def test_meta_schedule_local_runner_time_out():
"""Test meta schedule Local Runner time out"""
mod = MatmulModule
builder = LocalBuilder()
(builder_result,) = builder.build([BuilderInput(mod, Target("llvm"))])
assert builder_result.artifact_path is not None
assert builder_result.error_msg is None
runner_input = RunnerInput(
builder_result.artifact_path,
"llvm",
[
TensorInfo("float32", (MATMUL_N, MATMUL_N)),
TensorInfo("float32", (MATMUL_N, MATMUL_N)),
TensorInfo("float32", (MATMUL_N, MATMUL_N)),
],
)
def initializer():
@register_func("meta_schedule.runner.test_time_out")
def timeout_session_creator( # pylint: disable=unused-variable
device: Device, # pylint: disable=unused-argument
args_info: T_ARG_INFO_JSON_OBJ_LIST, # pylint: disable=unused-argument
alloc_repeat: int, # pylint: disable=unused-argument
) -> RPCSession:
time.sleep(2)
evaluator_config = EvaluatorConfig(
number=1,
repeat=1,
min_repeat_ms=0,
enable_cpu_cache_flush=False,
)
runner = LocalRunner(
timeout_sec=1,
evaluator_config=evaluator_config,
initializer=initializer,
f_alloc_argument="meta_schedule.runner.test_time_out",
)
# Run the module
(runner_future,) = runner.run([runner_input])
runner_result = runner_future.result()
assert runner_result.error_msg is not None and runner_result.error_msg.startswith(
"LocalRunner: Timeout, killed after"
)
assert runner_result.run_secs is None
_clean_build(builder_result.artifact_path)
def verify_meta_schedule_with_tensorrt(
mod,
params,
data_shape,
use_trt: bool = True,
):
# Build
builder = LocalBuilder(
f_build=build_relay_with_tensorrt if use_trt else build_relay,
timeout_sec=1000,
)
builder_input = BuilderInput(mod, Target("cuda"), params)
builder_result = builder.build([builder_input])[0]
assert builder_result.error_msg is None, builder_result.error_msg
assert builder_result.artifact_path is not None
# Run
runner_input = RunnerInput(
builder_result.artifact_path,
device_type="cuda",
args_info=[TensorInfo("float32", data_shape)],
)
runner = LocalRunner(
evaluator_config=EvaluatorConfig(
number=5,
repeat=2,
min_repeat_ms=0,
enable_cpu_cache_flush=False,
),
f_run_evaluator=run_with_graph_executor,
)
# Run the module
runner_future = runner.run([runner_input])[0]
runner_result = runner_future.result()
assert runner_result is not None
assert runner_result.error_msg is None, runner_result.error_msg
assert runner_result.run_secs is not None
for result in runner_result.run_secs:
if isinstance(result, FloatImm):
result = result.value
assert isinstance(result, float)
assert result >= 0.0
def test_meta_schedule_local_single_run():
"""Test meta schedule local runner for a single run"""
# Build the module
mod = MatmulModule
builder = LocalBuilder()
(builder_result, ) = builder.build([BuilderInput(mod, Target("llvm"))])
assert builder_result.artifact_path is not None
assert builder_result.error_msg is None
runner_input = RunnerInput(
builder_result.artifact_path,
"llvm",
[
TensorInfo("float32", (MATMUL_N, MATMUL_N)),
TensorInfo("float32", (MATMUL_N, MATMUL_N)),
TensorInfo("float32", (MATMUL_N, MATMUL_N)),
],
)
evaluator_config = EvaluatorConfig(
number=1,
repeat=1,
min_repeat_ms=0,
enable_cpu_cache_flush=False,
)
runner = LocalRunner(timeout_sec=100, evaluator_config=evaluator_config)
# Run the module
(runner_future, ) = runner.run([runner_input])
runner_result = runner_future.result()
assert runner_result.error_msg is None
for result in runner_result.run_secs:
if isinstance(result, FloatImm):
result = result.value
assert isinstance(result, float)
assert result >= 0.0
_clean_build(builder_result.artifact_path)
def test_meta_schedule_evolutionary_search(
): # pylint: disable = invalid-name]
num_trials_per_iter = 10
num_trials_total = 100
strategy = EvolutionarySearch(
num_trials_per_iter=num_trials_per_iter,
num_trials_total=num_trials_total,
population_size=5,
init_measured_ratio=0.1,
init_min_unmeasured=50,
genetic_num_iters=3,
genetic_mutate_prob=0.5,
genetic_max_fail_count=10,
eps_greedy=0.9,
)
context = TuneContext(
mod=Matmul,
space_generator=ScheduleFn(sch_fn=_schedule_matmul),
mutator_probs={
DummyMutator(): 1.0,
},
target=tvm.target.Target("llvm"),
num_threads=1, # because we are using a mutator from the python side
)
_scheduler = RoundRobin(
tasks=[context],
builder=LocalBuilder(),
runner=LocalRunner(),
database=DummyDatabase(),
cost_model=RandomModel(),
measure_callbacks=[],
)
context.space_generator.initialize_with_tune_context(context)
spaces = context.space_generator.generate_design_space(context.mod)
strategy.initialize_with_tune_context(context)
strategy.pre_tuning(spaces)
(correct_sch, ) = ScheduleFn(
sch_fn=_schedule_matmul).generate_design_space(Matmul)
num_trials_each_iter: List[int] = []
candidates = strategy.generate_measure_candidates()
while candidates is not None:
num_trials_each_iter.append(len(candidates))
runner_results: List[RunnerResult] = []
for candidate in candidates:
_is_trace_equal(
candidate.sch,
correct_sch,
remove_decisions=(isinstance(strategy, ReplayTrace)),
)
runner_results.append(
RunnerResult(run_secs=[0.11, 0.41, 0.54], error_msg=None))
strategy.notify_runner_results(context, candidates, runner_results)
candidates = strategy.generate_measure_candidates()
strategy.post_tuning()
print(num_trials_each_iter)
correct_count = 10 # For each iteration except the last one
assert num_trials_each_iter == [correct_count] * (
num_trials_total // correct_count) + (
[num_trials_total %
correct_count] if num_trials_total % correct_count != 0 else [])
del _scheduler
def test_meta_schedule_evolutionary_search(
): # pylint: disable = invalid-name]
@derived_object
class DummyMutator(PyMutator):
"""Dummy Mutator for testing"""
def initialize_with_tune_context(self, context: "TuneContext") -> None:
pass
def apply(self, trace: Trace, _) -> Optional[Trace]:
return Trace(trace.insts, {})
@derived_object
class DummyDatabase(PyDatabase):
"""Dummy Database for testing"""
def __init__(self):
super().__init__()
self.records = []
self.workload_reg = []
def has_workload(self, mod: IRModule) -> bool:
for workload in self.workload_reg:
if tvm.ir.structural_equal(workload.mod, mod):
return True
return False
def commit_tuning_record(self, record: TuningRecord) -> None:
self.records.append(record)
def commit_workload(self, mod: IRModule) -> Workload:
for workload in self.workload_reg:
if tvm.ir.structural_equal(workload.mod, mod):
return workload
workload = Workload(mod)
self.workload_reg.append(workload)
return workload
def get_top_k(self, workload: Workload,
top_k: int) -> List[TuningRecord]:
return list(
filter(
lambda x: x.workload == workload,
sorted(self.records,
key=lambda x: sum(x.run_secs) / len(x.run_secs)),
))[:int(top_k)]
def __len__(self) -> int:
return len(self.records)
def print_results(self) -> None:
print("\n".join([str(r) for r in self.records]))
num_trials_per_iter = 10
num_trials_total = 100
strategy = EvolutionarySearch(
num_trials_per_iter=num_trials_per_iter,
num_trials_total=num_trials_total,
population_size=5,
init_measured_ratio=0.1,
init_min_unmeasured=50,
genetic_num_iters=3,
genetic_mutate_prob=0.5,
genetic_max_fail_count=10,
eps_greedy=0.9,
)
context = TuneContext(
mod=Matmul,
space_generator=ScheduleFn(sch_fn=_schedule_matmul),
mutator_probs={
DummyMutator(): 1.0,
},
target=tvm.target.Target("llvm"),
num_threads=1, # because we are using a mutator from the python side
)
_scheduler = RoundRobin(
tasks=[context],
builder=LocalBuilder(),
runner=LocalRunner(),
database=DummyDatabase(),
cost_model=RandomModel(),
measure_callbacks=[],
)
context.space_generator.initialize_with_tune_context(context)
spaces = context.space_generator.generate_design_space(context.mod)
strategy.initialize_with_tune_context(context)
strategy.pre_tuning(spaces)
(correct_sch, ) = ScheduleFn(
sch_fn=_schedule_matmul).generate_design_space(Matmul)
num_trials_each_iter: List[int] = []
candidates = strategy.generate_measure_candidates()
while candidates is not None:
num_trials_each_iter.append(len(candidates))
runner_results: List[RunnerResult] = []
for candidate in candidates:
_is_trace_equal(
candidate.sch,
correct_sch,
remove_decisions=(isinstance(strategy, ReplayTrace)),
)
runner_results.append(
RunnerResult(run_secs=[0.11, 0.41, 0.54], error_msg=None))
strategy.notify_runner_results(context, candidates, runner_results)
candidates = strategy.generate_measure_candidates()
strategy.post_tuning()
print(num_trials_each_iter)
correct_count = 10 # For each iteration except the last one
assert num_trials_each_iter == [correct_count] * (
num_trials_total // correct_count) + (
[num_trials_total %
correct_count] if num_trials_total % correct_count != 0 else [])
del _scheduler
def test_meta_schedule_local_runner_add_test():
"""Test meta schedule local runner with add module"""
def _check_correct_add(args_before: List[np.array],
args_after: List[np.array]) -> None:
a_before, b_before, c_before = args_before
a_after, b_after, c_after = args_after
c_before = a_before + b_before
assert (a_before == a_after).all()
assert (b_before == b_after).all()
assert (c_before == c_after).all()
def test_alloc_argument(
device: Device,
args_info: T_ARG_INFO_JSON_OBJ_LIST, # pylint: disable=unused-argument
alloc_repeat: int,
) -> List[T_ARGUMENT_LIST]:
global repeated_args_before # pylint: disable=global-variable-undefined, invalid-name
repeated_args_before = []
repeated_args = local_default_alloc_argument(device, args_info,
alloc_repeat)
for args in repeated_args:
repeated_args_before.append([arg.asnumpy() for arg in args])
return repeated_args
def test_run_evaluator(
rt_mod: Module,
device: Device,
evaluator_config: EvaluatorConfig,
repeated_args: List[Any],
) -> List[float]:
global repeated_args_before # pylint: disable=global-variable-undefined, invalid-name
repeated_args_after = []
evaluator = rt_mod.time_evaluator(
func_name=rt_mod.entry_name,
dev=device,
number=evaluator_config.number,
repeat=evaluator_config.repeat,
min_repeat_ms=evaluator_config.min_repeat_ms,
f_preproc="cache_flush_cpu_non_first_arg"
if evaluator_config.enable_cpu_cache_flush else "",
)
repeated_costs: List[List[float]] = []
for args in repeated_args:
device.sync()
profile_result = evaluator(*args)
repeated_costs.append(profile_result.results)
repeated_args_after.append([arg.asnumpy() for arg in args])
costs = [
float(cost)
for cost in itertools.chain.from_iterable(repeated_costs)
]
for args_before, args_after in zip(repeated_args_before,
repeated_args_after):
_check_correct_add(args_before, args_after)
del repeated_args_before
return costs
# Build the module
mod = AddModule
builder = LocalBuilder()
(builder_result, ) = builder.build([BuilderInput(mod, Target("llvm"))])
assert builder_result.artifact_path is not None
assert builder_result.error_msg is None
runner_input = RunnerInput(
builder_result.artifact_path,
"llvm",
[
TensorInfo("float32", [MATMUL_M]),
TensorInfo("float32", [MATMUL_M]),
TensorInfo("float32", [MATMUL_M]),
],
)
evaluator_config = EvaluatorConfig(
number=1,
repeat=1,
min_repeat_ms=0,
enable_cpu_cache_flush=False,
)
runner = LocalRunner(
timeout_sec=100,
evaluator_config=evaluator_config,
f_alloc_argument=test_alloc_argument,
f_run_evaluator=test_run_evaluator,
)
# Run the module
(runner_future, ) = runner.run([runner_input])
runner_result = runner_future.result()
assert runner_result.error_msg is None
for result in runner_result.run_secs:
if isinstance(result, FloatImm):
result = result.value
assert isinstance(result, float)
assert result >= 0.0
_clean_build(builder_result.artifact_path)
def test_meta_schedule_local_multiple_runs():
"""Test meta schedule local runner for multiple runs"""
# Build the module
mods = [
MatmulModule,
MatmulReluModule,
BatchMatmulModule,
]
builder = LocalBuilder()
builder_inputs = [BuilderInput(mod, Target("llvm")) for mod in mods]
builder_results = builder.build(builder_inputs)
for builder_result in builder_results:
assert builder_result.artifact_path is not None
assert builder_result.error_msg is None
args_infos = [
[
TensorInfo("float32", (MATMUL_N, MATMUL_N)),
TensorInfo("float32", (MATMUL_N, MATMUL_N)),
TensorInfo("float32", (MATMUL_N, MATMUL_N)),
],
[
TensorInfo("float32", (MATMUL_N, MATMUL_N)),
TensorInfo("float32", (MATMUL_N, MATMUL_N)),
TensorInfo("float32", (MATMUL_N, MATMUL_N)),
],
[
TensorInfo("float32", [16, MATMUL_M, MATMUL_M]),
TensorInfo("float32", [16, MATMUL_M, MATMUL_M]),
TensorInfo("float32", [16, MATMUL_M, MATMUL_M]),
],
]
runner_inputs = [
RunnerInput(builder_results[i].artifact_path, "llvm", args_infos[i])
for i in range(len(mods))
]
evaluator_config = EvaluatorConfig(
number=1,
repeat=1,
min_repeat_ms=0,
enable_cpu_cache_flush=False,
)
runner = LocalRunner(timeout_sec=100, evaluator_config=evaluator_config)
# Run the module
runner_futures = runner.run(runner_inputs)
runner_results = [
runner_future.result() for runner_future in runner_futures
]
for runner_result in runner_results:
assert runner_result.error_msg is None
for result in runner_result.run_secs:
if isinstance(result, FloatImm):
result = result.value
assert isinstance(result, float)
assert result >= 0.0
for builder_result in builder_results:
_clean_build(builder_result.artifact_path)
def verify_meta_schedule_with_tensorrt(
mod, params, data_shape, use_meta_sched: bool = True, use_trt: bool = True, mode: str = "vm"
):
if use_meta_sched:
# With meta_schedule
dev = "cuda"
# Build
if use_trt:
from tvm.meta_schedule.testing import relay_build_with_tensorrt
builder = LocalBuilder(f_build=relay_build_with_tensorrt)
else:
def relay_build_without_tensorrt(
mod: Module,
target: Target,
params: dict,
) -> List[BuilderResult]:
return tvm.relay.build_module._build_module_no_factory(mod, "cuda", "llvm", params)
builder = LocalBuilder(f_build=relay_build_without_tensorrt)
builder_input = BuilderInput(mod, Target(dev, host="llvm"), params)
(builder_result,) = builder.build([builder_input])
assert builder_result.error_msg is None
assert builder_result.artifact_path is not None
# Run
evaluator_config = EvaluatorConfig(
number=5,
repeat=2,
min_repeat_ms=0,
enable_cpu_cache_flush=False,
)
runner_input = RunnerInput(
builder_result.artifact_path, "cuda", [TensorInfo("float32", data_shape)]
)
def eval_func(rt_mod, device, evaluator_config, repeated_args):
rt_mod = tvm.contrib.graph_executor.GraphModule(rt_mod["default"](device))
eval = rt_mod.module.time_evaluator(
func_name="run",
dev=device,
number=evaluator_config.number,
repeat=evaluator_config.repeat,
min_repeat_ms=evaluator_config.min_repeat_ms,
f_preproc="cache_flush_cpu_non_first_arg"
if evaluator_config.enable_cpu_cache_flush
else "",
)
repeated_costs: List[List[float]] = []
for args in repeated_args:
profile_result = eval(*args)
repeated_costs.append(profile_result.results)
costs = [float(cost) for cost in itertools.chain.from_iterable(repeated_costs)]
return costs
runner = LocalRunner(
evaluator_config=evaluator_config,
f_run_evaluator=eval_func,
)
# Run the module
(runner_future,) = runner.run([runner_input])
runner_result = runner_future.result()
assert runner_result is not None
assert runner_result.run_secs is not None
assert runner_result.error_msg is None
for result in runner_result.run_secs:
if isinstance(result, FloatImm):
result = result.value
assert isinstance(result, float)
assert result >= 0.0
else:
# Without meta_schedule
if use_trt:
mod, config = tensorrt.partition_for_tensorrt(mod)
with tvm.transform.PassContext(
opt_level=3, config={"relay.ext.tensorrt.options": config}
):
func = relay.create_executor(
mode, mod=mod, device=tvm.cuda(0), target="cuda"
).evaluate()
else:
with tvm.transform.PassContext(opt_level=3):
func = relay.create_executor(
mode, mod=mod, device=tvm.cuda(0), target="cuda", params=params
).evaluate()