GPU thread load balancing method and device, chip and electronic equipment

Abstract

The invention provides a GPU thread load balancing method and device, a medium and electronic equipment, and the method comprises the steps: fixing the total number of threads started by a GPU program by fixing the number of working groups started by the GPU program and the number of threads started by each working group; all calculation tasks needing to be processed are grouped and all put into a command queue, a global task queue is constructed, and the access authority of each working group to the global task queue is allowed; and according to the current calculation task load condition of the thread which is always started by the GPU, obtaining a calculation task which needs to be executed by each working group from the global task queue. According to the method, by establishing a dynamic mapping relation between threads and tasks and constructing a local queue and a global queue, dynamic allocation according to task loads is achieved, and finally load balancing between GPU threads is achieved.

Description

technical field [0001] The present invention relates to the field of GPU technology, and in particular, to a method, device, chip, and electronic device for balancing GPU thread loads. Background technique [0002] Graphics processing unit (GPU) is an important part of artificial intelligence. As an accelerator, it has been widely used in many fields such as high-performance computing and image processing. [0003] Existing GPU chips have large-scale and fine-grained parallel architecture characteristics, which are suitable for processing large-scale data parallel tasks; at the same time, in traditional GPU programming, the mapping between threads and tasks is static. The mapping relationship will not change. Existing GPUs are very friendly to applications with large-scale data parallelism, which can greatly improve the performance of applications. However, for scenarios that require a combination of graphics processing algorithms and non-graphical algorithms such as artif...

AI Technical Summary

Problems solved by technology

However, this method does not fundamentally eliminate the unbalanced GPU thread load, and introduces multiple kernels, which also increases kernel startup and synchronization overhead, which has limited performance improvement

The coarse-grained parallel method reduces the number of open GPU threads and increases the workload of each thread, so as to reduce the performance bottleneck caused by the complex and unbalanced GPU threads, but this method still does not fundamentally eliminate the gap between GPU threads. load imbalance problem

Images