NVIDIA OpenACC

Type: Programming Model Tags: NVIDIA, OpenACC, HPC SDK, GPU offload, compiler directives, Fortran, C, C++, CUDA Related: NVIDIA-HPC-SDK, NVIDIA-HPC-Compilers, CUDA-Fortran, NVIDIA-Fortran-CUDA-Interfaces, NVIDIA-Stdpar, NVIDIA-CUDA, CUDA-Unified-Memory, CUDA-Compatibility, CUDA-Graphs, Nsight-Systems, Nsight-Compute Sources: https://docs.nvidia.com/hpc-sdk/compilers/openacc-gs/index.html, https://docs.nvidia.com/hpc-sdk/compilers/hpc-compilers-user-guide/index.html, https://docs.nvidia.com/hpc-sdk/index.html Last Updated: 2026-04-29

Summary

NVIDIA OpenACC refers to the OpenACC implementation and guidance in NVIDIA-HPC-SDK. OpenACC itself is an industry-standard directive API, while NVIDIA’s docs show how to build and run OpenACC Fortran, C, and C++ programs with the NVIDIA HPC compilers, targeting NVIDIA GPUs or multicore CPUs.

Detail

Purpose

OpenACC helps existing scientific and HPC applications offload loops and code regions to GPUs with compiler directives instead of a full CUDA rewrite. NVIDIA’s implementation matters because it lets Fortran, C, and C++ applications use GPU acceleration through nvfortran, nvc++, and nvc while still interoperating with CUDA memory/runtime behavior where needed.

Current scope

• Enable OpenACC with the -acc compiler option.
• Target NVIDIA GPUs by default, or use -acc=multicore and -acc=host for CPU/host paths.
• Use -gpu suboptions to control CUDA version, compute capability targets, memory mode, and code generation.
• NVIDIA HPC SDK 26.3 includes CUDA Toolkit 12.9U1 and CUDA 13.0 components for compiler-selected GPU code generation.
• Current docs cover CUDA version/driver matching, compute capability targeting, PTX JIT compilation, C structs/classes, atomics, declare-data handling, error handling, runtime routines, profiling, and limitations.

NVIDIA context

This page is intentionally about NVIDIA’s OpenACC implementation in the HPC SDK, not the OpenACC standard as a whole. It connects CUDA compatibility, compiler targets, driver/toolkit matching, and profiler visibility to directive-based GPU offload.

Connections

• NVIDIA-HPC-SDK - distribution and documentation hub.
• NVIDIA-HPC-Compilers - compiler family that implements OpenACC.
• CUDA-Fortran - explicit Fortran CUDA model that can interoperate with OpenACC workflows.
• NVIDIA-Fortran-CUDA-Interfaces - documents how OpenACC host code can pass device pointers to CUDA library calls.
• NVIDIA-Stdpar - standard-language parallelism path that can interoperate with OpenACC data directives.
• NVIDIA-CUDA - GPU runtime and toolkit layer underneath OpenACC offload.
• CUDA-Unified-Memory - memory mode used by OpenACC guidance.
• CUDA-Compatibility - driver/toolkit matching matters for OpenACC-generated GPU code.
• CUDA-Graphs - current HPC compiler docs include OpenACC and CUDA Graphs context.
• Nsight-Systems and Nsight-Compute - profiling tools for OpenACC-generated kernels.

Resources

• OpenACC Getting Started Guide
• HPC Compilers User’s Guide