A method for adaptive selection of JPEG decoding CPU, GPU path based on runtime performance measurement

Abstract

This invention relates to a method for adaptive selection of CPU and GPU paths for JPEG decoding based on runtime performance measurement, belonging to the field of heterogeneous computing technology. The method includes pre-measuring the performance of a heterogeneous platform consisting of a CPU, GPU, and the PCIe bus connecting the CPUs to obtain the average time the CPU spends computing a single DCT block, the effective bidirectional bandwidth of the PCIe bus, and the fixed startup overhead of the GPU task; online measurement to obtain the actual computation time of the GPU executing a single DCT block; the CPU performs entropy decoding on the input JPEG image to obtain _DCT blocks; the CPU predicts the total time based on N and _, and the GPU predicts the total time based on _, _, and _; based on the CPU's predicted total time and the GPU's predicted total time, the method adaptively selects either the CPU path or the GPU path to decode _DCT blocks; if the GPU path is selected, when the current operating frequency of the GPU exceeds a preset threshold, the method updates the frequency in real time based on the current operating frequency of the GPU. This achieves adaptive selection of CPU or GPU for JPEG image decoding.

Images