用于神经网络处理器的数据处理方法、装置、计算机设备、存储介质及程序产品
By partitioning and transforming tensor data, the problem of tensor size misalignment in neural network processors is solved, achieving correctness and efficiency in data transmission and adapting to the complex scenarios of DDR memory interleaving processing.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-22
- Publication Date
- 2026-07-17
AI Technical Summary
When processing tensor data, neural network processors struggle to meet the alignment requirements between tensor size and the granularity of the hardware storage system, leading to issues with the real-time performance and accuracy of data conversion, especially when data arrives out of order during interleaved processing in DDR memory.
By partitioning and transforming tensor data, the data length is ensured to be aligned with the hardware storage system. A streaming conversion buffer is used to handle the out-of-order data problem during DDR memory interleaving and relocation. Multiple DMAs are used for data transfer and to ensure order preservation.
It achieves correct data conversion and full-bandwidth transmission in complex scenarios, improves the flexibility and adaptability of data processing, and supports different address storage granularities.
Smart Images
Figure CN122220257B_ABST
Abstract
Citation Information
Patent Citations
Data transmission method based on remote direct memory access, computer equipment and medium
CN120670364A
Space-depth conversion optimization with DMA and DPU executors
WO2025065721A1