The goal of this project is two fold. First, accelerate the forward propagation step of the Convolutional Neural Network (CNN) algorithm using CUDA and OpenCL. Second, evaluate its performance on NVIDIA, AMD GPUs, and an FPGA platform.
Provided is a model that has been trained using 60,000 examples (training set images) and the provided test data is 10,000 batched queries (test set images). The expected accuracy of the CNN is ~87% on the provided test dataset.
The dataset and model are from the MNIST database.
The project requires a C++ compiler, CUDA 8 Toolkit, and OpenCL 1.2 or higher.
The CUDA 8 Toolkit can be downloaded from the CUDA Download page. Instructions on how to install the CUDA Toolkit are available in the Quick Start page. Installation guides and the list of supported C compilers for Windows, Linux, and OSX are also found in the CUDA Toolkit Documentation Page. Aside from a C compiler and the CUDA 8 Toolkit, CMake 3.1 or later is required to generate build scripts for your target IDE and compiler.
To build the CUDA and OpenCL versions the Hunter package manager with Cmake needs to be used. Install the libraries needed (mainly HDF5).
Assuming that you have checked out the project into $PROJECT_DIR, execute the following sequence of commands:
cd $PROJECT_DIR/cuda_implementation/
mkdir build
cd build
cmake ../
This will download the required software needed for the project (see the hunter docs for more information). You may see some warning while the system is compiling HDF5, which you can ignore. Once CMake has been run, a Makefile is generated so you can then perform make to build the project.
make
The same sequence of commands need to be executed in the openCL_implementation directory to run the OpenCL version.
If you do not plan on using make, examine the cmake -G option which allows you to generate XCode, Visual Studio, ... project configurations. You may also need to change the build type to enable/disable debugging and/or optimizations.
If you need to use another library, you need have to modify the [CMakeLists.txt] and add the libraries to the target_link_libraries (and possibly the include_directories) section. Documentation on the CMake commands is found in the documentation page.
Since the FPGA version needs to use modified versions of the OpenCL libraries provided by Intel, it makes more sense to just use the Makefile provided by Intel.
The first step in building is figuring out which boards are available. The following command will list all of the available boards:
aoc --list-boards
cd $PROJECT_DIR/fpgaOpenCL_implementation/
aoc device/cnn.cl -o bin/cnn.aocx --board <board>
cd $PROJECT_DIR/fpgaOpenCL_implementation/
aoc -march=emulator device/cnn.cl -o bin/cnn.aocx --board <board>
Assuming that you have checked out the project into $PROJECT_DIR, execute the following sequence of commands:
cd $PROJECT_DIR/fpgaOpenCL_implementation/
make
This will generate a bin folder that will contain the executable that can run the FPGA version of the code.
Test your implementation with small batch size first to verify the correctness. You can parse the data/test100.hdf5 into smaller chunks using your preferred language(e.g. python). 2, 10 and 100 queries are provides in data/test2.hdf5, data/test10.hdf5 and data/test100.hdf5 in the data folder. Maker sure the data file you feed in has the same batch size as the batch_size you specify in the command line.
To run the code, you need to be in the build or bin directories of a given version of the project.
./cuda_CNN ../../data/test10.hdf5 ../../data/model.hdf5 10
./openCL_CNN ../../data/test10.hdf5 ../../data/model.hdf5 10
./fpga_CNN ../../../data/test10.hdf5 ../../../data/model.hdf5 10
CL_CONTEXT_EMULATOR_DEVICE_ALTERA=1 ./fpga_CNN ../../../data/test10.hdf5 ../../../data/model.hdf5 10
The following are the results on a Tesla P100 GPU as of 4/22/2017.
test10: 15.9631 milliseconds
test100: 17.6193 milliseconds
testfull: 524.852 milliseconds
test10: 3.28717 milliseconds
test100: 3.76949 milliseconds
testfull: 34.1698 milliseconds
Need to get access to system with an FPGA board to test this version of the code.
test10: 7423.32 milliseconds
test100: 68188.2 milliseconds
testfull: TIMES OUT
Note: Major optimizations need to be made before the results from the FPGA version can match the results from the OpenCL and CUDA versions of the code.
Please use the GitHub issue manager to report any issues or suggestions about the project.