Speculative Block Reuse in GPGPU

Abstract

Speculative Parallelization is a technique used in the superscalar processors to process and compute the data based on certain predictions like memory, control and data. To reduce memory stalls and to increase the instruction level parallelism, this paper provides a new methodology in using the blocks of GPGPU’s to solve many generalized problems as speculative block reuse to store the low or high frequency redundant computations to reuse without executing them repeatedly avoiding repeated execution of the instructions. If the prediction fails, the processor executes normally and it does not alter state of execution. The speculative block reuse method of this paper provides more accuracy in reusing the values from the shared memory.

Keywords— speculative parallelism, block reuse, computations,
GPGPU, finite automaton, transition states.