Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2020 Arm Ltd and Contributors. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include <backendsCommon/test/CommonTestUtils.hpp> |
| 7 | |
| 8 | #include <test/GraphUtils.hpp> |
| 9 | |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 10 | #include <doctest/doctest.h> |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 11 | |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 12 | TEST_SUITE("ClFallback") |
| 13 | { |
| 14 | TEST_CASE("ClImportEnabledFallbackToNeon") |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 15 | { |
| 16 | using namespace armnn; |
| 17 | |
| 18 | IRuntime::CreationOptions options; |
| 19 | IRuntimePtr runtime(IRuntime::Create(options)); |
| 20 | |
| 21 | // Builds up the structure of the network. |
| 22 | INetworkPtr net(INetwork::Create()); |
| 23 | |
| 24 | IConnectableLayer* input0 = net->AddInputLayer(0, "input0"); |
| 25 | IConnectableLayer* input1 = net->AddInputLayer(1, "input1"); |
| 26 | IConnectableLayer* input2 = net->AddInputLayer(2, "input2"); |
| 27 | IConnectableLayer* add = net->AddAdditionLayer("add"); |
| 28 | IConnectableLayer* sub = net->AddSubtractionLayer("sub"); |
| 29 | IConnectableLayer* output = net->AddOutputLayer(0, "output"); |
| 30 | |
| 31 | input0->GetOutputSlot(0).Connect(add->GetInputSlot(0)); |
| 32 | input1->GetOutputSlot(0).Connect(add->GetInputSlot(1)); |
| 33 | input2->GetOutputSlot(0).Connect(sub->GetInputSlot(0)); |
| 34 | add->GetOutputSlot(0).Connect(sub->GetInputSlot(1)); |
| 35 | sub->GetOutputSlot(0).Connect(output->GetInputSlot(0)); |
| 36 | |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 37 | TensorInfo info = TensorInfo({ 1, 2, 4, 2 }, DataType::Float32); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 38 | |
| 39 | input0->GetOutputSlot(0).SetTensorInfo(info); |
| 40 | input1->GetOutputSlot(0).SetTensorInfo(info); |
| 41 | input2->GetOutputSlot(0).SetTensorInfo(info); |
| 42 | add->GetOutputSlot(0).SetTensorInfo(info); |
| 43 | sub->GetOutputSlot(0).SetTensorInfo(info); |
| 44 | |
| 45 | std::vector<BackendId> backends = { Compute::GpuAcc, Compute::CpuAcc }; |
| 46 | // Use BackendSelectionHint to specify CpuAcc for Subtraction layer |
| 47 | sub->BackendSelectionHint(backends[1]); |
| 48 | |
| 49 | // optimize the network |
| 50 | OptimizerOptions optOptions; |
| 51 | optOptions.m_ImportEnabled = true; |
| 52 | IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optOptions); |
| 53 | |
Francis Murtagh | 3d2b4b2 | 2021-02-15 18:23:17 +0000 | [diff] [blame] | 54 | Graph& graph = GetGraphForTesting(optNet.get()); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 55 | |
| 56 | armnn::Layer* const layer0 = GetFirstLayerWithName(graph, "input0"); |
| 57 | armnn::Layer* const layer1 = GetFirstLayerWithName(graph, "input1"); |
| 58 | armnn::Layer* const layer2 = GetFirstLayerWithName(graph, "input2"); |
| 59 | armnn::Layer* const layer3 = GetFirstLayerWithName(graph, "add"); |
| 60 | armnn::Layer* const layer4 = GetFirstLayerWithName(graph, "[ add (0) -> sub (1) ]"); |
| 61 | armnn::Layer* const layer5 = GetFirstLayerWithName(graph, "sub"); |
| 62 | armnn::Layer* const layer6 = GetFirstLayerWithName(graph, "output"); |
| 63 | |
| 64 | // Checks order is valid. |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 65 | CHECK(CheckOrder(graph, layer0, layer1)); |
| 66 | CHECK(CheckOrder(graph, layer1, layer2)); |
| 67 | CHECK(CheckOrder(graph, layer2, layer3)); |
| 68 | CHECK(CheckOrder(graph, layer3, layer4)); |
| 69 | CHECK(CheckOrder(graph, layer4, layer5)); |
| 70 | CHECK(CheckOrder(graph, layer5, layer6)); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 71 | |
| 72 | // Use memory import between backends |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 73 | CHECK((layer4->GetType() == LayerType::MemCopy)); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 74 | |
| 75 | // Correctly use backend hint |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 76 | CHECK((layer5->GetBackendId() == Compute::CpuAcc )); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 77 | |
| 78 | // Load it into the runtime. It should pass. |
| 79 | NetworkId netId; |
| 80 | std::string ignoredErrorMessage; |
Francis Murtagh | 73d3e2e | 2021-04-29 14:23:04 +0100 | [diff] [blame] | 81 | INetworkProperties networkProperties(false, MemorySource::Malloc, MemorySource::Malloc); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 82 | runtime->LoadNetwork(netId, std::move(optNet), ignoredErrorMessage, networkProperties); |
| 83 | |
| 84 | // Creates structures for input & output |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 85 | std::vector<float> inputValue0 |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 86 | { |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 87 | 1.0f, 1.0f, 2.0f, 2.0f, 2.0f, 3.0f, 4.0f, 4.0f, 5.0f, 5.0f, 6.0f, 6.0f, 1.0f, 1.0f, 2.0f, 2.0f |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 88 | }; |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 89 | std::vector<float> inputValue1 |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 90 | { |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 91 | 0.0f, 1.0f, 1.0f, 2.0f, 3.0f, 3.0f, 3.0f, 4.0f, 4.0f, 5.0f, 5.0f, 6.0f, 0.0f, 1.0f, 1.0f, 2.0f |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 92 | }; |
| 93 | std::vector<float> inputData2 |
| 94 | { |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 95 | 12.0f, 11.0f, 10.0f, 9.0f, 8.0f, 7.0f, 6.0f, 5.0f, 4.0f, 3.0f, 2.0f, 1.0f, 12.0f, 11.0f, 10.0f, 9.0f |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 96 | }; |
| 97 | |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 98 | std::vector<float> outputData(16); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 99 | |
| 100 | std::vector<float> expectedOutput |
| 101 | { |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 102 | 11.0f, 9.0f, 7.0f, 5.0f, 3.0f, 1.0f, -1.0f, -3.0f, -5.0f, -7.0f, -9.0f, -11.0f, 11.0f, 9.0f, 7.0f, 5.0f |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 103 | }; |
| 104 | |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 105 | // Prepare aligned data |
| 106 | unsigned int numElements = info.GetNumElements(); |
| 107 | size_t totalBytes = numElements * sizeof(float); |
| 108 | const size_t alignment = 64; |
| 109 | size_t space = totalBytes + alignment + alignment; |
| 110 | auto inputData0 = std::make_unique<uint8_t[]>(space); |
| 111 | void* alignedInputPtr0 = inputData0.get(); |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 112 | CHECK(std::align(alignment, totalBytes, alignedInputPtr0, space)); |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 113 | |
| 114 | auto* intputPtr0 = reinterpret_cast<float*>(alignedInputPtr0); |
| 115 | std::copy(inputValue0.begin(), inputValue0.end(), intputPtr0); |
| 116 | |
| 117 | auto inputData1 = std::make_unique<uint8_t[]>(space); |
| 118 | void* alignedInputPtr1 = inputData1.get(); |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 119 | CHECK(std::align(alignment, totalBytes, alignedInputPtr1, space)); |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 120 | |
| 121 | auto* intputPtr1 = reinterpret_cast<float*>(alignedInputPtr1); |
| 122 | std::copy(inputValue1.begin(), inputValue1.end(), intputPtr1); |
| 123 | |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 124 | InputTensors inputTensors |
| 125 | { |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 126 | { 0, armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 0), alignedInputPtr0) }, |
| 127 | { 1, armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 1), alignedInputPtr1) }, |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 128 | { 2, armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 2), inputData2.data()) } |
| 129 | }; |
| 130 | OutputTensors outputTensors |
| 131 | { |
| 132 | { 0,armnn::Tensor(runtime->GetOutputTensorInfo(netId, 0), outputData.data()) } |
| 133 | }; |
| 134 | |
| 135 | runtime->GetProfiler(netId)->EnableProfiling(true); |
| 136 | |
| 137 | // Do the inference |
| 138 | runtime->EnqueueWorkload(netId, inputTensors, outputTensors); |
| 139 | |
| 140 | // Retrieve the Profiler.Print() output to get the workload execution |
| 141 | ProfilerManager& profilerManager = armnn::ProfilerManager::GetInstance(); |
| 142 | std::stringstream ss; |
| 143 | profilerManager.GetProfiler()->Print(ss);; |
| 144 | std::string dump = ss.str(); |
| 145 | |
| 146 | // Executed Subtraction using CpuAcc |
| 147 | std::size_t found = dump.find("NeonSubtractionWorkload_Execute"); |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 148 | CHECK(found != std::string::npos); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 149 | |
| 150 | // Contain CopyMemGeneric |
| 151 | found = dump.find("CopyMemGeneric"); |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 152 | CHECK(found != std::string::npos); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 153 | |
| 154 | // Check output is as expected |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 155 | CHECK(outputData == expectedOutput); |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 156 | |
| 157 | runtime->UnloadNetwork(netId); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 158 | } |
| 159 | |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 160 | TEST_CASE("ClImportDisabledFallbackToNeon") |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 161 | { |
| 162 | using namespace armnn; |
| 163 | |
| 164 | IRuntime::CreationOptions options; |
| 165 | IRuntimePtr runtime(IRuntime::Create(options)); |
| 166 | |
| 167 | // Builds up the structure of the network. |
| 168 | INetworkPtr net(INetwork::Create()); |
| 169 | |
| 170 | IConnectableLayer* input0 = net->AddInputLayer(0, "input0"); |
| 171 | IConnectableLayer* input1 = net->AddInputLayer(1, "input1"); |
| 172 | IConnectableLayer* input2 = net->AddInputLayer(2, "input2"); |
| 173 | IConnectableLayer* add = net->AddAdditionLayer("add"); |
| 174 | IConnectableLayer* sub = net->AddSubtractionLayer("sub"); |
| 175 | IConnectableLayer* output = net->AddOutputLayer(0, "output"); |
| 176 | |
| 177 | input0->GetOutputSlot(0).Connect(add->GetInputSlot(0)); |
| 178 | input1->GetOutputSlot(0).Connect(add->GetInputSlot(1)); |
| 179 | input2->GetOutputSlot(0).Connect(sub->GetInputSlot(0)); |
| 180 | add->GetOutputSlot(0).Connect(sub->GetInputSlot(1)); |
| 181 | sub->GetOutputSlot(0).Connect(output->GetInputSlot(0)); |
| 182 | |
| 183 | TensorInfo info = TensorInfo({ 1, 2, 3, 2 }, DataType::Float32); |
| 184 | |
| 185 | input0->GetOutputSlot(0).SetTensorInfo(info); |
| 186 | input1->GetOutputSlot(0).SetTensorInfo(info); |
| 187 | input2->GetOutputSlot(0).SetTensorInfo(info); |
| 188 | add->GetOutputSlot(0).SetTensorInfo(info); |
| 189 | sub->GetOutputSlot(0).SetTensorInfo(info); |
| 190 | |
| 191 | std::vector<BackendId> backends = { Compute::GpuAcc, Compute::CpuAcc }; |
| 192 | // Use BackendSelectionHint to specify CpuAcc for Subtraction layer |
| 193 | sub->BackendSelectionHint(backends[1]); |
| 194 | |
| 195 | // optimize the network |
| 196 | OptimizerOptions optOptions; |
| 197 | IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optOptions); |
| 198 | |
Francis Murtagh | 3d2b4b2 | 2021-02-15 18:23:17 +0000 | [diff] [blame] | 199 | Graph& graph = GetGraphForTesting(optNet.get()); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 200 | |
| 201 | armnn::Layer* const layer0 = GetFirstLayerWithName(graph, "input0"); |
| 202 | armnn::Layer* const layer1 = GetFirstLayerWithName(graph, "input1"); |
| 203 | armnn::Layer* const layer2 = GetFirstLayerWithName(graph, "input2"); |
| 204 | armnn::Layer* const layer3 = GetFirstLayerWithName(graph, "add"); |
| 205 | armnn::Layer* const layer4 = GetFirstLayerWithName(graph, "[ add (0) -> sub (1) ]"); |
| 206 | armnn::Layer* const layer5 = GetFirstLayerWithName(graph, "sub"); |
| 207 | armnn::Layer* const layer6 = GetFirstLayerWithName(graph, "output"); |
| 208 | |
| 209 | // Checks order is valid. |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 210 | CHECK(CheckOrder(graph, layer0, layer1)); |
| 211 | CHECK(CheckOrder(graph, layer1, layer2)); |
| 212 | CHECK(CheckOrder(graph, layer2, layer3)); |
| 213 | CHECK(CheckOrder(graph, layer3, layer4)); |
| 214 | CHECK(CheckOrder(graph, layer4, layer5)); |
| 215 | CHECK(CheckOrder(graph, layer5, layer6)); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 216 | |
| 217 | // Use memory import between backends |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 218 | CHECK((layer4->GetType() == LayerType::MemCopy)); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 219 | |
| 220 | // Correctly use backend hint |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 221 | CHECK((layer5->GetBackendId() == Compute::CpuAcc )); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 222 | |
| 223 | // Load it into the runtime. It should pass. |
| 224 | NetworkId netId; |
| 225 | runtime->LoadNetwork(netId, std::move(optNet)); |
| 226 | |
| 227 | // Creates structures for input & output |
| 228 | std::vector<float> inputData0 |
| 229 | { |
| 230 | 1.0f, 1.0f, 2.0f, 2.0f, 2.0f, 3.0f, 4.0f, 4.0f, 5.0f, 5.0f, 6.0f, 6.0f |
| 231 | }; |
| 232 | std::vector<float> inputData1 |
| 233 | { |
| 234 | 0.0f, 1.0f, 1.0f, 2.0f, 3.0f, 3.0f, 3.0f, 4.0f, 4.0f, 5.0f, 5.0f, 6.0f |
| 235 | }; |
| 236 | std::vector<float> inputData2 |
| 237 | { |
| 238 | 12.0f, 11.0f, 10.0f, 9.0f, 8.0f, 7.0f, 6.0f, 5.0f, 4.0f, 3.0f, 2.0f, 1.0f |
| 239 | }; |
| 240 | |
| 241 | std::vector<float> outputData(12); |
| 242 | |
| 243 | std::vector<float> expectedOutput |
| 244 | { |
| 245 | 11.0f, 9.0f, 7.0f, 5.0f, 3.0f, 1.0f, -1.0f, -3.0f, -5.0f, -7.0f, -9.0f, -11.0f |
| 246 | }; |
| 247 | |
| 248 | InputTensors inputTensors |
| 249 | { |
| 250 | { 0, armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 0), inputData0.data()) }, |
| 251 | { 1, armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 1), inputData1.data()) }, |
| 252 | { 2, armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 2), inputData2.data()) } |
| 253 | }; |
| 254 | OutputTensors outputTensors |
| 255 | { |
| 256 | { 0,armnn::Tensor(runtime->GetOutputTensorInfo(netId, 0), outputData.data()) } |
| 257 | }; |
| 258 | |
| 259 | runtime->GetProfiler(netId)->EnableProfiling(true); |
| 260 | |
| 261 | // Do the inference |
| 262 | runtime->EnqueueWorkload(netId, inputTensors, outputTensors); |
| 263 | |
| 264 | // Retrieve the Profiler.Print() output to get the workload execution |
| 265 | ProfilerManager& profilerManager = armnn::ProfilerManager::GetInstance(); |
| 266 | std::stringstream ss; |
| 267 | profilerManager.GetProfiler()->Print(ss);; |
| 268 | std::string dump = ss.str(); |
| 269 | |
| 270 | // Executed Subtraction using CpuAcc |
| 271 | std::size_t found = dump.find("NeonSubtractionWorkload_Execute"); |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 272 | CHECK(found != std::string::npos); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 273 | |
| 274 | // Contain CopyMemGeneric |
| 275 | found = dump.find("CopyMemGeneric"); |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 276 | CHECK(found != std::string::npos); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 277 | |
| 278 | // Check output is as expected |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 279 | CHECK(outputData == expectedOutput); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 280 | } |
| 281 | |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 282 | TEST_CASE("ClImportEnabledFallbackSubgraphToNeon") |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 283 | { |
| 284 | using namespace armnn; |
| 285 | |
| 286 | IRuntime::CreationOptions options; |
| 287 | IRuntimePtr runtime(IRuntime::Create(options)); |
| 288 | |
| 289 | // Builds up the structure of the network. |
| 290 | INetworkPtr net(INetwork::Create()); |
| 291 | |
| 292 | Pooling2dDescriptor desc; |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 293 | desc.m_PoolWidth = 2; |
| 294 | desc.m_PoolHeight = 2; |
| 295 | desc.m_StrideX = 2; |
| 296 | desc.m_StrideY = 2; |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 297 | |
| 298 | IConnectableLayer* input0 = net->AddInputLayer(0, "input0"); |
| 299 | IConnectableLayer* input1 = net->AddInputLayer(1, "input1"); |
| 300 | IConnectableLayer* input2 = net->AddInputLayer(2, "input2"); |
| 301 | IConnectableLayer* add = net->AddAdditionLayer("add"); |
| 302 | IConnectableLayer* sub = net->AddSubtractionLayer("sub"); |
| 303 | IConnectableLayer* pooling = net->AddPooling2dLayer(desc, "pooling"); |
| 304 | IConnectableLayer* output = net->AddOutputLayer(0, "output"); |
| 305 | |
| 306 | input0->GetOutputSlot(0).Connect(add->GetInputSlot(0)); |
| 307 | input1->GetOutputSlot(0).Connect(add->GetInputSlot(1)); |
| 308 | input2->GetOutputSlot(0).Connect(sub->GetInputSlot(0)); |
| 309 | add->GetOutputSlot(0).Connect(sub->GetInputSlot(1)); |
| 310 | sub->GetOutputSlot(0).Connect(pooling->GetInputSlot(0)); |
| 311 | pooling->GetOutputSlot(0).Connect(output->GetInputSlot(0)); |
| 312 | |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 313 | TensorInfo info = TensorInfo({ 1, 2, 4, 2 }, DataType::Float32); |
| 314 | TensorInfo poolingInfo = TensorInfo({ 1, 2, 2, 1 }, DataType::Float32); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 315 | |
| 316 | input0->GetOutputSlot(0).SetTensorInfo(info); |
| 317 | input1->GetOutputSlot(0).SetTensorInfo(info); |
| 318 | input2->GetOutputSlot(0).SetTensorInfo(info); |
| 319 | add->GetOutputSlot(0).SetTensorInfo(info); |
| 320 | sub->GetOutputSlot(0).SetTensorInfo(info); |
| 321 | pooling->GetOutputSlot(0).SetTensorInfo(poolingInfo); |
| 322 | |
| 323 | std::vector<BackendId> backends = { Compute::GpuAcc, Compute::CpuAcc }; |
| 324 | // Use BackendSelectionHint to specify CpuAcc for Subtraction layer |
| 325 | sub->BackendSelectionHint(backends[1]); |
| 326 | |
| 327 | // optimize the network |
| 328 | OptimizerOptions optOptions; |
| 329 | optOptions.m_ImportEnabled = true; |
| 330 | IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optOptions); |
| 331 | |
Francis Murtagh | 3d2b4b2 | 2021-02-15 18:23:17 +0000 | [diff] [blame] | 332 | Graph& graph = GetGraphForTesting(optNet.get()); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 333 | |
| 334 | armnn::Layer* const layer0 = GetFirstLayerWithName(graph, "input0"); |
| 335 | armnn::Layer* const layer1 = GetFirstLayerWithName(graph, "input1"); |
| 336 | armnn::Layer* const layer2 = GetFirstLayerWithName(graph, "input2"); |
| 337 | armnn::Layer* const layer3 = GetFirstLayerWithName(graph, "add"); |
| 338 | armnn::Layer* const layer4 = GetFirstLayerWithName(graph, "[ add (0) -> sub (1) ]"); |
| 339 | armnn::Layer* const layer5 = GetFirstLayerWithName(graph, "sub"); |
| 340 | armnn::Layer* const layer6 = GetFirstLayerWithName(graph, "[ sub (0) -> pooling (0) ]"); |
| 341 | armnn::Layer* const layer7 = GetFirstLayerWithName(graph, "pooling"); |
| 342 | armnn::Layer* const layer8 = GetFirstLayerWithName(graph, "output"); |
| 343 | |
| 344 | // Checks order is valid. |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 345 | CHECK(CheckOrder(graph, layer0, layer1)); |
| 346 | CHECK(CheckOrder(graph, layer1, layer2)); |
| 347 | CHECK(CheckOrder(graph, layer2, layer3)); |
| 348 | CHECK(CheckOrder(graph, layer3, layer4)); |
| 349 | CHECK(CheckOrder(graph, layer4, layer5)); |
| 350 | CHECK(CheckOrder(graph, layer5, layer6)); |
| 351 | CHECK(CheckOrder(graph, layer6, layer7)); |
| 352 | CHECK(CheckOrder(graph, layer7, layer8)); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 353 | |
| 354 | // Use memory import between backends |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 355 | CHECK((layer4->GetType() == LayerType::MemCopy)); |
| 356 | CHECK((layer6->GetType() == LayerType::MemCopy)); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 357 | |
| 358 | // Correctly use backend hint |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 359 | CHECK((layer5->GetBackendId() == Compute::CpuAcc )); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 360 | |
| 361 | // Load it into the runtime. It should pass. |
| 362 | NetworkId netId; |
| 363 | std::string ignoredErrorMessage; |
Francis Murtagh | 73d3e2e | 2021-04-29 14:23:04 +0100 | [diff] [blame] | 364 | INetworkProperties networkProperties(false, MemorySource::Malloc, MemorySource::Malloc); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 365 | runtime->LoadNetwork(netId, std::move(optNet), ignoredErrorMessage, networkProperties); |
| 366 | |
| 367 | // Creates structures for input & output |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 368 | std::vector<float> inputValue0 |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 369 | { |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 370 | 1.0f, 1.0f, 2.0f, 2.0f, 2.0f, 3.0f, 4.0f, 4.0f, 5.0f, 5.0f, 6.0f, 6.0f, 1.0f, 1.0f, 2.0f, 2.0f |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 371 | }; |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 372 | std::vector<float> inputValue1 |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 373 | { |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 374 | 0.0f, 1.0f, 1.0f, 2.0f, 3.0f, 3.0f, 3.0f, 4.0f, 4.0f, 5.0f, 5.0f, 6.0f, 0.0f, 1.0f, 1.0f, 2.0f |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 375 | }; |
| 376 | std::vector<float> inputData2 |
| 377 | { |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 378 | 12.0f, 11.0f, 10.0f, 9.0f, 8.0f, 7.0f, 6.0f, 5.0f, 4.0f, 3.0f, 2.0f, 1.0f, 12.0f, 11.0f, 10.0f, 9.0f |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 379 | }; |
| 380 | |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 381 | std::vector<float> outputData(4); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 382 | |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 383 | std::vector<float> expectedOutput{ 11.0f, 3.0f, -5.0f, 11.0f }; |
| 384 | |
| 385 | unsigned int numElements = info.GetNumElements(); |
| 386 | size_t totalBytes = numElements * sizeof(float); |
| 387 | const size_t alignment = 64; |
| 388 | size_t space = totalBytes + alignment + alignment; |
| 389 | auto inputData0 = std::make_unique<uint8_t[]>(space); |
| 390 | void* alignedInputPtr0 = inputData0.get(); |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 391 | CHECK(std::align(alignment, totalBytes, alignedInputPtr0, space)); |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 392 | |
| 393 | auto* intputPtr0 = reinterpret_cast<float*>(alignedInputPtr0); |
| 394 | std::copy(inputValue0.begin(), inputValue0.end(), intputPtr0); |
| 395 | |
| 396 | auto inputData1 = std::make_unique<uint8_t[]>(space); |
| 397 | void* alignedInputPtr1 = inputData1.get(); |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 398 | CHECK(std::align(alignment, totalBytes, alignedInputPtr1, space)); |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 399 | |
| 400 | auto* intputPtr1 = reinterpret_cast<float*>(alignedInputPtr1); |
| 401 | std::copy(inputValue1.begin(), inputValue1.end(), intputPtr1); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 402 | |
| 403 | InputTensors inputTensors |
| 404 | { |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 405 | { 0, armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 0), alignedInputPtr0) }, |
| 406 | { 1, armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 1), alignedInputPtr1) }, |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 407 | { 2, armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 2), inputData2.data()) } |
| 408 | }; |
| 409 | OutputTensors outputTensors |
| 410 | { |
| 411 | { 0,armnn::Tensor(runtime->GetOutputTensorInfo(netId, 0), outputData.data()) } |
| 412 | }; |
| 413 | |
| 414 | runtime->GetProfiler(netId)->EnableProfiling(true); |
| 415 | |
| 416 | // Do the inference |
| 417 | runtime->EnqueueWorkload(netId, inputTensors, outputTensors); |
| 418 | |
| 419 | // Retrieve the Profiler.Print() output to get the workload execution |
| 420 | ProfilerManager& profilerManager = armnn::ProfilerManager::GetInstance(); |
| 421 | std::stringstream ss; |
| 422 | profilerManager.GetProfiler()->Print(ss);; |
| 423 | std::string dump = ss.str(); |
| 424 | |
| 425 | // Executed Subtraction using CpuAcc |
| 426 | std::size_t found = dump.find("NeonSubtractionWorkload_Execute"); |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 427 | CHECK(found != std::string::npos); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 428 | |
| 429 | // Correctly switch back to GpuAcc |
| 430 | found = dump.find("ClPooling2dWorkload_Execute"); |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 431 | CHECK(found != std::string::npos); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 432 | |
| 433 | // Contain CopyMemGeneric |
| 434 | found = dump.find("CopyMemGeneric"); |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 435 | CHECK(found != std::string::npos); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 436 | |
| 437 | // Check output is as expected |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 438 | CHECK(outputData == expectedOutput); |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 439 | |
| 440 | runtime->UnloadNetwork(netId); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 441 | } |
| 442 | |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 443 | TEST_CASE("ClImportDisableFallbackSubgraphToNeon") |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 444 | { |
| 445 | using namespace armnn; |
| 446 | |
| 447 | IRuntime::CreationOptions options; |
| 448 | IRuntimePtr runtime(IRuntime::Create(options)); |
| 449 | |
| 450 | // Builds up the structure of the network. |
| 451 | INetworkPtr net(INetwork::Create()); |
| 452 | |
| 453 | Pooling2dDescriptor desc; |
| 454 | |
| 455 | IConnectableLayer* input0 = net->AddInputLayer(0, "input0"); |
| 456 | IConnectableLayer* input1 = net->AddInputLayer(1, "input1"); |
| 457 | IConnectableLayer* input2 = net->AddInputLayer(2, "input2"); |
| 458 | IConnectableLayer* add = net->AddAdditionLayer("add"); |
| 459 | IConnectableLayer* sub = net->AddSubtractionLayer("sub"); |
| 460 | IConnectableLayer* pooling = net->AddPooling2dLayer(desc, "pooling"); |
| 461 | IConnectableLayer* output = net->AddOutputLayer(0, "output"); |
| 462 | |
| 463 | input0->GetOutputSlot(0).Connect(add->GetInputSlot(0)); |
| 464 | input1->GetOutputSlot(0).Connect(add->GetInputSlot(1)); |
| 465 | input2->GetOutputSlot(0).Connect(sub->GetInputSlot(0)); |
| 466 | add->GetOutputSlot(0).Connect(sub->GetInputSlot(1)); |
| 467 | sub->GetOutputSlot(0).Connect(pooling->GetInputSlot(0)); |
| 468 | pooling->GetOutputSlot(0).Connect(output->GetInputSlot(0)); |
| 469 | |
| 470 | TensorInfo info = TensorInfo({ 1, 2, 3, 2 }, DataType::Float32); |
| 471 | TensorInfo poolingInfo = TensorInfo({ 1, 2, 1, 1 }, DataType::Float32); |
| 472 | |
| 473 | input0->GetOutputSlot(0).SetTensorInfo(info); |
| 474 | input1->GetOutputSlot(0).SetTensorInfo(info); |
| 475 | input2->GetOutputSlot(0).SetTensorInfo(info); |
| 476 | add->GetOutputSlot(0).SetTensorInfo(info); |
| 477 | sub->GetOutputSlot(0).SetTensorInfo(info); |
| 478 | pooling->GetOutputSlot(0).SetTensorInfo(poolingInfo); |
| 479 | |
| 480 | std::vector<BackendId> backends = { Compute::GpuAcc, Compute::CpuAcc }; |
| 481 | // Use BackendSelectionHint to specify CpuAcc for Subtraction layer |
| 482 | sub->BackendSelectionHint(backends[1]); |
| 483 | |
| 484 | // optimize the network |
| 485 | OptimizerOptions optOptions; |
| 486 | IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optOptions); |
| 487 | |
Francis Murtagh | 3d2b4b2 | 2021-02-15 18:23:17 +0000 | [diff] [blame] | 488 | Graph& graph = GetGraphForTesting(optNet.get()); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 489 | |
| 490 | armnn::Layer* const layer0 = GetFirstLayerWithName(graph, "input0"); |
| 491 | armnn::Layer* const layer1 = GetFirstLayerWithName(graph, "input1"); |
| 492 | armnn::Layer* const layer2 = GetFirstLayerWithName(graph, "input2"); |
| 493 | armnn::Layer* const layer3 = GetFirstLayerWithName(graph, "add"); |
| 494 | armnn::Layer* const layer4 = GetFirstLayerWithName(graph, "[ add (0) -> sub (1) ]"); |
| 495 | armnn::Layer* const layer5 = GetFirstLayerWithName(graph, "sub"); |
| 496 | armnn::Layer* const layer6 = GetFirstLayerWithName(graph, "[ sub (0) -> pooling (0) ]"); |
| 497 | armnn::Layer* const layer7 = GetFirstLayerWithName(graph, "pooling"); |
| 498 | armnn::Layer* const layer8 = GetFirstLayerWithName(graph, "output"); |
| 499 | |
| 500 | // Checks order is valid. |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 501 | CHECK(CheckOrder(graph, layer0, layer1)); |
| 502 | CHECK(CheckOrder(graph, layer1, layer2)); |
| 503 | CHECK(CheckOrder(graph, layer2, layer3)); |
| 504 | CHECK(CheckOrder(graph, layer3, layer4)); |
| 505 | CHECK(CheckOrder(graph, layer4, layer5)); |
| 506 | CHECK(CheckOrder(graph, layer5, layer6)); |
| 507 | CHECK(CheckOrder(graph, layer6, layer7)); |
| 508 | CHECK(CheckOrder(graph, layer7, layer8)); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 509 | |
| 510 | // Use memory import between backends |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 511 | CHECK((layer4->GetType() == LayerType::MemCopy)); |
| 512 | CHECK((layer6->GetType() == LayerType::MemCopy)); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 513 | |
| 514 | // Correctly use backend hint |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 515 | CHECK((layer5->GetBackendId() == Compute::CpuAcc )); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 516 | |
| 517 | // Load it into the runtime. It should pass. |
| 518 | NetworkId netId; |
| 519 | runtime->LoadNetwork(netId, std::move(optNet)); |
| 520 | |
| 521 | // Creates structures for input & output |
| 522 | std::vector<float> inputData0 |
| 523 | { |
| 524 | 1.0f, 1.0f, 2.0f, 2.0f, 2.0f, 3.0f, 4.0f, 4.0f, 5.0f, 5.0f, 6.0f, 6.0f |
| 525 | }; |
| 526 | std::vector<float> inputData1 |
| 527 | { |
| 528 | 0.0f, 1.0f, 1.0f, 2.0f, 3.0f, 3.0f, 3.0f, 4.0f, 4.0f, 5.0f, 5.0f, 6.0f |
| 529 | }; |
| 530 | std::vector<float> inputData2 |
| 531 | { |
| 532 | 12.0f, 11.0f, 10.0f, 9.0f, 8.0f, 7.0f, 6.0f, 5.0f, 4.0f, 3.0f, 2.0f, 1.0f |
| 533 | }; |
| 534 | |
| 535 | std::vector<float> outputData(2); |
| 536 | |
| 537 | std::vector<float> expectedOutput{ 11.0f, -1.0f }; |
| 538 | |
| 539 | InputTensors inputTensors |
| 540 | { |
| 541 | { 0, armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 0), inputData0.data()) }, |
| 542 | { 1, armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 1), inputData1.data()) }, |
| 543 | { 2, armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 2), inputData2.data()) } |
| 544 | }; |
| 545 | OutputTensors outputTensors |
| 546 | { |
| 547 | { 0,armnn::Tensor(runtime->GetOutputTensorInfo(netId, 0), outputData.data()) } |
| 548 | }; |
| 549 | |
| 550 | runtime->GetProfiler(netId)->EnableProfiling(true); |
| 551 | |
| 552 | // Do the inference |
| 553 | runtime->EnqueueWorkload(netId, inputTensors, outputTensors); |
| 554 | |
| 555 | // Retrieve the Profiler.Print() output to get the workload execution |
| 556 | ProfilerManager& profilerManager = armnn::ProfilerManager::GetInstance(); |
| 557 | std::stringstream ss; |
| 558 | profilerManager.GetProfiler()->Print(ss);; |
| 559 | std::string dump = ss.str(); |
| 560 | |
| 561 | // Executed Subtraction using CpuAcc |
| 562 | std::size_t found = dump.find("NeonSubtractionWorkload_Execute"); |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 563 | CHECK(found != std::string::npos); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 564 | |
| 565 | // Correctly switch back to GpuAcc |
| 566 | found = dump.find("ClPooling2dWorkload_Execute"); |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 567 | CHECK(found != std::string::npos); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 568 | |
| 569 | // Contain CopyMemGeneric |
| 570 | found = dump.find("CopyMemGeneric"); |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 571 | CHECK(found != std::string::npos); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 572 | |
| 573 | // Check output is as expected |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 574 | CHECK(outputData == expectedOutput); |
Narumol Prangnawarat | 265e53e | 2020-10-30 16:06:55 +0000 | [diff] [blame] | 575 | } |
| 576 | |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 577 | } |