Метод програмної конвеєризації для графічних співпроцесорів

Graphics Processing Units (GPUs) play a significant role in high-end computations, including artificial intelligence. However, the GPU hardware is often underloaded. This forces an increased volume of GPU hardware to maintain a high throughput for task execution. The low loading of the GPU resources remains an actual problem and it needs to be solved now. Therefore, it is essential to seek methods that enhance GPU loading. The research object is computational processes in modern processors, especially in GPUs. The purpose of this study is to review the software pipelining approach, its advantages and disadvantages, the techniques that can be used in it, including both instruction-level and decoupled versions, and to assess the effectiveness of this approach for the GPU. To satisfy the requirements, different analysis methods were used. First, the architectural requirements to apply software pipelining were reviewed. Second, the original formulation and historical development of the approach were examined. Third, different levels of parallelisation to implement software pipelining were explored. Finally, C-slowing was proposed as an optimisation technique to overcome the adversities of the underutilisation of computational resources. The research has revealed the abundance of proper software pipelining for GPU implementations. Whereas existing works review the possibilities of this technique, they are often overlooked in contrast to simpler multi-threading techniques. However, investigated researchers have defined the crucial limiting factor to computational resources as a constraint by memory overloading, specifically the pipelining registers. To address this, the C-slowing approach was suggested and theoretically evaluated. It demonstrated a possible increase of over 30% in GPU loading for the analysed algorithm, proving its applicability. In conclusion, the software pipelining approach shows decent potential to optimise GPU algorithms, requiring further investigation. C-slowing could be utilised to handle the problem of underutilisation of computation.