module.hpp

Go to the documentation of this file.

#pragma once
#ifndef CUDA_API_WRAPPERS_MODULE_HPP_
#define CUDA_API_WRAPPERS_MODULE_HPP_

#include "context.hpp"
#include "primary_context.hpp"
#include "kernel.hpp"
#include "memory.hpp"
#include "link_options.hpp"

#if __cplusplus >= 201703L
#include <filesystem>
#endif

namespace cuda {

class device_t;
class context_t;
class module_t;
class kernel_t;

namespace module {

// The CUDA driver's raw handle for modules
using handle_t = CUmodule;

namespace detail_ {

inline module_t wrap(
    device::id_t            device_id,
    context::handle_t       context_handle,
    handle_t                handle,
    bool                    take_ownership = false,
    bool                    holds_primary_context_refcount_unit = false) noexcept;

inline ::std::string identify(const module::handle_t &handle)
{
    return ::std::string("module ") + cuda::detail_::ptr_as_hex(handle);
}

inline ::std::string identify(const module::handle_t &handle, context::handle_t context_handle)
{
    return identify(handle) + " in " + context::detail_::identify(context_handle);
}

inline ::std::string identify(const module::handle_t &handle, context::handle_t context_handle, device::id_t device_id)
{
    return identify(handle) + " in " + context::detail_::identify(context_handle, device_id);
}

::std::string identify(const module_t &module);

inline void unload(handle_t handle, context::handle_t context_handle, device::id_t device_id)
{
    CAW_SET_SCOPE_CONTEXT(context_handle);
    auto status = cuModuleUnload(handle);
    throw_if_error_lazy(status, "Failed unloading " + identify(handle, context_handle, device_id));
}

#if CUDA_VERSION >= 12040
inline unique_span<kernel::handle_t> get_kernel_handles(handle_t module_handle, size_t num_kernels)
{
    auto result = make_unique_span<kernel::handle_t>(num_kernels);
    auto status = cuModuleEnumerateFunctions(result.data(), (unsigned int) num_kernels, module_handle);
    throw_if_error_lazy(status, "Failed enumerating the kernels in " + module::detail_::identify(module_handle));
    return result;
}
#endif

} // namespace detail_

template <typename Locus, typename ContiguousContainer,
    cuda::detail_::enable_if_t<cuda::detail_::is_kinda_like_contiguous_container<ContiguousContainer>::value, bool> = true >
module_t create(
    Locus&&                 locus,
    ContiguousContainer     module_data,
    const link::options_t&  link_options);

template <typename Locus, typename ContiguousContainer,
    cuda::detail_::enable_if_t<cuda::detail_::is_kinda_like_contiguous_container<ContiguousContainer>::value, bool> = true >
module_t create(
    Locus&&              locus,
    ContiguousContainer  module_data);

#if CUDA_VERSION >= 12030
using loading_mode_t = CUmoduleLoadingMode;

inline loading_mode_t loading_mode() {
    loading_mode_t result;
    auto status = cuModuleGetLoadingMode(&result);
    throw_if_error_lazy(status, "Failed obtaining CUDA module loading mode");
    return result;
}
#endif

} // namespace module

class module_t {

public: // getters
    module::handle_t handle() const { return handle_; }
    context::handle_t context_handle() const { return context_handle_; }
    device::id_t device_id() const { return device_id_; }

    context_t context() const;

    device_t device() const;

    cuda::kernel_t get_kernel(const char* name) const;

    cuda::kernel_t get_kernel(const ::std::string& name) const
    {
        return get_kernel(name.c_str());
    }

    memory::region_t get_global_region(const char* name) const
    {
        CUdeviceptr dptr;
        size_t size;
        auto result = cuModuleGetGlobal(&dptr, &size, handle_, name);
        throw_if_error_lazy(result, "Obtaining the address and size of a named global object");
        return { memory::as_pointer(dptr), size };
    }

#if CUDA_VERSION >= 12040
    size_t get_num_kernels() const
    {
        unsigned result;
        auto status = cuModuleGetFunctionCount(&result, handle_);
        throw_if_error_lazy(status, "Failed determining function count for " + module::detail_::identify(*this));
        return result;
    }

    unique_span<kernel_t> get_kernels() const
    {
        auto num_kernels = get_num_kernels();
        // It's ok if the number is 0!
        auto handles = module::detail_::get_kernel_handles(handle_, num_kernels);
        auto gen = [&](size_t i) { return kernel::wrap(device_id_, context_handle_, handles[i]); };
        return generate_unique_span<kernel_t>(handles.size(), gen);
    }
#endif // CUDA_VERSION >= 12040

    // TODO: Implement a surface reference and texture reference class rather than these raw pointers.

#if CUDA_VERSION < 12000
    CUsurfref get_surface(const char* name) const;

    CUtexref get_texture_reference(const char* name) const;
#endif

protected: // constructors

    module_t(
        device::id_t device_id,
        context::handle_t context,
        module::handle_t handle,
        bool owning,
        bool holds_primary_context_refcount_unit)
    noexcept
        : device_id_(device_id), context_handle_(context), handle_(handle), owning_(owning),
          holds_pc_refcount_unit_(holds_primary_context_refcount_unit)
    { }

public: // friendship

    friend module_t module::detail_::wrap(device::id_t, context::handle_t, module::handle_t, bool, bool) noexcept;

public: // constructors and destructor

    module_t(const module_t&) = delete;

    module_t(module_t&& other) noexcept :
        module_t(
            other.device_id_,
            other.context_handle_,
            other.handle_,
            other.owning_,
            other.holds_pc_refcount_unit_)
    {
        other.owning_ = false;
        other.holds_pc_refcount_unit_ = false;
    };

    // Note: It is up to the user of this class to ensure that it is unloaded _before_ the context
    // in which it was created; and one needs to be particularly careful about this point w.r.t.
    // primary contexts
    ~module_t() DESTRUCTOR_EXCEPTION_SPEC
    {
        if (owning_) {
#ifdef THROW_IN_DESTRUCTORS
            try
#endif
            {
                module::detail_::unload(handle_, context_handle_, device_id_);
            }
#ifdef THROW_IN_DESTRUCTORS
            catch (...) {}
#endif
        }
        if (holds_pc_refcount_unit_) {
            device::primary_context::detail_::decrease_refcount_in_dtor(device_id_);
        }
    }

public: // operators

    module_t& operator=(const module_t&) = delete;
    module_t& operator=(module_t&& other) noexcept
    {
        ::std::swap(device_id_, other.device_id_);
        ::std::swap(context_handle_, other.context_handle_);
        ::std::swap(handle_, other.handle_);
        ::std::swap(owning_, other.owning_);
        ::std::swap(holds_pc_refcount_unit_, holds_pc_refcount_unit_);
        return *this;
    }

protected: // data members
    device::id_t       device_id_;
    context::handle_t  context_handle_;
    module::handle_t   handle_;
    bool               owning_;
        // this field is mutable only for enabling move construction; other
        // than in that case it must not be altered
    bool holds_pc_refcount_unit_;
        // When context_handle_ is the handle of a primary context, this module
        // may be "keeping that context alive" through the refcount - in which
        // case it must release its refcount unit on destruction
};

namespace module {

namespace detail_ {

inline module_t load_from_file_in_current_context(
    device::id_t            current_context_device_id,
    context::handle_t       current_context_handle,
    const char *            path,
    bool                    holds_primary_context_refcount_unit = false)
{
    handle_t new_module_handle;
    auto status = cuModuleLoad(&new_module_handle, path);
    throw_if_error_lazy(status, ::std::string("Failed loading a module from file ") + path);
    bool do_take_ownership{true};
    return wrap(
        current_context_device_id,
        current_context_handle,
        new_module_handle,
        do_take_ownership,
        holds_primary_context_refcount_unit);
}

} // namespace detail_


inline module_t load_from_file(
    const context_t&        context,
    const char*             path)
{
    CAW_SET_SCOPE_CONTEXT(context.handle());
    return detail_::load_from_file_in_current_context(context.device_id(), context.handle(), path);
}

inline module_t load_from_file(
    const context_t&        context,
    const ::std::string&    path)
{
    return load_from_file(context, path.c_str());
}

module_t load_from_file(
    const device_t&         device,
    const char*             path);

inline module_t load_from_file(
    const device_t&         device,
    const ::std::string&    path)
{
    return load_from_file(device, path.c_str());
}

module_t load_from_file(const char* path);

inline module_t load_from_file(const ::std::string& path)
{
    return load_from_file(path.c_str());
}

#if __cplusplus >= 201703L
inline module_t load_from_file(
    const device_t&                 device,
    const ::std::filesystem::path&  path)
{
    return load_from_file(device, path.c_str());
}

inline module_t load_from_file(
    const ::std::filesystem::path&  path)
{
    return load_from_file(device::current::get(), path);
}
#endif

namespace detail_ {

inline module_t wrap(
    device::id_t            device_id,
    context::handle_t       context_handle,
    handle_t                module_handle,
    bool                    take_ownership,
    bool                    hold_pc_refcount_unit
) noexcept
{
    return module_t{device_id, context_handle, module_handle, take_ownership, hold_pc_refcount_unit};
}

module_t create(const context_t& context, const void* module_data, const link::options_t& link_options);

module_t create(const context_t& context, const void* module_data);

} // namespace detail_

// TODO: Use an optional to reduce the number of functions here... when the
// library starts requiring C++14.

namespace detail_ {

inline ::std::string identify(const module_t& module)
{
    return identify(module.handle(), module.context_handle(), module.device_id());
}

inline context_t get_context_for(const context_t& locus) { return locus; }
inline device::primary_context_t get_context_for(const device_t& locus);

} // namespace detail_

template <typename Locus, typename ContiguousContainer,
    cuda::detail_::enable_if_t<cuda::detail_::is_kinda_like_contiguous_container<ContiguousContainer>::value, bool>>
module_t create(
    Locus&&             locus,
    ContiguousContainer module_data)
{
    auto context = detail_::get_context_for(locus);
    return detail_::create(context, module_data.data());
}

// Note: The following may create the primary context of a device!
template <typename Locus, typename ContiguousContainer,
    cuda::detail_::enable_if_t<cuda::detail_::is_kinda_like_contiguous_container<ContiguousContainer>::value, bool>>
module_t create(
    Locus&&                 locus,
    ContiguousContainer     module_data,
    const link::options_t&  link_options)
{
    auto context = detail_::get_context_for(locus);
    return detail_::create(context, module_data.data(), link_options);
}

} // namespace module

} // namespace cuda

#endif // CUDA_API_WRAPPERS_MODULE_HPP_