NVRTC Native Runtime Libraries for CUDA 11

Warnings

• breaking NVRTC in CUDA 11.0 and later no longer implicitly adds `/usr/include` to the header file search path during compilation.
  Fix: If your CUDA C++ kernel code includes headers located in `/usr/include`, you must explicitly pass `-I /usr/include` (or the relevant path) to the `nvrtcCompileProgram()` API call or via the options argument in `pynvrtc`.
• gotcha This package (`nvidia-cuda-nvrtc-cu11`) provides only the native NVRTC shared libraries. It does not expose a direct Python API itself. Users must install a separate Python binding library, such as `pynvrtc` or `cuda-python`, to programmatically interact with NVRTC.
  Fix: Install a Python wrapper library (e.g., `pip install pynvrtc` or `pip install cuda-python`) to access NVRTC functionalities from Python. Attempting to `import nvidia_cuda_nvrtc_cu11` will result in an `ImportError`.
• gotcha Mismatch between the installed `nvidia-cuda-nvrtc-cu11` version (or the CUDA version it implies) and the NVIDIA GPU driver or other CUDA-dependent libraries (like PyTorch or CuPy) can lead to `ImportError: libnvrtc.so.<VERSION> not found`.
  Fix: Ensure your NVIDIA GPU driver is up-to-date and compatible with CUDA 11.x (check `nvidia-smi` output). If using other CUDA-enabled Python libraries, verify their compatibility with CUDA 11.x. Ensure `LD_LIBRARY_PATH` (Linux) or `PATH` (Windows) correctly includes the directory containing `libnvrtc.so` or `nvrtc64_*.dll` if not found automatically.
• gotcha NVRTC requires a compatible NVIDIA GPU and an installed NVIDIA display driver to function. While the NVRTC library itself can run on a system without a GPU, its utility for generating PTX for execution implies a GPU target.
  Fix: Ensure you have a CUDA-capable NVIDIA GPU and the latest compatible NVIDIA display drivers installed on your system. Verify driver version compatibility with your desired CUDA Toolkit version (e.g., CUDA 11.x).

Install

• pip install nvidia-cuda-nvrtc-cu11 Install via pip

Imports

• Program

  from pynvrtc.compiler import Program

  The `nvidia-cuda-nvrtc-cu11` package provides native shared libraries, not a direct Python API. Python bindings for NVRTC, such as the `pynvrtc` library, must be installed separately to interact with NVRTC functionality.
• NVRTCInterface

  from pynvrtc.interface import NVRTCInterface

  This provides a lower-level interface to the NVRTC API through the `pynvrtc` binding.

Quickstart

This quickstart demonstrates how to compile CUDA C++ source code into PTX using the `pynvrtc` Python binding, which relies on the native NVRTC libraries provided by `nvidia-cuda-nvrtc-cu11`. The `Program` class handles the compilation process. The resulting PTX code can then be loaded and executed on an NVIDIA GPU using lower-level CUDA driver APIs.

import os
from pynvrtc.compiler import Program, ProgramException

# Example CUDA C++ kernel source code
cuda_source_code = '''
extern "C" __global__
void add(int *a, int *b, int *c, int N)
{
    int idx = blockIdx.x * blockDim.x + threadIdx.x;
    if (idx < N)
    {
        c[idx] = a[idx] + b[idx];
    }
}
'''

try:
    # Compile the CUDA source code to PTX using the Program API
    # The nvidia-cuda-nvrtc-cu11 library is implicitly used by pynvrtc
    program = Program(cuda_source_code, 'add_kernel.cu')
    ptx_code = program.compile(['-arch=compute_60']) # Adjust arch for your GPU
    print("PTX code generated successfully. First 200 chars:\n", ptx_code[:200], '...')

    # In a real application, ptx_code would then be loaded and executed
    # using a CUDA driver API wrapper (e.g., from `cuda-python` or `pycuda`)
    # This part requires more setup (context, module, kernel launch) and is omitted for brevity.

except ProgramException as e:
    print(f"Error during NVRTC compilation: {e}")
except Exception as e:
    print(f"An unexpected error occurred: {e}")