Aron Virginas-Tar | 7010400 | 2018-10-24 15:33:28 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2017 Arm Ltd. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
Aron Virginas-Tar | 5605519 | 2018-11-12 18:10:43 +0000 | [diff] [blame] | 6 | #include "ClWorkloadFactoryHelper.hpp" |
| 7 | |
Aron Virginas-Tar | 7010400 | 2018-10-24 15:33:28 +0100 | [diff] [blame] | 8 | #include <armnn/ArmNN.hpp> |
Aron Virginas-Tar | c9cc804 | 2018-11-01 16:15:57 +0000 | [diff] [blame] | 9 | #include <Network.hpp> |
Aron Virginas-Tar | 7010400 | 2018-10-24 15:33:28 +0100 | [diff] [blame] | 10 | |
Aron Virginas-Tar | c9cc804 | 2018-11-01 16:15:57 +0000 | [diff] [blame] | 11 | #include <test/GraphUtils.hpp> |
Aron Virginas-Tar | 7010400 | 2018-10-24 15:33:28 +0100 | [diff] [blame] | 12 | |
Aron Virginas-Tar | c9cc804 | 2018-11-01 16:15:57 +0000 | [diff] [blame] | 13 | #include <cl/ClWorkloadFactory.hpp> |
Aron Virginas-Tar | 7010400 | 2018-10-24 15:33:28 +0100 | [diff] [blame] | 14 | |
| 15 | #include <boost/test/unit_test.hpp> |
| 16 | |
| 17 | BOOST_AUTO_TEST_SUITE(ClOptimizedNetwork) |
| 18 | |
| 19 | BOOST_AUTO_TEST_CASE(OptimizeValidateGpuDeviceSupportLayerNoFallback) |
| 20 | { |
| 21 | // build up the structure of the network |
| 22 | armnn::INetworkPtr net(armnn::INetwork::Create()); |
| 23 | |
| 24 | armnn::IConnectableLayer* input = net->AddInputLayer(0); |
| 25 | armnn::IConnectableLayer* output = net->AddOutputLayer(0); |
| 26 | |
| 27 | input->GetOutputSlot(0).Connect(output->GetInputSlot(0)); |
| 28 | input->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 1, 4, 4 }, armnn::DataType::Float32)); |
| 29 | |
| 30 | armnn::IRuntime::CreationOptions options; |
| 31 | armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options)); |
| 32 | |
| 33 | std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc }; |
| 34 | armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec()); |
| 35 | BOOST_CHECK(optNet); |
| 36 | // validate workloads |
Aron Virginas-Tar | 5caf907 | 2018-11-14 18:35:18 +0000 | [diff] [blame] | 37 | armnn::ClWorkloadFactory fact = |
| 38 | ClWorkloadFactoryHelper::GetFactory(ClWorkloadFactoryHelper::GetMemoryManager()); |
Aron Virginas-Tar | 7010400 | 2018-10-24 15:33:28 +0100 | [diff] [blame] | 39 | for (auto&& layer : static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph()) |
| 40 | { |
| 41 | BOOST_CHECK(layer->GetBackendId() == armnn::Compute::GpuAcc); |
| 42 | BOOST_CHECK_NO_THROW( |
| 43 | layer->CreateWorkload(static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph(), fact)); |
| 44 | } |
| 45 | } |
| 46 | |
| 47 | BOOST_AUTO_TEST_CASE(FP16TurboModeTestOnGpuAcc) |
| 48 | { |
| 49 | // Test to check when Fp16 Turbo mode set |
| 50 | // it converts the Fp32 network to Fp16 Network |
| 51 | // add Fp32ToFp16 conversion layer after the InputLayer |
| 52 | // add Fp16ToFp32 conversion layer after the OutputLayer |
| 53 | // checks the other layers if they are supported in Fp16 |
| 54 | // if they are not put the conversion layers before and after |
| 55 | // if they are not supported in Fp16 use Fp32 instead |
| 56 | // if there are inverse conversion layers remove them with optimization |
| 57 | // at the moment FloorLayer is not supported in Fp16 so it rolls back to Fp32 |
| 58 | // and inverse conversion layers are removed by the optimizer |
| 59 | armnn::Network net; |
| 60 | |
| 61 | // Defines layers. |
| 62 | auto input = net.AddInputLayer(0, "input layer"); |
| 63 | // ReLu1 |
| 64 | armnn::ActivationDescriptor activation1Descriptor; |
| 65 | activation1Descriptor.m_Function = armnn::ActivationFunction::BoundedReLu; |
| 66 | activation1Descriptor.m_A = 1.f; |
| 67 | activation1Descriptor.m_B = -1.f; |
| 68 | auto activation = net.AddActivationLayer(activation1Descriptor, "activation layer"); |
| 69 | auto output = net.AddOutputLayer(0, "output layer"); |
| 70 | |
| 71 | // Connects layers. |
| 72 | input->GetOutputSlot(0).Connect(activation->GetInputSlot(0)); |
| 73 | activation->GetOutputSlot(0).Connect(output->GetInputSlot(0)); |
| 74 | |
| 75 | armnn::TensorShape shape({4}); |
| 76 | armnn::TensorInfo info(shape, armnn::DataType::Float32); |
| 77 | input->GetOutputSlot(0).SetTensorInfo(info); |
| 78 | activation->GetOutputSlot(0).SetTensorInfo(info); |
| 79 | |
| 80 | armnn::IRuntime::CreationOptions options; |
| 81 | armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options)); |
| 82 | |
| 83 | std::vector<armnn::BackendId> backends = {armnn::Compute::GpuAcc}; |
| 84 | |
| 85 | armnn::OptimizerOptions optimizerOptions; |
| 86 | optimizerOptions.m_ReduceFp32ToFp16 = true; |
| 87 | |
| 88 | armnn::IOptimizedNetworkPtr optimizedNet = armnn::Optimize( |
| 89 | net, backends, runtime->GetDeviceSpec(), optimizerOptions); |
| 90 | |
| 91 | const armnn::Graph& graph = static_cast<armnn::OptimizedNetwork*>(optimizedNet.get())->GetGraph(); |
| 92 | |
| 93 | // Tests that all layers are present in the graph. |
| 94 | BOOST_TEST(graph.GetNumLayers() == 5); |
| 95 | |
| 96 | // Tests that the vertices exist and have correct names. |
| 97 | BOOST_TEST(GraphHasNamedLayer(graph, "input layer")); |
| 98 | BOOST_TEST(GraphHasNamedLayer(graph, "convert_fp32_to_fp16-0-input layer")); |
| 99 | BOOST_TEST(GraphHasNamedLayer(graph, "activation layer")); |
| 100 | BOOST_TEST(GraphHasNamedLayer(graph, "convert_fp16_to_fp32-0-output layer")); |
| 101 | BOOST_TEST(GraphHasNamedLayer(graph, "output layer")); |
| 102 | } |
| 103 | |
| 104 | BOOST_AUTO_TEST_SUITE_END(); |