Efficient calling method and device for acceleration library in PyTorch framework

• Main Authors: SUN YUFEI, SUI YICHENG, ZHANG ZHIQIANG, SHI CHANGQING, WANG HAOTIAN, ZHANG YUZHI
• Format: Patent
• Language: Chinese; English
• Published: 27.02.2024
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; ELECTRIC DIGITAL DATA PROCESSING; PHYSICS

The invention discloses an efficient calling method and device for an acceleration library in a PyTorch framework, and relates to the technical field of efficient calling of acceleration libraries. During handle binding, a handle binding class is created in advance based on a singleton mode, a handle instance is generated only during first binding, the handle instance is directly taken out subsequently, and during video memory space calculation, a first cache class is created; the method comprises the following steps: creating a first cache class when a convolution algorithm is selected, calculating a video memory space of the convolution kernel parameter class only during first calculation, directly extracting the video memory space of the convolution kernel parameter class subsequently, creating a second cache class when the convolution algorithm is selected, selecting the convolution algorithm of the convolution kernel parameter class only during first selection, and directly extracting the convolution alg