File size: 8,127 Bytes
9375c9a |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 |
// Copyright (C) 2015 Davis E. King ([email protected])
// License: Boost Software License See LICENSE.txt for the full license.
#ifndef DLIB_GPU_DaTA_H_
#define DLIB_GPU_DaTA_H_
#include "gpu_data_abstract.h"
#include <memory>
#include <cstring>
#include "cuda_errors.h"
#include "../serialize.h"
namespace dlib
{
// ----------------------------------------------------------------------------------------
class gpu_data
{
/*!
CONVENTION
- if (size() != 0) then
- data_host == a pointer to size() floats in CPU memory.
- if (data_device) then
- data_device == a pointer to size() floats in device memory.
- if (there might be an active async transfer from host to device) then
- have_active_transfer == true
- We use the host_current and device_current bools to keep track of which
copy of the data (or both) are most current. e.g. if the CPU has
modified the data and it hasn't been copied to the device yet then
host_current==true and device_current==false.
Similarly, we use device_in_use==true to indicate that device() has been
called and no operation to wait for all CUDA kernel completion has been
executed. So if device_in_use==true then there might be a CUDA kernel
executing that is using the device memory block contained in this object.
!*/
public:
gpu_data(
) : data_size(0), host_current(true), device_current(true),have_active_transfer(false),device_in_use(false), the_device_id(0)
{
}
// Not copyable
gpu_data(const gpu_data&) = delete;
gpu_data& operator=(const gpu_data&) = delete;
// but is movable
gpu_data(gpu_data&& item) : gpu_data() { swap(item); }
gpu_data& operator=(gpu_data&& item) { swap(item); return *this; }
int device_id() const { return the_device_id; }
#ifdef DLIB_USE_CUDA
void async_copy_to_device() const;
void set_size(size_t new_size);
#else
// Note that calls to host() or device() will block until any async transfers are complete.
void async_copy_to_device() const{}
void set_size(size_t new_size)
{
if (new_size == 0)
{
data_size = 0;
host_current = true;
device_current = true;
device_in_use = false;
data_host.reset();
data_device.reset();
}
else if (new_size != data_size)
{
data_size = new_size;
host_current = true;
device_current = true;
device_in_use = false;
data_host.reset(new float[new_size], std::default_delete<float[]>());
data_device.reset();
}
}
#endif
const float* host() const
{
copy_to_host();
return data_host.get();
}
float* host()
{
copy_to_host();
device_current = false;
return data_host.get();
}
float* host_write_only()
{
host_current = true;
device_current = false;
return data_host.get();
}
const float* device() const
{
#ifndef DLIB_USE_CUDA
DLIB_CASSERT(false, "CUDA NOT ENABLED");
#endif
copy_to_device();
device_in_use = true;
return data_device.get();
}
float* device()
{
#ifndef DLIB_USE_CUDA
DLIB_CASSERT(false, "CUDA NOT ENABLED");
#endif
copy_to_device();
host_current = false;
device_in_use = true;
return data_device.get();
}
float* device_write_only()
{
#ifndef DLIB_USE_CUDA
DLIB_CASSERT(false, "CUDA NOT ENABLED");
#endif
wait_for_transfer_to_finish();
host_current = false;
device_current = true;
device_in_use = true;
return data_device.get();
}
bool host_ready (
) const { return host_current; }
bool device_ready (
) const { return device_current && !have_active_transfer; }
size_t size() const { return data_size; }
void swap (gpu_data& item)
{
std::swap(data_size, item.data_size);
std::swap(host_current, item.host_current);
std::swap(device_current, item.device_current);
std::swap(have_active_transfer, item.have_active_transfer);
std::swap(data_host, item.data_host);
std::swap(data_device, item.data_device);
std::swap(cuda_stream, item.cuda_stream);
std::swap(the_device_id, item.the_device_id);
}
private:
#ifdef DLIB_USE_CUDA
void copy_to_device() const;
void copy_to_host() const;
void wait_for_transfer_to_finish() const;
#else
void copy_to_device() const{}
void copy_to_host() const{}
void wait_for_transfer_to_finish() const{}
#endif
size_t data_size;
mutable bool host_current;
mutable bool device_current;
mutable bool have_active_transfer;
mutable bool device_in_use;
std::shared_ptr<float> data_host;
std::shared_ptr<float> data_device;
std::shared_ptr<void> cuda_stream;
int the_device_id;
};
inline void serialize(const gpu_data& item, std::ostream& out)
{
int version = 1;
serialize(version, out);
serialize(item.size(), out);
auto data = item.host();
for (size_t i = 0; i < item.size(); ++i)
serialize(data[i], out);
}
inline void deserialize(gpu_data& item, std::istream& in)
{
int version;
deserialize(version, in);
if (version != 1)
throw serialization_error("Unexpected version found while deserializing dlib::gpu_data.");
size_t s;
deserialize(s, in);
item.set_size(s);
auto data = item.host();
for (size_t i = 0; i < item.size(); ++i)
deserialize(data[i], in);
}
#ifdef DLIB_USE_CUDA
void memcpy (gpu_data& dest, const gpu_data& src);
void memcpy (
gpu_data& dest,
size_t dest_offset,
const gpu_data& src,
size_t src_offset,
size_t num
);
#else
inline void memcpy (gpu_data& dest, const gpu_data& src)
{
DLIB_CASSERT(dest.size() == src.size());
if (src.size() == 0 || &dest == &src)
return;
std::memcpy(dest.host_write_only(), src.host(), sizeof(float)*src.size());
}
inline void memcpy (
gpu_data& dest,
size_t dest_offset,
const gpu_data& src,
size_t src_offset,
size_t num
)
{
DLIB_CASSERT(dest_offset + num <= dest.size());
DLIB_CASSERT(src_offset + num <= src.size());
if (num == 0)
return;
if (&dest == &src && std::max(dest_offset, src_offset) < std::min(dest_offset,src_offset)+num)
{
// if they perfectly alias each other then there is nothing to do
if (dest_offset == src_offset)
return;
else
std::memmove(dest.host()+dest_offset, src.host()+src_offset, sizeof(float)*num);
}
else
{
// if we write to the entire thing then we can use host_write_only()
if (dest_offset == 0 && num == dest.size())
std::memcpy(dest.host_write_only(), src.host()+src_offset, sizeof(float)*num);
else
std::memcpy(dest.host()+dest_offset, src.host()+src_offset, sizeof(float)*num);
}
}
#endif
// ----------------------------------------------------------------------------------------
}
#endif // DLIB_GPU_DaTA_H_
|