This workload implements significant aspects of SOTA deep learning architectures (convolutions only and for inference) for image classification using the MNIST dataset. MNIST is one of the oldest and most popular data sets in machine learning.

The 4 data files for the DL-MNIST workload that are needed to run it are:

The datafiles can be downloaded from (http://yann.lecun.com/exdb/mnist/)

Create a directory called 'datasets' in 'dl-mnist' directory. Copy the files into the 'datasets' directory.

mkdir dl-mnist/datasets
cp t10k-images.idx3-ubyte dl-mnist/datasets/.
cp t10k-labels.idx1-ubyte dl-mnist/datasets/.
cp train-images.idx3-ubyte.dat dl-mnist/datasets/.
cp train-labels.idx1-ubyte dl-mnist/datasets/.

To build the cuda program:

cd dl-mnist/CUDA
mkdir build
cd build
cmake ..
make
./dl-mnist-cuda -conv_algo CUDNN_FIND_BEST_ALGO

Note: There is a parameter for cmake called USE_SM. For A100, it's value is 80, and is the default. So it need not be passed in (as shown above). But for other machine types, this needs to be explicitly passed in. For example, running on a RTX6000, the parameter would be USE_SM=75. The full command would be cmake -DUSE_SM=75 ..

To build the sycl program:

cd dl-mnist/SYCL
mkdir build
cd build
CC=dpcpp CXX=dpcpp cmake -DGPU_AOT=PVC ..
make
./dl-mnist-sycl -conv_algo ONEDNN_AUTO

To build the sycl program:

cd dl-mnist/HIP
mkdir build
cd build
CC=hipcc CXX=hipcc cmake ..
make
./dl-mnist-hip -conv_algo MIOPEN_FIND_BEST_ALGO

To build the sycl program, you will need 1) Open source oneAPI DPC++ compiler setup to run on NVIDIA-BACKEND 2) cudnn and oneDNN would be required:
We use the flag -DUSE_SM=80 for A100. Use the correct value for your GPU.

cd dl-mnist/SYCL
mkdir build
cd build
CC=clang CXX=clang++ cmake -DUSE_NVIDIA_BACKEND=YES -DUSE_SM=80 -DDNNLROOT=/path/to/oneDNN/install/location ..
make
ONEAPI_DEVICE_SELECTOR=cuda:gpu ./dl-mnist-sycl -conv_algo ONEDNN_AUTO

To build the sycl program, you will need 1) Open source oneAPI DPC++ compiler setup to run on AMD-BACKEND 2) MIOpen and oneDNN would be required:
We use the flag -DUSE_AMDHIP_BACKEND=gfx90a for MI250. Use the correct value for your GPU.

cd dl-mnist/SYCL
mkdir build
cd build
CC=clang CXX=clang++ cmake -DUSE_AMDHIP_BACKEND=gfx90a -DDNNLROOT=/path/to/oneDNN/install/location ..
make
ONEAPI_DEVICE_SELECTOR=hip:gpu ./dl-mnist-sycl -conv_algo ONEDNN_AUTO

func - All function ENTER/EXIT
mem - Memory operations
conv - Convolution operations

All three are optional, and can be enabled by providing a comma separated list of them as options to '-trace' paramerer.
Example:
./dl-mnist-sycl -trace func,mem,conv

DL-MNIST uses oneDNN and cuDNN. They can be enabled by setting certain environment variables described below:
oneDNN - https://oneapi-src.github.io/oneDNN/dev_guide_verbose.html
cuDNN - https://docs.nvidia.com/deeplearning/cudnn/developer-guide/index.html#api-logging