如何解决GCC与CUDA版本不匹配问题，获取适配CUDA的正确GCC版本

Alright, let's tackle this GCC and CUDA version mismatch issue step by step—this is a super common problem when working with CUDA, so don't worry, we've got this.

Step 1: Confirm Your Current Versions #

First, let's get clear on what versions you're running. Open your terminal and run these commands:

• Check GCC version: gcc --version
• Check CUDA version: nvcc --version

Once you have these numbers, you need to match them to NVIDIA's official supported combinations. Here's a quick, reliable cheat sheet for common releases:

• CUDA 11.x: Supports GCC 5.x to 10.x (CUDA 11.4+ adds partial support for GCC 11)
• CUDA 12.x: Supports GCC 7.x to 12.x
• Older CUDA 10.x: Stick to GCC 4.x to 8.x

Step 2: Pick Your Fix Strategy #

You have two solid paths here—either adjust your GCC version to fit your CUDA install, or adjust CUDA to match your GCC. Let's break down both options.

Option 1: Adjust GCC Version (Upgrade/Downgrade) #

Most Linux distros let you run multiple GCC versions side by side. Here's how to do it on Debian/Ubuntu-based systems:

1. Install your desired GCC version (example: GCC 10 for CUDA 11.7):

   sudo apt update
   sudo apt install gcc-10 g++-10
   

2. Set it as the default compiler:
   • Temporary (only for your current terminal session):

     export CC=/usr/bin/gcc-10
     export CXX=/usr/bin/g++-10
     

   • Permanent (add to your shell config like ~/.bashrc or ~/.zshrc):

     echo "export CC=/usr/bin/gcc-10" >> ~/.bashrc
     echo "export CXX=/usr/bin/g++-10" >> ~/.bashrc
     source ~/.bashrc
     

For RHEL/CentOS/Fedora, use dnf instead of apt, with similar package names (e.g., gcc10).

Option 2: Adjust CUDA Version (Upgrade/Downgrade) #

If you'd rather keep your current GCC, install a CUDA release that supports it:

1. Uninstall your current CUDA toolkit (back up important projects first!):

   sudo apt-get --purge remove "*cuda*" "*cublas*" "*cufft*" "*cufile*" "*curand*" \
   "*cusolver*" "*cusparse*" "*npp*" "*nvjpeg*" "nsight*"
   sudo rm -rf /usr/local/cuda*
   

2. Download the matching CUDA version from NVIDIA's official repository (search for "CUDA toolkit [your desired version] Linux" and follow distro-specific steps).
3. Update your environment variables to point to the new install:

   echo "export PATH=/usr/local/cuda-[version]/bin:$PATH" >> ~/.bashrc
   echo "export LD_LIBRARY_PATH=/usr/local/cuda-[version]/lib64:$LD_LIBRARY_PATH" >> ~/.bashrc
   source ~/.bashrc
   

   Replace [version] with your installed CUDA number (e.g., 12.2).

Step 3: Verify the Fix #

Once you've made changes, test with a simple CUDA program to confirm everything works:

1. Create a test.cu file:

   #include <stdio.h>
   
   __global__ void hello() {
       printf("Hello from CUDA!\n");
   }
   
   int main() {
       hello<<<1,1>>>();
       cudaDeviceSynchronize();
       return 0;
   }
   

2. Compile it:

   nvcc test.cu -o test
   

3. Run the executable:

   ./test
   

If you see "Hello from CUDA!", you're good to go!

内容的提问来源于stack exchange，提问作者papatya2