Abstract: System, method, and accelerator to process a convolutional neural network. In accordance therewith, a tile structure having input data values is loaded for a convolution layer. Each tile of the tile structure corresponds to a respective feature map in a set of input feature maps. The tile structure of each iteration represents a different subset of data values in the input feature maps. Intermediate data values associated with a subset of the data values of the input feature maps in the current intermediate tile structure are reused, when the intermediate data values of a previous tile structure overlap values to be computed in the current tile structure. Intermediate non-overlapping data values that are associated with the subset of the data values in the current tile structure are computed using associated filters having weight data values. Available reused intermediate data values and computed intermediate data values are buffered as intermediate data.

Abstract: System, method, and accelerator to process a convolutional neural network. In accordance therewith, a tile structure having input data values is loaded for a convolution layer. Each tile of the tile structure corresponds to a respective feature map in a set of input feature maps. The tile structure of each iteration represents a different subset of data values in the input feature maps. Intermediate data values associated with a subset of the data values of the input feature maps in the current intermediate tile structure are reused, when the intermediate data values of a previous tile structure overlap values to be computed in the current tile structure. Intermediate non-overlapping data values that are associated with the subset of the data values in the current tile structure are computed using associated filters having weight data values. Available reused intermediate data values and computed intermediate data values are buffered as intermediate data.

Abstract: Computer-implemented systems and methods that provide an efficient technique for performing a large class of permutations on data vectors of length 2n, n>1, implemented with streaming width 2k (where 1?k?n?1). The technique applies to any permutation Q on 2n datawords that can be specified as a linear transform, i.e., as an n×n bit matrix (a matrix containing only 1s and 0s) P on the bit level. The relationship between Q and P is as follows: If Q maps (dataword) i to (dataword) j, then the bit representation of j is the bit-matrix-vector product of P with the bit representation of i. Given such a permutation specified by the matrix P and given the streaming width (k), an architectural framework (or datapath) is calculated to implement the permutation.

Abstract: Computer-implemented systems and methods that provide an efficient technique for performing a large class of permutations on data vectors of length 2n, n>1, implemented with streaming width 2k (where 1?k?n?1). The technique applies to any permutation Q on 2n datawords that can be specified as a linear transform, i.e., as an n×n bit matrix (a matrix containing only 1s and 0s) P on the bit level. The relationship between Q and P is as follows: If Q maps (dataword) i to (dataword) j, then the bit representation of j is the bit-matrix-vector product of P with the bit representation of i. Given such a permutation specified by the matrix P and given the streaming width (k), an architectural framework (or datapath) is calculated to implement the permutation.