SparsePropagator

class SparsePropagator

RAII wrapper for cuSPARSE SpMV propagation with prebuilt CSR matrices.

Functions

Name	Description
init	Initializes the cuSPARSE handle if it has not been created.
destroy	Releases allocated matrices and destroys the cuSPARSE handle.
buildCSR	Build CSR from (Xn, X, W) triplets.
propagate	Propagate preLayer state -> postLayer input using a named CSR.

Function Details

buildCSR

void buildCSR(const thrust::device_vector<int>& X, const thrust::device_vector<int>& Xn, const thrust::device_vector<float>& W, const std::string& name, int num_pre, int num_post, cudaStream_t stream = 0)

Build CSR from (Xn, X, W) triplets. Build CSR from (Xn, X, W) triplets.

X

Presynaptic indices.

Xn

Postsynaptic indices.

W

Connection weights.

name

Key used to reference the built matrix.

num_pre

Number of presynaptic neurons.

num_post

Number of postsynaptic neurons.

stream

CUDA stream for preprocessing.

destroy

void destroy()

Releases allocated matrices and destroys the cuSPARSE handle.

init

void init()

Initializes the cuSPARSE handle if it has not been created.

propagate

template<size_t stateVar, typename PreStateTypes, typename PostStateTypes> void propagate(genericLayer<PreStateTypes>& preLayer, genericLayer<PostStateTypes>& postLayer, const std::string& name, InputBehavior behavior = InputBehavior::INPUT_OVERRIDE, cudaStream_t stream = 0)

Propagate preLayer state -> postLayer input using a named CSR. Propagate preLayer state -> postLayer input using a named CSR.

stateVar

Index of the presynaptic state variable to propagate.

PreStateTypes

Tuple type for the presynaptic layer.

PostStateTypes

Tuple type for the postsynaptic layer.

preLayer

Source layer.

postLayer

Destination layer.

name

Key of the CSR matrix to use.

behavior

Input accumulation behavior.

stream

CUDA stream for the SpMV call.