aihwkit.simulator.tiles.analog_mvm module

Implementation of analog MVM for torch tiles.

class aihwkit.simulator.tiles.analog_mvm.AnalogMVM[source]

Bases: object

Torch implementation of (part of) the IO-managed forward / backward pass in RPUCuda.

classmethod check_support(io_pars)[source]

Check whether the IO settings are supported.

Throws an assertion error when there is an incompatibility

Parameters:

io_pars (IOParameters) – the IOParameters to be checked

Raises:

ConfigError – in case a feature is not supported

Return type:

None

classmethod matmul(weight, input_, io_pars, trans=False, is_test=False, **fwd_pars)[source]

Noisy, io-managed mat-mul.

Parameters:

  • weight (Tensor) – weight matrix (out_size, in_size)

  • input – activation (m, in_size / out_size)

  • io_pars (IOParameters) – Parameter defining the mat-mul nonlinearities

  • trans (bool) – transpose of the weight (so that in_size and out_size is transposed).

  • is_test (bool) – whether testing or training mode

  • fwd_pars (Any) – additional parameter dictionary

  • input_ (Tensor) –

Returns:

Result tensor

Return type:

Tensor