David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2021 Arm Ltd and Contributors. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #pragma once |
| 7 | |
| 8 | #include <aclCommon/ArmComputeTensorHandle.hpp> |
| 9 | #include <aclCommon/ArmComputeTensorUtils.hpp> |
| 10 | |
| 11 | #include <Half.hpp> |
| 12 | |
| 13 | #include <armnn/utility/PolymorphicDowncast.hpp> |
| 14 | |
| 15 | #include <arm_compute/runtime/CL/CLTensor.h> |
| 16 | #include <arm_compute/runtime/CL/CLSubTensor.h> |
| 17 | #include <arm_compute/runtime/IMemoryGroup.h> |
| 18 | #include <arm_compute/runtime/MemoryGroup.h> |
| 19 | #include <arm_compute/core/TensorShape.h> |
| 20 | #include <arm_compute/core/Coordinates.h> |
| 21 | |
Narumol Prangnawarat | 9ef3614 | 2022-01-25 15:15:34 +0000 | [diff] [blame] | 22 | #include <cl/IClTensorHandle.hpp> |
| 23 | |
Francis Murtagh | e73eda9 | 2021-05-21 13:36:54 +0100 | [diff] [blame] | 24 | #include <CL/cl_ext.h> |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 25 | #include <arm_compute/core/CL/CLKernelLibrary.h> |
| 26 | |
| 27 | namespace armnn |
| 28 | { |
| 29 | |
Narumol Prangnawarat | 9ef3614 | 2022-01-25 15:15:34 +0000 | [diff] [blame] | 30 | class ClImportTensorHandle : public IClTensorHandle |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 31 | { |
| 32 | public: |
| 33 | ClImportTensorHandle(const TensorInfo& tensorInfo, MemorySourceFlags importFlags) |
| 34 | : m_ImportFlags(importFlags) |
| 35 | { |
| 36 | armnn::armcomputetensorutils::BuildArmComputeTensor(m_Tensor, tensorInfo); |
| 37 | } |
| 38 | |
| 39 | ClImportTensorHandle(const TensorInfo& tensorInfo, |
| 40 | DataLayout dataLayout, |
| 41 | MemorySourceFlags importFlags) |
David Monahan | 6642b8a | 2021-11-04 16:31:46 +0000 | [diff] [blame] | 42 | : m_ImportFlags(importFlags), m_Imported(false) |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 43 | { |
| 44 | armnn::armcomputetensorutils::BuildArmComputeTensor(m_Tensor, tensorInfo, dataLayout); |
| 45 | } |
| 46 | |
| 47 | arm_compute::CLTensor& GetTensor() override { return m_Tensor; } |
| 48 | arm_compute::CLTensor const& GetTensor() const override { return m_Tensor; } |
| 49 | virtual void Allocate() override {} |
| 50 | virtual void Manage() override {} |
| 51 | |
| 52 | virtual const void* Map(bool blocking = true) const override |
| 53 | { |
| 54 | IgnoreUnused(blocking); |
| 55 | return static_cast<const void*>(m_Tensor.buffer() + m_Tensor.info()->offset_first_element_in_bytes()); |
| 56 | } |
| 57 | |
| 58 | virtual void Unmap() const override {} |
| 59 | |
| 60 | virtual ITensorHandle* GetParent() const override { return nullptr; } |
| 61 | |
| 62 | virtual arm_compute::DataType GetDataType() const override |
| 63 | { |
| 64 | return m_Tensor.info()->data_type(); |
| 65 | } |
| 66 | |
| 67 | virtual void SetMemoryGroup(const std::shared_ptr<arm_compute::IMemoryGroup>& memoryGroup) override |
| 68 | { |
| 69 | IgnoreUnused(memoryGroup); |
| 70 | } |
| 71 | |
| 72 | TensorShape GetStrides() const override |
| 73 | { |
| 74 | return armcomputetensorutils::GetStrides(m_Tensor.info()->strides_in_bytes()); |
| 75 | } |
| 76 | |
| 77 | TensorShape GetShape() const override |
| 78 | { |
| 79 | return armcomputetensorutils::GetShape(m_Tensor.info()->tensor_shape()); |
| 80 | } |
| 81 | |
| 82 | void SetImportFlags(MemorySourceFlags importFlags) |
| 83 | { |
| 84 | m_ImportFlags = importFlags; |
| 85 | } |
| 86 | |
| 87 | MemorySourceFlags GetImportFlags() const override |
| 88 | { |
| 89 | return m_ImportFlags; |
| 90 | } |
| 91 | |
| 92 | virtual bool Import(void* memory, MemorySource source) override |
| 93 | { |
| 94 | if (m_ImportFlags & static_cast<MemorySourceFlags>(source)) |
| 95 | { |
| 96 | if (source == MemorySource::Malloc) |
| 97 | { |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 98 | const cl_import_properties_arm importProperties[] = |
| 99 | { |
Narumol Prangnawarat | ff9a29d | 2021-05-10 11:02:58 +0100 | [diff] [blame] | 100 | CL_IMPORT_TYPE_ARM, |
| 101 | CL_IMPORT_TYPE_HOST_ARM, |
| 102 | 0 |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 103 | }; |
| 104 | |
Narumol Prangnawarat | ff9a29d | 2021-05-10 11:02:58 +0100 | [diff] [blame] | 105 | return ClImport(importProperties, memory); |
| 106 | } |
| 107 | if (source == MemorySource::DmaBuf) |
| 108 | { |
| 109 | const cl_import_properties_arm importProperties[] = |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 110 | { |
Narumol Prangnawarat | ff9a29d | 2021-05-10 11:02:58 +0100 | [diff] [blame] | 111 | CL_IMPORT_TYPE_ARM, |
| 112 | CL_IMPORT_TYPE_DMA_BUF_ARM, |
Francis Murtagh | f5d5e6c | 2021-07-26 13:19:33 +0100 | [diff] [blame] | 113 | CL_IMPORT_DMA_BUF_DATA_CONSISTENCY_WITH_HOST_ARM, |
| 114 | CL_TRUE, |
Narumol Prangnawarat | ff9a29d | 2021-05-10 11:02:58 +0100 | [diff] [blame] | 115 | 0 |
| 116 | }; |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 117 | |
Narumol Prangnawarat | ff9a29d | 2021-05-10 11:02:58 +0100 | [diff] [blame] | 118 | return ClImport(importProperties, memory); |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 119 | |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 120 | } |
Francis Murtagh | 9db96e0 | 2021-08-13 16:15:09 +0100 | [diff] [blame] | 121 | if (source == MemorySource::DmaBufProtected) |
| 122 | { |
| 123 | const cl_import_properties_arm importProperties[] = |
| 124 | { |
| 125 | CL_IMPORT_TYPE_ARM, |
| 126 | CL_IMPORT_TYPE_DMA_BUF_ARM, |
| 127 | CL_IMPORT_TYPE_PROTECTED_ARM, |
| 128 | CL_TRUE, |
| 129 | 0 |
| 130 | }; |
| 131 | |
| 132 | return ClImport(importProperties, memory, true); |
| 133 | |
| 134 | } |
David Monahan | 6642b8a | 2021-11-04 16:31:46 +0000 | [diff] [blame] | 135 | // Case for importing memory allocated by OpenCl externally directly into the tensor |
| 136 | else if (source == MemorySource::Gralloc) |
| 137 | { |
| 138 | // m_Tensor not yet Allocated |
| 139 | if (!m_Imported && !m_Tensor.buffer()) |
| 140 | { |
| 141 | // Importing memory allocated by OpenCl into the tensor directly. |
| 142 | arm_compute::Status status = |
| 143 | m_Tensor.allocator()->import_memory(cl::Buffer(static_cast<cl_mem>(memory))); |
| 144 | m_Imported = bool(status); |
| 145 | if (!m_Imported) |
| 146 | { |
| 147 | throw MemoryImportException(status.error_description()); |
| 148 | } |
| 149 | return m_Imported; |
| 150 | } |
| 151 | |
| 152 | // m_Tensor.buffer() initially allocated with Allocate(). |
| 153 | else if (!m_Imported && m_Tensor.buffer()) |
| 154 | { |
| 155 | throw MemoryImportException( |
| 156 | "ClImportTensorHandle::Import Attempting to import on an already allocated tensor"); |
| 157 | } |
| 158 | |
| 159 | // m_Tensor.buffer() previously imported. |
| 160 | else if (m_Imported) |
| 161 | { |
| 162 | // Importing memory allocated by OpenCl into the tensor directly. |
| 163 | arm_compute::Status status = |
| 164 | m_Tensor.allocator()->import_memory(cl::Buffer(static_cast<cl_mem>(memory))); |
| 165 | m_Imported = bool(status); |
| 166 | if (!m_Imported) |
| 167 | { |
| 168 | throw MemoryImportException(status.error_description()); |
| 169 | } |
| 170 | return m_Imported; |
| 171 | } |
| 172 | else |
| 173 | { |
| 174 | throw MemoryImportException("ClImportTensorHandle::Failed to Import Gralloc Memory"); |
| 175 | } |
| 176 | } |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 177 | else |
| 178 | { |
| 179 | throw MemoryImportException("ClImportTensorHandle::Import flag is not supported"); |
| 180 | } |
| 181 | } |
| 182 | else |
| 183 | { |
| 184 | throw MemoryImportException("ClImportTensorHandle::Incorrect import flag"); |
| 185 | } |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 186 | } |
| 187 | |
Nikhil Raj | 60ab976 | 2022-01-13 09:34:44 +0000 | [diff] [blame] | 188 | virtual bool CanBeImported(void* memory, MemorySource source) override |
| 189 | { |
| 190 | if (m_ImportFlags & static_cast<MemorySourceFlags>(source)) |
| 191 | { |
| 192 | if (source == MemorySource::Malloc) |
| 193 | { |
| 194 | const cl_import_properties_arm importProperties[] = |
| 195 | { |
| 196 | CL_IMPORT_TYPE_ARM, |
| 197 | CL_IMPORT_TYPE_HOST_ARM, |
| 198 | 0 |
| 199 | }; |
| 200 | |
| 201 | size_t totalBytes = m_Tensor.info()->total_size(); |
| 202 | |
| 203 | // Round the size of the mapping to match the CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE |
| 204 | // This does not change the size of the buffer, only the size of the mapping the buffer is mapped to |
| 205 | // We do this to match the behaviour of the Import function later on. |
| 206 | auto cachelineAlignment = |
| 207 | arm_compute::CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE>(); |
Narumol Prangnawarat | e2af6f4 | 2022-01-28 17:59:18 +0000 | [diff] [blame] | 208 | auto roundedSize = totalBytes; |
| 209 | if (totalBytes % cachelineAlignment != 0) |
| 210 | { |
| 211 | roundedSize = cachelineAlignment + totalBytes - (totalBytes % cachelineAlignment); |
| 212 | } |
Nikhil Raj | 60ab976 | 2022-01-13 09:34:44 +0000 | [diff] [blame] | 213 | |
| 214 | cl_int error = CL_SUCCESS; |
| 215 | cl_mem buffer; |
| 216 | buffer = clImportMemoryARM(arm_compute::CLKernelLibrary::get().context().get(), |
| 217 | CL_MEM_READ_WRITE, importProperties, memory, roundedSize, &error); |
| 218 | |
| 219 | // If we fail to map we know the import will not succeed and can return false. |
| 220 | // There is no memory to be released if error is not CL_SUCCESS |
| 221 | if (error != CL_SUCCESS) |
| 222 | { |
| 223 | return false; |
| 224 | } |
| 225 | else |
| 226 | { |
| 227 | // If import was successful we can release the mapping knowing import will succeed at workload |
| 228 | // execution and return true |
| 229 | error = clReleaseMemObject(buffer); |
| 230 | if (error == CL_SUCCESS) |
| 231 | { |
| 232 | return true; |
| 233 | } |
| 234 | else |
| 235 | { |
| 236 | // If we couldn't release the mapping this constitutes a memory leak and throw an exception |
| 237 | throw MemoryImportException("ClImportTensorHandle::Failed to unmap cl_mem buffer: " |
| 238 | + std::to_string(error)); |
| 239 | } |
| 240 | } |
| 241 | } |
| 242 | } |
| 243 | else |
| 244 | { |
| 245 | throw MemoryImportException("ClImportTensorHandle::Incorrect import flag"); |
| 246 | } |
| 247 | return false; |
| 248 | } |
| 249 | |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 250 | private: |
Francis Murtagh | 9db96e0 | 2021-08-13 16:15:09 +0100 | [diff] [blame] | 251 | bool ClImport(const cl_import_properties_arm* importProperties, void* memory, bool isProtected = false) |
Narumol Prangnawarat | ff9a29d | 2021-05-10 11:02:58 +0100 | [diff] [blame] | 252 | { |
Jan Eilers | c1c872f | 2021-07-22 13:17:04 +0100 | [diff] [blame] | 253 | size_t totalBytes = m_Tensor.info()->total_size(); |
| 254 | |
Nikhil Raj | 60ab976 | 2022-01-13 09:34:44 +0000 | [diff] [blame] | 255 | // Round the size of the mapping to match the CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE |
| 256 | // This does not change the size of the buffer, only the size of the mapping the buffer is mapped to |
Jan Eilers | c1c872f | 2021-07-22 13:17:04 +0100 | [diff] [blame] | 257 | auto cachelineAlignment = |
| 258 | arm_compute::CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE>(); |
Narumol Prangnawarat | e2af6f4 | 2022-01-28 17:59:18 +0000 | [diff] [blame] | 259 | auto roundedSize = totalBytes; |
| 260 | if (totalBytes % cachelineAlignment != 0) |
| 261 | { |
| 262 | roundedSize = cachelineAlignment + totalBytes - (totalBytes % cachelineAlignment); |
| 263 | } |
Jan Eilers | c1c872f | 2021-07-22 13:17:04 +0100 | [diff] [blame] | 264 | |
Narumol Prangnawarat | ff9a29d | 2021-05-10 11:02:58 +0100 | [diff] [blame] | 265 | cl_int error = CL_SUCCESS; |
Francis Murtagh | 9db96e0 | 2021-08-13 16:15:09 +0100 | [diff] [blame] | 266 | cl_mem buffer; |
| 267 | if (isProtected) |
| 268 | { |
| 269 | buffer = clImportMemoryARM(arm_compute::CLKernelLibrary::get().context().get(), |
| 270 | CL_MEM_HOST_NO_ACCESS, importProperties, memory, roundedSize, &error); |
| 271 | } |
| 272 | else |
| 273 | { |
| 274 | buffer = clImportMemoryARM(arm_compute::CLKernelLibrary::get().context().get(), |
| 275 | CL_MEM_READ_WRITE, importProperties, memory, roundedSize, &error); |
| 276 | } |
| 277 | |
Narumol Prangnawarat | ff9a29d | 2021-05-10 11:02:58 +0100 | [diff] [blame] | 278 | if (error != CL_SUCCESS) |
| 279 | { |
Francis Murtagh | f5d5e6c | 2021-07-26 13:19:33 +0100 | [diff] [blame] | 280 | throw MemoryImportException("ClImportTensorHandle::Invalid imported memory" + std::to_string(error)); |
Narumol Prangnawarat | ff9a29d | 2021-05-10 11:02:58 +0100 | [diff] [blame] | 281 | } |
| 282 | |
| 283 | cl::Buffer wrappedBuffer(buffer); |
| 284 | arm_compute::Status status = m_Tensor.allocator()->import_memory(wrappedBuffer); |
| 285 | |
| 286 | // Use the overloaded bool operator of Status to check if it is success, if not throw an exception |
| 287 | // with the Status error message |
| 288 | bool imported = (status.error_code() == arm_compute::ErrorCode::OK); |
| 289 | if (!imported) |
| 290 | { |
| 291 | throw MemoryImportException(status.error_description()); |
| 292 | } |
| 293 | |
| 294 | ARMNN_ASSERT(!m_Tensor.info()->is_resizable()); |
| 295 | return imported; |
| 296 | } |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 297 | // Only used for testing |
| 298 | void CopyOutTo(void* memory) const override |
| 299 | { |
| 300 | const_cast<armnn::ClImportTensorHandle*>(this)->Map(true); |
| 301 | switch(this->GetDataType()) |
| 302 | { |
| 303 | case arm_compute::DataType::F32: |
| 304 | armcomputetensorutils::CopyArmComputeITensorData(this->GetTensor(), |
| 305 | static_cast<float*>(memory)); |
| 306 | break; |
| 307 | case arm_compute::DataType::U8: |
| 308 | case arm_compute::DataType::QASYMM8: |
| 309 | armcomputetensorutils::CopyArmComputeITensorData(this->GetTensor(), |
| 310 | static_cast<uint8_t*>(memory)); |
| 311 | break; |
| 312 | case arm_compute::DataType::QSYMM8_PER_CHANNEL: |
| 313 | case arm_compute::DataType::QASYMM8_SIGNED: |
| 314 | armcomputetensorutils::CopyArmComputeITensorData(this->GetTensor(), |
| 315 | static_cast<int8_t*>(memory)); |
| 316 | break; |
| 317 | case arm_compute::DataType::F16: |
| 318 | armcomputetensorutils::CopyArmComputeITensorData(this->GetTensor(), |
| 319 | static_cast<armnn::Half*>(memory)); |
| 320 | break; |
| 321 | case arm_compute::DataType::S16: |
| 322 | case arm_compute::DataType::QSYMM16: |
| 323 | armcomputetensorutils::CopyArmComputeITensorData(this->GetTensor(), |
| 324 | static_cast<int16_t*>(memory)); |
| 325 | break; |
| 326 | case arm_compute::DataType::S32: |
| 327 | armcomputetensorutils::CopyArmComputeITensorData(this->GetTensor(), |
| 328 | static_cast<int32_t*>(memory)); |
| 329 | break; |
| 330 | default: |
| 331 | { |
| 332 | throw armnn::UnimplementedException(); |
| 333 | } |
| 334 | } |
| 335 | const_cast<armnn::ClImportTensorHandle*>(this)->Unmap(); |
| 336 | } |
| 337 | |
| 338 | // Only used for testing |
| 339 | void CopyInFrom(const void* memory) override |
| 340 | { |
| 341 | this->Map(true); |
| 342 | switch(this->GetDataType()) |
| 343 | { |
| 344 | case arm_compute::DataType::F32: |
| 345 | armcomputetensorutils::CopyArmComputeITensorData(static_cast<const float*>(memory), |
| 346 | this->GetTensor()); |
| 347 | break; |
| 348 | case arm_compute::DataType::U8: |
| 349 | case arm_compute::DataType::QASYMM8: |
| 350 | armcomputetensorutils::CopyArmComputeITensorData(static_cast<const uint8_t*>(memory), |
| 351 | this->GetTensor()); |
| 352 | break; |
| 353 | case arm_compute::DataType::F16: |
| 354 | armcomputetensorutils::CopyArmComputeITensorData(static_cast<const armnn::Half*>(memory), |
| 355 | this->GetTensor()); |
| 356 | break; |
| 357 | case arm_compute::DataType::S16: |
| 358 | case arm_compute::DataType::QSYMM8_PER_CHANNEL: |
| 359 | case arm_compute::DataType::QASYMM8_SIGNED: |
| 360 | armcomputetensorutils::CopyArmComputeITensorData(static_cast<const int8_t*>(memory), |
| 361 | this->GetTensor()); |
| 362 | break; |
| 363 | case arm_compute::DataType::QSYMM16: |
| 364 | armcomputetensorutils::CopyArmComputeITensorData(static_cast<const int16_t*>(memory), |
| 365 | this->GetTensor()); |
| 366 | break; |
| 367 | case arm_compute::DataType::S32: |
| 368 | armcomputetensorutils::CopyArmComputeITensorData(static_cast<const int32_t*>(memory), |
| 369 | this->GetTensor()); |
| 370 | break; |
| 371 | default: |
| 372 | { |
| 373 | throw armnn::UnimplementedException(); |
| 374 | } |
| 375 | } |
| 376 | this->Unmap(); |
| 377 | } |
| 378 | |
| 379 | arm_compute::CLTensor m_Tensor; |
| 380 | MemorySourceFlags m_ImportFlags; |
David Monahan | 6642b8a | 2021-11-04 16:31:46 +0000 | [diff] [blame] | 381 | bool m_Imported; |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 382 | }; |
| 383 | |
Narumol Prangnawarat | 9ef3614 | 2022-01-25 15:15:34 +0000 | [diff] [blame] | 384 | class ClImportSubTensorHandle : public IClTensorHandle |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 385 | { |
| 386 | public: |
Narumol Prangnawarat | 9ef3614 | 2022-01-25 15:15:34 +0000 | [diff] [blame] | 387 | ClImportSubTensorHandle(IClTensorHandle* parent, |
| 388 | const arm_compute::TensorShape& shape, |
| 389 | const arm_compute::Coordinates& coords) |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 390 | : m_Tensor(&parent->GetTensor(), shape, coords) |
| 391 | { |
| 392 | parentHandle = parent; |
| 393 | } |
| 394 | |
| 395 | arm_compute::CLSubTensor& GetTensor() override { return m_Tensor; } |
| 396 | arm_compute::CLSubTensor const& GetTensor() const override { return m_Tensor; } |
| 397 | |
| 398 | virtual void Allocate() override {} |
| 399 | virtual void Manage() override {} |
| 400 | |
| 401 | virtual const void* Map(bool blocking = true) const override |
| 402 | { |
| 403 | IgnoreUnused(blocking); |
| 404 | return static_cast<const void*>(m_Tensor.buffer() + m_Tensor.info()->offset_first_element_in_bytes()); |
| 405 | } |
| 406 | virtual void Unmap() const override {} |
| 407 | |
| 408 | virtual ITensorHandle* GetParent() const override { return parentHandle; } |
| 409 | |
| 410 | virtual arm_compute::DataType GetDataType() const override |
| 411 | { |
| 412 | return m_Tensor.info()->data_type(); |
| 413 | } |
| 414 | |
| 415 | virtual void SetMemoryGroup(const std::shared_ptr<arm_compute::IMemoryGroup>& memoryGroup) override |
| 416 | { |
| 417 | IgnoreUnused(memoryGroup); |
| 418 | } |
| 419 | |
| 420 | TensorShape GetStrides() const override |
| 421 | { |
| 422 | return armcomputetensorutils::GetStrides(m_Tensor.info()->strides_in_bytes()); |
| 423 | } |
| 424 | |
| 425 | TensorShape GetShape() const override |
| 426 | { |
| 427 | return armcomputetensorutils::GetShape(m_Tensor.info()->tensor_shape()); |
| 428 | } |
| 429 | |
| 430 | private: |
| 431 | // Only used for testing |
| 432 | void CopyOutTo(void* memory) const override |
| 433 | { |
| 434 | const_cast<ClImportSubTensorHandle*>(this)->Map(true); |
| 435 | switch(this->GetDataType()) |
| 436 | { |
| 437 | case arm_compute::DataType::F32: |
| 438 | armcomputetensorutils::CopyArmComputeITensorData(this->GetTensor(), |
| 439 | static_cast<float*>(memory)); |
| 440 | break; |
| 441 | case arm_compute::DataType::U8: |
| 442 | case arm_compute::DataType::QASYMM8: |
| 443 | armcomputetensorutils::CopyArmComputeITensorData(this->GetTensor(), |
| 444 | static_cast<uint8_t*>(memory)); |
| 445 | break; |
| 446 | case arm_compute::DataType::F16: |
| 447 | armcomputetensorutils::CopyArmComputeITensorData(this->GetTensor(), |
| 448 | static_cast<armnn::Half*>(memory)); |
| 449 | break; |
| 450 | case arm_compute::DataType::QSYMM8_PER_CHANNEL: |
| 451 | case arm_compute::DataType::QASYMM8_SIGNED: |
| 452 | armcomputetensorutils::CopyArmComputeITensorData(this->GetTensor(), |
| 453 | static_cast<int8_t*>(memory)); |
| 454 | break; |
| 455 | case arm_compute::DataType::S16: |
| 456 | case arm_compute::DataType::QSYMM16: |
| 457 | armcomputetensorutils::CopyArmComputeITensorData(this->GetTensor(), |
| 458 | static_cast<int16_t*>(memory)); |
| 459 | break; |
| 460 | case arm_compute::DataType::S32: |
| 461 | armcomputetensorutils::CopyArmComputeITensorData(this->GetTensor(), |
| 462 | static_cast<int32_t*>(memory)); |
| 463 | break; |
| 464 | default: |
| 465 | { |
| 466 | throw armnn::UnimplementedException(); |
| 467 | } |
| 468 | } |
| 469 | const_cast<ClImportSubTensorHandle*>(this)->Unmap(); |
| 470 | } |
| 471 | |
| 472 | // Only used for testing |
| 473 | void CopyInFrom(const void* memory) override |
| 474 | { |
| 475 | this->Map(true); |
| 476 | switch(this->GetDataType()) |
| 477 | { |
| 478 | case arm_compute::DataType::F32: |
| 479 | armcomputetensorutils::CopyArmComputeITensorData(static_cast<const float*>(memory), |
| 480 | this->GetTensor()); |
| 481 | break; |
| 482 | case arm_compute::DataType::U8: |
| 483 | case arm_compute::DataType::QASYMM8: |
| 484 | armcomputetensorutils::CopyArmComputeITensorData(static_cast<const uint8_t*>(memory), |
| 485 | this->GetTensor()); |
| 486 | break; |
| 487 | case arm_compute::DataType::F16: |
| 488 | armcomputetensorutils::CopyArmComputeITensorData(static_cast<const armnn::Half*>(memory), |
| 489 | this->GetTensor()); |
| 490 | break; |
| 491 | case arm_compute::DataType::QSYMM8_PER_CHANNEL: |
| 492 | case arm_compute::DataType::QASYMM8_SIGNED: |
| 493 | armcomputetensorutils::CopyArmComputeITensorData(static_cast<const int8_t*>(memory), |
| 494 | this->GetTensor()); |
| 495 | break; |
| 496 | case arm_compute::DataType::S16: |
| 497 | case arm_compute::DataType::QSYMM16: |
| 498 | armcomputetensorutils::CopyArmComputeITensorData(static_cast<const int16_t*>(memory), |
| 499 | this->GetTensor()); |
| 500 | break; |
| 501 | case arm_compute::DataType::S32: |
| 502 | armcomputetensorutils::CopyArmComputeITensorData(static_cast<const int32_t*>(memory), |
| 503 | this->GetTensor()); |
| 504 | break; |
| 505 | default: |
| 506 | { |
| 507 | throw armnn::UnimplementedException(); |
| 508 | } |
| 509 | } |
| 510 | this->Unmap(); |
| 511 | } |
| 512 | |
| 513 | mutable arm_compute::CLSubTensor m_Tensor; |
| 514 | ITensorHandle* parentHandle = nullptr; |
| 515 | }; |
| 516 | |
| 517 | } // namespace armnn |