardwareAxisIndex
- assert check_iname_tags(tags, UniqueTag)
+ tag, = get_iname_tags(tags, UniqueTag, max_num=1, min_num=1)
- if len(tags) > 1:
- raise LoopyError("cannot have more than one UniqueTag")
-
- tag, = tags
if isinstance(tag, GroupIndexTag):
hw_axis_expr = GroupHardwareAxisIndex(tag.axis)
elif isinstance(tag, LocalIndexTag):
@@ -271,11 +267,14 @@ def set_up_hw_parallel_loops(codegen_state, schedule_index, next_func,
else:
raise RuntimeError("unexpected hw tag type")
+ # TODO: get rid of None
+
other_inames_with_same_tag = [
other_iname for other_iname in kernel.all_inames()
- if check_iname_tags(kernel.iname_to_tags[other_iname], UniqueTag)
- and any(_tag.key == tag.key for _tag in kernel.iname_to_tags[other_iname])
- and other_iname != iname]
+ if (get_iname_tags(kernel.iname_to_tags[other_iname], UniqueTag)
+ and other_iname != iname
+ and any(_tag.key == tag.key
+ for _tag in kernel.iname_to_tags[other_iname]))]
# {{{ 'implement' hardware axis boundaries
diff --git a/loopy/kernel/__init__.py b/loopy/kernel/__init__.py
index a69f0727db132aa5b34372cd4fd9f245a2270fc5..872365fca2c87a0c4bab3ca5569acc471364d63a 100644
--- a/loopy/kernel/__init__.py
+++ b/loopy/kernel/__init__.py
@@ -44,7 +44,7 @@ from loopy.library.function import (
from loopy.diagnostic import CannotBranchDomainTree, LoopyError
from loopy.tools import natsorted
from loopy.diagnostic import StaticValueFindingError
-from loopy.kernel.data import check_iname_tags, get_iname_tags
+from loopy.kernel.data import get_iname_tags
# {{{ unique var names
@@ -197,7 +197,7 @@ class LoopKernel(ImmutableRecordWithoutPickling):
assumptions=None,
local_sizes={},
temporary_variables={},
- iname_to_tags=defaultdict(tuple),
+ iname_to_tags=defaultdict(set),
substitutions={},
function_manglers=[
default_function_mangler,
@@ -711,11 +711,9 @@ class LoopKernel(ImmutableRecordWithoutPickling):
from loopy.kernel.data import HardwareConcurrentTag
for iname in cond_inames:
- tags = self.iname_to_tags[iname]
- if check_iname_tags(tags, HardwareConcurrentTag):
- tags = get_iname_tags(tags, HardwareConcurrentTag)
- if len(tags) > 1:
- raise LoopyError("cannot have more than one HardwareConcurentTags")
+ tags = get_iname_tags(self.iname_to_tags[iname],
+ HardwareConcurrentTag, 1)
+ if tags:
tag, = tags
tag_key_uses[tag.key].append(iname)
@@ -725,8 +723,10 @@ class LoopKernel(ImmutableRecordWithoutPickling):
multi_use_inames = set()
for iname in cond_inames:
- for tag in self.iname_to_tags[iname]:
- if isinstance(tag, HardwareConcurrentTag) and tag.key in multi_use_keys:
+ tags = get_iname_tags(self.iname_to_tags[iname], HardwareConcurrentTag)
+ if tags:
+ tag, = get_iname_tags(tags, HardwareConcurrentTag, 1)
+ if tag.key in multi_use_keys:
multi_use_inames.add(iname)
break
@@ -960,22 +960,17 @@ class LoopKernel(ImmutableRecordWithoutPickling):
for iname in all_inames_by_insns:
tags = self.iname_to_tags[iname]
- if check_iname_tags(tags, GroupIndexTag):
+ if get_iname_tags(tags, GroupIndexTag):
tgt_dict = global_sizes
- elif check_iname_tags(tags, LocalIndexTag):
+ elif get_iname_tags(tags, LocalIndexTag):
tgt_dict = local_sizes
- elif check_iname_tags(tags, AutoLocalIndexTagBase) and not ignore_auto:
+ elif get_iname_tags(tags, AutoLocalIndexTagBase) and not ignore_auto:
raise RuntimeError("cannot find grid sizes if automatic "
"local index tags are present")
else:
continue
- tags = get_iname_tags(tags, (GroupIndexTag, LocalIndexTag))
-
- if len(tags) != 1:
- raise LoopyError("Multiple axis tag not allowed")
-
- tag, = tags
+ tag, = get_iname_tags(tags, (GroupIndexTag, LocalIndexTag), 1)
size = self.get_iname_bounds(iname).size
diff --git a/loopy/kernel/data.py b/loopy/kernel/data.py
index 9b66088e5b75ca6bad3d06d931112de29980b634..9250c5acf7304e19572a96ec214452a62d891254 100644
--- a/loopy/kernel/data.py
+++ b/loopy/kernel/data.py
@@ -55,19 +55,24 @@ class auto(object): # noqa
# {{{ iname tags
-def check_iname_tags(tags, tag_type):
- return any([isinstance(tag, tag_type) for tag in tags])
-
-
-def get_iname_tags(tags, tag_type):
- return tuple(tag for tag in tags if isinstance(tag, tag_type))
+def get_iname_tags(tags, tag_type, max_num=None, min_num=None):
+ result = set(tag for tag in tags if isinstance(tag, tag_type))
+ if max_num:
+ if len(result) > max_num:
+ raise LoopyError("cannot have more than {0} tags"
+ "of type(s): {1}".format(max_num, tag_type))
+ if min_num:
+ if len(result) < min_num:
+ raise LoopyError("must have more than {0} tags"
+ "of type(s): {1}".format(max_num, tag_type))
+ return result
class IndexTag(ImmutableRecord):
__slots__ = []
def __hash__(self):
- raise RuntimeError("use .key to hash index tags")
+ return hash(self.key)
def update_persistent_hash(self, key_hash, key_builder):
"""Custom hash computation function for use with
diff --git a/loopy/kernel/tools.py b/loopy/kernel/tools.py
index 505482dea66cf2a1849b262eb3d28b121f990b88..5be3375373ca322bb3126ebf1a4407235f0ca2eb 100644
--- a/loopy/kernel/tools.py
+++ b/loopy/kernel/tools.py
@@ -36,7 +36,7 @@ from islpy import dim_type
from loopy.diagnostic import LoopyError, warn_with_kernel
from pytools import memoize_on_first_arg
from loopy.tools import natsorted
-from loopy.kernel.data import check_iname_tags
+from loopy.kernel.data import get_iname_tags
import logging
logger = logging.getLogger(__name__)
@@ -632,7 +632,7 @@ def is_domain_dependent_on_inames(kernel, domain_index, inames):
# {{{ rank inames by stride
def get_auto_axis_iname_ranking_by_stride(kernel, insn):
- from loopy.kernel.data import ImageArg, ValueArg, check_iname_tags
+ from loopy.kernel.data import ImageArg, ValueArg, get_iname_tags
approximate_arg_values = {}
for arg in kernel.args:
@@ -677,8 +677,8 @@ def get_auto_axis_iname_ranking_by_stride(kernel, insn):
from loopy.kernel.data import AutoLocalIndexTagBase
auto_axis_inames = set(
- iname for iname in kernel.insn_inames(insn)
- if check_iname_tags(kernel.iname_to_tags[iname], AutoLocalIndexTagBase))
+ iname for iname in kernel.insn_inames(insn)
+ if get_iname_tags(kernel.iname_to_tags[iname], AutoLocalIndexTagBase))
# }}}
@@ -754,7 +754,7 @@ def assign_automatic_axes(kernel, axis=0, local_size=None):
# to set() from tuple()
from loopy.kernel.data import (AutoLocalIndexTagBase, LocalIndexTag,
- check_iname_tags, get_iname_tags)
+ get_iname_tags)
# Realize that at this point in time, axis lengths are already
# fixed. So we compute them once and pass them to our recursive
@@ -832,7 +832,7 @@ def assign_automatic_axes(kernel, axis=0, local_size=None):
do_tagged_check=False),
axis=recursion_axis, local_size=local_size)
- if not check_iname_tags(kernel.iname_to_tags[iname], AutoLocalIndexTagBase):
+ if not get_iname_tags(kernel.iname_to_tags[iname], AutoLocalIndexTagBase):
raise LoopyError("trying to reassign '%s'" % iname)
new_iname_to_tags = kernel.iname_to_tags.copy()
@@ -855,7 +855,7 @@ def assign_automatic_axes(kernel, axis=0, local_size=None):
auto_axis_inames = [
iname for iname in kernel.insn_inames(insn)
- if check_iname_tags(kernel.iname_to_tags[iname], AutoLocalIndexTagBase)]
+ if get_iname_tags(kernel.iname_to_tags[iname], AutoLocalIndexTagBase)]
if not auto_axis_inames:
continue
@@ -878,7 +878,7 @@ def assign_automatic_axes(kernel, axis=0, local_size=None):
if iname_ranking is not None:
for iname in iname_ranking:
prev_tags = kernel.iname_to_tags[iname]
- if check_iname_tags(prev_tags, AutoLocalIndexTagBase):
+ if get_iname_tags(prev_tags, AutoLocalIndexTagBase):
return assign_axis(axis, iname, axis)
else:
@@ -1134,7 +1134,7 @@ def get_visual_iname_order_embedding(kernel):
# nest.
ilp_inames = frozenset(iname
for iname in kernel.iname_to_tags
- if check_iname_tags(kernel.iname_to_tags[iname], IlpBaseTag))
+ if get_iname_tags(kernel.iname_to_tags[iname], IlpBaseTag))
iname_trie = SetTrie()
diff --git a/loopy/preprocess.py b/loopy/preprocess.py
index b20fbef91168ab4a5580b031d1cab4403bfec7c1..38b213c05632670ebf98fba0534fbffba5738fa5 100644
--- a/loopy/preprocess.py
+++ b/loopy/preprocess.py
@@ -34,7 +34,7 @@ from pytools.persistent_dict import WriteOncePersistentDict
from loopy.tools import LoopyKeyBuilder
from loopy.version import DATA_MODEL_VERSION
-from loopy.kernel.data import make_assignment, check_iname_tags, get_iname_tags
+from loopy.kernel.data import make_assignment, get_iname_tags
# for the benefit of loopy.statistics, for now
from loopy.type_inference import infer_unknown_types
@@ -136,7 +136,7 @@ def check_reduction_iname_uniqueness(kernel):
def _get_compute_inames_tagged(kernel, insn, tag_base):
return set(iname for iname in kernel.insn_inames(insn.id)
- if check_iname_tags(kernel.iname_to_tags[iname], tag_base))
+ if get_iname_tags(kernel.iname_to_tags[iname], tag_base))
def _get_assignee_inames_tagged(kernel, insn, tag_base, tv_names):
@@ -146,7 +146,7 @@ def _get_assignee_inames_tagged(kernel, insn, tag_base, tv_names):
insn.assignee_subscript_deps())
for iname in adeps & kernel.all_inames()
if aname in tv_names
- if check_iname_tags(kernel.iname_to_tags[iname], tag_base))
+ if get_iname_tags(kernel.iname_to_tags[iname], tag_base))
def find_temporary_scope(kernel):
@@ -291,20 +291,20 @@ def _classify_reduction_inames(kernel, inames):
from loopy.kernel.data import (
LocalIndexTagBase, UnrolledIlpTag, UnrollTag, VectorizeTag,
- ConcurrentTag, check_iname_tags)
+ ConcurrentTag, get_iname_tags)
for iname in inames:
iname_tags = kernel.iname_to_tags[iname]
- if check_iname_tags(iname_tags, (UnrollTag, UnrolledIlpTag)):
+ if get_iname_tags(iname_tags, (UnrollTag, UnrolledIlpTag)):
# These are nominally parallel, but we can live with
# them as sequential.
sequential.append(iname)
- elif check_iname_tags(iname_tags, LocalIndexTagBase):
+ elif get_iname_tags(iname_tags, LocalIndexTagBase):
local_par.append(iname)
- elif check_iname_tags(iname_tags, (ConcurrentTag, VectorizeTag)):
+ elif get_iname_tags(iname_tags, (ConcurrentTag, VectorizeTag)):
nonlocal_par.append(iname)
else:
@@ -1136,9 +1136,9 @@ def realize_reduction(kernel, insn_id_filter=None, unknown_types_ok=True,
outer_insn_inames = temp_kernel.insn_inames(insn)
- from loopy.kernel.data import LocalIndexTagBase, check_iname_tags
+ from loopy.kernel.data import LocalIndexTagBase, get_iname_tags
outer_local_inames = tuple(oiname for oiname in outer_insn_inames
- if check_iname_tags(kernel.iname_to_tags[oiname], LocalIndexTagBase))
+ if get_iname_tags(kernel.iname_to_tags[oiname], LocalIndexTagBase))
from pymbolic import var
outer_local_iname_vars = tuple(
@@ -1471,9 +1471,9 @@ def realize_reduction(kernel, insn_id_filter=None, unknown_types_ok=True,
outer_insn_inames = temp_kernel.insn_inames(insn)
- from loopy.kernel.data import LocalIndexTagBase, check_iname_tags
+ from loopy.kernel.data import LocalIndexTagBase, get_iname_tags
outer_local_inames = tuple(oiname for oiname in outer_insn_inames
- if check_iname_tags(kernel.iname_to_tags[oiname], LocalIndexTagBase)
+ if get_iname_tags(kernel.iname_to_tags[oiname], LocalIndexTagBase)
and oiname != sweep_iname)
from pymbolic import var
@@ -1740,7 +1740,7 @@ def realize_reduction(kernel, insn_id_filter=None, unknown_types_ok=True,
"by reductions is 'local'--found iname(s) '%s' "
"respectively tagged '%s'"
% (", ".join(bad_inames),
- ", ".join(tag.key for tag in kernel.iname_to_tags[iname]
+ ", ".join(str(kernel.iname_to_tags[iname])
for iname in bad_inames)))
if n_local_par == 0 and n_sequential == 0:
@@ -2150,7 +2150,7 @@ def preprocess_kernel(kernel, device=None):
from loopy.kernel.data import AutoLocalIndexTagBase
for iname, tags in six.iteritems(kernel.iname_to_tags):
- if (check_iname_tags(tags, AutoLocalIndexTagBase)
+ if (get_iname_tags(tags, AutoLocalIndexTagBase)
and iname in kernel.all_inames()):
raise LoopyError("kernel with automatically-assigned "
"local axes passed to preprocessing")
diff --git a/loopy/schedule/__init__.py b/loopy/schedule/__init__.py
index 68b3fb0cbad892745a775f6e1ef3f6b5e34a1c5a..bda316bc170b9099e41355a4a3670902f44461e6 100644
--- a/loopy/schedule/__init__.py
+++ b/loopy/schedule/__init__.py
@@ -213,12 +213,12 @@ def find_loop_nest_with_map(kernel):
result = {}
from loopy.kernel.data import (ConcurrentTag, IlpBaseTag, VectorizeTag,
- check_iname_tags)
+ get_iname_tags)
all_nonpar_inames = set(
- iname for iname, tags in six.iteritems(kernel.iname_to_tags)
- if tags and not check_iname_tags(tags,
- (ConcurrentTag, IlpBaseTag, VectorizeTag)))
+ iname for iname in kernel.all_inames()
+ if not get_iname_tags(kernel.iname_to_tags[iname],
+ (ConcurrentTag, IlpBaseTag, VectorizeTag)))
iname_to_insns = kernel.iname_to_insns()
@@ -241,7 +241,7 @@ def find_loop_nest_around_map(kernel):
iname_to_insns = kernel.iname_to_insns()
# examine pairs of all inames--O(n**2), I know.
- from loopy.kernel.data import IlpBaseTag, check_iname_tags
+ from loopy.kernel.data import IlpBaseTag, get_iname_tags
for inner_iname in all_inames:
result[inner_iname] = set()
for outer_iname in all_inames:
@@ -249,7 +249,7 @@ def find_loop_nest_around_map(kernel):
continue
tags = kernel.iname_to_tags[outer_iname]
- if check_iname_tags(tags, IlpBaseTag):
+ if get_iname_tags(tags, IlpBaseTag):
# ILP tags are special because they are parallel tags
# and therefore 'in principle' nest around everything.
# But they're realized by the scheduler as a loop
@@ -279,10 +279,10 @@ def find_loop_insn_dep_map(kernel, loop_nest_with_map, loop_nest_around_map):
result = {}
from loopy.kernel.data import (ConcurrentTag, IlpBaseTag, VectorizeTag,
- check_iname_tags)
+ get_iname_tags)
for insn in kernel.instructions:
for iname in kernel.insn_inames(insn):
- if check_iname_tags(kernel.iname_to_tags[iname], ConcurrentTag):
+ if get_iname_tags(kernel.iname_to_tags[iname], ConcurrentTag):
continue
iname_dep = result.setdefault(iname, set())
@@ -313,7 +313,8 @@ def find_loop_insn_dep_map(kernel, loop_nest_with_map, loop_nest_around_map):
continue
tags = kernel.iname_to_tags[dep_insn_iname]
- if check_iname_tags(tags, (ConcurrentTag, IlpBaseTag, VectorizeTag)):
+ if get_iname_tags(tags,
+ (ConcurrentTag, IlpBaseTag, VectorizeTag)):
# Parallel tags don't really nest, so we'll disregard
# them here.
continue
@@ -1878,19 +1879,19 @@ def generate_loop_schedules_inner(kernel, debug_args={}):
for insn_id in sched_item_to_insn_id(item))
from loopy.kernel.data import (IlpBaseTag, ConcurrentTag, VectorizeTag,
- check_iname_tags)
+ get_iname_tags)
ilp_inames = set(
iname
for iname, tags in six.iteritems(kernel.iname_to_tags)
- if check_iname_tags(tags, IlpBaseTag))
+ if get_iname_tags(tags, IlpBaseTag))
vec_inames = set(
iname
for iname, tags in six.iteritems(kernel.iname_to_tags)
- if check_iname_tags(tags, VectorizeTag))
+ if get_iname_tags(tags, VectorizeTag))
parallel_inames = set(
iname
for iname, tags in six.iteritems(kernel.iname_to_tags)
- if check_iname_tags(tags, ConcurrentTag))
+ if get_iname_tags(tags, ConcurrentTag))
loop_nest_with_map = find_loop_nest_with_map(kernel)
loop_nest_around_map = find_loop_nest_around_map(kernel)
diff --git a/loopy/statistics.py b/loopy/statistics.py
index 77c638128fe3dab63ee79f72bd14d70e0ca866bf..4b0643873d100af365cbf2782f0c084f83b55760 100755
--- a/loopy/statistics.py
+++ b/loopy/statistics.py
@@ -1182,12 +1182,11 @@ def get_unused_hw_axes_factor(knl, insn, disregard_local_axes, space=None):
l_used = set()
from loopy.kernel.data import (LocalIndexTag, GroupIndexTag,
- get_iname_tags, check_iname_tags)
+ get_iname_tags)
for iname in knl.insn_inames(insn):
- tags = get_iname_tags(knl.iname_to_tags[iname], (LocalIndexTag, GroupIndexTag))
+ tags = get_iname_tags(knl.iname_to_tags[iname],
+ (LocalIndexTag, GroupIndexTag), 1)
if tags:
- if len(tags) > 1:
- raise LoopyError("cannot have more than one UniqueTags")
tag, = tags
if isinstance(tag, LocalIndexTag):
l_used.add(tag.axis)
@@ -1221,9 +1220,9 @@ def count_insn_runs(knl, insn, count_redundant_work, disregard_local_axes=False)
insn_inames = knl.insn_inames(insn)
if disregard_local_axes:
- from loopy.kernel.data import LocalIndexTag, check_iname_tags
+ from loopy.kernel.data import LocalIndexTag, get_iname_tags
insn_inames = [iname for iname in insn_inames if not
- check_iname_tags(kernel.iname_to_tags[iname], LocalIndexTag)]
+ get_iname_tags(knl.iname_to_tags[iname], LocalIndexTag)]
inames_domain = knl.get_inames_domain(insn_inames)
domain = (inames_domain.project_out_except(
diff --git a/loopy/target/ispc.py b/loopy/target/ispc.py
index 3fd2bc1ae6a439aeab1ff290e990143207f948d2..b7edc517bee38c3f7b6a6cdce5419863db5ffed8 100644
--- a/loopy/target/ispc.py
+++ b/loopy/target/ispc.py
@@ -418,33 +418,33 @@ class ISPCASTBuilder(CASTBuilder):
new_terms = []
- from loopy.kernel.data import LocalIndexTag, check_iname_tags, get_iname_tags
+ from loopy.kernel.data import LocalIndexTag, get_iname_tags
from loopy.symbolic import get_dependencies
saw_l0 = False
for term in terms:
if (isinstance(term, Variable)
- and check_iname_tags(kernel.iname_to_tags[term.name], LocalIndexTag)):
- tags = get_iname_tags(kernel.iname_to_tags[term.name], LocalIndexTag)
- if len(tags) > 1:
- raise LoopyError("cannot have more than one LocalIndexTags")
- tag, = tags
+ and get_iname_tags(
+ kernel.iname_to_tags[term.name], LocalIndexTag)):
+ tag, = get_iname_tags(kernel.iname_to_tags[term.name],
+ LocalIndexTag, 1)
if tag.axis == 0:
if saw_l0:
- raise LoopyError("streaming store must have stride 1 "
- "in local index, got: %s" % subscript)
+ raise LoopyError(
+ "streaming store must have stride 1 in "
+ "local index, got: %s" % subscript)
saw_l0 = True
continue
else:
for dep in get_dependencies(term):
- if check_iname_tags(kernel.iname_to_tags[dep], LocalIndexTag):
- tags = get_iname_tags(kernel.iname_to_tags[dep], LocalIndexTag)
- if len(tags) > 1:
- raise LoopyError("cannot have more than one LocalIndexTags")
- tag, = tags
+ if get_iname_tags(
+ kernel.iname_to_tags[dep], LocalIndexTag):
+ tag, = get_iname_tags(kernel.iname_to_tags[dep],
+ LocalIndexTag, 1)
if tag.axis == 0:
- raise LoopyError("streaming store must have stride 1 "
- "in local index, got: %s" % subscript)
+ raise LoopyError(
+ "streaming store must have stride 1 in "
+ "local index, got: %s" % subscript)
new_terms.append(term)
diff --git a/loopy/transform/iname.py b/loopy/transform/iname.py
index 21c2f7eea5054db0ed393dd49c003ae8bcfa7ee6..365f2db773e63b69c75771cb45532c29416e03c4 100644
--- a/loopy/transform/iname.py
+++ b/loopy/transform/iname.py
@@ -33,7 +33,6 @@ from loopy.symbolic import (
RuleAwareIdentityMapper, RuleAwareSubstitutionMapper,
SubstitutionRuleMappingContext)
from loopy.diagnostic import LoopyError
-from loopy.kernel.data import check_iname_tags, get_iname_tags
__doc__ = """
@@ -177,9 +176,9 @@ def _split_iname_backend(kernel, split_iname,
"""
existing_tags = kernel.iname_to_tags[split_iname]
- from loopy.kernel.data import ForceSequentialTag, check_iname_tags
+ from loopy.kernel.data import ForceSequentialTag, get_iname_tags
if (do_tagged_check and existing_tags
- and not check_iname_tags(existing_tags, ForceSequentialTag)):
+ and not get_iname_tags(existing_tags, ForceSequentialTag)):
raise LoopyError("cannot split already tagged iname '%s'" % split_iname)
if split_iname not in kernel.all_inames():
@@ -648,8 +647,8 @@ def tag_inames(kernel, iname_to_tag, force=False, ignore_nonexistent=False):
iname_to_tag = [(iname, parse_tag(tag)) for iname, tag in iname_to_tag]
- from loopy.kernel.data import (ConcurrentTag, AutoLocalIndexTagBase,
- ForceSequentialTag)
+ from loopy.kernel.data import (ConcurrentTag, ForceSequentialTag,
+ get_iname_tags)
# {{{ globbing
@@ -680,27 +679,27 @@ def tag_inames(kernel, iname_to_tag, force=False, ignore_nonexistent=False):
knl_iname_to_tags = kernel.iname_to_tags.copy()
for iname, new_tag in six.iteritems(iname_to_tag):
+ if not new_tag:
+ continue
+
old_tags = kernel.iname_to_tags[iname]
if iname not in kernel.all_inames():
raise ValueError("cannot tag '%s'--not known" % iname)
- if isinstance(new_tag, ConcurrentTag) \
- and check_iname_tags(old_tags, ForceSequentialTag):
+ if (isinstance(new_tag, ConcurrentTag)
+ and get_iname_tags(old_tags, ForceSequentialTag)):
raise ValueError("cannot tag '%s' as parallel--"
"iname requires sequential execution" % iname)
- if isinstance(new_tag, ForceSequentialTag) \
- and check_iname_tags(old_tags, ConcurrentTag):
+ if (isinstance(new_tag, ForceSequentialTag)
+ and get_iname_tags(old_tags, ConcurrentTag)):
raise ValueError("'%s' is already tagged as parallel, "
"but is now prohibited from being parallel "
"(likely because of participation in a precompute or "
"a reduction)" % iname)
- if new_tag and all(tag.key != new_tag.key for tag in old_tags):
- old_tags = old_tags + (new_tag,)
-
- knl_iname_to_tags[iname] = old_tags
+ knl_iname_to_tags[iname] = old_tags.union([new_tag])
return kernel.copy(iname_to_tags=knl_iname_to_tags)
@@ -982,9 +981,9 @@ def get_iname_duplication_options(knl, use_boostable_into=False):
# Get the duplication options as a tuple of iname and a set
for iname, insns in _get_iname_duplication_options(insn_iname_sets):
# Check whether this iname has a parallel tag and discard it if so
- from loopy.kernel.data import ConcurrentTag, check_iname_tags
+ from loopy.kernel.data import ConcurrentTag, get_iname_tags
if (iname in knl.iname_to_tags
- and check_iname_tags(knl.iname_to_tags[iname], ConcurrentTag)):
+ and get_iname_tags(knl.iname_to_tags[iname], ConcurrentTag)):
continue
# If we find a duplication option and to not use boostable_into
@@ -1501,7 +1500,7 @@ def find_unused_axis_tag(kernel, kind, insn_match=None):
"""
used_axes = set()
- from loopy.kernel.data import GroupIndexTag, LocalIndexTag, check_iname_tags
+ from loopy.kernel.data import GroupIndexTag, LocalIndexTag, get_iname_tags
if isinstance(kind, str):
found = False
@@ -1521,7 +1520,7 @@ def find_unused_axis_tag(kernel, kind, insn_match=None):
for insn in insns:
for iname in kernel.insn_inames(insn):
dim_tags = kernel.iname_to_tags[iname]
- if check_iname_tags(dim_tags, kind):
+ if get_iname_tags(dim_tags, kind):
used_axes.add(kind.axis)
i = 0
diff --git a/loopy/transform/privatize.py b/loopy/transform/privatize.py
index d1c112eca6a44cdd0bbd2b2826c291cd52ba8a0a..4810784039899a471bbd23c4922438e75275db25 100644
--- a/loopy/transform/privatize.py
+++ b/loopy/transform/privatize.py
@@ -41,7 +41,6 @@ __doc__ = """
# {{{ privatize temporaries with iname
from loopy.symbolic import IdentityMapper
-from loopy.kernel.data import check_iname_tags, get_iname_tags
class ExtraInameIndexInserter(IdentityMapper):
@@ -85,6 +84,7 @@ def privatize_temporaries_with_inames(
Example::
+<<<<<<< HEAD:loopy/transform/privatize.py
<<<<<<< HEAD:loopy/transform/privatize.py
for imatrix, i
acc = 0
@@ -95,6 +95,9 @@ def privatize_temporaries_with_inames(
=======
from loopy.kernel.data import IlpBaseTag, VectorizeTag, check_iname_tags
>>>>>>> d4c1d2e... change tags from set to tuple:loopy/transform/ilp.py
+=======
+ from loopy.kernel.data import IlpBaseTag, VectorizeTag, get_iname_tags
+>>>>>>> 38a4424... change tags from tuple to set:loopy/transform/ilp.py
might become::
@@ -192,7 +195,7 @@ def privatize_temporaries_with_inames(
dim_tags = ["c"] * (len(shape) + len(extra_shape))
for i, iname in enumerate(inames):
- if check_iname_tags(kernel.iname_to_tags[iname], VectorizeTag):
+ if get_iname_tags(kernel.iname_to_tags[iname], VectorizeTag):
dim_tags[len(shape) + i] = "vec"
new_temp_vars[tv.name] = tv.copy(shape=shape + extra_shape,
diff --git a/loopy/transform/save.py b/loopy/transform/save.py
index 962a83cd1098b48c7bcf21190609209dc79904b2..cd4c10272651128c4fe311c9e4af66798fa448df 100644
--- a/loopy/transform/save.py
+++ b/loopy/transform/save.py
@@ -403,26 +403,20 @@ class TemporarySaver(object):
continue
from loopy.kernel.data import (GroupIndexTag, LocalIndexTag,
- ConcurrentTag, get_iname_tags, check_iname_tags)
+ ConcurrentTag, get_iname_tags)
- if check_iname_tags(tags, GroupIndexTag):
- tags = get_iname_tags(tags, GroupIndexTag)
- if len(tags) > 1:
- raise LoopyError("cannot have more than one GroupIndexTags")
- tag, = tags
+ if get_iname_tags(tags, GroupIndexTag):
+ tag, = get_iname_tags(tags, GroupIndexTag, 1)
my_group_tags.append(tag)
- elif check_iname_tags(tags, LocalIndexTag):
- tags = get_iname_tags(tags, LocalIndexTag)
- if len(tags) > 1:
- raise LoopyError("cannot have more than one LocalIndexTags")
- tag, = tags
+ elif get_iname_tags(tags, LocalIndexTag):
+ tag, = get_iname_tags(tags, LocalIndexTag, 1)
my_local_tags.append(tag)
- elif check_iname_tags(tags, ConcurrentTag):
+ elif get_iname_tags(tags, ConcurrentTag):
raise LoopyError(
"iname '%s' is tagged with '%s' - only "
"group and local tags are supported for "
"auto save/reload of temporaries" %
- (iname, ", ".join(str(tag) for tag in tags)))
+ (iname, tags))
if group_tags is None:
group_tags = _sortedtags(my_group_tags)