Finn Williams | b76eaed | 2021-03-31 16:22:40 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2021 Arm Ltd and Contributors. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include <armnn/Exceptions.hpp> |
| 7 | |
Colm Donelan | 0c47974 | 2021-12-10 12:43:54 +0000 | [diff] [blame] | 8 | #include <armnn/backends/TensorHandle.hpp> |
| 9 | #include <armnn/backends/Workload.hpp> |
Finn Williams | b76eaed | 2021-03-31 16:22:40 +0100 | [diff] [blame] | 10 | |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 11 | #include <doctest/doctest.h> |
Finn Williams | b76eaed | 2021-03-31 16:22:40 +0100 | [diff] [blame] | 12 | |
Jim Flynn | 2776183 | 2022-03-20 21:52:17 +0000 | [diff] [blame^] | 13 | #include <thread> |
| 14 | |
Finn Williams | b76eaed | 2021-03-31 16:22:40 +0100 | [diff] [blame] | 15 | using namespace armnn; |
| 16 | |
Finn Williams | b76eaed | 2021-03-31 16:22:40 +0100 | [diff] [blame] | 17 | |
| 18 | namespace |
| 19 | { |
| 20 | |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 21 | TEST_SUITE("WorkloadAsyncExecuteTests") |
| 22 | { |
| 23 | |
Finn Williams | b76eaed | 2021-03-31 16:22:40 +0100 | [diff] [blame] | 24 | struct Workload0 : BaseWorkload<ElementwiseUnaryQueueDescriptor> |
| 25 | { |
| 26 | Workload0(const ElementwiseUnaryQueueDescriptor& descriptor, const WorkloadInfo& info) |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 27 | : BaseWorkload(descriptor, info) |
Finn Williams | b76eaed | 2021-03-31 16:22:40 +0100 | [diff] [blame] | 28 | { |
| 29 | } |
| 30 | |
| 31 | Workload0() : BaseWorkload(ElementwiseUnaryQueueDescriptor(), WorkloadInfo()) |
| 32 | { |
| 33 | } |
| 34 | |
| 35 | void Execute() const |
| 36 | { |
| 37 | int* inVals = static_cast<int*>(m_Data.m_Inputs[0][0].Map()); |
| 38 | int* outVals = static_cast<int*>(m_Data.m_Outputs[0][0].Map()); |
| 39 | |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 40 | for (unsigned int i = 0; |
| 41 | i < m_Data.m_Inputs[0][0].GetShape().GetNumElements(); |
| 42 | ++i) |
Finn Williams | b76eaed | 2021-03-31 16:22:40 +0100 | [diff] [blame] | 43 | { |
| 44 | outVals[i] = inVals[i] * outVals[i]; |
| 45 | inVals[i] = outVals[i]; |
| 46 | } |
| 47 | } |
| 48 | |
| 49 | void ExecuteAsync(WorkingMemDescriptor& desc) |
| 50 | { |
| 51 | int* inVals = static_cast<int*>(desc.m_Inputs[0][0].Map()); |
| 52 | int* outVals = static_cast<int*>(desc.m_Outputs[0][0].Map()); |
| 53 | |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 54 | for (unsigned int i = 0; |
| 55 | i < desc.m_Inputs[0][0].GetShape().GetNumElements(); |
| 56 | ++i) |
Finn Williams | b76eaed | 2021-03-31 16:22:40 +0100 | [diff] [blame] | 57 | { |
| 58 | outVals[i] = inVals[i] + outVals[i]; |
| 59 | inVals[i] = outVals[i]; |
| 60 | } |
| 61 | } |
| 62 | |
| 63 | QueueDescriptor* GetQueueDescriptor() |
| 64 | { |
| 65 | return &m_Data; |
| 66 | } |
| 67 | }; |
| 68 | |
| 69 | struct Workload1 : BaseWorkload<ElementwiseUnaryQueueDescriptor> |
| 70 | { |
| 71 | Workload1(const ElementwiseUnaryQueueDescriptor& descriptor, const WorkloadInfo& info) |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 72 | : BaseWorkload(descriptor, info) |
Finn Williams | b76eaed | 2021-03-31 16:22:40 +0100 | [diff] [blame] | 73 | { |
| 74 | } |
| 75 | |
| 76 | void Execute() const |
| 77 | { |
| 78 | int* inVals = static_cast<int*>(m_Data.m_Inputs[0][0].Map()); |
| 79 | int* outVals = static_cast<int*>(m_Data.m_Outputs[0][0].Map()); |
| 80 | |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 81 | for (unsigned int i = 0; |
| 82 | i < m_Data.m_Inputs[0][0].GetShape().GetNumElements(); |
| 83 | ++i) |
Finn Williams | b76eaed | 2021-03-31 16:22:40 +0100 | [diff] [blame] | 84 | { |
| 85 | outVals[i] = inVals[i] * outVals[i]; |
| 86 | inVals[i] = outVals[i]; |
| 87 | } |
| 88 | } |
| 89 | }; |
| 90 | |
| 91 | void ValidateTensor(ITensorHandle* tensorHandle, int expectedValue) |
| 92 | { |
| 93 | int* actualOutput = static_cast<int*>(tensorHandle->Map()); |
| 94 | |
| 95 | bool allValuesCorrect = true; |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 96 | for (unsigned int i = 0; |
| 97 | i < tensorHandle->GetShape().GetNumElements(); |
| 98 | ++i) |
Finn Williams | b76eaed | 2021-03-31 16:22:40 +0100 | [diff] [blame] | 99 | { |
| 100 | if (actualOutput[i] != expectedValue) |
| 101 | { |
| 102 | allValuesCorrect = false; |
| 103 | } |
| 104 | } |
| 105 | |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 106 | CHECK(allValuesCorrect); |
Finn Williams | b76eaed | 2021-03-31 16:22:40 +0100 | [diff] [blame] | 107 | } |
| 108 | |
| 109 | template<typename Workload> |
| 110 | std::unique_ptr<Workload> CreateWorkload(TensorInfo info, ITensorHandle* inputTensor, ITensorHandle* outputTensor) |
| 111 | { |
| 112 | WorkloadInfo workloadInfo; |
| 113 | workloadInfo.m_InputTensorInfos = std::vector<TensorInfo>{info}; |
| 114 | workloadInfo.m_OutputTensorInfos = std::vector<TensorInfo>{info}; |
| 115 | |
| 116 | ElementwiseUnaryQueueDescriptor elementwiseUnaryQueueDescriptor; |
| 117 | elementwiseUnaryQueueDescriptor.m_Inputs = std::vector<ITensorHandle*>{inputTensor}; |
| 118 | elementwiseUnaryQueueDescriptor.m_Outputs = std::vector<ITensorHandle*>{outputTensor}; |
| 119 | |
| 120 | return std::make_unique<Workload>(elementwiseUnaryQueueDescriptor, workloadInfo); |
| 121 | } |
| 122 | |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 123 | TEST_CASE("TestAsyncExecute") |
Finn Williams | b76eaed | 2021-03-31 16:22:40 +0100 | [diff] [blame] | 124 | { |
Cathal Corbett | 5b8093c | 2021-10-22 11:12:07 +0100 | [diff] [blame] | 125 | TensorInfo info({5}, DataType::Signed32, 0.0, 0, true); |
Finn Williams | b76eaed | 2021-03-31 16:22:40 +0100 | [diff] [blame] | 126 | |
| 127 | int inVals[5]{2, 2, 2, 2, 2}; |
| 128 | int outVals[5]{1, 1, 1, 1, 1}; |
| 129 | |
| 130 | int expectedExecuteval = 2; |
| 131 | int expectedExecuteAsyncval = 3; |
| 132 | |
| 133 | ConstTensor constInputTensor(info, inVals); |
| 134 | ConstTensor constOutputTensor(info, outVals); |
| 135 | |
James Conroy | 1f58f03 | 2021-04-27 17:13:27 +0100 | [diff] [blame] | 136 | ScopedTensorHandle syncInput0(constInputTensor); |
| 137 | ScopedTensorHandle syncOutput0(constOutputTensor); |
Finn Williams | b76eaed | 2021-03-31 16:22:40 +0100 | [diff] [blame] | 138 | |
| 139 | std::unique_ptr<Workload0> workload0 = CreateWorkload<Workload0>(info, &syncInput0, &syncOutput0); |
| 140 | |
| 141 | workload0.get()->Execute(); |
| 142 | |
James Conroy | 1f58f03 | 2021-04-27 17:13:27 +0100 | [diff] [blame] | 143 | ScopedTensorHandle asyncInput0(constInputTensor); |
| 144 | ScopedTensorHandle asyncOutput0(constOutputTensor); |
Finn Williams | b76eaed | 2021-03-31 16:22:40 +0100 | [diff] [blame] | 145 | |
| 146 | WorkingMemDescriptor workingMemDescriptor0; |
| 147 | workingMemDescriptor0.m_Inputs = std::vector<ITensorHandle*>{&asyncInput0}; |
| 148 | workingMemDescriptor0.m_Outputs = std::vector<ITensorHandle*>{&asyncOutput0}; |
| 149 | |
| 150 | workload0.get()->ExecuteAsync(workingMemDescriptor0); |
| 151 | |
| 152 | // Inputs are also changed by the execute/executeAsync calls to make sure there is no interference with them |
| 153 | ValidateTensor(workingMemDescriptor0.m_Outputs[0], expectedExecuteAsyncval); |
| 154 | ValidateTensor(workingMemDescriptor0.m_Inputs[0], expectedExecuteAsyncval); |
| 155 | |
| 156 | ValidateTensor(&workload0.get()->GetQueueDescriptor()->m_Outputs[0][0], expectedExecuteval); |
| 157 | ValidateTensor(&workload0.get()->GetQueueDescriptor()->m_Inputs[0][0], expectedExecuteval); |
| 158 | } |
| 159 | |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 160 | TEST_CASE("TestDefaultAsyncExecute") |
Finn Williams | b76eaed | 2021-03-31 16:22:40 +0100 | [diff] [blame] | 161 | { |
Cathal Corbett | 5b8093c | 2021-10-22 11:12:07 +0100 | [diff] [blame] | 162 | TensorInfo info({5}, DataType::Signed32, 0.0f, 0, true); |
Finn Williams | b76eaed | 2021-03-31 16:22:40 +0100 | [diff] [blame] | 163 | |
| 164 | std::vector<int> inVals{2, 2, 2, 2, 2}; |
| 165 | std::vector<int> outVals{1, 1, 1, 1, 1}; |
| 166 | std::vector<int> defaultVals{0, 0, 0, 0, 0}; |
| 167 | |
| 168 | int expectedExecuteval = 2; |
| 169 | |
| 170 | ConstTensor constInputTensor(info, inVals); |
| 171 | ConstTensor constOutputTensor(info, outVals); |
| 172 | ConstTensor defaultTensor(info, &defaultVals); |
| 173 | |
James Conroy | 1f58f03 | 2021-04-27 17:13:27 +0100 | [diff] [blame] | 174 | ScopedTensorHandle defaultInput = ScopedTensorHandle(defaultTensor); |
| 175 | ScopedTensorHandle defaultOutput = ScopedTensorHandle(defaultTensor); |
Finn Williams | b76eaed | 2021-03-31 16:22:40 +0100 | [diff] [blame] | 176 | |
| 177 | std::unique_ptr<Workload1> workload1 = CreateWorkload<Workload1>(info, &defaultInput, &defaultOutput); |
| 178 | |
James Conroy | 1f58f03 | 2021-04-27 17:13:27 +0100 | [diff] [blame] | 179 | ScopedTensorHandle asyncInput(constInputTensor); |
| 180 | ScopedTensorHandle asyncOutput(constOutputTensor); |
Finn Williams | b76eaed | 2021-03-31 16:22:40 +0100 | [diff] [blame] | 181 | |
| 182 | WorkingMemDescriptor workingMemDescriptor; |
| 183 | workingMemDescriptor.m_Inputs = std::vector<ITensorHandle*>{&asyncInput}; |
| 184 | workingMemDescriptor.m_Outputs = std::vector<ITensorHandle*>{&asyncOutput}; |
| 185 | |
| 186 | workload1.get()->ExecuteAsync(workingMemDescriptor); |
| 187 | |
| 188 | // workload1 has no AsyncExecute implementation and so should use the default workload AsyncExecute |
| 189 | // implementation which will call workload1.Execute() in a thread safe manner |
| 190 | ValidateTensor(workingMemDescriptor.m_Outputs[0], expectedExecuteval); |
| 191 | ValidateTensor(workingMemDescriptor.m_Inputs[0], expectedExecuteval); |
| 192 | } |
| 193 | |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 194 | TEST_CASE("TestDefaultAsyncExeuteWithThreads") |
Finn Williams | b76eaed | 2021-03-31 16:22:40 +0100 | [diff] [blame] | 195 | { |
| 196 | // Use a large vector so the threads have a chance to interact |
| 197 | unsigned int vecSize = 1000; |
Cathal Corbett | 5b8093c | 2021-10-22 11:12:07 +0100 | [diff] [blame] | 198 | TensorInfo info({vecSize}, DataType::Signed32, 0.0f, 0, true); |
Finn Williams | b76eaed | 2021-03-31 16:22:40 +0100 | [diff] [blame] | 199 | |
| 200 | std::vector<int> inVals1(vecSize, 2); |
| 201 | std::vector<int> outVals1(vecSize, 1); |
| 202 | std::vector<int> inVals2(vecSize, 5); |
| 203 | std::vector<int> outVals2(vecSize, -1); |
| 204 | |
| 205 | std::vector<int> defaultVals(vecSize, 0); |
| 206 | |
| 207 | int expectedExecuteval1 = 4; |
| 208 | int expectedExecuteval2 = 25; |
| 209 | ConstTensor constInputTensor1(info, inVals1); |
| 210 | ConstTensor constOutputTensor1(info, outVals1); |
| 211 | |
| 212 | ConstTensor constInputTensor2(info, inVals2); |
| 213 | ConstTensor constOutputTensor2(info, outVals2); |
| 214 | |
Ryan OShea | 49428b7 | 2022-02-07 11:16:23 +0000 | [diff] [blame] | 215 | ConstTensor defaultTensor(info, defaultVals.data()); |
Finn Williams | b76eaed | 2021-03-31 16:22:40 +0100 | [diff] [blame] | 216 | |
James Conroy | 1f58f03 | 2021-04-27 17:13:27 +0100 | [diff] [blame] | 217 | ScopedTensorHandle defaultInput = ScopedTensorHandle(defaultTensor); |
| 218 | ScopedTensorHandle defaultOutput = ScopedTensorHandle(defaultTensor); |
Finn Williams | b76eaed | 2021-03-31 16:22:40 +0100 | [diff] [blame] | 219 | std::unique_ptr<Workload1> workload = CreateWorkload<Workload1>(info, &defaultInput, &defaultOutput); |
| 220 | |
James Conroy | 1f58f03 | 2021-04-27 17:13:27 +0100 | [diff] [blame] | 221 | ScopedTensorHandle asyncInput1(constInputTensor1); |
| 222 | ScopedTensorHandle asyncOutput1(constOutputTensor1); |
Finn Williams | b76eaed | 2021-03-31 16:22:40 +0100 | [diff] [blame] | 223 | |
| 224 | WorkingMemDescriptor workingMemDescriptor1; |
| 225 | workingMemDescriptor1.m_Inputs = std::vector<ITensorHandle*>{&asyncInput1}; |
| 226 | workingMemDescriptor1.m_Outputs = std::vector<ITensorHandle*>{&asyncOutput1}; |
| 227 | |
| 228 | |
James Conroy | 1f58f03 | 2021-04-27 17:13:27 +0100 | [diff] [blame] | 229 | ScopedTensorHandle asyncInput2(constInputTensor2); |
| 230 | ScopedTensorHandle asyncOutput2(constOutputTensor2); |
Finn Williams | b76eaed | 2021-03-31 16:22:40 +0100 | [diff] [blame] | 231 | |
| 232 | WorkingMemDescriptor workingMemDescriptor2; |
| 233 | workingMemDescriptor2.m_Inputs = std::vector<ITensorHandle*>{&asyncInput2}; |
| 234 | workingMemDescriptor2.m_Outputs = std::vector<ITensorHandle*>{&asyncOutput2}; |
| 235 | |
| 236 | std::thread thread1 = std::thread([&]() |
| 237 | { |
| 238 | workload.get()->ExecuteAsync(workingMemDescriptor1); |
| 239 | workload.get()->ExecuteAsync(workingMemDescriptor1); |
| 240 | }); |
| 241 | |
| 242 | std::thread thread2 = std::thread([&]() |
| 243 | { |
| 244 | workload.get()->ExecuteAsync(workingMemDescriptor2); |
| 245 | workload.get()->ExecuteAsync(workingMemDescriptor2); |
| 246 | }); |
| 247 | |
| 248 | thread1.join(); |
| 249 | thread2.join(); |
| 250 | |
| 251 | ValidateTensor(workingMemDescriptor1.m_Outputs[0], expectedExecuteval1); |
| 252 | ValidateTensor(workingMemDescriptor1.m_Inputs[0], expectedExecuteval1); |
| 253 | |
| 254 | ValidateTensor(workingMemDescriptor2.m_Outputs[0], expectedExecuteval2); |
| 255 | ValidateTensor(workingMemDescriptor2.m_Inputs[0], expectedExecuteval2); |
| 256 | } |
| 257 | |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 258 | } |
Finn Williams | b76eaed | 2021-03-31 16:22:40 +0100 | [diff] [blame] | 259 | |
Jim Flynn | 2776183 | 2022-03-20 21:52:17 +0000 | [diff] [blame^] | 260 | } |