Program Listing for File cudaKernelProcesor.hpp

↰ Return to documentation for file (librapid/include/librapid/cuda/cudaKernelProcesor.hpp)

#ifndef LIBRAPID_CUDA_HEADER_LOADER_HPP
#define LIBRAPID_CUDA_HEADER_LOADER_HPP

#if defined(LIBRAPID_HAS_CUDA)

namespace librapid::cuda {
    const std::string &loadKernel(const std::string &path, bool relative = true);

    jitify::Program generateCudaProgram(const std::string &kernel);

    template<typename... Templates, typename... Args>
    void runKernelString(const std::string &kernel, const std::string &kernelName, size_t elements,
                         Args... arguments) {
        jitify::Program program = generateCudaProgram(kernel);
        unsigned int threadsPerBlock, blocksPerGrid;

        // Use 1 to 512 threads per block
        if (elements < 512) {
            threadsPerBlock = static_cast<unsigned int>(elements);
            blocksPerGrid   = 1;
        } else {
            threadsPerBlock = 512;
            blocksPerGrid   = static_cast<unsigned int>(
              ceil(static_cast<double>(elements) / static_cast<double>(threadsPerBlock)));
        }

        dim3 grid(blocksPerGrid);
        dim3 block(threadsPerBlock);

#    if defined(LIBRAPID_DEBUG)
        try {
#    endif // LIBRAPID_DEBUG

            jitifyCall(program.kernel(kernelName)
                         .instantiate(jitify::reflection::Type<Templates>()...)
                         .configure(grid, block, 0, global::cudaStream)
                         .launch(arguments...));

#    if defined(LIBRAPID_DEBUG)
        } catch (const std::exception &e) {
            auto format = fmt::emphasis::bold | fmt::fg(fmt::color::red);
            fmt::print(format, "Error            : {}\n", e.what());
            fmt::print(format, "Kernel name      : {}\n", kernelName);
            fmt::print(format, "Elements         : {}\n", elements);
            fmt::print(format, "Threads per block: {}\n", threadsPerBlock);
            fmt::print(format, "Blocks per grid  : {}\n", blocksPerGrid);
            fmt::print(format, "Arguments        : {}\n", sizeof...(Args));

            // Print all arguments
            auto printer = [](auto x, auto format) {
                fmt::print(fmt::fg(fmt::color::purple), "\nArgument:\n");

                // True if x can be printed with fmt
                constexpr bool isPrintable = fmt::is_formattable<decltype(x)>::value;

                if constexpr (isPrintable) {
                    fmt::print(format, "\tValue: {}\n", x);
                } else {
                    fmt::print(format, "\tValue: [ CANNOT PRINT ]\n");
                }
                fmt::print(format, "\tType : {}\n", typeid(x).name());
            };

            (printer(arguments, fmt::emphasis::bold | fmt::fg(fmt::color::dark_orange)), ...);
            (printer(typeid(Templates).name(), fmt::emphasis::bold | fmt::fg(fmt::color::plum)),
             ...);

            throw;
        }
#    endif // LIBRAPID_DEBUG
    }

    template<typename... Templates, typename... Args>
    void runKernel(const std::string &name, const std::string &kernelName, size_t elements,
                   Args... arguments) {
        runKernelString<Templates...>(loadKernel(name), kernelName, elements, arguments...);
    }
} // namespace librapid::cuda

#endif // LIBRAPID_HAS_CUDA

#endif // LIBRAPID_CUDA_HEADER_LOADER_HPP