US 12,277,494 B2
Multi-dimensional tensor support extension in neural network processor
Christopher L. Mills, Saratoga, CA (US)
Assigned to APPLE INC., Cupertino, CA (US)
Filed by Apple Inc., Cupertino, CA (US)
Filed on Nov. 19, 2020, as Appl. No. 16/953,033.
Prior Publication US 2022/0156575 A1, May 19, 2022
Int. Cl. G06N 3/08 (2023.01); G06N 3/045 (2023.01); G06N 3/063 (2023.01)
CPC G06N 3/08 (2013.01) [G06N 3/045 (2023.01); G06N 3/063 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A neural processor circuit, comprising:
a tensor access operation circuit coupled to a system memory external to the neural processor circuit, the tensor access operation circuit configured to:
select, based on a mode signal, an operation mode to access the system memory in one of two mutually exclusive modes;
indirectly access, in response to the selected operation mode including a first mode of the two mutually exclusive modes, at least a region of a source tensor in the system memory based on a first index for referencing a first index tensor in the system memory generated by a first circuit using a source index provided to the first circuit in the first mode;
indirectly access, in response to the selected operation mode including a second mode of the two mutually exclusive modes, at least the region of the source tensor in the system memory based on crop descriptors generated by a second circuit and used for computation of an indirect source index, wherein the crop descriptors are generated by the second circuit receiving as input a second index for referencing a second index tensor, the second index is generated by a third circuit using the source index provided to the third circuit in the second mode, the source tensor comprising one or more source components having a rank; and
map the one or more source components into an input tensor having another rank;
a data processor circuit coupled to the tensor access operation circuit, the data processor circuit configured to store an output version of the input tensor; and
at least one neural engine circuit coupled to the data processor circuit, the at least one neural engine circuit configured to:
receive, from the data processor circuit, the output version of the input tensor as a plurality of units of input data; and
perform at least one convolution operation on the plurality of units of the input data to generate output data.