1 /* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
2
3 Licensed under the Apache License, Version 2.0 (the "License");
4 you may not use this file except in compliance with the License.
5 You may obtain a copy of the License at
6
7 http://www.apache.org/licenses/LICENSE-2.0
8
9 Unless required by applicable law or agreed to in writing, software
10 distributed under the License is distributed on an "AS IS" BASIS,
11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 See the License for the specific language governing permissions and
13 limitations under the License.
14 ==============================================================================*/
15
16 #include "tensorflow/core/kernels/hexagon/graph_transfer_utils.h"
17 #include "tensorflow/core/framework/graph.pb.h"
18 #include "tensorflow/core/framework/remote_fused_graph_execute_info.pb.h"
19
20 #include "tensorflow/cc/framework/scope.h"
21 #include "tensorflow/cc/ops/const_op.h"
22 #include "tensorflow/core/framework/tensor_shape.pb.h"
23 #include "tensorflow/core/graph/node_builder.h"
24 #include "tensorflow/core/platform/logging.h"
25 namespace tensorflow {
26
27 // function alias
28 constexpr auto AddOutputTensorShapeTypeByTensorShapeMap =
29 &RemoteFusedGraphExecuteUtils::AddOutputTensorShapeTypeByTensorShapeMap;
30
31 /* static */ std::priority_queue<std::tuple<float, int, string>>
GetTopNFloatResults(const float * const data,const string * const labels,const int element_count)32 GraphTransferUtils::GetTopNFloatResults(const float* const data,
33 const string* const labels,
34 const int element_count) {
35 CHECK(data != nullptr);
36 CHECK(labels != nullptr);
37 std::priority_queue<std::tuple<float, int, string>> queue;
38 for (int i = 0; i < element_count; ++i) {
39 queue.emplace(data[i], i, labels[i]);
40 }
41 return queue;
42 }
43
DumpTopNFloatResults(const float * const data,const string * const labels,const int element_count,const int top_n)44 /* static */ void GraphTransferUtils::DumpTopNFloatResults(
45 const float* const data, const string* const labels,
46 const int element_count, const int top_n) {
47 std::priority_queue<std::tuple<float, int, string>> queue =
48 GetTopNFloatResults(data, labels, element_count);
49 LOG(INFO) << "=== Dump ranking ===";
50 for (int i = 0; i < top_n; ++i) {
51 const std::tuple<float, int, string>& entry = queue.top();
52 LOG(INFO) << i << ": " << std::get<1>(entry) << ", " << std::get<2>(entry)
53 << ", " << std::get<0>(entry);
54 queue.pop();
55 }
56 }
57
58 /* static */ RemoteFusedGraphExecuteInfo
BuildRemoteFusedGraphExecuteInfo(const GraphDef & graph_def,const std::vector<std::pair<string,Tensor>> & inputs,const std::vector<string> & outputs,const RemoteFusedGraphExecuteUtils::TensorShapeMap & tensor_shape_map)59 GraphTransferUtils::BuildRemoteFusedGraphExecuteInfo(
60 const GraphDef& graph_def,
61 const std::vector<std::pair<string, Tensor>>& inputs,
62 const std::vector<string>& outputs,
63 const RemoteFusedGraphExecuteUtils::TensorShapeMap& tensor_shape_map) {
64 RemoteFusedGraphExecuteInfo execute_info;
65 execute_info.set_executor_name("build_hexagon_remote_fused_graph_executor");
66
67 // copy graph
68 *execute_info.mutable_remote_graph() = graph_def;
69
70 for (const std::pair<string, Tensor>& input : inputs) {
71 execute_info.add_graph_input_node_name(input.first);
72 RemoteFusedGraphExecuteInfo::TensorShapeTypeProto& tensor_shape_type =
73 *execute_info.add_default_graph_input_tensor_shape();
74 tensor_shape_type.set_dtype(input.second.dtype());
75 TensorShapeProto& tensor_shape_proto = *tensor_shape_type.mutable_shape();
76 for (const int64 dim : input.second.shape().dim_sizes()) {
77 tensor_shape_proto.add_dim()->set_size(dim);
78 }
79 }
80
81 for (const string& output_name : outputs) {
82 const std::pair<DataType, TensorShape>* tensor_shape_type =
83 RemoteFusedGraphExecuteUtils::GetTensorShapeType(tensor_shape_map,
84 output_name);
85 CHECK_NOTNULL(tensor_shape_type);
86 execute_info.add_graph_output_node_name(output_name);
87 RemoteFusedGraphExecuteInfo::TensorShapeTypeProto& tensor_shape_type_proto =
88 *execute_info.add_default_graph_output_tensor_shape();
89 tensor_shape_type_proto.set_dtype(tensor_shape_type->first);
90 TensorShapeProto& tensor_shape_proto =
91 *tensor_shape_type_proto.mutable_shape();
92 for (const int64 dim : tensor_shape_type->second.dim_sizes()) {
93 tensor_shape_proto.add_dim()->set_size(dim);
94 }
95 }
96
97 return execute_info;
98 }
99
BuildFusedGraphDef(const IRemoteFusedGraphOpsDefinitions & ops_definitions,const string & remote_graph_execute_name,const std::vector<std::pair<string,Tensor>> & inputs,const std::vector<string> & outputs,GraphDef * original_def)100 /* static */ GraphDef GraphTransferUtils::BuildFusedGraphDef(
101 const IRemoteFusedGraphOpsDefinitions& ops_definitions,
102 const string& remote_graph_execute_name,
103 const std::vector<std::pair<string, Tensor>>& inputs,
104 const std::vector<string>& outputs, GraphDef* original_def) {
105 RemoteFusedGraphExecuteUtils::TensorShapeMap tensor_shape_map;
106 Status status = RemoteFusedGraphExecuteUtils::DryRunInferenceForAllNode(
107 *original_def, inputs, true /* initialize_by_zero */, &tensor_shape_map);
108 for (NodeDef& node_def : *original_def->mutable_node()) {
109 TF_CHECK_OK(
110 AddOutputTensorShapeTypeByTensorShapeMap(tensor_shape_map, &node_def));
111 }
112 CHECK(status.ok());
113
114 Scope root = Scope::NewRootScope();
115 std::vector<Output> output_list;
116 DataTypeVector input_types;
117 for (const std::pair<string, Tensor>& input_node_info : inputs) {
118 const Scope& scope = root.WithOpName(input_node_info.first);
119 Node* ret;
120 const auto unique_name = scope.GetUniqueNameForOp("Placeholder");
121 auto builder = NodeBuilder(unique_name, "Placeholder")
122 .Attr("dtype", input_node_info.second.dtype())
123 .Attr("shape", input_node_info.second.shape());
124 scope.UpdateBuilder(&builder);
125 scope.UpdateStatus(builder.Finalize(scope.graph(), &ret));
126 TF_CHECK_OK(scope.status());
127 output_list.emplace_back(Output(ret, 0));
128 input_types.push_back(input_node_info.second.dtype());
129 }
130
131 const RemoteFusedGraphExecuteInfo execute_info =
132 BuildRemoteFusedGraphExecuteInfo(*original_def, inputs, outputs,
133 tensor_shape_map);
134
135 DataTypeVector output_types;
136 // Sanity-check to confirm all output data types are same.
137 for (const string& output_node_name : outputs) {
138 const std::pair<DataType, TensorShape>* tst =
139 RemoteFusedGraphExecuteUtils::GetTensorShapeType(tensor_shape_map,
140 output_node_name);
141 CHECK_NE(tst, nullptr);
142 output_types.push_back(tst->first);
143 }
144
145 const Scope& scope = root.WithOpName(remote_graph_execute_name);
146 CHECK(scope.ok());
147 auto node_out_list = ops::AsNodeOutList(scope, InputList(output_list));
148 Node* node;
149 const auto unique_name = scope.GetUniqueNameForOp("RemoteFusedGraphExecute");
150
151 auto builder = NodeBuilder(unique_name, "RemoteFusedGraphExecute")
152 .Input(node_out_list)
153 .Attr("Tinputs", input_types)
154 .Attr("Toutputs", output_types)
155 .Attr("serialized_remote_fused_graph_execute_info",
156 StringPiece(execute_info.SerializeAsString()));
157 CHECK(scope.ok());
158 scope.UpdateBuilder(&builder);
159 scope.UpdateStatus(builder.Finalize(scope.graph(), &node));
160 CHECK(scope.ok()) << scope.status();
161
162 GraphDef fusedGraphDef;
163 TF_CHECK_OK(root.ToGraphDef(&fusedGraphDef));
164 return fusedGraphDef;
165 }
166
167 } // namespace tensorflow
168