Towards Automatic Restrictification of CUDA Kernel Arguments

Abstract

Many procedural languages, such as C and C++, have pointers. Pointers are powerful and convenient, but pointer aliasing still hinders compiler optimizations, despite several years of research on pointer aliasing analysis. Because alias analysis is a difficult task and results are not always accurate, the ISO C standard 99 has added a keyword, named restrict to allow the programmer to specify non-aliasing as an aid to the compiler's optimizer and to thereby possibly improve performance. The task of annotating pointers with the restrict keyword is still left to the programmer. This task is, in general, tedious and prone to errors especially since the C does not perform any verification to ensure that restrict keyword is not misplaced. In this paper we present a static analysis tool that (i) finds CUDA kernels call sites in which actual parameters do not alias; (ii) clones the kernels called at such sites; (iii) after performing an alias analysis in these kernels, adds the restrict keyword to their arguments; and (iv) replaces the original kernel call by a call to the optimized clone whenever possible.