19801201/spinalhdl_cnn_accelerator

CNN accelerator implemented with Spinal HDL

About SpinalHDL_CNN_Accelerator

中文说明

Implementing accelerator supporting various neural networks in SpinalHDL

• This repository is implementing common operators in CNN in SpinalHDL 1.7.3.It can be guided by compiler to generate accelerators for various neural networks.
• This repository aims to implement a general accelerator at the SpinalHDL level and a special accelerator at the Verilog level.
• This repository implements an Npu module, which can be put into the project,as shown in the figure.

• The resource occupation in FPGA is shown in the figure below:

Description

This repository hosts a CNN Accelerator implementation written in SpinalHDL. Here are some specs :

• Easy to use
• Complete implementation of convolution, quantization and shape
• Configurable parametric interface
• Optional convolution kernel type
• The FPGA resource occupation is optimized at the code level
• Implementation of rich tool class code
• Access control of DMA and AXI
• Automatically generate TCL files for instantiating Xilinx IP
• Top-level files are easy to run
• Implementation of the simulation of convolution

RTL code generation

You can find three runnable top modules in:

• src/main/scala/top.scala
• src/main/scala/Npu.scala
• src/test/scala/TbConv.scala

NOTES:

• It could take time when you run it.
• top.scala is used to implement convolution.
• Npu.scala is used to implement complete project process.
• TbConv.scala is used to implement the simulation of convolution.
• Before you run the TbConv.scala,you should prepare files with feature and weight in advance.

spinalhdl_cnn_accelerator's Issues

请问实现一个神经网络（比如说yolo，resnet），开发周期大概有多长呢

我还没有实际操作，在收集算法的fpga实现时，发现有人推荐Spianl HDL，但是不知道大概一个神经网络的FPGA实现需要多久

Can you provide a reproduction process?

Can you provide a reproduction process? The amount of code is too large, and I don't know which module to start debugging and reproducing.

How to start?

我已经安装了能够运行spinalHDL官方模板的环境，在Intellij上运行，但是我如何运行该加速器？在命令行中使用sbt run 导入依赖库会卡在这一步
[info] set current project to spinal_yolo_dev (in build file:/D:/project/spianl/spianl_CNN_nishuo/SpinalHDL_CNN_Accelerator/)
| => spinalhdl_cnn_accelerator / update 102s