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 | c9cc804 | 2018-11-01 16:15:57 +0000 | [diff] [blame] | 8 | #include <Network.hpp> |
Aron Virginas-Tar | 7010400 | 2018-10-24 15:33:28 +0100 | [diff] [blame] | 9 | |
Aron Virginas-Tar | c9cc804 | 2018-11-01 16:15:57 +0000 | [diff] [blame] | 10 | #include <test/GraphUtils.hpp> |
Aron Virginas-Tar | 7010400 | 2018-10-24 15:33:28 +0100 | [diff] [blame] | 11 | |
Aron Virginas-Tar | c9cc804 | 2018-11-01 16:15:57 +0000 | [diff] [blame] | 12 | #include <cl/ClWorkloadFactory.hpp> |
Finn Williams | 4064632 | 2021-02-11 16:16:42 +0000 | [diff] [blame] | 13 | #include <cl/ClBackendContext.hpp> |
| 14 | |
| 15 | #include <Filesystem.hpp> |
| 16 | |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 17 | #include <doctest/doctest.h> |
Aron Virginas-Tar | 7010400 | 2018-10-24 15:33:28 +0100 | [diff] [blame] | 18 | |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 19 | TEST_SUITE("ClOptimizedNetwork") |
| 20 | { |
| 21 | TEST_CASE("OptimizeValidateGpuDeviceSupportLayerNoFallback") |
Aron Virginas-Tar | 7010400 | 2018-10-24 15:33:28 +0100 | [diff] [blame] | 22 | { |
| 23 | // build up the structure of the network |
| 24 | armnn::INetworkPtr net(armnn::INetwork::Create()); |
| 25 | |
| 26 | armnn::IConnectableLayer* input = net->AddInputLayer(0); |
| 27 | armnn::IConnectableLayer* output = net->AddOutputLayer(0); |
| 28 | |
| 29 | input->GetOutputSlot(0).Connect(output->GetInputSlot(0)); |
| 30 | input->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 1, 4, 4 }, armnn::DataType::Float32)); |
| 31 | |
| 32 | armnn::IRuntime::CreationOptions options; |
| 33 | armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options)); |
| 34 | |
| 35 | std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc }; |
| 36 | armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec()); |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 37 | CHECK(optNet); |
Aron Virginas-Tar | 7010400 | 2018-10-24 15:33:28 +0100 | [diff] [blame] | 38 | // validate workloads |
Aron Virginas-Tar | 5caf907 | 2018-11-14 18:35:18 +0000 | [diff] [blame] | 39 | armnn::ClWorkloadFactory fact = |
| 40 | ClWorkloadFactoryHelper::GetFactory(ClWorkloadFactoryHelper::GetMemoryManager()); |
Francis Murtagh | 3d2b4b2 | 2021-02-15 18:23:17 +0000 | [diff] [blame] | 41 | |
| 42 | const armnn::Graph& theGraph = GetGraphForTesting(optNet.get()); |
| 43 | for (auto&& layer : theGraph) |
Aron Virginas-Tar | 7010400 | 2018-10-24 15:33:28 +0100 | [diff] [blame] | 44 | { |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 45 | CHECK(layer->GetBackendId() == armnn::Compute::GpuAcc); |
| 46 | CHECK_NOTHROW( |
Derek Lamberti | 94a88d2 | 2019-12-10 21:12:59 +0000 | [diff] [blame] | 47 | layer->CreateWorkload(fact)); |
Aron Virginas-Tar | 7010400 | 2018-10-24 15:33:28 +0100 | [diff] [blame] | 48 | } |
| 49 | } |
| 50 | |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 51 | TEST_CASE("FP16TurboModeTestOnGpuAcc") |
Aron Virginas-Tar | 7010400 | 2018-10-24 15:33:28 +0100 | [diff] [blame] | 52 | { |
| 53 | // Test to check when Fp16 Turbo mode set |
| 54 | // it converts the Fp32 network to Fp16 Network |
| 55 | // add Fp32ToFp16 conversion layer after the InputLayer |
| 56 | // add Fp16ToFp32 conversion layer after the OutputLayer |
| 57 | // checks the other layers if they are supported in Fp16 |
| 58 | // if they are not put the conversion layers before and after |
| 59 | // if they are not supported in Fp16 use Fp32 instead |
| 60 | // if there are inverse conversion layers remove them with optimization |
| 61 | // at the moment FloorLayer is not supported in Fp16 so it rolls back to Fp32 |
| 62 | // and inverse conversion layers are removed by the optimizer |
Francis Murtagh | 3d2b4b2 | 2021-02-15 18:23:17 +0000 | [diff] [blame] | 63 | armnn::INetworkPtr net(armnn::INetwork::Create()); |
Aron Virginas-Tar | 7010400 | 2018-10-24 15:33:28 +0100 | [diff] [blame] | 64 | |
| 65 | // Defines layers. |
Francis Murtagh | 3d2b4b2 | 2021-02-15 18:23:17 +0000 | [diff] [blame] | 66 | auto input = net->AddInputLayer(0, "input layer"); |
Aron Virginas-Tar | 7010400 | 2018-10-24 15:33:28 +0100 | [diff] [blame] | 67 | // ReLu1 |
| 68 | armnn::ActivationDescriptor activation1Descriptor; |
| 69 | activation1Descriptor.m_Function = armnn::ActivationFunction::BoundedReLu; |
| 70 | activation1Descriptor.m_A = 1.f; |
| 71 | activation1Descriptor.m_B = -1.f; |
Francis Murtagh | 3d2b4b2 | 2021-02-15 18:23:17 +0000 | [diff] [blame] | 72 | auto activation = net->AddActivationLayer(activation1Descriptor, "activation layer"); |
| 73 | auto output = net->AddOutputLayer(0, "output layer"); |
Aron Virginas-Tar | 7010400 | 2018-10-24 15:33:28 +0100 | [diff] [blame] | 74 | |
| 75 | // Connects layers. |
| 76 | input->GetOutputSlot(0).Connect(activation->GetInputSlot(0)); |
| 77 | activation->GetOutputSlot(0).Connect(output->GetInputSlot(0)); |
| 78 | |
| 79 | armnn::TensorShape shape({4}); |
| 80 | armnn::TensorInfo info(shape, armnn::DataType::Float32); |
| 81 | input->GetOutputSlot(0).SetTensorInfo(info); |
| 82 | activation->GetOutputSlot(0).SetTensorInfo(info); |
| 83 | |
| 84 | armnn::IRuntime::CreationOptions options; |
| 85 | armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options)); |
| 86 | |
| 87 | std::vector<armnn::BackendId> backends = {armnn::Compute::GpuAcc}; |
| 88 | |
| 89 | armnn::OptimizerOptions optimizerOptions; |
| 90 | optimizerOptions.m_ReduceFp32ToFp16 = true; |
| 91 | |
| 92 | armnn::IOptimizedNetworkPtr optimizedNet = armnn::Optimize( |
Francis Murtagh | 3d2b4b2 | 2021-02-15 18:23:17 +0000 | [diff] [blame] | 93 | *net, backends, runtime->GetDeviceSpec(), optimizerOptions); |
Aron Virginas-Tar | 7010400 | 2018-10-24 15:33:28 +0100 | [diff] [blame] | 94 | |
Francis Murtagh | 3d2b4b2 | 2021-02-15 18:23:17 +0000 | [diff] [blame] | 95 | const armnn::Graph& graph = GetGraphForTesting(optimizedNet.get()); |
Aron Virginas-Tar | 7010400 | 2018-10-24 15:33:28 +0100 | [diff] [blame] | 96 | |
| 97 | // Tests that all layers are present in the graph. |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 98 | CHECK(graph.GetNumLayers() == 5); |
Aron Virginas-Tar | 7010400 | 2018-10-24 15:33:28 +0100 | [diff] [blame] | 99 | |
| 100 | // Tests that the vertices exist and have correct names. |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 101 | CHECK(GraphHasNamedLayer(graph, "input layer")); |
| 102 | CHECK(GraphHasNamedLayer(graph, "convert_fp32_to_fp16-0-input layer")); |
| 103 | CHECK(GraphHasNamedLayer(graph, "activation layer")); |
| 104 | CHECK(GraphHasNamedLayer(graph, "convert_fp16_to_fp32-0-output layer")); |
| 105 | CHECK(GraphHasNamedLayer(graph, "output layer")); |
Aron Virginas-Tar | 7010400 | 2018-10-24 15:33:28 +0100 | [diff] [blame] | 106 | } |
| 107 | |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 108 | TEST_CASE("FastMathEnabledTestOnGpuAcc") |
Sadik Armagan | 045f6be | 2020-09-10 13:37:32 +0100 | [diff] [blame] | 109 | { |
| 110 | armnn::INetworkPtr net(armnn::INetwork::Create()); |
| 111 | |
| 112 | armnn::IConnectableLayer* input = net->AddInputLayer(0); |
| 113 | armnn::IConnectableLayer* output = net->AddOutputLayer(0); |
| 114 | |
| 115 | input->GetOutputSlot(0).Connect(output->GetInputSlot(0)); |
| 116 | input->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 1, 4, 4 }, armnn::DataType::Float32)); |
| 117 | |
| 118 | armnn::IRuntime::CreationOptions options; |
| 119 | armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options)); |
| 120 | |
| 121 | std::vector<armnn::BackendId> backends = {armnn::Compute::GpuAcc}; |
| 122 | armnn::OptimizerOptions optimizerOptions; |
| 123 | armnn::BackendOptions modelOptions("GpuAcc", {{"FastMathEnabled", true}}); |
| 124 | optimizerOptions.m_ModelOptions.push_back(modelOptions); |
| 125 | |
| 126 | armnn::IOptimizedNetworkPtr optimizedNet = armnn::Optimize( |
| 127 | *net, backends, runtime->GetDeviceSpec(), optimizerOptions); |
| 128 | |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 129 | CHECK(optimizedNet); |
Sadik Armagan | 045f6be | 2020-09-10 13:37:32 +0100 | [diff] [blame] | 130 | |
Francis Murtagh | 3d2b4b2 | 2021-02-15 18:23:17 +0000 | [diff] [blame] | 131 | auto modelOptionsOut = GetModelOptionsForTesting(optimizedNet.get()); |
Sadik Armagan | 045f6be | 2020-09-10 13:37:32 +0100 | [diff] [blame] | 132 | |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 133 | CHECK(modelOptionsOut.size() == 1); |
| 134 | CHECK(modelOptionsOut[0].GetOption(0).GetName() == "FastMathEnabled"); |
| 135 | CHECK(modelOptionsOut[0].GetOption(0).GetValue().AsBool() == true); |
Sadik Armagan | 045f6be | 2020-09-10 13:37:32 +0100 | [diff] [blame] | 136 | } |
| 137 | |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 138 | TEST_CASE("CheckMLGOTuningFile") |
Finn Williams | 4064632 | 2021-02-11 16:16:42 +0000 | [diff] [blame] | 139 | { |
| 140 | class ClBackendContextTestClass : public armnn::ClBackendContext |
| 141 | { |
| 142 | public: |
| 143 | ClBackendContextTestClass(const armnn::IRuntime::CreationOptions &options) : ClBackendContext(options) |
| 144 | {} |
| 145 | |
| 146 | bool call_reload_from_file() |
| 147 | { |
| 148 | return m_MLGOTuner.reload_from_file(m_MLGOTuningFile); |
| 149 | } |
| 150 | }; |
| 151 | |
| 152 | const std::string validText{ |
| 153 | "<header>\n" |
| 154 | "gemm-version, [1,2,1]\n" |
| 155 | "ip-type,gpu\n" |
| 156 | "</header>\n" |
| 157 | "<heuristics-table>\n" |
| 158 | "0, g71 , 8, f32, best-performance, static, gemm-type, [m,n,k,n]\n" |
| 159 | "1, g71 , 8, f32, best-performance, static, gemm-config-reshaped-only-rhs, [m,n,k,n]\n" |
| 160 | "2, g71 , 8, f32, best-performance, static, gemm-config-reshaped, [m,n,k,n]\n" |
| 161 | "3, g71 , 8, qasymm8, best-performance, static, gemm-type, [m,n,k,n]\n" |
| 162 | "4, g71 , 8, qasymm8, best-performance, static, gemm-config-reshaped-only-rhs, [m,n,k,n]\n" |
| 163 | "5, g71 , 8, qasymm8, best-performance, static, gemm-config-native, [m,n,k,n]\n" |
| 164 | "</heuristics-table>\n" |
| 165 | "<heuristic, 0>\n" |
| 166 | "b , 0, var, r_mn, >=, num, 2., 1, 2\n" |
| 167 | "l , 1, gemm-type, reshaped\n" |
| 168 | "l , 2, gemm-type, reshaped-only-rhs\n" |
| 169 | "</heuristic>\n" |
| 170 | "<heuristic, 1>\n" |
| 171 | "l ,0,gemm-config-reshaped-only-rhs, [2, 4,4,4,1,1,0]\n" |
| 172 | "</heuristic>\n" |
| 173 | "<heuristic, 2>\n" |
| 174 | "l ,0,gemm-config-reshaped,[4,2,8,16,16,1,0,1,0]\n" |
| 175 | "</heuristic>\n" |
| 176 | "<heuristic, 3>\n" |
| 177 | "l , 0, gemm-type, native\n" |
| 178 | "</heuristic>\n" |
| 179 | "<heuristic, 4>\n" |
| 180 | "l ,0,gemm-config-reshaped-only-rhs, [2, 4,4,4,1,1,0]\n" |
| 181 | "</heuristic>\n" |
| 182 | "<heuristic, 5>\n" |
| 183 | "l ,0,gemm-config-native,[4,2,8]\n" |
| 184 | "</heuristic>\n"}; |
| 185 | |
| 186 | const std::string invalidText{"ʕノ•ᴥ•ʔノ ︵ ┻━┻"}; |
| 187 | |
| 188 | fs::path validFile = armnnUtils::Filesystem::NamedTempFile("validFile.mlgo"); |
| 189 | fs::path invalidFile = armnnUtils::Filesystem::NamedTempFile("invalidFile.mlgo"); |
| 190 | |
| 191 | try |
| 192 | { |
| 193 | std::ofstream ofs1{validFile}; |
| 194 | ofs1 << validText << std::endl; |
| 195 | ofs1.close(); |
| 196 | |
| 197 | std::ofstream ofs2{invalidFile}; |
| 198 | ofs2 << invalidText << std::endl; |
| 199 | ofs2.close(); |
| 200 | } |
| 201 | catch (std::exception &e) |
| 202 | { |
| 203 | std::cerr << "Unable to write to file at location [" << validFile.c_str() << "] : " << e.what() << std::endl; |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 204 | CHECK(false); |
Finn Williams | 4064632 | 2021-02-11 16:16:42 +0000 | [diff] [blame] | 205 | } |
| 206 | |
| 207 | armnn::IRuntime::CreationOptions creationOptions1; |
| 208 | armnn::BackendOptions validOptions |
| 209 | { |
| 210 | "GpuAcc", |
| 211 | { |
| 212 | {"MLGOTuningFilePath", validFile.c_str()} |
| 213 | } |
| 214 | }; |
| 215 | |
| 216 | creationOptions1.m_BackendOptions.emplace_back(validOptions); |
| 217 | ClBackendContextTestClass clBackendContext1(creationOptions1); |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 218 | CHECK(clBackendContext1.call_reload_from_file()); |
Finn Williams | 4064632 | 2021-02-11 16:16:42 +0000 | [diff] [blame] | 219 | |
| 220 | armnn::BackendOptions invalidOptions |
| 221 | { |
| 222 | "GpuAcc", |
| 223 | { |
| 224 | {"MLGOTuningFilePath", invalidFile.c_str()} |
| 225 | } |
| 226 | }; |
| 227 | |
| 228 | armnn::IRuntime::CreationOptions creationOptions2; |
| 229 | creationOptions2.m_BackendOptions.emplace_back(invalidOptions); |
| 230 | ClBackendContextTestClass clBackendContext2(creationOptions2); |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 231 | CHECK(clBackendContext2.call_reload_from_file() == false); |
Finn Williams | 4064632 | 2021-02-11 16:16:42 +0000 | [diff] [blame] | 232 | |
| 233 | armnn::BackendOptions invalidPathOptions |
| 234 | { |
| 235 | "GpuAcc", |
| 236 | { |
| 237 | {"MLGOTuningFilePath", "not_a_real_file_path"} |
| 238 | } |
| 239 | }; |
| 240 | |
| 241 | armnn::IRuntime::CreationOptions creationOptions3; |
| 242 | creationOptions3.m_BackendOptions.emplace_back(invalidPathOptions); |
| 243 | ClBackendContextTestClass clBackendContext3(creationOptions3); |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 244 | CHECK(clBackendContext3.call_reload_from_file() == false); |
Finn Williams | 4064632 | 2021-02-11 16:16:42 +0000 | [diff] [blame] | 245 | } |
| 246 | |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 247 | } |