CUDA is a parallel programming model and software environment developed by NVIDIA. It provides programmers with a set of instructions that enable GPU acceleration for data-parallel computations. The computing performance of many applications can be dramatically increased by using CUDA directly or by linking to GPU-accelerated libraries.<\/p>\n
To link and run applications using CUDA you will need to make some changes to your path and environment. Load the appropriate version of cuda:<\/p>\n
module load<\/span> cuda\/11.3<\/code><\/pre>\nThe list of available versions of cuda can be obtained by executing the module avail cuda<\/span><\/code> command.<\/p>\nCompiling a simple CUDA C\/C++ program<\/h2>\n
Consider the following simple CUDA program gpu_info.cu<\/a> that prints out information about GPUs installed on the system.<\/p>\nDownload the source code of gpu_info.cu<\/a> and transfer it to the directory where you are working on the SCC.
\nThen execute the following command to compile gpu_info.cu<\/code>. Depending on the version of the cuda<\/code> module that you are using you may need to load a newer version of the gcc compiler, such as gcc\/8.3.0<\/code>. The CUDA nvcc<\/code> compiler will print an error if a newer compiler is required:<\/p>\nscc1%<\/span> nvcc -o<\/span> gpu_info gpu_info.cu<\/span><\/code><\/pre>\nRunning a CUDA program interactively on a GPU-enabled node<\/h2>\n
To execute a CUDA code, you have to login via interactive batch to a GPU-enabled node on the SCC. To request an interactive session with access to 1 GPU:<\/p>\n
scc1%<\/span> qrsh<\/span> -l gpus=1<\/span><\/code><\/pre>\nTo run a CUDA program interactively, you then type in the name of the program at the command prompt:<\/p>\n
gpunode%<\/span><\/span> gpu_info<\/span><\/code><\/pre>\nSubmit a CUDA program Batch Job<\/h2>\n
The following line shows how to submit the gpu_info<\/code> program to run in batch mode on a single CPU with access to a single GPU:<\/p>\nscc1%<\/span> qsub -l gpus=1<\/span> -b y gpu_info<\/span><\/code><\/pre>\nwhere the \u2013l gpus=<\/span>#<\/span><\/code> option indicates the number of GPUs. To learn about all options that could be used for submitting a job, please visit the running jobs<\/a> and GPU computing<\/a> pages.<\/p>\nCUDA Libraries<\/h2>\n
Several scientific libraries that make use of CUDA are available:<\/p>\n
\n- cuBLAS<\/a> \u2013 Linear Algebra Subroutines. A GPU accelerated version of the complete standard BLAS library.<\/li>\n
- cuFFT<\/a> \u2013 Fast Fourier Transform library. Provides a simple interface for computing FFTs up to 10x faster.<\/li>\n
- cuRAND<\/a> \u2013 Random Number Generation library. Delivers high performance random number generation.<\/li>\n
- cuSparse<\/a> \u2013 Sparse Matrix library. Provides a collection of basic linear algebra subroutines used for sparse matrices.<\/li>\n
- NPP<\/a> \u2013 Performance Primitives library. A collection of image and signal processing primitives.<\/li>\n<\/ul>\n
Architecture specific options<\/h2>\n
Architecture specific features can be enabled using the \u2013arch sm_##<\/span><\/code> flag during compilation. The “sm<\/code>” stands for “streaming multiprocessor” and the number following sm_<\/b> indicates the features supported by the architecture. For example, for a CUDA program running on the SCC you can add the \u2013arch sm_60<\/code> flag to allow for functionality available on GPUs that have Compute Capability 6.0 (Pascal architecture). See the CUDA Toolkit documentation<\/a> for more information on this.<\/p>\nAdditional CUDA training resources<\/h2>\n
NVIDIA provides resources for learning CUDA programming at
\nhttps:\/\/developer.nvidia.com\/cuda-training<\/a>.<\/p>\nCUDA Consulting<\/h2>\n