You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
 
inspur-inna dd2319cd56
update README.md.
6 years ago
Image add inspur logo 6 years ago
docs use tvm.Relay replace nnvm 6 years ago
inna use tvm.Relay replace nnvm 6 years ago
inna.egg-info use tvm.Relay replace nnvm 6 years ago
tests use tvm.Relay replace nnvm 6 years ago
tools use tvm.Relay replace nnvm 6 years ago
.gitignore add .gitignore 6 years ago
.gitmodules remove submodule 6 years ago
LICENSE Initial commit 6 years ago
README.md update README.md. 6 years ago
requirements.txt use tvm.Relay replace nnvm 6 years ago
setup.py use tvm.Relay replace nnvm 6 years ago

README.md

Image text

基于FPGA的CNN自适应映射技术——inna1.0

基于FPGA板卡设计深度学习加速器并进行优化在整体性能和功耗方面拟达到业界领先水平映射技术采用宏指令的Look-Aside Acceleration框架实现了一键式快速部署、软硬件协同优化、支持多种卷积、执行过程无需主机干预。本项目为映射技术的软件端拟实现CNN映射编译器和CNN量化器首先由TensorFlow产生的模型文件解析产生CNN的计算图模型CNN映射编译器会根据解析的计算图和现有的CNN加速库单元选择相应的CNN库单元生成相应的硬件结构和相应的调度器的配置参数以达到计算、片上存储、片上带宽和片外带宽的均衡从而达到最优的计算性能CNN量化器可根据模型的权重文件对各层数据进行8位定点量化以便于FPGA的DSP计算从而在保证精度的前提下降低存储开销提高处理速度降低功耗。

Install

inna install

TVM need LLVMLLVM install in Ubuntuother system require source code compilation

apt search llvm
apt install llvm-6.0
apt install clang-6.0

Install miniconda for python=3.6install_inna.sh include TVM install scriptrefer to TVM https://tvm.apache.org/docs/install/from_source.html

conda create -n inna python=3.6 ipykernel -y
conda activate inna
cd inna/tools && ./install_inna.sh

Run

Compiler

Compiler in TensorFlow or Mxnet or keras or onnx.

$ python compiler.py

Quantizer

Quantizer in TensorFlow.

$ python quantize.py

Runtime

Runtime in ours.

$ python runtime.py