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
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
|
/******************************************************************************
* Copyright (c) 2011-2018, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are not permitted.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
* WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
* DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
* DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
* LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
* ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
* SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*
******************************************************************************/
#pragma once
/**
* \file
* \brief C++ interface to CUDA device memory management functions.
*/
#include <memory>
#include <cutlass/util/debug.h>
#include <cutlass/util/platform.h>
#include <exceptions.h>
namespace cutlass {
namespace device_memory {
/******************************************************************************
* Allocation lifetime
******************************************************************************/
/// Allocate a buffer of \p count elements of type \p T on the current CUDA device
template <typename T>
T* allocate(size_t count = 1) {
T* ptr = 0;
size_t bytes = sizeof(T) * count;
cudaError_t cuda_error = CUDA_PERROR(cudaMalloc((void**)&ptr, bytes));
if (cuda_error != cudaSuccess) {
throw cuda_exception("Failed to allocate memory", cuda_error);
}
return ptr;
}
/// Free the buffer pointed to by \p ptr
template <typename T>
void free(T* ptr) {
if (ptr) {
cudaError_t cuda_error = CUDA_PERROR(cudaFree(ptr));
if (cuda_error != cudaSuccess) {
throw cuda_exception("Failed to free device memory", cuda_error);
}
}
}
/******************************************************************************
* Data movement
******************************************************************************/
template <typename T>
void copy(T* dst, T const* src, size_t count, cudaMemcpyKind kind) {
size_t bytes = count * sizeof(T);
cudaError_t cuda_error = CUDA_PERROR(cudaMemcpy(dst, src, bytes, kind));
if (cuda_error != cudaSuccess) {
throw cuda_exception("cudaMemcpy() failed", cuda_error);
}
}
template <typename T>
void copy_to_device(T* dst, T const* src, size_t count = 1) {
copy(dst, src, count, cudaMemcpyHostToDevice);
}
template <typename T>
void copy_to_host(T* dst, T const* src, size_t count = 1) {
copy(dst, src, count, cudaMemcpyDeviceToHost);
}
template <typename T>
void copy_device_to_device(T* dst, T const* src, size_t count = 1) {
copy(dst, src, count, cudaMemcpyDeviceToDevice);
}
/// Copies elements from device memory to host-side range
template <typename OutputIterator, typename T>
void insert_to_host(OutputIterator begin, OutputIterator end, T const* device_begin) {
size_t elements = end - begin;
copy_to_host(&*begin, device_begin, elements);
}
/// Copies elements to device memory from host-side range
template <typename T, typename InputIterator>
void insert_to_device(T* device_begin, InputIterator begin, InputIterator end) {
size_t elements = end - begin;
copy_to_device(device_begin, &*begin, elements);
}
/******************************************************************************
* "Smart" device memory allocation
******************************************************************************/
/// Device allocation abstraction that tracks size and capacity
template <typename T>
struct allocation {
/// Delete functor for CUDA device memory
struct deleter {
void operator()(T* ptr) {
cudaError_t cuda_error = CUDA_PERROR(cudaFree(ptr));
if (cuda_error != cudaSuccess) {
// noexcept
// throw cuda_exception("cudaFree() failed", cuda_error);
return;
}
}
};
/// Number of elements of T allocated on the current CUDA device
size_t capacity;
/// Smart pointer
platform::unique_ptr<T, deleter> smart_ptr;
//
//
//
/// Constructor: allocates no memory
allocation() : capacity(0) {}
/// Constructor: allocates \p capacity elements on the current CUDA device
allocation(size_t _capacity) : smart_ptr(allocate<T>(_capacity)), capacity(_capacity) {}
/// Destructor
~allocation() { reset(); }
/// Returns a pointer to the managed object
T* get() const { return smart_ptr.get(); }
/// Releases the ownership of the managed object (without deleting) and resets capacity to zero
T* release() {
capacity = 0;
return smart_ptr.release();
}
/// Deletes the managed object and resets capacity to zero
void reset() {
capacity = 0;
smart_ptr.reset();
}
/// Deletes managed object, if owned, and replaces its reference with a given pointer and capacity
void reset(T* _ptr, size_t _capacity) {
smart_ptr.reset(_ptr);
capacity = _capacity;
}
/// Returns a pointer to the object owned by *this
T* operator->() const { return smart_ptr.get(); }
/// Returns the deleter object which would be used for destruction of the managed object.
deleter& get_deleter() { return smart_ptr.get_deleter(); }
/// Returns the deleter object which would be used for destruction of the managed object (const)
const deleter& get_deleter() const { return smart_ptr.get_deleter(); }
};
} // namespace device_memory
} // namespace cutlass
|
