Conv2D Implementation with OpenCL

This project demonstrates various convolutional neural network (CNN) operations using OpenCL for GPU acceleration. The operations include convolution, ReLU activation, max pooling, batch normalization, tensor concatenation, upsampling and sigmoid activation.

Features

• Tensor Convolution
• ReLU Activation
• Max Pooling
• Batch Normalization
• Tensor Concatenation
• Upsampling

Prerequisites

• OpenCL SDK
• OpenCV
• cnpy
• Meson

Installation of OpenCL in ubuntu

foo@bar:~$ sudo apt update
foo@bar:~$ sudo apt install intel-opencl-icd

Checking the installation of OpenCL Drivers

foo@bar:~$ clinfo
foo@bar:~$
Number of platforms                               1
Platform Name                                   Intel(R) OpenCL HD Graphics
Platform Vendor                                 Intel(R) Corporation
Platform Version                                OpenCL 3.0 
Platform Profile                                FULL_PROFILE

Installation and Configuration

1. Install the required libraries:

   sudo apt-get install opencl-headers ocl-icd-opencl-dev
   sudo apt-get install libopencv-dev
   sudo apt-get install libboost-all-dev
   sudo apt-get install cmake,meson
   pip install -r requirements.txt

2. Clone the repository:

   git clone https://github.com/raghulrajn/OpenCL
   cd gpu

3. Clone cnpy:

   git clone https://github.com/rogersce/cnpy.git
   cp cnpy gpc/src

4. Download pretrained-kernel from ZF_UNET_224

   python3 extractWeights zf_unet_224.h5

   Weights and biases are extracted from the model and will be saved in pretrainedKernels folder

Required final structure

├── pretrainedKernels
├── gpu
    ├── lib
    │   ├── Core
    │   ├── lib
    │   ├── OpenCL
    │   └── vx
    ├── src
    │   └── cnpy
    │   ├── conv2d.cl
    │   ├── conv2d.cpp
    ├── meson.build
    ├── run.sh
    ├── results
    ├── sampleImages
    └── utils

Run the code

cd gpu
python3 genImage.py
chmod +x run.sh
./run.sh <PATH_TO_IMG NPY FILE>

Results after CPU and GPU execution are stored in results folder

Code Structure

• conv2d.cpp: Main source file containing the implementation of Conv2d, Maxpool, Upsampling and UNET operations using OpenCL.
• conv2d.cl: Kernel code for all GPU operations
• pytorch.py: Python code for convolution, Relu, maxpool etc. Results are stored in npy folder
• compare.py: Python code to crosscheck CPU, GPU and Pytorch results
• genImage.py: Python code to create sample images for inference. png image and npy file of the same will generated

Performance comparison

GPU performance

On Intel Iris Xe GPU

On AMD GPU

Results