A direct method for optimal VLSI realization of deeply nested n-D loop problems

• Authors: B. Bala Tripura Sundari;T. R. Padmanabhan
• Venue: Microprocessors & Microsystems
• Year: 2013

Many computationally intensive algorithms are often represented as n-dimensional (n-D) nested loop algorithms. Systolic-array-based projections and their modifications involving multidimensional vector space representations have been used to realise the optimal VLSI design of deeply nested loop problems. The approaches employed so far involve an extensive search of the feasible solution space through heuristic methods and yield near optimal solutions. This paper presents a method of identifying the optimal solution directly and through a logical procedure. The new allocation method is shown to evolve around the computational expression and the sub-space in which it lies. The array of processing elements termed as the PE array is allocated to the indentified computational sub-space which is strictly of lower dimension than the n-D problem space. The proposed new optimal allocation procedure is first explained using the 3-D matrix/matrix multiplication (MMM) problem. The effectiveness of the method for higher dimensional problem is demonstrated through the illustrative example flow of 6-D full search block motion (FSBM) algorithm. The various design possibilities of the above mapping procedure are explored analytically and the cost constraints termed the figure of merit (FoM) of the design are evolved for the various design trade-offs for MMM and 6-D FSBM problem. An experimental methodology is developed using a hyper-graph model to represent the PE allocation to a particular sub-space of the n-D problem space. The advantage of our mapping procedure is illustrated by considering two cases namely, first an allocation represented by a vertex cover that covers the nodes of the identified computational (n-x)-D sub-space, where x