def optguards(self, loop, user_code=False):
info = FakeLoopInfo(loop)
info.snapshot(loop)
for op in loop.operations:
if op.is_guard():
op.setdescr(compile.CompileLoopVersionDescr())
dep = DependencyGraph(loop)
opt = GuardStrengthenOpt(dep.index_vars)
opt.propagate_all_forward(info, loop, user_code)
return opt
def optguards(self, loop, user_code=False):
info = FakeLoopInfo(loop)
info.snapshot(loop)
for op in loop.operations:
if op.is_guard():
op.setdescr(compile.CompileLoopVersionDescr())
dep = DependencyGraph(loop)
opt = GuardStrengthenOpt(dep.index_vars)
opt.propagate_all_forward(info, loop, user_code)
return opt
def optimize_vector(trace, metainterp_sd, jitdriver_sd, warmstate,
loop_info, loop_ops, jitcell_token=None):
""" Enter the world of SIMD. Bails if it cannot transform the trace. """
user_code = not jitdriver_sd.vec and warmstate.vec_all
e = len(loop_ops)-1
assert e > 0
assert rop.is_final(loop_ops[e].getopnum())
loop = VectorLoop(loop_info.label_op, loop_ops[:e], loop_ops[-1])
if user_code and user_loop_bail_fast_path(loop, warmstate):
return loop_info, loop_ops
# the original loop (output of optimize_unroll)
info = LoopVersionInfo(loop_info)
version = info.snapshot(loop)
loop.setup_vectorization()
try:
debug_start("vec-opt-loop")
metainterp_sd.logger_noopt.log_loop([], loop.finaloplist(label=True), -2, None, None, "pre vectorize")
metainterp_sd.profiler.count(Counters.OPT_VECTORIZE_TRY)
#
start = time.clock()
opt = VectorizingOptimizer(metainterp_sd, jitdriver_sd, warmstate.vec_cost)
index_vars = opt.run_optimization(info, loop)
gso = GuardStrengthenOpt(index_vars)
gso.propagate_all_forward(info, loop, user_code)
end = time.clock()
#
metainterp_sd.profiler.count(Counters.OPT_VECTORIZED)
metainterp_sd.logger_noopt.log_loop([], loop.finaloplist(label=True), -2, None, None, "post vectorize")
nano = int((end-start)*10.0**9)
debug_print("# vecopt factor: %d opcount: (%d -> %d) took %dns" % \
(opt.unroll_count+1, len(version.loop.operations), len(loop.operations), nano))
debug_stop("vec-opt-loop")
#
info.label_op = loop.label
return info, loop.finaloplist(jitcell_token=jitcell_token, reset_label_token=False)
except NotAVectorizeableLoop:
debug_stop("vec-opt-loop")
# vectorization is not possible
return loop_info, version.loop.finaloplist()
except NotAProfitableLoop:
debug_stop("vec-opt-loop")
# cost model says to skip this loop
return loop_info, version.loop.finaloplist()
except Exception as e:
debug_stop("vec-opt-loop")
debug_print("failed to vectorize loop. THIS IS A FATAL ERROR!")
if we_are_translated():
from rpython.rtyper.lltypesystem import lltype
from rpython.rtyper.lltypesystem.lloperation import llop
llop.debug_print_traceback(lltype.Void)
else:
raise
finally:
loop.teardown_vectorization()
return loop_info, loop_ops
def run_optimization(self, metainterp_sd, info, loop, jitcell_token,
user_code):
self.orig_label_args = loop.label.getarglist_copy()
self.linear_find_smallest_type(loop)
byte_count = self.smallest_type_bytes
vsize = self.vector_ext.vec_size()
# stop, there is no chance to vectorize this trace
# we cannot optimize normal traces (if there is no label)
if vsize == 0:
debug_print("vector size is zero")
raise NotAVectorizeableLoop
if byte_count == 0:
debug_print("could not find smallest type")
raise NotAVectorizeableLoop
if loop.label.getopnum() != rop.LABEL:
debug_print("not a loop, can only vectorize loops")
raise NotAVectorizeableLoop
# find index guards and move to the earliest position
graph = self.analyse_index_calculations(loop)
if graph is not None:
state = SchedulerState(metainterp_sd.cpu, graph)
self.schedule(state) # reorder the trace
# unroll
self.unroll_count = self.get_unroll_count(vsize)
align_unroll = self.unroll_count==1 and \
self.vector_ext.should_align_unroll
self.unroll_loop_iterations(loop,
self.unroll_count,
align_unroll_once=align_unroll)
# vectorize
graph = DependencyGraph(loop)
self.find_adjacent_memory_refs(graph)
self.extend_packset()
self.combine_packset()
costmodel = GenericCostModel(self.cpu, self.cost_threshold)
state = VecScheduleState(graph, self.packset, self.cpu, costmodel)
self.schedule(state)
if not state.profitable():
raise NotAProfitableLoop
gso = GuardStrengthenOpt(graph.index_vars)
gso.propagate_all_forward(info, loop, user_code)
# re-schedule the trace -> removes many pure operations
graph = DependencyGraph(loop)
state = SchedulerState(self.cpu, graph)
state.schedule()
info.extra_before_label = loop.align_operations
for op in loop.align_operations:
op.set_forwarded(None)
return loop.finaloplist(jitcell_token=jitcell_token,
reset_label_token=False)
def run_optimization(self, metainterp_sd, info, loop, jitcell_token, user_code):
self.orig_label_args = loop.label.getarglist_copy()
self.linear_find_smallest_type(loop)
byte_count = self.smallest_type_bytes
vsize = self.vector_ext.vec_size()
# stop, there is no chance to vectorize this trace
# we cannot optimize normal traces (if there is no label)
if vsize == 0:
debug_print("vector size is zero")
raise NotAVectorizeableLoop
if byte_count == 0:
debug_print("could not find smallest type")
raise NotAVectorizeableLoop
if loop.label.getopnum() != rop.LABEL:
debug_print("not a loop, can only vectorize loops")
raise NotAVectorizeableLoop
# find index guards and move to the earliest position
graph = self.analyse_index_calculations(loop)
if graph is not None:
state = SchedulerState(metainterp_sd.cpu, graph)
self.schedule(state) # reorder the trace
# unroll
self.unroll_count = self.get_unroll_count(vsize)
align_unroll = self.unroll_count==1 and \
self.vector_ext.should_align_unroll
self.unroll_loop_iterations(loop, self.unroll_count,
align_unroll_once=align_unroll)
# vectorize
graph = DependencyGraph(loop)
self.find_adjacent_memory_refs(graph)
self.extend_packset()
self.combine_packset()
costmodel = GenericCostModel(self.cpu, self.cost_threshold)
state = VecScheduleState(graph, self.packset, self.cpu, costmodel)
self.schedule(state)
if not state.profitable():
raise NotAProfitableLoop
gso = GuardStrengthenOpt(graph.index_vars)
gso.propagate_all_forward(info, loop, user_code)
# re-schedule the trace -> removes many pure operations
graph = DependencyGraph(loop)
state = SchedulerState(self.cpu, graph)
state.schedule()
info.extra_before_label = loop.align_operations
for op in loop.align_operations:
op.set_forwarded(None)
return loop.finaloplist(jitcell_token=jitcell_token, reset_label_token=False)