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
|
#include "THCTensorMath.h"
#include "THCGeneral.h"
#include "THCHalf.h"
#include "THCTensorCopy.h"
#include "THCApply.cuh"
#include "THCNumerics.cuh"
template <typename T>
struct TensorAddConstantOp {
TensorAddConstantOp(T v) : val(v) {}
__device__ __forceinline__ void operator()(T* out, T* in) {
*out = *in + val;
}
__device__ __forceinline__ void operator()(T* v) {
*v += val;
}
const T val;
};
#ifdef CUDA_HALF_TENSOR
template <>
struct TensorAddConstantOp<half> {
TensorAddConstantOp(half v) : val(v) {}
__device__ __forceinline__ void operator()(half* out, half* in) {
#ifdef CUDA_HALF_INSTRUCTIONS
*out = __hadd(*in, val);
#else
float fin = __half2float(*in);
float fval = __half2float(val);
float fout = fin + fval;
*out = __float2half(fout);
#endif
}
__device__ __forceinline__ void operator()(half* v) {
#ifdef CUDA_HALF_INSTRUCTIONS
*v = __hadd(*v, val);
#else
float fv = __half2float(*v);
float fval = __half2float(val);
fv += fval;
*v = __float2half(fv);
#endif
}
const half val;
};
#endif // CUDA_HALF_TENSOR
template <typename T>
struct TensorMulConstantOp {
TensorMulConstantOp(T v) : val(v) {}
__device__ __forceinline__ void operator()(T* out, T* in) {
*out = *in * val;
}
__device__ __forceinline__ void operator()(T* v) {
*v *= val;
}
const T val;
};
#ifdef CUDA_HALF_TENSOR
template <>
struct TensorMulConstantOp<half> {
TensorMulConstantOp(half v) : val(v) {}
__device__ __forceinline__ void operator()(half* out, half* in) {
#ifdef CUDA_HALF_INSTRUCTIONS
*out = __hmul(*in, val);
#else
float fin = __half2float(*in);
float fval = __half2float(val);
float fout = fin * fval;
*out = __float2half(fout);
#endif
}
__device__ __forceinline__ void operator()(half* v) {
#ifdef CUDA_HALF_INSTRUCTIONS
*v = __hmul(*v, val);
#else
float fv = __half2float(*v);
float fval = __half2float(val);
fv *= fval;
*v = __float2half(fv);
#endif
}
const half val;
};
#endif // CUDA_HALF_TENSOR
template <int Upper>
struct TensorTriOp {
TensorTriOp(float *start_, long stride0_, long stride1_, long k_)
: start(start_), stride0(stride0_), stride1(stride1_), k(k_) {}
__device__ __forceinline__ int mask(float *in) {
ptrdiff_t n = in - start;
long row, col;
if (stride0 > stride1)
{
row = (long) (n / stride0);
col = (long) ((n % stride0) / stride1);
}
else
{
row = (long) ((n % stride1) / stride0);
col = (long) (n / stride1);
}
return Upper ? (col - row >= k) : (col - row <= k);
}
__device__ __forceinline__ void operator()(float* out, float* in) {
*out = mask(in) ? *in : 0;
}
__device__ __forceinline__ void operator()(float* v) {
if (!mask(v))
*v = 0;
}
const float *start;
const long stride0, stride1, k;
};
void THCudaTensor_tril(THCState *state, THCudaTensor *self_, THCudaTensor *src_, long k)
{
THAssert(THCudaTensor_checkGPU(state, 2, self_, src_));
THArgCheck(src_->nDimension == 2, 1, "expected a matrix");
THCudaTensor *src = src_;
if (self_ == src_)
src = THCudaTensor_newContiguous(state, src_);
long stride0 = src->stride[0];
long stride1 = src->stride[1];
float *start = THCudaTensor_data(state, src) + src->storageOffset;
TensorTriOp<0> op(start, stride0, stride1, k);
if (self_ == src_) {
if (!THC_pointwiseApply1(state, src, op)) {
THArgCheck(false, 2, CUTORCH_DIM_WARNING);
}
} else {
THCudaTensor_resizeAs(state, self_, src);
if (!THC_pointwiseApply2(state, self_, src, op)) {
THArgCheck(false, 2, CUTORCH_DIM_WARNING);
}
}
if (self_ == src_)
THCudaTensor_freeCopyTo(state, src, src_);
THCudaCheck(cudaGetLastError());
}
void THCudaTensor_triu(THCState *state, THCudaTensor *self_, THCudaTensor *src_, long k)
{
THAssert(THCudaTensor_checkGPU(state, 2, self_, src_));
THArgCheck(src_->nDimension == 2, 1, "expected a matrix");
THCudaTensor *src = src_;
if (self_ == src_)
src = THCudaTensor_newContiguous(state, src_);
long stride0 = src->stride[0];
long stride1 = src->stride[1];
float *start = THCudaTensor_data(state, src) + src->storageOffset;
TensorTriOp<1> op(start, stride0, stride1, k);
if (self_ == src_) {
if (!THC_pointwiseApply1(state, src, op)) {
THArgCheck(false, 2, CUTORCH_DIM_WARNING);
}
} else {
THCudaTensor_resizeAs(state, self_, src);
if (!THC_pointwiseApply2(state, self_, src, op)) {
THArgCheck(false, 2, CUTORCH_DIM_WARNING);
}
}
if (self_ == src_)
THCudaTensor_freeCopyTo(state, src, src_);
THCudaCheck(cudaGetLastError());
}
#include "generic/THCTensorMathPairwise.cu"
#include "THCGenerateAllTypes.h"
|
