CUDA

"CUDA® is a parallel computing platform and programming model developed by NVIDIA for general computing on graphical processing units (GPUs)."

It is reasonable to think of CUDA as a set of libraries and associated C, C++, and Fortran compilers that enable you to write code for GPUs. See OpenACC Tutorial for another set of GPU programming tools.

Quick start guide

Compiling

Here we show a simple example of how to use the CUDA C/C++ language compiler, nvcc, and run code created with it. For a longer tutorial in CUDA programming, see CUDA tutorial.

First, load a CUDA module.

$ module purge
$ module load cuda

The following program will add two numbers together on a GPU. Save the file as add.cu. The cu file extension is important!.

File : add.cu

#include <iostream>

__global__ void add (int *a, int *b, int *c){
  *c = *a + *b;
}

int main(void){
  int a, b, c;
  int *dev_a, *dev_b, *dev_c;
  int size = sizeof(int);
  
  //  allocate device copies of a,b, c
  cudaMalloc ( (void**) &dev_a, size);
  cudaMalloc ( (void**) &dev_b, size);
  cudaMalloc ( (void**) &dev_c, size);
  
  a=2; b=7;
  //  copy inputs to device
  cudaMemcpy (dev_a, &a, size, cudaMemcpyHostToDevice);
  cudaMemcpy (dev_b, &b, size, cudaMemcpyHostToDevice);
  
  // launch add() kernel on GPU, passing parameters
  add <<< 1, 1 >>> (dev_a, dev_b, dev_c);
  
  // copy device result back to host
  cudaMemcpy (&c, dev_c, size, cudaMemcpyDeviceToHost);
  std::cout<<a<<"+"<<b<<"="<<c<<std::endl;
  
  cudaFree ( dev_a ); cudaFree ( dev_b ); cudaFree ( dev_c );
}

Compile the program with nvcc to create an executable named add.

$ nvcc add.cu -o add

Submitting jobs

To run the program, create a Slurm job script as shown below. Be sure to replace def-someuser with your specific account (see accounts and projects). For options relating to scheduling jobs with GPUs see Using GPUs with Slurm.

File : gpu_job.sh

#!/bin/bash
#SBATCH --account=def-someuser
#SBATCH --gres=gpu:1              # Number of GPUs (per node)
#SBATCH --mem=400M                # memory (per node)
#SBATCH --time=0-00:10            # time (DD-HH:MM)
./add #name of your program

Submit your GPU job to the scheduler with this command.

$ sbatch gpu_job.sh
Submitted batch job 3127733

For more information about the sbatch command and running and monitoring jobs see Running jobs.

Once your job has finished you should see an output file similar to this.

$ cat slurm-3127733.out
2+7=9

If you run this without a GPU present you might see output like 2+7=0.

Linking libraries

If you have a program that needs to link some libraries included with CUDA, for example cuBLAS, compile with the following flags

nvcc -lcublas -Xlinker=-rpath,$CUDA_PATH/lib64

To learn more about how the above program works and how to make the use of a GPUs parallelism see CUDA tutorial.

Troubleshooting

"Compute Capability"

NVidia has created a technical term "compute capabilty" which they describe as follows:

The compute capability of a device is represented by a version number, also sometimes called its "SM version". This version number identifies the features supported by the GPU hardware and is used by applications at runtime to determine which hardware features and/or instructions are available on the present GPU." (CUDA Toolkit Documentation, section 2.6)

The following errors are connected with "compute capability":

nvcc fatal : Unsupported gpu architecture 'compute_XX'

no kernel image is available for execution on the device (209)

If you encounter either of these errors, you may be able to fix it by adding the correct FLAG to the nvcc call:

-gencode arch=compute_XX,code=[sm_XX,compute_XX]

Or if you are using cmake instead of nvcc directly, provide the following flag:

cmake .. -DCMAKE_CUDA_ARCHITECTURES=XX

where “XX” is the "compute capability" of the Nvidia GPU that you expect to run the application on. To find the value to replace “XX“, see the Available Hardware table on the page Using GPUs with Slurm.

For example, if you will run your code on a Narval A100 node, its "compute capability" is 80. The correct FLAG to use when compiling with nvcc is

-gencode arch=compute_80,code=[sm_80,compute_80]

The flag to supply to cmake is:

cmake .. -DCMAKE_CUDA_ARCHITECTURES=80