Narumol Prangnawarat | 1a26896 | 2020-07-27 15:52:13 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2020 Arm Ltd and Contributors. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
Sadik Armagan | 76615a5 | 2020-08-04 14:01:05 +0100 | [diff] [blame] | 5 | |
| 6 | #include <Graph.hpp> |
| 7 | #include <Network.hpp> |
| 8 | |
Narumol Prangnawarat | 1a26896 | 2020-07-27 15:52:13 +0100 | [diff] [blame] | 9 | #include <neon/NeonTensorHandle.hpp> |
| 10 | #include <neon/NeonTensorHandleFactory.hpp> |
| 11 | |
Sadik Armagan | 76615a5 | 2020-08-04 14:01:05 +0100 | [diff] [blame] | 12 | #include <armnn/utility/PolymorphicDowncast.hpp> |
| 13 | |
| 14 | #include <test/GraphUtils.hpp> |
Narumol Prangnawarat | b8d771a | 2020-08-14 11:51:12 +0100 | [diff] [blame] | 15 | #include <arm_compute/runtime/Allocator.h> |
Sadik Armagan | 76615a5 | 2020-08-04 14:01:05 +0100 | [diff] [blame] | 16 | |
Narumol Prangnawarat | 1a26896 | 2020-07-27 15:52:13 +0100 | [diff] [blame] | 17 | #include <boost/test/unit_test.hpp> |
| 18 | |
| 19 | BOOST_AUTO_TEST_SUITE(NeonTensorHandleTests) |
| 20 | using namespace armnn; |
| 21 | |
| 22 | BOOST_AUTO_TEST_CASE(NeonTensorHandleGetCapabilitiesNoPadding) |
| 23 | { |
| 24 | std::shared_ptr<NeonMemoryManager> memoryManager = std::make_shared<NeonMemoryManager>(); |
| 25 | NeonTensorHandleFactory handleFactory(memoryManager); |
| 26 | |
| 27 | INetworkPtr network(INetwork::Create()); |
| 28 | |
| 29 | // Add the layers |
| 30 | IConnectableLayer* input = network->AddInputLayer(0); |
| 31 | SoftmaxDescriptor descriptor; |
| 32 | descriptor.m_Beta = 1.0f; |
| 33 | IConnectableLayer* softmax = network->AddSoftmaxLayer(descriptor); |
| 34 | IConnectableLayer* output = network->AddOutputLayer(2); |
| 35 | |
| 36 | // Establish connections |
| 37 | input->GetOutputSlot(0).Connect(softmax->GetInputSlot(0)); |
| 38 | softmax->GetOutputSlot(0).Connect(output->GetInputSlot(0)); |
| 39 | |
| 40 | // No padding required for input |
| 41 | std::vector<Capability> capabilities = handleFactory.GetCapabilities(input, |
| 42 | softmax, |
| 43 | CapabilityClass::PaddingRequired); |
| 44 | BOOST_TEST(capabilities.empty()); |
| 45 | |
| 46 | // No padding required for Softmax |
| 47 | capabilities = handleFactory.GetCapabilities(softmax, output, CapabilityClass::PaddingRequired); |
| 48 | BOOST_TEST(capabilities.empty()); |
| 49 | |
| 50 | // No padding required for output |
| 51 | capabilities = handleFactory.GetCapabilities(output, nullptr, CapabilityClass::PaddingRequired); |
| 52 | BOOST_TEST(capabilities.empty()); |
| 53 | } |
| 54 | |
| 55 | BOOST_AUTO_TEST_CASE(NeonTensorHandleGetCapabilitiesPadding) |
| 56 | { |
| 57 | std::shared_ptr<NeonMemoryManager> memoryManager = std::make_shared<NeonMemoryManager>(); |
| 58 | NeonTensorHandleFactory handleFactory(memoryManager); |
| 59 | |
| 60 | INetworkPtr network(INetwork::Create()); |
| 61 | |
| 62 | // Add the layers |
| 63 | IConnectableLayer* input = network->AddInputLayer(0); |
| 64 | Pooling2dDescriptor descriptor; |
| 65 | IConnectableLayer* pooling = network->AddPooling2dLayer(descriptor); |
| 66 | IConnectableLayer* output = network->AddOutputLayer(2); |
| 67 | |
| 68 | // Establish connections |
| 69 | input->GetOutputSlot(0).Connect(pooling->GetInputSlot(0)); |
| 70 | pooling->GetOutputSlot(0).Connect(output->GetInputSlot(0)); |
| 71 | |
| 72 | // No padding required for input |
| 73 | std::vector<Capability> capabilities = handleFactory.GetCapabilities(input, |
| 74 | pooling, |
| 75 | CapabilityClass::PaddingRequired); |
| 76 | BOOST_TEST(capabilities.empty()); |
| 77 | |
| 78 | // No padding required for output |
| 79 | capabilities = handleFactory.GetCapabilities(output, nullptr, CapabilityClass::PaddingRequired); |
| 80 | BOOST_TEST(capabilities.empty()); |
| 81 | |
| 82 | // Padding required for Pooling2d |
| 83 | capabilities = handleFactory.GetCapabilities(pooling, output, CapabilityClass::PaddingRequired); |
| 84 | BOOST_TEST(capabilities.size() == 1); |
| 85 | BOOST_TEST((capabilities[0].m_CapabilityClass == CapabilityClass::PaddingRequired)); |
| 86 | BOOST_TEST(capabilities[0].m_Value); |
| 87 | } |
| 88 | |
Sadik Armagan | 76615a5 | 2020-08-04 14:01:05 +0100 | [diff] [blame] | 89 | BOOST_AUTO_TEST_CASE(ConcatOnXorYSubTensorsNoPaddinRequiredTest) |
| 90 | { |
| 91 | armnn::INetworkPtr net(armnn::INetwork::Create()); |
| 92 | |
| 93 | // Set up tensor infos |
| 94 | const armnn::TensorInfo inputInfo = armnn::TensorInfo({2, 3, 2, 2}, armnn::DataType::Float32); |
| 95 | const armnn::TensorInfo intermediateInfo = armnn::TensorInfo({2, 3, 2, 2}, armnn::DataType::Float32); |
| 96 | const armnn::TensorInfo outputInfo = armnn::TensorInfo({2, 3, 4, 2}, armnn::DataType::Float32); |
| 97 | |
| 98 | armnn::ElementwiseUnaryDescriptor descriptor(armnn::UnaryOperation::Abs); |
| 99 | |
| 100 | // Create the network |
| 101 | armnn::IConnectableLayer* const input0Layer = net->AddInputLayer(0, "input_0"); |
| 102 | input0Layer->GetOutputSlot(0).SetTensorInfo(inputInfo); |
| 103 | armnn::IConnectableLayer* elementwiseUnaryLayer0 = net->AddElementwiseUnaryLayer(descriptor, "elementwiseUnary_0"); |
| 104 | elementwiseUnaryLayer0->GetOutputSlot(0).SetTensorInfo(intermediateInfo); |
| 105 | input0Layer->GetOutputSlot(0).Connect(elementwiseUnaryLayer0->GetInputSlot(0)); |
| 106 | |
| 107 | armnn::IConnectableLayer* const input1Layer = net->AddInputLayer(1, "input_1"); |
| 108 | input1Layer->GetOutputSlot(0).SetTensorInfo(inputInfo); |
| 109 | armnn::IConnectableLayer* elementwiseUnaryLayer1 = net->AddElementwiseUnaryLayer(descriptor, "elementwiseUnary_1"); |
| 110 | elementwiseUnaryLayer1->GetOutputSlot(0).SetTensorInfo(intermediateInfo); |
| 111 | input1Layer->GetOutputSlot(0).Connect(elementwiseUnaryLayer1->GetInputSlot(0)); |
| 112 | |
| 113 | std::array<armnn::TensorShape, 2> concatInputShapes = { intermediateInfo.GetShape(), intermediateInfo.GetShape() }; |
| 114 | armnn::IConnectableLayer* const concatLayer = net->AddConcatLayer(armnn::CreateDescriptorForConcatenation( |
| 115 | concatInputShapes.begin(), concatInputShapes.end(), 2), "concatenation"); |
| 116 | concatLayer->GetOutputSlot(0).SetTensorInfo(outputInfo); |
| 117 | elementwiseUnaryLayer0->GetOutputSlot(0).Connect(concatLayer->GetInputSlot(0)); |
| 118 | elementwiseUnaryLayer1->GetOutputSlot(0).Connect(concatLayer->GetInputSlot(1)); |
| 119 | |
| 120 | armnn::IConnectableLayer* const outputLayer = net->AddOutputLayer(0, "output"); |
| 121 | concatLayer->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0)); |
| 122 | |
| 123 | armnn::IRuntime::CreationOptions options; |
| 124 | armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options)); |
| 125 | |
| 126 | std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc }; |
| 127 | armnn::IOptimizedNetworkPtr optimizedNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec()); |
| 128 | |
| 129 | const armnn::Graph& theGraph = static_cast<armnn::OptimizedNetwork*>(optimizedNet.get())->GetGraph(); |
| 130 | |
| 131 | // Load graph into runtime |
| 132 | armnn::NetworkId networkIdentifier; |
| 133 | runtime->LoadNetwork(networkIdentifier, std::move(optimizedNet)); |
| 134 | |
| 135 | // now check the concat how many sub-tensors it is using.. |
| 136 | auto TraceSubTensorHandleAncestry = [](armnn::ITensorHandle* const subTensorHandle) |
| 137 | { |
| 138 | if (subTensorHandle && subTensorHandle->GetParent()) |
| 139 | { |
| 140 | return true; |
| 141 | } |
| 142 | return false; |
| 143 | }; |
| 144 | |
| 145 | for (auto&& layer : theGraph) |
| 146 | { |
| 147 | if(layer->GetType() == armnn::LayerType::Concat) |
| 148 | { |
| 149 | unsigned int numberOfSubTensors = 0; |
| 150 | for (unsigned int i = 0; i < layer->GetNumInputSlots(); ++i) |
| 151 | { |
| 152 | const armnn::OutputSlot* slot = layer->GetInputSlot(i).GetConnectedOutputSlot(); |
| 153 | if (TraceSubTensorHandleAncestry(slot->GetOutputHandler().GetData())) |
| 154 | { |
| 155 | ++numberOfSubTensors; |
| 156 | } |
| 157 | } |
| 158 | // sub-tensors should be supported in this configuration |
| 159 | BOOST_CHECK(numberOfSubTensors > 0); |
| 160 | } |
| 161 | } |
| 162 | } |
| 163 | |
Narumol Prangnawarat | b8d771a | 2020-08-14 11:51:12 +0100 | [diff] [blame] | 164 | BOOST_AUTO_TEST_CASE(NeonTensorHandleFactoryMemoryManaged) |
| 165 | { |
| 166 | std::shared_ptr<NeonMemoryManager> memoryManager = std::make_shared<NeonMemoryManager>( |
| 167 | std::make_unique<arm_compute::Allocator>(), |
| 168 | BaseMemoryManager::MemoryAffinity::Offset); |
| 169 | NeonTensorHandleFactory handleFactory(memoryManager); |
| 170 | TensorInfo info({ 1, 1, 2, 1 }, DataType::Float32); |
| 171 | |
| 172 | // create TensorHandle with memory managed |
| 173 | auto handle = handleFactory.CreateTensorHandle(info, true); |
| 174 | handle->Manage(); |
| 175 | handle->Allocate(); |
| 176 | |
| 177 | memoryManager->Acquire(); |
| 178 | { |
| 179 | float* buffer = reinterpret_cast<float*>(handle->Map()); |
| 180 | BOOST_CHECK(buffer != nullptr); // Yields a valid pointer |
| 181 | buffer[0] = 1.5f; |
| 182 | buffer[1] = 2.5f; |
| 183 | BOOST_CHECK(buffer[0] == 1.5f); // Memory is writable and readable |
| 184 | BOOST_CHECK(buffer[1] == 2.5f); // Memory is writable and readable |
| 185 | } |
| 186 | memoryManager->Release(); |
| 187 | |
| 188 | memoryManager->Acquire(); |
| 189 | { |
| 190 | float* buffer = reinterpret_cast<float*>(handle->Map()); |
| 191 | BOOST_CHECK(buffer != nullptr); // Yields a valid pointer |
| 192 | buffer[0] = 3.5f; |
| 193 | buffer[1] = 4.5f; |
| 194 | BOOST_CHECK(buffer[0] == 3.5f); // Memory is writable and readable |
| 195 | BOOST_CHECK(buffer[1] == 4.5f); // Memory is writable and readable |
| 196 | } |
| 197 | memoryManager->Release(); |
| 198 | |
| 199 | float testPtr[2] = { 2.5f, 5.5f }; |
| 200 | // Cannot import as import is disabled |
| 201 | BOOST_CHECK(!handle->Import(static_cast<void*>(testPtr), MemorySource::Malloc)); |
| 202 | } |
| 203 | |
| 204 | BOOST_AUTO_TEST_CASE(NeonTensorHandleFactoryImport) |
| 205 | { |
| 206 | std::shared_ptr<NeonMemoryManager> memoryManager = std::make_shared<NeonMemoryManager>( |
| 207 | std::make_unique<arm_compute::Allocator>(), |
| 208 | BaseMemoryManager::MemoryAffinity::Offset); |
| 209 | NeonTensorHandleFactory handleFactory(memoryManager); |
| 210 | TensorInfo info({ 1, 1, 2, 1 }, DataType::Float32); |
| 211 | |
| 212 | // create TensorHandle without memory managed |
| 213 | auto handle = handleFactory.CreateTensorHandle(info, false); |
| 214 | handle->Manage(); |
| 215 | handle->Allocate(); |
| 216 | memoryManager->Acquire(); |
| 217 | |
| 218 | // No buffer allocated when import is enabled |
| 219 | BOOST_CHECK((PolymorphicDowncast<NeonTensorHandle*>(handle.get()))->GetTensor().buffer() == nullptr); |
| 220 | |
| 221 | float testPtr[2] = { 2.5f, 5.5f }; |
| 222 | // Correctly import |
| 223 | BOOST_CHECK(handle->Import(static_cast<void*>(testPtr), MemorySource::Malloc)); |
| 224 | float* buffer = reinterpret_cast<float*>(handle->Map()); |
| 225 | BOOST_CHECK(buffer != nullptr); // Yields a valid pointer after import |
| 226 | BOOST_CHECK(buffer == testPtr); // buffer is pointing to testPtr |
| 227 | // Memory is writable and readable with correct value |
| 228 | BOOST_CHECK(buffer[0] == 2.5f); |
| 229 | BOOST_CHECK(buffer[1] == 5.5f); |
| 230 | buffer[0] = 3.5f; |
| 231 | buffer[1] = 10.0f; |
| 232 | BOOST_CHECK(buffer[0] == 3.5f); |
| 233 | BOOST_CHECK(buffer[1] == 10.0f); |
| 234 | memoryManager->Release(); |
| 235 | } |
| 236 | |
Narumol Prangnawarat | 1a26896 | 2020-07-27 15:52:13 +0100 | [diff] [blame] | 237 | BOOST_AUTO_TEST_SUITE_END() |