1 /**
2 * Copyright 2019 Huawei Technologies Co., Ltd
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 * http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17 #include <string>
18 #include <list>
19 #include <vector>
20 #include "common/common_test.h"
21 #include "frontend/parallel/strategy.h"
22 #include "frontend/parallel/ops_info/activation_info.h"
23 #include "frontend/parallel/device_manager.h"
24
25 namespace mindspore {
26 namespace parallel {
27
28 class Activation;
29 class Softmax;
30 using ActivationPtr = std::shared_ptr<ActivationInfo>;
31 using SoftmaxPtr = std::shared_ptr<Softmax>;
32 ActivationPtr act_ptr_;
33 SoftmaxPtr soft_ptr_;
34
35 class TestActivation : public UT::Common {
36 public:
TestActivation()37 TestActivation() {}
38 void SetUp();
TearDown()39 void TearDown() {}
40 };
41
SetUp()42 void TestActivation::SetUp() {
43 RankList dev_list;
44
45 for (int32_t i = 0; i < 1050; i++) {
46 dev_list.push_back(i);
47 }
48
49 RankList stage_map;
50 stage_map.push_back(1024);
51 stage_map.push_back(26);
52
53 int32_t local_dev = 0;
54
55 // create a new g_device_manager
56 g_device_manager = std::make_shared<DeviceManager>();
57 g_device_manager->Init(dev_list, local_dev, stage_map, "hccl");
58
59 ValuePtr relu = MakeValue(std::string("relu"));
60 std::unordered_map<std::string, ValuePtr> relu_attr = {{"activation_type", relu}};
61 ValuePtr sm = MakeValue(std::string("softmax"));
62 ValuePtr axix = MakeValue(std::int64_t(2));
63 std::unordered_map<std::string, ValuePtr> softmax_attr = {{"activation_type", sm}, {"axis", axix}};
64
65 Shapes relu_inputs_shape = {{2, 4, 8, 16}};
66 Shapes relu_outputs_shape = {{2, 4, 8, 16}};
67 Shapes sm_inputs_shape = {{8, 8, 8, 16}};
68 Shapes sm_outputs_shape = {{8, 8, 8, 16}};
69
70 act_ptr_ = std::make_shared<ActivationInfo>("relu_info", relu_inputs_shape, relu_outputs_shape, relu_attr);
71 soft_ptr_ = std::make_shared<Softmax>("softmax_info", sm_inputs_shape, sm_outputs_shape, softmax_attr);
72 }
73
TEST_F(TestActivation,test_activation_strategies)74 TEST_F(TestActivation, test_activation_strategies) {
75 ASSERT_EQ(act_ptr_->GenerateStrategies(0), Status::SUCCESS);
76 std::vector<std::shared_ptr<StrategyWithCost>> sc = act_ptr_->GetStrategyCost();
77 for (const auto& swc : sc) {
78 ASSERT_NE(swc, nullptr);
79 ASSERT_GT(swc->cost_list.size(), 0);
80 StrategyPtr sp = swc->strategy_ptr;
81 ASSERT_NE(sp, nullptr);
82 Cost cost = *(swc->cost_list[0]);
83
84 act_ptr_->InitForCostModel(sp);
85 std::vector<TensorInfo> inputs_info = act_ptr_->inputs_tensor_info();
86 std::vector<TensorInfo> outputs_info = act_ptr_->outputs_tensor_info();
87 ASSERT_DOUBLE_EQ(act_ptr_->operator_cost()->GetComputationCost(inputs_info, outputs_info, sp->GetInputStage()),
88 cost.computation_cost_);
89 ASSERT_DOUBLE_EQ(act_ptr_->operator_cost()->GetCommCost(inputs_info, outputs_info, sp->GetInputStage()),
90 cost.communication_cost_);
91 }
92 }
93
TEST_F(TestActivation,test_softmax_strategies)94 TEST_F(TestActivation, test_softmax_strategies) {
95 ASSERT_EQ(soft_ptr_->GenerateStrategies(0), Status::SUCCESS);
96 std::vector<std::shared_ptr<StrategyWithCost>> sc = soft_ptr_->GetStrategyCost();
97 for (const auto& swc : sc) {
98 ASSERT_NE(swc, nullptr);
99 ASSERT_GT(swc->cost_list.size(), 0);
100 StrategyPtr sp = swc->strategy_ptr;
101 ASSERT_NE(sp, nullptr);
102 Cost cost = *(swc->cost_list[0]);
103
104 Strategys stra = sp->GetInputDim();
105 ASSERT_GT(stra.size(), 0);
106 Dimensions input0_stra = stra[0];
107 ASSERT_GT(input0_stra.size(), 2);
108 ASSERT_EQ(input0_stra[2], 1);
109 soft_ptr_->InitForCostModel(sp);
110 std::vector<TensorInfo> inputs_info = soft_ptr_->inputs_tensor_info();
111 std::vector<TensorInfo> outputs_info = soft_ptr_->outputs_tensor_info();
112 ASSERT_DOUBLE_EQ(soft_ptr_->operator_cost()->GetComputationCost(inputs_info, outputs_info, sp->GetInputStage()),
113 cost.computation_cost_);
114 ASSERT_DOUBLE_EQ(soft_ptr_->operator_cost()->GetCommCost(inputs_info, outputs_info, sp->GetInputStage()),
115 cost.communication_cost_);
116 }
117 }
118
119 } // namespace parallel
120 } // namespace mindspore
121