tvm.autotvm

The auto-tuning module of tvm

This module includes:

  • Tuning space definition API

  • Efficient auto-tuners

  • Tuning result and database support

  • Distributed measurement to scale up tuning

tvm.autotvm.apply_history_best(records)

Apply the history best config

Parameters

records (str or iterator of (autotvm.measure.MeasureInput, autotvm.measure.MeasureResult)) – Collection of tuning records. If is str, then it should be the filename of a records log file. Each row of this file is an encoded record pair. Otherwise, it is an iterator.

tvm.autotvm.measure

User facing API for specifying how to measure the generated code

class tvm.autotvm.measure.MeasureInput(target, task, config)

Stores all the necessary inputs for a measurement.

Parameters
class tvm.autotvm.measure.MeasureResult(costs, error_no, all_cost, timestamp)

Stores all the results of a measurement

Parameters
  • costs (Array of float or Array of Exception) – If no error occurs during measurement, it is an array of measured running times. If an error occurs during measurement, it is an array of the exception objections.

  • error_no (int) – Denote error type, defined by MeasureErrorNo

  • all_cost (float) – All cost of this measure, including rpc, compilation, test runs

  • timestamp (float) – The absolute time stamp when we finish measurement.

tvm.autotvm.measure.measure_option(builder, runner)

Set options for measure. To measure a config, we will build it and run it. So we have to set options for these two steps. They have their own options on timeout, parallel, etc.

Parameters
  • builder (Builder) – Specify how to build programs

  • runner (Runner) – Specify how to run programs

Examples

# example setting for using local devices >>> measure_option = autotvm.measure_option( >>> builder=autotvm.LocalBuilder(), # use all local cpu cores for compilation >>> runner=autotvm.LocalRunner( # measure them sequentially >>> number=10, >>> timeout=5) >>> )

# example setting for using remote devices >>> measure_option = autotvm.measure_option( >>> builder=autotvm.LocalBuilder(), # use all local cpu cores for compilation >>> runner=autotvm.RPCRunner( >>> ‘rasp3b’, ‘locahost’, 9190, # device key, host and port of the rpc tracker >>> number=4, >>> timeout=4) # timeout of a run on the device. RPC request waiting time is excluded. >>>)

Note

To make measurement results accurate, you should pick the correct value for the argument number and repeat in Runner(). Some devices need a certain minimum running time to “warm up,” such as GPUs that need time to reach a performance power state. Using min_repeat_ms can dynamically adjusts number, so it is recommended. The typical value for NVIDIA GPU is 150 ms.

tvm.autotvm.measure.create_measure_batch(task, option)

Get a standard measure_batch function.

Parameters
  • task (tvm.autotvm.task.Task) – The tuning task

  • option (dict) – The option for measuring generated code. You should use the return value of function measure_option for this argument.

Returns

measure_batch – a callback function to measure a batch of configs

Return type

callable

class tvm.autotvm.measure.measure_methods.LocalBuilder(timeout=10, n_parallel=None, build_kwargs=None, build_func='default', do_fork=False)

Run compilation on local machine

Parameters
  • timeout (float) – The timeout of a compilation

  • n_parallel (int) – The number of tasks run in parallel. “None” will use all cpu cores

  • build_kwargs (dict) – If supplied, additional kwargs passed to build_func. Overrides any build_kwargs supplied by the Runner.

  • build_func (callable or str) – If is ‘default’, use default build function If is ‘ndk’, use function for android ndk If id ‘stackvm’, use function for stackvm If is callable, use it as custom build function, expect lib_format field.

  • do_fork (bool) – If False, do not fork when building. Requires n_parallel=1.

class tvm.autotvm.measure.measure_methods.RPCRunner(key, host, port, priority=1, timeout=10, n_parallel=None, number=4, repeat=3, min_repeat_ms=0, cooldown_interval=0.1, enable_cpu_cache_flush=False, module_loader=None)

Run generated code on remove devices. This function will ask a RPC Tracker to get device for measurement.

Parameters
  • timeout (float) – The timeout of a RPCRunner measurement task

  • n_parallel (int) – The number of tasks run in parallel. “None” will use all cpu cores

  • key (str) – The key of the device registered in the tracker

  • host (str) – The host address of RPC Tracker

  • port (int) – The port of RPC Tracker

  • 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 “1” 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, optional) – The cool down interval between two measurements.

  • 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) – If given, a context manager that loads the module to be timed into the remote runtime. If not given, default_module_loader is used.

class tvm.autotvm.measure.measure_methods.LocalRunner(timeout=10, number=4, repeat=3, min_repeat_ms=0, cooldown_interval=0.1, enable_cpu_cache_flush=False, module_loader=None)

Run generated code on local devices.

Parameters
  • timeout (float) – The timeout of a compilation

  • 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, optional) – The cool down interval between two measurements.

  • 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.

Note

This is a “fake” local mode. We start a silent rpc tracker and rpc server for the user. In this way we reuse timeout/isolation mechanism in RPC infrastructure.

tvm.autotvm.tuner

A tuner takes a task as input. It proposes some promising ConfigEntity in the ConfigSpace and measure them on the real hardware. Then it proposed the next batch of ConfigEntity according to the measure results. This tuning loop is repeated.

class tvm.autotvm.tuner.Tuner(task, **kwargs)

Base class for tuners

Parameters

task (autotvm.task.Task) – Tuning Task

has_next()

Whether has next untried config in the space

Returns

has_next

Return type

bool

next_batch(batch_size)

get the next batch of configs to be measure on real hardware

Parameters

batch_size (int) – The size of the batch

Returns

Return type

a batch of configs

update(inputs, results)

Update parameters of the tuner according to measurement results

Parameters
  • inputs (Array of autotvm.measure.MeasureInput) – The input for measurement

  • results (Array of autotvm.measure.MeasureResult) – result for measurement

tune(n_trial, measure_option, early_stopping=None, callbacks=(), si_prefix='G')

Begin tuning

Parameters
  • n_trial (int) – Maximum number of configs to try (measure on real hardware)

  • measure_option (dict) – The options for how to measure generated code. You should use the return value ot autotvm.measure_option for this argument.

  • early_stopping (int, optional) – Early stop the tuning when not finding better configs in this number of trials

  • callbacks (List of callable) – A list of callback functions. The signature of callback function is (Tuner, List of MeasureInput, List of MeasureResult) with no return value. These callback functions will be called on every measurement pair. See autotvm/tuner/callback.py for some examples.

  • si_prefix (str) – One of tvm.autotvm.utils.SI_PREFIXES. The SI prefix to use when reporting FLOPS.

reset()

reset the status of tuner

load_history(data_set, min_seed_records=500)

load history data for transfer learning

Parameters
  • data_set (Array of (autotvm.measure.MeasureInput, autotvm.measure.MeasureResult) pair) – Previous tuning records

  • min_seed_records (int) – Defaults to 500. Indicates the minimum number of records to train the tuner with. If there are less than min_seed_records number of records in data_set, no training of the tuner will be done.

class tvm.autotvm.tuner.RandomTuner(task, range_idx=None)

Enumerate the search space in a random order

Parameters
  • task (autotvm.task.Task) – Tuning Task

  • range_idx (Optional[Tuple[int, int]]) – A tuple of index range to random

has_next()

Whether has next untried config in the space

Returns

has_next

Return type

bool

load_history(data_set, min_seed_records=500)

load history data for transfer learning

Parameters
  • data_set (Array of (autotvm.measure.MeasureInput, autotvm.measure.MeasureResult) pair) – Previous tuning records

  • min_seed_records (int) – Defaults to 500. Indicates the minimum number of records to train the tuner with. If there are less than min_seed_records number of records in data_set, no training of the tuner will be done.

reset()

reset the status of tuner

tune(n_trial, measure_option, early_stopping=None, callbacks=(), si_prefix='G')

Begin tuning

Parameters
  • n_trial (int) – Maximum number of configs to try (measure on real hardware)

  • measure_option (dict) – The options for how to measure generated code. You should use the return value ot autotvm.measure_option for this argument.

  • early_stopping (int, optional) – Early stop the tuning when not finding better configs in this number of trials

  • callbacks (List of callable) – A list of callback functions. The signature of callback function is (Tuner, List of MeasureInput, List of MeasureResult) with no return value. These callback functions will be called on every measurement pair. See autotvm/tuner/callback.py for some examples.

  • si_prefix (str) – One of tvm.autotvm.utils.SI_PREFIXES. The SI prefix to use when reporting FLOPS.

update(inputs, results)

Update parameters of the tuner according to measurement results

Parameters
  • inputs (Array of autotvm.measure.MeasureInput) – The input for measurement

  • results (Array of autotvm.measure.MeasureResult) – result for measurement

next_batch(batch_size)

get the next batch of configs to be measure on real hardware

Parameters

batch_size (int) – The size of the batch

Returns

Return type

a batch of configs

class tvm.autotvm.tuner.GridSearchTuner(task, range_idx=None)

Enumerate the search space in a grid search order

next_batch(batch_size)

get the next batch of configs to be measure on real hardware

Parameters

batch_size (int) – The size of the batch

Returns

Return type

a batch of configs

has_next()

Whether has next untried config in the space

Returns

has_next

Return type

bool

load_history(data_set, min_seed_records=500)

load history data for transfer learning

Parameters
  • data_set (Array of (autotvm.measure.MeasureInput, autotvm.measure.MeasureResult) pair) – Previous tuning records

  • min_seed_records (int) – Defaults to 500. Indicates the minimum number of records to train the tuner with. If there are less than min_seed_records number of records in data_set, no training of the tuner will be done.

reset()

reset the status of tuner

tune(n_trial, measure_option, early_stopping=None, callbacks=(), si_prefix='G')

Begin tuning

Parameters
  • n_trial (int) – Maximum number of configs to try (measure on real hardware)

  • measure_option (dict) – The options for how to measure generated code. You should use the return value ot autotvm.measure_option for this argument.

  • early_stopping (int, optional) – Early stop the tuning when not finding better configs in this number of trials

  • callbacks (List of callable) – A list of callback functions. The signature of callback function is (Tuner, List of MeasureInput, List of MeasureResult) with no return value. These callback functions will be called on every measurement pair. See autotvm/tuner/callback.py for some examples.

  • si_prefix (str) – One of tvm.autotvm.utils.SI_PREFIXES. The SI prefix to use when reporting FLOPS.

update(inputs, results)

Update parameters of the tuner according to measurement results

Parameters
  • inputs (Array of autotvm.measure.MeasureInput) – The input for measurement

  • results (Array of autotvm.measure.MeasureResult) – result for measurement

class tvm.autotvm.tuner.GATuner(task, pop_size=100, elite_num=3, mutation_prob=0.1)

Tuner with genetic algorithm. This tuner does not have a cost model so it always run measurement on real machines. This tuner expands the ConfigEntity as gene.

Parameters
  • pop_size (int) – number of genes in one generation

  • elite_num (int) – number of elite to keep

  • mutation_prob (float) – probability of mutation of a knob in a gene

next_batch(batch_size)

get the next batch of configs to be measure on real hardware

Parameters

batch_size (int) – The size of the batch

Returns

Return type

a batch of configs

update(inputs, results)

Update parameters of the tuner according to measurement results

Parameters
  • inputs (Array of autotvm.measure.MeasureInput) – The input for measurement

  • results (Array of autotvm.measure.MeasureResult) – result for measurement

has_next()

Whether has next untried config in the space

Returns

has_next

Return type

bool

reset()

reset the status of tuner

tune(n_trial, measure_option, early_stopping=None, callbacks=(), si_prefix='G')

Begin tuning

Parameters
  • n_trial (int) – Maximum number of configs to try (measure on real hardware)

  • measure_option (dict) – The options for how to measure generated code. You should use the return value ot autotvm.measure_option for this argument.

  • early_stopping (int, optional) – Early stop the tuning when not finding better configs in this number of trials

  • callbacks (List of callable) – A list of callback functions. The signature of callback function is (Tuner, List of MeasureInput, List of MeasureResult) with no return value. These callback functions will be called on every measurement pair. See autotvm/tuner/callback.py for some examples.

  • si_prefix (str) – One of tvm.autotvm.utils.SI_PREFIXES. The SI prefix to use when reporting FLOPS.

load_history(data_set, min_seed_records=500)

load history data for transfer learning

Parameters
  • data_set (Array of (autotvm.measure.MeasureInput, autotvm.measure.MeasureResult) pair) – Previous tuning records

  • min_seed_records (int) – Defaults to 500. Indicates the minimum number of records to train the tuner with. If there are less than min_seed_records number of records in data_set, no training of the tuner will be done.

class tvm.autotvm.tuner.XGBTuner(task, plan_size=64, feature_type='itervar', loss_type='rank', num_threads=None, optimizer='sa', diversity_filter_ratio=None, log_interval=50)

Tuner that uses xgboost as cost model

Parameters
  • task (Task) – The tuning task

  • plan_size (int) – The size of a plan. After plan_size trials, the tuner will refit a new cost model and do planing for the next plan_size trials.

  • feature_type (str, optional) –

    If is ‘itervar’, use features extracted from IterVar (loop variable). If is ‘knob’, use flatten ConfigEntity directly. If is ‘curve’, use sampled curve feature (relation feature).

    Note on choosing feature type: For single task tuning, ‘itervar’ and ‘knob’ are good. ‘itervar’ is more accurate but ‘knob’ is much faster. There are some constraints on ‘itervar’, if you meet problems with feature extraction when using ‘itervar’, you can switch to ‘knob’.

    For cross-shape tuning (e.g. many convolutions with different shapes), ‘itervar’ and ‘curve’ has better transferability, ‘knob’ is faster.

    For cross-device or cross-operator tuning, you can use ‘curve’ only.

  • loss_type (str) – If is ‘reg’, use regression loss to train cost model. The cost model predicts the normalized flops. If is ‘rank’, use pairwise rank loss to train cost model. The cost model predicts relative rank score.

  • num_threads (int, optional) – The number of threads.

  • optimizer (str or ModelOptimizer, optional) – If is ‘sa’, use a default simulated annealing optimizer. Otherwise it should be a ModelOptimizer object.

  • diversity_filter_ratio (int or float, optional) – If is not None, the tuner will first select top-(plan_size * diversity_filter_ratio) candidates according to the cost model and then pick batch_size of them according to the diversity metric.

  • log_interval (int = 50) – The verbose level. If is 0, output nothing. Otherwise, output debug information every verbose iterations.

tune(*args, **kwargs)

Begin tuning

Parameters
  • n_trial (int) – Maximum number of configs to try (measure on real hardware)

  • measure_option (dict) – The options for how to measure generated code. You should use the return value ot autotvm.measure_option for this argument.

  • early_stopping (int, optional) – Early stop the tuning when not finding better configs in this number of trials

  • callbacks (List of callable) – A list of callback functions. The signature of callback function is (Tuner, List of MeasureInput, List of MeasureResult) with no return value. These callback functions will be called on every measurement pair. See autotvm/tuner/callback.py for some examples.

  • si_prefix (str) – One of tvm.autotvm.utils.SI_PREFIXES. The SI prefix to use when reporting FLOPS.

has_next()

Whether has next untried config in the space

Returns

has_next

Return type

bool

load_history(data_set, min_seed_records=500)

load history data for transfer learning

Parameters
  • data_set (Array of (autotvm.measure.MeasureInput, autotvm.measure.MeasureResult) pair) – Previous tuning records

  • min_seed_records (int) – Defaults to 500. Indicates the minimum number of records to train the tuner with. If there are less than min_seed_records number of records in data_set, no training of the tuner will be done.

next_batch(batch_size)

get the next batch of configs to be measure on real hardware

Parameters

batch_size (int) – The size of the batch

Returns

Return type

a batch of configs

reset()

reset the status of tuner

update(inputs, results)

Update parameters of the tuner according to measurement results

Parameters
  • inputs (Array of autotvm.measure.MeasureInput) – The input for measurement

  • results (Array of autotvm.measure.MeasureResult) – result for measurement

Namespace of callback utilities of AutoTVM

tvm.autotvm.tuner.callback.log_to_file(file_out, protocol='json')

Log the tuning records into file. The rows of the log are stored in the format of autotvm.record.encode.

Parameters
  • file_out (File or str) – The file to log to.

  • protocol (str, optional) – The log protocol. Can be ‘json’ or ‘pickle’

Returns

callback – Callback function to do the logging.

Return type

callable

tvm.autotvm.tuner.callback.log_to_database(db)

Save the tuning records to a database object.

Parameters

db (Database) – The database

class tvm.autotvm.tuner.callback.Monitor

A monitor to collect statistic during tuning

trial_scores()

get scores (currently is flops) of all trials

trial_timestamps()

get wall clock time stamp of all trials

tvm.autotvm.tuner.callback.progress_bar(total, prefix='', si_prefix='G')

Display progress bar for tuning

Parameters
  • total (int) – The total number of trials

  • prefix (str) – The prefix of output message

  • si_prefix (str) – SI prefix for flops

tvm.autotvm.task

Task is a tunable composition of template functions.

Tuner takes a tunable task and optimizes the joint configuration space of all the template functions in the task. This module defines the task data structure, as well as a collection(zoo) of typical tasks of interest.

Definition of task function.

Task can be constructed from tuple of func, args, and kwargs. func is a state-less function, or a string that registers the standard task.

tvm.autotvm.task.task.serialize_args(args)

serialize arguments of a topi function to a hashable tuple.

Parameters

args (list of hashable or Tensor) –

tvm.autotvm.task.task.deserialize_args(args)

The inverse function of serialize_args.

Parameters

args (list of hashable or Tensor) –

tvm.autotvm.task.task.args_to_workload(args, task_name=None)

Convert argument list to hashable workload tuple. This function will convert list to tuple, tvm node to python value and flatten te.tensor.Tensor to a tuple

Parameters
  • task_name (str) – The AutoTVM task name

  • args (list of args) – The arguments to the function

Returns

ret – The hashable value

Return type

hashable

class tvm.autotvm.task.task.Task(name, args)

A Tunable Task

Parameters
  • name (str) – The name of the task.

  • args (Tuple) – Positional argument of func

instantiate(config)

Instantiate this task function (template) with a config. Returns corresponding schedule.

Parameters

config (template.ConfigEntity) – parameter config for this template

Returns

  • sch (tvm.te.schedule.Schedule) – The tvm schedule

  • arg_bufs (Array of te.tensor.Tensor) – The input/output buffers

class tvm.autotvm.task.task.TaskTemplate

Task template is used to creates a tunable AutoTVM task.

It can be defined by a pair of compute and schedule function using _register_task_compute and _register_task_schedule, or by a customized task creation function that is more flexible using _register_customized_task.

Note that when customized func is registered, compute and schedule function will be ignored

class tvm.autotvm.task.task.MissingTask(taskname: str)

Dummy task template for a task lookup which cannot be resolved. This can occur if the task being requested from _lookup_task() has not been imported in this run.

tvm.autotvm.task.task.template(task_name, func=None)

Decorate a function as a tunable schedule template.

Parameters
  • task_name (str) – The task name

  • func (None or callable) – A callable template function. If it is None, return a decorator. If is callable, decorate this function.

Returns

func – The decorated function

Return type

callable

Examples

The following code is a tunable template for a blocked matrix multiplication

@autotvm.template("matmul")
def matmul(N, L, M, dtype):
    A = te.placeholder((N, L), name='A', dtype=dtype)
    B = te.placeholder((L, M), name='B', dtype=dtype)

    k = te.reduce_axis((0, L), name='k')
    C = te.compute((N, M), lambda i, j: te.sum(A[i, k] * B[k, j], axis=k), name='C')
    s = te.create_schedule(C.op)

    # schedule
    y, x = s[C].op.axis
    k = s[C].op.reduce_axis[0]

    ##### define space begin #####
    cfg = autotvm.get_config()
    cfg.define_split("tile_y", y, num_outputs=2)
    cfg.define_split("tile_x", x, num_outputs=2)
    ##### define space end #####

    # schedule according to config
    yo, yi = cfg["tile_y"].apply(s, C, y)
    xo, xi = cfg["tile_x"].apply(s, C, x)

    s[C].reorder(yo, xo, k, yi, xi)

    return s, [A, B, C]
tvm.autotvm.task.task.create(task_name, args, target, target_host=None)

Create a tuning task and initialize its search space

Parameters
  • task_name (str) – The AutoTVM task name

  • args (List) – Positional arguments

  • target (Target) – The compilation target

  • target_host (Target, optional) – The compilation target for host side

Returns

tsk – a task object

Return type

Task

tvm.autotvm.task.task.get_config()

Get current config object

Returns

cfg – The current config

Return type

ConfigSpace or ConfigEntity

exception tvm.autotvm.task.task.FlopCalculationError

Error happens when estimating FLOP for a compute op

tvm.autotvm.task.task.compute_flop(sch)

Calculate number of FLOP (floating number operations) of the compute ops in a schedule

Parameters

sch (tvm.te.schedule.Schedule) – schedule

Returns

flop – number of FLOP in this schedule

Return type

int

Template configuration space.

Each template function can be parameterized by a ConfigSpace. The space is declared when we invoke the template function with ConfigSpace. During evaluation, we pass in a ConfigEntity, which contains a specific entity in the space. This entity contains deterministic parameters.

class tvm.autotvm.task.space.Axis(space, index)
property index

Alias for field number 1

property space

Alias for field number 0

exception tvm.autotvm.task.space.InstantiationError

Actively detected error in instantiating a template with a config, raised by cfg.raise_error e.g. too many unrolling, too many threads in a block

class tvm.autotvm.task.space.TransformSpace

Base class for transform space TransformSpace is the node in the computation graph of axes

Note

We can regard our schedule code as a transformation graph of axes. Starting from raw axes in the definition of te.compute, we can transform these axes by some operators. The operator includes ‘split’, ‘reorder’ and ‘annotate’. Each operator has some tunable parameters (e.g. the split factor). Then the tuning process is just to find good parameters of these op.

So all the combinations of the parameters of these op form our search space.

Naming convention: We call the set of all possible values as XXXSpace. (XXX can be Split, Reorder, Config …) We call a specific entity in a space as XXXEntity.

static get_num_output()

get number of output axes after this transform

Returns

n – number of output axes

Return type

int

class tvm.autotvm.task.space.VirtualAxis(var, name=None)

Axis placeholder in template

Parameters
  • var (int or tvm.te.schedule.IterVar) – If is int, return a virtual axis whose length is the provided argument. If is IterVar, return a virtual axis whose length is extracted from the IterVar’s extent domain.

  • name (str) –

static get_num_output(var, name=None)

get number of output axes after this transform

Returns

n – number of output axes

Return type

int

tvm.autotvm.task.space.get_factors(n)

return all factors of an integer

Parameters

n (int) – integer to factorize

Returns

factors – List of all factors

Return type

list

tvm.autotvm.task.space.get_pow2s(n)

return all power-of-two numbers that are less or equal than the integer

Parameters

n (int) – integer for reference

Returns

factors – List of all power-of-two numbers

Return type

list

class tvm.autotvm.task.space.SplitSpace(axes, policy, **kwargs)

Split an axis for several times

static get_num_output(axes, policy, **kwargs)

get number of output axes after this transform

Returns

n – number of output axes

Return type

int

class tvm.autotvm.task.space.SplitEntity(size)

A split operation with detailed parameters that can apply to an axis

Parameters

size (Array of int) – the size of every axis after split. e.g. an axis of extent 128, we split it into 3 axes, a possible size is [4, 4, 8] (4x4x8 = 128).

apply(sch, op, axis)

Apply split to an axis

Parameters
  • sch (tvm.te.schedule.Schedule) – The tvm schedule

  • op (tvm.te.Operation) – The stage to be applied

  • axis (tvm.te.schedule.IterVar) – axis to split

Returns

axes – The transformed axes.

Return type

list of Axis

class tvm.autotvm.task.space.ReorderSpace(axes, policy, **kwargs)

The parameter space for ordering an array of axes

static get_num_output(axes, policy, **kwargs)

get number of output axes after this transform

Returns

n – number of output axes

Return type

int

class tvm.autotvm.task.space.ReorderEntity(perm)

A reorder operation with detailed parameters that can apply to axes

Parameters

perm (Array of int) – define the permutation

apply(sch, op, axes)

Apply reorder to an array of axes

Parameters
  • sch (tvm.te.schedule.Schedule) – The tvm schedule

  • op (tvm.te.Operation) – The stage to be applied

  • axis (tvm.te.schedule.IterVar) – axis to split

Returns

axes – The transformed axes.

Return type

list of Axis

class tvm.autotvm.task.space.AnnotateSpace(axes, policy, **kwargs)

The parameter space for annotating an array of axes

static get_num_output(axes, policy, **kwargs)

get number of output axes after this transform

Returns

n – number of output axes

Return type

int

class tvm.autotvm.task.space.AnnotateEntity(anns)

An annotation operation with detailed parameters that can apply to axes

Parameters

anns (Array of string) – The annotations of axes

apply(sch, op, axes, axis_lens=None, max_unroll=None, vec_size=None, cfg=None, source=None)

Apply annotation to an array of axes

Parameters
  • sch (tvm.te.schedule.Schedule) – The tvm schedule

  • op (tvm.te.Operation) – The stage to be applied

  • axes (Array of tvm.te.schedule.IterVar) – axis to split

  • axis_lens (Array of int, optional) – the length of axes

  • max_unroll (int, optional) – maximum unroll step

  • vec_size (Array of int, optional) – valid vector lanes for vectorization

  • cfg (ConfigEntity, optional) – cfg for recording error

  • source (Array of Array tensor, optional) – source tensor for attaching cache

Returns

axes – The transformed axes

Return type

list of tvm.te.schedule.IterVar

class tvm.autotvm.task.space.OtherOptionSpace(axes, policy, **kwargs)

The parameter space for general option

static get_num_output(axes, policy, **kwargs)

get number of output axes after this transform

Returns

n – number of output axes

Return type

int

class tvm.autotvm.task.space.OtherOptionEntity(val)

The parameter entity for general option, with a detailed value

class tvm.autotvm.task.space.ConfigSpace

The configuration space of a schedule. Pass it as config in template to collect transformation space and build transform graph of axes

static axis(var)

get a virtual axis (axis placeholder)

Parameters

var (int or tvm.te.schedule.IterVar) – If is int, return an axis whose length is the provided argument. If is IterVar, return an axis whose length is extracted from the IterVar’s extent domain.

static reduce_axis(var)

get a virtual axis (axis placeholder)

Parameters

var (int or tvm.te.schedule.IterVar) – If is int, return an axis whose length is the provided argument. If is IterVar, return an axis whose length is extracted from the IterVar’s extent domain.

define_split(name, axis, policy='factors', **kwargs)

Define a new tunable knob which splits an axis into a list of axes

Parameters
  • name (str) – name to index the entity of this space

  • axis (tvm.te.schedule.IterVar) – axis to split

  • policy (str) – name of policy. If is ‘factors’, the tuner will try all divisible factors. If is ‘power2’, the tuner will try power-of-two factors less or equal to the length. If is ‘verbose’, the tuner will try all candidates in above two policies. If is ‘candidate’, try given candidates.

  • **kwargs

    extra arguments for policy

    max_factor:

    the maximum split factor (int).

    filter:

    see examples below for how to use filter (Callable[[int], bool]).

    num_outputs:

    the total number of axis after split (int).

    no_tail:

    should we only include divisible numbers as split factors (bool).

    candidate`:

    (policy=candidate) manual candidate list (List).

Examples

>>> # use custom candidates
>>> cfg.define_split('tile_x', x, policy='candidate', candidate=[[1, 4, 4], [4, 1, 4]])
>>> # use a filter that only accepts the split scheme whose inner most tile is less then 4
>>> cfg.define_split('tile_y', y, policy='factors', filter=lambda x: x.size[-1] <= 4)
define_reorder(name, axes, policy, **kwargs)

Define a new tunable knob which reorders a list of axes

Parameters
  • name (str) – name to index the entity of this space

  • axes (Array of tvm.te.schedule.IterVar) – axes to reorder

  • policy (str) – name of policy If is ‘identity’, do an identity permutation. If is ‘all’, try all permutations. If is ‘interval_all’, try all permutations of an interval of axes. If is ‘candidate’, try listed candidate. If is ‘interleave’, interleave chains of spatial axes and chains of reduction axes.

  • kwargs (dict) – extra arguments for policy

define_annotate(name, axes, policy, **kwargs)

Define a new tunable knob which annotates a list of axes

Parameters
  • name (str) – name to index the entity of this space

  • axes (Array of tvm.te.schedule.IterVar) – axes to annotate

  • policy (str) – name of policy If is ‘unroll’, unroll the axes. If is ‘try_unroll’, try to unroll the axes. If is ‘try_unroll_vec’, try to unroll or vectorize the axes. If is ‘bind_gpu’, bind the first few axes to gpu threads. If is ‘locate_cache’, choose n axes to attach shared/local cache.

  • kwargs (dict) – extra arguments for policy

define_knob(name, candidate)

Define a tunable knob with a list of candidates

Parameters
  • name (str) – name key of that option

  • candidate (list) – list of candidates

add_flop(flop)

Add float operation statistics for this tuning task

Parameters

flop (int or float or IntImm or FloatImm) – number of float operations

raise_error(msg)

register error in config Using this to actively detect error when scheduling. Otherwise these error will occur during runtime, which will cost more time.

Parameters

msg (str) –

valid()

Check whether the config meets all the constraints

Note

This check should be called after instantiation of task, because the ConfigEntity/ConfigSpace collects errors during instantiation

Returns

valid – whether the config meets all the constraints

Return type

bool

get(index)

Get a config entity with detailed parameters from this space

Parameters

index (int) – index in the space

class tvm.autotvm.task.space.ConfigEntity(index, code_hash, entity_map, constraints)

A configuration with detailed parameters

Parameters
  • index (int) – index of this config in space

  • code_hash (str) – hash of schedule code

  • entity_map (dict) – map name to transform entity

  • constraints (list) – List of constraints

get_flatten_feature()

flatten entities to a numerical one-dimensional feature vector

Returns

fea – one dimensional float32 array

Return type

np.array

get_other_option()
Returns

other_option – other tunable parameters (tunable parameters defined by cfg.define_knob)

Return type

dict

to_json_dict()

convert to a json serializable dictionary

Returns

json_dict – a json serializable dictionary

Return type

dict

static from_json_dict(json_dict)

Build a ConfigEntity from json serializable dictionary

Parameters

json_dict (dict) – Json serializable dictionary. This should be the return value of to_json_dict.

Returns

config – The corresponding config object

Return type

ConfigEntity

class tvm.autotvm.task.space.FallbackConfigEntity

The config entity created to support fallback

fallback_split(name, constraints)

Fallback a split knob

Parameters
  • name (str) – name of the knob

  • constraints (List of int) – The maximum tile size for every dimension. Value -1 means no constraint.

Examples

If you use cfg.define_split(‘tile_0’, 128, num_outputs=3), Then cfg.fallback_split(‘tile_0’, [-1, 8, 4]) will give you cfg[‘tile_0’].size = [4, 8, 4]

If you use cfg.define_split(‘tile_0’, 49, num_outputs=3), Then cfg.fallback_split(‘tile_0’, [-1, 8, 4]) will give you cfg[‘tile_0’].size = [7, 7, 1]

fallback_with_reference_log(ref_log)

A data driven fallback mechanism. We use tuned parameters from TopHub as reference data. For an unseen shape, we find the most similar tuned one from TopHub and mimic its parameters. Note that we are not matching by workload (e.g., input size, kernel size), but instead matching by configuration space. The idea is that if two workloads have similar configuration space, their optimal configurations are also likely to be similar.

Parameters

ref_log (List of (autotvm.measure.MeasureInput, autotvm.measure.MeasureResult)) – The reference log

Template dispatcher module.

A dispatcher is a function that can contains multiple behaviors. Its specific behavior is can be controlled by DispatchContext.

DispatchContext is used in two ways, usually via different implementation of the DispatchContext base class.

  • During search, we can use it to pass the current proposal from tuner.

  • During evaluation, we can use it to set pick the best policy.

class tvm.autotvm.task.dispatcher.DispatchContext

Base class of dispatch context.

DispatchContext enables the target and workload specific dispatch mechanism for templates.

query(target, workload)

Query the context to get the specific config for a template. If cannot find the result inside this context, this function will query it from the upper contexts.

Parameters
  • target (Target) – The current target

  • workload (Workload) – The current workload.

Returns

cfg – The specific configuration.

Return type

ConfigSpace

update(target, workload, cfg)

Update context with a specific config.

Parameters
  • target (Target) – The current target

  • workload (Workload) – The current workload.

  • cfg (ConfigSpace) – The specific configuration.

Note

This interface is for cases when TVM decides to replace an operator in the graph. For example, AlterOpLayout pass (enables when opt_level = 3) replaces NCHW convolution with NCHW[x]c implementation on x86 CPUs. Thus in TOPI, we first query schedule using original NCHW workload, then update the dispatcher with the new NCHW[x]c workload. So that later on, NCHW[x]c convolution can get schedule from the dispatcher using its own workload directly.

@conv2d_alter_layout.register("cpu")
def _alter_conv2d_layout(attrs, inputs, tinfo):
    workload = get_conv2d_workload(...)
    dispatch_ctx = autotvm.task.DispatchContext.current
    target = tvm.target.Target.current()
    config = dispatch_ctx.query(target, workload)

    # Get conv2d_NCHWc workload from config
    # new_workload = ...
    # new_inputs = ...
    # new_attrs = ...

    # Store altered operator's config
    dispatch_ctx.update(target, new_workload, config)
    return sym.contrib.conv2d_NCHWc(*new_inputs, **new_attrs)

We directly store config back because conv2d_NCHW and conv2d_NCHWc share the same schedule parameters. One can construct a new ConfigEntity if this is not the case.

class tvm.autotvm.task.dispatcher.ApplyConfig(config)

Apply a deterministic config entity for all queries.

Parameters

config (ConfigSpace or ConfigEntity) – The specific configuration we care about.

update(target, workload, cfg)

Override update

class tvm.autotvm.task.dispatcher.ApplyHistoryBest(records)

Apply the history best config

Parameters

records (str or iterator of (autotvm.measure.MeasureInput, autotvm.measure.MeasureResult)) – Collection of tuning records. If is str, then it should be the filename of a records log file. Each row of this file is an encoded record pair. Otherwise, it is an iterator.

load(records)

Load records to this dispatch context

Parameters

records (str or iterator of (autotvm.measure.MeasureInput, autotvm.measure.MeasureResult)) – Collection of tuning records. If is str, then it should be the filename of a records log file. Each row of this file is an encoded record pair. Otherwise, it is an iterator.

update(target, workload, cfg)

Update context with a specific config.

Parameters
  • target (Target) – The current target

  • workload (Workload) – The current workload.

  • cfg (ConfigSpace) – The specific configuration.

Note

This interface is for cases when TVM decides to replace an operator in the graph. For example, AlterOpLayout pass (enables when opt_level = 3) replaces NCHW convolution with NCHW[x]c implementation on x86 CPUs. Thus in TOPI, we first query schedule using original NCHW workload, then update the dispatcher with the new NCHW[x]c workload. So that later on, NCHW[x]c convolution can get schedule from the dispatcher using its own workload directly.

@conv2d_alter_layout.register("cpu")
def _alter_conv2d_layout(attrs, inputs, tinfo):
    workload = get_conv2d_workload(...)
    dispatch_ctx = autotvm.task.DispatchContext.current
    target = tvm.target.Target.current()
    config = dispatch_ctx.query(target, workload)

    # Get conv2d_NCHWc workload from config
    # new_workload = ...
    # new_inputs = ...
    # new_attrs = ...

    # Store altered operator's config
    dispatch_ctx.update(target, new_workload, config)
    return sym.contrib.conv2d_NCHWc(*new_inputs, **new_attrs)

We directly store config back because conv2d_NCHW and conv2d_NCHWc share the same schedule parameters. One can construct a new ConfigEntity if this is not the case.

class tvm.autotvm.task.dispatcher.FallbackContext

A fallback dispatch context.

Any tunable template can be called under this context. This is the root context.

clear_cache(target, workload)

Clear fallback cache. Pass the same argument as _query_inside to this function to clean the cache.

Parameters
  • target (Target) – The current target

  • workload (Workload) – The current workload.

update(target, workload, cfg)

Update context with a specific config.

Parameters
  • target (Target) – The current target

  • workload (Workload) – The current workload.

  • cfg (ConfigSpace) – The specific configuration.

Note

This interface is for cases when TVM decides to replace an operator in the graph. For example, AlterOpLayout pass (enables when opt_level = 3) replaces NCHW convolution with NCHW[x]c implementation on x86 CPUs. Thus in TOPI, we first query schedule using original NCHW workload, then update the dispatcher with the new NCHW[x]c workload. So that later on, NCHW[x]c convolution can get schedule from the dispatcher using its own workload directly.

@conv2d_alter_layout.register("cpu")
def _alter_conv2d_layout(attrs, inputs, tinfo):
    workload = get_conv2d_workload(...)
    dispatch_ctx = autotvm.task.DispatchContext.current
    target = tvm.target.Target.current()
    config = dispatch_ctx.query(target, workload)

    # Get conv2d_NCHWc workload from config
    # new_workload = ...
    # new_inputs = ...
    # new_attrs = ...

    # Store altered operator's config
    dispatch_ctx.update(target, new_workload, config)
    return sym.contrib.conv2d_NCHWc(*new_inputs, **new_attrs)

We directly store config back because conv2d_NCHW and conv2d_NCHWc share the same schedule parameters. One can construct a new ConfigEntity if this is not the case.

tvm.autotvm.task.dispatcher.clear_fallback_cache(target, workload)

Clear fallback cache. Pass the same argument as _query_inside to this function to clean the cache.

Parameters
  • target (Target) – The current target

  • workload (Workload) – The current workload.

Note

This is used in alter_op_layout to clear the bad cache created before call topi compute function

class tvm.autotvm.task.dispatcher.ApplyGraphBest(records)

Load the graph level tuning optimal schedules.

The input records should be in the ascending order of node index for target operator. Usually this can be obtained with graph tuner.

This context maintains an internal counter to indicate the current node index.

update(target, workload, cfg)

Update context with a specific config.

Parameters
  • target (Target) – The current target

  • workload (Workload) – The current workload.

  • cfg (ConfigSpace) – The specific configuration.

Note

This interface is for cases when TVM decides to replace an operator in the graph. For example, AlterOpLayout pass (enables when opt_level = 3) replaces NCHW convolution with NCHW[x]c implementation on x86 CPUs. Thus in TOPI, we first query schedule using original NCHW workload, then update the dispatcher with the new NCHW[x]c workload. So that later on, NCHW[x]c convolution can get schedule from the dispatcher using its own workload directly.

@conv2d_alter_layout.register("cpu")
def _alter_conv2d_layout(attrs, inputs, tinfo):
    workload = get_conv2d_workload(...)
    dispatch_ctx = autotvm.task.DispatchContext.current
    target = tvm.target.Target.current()
    config = dispatch_ctx.query(target, workload)

    # Get conv2d_NCHWc workload from config
    # new_workload = ...
    # new_inputs = ...
    # new_attrs = ...

    # Store altered operator's config
    dispatch_ctx.update(target, new_workload, config)
    return sym.contrib.conv2d_NCHWc(*new_inputs, **new_attrs)

We directly store config back because conv2d_NCHW and conv2d_NCHWc share the same schedule parameters. One can construct a new ConfigEntity if this is not the case.

Decorators for registering tunable templates to TOPI.

These decorators can make your simple implementation be able to use different configurations for different workloads. Here we directly use all arguments to the TOPI call as “workload”, so make sure all the arguments (except tvm.te.Tensor) in you calls are hashable. For tvm.te.Tensor, we will serialize it to a hashable tuple.

See tvm/topi/python/topi/arm_cpu/depthwise_conv2d.py for example usage.

class tvm.autotvm.task.topi_integration.TaskExtractEnv(allow_duplicate=False)

Global environment for extracting tuning tasks from graph

reset(wanted_relay_ops=None)

Reset task collections

Parameters

wanted_relay_ops (List of tvm.ir.Op) – The relay ops to be extracted

add_task(task_name, args)

Add AutoTVM task

Parameters
  • task_name (str) – AutoTVM task name.

  • args (tuple) – Arguments to the TOPI function.

get_tasks()

Get collected tasks

Returns

tasks – A list of tasks extracted from the graph

Return type

List of tuple(name, args)

static get(allow_duplicate=False)

Get the single instance of TaskExtractEnv

Parameters

allow_duplicate (boolean) – Whether to fetch all workloads in the network, even though some of them are the same. This is useful for graph tuning.

Returns

env – The single instance of TaskExtractEnv

Return type

TaskExtractEnv

tvm.autotvm.task.topi_integration.register_topi_compute(task_name, func=None)

Register a tunable template for a topi compute function.

The registration will wrap this topi compute to take cfg as the first argument, followed by the original argument list. It uses all its argument as workload and stores this “workload” to its final ComputeOp, which can be used to reconstruct “workload” in the following topi_schedule call.

Parameters
  • task_name (str) – The AutoTVM task name

  • func (None or callable) – If it is None, return a decorator. If is callable, decorate this function.

Returns

decorator – A decorator

Return type

callable

Examples

See tvm/topi/python/topi/arm_cpu/depthwise_conv2d.py for example usage.

tvm.autotvm.task.topi_integration.register_topi_schedule(task_name, func=None)

Register a tunable template for a topi schedule function.

The registration will wrap this topi schedule to take cfg as the first argument, followed by the original argument list.

Note that this function will try to find “workload” from all the ComputeOp in the input. You can attach “workload” to your compute op by using register_topi_compute.

The task name has to be the same as that of the corresponding topi compute function.

Parameters
  • task_name (str) – The AutoTVM task name

  • func (None or callable) – If it is None, return a decorator. If is callable, decorate this function.

Returns

decorator – A decorator

Return type

callable

Examples

See tvm/topi/python/topi/arm_cpu/depthwise_conv2d.py for example usage.

tvm.autotvm.task.topi_integration.get_workload(outs, task_name=None)

Retrieve the workload from outputs

tvm.autotvm.record

Tuning record and serialization format

tvm.autotvm.record.measure_str_key(inp, include_config=True)

get unique str key for MeasureInput

Parameters
Returns

key – The str representation of key

Return type

str

tvm.autotvm.record.encode(inp, result, protocol='json')

encode (MeasureInput, MeasureResult) pair to a string

Parameters
Returns

row – a row in the logger file

Return type

str

tvm.autotvm.record.decode(row, protocol='json')

Decode encoded record string to python object

Parameters
  • row (str) – a row in the logger file

  • protocol (str) – log protocol, json or pickle

Returns

ret – The tuple of input and result, or None if input uses old version log format.

Return type

tuple(autotvm.measure.MeasureInput, autotvm.measure.MeasureResult), or None

tvm.autotvm.record.load_from_file(filename)

Generator: load records from file. This is a generator that yields the records.

Parameters

filename (str) –

Yields
  • input (autotvm.measure.MeasureInput)

  • result (autotvm.measure.MeasureResult)

tvm.autotvm.record.split_workload(in_file, clean=True)

Split a log file into separate files, each of which contains only a single workload This function can also delete duplicated records in log file

Parameters
  • in_file (str) – input filename

  • clean (bool) – whether delete duplicated items

tvm.autotvm.record.pick_best(in_file, out_file)

Pick the best entries from a file and store them to another file. This function distills the useful log entries from a large log file. If out_file already exists, the best entries from both in_file and out_file will be saved.

Parameters
  • in_file (str) – The filename of input

  • out_file (str or file) – The filename of output