layer_norm_cuda.cpp 6.42 KB
Newer Older
1
2
3
#include <torch/extension.h>
#include <vector>
#include <cassert>
4
#include "compat.h"
5
6
7
8

namespace {
void compute_n1_n2(
    at::Tensor input,
9
    #ifdef VERSION_GE_1_1
10
    at::IntArrayRef normalized_shape,
11
12
13
    #else
    at::IntList normalized_shape,
    #endif
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
    int& n1,
    int& n2)
{
    int idiff = input.ndimension() - normalized_shape.size();
    n2 = 1;
    for (int i = 0;  i < (int)normalized_shape.size();  ++i) {
	    assert( input.sizes()[i+idiff] == normalized_shape[i] );
	    n2 *= normalized_shape[i];
    }
    n1 = 1;
    for (int i = 0;  i < idiff;  ++i) {
	    n1 *= input.sizes()[i];
    }
}

void check_args(
30
    #ifdef VERSION_GE_1_1
31
    at::IntArrayRef normalized_shape,
32
33
34
    #else
    at::IntList normalized_shape,
    #endif
35
36
37
38
    at::Tensor gamma,
    at::Tensor beta
    )
{
39
40
    TORCH_CHECK(!gamma.defined() || gamma.sizes().equals(normalized_shape));
    TORCH_CHECK(!beta.defined() || beta.sizes().equals(normalized_shape));
41
42
43
44
}

void check_args(
    at::Tensor input,
45
    #ifdef VERSION_GE_1_1
46
    at::IntArrayRef normalized_shape,
47
48
49
    #else
    at::IntList normalized_shape,
    #endif
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
    int& n1,
    int& n2
    )
{
    int64_t normalized_ndim = normalized_shape.size();

    if (normalized_ndim < 1) {
      std::stringstream ss;
      ss << "Expected normalized_shape to be at least 1-dimensional, i.e., "
         << "containing at least one element, but got normalized_shape="
         << normalized_shape;
      throw std::runtime_error(ss.str());
    }

    auto input_shape = input.sizes();
    auto input_ndim = input.dim();

    if (input_ndim < normalized_ndim ||
        !input_shape.slice(input_ndim - normalized_ndim).equals(normalized_shape)) {
      std::stringstream ss;
      ss << "Given normalized_shape=" << normalized_shape
         << ", expected input with shape [*";
      for (auto size : normalized_shape) {
        ss << ", " << size;
      }
      ss << "], but got input of size" << input_shape;
      throw std::runtime_error(ss.str());
    }

    compute_n1_n2(input,normalized_shape,n1,n2);
}


void check_args(
    at::Tensor input,
85
    #ifdef VERSION_GE_1_1
86
    at::IntArrayRef normalized_shape,
87
88
89
    #else
    at::IntList normalized_shape,
    #endif
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
    at::Tensor gamma,
    at::Tensor beta,
    int& n1,
    int& n2
    )
{
    check_args(input,normalized_shape,n1,n2);
    check_args(normalized_shape,gamma,beta);
}
}

void cuda_layer_norm(
    at::Tensor* output,
    at::Tensor* mean,
    at::Tensor* invvar,
    at::Tensor* input,
    int n1,
    int n2,
108
    #ifdef VERSION_GE_1_1
109
    at::IntArrayRef normalized_shape,
110
111
112
    #else
    at::IntList normalized_shape,
    #endif
113
114
115
116
    at::Tensor* gamma,
    at::Tensor* beta,
    double epsilon);

117
118
#define CHECK_CUDA(x) TORCH_CHECK(x.type().is_cuda(), #x " must be a CUDA tensor")
#define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")
119
120
121
122
#define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x)

std::vector<at::Tensor> layer_norm(
    at::Tensor input,
123
    #ifdef VERSION_GE_1_1
124
    at::IntArrayRef normalized_shape,
125
126
127
    #else
    at::IntList normalized_shape,
    #endif
128
129
130
131
132
    double epsilon) {
  CHECK_INPUT(input);
  int n1,n2;
  check_args(input,normalized_shape,n1,n2);
  at::Tensor output = at::empty_like(input);
133
  at::Tensor mean = at::empty({n1}, input.options().dtype(input.scalar_type()==at::ScalarType::Half ? at::ScalarType::Float : input.scalar_type()));
134
135
136
137
138
139
140
  at::Tensor invvar = at::empty_like(mean);
  cuda_layer_norm(&output,&mean,&invvar,&input,n1,n2,
      normalized_shape,NULL,NULL,epsilon);
  return {output, mean, invvar};
}
std::vector<at::Tensor> layer_norm_affine(
    at::Tensor input,
141
    #ifdef VERSION_GE_1_1
142
    at::IntArrayRef normalized_shape,
143
144
145
    #else
    at::IntList normalized_shape,
    #endif
146
147
148
149
150
151
152
153
154
    at::Tensor gamma,
    at::Tensor beta,
    double epsilon) {
  CHECK_INPUT(input);
  CHECK_INPUT(gamma);
  CHECK_INPUT(beta);
  int n1,n2;
  check_args(input,normalized_shape,gamma,beta,n1,n2);
  at::Tensor output = at::empty_like(input);
155
  at::Tensor mean = at::empty({n1}, input.options().dtype(input.scalar_type()==at::ScalarType::Half ? at::ScalarType::Float : input.scalar_type()));
156
157
158
159
160
161
162
163
164
165
166
167
168
  at::Tensor invvar = at::empty_like(mean);
  cuda_layer_norm(&output,&mean,&invvar,&input,n1,n2,
      normalized_shape,&gamma,&beta,epsilon);
  return {output, mean, invvar};
}

void cuda_layer_norm_gradient(
    at::Tensor* dout,
    at::Tensor* mean,
    at::Tensor* invvar,
    at::Tensor* input,
    int n1,
    int n2,
169
    #ifdef VERSION_GE_1_1
170
    at::IntArrayRef normalized_shape,
171
172
173
    #else
    at::IntList normalized_shape,
    #endif
174
175
176
177
178
179
180
181
182
183
184
185
186
    at::Tensor* gamma,
    at::Tensor* beta,
    double epsilon,
    at::Tensor* grad_input,
    at::Tensor* grad_gamma,
    at::Tensor* grad_beta
    );

at::Tensor layer_norm_gradient(
    at::Tensor dout,
    at::Tensor mean,
    at::Tensor invvar,
    at::Tensor input,
187
    #ifdef VERSION_GE_1_1
188
    at::IntArrayRef normalized_shape,
189
190
191
    #else
    at::IntList normalized_shape,
    #endif
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
    double epsilon) {
  CHECK_INPUT(dout);
  CHECK_INPUT(mean);
  CHECK_INPUT(invvar);
  CHECK_INPUT(input);
  int n1,n2;
  check_args(input,normalized_shape,n1,n2);
  at::Tensor grad_input = at::empty_like(input);
  cuda_layer_norm_gradient(&dout,&mean,&invvar,&input,n1,n2,
      normalized_shape,NULL,NULL,epsilon,
      &grad_input,NULL,NULL);
  return grad_input;
}
std::vector<at::Tensor> layer_norm_gradient_affine(
    at::Tensor dout,
    at::Tensor mean,
    at::Tensor invvar,
    at::Tensor input,
210
    #ifdef VERSION_GE_1_1
211
    at::IntArrayRef normalized_shape,
212
213
214
    #else
    at::IntList normalized_shape,
    #endif
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
    at::Tensor gamma,
    at::Tensor beta,
    double epsilon) {
  CHECK_INPUT(dout);
  CHECK_INPUT(mean);
  CHECK_INPUT(invvar);
  CHECK_INPUT(input);
  CHECK_INPUT(gamma);
  CHECK_INPUT(beta);
  int n1,n2;
  check_args(input,normalized_shape,gamma,beta,n1,n2);
  at::Tensor grad_input = at::empty_like(input);
  at::Tensor grad_gamma = at::empty_like(gamma);
  at::Tensor grad_beta = at::empty_like(beta);
  cuda_layer_norm_gradient(&dout,&mean,&invvar,&input,n1,n2,
      normalized_shape,&gamma,&beta,epsilon,
      &grad_input,&grad_gamma,&grad_beta);
  return {grad_input, grad_gamma, grad_beta};
}

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
  m.def("forward_affine", &layer_norm_affine, "LayerNorm forward (CUDA)");
  m.def("forward", &layer_norm, "LayerNorm forward (CUDA)");
  m.def("backward_affine", &layer_norm_gradient_affine, "LayerNorm backward (CUDA)");
  m.def("backward", &layer_norm_gradient, "LayerNorm backward (CUDA)");
}