Compiler Optimizations to Execute MATLAB Programs on Memory Constrained GPUs

MATLAB is an array language that is being increasingly used for prototyping and developing code for numerical and scientific applications. Many MATLAB programs have a lot of data parallelism. However, they also contain con-trol flow dominated scalar regions that impact the pro-gram’s execution time. As modern computer systems have throughput-oriented accelerators such as graphics process-ing units (GPUs) in addition to traditional CPU cores, our earlier work proposed an approach that maps the control flow dominated regions of a MATLAB program to the CPU and the data parallel regions to the GPU thus significantly improving program performance. In this paper, we address the problem of identifying what variables can coexist in GPU memory simultaneously given the GPU memory constraints. We formulate this problem as that of identifying maximal cliques in an interference graph. We approximate the interference graph using an in-terval graph and then leverage the linear scan algorithm to efficiently determine whether generated code meets the re-quired constraints. Furthermore, we present two program transformations that optimize memory accesses on the GPU using the software managed scratchpad memory available in GPUs. We have prototyped the proposed compiler transfor-mations in MEGHA. Our experiments show that code gener-ated using our scheme performs 1.7X better than CPU only code on average and outperforms a naı̈ve scheme by 1.5X on average. Also, the memory access optimizations yield aver-age speedups of 2.5X and 3.1X on the GeForce 8800 GTS