| 摘要 |
A pipelined linear array of processor elements (PEs) for performing matrix computations in an efficient manner. The linear array generally includes a head PE and a set of regular PEs, the head PE being a functional superset of the regular PE, with interconnections between nearest neighbor PEs in the array and a feedback path from a non-neighbor regular PE back to the head PE. Each PE includes arithmetic circuitry for performing multiply, combine and accumulate operations, and a register file for storing inputs and outputs of the arithmetic circuitry. The head PE further includes a non-linear function generator. Each PE is pipelined such that the latency for an arithmetic operation to complete is a multiple of the period with which new operations can be initiated. A Very Large Instruction Word (VLIW) program or other type of program may be used to control the array. The array is particularly efficient at performing complex matrix operations, such as, e.g., the solution of a set of linear equations, matrix inversion, matrix-matrix multiplication, and computation of covariance and cross correlation. |
