diff --git a/MEMO b/MEMO
index 05eaf2de75a810bd9ff726646ac8ae53d56bd382..c13497a091cdbcd4e70448f9277e56bceadcb977 100644
--- a/MEMO
+++ b/MEMO
@@ -36,19 +36,17 @@ Things to consider
- Loopy as a data model for implementing custom rewritings
+- We won't generate WAWs barrier-needing dependencies
+ from one instruction to itself.
+
To-do
^^^^^
- Just touching a variable written to by a non-idempotent
instruction makes that instruction also not idempotent
-- assert dependencies <= parent_inames in loopy/__init__.py
- ???
-
- user interface for dim length prescription
-- Way too many barriers in SEM test.
-
- Deal with equality constraints.
(These arise, e.g., when partitioning a loop of length 16 into 16s.)
@@ -91,6 +89,9 @@ Future ideas
Dealt with
^^^^^^^^^^
+- assert dependencies <= parent_inames in loopy/__init__.py
+ -> Yes, this must be the case.
+
- Give a good error message if a parameter assignment in get_problems()
is missing.
diff --git a/loopy/__init__.py b/loopy/__init__.py
index 46f6a76555afcf51c65cb4b2fcc7d7283fbf397a..f0fe3f0fda31cbc07d0d3b295dc21d8c0d12f011 100644
--- a/loopy/__init__.py
+++ b/loopy/__init__.py
@@ -234,8 +234,7 @@ def realize_cse(kernel, cse_tag, dtype, duplicate_inames=[], parallel_inames=Non
dependencies = IndexVariableFinder(
include_reduction_inames=False)(expr.child)
- # FIXME: can happen with
- # assert dependencies <= parent_inames
+ assert dependencies <= parent_inames
for iname in parent_inames:
if iname in duplicate_inames:
diff --git a/loopy/codegen/control.py b/loopy/codegen/control.py
index 788625e0ee7cf6e0229cdf30a7b2776bacdfbad6..f4e1d1c0a619ae6fdcd863bc30c719f43b1de57e 100644
--- a/loopy/codegen/control.py
+++ b/loopy/codegen/control.py
@@ -55,8 +55,14 @@ def generate_code_for_sched_index(kernel, sched_index, codegen_state):
elif isinstance(sched_item, Barrier):
from loopy.codegen import GeneratedInstruction
from cgen import Statement as S
+
+ if sched_item.comment:
+ comment = " /* %s */" % sched_item.comment
+ else:
+ comment = ""
+
return GeneratedInstruction(
- ast=S("barrier(CLK_LOCAL_MEM_FENCE)"),
+ ast=S("barrier(CLK_LOCAL_MEM_FENCE)%s" % comment),
implemented_domain=None)
elif isinstance(sched_item, RunInstruction):
diff --git a/loopy/kernel.py b/loopy/kernel.py
index 1cc356f77d4d19a4c9c334a1de2bce49a83b4f63..db1a326a7c79c2be2527eb7879da1bab597e80ce 100644
--- a/loopy/kernel.py
+++ b/loopy/kernel.py
@@ -310,6 +310,12 @@ class Instruction(Record):
return result
+ @memoize_method
+ def get_read_var_names(self):
+ from loopy.symbolic import DependencyMapper
+ return set(var.name for var in
+ DependencyMapper(composite_leaves=False)(self.expression))
+
# }}}
# {{{ reduction operations
@@ -774,6 +780,13 @@ class LoopKernel(Record):
return tup_to_exprs(grid_size), tup_to_exprs(group_size)
+ @memoize_method
+ def local_var_names(self):
+ return set(
+ tv.name
+ for tv in self.temporary_variables.itervalues()
+ if tv.is_local)
+
def local_mem_use(self):
return sum(lv.nbytes for lv in self.temporary_variables.itervalues()
if lv.is_local)
diff --git a/loopy/preprocess.py b/loopy/preprocess.py
index 990a4fc7f2cb526b73a7afa514f9e571c85c0c40..85ecca37d997a8f98abab945291acaf2f63be6eb 100644
--- a/loopy/preprocess.py
+++ b/loopy/preprocess.py
@@ -193,8 +193,7 @@ def find_accessors(kernel, readers):
for insn in kernel.instructions:
if readers:
- from loopy.symbolic import DependencyMapper
- var_names = DependencyMapper()(insn.expression) & admissible_vars
+ var_names = insn.get_read_var_names() & admissible_vars
else:
var_name = insn.get_assignee_var_name()
diff --git a/loopy/schedule.py b/loopy/schedule.py
index 07ce6c8cbff67ce04ca8fc06fa6545caec4932f1..d9e52dfa8aedacf003f3b3190dc57ff5dacb2297 100644
--- a/loopy/schedule.py
+++ b/loopy/schedule.py
@@ -17,7 +17,7 @@ class RunInstruction(Record):
__slots__ = ["insn_id"]
class Barrier(Record):
- __slots__ = []
+ __slots__ = ["comment"]
# }}}
@@ -43,12 +43,52 @@ def gather_schedule_subloop(schedule, start_idx):
-def has_dependent_in_schedule(kernel, insn_id, schedule):
- from pytools import any
- return any(sched_item
- for sched_item in schedule
- if isinstance(sched_item, RunInstruction)
- and kernel.id_to_insn[sched_item.insn_id].insn_deps)
+
+def get_barrier_needing_dependency(kernel, target, source):
+ from loopy.kernel import Instruction
+ if not isinstance(source, Instruction):
+ source = kernel.id_to_insn[source]
+ if not isinstance(target, Instruction):
+ target = kernel.id_to_insn[target]
+
+ local_vars = kernel.local_var_names()
+
+ tgt_write = set([target.get_assignee_var_name()]) & local_vars
+ tgt_read = target.get_read_var_names() & local_vars
+
+ src_write = set([source.get_assignee_var_name()]) & local_vars
+ src_read = source.get_read_var_names() & local_vars
+
+ waw = tgt_write & src_write
+ raw = tgt_read & src_write
+ war = tgt_write & src_read
+
+ for var_name in raw | war:
+ assert source.id in target.insn_deps
+ return (target, source, var_name)
+
+ if source is target:
+ return None
+
+ for var_name in waw:
+ assert (source.id in target.insn_deps
+ or source is target)
+ return (target, source, var_name)
+
+ return None
+
+
+
+
+
+
+def get_barrier_dependent_in_schedule(kernel, source, schedule):
+ for sched_item in schedule:
+ if isinstance(sched_item, RunInstruction):
+ temp_res = get_barrier_needing_dependency(
+ kernel, sched_item.insn_id, source)
+ if temp_res:
+ return temp_res
@@ -289,8 +329,7 @@ def insert_barriers(kernel, schedule, level=0):
# dependencies for which the 'normal' mechanism below will generate
# barriers.
- def issue_barrier(is_pre_barrier):
- owed_barriers.clear()
+ def issue_barrier(is_pre_barrier, dep):
if result and isinstance(result[-1], Barrier):
return
@@ -298,8 +337,18 @@ def insert_barriers(kernel, schedule, level=0):
if loop_had_barrier[0] or level == 0:
return
+ owed_barriers.clear()
+
+ cmt = None
+ if dep is not None:
+ target, source, var = dep
+ if is_pre_barrier:
+ cmt = "pre-barrier: %s" % var
+ else:
+ cmt = "dependency: %s" % var
+
loop_had_barrier[0] = True
- result.append(Barrier())
+ result.append(Barrier(comment=cmt))
i = 0
while i < len(schedule):
@@ -313,8 +362,9 @@ def insert_barriers(kernel, schedule, level=0):
# (i.e. if anything *in* the loop depends on something beforehand)
for insn_id in owed_barriers:
- if has_dependent_in_schedule(kernel, insn_id, subloop):
- issue_barrier(is_pre_barrier=False)
+ dep = get_barrier_dependent_in_schedule(kernel, insn_id, subloop)
+ if dep:
+ issue_barrier(is_pre_barrier=False, dep=dep)
break
# }}}
@@ -326,9 +376,10 @@ def insert_barriers(kernel, schedule, level=0):
if not loop_had_barrier[0]:
for insn_id in sub_owed_barriers:
- if has_dependent_in_schedule(
- kernel, insn_id, schedule):
- issue_barrier(is_pre_barrier=True)
+ dep = get_barrier_dependent_in_schedule(
+ kernel, insn_id, schedule)
+ if dep:
+ issue_barrier(is_pre_barrier=True, dep=dep)
# }}}
@@ -347,20 +398,23 @@ def insert_barriers(kernel, schedule, level=0):
# {{{ issue dependency-based barriers for this instruction
- if set(insn.insn_deps) & owed_barriers:
- issue_barrier(is_pre_barrier=False)
+ for dep_src_insn_id in set(insn.insn_deps) & owed_barriers:
+ dep = get_barrier_needing_dependency(kernel, insn, dep_src_insn_id)
+ if dep:
+ issue_barrier(is_pre_barrier=False, dep=dep)
# }}}
assignee_temp_var = kernel.temporary_variables.get(
insn.get_assignee_var_name())
if assignee_temp_var is not None and assignee_temp_var.is_local:
+ dep = get_barrier_dependent_in_schedule(kernel, insn.id, schedule)
+
if level == 0:
- assert has_dependent_in_schedule(
- kernel, insn.id, schedule)
+ assert dep
- if has_dependent_in_schedule(kernel, insn.id, schedule):
- issue_barrier(is_pre_barrier=True)
+ if dep:
+ issue_barrier(is_pre_barrier=True, dep=dep)
result.append(sched_item)
owed_barriers.add(insn.id)