libutils/DeviceMemory.cuh

   1 /*
   2         Copyright (C) 2012  fmaj7b5.info
   3
   4         This program is free software: you can redistribute it and/or modify
   5         it under the terms of the GNU General Public License as published by
   6         the Free Software Foundation, either version 2 of the License, or
   7         (at your option) any later version.
   8
   9         This program is distributed in the hope that it will be useful,
  10         but WITHOUT ANY WARRANTY; without even the implied warranty of
  11         MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12         GNU General Public License for more details.
  13
  14         You should have received a copy of the GNU General Public License
  15         along with this program.  If not, see <http://www.gnu.org/licenses/>.
  16 */
  17
  18 #ifndef FM7b5_DEVICE_MEMORY_CUH
  19 #define FM7b5_DEVICE_MEMORY_CUH
  20
  21 #include <stdexcept>
  22 #include <type_traits>
  23 #include <cuda.h>
  24
  25 #include "MDView.cuh"
  26
  27 namespace FM7b5
  28 {
  29         namespace memory
  30         {
  31                 //! Linear 1-D memory
  32                 template <class T>
  33                 class Linear
  34                 {
  35                 public:
  36                         typedef T value_type;
  37                         typedef typename std::add_pointer<T>::type pointer_type;
  38                         typedef typename std::add_const<pointer_type>::type const_pointer_type;
  39                         typedef typename std::add_reference<T>::type reference_type;
  40
  41                         Linear(const size_t count = 0)
  42                                 : m_count(count)
  43                         {
  44                                 m_extent = sizeof(value_type);
  45                                 md_data = static_cast<pointer_type>(allocate(m_extent * count));
  46                         }
  47
  48                         ~Linear()
  49                         {
  50                                 deallocate(md_data);
  51                         }
  52
  53                         void* allocate(const size_t byte)
  54                         {
  55                                 void* p(nullptr);
  56                                 cudaError_t status;
  57
  58                                 status = cudaMalloc(&p, byte);
  59                                 if (status != cudaSuccess) {
  60                                         throw std::bad_alloc(cudaGetErrorString(status));
  61                                 }
  62
  63                                 return p;
  64                         }
  65
  66                         void deallocate(void* p)
  67                         {
  68                                 cudaFree(p);
  69                         }
  70
  71                         void copy_from(const void* const src, const size_t byte)
  72                         {
  73                                 cudaMemcpy(md_data, src, byte, cudaMemcpyHostToDevice);
  74                         }
  75
  76                         void copy_to(void* const dst, const size_t byte)
  77                         {
  78                                 cudaMemcpy(dst, md_data, byte, cudaMemcpyDeviceToHost);
  79                         }
  80
  81                         size_t size() const { return m_extent * m_count; }
  82                         size_t count() const { return m_count; }
  83                         pointer_type data() const { return md_data; }
  84
  85                         MDView<value_type, 1> ref() const
  86                         {
  87                                 return MDView<value_type, 1>(md_data, &m_extent);
  88                         }
  89
  90                 protected:
  91                         size_t m_extent;
  92                         size_t m_count;
  93                         pointer_type md_data;
  94                 };
  95
  96                 //! Linear 2-D memory
  97                 template <class T>
  98                 class LinearPitch
  99                 {
 100                 public:
 101                         typedef T value_type;
 102                         typedef typename std::add_pointer<T>::type pointer_type;
 103                         typedef typename std::add_const<pointer_type>::type const_pointer_type;
 104                         typedef typename std::add_reference<T>::type reference_type;
 105
 106                         LinearPitch(const size_t width, const size_t height)
 107                                 : m_width(width), m_height(height)
 108                         {
 109                                 md_data = static_cast<pointer_type>(allocate(width, height, &m_extent[0]));
 110                                 m_extent[1] = sizeof(value_type);
 111                         }
 112
 113                         ~LinearPitch()
 114                         {
 115                                 deallocate(md_data);
 116                         }
 117
 118                         void* allocate(const size_t width, const size_t height, size_t* pitch)
 119                         {
 120                                 void* p(nullptr);
 121                                 cudaError_t status;
 122
 123                                 status = cudaMallocPitch(&p, pitch, sizeof(value_type) * width, height);
 124                                 if (status != cudaSuccess) {
 125                                         throw std::bad_alloc(cudaGetErrorString(status));
 126                                 }
 127
 128                                 return p;
 129                         }
 130
 131                         void deallocate(void* p)
 132                         {
 133                                 cudaFree(p);
 134                         }
 135
 136                         void copy_from(const void* const src, const size_t width, const size_t height, size_t spitch = 0)
 137                         {
 138                                 if (spitch < 1) {
 139                                         spitch = width;
 140                                 }
 141                                 cudaMemcpy2D(md_data, pitch(), src, spitch, width, height, cudaMemcpyHostToDevice);
 142                         }
 143
 144                         void copy_to(void* const dst, const size_t width, const size_t height, size_t dpitch = 0)
 145                         {
 146                                 if (dpitch < 1) {
 147                                         dpitch = width;
 148                                 }
 149                                 cudaMemcpy2D(dst, dpitch, md_data, pitch(), width, height, cudaMemcpyDeviceToHost);
 150                         }
 151
 152                         size_t size() const { return m_extent[1] * m_height; }
 153                         size_t width() const { return m_width; }
 154                         size_t height() const { return m_height; }
 155                         size_t pitch() const { return m_extent[0]; }
 156                         pointer_type data() const { return md_data; }
 157
 158                         MDView<value_type, 2> ref() const
 159                         {
 160                                 return MDView<value_type, 2>(md_data, m_extent);
 161                         }
 162
 163                 protected:
 164                         size_t m_width;
 165                         size_t m_height;
 166                         size_t m_extent[2];
 167                         pointer_type md_data;
 168                 };
 169         }
 170 }
 171
 172 #endif /* FM7b5_DEVICE_MEMORY_CUH */