OpenMP Offload/Blas Examples

Example C and Fortran code showing how to offload blas calls from OpenMP regions, using cuBLAS, NVBLAS, and MKL.

There are three directories:

• cublas
• nvblas
• mkl

These contain Makefiles and examples of calling DGEMM from an OpenMP offload region with cuBLAS, NVBLAS, and MKL. Note: The NVBLAS Makefile is hard-coded for Summit.

To run them, cd into cublas, nvblas, or mkl, then cd into a "c" or "fortran" subdirectory, and compile with make.

• dgemm_nvblas uses NVBLAS
• dgemm_cublas explicitly uses cuBLAS
• dgemm_cublas_fortran calls cuBLAS from Fortran using iso_c bindings.
• dgemm_mkl uses MKL with the Intel extension target variant dispatch

A snippet of the code, showing the interface is:

For cublas:

#pragma omp target data map(to:aa[0:SIZE*SIZE],bb[0:SIZE*SIZE],alpha,beta) map(tofrom:cc_gpu[0:SIZE*SIZE]) use_device_ptr(aa,bb,cc_gpu)
 {
   cublasDgemm(handle,CUBLAS_OP_N, CUBLAS_OP_N,SIZE, SIZE, SIZE, &alpha, aa, SIZE, bb, SIZE, &beta, cc_gpu, SIZE);
  }

For nvblas:

#pragma omp target data map(to:aa[0:SIZE*SIZE],bb[0:SIZE*SIZE],alpha,beta) map(tofrom:cc_gpu[0:SIZE*SIZE]) use_device_ptr(aa,bb,cc_gpu)
 {
   dgemm("N","N",&size, &size, &size, &alpha, aa, &size, bb, &size, &beta, cc_gpu, &size);
  }

For mkl:

#pragma omp target variant dispatch use_device_ptr(aa, bb, cc_gpu) 
      cblas_dgemm( CblasColMajor, CblasNoTrans, CblasNoTrans,SIZE,SIZE,SIZE, alpha, aa,
		   SIZE, bb, SIZE, beta, cc_gpu, SIZE);

Building CuBLAS and NVBLAS examples on Summit

To run, the simplest is to submit an interactive job:

1. Submit an interactive job:

   $ bsub -W 1:00 -nnodes 1 -P Project -Is $SHELL
   

2. Set the environment:

   $ source environment_setup.sh
   

   This loads the cuda module, which is needed for tests with nvprof

3. Move into a subdirectory (e.g. nvblas/c or cublas/fortran)

   $ cd cublas/fortran
   

4. Compile the file

   $ make 
   

5. Use jsun to run each simple example:

   $ jsrun -n 1 -a 1 -c 1 -g 1 ./dgemm_nvblas
   
      -n is the nubmer of resource sets,
      -a is the number of MPI ranks,
      -c is the number of physical cores,
      -g is the number of GPUs
   
   $ jsrun -n 1 -a 1 -c 1 -g 1 nvprof --print-gpu-trace ./dgemm_cublas
   

   etc.