def test_tuple_passing():
x = relay.var(
"x",
type_annotation=relay.ty.TupleType(
[relay.ty.TensorType((), "int64"), relay.ty.TensorType((), "int64")]
),
)
fn = relay.Function([x], relay.expr.TupleGetItem(x, 0))
mod = tvm.IRModule({})
gv = relay.GlobalVar("main")
mod[gv] = fn
mod = relay.transform.InferType()(mod)
dev = tvm.cpu()
target = tvm.target.Target("llvm")
exec = relay.create_executor(mod=mod, device=dev, target=target)
f = exec.evaluate(gv)
# First use a Python tuple.
out = f((10, 8))
tvm.testing.assert_allclose(out.numpy(), np.array(10))
# Second use a tuple value.
value_tuple = container.tuple_object([nd.array(np.array(11)), nd.array(np.array(12))])
out = f(value_tuple)
tvm.testing.assert_allclose(out.numpy(), np.array(11))
def test_tuple_passing():
x = relay.var('x',
type_annotation=relay.ty.TupleType([
relay.ty.TensorType((), 'int64'),
relay.ty.TensorType((), 'int64')
]))
fn = relay.Function([x], relay.expr.TupleGetItem(x, 0))
mod = relay.Module({})
gv = relay.GlobalVar('main')
mod[gv] = fn
mod = relay.transform.InferType()(mod)
ctx = tvm.cpu()
target = tvm.target.create('llvm')
exec = relay.create_executor(mod=mod, ctx=ctx, target=target)
f = exec.evaluate(gv)
# First use a Python tuple.
out = f((10, 8))
tvm.testing.assert_allclose(out.asnumpy(), np.array(10))
# Second use a tuple value.
value_tuple = container.tuple_object(
[nd.array(np.array(11)),
nd.array(np.array(12))])
out = f(value_tuple)
tvm.testing.assert_allclose(out.asnumpy(), np.array(11))
def adaptive_evaluator(epsilon, remote, build_result, measure_input, ref_input,
number, repeat, min_repeat_ms):
# print("####in adaptive evaluator###")
func = remote.load_module(os.path.split(build_result.filename)[1])
ctx = remote.context(str(measure_input.target), 0)
flop = measure_input.task.flop
# set input
if ref_input:
args = [nd.array(x, ctx=ctx) for x in ref_input]
else:
# create empty arrays on the remote device and copy them once.
# This can avoid some memory issues that make the measurement results unreliable.
args = [
nd.empty(x[0], dtype=x[1], ctx=ctx) for x in build_result.arg_info
]
args = [nd.array(x, ctx=ctx) for x in args]
ctx.sync()
# break the number*repeat into several batch
# print("number=%d, repeat=%d" % (number, repeat))
if repeat * number < 300: # no need to do adaptive evaluator
time_f = func.time_evaluator(func.entry_name,
ctx,
number=number,
repeat=repeat,
min_repeat_ms=min_repeat_ms)
eva_res = time_f(*args)
costs = eva_res.results
else:
b_size = 50
costs = []
sum_num = 0
max_iter = number * repeat
pis = []
bi = 1
rep = 0
flag = True
while flag and sum_num < max_iter:
time_f = func.time_evaluator(func.entry_name,
ctx,
number=b_size,
repeat=1,
min_repeat_ms=min_repeat_ms)
b_mean = time_f(*args).mean
# print("b_mean:" + str(b_mean))
costs.append(b_mean)
sum_num = sum_num + b_size
# calculate the flops of per batch: flop/time_mean_batch_i
pi = flop / b_mean
pis.append(pi)
pis_array = np.array(pis)
if len(pis_array) > 4: # remove the min and max to reduce variance
pis_array.sort()
pis_array = pis_array[1:-1]
# calculate the coefficient of variation
cv = pis_array.std() / pis_array.mean()
if bi > 2 and cv < epsilon:
# print("\nindex is %d, break at batch#%d, cv=%.10f, cost is %.8f." % (measure_input.config.index, bi, cv, b_mean))
flag = False
bi = bi + 1
return costs
def broadcast_to(children, attrs, odtype='float32'):
# TODO(@jroesch) export broadcast to?
data = children[0]
shape = attrs.get_int_tuple('shape')
array = numpy.zeros(shape).astype(odtype)
rconst = relay.Constant(nd.array(array))
return op.broadcast_to_like(data, rconst)
def set_task(self, task):
self.task = task
if check_remote(task.target, self.key, self.host, self.port):
logger.info("Get devices for measurement successfully!")
else:
raise RuntimeError(
"Cannot get remote devices from the tracker. "
"Please check the status of tracker by "
"'python -m tvm.exec.query_rpc_tracker --port [THE PORT YOU USE]' "
"and make sure you have free devices on the queue status.")
if self.check_correctness:
# use llvm cpu to generate a reference input/output
# this option works for tuning topi, but might not work for you custom op
with _target.create("llvm"):
s, arg_bufs = task.instantiate(task.config_space.get(0))
self.ref_input = [
np.random.uniform(size=get_const_tuple(x.shape)).astype(
x.dtype) for x in arg_bufs
]
func = build(s, arg_bufs, "llvm")
tvm_buf = [nd.array(x) for x in self.ref_input]
func(*tvm_buf)
self.ref_output = [x.asnumpy() for x in tvm_buf]
def test_adt_constructor():
arr = nd.array([1, 2, 3])
fields = [arr, arr]
y = _container.ADT(0, [arr, arr])
assert len(y) == 2
assert isinstance(y, _container.ADT)
y[0:1][-1] == arr
assert y.tag == 0
assert isinstance(arr, nd.NDArray)
def test_tuple_object():
x = relay.var('x',
type_annotation=relay.ty.TupleType([
relay.ty.TensorType((), 'int32'),
relay.ty.TensorType((), 'int32')
]))
fn = relay.Function([x], relay.expr.TupleGetItem(x, 0))
mod = tvm.IRModule.from_expr(fn)
exe = relay.create_executor(kind="vm",
mod=mod,
ctx=nd.cpu(),
target="llvm")
f = exe.evaluate()
value_tuple = _container.tuple_object(
[nd.array(np.array(11)),
nd.array(np.array(12))])
# pass an ADT object to evaluate
out = f(value_tuple)
tvm.testing.assert_allclose(out.asnumpy(), np.array(11))
def test_function_taking_adt_ref_tuple():
mod = tvm.IRModule()
prelude = relay.prelude.Prelude(mod)
intrp = create_executor("debug", mod)
_, cons, nil = prelude.mod.get_type("List")
nil_value = ConstructorValue(nil.tag, [], nil)
cons_value = ConstructorValue(
cons.tag,
[nd.array(np.random.rand(1, 10).astype("float32")), nil_value],
cons,
)
ref_value = RefValue(nd.array(np.random.rand(1, 10).astype("float32")))
tuple_value = container.tuple_object(
[nd.array(np.random.rand(1, 10).astype("float32")) for _ in range(10)])
id_func = intrp.evaluate(prelude.id)
res_nil = id_func(nil_value)
assert res_nil.tag == nil_value.tag
assert len(res_nil.fields) == 0
res_cons = id_func(cons_value)
assert res_cons.tag == cons_value.tag
assert len(res_cons.fields) == len(cons_value.fields)
tvm.testing.assert_allclose(res_cons.fields[0].asnumpy(),
cons_value.fields[0].asnumpy())
assert isinstance(res_cons.fields[1], ConstructorValue)
assert res_cons.fields[1].tag == nil.tag
assert len(res_cons.fields[1].fields) == 0
res_ref = id_func(ref_value)
tvm.testing.assert_allclose(res_ref.value.asnumpy(),
ref_value.value.asnumpy())
res_tuple = id_func(tuple_value)
for i in range(10):
tvm.testing.assert_allclose(res_tuple[i].asnumpy(),
tuple_value[i].asnumpy())
def run_through_rpc(measure_input,
build_result,
number,
repeat,
min_repeat_ms,
cooldown_interval,
remote_args,
ref_input=None,
ref_output=None):
"""Run a generated library through rpc
Parameters
----------
measure_input: MeasureInput
The raw measure input
build_result: BuildResult
The result returned from Builder. This contains the path to the generated library.
number: int
The number of times to run the generated code for taking average.
We call these runs as one `repeat` of measurement.
repeat : int, optional
The number of times to repeat the measurement.
In total, the generated code will be run (1 + number x repeat) times,
where the first one is warm up and will be discarded.
The returned result contains `repeat` costs,
each of which is an average of `number` costs.
min_repeat_ms: int, optional
The minimum duration of one `repeat` in milliseconds.
By default, one `repeat` contains `number` runs. If this parameter is set,
the parameters `number` will be dynamically adjusted to meet the
minimum duration requirement of one `repeat`.
i.e., When the run time of one `repeat` falls below this time, the `number` parameter
will be automatically increased.
cooldown_interval: float
The cool down interval between two measurements
remote_args: Tuple
The argument for request_remote
ref_input: List of np.ndarray
The reference input used for checking correctness
ref_output: List of np.ndarray
The reference output used for checking correctness
"""
if isinstance(build_result, MeasureResult):
return build_result
tic = time.time()
errno = MeasureErrorNo.NO_ERROR
try:
# upload built module
remote = request_remote(*remote_args)
# Program the FPGA every single time when targeting VTA
if hasattr(measure_input.target, 'device_name') and \
measure_input.target.device_name == 'vta':
# pylint: disable=import-outside-toplevel
from vta import program_fpga, reconfig_runtime
program_fpga(remote, None)
reconfig_runtime(remote)
remote.upload(build_result.filename)
func = remote.load_module(os.path.split(build_result.filename)[1])
ctx = remote.context(str(measure_input.target), 0)
time_f = func.time_evaluator(func.entry_name,
ctx,
number=number,
repeat=repeat,
min_repeat_ms=min_repeat_ms)
# set input
if ref_input:
args = [nd.array(x, ctx=ctx) for x in ref_input]
else:
# create empty arrays on the remote device and copy them once.
# This can avoid some memory issues that make the measurement results unreliable.
args = [
nd.empty(x[0], dtype=x[1], ctx=ctx)
for x in build_result.arg_info
]
args = [nd.array(x, ctx=ctx) for x in args]
ctx.sync()
costs = time_f(*args).results
# clean up remote files
remote.remove(build_result.filename)
remote.remove(os.path.splitext(build_result.filename)[0] + '.so')
remote.remove('')
if len(costs
) > 2: # remove largest and smallest value to reduce variance
costs = list(costs)
costs.sort()
costs = tuple(costs[1:-1])
# check correctness of output
if ref_output:
for expected, real in zip(ref_output, args):
if not np.allclose(expected, real.asnumpy(), rtol=1e-4):
logger.warning("Wrong Answer!")
errno = MeasureErrorNo.WRONG_ANSWER
except TVMError as exc:
msg = str(exc)
if "Stack trace returned" in msg:
msg = msg[:msg.index("Stack trace returned")]
if "CUDA Source" in msg:
msg = msg[:msg.index("CUDA Source")]
costs = (RuntimeError(msg[:1024]), )
errno = MeasureErrorNo.RUNTIME_DEVICE
tstamp = time.time()
time.sleep(cooldown_interval)
return MeasureResult(costs, errno, tstamp - tic + build_result.time_cost,
tstamp)
def extract_task_from_relay(
mod: IRModule,
target: Target,
params: Optional[Dict[str, NDArray]] = None,
*,
opt_level: int = 3,
pass_config: Optional[Dict[str, Any]] = None,
disabled_pass: Optional[List[str]] = None,
) -> List[ExtractedTask]:
"""Extract tuning tasks from a relay program.
Parameters
----------
mod : IRModule
The module or function to tune
target : tvm.target.Target
The compilation target
params : Optional[Dict[str, tvm.runtime.NDArray]]
The associated parameters of the program
opt_level : int
The optimization level of the compiler
pass_config : Optional[Dict[str, Any]]
The pass config of the compiler
disabled_pass : Optional[List[str]]
The list of disabled passes of the compiler
Returns
-------
tasks: List[ExtractedTask]
The tasks extracted from this network
"""
# pylint: disable=import-outside-toplevel
from tvm.relay import Function as RelayFunc
# pylint: enable=import-outside-toplevel
extract_task_func = get_global_func(
"relay.backend.MetaScheduleExtractTask",
allow_missing=False,
)
if isinstance(mod, RelayFunc):
mod = IRModule.from_expr(mod)
if not isinstance(target, Target):
target = Target(target)
if disabled_pass is None:
disabled_pass = []
if pass_config is None:
pass_config = {"relay.backend.use_meta_schedule": True}
if params is None:
params = {}
relay_params = {}
for name, param in params.items():
if isinstance(param, np.ndarray):
param = nd.array(param)
relay_params[name] = param
with autotvm_silencer(), target, transform.PassContext(
opt_level=opt_level,
config=pass_config,
disabled_pass=disabled_pass,
):
return list(extract_task_func(mod, target, relay_params))
def extract_task_from_relay(
mod: IRModule,
target: Target,
params: Optional[Dict[str, NDArray]] = None,
*,
opt_level: int = 3,
pass_config: Optional[Dict[str, Any]] = None,
disabled_pass: Optional[List[str]] = None,
te_filter_func: Union[str, None, Callable[[List[Tensor]],
PrimFunc]] = None,
) -> List[ExtractedTask]:
"""Extract tuning tasks from a relay program.
Parameters
----------
mod : IRModule
The module or function to tune
target : tvm.target.Target
The compilation target
params : Optional[Dict[str, tvm.runtime.NDArray]]
The associated parameters of the program
opt_level : int
The optimization level of the compiler
pass_config : Optional[Dict[str, Any]]
The pass config of the compiler
disabled_pass : Optional[List[str]]
The list of disabled passes of the compiler
te_filter_func : Callable[[List[tvm.te.Tensor]], bool]
The filter function to filter out the extracted tasks
If it's a string, it's the name of the filtering function. Built in functions are
- "meta_schedule.DefaultTaskFilter"
- "meta_schedule.DefaultTaskFilterAllowExtern"
If it's None, it's the default filtering function
If it's a callable, it's the filtering function
Returns
-------
tasks: List[ExtractedTask]
The tasks extracted from this network
"""
# pylint: disable=import-outside-toplevel
from tvm import autotvm
from tvm.relay import Function as RelayFunc
# pylint: enable=import-outside-toplevel
if isinstance(te_filter_func, str):
te_filter_func = get_global_func(te_filter_func)
extract_task_func = get_global_func(
"relay.backend.MetaScheduleExtractTask",
allow_missing=False,
)
if isinstance(mod, RelayFunc):
mod = IRModule.from_expr(mod)
if not isinstance(target, Target):
target = Target(target)
if disabled_pass is None:
disabled_pass = []
if pass_config is None:
pass_config = {"relay.backend.use_meta_schedule": True}
if params is None:
params = {}
relay_params = {}
for name, param in params.items():
if isinstance(param, np.ndarray):
param = nd.array(param)
relay_params[name] = param
with target, autotvm_silencer(), transform.PassContext(
opt_level=opt_level,
config=pass_config,
disabled_pass=disabled_pass,
):
if target.kind.name != "cuda" and isinstance(
autotvm.DispatchContext.current, autotvm.FallbackContext):
tophub_context = autotvm.tophub.context(target)
else:
tophub_context = autotvm.utils.EmptyContext()
with tophub_context:
return list(
extract_task_func(mod, target, relay_params, te_filter_func))
def run_through_rpc(
measure_input,
build_result,
number,
repeat,
min_repeat_ms,
cooldown_interval,
remote_args,
ref_input=None,
ref_output=None,
enable_cpu_cache_flush=False,
):
"""Run a generated library through rpc
Parameters
----------
measure_input: MeasureInput
The raw measure input
build_result: BuildResult
The result returned from Builder. This contains the path to the generated library.
number: int
The number of times to run the generated code for taking average.
We call these runs as one `repeat` of measurement.
repeat : int, optional
The number of times to repeat the measurement.
In total, the generated code will be run (1 + number x repeat) times,
where the first one is warm up and will be discarded.
The returned result contains `repeat` costs,
each of which is an average of `number` costs.
min_repeat_ms: int, optional
The minimum duration of one `repeat` in milliseconds.
By default, one `repeat` contains `number` runs. If this parameter is set,
the parameters `number` will be dynamically adjusted to meet the
minimum duration requirement of one `repeat`.
i.e., When the run time of one `repeat` falls below this time, the `number` parameter
will be automatically increased.
cooldown_interval: float
The cool down interval between two measurements
remote_args: Tuple
The argument for request_remote
ref_input: List of np.ndarray
The reference input used for checking correctness
ref_output: List of np.ndarray
The reference output used for checking correctness
enable_cpu_cache_flush: bool
Whether to flush cache on CPU between repeated measurements.
Flushing cache can make the measured latency of one operator closer to
its actual latency during end-to-end inference.
To make this option effective, the argument `number` should also be set to 1.
This is only has effect on CPU task.
"""
if isinstance(build_result, MeasureResult):
return build_result
tic = time.time()
errno = MeasureErrorNo.NO_ERROR
try:
# upload built module
remote = request_remote(*remote_args)
# Program the FPGA every single time when targeting VTA
if (
hasattr(measure_input.target, "device_name")
and measure_input.target.device_name == "vta"
):
# pylint: disable=import-outside-toplevel
from vta import program_fpga, reconfig_runtime
program_fpga(remote, None)
reconfig_runtime(remote)
remote.upload(build_result.filename)
func = remote.load_module(os.path.split(build_result.filename)[1])
ctx = remote.context(str(measure_input.target), 0)
# Limitation:
# We can not get PackFunction directly in the remote mode as it is wrapped
# under the std::function. We could lift the restriction later once we fold
# the PackedFunc as an object. Currently, we pass function name to work
# around it.
f_prepare = "cache_flush_cpu_non_first_arg" if enable_cpu_cache_flush else ""
time_f = func.time_evaluator(
func.entry_name,
ctx,
number=number,
repeat=repeat,
min_repeat_ms=min_repeat_ms,
f_preproc=f_prepare,
)
# set input
if ref_input:
args = [nd.array(x, ctx=ctx) for x in ref_input]
else:
try:
random_fill = remote.get_function("tvm.contrib.random.random_fill")
except AttributeError:
raise AttributeError(
"Please make sure USE_RANDOM is ON in the config.cmake " "on the remote devices"
)
args = [nd.empty(x[0], dtype=x[1], ctx=ctx) for x in build_result.arg_info]
for arg in args:
random_fill(arg)
ctx.sync()
costs = time_f(*args).results
# clean up remote files
remote.remove(build_result.filename)
remote.remove(os.path.splitext(build_result.filename)[0] + ".so")
remote.remove("")
if len(costs) > 2: # remove largest and smallest value to reduce variance
costs = list(costs)
costs.sort()
costs = tuple(costs[1:-1])
# check correctness of output
if ref_output:
for expected, real in zip(ref_output, args):
if not np.allclose(expected, real.asnumpy(), rtol=1e-4):
logger.warning("Wrong Answer!")
errno = MeasureErrorNo.WRONG_ANSWER
except TVMError as exc:
msg = str(exc)
if "Stack trace returned" in msg:
msg = msg[: msg.index("Stack trace returned")]
if "CUDA Source" in msg:
msg = msg[: msg.index("CUDA Source")]
costs = (RuntimeError(msg[:1024]),)
errno = MeasureErrorNo.RUNTIME_DEVICE
tstamp = time.time()
time.sleep(cooldown_interval)
return MeasureResult(costs, errno, tstamp - tic + build_result.time_cost, tstamp)
def run_through_rpc(
measure_input,
build_result,
number,
repeat,
min_repeat_ms,
cooldown_interval,
remote_kwargs,
ref_input,
enable_cpu_cache_flush=False,
module_loader=None,
):
"""Run a generated library through rpc
Parameters
----------
measure_input: MeasureInput
The raw measure input
build_result: BuildResult
The result returned from Builder. This contains the path to the generated library.
number: int
The number of times to run the generated code for taking average.
We call these runs as one `repeat` of measurement.
repeat : int, optional
The number of times to repeat the measurement.
In total, the generated code will be run (1 + number x repeat) times,
where the first one is warm up and will be discarded.
The returned result contains `repeat` costs,
each of which is an average of `number` costs.
min_repeat_ms: int, optional
The minimum duration of one `repeat` in milliseconds.
By default, one `repeat` contains `number` runs. If this parameter is set,
the parameters `number` will be dynamically adjusted to meet the
minimum duration requirement of one `repeat`.
i.e., When the run time of one `repeat` falls below this time, the `number` parameter
will be automatically increased.
cooldown_interval: float
The cool down interval between two measurements
remote_kwargs: dict
Passed to module_loader(). Ultimately, keyword args to request_remote().
ref_input: List of np.ndarray
The reference input used for tuning. Empty for randomly filled input.
enable_cpu_cache_flush: bool
Whether to flush cache on CPU between repeated measurements.
Flushing cache can make the measured latency of one operator closer to
its actual latency during end-to-end inference.
To make this option effective, the argument `number` should also be set to 1.
This is only has effect on CPU task.
module_loader: ModuleLoader
A function that returns a ContextManager used to establish and teardown the remote session.
"""
if isinstance(build_result, MeasureResult):
return build_result
tic = time.time()
errno = MeasureErrorNo.NO_ERROR
try:
# upload built module
with module_loader(remote_kwargs, build_result) as (remote, mod):
dev = remote.device(str(measure_input.target), 0)
# Limitation:
# We can not get PackFunction directly in the remote mode as it is wrapped
# under the std::function. We could lift the restriction later once we fold
# the PackedFunc as an object. Currently, we pass function name to work
# around it.
f_prepare = "cache_flush_cpu_non_first_arg" if enable_cpu_cache_flush else ""
time_f = mod.time_evaluator(
mod.entry_name,
dev,
number=number,
repeat=repeat,
min_repeat_ms=min_repeat_ms,
f_preproc=f_prepare,
)
if ref_input:
args = [nd.array(x, device=dev) for x in ref_input]
else:
try:
random_fill = remote.get_function(
"tvm.contrib.random.random_fill")
except AttributeError:
raise AttributeError(
"Please make sure USE_RANDOM is ON in the config.cmake "
"on the remote devices")
args = [
nd.empty(x[0], x[1], dev) for x in build_result.arg_info
]
if "scatter" not in measure_input.task.name:
# the index tensor of scatter op cannot be randomly initialized
for arg in args:
random_fill(arg)
dev.sync()
costs = time_f(*args).results
if len(costs
) > 2: # remove largest and smallest value to reduce variance
costs = list(costs)
costs.sort()
costs = tuple(costs[1:-1])
except TVMError as exc:
msg = str(exc)
if "Stack trace returned" in msg:
msg = msg[:msg.index("Stack trace returned")]
if "CUDA Source" in msg:
msg = msg[:msg.index("CUDA Source")]
costs = (RuntimeError(msg[:1024]), )
errno = MeasureErrorNo.RUNTIME_DEVICE
tstamp = time.time()
time.sleep(cooldown_interval)
return MeasureResult(costs, errno, tstamp - tic + build_result.time_cost,
tstamp)