Methods and systems for performing a sparse submanifold convolution on a GPU

A method of implementing a sparse submanifold convolution on a graphics processing unit, the method comprising:

• receiving, at the graphics processing unit, an input tensor in a dense format;
• identifying, at the graphics processing unit, active positions of the input tensor;
• for each identified position of the input tensor:
• performing, by a separate thread at the graphics processing unit, an indexed unfold operation on the input tensor based on the identified active position of the input tensor to generate a column of an input matrix, thereby obtaining an input matrix comprising a column for each active window of the input tensor in which each column comprises elements of the input tensor in the corresponding active window of the input tensor;
• and
• performing, at the graphics processing unit, a matrix multiplication between a weight matrix and the input matrix to generate an output matrix wherein the matrix multiplication is performed by processing each column of the input matrix in a separate thread of the graphics processing unit, the output matrix comprising elements of an output tensor of the sparse submanifold convolution based on the active windows.