Method and device for detecting GPU memory competition based on CPU end

Abstract

This invention discloses a method and apparatus for detecting GPU memory contention on the CPU side. The method includes: embedding a callback function at the API entry point of a GPU application during runtime to dynamically capture the context information corresponding to each GPU call initiated by the GPU application; constructing a CPU simulation execution environment on the CPU side that is semantically equivalent to parallel execution on the GPU side based on the context information; enabling TSAN in this environment to monitor the concurrent access behavior of multiple CPU threads to the same mapped memory address; and when the monitoring results indicate that TSAN has detected data contention, mapping the contention information associated with the contention event back to the context information to complete the detection of GPU memory contention. This invention solves the technical problem in related technologies where, in GPU parallel computing scenarios, when multiple kernel functions concurrently access the same memory address through different streams, causing data contention, it is impossible to effectively locate and diagnose such contention, leading to program instability.

Images