Section: New Results

Tiling Bitwise Computations Using Look-up Instructions

Participants : Florent Bouchez - Tichadou, Cyril Six [Inria, Internship, from Feb 2015 until Jun 2015] .

The BWLU is an instruction of a Very Long Instruction Word processor (VLIW) that performs a series of bit-independent computations in only one step through the use of a “look-up table” (LUT). The Bit-Wise Look-Up table instruction (BWLU) takes as input four registers as well as a 32-bit integer (the “table”), and is able to output two bit-independent computations based on the input registers into two output registers.

The goal is to make the best use possible of this instruction by replacing during compilation as much as possible groups of bitwise computation using BWLUs so as to reduce the number of instructions required to perform a computation. The problem is represented by a data-flow graph representing a computation, and the goal is use BLWUs as tiles to “match” groups of bitwise instruction.

We proved the problem NP-complete for a general data-flow graph, so it is not practical to try to find the optimal solution.

It is easy to devise a greedy algorithm that will produce a solution, but we wanted a way to check whether the solutions found where far from the optimal. An optimal algorithm is of course exponential in the size of the input graph, however, we devised a complete space exploration algorithm based on dynamic programming that manages to find the optimal solution for data-flow graphs with small width or height.