def generate_loop_schedules(kernel, debug_args={}):
from loopy.kernel import kernel_state
if kernel.state != kernel_state.PREPROCESSED:
raise LoopyError("cannot schedule a kernel that has not been " "preprocessed")
from loopy.check import pre_schedule_checks
pre_schedule_checks(kernel)
schedule_count = 0
debug = ScheduleDebugger(**debug_args)
from loopy.kernel.data import IlpBaseTag, ParallelTag, VectorizeTag
ilp_inames = set(iname for iname in kernel.all_inames() if isinstance(kernel.iname_to_tag.get(iname), IlpBaseTag))
vec_inames = set(iname for iname in kernel.all_inames() if isinstance(kernel.iname_to_tag.get(iname), VectorizeTag))
parallel_inames = set(
iname for iname in kernel.all_inames() if isinstance(kernel.iname_to_tag.get(iname), ParallelTag)
)
loop_nest_with_map = find_loop_nest_with_map(kernel)
loop_nest_around_map = find_loop_nest_around_map(kernel)
sched_state = SchedulerState(
kernel=kernel,
loop_nest_around_map=loop_nest_around_map,
loop_insn_dep_map=find_loop_insn_dep_map(
kernel, loop_nest_with_map=loop_nest_with_map, loop_nest_around_map=loop_nest_around_map
),
breakable_inames=ilp_inames,
ilp_inames=ilp_inames,
vec_inames=vec_inames,
# time-varying part
active_inames=(),
entered_inames=frozenset(),
schedule=(),
unscheduled_insn_ids=set(insn.id for insn in kernel.instructions),
scheduled_insn_ids=frozenset(),
# ilp and vec are not parallel for the purposes of the scheduler
parallel_inames=parallel_inames - ilp_inames - vec_inames,
group_insn_counts=group_insn_counts(kernel),
active_group_counts={},
)
generators = [
generate_loop_schedules_internal(sched_state, debug=debug, allow_boost=None),
generate_loop_schedules_internal(sched_state, debug=debug),
]
def print_longest_dead_end():
if debug.interactive:
print("Loo.py will now show you the scheduler state at the point")
print("where the longest (dead-end) schedule was generated, in the")
print("the hope that some of this makes sense and helps you find")
print("the issue.")
print()
print("To disable this interactive behavior, pass")
print(" debug_args=dict(interactive=False)")
print("to generate_loop_schedules().")
print(75 * "-")
six.moves.input("Enter:")
print()
print()
debug.debug_length = len(debug.longest_rejected_schedule)
while True:
try:
for _ in generate_loop_schedules_internal(sched_state, debug=debug):
pass
except ScheduleDebugInput as e:
debug.debug_length = int(str(e))
continue
break
try:
for gen in generators:
for gen_sched in gen:
# gen_sched = insert_barriers(kernel, gen_sched,
# reverse=False, kind="global")
# for sched_item in gen_sched:
# if (
# isinstance(sched_item, Barrier)
# and sched_item.kind == "global"):
# raise LoopyError("kernel requires a global barrier %s"
# % sched_item.comment)
debug.stop()
logger.info("%s: barrier insertion: start" % kernel.name)
gen_sched = insert_barriers(kernel, gen_sched, reverse=False, kind="local")
logger.info("%s: barrier insertion: done" % kernel.name)
yield kernel.copy(schedule=gen_sched, state=kernel_state.SCHEDULED)
debug.start()
schedule_count += 1
# if no-boost mode yielded a viable schedule, stop now
if schedule_count:
break
except KeyboardInterrupt:
print()
print(75 * "-")
print("Interrupted during scheduling")
print(75 * "-")
print_longest_dead_end()
raise
debug.done_scheduling()
if not schedule_count:
print(75 * "-")
print("ERROR: Sorry--loo.py did not find a schedule for your kernel.")
print(75 * "-")
print_longest_dead_end()
raise RuntimeError("no valid schedules found")
logger.info("%s: schedule done" % kernel.name)
def generate_loop_schedules(kernel, debug_args={}):
from loopy.kernel import kernel_state
if kernel.state != kernel_state.PREPROCESSED:
raise LoopyError("cannot schedule a kernel that has not been "
"preprocessed")
loop_priority = kernel.loop_priority
from loopy.check import pre_schedule_checks
pre_schedule_checks(kernel)
schedule_count = 0
debug = ScheduleDebugger(**debug_args)
from loopy.kernel.data import IlpBaseTag, ParallelTag, VectorizeTag
ilp_inames = set(
iname
for iname in kernel.all_inames()
if isinstance(kernel.iname_to_tag.get(iname), IlpBaseTag))
vec_inames = set(
iname
for iname in kernel.all_inames()
if isinstance(kernel.iname_to_tag.get(iname), VectorizeTag))
parallel_inames = set(
iname for iname in kernel.all_inames()
if isinstance(kernel.iname_to_tag.get(iname), ParallelTag))
sched_state = SchedulerState(
kernel=kernel,
loop_nest_map=loop_nest_map(kernel),
breakable_inames=ilp_inames,
ilp_inames=ilp_inames,
vec_inames=vec_inames,
# ilp and vec are not parallel for the purposes of the scheduler
parallel_inames=parallel_inames - ilp_inames - vec_inames)
generators = [
generate_loop_schedules_internal(sched_state, loop_priority,
debug=debug, allow_boost=None),
generate_loop_schedules_internal(sched_state, loop_priority,
debug=debug)]
for gen in generators:
for gen_sched in gen:
# gen_sched = insert_barriers(kernel, gen_sched,
# reverse=False, kind="global")
# for sched_item in gen_sched:
# if isinstance(sched_item, Barrier) and sched_item.kind == "global":
# raise LoopyError("kernel requires a global barrier %s"
# % sched_item.comment)
gen_sched = insert_barriers(kernel, gen_sched,
reverse=False, kind="local")
debug.stop()
yield kernel.copy(
schedule=gen_sched,
state=kernel_state.SCHEDULED)
debug.start()
schedule_count += 1
# if no-boost mode yielded a viable schedule, stop now
if schedule_count:
break
debug.done_scheduling()
if not schedule_count:
if debug.interactive:
print(75*"-")
print("ERROR: Sorry--loo.py did not find a schedule for your kernel.")
print(75*"-")
print("Loo.py will now show you the scheduler state at the point")
print("where the longest (dead-end) schedule was generated, in the")
print("the hope that some of this makes sense and helps you find")
print("the issue.")
print()
print("To disable this interactive behavior, pass")
print(" debug_args=dict(interactive=False)")
print("to generate_loop_schedules().")
print(75*"-")
six.moves.input("Enter:")
print()
print()
debug.debug_length = len(debug.longest_rejected_schedule)
for _ in generate_loop_schedules_internal(sched_state, loop_priority,
debug=debug):
pass
raise RuntimeError("no valid schedules found")
logger.info("%s: schedule done" % kernel.name)
