David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 1 | // |
John Mcloughlin | c5ee0d7 | 2023-03-24 12:07:25 +0000 | [diff] [blame] | 2 | // Copyright © 2021, 2023 Arm Ltd and Contributors. All rights reserved. |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include <arm_compute/runtime/CL/functions/CLActivationLayer.h> |
| 7 | |
| 8 | #include <cl/ClImportTensorHandle.hpp> |
| 9 | #include <cl/ClImportTensorHandleFactory.hpp> |
| 10 | #include <cl/test/ClContextControlFixture.hpp> |
| 11 | |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 12 | #include <doctest/doctest.h> |
| 13 | |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 14 | #include <armnn/IRuntime.hpp> |
| 15 | #include <armnn/INetwork.hpp> |
Cathal Corbett | a3f4fba | 2022-03-21 09:27:08 +0000 | [diff] [blame] | 16 | #include "Network.hpp" |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 17 | |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 18 | using namespace armnn; |
| 19 | |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 20 | TEST_SUITE("ClImportTensorHandleTests") |
| 21 | { |
| 22 | TEST_CASE_FIXTURE(ClContextControlFixture, "ClMallocImport") |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 23 | { |
| 24 | ClImportTensorHandleFactory handleFactory(static_cast<MemorySourceFlags>(MemorySource::Malloc), |
| 25 | static_cast<MemorySourceFlags>(MemorySource::Malloc)); |
| 26 | |
| 27 | TensorInfo info({ 1, 24, 16, 3 }, DataType::Float32); |
| 28 | unsigned int numElements = info.GetNumElements(); |
| 29 | |
| 30 | // create TensorHandle for memory import |
| 31 | auto handle = handleFactory.CreateTensorHandle(info); |
| 32 | |
| 33 | // Get CLtensor |
| 34 | arm_compute::CLTensor& tensor = PolymorphicDowncast<ClImportTensorHandle*>(handle.get())->GetTensor(); |
| 35 | |
| 36 | // Create and configure activation function |
| 37 | const arm_compute::ActivationLayerInfo act_info(arm_compute::ActivationLayerInfo::ActivationFunction::RELU); |
| 38 | arm_compute::CLActivationLayer act_func; |
| 39 | act_func.configure(&tensor, nullptr, act_info); |
| 40 | |
| 41 | // Allocate user memory |
| 42 | const size_t totalBytes = tensor.info()->total_size(); |
| 43 | const size_t alignment = |
| 44 | arm_compute::CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE>(); |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 45 | size_t space = totalBytes + alignment + alignment; |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 46 | auto testData = std::make_unique<uint8_t[]>(space); |
| 47 | void* alignedPtr = testData.get(); |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 48 | CHECK(std::align(alignment, totalBytes, alignedPtr, space)); |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 49 | |
| 50 | // Import memory |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 51 | CHECK(handle->Import(alignedPtr, armnn::MemorySource::Malloc)); |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 52 | |
| 53 | // Input with negative values |
| 54 | auto* typedPtr = reinterpret_cast<float*>(alignedPtr); |
| 55 | std::fill_n(typedPtr, numElements, -5.0f); |
| 56 | |
| 57 | // Execute function and sync |
| 58 | act_func.run(); |
| 59 | arm_compute::CLScheduler::get().sync(); |
| 60 | |
| 61 | // Validate result by checking that the output has no negative values |
| 62 | for(unsigned int i = 0; i < numElements; ++i) |
| 63 | { |
Jan Eilers | c1c872f | 2021-07-22 13:17:04 +0100 | [diff] [blame] | 64 | CHECK(typedPtr[i] == 0); |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 65 | } |
| 66 | } |
| 67 | |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 68 | TEST_CASE_FIXTURE(ClContextControlFixture, "ClIncorrectMemorySourceImport") |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 69 | { |
| 70 | ClImportTensorHandleFactory handleFactory(static_cast<MemorySourceFlags>(MemorySource::Malloc), |
| 71 | static_cast<MemorySourceFlags>(MemorySource::Malloc)); |
| 72 | |
| 73 | TensorInfo info({ 1, 24, 16, 3 }, DataType::Float32); |
| 74 | |
| 75 | // create TensorHandle for memory import |
| 76 | auto handle = handleFactory.CreateTensorHandle(info); |
| 77 | |
| 78 | // Get CLtensor |
| 79 | arm_compute::CLTensor& tensor = PolymorphicDowncast<ClImportTensorHandle*>(handle.get())->GetTensor(); |
| 80 | |
| 81 | // Allocate user memory |
| 82 | const size_t totalBytes = tensor.info()->total_size(); |
| 83 | const size_t alignment = |
| 84 | arm_compute::CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE>(); |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 85 | size_t space = totalBytes + alignment + alignment; |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 86 | auto testData = std::make_unique<uint8_t[]>(space); |
| 87 | void* alignedPtr = testData.get(); |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 88 | CHECK(std::align(alignment, totalBytes, alignedPtr, space)); |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 89 | |
| 90 | // Import memory |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 91 | CHECK_THROWS_AS(handle->Import(alignedPtr, armnn::MemorySource::Undefined), MemoryImportException); |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 92 | } |
| 93 | |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 94 | TEST_CASE_FIXTURE(ClContextControlFixture, "ClInvalidMemorySourceImport") |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 95 | { |
| 96 | MemorySource invalidMemSource = static_cast<MemorySource>(256); |
| 97 | ClImportTensorHandleFactory handleFactory(static_cast<MemorySourceFlags>(invalidMemSource), |
| 98 | static_cast<MemorySourceFlags>(invalidMemSource)); |
| 99 | |
| 100 | TensorInfo info({ 1, 2, 2, 1 }, DataType::Float32); |
| 101 | |
| 102 | // create TensorHandle for memory import |
| 103 | auto handle = handleFactory.CreateTensorHandle(info); |
| 104 | |
| 105 | // Allocate user memory |
| 106 | std::vector<float> inputData |
| 107 | { |
| 108 | 1.0f, 2.0f, 3.0f, 4.0f |
| 109 | }; |
| 110 | |
| 111 | // Import non-support memory |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 112 | CHECK_THROWS_AS(handle->Import(inputData.data(), invalidMemSource), MemoryImportException); |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 113 | } |
| 114 | |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 115 | TEST_CASE_FIXTURE(ClContextControlFixture, "ClImportEndToEnd") |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 116 | { |
| 117 | // Create runtime in which test will run |
| 118 | IRuntime::CreationOptions options; |
| 119 | IRuntimePtr runtime(armnn::IRuntime::Create(options)); |
| 120 | |
| 121 | // build up the structure of the network |
| 122 | INetworkPtr net(INetwork::Create()); |
| 123 | |
| 124 | IConnectableLayer* input = net->AddInputLayer(0, "Input"); |
| 125 | |
| 126 | ActivationDescriptor descriptor; |
| 127 | descriptor.m_Function = ActivationFunction::ReLu; |
| 128 | IConnectableLayer* activation = net->AddActivationLayer(descriptor, "Activation"); |
| 129 | |
| 130 | IConnectableLayer* output = net->AddOutputLayer(0, "Output"); |
| 131 | |
| 132 | input->GetOutputSlot(0).Connect(activation->GetInputSlot(0)); |
| 133 | activation->GetOutputSlot(0).Connect(output->GetInputSlot(0)); |
| 134 | |
| 135 | TensorInfo tensorInfo = TensorInfo({ 1, 24, 16, 3 }, DataType::Float32); |
| 136 | unsigned int numElements = tensorInfo.GetNumElements(); |
| 137 | size_t totalBytes = numElements * sizeof(float); |
| 138 | |
| 139 | input->GetOutputSlot(0).SetTensorInfo(tensorInfo); |
| 140 | activation->GetOutputSlot(0).SetTensorInfo(tensorInfo); |
| 141 | |
| 142 | // Optimize the network |
John Mcloughlin | c5ee0d7 | 2023-03-24 12:07:25 +0000 | [diff] [blame] | 143 | OptimizerOptionsOpaque optOptions; |
| 144 | optOptions.SetImportEnabled(true); |
| 145 | optOptions.SetExportEnabled(true); |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 146 | std::vector<armnn::BackendId> backends = {armnn::Compute::GpuAcc}; |
| 147 | IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optOptions); |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 148 | CHECK(optNet); |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 149 | |
| 150 | // Loads it into the runtime. |
| 151 | NetworkId netId; |
| 152 | std::string ignoredErrorMessage; |
| 153 | // Enable Importing |
| 154 | INetworkProperties networkProperties(false, MemorySource::Malloc, MemorySource::Malloc); |
| 155 | runtime->LoadNetwork(netId, std::move(optNet), ignoredErrorMessage, networkProperties); |
| 156 | |
| 157 | // Creates structures for input & output |
| 158 | const size_t alignment = |
| 159 | arm_compute::CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE>(); |
| 160 | size_t space = totalBytes + alignment + alignment; |
| 161 | auto inputData = std::make_unique<uint8_t[]>(space); |
| 162 | void* alignedInputPtr = inputData.get(); |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 163 | CHECK(std::align(alignment, totalBytes, alignedInputPtr, space)); |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 164 | |
| 165 | // Input with negative values |
| 166 | auto* intputPtr = reinterpret_cast<float*>(alignedInputPtr); |
| 167 | std::fill_n(intputPtr, numElements, -5.0f); |
| 168 | |
| 169 | auto outputData = std::make_unique<uint8_t[]>(space); |
| 170 | void* alignedOutputPtr = outputData.get(); |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 171 | CHECK(std::align(alignment, totalBytes, alignedOutputPtr, space)); |
Narumol Prangnawarat | 878e0f9 | 2021-05-11 19:51:14 +0100 | [diff] [blame] | 172 | auto* outputPtr = reinterpret_cast<float*>(alignedOutputPtr); |
| 173 | std::fill_n(outputPtr, numElements, -10.0f); |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 174 | |
Cathal Corbett | 5b8093c | 2021-10-22 11:12:07 +0100 | [diff] [blame] | 175 | TensorInfo inputTensorInfo = runtime->GetInputTensorInfo(netId, 0); |
| 176 | inputTensorInfo.SetConstant(true); |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 177 | InputTensors inputTensors |
| 178 | { |
Cathal Corbett | 5b8093c | 2021-10-22 11:12:07 +0100 | [diff] [blame] | 179 | {0,armnn::ConstTensor(inputTensorInfo, alignedInputPtr)}, |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 180 | }; |
| 181 | OutputTensors outputTensors |
| 182 | { |
| 183 | {0,armnn::Tensor(runtime->GetOutputTensorInfo(netId, 0), alignedOutputPtr)} |
| 184 | }; |
| 185 | |
| 186 | runtime->GetProfiler(netId)->EnableProfiling(true); |
| 187 | |
| 188 | // Do the inference |
| 189 | runtime->EnqueueWorkload(netId, inputTensors, outputTensors); |
| 190 | |
| 191 | // Retrieve the Profiler.Print() output to get the workload execution |
| 192 | ProfilerManager& profilerManager = armnn::ProfilerManager::GetInstance(); |
| 193 | std::stringstream ss; |
| 194 | profilerManager.GetProfiler()->Print(ss);; |
| 195 | std::string dump = ss.str(); |
| 196 | |
| 197 | // Contains ActivationWorkload |
| 198 | std::size_t found = dump.find("ActivationWorkload"); |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 199 | CHECK(found != std::string::npos); |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 200 | |
| 201 | // Contains SyncMemGeneric |
| 202 | found = dump.find("SyncMemGeneric"); |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 203 | CHECK(found != std::string::npos); |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 204 | |
| 205 | // Does not contain CopyMemGeneric |
| 206 | found = dump.find("CopyMemGeneric"); |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 207 | CHECK(found == std::string::npos); |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 208 | |
Narumol Prangnawarat | 878e0f9 | 2021-05-11 19:51:14 +0100 | [diff] [blame] | 209 | runtime->UnloadNetwork(netId); |
| 210 | |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 211 | // Check output is as expected |
| 212 | // Validate result by checking that the output has no negative values |
| 213 | auto* outputResult = reinterpret_cast<float*>(alignedOutputPtr); |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 214 | CHECK(outputResult); |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 215 | for(unsigned int i = 0; i < numElements; ++i) |
| 216 | { |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 217 | CHECK(outputResult[i] >= 0); |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 218 | } |
Narumol Prangnawarat | e5f0b24 | 2021-05-07 17:52:36 +0100 | [diff] [blame] | 219 | } |
| 220 | |
Nikhil Raj | 60ab976 | 2022-01-13 09:34:44 +0000 | [diff] [blame] | 221 | TEST_CASE_FIXTURE(ClContextControlFixture, "ClCanBeImported") |
| 222 | { |
| 223 | ClImportTensorHandleFactory handleFactory(static_cast<MemorySourceFlags>(MemorySource::Malloc), |
| 224 | static_cast<MemorySourceFlags>(MemorySource::Malloc)); |
| 225 | |
| 226 | TensorInfo info({ 1, 24, 16, 3 }, DataType::Float32); |
| 227 | |
| 228 | // create TensorHandle for memory import |
David Monahan | 3826ab6 | 2022-02-21 12:26:16 +0000 | [diff] [blame] | 229 | auto handle = handleFactory.CreateTensorHandle(info, DataLayout::NHWC); |
Nikhil Raj | 60ab976 | 2022-01-13 09:34:44 +0000 | [diff] [blame] | 230 | |
| 231 | // Get CLtensor |
| 232 | arm_compute::CLTensor& tensor = PolymorphicDowncast<ClImportTensorHandle*>(handle.get())->GetTensor(); |
| 233 | |
| 234 | // Allocate user memory |
| 235 | const size_t totalBytes = tensor.info()->total_size(); |
| 236 | const size_t alignment = |
| 237 | arm_compute::CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE>(); |
| 238 | size_t space = totalBytes + alignment + alignment; |
| 239 | auto testData = std::make_unique<uint8_t[]>(space); |
| 240 | void* alignedPtr = testData.get(); |
| 241 | CHECK(std::align(alignment, totalBytes, alignedPtr, space)); |
| 242 | |
| 243 | // Import memory |
| 244 | CHECK_THROWS_AS(handle->CanBeImported(alignedPtr, armnn::MemorySource::Undefined), MemoryImportException); |
| 245 | |
| 246 | } |
| 247 | |
| 248 | TEST_CASE("ClCanBeImportedAlignedMemory") |
| 249 | { |
| 250 | ClImportTensorHandleFactory handleFactory(static_cast<MemorySourceFlags>(MemorySource::Malloc), |
| 251 | static_cast<MemorySourceFlags>(MemorySource::Malloc)); |
| 252 | |
| 253 | TensorInfo info({ 1, 1, 1, 1 }, DataType::Float32); |
| 254 | |
| 255 | // create TensorHandle (Memory Managed status is irrelevant) |
David Monahan | 3826ab6 | 2022-02-21 12:26:16 +0000 | [diff] [blame] | 256 | auto handle = handleFactory.CreateTensorHandle(info, DataLayout::NHWC); |
Nikhil Raj | 60ab976 | 2022-01-13 09:34:44 +0000 | [diff] [blame] | 257 | // Get CLtensor |
| 258 | arm_compute::CLTensor& tensor = PolymorphicDowncast<ClImportTensorHandle*>(handle.get())->GetTensor(); |
| 259 | |
| 260 | // Create an aligned buffer |
| 261 | const size_t totalBytes = tensor.info()->total_size(); |
| 262 | const size_t alignment = |
| 263 | arm_compute::CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE>(); |
| 264 | size_t space = totalBytes + alignment + alignment; |
| 265 | auto testData = std::make_unique<uint8_t[]>(space); |
| 266 | void* alignedPtr = testData.get(); |
| 267 | CHECK(std::align(alignment, totalBytes, alignedPtr, space)); |
| 268 | |
| 269 | // Check aligned buffers return true |
| 270 | CHECK(handle->CanBeImported(alignedPtr, MemorySource::Malloc) == true); |
| 271 | |
| 272 | // Due to the nature of how GPU memory is mapped it is entirely possible for memory which is misaligned on cpu |
| 273 | // to be successfully import on GPU. As such there is no way to create a misaligned pointer that will always fail. |
| 274 | // Rather it will succeed on some devices and fail on others. As long as a correctly aligned buffer returns true |
| 275 | // we can be confident that it will be successfully imported. All other cases will need to be handled by the user. |
| 276 | } |
| 277 | |
Narumol Prangnawarat | e2af6f4 | 2022-01-28 17:59:18 +0000 | [diff] [blame] | 278 | TEST_CASE_FIXTURE(ClContextControlFixture, "ClForceImportConv2dEndToEnd") |
| 279 | { |
| 280 | // Create runtime in which test will run |
| 281 | IRuntime::CreationOptions options; |
| 282 | IRuntimePtr runtime(armnn::IRuntime::Create(options)); |
| 283 | |
| 284 | // build up the structure of the network |
| 285 | INetworkPtr network(INetwork::Create()); |
| 286 | |
| 287 | armnn::TensorInfo inputInfo({ 1, 3, 4, 1 }, DataType::Float32); |
| 288 | armnn::TensorInfo kernelInfo({ 1, 3, 3, 1 }, DataType::Float32); |
| 289 | armnn::TensorInfo outputInfo({ 1, 3, 4, 1 }, DataType::Float32); |
| 290 | |
| 291 | kernelInfo.SetConstant(true); |
| 292 | |
| 293 | std::vector<float> kernel = |
| 294 | { |
| 295 | 4, 5, 6, |
| 296 | 0, 0, 0, |
| 297 | 3, 2, 1 |
| 298 | }; |
| 299 | |
| 300 | const std::vector<float> expectedOutput = |
| 301 | { |
| 302 | 23, 41, 33, 21, |
| 303 | 44, 65, 76, 52, |
| 304 | 82, 85, 79, 42 |
| 305 | }; |
| 306 | |
| 307 | unsigned int numElements = inputInfo.GetNumElements(); |
| 308 | size_t totalBytes = numElements * sizeof(float); |
| 309 | |
| 310 | IConnectableLayer* const inputLayer = network->AddInputLayer(0, "input"); |
| 311 | ARMNN_ASSERT(inputLayer); |
| 312 | |
| 313 | armnn::ConstTensor weights(kernelInfo, kernel); |
| 314 | |
| 315 | armnn::Convolution2dDescriptor convDesc2d; |
| 316 | convDesc2d.m_StrideX = 1; |
| 317 | convDesc2d.m_StrideY = 1; |
| 318 | convDesc2d.m_PadLeft = 1; |
| 319 | convDesc2d.m_PadRight = 1; |
| 320 | convDesc2d.m_PadTop = 1; |
| 321 | convDesc2d.m_PadBottom = 1; |
| 322 | convDesc2d.m_DataLayout = DataLayout::NHWC; |
Keith Davis | b4dd5cc | 2022-04-07 11:32:00 +0100 | [diff] [blame] | 323 | |
Keith Davis | 721e629 | 2022-05-17 10:06:53 +0100 | [diff] [blame] | 324 | armnn::IConnectableLayer* const convLayer = network->AddConvolution2dLayer(convDesc2d, "conv"); |
| 325 | armnn::IConnectableLayer* weightsLayer = network->AddConstantLayer(weights); |
| 326 | |
Narumol Prangnawarat | e2af6f4 | 2022-01-28 17:59:18 +0000 | [diff] [blame] | 327 | ARMNN_ASSERT(convLayer); |
| 328 | |
Keith Davis | 721e629 | 2022-05-17 10:06:53 +0100 | [diff] [blame] | 329 | weightsLayer->GetOutputSlot(0).SetTensorInfo(weights.GetInfo()); |
| 330 | weightsLayer->GetOutputSlot(0).Connect(convLayer->GetInputSlot(1u)); |
| 331 | |
Narumol Prangnawarat | e2af6f4 | 2022-01-28 17:59:18 +0000 | [diff] [blame] | 332 | inputLayer->GetOutputSlot(0).Connect(convLayer->GetInputSlot(0)); |
| 333 | inputLayer->GetOutputSlot(0).SetTensorInfo(inputInfo); |
| 334 | |
| 335 | IConnectableLayer* output = network->AddOutputLayer(0, "output"); |
| 336 | convLayer->GetOutputSlot(0).Connect(output->GetInputSlot(0)); |
| 337 | convLayer->GetOutputSlot(0).SetTensorInfo(outputInfo); |
| 338 | |
| 339 | // Optimize the network |
John Mcloughlin | c5ee0d7 | 2023-03-24 12:07:25 +0000 | [diff] [blame] | 340 | OptimizerOptionsOpaque optOptions; |
| 341 | optOptions.SetImportEnabled(false); |
| 342 | optOptions.SetExportEnabled(false); |
Narumol Prangnawarat | e2af6f4 | 2022-01-28 17:59:18 +0000 | [diff] [blame] | 343 | std::vector<armnn::BackendId> backends = {armnn::Compute::GpuAcc}; |
| 344 | IOptimizedNetworkPtr optNet = Optimize(*network, backends, runtime->GetDeviceSpec(), optOptions); |
| 345 | CHECK(optNet); |
| 346 | |
| 347 | // Loads it into the runtime. |
| 348 | NetworkId netId; |
| 349 | std::string ignoredErrorMessage; |
| 350 | // Enable Importing |
| 351 | INetworkProperties networkProperties(false, MemorySource::Undefined, MemorySource::Undefined); |
| 352 | runtime->LoadNetwork(netId, std::move(optNet), ignoredErrorMessage, networkProperties); |
| 353 | |
| 354 | // Creates structures for input & output |
| 355 | const size_t alignment = |
| 356 | arm_compute::CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE>(); |
| 357 | size_t space = totalBytes + alignment + alignment; |
| 358 | auto inputData = std::make_unique<uint8_t[]>(space); |
| 359 | void* alignedInputPtr = inputData.get(); |
| 360 | CHECK(std::align(alignment, totalBytes, alignedInputPtr, space)); |
| 361 | |
| 362 | // Input with negative values |
| 363 | auto* inputPtr = reinterpret_cast<float*>(alignedInputPtr); |
| 364 | inputPtr[0] = 1; |
| 365 | inputPtr[1] = 5; |
| 366 | inputPtr[2] = 2; |
| 367 | inputPtr[3] = 3; |
| 368 | inputPtr[4] = 8; |
| 369 | inputPtr[5] = 7; |
| 370 | inputPtr[6] = 3; |
| 371 | inputPtr[7] = 6; |
| 372 | inputPtr[8] = 3; |
| 373 | inputPtr[9] = 3; |
| 374 | inputPtr[10] = 9; |
| 375 | inputPtr[11] = 1; |
| 376 | |
| 377 | |
| 378 | auto outputData = std::make_unique<uint8_t[]>(space); |
| 379 | void* alignedOutputPtr = outputData.get(); |
| 380 | CHECK(std::align(alignment, totalBytes, alignedOutputPtr, space)); |
| 381 | auto* outputPtr = reinterpret_cast<float*>(alignedOutputPtr); |
| 382 | std::fill_n(outputPtr, numElements, -10.0f); |
| 383 | |
| 384 | TensorInfo inputTensorInfo = runtime->GetInputTensorInfo(netId, 0); |
| 385 | inputTensorInfo.SetConstant(true); |
| 386 | InputTensors inputTensors |
| 387 | { |
| 388 | {0,armnn::ConstTensor(inputTensorInfo, alignedInputPtr)}, |
| 389 | }; |
| 390 | OutputTensors outputTensors |
| 391 | { |
| 392 | {0,armnn::Tensor(runtime->GetOutputTensorInfo(netId, 0), alignedOutputPtr)} |
| 393 | }; |
| 394 | |
| 395 | runtime->GetProfiler(netId)->EnableProfiling(true); |
| 396 | |
| 397 | INFO("Run ImportInputs"); |
| 398 | std::vector<ImportedInputId> importedInputIds = |
| 399 | runtime->ImportInputs(netId, inputTensors, MemorySource::Malloc); |
Colm Donelan | d7ceec5 | 2022-07-06 12:09:05 +0100 | [diff] [blame] | 400 | // We expect the import to have succeeded. |
| 401 | CHECK(importedInputIds.size() == 1); |
Narumol Prangnawarat | e2af6f4 | 2022-01-28 17:59:18 +0000 | [diff] [blame] | 402 | std::vector<ImportedOutputId> importedOutputIds = |
| 403 | runtime->ImportOutputs(netId, outputTensors, MemorySource::Malloc); |
Colm Donelan | d7ceec5 | 2022-07-06 12:09:05 +0100 | [diff] [blame] | 404 | // We expect the import to have succeeded. |
| 405 | CHECK(importedOutputIds.size() == 1); |
Narumol Prangnawarat | e2af6f4 | 2022-01-28 17:59:18 +0000 | [diff] [blame] | 406 | // Do the inference |
Colm Donelan | d7ceec5 | 2022-07-06 12:09:05 +0100 | [diff] [blame] | 407 | runtime->EnqueueWorkload(netId, InputTensors(), OutputTensors(), importedInputIds, importedOutputIds); |
Narumol Prangnawarat | e2af6f4 | 2022-01-28 17:59:18 +0000 | [diff] [blame] | 408 | |
| 409 | // Retrieve the Profiler.Print() output to get the workload execution |
| 410 | ProfilerManager& profilerManager = armnn::ProfilerManager::GetInstance(); |
| 411 | std::stringstream ss; |
| 412 | profilerManager.GetProfiler()->Print(ss);; |
| 413 | std::string dump = ss.str(); |
| 414 | |
| 415 | // Contains Convolution2dWorkload |
| 416 | std::size_t found = dump.find("Convolution2dWorkload"); |
| 417 | CHECK(found != std::string::npos); |
| 418 | |
| 419 | // Contains SyncMemGeneric |
| 420 | found = dump.find("SyncMemGeneric"); |
| 421 | CHECK(found != std::string::npos); |
| 422 | |
| 423 | // Does not contain CopyMemGeneric |
| 424 | found = dump.find("CopyMemGeneric"); |
| 425 | CHECK(found == std::string::npos); |
| 426 | |
| 427 | runtime->UnloadNetwork(netId); |
| 428 | |
| 429 | // Check output is as expected |
| 430 | // Validate result by checking that the output has no negative values |
| 431 | auto* outputResult = reinterpret_cast<float*>(alignedOutputPtr); |
| 432 | CHECK(outputResult); |
| 433 | |
| 434 | // Check the output is correct |
| 435 | CHECK(std::equal(outputResult, outputResult + numElements, expectedOutput.begin(), expectedOutput.end())); |
| 436 | } |
| 437 | |
Cathal Corbett | a3f4fba | 2022-03-21 09:27:08 +0000 | [diff] [blame] | 438 | TEST_CASE_FIXTURE(ClContextControlFixture, "ClForceImportConvertFp16toFp32EndToEnd") |
| 439 | { |
| 440 | using namespace half_float::literal; |
| 441 | |
| 442 | // Create runtime in which test will run |
| 443 | IRuntime::CreationOptions options; |
| 444 | IRuntimePtr runtime(armnn::IRuntime::Create(options)); |
| 445 | |
| 446 | // build up the structure of the network |
| 447 | NetworkImpl network; |
| 448 | |
| 449 | armnn::TensorInfo inputInfo({1, 3, 2, 3}, armnn::DataType::Float16); |
| 450 | armnn::TensorInfo outputTensorInfo({1, 3, 2, 3}, armnn::DataType::Float32); |
| 451 | |
| 452 | std::vector<float> expectedOutput = |
| 453 | { |
| 454 | -37.5f, -15.2f, -8.76f, -2.0f, -1.5f, -1.3f, -0.5f, -0.4f, 0.0f, |
| 455 | 1.0f, 0.4f, 0.5f, 1.3f, 1.5f, 2.0f, 8.76f, 15.2f, 37.5f |
| 456 | }; |
| 457 | |
| 458 | unsigned int numElements = inputInfo.GetNumElements(); |
| 459 | size_t totalBytesInput = numElements * sizeof(Half); |
| 460 | size_t totalBytesOutput = numElements * sizeof(float); |
| 461 | |
| 462 | IConnectableLayer* const inputLayer = network.AddInputLayer(0, "input"); |
| 463 | ARMNN_ASSERT(inputLayer); |
| 464 | |
| 465 | armnn::IConnectableLayer* const convLayer = network.AddConvertFp16ToFp32Layer("convert"); |
| 466 | ARMNN_ASSERT(convLayer); |
| 467 | |
| 468 | inputLayer->GetOutputSlot(0).Connect(convLayer->GetInputSlot(0)); |
| 469 | inputLayer->GetOutputSlot(0).SetTensorInfo(inputInfo); |
| 470 | |
| 471 | IConnectableLayer* output = network.AddOutputLayer(0, "output"); |
| 472 | convLayer->GetOutputSlot(0).Connect(output->GetInputSlot(0)); |
| 473 | convLayer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo); |
| 474 | |
| 475 | // Optimize the network |
John Mcloughlin | c5ee0d7 | 2023-03-24 12:07:25 +0000 | [diff] [blame] | 476 | OptimizerOptionsOpaque optOptions; |
| 477 | optOptions.SetImportEnabled(false); |
| 478 | optOptions.SetExportEnabled(false); |
Cathal Corbett | a3f4fba | 2022-03-21 09:27:08 +0000 | [diff] [blame] | 479 | std::vector<armnn::BackendId> backends = {armnn::Compute::GpuAcc}; |
| 480 | IOptimizedNetworkPtr optNet = Optimize(network.GetGraph(), backends, runtime->GetDeviceSpec(), optOptions); |
| 481 | CHECK(optNet); |
| 482 | |
| 483 | // Loads it into the runtime. |
| 484 | NetworkId netId; |
| 485 | std::string ignoredErrorMessage; |
| 486 | // Enable Importing |
| 487 | INetworkProperties networkProperties(false, MemorySource::Undefined, MemorySource::Undefined); |
| 488 | runtime->LoadNetwork(netId, std::move(optNet), ignoredErrorMessage, networkProperties); |
| 489 | |
| 490 | // Creates structures for input & output |
| 491 | const size_t alignment = |
| 492 | arm_compute::CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE>(); |
| 493 | size_t spaceInput = totalBytesInput + alignment + alignment; |
| 494 | size_t spaceOutput = totalBytesOutput + alignment + alignment; |
| 495 | auto inputData = std::make_unique<uint8_t[]>(spaceInput); |
| 496 | void* alignedInputPtr = inputData.get(); |
| 497 | CHECK(std::align(alignment, totalBytesInput, alignedInputPtr, spaceInput)); |
| 498 | |
| 499 | // Input with negative values |
| 500 | auto* inputPtr = reinterpret_cast<Half*>(alignedInputPtr); |
| 501 | inputPtr[0] = -37.5_h; |
| 502 | inputPtr[1] = -15.2_h; |
| 503 | inputPtr[2] = -8.76_h; |
| 504 | inputPtr[3] = -2.0_h; |
| 505 | inputPtr[4] = -1.5_h; |
| 506 | inputPtr[5] = -1.3_h; |
| 507 | inputPtr[6] = -0.5_h; |
| 508 | inputPtr[7] = -0.4_h; |
| 509 | inputPtr[8] = 0.0_h; |
| 510 | inputPtr[9] = 1.0_h; |
| 511 | inputPtr[10] = 0.4_h; |
| 512 | inputPtr[11] = 0.5_h; |
| 513 | inputPtr[12] = 1.3_h; |
| 514 | inputPtr[13] = 1.5_h; |
| 515 | inputPtr[14] = 2.0_h; |
| 516 | inputPtr[15] = 8.76_h; |
| 517 | inputPtr[16] = 15.2_h; |
| 518 | inputPtr[17] = 37.5_h; |
| 519 | |
| 520 | auto outputData = std::make_unique<uint8_t[]>(spaceOutput); |
| 521 | void* alignedOutputPtr = outputData.get(); |
| 522 | CHECK(std::align(alignment, totalBytesOutput, alignedOutputPtr, spaceOutput)); |
| 523 | auto* outputPtr = reinterpret_cast<float*>(alignedOutputPtr); |
| 524 | std::fill_n(outputPtr, numElements, -10.0f); |
| 525 | |
| 526 | TensorInfo inputTensorInfo = runtime->GetInputTensorInfo(netId, 0); |
| 527 | inputTensorInfo.SetConstant(true); |
| 528 | InputTensors inputTensors |
| 529 | { |
| 530 | {0,armnn::ConstTensor(inputTensorInfo, alignedInputPtr)}, |
| 531 | }; |
| 532 | OutputTensors outputTensors |
| 533 | { |
| 534 | {0,armnn::Tensor(runtime->GetOutputTensorInfo(netId, 0), alignedOutputPtr)} |
| 535 | }; |
| 536 | |
| 537 | runtime->GetProfiler(netId)->EnableProfiling(true); |
| 538 | |
| 539 | INFO("Run ImportInputs"); |
| 540 | std::vector<ImportedInputId> importedInputIds = |
| 541 | runtime->ImportInputs(netId, inputTensors, MemorySource::Malloc); |
Colm Donelan | d7ceec5 | 2022-07-06 12:09:05 +0100 | [diff] [blame] | 542 | // We expect the import to have succeeded. |
| 543 | CHECK(importedInputIds.size() == 1); |
Cathal Corbett | a3f4fba | 2022-03-21 09:27:08 +0000 | [diff] [blame] | 544 | std::vector<ImportedOutputId> importedOutputIds = |
| 545 | runtime->ImportOutputs(netId, outputTensors, MemorySource::Malloc); |
Colm Donelan | d7ceec5 | 2022-07-06 12:09:05 +0100 | [diff] [blame] | 546 | // We expect the import to have succeeded. |
| 547 | CHECK(importedOutputIds.size() == 1); |
Cathal Corbett | a3f4fba | 2022-03-21 09:27:08 +0000 | [diff] [blame] | 548 | |
| 549 | // Do the inference |
Colm Donelan | d7ceec5 | 2022-07-06 12:09:05 +0100 | [diff] [blame] | 550 | runtime->EnqueueWorkload(netId, InputTensors(), OutputTensors(), importedInputIds, importedOutputIds); |
Cathal Corbett | a3f4fba | 2022-03-21 09:27:08 +0000 | [diff] [blame] | 551 | |
| 552 | // Retrieve the Profiler.Print() output to get the workload execution |
| 553 | ProfilerManager& profilerManager = armnn::ProfilerManager::GetInstance(); |
| 554 | std::stringstream ss; |
| 555 | profilerManager.GetProfiler()->Print(ss);; |
| 556 | std::string dump = ss.str(); |
| 557 | |
| 558 | // Contains Convolution2dWorkload |
| 559 | std::size_t found = dump.find("ConvertFp16ToFp32Workload"); |
| 560 | CHECK(found != std::string::npos); |
| 561 | |
| 562 | // Contains SyncMemGeneric |
| 563 | found = dump.find("SyncMemGeneric"); |
| 564 | CHECK(found != std::string::npos); |
| 565 | |
| 566 | // Does not contain CopyMemGeneric |
| 567 | found = dump.find("CopyMemGeneric"); |
| 568 | CHECK(found == std::string::npos); |
| 569 | |
| 570 | runtime->UnloadNetwork(netId); |
| 571 | |
| 572 | // Check output is as expected |
| 573 | // Validate result by checking that the output has no negative values |
| 574 | auto* outputResult = reinterpret_cast<float*>(alignedOutputPtr); |
| 575 | CHECK(outputResult); |
| 576 | |
| 577 | // Check the output is correct |
| 578 | for (size_t i = 0; i < numElements; ++i) |
| 579 | { |
| 580 | DOCTEST_CHECK_MESSAGE(outputResult[i] == doctest::Approx(expectedOutput[i]).epsilon(0.0004), |
| 581 | "outputValue[" << i << "]: " << outputResult[i] << " != " << expectedOutput[i]); |
| 582 | } |
| 583 | } |
| 584 | |
| 585 | |
| 586 | TEST_CASE_FIXTURE(ClContextControlFixture, "ClForceImportConvertFp32toFp16EndToEnd") |
| 587 | { |
| 588 | using namespace half_float::literal; |
| 589 | |
| 590 | // Create runtime in which test will run |
| 591 | IRuntime::CreationOptions options; |
| 592 | IRuntimePtr runtime(armnn::IRuntime::Create(options)); |
| 593 | |
| 594 | // build up the structure of the network |
| 595 | NetworkImpl network; |
| 596 | |
| 597 | armnn::TensorInfo inputInfo({1, 3, 2, 3}, armnn::DataType::Float32); |
| 598 | armnn::TensorInfo outputTensorInfo({1, 3, 2, 3}, armnn::DataType::Float16); |
| 599 | |
| 600 | std::vector<Half> expectedOutput = |
| 601 | { |
| 602 | -37.5_h, -15.2_h, -8.76_h, -2.0_h, -1.5_h, -1.3_h, -0.5_h, -0.4_h, 0.0_h, |
| 603 | 1.0_h, 0.4_h, 0.5_h, 1.3_h, 1.5_h, 2.0_h, 8.76_h, 15.2_h, 37.5_h |
| 604 | }; |
| 605 | |
| 606 | unsigned int numElements = inputInfo.GetNumElements(); |
| 607 | size_t totalBytesInput = numElements * sizeof(float); |
| 608 | size_t totalBytesOutput = numElements * sizeof(Half); |
| 609 | |
| 610 | IConnectableLayer* const inputLayer = network.AddInputLayer(0, "input"); |
| 611 | ARMNN_ASSERT(inputLayer); |
| 612 | |
| 613 | armnn::IConnectableLayer* const convLayer = network.AddConvertFp32ToFp16Layer("convert"); |
| 614 | ARMNN_ASSERT(convLayer); |
| 615 | |
| 616 | inputLayer->GetOutputSlot(0).Connect(convLayer->GetInputSlot(0)); |
| 617 | inputLayer->GetOutputSlot(0).SetTensorInfo(inputInfo); |
| 618 | |
| 619 | IConnectableLayer* output = network.AddOutputLayer(0, "output"); |
| 620 | convLayer->GetOutputSlot(0).Connect(output->GetInputSlot(0)); |
| 621 | convLayer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo); |
| 622 | |
| 623 | // Optimize the network |
John Mcloughlin | c5ee0d7 | 2023-03-24 12:07:25 +0000 | [diff] [blame] | 624 | OptimizerOptionsOpaque optOptions; |
| 625 | optOptions.SetImportEnabled(false); |
| 626 | optOptions.SetExportEnabled(false); |
Cathal Corbett | a3f4fba | 2022-03-21 09:27:08 +0000 | [diff] [blame] | 627 | std::vector<armnn::BackendId> backends = {armnn::Compute::GpuAcc}; |
| 628 | IOptimizedNetworkPtr optNet = Optimize(network.GetGraph(), backends, runtime->GetDeviceSpec(), optOptions); |
| 629 | CHECK(optNet); |
| 630 | |
| 631 | // Loads it into the runtime. |
| 632 | NetworkId netId; |
| 633 | std::string ignoredErrorMessage; |
| 634 | // Enable Importing |
| 635 | INetworkProperties networkProperties(false, MemorySource::Undefined, MemorySource::Undefined); |
| 636 | runtime->LoadNetwork(netId, std::move(optNet), ignoredErrorMessage, networkProperties); |
| 637 | |
| 638 | // Creates structures for input & output |
| 639 | const size_t alignment = |
| 640 | arm_compute::CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE>(); |
| 641 | size_t spaceInput = totalBytesInput + alignment + alignment; |
| 642 | size_t spaceOutput = totalBytesOutput + alignment + alignment; |
| 643 | auto inputData = std::make_unique<uint8_t[]>(spaceInput); |
| 644 | void* alignedInputPtr = inputData.get(); |
| 645 | CHECK(std::align(alignment, totalBytesInput, alignedInputPtr, spaceInput)); |
| 646 | |
| 647 | // Input with negative values |
| 648 | auto* inputPtr = reinterpret_cast<float*>(alignedInputPtr); |
| 649 | inputPtr[0] = -37.5f; |
| 650 | inputPtr[1] = -15.2f; |
| 651 | inputPtr[2] = -8.76f; |
| 652 | inputPtr[3] = -2.0f; |
| 653 | inputPtr[4] = -1.5f; |
| 654 | inputPtr[5] = -1.3f; |
| 655 | inputPtr[6] = -0.5f; |
| 656 | inputPtr[7] = -0.4f; |
| 657 | inputPtr[8] = 0.0f; |
| 658 | inputPtr[9] = 1.0f; |
| 659 | inputPtr[10] = 0.4f; |
| 660 | inputPtr[11] = 0.5f; |
| 661 | inputPtr[12] = 1.3f; |
| 662 | inputPtr[13] = 1.5f; |
| 663 | inputPtr[14] = 2.0f; |
| 664 | inputPtr[15] = 8.76f; |
| 665 | inputPtr[16] = 15.2f; |
| 666 | inputPtr[17] = 37.5f; |
| 667 | |
| 668 | auto outputData = std::make_unique<uint8_t[]>(spaceOutput); |
| 669 | void* alignedOutputPtr = outputData.get(); |
| 670 | CHECK(std::align(alignment, totalBytesOutput, alignedOutputPtr, spaceOutput)); |
| 671 | auto* outputPtr = reinterpret_cast<Half*>(alignedOutputPtr); |
| 672 | std::fill_n(outputPtr, numElements, -10.0f); |
| 673 | |
| 674 | TensorInfo inputTensorInfo = runtime->GetInputTensorInfo(netId, 0); |
| 675 | inputTensorInfo.SetConstant(true); |
| 676 | InputTensors inputTensors |
| 677 | { |
| 678 | {0,armnn::ConstTensor(inputTensorInfo, alignedInputPtr)}, |
| 679 | }; |
| 680 | OutputTensors outputTensors |
| 681 | { |
| 682 | {0,armnn::Tensor(runtime->GetOutputTensorInfo(netId, 0), alignedOutputPtr)} |
| 683 | }; |
| 684 | |
| 685 | runtime->GetProfiler(netId)->EnableProfiling(true); |
| 686 | |
| 687 | INFO("Run ImportInputs"); |
| 688 | std::vector<ImportedInputId> importedInputIds = |
| 689 | runtime->ImportInputs(netId, inputTensors, MemorySource::Malloc); |
Colm Donelan | d7ceec5 | 2022-07-06 12:09:05 +0100 | [diff] [blame] | 690 | // We expect the import to have succeeded. |
| 691 | CHECK(importedInputIds.size() == 1); |
Cathal Corbett | a3f4fba | 2022-03-21 09:27:08 +0000 | [diff] [blame] | 692 | std::vector<ImportedOutputId> importedOutputIds = |
| 693 | runtime->ImportOutputs(netId, outputTensors, MemorySource::Malloc); |
Colm Donelan | d7ceec5 | 2022-07-06 12:09:05 +0100 | [diff] [blame] | 694 | // We expect the import to have succeeded. |
| 695 | CHECK(importedOutputIds.size() == 1); |
Cathal Corbett | a3f4fba | 2022-03-21 09:27:08 +0000 | [diff] [blame] | 696 | |
| 697 | // Do the inference |
Colm Donelan | d7ceec5 | 2022-07-06 12:09:05 +0100 | [diff] [blame] | 698 | runtime->EnqueueWorkload(netId, InputTensors(), OutputTensors(), importedInputIds, importedOutputIds); |
Cathal Corbett | a3f4fba | 2022-03-21 09:27:08 +0000 | [diff] [blame] | 699 | |
| 700 | // Retrieve the Profiler.Print() output to get the workload execution |
| 701 | ProfilerManager& profilerManager = armnn::ProfilerManager::GetInstance(); |
| 702 | std::stringstream ss; |
| 703 | profilerManager.GetProfiler()->Print(ss);; |
| 704 | std::string dump = ss.str(); |
| 705 | |
| 706 | // Contains Convolution2dWorkload |
| 707 | std::size_t found = dump.find("ConvertFp32ToFp16Workload"); |
| 708 | CHECK(found != std::string::npos); |
| 709 | |
| 710 | // Contains SyncMemGeneric |
| 711 | found = dump.find("SyncMemGeneric"); |
| 712 | CHECK(found != std::string::npos); |
| 713 | |
| 714 | // Does not contain CopyMemGeneric |
| 715 | found = dump.find("CopyMemGeneric"); |
| 716 | CHECK(found == std::string::npos); |
| 717 | |
| 718 | runtime->UnloadNetwork(netId); |
| 719 | |
| 720 | // Check output is as expected |
| 721 | // Validate result by checking that the output has no negative values |
| 722 | auto* outputResult = reinterpret_cast<Half*>(alignedOutputPtr); |
| 723 | CHECK(outputResult); |
| 724 | |
| 725 | // Check the output is correct |
| 726 | CHECK(std::equal(outputResult, outputResult + numElements, expectedOutput.begin(), expectedOutput.end())); |
| 727 | } |
| 728 | |
| 729 | TEST_CASE_FIXTURE(ClContextControlFixture, "ClForceImportSimpleConvertFp32toFp16EndToEnd") |
| 730 | { |
| 731 | using namespace half_float::literal; |
| 732 | |
| 733 | // Create runtime in which test will run |
| 734 | IRuntime::CreationOptions options; |
| 735 | IRuntimePtr runtime(armnn::IRuntime::Create(options)); |
| 736 | |
| 737 | // build up the structure of the network |
| 738 | NetworkImpl network; |
| 739 | |
| 740 | armnn::TensorInfo inputInfo({1}, armnn::DataType::Float32); |
| 741 | armnn::TensorInfo outputTensorInfo({1}, armnn::DataType::Float16); |
| 742 | |
| 743 | std::vector<Half> expectedOutput = { 1.0_h }; |
| 744 | |
| 745 | unsigned int numElements = inputInfo.GetNumElements(); |
| 746 | size_t totalBytesInput = numElements * sizeof(float); |
| 747 | size_t totalBytesOutput = numElements * sizeof(Half); |
| 748 | |
| 749 | IConnectableLayer* const inputLayer = network.AddInputLayer(0, "input"); |
| 750 | ARMNN_ASSERT(inputLayer); |
| 751 | |
| 752 | armnn::IConnectableLayer* const convLayer = network.AddConvertFp32ToFp16Layer("convert"); |
| 753 | ARMNN_ASSERT(convLayer); |
| 754 | |
| 755 | inputLayer->GetOutputSlot(0).Connect(convLayer->GetInputSlot(0)); |
| 756 | inputLayer->GetOutputSlot(0).SetTensorInfo(inputInfo); |
| 757 | |
| 758 | IConnectableLayer* output = network.AddOutputLayer(0, "output"); |
| 759 | convLayer->GetOutputSlot(0).Connect(output->GetInputSlot(0)); |
| 760 | convLayer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo); |
| 761 | |
| 762 | // Optimize the network |
John Mcloughlin | c5ee0d7 | 2023-03-24 12:07:25 +0000 | [diff] [blame] | 763 | OptimizerOptionsOpaque optOptions; |
| 764 | optOptions.SetImportEnabled(false); |
| 765 | optOptions.SetExportEnabled(false); |
Cathal Corbett | a3f4fba | 2022-03-21 09:27:08 +0000 | [diff] [blame] | 766 | std::vector<armnn::BackendId> backends = {armnn::Compute::GpuAcc}; |
| 767 | IOptimizedNetworkPtr optNet = Optimize(network.GetGraph(), backends, runtime->GetDeviceSpec(), optOptions); |
| 768 | CHECK(optNet); |
| 769 | |
| 770 | // Loads it into the runtime. |
| 771 | NetworkId netId; |
| 772 | std::string ignoredErrorMessage; |
| 773 | // Enable Importing |
| 774 | INetworkProperties networkProperties(false, MemorySource::Undefined, MemorySource::Undefined); |
| 775 | runtime->LoadNetwork(netId, std::move(optNet), ignoredErrorMessage, networkProperties); |
| 776 | |
| 777 | // Creates structures for input & output |
| 778 | const size_t alignment = |
| 779 | arm_compute::CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE>(); |
| 780 | size_t spaceInput = totalBytesInput + alignment + alignment; |
| 781 | size_t spaceOutput = totalBytesOutput + alignment + alignment; |
| 782 | auto inputData = std::make_unique<uint8_t[]>(spaceInput); |
| 783 | void* alignedInputPtr = inputData.get(); |
| 784 | CHECK(std::align(alignment, totalBytesInput, alignedInputPtr, spaceInput)); |
| 785 | |
| 786 | // Input with negative values |
| 787 | auto* inputPtr = reinterpret_cast<float*>(alignedInputPtr); |
| 788 | inputPtr[0] = 1.0f; |
| 789 | |
| 790 | auto outputData = std::make_unique<uint8_t[]>(spaceOutput); |
| 791 | void* alignedOutputPtr = outputData.get(); |
| 792 | CHECK(std::align(alignment, totalBytesOutput, alignedOutputPtr, spaceOutput)); |
| 793 | auto* outputPtr = reinterpret_cast<Half*>(alignedOutputPtr); |
| 794 | std::fill_n(outputPtr, numElements, -10.0f); |
| 795 | |
| 796 | TensorInfo inputTensorInfo = runtime->GetInputTensorInfo(netId, 0); |
| 797 | inputTensorInfo.SetConstant(true); |
| 798 | InputTensors inputTensors |
| 799 | { |
| 800 | {0,armnn::ConstTensor(inputTensorInfo, alignedInputPtr)}, |
| 801 | }; |
| 802 | OutputTensors outputTensors |
| 803 | { |
| 804 | {0,armnn::Tensor(runtime->GetOutputTensorInfo(netId, 0), alignedOutputPtr)} |
| 805 | }; |
| 806 | |
| 807 | runtime->GetProfiler(netId)->EnableProfiling(true); |
| 808 | |
| 809 | INFO("Run ImportInputs"); |
| 810 | std::vector<ImportedInputId> importedInputIds = |
| 811 | runtime->ImportInputs(netId, inputTensors, MemorySource::Malloc); |
Colm Donelan | d7ceec5 | 2022-07-06 12:09:05 +0100 | [diff] [blame] | 812 | CHECK(importedInputIds.size() == 1); |
Cathal Corbett | a3f4fba | 2022-03-21 09:27:08 +0000 | [diff] [blame] | 813 | std::vector<ImportedOutputId> importedOutputIds = |
| 814 | runtime->ImportOutputs(netId, outputTensors, MemorySource::Malloc); |
Colm Donelan | d7ceec5 | 2022-07-06 12:09:05 +0100 | [diff] [blame] | 815 | CHECK(importedOutputIds.size() == 1); |
Cathal Corbett | a3f4fba | 2022-03-21 09:27:08 +0000 | [diff] [blame] | 816 | |
| 817 | // Do the inference |
Colm Donelan | d7ceec5 | 2022-07-06 12:09:05 +0100 | [diff] [blame] | 818 | runtime->EnqueueWorkload(netId, InputTensors(), OutputTensors(), importedInputIds, importedOutputIds); |
Cathal Corbett | a3f4fba | 2022-03-21 09:27:08 +0000 | [diff] [blame] | 819 | |
| 820 | // Retrieve the Profiler.Print() output to get the workload execution |
| 821 | ProfilerManager& profilerManager = armnn::ProfilerManager::GetInstance(); |
| 822 | std::stringstream ss; |
| 823 | profilerManager.GetProfiler()->Print(ss);; |
| 824 | std::string dump = ss.str(); |
| 825 | |
| 826 | // Contains Convolution2dWorkload |
| 827 | std::size_t found = dump.find("ConvertFp32ToFp16Workload"); |
| 828 | CHECK(found != std::string::npos); |
| 829 | |
| 830 | // Contains SyncMemGeneric |
| 831 | found = dump.find("SyncMemGeneric"); |
| 832 | CHECK(found != std::string::npos); |
| 833 | |
| 834 | // Does not contain CopyMemGeneric |
| 835 | found = dump.find("CopyMemGeneric"); |
| 836 | CHECK(found == std::string::npos); |
| 837 | |
| 838 | runtime->UnloadNetwork(netId); |
| 839 | |
| 840 | // Check output is as expected |
| 841 | // Validate result by checking that the output has no negative values |
| 842 | auto* outputResult = reinterpret_cast<Half*>(alignedOutputPtr); |
| 843 | CHECK(outputResult); |
| 844 | |
| 845 | // Check the output is correct |
| 846 | CHECK(std::equal(outputResult, outputResult + numElements, expectedOutput.begin(), expectedOutput.end())); |
| 847 | } |
| 848 | |
David Monahan | 041f17a | 2022-03-03 10:56:17 +0000 | [diff] [blame] | 849 | TEST_CASE_FIXTURE(ClContextControlFixture, "ClForceImportRepeatedInferencesEndToEndTest") |
| 850 | { |
| 851 | /* |
| 852 | * This is a test to check the functionality of the Forced Import functionality when using repeated inferences that |
| 853 | * require switching from importing to copy. For the first inference we create aligned Pointers and check they are |
Colm Donelan | d7ceec5 | 2022-07-06 12:09:05 +0100 | [diff] [blame] | 854 | * imported correctly. For the second we use similar pointers but don't use PreImporting. |
David Monahan | 041f17a | 2022-03-03 10:56:17 +0000 | [diff] [blame] | 855 | */ |
| 856 | // Create runtime in which test will run |
| 857 | IRuntime::CreationOptions options; |
| 858 | IRuntimePtr runtime(armnn::IRuntime::Create(options)); |
| 859 | |
| 860 | // build up the structure of the network |
| 861 | INetworkPtr network(INetwork::Create()); |
| 862 | |
| 863 | armnn::TensorInfo inputInfo({ 1, 3, 4, 1 }, DataType::Float32); |
| 864 | armnn::TensorInfo kernelInfo({ 1, 3, 3, 1 }, DataType::Float32); |
| 865 | armnn::TensorInfo outputInfo({ 1, 3, 4, 1 }, DataType::Float32); |
| 866 | |
| 867 | kernelInfo.SetConstant(true); |
| 868 | |
| 869 | std::vector<float> kernel = |
| 870 | { |
| 871 | 4, 5, 6, |
| 872 | 0, 0, 0, |
| 873 | 3, 2, 1 |
| 874 | }; |
| 875 | |
| 876 | const std::vector<float> expectedOutput = |
| 877 | { |
| 878 | 23, 41, 33, 21, |
| 879 | 44, 65, 76, 52, |
| 880 | 82, 85, 79, 42 |
| 881 | }; |
| 882 | |
| 883 | unsigned int numElements = inputInfo.GetNumElements(); |
| 884 | size_t totalBytes = numElements * sizeof(float); |
| 885 | |
| 886 | IConnectableLayer* const inputLayer = network->AddInputLayer(0, "input"); |
| 887 | ARMNN_ASSERT(inputLayer); |
| 888 | |
| 889 | armnn::ConstTensor weights(kernelInfo, kernel); |
| 890 | |
| 891 | armnn::Convolution2dDescriptor convDesc2d; |
| 892 | convDesc2d.m_StrideX = 1; |
| 893 | convDesc2d.m_StrideY = 1; |
| 894 | convDesc2d.m_PadLeft = 1; |
| 895 | convDesc2d.m_PadRight = 1; |
| 896 | convDesc2d.m_PadTop = 1; |
| 897 | convDesc2d.m_PadBottom = 1; |
| 898 | convDesc2d.m_DataLayout = DataLayout::NHWC; |
Keith Davis | 721e629 | 2022-05-17 10:06:53 +0100 | [diff] [blame] | 899 | armnn::IConnectableLayer* const convLayer = network->AddConvolution2dLayer(convDesc2d, "conv"); |
David Monahan | 041f17a | 2022-03-03 10:56:17 +0000 | [diff] [blame] | 900 | ARMNN_ASSERT(convLayer); |
| 901 | |
Keith Davis | 721e629 | 2022-05-17 10:06:53 +0100 | [diff] [blame] | 902 | armnn::IConnectableLayer* weightsLayer = network->AddConstantLayer(weights); |
| 903 | |
| 904 | weightsLayer->GetOutputSlot(0).SetTensorInfo(weights.GetInfo()); |
| 905 | weightsLayer->GetOutputSlot(0).Connect(convLayer->GetInputSlot(1u)); |
| 906 | |
David Monahan | 041f17a | 2022-03-03 10:56:17 +0000 | [diff] [blame] | 907 | inputLayer->GetOutputSlot(0).Connect(convLayer->GetInputSlot(0)); |
| 908 | inputLayer->GetOutputSlot(0).SetTensorInfo(inputInfo); |
| 909 | |
| 910 | IConnectableLayer* output = network->AddOutputLayer(0, "output"); |
| 911 | convLayer->GetOutputSlot(0).Connect(output->GetInputSlot(0)); |
| 912 | convLayer->GetOutputSlot(0).SetTensorInfo(outputInfo); |
| 913 | |
| 914 | // Optimize the network |
John Mcloughlin | c5ee0d7 | 2023-03-24 12:07:25 +0000 | [diff] [blame] | 915 | OptimizerOptionsOpaque optOptions; |
| 916 | optOptions.SetImportEnabled(false); |
| 917 | optOptions.SetExportEnabled(false); |
David Monahan | 041f17a | 2022-03-03 10:56:17 +0000 | [diff] [blame] | 918 | std::vector<armnn::BackendId> backends = {armnn::Compute::GpuAcc}; |
| 919 | IOptimizedNetworkPtr optNet = Optimize(*network, backends, runtime->GetDeviceSpec(), optOptions); |
| 920 | CHECK(optNet); |
| 921 | |
| 922 | // Loads it into the runtime. |
| 923 | NetworkId netId; |
| 924 | std::string ignoredErrorMessage; |
| 925 | // Enable Importing |
| 926 | INetworkProperties networkProperties(false, MemorySource::Undefined, MemorySource::Undefined); |
| 927 | runtime->LoadNetwork(netId, std::move(optNet), ignoredErrorMessage, networkProperties); |
| 928 | |
| 929 | // Creates structures for input & output |
| 930 | const size_t alignment = |
| 931 | arm_compute::CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE>(); |
| 932 | size_t space = totalBytes + alignment + alignment; |
| 933 | auto inputData = std::make_unique<uint8_t[]>(space); |
| 934 | void* alignedInputPtr = inputData.get(); |
| 935 | CHECK(std::align(alignment, totalBytes, alignedInputPtr, space)); |
| 936 | |
| 937 | // Fill input with values |
| 938 | auto* inputPtr = reinterpret_cast<float*>(alignedInputPtr); |
| 939 | inputPtr[0] = 1; |
| 940 | inputPtr[1] = 5; |
| 941 | inputPtr[2] = 2; |
| 942 | inputPtr[3] = 3; |
| 943 | inputPtr[4] = 8; |
| 944 | inputPtr[5] = 7; |
| 945 | inputPtr[6] = 3; |
| 946 | inputPtr[7] = 6; |
| 947 | inputPtr[8] = 3; |
| 948 | inputPtr[9] = 3; |
| 949 | inputPtr[10] = 9; |
| 950 | inputPtr[11] = 1; |
| 951 | |
| 952 | |
| 953 | auto outputData = std::make_unique<uint8_t[]>(space); |
| 954 | void* alignedOutputPtr = outputData.get(); |
| 955 | CHECK(std::align(alignment, totalBytes, alignedOutputPtr, space)); |
| 956 | auto* outputPtr = reinterpret_cast<float*>(alignedOutputPtr); |
| 957 | std::fill_n(outputPtr, numElements, -10.0f); |
| 958 | |
| 959 | TensorInfo inputTensorInfo = runtime->GetInputTensorInfo(netId, 0); |
| 960 | inputTensorInfo.SetConstant(true); |
| 961 | InputTensors inputTensors |
| 962 | { |
| 963 | {0,armnn::ConstTensor(inputTensorInfo, alignedInputPtr)}, |
| 964 | }; |
| 965 | OutputTensors outputTensors |
| 966 | { |
| 967 | {0,armnn::Tensor(runtime->GetOutputTensorInfo(netId, 0), alignedOutputPtr)} |
| 968 | }; |
| 969 | |
| 970 | runtime->GetProfiler(netId)->EnableProfiling(true); |
| 971 | |
| 972 | INFO("Run ImportInputs"); |
| 973 | std::vector<ImportedInputId> importedInputIds = |
| 974 | runtime->ImportInputs(netId, inputTensors, MemorySource::Malloc); |
Colm Donelan | d7ceec5 | 2022-07-06 12:09:05 +0100 | [diff] [blame] | 975 | // We expect the import to have succeeded. |
| 976 | CHECK(importedInputIds.size() == 1); |
David Monahan | 041f17a | 2022-03-03 10:56:17 +0000 | [diff] [blame] | 977 | std::vector<ImportedOutputId> importedOutputIds = |
| 978 | runtime->ImportOutputs(netId, outputTensors, MemorySource::Malloc); |
Colm Donelan | d7ceec5 | 2022-07-06 12:09:05 +0100 | [diff] [blame] | 979 | // We expect the import to have succeeded. |
| 980 | CHECK(importedOutputIds.size() == 1); |
David Monahan | 041f17a | 2022-03-03 10:56:17 +0000 | [diff] [blame] | 981 | |
| 982 | // Do the inference |
Colm Donelan | d7ceec5 | 2022-07-06 12:09:05 +0100 | [diff] [blame] | 983 | runtime->EnqueueWorkload(netId, InputTensors(), OutputTensors(), importedInputIds, importedOutputIds); |
David Monahan | 041f17a | 2022-03-03 10:56:17 +0000 | [diff] [blame] | 984 | |
| 985 | // Retrieve the Profiler.AnalyzeEventsAndWriteResults() output to get the workload execution |
| 986 | ProfilerManager& profilerManager = armnn::ProfilerManager::GetInstance(); |
| 987 | std::stringstream ss; |
| 988 | profilerManager.GetProfiler()->AnalyzeEventsAndWriteResults(ss); |
| 989 | std::string dump = ss.str(); |
| 990 | |
| 991 | // Contains Convolution2dWorkload |
| 992 | std::size_t found = dump.find("Convolution2dWorkload"); |
| 993 | CHECK(found != std::string::npos); |
| 994 | |
| 995 | // Contains SyncMemGeneric |
| 996 | found = dump.find("SyncMemGeneric"); |
| 997 | CHECK(found != std::string::npos); |
| 998 | |
| 999 | // Does not contain CopyMemGeneric |
| 1000 | found = dump.find("CopyMemGeneric"); |
| 1001 | CHECK(found == std::string::npos); |
| 1002 | |
| 1003 | // Sync the outputs so we can read the data |
| 1004 | arm_compute::CLScheduler::get().sync(); |
| 1005 | |
| 1006 | // Check output is as expected |
| 1007 | auto* outputResult = reinterpret_cast<float*>(alignedOutputPtr); |
| 1008 | CHECK(outputResult); |
| 1009 | CHECK(std::equal(outputResult, outputResult + numElements, expectedOutput.begin(), expectedOutput.end())); |
| 1010 | |
| 1011 | // Repeat the inference, with new tensors and without using PreImporting to force it to fall back to copying |
| 1012 | |
| 1013 | // Creates structures for input & output |
| 1014 | auto inputDataCopy = std::make_unique<uint8_t[]>(space); |
| 1015 | void* copyInputPtr = inputDataCopy.get(); |
| 1016 | |
| 1017 | // Fill input with values |
| 1018 | auto* inputCopyPtr = reinterpret_cast<float*>(copyInputPtr); |
| 1019 | inputCopyPtr[0] = 1; |
| 1020 | inputCopyPtr[1] = 5; |
| 1021 | inputCopyPtr[2] = 2; |
| 1022 | inputCopyPtr[3] = 3; |
| 1023 | inputCopyPtr[4] = 8; |
| 1024 | inputCopyPtr[5] = 7; |
| 1025 | inputCopyPtr[6] = 3; |
| 1026 | inputCopyPtr[7] = 6; |
| 1027 | inputCopyPtr[8] = 3; |
| 1028 | inputCopyPtr[9] = 3; |
| 1029 | inputCopyPtr[10] = 9; |
| 1030 | inputCopyPtr[11] = 1; |
| 1031 | |
| 1032 | // Output pre-filled with -10.0f |
| 1033 | auto outputDataCopy = std::make_unique<uint8_t[]>(space); |
| 1034 | void* copyOutputPtr = outputDataCopy.get(); |
| 1035 | auto* outputCopyPtr = reinterpret_cast<float*>(copyOutputPtr); |
| 1036 | std::fill_n(outputCopyPtr, numElements, -10.0f); |
| 1037 | |
| 1038 | InputTensors inputTensorsCopy |
| 1039 | { |
| 1040 | {0,armnn::ConstTensor(inputTensorInfo, copyInputPtr)}, |
| 1041 | }; |
| 1042 | OutputTensors outputTensorsCopy |
| 1043 | { |
| 1044 | {0,armnn::Tensor(runtime->GetOutputTensorInfo(netId, 0), copyOutputPtr)} |
| 1045 | }; |
| 1046 | |
| 1047 | // Do the inference without any pre-imported input/output ids |
| 1048 | runtime->EnqueueWorkload(netId, inputTensorsCopy, outputTensorsCopy); |
| 1049 | // Sync the outputs so we can read the data |
| 1050 | arm_compute::CLScheduler::get().sync(); |
| 1051 | |
| 1052 | // Check the output is correct |
| 1053 | outputResult = reinterpret_cast<float*>(copyOutputPtr); |
| 1054 | CHECK(outputResult); |
| 1055 | CHECK(std::equal(outputResult, outputResult + numElements, expectedOutput.begin(), expectedOutput.end())); |
| 1056 | |
| 1057 | // Query the profiler again, this will contain the results of both inferences |
| 1058 | profilerManager.GetProfiler()->AnalyzeEventsAndWriteResults(ss); |
| 1059 | dump = ss.str(); |
| 1060 | |
| 1061 | // Contains Convolution2dWorkload |
| 1062 | found = dump.find("Convolution2dWorkload"); |
| 1063 | CHECK(found != std::string::npos); |
| 1064 | |
| 1065 | // Should still contain the SyncMemGeneric |
| 1066 | found = dump.find("SyncMemGeneric"); |
| 1067 | CHECK(found != std::string::npos); |
| 1068 | |
| 1069 | // Should now also contain a CopyMemGeneric |
| 1070 | found = dump.find("CopyMemGeneric"); |
| 1071 | CHECK(found != std::string::npos); |
| 1072 | runtime->UnloadNetwork(netId); |
| 1073 | } |
| 1074 | |
| 1075 | TEST_CASE_FIXTURE(ClContextControlFixture, "ClForceImportRepeatedInferencesInvertedEndToEndTest") |
| 1076 | { |
| 1077 | /* |
| 1078 | * This test is similar to the test above but instead of importing and then copying, we start by copying and then do |
| 1079 | * the import. |
| 1080 | */ |
| 1081 | // Create runtime in which test will run |
| 1082 | IRuntime::CreationOptions options; |
| 1083 | IRuntimePtr runtime(armnn::IRuntime::Create(options)); |
| 1084 | |
| 1085 | // build up the structure of the network |
| 1086 | INetworkPtr network(INetwork::Create()); |
| 1087 | |
| 1088 | armnn::TensorInfo inputInfo({ 1, 3, 4, 1 }, DataType::Float32); |
| 1089 | armnn::TensorInfo kernelInfo({ 1, 3, 3, 1 }, DataType::Float32); |
| 1090 | armnn::TensorInfo outputInfo({ 1, 3, 4, 1 }, DataType::Float32); |
| 1091 | |
| 1092 | kernelInfo.SetConstant(true); |
| 1093 | |
| 1094 | std::vector<float> kernel = |
| 1095 | { |
| 1096 | 4, 5, 6, |
| 1097 | 0, 0, 0, |
| 1098 | 3, 2, 1 |
| 1099 | }; |
| 1100 | |
| 1101 | const std::vector<float> expectedOutput = |
| 1102 | { |
| 1103 | 23, 41, 33, 21, |
| 1104 | 44, 65, 76, 52, |
| 1105 | 82, 85, 79, 42 |
| 1106 | }; |
| 1107 | |
| 1108 | unsigned int numElements = inputInfo.GetNumElements(); |
| 1109 | size_t totalBytes = numElements * sizeof(float); |
| 1110 | |
| 1111 | IConnectableLayer* const inputLayer = network->AddInputLayer(0, "input"); |
| 1112 | ARMNN_ASSERT(inputLayer); |
| 1113 | |
| 1114 | armnn::ConstTensor weights(kernelInfo, kernel); |
| 1115 | |
| 1116 | armnn::Convolution2dDescriptor convDesc2d; |
| 1117 | convDesc2d.m_StrideX = 1; |
| 1118 | convDesc2d.m_StrideY = 1; |
| 1119 | convDesc2d.m_PadLeft = 1; |
| 1120 | convDesc2d.m_PadRight = 1; |
| 1121 | convDesc2d.m_PadTop = 1; |
| 1122 | convDesc2d.m_PadBottom = 1; |
| 1123 | convDesc2d.m_DataLayout = DataLayout::NHWC; |
Keith Davis | 721e629 | 2022-05-17 10:06:53 +0100 | [diff] [blame] | 1124 | |
| 1125 | armnn::IConnectableLayer* const convLayer = network->AddConvolution2dLayer(convDesc2d, "conv"); |
David Monahan | 041f17a | 2022-03-03 10:56:17 +0000 | [diff] [blame] | 1126 | ARMNN_ASSERT(convLayer); |
| 1127 | |
Keith Davis | 721e629 | 2022-05-17 10:06:53 +0100 | [diff] [blame] | 1128 | armnn::IConnectableLayer* weightsLayer = network->AddConstantLayer(weights); |
| 1129 | |
| 1130 | weightsLayer->GetOutputSlot(0).SetTensorInfo(weights.GetInfo()); |
| 1131 | weightsLayer->GetOutputSlot(0).Connect(convLayer->GetInputSlot(1u)); |
| 1132 | |
David Monahan | 041f17a | 2022-03-03 10:56:17 +0000 | [diff] [blame] | 1133 | inputLayer->GetOutputSlot(0).Connect(convLayer->GetInputSlot(0)); |
| 1134 | inputLayer->GetOutputSlot(0).SetTensorInfo(inputInfo); |
| 1135 | |
| 1136 | IConnectableLayer* output = network->AddOutputLayer(0, "output"); |
| 1137 | convLayer->GetOutputSlot(0).Connect(output->GetInputSlot(0)); |
| 1138 | convLayer->GetOutputSlot(0).SetTensorInfo(outputInfo); |
| 1139 | |
| 1140 | // Optimize the network |
John Mcloughlin | c5ee0d7 | 2023-03-24 12:07:25 +0000 | [diff] [blame] | 1141 | OptimizerOptionsOpaque optOptions; |
| 1142 | optOptions.SetImportEnabled(false); |
| 1143 | optOptions.SetExportEnabled(false); |
David Monahan | 041f17a | 2022-03-03 10:56:17 +0000 | [diff] [blame] | 1144 | std::vector<armnn::BackendId> backends = {armnn::Compute::GpuAcc}; |
| 1145 | IOptimizedNetworkPtr optNet = Optimize(*network, backends, runtime->GetDeviceSpec(), optOptions); |
| 1146 | CHECK(optNet); |
| 1147 | |
| 1148 | // Loads it into the runtime. |
| 1149 | NetworkId netId; |
| 1150 | std::string ignoredErrorMessage; |
| 1151 | // Enable Importing |
| 1152 | INetworkProperties networkProperties(false, MemorySource::Undefined, MemorySource::Undefined); |
| 1153 | runtime->LoadNetwork(netId, std::move(optNet), ignoredErrorMessage, networkProperties); |
| 1154 | |
| 1155 | // Creates structures for input & output |
| 1156 | const size_t alignment = |
| 1157 | arm_compute::CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE>(); |
| 1158 | size_t space = totalBytes + alignment + alignment; |
| 1159 | auto inputData = std::make_unique<uint8_t[]>(space); |
| 1160 | void* copyInputPtr = inputData.get(); |
| 1161 | |
| 1162 | // Fill input with values |
| 1163 | auto* inputPtr = reinterpret_cast<float*>(copyInputPtr); |
| 1164 | inputPtr[0] = 1; |
| 1165 | inputPtr[1] = 5; |
| 1166 | inputPtr[2] = 2; |
| 1167 | inputPtr[3] = 3; |
| 1168 | inputPtr[4] = 8; |
| 1169 | inputPtr[5] = 7; |
| 1170 | inputPtr[6] = 3; |
| 1171 | inputPtr[7] = 6; |
| 1172 | inputPtr[8] = 3; |
| 1173 | inputPtr[9] = 3; |
| 1174 | inputPtr[10] = 9; |
| 1175 | inputPtr[11] = 1; |
| 1176 | |
| 1177 | // Create output buffer and fill it with -10.0f |
| 1178 | auto outputData = std::make_unique<uint8_t[]>(space); |
| 1179 | void* copyOutputPtr = outputData.get(); |
| 1180 | auto* outputPtr = reinterpret_cast<float*>(copyOutputPtr); |
| 1181 | std::fill_n(outputPtr, numElements, -10.0f); |
| 1182 | |
| 1183 | TensorInfo inputTensorInfo = runtime->GetInputTensorInfo(netId, 0); |
| 1184 | inputTensorInfo.SetConstant(true); |
| 1185 | InputTensors inputTensors |
| 1186 | { |
| 1187 | {0,armnn::ConstTensor(inputTensorInfo, copyInputPtr)}, |
| 1188 | }; |
| 1189 | OutputTensors outputTensors |
| 1190 | { |
| 1191 | {0,armnn::Tensor(runtime->GetOutputTensorInfo(netId, 0), copyOutputPtr)} |
| 1192 | }; |
| 1193 | |
| 1194 | runtime->GetProfiler(netId)->EnableProfiling(true); |
| 1195 | |
| 1196 | // Do the inference without any pre-imported inputs/outputs |
| 1197 | runtime->EnqueueWorkload(netId, inputTensors, outputTensors); |
| 1198 | |
| 1199 | // Retrieve the Profiler.AnalyzeEventsAndWriteResults() output to get the workload execution |
| 1200 | ProfilerManager& profilerManager = armnn::ProfilerManager::GetInstance(); |
| 1201 | std::stringstream ss; |
| 1202 | profilerManager.GetProfiler()->AnalyzeEventsAndWriteResults(ss); |
| 1203 | std::string dump = ss.str(); |
| 1204 | |
| 1205 | // Contains Convolution2dWorkload |
| 1206 | std::size_t found = dump.find("Convolution2dWorkload"); |
| 1207 | CHECK(found != std::string::npos); |
| 1208 | |
| 1209 | // Does not contain SyncMemGeneric |
| 1210 | found = dump.find("SyncMemGeneric"); |
| 1211 | CHECK(found == std::string::npos); |
| 1212 | |
| 1213 | // Does contain CopyMemGeneric |
| 1214 | found = dump.find("CopyMemGeneric"); |
| 1215 | CHECK(found != std::string::npos); |
| 1216 | |
| 1217 | // Sync the outputs so we can read the data |
| 1218 | arm_compute::CLScheduler::get().sync(); |
| 1219 | |
| 1220 | // Check output is as expected |
| 1221 | auto* outputResult = reinterpret_cast<float*>(copyOutputPtr); |
| 1222 | CHECK(outputResult); |
| 1223 | CHECK(std::equal(outputResult, outputResult + numElements, expectedOutput.begin(), expectedOutput.end())); |
| 1224 | |
| 1225 | // Repeat the inference, with new tensors and while using pre-importing to force it to import |
| 1226 | |
| 1227 | // Creates structures for input & output |
| 1228 | auto inputDataImport = std::make_unique<uint8_t[]>(space); |
| 1229 | void* alignedInputImportPtr = inputDataImport.get(); |
| 1230 | CHECK(std::align(alignment, totalBytes, alignedInputImportPtr, space)); |
| 1231 | |
| 1232 | // Fill input with values |
| 1233 | auto* inputImportPtr = reinterpret_cast<float*>(alignedInputImportPtr); |
| 1234 | inputImportPtr[0] = 1; |
| 1235 | inputImportPtr[1] = 5; |
| 1236 | inputImportPtr[2] = 2; |
| 1237 | inputImportPtr[3] = 3; |
| 1238 | inputImportPtr[4] = 8; |
| 1239 | inputImportPtr[5] = 7; |
| 1240 | inputImportPtr[6] = 3; |
| 1241 | inputImportPtr[7] = 6; |
| 1242 | inputImportPtr[8] = 3; |
| 1243 | inputImportPtr[9] = 3; |
| 1244 | inputImportPtr[10] = 9; |
| 1245 | inputImportPtr[11] = 1; |
| 1246 | |
| 1247 | // Output pre-filled with -10.0f |
| 1248 | auto outputDataImport = std::make_unique<uint8_t[]>(space); |
| 1249 | void* alignedOutputImportPtr = outputDataImport.get(); |
| 1250 | CHECK(std::align(alignment, totalBytes, alignedOutputImportPtr, space)); |
| 1251 | auto* outputImportPtr = reinterpret_cast<float*>(alignedOutputImportPtr); |
| 1252 | std::fill_n(outputImportPtr, numElements, -10.0f); |
| 1253 | |
| 1254 | InputTensors inputTensorsImport |
| 1255 | { |
| 1256 | {0,armnn::ConstTensor(inputTensorInfo, alignedInputImportPtr)}, |
| 1257 | }; |
| 1258 | OutputTensors outputTensorsImport |
| 1259 | { |
| 1260 | {0,armnn::Tensor(runtime->GetOutputTensorInfo(netId, 0), alignedOutputImportPtr)} |
| 1261 | }; |
| 1262 | |
| 1263 | INFO("Run ImportInputs"); |
| 1264 | std::vector<ImportedInputId> importedInputIds = |
| 1265 | runtime->ImportInputs(netId, inputTensorsImport, MemorySource::Malloc); |
Colm Donelan | d7ceec5 | 2022-07-06 12:09:05 +0100 | [diff] [blame] | 1266 | CHECK(importedInputIds.size() == 1); |
David Monahan | 041f17a | 2022-03-03 10:56:17 +0000 | [diff] [blame] | 1267 | std::vector<ImportedOutputId> importedOutputIds = |
| 1268 | runtime->ImportOutputs(netId, outputTensorsImport, MemorySource::Malloc); |
Colm Donelan | d7ceec5 | 2022-07-06 12:09:05 +0100 | [diff] [blame] | 1269 | CHECK(importedOutputIds.size() == 1); |
David Monahan | 041f17a | 2022-03-03 10:56:17 +0000 | [diff] [blame] | 1270 | |
| 1271 | // Do the inference with pre-imported inputs/outputs |
Colm Donelan | d7ceec5 | 2022-07-06 12:09:05 +0100 | [diff] [blame] | 1272 | runtime->EnqueueWorkload(netId, InputTensors(), OutputTensors(), importedInputIds, importedOutputIds); |
David Monahan | 041f17a | 2022-03-03 10:56:17 +0000 | [diff] [blame] | 1273 | // Sync the outputs so we can read the data |
| 1274 | arm_compute::CLScheduler::get().sync(); |
| 1275 | |
| 1276 | // Check the output is correct |
| 1277 | outputResult = reinterpret_cast<float*>(alignedOutputImportPtr); |
| 1278 | CHECK(outputResult); |
| 1279 | CHECK(std::equal(outputResult, outputResult + numElements, expectedOutput.begin(), expectedOutput.end())); |
| 1280 | |
| 1281 | |
| 1282 | // Query the profiler again, this will contain the results of both inferences |
| 1283 | profilerManager.GetProfiler()->AnalyzeEventsAndWriteResults(ss); |
| 1284 | dump = ss.str(); |
| 1285 | |
| 1286 | // Contains Convolution2dWorkload |
| 1287 | found = dump.find("Convolution2dWorkload"); |
| 1288 | CHECK(found != std::string::npos); |
| 1289 | |
| 1290 | // Should now contain the SyncMemGeneric |
| 1291 | found = dump.find("SyncMemGeneric"); |
| 1292 | CHECK(found != std::string::npos); |
| 1293 | |
| 1294 | // Should still contain a CopyMemGeneric from the first inference |
| 1295 | found = dump.find("CopyMemGeneric"); |
| 1296 | CHECK(found != std::string::npos); |
| 1297 | runtime->UnloadNetwork(netId); |
| 1298 | } |
| 1299 | |
Sadik Armagan | 1625efc | 2021-06-10 18:24:34 +0100 | [diff] [blame] | 1300 | } |