SC Conference - Activity Details

Michael Wolfe  (Portland Group)

GPU Software Tools

Wednesday,  10:30AM - 11:00AM

Room 19A/19B

Current strategies for programming GPUs require a program rewrite using a new language or new datatypes and methods. PGI offers an alternative for programming x64+GPU platforms leveraging the same techniques used by vectorizing compilers and OpenMP. GPU-enabled PGI compilers offload loops and functions within an x64 executable using an integrated x64+GPU programming model with minimal impact on existing source code and makefiles. As with OpenMP, pragmas and directives communicate to the compiler which loops or regions of code should be moved onto the GPU. As with vectorization, compiler analysis maps parallelism in standard Fortran and C programs onto the GPU architecture. In a step beyond, compiler feedback helps the programmer drive the performance optimization process. This talk presents PGI GPU pragmas and directives, presents the compiler analysis and translation steps necessary to map x64 Fortran or C programs onto an Accelerator, and demonstrates how compiler feedback can drive performance optimization.