def tune_network(network, target): # Extract tasks mod, params = get_network(network) target = tvm.target.Target(target) tasks, task_weights = auto_scheduler.extract_tasks(mod["main"], params, target) with tempfile.NamedTemporaryFile() as fp: log_file = fp.name # Tuning measure_ctx = auto_scheduler.LocalRPCMeasureContext(timeout=60) tuner = auto_scheduler.TaskScheduler(tasks, task_weights) tune_option = auto_scheduler.TuningOptions( num_measure_trials=100, num_measures_per_round=2, early_stopping=1, runner=measure_ctx.runner, builder=auto_scheduler.LocalBuilder(timeout=60), measure_callbacks=[auto_scheduler.RecordToFile(log_file)], ) tuner.tune(tune_option, search_policy="sketch.random") del measure_ctx # Compile with the history best with auto_scheduler.ApplyHistoryBest(log_file): with tvm.transform.PassContext( opt_level=3, config={"relay.backend.use_auto_scheduler": True}): lib = relay.build(mod, target=target, params=params)
def test_tuning_cuda(): auto_scheduler.enable_relay_integration() # Extract tasks mod, params = get_network("mlp") target = tvm.target.Target("cuda") tasks, task_weights = auto_scheduler.extract_tasks(mod["main"], params, target) objective = lambda costs: sum(c * w for c, w in zip(costs, task_weights)) with tempfile.NamedTemporaryFile() as fp: log_file = fp.name # Tuning measure_ctx = auto_scheduler.LocalRPCMeasureContext(timeout=100) tuner = auto_scheduler.TaskScheduler(tasks, objective) tune_option = auto_scheduler.TuningOptions( num_measure_trials=2, num_measures_per_round=1, runner=measure_ctx.runner, measure_callbacks=[auto_scheduler.RecordToFile(log_file)], ) tuner.tune(tune_option, search_policy="sketch.random") del measure_ctx # Compile with the history best with auto_scheduler.ApplyHistoryBest(log_file): with tvm.transform.PassContext(opt_level=3): lib = relay.build(mod, target=target, params=params) # Todo(merrymercy): compile without any history to test the fallback mechanism auto_scheduler.enable_relay_integration(False)
def tune_network(network, target): # Extract tasks mod, params = get_network(network) target = tvm.target.Target(target) tasks, task_weights = auto_scheduler.extract_tasks(mod["main"], params, target) with tempfile.NamedTemporaryFile() as fp: log_file = fp.name # Tuning measure_ctx = auto_scheduler.LocalRPCMeasureContext(timeout=60) tuner = auto_scheduler.TaskScheduler(tasks, task_weights) tune_option = auto_scheduler.TuningOptions( num_measure_trials=100, num_measures_per_round=2, early_stopping=1, runner=measure_ctx.runner, builder=auto_scheduler.LocalBuilder(timeout=60), measure_callbacks=[auto_scheduler.RecordToFile(log_file)], ) tuner.tune(tune_option, search_policy="sketch.random") del measure_ctx # Compile with the history best with auto_scheduler.ApplyHistoryBest(log_file): with tvm.transform.PassContext( opt_level=3, config={"relay.backend.use_auto_scheduler": True}): lib = relay.build(mod, target=target, params=params) # Compile without auto-scheduler and any other optimization for correctness check with tvm.transform.PassContext(opt_level=0): lib2 = relay.build(mod, target=target, params=params) # Check the correctness def get_output(data, lib): ctx = tvm.gpu() module = graph_runtime.GraphModule(lib["default"](ctx)) module.set_input("data", data) module.run() return module.get_output(0).asnumpy() np.random.seed(0) if network == "mlp": data = np.random.uniform(size=(1, 32)) elif network == "winograd-test": data = np.random.uniform(size=(1, 23, 40, 32)) else: raise ValueError("Unknown network: " + network) actual_output = get_output(data, lib) expected_output = get_output(data, lib2) tvm.testing.assert_allclose(actual_output, expected_output, rtol=1e-4, atol=1e-4)
def tune_network(network, target): auto_scheduler.enable_relay_integration() # Extract tasks mod, params = get_network(network) target = tvm.target.Target(target) tasks, task_weights = auto_scheduler.extract_tasks(mod["main"], params, target) with tempfile.NamedTemporaryFile() as fp: log_file = fp.name # Tuning measure_ctx = auto_scheduler.LocalRPCMeasureContext(timeout=60) tuner = auto_scheduler.TaskScheduler(tasks, task_weights) tune_option = auto_scheduler.TuningOptions( num_measure_trials=100, num_measures_per_round=2, early_stopping=1, runner=measure_ctx.runner, builder=auto_scheduler.LocalBuilder(timeout=60), measure_callbacks=[auto_scheduler.RecordToFile(log_file)], ) tuner.tune(tune_option, search_policy="sketch.random") del measure_ctx # Compile with the history best with auto_scheduler.ApplyHistoryBest(log_file): with tvm.transform.PassContext(opt_level=3): lib = relay.build(mod, target=target, params=params) # Todo(merrymercy): when the cpu backend is upstreamed, do the following things: # 1. compile without history to test the fallback mechanism # 2. check the correctness of layout rewrite / winograd pre-transform auto_scheduler.enable_relay_integration(False)
def tune_network(network, target): # Extract tasks mod, params = get_network(network) target = tvm.target.Target(target) tasks, task_weights = auto_scheduler.extract_tasks(mod["main"], params, target) with tempfile.NamedTemporaryFile() as fp: log_file = fp.name # Tuning measure_ctx = auto_scheduler.LocalRPCMeasureContext(timeout=60, device=0) tuner = auto_scheduler.TaskScheduler(tasks, task_weights, callbacks=[]) tune_option = auto_scheduler.TuningOptions( num_measure_trials=100, num_measures_per_round=2, early_stopping=1, runner=measure_ctx.runner, builder=auto_scheduler.LocalBuilder(timeout=60), measure_callbacks=[auto_scheduler.RecordToFile(log_file)], ) tuner.tune(tune_option, search_policy="sketch.random") del measure_ctx # Compile with the history best with auto_scheduler.ApplyHistoryBest(log_file): with tvm.transform.PassContext( opt_level=3, config={"relay.backend.use_auto_scheduler": True} ): lib = relay.build(mod, target=target, params=params) # Also test that multiple log files can be loaded. with auto_scheduler.ApplyHistoryBest([log_file, log_file]) as best: assert isinstance( best, auto_scheduler.dispatcher.ApplyHistoryBest ), "Unable to load multiple log files jointly." # Confirm iterables can be directly loaded. loaded_recs = auto_scheduler.dispatcher.load_records(log_file) with auto_scheduler.ApplyHistoryBest(iter(loaded_recs)) as best: assert isinstance( best, auto_scheduler.dispatcher.ApplyHistoryBest ), "Unable to ingest logs from an interator." # Sample a schedule when missing with auto_scheduler.ApplyHistoryBestOrSample(None, num_measure=2): with tvm.transform.PassContext( opt_level=3, config={"relay.backend.use_auto_scheduler": True} ): lib2 = relay.build(mod, target=target, params=params) # Compile without auto-scheduler and any other optimization for correctness check with tvm.transform.PassContext(opt_level=0): ref_lib = relay.build(mod, target=target, params=params) # Check the correctness def get_output(data, lib): dev = tvm.cuda() module = graph_executor.GraphModule(lib["default"](dev)) module.set_input("data", data) module.run() return module.get_output(0).numpy() np.random.seed(0) if network == "mlp": data = np.random.uniform(size=(1, 32)) elif network == "winograd-test": data = np.random.uniform(size=(1, 23, 40, 32)) else: raise ValueError("Unknown network: " + network) actual_output1 = get_output(data, lib) actual_output2 = get_output(data, lib2) expected_output = get_output(data, ref_lib) tvm.testing.assert_allclose(actual_output1, expected_output, rtol=1e-4, atol=1e-4) tvm.testing.assert_allclose(actual_output2, expected_output, rtol=1e-4, atol=1e-4)