VLSI array processors
Generalized Cannon's algorithm for parallel matrix multiplication
ICS '97 Proceedings of the 11th international conference on Supercomputing
A Processor-Time-Minimal Systolic Array for Cubical Mesh Algorithms
IEEE Transactions on Parallel and Distributed Systems
Constructive Methods for Scheduling Uniform Loop Nests
IEEE Transactions on Parallel and Distributed Systems
SUMMA: Scalable Universal Matrix Multiplication Algorithm
SUMMA: Scalable Universal Matrix Multiplication Algorithm
A cellular computer to implement the kalman filter algorithm
A cellular computer to implement the kalman filter algorithm
The general matrix multiply-add operation on 2D torus
IPDPS'06 Proceedings of the 20th international conference on Parallel and distributed processing
| Hi-index | 0.00 |
In this paper, we represent the computation space of the (n×n)-matrix multiplication problem C=C+AċB as a 3D torus. All possible time-minimal scheduling vectors needed to activate the computations inside the corresponding 3D index points at each step of computing are determined. Using the projection method to allocate the scheduled computations to the processing elements, the resulting array processor that minimizes the computing time is a 2D torus with n×n processing elements. For each optimal time scheduling function, three optimal array allocations are obtained from projection. All the resulting allocations of all the optimal scheduling vectors can be classified into three groups. In one group, matrix C remains and both matrices A and B are shifted between neighbor processors. The well-known Cannon's algorithm belongs to this group. In another group, matrix A remains and both matrices B and C are shifted. In the third group, matrix B remains while both matrices A and C are shifted. The obtained array processor allocations need n compute-shift steps to multiply n×n dense matrices.