Evolving GPU Machine Code

摘要

Parallel Graphics Processing Unit (GPU) implementations of GPhave appeared in the literature using three main methodologies:(i) compilation, which generates the individuals in GPUcode and requires compilation; (ii) pseudo-assembly,which generates the individuals in an intermediary assembly codeand also requires compilation; and (iii) interpretation,which interprets the codes. This paper proposes a newmethodology that uses the concepts of quantum computing anddirectly handles the GPU machine code instructions. Ourmethodology utilizes a probabilistic representation of anindividual to improve the global search capability. In addition,the evolution in machine code eliminates both the overhead ofcompiling the code and the cost of parsing the program duringevaluation. We obtained up to 2.74 trillion GP operations persecond for the 20-bit Boolean Multiplexer benchmark. We alsocompared our approach with the other three GPU-basedacceleration methodologies implemented for quantum-inspiredlinear GP. Significant gains in performance were obtained. color="gray">