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
|
#pragma once
#include "../graph_helpers.h"
#include "../node_interface.h"
namespace cudnn_frontend {
namespace graph {
class BatchNormNode : public NodeCRTP<BatchNormNode> {
public:
Batchnorm_attributes attributes;
BatchNormNode(Batchnorm_attributes&& attributes_, detail::Context const& context)
: NodeCRTP(context), attributes(std::move(attributes_)) {}
Type
getType() override final {
return Type::BATCHNORM;
}
error_t
infer_properties_node() override final {
CUDNN_FE_LOG_LABEL_ENDL("INFO: Inferencing properties for batchnorm node " << attributes.name);
attributes.fill_from_context(context);
auto X = attributes.inputs[Batchnorm_attributes::input_names::X];
auto Y = attributes.outputs[Batchnorm_attributes::output_names::Y];
// Only infer dims and strides if user did not set them
if (Y->get_dim().empty()) {
Y->set_dim(X->get_dim());
}
if (Y->get_stride().empty()) {
Y->set_stride(X->get_stride());
}
// Set channel length tensors
auto const x_tensor_dim = X->get_dim();
auto infer_per_channel_tensors = [&x_tensor_dim](std::shared_ptr<Tensor_attributes>& T) {
auto tensor_dim = T->get_dim();
// Only infer dims and strides if user did not set them
if (tensor_dim.empty()) {
tensor_dim.resize(x_tensor_dim.size(), 1);
tensor_dim[1] = x_tensor_dim[1];
T->set_dim(tensor_dim);
}
if (T->get_stride().empty()) {
auto const& T_dim = T->get_dim();
// Default to NHWC
auto const& stride_order = detail::generate_NHWC_stride_order(T_dim.size());
T->set_stride(detail::generate_stride(T_dim, stride_order));
}
};
infer_per_channel_tensors(attributes.outputs[Batchnorm_attributes::output_names::MEAN]);
infer_per_channel_tensors(attributes.outputs[Batchnorm_attributes::output_names::INV_VARIANCE]);
auto has_running_stats = attributes.inputs[Batchnorm_attributes::input_names::PREV_RUNNING_MEAN] ||
attributes.inputs[Batchnorm_attributes::input_names::PREV_RUNNING_VAR];
if (has_running_stats) {
infer_per_channel_tensors(attributes.outputs[Batchnorm_attributes::output_names::NEXT_RUNNING_MEAN]);
infer_per_channel_tensors(attributes.outputs[Batchnorm_attributes::output_names::NEXT_RUNNING_VAR]);
}
return {error_code_t::OK, ""};
}
error_t
create_cudnn_operations(
std::unordered_set<Tensor_attributes::uid_t>& uids_involved_in_operations,
std::vector<std::shared_ptr<cudnn_frontend::Operation>>& operations,
managed_backend_descriptor_t& raw_operations,
std::unordered_map<int64_t, std::shared_ptr<cudnn_frontend::Tensor>>& tensors) const override final {
CUDNN_FRONTEND_UNUSED(raw_operations);
CUDNN_FE_LOG_LABEL("INFO: Building BatchNormNode operations " << attributes.name << " ");
// Create operation by directly calling cuDNN backend API
Operation_v8 batchnorm_operation;
_CUDNN_CHECK_CUDNN_ERROR(
batchnorm_operation.initialize_managed_backend_pointer(CUDNN_BACKEND_OPERATION_NORM_FORWARD_DESCRIPTOR));
// Set norm mode to BATCH_NORM
cudnnBackendNormMode_t cudnn_norm_mode;
_CUDNN_CHECK_CUDNN_ERROR(detail::convert_to_cudnn_type(NormMode_t::BATCH_NORM, cudnn_norm_mode));
_CUDNN_CHECK_CUDNN_ERROR(detail::set_attribute(batchnorm_operation.get_raw_desc(),
CUDNN_ATTR_OPERATION_NORM_FWD_MODE,
CUDNN_TYPE_NORM_MODE,
1,
&cudnn_norm_mode));
// Set forward phase to TRAINING
cudnnBackendNormFwdPhase_t cudnn_norm_fwd_phase;
_CUDNN_CHECK_CUDNN_ERROR(detail::convert_to_cudnn_type(NormFwdPhase_t::TRAINING, cudnn_norm_fwd_phase));
_CUDNN_CHECK_CUDNN_ERROR(detail::set_attribute(batchnorm_operation.get_raw_desc(),
CUDNN_ATTR_OPERATION_NORM_FWD_PHASE,
CUDNN_TYPE_NORM_FWD_PHASE,
1,
&cudnn_norm_fwd_phase));
// Set input tensor X
CUDNN_FE_VALIDATE_AND_ASSIGN_INPUT_TENSOR(X, Batchnorm_attributes::input_names::X);
auto x_desc = tensors.at(X->second->get_uid())->get_raw_desc();
_CUDNN_CHECK_CUDNN_ERROR(detail::set_attribute(batchnorm_operation.get_raw_desc(),
CUDNN_ATTR_OPERATION_NORM_FWD_XDESC,
CUDNN_TYPE_BACKEND_DESCRIPTOR,
1,
&x_desc));
// Set saved mean and inv_variance
CUDNN_FE_VALIDATE_AND_ASSIGN_OUTPUT_TENSOR(MEAN, Batchnorm_attributes::output_names::MEAN);
auto mean_desc = tensors.at(MEAN->second->get_uid())->get_raw_desc();
_CUDNN_CHECK_CUDNN_ERROR(detail::set_attribute(batchnorm_operation.get_raw_desc(),
CUDNN_ATTR_OPERATION_NORM_FWD_MEAN_DESC,
CUDNN_TYPE_BACKEND_DESCRIPTOR,
1,
&mean_desc));
CUDNN_FE_VALIDATE_AND_ASSIGN_OUTPUT_TENSOR(INV_VARIANCE, Batchnorm_attributes::output_names::INV_VARIANCE);
auto inv_var_desc = tensors.at(INV_VARIANCE->second->get_uid())->get_raw_desc();
_CUDNN_CHECK_CUDNN_ERROR(detail::set_attribute(batchnorm_operation.get_raw_desc(),
CUDNN_ATTR_OPERATION_NORM_FWD_INV_VARIANCE_DESC,
CUDNN_TYPE_BACKEND_DESCRIPTOR,
1,
&inv_var_desc));
// Set scale and bias tensors
CUDNN_FE_VALIDATE_AND_ASSIGN_INPUT_TENSOR(SCALE, Batchnorm_attributes::input_names::SCALE);
auto scale_desc = tensors.at(SCALE->second->get_uid())->get_raw_desc();
_CUDNN_CHECK_CUDNN_ERROR(detail::set_attribute(batchnorm_operation.get_raw_desc(),
CUDNN_ATTR_OPERATION_NORM_FWD_SCALE_DESC,
CUDNN_TYPE_BACKEND_DESCRIPTOR,
1,
&scale_desc));
CUDNN_FE_VALIDATE_AND_ASSIGN_INPUT_TENSOR(BIAS, Batchnorm_attributes::input_names::BIAS);
auto bias_desc = tensors.at(BIAS->second->get_uid())->get_raw_desc();
_CUDNN_CHECK_CUDNN_ERROR(detail::set_attribute(batchnorm_operation.get_raw_desc(),
CUDNN_ATTR_OPERATION_NORM_FWD_BIAS_DESC,
CUDNN_TYPE_BACKEND_DESCRIPTOR,
1,
&bias_desc));
// Set epsilon tensor
CUDNN_FE_VALIDATE_AND_ASSIGN_INPUT_TENSOR(EPSILON, Batchnorm_attributes::input_names::EPSILON);
auto epsilon_desc = tensors.at(EPSILON->second->get_uid())->get_raw_desc();
_CUDNN_CHECK_CUDNN_ERROR(detail::set_attribute(batchnorm_operation.get_raw_desc(),
CUDNN_ATTR_OPERATION_NORM_FWD_EPSILON_DESC,
CUDNN_TYPE_BACKEND_DESCRIPTOR,
1,
&epsilon_desc));
// Check for running stats
bool has_running_stats = true;
auto it = attributes.inputs.find(Batchnorm_attributes::input_names::PREV_RUNNING_MEAN);
if (it == attributes.inputs.end() || it->second == nullptr) {
has_running_stats = false;
}
if (has_running_stats) {
// Set momentum (exp decay factor)
CUDNN_FE_VALIDATE_AND_ASSIGN_INPUT_TENSOR(MOMENTUM, Batchnorm_attributes::input_names::MOMENTUM);
auto momentum_desc = tensors.at(MOMENTUM->second->get_uid())->get_raw_desc();
_CUDNN_CHECK_CUDNN_ERROR(detail::set_attribute(batchnorm_operation.get_raw_desc(),
CUDNN_ATTR_OPERATION_NORM_FWD_EXP_AVG_FACTOR_DESC,
CUDNN_TYPE_BACKEND_DESCRIPTOR,
1,
&momentum_desc));
// Set prev running mean and var
CUDNN_FE_VALIDATE_AND_ASSIGN_INPUT_TENSOR(PREV_RUNNING_MEAN,
Batchnorm_attributes::input_names::PREV_RUNNING_MEAN);
auto prev_mean_desc = tensors.at(PREV_RUNNING_MEAN->second->get_uid())->get_raw_desc();
_CUDNN_CHECK_CUDNN_ERROR(detail::set_attribute(batchnorm_operation.get_raw_desc(),
CUDNN_ATTR_OPERATION_NORM_FWD_INPUT_RUNNING_MEAN_DESC,
CUDNN_TYPE_BACKEND_DESCRIPTOR,
1,
&prev_mean_desc));
CUDNN_FE_VALIDATE_AND_ASSIGN_INPUT_TENSOR(PREV_RUNNING_VAR,
Batchnorm_attributes::input_names::PREV_RUNNING_VAR);
auto prev_var_desc = tensors.at(PREV_RUNNING_VAR->second->get_uid())->get_raw_desc();
_CUDNN_CHECK_CUDNN_ERROR(detail::set_attribute(batchnorm_operation.get_raw_desc(),
CUDNN_ATTR_OPERATION_NORM_FWD_INPUT_RUNNING_VAR_DESC,
CUDNN_TYPE_BACKEND_DESCRIPTOR,
1,
&prev_var_desc));
// Set next running mean and var
CUDNN_FE_VALIDATE_AND_ASSIGN_OUTPUT_TENSOR(NEXT_RUNNING_MEAN,
Batchnorm_attributes::output_names::NEXT_RUNNING_MEAN);
auto next_mean_desc = tensors.at(NEXT_RUNNING_MEAN->second->get_uid())->get_raw_desc();
_CUDNN_CHECK_CUDNN_ERROR(detail::set_attribute(batchnorm_operation.get_raw_desc(),
CUDNN_ATTR_OPERATION_NORM_FWD_OUTPUT_RUNNING_MEAN_DESC,
CUDNN_TYPE_BACKEND_DESCRIPTOR,
1,
&next_mean_desc));
CUDNN_FE_VALIDATE_AND_ASSIGN_OUTPUT_TENSOR(NEXT_RUNNING_VAR,
Batchnorm_attributes::output_names::NEXT_RUNNING_VAR);
auto next_var_desc = tensors.at(NEXT_RUNNING_VAR->second->get_uid())->get_raw_desc();
_CUDNN_CHECK_CUDNN_ERROR(detail::set_attribute(batchnorm_operation.get_raw_desc(),
CUDNN_ATTR_OPERATION_NORM_FWD_OUTPUT_RUNNING_VAR_DESC,
CUDNN_TYPE_BACKEND_DESCRIPTOR,
1,
&next_var_desc));
}
// Set output tensor Y
CUDNN_FE_VALIDATE_AND_ASSIGN_OUTPUT_TENSOR(Y, Batchnorm_attributes::output_names::Y);
auto y_desc = tensors.at(Y->second->get_uid())->get_raw_desc();
_CUDNN_CHECK_CUDNN_ERROR(detail::set_attribute(batchnorm_operation.get_raw_desc(),
CUDNN_ATTR_OPERATION_NORM_FWD_YDESC,
CUDNN_TYPE_BACKEND_DESCRIPTOR,
1,
&y_desc));
// Set peer stat tensors if any
if (!attributes.peer_stats.empty()) {
std::vector<cudnnBackendDescriptor_t> peer_stat_descs;
for (auto const& peer_stat : attributes.peer_stats) {
peer_stat_descs.push_back(tensors.at(peer_stat->get_uid())->get_raw_desc());
}
_CUDNN_CHECK_CUDNN_ERROR(detail::set_attribute(batchnorm_operation.get_raw_desc(),
CUDNN_ATTR_OPERATION_NORM_FWD_PEER_STAT_DESCS,
CUDNN_TYPE_BACKEND_DESCRIPTOR,
peer_stat_descs.size(),
peer_stat_descs.data()));
}
_CUDNN_CHECK_CUDNN_ERROR(detail::finalize(batchnorm_operation.get_raw_desc()));
operations.push_back(std::make_shared<Operation_v8>(std::move(batchnorm_operation)));
auto const& non_virtual_uids = attributes.get_non_virtual_uids();
uids_involved_in_operations.insert(non_virtual_uids.begin(), non_virtual_uids.end());
return {error_code_t::OK, ""};
}
#ifndef CUDNN_FRONTEND_SKIP_JSON_LIB
virtual void
serialize(json& j) const override final {
j = attributes;
j.update(R"( {"tag": "BATCHNORM"})"_json);
}
#endif
};
} // namespace graph
} // namespace cudnn_frontend
|