tvm.auto_scheduler#
Namespace for TVM Auto-scheduler.
Classes:
|
Apply the history best config |
|
Apply the history best config, or sample a valid schedule if no config is found. |
|
The auto-scheduler's computational graph and related program analyses. |
Base class of dispatch context. |
|
|
A simple example of the search policy which always returns the initial naive schedule (state). |
|
The parameters of target hardware used to guide the search policy. |
Options for applying layout rewrite. |
|
|
LocalBuilder use local CPU cores to build programs in parallel. |
|
A context wrapper for running RPCRunner locally. |
|
LocalRunner that uses local CPU/GPU to measures the time cost of programs. |
|
Store the input of a measurement. |
|
Store the results of a measurement. |
|
A SearchCallback for SketchSearchPolicy that allows users to add custom sketch rule. |
|
A SearchCallback to load measured states from the log file for a search policy. |
|
RPCRunner that uses RPC call to measures the time cost of programs on remote devices. |
A model that returns random estimation for all inputs |
|
|
Reader of the json log file. |
|
A measurement callback that writes measurement records into a file. |
|
The computation information and hardware parameters for a schedule search task. |
|
The search policy that searches in a hierarchical search space defined by sketches. |
|
Allocate the time resources when tuning multiple tasks together. |
|
This controls the options of performance tuning. |
|
Train a XGBoost model to predict the normalized throughputs of programs. |
Functions:
|
THIS API IS DEPRECATED. |
|
THIS API IS DEPRECATED. |
|
Extract tuning tasks from a relay program. |
Get the orginal shape from a rewritten layout string. |
|
Return whether the auto-scheduler is enabled. |
|
|
Return the best measurement pair form a log file. |
|
Load measurement records from a file. |
|
Make a workload key by function and arguments. |
|
Register a function that checks the input buffer map. |
|
Register a function that generates a certain workload. |
|
Remove the safety check in the indexing function for a tensor. |
|
Rewrite the body of a ComputeOp according to a new layout of a placeholder |
|
Rewrite the tensor shape |
|
Append measure records to file. |
- class tvm.auto_scheduler.ApplyHistoryBest(records, n_lines=None, include_compatible=False)[源代码]#
Apply the history best config
Parameters#
- recordsstr, list of str, or iterator of (auto_scheduler.measure.MeasureInput, auto_scheduler.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. If it is an iterator, it can either be a set of str filenames which will be applied jointly, or a set of (input, result) tuples.
- n_lines: Optional[int]
if it is not None, only load the first n_lines lines of log.
- include_compatible: bool
When set to True, compatible records will also be considered.
Methods:
get_workload_entry
(best_records, target_key, ...)Get the entry of the target key and workload key hash in the given best record map.
load
(records[, n_lines])Load records to this dispatch context
update
(target, workload_key, state)Update the config for a workload
- static get_workload_entry(best_records, target_key, workload_key)[源代码]#
Get the entry of the target key and workload key hash in the given best record map.
Parameters#
- best_records: Dict[str, Dict[str, Dict[str, Any]]]
The best record map.
- target_key: str
The first key to the best_records.
- workload_key: str
The workload key that can be decoded to workload hash and args.
Returns#
- entry: Dict[str, Any]
The entry in best_records with target key and workload hash.
- workload_hash: str
The workload hash decoded from workload_key.
- workload_args: Tuple[Any, …]
The hashable tuple of workload args decoded from workload_key.
- load(records, n_lines=None)[源代码]#
Load records to this dispatch context
Parameters#
- recordsstr or iterator of (auto_scheduler.measure.MeasureInput, auto_scheduler.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.
- n_lines: Optional[int]
if it is not None, only load the first n_lines lines of log
- class tvm.auto_scheduler.ApplyHistoryBestOrSample(records, sample_simple_workloads=False, cost_model_file=None, num_measure=-1)[源代码]#
Apply the history best config, or sample a valid schedule if no config is found.
Parameters#
- recordsstr or iterator of (auto_scheduler.measure.MeasureInput, auto_scheduler.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.
- sample_simple_workloads: bool
When False, sampling will not apply to simple workloads (w/o reduction).
- cost_model_file: str
The filename of the pre-trained XGBoost cost model. If not present, then random model will be used.
- num_measure: int
Meausre the top-N rank of sampled schedules on the device. The default -1 means no measurement and simply return the top-1 schedule ranked by the cost model.
Methods:
query
(target, workload_key, has_complex_op, ...)Query the context to get the specific config for a workload.
- query(target, workload_key, has_complex_op, dag, func_name)[源代码]#
Query the context to get the specific config for a workload. If this function cannot find the result inside this context, it will query the result from the upper contexts.
Parameters#
- target: Target
The current target
- workload_keystr
The workload key
- has_complex_op: bool
Whether this workload has at least one complex op.
- dag: ComputeDAG
The ComputeDAG of the workload.
- func_name: str
The function name of this workload.
Returns#
- stateStateObject
The state that stores schedule configuration for the workload
- class tvm.auto_scheduler.ComputeDAG(compute_or_sche)[源代码]#
The auto-scheduler’s computational graph and related program analyses.
We convert a compute declaration described by tvm.compute (could be a single operator or a subgraph) to a ComputeDAG. It keeps the input/output tensors, all operations in the DAG, and some static analysis results for the DAG (e.g. the total float operation count, consumer/producer relations of operations, whether an operation stage should be tiled/compute inlined). These analyses can help the search policy to make decisions during the search. ComputeDAG is also responsible for the interaction between auto-scheduler’s LoopState and TVM schedule (e.g. applying the LoopState transform steps to a TVM schedule, providing LoopState with extra information got from TVM schedule).
Parameters#
- computeUnion[List[Tensor], str, tvm.te.Schedule]
Input/output tensors or workload key for a compute declaration.
Methods:
apply_steps_from_state
(state[, layout_rewrite])Apply the history transform steps from a State to get a TVM schedule.
Get the init state of this ComputeDAG.
infer_bound_from_state
(state)Infer and fill the bound of all iterators of a state.
print_python_code_from_state
(state)Print transform steps in the history of a State as TVM's python schedule code.
rewrite_layout_from_state
(state)Rewrite the layout of the DAG according to the history transform steps of a state.
Return the workload key of this compute DAG.
- apply_steps_from_state(state, layout_rewrite=0)[源代码]#
Apply the history transform steps from a State to get a TVM schedule.
Parameters#
- stateUnion[State, StateObject]
The state from which we get transform steps.
- layout_rewrite: LayoutRewriteOption = NoRewrite
Rewrite the layout of placeholders specified by “layout_free_placeholders” attr to make it most friendly for the generated schedule to read from.
Returns#
A te.schedule and the a list of te.Tensor to be used in tvm.lower or tvm.build.
- get_init_state()[源代码]#
Get the init state of this ComputeDAG.
Returns#
- stateState
The initial State without any transform steps.
- infer_bound_from_state(state)[源代码]#
Infer and fill the bound of all iterators of a state.
The states may lose complete bound information after some transform steps (e.g., compute_at). We can call this function to infer and fill all the bound information. This function calls TVM InferBound pass internally to get the bound. The returned state of this function is guaranteed to have complete iterator extent information.
Parameters#
- stateUnion[State, StateObject]
The state from which we get transform steps.
Returns#
- updated_stateState
The State with complete bound information.
- print_python_code_from_state(state)[源代码]#
Print transform steps in the history of a State as TVM’s python schedule code.
This is used to print transformation steps for debugging. Use apply_steps_from_state if you want to get a schedule for code generation.
Parameters#
- stateUnion[State, StateObject]
The state from which we get transform steps.
Returns#
- strStr
The Python schedule code.
- class tvm.auto_scheduler.DispatchContext[源代码]#
Base class of dispatch context.
Methods:
_query_inside
(target, workload_key, func_name)Query the context to get the specific config for a workload.
query
(target, workload_key, has_complex_op, ...)Query the context to get the specific config for a workload.
update
(target, workload_key, state)Update the config for a workload
- _query_inside(target, workload_key, func_name)[源代码]#
Query the context to get the specific config for a workload. This function only query config inside this context.
Parameters#
- target: Target
The current target
- workload_keystr
The current workload_key.
- func_name: str
The function name of this workload.
Returns#
- stateStateObject
The schedule configuration for the workload
- query(target, workload_key, has_complex_op, dag, func_name)[源代码]#
Query the context to get the specific config for a workload. If this function cannot find the result inside this context, it will query the result from the upper contexts.
Parameters#
- target: Target
The current target
- workload_keystr
The workload key
- has_complex_op: bool
Whether this workload has at least one complex op.
- dag: ComputeDAG
The ComputeDAG of the workload.
- func_name: str
The function name of this workload.
Returns#
- stateStateObject
The state that stores schedule configuration for the workload
- class tvm.auto_scheduler.EmptyPolicy(task, init_search_callbacks=None)[源代码]#
A simple example of the search policy which always returns the initial naive schedule (state).
Parameters#
- taskSearchTask
The SearchTask for the computation declaration.
- init_search_callbacksOptional[List[SearchCallback]]
Callback functions called before the search process.
- class tvm.auto_scheduler.HardwareParams(num_cores=None, vector_unit_bytes=None, cache_line_bytes=None, max_shared_memory_per_block=None, max_local_memory_per_block=None, max_threads_per_block=None, max_vthread_extent=None, warp_size=None, target=None, target_host=None)[源代码]#
The parameters of target hardware used to guide the search policy.
When a parameter isn’t provided, it will instead use the current machine’s default value if target is specified. TODO(jcf94): This is considered to be merged with the new Target specification: https://discuss.tvm.apache.org/t/rfc-tvm-target-specification/6844 Parameters ———- num_cores : int, optional
The number of device cores.
- vector_unit_bytesint, optional
The width of vector units in bytes.
- cache_line_bytesint, optional
The size of cache line in bytes.
- max_shared_memory_per_blockint, optional
The max shared memory per block in bytes.
- max_local_memory_per_blockint, optional
The max local memory per block in bytes.
- max_threads_per_blockint, optional
The max number of threads per block.
- max_vthread_extentint, optional
The max vthread extent.
- warp_sizeint, optional
The thread numbers of a warp.
- targetstr or Target, optional
The compilation target. Used to determine default values if provided.
- target_hoststr or Target, optional
The compilation target host. Used to determine default values if provided.
Methods:
__str__
()Pretty printing for hardware parameter configuration.
- class tvm.auto_scheduler.LayoutRewriteOption[源代码]#
Options for applying layout rewrite.
The NO_REWRITE and INSERT_TRANSFORM_STAGE are expected to be used when tuning a standalone op, and the REWRITE_FOR_PRE_TRANSFORMED is expected to be used when tuning ops inside a network.
Methods:
get_target_default
(target[, ...])Get the default layout rewrite option for the specified target.
- static get_target_default(target, in_relay_integration=False)[源代码]#
Get the default layout rewrite option for the specified target. Currently we only enable layout rewrite for cpu / mali backend for now
Parameters#
- target: tvm.target.Target
The compilation target.
- in_relay_integration: bool
If this check is ask for relay integration.
Returns#
- layout_rewrite_option: LayoutRewriteOption
The default layout rewrite option for the specified target.
- class tvm.auto_scheduler.LocalBuilder(timeout=15, n_parallel=4, build_func='default')[源代码]#
LocalBuilder use local CPU cores to build programs in parallel.
Parameters#
- timeoutint = 15
The timeout limit (in second) for each build thread. This is used in a wrapper of the multiprocessing.Process.join().
- n_parallelint = multiprocessing.cpu_count()
Number of threads used to build in parallel.
- build_func: callable or str = “default”
If is ‘default’, use default build function If is ‘ndk’, use function for android ndk If is callable, use it as custom build function, expect lib_format field.
- class tvm.auto_scheduler.LocalRPCMeasureContext(priority=1, n_parallel=1, timeout=10, number=3, repeat=1, min_repeat_ms=0, cooldown_interval=0.0, enable_cpu_cache_flush=False, device=0)[源代码]#
A context wrapper for running RPCRunner locally. This will launch a local RPC Tracker and local RPC Server.
Parameters#
- priorityint = 1
The priority of this run request, larger is more prior.
- n_parallelint = 1
The number of tasks run in parallel.
- timeoutint = 10
The timeout limit (in second) for each run. This is used in a wrapper of the multiprocessing.Process.join().
- numberint = 3
The number of times to run the generated code for taking average. We call these runs as one repeat of measurement.
- repeatint = 1
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_msint = 0
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_intervalfloat = 0.0
The cool down interval between two measurements in seconds.
- enable_cpu_cache_flush: bool = False
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.
- device: int = 0
Which device to run on if multiple are available.
- class tvm.auto_scheduler.LocalRunner(timeout=10, number=3, repeat=1, min_repeat_ms=100, cooldown_interval=0.0, enable_cpu_cache_flush=False, device=0)[源代码]#
LocalRunner that uses local CPU/GPU to measures the time cost of programs.
Parameters#
- timeoutint = 10
The timeout limit (in second) for each run. This is used in a wrapper of the multiprocessing.Process.join().
- numberint = 3
The number of times to run the generated code for taking average. We call these runs as one repeat of measurement.
- repeatint = 1
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_msint = 100
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_intervalfloat = 0.0
The cool down interval between two measurements in seconds.
- enable_cpu_cache_flush: bool = False
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.
- device: int = 0
Which device to run on if multiple are available.
- class tvm.auto_scheduler.MeasureInput(task, state)[源代码]#
Store the input of a measurement.
Parameters#
- taskSearchTask
The SearchTask of this measurement.
- stateUnion[State, StateObject]
The State to be measured.
Methods:
Custom serialization to workaround MeasureInput not exposing all its members to the TVM ffi interface.
- class tvm.auto_scheduler.MeasureResult(costs, error_no, error_msg, all_cost, timestamp)[源代码]#
Store the results of a measurement.
Parameters#
- costsList[float]
The time costs of execution.
- error_noint
The error code.
- error_msgOptional[str]
The error message if there is any error.
- all_costfloat
The time cost of build and run.
- timestampfloat
The time stamps of this measurement.
- class tvm.auto_scheduler.PreloadCustomSketchRule(meet_condition_func, apply_func, rule_name='CustomSketchRule')[源代码]#
A SearchCallback for SketchSearchPolicy that allows users to add custom sketch rule.
Notes#
This is an advanced feature. Make sure you’re clear how it works and this should only be used in SketchSearchPolicy.
Parameters#
- meet_condition_func: Callable
A function with (policy, state, stage_id) -> int. Should return one of the result enumeration.
- apply_func: Callable
A function with (policy, state, stage_id) -> [[State, int], …].
- rule_name: str = “CustomSketchRule”
The name of this custom sketch rule.
- class tvm.auto_scheduler.PreloadMeasuredStates(filename)[源代码]#
A SearchCallback to load measured states from the log file for a search policy.
- This can resume the state of the search policy:
Making sure an already measured state in former searches will never be measured again.
The history states can be used to speed up the search process(e.g. SketchPolicy uses history states as starting point to perform Evolutionary Search).
Parameters#
- filenamestr
The name of the record file.
- class tvm.auto_scheduler.RPCRunner(key, host, port, priority=1, n_parallel=1, timeout=10, number=3, repeat=1, min_repeat_ms=100, cooldown_interval=0.0, enable_cpu_cache_flush=False, device=0)[源代码]#
RPCRunner that uses RPC call to measures the time cost of programs on remote devices. Or sometime we may need to use RPC even in local running to insulate the thread environment. (e.g. running CUDA programs)
Parameters#
- keystr
The key of the device registered in the RPC tracker.
- hoststr
The host address of the RPC Tracker.
- portint
The port of RPC Tracker.
- priorityint = 1
The priority of this run request, larger is more prior.
- n_parallelint = 1
The number of tasks run in parallel.
- timeoutint = 10
The timeout limit (in second) for each run. This is used in a wrapper of the multiprocessing.Process.join().
- numberint = 3
The number of times to run the generated code for taking average. We call these runs as one repeat of measurement.
- repeatint = 1
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_msint = 100
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_intervalfloat = 0.0
The cool down interval between two measurements in seconds.
- enable_cpu_cache_flush: bool = False
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.
- device: int = 0
Which device to run on if multiple are available.
- class tvm.auto_scheduler.RandomModel[源代码]#
A model that returns random estimation for all inputs
Methods:
predict
(search_task, states)Predict the scores of states
update
(inputs, results)Update the cost model according to new measurement results (training data).
- class tvm.auto_scheduler.RecordReader(filename)[源代码]#
Reader of the json log file.
Parameters#
- filenamestr
File name for this reader to load log from.
Methods:
check_workload_key
(inputs)Check and throw warnings for records with old format workload key.
read_lines
([max_lines, skip_lines])Read multiple lines from the log file.
- check_workload_key(inputs)[源代码]#
Check and throw warnings for records with old format workload key.
Parameters#
- inputs: List[MeasureInput]
The measure inputs to be checked.
Notes#
This checker could be deprecated in the future.
- read_lines(max_lines=None, skip_lines=0)[源代码]#
Read multiple lines from the log file.
Parameters#
- max_linesOptional[int]
The maximum number of lines. None to read all lines.
- skip_linesint = 0
Skip the first n lines.
Returns#
- inputsList[auto_scheduler.measure.MeasureInput]
The MeasureInputs loaded from the log file.
- resultsList[auto_scheduler.measure.MeasureResult]
The MeasureResults loaded from the log file.
Notes#
Some unimportant and expensive fields in the returned MeasureInput are not deserialized for faster read speed (e.g. input.task.compute_dag, input.state.stages). If you want to use them, you can call the
recover_measure_input
below to rebuild these fields.
- class tvm.auto_scheduler.RecordToFile(filename)[源代码]#
A measurement callback that writes measurement records into a file.
Parameters#
- filenamestr
File name for this callback to write log to.
- class tvm.auto_scheduler.SearchTask(func=None, args=None, compute_dag=None, workload_key=None, target=None, target_host=None, hardware_params=None, layout_rewrite_option=None, task_inputs=None, task_inputs_overwrite=False, task_inputs_save_to_file=False, desc='')[源代码]#
The computation information and hardware parameters for a schedule search task.
Parameters#
- funcUnion[Function, str]
The function that returns the compute declaration Tensors. Can be the a function or the function name.
- argsUnion[Tuple[Any, …], List[Any]]
The args of the function.
- compute_dagComputeDAG
The ComputeDAG for the corresponding compute declaration.
- workload_keystr
The workload key for the corresponding compute declaration.
- targetany target-like object, see Target.canon_target
The target device of this search task.
- target_hostNone or any target-like object, see Target.canon_target
The target host device of this search task.
- hardware_paramsOptional[HardwareParams]
Hardware parameters used in this search task.
- layout_rewrite_optionOptional[LayoutRewriteOption]
The layout rewrite option used for measuring programs. If None, the default value will be set depending on the specified target. Auto_scheduler will find a better schedule for the specified layout rewrite option. The NO_REWRITE and INSERT_TRANSFORM_STAGE are expected to be used when tuning a standalone op, and the REWRITE_FOR_PRE_TRANSFORMED is expected to be used when tuning ops inside a network.
- task_inputsUnion[Dict[str, tvm.nd.NDArray], List[str]]
A dict maps the input names to input tensors or a list of input names. Some special Tensor used as inputs in program measuring. Usually we do not need to care about it, but for special workloads like Sparse computation the Sparse Tensor input are meaningful that we cannot use random input directly.
- task_inputs_overwritebool = False
Whether to overwrite the data if a name has already in the global table.
- task_inputs_save_to_filebool = False
Whether to save the data to a local file as well. This can be reused to resume the last tuning process.
- desc: str = “”
The description string of this task.
Examples#
# We support two ways to create a search task # Way 1: create a task by a workload generation function. # The `workload_func` is a function decorated by @auto_scheduler.register_workload task = SearchTask(func=workload_func, args=args, target=target) # Way 2: create a task by a workload_key. # The `workload_key` is a string, which can be either a hash key or a json-serialized # tuple(func, args). task = SearchTask(workload_key=workload_key, target=target)
Methods:
apply_best
(log_file[, include_compatible, ...])Apply the history best from a log file and return the schedule.
print_best
(log_file[, print_mode])Print the best schedule as python schedule API code or CUDA source code.
tune
(tuning_options[, search_policy, ...])Run auto scheduling search for a task
- apply_best(log_file, include_compatible=False, layout_rewrite_option=None)[源代码]#
Apply the history best from a log file and return the schedule.
Parameters#
- log_filestr
The name of the log file.
- include_compatible: bool
When set to True, all compatible records in the log file will be considered.
- layout_rewrite_optionOptional[LayoutRewriteOption]
The layout rewrite option.
Returns#
A te.Schedule and the a list of te.Tensor to be used in tvm.lower or tvm.build.
- print_best(log_file, print_mode='schedule')[源代码]#
Print the best schedule as python schedule API code or CUDA source code.
Parameters#
- log_filestr
The name of the log file
- print_mode: str
if “schedule”, print the best schedule as python schedule API code. if “cuda”, print the best schedule as CUDA source code.
Returns#
- code: str
The best schedule code in python API or CUDA source code
- class tvm.auto_scheduler.SketchPolicy(task, program_cost_model=auto_scheduler.RandomModel(0x5637198f9808), params=None, seed=None, verbose=1, init_search_callbacks=None)[源代码]#
The search policy that searches in a hierarchical search space defined by sketches. The policy randomly samples programs from the space defined by sketches and use evolutionary search to fine-tune them.
Parameters#
- taskSearchTask
The SearchTask for the computation declaration.
- program_cost_modelCostModel = RandomModel()
The cost model to estimate the complete schedules.
- paramsOptional[Dict[str, Any]]
Parameters of the search policy. See src/auto_scheduler/search_policy/sketch_search_policy.h for the definitions. See DEFAULT_PARAMS below to find the default values.
- seedOptional[int]
Random seed.
- verboseint = 1
Verbosity level. 0 for silent, 1 to output information during schedule search.
- init_search_callbacksOptional[List[SearchCallback]]
Callback functions called before the search process, usually used to do extra initializations. Possible callbacks:
auto_scheduler.PreloadMeasuredStates
auto_scheduler.PreloadCustomSketchRule
Methods:
evolutionary_search
(init_populations, out_size)Perform evolutionary search.
generate_sketches
([print_for_debug])Generate the sketches.
Sample initial population.
- evolutionary_search(init_populations, out_size)[源代码]#
Perform evolutionary search. This python interface is mainly used for debugging and testing. The actual search is all done in c++.
Parameters#
- init_populations: List[State]
The initial population states
- out_sizeint
The size of generated states
Returns#
- states: List[State]
The generated states
- generate_sketches(print_for_debug=False)[源代码]#
Generate the sketches. This python interface is mainly used for debugging and testing. The actual search is all done in c++.
Parameters#
- print_for_debugbool = False
Whether print out the sketches for debug.
Returns#
- sketchesList[State]
The generated sketches of this search task.
- class tvm.auto_scheduler.TaskScheduler(tasks, task_weights=None, objective_func=None, strategy='gradient', load_model_file=None, load_log_file=None, alpha=0.2, beta=2, gamma=0.5, backward_window_size=3, callbacks=None)[源代码]#
Allocate the time resources when tuning multiple tasks together. This implements two strategies: “round-robin” and “gradient”.
Parameters#
- tasks: List[SearchTask]
All tasks to tune
- task_weights: Optional[List[float]]
The weights of tasks. If provided, the task scheduler will set the objective function to sum(weight[t] * latency[t]), where weight[t] is the weight of a task and the lantecy[t] is the lantecy of the task. If not provided, the task scheduer will assign equal weights to all tasks (i.e., the objective function is sum(latency[t])).
- objective_func: Optional[Callable[List[float] -> float]]
The objective function to be minimized. The objective function accepts the current latencies of all tasks and returns the objective. If not provided, the objective is the weighted sum of the latencies of all tasks.
- strategy: str = “gradient”
The scheduling strategy. “round-robin”: Tune tasks in round robin order. “gradient” : Tune tasks with gradient descent.
- load_model_file: Optional[str]
Load pre-trained model from this file. If this is None, the cost model will be trained from scratch.
- load_log_file: Optional[str]
Load measurement records from this file. If it is not None, the status of the task scheduler, search policies and cost models will be restored according to this file.
- verbose: int = 1
The level of verbosity. 0 means silent.
- alpha: float = 0.2
The parameter used for ‘gradient’ strategy
- beta: float = 2
The parameter used for ‘gradient’ strategy
- backward_window_size: int = 3
The parameter used for ‘gradient’ strategy
- callbacks: Optional[List[TaskSchedulerCallback]]
The task scheduler callbacks that will be called before and after tuning a task. If None, PrintTableInfo and LogEstimatedLatency callback will be used.
Methods:
_adjust_similarity_group
(task_idx)adjust the similarity group for the selected task
_compute_score
(costs)compute the objective function
_restore_status
(log_file, num_measures_per_round)restore task_cts and best_costs from a log file
_tune_task
(task_idx)Tune the select task for one round
tune
(tune_option[, search_policy, ...])Tune a batch of tasks together.
- _restore_status(log_file, num_measures_per_round)[源代码]#
restore task_cts and best_costs from a log file
- tune(tune_option, search_policy='default', search_policy_params=None, adaptive_training=False, per_task_early_stopping=None)[源代码]#
Tune a batch of tasks together.
Parameters#
- tune_option: TuningOptions
The tuning options applied to all tasks.
- search_policy:Union[str, List[SearchPolicy]] = “default”
The list of search policies. If it is str, “default” for the default policy (SketchPolicy + XGBModel), “sketch.xgb” for SketchPolicy + XGBModel, “sketch.random” for SketchPolicy + RandomModel.
- search_policy_paramsOptional[Dict[str, Any]]
The parameters of the search policy
- adaptive_trainingbool = False
Option used by XGBModel to reduce the model training frequency when there’re too many logs.
- per_task_early_stoppingOptional[int]
Stop tuning a task early if getting no improvement after n measurements.
- class tvm.auto_scheduler.TuningOptions(num_measure_trials=0, early_stopping=None, num_measures_per_round=64, verbose=1, builder='local', runner='local', measure_callbacks=None)[源代码]#
This controls the options of performance tuning.
Parameters#
- num_measure_trials: int = 0
The number of measurement trials. The search policy measures num_measure_trials schedules in total and returns the best one among them. With num_measure_trials == 0, the policy will do the schedule search but won’t involve measurement. This can be used to get a runnable schedule quickly without auto-tuning.
- early_stopping: Optional[int]
Stop the tuning early if getting no improvement after n measurements.
- num_measures_per_round: int = 64
The number of schedules to be measured at each search round. The whole schedule search process will try a total number of num_measure_trials in several rounds.
- verbose: int = 1
Verbosity level. 0 for silent, 1 to output information during schedule search.
- builder: Union[ProgramBuilder, str] = ‘local’
ProgramBuilder which builds the program.
- runner: Union[ProgramRunner, str] = ‘local’
ProgramRunner which runs the program and measures time costs.
- measure_callbacks: Optional[List[MeasureCallback]]
Callback functions called after each measurement. Candidates: - auto_scheduler.RecordToFile
- class tvm.auto_scheduler.XGBModel(verbose_eval=25, num_warmup_sample=100, seed=None, model_file=None, adaptive_training=False)[源代码]#
Train a XGBoost model to predict the normalized throughputs of programs. Let the normalized throughput be the score of a program (higher is better). We predict the (approximate) score of a program = the sum of the scores of all stages in this program. i.e. score(P) = score_s0 + score_s1 + … + score_sn, where score_si is the score of Stage i in Program P. We extract feature for each stage and let the xgboost predict the score for each stage. We then sum up the predictions as the score of the whole program. We use RMSE as the loss function. i.e. loss(P, y) = 1/2 * (score(P) - y)^2, where P is the program and y is the normalized throughput according to the ground truth (measurement). XGBoost does not support this loss function because score(P) is a sum of the prediction of several samples, so we implemented a custom loss function and call it pack-sum-rmse. It is called “pack-sum” because we combine several samples into a “pack” and sum up their predictions.
Parameters#
- verbose_eval: int = 25
Print training log every verbose_eval iterations.
- num_warmup_sample: int = 100
The minimum number of samples to start to use the trained model. If the number of samples is less than this number, the model outputs random predictions.
- seed: Optional[int]
The random seed
- model_file: Optional[str]
If is not None, save model to this file after every update.
- adaptive_training: bool = False
Whether to use adaptive training, which reduces the training frequency when there are too many logs.
Methods:
load
(file_name)Load the model from a file Parameters ---------- file_name: str The filename
predict
(task, states)Predict the scores of states Parameters ---------- search_task : SearchTask The search task of states statse : List[State] The input states Returns ------- scores: List[float] The predicted scores for all states
predict_stages
(task, states)Predict the scores of all stages in states.
save
(file_name)Save the model to a file Parameters ---------- file_name: str The filename
update
(inputs, results)Update the cost model according to new measurement results (training data). XGBoost does not support incremental training, so we re-train a new model every time. Parameters ---------- inputs : List[MeasureInput] The measurement inputs results : List[MeasureResult] The measurement results.
update_from_file
(file_name[, n_lines])Load measure records from a log file to update the cost model. This function can be used to pre-train the cost model with history log files. Parameters ---------- file_name: str The filename n_lines: Optional[int] Only load first n lines of the log file.
- load(file_name)[源代码]#
Load the model from a file Parameters ———- file_name: str
The filename
- 参数:
file_name (str)
- predict(task, states)[源代码]#
Predict the scores of states Parameters ———- search_task : SearchTask
The search task of states
- statseList[State]
The input states
Returns#
- scores: List[float]
The predicted scores for all states
- predict_stages(task, states)[源代码]#
Predict the scores of all stages in states. This is the breakdown version of predict.
Parameters#
- search_taskSearchTask
The search task of states
- statseList[State]
The input states
Returns#
- scores: List[float]
The predicted scores for all stages in all states in the packed format
Note#
For faster data copy between c++ and python, the python part returns scores in a single flatten array using a packed format. The c++ part then unpacks the flatten array. The packed format is: {
float scores[N]; // scores[i] is the score for states[i]. int n_stage_0; // the number of stages in states[0] float stage_scores_0[[n_stage_0] // the scores for all stages in states[0] int n_stage_1; // the number of stages in states[1] float stage_scores_1[n_stage_1]; // the scores for all stages in states[1] … int n_stage_i; // the number of stages in states[i] float stage_scores_1[n_stage_i]; // the scores for all stages in states[i] … // untill i == N - 1
} To implement this format, we also store int as float, so we can store all numbers into a single float array.
- save(file_name)[源代码]#
Save the model to a file Parameters ———- file_name: str
The filename
- 参数:
file_name (str)
- update(inputs, results)[源代码]#
Update the cost model according to new measurement results (training data). XGBoost does not support incremental training, so we re-train a new model every time. Parameters ———- inputs : List[MeasureInput]
The measurement inputs
- resultsList[MeasureResult]
The measurement results
- tvm.auto_scheduler.auto_schedule(task, search_policy=None, tuning_options=auto_scheduler.TuningOptions(0x5637198979b0))[源代码]#
THIS API IS DEPRECATED.
Run auto scheduling search for a task.
Parameters#
- taskSearchTask
The SearchTask for the computation declaration.
- search_policyOptional[SearchPolicy]
The search policy to be used for schedule search.
- tuning_optionsOptional[TuningOptions]
Tuning and measurement options.
Returns#
A te.Schedule and the a list of te.Tensor to be used in tvm.lower or tvm.build.
- tvm.auto_scheduler.create_task(func, args, target, target_host=None, hardware_params=None)[源代码]#
THIS API IS DEPRECATED.
Create a search task.
Parameters#
- funcUnion[Function, str]
The function that returns the compute declaration Tensors. Can be the a function or the function name.
- argsUnion[Tuple[Any, …], List[Any]]
The args of the function.
- targetUnion[tvm.target.Target, str]
The target device of this search task.
- target_hostOptional[Union[tvm.target.Target, str]]
The target host device of this search task.
- hardware_paramsOptional[HardwareParams]
Hardware parameters used in this search task.
Returns#
SearchTask: the created task
- tvm.auto_scheduler.extract_tasks(mod, params, target, target_host=None, hardware_params=None, include_simple_tasks=False, dump_workload_to_dag_log=None, opt_level=3, other_targets=None)[源代码]#
Extract tuning tasks from a relay program.
Parameters#
- mod: tvm.IRModule or relay.function.Function
The module or function to tune
- params: dict of str to numpy array
The associated parameters of the program
- target: Union[tvm.target.Target, str]
The compilation target
- target_host: Optional[Union[tvm.target.Target, str]]
The host compilation target
- hardware_paramsOptional[HardwareParams]
Hardware parameters used for the search tasks
- include_simple_tasks: bool
Whether to extract simple tasks that do not include complicated ops.
- dump_workload_to_dag_log: Optional[str]
A file to dump an association between the workload keys and the actual DAG
- opt_levelOptional[int]
The optimization level of the task extractions.
- other_targets: Optional[List[tvm.target.Target]]
Other targets for call_all_topi_funcs, e.g., cutlass target.
Returns#
- tasks: List[SearchTask]
The tasks in this network
- weights: List[int]
The weight (i.e. the number of appearance) of extracted tasks
- tvm.auto_scheduler.get_shape_from_rewritten_layout(rewritten_layout, axis_names)[源代码]#
Get the orginal shape from a rewritten layout string.
Parameters#
- rewritten_layout: str
The layout after rewrite
- axis_names: List[str]
Specify the order of axes by names
Returns#
- shape: List[PrimExpr]
The original shape
- tvm.auto_scheduler.is_auto_scheduler_enabled()[源代码]#
Return whether the auto-scheduler is enabled.
Parameters#
- enabled: bool
Whether the auto-scheduler is enabled
- tvm.auto_scheduler.load_best_record(filename, workload_key=None, target=None, include_compatible=False)[源代码]#
Return the best measurement pair form a log file. This may return none results if there is no legal measure pair with the specified workload_key/target found from the log file.
Parameters#
- filenamestr
File name to load log from.
- workload_keyOptional[str]
The workload key of the compute declaration. With None, this returns the best measure pair of all workloads.
- targetOptional[tvm.target.Target]
The target device. With None, this returns the best measure pair of all target devices.
- include_compatible: bool
When set to True, all compatible records in the log file will be considered.
Returns#
- inputauto_scheduler.measure.MeasureInput
The best State’s MeasureInput from this log fine.
- resultauto_scheduler.measure.MeasureResult
The best State’s MeasureResult from this log fine.
- tvm.auto_scheduler.load_records(filename)[源代码]#
Load measurement records from a file.
Parameters#
- filenamestr
File name to load log from.
Returns#
logs : List[auto_scheduler.measure.MeasureInput, auto_scheduler.measure.MeasureResult]
Notes#
Some unimportant and expensive fields in the returned MeasureInput are not deserialized for faster read speed (e.g., input.task.compute_dag, input.state.stages). If you want to use them, you can call the
recover_measure_input
below to rebuild these fields.
- tvm.auto_scheduler.make_workload_key(func, args)[源代码]#
Make a workload key by function and arguments.
Parameters#
- funcUnion[Function, str]
The function that returns the compute declaration Tensors. Can be the a function or the function name.
- argsArgs
The args of the function.
Returns#
- workload_keystr
The workload key of the function.
- tvm.auto_scheduler.register_task_input_check_func(func_name, f=None, override=False)[源代码]#
Register a function that checks the input buffer map.
The input function should take a list of Tensor wich indicate the Input/output Tensor of a TVM subgraph and return a Map from the input Tensor to its buffer name.
Parameters#
- func_nameUnion[Function, str]
The check function that returns the compute declaration Tensors or its function name.
- fOptional[Function]
The check function to be registered.
- overrideboolean = False
Whether to override existing entry.
Examples#
@auto_scheduler.register_task_input_check_func def check_task_input_by_placeholder_name(args : List[Tensor]): tensor_input_map = {} for arg in args: if isinstance(arg.op, tvm.te.PlaceholderOp): if arg.op.name != "placeholder": tensor_input_map[arg] = arg.op.name return tensor_input_map
- tvm.auto_scheduler.register_workload(func_name, f=None, override=False)[源代码]#
Register a function that generates a certain workload.
The input function should take hashable and jsonable arguments (int, float, tuple of int, tvm.tensor.Tensor, …) and return a list of tvm.tensor.Tensor.
Parameters#
- func_nameUnion[Function, str]
The generation function that returns the compute declaration Tensors or its function name.
- fOptional[Function]
The generation function to be registered.
- overrideboolean = False
Whether to override existing entry.
Examples#
@auto_scheduler.register_workload def matmul(N, M, K): A = te.placeholder((N, K), name='A') B = te.placeholder((K, M), name='B') k = te.reduce_axis((0, K), name='k') C = te.compute((N, M), lambda i, j: te.sum(A[i][k] * B[k][j], axis=[k]), name='C') return [A, B, C]
- tvm.auto_scheduler.remove_index_check(tensor)[源代码]#
Remove the safety check in the indexing function for a tensor. This is done by monkey patching its indexing function. After removing the check, we are allowed to create a temporary wrong IR and fix it later in other places.
Parameters#
- tensor: Tensor
The tensor to remove index check.