Generate CUDA -gencode parameters for available GPU devices. Usage: chmod +x get_cuda_gencode.cpp && ./get_cuda_gencode.cpp

get_cuda_gencode.cpp

//usr/bin/env nvcc --run "$0" -o /tmp/get_cuda_gencode; rm /tmp/get_cuda_gencode; exit
#include <cuda_runtime_api.h>
#include <iostream>
#include <set>

int main(int argc, char *argv[]) {
    cudaDeviceProp prop;
    cudaError_t status;
    int device_count;
    status = cudaGetDeviceCount(&device_count);
    if (status != cudaSuccess) {
        std::cerr << "cudaGetDeviceCount() failed: " << cudaGetErrorString(status) << std::endl;
        return -1;
    }

    if (device_count == 0) {
        std::cerr << "No CUDA devices found" << std::endl;
        return -1;
    }

    std::set<int> versions;
    for (int device_index = 0; device_index < device_count; device_index++) {
        status = cudaGetDeviceProperties(&prop, device_index);
        if (status != cudaSuccess) {
            std::cerr << "cudaGetDeviceCount() for device " << device_index << " failed: "
                      << cudaGetErrorString(status) << std::endl;
            return -1;
        }
        versions.insert(prop.major * 10 + prop.minor);
    }

    bool first = true;
    for (auto &v : versions) {
        if (!first)
            std::cout << ' ';
        first = false;
        std::cout << "-gencode arch=compute_" << v << ",code=sm_" << v << " ";
    }
    std::cout << std::endl;
}