hidet.ir.task¶
Classes:
|
A task defines the operator computation. |
- class hidet.ir.task.Task(name, inputs, outputs, *, inverse_map=None, attributes=None, share_map=None)[source]¶
A task defines the operator computation.
- name¶
The name of the task.
- Type:
str
- inputs¶
The input tensors of this task.
- Type:
List[TensorInput]
- outputs¶
The output tensors of this task. They are derived by applying computation on the input tensors.
- Type:
List[TensorNode]
- inverse_map¶
The inverse map. It records how the input tensors are derived from the output tensor (when there is only one output tensor). This is used to support epilogue fusion.
- Type:
Dict[TensorInput, InverseMap]
- attrs¶
The attributes of this task.
- Type:
Dict[str, Union[str, float, int, bool]]
- assertions¶
The assertions of this task. Each assertion is a tuple of an expression and an optional message. The expression is evaluated at runtime. If the expression is not true, the program will abort and print the message.
- Type:
List[Tuple[Expr, Optional[str]]]
The share map. If one output tensor shares memory with one input tensor, it is specified in this map. For example, share_map = {0: 0, 1: 2} means that the output tensor 0 shares the memory with input tensor 0, and output tensor 1 shares the memory with input tensor 2.
- Type:
Dict[int, int]
- symbols¶
The list of symbols used in this task.
- Type:
List[SymbolVar]
Methods:
generate_arguments
(inputs, outputs)Generate arguments for the compiled function of this task given the tensor parameters.
dummy_arguments
(device)Generate dummy arguments for the compiled function of this task.
build
(target[, load])Build the task for the given target to a callable function.
- generate_arguments(inputs, outputs)[source]¶
Generate arguments for the compiled function of this task given the tensor parameters.
- dummy_arguments(device)[source]¶
Generate dummy arguments for the compiled function of this task.
- Parameters:
device (str) – The target device.
- Returns:
args – The arguments for the compiled function.
- Return type:
Sequence[Tensor or int]