diff --git a/doc/reference.rst b/doc/reference.rst
index ac443cf624c0f94871e1c481c54b852fe2f6fa14..a1cb0efab031476bd469cbacea8b7112603e37d9 100644
--- a/doc/reference.rst
+++ b/doc/reference.rst
@@ -176,7 +176,18 @@ These are usually key-value pairs. The following attributes are recognized:
(often numbers) to the given ``id_prefix``.
* ``inames=i:j:k`` forces the instruction to reside within the loops over
- :ref:`inames` ``i``, ``j`` and ``k``.
+ :ref:`inames` ``i``, ``j`` and ``k`` (and only those).
+
+ .. note::
+
+ The default for the inames that the instruction depends on is
+ the inames used in the instruction itself plus the common
+ subset of inames shared by writers of all variables read by the
+ instruction.
+
+ You can add a plus sign ("``+``") to the front of this option
+ value to indicate that you would like the inames you specify here
+ to be in addition to the ones found by the heuristic described above.
* ``dep=id1:id2`` creates a dependency of this instruction on the
instructions with identifiers ``id1`` and ``id2``. This requires that the
@@ -184,7 +195,7 @@ These are usually key-value pairs. The following attributes are recognized:
instructions' generated code.
Identifiers here are allowed to be wildcards as defined by
- :mod:`fnmatchcase`.
+ the Python module :mod:`fnmatchcase`.
.. note::
diff --git a/loopy/kernel/creation.py b/loopy/kernel/creation.py
index acf361dfdf0c5bc8a7d5d3fe78b28442f4a5c8bf..252a588b2528545b5d94178b937ecf621146060b 100644
--- a/loopy/kernel/creation.py
+++ b/loopy/kernel/creation.py
@@ -137,6 +137,7 @@ def expand_defines_in_expr(expr, defines):
# }}}
+
# {{{ parse instructions
INSN_RE = re.compile(
@@ -181,6 +182,7 @@ def parse_insn(insn):
insn_deps = None
insn_id = None
priority = 0
+ forced_iname_deps_is_final = False
forced_iname_deps = frozenset()
predicates = frozenset()
@@ -208,6 +210,12 @@ def parse_insn(insn):
insn_deps = frozenset(dep.strip() for dep in opt_value.split(":")
if dep.strip())
elif opt_key == "inames":
+ if opt_value.startswith("+"):
+ forced_iname_deps_is_final = False
+ opt_value = (opt_value[1:]).strip()
+ else:
+ forced_iname_deps_is_final = True
+
forced_iname_deps = frozenset(opt_value.split(":"))
elif opt_key == "if":
predicates = frozenset(opt_value.split(":"))
@@ -232,6 +240,7 @@ def parse_insn(insn):
return ExpressionInstruction(
id=insn_id,
insn_deps=insn_deps,
+ forced_iname_deps_is_final=forced_iname_deps_is_final,
forced_iname_deps=forced_iname_deps,
assignee=lhs, expression=rhs,
temp_var_type=temp_var_type,
@@ -288,6 +297,7 @@ def parse_if_necessary(insn, defines):
# }}}
+
# {{{ domain parsing
EMPTY_SET_DIMS_RE = re.compile(r"^\s*\{\s*\:")
diff --git a/loopy/kernel/data.py b/loopy/kernel/data.py
index df9b7f3ad7d871eac301ac296034045278df287c..9e9538b2ee47e155309a5bdb6e88b68931262c8c 100644
--- a/loopy/kernel/data.py
+++ b/loopy/kernel/data.py
@@ -356,10 +356,16 @@ class InstructionBase(Record):
a :class:`frozenset` of variable names whose truth values (as defined
by C) determine whether this instruction should be run
+ .. attribute:: forced_iname_deps_is_final
+
+ A :class:`bool` determining whether :attr:`forced_iname_deps` constitutes
+ the *entire* list of iname dependencies.
+
.. attribute:: forced_iname_deps
A :class:`frozenset` of inames that are added to the list of iname
- dependencies.
+ dependencies *or* constitute the entire list of iname dependencies,
+ depending on the value of :attr:`forced_iname_deps_is_final`.
.. attribute:: priority
@@ -379,18 +385,24 @@ class InstructionBase(Record):
Also allowed to be *None*.
"""
- fields = set("id insn_deps predicates forced_iname_deps "
+ fields = set("id insn_deps predicates "
+ "forced_iname_deps_is_final forced_iname_deps "
"priority boostable boostable_into".split())
- def __init__(self, id, insn_deps, forced_iname_deps, priority,
+ def __init__(self, id, insn_deps,
+ forced_iname_deps_is_final, forced_iname_deps, priority,
boostable, boostable_into, predicates):
+ if forced_iname_deps_is_final is None:
+ forced_iname_deps_is_final = False
+
assert isinstance(forced_iname_deps, frozenset)
assert isinstance(insn_deps, frozenset) or insn_deps is None
Record.__init__(self,
id=id,
insn_deps=insn_deps,
+ forced_iname_deps_is_final=forced_iname_deps_is_final,
forced_iname_deps=forced_iname_deps,
priority=priority,
boostable=boostable,
@@ -502,12 +514,16 @@ class ExpressionInstruction(InstructionBase):
def __init__(self,
assignee, expression,
- id=None, forced_iname_deps=frozenset(), insn_deps=None,
+ id=None,
+ forced_iname_deps_is_final=None,
+ forced_iname_deps=frozenset(),
+ insn_deps=None,
boostable=None, boostable_into=None,
temp_var_type=None, priority=0, predicates=frozenset()):
InstructionBase.__init__(self,
id=id,
+ forced_iname_deps_is_final=forced_iname_deps_is_final,
forced_iname_deps=forced_iname_deps,
insn_deps=insn_deps, boostable=boostable,
boostable_into=boostable_into,
@@ -642,8 +658,10 @@ class CInstruction(InstructionBase):
def __init__(self,
iname_exprs, code,
read_variables=frozenset(), assignees=frozenset(),
- id=None, insn_deps=None, forced_iname_deps=frozenset(), priority=0,
- boostable=None, boostable_into=None, predicates=frozenset()):
+ id=None, insn_deps=None,
+ forced_iname_deps_is_final=None, forced_iname_deps=frozenset(),
+ priority=0, boostable=None, boostable_into=None,
+ predicates=frozenset()):
"""
:arg iname_exprs: Like :attr:`iname_exprs`, but instead of tuples,
simple strings pepresenting inames are also allowed. A single
@@ -656,6 +674,7 @@ class CInstruction(InstructionBase):
InstructionBase.__init__(self,
id=id,
+ forced_iname_deps_is_final=forced_iname_deps_is_final,
forced_iname_deps=forced_iname_deps,
insn_deps=insn_deps, boostable=boostable,
boostable_into=boostable_into,
diff --git a/loopy/kernel/tools.py b/loopy/kernel/tools.py
index b7793143923286941940afe3e70cd2794ce7006a..57ca8890c29a3736da6db49905ebc6894a333504 100644
--- a/loopy/kernel/tools.py
+++ b/loopy/kernel/tools.py
@@ -115,9 +115,12 @@ def find_all_insn_inames(kernel):
all_write_deps[insn.id] = write_deps = insn.write_dependency_names()
deps = read_deps | write_deps
- iname_deps = (
- deps & kernel.all_inames()
- | insn.forced_iname_deps)
+ if insn.forced_iname_deps_is_final:
+ iname_deps = insn.forced_iname_deps
+ else:
+ iname_deps = (
+ deps & kernel.all_inames()
+ | insn.forced_iname_deps)
assert isinstance(read_deps, frozenset), type(insn)
assert isinstance(write_deps, frozenset), type(insn)
@@ -151,6 +154,9 @@ def find_all_insn_inames(kernel):
did_something = False
for insn in kernel.instructions:
+ if insn.forced_iname_deps_is_final:
+ continue
+
# {{{ depdency-based propagation
# For all variables that insn depends on, find the intersection
diff --git a/test/test_loopy.py b/test/test_loopy.py
index 8092e9c6f87a0c40228a427dcbc20511b3531d10..c6b5e6f3eb176ad61add44ff899cfd7fe2470f02 100644
--- a/test/test_loopy.py
+++ b/test/test_loopy.py
@@ -1400,21 +1400,21 @@ def test_rob_stroud_bernstein(ctx_factory):
0 <= i2 < nqp1d and \
0 <= alpha1 <= deg and 0 <= alpha2 <= deg-alpha1 }",
"""
- <> xi = qpts[1, i2] {inames=el}
+ <> xi = qpts[1, i2] {inames=+el}
<> s = 1-xi
<> r = xi/s
- <> aind = 0 {id=aind_init,inames=i2:el}
+ <> aind = 0 {id=aind_init,inames=+i2:el}
<> w = s**(deg-alpha1) {id=init_w}
tmp[el,alpha1,i2] = tmp[el,alpha1,i2] + w * coeffs[aind] \
- {id=write_tmp,inames=alpha2}
+ {id=write_tmp,inames=+alpha2}
w = w * r * ( deg - alpha1 - alpha2 ) / (1 + alpha2) \
{id=update_w,dep=init_w:write_tmp}
aind = aind + 1 \
{id=aind_incr,\
dep=aind_init:write_tmp:update_w, \
- inames=el:i2:alpha1:alpha2}
+ inames=+el:i2:alpha1:alpha2}
""",
[
# Must declare coeffs to have "no" shape, to keep loopy