Method and apparatus for generating architecture specific convolution gradient kernels

A method for accelerating a convolution operation, the method comprising:

• receiving, by a task scheduler from an input/output (I/O) interface, a first data set and a second data set;
• transforming, by the task scheduler, the first data set into a first converted data set, by applying a first tensor transformation, the first converted data set having a first format;
• transforming, by the scheduler, the second data set into a second converted data set, by applying a second tensor transformation, the first tensor transformation and the second tensor transformation being applied to reformulate a same backward pass gradient calculation, the second converted data set having a second format;
• loading, by the task scheduler into a convolution functional unit of a core, the first converted data set and the second converted data set, the convolution functional unit configured to receive the first converted data set as an input in the first format and the second converted data set as a further input in the second format;
• and
• receiving, by the task scheduler from the convolution functional unit, a result of the convolution operation in a third format.

A method for transforming tensors, the method comprising:

• receiving an input tensor having an input format, the input format having a set of input iterators, and storing the input tensor into an input memory;
• allocating an output memory to store an output tensor having an output format, the output format having a set of output iterators;
• and
• iterating over the output memory to copy an input component of the input tensor to a portion of the output memory corresponding to an output component of the output tensor, an index of the input component defined as a function of the output set of iterators.