def run_optimization(self, info, loop):
self.orig_label_args = loop.label.getarglist_copy()
self.linear_find_smallest_type(loop)
byte_count = self.smallest_type_bytes
vsize = self.cpu.vector_register_size
if vsize == 0 or byte_count == 0 or loop.label.getopnum() != rop.LABEL:
# stop, there is no chance to vectorize this trace
# we cannot optimize normal traces (if there is no label)
raise NotAVectorizeableLoop()
# find index guards and move to the earliest position
graph = self.analyse_index_calculations(loop)
if graph is not None:
state = SchedulerState(graph)
self.schedule(state) # reorder the trace
# unroll
self.unroll_count = self.get_unroll_count(vsize)
self.unroll_loop_iterations(loop, self.unroll_count)
# vectorize
graph = DependencyGraph(loop)
self.find_adjacent_memory_refs(graph)
self.extend_packset()
self.combine_packset()
# TODO move cost model to CPU
costmodel = X86_CostModel(self.cpu, self.cost_threshold)
state = VecScheduleState(graph, self.packset, self.cpu, costmodel)
self.schedule(state)
if not state.profitable():
raise NotAProfitableLoop()
return graph.index_vars
def vectoroptimizer_unrolled(self, loop, unroll_factor=-1):
opt = self.vectoroptimizer(loop)
opt.linear_find_smallest_type(loop)
loop.setup_vectorization()
if unroll_factor == -1 and opt.smallest_type_bytes == 0:
raise NotAVectorizeableLoop()
if unroll_factor == -1:
unroll_factor = opt.get_unroll_count(ARCH_VEC_REG_SIZE)
print ""
print "unroll factor: ", unroll_factor, opt.smallest_type_bytes
self.show_dot_graph(DependencyGraph(loop),
"original_" + self.test_name)
graph = opt.analyse_index_calculations(loop)
if graph is not None:
cycle = graph.cycles()
if cycle is not None:
print "CYCLE found %s" % cycle
self.show_dot_graph(graph, "early_exit_" + self.test_name)
assert cycle is None
state = SchedulerState(graph)
opt.schedule(state)
opt.unroll_loop_iterations(loop, unroll_factor)
self.debug_print_operations(loop)
graph = DependencyGraph(loop)
self.last_graph = graph # legacy for test_dependency
self.show_dot_graph(graph, self.test_name)
def gmr(i):
return graph.memory_refs[graph.nodes[i]]
graph.getmemref = gmr
return opt, graph
def vectorize(self, loop, unroll_factor=-1):
info = FakeLoopInfo(loop)
info.snapshot(loop)
opt, graph = self.vectoroptimizer_unrolled(loop, unroll_factor)
opt.find_adjacent_memory_refs(graph)
opt.extend_packset()
opt.combine_packset()
costmodel = GenericCostModel(self.cpu, 0)
state = VecScheduleState(graph, opt.packset, self.cpu, costmodel)
opt.schedule(state)
if not costmodel.profitable():
raise NotAProfitableLoop()
gso = GuardStrengthenOpt(graph.index_vars)
gso.propagate_all_forward(info, loop)
#
# re-schedule
graph = DependencyGraph(loop)
state = SchedulerState(self.cpu, graph)
state.prepare()
Scheduler().walk_and_emit(state)
state.post_schedule()
#
oplist = loop.operations
loop.operations = loop.prefix[:]
if loop.prefix_label:
loop.operations += [loop.prefix_label]
loop.operations += oplist
return opt
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)
def schedule(self, loop, unroll_factor=-1, with_guard_opt=False):
info = FakeLoopInfo(loop)
info.snapshot(loop)
opt, graph = self.vectoroptimizer_unrolled(loop, unroll_factor)
opt.find_adjacent_memory_refs(graph)
opt.extend_packset()
opt.combine_packset()
costmodel = FakeCostModel(self.cpu)
state = VecScheduleState(graph, opt.packset, self.cpu, costmodel)
opt.schedule(state)
if with_guard_opt:
gso = GuardStrengthenOpt(graph.index_vars)
gso.propagate_all_forward(info, loop)
# re-schedule
graph = DependencyGraph(loop)
state = SchedulerState(self.cpu, graph)
state.prepare()
Scheduler().walk_and_emit(state)
state.post_schedule()
return opt
def earlyexit(self, loop):
opt = self.vectoroptimizer(loop)
graph = opt.analyse_index_calculations(loop)
state = SchedulerState(graph)
opt.schedule(state)
return graph.loop
