xlnstorch.benchmark.UnaryBench

class xlnstorch.benchmark.UnaryBench(func, shape, lns=True, f=None, b=None, backward=False, device='cpu', kwargs=None)

A benchmark for unary operations in xlnstorch or torch. Inputs are generated from the standard normal distribution.

Parameters:

• func (Callable) – The unary function to benchmark, e.g., torch.sign or torch.relu.

• shape (Tuple) – The shape of the input tensor.

• lns (bool, optional) – If True, uses xlnstorch.randn for generating inputs, otherwise uses torch.randn.

• f (int, optional) – The precision parameter for the input LNSTensor.

• b (float, optional) – The base parameter for the input LNSTensor.

• backward (bool, optional) – If True, the input tensor will require gradients for backward pass.

• device (torch.device or str, optional) – The device on which to create the input tensor (default is “cpu”).

• kwargs (Dict, optional) – Additional keyword arguments to pass to the function being benchmarked.

__init__(func, shape, lns=True, f=None, b=None, backward=False, device='cpu', kwargs=None)

Parameters:

• func (Callable)

• shape (Tuple)

• lns (bool)

• f (int | None)

• b (float | None)

• backward (bool)

• device (device | str)

• kwargs (Dict | None)

Methods

__init__(func, shape[, lns, f, b, backward, ...])
after_epoch(idx)	Symmetric counterpart of before_epoch().
before_epoch(idx)	Called before the idx-th epoch.
forward(x)	The workload under test (e.g. model(*args, **kwargs)).
make_inputs()	Produce input tensors for one iteration of the benchmark.
post_forward(output)	Optional light-weight processing of the forward result.

Attributes

device