Laurent Carlier | 749294b | 2020-06-01 09:03:17 +0100 | [diff] [blame] | 1 | // |
Teresa Charlin | 8398edc | 2020-07-20 14:23:02 +0100 | [diff] [blame] | 2 | // Copyright © 2017 Arm Ltd and Contributors. All rights reserved. |
David Beck | ecb56cd | 2018-09-05 12:52:57 +0100 | [diff] [blame] | 3 | // SPDX-License-Identifier: MIT |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 4 | // |
| 5 | #include "ClWorkloadFactory.hpp" |
David Beck | 79141b9 | 2018-10-23 16:09:36 +0100 | [diff] [blame] | 6 | #include "ClBackendId.hpp" |
Sadik Armagan | 04a7297 | 2020-09-14 15:44:18 +0100 | [diff] [blame] | 7 | #include "ClBackendModelContext.hpp" |
Matthew Sloyan | 80fbcd5 | 2021-01-07 13:28:47 +0000 | [diff] [blame] | 8 | #include "ClContextDeserializer.hpp" |
| 9 | #include "ClContextSerializer.hpp" |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 10 | |
Aron Virginas-Tar | 5caf907 | 2018-11-14 18:35:18 +0000 | [diff] [blame] | 11 | #include <Layer.hpp> |
| 12 | |
David Beck | 0dbe0ee | 2018-09-24 15:59:27 +0100 | [diff] [blame] | 13 | #include <armnn/Exceptions.hpp> |
Sadik Armagan | 6b9eba2 | 2021-11-01 07:27:50 +0000 | [diff] [blame] | 14 | #include <armnn/Logging.hpp> |
David Beck | 0dbe0ee | 2018-09-24 15:59:27 +0100 | [diff] [blame] | 15 | #include <armnn/Utils.hpp> |
Jan Eilers | 8eb2560 | 2020-03-09 12:13:48 +0000 | [diff] [blame] | 16 | #include <armnn/utility/IgnoreUnused.hpp> |
Matthew Sloyan | 171214c | 2020-09-09 09:07:37 +0100 | [diff] [blame] | 17 | #include <armnn/utility/NumericCast.hpp> |
Jan Eilers | bb446e5 | 2020-04-02 13:56:54 +0100 | [diff] [blame] | 18 | #include <armnn/utility/PolymorphicDowncast.hpp> |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 19 | |
Aron Virginas-Tar | 5caf907 | 2018-11-14 18:35:18 +0000 | [diff] [blame] | 20 | #include <backendsCommon/MakeWorkloadHelper.hpp> |
Colm Donelan | 0c47974 | 2021-12-10 12:43:54 +0000 | [diff] [blame] | 21 | #include <armnn/backends/MemCopyWorkload.hpp> |
Derek Lamberti | f674aa0 | 2019-08-01 15:56:25 +0100 | [diff] [blame] | 22 | #include <backendsCommon/MemImportWorkload.hpp> |
Colm Donelan | 0c47974 | 2021-12-10 12:43:54 +0000 | [diff] [blame] | 23 | #include <armnn/backends/TensorHandle.hpp> |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 24 | |
Aron Virginas-Tar | c9cc804 | 2018-11-01 16:15:57 +0000 | [diff] [blame] | 25 | #include <cl/ClTensorHandle.hpp> |
| 26 | #include <cl/workloads/ClWorkloads.hpp> |
| 27 | #include <cl/workloads/ClWorkloadUtils.hpp> |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 28 | |
Aron Virginas-Tar | 5caf907 | 2018-11-14 18:35:18 +0000 | [diff] [blame] | 29 | #include <arm_compute/core/CL/CLKernelLibrary.h> |
| 30 | #include <arm_compute/runtime/CL/CLBufferAllocator.h> |
| 31 | #include <arm_compute/runtime/CL/CLScheduler.h> |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 32 | |
Rob Hughes | 9542f90 | 2021-07-14 09:48:54 +0100 | [diff] [blame] | 33 | #include <armnnUtils/Filesystem.hpp> |
Matthew Sloyan | 80fbcd5 | 2021-01-07 13:28:47 +0000 | [diff] [blame] | 34 | #include <fstream> |
Sadik Armagan | dea8fb6 | 2020-11-26 10:38:11 +0000 | [diff] [blame] | 35 | |
Sadik Armagan | b7851f9 | 2021-10-06 16:37:02 +0100 | [diff] [blame] | 36 | #include <sys/stat.h> |
| 37 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 38 | namespace armnn |
| 39 | { |
| 40 | |
David Beck | 79141b9 | 2018-10-23 16:09:36 +0100 | [diff] [blame] | 41 | namespace |
| 42 | { |
| 43 | static const BackendId s_Id{ClBackendId()}; |
| 44 | } |
| 45 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 46 | bool ClWorkloadFactory::IsLayerSupported(const Layer& layer, |
David Beck | 29c75de | 2018-10-23 13:35:58 +0100 | [diff] [blame] | 47 | Optional<DataType> dataType, |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 48 | std::string& outReasonIfUnsupported) |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 49 | { |
David Beck | 79141b9 | 2018-10-23 16:09:36 +0100 | [diff] [blame] | 50 | return IWorkloadFactory::IsLayerSupported(s_Id, layer, dataType, outReasonIfUnsupported); |
| 51 | } |
| 52 | |
Sadik Armagan | 04a7297 | 2020-09-14 15:44:18 +0100 | [diff] [blame] | 53 | bool ClWorkloadFactory::IsLayerSupported(const IConnectableLayer& layer, |
| 54 | Optional<DataType> dataType, |
| 55 | std::string& outReasonIfUnsupported, |
| 56 | const ModelOptions& modelOptions) |
| 57 | { |
| 58 | return IWorkloadFactory::IsLayerSupported(s_Id, layer, dataType, outReasonIfUnsupported, modelOptions); |
| 59 | } |
| 60 | |
David Beck | 79141b9 | 2018-10-23 16:09:36 +0100 | [diff] [blame] | 61 | const BackendId& ClWorkloadFactory::GetBackendId() const |
| 62 | { |
| 63 | return s_Id; |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 64 | } |
| 65 | |
Sadik Armagan | dea8fb6 | 2020-11-26 10:38:11 +0000 | [diff] [blame] | 66 | void ClWorkloadFactory::AfterWorkloadsCreated() |
| 67 | { |
| 68 | if(m_ModelContextPtr) |
| 69 | { |
| 70 | auto modelOptions = dynamic_cast<ClBackendModelContext*>(m_ModelContextPtr.get()); |
| 71 | if (modelOptions->SaveCachedNetwork()) |
| 72 | { |
Sadik Armagan | b7851f9 | 2021-10-06 16:37:02 +0100 | [diff] [blame] | 73 | ClContextSerializer serializer; |
| 74 | serializer.Serialize(m_CLCompileContext); |
| 75 | auto cachedFd = modelOptions->GetCachedFileDescriptor(); |
| 76 | if (cachedFd != -1) |
| 77 | { |
| 78 | std::vector<uint8_t> compiledContextData; |
| 79 | std::stringstream stream; |
| 80 | bool serialized = serializer.SaveSerializedToStream(stream); |
| 81 | if (serialized) |
| 82 | { |
| 83 | std::string const serializedString{stream.str()}; |
| 84 | std::copy(serializedString.begin(), |
| 85 | serializedString.end(), |
| 86 | std::back_inserter(compiledContextData)); |
Sadik Armagan | 6b9eba2 | 2021-11-01 07:27:50 +0000 | [diff] [blame] | 87 | auto success = write(cachedFd, compiledContextData.data(), compiledContextData.size()); |
| 88 | if (success == -1) |
| 89 | { |
| 90 | ARMNN_LOG(info) << "ClWorkloadFactory:: Could not cache the compiled context!"; |
| 91 | } |
Sadik Armagan | b7851f9 | 2021-10-06 16:37:02 +0100 | [diff] [blame] | 92 | } |
| 93 | } |
| 94 | |
Sadik Armagan | dea8fb6 | 2020-11-26 10:38:11 +0000 | [diff] [blame] | 95 | // Save map to a filepath provided in ModelOptions |
| 96 | auto filePath = modelOptions->GetCachedNetworkFilePath(); |
| 97 | if (filePath != "" && fs::exists(filePath) && fs::is_regular_file(filePath)) |
| 98 | { |
Matthew Sloyan | 80fbcd5 | 2021-01-07 13:28:47 +0000 | [diff] [blame] | 99 | // Serialize ClContext to the file specified |
Matthew Sloyan | 80fbcd5 | 2021-01-07 13:28:47 +0000 | [diff] [blame] | 100 | std::ofstream file(filePath, std::ios::out | std::ios::binary); |
| 101 | serializer.SaveSerializedToStream(file); |
Sadik Armagan | dea8fb6 | 2020-11-26 10:38:11 +0000 | [diff] [blame] | 102 | } |
| 103 | } |
| 104 | } |
| 105 | } |
| 106 | |
Aron Virginas-Tar | a8e06ed | 2018-10-19 16:46:15 +0100 | [diff] [blame] | 107 | template <typename FloatWorkload, typename Uint8Workload, typename QueueDescriptorType, typename... Args> |
| 108 | std::unique_ptr<IWorkload> ClWorkloadFactory::MakeWorkload(const QueueDescriptorType& descriptor, |
| 109 | const WorkloadInfo& info, |
| 110 | Args&&... args) |
| 111 | { |
| 112 | try |
| 113 | { |
| 114 | return MakeWorkloadHelper<FloatWorkload, Uint8Workload>(descriptor, info, std::forward<Args>(args)...); |
| 115 | } |
| 116 | catch (const cl::Error& clError) |
| 117 | { |
| 118 | throw WrapClError(clError, CHECK_LOCATION()); |
| 119 | } |
| 120 | } |
| 121 | |
| 122 | template <typename Workload, typename QueueDescriptorType, typename... Args> |
| 123 | std::unique_ptr<IWorkload> ClWorkloadFactory::MakeWorkload(const QueueDescriptorType& descriptor, |
| 124 | const WorkloadInfo& info, |
| 125 | Args&&... args) |
| 126 | { |
| 127 | try |
| 128 | { |
| 129 | return std::make_unique<Workload>(descriptor, info, std::forward<Args>(args)...); |
| 130 | } |
| 131 | catch (const cl::Error& clError) |
| 132 | { |
| 133 | throw WrapClError(clError, CHECK_LOCATION()); |
| 134 | } |
| 135 | } |
| 136 | |
Sadik Armagan | dea8fb6 | 2020-11-26 10:38:11 +0000 | [diff] [blame] | 137 | void ClWorkloadFactory::InitializeCLCompileContext() |
| 138 | { |
| 139 | // Initialize our m_CLCompileContext using default device and context |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 140 | auto context = arm_compute::CLKernelLibrary::get().context(); |
| 141 | auto device = arm_compute::CLKernelLibrary::get().get_device(); |
Sadik Armagan | dea8fb6 | 2020-11-26 10:38:11 +0000 | [diff] [blame] | 142 | m_CLCompileContext = arm_compute::CLCompileContext(context, device); |
| 143 | |
| 144 | if (m_ModelContextPtr) |
| 145 | { |
| 146 | // Load saved programs if the user has set a filepath |
| 147 | auto modelOptions = dynamic_cast<ClBackendModelContext*>(m_ModelContextPtr.get()); |
| 148 | auto filePath = modelOptions->GetCachedNetworkFilePath(); |
Sadik Armagan | b7851f9 | 2021-10-06 16:37:02 +0100 | [diff] [blame] | 149 | if (!(modelOptions->SaveCachedNetwork())) |
Sadik Armagan | dea8fb6 | 2020-11-26 10:38:11 +0000 | [diff] [blame] | 150 | { |
Matthew Sloyan | 80fbcd5 | 2021-01-07 13:28:47 +0000 | [diff] [blame] | 151 | ClContextDeserializer deserializer; |
Sadik Armagan | b7851f9 | 2021-10-06 16:37:02 +0100 | [diff] [blame] | 152 | auto cachedFd = modelOptions->GetCachedFileDescriptor(); |
| 153 | if (cachedFd != -1) |
| 154 | { |
| 155 | struct stat statBuffer; |
| 156 | if (fstat(cachedFd, &statBuffer) == 0) |
| 157 | { |
| 158 | long dataSize = static_cast<long>(statBuffer.st_size); |
| 159 | if( dataSize > 0) |
| 160 | { |
| 161 | auto offset = lseek(cachedFd, 0, SEEK_CUR); |
| 162 | if (offset == 0) |
| 163 | { |
| 164 | std::vector <uint8_t> compiledContextData(static_cast<unsigned int>(dataSize)); |
Sadik Armagan | 6b9eba2 | 2021-11-01 07:27:50 +0000 | [diff] [blame] | 165 | auto success = pread(cachedFd, compiledContextData.data(), compiledContextData.size(), 0); |
| 166 | if (success != -1) |
| 167 | { |
| 168 | deserializer.DeserializeFromBinary(m_CLCompileContext, |
| 169 | context, |
| 170 | device, |
| 171 | compiledContextData); |
| 172 | } |
Sadik Armagan | b7851f9 | 2021-10-06 16:37:02 +0100 | [diff] [blame] | 173 | } |
| 174 | } |
| 175 | |
| 176 | } |
| 177 | } |
| 178 | |
| 179 | if (filePath != "" && fs::exists(filePath) && fs::is_regular_file(filePath)) |
| 180 | { |
| 181 | // Deserialize binary file and load into m_CLCompileContext |
| 182 | deserializer.Deserialize(m_CLCompileContext, context, device, filePath); |
| 183 | } |
Sadik Armagan | dea8fb6 | 2020-11-26 10:38:11 +0000 | [diff] [blame] | 184 | } |
| 185 | } |
| 186 | } |
| 187 | |
Aron Virginas-Tar | 5605519 | 2018-11-12 18:10:43 +0000 | [diff] [blame] | 188 | ClWorkloadFactory::ClWorkloadFactory(const std::shared_ptr<ClMemoryManager>& memoryManager) |
Sadik Armagan | 04a7297 | 2020-09-14 15:44:18 +0100 | [diff] [blame] | 189 | : m_MemoryManager(memoryManager), m_ModelContextPtr(IBackendInternal::IBackendSpecificModelContextPtr{}) |
| 190 | { |
Sadik Armagan | dea8fb6 | 2020-11-26 10:38:11 +0000 | [diff] [blame] | 191 | InitializeCLCompileContext(); |
Sadik Armagan | 04a7297 | 2020-09-14 15:44:18 +0100 | [diff] [blame] | 192 | } |
| 193 | |
| 194 | ClWorkloadFactory::ClWorkloadFactory(const std::shared_ptr<ClMemoryManager>& memoryManager, |
| 195 | const IBackendInternal::IBackendSpecificModelContextPtr& modelContextPtr) |
| 196 | : m_MemoryManager(memoryManager), m_ModelContextPtr(modelContextPtr) |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 197 | { |
Sadik Armagan | dea8fb6 | 2020-11-26 10:38:11 +0000 | [diff] [blame] | 198 | InitializeCLCompileContext(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 199 | } |
| 200 | |
David Monahan | 3fb7e10 | 2019-08-20 11:25:29 +0100 | [diff] [blame] | 201 | std::unique_ptr<ITensorHandle> ClWorkloadFactory::CreateTensorHandle(const TensorInfo& tensorInfo, |
| 202 | const bool IsMemoryManaged) const |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 203 | { |
Jan Eilers | 8eb2560 | 2020-03-09 12:13:48 +0000 | [diff] [blame] | 204 | IgnoreUnused(IsMemoryManaged); |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 205 | std::unique_ptr<ClTensorHandle> tensorHandle = std::make_unique<ClTensorHandle>(tensorInfo); |
Aron Virginas-Tar | 5605519 | 2018-11-12 18:10:43 +0000 | [diff] [blame] | 206 | tensorHandle->SetMemoryGroup(m_MemoryManager->GetInterLayerMemoryGroup()); |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 207 | |
| 208 | return tensorHandle; |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 209 | } |
| 210 | |
Francis Murtagh | 351d13d | 2018-09-24 15:01:18 +0100 | [diff] [blame] | 211 | std::unique_ptr<ITensorHandle> ClWorkloadFactory::CreateTensorHandle(const TensorInfo& tensorInfo, |
David Monahan | 3fb7e10 | 2019-08-20 11:25:29 +0100 | [diff] [blame] | 212 | DataLayout dataLayout, |
| 213 | const bool IsMemoryManaged) const |
Francis Murtagh | 351d13d | 2018-09-24 15:01:18 +0100 | [diff] [blame] | 214 | { |
Jan Eilers | 8eb2560 | 2020-03-09 12:13:48 +0000 | [diff] [blame] | 215 | IgnoreUnused(IsMemoryManaged); |
Francis Murtagh | 351d13d | 2018-09-24 15:01:18 +0100 | [diff] [blame] | 216 | std::unique_ptr<ClTensorHandle> tensorHandle = std::make_unique<ClTensorHandle>(tensorInfo, dataLayout); |
Aron Virginas-Tar | 5605519 | 2018-11-12 18:10:43 +0000 | [diff] [blame] | 217 | tensorHandle->SetMemoryGroup(m_MemoryManager->GetInterLayerMemoryGroup()); |
Francis Murtagh | 351d13d | 2018-09-24 15:01:18 +0100 | [diff] [blame] | 218 | |
| 219 | return tensorHandle; |
| 220 | } |
| 221 | |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 222 | std::unique_ptr<ITensorHandle> ClWorkloadFactory::CreateSubTensorHandle(ITensorHandle& parent, |
| 223 | TensorShape const& subTensorShape, |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 224 | unsigned int const* subTensorOrigin) const |
| 225 | { |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 226 | arm_compute::Coordinates coords; |
| 227 | arm_compute::TensorShape shape = armcomputetensorutils::BuildArmComputeTensorShape(subTensorShape); |
| 228 | |
| 229 | coords.set_num_dimensions(subTensorShape.GetNumDimensions()); |
| 230 | for (unsigned int i = 0; i < subTensorShape.GetNumDimensions(); i++) |
| 231 | { |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 232 | // Arm compute indexes tensor coords in reverse order. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 233 | unsigned int revertedIndex = subTensorShape.GetNumDimensions() - i - 1; |
Matthew Sloyan | 171214c | 2020-09-09 09:07:37 +0100 | [diff] [blame] | 234 | coords.set(i, armnn::numeric_cast<int>(subTensorOrigin[revertedIndex])); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 235 | } |
| 236 | |
Derek Lamberti | 0790dce | 2019-04-15 18:37:35 +0100 | [diff] [blame] | 237 | const arm_compute::TensorShape parentShape = armcomputetensorutils::BuildArmComputeTensorShape(parent.GetShape()); |
| 238 | if (!::arm_compute::error_on_invalid_subtensor(__func__, __FILE__, __LINE__, parentShape, coords, shape)) |
| 239 | { |
| 240 | return nullptr; |
| 241 | } |
| 242 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 243 | return std::make_unique<ClSubTensorHandle>( |
Jan Eilers | bb446e5 | 2020-04-02 13:56:54 +0100 | [diff] [blame] | 244 | PolymorphicDowncast<IClTensorHandle*>(&parent), shape, coords); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 245 | } |
| 246 | |
Teresa Charlin | 611c7fb | 2022-01-07 09:47:29 +0000 | [diff] [blame] | 247 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateWorkload(LayerType type, |
| 248 | const QueueDescriptor& descriptor, |
| 249 | const WorkloadInfo& info) const |
| 250 | { |
| 251 | switch(type) |
| 252 | { |
| 253 | case LayerType::Activation : |
| 254 | { |
| 255 | auto activationQueueDescriptor = PolymorphicDowncast<const ActivationQueueDescriptor*>(&descriptor); |
| 256 | return MakeWorkload<ClActivationWorkload>(*activationQueueDescriptor, info, m_CLCompileContext); |
| 257 | } |
| 258 | case LayerType::Addition : |
| 259 | { |
| 260 | auto additionQueueDescriptor = PolymorphicDowncast<const AdditionQueueDescriptor*>(&descriptor); |
| 261 | return MakeWorkload<ClAdditionWorkload>(*additionQueueDescriptor, info, m_CLCompileContext); |
| 262 | } |
| 263 | case LayerType::ArgMinMax : |
| 264 | { |
| 265 | auto argMinMaxQueueDescriptor = PolymorphicDowncast<const ArgMinMaxQueueDescriptor*>(&descriptor); |
Cathal Corbett | ea660e1 | 2022-06-08 17:36:13 +0100 | [diff] [blame] | 266 | return MakeWorkload<ClArgMinMaxWorkload>(*argMinMaxQueueDescriptor, info, m_CLCompileContext); |
Teresa Charlin | 611c7fb | 2022-01-07 09:47:29 +0000 | [diff] [blame] | 267 | } |
| 268 | case LayerType::BatchNormalization : |
| 269 | { |
| 270 | auto batchNormalizationQueueDescriptor |
| 271 | = PolymorphicDowncast<const BatchNormalizationQueueDescriptor*>(&descriptor); |
| 272 | return MakeWorkload<ClBatchNormalizationFloatWorkload, NullWorkload> |
| 273 | (*batchNormalizationQueueDescriptor, info, m_CLCompileContext); |
| 274 | } |
| 275 | case LayerType::BatchToSpaceNd : |
| 276 | { |
| 277 | auto batchToSpaceNdQueueDescriptor |
| 278 | = PolymorphicDowncast<const BatchToSpaceNdQueueDescriptor*>(&descriptor); |
| 279 | return MakeWorkload<ClBatchToSpaceNdWorkload>(*batchToSpaceNdQueueDescriptor, info, m_CLCompileContext); |
| 280 | } |
| 281 | case LayerType::Cast : |
| 282 | { |
| 283 | auto castQueueDescriptor = PolymorphicDowncast<const CastQueueDescriptor*>(&descriptor); |
| 284 | return MakeWorkload<ClCastWorkload>(*castQueueDescriptor, info, m_CLCompileContext); |
| 285 | } |
| 286 | case LayerType::ChannelShuffle : |
| 287 | { |
| 288 | auto channelShuffleQueueDescriptor |
| 289 | = PolymorphicDowncast<const ChannelShuffleQueueDescriptor*>(&descriptor); |
| 290 | return MakeWorkload<ClChannelShuffleWorkload>(*channelShuffleQueueDescriptor, info, m_CLCompileContext); |
| 291 | } |
| 292 | case LayerType::Comparison : |
| 293 | { |
| 294 | auto comparisonQueueDescriptor = PolymorphicDowncast<const ComparisonQueueDescriptor*>(&descriptor); |
| 295 | return MakeWorkload<ClComparisonWorkload>(*comparisonQueueDescriptor, info, m_CLCompileContext); |
| 296 | } |
| 297 | case LayerType::Concat : |
| 298 | { |
| 299 | auto concatQueueDescriptor = PolymorphicDowncast<const ConcatQueueDescriptor*>(&descriptor); |
| 300 | return MakeWorkload<ClConcatWorkload>(*concatQueueDescriptor, info, m_CLCompileContext); |
| 301 | } |
| 302 | case LayerType::Constant : |
| 303 | { |
| 304 | auto constantQueueDescriptor = PolymorphicDowncast<const ConstantQueueDescriptor*>(&descriptor); |
| 305 | return MakeWorkload<ClConstantWorkload>(*constantQueueDescriptor, info, m_CLCompileContext); |
| 306 | } |
| 307 | case LayerType::ConvertFp16ToFp32 : |
| 308 | { |
| 309 | auto convertFp16ToFp32QueueDescriptor |
| 310 | = PolymorphicDowncast<const ConvertFp16ToFp32QueueDescriptor*>(&descriptor); |
| 311 | return MakeWorkload<ClConvertFp16ToFp32Workload>(*convertFp16ToFp32QueueDescriptor, |
| 312 | info, |
| 313 | m_CLCompileContext); |
| 314 | } |
| 315 | case LayerType::ConvertFp32ToFp16 : |
| 316 | { |
| 317 | auto convertFp32ToFp16QueueDescriptor |
| 318 | = PolymorphicDowncast<const ConvertFp32ToFp16QueueDescriptor*>(&descriptor); |
| 319 | return MakeWorkload<ClConvertFp32ToFp16Workload>(*convertFp32ToFp16QueueDescriptor, |
| 320 | info, |
| 321 | m_CLCompileContext); |
| 322 | } |
| 323 | case LayerType::Convolution2d : |
| 324 | { |
| 325 | auto convolution2dQueueDescriptor = PolymorphicDowncast<const Convolution2dQueueDescriptor*>(&descriptor); |
| 326 | |
| 327 | bool isFastMathEnabled = false; |
| 328 | if (m_ModelContextPtr) |
| 329 | { |
| 330 | if (m_ModelContextPtr.get() != nullptr) |
| 331 | { |
| 332 | auto modelOptions = dynamic_cast<ClBackendModelContext*>(m_ModelContextPtr.get()); |
| 333 | if (modelOptions) |
| 334 | { |
| 335 | isFastMathEnabled = modelOptions->IsFastMathEnabled(); |
| 336 | } |
| 337 | } |
| 338 | } |
| 339 | return MakeWorkload<ClConvolution2dWorkload>(*convolution2dQueueDescriptor, |
| 340 | info, |
| 341 | m_MemoryManager->GetIntraLayerManager(), |
| 342 | m_CLCompileContext, |
| 343 | isFastMathEnabled); |
| 344 | } |
| 345 | case LayerType::Convolution3d : |
| 346 | { |
| 347 | auto convolution3dQueueDescriptor = PolymorphicDowncast<const Convolution3dQueueDescriptor*>(&descriptor); |
| 348 | |
| 349 | bool isFastMathEnabled = false; |
| 350 | if (m_ModelContextPtr) |
| 351 | { |
| 352 | if (m_ModelContextPtr.get() != nullptr) |
| 353 | { |
| 354 | auto modelOptions = dynamic_cast<ClBackendModelContext*>(m_ModelContextPtr.get()); |
| 355 | if (modelOptions) |
| 356 | { |
| 357 | isFastMathEnabled = modelOptions->IsFastMathEnabled(); |
| 358 | } |
| 359 | } |
| 360 | } |
| 361 | return MakeWorkload<ClConvolution3dWorkload>(*convolution3dQueueDescriptor, |
| 362 | info, |
| 363 | m_MemoryManager->GetIntraLayerManager(), |
| 364 | m_CLCompileContext, |
| 365 | isFastMathEnabled); |
| 366 | } |
| 367 | case LayerType::Debug : |
| 368 | { |
| 369 | auto debugQueueDescriptor = PolymorphicDowncast<const DebugQueueDescriptor*>(&descriptor); |
| 370 | return MakeWorkload<NullWorkload, NullWorkload>(*debugQueueDescriptor, info, m_CLCompileContext); |
| 371 | } |
| 372 | case LayerType::DepthToSpace : |
| 373 | { |
| 374 | auto depthToSpaceQueueDescriptor = PolymorphicDowncast<const DepthToSpaceQueueDescriptor*>(&descriptor); |
| 375 | return MakeWorkload<ClDepthToSpaceWorkload>(*depthToSpaceQueueDescriptor, info, m_CLCompileContext); |
| 376 | } |
| 377 | case LayerType::DepthwiseConvolution2d : |
| 378 | { |
| 379 | auto depthwiseConvolution2dQueueDescriptor |
| 380 | = PolymorphicDowncast<const DepthwiseConvolution2dQueueDescriptor*>(&descriptor); |
| 381 | return MakeWorkload<ClDepthwiseConvolutionWorkload>(*depthwiseConvolution2dQueueDescriptor, |
| 382 | info, |
| 383 | m_CLCompileContext); |
| 384 | } |
| 385 | case LayerType::Dequantize : |
| 386 | { |
| 387 | auto dequantizeQueueDescriptor = PolymorphicDowncast<const DequantizeQueueDescriptor*>(&descriptor); |
| 388 | return MakeWorkload<ClDequantizeWorkload>(*dequantizeQueueDescriptor, info, m_CLCompileContext); |
| 389 | } |
| 390 | case LayerType::DetectionPostProcess : |
| 391 | { |
| 392 | auto detectionPostProcessQueueDescriptor |
| 393 | = PolymorphicDowncast<const DetectionPostProcessQueueDescriptor*>(&descriptor); |
| 394 | return MakeWorkload<NullWorkload, NullWorkload>(*detectionPostProcessQueueDescriptor, |
| 395 | info, |
| 396 | m_CLCompileContext); |
| 397 | } |
| 398 | case LayerType::Division : |
| 399 | { |
| 400 | auto divisionQueueDescriptor = PolymorphicDowncast<const DivisionQueueDescriptor*>(&descriptor); |
| 401 | return std::make_unique<ClDivisionWorkload>(*divisionQueueDescriptor, info, m_CLCompileContext); |
| 402 | } |
| 403 | case LayerType::ElementwiseUnary : |
| 404 | { |
| 405 | auto elementwiseUnaryQueueDescriptor |
| 406 | = PolymorphicDowncast<const ElementwiseUnaryQueueDescriptor*>(&descriptor); |
| 407 | |
| 408 | switch(elementwiseUnaryQueueDescriptor->m_Parameters.m_Operation) |
| 409 | { |
| 410 | case UnaryOperation::Abs: |
| 411 | { |
| 412 | AbsQueueDescriptor absQueueDescriptor; |
| 413 | absQueueDescriptor.m_Inputs = elementwiseUnaryQueueDescriptor->m_Inputs; |
| 414 | absQueueDescriptor.m_Outputs = elementwiseUnaryQueueDescriptor->m_Outputs; |
| 415 | |
| 416 | return std::make_unique<ClAbsWorkload>(absQueueDescriptor, info, m_CLCompileContext); |
| 417 | } |
| 418 | case UnaryOperation::Exp: |
| 419 | return std::make_unique<ClExpWorkload>(*elementwiseUnaryQueueDescriptor, info, m_CLCompileContext); |
| 420 | case UnaryOperation::Log: |
| 421 | return std::make_unique<ClLogWorkload>(*elementwiseUnaryQueueDescriptor, info, m_CLCompileContext); |
| 422 | case UnaryOperation::LogicalNot: |
| 423 | return std::make_unique<ClLogicalNotWorkload>(*elementwiseUnaryQueueDescriptor, |
| 424 | info, |
| 425 | m_CLCompileContext); |
| 426 | case UnaryOperation::Neg: |
| 427 | return std::make_unique<ClNegWorkload>(*elementwiseUnaryQueueDescriptor, info, m_CLCompileContext); |
| 428 | case UnaryOperation::Rsqrt: |
| 429 | { |
| 430 | RsqrtQueueDescriptor rsqrtQueueDescriptor; |
| 431 | rsqrtQueueDescriptor.m_Inputs = elementwiseUnaryQueueDescriptor->m_Inputs; |
| 432 | rsqrtQueueDescriptor.m_Outputs = elementwiseUnaryQueueDescriptor->m_Outputs; |
| 433 | |
| 434 | return std::make_unique<ClRsqrtWorkload>(rsqrtQueueDescriptor, info, m_CLCompileContext); |
| 435 | } |
| 436 | case UnaryOperation::Sin: |
| 437 | return std::make_unique<ClSinWorkload>(*elementwiseUnaryQueueDescriptor, info, m_CLCompileContext); |
Teresa Charlin | 0302723 | 2022-05-09 17:27:08 +0100 | [diff] [blame] | 438 | case UnaryOperation::Sqrt: |
| 439 | return std::make_unique<ClSqrtWorkload>(*elementwiseUnaryQueueDescriptor, info, m_CLCompileContext); |
Teresa Charlin | 611c7fb | 2022-01-07 09:47:29 +0000 | [diff] [blame] | 440 | default: |
| 441 | return nullptr; |
| 442 | } |
| 443 | } |
| 444 | case LayerType::Fill : |
| 445 | { |
| 446 | auto fillQueueDescriptor = PolymorphicDowncast<const FillQueueDescriptor*>(&descriptor); |
| 447 | return std::make_unique<ClFillWorkload>(*fillQueueDescriptor, info, m_CLCompileContext); |
| 448 | } |
| 449 | case LayerType::Floor : |
| 450 | { |
| 451 | auto floorQueueDescriptor = PolymorphicDowncast<const FloorQueueDescriptor*>(&descriptor); |
| 452 | return MakeWorkload<ClFloorFloatWorkload, NullWorkload>(*floorQueueDescriptor, info, m_CLCompileContext); |
| 453 | } |
| 454 | case LayerType::FullyConnected : |
| 455 | { |
| 456 | auto fullyConnectedQueueDescriptor |
| 457 | = PolymorphicDowncast<const FullyConnectedQueueDescriptor*>(&descriptor); |
| 458 | return MakeWorkload<ClFullyConnectedWorkload>(*fullyConnectedQueueDescriptor, |
| 459 | info, |
| 460 | m_MemoryManager->GetIntraLayerManager(), |
| 461 | m_CLCompileContext); |
| 462 | } |
| 463 | case LayerType::Gather : |
| 464 | { |
| 465 | auto gatherQueueDescriptor = PolymorphicDowncast<const GatherQueueDescriptor*>(&descriptor); |
| 466 | return MakeWorkload<ClGatherWorkload>(*gatherQueueDescriptor, info, m_CLCompileContext); |
| 467 | } |
Teresa Charlin | 989e2f6 | 2022-04-27 16:26:11 +0100 | [diff] [blame] | 468 | case LayerType::GatherNd : |
| 469 | { |
| 470 | auto gatherNdQueueDescriptor = PolymorphicDowncast<const GatherNdQueueDescriptor*>(&descriptor); |
| 471 | return MakeWorkload<ClGatherNdWorkload>(*gatherNdQueueDescriptor, info, m_CLCompileContext); |
| 472 | } |
Teresa Charlin | 611c7fb | 2022-01-07 09:47:29 +0000 | [diff] [blame] | 473 | case LayerType::Input : |
| 474 | { |
| 475 | auto inputQueueDescriptor = PolymorphicDowncast<const InputQueueDescriptor*>(&descriptor); |
| 476 | return std::make_unique<CopyMemGenericWorkload>(*inputQueueDescriptor, info); |
| 477 | } |
| 478 | case LayerType::InstanceNormalization : |
| 479 | { |
| 480 | auto instanceNormalizationQueueDescriptor |
| 481 | = PolymorphicDowncast<const InstanceNormalizationQueueDescriptor*>(&descriptor); |
| 482 | return MakeWorkload<ClInstanceNormalizationWorkload>(*instanceNormalizationQueueDescriptor, |
| 483 | info, |
| 484 | m_CLCompileContext); |
| 485 | } |
| 486 | case LayerType::L2Normalization : |
| 487 | { |
| 488 | auto l2NormalizationQueueDescriptor |
| 489 | = PolymorphicDowncast<const L2NormalizationQueueDescriptor*>(&descriptor); |
| 490 | return MakeWorkload<ClL2NormalizationFloatWorkload, NullWorkload>(*l2NormalizationQueueDescriptor, |
| 491 | info, |
| 492 | m_CLCompileContext); |
| 493 | } |
| 494 | case LayerType::LogicalBinary : |
| 495 | { |
| 496 | auto logicalBinaryQueueDescriptor = PolymorphicDowncast<const LogicalBinaryQueueDescriptor*>(&descriptor); |
| 497 | |
| 498 | switch(logicalBinaryQueueDescriptor->m_Parameters.m_Operation) |
| 499 | { |
| 500 | case LogicalBinaryOperation::LogicalAnd: |
| 501 | return std::make_unique<ClLogicalAndWorkload>(*logicalBinaryQueueDescriptor, |
| 502 | info, |
| 503 | m_CLCompileContext); |
| 504 | case LogicalBinaryOperation::LogicalOr: |
| 505 | return std::make_unique<ClLogicalOrWorkload>(*logicalBinaryQueueDescriptor, |
| 506 | info, |
| 507 | m_CLCompileContext); |
| 508 | default: |
| 509 | return nullptr; |
| 510 | } |
| 511 | } |
| 512 | case LayerType::LogSoftmax : |
| 513 | { |
| 514 | auto logSoftmaxQueueDescriptor = PolymorphicDowncast<const LogSoftmaxQueueDescriptor*>(&descriptor); |
| 515 | |
| 516 | return MakeWorkload<ClLogSoftmaxWorkload>(*logSoftmaxQueueDescriptor, |
| 517 | info, |
| 518 | m_MemoryManager->GetIntraLayerManager(), |
| 519 | m_CLCompileContext); |
| 520 | } |
| 521 | case LayerType::Lstm : |
| 522 | { |
| 523 | auto lstmQueueDescriptor = PolymorphicDowncast<const LstmQueueDescriptor*>(&descriptor); |
| 524 | return MakeWorkload<ClLstmFloatWorkload, NullWorkload>(*lstmQueueDescriptor, info, m_CLCompileContext); |
| 525 | } |
| 526 | case LayerType::Maximum : |
| 527 | { |
| 528 | auto maximumQueueDescriptor = PolymorphicDowncast<const MaximumQueueDescriptor*>(&descriptor); |
| 529 | return MakeWorkload<ClMaximumWorkload>(*maximumQueueDescriptor, info, m_CLCompileContext); |
| 530 | } |
| 531 | case LayerType::Mean : |
| 532 | { |
| 533 | auto meanQueueDescriptor = PolymorphicDowncast<const MeanQueueDescriptor*>(&descriptor); |
| 534 | return MakeWorkload<ClMeanWorkload>(*meanQueueDescriptor, info, m_CLCompileContext); |
| 535 | } |
| 536 | case LayerType::MemCopy : |
| 537 | { |
| 538 | auto memCopyQueueDescriptor = PolymorphicDowncast<const MemCopyQueueDescriptor*>(&descriptor); |
| 539 | if (memCopyQueueDescriptor->m_Inputs.empty() || !memCopyQueueDescriptor->m_Inputs[0]) |
| 540 | { |
| 541 | throw InvalidArgumentException("ClWorkloadFactory: Invalid null input for MemCopy workload"); |
| 542 | } |
| 543 | return MakeWorkload<CopyMemGenericWorkload>(*memCopyQueueDescriptor, info); |
| 544 | } |
| 545 | case LayerType::MemImport : |
| 546 | { |
| 547 | auto memImportQueueDescriptor = PolymorphicDowncast<const MemImportQueueDescriptor*>(&descriptor); |
| 548 | if (memImportQueueDescriptor->m_Inputs.empty() || !memImportQueueDescriptor->m_Inputs[0]) |
| 549 | { |
| 550 | throw InvalidArgumentException("ClWorkloadFactory: Invalid null input for MemImport workload"); |
| 551 | } |
| 552 | return std::make_unique<ImportMemGenericWorkload>(*memImportQueueDescriptor, info); |
| 553 | } |
| 554 | case LayerType::Minimum : |
| 555 | { |
| 556 | auto minimumQueueDescriptor = PolymorphicDowncast<const MinimumQueueDescriptor*>(&descriptor); |
| 557 | return MakeWorkload<ClMinimumWorkload>(*minimumQueueDescriptor, info, m_CLCompileContext); |
| 558 | } |
| 559 | case LayerType::Multiplication : |
| 560 | { |
| 561 | auto multiplicationQueueDescriptor = PolymorphicDowncast<const MultiplicationQueueDescriptor*>(&descriptor); |
| 562 | return MakeWorkload<ClMultiplicationWorkload>(*multiplicationQueueDescriptor, info, m_CLCompileContext); |
| 563 | } |
| 564 | case LayerType::Normalization : |
| 565 | { |
| 566 | auto normalizationQueueDescriptor = PolymorphicDowncast<const NormalizationQueueDescriptor*>(&descriptor); |
| 567 | return MakeWorkload<ClNormalizationFloatWorkload, NullWorkload>(*normalizationQueueDescriptor, |
| 568 | info, |
| 569 | m_CLCompileContext); |
| 570 | } |
| 571 | case LayerType::Output : |
| 572 | { |
| 573 | auto outputQueueDescriptor = PolymorphicDowncast<const OutputQueueDescriptor*>(&descriptor); |
| 574 | return std::make_unique<CopyMemGenericWorkload>(*outputQueueDescriptor, info); |
| 575 | } |
| 576 | case LayerType::Pad : |
| 577 | { |
| 578 | auto padQueueDescriptor = PolymorphicDowncast<const PadQueueDescriptor*>(&descriptor); |
| 579 | return MakeWorkload<ClPadWorkload>(*padQueueDescriptor, info, m_CLCompileContext); |
| 580 | } |
| 581 | case LayerType::Permute : |
| 582 | { |
| 583 | auto permuteQueueDescriptor = PolymorphicDowncast<const PermuteQueueDescriptor*>(&descriptor); |
| 584 | return MakeWorkload<ClPermuteWorkload>(*permuteQueueDescriptor, info, m_CLCompileContext); |
| 585 | } |
| 586 | case LayerType::Pooling2d : |
| 587 | { |
| 588 | auto pooling2dQueueDescriptor = PolymorphicDowncast<const Pooling2dQueueDescriptor*>(&descriptor); |
| 589 | return MakeWorkload<ClPooling2dWorkload>(*pooling2dQueueDescriptor, info, m_CLCompileContext); |
| 590 | } |
Ryan OShea | bab8fa9 | 2022-03-09 10:29:02 +0000 | [diff] [blame] | 591 | case LayerType::Pooling3d : |
| 592 | { |
| 593 | auto pooling3dQueueDescriptor = PolymorphicDowncast<const Pooling3dQueueDescriptor*>(&descriptor); |
| 594 | return MakeWorkload<ClPooling3dWorkload>(*pooling3dQueueDescriptor, info, m_CLCompileContext); |
| 595 | } |
Teresa Charlin | 611c7fb | 2022-01-07 09:47:29 +0000 | [diff] [blame] | 596 | case LayerType::PreCompiled : |
| 597 | { |
| 598 | auto preCompiledQueueDescriptor = PolymorphicDowncast<const PreCompiledQueueDescriptor*>(&descriptor); |
| 599 | return MakeWorkload<NullWorkload, NullWorkload>(*preCompiledQueueDescriptor, info, m_CLCompileContext); |
| 600 | } |
| 601 | case LayerType::Prelu : |
| 602 | { |
| 603 | auto preluQueueDescriptor = PolymorphicDowncast<const PreluQueueDescriptor*>(&descriptor); |
| 604 | return MakeWorkload<ClPreluWorkload>(*preluQueueDescriptor, info, m_CLCompileContext); |
| 605 | } |
| 606 | case LayerType::QLstm : |
| 607 | { |
| 608 | auto qLstmQueueDescriptor = PolymorphicDowncast<const QLstmQueueDescriptor*>(&descriptor); |
| 609 | return std::make_unique<ClQLstmWorkload>(*qLstmQueueDescriptor, info, m_CLCompileContext); |
| 610 | } |
| 611 | case LayerType::Quantize : |
| 612 | { |
| 613 | auto quantizeQueueDescriptor = PolymorphicDowncast<const QuantizeQueueDescriptor*>(&descriptor); |
| 614 | return MakeWorkload<ClQuantizeWorkload>(*quantizeQueueDescriptor, info, m_CLCompileContext); |
| 615 | } |
| 616 | case LayerType::QuantizedLstm : |
| 617 | { |
| 618 | auto quantizedLstmQueueDescriptor = PolymorphicDowncast<const QuantizedLstmQueueDescriptor*>(&descriptor); |
| 619 | return MakeWorkload<ClQuantizedLstmWorkload>(*quantizedLstmQueueDescriptor, info, m_CLCompileContext); |
| 620 | } |
| 621 | case LayerType::Rank : |
| 622 | { |
| 623 | auto rankQueueDescriptor = PolymorphicDowncast<const RankQueueDescriptor*>(&descriptor); |
| 624 | return std::make_unique<ClRankWorkload>(*rankQueueDescriptor, info); |
| 625 | } |
| 626 | case LayerType::Reduce : |
| 627 | { |
| 628 | auto reduceQueueDescriptor = PolymorphicDowncast<const ReduceQueueDescriptor*>(&descriptor); |
| 629 | return std::make_unique<ClReduceWorkload>(*reduceQueueDescriptor, info); |
| 630 | } |
| 631 | case LayerType::Reshape : |
| 632 | { |
| 633 | auto reshapeQueueDescriptor = PolymorphicDowncast<const ReshapeQueueDescriptor*>(&descriptor); |
| 634 | return MakeWorkload<ClReshapeWorkload>(*reshapeQueueDescriptor, info, m_CLCompileContext); |
| 635 | } |
| 636 | case LayerType::Resize : |
| 637 | { |
| 638 | auto resizeQueueDescriptor = PolymorphicDowncast<const ResizeQueueDescriptor*>(&descriptor); |
| 639 | return MakeWorkload<ClResizeWorkload>(*resizeQueueDescriptor, info, m_CLCompileContext); |
| 640 | } |
| 641 | case LayerType::Slice : |
| 642 | { |
| 643 | auto sliceQueueDescriptor = PolymorphicDowncast<const SliceQueueDescriptor*>(&descriptor); |
| 644 | return MakeWorkload<ClSliceWorkload>(*sliceQueueDescriptor, info, m_CLCompileContext); |
| 645 | } |
| 646 | case LayerType::Softmax : |
| 647 | { |
| 648 | auto softmaxQueueDescriptor = PolymorphicDowncast<const SoftmaxQueueDescriptor*>(&descriptor); |
| 649 | return std::make_unique<ClSoftmaxWorkload>(*softmaxQueueDescriptor, |
| 650 | info, |
| 651 | m_MemoryManager->GetIntraLayerManager(), |
| 652 | m_CLCompileContext); |
| 653 | } |
| 654 | case LayerType::SpaceToBatchNd : |
| 655 | { |
| 656 | auto spaceToBatchNdQueueDescriptor |
| 657 | = PolymorphicDowncast<const SpaceToBatchNdQueueDescriptor*>(&descriptor); |
| 658 | return MakeWorkload<ClSpaceToBatchNdWorkload>(*spaceToBatchNdQueueDescriptor, info, m_CLCompileContext); |
| 659 | } |
| 660 | case LayerType::SpaceToDepth : |
| 661 | { |
| 662 | auto spaceToDepthQueueDescriptor = PolymorphicDowncast<const SpaceToDepthQueueDescriptor*>(&descriptor); |
| 663 | return MakeWorkload<ClSpaceToDepthWorkload>(*spaceToDepthQueueDescriptor, info, m_CLCompileContext); |
| 664 | } |
| 665 | case LayerType::Splitter : |
| 666 | { |
| 667 | auto splitterQueueDescriptor = PolymorphicDowncast<const SplitterQueueDescriptor*>(&descriptor); |
| 668 | return MakeWorkload<ClSplitterWorkload>(*splitterQueueDescriptor, info, m_CLCompileContext); |
| 669 | } |
| 670 | case LayerType::Stack : |
| 671 | { |
| 672 | auto stackQueueDescriptor = PolymorphicDowncast<const StackQueueDescriptor*>(&descriptor); |
| 673 | return MakeWorkload<ClStackWorkload>(*stackQueueDescriptor, info, m_CLCompileContext); |
| 674 | } |
| 675 | case LayerType::StridedSlice : |
| 676 | { |
| 677 | auto stridedSliceQueueDescriptor = PolymorphicDowncast<const StridedSliceQueueDescriptor*>(&descriptor); |
| 678 | return MakeWorkload<ClStridedSliceWorkload>(*stridedSliceQueueDescriptor, info, m_CLCompileContext); |
| 679 | } |
| 680 | case LayerType::Subtraction : |
| 681 | { |
| 682 | auto subtractionQueueDescriptor = PolymorphicDowncast<const SubtractionQueueDescriptor*>(&descriptor); |
| 683 | return MakeWorkload<ClSubtractionWorkload>(*subtractionQueueDescriptor, info, m_CLCompileContext); |
| 684 | } |
| 685 | case LayerType::Transpose : |
| 686 | { |
| 687 | auto transposeQueueDescriptor = PolymorphicDowncast<const TransposeQueueDescriptor*>(&descriptor); |
| 688 | return MakeWorkload<ClTransposeWorkload>(*transposeQueueDescriptor, info, m_CLCompileContext); |
| 689 | } |
| 690 | case LayerType::TransposeConvolution2d : |
| 691 | { |
| 692 | auto transposeConvolution2dQueueDescriptor |
| 693 | = PolymorphicDowncast<const TransposeConvolution2dQueueDescriptor*>(&descriptor); |
| 694 | return MakeWorkload<ClTransposeConvolution2dWorkload>(*transposeConvolution2dQueueDescriptor, |
| 695 | info, |
| 696 | m_MemoryManager->GetIntraLayerManager(), |
| 697 | m_CLCompileContext); |
| 698 | } |
Cathal Corbett | 4952a3e | 2022-03-03 15:14:18 +0000 | [diff] [blame] | 699 | case LayerType::UnidirectionalSequenceLstm : |
| 700 | { |
| 701 | auto desc = PolymorphicDowncast<const UnidirectionalSequenceLstmQueueDescriptor*>(&descriptor); |
| 702 | return MakeWorkloadHelper<ClUnidirectionalSequenceLstmFloatWorkload, NullWorkload>(*desc, |
| 703 | info, |
| 704 | m_CLCompileContext); |
| 705 | } |
Teresa Charlin | 611c7fb | 2022-01-07 09:47:29 +0000 | [diff] [blame] | 706 | default: |
| 707 | return nullptr; |
| 708 | } |
| 709 | } |
| 710 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 711 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateActivation(const ActivationQueueDescriptor& descriptor, |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 712 | const WorkloadInfo& info) const |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 713 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 714 | return MakeWorkload<ClActivationWorkload>(descriptor, info, m_CLCompileContext); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 715 | } |
| 716 | |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 717 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateAddition(const AdditionQueueDescriptor& descriptor, |
| 718 | const WorkloadInfo& info) const |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 719 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 720 | return MakeWorkload<ClAdditionWorkload>(descriptor, info, m_CLCompileContext); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 721 | } |
| 722 | |
James Conroy | 2dc0572 | 2019-09-19 17:00:31 +0100 | [diff] [blame] | 723 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateArgMinMax(const ArgMinMaxQueueDescriptor& descriptor, |
| 724 | const WorkloadInfo& info) const |
| 725 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 726 | return std::make_unique<ClArgMinMaxWorkload>(descriptor, info, m_CLCompileContext); |
James Conroy | 2dc0572 | 2019-09-19 17:00:31 +0100 | [diff] [blame] | 727 | } |
| 728 | |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 729 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateBatchNormalization( |
| 730 | const BatchNormalizationQueueDescriptor& descriptor, |
| 731 | const WorkloadInfo& info) const |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 732 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 733 | return MakeWorkload<ClBatchNormalizationFloatWorkload, NullWorkload>(descriptor, info, m_CLCompileContext); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 734 | } |
| 735 | |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 736 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateBatchToSpaceNd(const BatchToSpaceNdQueueDescriptor& descriptor, |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 737 | const WorkloadInfo& info) const |
| 738 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 739 | return MakeWorkload<ClBatchToSpaceNdWorkload>(descriptor, info, m_CLCompileContext); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 740 | } |
| 741 | |
Sadik Armagan | f40d6d4 | 2021-04-22 09:12:11 +0100 | [diff] [blame] | 742 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateCast(const CastQueueDescriptor& descriptor, |
| 743 | const WorkloadInfo& info) const |
| 744 | { |
| 745 | return MakeWorkload<ClCastWorkload>(descriptor, info, m_CLCompileContext); |
| 746 | } |
| 747 | |
Teresa Charlin | 1222dbd | 2021-09-02 13:58:52 +0100 | [diff] [blame] | 748 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateChannelShuffle(const ChannelShuffleQueueDescriptor& descriptor, |
| 749 | const WorkloadInfo& info) const |
| 750 | { |
| 751 | return MakeWorkload<ClChannelShuffleWorkload>(descriptor, info, m_CLCompileContext); |
| 752 | } |
| 753 | |
Aron Virginas-Tar | 77bfb5e | 2019-10-16 17:45:38 +0100 | [diff] [blame] | 754 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateComparison(const ComparisonQueueDescriptor& descriptor, |
| 755 | const WorkloadInfo& info) const |
| 756 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 757 | return MakeWorkload<ClComparisonWorkload>(descriptor, info, m_CLCompileContext); |
Aron Virginas-Tar | 77bfb5e | 2019-10-16 17:45:38 +0100 | [diff] [blame] | 758 | } |
| 759 | |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 760 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateConcat(const ConcatQueueDescriptor& descriptor, |
| 761 | const WorkloadInfo& info) const |
Jim Flynn | 4ed6c83 | 2019-05-20 11:02:46 +0100 | [diff] [blame] | 762 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 763 | return MakeWorkload<ClConcatWorkload>(descriptor, info, m_CLCompileContext); |
Jim Flynn | 4ed6c83 | 2019-05-20 11:02:46 +0100 | [diff] [blame] | 764 | } |
| 765 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 766 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateConstant(const ConstantQueueDescriptor& descriptor, |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 767 | const WorkloadInfo& info) const |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 768 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 769 | return MakeWorkload<ClConstantWorkload>(descriptor, info, m_CLCompileContext); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 770 | } |
| 771 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 772 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateConvertFp16ToFp32( |
| 773 | const ConvertFp16ToFp32QueueDescriptor& descriptor, |
| 774 | const WorkloadInfo& info) const |
| 775 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 776 | return MakeWorkload<ClConvertFp16ToFp32Workload>(descriptor, info, m_CLCompileContext); |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 777 | } |
| 778 | |
| 779 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateConvertFp32ToFp16( |
| 780 | const ConvertFp32ToFp16QueueDescriptor& descriptor, |
| 781 | const WorkloadInfo& info) const |
| 782 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 783 | return MakeWorkload<ClConvertFp32ToFp16Workload>(descriptor, info, m_CLCompileContext); |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 784 | } |
| 785 | |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 786 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateConvolution2d(const Convolution2dQueueDescriptor& descriptor, |
| 787 | const WorkloadInfo& info) const |
| 788 | { |
Sadik Armagan | 04a7297 | 2020-09-14 15:44:18 +0100 | [diff] [blame] | 789 | bool isFastMathEnabled = false; |
| 790 | if (m_ModelContextPtr) |
| 791 | { |
| 792 | if (m_ModelContextPtr.get() != nullptr) |
| 793 | { |
| 794 | auto modelOptions = dynamic_cast<ClBackendModelContext*>(m_ModelContextPtr.get()); |
| 795 | if (modelOptions) |
| 796 | { |
| 797 | isFastMathEnabled = modelOptions->IsFastMathEnabled(); |
| 798 | } |
| 799 | } |
| 800 | } |
| 801 | return MakeWorkload<ClConvolution2dWorkload>(descriptor, |
| 802 | info, |
| 803 | m_MemoryManager->GetIntraLayerManager(), |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 804 | m_CLCompileContext, |
Sadik Armagan | 04a7297 | 2020-09-14 15:44:18 +0100 | [diff] [blame] | 805 | isFastMathEnabled); |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 806 | } |
| 807 | |
Teresa Charlin | 615ad6c | 2021-10-26 12:22:20 +0100 | [diff] [blame] | 808 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateConvolution3d(const Convolution3dQueueDescriptor& descriptor, |
| 809 | const WorkloadInfo& info) const |
| 810 | { |
| 811 | bool isFastMathEnabled = false; |
| 812 | if (m_ModelContextPtr) |
| 813 | { |
| 814 | if (m_ModelContextPtr.get() != nullptr) |
| 815 | { |
| 816 | auto modelOptions = dynamic_cast<ClBackendModelContext*>(m_ModelContextPtr.get()); |
| 817 | if (modelOptions) |
| 818 | { |
| 819 | isFastMathEnabled = modelOptions->IsFastMathEnabled(); |
| 820 | } |
| 821 | } |
| 822 | } |
| 823 | return MakeWorkload<ClConvolution3dWorkload>(descriptor, |
| 824 | info, |
| 825 | m_MemoryManager->GetIntraLayerManager(), |
| 826 | m_CLCompileContext, |
| 827 | isFastMathEnabled); |
| 828 | } |
| 829 | |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 830 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateDebug(const DebugQueueDescriptor& descriptor, |
| 831 | const WorkloadInfo& info) const |
| 832 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 833 | return MakeWorkload<NullWorkload, NullWorkload>(descriptor, info, m_CLCompileContext); |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 834 | } |
| 835 | |
| 836 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateDepthToSpace(const DepthToSpaceQueueDescriptor& descriptor, |
| 837 | const WorkloadInfo& info) const |
| 838 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 839 | return MakeWorkload<ClDepthToSpaceWorkload>(descriptor, info, m_CLCompileContext); |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 840 | } |
| 841 | |
| 842 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateDepthwiseConvolution2d( |
| 843 | const DepthwiseConvolution2dQueueDescriptor& descriptor, |
| 844 | const WorkloadInfo& info) const |
| 845 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 846 | return MakeWorkload<ClDepthwiseConvolutionWorkload>(descriptor, info, m_CLCompileContext); |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 847 | } |
| 848 | |
| 849 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateDequantize(const DequantizeQueueDescriptor& descriptor, |
| 850 | const WorkloadInfo& info) const |
| 851 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 852 | return MakeWorkload<ClDequantizeWorkload>(descriptor, info, m_CLCompileContext); |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 853 | } |
| 854 | |
| 855 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateDetectionPostProcess( |
| 856 | const DetectionPostProcessQueueDescriptor& descriptor, |
| 857 | const WorkloadInfo& info) const |
| 858 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 859 | return MakeWorkload<NullWorkload, NullWorkload>(descriptor, info, m_CLCompileContext); |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 860 | } |
| 861 | |
| 862 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateDivision(const DivisionQueueDescriptor& descriptor, |
| 863 | const WorkloadInfo& info) const |
| 864 | { |
Teresa Charlin | e11e63d | 2021-04-21 12:56:45 +0100 | [diff] [blame] | 865 | return std::make_unique<ClDivisionWorkload>(descriptor, info, m_CLCompileContext); |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 866 | } |
| 867 | |
josh minor | 4a3c610 | 2020-01-06 16:40:46 -0600 | [diff] [blame] | 868 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateElementwiseUnary(const ElementwiseUnaryQueueDescriptor& descriptor, |
| 869 | const WorkloadInfo& info) const |
| 870 | { |
Sadik Armagan | 9fabf43 | 2020-05-27 13:40:58 +0100 | [diff] [blame] | 871 | switch(descriptor.m_Parameters.m_Operation) |
josh minor | 4a3c610 | 2020-01-06 16:40:46 -0600 | [diff] [blame] | 872 | { |
Sadik Armagan | 9fabf43 | 2020-05-27 13:40:58 +0100 | [diff] [blame] | 873 | case UnaryOperation::Abs: |
James Conroy | fe3ec94 | 2020-11-18 14:20:53 +0000 | [diff] [blame] | 874 | { |
| 875 | AbsQueueDescriptor absQueueDescriptor; |
| 876 | absQueueDescriptor.m_Inputs = descriptor.m_Inputs; |
| 877 | absQueueDescriptor.m_Outputs = descriptor.m_Outputs; |
josh minor | 4a3c610 | 2020-01-06 16:40:46 -0600 | [diff] [blame] | 878 | |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 879 | return std::make_unique<ClAbsWorkload>(absQueueDescriptor, info, m_CLCompileContext); |
James Conroy | fe3ec94 | 2020-11-18 14:20:53 +0000 | [diff] [blame] | 880 | } |
Sadik Armagan | 9fabf43 | 2020-05-27 13:40:58 +0100 | [diff] [blame] | 881 | case UnaryOperation::Exp: |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 882 | return std::make_unique<ClExpWorkload>(descriptor, info, m_CLCompileContext); |
Teresa Charlin | 50de4fa | 2021-05-31 18:47:33 +0100 | [diff] [blame] | 883 | case UnaryOperation::Log: |
| 884 | return std::make_unique<ClLogWorkload>(descriptor, info, m_CLCompileContext); |
| 885 | case UnaryOperation::LogicalNot: |
| 886 | return std::make_unique<ClLogicalNotWorkload>(descriptor, info, m_CLCompileContext); |
Sadik Armagan | 9fabf43 | 2020-05-27 13:40:58 +0100 | [diff] [blame] | 887 | case UnaryOperation::Neg: |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 888 | return std::make_unique<ClNegWorkload>(descriptor, info, m_CLCompileContext); |
Sadik Armagan | 9fabf43 | 2020-05-27 13:40:58 +0100 | [diff] [blame] | 889 | case UnaryOperation::Rsqrt: |
James Conroy | fe3ec94 | 2020-11-18 14:20:53 +0000 | [diff] [blame] | 890 | { |
| 891 | RsqrtQueueDescriptor rsqrtQueueDescriptor; |
| 892 | rsqrtQueueDescriptor.m_Inputs = descriptor.m_Inputs; |
| 893 | rsqrtQueueDescriptor.m_Outputs = descriptor.m_Outputs; |
josh minor | 4a3c610 | 2020-01-06 16:40:46 -0600 | [diff] [blame] | 894 | |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 895 | return std::make_unique<ClRsqrtWorkload>(rsqrtQueueDescriptor, info, m_CLCompileContext); |
James Conroy | fe3ec94 | 2020-11-18 14:20:53 +0000 | [diff] [blame] | 896 | } |
Teresa Charlin | 50de4fa | 2021-05-31 18:47:33 +0100 | [diff] [blame] | 897 | case UnaryOperation::Sin: |
| 898 | return std::make_unique<ClSinWorkload>(descriptor, info, m_CLCompileContext); |
Sadik Armagan | 9fabf43 | 2020-05-27 13:40:58 +0100 | [diff] [blame] | 899 | default: |
| 900 | return nullptr; |
josh minor | 4a3c610 | 2020-01-06 16:40:46 -0600 | [diff] [blame] | 901 | } |
josh minor | 4a3c610 | 2020-01-06 16:40:46 -0600 | [diff] [blame] | 902 | } |
| 903 | |
Sadik Armagan | 66aecb0 | 2020-06-24 11:42:20 +0100 | [diff] [blame] | 904 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateFill(const FillQueueDescriptor& descriptor, |
| 905 | const WorkloadInfo& info) const |
| 906 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 907 | return std::make_unique<ClFillWorkload>(descriptor, info, m_CLCompileContext); |
Sadik Armagan | 66aecb0 | 2020-06-24 11:42:20 +0100 | [diff] [blame] | 908 | } |
| 909 | |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 910 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateFloor(const FloorQueueDescriptor& descriptor, |
| 911 | const WorkloadInfo& info) const |
| 912 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 913 | return MakeWorkload<ClFloorFloatWorkload, NullWorkload>(descriptor, info, m_CLCompileContext); |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 914 | } |
| 915 | |
| 916 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateFullyConnected(const FullyConnectedQueueDescriptor& descriptor, |
| 917 | const WorkloadInfo& info) const |
| 918 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 919 | return MakeWorkload<ClFullyConnectedWorkload>(descriptor, |
| 920 | info, |
| 921 | m_MemoryManager->GetIntraLayerManager(), |
| 922 | m_CLCompileContext); |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 923 | } |
| 924 | |
| 925 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateGather(const GatherQueueDescriptor& descriptor, |
| 926 | const WorkloadInfo& info) const |
| 927 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 928 | return MakeWorkload<ClGatherWorkload>(descriptor, info, m_CLCompileContext); |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 929 | } |
| 930 | |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 931 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateInput(const InputQueueDescriptor& descriptor, |
| 932 | const WorkloadInfo& info) const |
| 933 | { |
| 934 | return std::make_unique<CopyMemGenericWorkload>(descriptor, info); |
| 935 | } |
| 936 | |
| 937 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateInstanceNormalization( |
| 938 | const InstanceNormalizationQueueDescriptor& descriptor, |
| 939 | const WorkloadInfo& info) const |
| 940 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 941 | return MakeWorkload<ClInstanceNormalizationWorkload>(descriptor, info, m_CLCompileContext); |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 942 | } |
| 943 | |
| 944 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateL2Normalization(const L2NormalizationQueueDescriptor& descriptor, |
| 945 | const WorkloadInfo& info) const |
| 946 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 947 | return MakeWorkload<ClL2NormalizationFloatWorkload, NullWorkload>(descriptor, info, m_CLCompileContext); |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 948 | } |
| 949 | |
James Conroy | fe3ec94 | 2020-11-18 14:20:53 +0000 | [diff] [blame] | 950 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateLogicalBinary(const LogicalBinaryQueueDescriptor& descriptor, |
| 951 | const WorkloadInfo& info) const |
| 952 | { |
| 953 | switch(descriptor.m_Parameters.m_Operation) |
| 954 | { |
| 955 | case LogicalBinaryOperation::LogicalAnd: |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 956 | return std::make_unique<ClLogicalAndWorkload>(descriptor, info, m_CLCompileContext); |
James Conroy | fe3ec94 | 2020-11-18 14:20:53 +0000 | [diff] [blame] | 957 | case LogicalBinaryOperation::LogicalOr: |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 958 | return std::make_unique<ClLogicalOrWorkload>(descriptor, info, m_CLCompileContext); |
James Conroy | fe3ec94 | 2020-11-18 14:20:53 +0000 | [diff] [blame] | 959 | default: |
| 960 | return nullptr; |
| 961 | } |
| 962 | } |
| 963 | |
Teresa Charlin | 8398edc | 2020-07-20 14:23:02 +0100 | [diff] [blame] | 964 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateLogSoftmax(const LogSoftmaxQueueDescriptor& descriptor, |
| 965 | const WorkloadInfo& info) const |
| 966 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 967 | return MakeWorkload<ClLogSoftmaxWorkload>(descriptor, |
| 968 | info, |
| 969 | m_MemoryManager->GetIntraLayerManager(), |
| 970 | m_CLCompileContext); |
Teresa Charlin | 8398edc | 2020-07-20 14:23:02 +0100 | [diff] [blame] | 971 | } |
| 972 | |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 973 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateLstm(const LstmQueueDescriptor& descriptor, |
| 974 | const WorkloadInfo& info) const |
| 975 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 976 | return MakeWorkload<ClLstmFloatWorkload, NullWorkload>(descriptor, info, m_CLCompileContext); |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 977 | } |
| 978 | |
Nattapat Chaimanowong | 5a4304a | 2018-11-28 10:44:37 +0000 | [diff] [blame] | 979 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateMaximum(const MaximumQueueDescriptor& descriptor, |
| 980 | const WorkloadInfo& info) const |
| 981 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 982 | return MakeWorkload<ClMaximumWorkload>(descriptor, info, m_CLCompileContext); |
Nattapat Chaimanowong | 5a4304a | 2018-11-28 10:44:37 +0000 | [diff] [blame] | 983 | } |
| 984 | |
narpra01 | a6bf912 | 2018-09-10 09:50:09 +0100 | [diff] [blame] | 985 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateMean(const MeanQueueDescriptor& descriptor, |
| 986 | const WorkloadInfo& info) const |
| 987 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 988 | return MakeWorkload<ClMeanWorkload>(descriptor, info, m_CLCompileContext); |
narpra01 | a6bf912 | 2018-09-10 09:50:09 +0100 | [diff] [blame] | 989 | } |
| 990 | |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 991 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateMemCopy(const MemCopyQueueDescriptor& descriptor, |
| 992 | const WorkloadInfo& info) const |
jimfly01 | 2c9322a | 2018-09-19 10:59:49 +0100 | [diff] [blame] | 993 | { |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 994 | if (descriptor.m_Inputs.empty() || !descriptor.m_Inputs[0]) |
| 995 | { |
| 996 | throw InvalidArgumentException("ClWorkloadFactory: Invalid null input for MemCopy workload"); |
| 997 | } |
| 998 | |
| 999 | return MakeWorkload<CopyMemGenericWorkload>(descriptor, info); |
jimfly01 | 2c9322a | 2018-09-19 10:59:49 +0100 | [diff] [blame] | 1000 | } |
| 1001 | |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 1002 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateMemImport(const MemImportQueueDescriptor& descriptor, |
| 1003 | const WorkloadInfo& info) const |
FrancisMurtagh | 2099595 | 2018-12-17 12:11:36 +0000 | [diff] [blame] | 1004 | { |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 1005 | if (descriptor.m_Inputs.empty() || !descriptor.m_Inputs[0]) |
| 1006 | { |
| 1007 | throw InvalidArgumentException("ClWorkloadFactory: Invalid null input for MemImport workload"); |
| 1008 | } |
| 1009 | |
| 1010 | return std::make_unique<ImportMemGenericWorkload>(descriptor, info); |
FrancisMurtagh | 2099595 | 2018-12-17 12:11:36 +0000 | [diff] [blame] | 1011 | } |
| 1012 | |
kevmay01 | 9053969 | 2018-11-29 08:40:19 +0000 | [diff] [blame] | 1013 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateMinimum(const MinimumQueueDescriptor& descriptor, |
| 1014 | const WorkloadInfo& info) const |
| 1015 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 1016 | return MakeWorkload<ClMinimumWorkload>(descriptor, info, m_CLCompileContext); |
kevmay01 | 9053969 | 2018-11-29 08:40:19 +0000 | [diff] [blame] | 1017 | } |
| 1018 | |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 1019 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateMultiplication(const MultiplicationQueueDescriptor& descriptor, |
| 1020 | const WorkloadInfo& info) const |
| 1021 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 1022 | return MakeWorkload<ClMultiplicationWorkload>(descriptor, info, m_CLCompileContext); |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 1023 | } |
| 1024 | |
| 1025 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateNormalization(const NormalizationQueueDescriptor& descriptor, |
| 1026 | const WorkloadInfo& info) const |
| 1027 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 1028 | return MakeWorkload<ClNormalizationFloatWorkload, NullWorkload>(descriptor, info, m_CLCompileContext); |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 1029 | } |
| 1030 | |
| 1031 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateOutput(const OutputQueueDescriptor& descriptor, |
| 1032 | const WorkloadInfo& info) const |
| 1033 | { |
| 1034 | return std::make_unique<CopyMemGenericWorkload>(descriptor, info); |
| 1035 | } |
| 1036 | |
| 1037 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreatePad(const PadQueueDescriptor& descriptor, |
| 1038 | const WorkloadInfo& info) const |
| 1039 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 1040 | return MakeWorkload<ClPadWorkload>(descriptor, info, m_CLCompileContext); |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 1041 | } |
| 1042 | |
| 1043 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreatePermute(const PermuteQueueDescriptor& descriptor, |
Matteo Martincigh | 59a950c | 2018-12-13 12:48:25 +0000 | [diff] [blame] | 1044 | const WorkloadInfo& info) const |
| 1045 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 1046 | return MakeWorkload<ClPermuteWorkload>(descriptor, info, m_CLCompileContext); |
Matteo Martincigh | 59a950c | 2018-12-13 12:48:25 +0000 | [diff] [blame] | 1047 | } |
| 1048 | |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 1049 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreatePooling2d(const Pooling2dQueueDescriptor& descriptor, |
| 1050 | const WorkloadInfo& info) const |
Nattapat Chaimanowong | a9a1cf1 | 2018-12-03 16:06:49 +0000 | [diff] [blame] | 1051 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 1052 | return MakeWorkload<ClPooling2dWorkload>(descriptor, info, m_CLCompileContext); |
Mohamed Nour Abouelseoud | a1d3c6a | 2018-12-27 12:39:16 +0000 | [diff] [blame] | 1053 | } |
| 1054 | |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 1055 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreatePreCompiled(const PreCompiledQueueDescriptor& descriptor, |
| 1056 | const WorkloadInfo& info) const |
| 1057 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 1058 | return MakeWorkload<NullWorkload, NullWorkload>(descriptor, info, m_CLCompileContext); |
Matteo Martincigh | 4912402 | 2019-01-11 13:25:59 +0000 | [diff] [blame] | 1059 | } |
| 1060 | |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 1061 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreatePrelu(const PreluQueueDescriptor &descriptor, |
| 1062 | const WorkloadInfo &info) const |
narpra01 | b89b05f | 2019-01-16 09:53:09 +0000 | [diff] [blame] | 1063 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 1064 | return MakeWorkload<ClPreluWorkload>(descriptor, info, m_CLCompileContext); |
narpra01 | b89b05f | 2019-01-16 09:53:09 +0000 | [diff] [blame] | 1065 | } |
| 1066 | |
Ryan OShea | 2323af4 | 2020-05-13 16:36:19 +0100 | [diff] [blame] | 1067 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateQLstm(const QLstmQueueDescriptor& descriptor, |
| 1068 | const WorkloadInfo& info) const |
| 1069 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 1070 | return std::make_unique<ClQLstmWorkload>(descriptor, info, m_CLCompileContext); |
Ryan OShea | 2323af4 | 2020-05-13 16:36:19 +0100 | [diff] [blame] | 1071 | } |
| 1072 | |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 1073 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateQuantize(const QuantizeQueueDescriptor& descriptor, |
| 1074 | const WorkloadInfo& info) const |
Aron Virginas-Tar | 7a3e2fe | 2019-06-27 18:54:47 +0100 | [diff] [blame] | 1075 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 1076 | return MakeWorkload<ClQuantizeWorkload>(descriptor, info, m_CLCompileContext); |
James Conroy | d2aa85e | 2019-07-01 17:12:40 +0100 | [diff] [blame] | 1077 | } |
| 1078 | |
Ferran Balaguer | 737d9ff | 2019-08-01 09:58:08 +0100 | [diff] [blame] | 1079 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateQuantizedLstm(const QuantizedLstmQueueDescriptor& descriptor, |
| 1080 | const WorkloadInfo& info) const |
| 1081 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 1082 | return MakeWorkload<ClQuantizedLstmWorkload>(descriptor, info, m_CLCompileContext); |
Ferran Balaguer | 737d9ff | 2019-08-01 09:58:08 +0100 | [diff] [blame] | 1083 | } |
| 1084 | |
David Monahan | c11ba46 | 2020-12-03 11:09:46 +0000 | [diff] [blame] | 1085 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateRank(const RankQueueDescriptor& descriptor, |
| 1086 | const WorkloadInfo& info) const |
| 1087 | { |
| 1088 | return std::make_unique<ClRankWorkload>(descriptor, info); |
| 1089 | } |
| 1090 | |
Sadik Armagan | a274748 | 2021-02-09 10:28:54 +0000 | [diff] [blame] | 1091 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateReduce(const ReduceQueueDescriptor& descriptor, |
| 1092 | const WorkloadInfo& info) const |
| 1093 | { |
| 1094 | return std::make_unique<ClReduceWorkload>(descriptor, info); |
| 1095 | } |
| 1096 | |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 1097 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateReshape(const ReshapeQueueDescriptor& descriptor, |
| 1098 | const WorkloadInfo& info) const |
| 1099 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 1100 | return MakeWorkload<ClReshapeWorkload>(descriptor, info, m_CLCompileContext); |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 1101 | } |
| 1102 | |
| 1103 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateResize(const ResizeQueueDescriptor& descriptor, |
| 1104 | const WorkloadInfo& info) const |
| 1105 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 1106 | return MakeWorkload<ClResizeWorkload>(descriptor, info, m_CLCompileContext); |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 1107 | } |
| 1108 | |
Aron Virginas-Tar | 94c4fef | 2019-11-25 15:37:08 +0000 | [diff] [blame] | 1109 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateSlice(const SliceQueueDescriptor& descriptor, |
| 1110 | const WorkloadInfo& info) const |
| 1111 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 1112 | return MakeWorkload<ClSliceWorkload>(descriptor, info, m_CLCompileContext); |
Aron Virginas-Tar | 94c4fef | 2019-11-25 15:37:08 +0000 | [diff] [blame] | 1113 | } |
| 1114 | |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 1115 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateSoftmax(const SoftmaxQueueDescriptor& descriptor, |
| 1116 | const WorkloadInfo& info) const |
| 1117 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 1118 | return std::make_unique<ClSoftmaxWorkload>(descriptor, |
| 1119 | info, |
| 1120 | m_MemoryManager->GetIntraLayerManager(), |
| 1121 | m_CLCompileContext); |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 1122 | } |
| 1123 | |
| 1124 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateSpaceToBatchNd(const SpaceToBatchNdQueueDescriptor& descriptor, |
| 1125 | const WorkloadInfo& info) const |
| 1126 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 1127 | return MakeWorkload<ClSpaceToBatchNdWorkload>(descriptor, info, m_CLCompileContext); |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 1128 | } |
| 1129 | |
| 1130 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateSpaceToDepth(const SpaceToDepthQueueDescriptor& descriptor, |
| 1131 | const WorkloadInfo& info) const |
| 1132 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 1133 | return MakeWorkload<ClSpaceToDepthWorkload>(descriptor, info, m_CLCompileContext); |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 1134 | } |
| 1135 | |
| 1136 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateSplitter(const SplitterQueueDescriptor& descriptor, |
| 1137 | const WorkloadInfo& info) const |
| 1138 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 1139 | return MakeWorkload<ClSplitterWorkload>(descriptor, info, m_CLCompileContext); |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 1140 | } |
| 1141 | |
Matthew Jackson | d516610 | 2019-07-31 14:06:28 +0100 | [diff] [blame] | 1142 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateStack(const StackQueueDescriptor& descriptor, |
| 1143 | const WorkloadInfo& info) const |
| 1144 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 1145 | return MakeWorkload<ClStackWorkload>(descriptor, info, m_CLCompileContext); |
Matthew Jackson | d516610 | 2019-07-31 14:06:28 +0100 | [diff] [blame] | 1146 | } |
| 1147 | |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 1148 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateStridedSlice(const StridedSliceQueueDescriptor& descriptor, |
Aron Virginas-Tar | b280196 | 2019-09-30 11:24:53 +0100 | [diff] [blame] | 1149 | const WorkloadInfo& info) const |
| 1150 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 1151 | return MakeWorkload<ClStridedSliceWorkload>(descriptor, info, m_CLCompileContext); |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 1152 | } |
| 1153 | |
| 1154 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateSubtraction(const SubtractionQueueDescriptor& descriptor, |
| 1155 | const WorkloadInfo& info) const |
| 1156 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 1157 | return MakeWorkload<ClSubtractionWorkload>(descriptor, info, m_CLCompileContext); |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 1158 | } |
| 1159 | |
Mike Kelly | c9ea45a | 2020-02-28 18:11:58 +0000 | [diff] [blame] | 1160 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateTranspose(const TransposeQueueDescriptor& descriptor, |
| 1161 | const WorkloadInfo& info) const |
| 1162 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 1163 | return MakeWorkload<ClTransposeWorkload>(descriptor, info, m_CLCompileContext); |
Mike Kelly | c9ea45a | 2020-02-28 18:11:58 +0000 | [diff] [blame] | 1164 | } |
| 1165 | |
Aron Virginas-Tar | 8168f40 | 2019-10-04 13:10:16 +0100 | [diff] [blame] | 1166 | std::unique_ptr<IWorkload> ClWorkloadFactory::CreateTransposeConvolution2d( |
| 1167 | const TransposeConvolution2dQueueDescriptor& descriptor, |
| 1168 | const WorkloadInfo& info) const |
| 1169 | { |
Sadik Armagan | e944475 | 2020-12-02 11:28:58 +0000 | [diff] [blame] | 1170 | return MakeWorkload<ClTransposeConvolution2dWorkload>(descriptor, |
| 1171 | info, |
| 1172 | m_MemoryManager->GetIntraLayerManager(), |
| 1173 | m_CLCompileContext); |
Aron Virginas-Tar | b280196 | 2019-09-30 11:24:53 +0100 | [diff] [blame] | 1174 | } |
| 1175 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1176 | } // namespace armnn |