A method and apparatus for accelerating a convolution network based on a macro instruction set

Abstract

The invention discloses a convolution network acceleration method based on a macro instruction set, which is based on a convolution network accelerator. The method comprises steps: a macro instructionset sent by a host is received; each macro instruction in the macro instruction set includes macro instruction parameters for executing different operations; the mapping and scheduling from convolution network to convolution network accelerator is realized by parsing every macro instruction in macro instruction set. It can be seen that the present invention discloses a macro instruction set of aconvolution network accelerator, the macro instruction set consists of every part of convolution network into a macro instruction, which realizes the efficient mapping and scheduling of convolution network to the hardware of convolution network accelerator with different specifications and types, simplifies the mapping process of convolution network in depth neural network, and has a wide range ofapplication scenarios. The invention also discloses a convolution network accelerating device based on a macro instruction set, which can also realize the technical effect.

Description

technical field [0001] The present invention relates to the technical field of convolutional network acceleration, and more specifically, to a method and device for accelerating a convolutional network based on a macro instruction set. Background technique [0002] In the current design of neural network accelerators, processes such as convolution, pooling, and normalization are treated as independent components, and independent instructions are designed respectively. Simultaneous operation and out-of-order execution of multiple components will lead to complex scheduling problems. And increase the difficulty of deep neural network compilation and mapping. [0003] Therefore, how to simplify the mapping and scheduling process of the convolutional network in the neural network is a problem to be solved by those skilled in the art. Contents of the invention [0004] The object of the present invention is to provide a convolutional network acceleration method and device based...

AI Technical Summary

Problems solved by technology

[0002] In the current design of neural network accelerators, processes such as convolution, pooling, and normalization are treated as independent components, and independent instructions are designed respectively. Simultaneous operation and out-of-order execution of multiple components will lead to complex scheduling problems. And increased the difficulty of deep neural network compilation and mapping

Images