Video Processing on GPU: Analysis of Data Transfer Overhead

摘要

In this work, we study one of the major problems in exploring the power of GPUs to accelerate video processing applications: countless frames have to be transferred back and forth between the CPU and GPU. We evaluate four different data transfer approaches currently available on modern GPUs: Standard Allocation, Pinned Memory, Data Stream, and Zero-Copy. Our results show that Data Stream is the most efficient strategy, but requires more programming effort. Zero-Copy, on the other hand, demonstrates inferior performance due to the significant latency incurred by the PCIe bus transfers for every memory access.