Computational graph execution method and device and related equipment

Abstract

The invention discloses a computational graph execution method. The method comprises the steps of obtaining a target computational graph; layering the target calculation graph by using a preset layering algorithm to obtain an optimal operator layering combination; grading each algorithm hierarchy in the optimal operator hierarchy combination according to a preset grading strategy to obtain each pipeline grade; performing serial processing on each operator hierarchy belonging to the same pipeline hierarchy according to the pipeline hierarchy level, and performing serial processing on each pipeline hierarchy to obtain an operation result, wherein each running water stage is in a running water parallel operation state. By applying the technical scheme provided by the invention, load balancing can be realized in the computational graph scheduling process, and the resource utilization rate is improved. The invention further discloses a computational graph execution device, computer equipment and a computer readable storage medium which all have the above beneficial effects.

Description

technical field [0001] The present application relates to the technical field of deep learning, in particular to a calculation graph execution method, and also relates to a calculation graph execution device, computer equipment, and a computer-readable storage medium. Background technique [0002] With the development of deep learning, the neural network model is becoming more and more complex, and the amount of calculation is increasing. How to execute the entire calculation graph more efficiently becomes more and more important. In different scenarios, there are high requirements on the real-time performance of data processing, and the execution order of operators in the entire calculation graph has a crucial impact on performance. At present, in the traditional deep learning framework, the optimization of computation graph execution scheduling is emerging in an endless stream. [0003] Currently, computation graph execution scheduling in various deep learning frameworks ...

AI Technical Summary

Problems solved by technology

Sequential scheduling only schedules one operator at a time, which will cause serious waste of resources

Greedy scheduling schedules operators that can be scheduled every time, but the number of operators scheduled at different times is different. This situation will lead to resource competition or waste of resources at different stages, and there is a serious imbalance in the scheduling strategy.

It can be seen that these two algorithms do not consider the parallelism and load balancing of scheduling among operators, and have the disadvantages of low resource utilization and low performance.

[0004] For example, refer to figure 1 and figure 2 , figure 1 It is a schematic diagram of a calculation graph provided by this application, figure 2 In order to adopt the greedy scheduling strategy for figure 1 The schematic diagram of the scheduling of the calculation graph shown in the figure, in the first stage, (a) (b) (c) operators are scheduled at the same time. At this time, there are many scheduling tasks, the equipment utilization rate reaches 95%, and there will be serious resource competition among tasks. In the third stage, only one operator is scheduled. At this time, there are fewer scheduling tasks, and the equipment cannot be fully utilized, and the utilization rate is only 10%.

Images