Neural network accelerators
AssigneeHewlett Packard Enterprise Development LP
InventorsVaithyalingam NAGENDRAN, Jitendra Onkar KOLHE, Soumitra CHATTERJEE, Shounak BANDOPADHYAY
Life of this patent
Estimating the term…
Abstract
Examples of performing tensor operations by a neural network-based computing system, are described. In an example, a first output working set generated by a first operation, wherein the first output working set is a set of processed partitioned tensors, is obtained. The first output working set is then copied to the output working set, for retrieving by the second operation.
Claims20
A system comprising:
- a processor;
- a machine-readable storage medium comprising instructions executable by the processor to:
- generate a plurality of partitioned tensors based on a received tensor, wherein the received tensor is representative of digital data;
- generate a working set corresponding to a set of partitioned tensors from amongst the plurality of partitioned tensors;
- identify a first operation to be performed on the working set;
- define an input working space and an output working space corresponding to the first operation, the input working space and the output working space are to be defined in a portion of a memory of a neural network accelerator;
- copy the working set to the input working space corresponding to the first operation;
- execute the first operation on the working set in the input working space to generate a first output working set;
- and
- copy the first output working set to the output working space of the first operation.
A method for processing a tensor by a neural network accelerator, the method comprising:
- obtaining a first output working set generated by a first operation, wherein the first output working set is a set of processed partitioned tensors;
- copying the first output working set to a second input working space, wherein the second input working space is a memory portion of the neural network accelerator, corresponding to a second operation;
- executing the second operation on the first output working set stored in the second input working space to generate a second output working set;
- and
- copying the second output working set to a second output working space for retrieving by a subsequent operation.
A non-transitory computer-readable medium comprising instructions for performing a tensor operation using a neural network accelerator, the instructions being executable by a processing resource to:
- define a working set comprising a set of partitioned tensors, wherein the partitioned tensors are obtained from a tensor, the tensor being representative of digital data;
- identify a first operation and second operation, wherein the first operation and the second operations form part of the tensor operation;
- define an input working space and an output working space for the first operation and the second operation;
- wherein the input working space and the output working space are a portion of memory within the neural network accelerator, and is to accommodate an input working set and an output working set, respectively;
- execute the first operation on a core of the neural network accelerator to generate a first output working set based on processing the working set within the input working space of the first operation;
- and
- providing the first output working set from the output working space of the first operation to the input working space of the second operation.
Dependencies are inferred from the claim text.