Abstract

This work presents a dynamically growable GPU array (GGArray) fully implemented in GPU that facilitates the programming of GPU applications with dynamic memory. The GGArray is based on an array of LFVectors, taking advantage of the GPU architecture and its synchronization at block level. The GGArray is compared to other state of the art approaches such as a pre-allocated static array and a semi-static array that needs to be resized through communication with the host. Experimental evaluation shows that the GGArray achieves an efficient memory usage close to the optimal and not greater than 2 x the needed memory, as well as a competitive insertion/resize performance, but it is slower for regular parallel memory accesses. Given these results, the GGArray is a potentially useful structure for applications with high uncertainty on the memory usage as it does not require pre-allocating GPU VRAM for the worst case scenario. It can also be useful in applications that exhibit phases in terms of memory behavior, such as an insertion phase followed by a regular r/w GPU phase. In these cases, the GGArray can be used for the first phase and then data can be flattened for the second phase in order to allow the regular and faster GPU memory accesses to take place. These results constitute a step towards achieving a parallel efficient C++ like vector for modern GPU architectures.