发明名称 MATRIX OPERANDS FOR LINEAR ALGEBRA OPERATIONS
摘要 Described herein are methods, systems, and apparatuses to utilize a matrix operation by accessing each of the operation's matrix operands via a respective single memory handle. This use of a single memory handle for each matrix operand eliminates significant overhead in memory allocation, data tracking, and subroutine complexity present in prior art solutions. The result of the matrix operation can also be accessible via a single memory handle identifying the matrix elements of the result.
申请公布号 US2017060811(A1) 申请公布日期 2017.03.02
申请号 US201514697728 申请日期 2015.04.28
申请人 Intel Corporation 发明人 Yang Andrew;Kloss Carey;Arora Prashant;Werner Tony;Rao Naveen Gandham;Khosrowshahi Amir
分类号 G06F17/16;G06N3/08;G06F12/02 主分类号 G06F17/16
代理机构 代理人
主权项 1. An apparatus comprising: a memory; and one or more integrated circuits (ICs) communicatively coupled to the memory and comprising: controller circuity to: receive a matrix operation, the matrix operation to identify a plurality of matrix operands, at least some of the matrix operands comprising at least two-dimensional (2D) matrix operands and including a set of matrix elements arranged in at least row and column directions; andload the matrix elements for the plurality of matrix operands onto the memory, wherein each of the 2D matrix operands are to be loaded into one or more blocks of the memory comprising at least 2D blocks of memory, and wherein each of the matrix operands are to be accessible via a single memory handle identifying dimensions of the matrix operands and the block(s) of the memory including each operand's set of matrix elements; andone or more processing units to: execute the matrix operation by accessing each of the matrix operands via the respective single memory handle associated with each operand; andoutput a result of the matrix operation as a matrix operand to be stored in the memory.
地址 Santa Clara CA US