diff --git a/loopy/codegen/instruction.py b/loopy/codegen/instruction.py
index 6224d9709f5f796f84c3fd177125e0703d92d173..3ef7c8f6ad6c8af09dd01bf9e1341179d2be0be7 100644
--- a/loopy/codegen/instruction.py
+++ b/loopy/codegen/instruction.py
@@ -126,10 +126,13 @@ def generate_assignment_instruction_code(codegen_state, insn):
# }}}
- from pymbolic.primitives import Variable, Subscript
+ from pymbolic.primitives import Variable, Subscript, Lookup
from loopy.symbolic import LinearSubscript
lhs = insn.assignee
+ if isinstance(lhs, Lookup):
+ lhs = lhs.aggregate
+
if isinstance(lhs, Variable):
assignee_var_name = lhs.name
assignee_indices = ()
@@ -145,6 +148,8 @@ def generate_assignment_instruction_code(codegen_state, insn):
else:
raise RuntimeError("invalid lvalue '%s'" % lhs)
+ del lhs
+
result = codegen_state.ast_builder.emit_assignment(codegen_state, insn)
# {{{ tracing
diff --git a/loopy/codegen/loop.py b/loopy/codegen/loop.py
index a334462049634fff1e3137ffd09acd3ef254bb51..0110a06095fa0bd690045f050136027d7bed3a28 100644
--- a/loopy/codegen/loop.py
+++ b/loopy/codegen/loop.py
@@ -465,12 +465,17 @@ def generate_sequential_loop_dim_code(codegen_state, sched_index):
else:
inner_ast = inner.current_ast(codegen_state)
+
+ from loopy.isl_helpers import simplify_pw_aff
+
result.append(
inner.with_new_ast(
codegen_state,
astb.emit_sequential_loop(
codegen_state, loop_iname, kernel.index_dtype,
- pw_aff_to_expr(lbound), pw_aff_to_expr(ubound), inner_ast)))
+ pw_aff_to_expr(simplify_pw_aff(lbound, kernel.assumptions)),
+ pw_aff_to_expr(simplify_pw_aff(ubound, kernel.assumptions)),
+ inner_ast)))
return merge_codegen_results(codegen_state, result)
diff --git a/loopy/isl_helpers.py b/loopy/isl_helpers.py
index 602830de38e457c5ff4a55d7685dc346a7b4de35..0ebe90fbca0d31c05eaee64321e2b73709292331 100644
--- a/loopy/isl_helpers.py
+++ b/loopy/isl_helpers.py
@@ -142,6 +142,55 @@ def iname_rel_aff(space, iname, rel, aff):
raise ValueError("unknown value of 'rel': %s" % rel)
+# {{{ simplify_pw_aff
+
+def simplify_pw_aff(pw_aff, context=None):
+ if context is not None:
+ pw_aff = pw_aff.gist_params(context)
+
+ old_pw_aff = pw_aff
+
+ while True:
+ restart = False
+ did_something = False
+
+ pieces = pw_aff.get_pieces()
+ for i, (dom_i, aff_i) in enumerate(pieces):
+ for j, (dom_j, aff_j) in enumerate(pieces):
+ if i == j:
+ continue
+
+ if aff_i.gist(dom_j).is_equal(aff_j):
+ # aff_i is sufficient to conver aff_j, eliminate aff_j
+ new_pieces = pieces[:]
+ if i < j:
+ new_pieces.pop(j)
+ new_pieces.pop(i)
+ else:
+ new_pieces.pop(i)
+ new_pieces.pop(j)
+
+ pw_aff = isl.PwAff.alloc(dom_i | dom_j, aff_i)
+ for dom, aff in new_pieces:
+ pw_aff = pw_aff.union_max(isl.PwAff.alloc(dom, aff))
+
+ restart = True
+ did_something = True
+ break
+
+ if restart:
+ break
+
+ if not did_something:
+ break
+
+ assert pw_aff.get_aggregate_domain() <= pw_aff.eq_set(old_pw_aff)
+
+ return pw_aff
+
+# }}}
+
+
# {{{ static_*_of_pw_aff
def static_extremum_of_pw_aff(pw_aff, constants_only, set_method, what, context):
diff --git a/loopy/kernel/creation.py b/loopy/kernel/creation.py
index 6eedfcc20e7e59f129c8f19e2d96c07a80714533..14b18150f5b84218f39ba23662eb6106ffb596a0 100644
--- a/loopy/kernel/creation.py
+++ b/loopy/kernel/creation.py
@@ -448,7 +448,7 @@ def parse_insn(groups, insn_options):
"the following error occurred:" % groups["rhs"])
raise
- from pymbolic.primitives import Variable, Subscript
+ from pymbolic.primitives import Variable, Subscript, Lookup
from loopy.symbolic import TypeAnnotation
if not isinstance(lhs, tuple):
@@ -469,11 +469,15 @@ def parse_insn(groups, insn_options):
else:
temp_var_types.append(None)
+ inner_lhs_i = lhs_i
+ if isinstance(inner_lhs_i, Lookup):
+ inner_lhs_i = inner_lhs_i.aggregate
+
from loopy.symbolic import LinearSubscript
- if isinstance(lhs_i, Variable):
- assignee_names.append(lhs_i.name)
- elif isinstance(lhs_i, (Subscript, LinearSubscript)):
- assignee_names.append(lhs_i.aggregate.name)
+ if isinstance(inner_lhs_i, Variable):
+ assignee_names.append(inner_lhs_i.name)
+ elif isinstance(inner_lhs_i, (Subscript, LinearSubscript)):
+ assignee_names.append(inner_lhs_i.aggregate.name)
else:
raise LoopyError("left hand side of assignment '%s' must "
"be variable or subscript" % (lhs_i,))
@@ -1638,11 +1642,11 @@ def _resolve_dependencies(knl, insn, deps):
new_deps.append(other_insn.id)
found_any = True
- if not found_any:
+ if not found_any and knl.options.check_dep_resolution:
raise LoopyError("instruction '%s' declared a depency on '%s', "
"which did not resolve to any instruction present in the "
- "kernel '%s'"
- % (insn.id, dep, knl.name))
+ "kernel '%s'. Set the kernel option 'check_dep_resolution'"
+ "to False to disable this check." % (insn.id, dep, knl.name))
for dep_id in new_deps:
if dep_id not in knl.id_to_insn:
diff --git a/loopy/kernel/instruction.py b/loopy/kernel/instruction.py
index fdd8f1d3764ec03ca40a8338dc512b8cd2ae38cf..0d22dbb88ed99c7c92480d1d39b924cc2198cc3f 100644
--- a/loopy/kernel/instruction.py
+++ b/loopy/kernel/instruction.py
@@ -455,9 +455,12 @@ class InstructionBase(ImmutableRecord):
def _get_assignee_var_name(expr):
- from pymbolic.primitives import Variable, Subscript
+ from pymbolic.primitives import Variable, Subscript, Lookup
from loopy.symbolic import LinearSubscript
+ if isinstance(expr, Lookup):
+ expr = expr.aggregate
+
if isinstance(expr, Variable):
return expr.name
@@ -477,9 +480,12 @@ def _get_assignee_var_name(expr):
def _get_assignee_subscript_deps(expr):
- from pymbolic.primitives import Variable, Subscript
+ from pymbolic.primitives import Variable, Subscript, Lookup
from loopy.symbolic import LinearSubscript, get_dependencies
+ if isinstance(expr, Lookup):
+ expr = expr.aggregate
+
if isinstance(expr, Variable):
return frozenset()
elif isinstance(expr, Subscript):
@@ -770,9 +776,9 @@ class Assignment(MultiAssignmentBase):
if isinstance(expression, str):
expression = parse(expression)
- from pymbolic.primitives import Variable, Subscript
+ from pymbolic.primitives import Variable, Subscript, Lookup
from loopy.symbolic import LinearSubscript
- if not isinstance(assignee, (Variable, Subscript, LinearSubscript)):
+ if not isinstance(assignee, (Variable, Subscript, LinearSubscript, Lookup)):
raise LoopyError("invalid lvalue '%s'" % assignee)
self.assignee = assignee
@@ -993,6 +999,11 @@ class CallInstruction(MultiAssignmentBase):
if field_name in ["assignees", "expression"]:
key_builder.update_for_pymbolic_expression(
key_hash, getattr(self, field_name))
+ elif field_name == "predicates":
+ preds = sorted(self.predicates, key=str)
+ for pred in preds:
+ key_builder.update_for_pymbolic_expression(
+ key_hash, pred)
else:
key_builder.rec(key_hash, getattr(self, field_name))
diff --git a/loopy/kernel/tools.py b/loopy/kernel/tools.py
index 539bfbed06572b07491c215770a0330963764d1d..2033425236836ecf000d6c341c46dcb8b087a29a 100644
--- a/loopy/kernel/tools.py
+++ b/loopy/kernel/tools.py
@@ -1341,17 +1341,30 @@ def draw_dependencies_as_unicode_arrows(
uniform_length = min(n_columns[0], max_columns)
+ added_ellipsis = [False]
+
def conform_to_uniform_length(s):
if len(s) <= uniform_length:
return s + " "*(uniform_length-len(s))
else:
- return s[:uniform_length] + "..."
+ added_ellipsis[0] = True
+ return s[:uniform_length] + u"…"
- return [
+ rows = [
(conform_to_uniform_length(row),
conform_to_uniform_length(extender))
for row, extender in rows]
+ if added_ellipsis[0]:
+ uniform_length += 1
+
+ rows = [
+ (conform_to_uniform_length(row),
+ conform_to_uniform_length(extender))
+ for row, extender in rows]
+
+ return rows
+
# }}}
# vim: foldmethod=marker
diff --git a/loopy/options.py b/loopy/options.py
index 7c778681dced904b31e0cc39cff529b8c026640d..25bb7014ce07a30c49f7f78d5a6325eaba36291d 100644
--- a/loopy/options.py
+++ b/loopy/options.py
@@ -95,6 +95,11 @@ class Options(ImmutableRecord):
determining whether an iname duplication is necessary
for the kernel to be schedulable.
+ .. attribute:: check_dep_resolution
+
+ Whether loopy should issue an error if a dependency
+ expression does not match any instructions in the kernel.
+
.. rubric:: Invocation-related options
.. attribute:: skip_arg_checks
@@ -200,6 +205,7 @@ class Options(ImmutableRecord):
allow_terminal_colors_def),
disable_global_barriers=kwargs.get("disable_global_barriers",
False),
+ check_dep_resolution=kwargs.get("check_dep_resolution", True),
)
# {{{ legacy compatibility
diff --git a/loopy/preprocess.py b/loopy/preprocess.py
index 2b6d97c38a12b47e5b4653297c18b24c40ed938b..0d8e771954cf26cc11747e745946389420fa5e1b 100644
--- a/loopy/preprocess.py
+++ b/loopy/preprocess.py
@@ -539,7 +539,7 @@ def realize_reduction(kernel, insn_id_filter=None, unknown_types_ok=True):
new_insn_add_depends_on.add(prev_id)
new_insn_add_no_sync_with.add((prev_id, "any"))
- new_insn_add_within_inames.add(stage_exec_iname or base_exec_iname)
+ new_insn_add_within_inames.add(base_exec_iname or stage_exec_iname)
if nresults == 1:
assert len(acc_vars) == 1
diff --git a/loopy/transform/batch.py b/loopy/transform/batch.py
index 88e3898e2cceeeb62edea306283fcb718c3b088d..e7a86300f9d040cba1688e5bb0f3dcbbd926f783 100644
--- a/loopy/transform/batch.py
+++ b/loopy/transform/batch.py
@@ -63,7 +63,7 @@ class _BatchVariableChanger(RuleAwareIdentityMapper):
if not self.needs_batch_subscript(expr.aggregate.name):
return super(_BatchVariableChanger, self).map_subscript(expr, expn_state)
- idx = expr.index
+ idx = self.rec(expr.index, expn_state)
if not isinstance(idx, tuple):
idx = (idx,)
@@ -73,7 +73,7 @@ class _BatchVariableChanger(RuleAwareIdentityMapper):
if not self.needs_batch_subscript(expr.name):
return super(_BatchVariableChanger, self).map_variable(expr, expn_state)
- return expr.aggregate[self.batch_iname_expr]
+ return expr[self.batch_iname_expr]
def _add_unique_dim_name(name, dim_names):
diff --git a/loopy/transform/iname.py b/loopy/transform/iname.py
index c35b5064365293ac78cdd01af537c9d28bd67193..ea90abfe27c8de69daf39021b3d0ea5463a2e4c8 100644
--- a/loopy/transform/iname.py
+++ b/loopy/transform/iname.py
@@ -986,8 +986,9 @@ def get_iname_duplication_options(knl, use_boostable_into=False):
# Emit a warning that we needed boostable_into
from warnings import warn
from loopy.diagnostic import LoopyWarning
- warn("Kernel '%s' required the deprecated 'boostable_into"
- "field in order to be schedulable!" % knl.name, LoopyWarning)
+ warn("Kernel '%s' required the deprecated 'boostable_into' "
+ "instruction attribute in order to be schedulable!" % knl.name,
+ LoopyWarning)
# Return to avoid yielding the duplication
# options without boostable_into
@@ -1198,7 +1199,8 @@ class _ReductionSplitter(RuleAwareIdentityMapper):
return Reduction(expr.operation, tuple(self.inames),
Reduction(expr.operation, tuple(leftover_inames),
self.rec(expr.expr, expn_state),
- expr.allow_simultaneous))
+ expr.allow_simultaneous),
+ expr.allow_simultaneous)
else:
assert False
else:
diff --git a/loopy/version.py b/loopy/version.py
index 5c6ad47f8571ceb9100f4f7f8dece9d80a35d10c..77d0e21bdd2ef5383c5f874656c25fe1ede21a70 100644
--- a/loopy/version.py
+++ b/loopy/version.py
@@ -32,4 +32,4 @@ except ImportError:
else:
_islpy_version = islpy.version.VERSION_TEXT
-DATA_MODEL_VERSION = "v59-islpy%s" % _islpy_version
+DATA_MODEL_VERSION = "v60-islpy%s" % _islpy_version
diff --git a/test/test_loopy.py b/test/test_loopy.py
index ec6dd5d3fc802918297942a2c907af9df615f239..82994c386062cf4eb4bf691f8464c56ae0d398c0 100644
--- a/test/test_loopy.py
+++ b/test/test_loopy.py
@@ -2152,6 +2152,41 @@ def test_global_barrier_error_if_unordered():
knl.global_barrier_order
+def test_struct_assignment(ctx_factory):
+ ctx = ctx_factory()
+ queue = cl.CommandQueue(ctx)
+
+ bbhit = np.dtype([
+ ("tmin", np.float32),
+ ("tmax", np.float32),
+ ("bi", np.int32),
+ ("hit", np.int32)])
+
+ bbhit, bbhit_c_decl = cl.tools.match_dtype_to_c_struct(
+ ctx.devices[0], "bbhit", bbhit)
+ bbhit = cl.tools.get_or_register_dtype('bbhit', bbhit)
+
+ preamble = bbhit_c_decl
+
+ knl = lp.make_kernel(
+ "{ [i]: 0<=i<N }",
+ """
+ for i
+ result[i].hit = i % 2
+ result[i].tmin = i
+ result[i].tmax = i+10
+ result[i].bi = i
+ end
+ """,
+ [
+ lp.GlobalArg("result", shape=("N",), dtype=bbhit),
+ "..."],
+ preambles=[("000", preamble)])
+
+ knl = lp.set_options(knl, write_cl=True)
+ knl(queue, N=200)
+
+
if __name__ == "__main__":
if len(sys.argv) > 1:
exec(sys.argv[1])
diff --git a/test/test_reduction.py b/test/test_reduction.py
index 5887df7a628c46fbf09539fdd48c08aaacd8e409..86e72c0c6644b7b9837a6d74da756c58344b1d6f 100644
--- a/test/test_reduction.py
+++ b/test/test_reduction.py
@@ -181,7 +181,7 @@ def test_recursive_nested_dependent_reduction(ctx_factory):
# FIXME: Actually test functionality.
-@pytest.mark.parametrize("size", [128, 5, 113, 67])
+@pytest.mark.parametrize("size", [128, 5, 113, 67, 1])
def test_local_parallel_reduction(ctx_factory, size):
ctx = ctx_factory()