Automatic Conversion of Source Code for C to CUDA C

Abstract

As today in so many elds, computation is the main part of the algorithm and takes too much time in execution of the algorithm, it is necessary to parallelize the computation or reduce execution time. GPUs are widely used in (HPC) High Performance Computing. To achieve speedup, either we can increase clock frequency or multiple computation cores on the same chip. The clock speeds have reached the physical limit, so the use of many cores is the only way left to achieve speedup. As the GPU is growing demand of the Game Industry and large scienti c computations, e orts have been made to take advantages to gain maximum utilization of the GPUs in computation. Though GPUs are widely used in Supercomputers today, they are not code transparent because one has to sit and code the algorithms in CUDA C to run them on GPU. So if we can have some middleware that converts the C programs to CUDA, the end user gets transparency. I tried to develop a prototype compiler using ANTLR in visual studio that converts the C programs in CUDA C language. The thesis describes the literature survey in CUDA, dif- ferent performance optimization strategies to reduce execution time, the Pattern approach to develop a translator for source code to source code translation on the basis of selection of codes using patterns, platforms to code such translator and platform comparison and choice and algorithm of translation. The Compiler Architecture and its implementation details are widely described in thesis. The thesis describes implementation of the com- plete C2CUDATranslator, testing and analysis of the developed compiler. The compiler takes input of C program and generates CUDA program. The thesis demonstrates the pattern approach for language to language translation and the compiler ow architecture. C2CUDAranslator covers a new way or a framework to implement new analysis algorithms to detect dependencies in the code. The thesis also covers neural network design for com- piler learning and optimization of the translated code. The Neural Network helps compiler to take decision for selection of transformation and translation. Finally, the thesis covers outcome of the compiler, converted programs list, evaluation using parboil benchmark suite, performance graph of converted programs. It is concluded that the C2CUDATranslator saves 95% of the development time in selected cases.