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
|
#ifndef THC_GENERIC_FILE
#define THC_GENERIC_FILE "generic/THCTensorCopy.c"
#else
/* specific methods */
#ifndef THC_REAL_IS_HALF
void THCTensor_(copyCPU)(THCState *state, THCTensor *self, struct THTensor *src)
{
THArgCheck(THCTensor_(nElement)(state, self) == THTensor_(nElement)(src), 2, "sizes do not match");
{
THCTensor *selfc = THCTensor_(newContiguous)(state, self);
src = THTensor_(newContiguous)(src);
THCudaCheck(cudaMemcpy(THCTensor_(data)(state,selfc),
THTensor_(data)(src),
THTensor_(nElement)(src) * sizeof(real),
cudaMemcpyHostToDevice));
THTensor_(free)(src);
THCTensor_(freeCopyTo)(state, selfc, self);
}
}
#endif
#ifndef THC_REAL_IS_HALF
#define IMPLEMENT_TH_CUDA_TENSOR_COPY(TYPEC) \
void THCTensor_(copy##TYPEC)(THCState *state, THCTensor *self, struct TH##TYPEC##Tensor *src) \
{ \
THArgCheck(THCTensor_(nElement)(state, self) == TH##TYPEC##Tensor_nElement(src), 2, "sizes do not match"); \
if(THCTypeIdx_(Real) == THCTypeIdx_(TYPEC)) { \
THCTensor_(copyCPU)(state, self, (THTensor*) src); /* cast just removes warnings */ \
} else { \
THLongStorage *size = TH##TYPEC##Tensor_newSizeOf(src); \
THTensor *srcf = THTensor_(newWithSize)(size, NULL); \
\
THTensor_(copy##TYPEC)(srcf, src); \
THCTensor_(copyCPU)(state, self, srcf); \
\
THLongStorage_free(size); \
THTensor_(free)(srcf); \
} \
}
#else
#define IMPLEMENT_TH_CUDA_TENSOR_COPY(TYPEC) \
void THCTensor_(copy##TYPEC)(THCState *state, THCTensor *self, struct TH##TYPEC##Tensor *src) \
{ \
THArgCheck(THCTensor_(nElement)(state, self) == TH##TYPEC##Tensor_nElement(src), 2, "sizes do not match"); \
THLongStorage *size = TH##TYPEC##Tensor_newSizeOf(src); \
THCudaTensor *buffer = THCudaTensor_newWithSize(state, size, NULL); \
THCudaTensor_copy##TYPEC(state, buffer, src); \
THCudaHalfTensor_copyCudaFloat(state, self, buffer); \
THCudaTensor_free(state, buffer); \
THLongStorage_free(size); \
}
#endif
IMPLEMENT_TH_CUDA_TENSOR_COPY(Byte)
IMPLEMENT_TH_CUDA_TENSOR_COPY(Char)
IMPLEMENT_TH_CUDA_TENSOR_COPY(Short)
IMPLEMENT_TH_CUDA_TENSOR_COPY(Int)
IMPLEMENT_TH_CUDA_TENSOR_COPY(Long)
IMPLEMENT_TH_CUDA_TENSOR_COPY(Float)
IMPLEMENT_TH_CUDA_TENSOR_COPY(Double)
/* copyCuda */
#ifndef THC_REAL_IS_HALF
void THTensor_(copyCuda)(THCState *state, THTensor *self, struct THCTensor *src)
{
THArgCheck(THTensor_(nElement)(self) == THCTensor_(nElement)(state, src), 2, "sizes do not match");
{
THTensor *selfc = THTensor_(newContiguous)(self);
src = THCTensor_(newContiguous)(state, src);
THCudaCheck(cudaMemcpy(THTensor_(data)(selfc),
THCTensor_(data)(state, src),
THCTensor_(nElement)(state, src) * sizeof(real),
cudaMemcpyDeviceToHost));
THCTensor_(free)(state, src);
THTensor_(freeCopyTo)(selfc, self);
}
}
#endif
#ifndef THC_REAL_IS_HALF
#define IMPLEMENT_TH_CUDA_TENSOR_COPY_TO(TYPEC) \
void TH_CONCAT_4(TH,TYPEC,Tensor_copyCuda,Real)(THCState *state, TH##TYPEC##Tensor *self, struct THCTensor *src) \
{ \
THArgCheck(TH##TYPEC##Tensor_nElement(self) == THCTensor_(nElement)(state, src), 2, "sizes do not match"); \
if(THCTypeIdx_(Real) == THCTypeIdx_(TYPEC)) { \
THTensor_(copyCuda)(state, (THTensor*) self, src); /* cast just removes compiler warning */ \
} else { \
THLongStorage *size = THCTensor_(newSizeOf)(state, src); \
THTensor *srcf = THTensor_(newWithSize)(size, NULL); \
\
THTensor_(copyCuda)(state, srcf, src); \
TH_CONCAT_4(TH,TYPEC,Tensor_copy,Real)(self, srcf); \
\
THLongStorage_free(size); \
THTensor_(free)(srcf); \
} \
}
#else
#define IMPLEMENT_TH_CUDA_TENSOR_COPY_TO(TYPEC) \
void TH_CONCAT_4(TH,TYPEC,Tensor_copyCuda,Real)(THCState *state, TH##TYPEC##Tensor *self, struct THCTensor *src) \
{ \
THArgCheck(TH##TYPEC##Tensor_nElement(self) == THCTensor_(nElement)(state, src), 2, "sizes do not match"); \
THLongStorage *size = THCTensor_(newSizeOf)(state, src); \
THCudaTensor *buffer = THCudaTensor_newWithSize(state, size, NULL); \
THCudaTensor_copyCudaHalf(state, buffer, src); \
TH_CONCAT_3(TH,TYPEC,Tensor_copyCudaFloat)(state, self, buffer); \
THCudaTensor_free(state, buffer); \
THLongStorage_free(size); \
}
#endif
IMPLEMENT_TH_CUDA_TENSOR_COPY_TO(Byte)
IMPLEMENT_TH_CUDA_TENSOR_COPY_TO(Char)
IMPLEMENT_TH_CUDA_TENSOR_COPY_TO(Short)
IMPLEMENT_TH_CUDA_TENSOR_COPY_TO(Int)
IMPLEMENT_TH_CUDA_TENSOR_COPY_TO(Long)
IMPLEMENT_TH_CUDA_TENSOR_COPY_TO(Float)
IMPLEMENT_TH_CUDA_TENSOR_COPY_TO(Double)
void THCTensor_(copyCuda)(THCState *state, THCTensor *self, THCTensor *src)
{
THCTensor_(copy)(state, self, src);
}
#ifndef THC_REAL_IS_HALF
void THCTensor_(copyAsyncCPU)(THCState *state, THCTensor *self, struct THTensor *src)
{
THArgCheck(THCTensor_(nElement)(state, self) == THTensor_(nElement)(src), 2, "sizes do not match");
THArgCheck(THCTensor_(isContiguous)(state, self), 2, "Target tensor must be contiguous");
THArgCheck(THTensor_(isContiguous)(src), 3, "Source tensor must be contiguous");
if (THCTensor_(nElement)(state, self) == 0) return;
// Perform the copy wrt the current stream on the CudaTensor's device.
int tensorDevice = THCTensor_(getDevice)(state, self);
int currentDevice;
THCudaCheck(cudaGetDevice(¤tDevice));
if (currentDevice != tensorDevice) {
THCudaCheck(cudaSetDevice(tensorDevice));
}
THCudaCheck(cudaMemcpyAsync(THCTensor_(data)(state, self),
THTensor_(data)(src),
THTensor_(nElement)(src) * sizeof(real),
cudaMemcpyHostToDevice,
THCState_getDeviceStream(state, tensorDevice,
THCState_getCurrentStreamIndex(state))));
if (currentDevice != tensorDevice) {
THCudaCheck(cudaSetDevice(currentDevice));
}
}
void THTensor_(copyAsyncCuda)(THCState *state, THTensor *self, struct THCTensor *src)
{
THArgCheck(THTensor_(nElement)(self) == THCTensor_(nElement)(state, src), 2, "sizes do not match");
THArgCheck(THTensor_(isContiguous)(self), 2, "Target tensor must be contiguous");
THArgCheck(THCTensor_(isContiguous)(state, src), 3, "Source tensor must be contiguous");
if (THTensor_(nElement)(self) == 0) return;
// Perform the copy wrt the current stream on the CudaTensor's device.
int tensorDevice = THCTensor_(getDevice)(state, src);
int currentDevice;
THCudaCheck(cudaGetDevice(¤tDevice));
if (currentDevice != tensorDevice) {
THCudaCheck(cudaSetDevice(tensorDevice));
}
THCudaCheck(cudaMemcpyAsync(THTensor_(data)(self),
THCTensor_(data)(state, src),
THCTensor_(nElement)(state, src) * sizeof(real),
cudaMemcpyDeviceToHost,
THCState_getDeviceStream(state, tensorDevice,
THCState_getCurrentStreamIndex(state))));
if (currentDevice != tensorDevice) {
THCudaCheck(cudaSetDevice(currentDevice));
}
}
#endif
#undef IMPLEMENT_TH_CUDA_TENSOR_COPY
#undef IMPLEMENT_TH_CUDA_TENSOR_COPY_TO
#endif
|
