ipc.hpp

Go to the documentation of this file.

#pragma once
#ifndef CUDA_API_WRAPPERS_IPC_HPP_
#define CUDA_API_WRAPPERS_IPC_HPP_

#include "context.hpp"
#include "types.hpp"
#include "error.hpp"

#include <string>

namespace cuda {

class device_t;
class event_t;

namespace memory {

class pool_t;

namespace ipc {

using ptr_handle_t = CUipcMemHandle;

class imported_ptr_t;
imported_ptr_t wrap(void * ptr, bool owning) noexcept;

namespace detail_ {

inline void* import(const ptr_handle_t& handle)
{
    CUdeviceptr device_ptr;
    auto status = cuIpcOpenMemHandle(&device_ptr, handle, CU_IPC_MEM_LAZY_ENABLE_PEER_ACCESS);
    throw_if_error_lazy(status, "Failed obtaining a device pointer from an IPC memory handle");
    return memory::as_pointer(device_ptr);
}

inline status_t unmap_nothrow(void* ipc_mapped_ptr) noexcept
{
    return cuIpcCloseMemHandle(device::address(ipc_mapped_ptr));
}

inline void unmap(void* ipc_mapped_ptr)
{
    auto status = unmap_nothrow(ipc_mapped_ptr);
    throw_if_error_lazy(status, "Failed unmapping IPC memory mapped to " + cuda::detail_::ptr_as_hex(ipc_mapped_ptr));
}

} // namespace detail_

inline ptr_handle_t export_(void* device_ptr)
{
    ptr_handle_t handle;
    auto status = cuIpcGetMemHandle(&handle, device::address(device_ptr));
    throw_if_error_lazy(status, "Failed producing an IPC memory handle for device pointer "
        + cuda::detail_::ptr_as_hex(device_ptr));
    return handle;
}

class imported_ptr_t {
protected: // constructors & destructor
    imported_ptr_t(void* ptr, bool owning) : ptr_(ptr), owning_(owning)
    {
        if (ptr_ == nullptr) {
            throw ::std::logic_error("IPC memory handle yielded a null pointer");
        }
    }

public: // constructors & destructors
    friend imported_ptr_t wrap(void * ptr, bool owning) noexcept;

    ~imported_ptr_t() DESTRUCTOR_EXCEPTION_SPEC
    {
        if (not owning_) { return; }
#if THROW_IN_DESTRUCTORS
        detail_::unmap(ptr_);
#else
        detail_::unmap_nothrow(ptr_);
#endif
    }

public: // operators

    imported_ptr_t(const imported_ptr_t& other) = delete;
    imported_ptr_t& operator=(const imported_ptr_t& other) = delete;
    imported_ptr_t& operator=(imported_ptr_t&& other) noexcept
    {
        ::std::swap(ptr_, other.ptr_);
        ::std::swap(owning_, other.owning_);
        return *this;
    }
    imported_ptr_t(imported_ptr_t&& other) noexcept = default;

public: // getters

    template <typename T = void>
    T* get() const noexcept
    {
        // If you're wondering why this cast is necessary - some IDEs/compilers
        // have the notion that if the method is const, `ptr_` is a const void* within it
        return static_cast<T*>(const_cast<void*>(ptr_));
    }

    bool is_owning() const noexcept { return owning_; }

protected: // data members
    void*  ptr_;
    bool   owning_;
}; // class imported_ptr_t

inline imported_ptr_t wrap(void * ptr, bool owning) noexcept
{
    return imported_ptr_t(ptr, owning);
}

inline imported_ptr_t import(const ptr_handle_t& ptr_handle)
{
    auto raw_ptr = detail_::import(ptr_handle);
    return wrap(raw_ptr, do_take_ownership);
}

} // namespace ipc

#if CUDA_VERSION >= 11020
namespace pool {

namespace ipc {

using handle_t = void *;

template <shared_handle_kind_t Kind>
shared_handle_t<Kind> export_(const pool_t& pool);

namespace detail_ {

template <shared_handle_kind_t Kind>
pool::handle_t import(const shared_handle_t<Kind>& shared_pool_handle)
{
    memory::pool::handle_t result;
    static constexpr const unsigned long long flags { 0 };
    void * ptr_to_handle = static_cast<void*>(const_cast<shared_handle_t<Kind>*>(&shared_pool_handle));
    auto status = cuMemPoolImportFromShareableHandle(
        &result, ptr_to_handle, static_cast<CUmemAllocationHandleType>(Kind), flags);
    throw_if_error_lazy(status, "Importing an IPC-shared memory pool handle");
    return result;
}

} // namespace detail_

template <shared_handle_kind_t Kind>
pool_t import(const device_t& device, const shared_handle_t<Kind>& shared_pool_handle);

inline ptr_handle_t export_ptr(void* pool_allocated) {
    ptr_handle_t handle;
    auto status = cuMemPoolExportPointer(&handle, device::address(pool_allocated));
    throw_if_error_lazy(status,
        "Failed producing an IPC handle for memory-pool-allocated pointer "
        + cuda::detail_::ptr_as_hex(pool_allocated));
    return handle;
}

namespace detail_ {

inline void* import_ptr(const pool::handle_t pool_handle, const ptr_handle_t& handle)
{
    CUdeviceptr imported;
    auto status = cuMemPoolImportPointer(&imported, pool_handle, const_cast<ptr_handle_t*>(&handle));
    throw_if_error_lazy(status, "Failed importing an IPC-exported a pool-allocated pointer");
    return as_pointer(imported);
}

} // namespace detail_

class imported_ptr_t;

imported_ptr_t import_ptr(const pool_t& shared_pool, const ptr_handle_t& ptr_handle);
imported_ptr_t import_ptr(const pool_t& shared_pool, const ptr_handle_t& ptr_handle, const stream_t& freeing_stream);

} // namespace ipc

} // namespace pool
#endif // CUDA_VERSION >= 11020

} // namespace memory

namespace event {
namespace ipc {

using handle_t = CUipcEventHandle;

namespace detail_ {

inline handle_t export_(event::handle_t event_handle)
{
    handle_t ipc_handle;
    auto status = cuIpcGetEventHandle(&ipc_handle, event_handle);
    throw_if_error_lazy(status, "Failed obtaining an IPC event handle for " +
        event::detail_::identify(event_handle));
    return ipc_handle;
}

inline event::handle_t import(const handle_t& handle)
{
    event::handle_t event_handle;
    auto status = cuIpcOpenEventHandle(&event_handle, handle);
    throw_if_error_lazy(status, "Failed obtaining an event handle from an IPC event handle");
    return event_handle;
}

} // namespace detail_

inline handle_t export_(const event_t& event);


inline event_t import(const device_t& device, const handle_t& event_ipc_handle);

inline event_t import(const context_t& context, const handle_t& event_ipc_handle);

} // namespace ipc
} // namespace event
} // namespace cuda

#endif // CUDA_API_WRAPPERS_IPC_HPP_