#include <iostream>
#include <fstream>

#define CUDA_SAFECALL(fnName, ...)                   \
do {                                                 \
    cudaError_t fnName##Error = fnName(__VA_ARGS__); \
    if (fnName##Error != cudaSuccess) {              \
        std::cerr << "failed to call " << #fnName << ": return code " << fnName##Error  << ", which means " << cudaGetErrorString(fnName##Error) << std::endl; \
        exit(fnName##Error);                         \
    }                                                \
} while (0)

#define CUDA_CHECKRESULT(func) \
do {                           \
    cudaError_t res = (func);  \
    if (res != cudaSuccess) {  \
        std::cerr << "failed function call " << #func << ": return code " << res << ", which means " << cudaGetErrorString(res) << std::endl; \
        exit(res);             \
    }                          \
} while (0)

static void run(std::ostream&);

void run(std::ostream& printer) {
    int count;
    printer << "C++ Version: " << __cplusplus << std::endl;
    CUDA_CHECKRESULT(cudaGetDeviceCount(&count));
    printer << std::endl << "Number of CUDA devices: " << count << std::endl;
    printer << std::endl << "-------------------------Device Properties-------------------------" << std::endl;
    for (int dev = 0; dev < count; ++dev) {
        printer << std::endl << "**************************** GPU " << dev << " ****************************"
                << std::endl << std::endl;
        cudaDeviceProp prop{};
        CUDA_SAFECALL(cudaGetDeviceProperties, &prop, dev);
        printer << "Name:                  " << prop.name << std::endl;
        printer << "Global Memory:         " << ((double) prop.totalGlobalMem / 1024 / 1048576) << " Gb" << std::endl;
        printer << "Shared Memory/block:   " << ((double) prop.sharedMemPerBlock / 1024) << " Kb" << std::endl;
        printer << "Registers/block:       " << prop.regsPerBlock << std::endl;
        printer << "Warp Size:             " << prop.warpSize << std::endl;
        printer << "Max Threads/block:     " << prop.maxThreadsPerBlock << std::endl;
        printer << "Max Block Dimensions:  [ " << prop.maxThreadsDim[0] << " x " << prop.maxThreadsDim[1] << " x "
                << prop.maxThreadsDim[2] << " ]" << std::endl;
        printer << "Max Grid Dimensions:   [ " << prop.maxGridSize[0] << " x " << prop.maxGridSize[1] << " x "
                << prop.maxGridSize[2] << " ]" << std::endl;
        printer << "Constant Memory:       " << ((double) prop.totalConstMem / 1024) << " Kb" << std::endl;
        printer << "Compute Capability:    " << prop.major << "." << prop.minor << std::endl;
        printer << "Clock Rate:            " << ((double) prop.clockRate / 1000000) << " GHz" << std::endl;
    }
}

int main(int argc, char** argv) {
    if (argc >= 2) {
        std::ofstream filestream(argv[1]);
        if (!filestream) {
            std::cerr << "Unable to open file '" << argv[1] << "' for writing, exiting!" << std::endl;
            return -1;
        }
        run(filestream);
    } else {
        run(std::cout);
    }
}