diff --git a/MEMO b/MEMO
index fc33a978c1d7dcad9ca3282d302fe87cb3822161..eb389e7a9c6f31fe35f9fb3564969215e439208d 100644
--- a/MEMO
+++ b/MEMO
@@ -87,6 +87,8 @@ Future ideas
Dealt with
^^^^^^^^^^
+- Scheduler tries too many boostability-related options
+
- Automatically generate testing code vs. sequential.
- If isl can prove that all operands are positive, may use '/' instead of
@@ -209,4 +211,3 @@ g: is tagged as group idx
Raise error if dl is targeting a private variable, regardless of whether it's
a dependency or not.
-
diff --git a/loopy/kernel.py b/loopy/kernel.py
index 09b5c9b9cb7d7334d4eda8109143f0484d866c8d..b301ac9856e4a42928d553f88bab2d31906e98f1 100644
--- a/loopy/kernel.py
+++ b/loopy/kernel.py
@@ -738,6 +738,16 @@ class LoopKernel(Record):
else:
return self.all_insn_inames()[insn.id]
+ @memoize_method
+ def iname_to_insns(self):
+ result = dict(
+ (iname, set()) for iname in self.all_inames())
+ for insn in self.instructions:
+ for iname in self.insn_inames(insn):
+ result[iname].add(insn.id)
+
+ return result
+
@property
@memoize_method
def sequential_inames(self):
@@ -990,6 +1000,22 @@ class LoopKernel(Record):
return sum(lv.nbytes for lv in self.temporary_variables.itervalues()
if lv.is_local)
+ @memoize_method
+ def loop_nest_map(self):
+ """Returns a dictionary mapping inames to other inames that are
+ always nested around them.
+ """
+ result = {}
+ iname_to_insns = self.iname_to_insns()
+
+ for inner_iname in self.all_inames():
+ result[inner_iname] = set()
+ for outer_iname in self.all_inames():
+ if iname_to_insns[inner_iname] < iname_to_insns[outer_iname]:
+ result[inner_iname].add(outer_iname)
+
+ return result
+
def __str__(self):
lines = []
diff --git a/loopy/schedule.py b/loopy/schedule.py
index 5523583dfacf48f4cae835ed4c54fb5f26c80138..929792fdab4d404be7a78d9a52155a9fd2ec7379 100644
--- a/loopy/schedule.py
+++ b/loopy/schedule.py
@@ -266,12 +266,12 @@ def generate_loop_schedules_internal(kernel, loop_priority, schedule=[], allow_b
and len(schedule) >= debug.debug_length):
debug_mode = True
- #print dump_schedule(schedule), len(schedule)
if debug_mode:
print kernel
print "--------------------------------------------"
print "CURRENT SCHEDULE:"
- print dump_schedule(schedule)
+ print dump_schedule(schedule), len(schedule)
+ print "boost allowed:", allow_boost
print "--------------------------------------------"
#if len(schedule) == 2:
@@ -289,6 +289,8 @@ def generate_loop_schedules_internal(kernel, loop_priority, schedule=[], allow_b
reachable_insn_ids = set()
for insn_id in all_insn_ids - scheduled_insn_ids:
+ if debug_mode:
+ print insn_id
insn = kernel.id_to_insn[insn_id]
schedule_now = set(insn.insn_deps) <= scheduled_insn_ids
@@ -299,52 +301,28 @@ def generate_loop_schedules_internal(kernel, loop_priority, schedule=[], allow_b
insn.id, ",".join(set(insn.insn_deps) - scheduled_insn_ids))
continue
+ want = kernel.insn_inames(insn) - parallel_inames
+ have = active_inames_set - parallel_inames
+
# If insn is boostable, it may be placed inside a more deeply
# nested loop without harm.
- # But if it can be scheduled on the way *out* of the currently
- # active loops, now is not the right moment.
-
- schedulable_at_loop_levels = []
-
if allow_boost:
- try_loop_counts = xrange(len(active_inames), -1, -1)
- else:
- try_loop_counts = [len(active_inames)]
-
- for active_loop_count in try_loop_counts:
- outer_active_inames = set(active_inames[:active_loop_count])
-
- want = kernel.insn_inames(insn) - parallel_inames
- have = outer_active_inames - parallel_inames - insn.boostable_into
-
- if allow_boost:
- have -= insn.boostable_into
+ have = have - insn.boostable_into
- if want == have:
- schedulable_at_loop_levels.append(active_loop_count)
-
- if schedulable_at_loop_levels != [len(active_inames)]:
+ if want != have:
schedule_now = False
+
if debug_mode:
- if schedulable_at_loop_levels:
- print ("instruction '%s' will be scheduled when more "
- "loops have been exited" % insn.id)
- else:
- want = (kernel.insn_inames(insn) - parallel_inames)
- have = (active_inames_set - parallel_inames)
- if want-have:
- print ("instruction '%s' is missing inames '%s'"
- % (insn.id, ",".join(want-have)))
- if have-want:
- print ("instruction '%s' won't work under inames '%s'"
- % (insn.id, ",".join(have-want)))
+ if want-have:
+ print ("instruction '%s' is missing inames '%s'"
+ % (insn.id, ",".join(want-have)))
+ if have-want:
+ print ("instruction '%s' won't work under inames '%s'"
+ % (insn.id, ",".join(have-want)))
# {{{ determine reachability
- want = kernel.insn_inames(insn) - parallel_inames
- have = active_inames_set - parallel_inames
-
if (not schedule_now and have <= want):
reachable_insn_ids.add(insn_id)
else:
@@ -355,6 +333,8 @@ def generate_loop_schedules_internal(kernel, loop_priority, schedule=[], allow_b
# }}}
if schedule_now:
+ if debug_mode:
+ print "scheduling '%s'" % insn.id
scheduled_insn_ids.add(insn.id)
schedule = schedule + [RunInstruction(insn_id=insn.id)]
made_progress = True
@@ -408,6 +388,9 @@ def generate_loop_schedules_internal(kernel, loop_priority, schedule=[], allow_b
useful_loops = []
for iname in available_loops:
+ if not kernel.loop_nest_map()[iname] <= active_inames_set | parallel_inames:
+ continue
+
# {{{ determine if that gets us closer to being able to schedule an insn
useful = False
@@ -415,8 +398,10 @@ def generate_loop_schedules_internal(kernel, loop_priority, schedule=[], allow_b
hypothetically_active_loops = active_inames_set | set([iname])
for insn_id in reachable_insn_ids:
insn = kernel.id_to_insn[insn_id]
- if (hypothetically_active_loops
- <= (kernel.insn_inames(insn) | insn.boostable_into)):
+
+ want = kernel.insn_inames(insn) | insn.boostable_into
+
+ if hypothetically_active_loops <= want:
useful = True
break