Crossbar allocation for matrix-vector multiplications

A method comprising:

• identifying a repeating submatrix pattern of a plurality of submatrix patterns corresponding to a plurality of matrix partitions of a matrix, wherein the repeating submatrix pattern corresponds to a first matrix partition and a second matrix partition of the plurality of matrix partitions;
• and
• compiling instructions to be executed by a processing core to perform operations corresponding to a multiplication of the matrix with an input vector to provide an output vector, wherein the input vector comprises a first vector partition and a second vector partition, and compiling the instructions comprises providing instructions to cause the processing core to:
• in a first pass, cause matrix values corresponding to the repeating submatrix pattern to be loaded into an analog multiplication crossbar, and read first data, provided by the analog multiplication crossbar, representing a product of the first matrix partition and the first vector partition;
• in a second pass, read second data, provided by the analog multiplication crossbar, representing a product of the second matrix partition and the second vector partition;
• and
• combine the first data and the second data to generate at least part of the output vector.

An apparatus comprising:

• a processing core;
• a plurality of analog multiplication crossbars associated with the processing core;
• and
• a memory to store instructions that, when executed by the processing core, cause the processing core to:
• cause matrix values to be loaded into a given analog multiplication crossbar of a plurality of analog multiplication crossbars, and read first data, provided by the given analog multiplication crossbar, representing a product of a first matrix partition of a plurality of matrix partitions of an input matrix and a first vector partition of a plurality of partitions of an input vector;
• read second data, provided by the given analog multiplication crossbar, representing a product of a second matrix partition of the plurality of matrix partitions and a second vector partition of the plurality of vector partitions of the input vector;
• and
• combine the first data and the second data to generate at least part of an output vector, wherein the output vector corresponds to a product of the input vector and the input matrix.