-
Notifications
You must be signed in to change notification settings - Fork 3
/
host_noncached_unique_ptr.h
65 lines (55 loc) · 2.54 KB
/
host_noncached_unique_ptr.h
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
#ifndef HeterogeneousCore_CUDAUtilities_interface_host_noncached_unique_ptr_h
#define HeterogeneousCore_CUDAUtilities_interface_host_noncached_unique_ptr_h
#include <memory>
#include <cuda/api_wrappers.h>
#include <cuda_runtime.h>
namespace cudautils {
namespace host {
namespace noncached {
namespace impl {
// Additional layer of types to distinguish from host::unique_ptr
class HostDeleter {
public:
void operator()(void *ptr) {
cuda::throw_if_error(cudaFreeHost(ptr));
}
};
}
template <typename T>
using unique_ptr = std::unique_ptr<T, impl::HostDeleter>;
namespace impl {
template <typename T>
struct make_host_unique_selector { using non_array = cudautils::host::noncached::unique_ptr<T>; };
template <typename T>
struct make_host_unique_selector<T[]> { using unbounded_array = cudautils::host::noncached::unique_ptr<T[]>; };
template <typename T, size_t N>
struct make_host_unique_selector<T[N]> { struct bounded_array {}; };
}
}
}
/**
* The difference wrt. make_host_unique is that these
* do not cache, so they should not be called per-event.
*/
template <typename T>
typename host::noncached::impl::make_host_unique_selector<T>::non_array
make_host_noncached_unique(unsigned int flags = cudaHostAllocDefault) {
static_assert(std::is_trivially_constructible<T>::value, "Allocating with non-trivial constructor on the pinned host memory is not supported");
void *mem;
cuda::throw_if_error(cudaHostAlloc(&mem, sizeof(T), flags));
return typename cudautils::host::noncached::impl::make_host_unique_selector<T>::non_array(reinterpret_cast<T *>(mem));
}
template <typename T>
typename host::noncached::impl::make_host_unique_selector<T>::unbounded_array
make_host_noncached_unique(size_t n, unsigned int flags = cudaHostAllocDefault) {
using element_type = typename std::remove_extent<T>::type;
static_assert(std::is_trivially_constructible<element_type>::value, "Allocating with non-trivial constructor on the pinned host memory is not supported");
void *mem;
cuda::throw_if_error(cudaHostAlloc(&mem, n*sizeof(element_type), flags));
return typename cudautils::host::noncached::impl::make_host_unique_selector<T>::unbounded_array(reinterpret_cast<element_type *>(mem));
}
template <typename T, typename ...Args>
typename cudautils::host::noncached::impl::make_host_unique_selector<T>::bounded_array
make_host_noncached_unique(Args&&...) = delete;
}
#endif