OpenACC Routine Directive Propagation Using Interprocedural Analysis

Abstract

Accelerator programming today requires the programmer to specify what data to place in device memory, and what code to run on the accelerator device. When programming with OpenACC, directives and clauses are used to tell the compiler what data to copy to and from the device, and what code to compile for and run on the device. In particular, the programmer inserts directives around code regions, typically loops, to identify compute constructs to be compiled for and run on the device. If the compute construct calls a procedure, that procedure also needs to be marked for device compilation, as does any routine called in that procedure, and so on transitively. In addition, the marking needs to include the kind of parallelism that is exploited within the procedure, or within routines called by the procedure. When using separate compilation, the marking where the procedure is defined must be replicated in any file where it is called. This causes much frustration when first porting existing programs to GPU programming using OpenACC.

This paper presents an approach to partially automate this process. The approach relies on interprocedural analysis (IPA) to analyze OpenACC regions and procedure definitions, and to propagate the necessary information forward and backward across procedure calls spanning all the linked files, generating the required accelerator code through recompilation at link time. This approach can also perform correctness checks to prevent compilation or runtime errors. This method is implemented in the PGI OpenACC compiler.

A. Shivam—Work done at NVIDIA/PGI.