diff --git a/doc/reference.rst b/doc/reference.rst
index c829462d37855b54bac993174c37a2faef2a1dbf..bc8b9c484b2b6e28c13ad82c6df8c01316c38bdc 100644
--- a/doc/reference.rst
+++ b/doc/reference.rst
@@ -296,6 +296,9 @@ These are usually key-value pairs. The following attributes are recognized:
variables (which must be scalar variables) evaluate to ``true`` (as
defined by C).
+* ``tags=tag1:tag2`` Apply tags to this instruction that can then be used
+ for :ref:`context-matching`.
+
.. autoclass:: ExpressionInstruction
@@ -334,9 +337,13 @@ function, which is responsible for creating kernels:
Transforming Kernels
--------------------
+.. _context-matching:
+
Matching contexts
^^^^^^^^^^^^^^^^^
+TODO: Matching instruction tags
+
.. automodule:: loopy.context_matching
.. autofunction:: parse_id_match
diff --git a/loopy/context_matching.py b/loopy/context_matching.py
index a80aa04c9cc85c7bc6e925c131d836c48a0273de..eb85c1dd0a4c0b7b55607ec5358d9e3ef18470cb 100644
--- a/loopy/context_matching.py
+++ b/loopy/context_matching.py
@@ -27,6 +27,8 @@ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
THE SOFTWARE.
"""
+import types
+
# {{{ id match objects
@@ -40,26 +42,30 @@ class RegexIdentifierMatch(object):
self.id_re = id_re
self.tag_re = tag_re
- def __call__(self, identifier, tag):
+ def __call__(self, identifier, tags):
+ assert isinstance(tags, (tuple, types.NoneType))
+
if self.tag_re is None:
return self.id_re.match(identifier) is not None
else:
- if tag is None:
- tag = ""
+ if not tags:
+ tags = ("",)
return (
self.id_re.match(identifier) is not None
- and self.tag_re.match(tag) is not None)
+ and any(
+ self.tag_re.match(tag) is not None
+ for tag in tags))
class AlternativeMatch(object):
def __init__(self, matches):
self.matches = matches
- def __call__(self, identifier, tag):
+ def __call__(self, identifier, tags):
from pytools import any
return any(
- mtch(identifier, tag) for mtch in self.matches)
+ mtch(identifier, tags) for mtch in self.matches)
# }}}
@@ -206,5 +212,4 @@ def parse_stack_match(smatch):
# }}}
-
# vim: foldmethod=marker
diff --git a/loopy/kernel/creation.py b/loopy/kernel/creation.py
index b932d448a47cd328755c20b7e567642276f1565a..0727000f9d1b5cda84359c0d7ac3e49e5694f028 100644
--- a/loopy/kernel/creation.py
+++ b/loopy/kernel/creation.py
@@ -198,6 +198,7 @@ def parse_insn(insn):
forced_iname_deps_is_final = False
forced_iname_deps = frozenset()
predicates = frozenset()
+ tags = ()
if groups["options"] is not None:
for option in groups["options"].split(","):
@@ -244,8 +245,15 @@ def parse_insn(insn):
forced_iname_deps_is_final = True
forced_iname_deps = frozenset(opt_value.split(":"))
+
elif opt_key == "if":
predicates = frozenset(opt_value.split(":"))
+
+ elif opt_key == "tags":
+ tags = tuple(
+ tag.strip() for tag in opt_value.split(":")
+ if tag.strip())
+
else:
raise ValueError("unrecognized instruction option '%s'"
% opt_key)
@@ -273,7 +281,8 @@ def parse_insn(insn):
assignee=lhs, expression=rhs,
temp_var_type=temp_var_type,
priority=priority,
- predicates=predicates), inames_to_dup
+ predicates=predicates,
+ tags=tags), inames_to_dup
elif subst_match is not None:
from pymbolic.primitives import Variable, Call
@@ -450,16 +459,16 @@ class ArgumentGuesser:
(assignee_var_name, _), = insn.assignees_and_indices()
self.all_written_names.add(assignee_var_name)
self.all_names.update(get_dependencies(
- self.submap(insn.assignee, insn.id)))
+ self.submap(insn.assignee, insn.id, insn.tags)))
self.all_names.update(get_dependencies(
- self.submap(insn.expression, insn.id)))
+ self.submap(insn.expression, insn.id, insn.tags)))
def find_index_rank(self, name):
irf = IndexRankFinder(name)
for insn in self.instructions:
insn.with_transformed_expressions(
- lambda expr: irf(self.submap(expr, insn.id)))
+ lambda expr: irf(self.submap(expr, insn.id, insn.tags)))
if not irf.index_ranks:
return 0
@@ -860,9 +869,9 @@ def guess_arg_shape_if_requested(kernel, default_order):
try:
for insn in kernel.instructions:
if isinstance(insn, lp.ExpressionInstruction):
- armap(submap(insn.assignee, insn.id),
+ armap(submap(insn.assignee, insn.id, insn.tags),
kernel.insn_inames(insn))
- armap(submap(insn.expression, insn.id),
+ armap(submap(insn.expression, insn.id, insn.tags),
kernel.insn_inames(insn))
except TypeError as e:
from loopy.diagnostic import LoopyError
diff --git a/loopy/kernel/data.py b/loopy/kernel/data.py
index a7e4720147b4f6e8d7ab4afbd6b83db2dc339bb2..a563b9ef39560dacfa77943e037d4a25b2213c15 100644
--- a/loopy/kernel/data.py
+++ b/loopy/kernel/data.py
@@ -457,6 +457,11 @@ class InstructionBase(Record):
A :class:`set` of inames into which the instruction
may need to be boosted, as a heuristic help for the scheduler.
Also allowed to be *None*.
+
+ .. attribute:: tags
+
+ A tuple of string identifiers that can be used to identify groups
+ of statements.
"""
fields = set("id insn_deps insn_deps_is_final predicates "
@@ -465,7 +470,7 @@ class InstructionBase(Record):
def __init__(self, id, insn_deps, insn_deps_is_final,
forced_iname_deps_is_final, forced_iname_deps, priority,
- boostable, boostable_into, predicates):
+ boostable, boostable_into, predicates, tags):
if forced_iname_deps_is_final is None:
forced_iname_deps_is_final = False
@@ -473,6 +478,9 @@ class InstructionBase(Record):
if insn_deps_is_final is None:
insn_deps_is_final = False
+ if tags is None:
+ tags = ()
+
assert isinstance(forced_iname_deps, frozenset)
assert isinstance(insn_deps, frozenset) or insn_deps is None
@@ -485,7 +493,8 @@ class InstructionBase(Record):
priority=priority,
boostable=boostable,
boostable_into=boostable_into,
- predicates=predicates)
+ predicates=predicates,
+ tags=tags)
# {{{ abstract interface
@@ -551,6 +560,8 @@ class InstructionBase(Record):
result.append("deps="+":".join(self.insn_deps))
if self.priority:
result.append("priority=%d" % self.priority)
+ if self.tags:
+ result.append("tags=%s" % ":".join(self.tags))
return result
@@ -629,7 +640,7 @@ class ExpressionInstruction(InstructionBase):
forced_iname_deps=frozenset(),
insn_deps=None,
insn_deps_is_final=None,
- boostable=None, boostable_into=None,
+ boostable=None, boostable_into=None, tags=None,
temp_var_type=None, priority=0, predicates=frozenset()):
InstructionBase.__init__(self,
@@ -641,7 +652,8 @@ class ExpressionInstruction(InstructionBase):
boostable=boostable,
boostable_into=boostable_into,
priority=priority,
- predicates=predicates)
+ predicates=predicates,
+ tags=tags)
from loopy.symbolic import parse
if isinstance(assignee, str):
@@ -795,7 +807,7 @@ class CInstruction(InstructionBase):
id=None, insn_deps=None, insn_deps_is_final=None,
forced_iname_deps_is_final=None, forced_iname_deps=frozenset(),
priority=0, boostable=None, boostable_into=None,
- predicates=frozenset()):
+ predicates=frozenset(), tags=None):
"""
:arg iname_exprs: Like :attr:`iname_exprs`, but instead of tuples,
simple strings pepresenting inames are also allowed. A single
@@ -814,7 +826,7 @@ class CInstruction(InstructionBase):
insn_deps_is_final=insn_deps_is_final,
boostable=boostable,
boostable_into=boostable_into,
- priority=priority, predicates=predicates)
+ priority=priority, predicates=predicates, tags=tags)
# {{{ normalize iname_exprs
@@ -914,6 +926,7 @@ class CInstruction(InstructionBase):
key_builder.update_for_pymbolic_expression(key_hash, val)
else:
key_builder.rec(key_hash, getattr(self, field_name))
+
# }}}
# }}}
diff --git a/loopy/precompute.py b/loopy/precompute.py
index 485c9c954552bc42da237021556abfbee3ac9f02..21e6e5ddf394301ce9fcf9c54c682adf1dee3f60 100644
--- a/loopy/precompute.py
+++ b/loopy/precompute.py
@@ -407,7 +407,8 @@ class InvocationGatherer(ExpandingIdentityMapper):
arg_deps = set()
for arg_val in six.itervalues(arg_context):
arg_deps = (arg_deps
- | get_dependencies(self.subst_expander(arg_val, insn_id=None)))
+ | get_dependencies(self.subst_expander(
+ arg_val, insn_id=None, insn_tags=None)))
if not arg_deps <= self.kernel.all_inames():
from warnings import warn
@@ -680,7 +681,7 @@ def precompute(kernel, subst_use, sweep_inames=[], within=None,
import loopy as lp
for insn in kernel.instructions:
if isinstance(insn, lp.ExpressionInstruction):
- invg(insn.expression, insn.id)
+ invg(insn.expression, insn.id, insn.tags)
for invdesc in invg.invocation_descriptors:
invocation_descriptors.append(
@@ -717,7 +718,8 @@ def precompute(kernel, subst_use, sweep_inames=[], within=None,
submap = SubstitutionRuleExpander(kernel.substitutions)
value_inames = get_dependencies(
- submap(subst.expression, insn_id=None)) & kernel.all_inames()
+ submap(subst.expression, insn_id=None, insn_tags=None)
+ ) & kernel.all_inames()
if value_inames - expanding_usage_arg_deps < extra_storage_axes:
raise RuntimeError("unreferenced sweep inames specified: "
+ ", ".join(extra_storage_axes
diff --git a/loopy/preprocess.py b/loopy/preprocess.py
index 0adea202a1a9319c64a13e85f819a60492b9158a..8edddc2bdfa78cec6674e702661b1d6f6f1a4057 100644
--- a/loopy/preprocess.py
+++ b/loopy/preprocess.py
@@ -62,7 +62,8 @@ def _infer_var_type(kernel, var_name, type_inf_mapper, subst_expander):
if not isinstance(writer_insn, lp.ExpressionInstruction):
continue
- expr = subst_expander(writer_insn.expression, insn_id=writer_insn_id)
+ expr = subst_expander(writer_insn.expression,
+ insn_id=writer_insn_id, insn_tags=writer_insn.tags)
try:
debug(" via expr %s" % expr)
diff --git a/loopy/symbolic.py b/loopy/symbolic.py
index aec1c15efc3c5079c793bfb5e1aa0d05470f2977..3b3eba49e8cec0d62fd2a57bcb19d795dcb40167 100644
--- a/loopy/symbolic.py
+++ b/loopy/symbolic.py
@@ -429,8 +429,14 @@ class ExpandingIdentityMapper(IdentityMapper):
rule = self.old_subst_rules[name]
rec_arguments = self.rec(arguments, expn_state)
+
+ if tag is None:
+ tags = None
+ else:
+ tags = (tag,)
+
new_expn_state = expn_state.copy(
- stack=expn_state.stack + ((name, tag),),
+ stack=expn_state.stack + ((name, tags),),
arg_context=self.make_new_arg_context(
name, rule.arguments, rec_arguments, expn_state.arg_context))
@@ -448,9 +454,9 @@ class ExpandingIdentityMapper(IdentityMapper):
else:
return sym
- def __call__(self, expr, insn_id):
+ def __call__(self, expr, insn_id, insn_tags):
if insn_id is not None:
- stack = ((insn_id, None),)
+ stack = ((insn_id, insn_tags),)
else:
stack = ()
@@ -520,7 +526,7 @@ class ExpandingIdentityMapper(IdentityMapper):
# may perform tasks entirely unrelated to subst rules, so
# we must map assignees, too.
- insn.with_transformed_expressions(self, insn.id)
+ insn.with_transformed_expressions(self, insn.id, insn.tags)
for insn in kernel.instructions]
new_substs, renames = self._get_new_substitutions_and_renames()
@@ -564,7 +570,13 @@ class SubstitutionRuleExpander(ExpandingIdentityMapper):
self.ctx_match = ctx_match
def map_substitution(self, name, tag, arguments, expn_state):
- new_stack = expn_state.stack + ((name, tag),)
+ if tag is None:
+ tags = None
+ else:
+ tags = (tag,)
+
+ new_stack = expn_state.stack + ((name, tags),)
+
if self.ctx_match(new_stack):
# expand
rule = self.old_subst_rules[name]