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
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
|
#include "caffe2/core/context_gpu.h"
#include "caffe2/operators/prelu_op.h"
#include "caffe2/utils/cub_namespace.cuh"
#include <cub/block/block_reduce.cuh>
namespace caffe2 {
namespace {
template <typename T>
__global__ void PReluKernel(const int N, const T* X, const T* W, T* Y) {
CUDA_1D_KERNEL_LOOP(i, N) {
Y[i] = (X[i] > 0) * X[i] + (X[i] < 0) * X[i] * W[0];
}
}
template <typename T>
__global__ void PReluKernelNCHW(
const int N,
const int C,
const int dim,
const T* X,
const T* W,
T* Y) {
CUDA_1D_KERNEL_LOOP(i, N * C * dim) {
int c = (i / dim) % C;
Y[i] = (X[i] > 0) * X[i] + (X[i] < 0) * X[i] * W[c];
}
}
template <typename T>
__global__ void
PReluKernelNHWC(const int nitems, const int C, const T* X, const T* W, T* Y) {
CUDA_1D_KERNEL_LOOP(i, nitems) {
int c = i % C;
Y[i] = (X[i] > 0) * X[i] + (X[i] < 0) * X[i] * W[c];
}
}
template <typename T>
__global__ void
PReluGradientKernel(const int N, const T* X, const T* W, const T* dY, T* dX) {
CUDA_1D_KERNEL_LOOP(i, N) {
dX[i] = (X[i] > 0) * dY[i] + (X[i] <= 0) * dY[i] * W[0];
}
}
template <typename T>
__global__ void PReluGradientKernelNCHW(
const int N,
const int C,
const int dim,
const T* X,
const T* W,
const T* dY,
T* dX) {
CUDA_1D_KERNEL_LOOP(i, N * C * dim) {
int c = (i / dim) % C;
dX[i] = (X[i] > 0) * dY[i] + (X[i] <= 0) * dY[i] * W[c];
}
}
template <typename T>
__global__ void PReluGradientKernelNHWC(
const int nitems,
const int C,
const T* X,
const T* W,
const T* dY,
T* dX) {
CUDA_1D_KERNEL_LOOP(i, nitems) {
int c = i % C;
dX[i] = (X[i] > 0) * dY[i] + (X[i] <= 0) * dY[i] * W[c];
}
}
template <typename T>
__global__ void PReluSharedWGradientKernelNCHW(
const int num_items,
const T* Xdata,
const T* dYdata,
T* dW) {
T wsum = 0.0;
for (int i = threadIdx.x; i < num_items; i += blockDim.x) {
wsum += (Xdata[i] <= 0) * dYdata[i] * Xdata[i];
}
typedef cub::BlockReduce<T, CAFFE_CUDA_NUM_THREADS> BlockReduce;
__shared__ typename BlockReduce::TempStorage temp_storage;
T sum = BlockReduce(temp_storage).Sum(wsum);
if (threadIdx.x == 0) {
*dW = sum;
}
}
template <typename T>
__global__ void PReluWGradientKernelNCHW(
const int C,
const int N,
const int num_items,
const T* Xdata,
const T* dYdata,
T* dW) {
int c = blockIdx.x;
T wsum = 0.0;
int items_per_channel = num_items / C;
int items_per_sample_channel = items_per_channel / N;
for (int i = threadIdx.x; i < items_per_channel; i += blockDim.x) {
// TODO: simplify
int n = i / items_per_sample_channel;
int ii = n * items_per_sample_channel * C + c * items_per_sample_channel +
i % items_per_sample_channel;
wsum += (Xdata[ii] <= 0) * dYdata[ii] * Xdata[ii];
}
typedef cub::BlockReduce<T, CAFFE_CUDA_NUM_THREADS> BlockReduce;
__shared__ typename BlockReduce::TempStorage temp_storage;
T sum = BlockReduce(temp_storage).Sum(wsum);
if (threadIdx.x == 0) {
dW[c] = sum;
}
}
template <typename T>
__global__ void PReluWGradientKernelNHWC(
const int C,
const int num_items,
const T* Xdata,
const T* dYdata,
T* dW) {
const auto c = blockIdx.x;
T wsum = 0.0;
const auto items_per_channel = num_items / C;
for (int i = threadIdx.x; i < items_per_channel; i += blockDim.x) {
const auto ii = i * C + c;
wsum += (Xdata[ii] <= 0) * dYdata[ii] * Xdata[ii];
}
typedef cub::BlockReduce<T, CAFFE_CUDA_NUM_THREADS> BlockReduce;
__shared__ typename BlockReduce::TempStorage temp_storage;
T sum = BlockReduce(temp_storage).Sum(wsum);
if (threadIdx.x == 0) {
dW[c] = sum;
}
}
} // namespace
template <>
bool PReluOp<float, CUDAContext>::RunOnDevice() {
const auto& X = Input(0);
const auto& W = Input(1);
auto* Y = Output(0, X.sizes(), at::dtype<float>());
const auto* Xdata = X.data<float>();
const auto* Wdata = W.data<float>();
auto* Ydata = Y->template mutable_data<float>();
const auto C = order_ == StorageOrder::NCHW ? X.dim(1) : X.dim(X.dim() - 1);
const auto C_shared = (W.numel() == 1);
if (!C_shared) {
CAFFE_ENFORCE_EQ(C, W.numel());
}
if (C_shared) {
PReluKernel<<<
CAFFE_GET_BLOCKS(X.numel()),
CAFFE_CUDA_NUM_THREADS,
0,
context_.cuda_stream()>>>(X.numel(), Xdata, Wdata, Ydata);
C10_CUDA_KERNEL_LAUNCH_CHECK();
return true;
}
// non-shared case.
switch (order_) {
case StorageOrder::NCHW: {
const auto N = X.dim(0);
const auto dim = X.size_from_dim(2);
CHECK(N * C * dim == X.numel());
PReluKernelNCHW<<<
CAFFE_GET_BLOCKS(X.numel()),
CAFFE_CUDA_NUM_THREADS,
0,
context_.cuda_stream()>>>(N, C, dim, Xdata, Wdata, Ydata);
C10_CUDA_KERNEL_LAUNCH_CHECK();
break;
}
case StorageOrder::NHWC: {
PReluKernelNHWC<<<
CAFFE_GET_BLOCKS(X.numel()),
CAFFE_CUDA_NUM_THREADS,
0,
context_.cuda_stream()>>>(X.numel(), C, Xdata, Wdata, Ydata);
C10_CUDA_KERNEL_LAUNCH_CHECK();
break;
}
default:
CAFFE_THROW("Unknown storage order: ", order_);
}
return true;
}
template <>
bool PReluGradientOp<float, CUDAContext>::RunOnDevice() {
auto& Y = Input(0);
auto& dY = Input(1);
auto& X = Input(2);
auto& W = Input(3);
CAFFE_ENFORCE(&Y != &X, "Cannot backpropagate through an in-place PReLU");
TORCH_DCHECK_EQ(dY.numel(), Y.numel());
auto* dX = Output(0, Y.sizes(), at::dtype<float>());
auto* dW = Output(1, W.sizes(), at::dtype<float>());
const auto C = order_ == StorageOrder::NCHW ? X.dim(1) : X.dim(X.dim() - 1);
const auto C_shared = (W.numel() == 1);
const float* Ydata = Y.data<float>();
const float* dYdata = dY.data<float>();
const float* Xdata = X.data<float>();
const float* Wdata = W.data<float>();
float* dXdata = dX->template mutable_data<float>();
float* dWdata = dW->template mutable_data<float>();
int N = Y.dim(0);
if (C_shared) {
PReluSharedWGradientKernelNCHW<<<
1,
CAFFE_CUDA_NUM_THREADS,
0,
context_.cuda_stream()>>>(X.numel(), Xdata, dYdata, dWdata);
C10_CUDA_KERNEL_LAUNCH_CHECK();
PReluGradientKernel<<<
CAFFE_GET_BLOCKS(X.numel()),
CAFFE_CUDA_NUM_THREADS,
0,
context_.cuda_stream()>>>(X.numel(), Xdata, Wdata, dYdata, dXdata);
C10_CUDA_KERNEL_LAUNCH_CHECK();
return true;
}
// non-shared case.
switch (order_) {
case StorageOrder::NCHW: {
const auto dim = Y.size_from_dim(2);
PReluWGradientKernelNCHW<<<
C,
CAFFE_CUDA_NUM_THREADS,
0,
context_.cuda_stream()>>>(C, N, X.numel(), Xdata, dYdata, dWdata);
C10_CUDA_KERNEL_LAUNCH_CHECK();
PReluGradientKernelNCHW<<<
CAFFE_GET_BLOCKS(X.numel()),
CAFFE_CUDA_NUM_THREADS,
0,
context_.cuda_stream()>>>(N, C, dim, Xdata, Wdata, dYdata, dXdata);
C10_CUDA_KERNEL_LAUNCH_CHECK();
break;
}
case StorageOrder::NHWC: {
PReluWGradientKernelNHWC<<<
C,
CAFFE_CUDA_NUM_THREADS,
0,
context_.cuda_stream()>>>(C, X.numel(), Xdata, dYdata, dWdata);
C10_CUDA_KERNEL_LAUNCH_CHECK();
PReluGradientKernelNHWC<<<
CAFFE_GET_BLOCKS(Y.numel()),
CAFFE_CUDA_NUM_THREADS,
0,
context_.cuda_stream()>>>(X.numel(), C, Xdata, Wdata, dYdata, dXdata);
C10_CUDA_KERNEL_LAUNCH_CHECK();
break;
}
default:
CAFFE_THROW("Unknown storage order: ", order_);
}
return true;
}
REGISTER_CUDA_OPERATOR(PRelu, PReluOp<float, CUDAContext>);
REGISTER_CUDA_OPERATOR(PReluGradient, PReluGradientOp<float, CUDAContext>);
} // namespace caffe2
|
