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
|
#ifdef USE_CUDNN
#include <vector>
#include "caffe/filler.hpp"
#include "caffe/layer.hpp"
#include "caffe/util/im2col.hpp"
#include "caffe/util/math_functions.hpp"
#include "caffe/vision_layers.hpp"
namespace caffe {
__global__ void sync_conv_groups() { }
template <typename Dtype>
void CuDNNConvolutionLayer<Dtype>::Forward_gpu(
const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) {
const int* kernel_shape_data = this->kernel_shape_.cpu_data();
const int kernel_h = kernel_shape_data[0];
const int kernel_w = kernel_shape_data[1];
const size_t workspace_limit_bytes =
kernel_h * kernel_w * this->channels_ * sizeof(int) + 1;
const Dtype* weight = this->blobs_[0]->gpu_data();
for (int i = 0; i < bottom.size(); ++i) {
const Dtype* bottom_data = bottom[i]->gpu_data();
Dtype* top_data = top[i]->mutable_gpu_data();
// Forward through cuDNN in parallel over groups.
for (int g = 0; g < this->group_; g++) {
cudnnConvolutionFwdAlgo_t algo;
// pick the convolution algorithm
// TODO(shelhamer) this should be done during reshape
// TODO(shelhamer) the choice of automatic or manual algorithm picking
// should be exposed in proto
CUDNN_CHECK(cudnnGetConvolutionForwardAlgorithm(handle_[g],
bottom_descs_[i],
filter_desc_,
conv_descs_[i],
top_descs_[i],
CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT,
workspace_limit_bytes, // memoryLimitInBytes,
&algo));
// get minimum size of the workspace needed for the desired algorithm
size_t workspaceSizeInBytes_temp = 0;
CUDNN_CHECK(cudnnGetConvolutionForwardWorkspaceSize(handle_[g],
bottom_descs_[i],
filter_desc_,
conv_descs_[i],
top_descs_[i],
algo,
&workspaceSizeInBytes_temp));
if (workspaceSizeInBytes_temp > workspaceSizeInBytes) {
workspaceSizeInBytes = workspaceSizeInBytes_temp;
// free the existing workspace and allocate a new (larger) one
cudaFree(this->workspace);
cudaError_t err = cudaMalloc(&(this->workspace), workspaceSizeInBytes);
if (err != cudaSuccess) {
// force zero memory path
algo = CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM;
workspace = NULL;
workspaceSizeInBytes = 0;
}
}
// Filters.
CUDNN_CHECK(cudnnConvolutionForward(handle_[g],
cudnn::dataType<Dtype>::one,
bottom_descs_[i], bottom_data + bottom_offset_ * g,
filter_desc_, weight + this->weight_offset_ * g,
conv_descs_[i],
algo, workspace, workspaceSizeInBytes,
cudnn::dataType<Dtype>::zero,
top_descs_[i], top_data + top_offset_ * g));
// Bias.
if (this->bias_term_) {
const Dtype* bias_data = this->blobs_[1]->gpu_data();
CUDNN_CHECK(cudnnAddTensor(handle_[g], CUDNN_ADD_SAME_C,
cudnn::dataType<Dtype>::one,
bias_desc_, bias_data + bias_offset_ * g,
cudnn::dataType<Dtype>::one,
top_descs_[i], top_data + top_offset_ * g));
}
}
// Synchronize the work across groups, each of which went into its own
// stream, by launching an empty kernel into the default (null) stream.
// NOLINT_NEXT_LINE(whitespace/operators)
sync_conv_groups<<<1, 1>>>();
}
}
template <typename Dtype>
void CuDNNConvolutionLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top,
const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {
const Dtype* weight = NULL;
Dtype* weight_diff = NULL;
if (this->param_propagate_down_[0]) {
weight = this->blobs_[0]->gpu_data();
weight_diff = this->blobs_[0]->mutable_gpu_diff();
}
Dtype* bias_diff = NULL;
if (this->bias_term_ && this->param_propagate_down_[1]) {
bias_diff = this->blobs_[1]->mutable_gpu_diff();
}
for (int i = 0; i < top.size(); ++i) {
const Dtype* top_diff = top[i]->gpu_diff();
// Backward through cuDNN in parallel over groups and gradients.
for (int g = 0; g < this->group_; g++) {
// Gradient w.r.t. bias.
if (this->bias_term_ && this->param_propagate_down_[1]) {
CUDNN_CHECK(cudnnConvolutionBackwardBias(handle_[0*this->group_ + g],
cudnn::dataType<Dtype>::one,
top_descs_[i], top_diff + top_offset_ * g,
cudnn::dataType<Dtype>::one,
bias_desc_, bias_diff + bias_offset_ * g));
}
// Gradient w.r.t. weights.
if (this->param_propagate_down_[0]) {
const Dtype* bottom_data = bottom[i]->gpu_data();
CUDNN_CHECK(cudnnConvolutionBackwardFilter(handle_[1*this->group_ + g],
cudnn::dataType<Dtype>::one,
bottom_descs_[i], bottom_data + bottom_offset_ * g,
top_descs_[i], top_diff + top_offset_ * g,
conv_descs_[i],
cudnn::dataType<Dtype>::one,
filter_desc_, weight_diff + this->weight_offset_ * g));
}
// Gradient w.r.t. bottom data.
if (propagate_down[i]) {
if (weight == NULL) {
weight = this->blobs_[0]->gpu_data();
}
Dtype* bottom_diff = bottom[i]->mutable_gpu_diff();
CUDNN_CHECK(cudnnConvolutionBackwardData(handle_[2*this->group_ + g],
cudnn::dataType<Dtype>::one,
filter_desc_, weight + this->weight_offset_ * g,
top_descs_[i], top_diff + top_offset_ * g,
conv_descs_[i],
cudnn::dataType<Dtype>::zero,
bottom_descs_[i], bottom_diff + bottom_offset_ * g));
}
}
// Synchronize the work across groups, each of which went into its own
// stream, by launching an empty kernel into the default (null) stream.
// NOLINT_NEXT_LINE(whitespace/operators)
sync_conv_groups<<<1, 1>>>();
}
}
INSTANTIATE_LAYER_GPU_FUNCS(CuDNNConvolutionLayer);
} // namespace caffe
#endif
|
