tvm.auto_scheduler¶
Namespace for TVM Auto-scheduler.
Classes:
|
Apply the history best config |
|
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 TODO(jcf94): This is considered to be merged with the new Target specification: https://discuss.tvm.apache.org/t/rfc-tvm-target-specification/6844 :param num_cores: The number of device cores. |
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 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 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 |
|
Append measure records to file. |
-
class
tvm.auto_scheduler.ComputeDAG(compute_or_sche)¶ 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
compute (Union[List[Tensor], str, 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.
hash_key()Return the hash key of this compute DAG.
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.
-
get_init_state()¶ Get the init state of this ComputeDAG.
- Returns
state – The initial State without any transform steps.
- Return type
State
-
apply_steps_from_state(state, layout_rewrite=0)¶ Apply the history transform steps from a State to get a TVM schedule.
- Parameters
state (Union[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
- Return type
A te.schedule and the a list of te.Tensor to be used in tvm.lower or tvm.build.
-
print_python_code_from_state(state)¶ 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
state (Union[State, StateObject]) – The state from which we get transform steps.
- Returns
str – The Python schedule code.
- Return type
Str
-
infer_bound_from_state(state)¶ 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
state (Union[State, StateObject]) – The state from which we get transform steps.
- Returns
updated_state – The State with complete bound information.
- Return type
State
-
rewrite_layout_from_state(state)¶ Rewrite the layout of the DAG according to the history transform steps of a state.
- Parameters
state (Union[State, StateObject]) – The state from which we get transform steps.
- Returns
updated_dag – The compute dag with rewritten layout.
- Return type
-
class
tvm.auto_scheduler.LayoutRewriteOption¶ 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)¶ 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 – The default layout rewrite option for the specified target.
- Return type
-
static
-
tvm.auto_scheduler.get_shape_from_rewritten_layout(rewritten_layout, axis_names)¶ Get the orginal shape from a rewritten layout string.
-
class
tvm.auto_scheduler.RandomModel¶ 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).
-
update(inputs, results)¶ Update the cost model according to new measurement results (training data).
-
predict(search_task, states)¶ Predict the scores of states
- Parameters
search_task (SearchTask) – The search task of states
states (List[State]) – The input states
- Returns
scores – The predicted scores for all states
- Return type
-
-
class
tvm.auto_scheduler.XGBModel(verbose_eval=25, num_warmup_sample=100, seed=None, model_file=None, adapative_training=False)¶ 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.
adapative_training (bool = False) – Whether to use adapatie training, which reduces the training frequency when there are too many logs.
Methods:
load(file_name)Load the model from a file :param file_name: The filename :type file_name: str
predict(task, states)Predict the scores of states :param search_task: The search task of states :type search_task: SearchTask :param statse: The input states :type statse: List[State]
predict_stages(task, states)Predict the scores of all stages in states.
save(file_name)Save the model to a file :param file_name: The filename :type file_name: str
update(inputs, results)Update the cost model according to new measurement results (training data).
update_from_file(file_name[, n_lines])Load measure records from a log file to update the cost model.
-
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. :param inputs: The measurement inputs :type inputs: List[MeasureInput] :param results: The measurement results :type results: List[MeasureResult]
-
predict(task, states)¶ Predict the scores of states :param search_task: The search task of states :type search_task: SearchTask :param statse: The input states :type statse: List[State]
-
predict_stages(task, states)¶ Predict the scores of all stages in states. This is the breakdown version of predict.
- Parameters
search_task (SearchTask) – The search task of states
statse (List[State]) – The input states
- Returns
scores – The predicted scores for all stages in all states in the packed format
- Return type
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.
-
update_from_file(file_name, n_lines=None)¶ 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. :param file_name: The filename :type file_name: str :param n_lines: Only load first n lines of the log file :type n_lines: Optional[int]
-
class
tvm.auto_scheduler.DispatchContext¶ Base class of dispatch context.
Methods:
query(target, workload_key, has_complex_op, dag)Query the context to get the specific config for a workload.
update(target, workload_key, state)Update the config for a workload
-
query(target, workload_key, has_complex_op, dag)¶ Query the context to get the specific config for a workload. If cannot find the result inside this context, this function will query it from the upper contexts.
- Parameters
target (Target) – The current target
workload_key (str) – The workload key
has_complex_op (bool) – Whether this workload has at least one complex op.
dag (ComputeDAG) – The ComputeDAG of the workload.
- Returns
state – The state that stores schedule configuration for the workload
- Return type
StateObject
-
-
class
tvm.auto_scheduler.ApplyHistoryBest(records, n_lines=None)¶ Apply the history best config
- Parameters
records (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. Otherwise, it is an iterator.
n_lines (Optional[int]) – if it is not None, only load the first n_lines lines of log
Methods:
load(records[, n_lines])Load records to this dispatch context
update(target, workload_key, state)Update the config for a workload
-
load(records, n_lines=None)¶ Load records to this dispatch context
- Parameters
records (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. 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.MeasureInput(task, state)¶ Store the input of a measurement.
- Parameters
task (SearchTask) – The SearchTask of this measurement.
state (Union[State, StateObject]) – The State to be measured.
Methods:
Custom serialization to workaround MeasureInput not exposing all its members to the TVM ffi interface.
-
serialize()¶ Custom serialization to workaround MeasureInput not exposing all its members to the TVM ffi interface.
Note that we do not implement __getstate__ as it does not seem to work with initialization of the workload registry (maybe because of initialization order?).
-
class
tvm.auto_scheduler.MeasureResult(costs, error_no, error_msg, all_cost, timestamp)¶ Store the results of a measurement.
-
class
tvm.auto_scheduler.LocalBuilder(timeout=15, n_parallel=8, build_func='default')¶ LocalBuilder use local CPU cores to build programs in parallel.
- Parameters
timeout (int = 15) – The timeout limit (in second) for each build thread. This is used in a wrapper of the multiprocessing.Process.join().
n_parallel (int = 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.LocalRunner(timeout=10, number=3, repeat=1, min_repeat_ms=100, cooldown_interval=0.0, enable_cpu_cache_flush=False)¶ LocalRunner that uses local CPU/GPU to measures the time cost of programs.
- Parameters
timeout (int = 10) – The timeout limit (in second) for each run. This is used in a wrapper of the multiprocessing.Process.join().
number (int = 3) – The number of times to run the generated code for taking average. We call these runs as one repeat of measurement.
repeat (int = 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_ms (int = 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_interval (float = 0.0) – The cool down interval between two measurements.
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.
-
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)¶ 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
key (str) – The key of the device registered in the RPC tracker.
host (str) – The host address of the RPC Tracker.
port (int) – The port of RPC Tracker.
priority (int = 1) – The priority of this run request, larger is more prior.
n_parallel (int = 1) – The number of tasks run in parallel.
timeout (int = 10) – The timeout limit (in second) for each run. This is used in a wrapper of the multiprocessing.Process.join().
number (int = 3) – The number of times to run the generated code for taking average. We call these runs as one repeat of measurement.
repeat (int = 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_ms (int = 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_interval (float = 0.0) – The cool down interval between two measurements.
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.
-
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)¶ A context wrapper for running RPCRunner locally. This will launch a local RPC Tracker and local RPC Server.
- Parameters
priority (int = 1) – The priority of this run request, larger is more prior.
n_parallel (int = 1) – The number of tasks run in parallel.
timeout (int = 10) – The timeout limit (in second) for each run. This is used in a wrapper of the multiprocessing.Process.join().
number (int = 3) – The number of times to run the generated code for taking average. We call these runs as one repeat of measurement.
repeat (int = 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_ms (int = 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_interval (float = 0.0) – The cool down interval between two measurements.
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.
-
class
tvm.auto_scheduler.RecordToFile(filename)¶ A measurement callback that writes measurement records into a file.
- Parameters
filename (str) – File name for this callback to write log to.
-
class
tvm.auto_scheduler.RecordReader(filename)¶ Reader of the json log file.
- Parameters
filename (str) – File name for this reader to load log from.
Methods:
read_lines([max_lines, skip_lines])Read multiple lines from the log file.
-
read_lines(max_lines=None, skip_lines=0)¶ Read multiple lines from the log file.
- Parameters
max_lines (Optional[int]) – The maximum number of lines. None to read all lines.
skip_lines (int = 0) – Skip the first n lines.
- Returns
inputs (List[auto_scheduler.measure.MeasureInput]) – The MeasureInputs loaded from the log file.
results (List[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_inputbelow to rebuild these fields.
-
tvm.auto_scheduler.load_best_record(filename, workload_key=None, target=None)¶ 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
filename (str) – File name to load log from.
workload_key (Optional[str]) – The workload key of the compute declaration. With None, this returns the best measure pair of all workloads.
target (Optional[tvm.target.Target]) – The target device. With None, this returns the best measure pair of all target devices.
- Returns
input (auto_scheduler.measure.MeasureInput) – The best State’s MeasureInput from this log fine.
result (auto_scheduler.measure.MeasureResult) – The best State’s MeasureResult from this log fine.
-
tvm.auto_scheduler.load_records(filename)¶ Load measurement records from a file.
- Parameters
filename (str) – File name to load log from.
- Returns
logs
- Return type
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_inputbelow to rebuild these fields.
-
tvm.auto_scheduler.save_records(filename, inputs, results)¶ Append measure records to file.
-
tvm.auto_scheduler.extract_tasks(mod, params, target, target_host=None, hardware_params=None, include_simple_tasks=False)¶ 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_params (Optional[HardwareParams]) – Hardware parameters used for the search tasks
include_simple_tasks (bool) – Whether to extract simple tasks that do not include complicated ops.
- 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.remove_index_check(tensor)¶ 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.
-
tvm.auto_scheduler.rewrite_compute_body(compute_tensor, new_layout)¶ Rewrite the body of a ComputeOp according to a new layout of a placeholder
-
tvm.auto_scheduler.is_auto_scheduler_enabled()¶ Return whether the auto-scheduler is enabled.
- Parameters
enabled (bool) – Whether the auto-scheduler is enabled
-
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)¶ The computation information and hardware parameters for a schedule search task.
- Parameters
func (Union[Function, str]) – The function that returns the compute declaration Tensors. Can be the a function or the function name.
args (Union[Tuple[Any, ..], List[Any]]) – The args of the function.
compute_dag (ComputeDAG) – The ComputeDAG for the corresponding compute declaration.
workload_key (str) – The workload key for the corresponding compute declaration.
target (tvm.target.Target) – The target device of this search task.
target_host (Optional[tvm.target.Target]) – The target host device of this search task.
hardware_params (Optional[HardwareParams]) – Hardware parameters used in this search task.
layout_rewrite_option (Optional[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.
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[, layout_rewrite_option])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
-
tune(tuning_options, search_policy=None)¶ Run auto scheduling search for a task
- Parameters
tuning_options (TuningOptions) – Tuning and measurement options.
search_policy (Optional[SearchPolicy]) – The search policy to be used for schedule search.
-
apply_best(log_file, layout_rewrite_option=None)¶ Apply the history best from a log file and return the schedule.
- Parameters
log_file (str) – The name of the log file.
layout_rewrite_option (Optional[LayoutRewriteOption]) – The layout rewrite option.
- Returns
- Return type
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')¶ Print the best schedule as python schedule API code or CUDA source code.
-
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)¶ 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.HardwareParams(num_cores, vector_unit_bytes, cache_line_bytes, max_shared_memory_per_block, max_local_memory_per_block, max_threads_per_block, max_vthread_extent, warp_size)¶ The parameters of target hardware used to guide the search policy TODO(jcf94): This is considered to be merged with the new Target specification: https://discuss.tvm.apache.org/t/rfc-tvm-target-specification/6844 :param num_cores: The number of device cores. :type num_cores: int :param vector_unit_bytes: The width of vector units in bytes. :type vector_unit_bytes: int :param cache_line_bytes: The size of cache line in bytes. :type cache_line_bytes: int :param max_shared_memory_per_block: The max shared memory per block in bytes. :type max_shared_memory_per_block: int :param max_local_memory_per_block: The max local memory per block in bytes. :type max_local_memory_per_block: int :param max_threads_per_block: The max number of threads per block. :type max_threads_per_block: int :param max_vthread_extent: The max vthread extent. :type max_vthread_extent: int :param warp_size: The thread numbers of a warp. :type warp_size: int
-
tvm.auto_scheduler.create_task(func, args, target, target_host=None, hardware_params=None)¶ THIS API IS DEPRECATED.
Create a search task.
- Parameters
func (Union[Function, str]) – The function that returns the compute declaration Tensors. Can be the a function or the function name.
args (Union[Tuple[Any, ..], List[Any]]) – The args of the function.
target (Union[tvm.target.Target, str]) – The target device of this search task.
target_host (Optional[Union[tvm.target.Target, str]]) – The target host device of this search task.
hardware_params (Optional[HardwareParams]) – Hardware parameters used in this search task.
- Returns
SearchTask
- Return type
the created task
-
tvm.auto_scheduler.auto_schedule(task, search_policy=None, tuning_options=auto_scheduler.TuningOptions(39301472))¶ THIS API IS DEPRECATED.
Run auto scheduling search for a task.
- Parameters
task (SearchTask) – The SearchTask for the computation declaration.
search_policy (Optional[SearchPolicy]) – The search policy to be used for schedule search.
tuning_options (Optional[TuningOptions]) – Tuning and measurement options.
- Returns
- Return type
A te.Schedule and the a list of te.Tensor to be used in tvm.lower or tvm.build.
-
class
tvm.auto_scheduler.EmptyPolicy(task, init_search_callbacks=None)¶ A simple example of the search policy which always returns the initial naive schedule (state).
- Parameters
task (SearchTask) – The SearchTask for the computation declaration.
init_search_callbacks (Optional[List[SearchCallback]]) – Callback functions called before the search process.
-
class
tvm.auto_scheduler.SketchPolicy(task, program_cost_model=auto_scheduler.RandomModel(39114648), params=None, seed=None, verbose=1, init_search_callbacks=None)¶ 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
task (SearchTask) – The SearchTask for the computation declaration.
program_cost_model (CostModel = RandomModel()) – The cost model to estimate the complete schedules.
params (Optional[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.
seed (Optional[int]) – Random seed.
verbose (int = 1) – Verbosity level. 0 for silent, 1 to output information during schedule search.
init_search_callbacks (Optional[List[SearchCallback]]) –
Callback functions called before the search process, usually used to do extra initializations. Possible callbacks:
auto_scheduler.PreloadMeasuredStates
auto_scheduler.PreloadCustomSketchRule
TODO(jcf94): Add these search callback implementations.
Methods:
evolutionary_search(init_populations, out_size)Perform evolutionary search.
generate_sketches([print_for_debug])Generate the sketches.
Sample initial population.
-
generate_sketches(print_for_debug=False)¶ Generate the sketches. This python interface is mainly used for debugging and testing. The actual search is all done in c++.
- Parameters
print_for_debug (bool = False) – Whether print out the sketches for debug.
- Returns
sketches – The generated sketches of this search task.
- Return type
List[State]
-
sample_initial_population()¶ Sample initial population. This python interface is mainly used for debugging and testing. The actual search is all done in c++.
- Returns
states – The sampled states
- Return type
List[State]
-
evolutionary_search(init_populations, out_size)¶ Perform evolutionary search. This python interface is mainly used for debugging and testing. The actual search is all done in c++.
-
class
tvm.auto_scheduler.PreloadMeasuredStates(filename)¶ 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
filename (str) – The name of the record file.
-
class
tvm.auto_scheduler.TaskScheduler(tasks, task_weights=None, objective_func=None, strategy='gradient', load_model_file: str = None, load_log_file: str = None, alpha: float = 0.2, beta: float = 2, gamma: float = 0.5, backward_window_size: int = 3, callbacks=None)¶ 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:
tune(tune_option[, search_policy, …])Tune a batch of tasks together.
-
tune(tune_option, search_policy='default', search_policy_params=None)¶ Tune a batch of tasks together.
- Parameters
tune_option (TuningOptions) – The options of tuning
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_params (Optional[Dict[str, Any]]) – The parameters of the search policy
-
tvm.auto_scheduler.register_workload(func_name, f=None, override=False)¶ 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
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: tvm.sum(A[i][k] * B[k][j], axis=[k]), name='C') return [A, B, C]
-
tvm.auto_scheduler.make_workload_key(func, args)¶ Make a workload key by function and arguments.