calix.hagen.multiplication.Multiplication

class calix.hagen.multiplication.Multiplication(synram: pyhalco_hicann_dls_vx_v3.SynramOnDLS = SynramOnDLS(0), *, num_sends: int = 1, wait_period: int = 4, signed_mode: bool = True)

Bases: object

Handles multiplication of vector and matrix with integration on the synaptic input lines.

Requirements: * CADCs are calibrated and connected to the neuron readout. You can

use calix.common.cadc.calibrate() and calix.hagen.neuron_helpers.configure_chip() to achieve this.

  • Synapse DAC bias is calibrated.

  • Synapse drivers are calibrated.

Variables

  • synram_selection_bit – Position of bit in event label that selects whether an SPL1 event reaches the top or bottom synram. You can choose one of bits 11 to 13 (default).

  • _synram_coord – Coordinate of synapse array to use.

  • num_sends – Number of sends of the vector, values greater than one indicate repetitions of the input events.

  • wait_period – Number of clock cycles to wait between events.

  • signed_mode – Decide whether the multiplication is using signed weights. This affects the shape of the matrix: in unsigned mode, i.e. signed_mode = False, weights are shaped (256, 256). If using signed weights, the shape is reduced to (128, 256), i.e. 256 inputs are mapped to 128 results.

  • cached_reset_synin – Cached builder containing instructions to reset synaptic input lines. Contains a call to reset_synin().

__init__(synram: pyhalco_hicann_dls_vx_v3.SynramOnDLS = SynramOnDLS(0), *, num_sends: int = 1, wait_period: int = 4, signed_mode: bool = True)

Initialize self. See help(type(self)) for accurate signature.

Methods

__init__([synram, num_sends, wait_period, …])

Initialize self.

auto_multiply(connection, vectors, matrix, *)

Multiply given vectors with the given matrix.

configure_for_integration(config)

Set static configuration for integration on excitatory synaptic input lines.

get_synapse_matrix(matrix)

Return a suitable lola SynapseMatrix depending on the requested weights.

multiply(connection, vectors, matrix)

Multiply given vectors with the given matrix, return results.

preconfigure(connection)

Configure synapse drivers, neurons, crossbar and external DAC.

preconfigure_crossbar(builder)

Configure the crossbar such that the upper bits in an event label select target hemisphere and PADI bus.

preconfigure_dac(builder)

Connect the external DAC to the synapse debug lines, supplying a voltage of 1.2 V.

reset_synin(builder)

Connect synapse lines to the debug lines shortly.

Attributes

synram_coord

Get the synram to perform MAC operations on.

synram_selection_bit

auto_multiply(connection: pyhxcomm_vx.ConnectionHandle, vectors: numpy.ndarray, matrix: numpy.ndarray, *, n_row_repeats: Union[int, str] = 1) → numpy.ndarray

Multiply given vectors with the given matrix.

Handle input shape of the matrix automatically, i.e. split large matrices (containing many inputs) up into multiple runs or repeat small matrices multiple times in order to increase signal. The outer dimension (i.e. number of outputs) has to fit on the width of the synapse array, it is not handled here.

Parameters

  • connection – Connection to the chip to run on.

  • vectors – Array of vectors to be multiplied. We expect row vectors, with the inner dimension (columns) holding the individual entries of each vector.

  • matrix – Weight matrix to be multiplied.

  • n_row_repeats – Number of repetitions of the weight matrix. In contrast to vector resends, the matrix rows are repeated, which also compensates fixed-pattern noise. Defaults to 1. Select “auto” to fill up the synapse array with a small matrix.

Returns

Numpy array with the result of multiplication.

static configure_for_integration(config: pylola_vx_v3.AtomicNeuron) → None

Set static configuration for integration on excitatory synaptic input lines.

Parameters

config – Configuration container to be altered.

get_synapse_matrix(matrix: numpy.ndarray) → pylola_vx_v3.SynapseMatrix

Return a suitable lola SynapseMatrix depending on the requested weights.

Parameters

matrix – Weight matrix as a numpy array.

Returns

lola.SynapseMatrix with addresses and weights set.

multiply(connection: pyhxcomm_vx.ConnectionHandle, vectors: numpy.ndarray, matrix: numpy.ndarray) → numpy.ndarray

Multiply given vectors with the given matrix, return results.

The matrix shape has to match the synapse array exactly, there is no splitting or merging of matrix shapes.

Parameters

  • connection – Connection to the chip to run on.

  • vectors – Array of vectors to be multiplied. We expect row vectors, with the inner dimension (columns) holding the individual entries of each vector.

  • matrix – Array of weights.

Returns

Array of results of the MAC operation.

preconfigure(connection: pyhxcomm_vx.ConnectionHandle) → None

Configure synapse drivers, neurons, crossbar and external DAC.

The previously configured hagen dac offsets and STP ramp offsets are kept unchanged.

Parameters

connection – Connection to the chip to run on.

classmethod preconfigure_crossbar(builder: calix.common.base.WriteRecordingPlaybackProgramBuilder) → None

Configure the crossbar such that the upper bits in an event label select target hemisphere and PADI bus. Also enable SPL1 events in the PADI buses.

Parameters

builder – Builder to append instructions to.

static preconfigure_dac(builder: calix.common.base.WriteRecordingPlaybackProgramBuilder) → None

Connect the external DAC to the synapse debug lines, supplying a voltage of 1.2 V.

The potential will be used for resetting potentials on the synapse lines before integration.

Parameters

builder – Builder to append instructions to.

reset_synin(builder: calix.common.base.WriteRecordingPlaybackProgramBuilder)

Connect synapse lines to the debug lines shortly.

The switches for excitatory and inhibitory synaptic currents are enabled.

Parameters

builder – Builder to append instructions to.

property synram_coord

Get the synram to perform MAC operations on.

synram_selection_bit = 13