|
| 1 | +/* Copyright 2015 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/common_runtime/direct_session.h" |
| 17 | + |
| 18 | +#include <map> |
| 19 | +#include <memory> |
| 20 | +#include <string> |
| 21 | +#include <thread> // NOLINT |
| 22 | +#include <unistd.h> |
| 23 | +#include <unordered_map> |
| 24 | +#include <vector> |
| 25 | + |
| 26 | +#include "absl/memory/memory.h" |
| 27 | +#include "absl/strings/match.h" |
| 28 | +#include "tensorflow/core/common_runtime/device_factory.h" |
| 29 | +#include "tensorflow/core/common_runtime/device_mgr.h" |
| 30 | +#include "tensorflow/core/common_runtime/function_testlib.h" |
| 31 | +#include "tensorflow/core/framework/allocator.h" |
| 32 | +#include "tensorflow/core/framework/graph.pb.h" |
| 33 | +#include "tensorflow/core/framework/op_kernel.h" |
| 34 | +#include "tensorflow/core/framework/tensor.h" |
| 35 | +#include "tensorflow/core/framework/tensor_testutil.h" |
| 36 | +#include "tensorflow/core/framework/types.pb.h" |
| 37 | +#include "tensorflow/core/graph/costmodel.h" |
| 38 | +#include "tensorflow/core/graph/graph.h" |
| 39 | +#include "tensorflow/core/graph/node_builder.h" |
| 40 | +#include "tensorflow/core/graph/testlib.h" |
| 41 | +#include "tensorflow/core/kernels/ops_util.h" |
| 42 | +#include "tensorflow/core/lib/core/errors.h" |
| 43 | +#include "tensorflow/core/lib/core/status.h" |
| 44 | +#include "tensorflow/core/lib/core/status_test_util.h" |
| 45 | +#include "tensorflow/core/lib/core/threadpool.h" |
| 46 | +#include "tensorflow/core/lib/strings/str_util.h" |
| 47 | +#include "tensorflow/core/platform/protobuf.h" |
| 48 | +#include "tensorflow/core/platform/test.h" |
| 49 | +#include "tensorflow/core/platform/test_benchmark.h" |
| 50 | +#include "tensorflow/core/protobuf/rewriter_config.pb.h" |
| 51 | +#include "tensorflow/core/public/session.h" |
| 52 | +#include "tensorflow/core/public/session_options.h" |
| 53 | +#include "tensorflow/core/util/device_name_utils.h" |
| 54 | + |
| 55 | +#if GOOGLE_CUDA |
| 56 | +#include "third_party/gpus/cuda/include/cuda.h" |
| 57 | +#include "third_party/gpus/cuda/include/cuda_runtime_api.h" |
| 58 | +#elif TENSORFLOW_USE_ROCM |
| 59 | +#include "rocm/include/hip/hip_runtime.h" |
| 60 | +#endif // GOOGLE_CUDA |
| 61 | + |
| 62 | +namespace tensorflow { |
| 63 | +namespace { |
| 64 | + |
| 65 | +CallableOptions MakeCallableOptions(gtl::ArraySlice<string> feeds, |
| 66 | + gtl::ArraySlice<string> fetches, |
| 67 | + gtl::ArraySlice<string> targets) { |
| 68 | + CallableOptions ret; |
| 69 | + for (const string& feed : feeds) { |
| 70 | + ret.add_feed(feed); |
| 71 | + } |
| 72 | + for (const string& fetch : fetches) { |
| 73 | + ret.add_fetch(fetch); |
| 74 | + } |
| 75 | + for (const string& target : targets) { |
| 76 | + ret.add_target(target); |
| 77 | + } |
| 78 | + return ret; |
| 79 | +} |
| 80 | + |
| 81 | +SessionOptions DefaultSessionOptions() { |
| 82 | + SessionOptions options; |
| 83 | + (*options.config.mutable_device_count())["CPU"] = 2; |
| 84 | + return options; |
| 85 | +} |
| 86 | + |
| 87 | +std::unique_ptr<SessionGroup> CreateSessionGroup() { |
| 88 | + SessionGroup* sg = nullptr; |
| 89 | + SessionGroupMetadata metadata; |
| 90 | + NewSessionGroup(DefaultSessionOptions(), &sg, metadata); |
| 91 | + return std::unique_ptr<SessionGroup>(sg); |
| 92 | +} |
| 93 | + |
| 94 | +std::unique_ptr<Session> CreateSession() { |
| 95 | + return std::unique_ptr<Session>(NewSession(DefaultSessionOptions())); |
| 96 | +} |
| 97 | + |
| 98 | +class DirectSessionMinusAXTest : public ::testing::Test { |
| 99 | + public: |
| 100 | + void Initialize(std::initializer_list<float> a_values) { |
| 101 | + Graph graph(OpRegistry::Global()); |
| 102 | + |
| 103 | + Tensor a_tensor(DT_FLOAT, TensorShape({2, 2})); |
| 104 | + test::FillValues<float>(&a_tensor, a_values); |
| 105 | + Node* a = test::graph::Constant(&graph, a_tensor); |
| 106 | + a->set_assigned_device_name("/job:localhost/replica:0/task:0/cpu:0"); |
| 107 | + a_ = a->name(); |
| 108 | + |
| 109 | + Tensor x_tensor(DT_FLOAT, TensorShape({2, 1})); |
| 110 | + test::FillValues<float>(&x_tensor, {1, 1}); |
| 111 | + Node* x = test::graph::Constant(&graph, x_tensor); |
| 112 | + x->set_assigned_device_name("/job:localhost/replica:0/task:0/cpu:1"); |
| 113 | + x_ = x->name(); |
| 114 | + |
| 115 | + // y = A * x |
| 116 | + Node* y = test::graph::Matmul(&graph, a, x, false, false); |
| 117 | + y->set_assigned_device_name("/job:localhost/replica:0/task:0/cpu:0"); |
| 118 | + y_ = y->name(); |
| 119 | + |
| 120 | + Node* y_neg = test::graph::Unary(&graph, "Neg", y); |
| 121 | + y_neg_ = y_neg->name(); |
| 122 | + y_neg->set_assigned_device_name("/job:localhost/replica:0/task:0/cpu:1"); |
| 123 | + |
| 124 | + Node* z = test::graph::Unary(&graph, "Identity", y_neg); |
| 125 | + z_ = z->name(); |
| 126 | + z->set_assigned_device_name("/job:localhost/replica:0/task:0/cpu:1"); |
| 127 | + |
| 128 | + graph.ToGraphDef(&def_); |
| 129 | + } |
| 130 | + |
| 131 | + string a_; |
| 132 | + string x_; |
| 133 | + string y_; |
| 134 | + string y_neg_; |
| 135 | + string z_; |
| 136 | + GraphDef def_; |
| 137 | +}; |
| 138 | + |
| 139 | +TEST_F(DirectSessionMinusAXTest, TestDirectSessionGroup) { |
| 140 | + for (int i = 0; i < 1000; ++i) { |
| 141 | + Initialize({3, 2, -1, 0}); |
| 142 | + auto sg = CreateSessionGroup(); |
| 143 | + ASSERT_TRUE(sg != nullptr); |
| 144 | + TF_ASSERT_OK(sg->Create(def_)); |
| 145 | + std::vector<std::pair<string, Tensor>> inputs; |
| 146 | + |
| 147 | + // Request two targets: one fetch output and one non-fetched output. |
| 148 | + std::vector<string> output_names = {y_ + ":0"}; |
| 149 | + std::vector<string> target_nodes = {y_neg_}; |
| 150 | + std::vector<Tensor> outputs; |
| 151 | + Status s = sg->Run(inputs, output_names, target_nodes, &outputs); |
| 152 | + TF_ASSERT_OK(s); |
| 153 | + |
| 154 | + ASSERT_EQ(1, outputs.size()); |
| 155 | + // The first output should be initialized and have the correct |
| 156 | + // output. |
| 157 | + auto mat = outputs[0].matrix<float>(); |
| 158 | + ASSERT_TRUE(outputs[0].IsInitialized()); |
| 159 | + EXPECT_FLOAT_EQ(5.0, mat(0, 0)); |
| 160 | + |
| 161 | + //usleep(10000); |
| 162 | + } |
| 163 | +} |
| 164 | + |
| 165 | +TEST_F(DirectSessionMinusAXTest, TestDirectSession) { |
| 166 | + for (int i = 0; i < 1000; ++i) { |
| 167 | + Initialize({3, 2, -1, 0}); |
| 168 | + auto sg = CreateSession(); |
| 169 | + ASSERT_TRUE(sg != nullptr); |
| 170 | + TF_ASSERT_OK(sg->Create(def_)); |
| 171 | + std::vector<std::pair<string, Tensor>> inputs; |
| 172 | + |
| 173 | + // Request two targets: one fetch output and one non-fetched output. |
| 174 | + std::vector<string> output_names = {y_ + ":0"}; |
| 175 | + std::vector<string> target_nodes = {y_neg_}; |
| 176 | + std::vector<Tensor> outputs; |
| 177 | + Status s = sg->Run(inputs, output_names, target_nodes, &outputs); |
| 178 | + TF_ASSERT_OK(s); |
| 179 | + |
| 180 | + ASSERT_EQ(1, outputs.size()); |
| 181 | + // The first output should be initialized and have the correct |
| 182 | + // output. |
| 183 | + auto mat = outputs[0].matrix<float>(); |
| 184 | + ASSERT_TRUE(outputs[0].IsInitialized()); |
| 185 | + EXPECT_FLOAT_EQ(5.0, mat(0, 0)); |
| 186 | + |
| 187 | + //usleep(10000); |
| 188 | + } |
| 189 | +} |
| 190 | + |
| 191 | +} |
| 192 | +} // namespace tensorflow |
0 commit comments