Matrix convolution calculation method, interface, coprocessor and system based on risc-v architecture

Abstract

The invention discloses a set of instructions, interfaces, coprocessor methods and system integrity mechanisms for matrix convolution calculations based on the RISC-V instruction set architecture, and efficiently implements traditional matrix convolution calculations by combining software and hardware , using the scalability of the RISC-V instruction set, by designing a small number of instructions and a dedicated convolution calculation unit (ie, a coprocessor), the number of memory accesses and execution cycles of the matrix convolution calculation instructions are reduced, and the application layer software calculation is reduced. It improves the efficiency of large-scale matrix convolution calculations, improves the calculation speed of matrix convolutions, and is also conducive to flexible calls by upper-level developers and simplifies coding design. At the same time, processors designed using the RISC-V instruction set also have huge advantages over ARM, X86 and other architectures in terms of power consumption, volume and flexibility, and can adapt to different application scenarios, and have broad prospects in the field of artificial intelligence.

Description

technical field [0001] The invention relates to matrix convolution calculation, in particular to a matrix convolution calculation method, interface, coprocessor and system based on RISC-V architecture. Background technique [0002] With the development of artificial intelligence, image recognition, neural network, deep learning and other technologies have been widely and deeply applied. There are a large number of matrix convolution calculations in the underlying implementation. Matrix convolution is the most basic and core algorithm of these technologies. The calculation of matrix convolution is the multiplication and accumulation of a large amount of data. The traditional C / C++ language implementation requires multi-layer loop calculations, multiple memory accesses to read data, and runs on systems with traditional architectures. The calculation speed, power consumption and volume It has been unable to meet the requirements of modern artificial intelligence. [0003] The ...

AI Technical Summary

Problems solved by technology

The calculation of matrix convolution is the multiplication and accumulation of a large amount of data. The traditional C / C++ language implementation requires multi-layer loop calculations, multiple memory accesses to read data, and runs on systems with traditional architectures. The calculation speed, power consumption and volume Can no longer meet the requirements of modern artificial intelligence

[0003] The existing technology generally adopts the implementation method of high-level language, which has many memory access instructions, long execution time, poor real-time performance, and low energy efficiency ratio, which cannot meet the requirements of modern image processing.

Even if the matrix convolution operation implemented by ASIC chips is relatively fragmented in design, the versatility is not strong, and the flexibility is not enough.

Images