US 12,327,326 B2
Method and apparatus for efficient non-integer scaling in neural network accelerators
Michael Scott Lee, North Vancouver (CA); Scott Chin, Vancouver (CA); Bradley Quinton, Vancouver (CA); and Trent McClements, Burnaby (CA)
Assigned to Singulos Research Inc., Burnaby (CA)
Filed by SINGULOS RESEARCH INC., Burnaby (CA)
Filed on Jun. 15, 2022, as Appl. No. 17/841,057.
Claims priority of provisional application 63/210,847, filed on Jun. 15, 2021.
Prior Publication US 2022/0405881 A1, Dec. 22, 2022
Int. Cl. G06T 3/10 (2024.01); G06T 3/4007 (2024.01); G06T 3/4046 (2024.01); G06T 7/11 (2017.01)
CPC G06T 3/4046 (2013.01) [G06T 3/4007 (2013.01); G06T 7/11 (2017.01); G06T 2207/20021 (2013.01); G06T 2207/20084 (2013.01)] 22 Claims
OG exemplary drawing
 
1. A method for performing data scaling on spatial data, the data scaling having a scaling factor given by N/D where N and D are positive integers, the method comprising:
receiving input spatial data;
dividing elements of the input spatial data into non-overlapping, contiguous input regions sized D×D elements;
obtaining interpolation weight sets for N*N kernels, each kernel having an associated interpolation weight set, each kernel associated with an element of an scaled output region sized N×N elements, where each scaled output region is associated with a respective one of the input regions; and
generating rescaled spatial data by, for each of the input regions, performing an interpolation operation that includes, using the interpolation weight set of each kernel, computing a weighted sum of the elements of input region utilizing the weight set to generate the element of the scaled output region associated with that kernel.