Combining memory-mapped I/O and CUDA mapped memory

cudamap.cc

#include <sys/mman.h>
#include <fcntl.h>
#include <unistd.h>
#include <cuda_runtime.h>

#include <cerrno>
#include <cstring>
#include <memory>
#include <stdexcept>

struct cuda_mmap_deleter
{
	std::size_t filesize;

	cuda_mmap_deleter (std::size_t length)
		: filesize (length)
	{
	}

	void operator () (char* ptr)
	{
		cudaHostUnregister (ptr);
		munmap (static_cast <void*> (ptr), filesize);
	}
};

std::unique_ptr <char [], cuda_mmap_deleter> map_file (std::string filename,
                                                       std::size_t filesize)
{
	int fd = open (filename.c_str (), O_CREAT | O_RDWR, S_IRUSR | S_IWUSR);
	if (fd == -1)
		throw std::runtime_error (std::strerror (errno) + std::string (" (open)"));

	ftruncate (fd, filesize);
	auto fmem = mmap (nullptr, filesize, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
	close (fd);
	if (fmem == MAP_FAILED)
		throw std::runtime_error (std::strerror (errno) + std::string (" (mmap)"));

	decltype (map_file ("", 0))
	mapped_file (static_cast <char*> (fmem), filesize);

	auto cuda_retval = cudaHostRegister (fmem, filesize, cudaHostRegisterMapped);
	if (cuda_retval != cudaSuccess)
		throw std::runtime_error (cudaGetErrorString (cuda_retval) + std::string (" (cudaHostRegister)"));

	return mapped_file;
}