⚠ INFORMATION
This page was translated by an AI (LLM) with a cursory human check and is awaiting full review.

Compilation of OpenACC code (Fortran, C/C++)

On Jean Zay, the NVIDIA HPC compilers provide the possibility to use OpenACC.

Compilation options for NVIDIA HPC

The documentation for NVIDIA compilers is available on the NVIDIA website.

Taking OpenACC into account

The compilation options for OpenACC support by the NVIDIA HPC compiler are as follows:

• -acc: this option enables OpenACC support. It can be given the following sub-options:
  • [no]autopar: Enable (or not) automatic parallelisation for the ACC PARALLEL directive. The default is to enable it.
  • [no]routineseq: Compile (or not) all routines for the accelerator. The default is not to compile each routine as sequential.
  • strict: Display warning messages in case of use of non-OpenACC directives for the accelerator.
  • verystrict: Stop the compilation in case of use of non-OpenACC directives for the accelerator.
  • sync: Ignore async clauses.
  • [no]wait: Wait (or not) for the end of each computation kernel on the accelerator. Kernel launch is blocked by default, unless async is used.
  • Example:

    nvfortran -acc=noautopar,sync -o prog_ACC prog_ACC.f90

`

• -ta: This option enables offloading computation to the accelerator. It implies the -acc option.
  • It is useful for choosing the target architecture for which the code will be compiled.
  • To use the V100 GPUs of Jean Zay, you must use the tesla sub-option of -ta and the compute capability cc70. For example:

    nvfortran -ta=tesla:cc70 -o prog_gpu prog_gpu.f90

  • Some useful examples of tesla sub-options:
    • managed: Allows creating a shared view of GPU and CPU memory.
    • pinned: Enables memory pinning on the CPU. This can improve data transfer performance.
    • autocompare: Enables comparison of results on CPU and GPU.

The other options are available in the Web documentation NVIDIA.