基于不规则卷积算子自适应分发的RISC-V模型推理加速方法、装置、设备及介质
By acquiring fine-grained shape and hardware characteristic information of the convolution operator, adjusting the operator distribution code, and conducting performance tests, the problem of low execution efficiency of irregular convolution operators on the RISC-V hardware platform was solved, and the model inference speed was improved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- INST OF SOFTWARE - CHINESE ACAD OF SCI
- Filing Date
- 2026-03-05
- Publication Date
- 2026-07-17
AI Technical Summary
Existing rule-based operator distribution methods cannot effectively adapt to the fine-grained shape information of irregular convolution operators and the characteristics of RISC-V hardware platforms, resulting in low computational and memory access efficiency and limiting the inference speed of deep learning models on edge devices.
By acquiring fine-grained shape information and hardware characteristic information of the convolution operator, the operator distribution code is adjusted using a large language model to generate candidate operator distribution codes. Performance tests are then conducted on the target hardware platform to select the operator distribution code with the highest execution efficiency.
It significantly improves the execution efficiency of irregular convolution operators on the RISC-V hardware platform, optimizes computation and memory access performance, and increases end-to-end model inference speed.
Smart Images
Figure CN121764697B_ABST