Automaticand efficient data host-device communication for many-core coprocessors

Abstract

Orchestrating data transfers between CPU and a coprocessor manually is cumbersome, particularly for multi-dimensional arrays and other data structures with multi-level pointers common in scientific computations. This paper describes a system that includes both compiletime and runtime solutions for this problem, with the overarching goal of improving programmer productivity while maintaining performance. We find that the standard linearization method performs poorly for non-uniform dimensions on the coprocessor due to redundant data transfers and suppression of important compiler optimizations such as vectorization. The key contribution of this paper is a novel approach for heap linearization that avoids modifying memory accesses to enable vectorization, referred to as partial linearization with pointer reset. We implement partial linearization with pointer reset as the compile time solution, whereas runtime solution is implemented as an enhancement to MYO library. We evaluate our approach with respect to multiple C benchmarks. Experimental results demonstrate that our best compiletime solution can perform 2.5x-5x faster than original runtime solution, and the CPU-MIC code with it can achieve 1.5x-2.5x speedup over the 16-thread CPU version.