OGLE

opengl2.0 based deeplearning inference engine, used for stricted environment, like web, opencl-disable arm

Introduction

OGLE is mainly for deeplearning inferece, it has two parts,

1. Model Converter
2. opengl based graph runtime you need to convert model you trained to onnx format first. then use converter to convert model from onnx to dlx format, only dlx model can used for graph runtime

Install

Prerequisite

1. opengl libraries
2. common utils libraries, like glog gtest gflag
3. opencv, protobuf please build from source

compile

# build converter
cd converter
mkdir build && cd build && cmake .. && make -j`nproco`

# build runtime
cd ..
cd ./ogl_runtime/opengl/nn
# generate kernel files
python make_shaders.py glsl all_shaders.h all_shaders.cc
cd -

mkdir build && cd build && cmake .. && make -j`nproco`

How to use

• convert model first

cd converter/build
export SRC=model.onnx
exoprt DST=demo.dlx
./Converter ${SRC} ${DST}
# then you can get the demo.dlx in current directory

• to use opengl based graph runtime, please refer to ogl_runtime/opengl/examples/ssd/main.cc for example.

# demo
cd ogl_runtime/build
./opengl/examples/ssd_detector

• trained model download from google drive or baiduyunpan , passwd: ef7i, then put demo.dlx into build directory, you can use the person detector model to detect in camera or image. the follow image is used for demostration.

Benchmark

framework	model	time	platform	mem
OGLE	mobilenet ssd	8ms	1080TI	187m
MNN	mobilenet ssd	9-10ms	1080TI	174m

Develop

1. add new op please read ogl_runtime/opengl/nn/glsl and ogl_runtime/opengl/nn/kernels for reference. then you need to rerun python make_shaders.py glsl all_shaders.h all_shaders.cc again