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 | |
Francis Murtagh | e73eda9 | 2021-05-21 13:36:54 +0100 | [diff] [blame] | 22 | #include <CL/cl_ext.h> |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 23 | #include <arm_compute/core/CL/CLKernelLibrary.h> |
| 24 | |
| 25 | namespace armnn |
| 26 | { |
| 27 | |
| 28 | class IClImportTensorHandle : public IAclTensorHandle |
| 29 | { |
| 30 | public: |
| 31 | virtual arm_compute::ICLTensor& GetTensor() = 0; |
| 32 | virtual arm_compute::ICLTensor const& GetTensor() const = 0; |
| 33 | virtual arm_compute::DataType GetDataType() const = 0; |
| 34 | virtual void SetMemoryGroup(const std::shared_ptr<arm_compute::IMemoryGroup>& memoryGroup) = 0; |
| 35 | }; |
| 36 | |
| 37 | class ClImportTensorHandle : public IClImportTensorHandle |
| 38 | { |
| 39 | public: |
| 40 | ClImportTensorHandle(const TensorInfo& tensorInfo, MemorySourceFlags importFlags) |
| 41 | : m_ImportFlags(importFlags) |
| 42 | { |
| 43 | armnn::armcomputetensorutils::BuildArmComputeTensor(m_Tensor, tensorInfo); |
| 44 | } |
| 45 | |
| 46 | ClImportTensorHandle(const TensorInfo& tensorInfo, |
| 47 | DataLayout dataLayout, |
| 48 | MemorySourceFlags importFlags) |
| 49 | : m_ImportFlags(importFlags) |
| 50 | { |
| 51 | armnn::armcomputetensorutils::BuildArmComputeTensor(m_Tensor, tensorInfo, dataLayout); |
| 52 | } |
| 53 | |
| 54 | arm_compute::CLTensor& GetTensor() override { return m_Tensor; } |
| 55 | arm_compute::CLTensor const& GetTensor() const override { return m_Tensor; } |
| 56 | virtual void Allocate() override {} |
| 57 | virtual void Manage() override {} |
| 58 | |
| 59 | virtual const void* Map(bool blocking = true) const override |
| 60 | { |
| 61 | IgnoreUnused(blocking); |
| 62 | return static_cast<const void*>(m_Tensor.buffer() + m_Tensor.info()->offset_first_element_in_bytes()); |
| 63 | } |
| 64 | |
| 65 | virtual void Unmap() const override {} |
| 66 | |
| 67 | virtual ITensorHandle* GetParent() const override { return nullptr; } |
| 68 | |
| 69 | virtual arm_compute::DataType GetDataType() const override |
| 70 | { |
| 71 | return m_Tensor.info()->data_type(); |
| 72 | } |
| 73 | |
| 74 | virtual void SetMemoryGroup(const std::shared_ptr<arm_compute::IMemoryGroup>& memoryGroup) override |
| 75 | { |
| 76 | IgnoreUnused(memoryGroup); |
| 77 | } |
| 78 | |
| 79 | TensorShape GetStrides() const override |
| 80 | { |
| 81 | return armcomputetensorutils::GetStrides(m_Tensor.info()->strides_in_bytes()); |
| 82 | } |
| 83 | |
| 84 | TensorShape GetShape() const override |
| 85 | { |
| 86 | return armcomputetensorutils::GetShape(m_Tensor.info()->tensor_shape()); |
| 87 | } |
| 88 | |
| 89 | void SetImportFlags(MemorySourceFlags importFlags) |
| 90 | { |
| 91 | m_ImportFlags = importFlags; |
| 92 | } |
| 93 | |
| 94 | MemorySourceFlags GetImportFlags() const override |
| 95 | { |
| 96 | return m_ImportFlags; |
| 97 | } |
| 98 | |
| 99 | virtual bool Import(void* memory, MemorySource source) override |
| 100 | { |
| 101 | if (m_ImportFlags & static_cast<MemorySourceFlags>(source)) |
| 102 | { |
| 103 | if (source == MemorySource::Malloc) |
| 104 | { |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 105 | const cl_import_properties_arm importProperties[] = |
| 106 | { |
Narumol Prangnawarat | ff9a29d | 2021-05-10 11:02:58 +0100 | [diff] [blame] | 107 | CL_IMPORT_TYPE_ARM, |
| 108 | CL_IMPORT_TYPE_HOST_ARM, |
| 109 | 0 |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 110 | }; |
| 111 | |
Narumol Prangnawarat | ff9a29d | 2021-05-10 11:02:58 +0100 | [diff] [blame] | 112 | return ClImport(importProperties, memory); |
| 113 | } |
| 114 | if (source == MemorySource::DmaBuf) |
| 115 | { |
| 116 | const cl_import_properties_arm importProperties[] = |
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 | CL_IMPORT_TYPE_ARM, |
| 119 | CL_IMPORT_TYPE_DMA_BUF_ARM, |
| 120 | 0 |
| 121 | }; |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 122 | |
Narumol Prangnawarat | ff9a29d | 2021-05-10 11:02:58 +0100 | [diff] [blame] | 123 | return ClImport(importProperties, memory); |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 124 | |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 125 | } |
| 126 | else |
| 127 | { |
| 128 | throw MemoryImportException("ClImportTensorHandle::Import flag is not supported"); |
| 129 | } |
| 130 | } |
| 131 | else |
| 132 | { |
| 133 | throw MemoryImportException("ClImportTensorHandle::Incorrect import flag"); |
| 134 | } |
| 135 | return false; |
| 136 | } |
| 137 | |
| 138 | private: |
Narumol Prangnawarat | ff9a29d | 2021-05-10 11:02:58 +0100 | [diff] [blame] | 139 | bool ClImport(const cl_import_properties_arm* importProperties, void* memory) |
| 140 | { |
| 141 | const size_t totalBytes = m_Tensor.info()->total_size(); |
| 142 | cl_int error = CL_SUCCESS; |
| 143 | cl_mem buffer = clImportMemoryARM(arm_compute::CLKernelLibrary::get().context().get(), |
| 144 | CL_MEM_READ_WRITE, importProperties, memory, totalBytes, &error); |
| 145 | if (error != CL_SUCCESS) |
| 146 | { |
| 147 | throw MemoryImportException("ClImportTensorHandle::Invalid imported memory"); |
| 148 | } |
| 149 | |
| 150 | cl::Buffer wrappedBuffer(buffer); |
| 151 | arm_compute::Status status = m_Tensor.allocator()->import_memory(wrappedBuffer); |
| 152 | |
| 153 | // Use the overloaded bool operator of Status to check if it is success, if not throw an exception |
| 154 | // with the Status error message |
| 155 | bool imported = (status.error_code() == arm_compute::ErrorCode::OK); |
| 156 | if (!imported) |
| 157 | { |
| 158 | throw MemoryImportException(status.error_description()); |
| 159 | } |
| 160 | |
| 161 | ARMNN_ASSERT(!m_Tensor.info()->is_resizable()); |
| 162 | return imported; |
| 163 | } |
David Monahan | e4a41dc | 2021-04-14 16:55:36 +0100 | [diff] [blame] | 164 | // Only used for testing |
| 165 | void CopyOutTo(void* memory) const override |
| 166 | { |
| 167 | const_cast<armnn::ClImportTensorHandle*>(this)->Map(true); |
| 168 | switch(this->GetDataType()) |
| 169 | { |
| 170 | case arm_compute::DataType::F32: |
| 171 | armcomputetensorutils::CopyArmComputeITensorData(this->GetTensor(), |
| 172 | static_cast<float*>(memory)); |
| 173 | break; |
| 174 | case arm_compute::DataType::U8: |
| 175 | case arm_compute::DataType::QASYMM8: |
| 176 | armcomputetensorutils::CopyArmComputeITensorData(this->GetTensor(), |
| 177 | static_cast<uint8_t*>(memory)); |
| 178 | break; |
| 179 | case arm_compute::DataType::QSYMM8_PER_CHANNEL: |
| 180 | case arm_compute::DataType::QASYMM8_SIGNED: |
| 181 | armcomputetensorutils::CopyArmComputeITensorData(this->GetTensor(), |
| 182 | static_cast<int8_t*>(memory)); |
| 183 | break; |
| 184 | case arm_compute::DataType::F16: |
| 185 | armcomputetensorutils::CopyArmComputeITensorData(this->GetTensor(), |
| 186 | static_cast<armnn::Half*>(memory)); |
| 187 | break; |
| 188 | case arm_compute::DataType::S16: |
| 189 | case arm_compute::DataType::QSYMM16: |
| 190 | armcomputetensorutils::CopyArmComputeITensorData(this->GetTensor(), |
| 191 | static_cast<int16_t*>(memory)); |
| 192 | break; |
| 193 | case arm_compute::DataType::S32: |
| 194 | armcomputetensorutils::CopyArmComputeITensorData(this->GetTensor(), |
| 195 | static_cast<int32_t*>(memory)); |
| 196 | break; |
| 197 | default: |
| 198 | { |
| 199 | throw armnn::UnimplementedException(); |
| 200 | } |
| 201 | } |
| 202 | const_cast<armnn::ClImportTensorHandle*>(this)->Unmap(); |
| 203 | } |
| 204 | |
| 205 | // Only used for testing |
| 206 | void CopyInFrom(const void* memory) override |
| 207 | { |
| 208 | this->Map(true); |
| 209 | switch(this->GetDataType()) |
| 210 | { |
| 211 | case arm_compute::DataType::F32: |
| 212 | armcomputetensorutils::CopyArmComputeITensorData(static_cast<const float*>(memory), |
| 213 | this->GetTensor()); |
| 214 | break; |
| 215 | case arm_compute::DataType::U8: |
| 216 | case arm_compute::DataType::QASYMM8: |
| 217 | armcomputetensorutils::CopyArmComputeITensorData(static_cast<const uint8_t*>(memory), |
| 218 | this->GetTensor()); |
| 219 | break; |
| 220 | case arm_compute::DataType::F16: |
| 221 | armcomputetensorutils::CopyArmComputeITensorData(static_cast<const armnn::Half*>(memory), |
| 222 | this->GetTensor()); |
| 223 | break; |
| 224 | case arm_compute::DataType::S16: |
| 225 | case arm_compute::DataType::QSYMM8_PER_CHANNEL: |
| 226 | case arm_compute::DataType::QASYMM8_SIGNED: |
| 227 | armcomputetensorutils::CopyArmComputeITensorData(static_cast<const int8_t*>(memory), |
| 228 | this->GetTensor()); |
| 229 | break; |
| 230 | case arm_compute::DataType::QSYMM16: |
| 231 | armcomputetensorutils::CopyArmComputeITensorData(static_cast<const int16_t*>(memory), |
| 232 | this->GetTensor()); |
| 233 | break; |
| 234 | case arm_compute::DataType::S32: |
| 235 | armcomputetensorutils::CopyArmComputeITensorData(static_cast<const int32_t*>(memory), |
| 236 | this->GetTensor()); |
| 237 | break; |
| 238 | default: |
| 239 | { |
| 240 | throw armnn::UnimplementedException(); |
| 241 | } |
| 242 | } |
| 243 | this->Unmap(); |
| 244 | } |
| 245 | |
| 246 | arm_compute::CLTensor m_Tensor; |
| 247 | MemorySourceFlags m_ImportFlags; |
| 248 | }; |
| 249 | |
| 250 | class ClImportSubTensorHandle : public IClImportTensorHandle |
| 251 | { |
| 252 | public: |
| 253 | ClImportSubTensorHandle(IClImportTensorHandle* parent, |
| 254 | const arm_compute::TensorShape& shape, |
| 255 | const arm_compute::Coordinates& coords) |
| 256 | : m_Tensor(&parent->GetTensor(), shape, coords) |
| 257 | { |
| 258 | parentHandle = parent; |
| 259 | } |
| 260 | |
| 261 | arm_compute::CLSubTensor& GetTensor() override { return m_Tensor; } |
| 262 | arm_compute::CLSubTensor const& GetTensor() const override { return m_Tensor; } |
| 263 | |
| 264 | virtual void Allocate() override {} |
| 265 | virtual void Manage() override {} |
| 266 | |
| 267 | virtual const void* Map(bool blocking = true) const override |
| 268 | { |
| 269 | IgnoreUnused(blocking); |
| 270 | return static_cast<const void*>(m_Tensor.buffer() + m_Tensor.info()->offset_first_element_in_bytes()); |
| 271 | } |
| 272 | virtual void Unmap() const override {} |
| 273 | |
| 274 | virtual ITensorHandle* GetParent() const override { return parentHandle; } |
| 275 | |
| 276 | virtual arm_compute::DataType GetDataType() const override |
| 277 | { |
| 278 | return m_Tensor.info()->data_type(); |
| 279 | } |
| 280 | |
| 281 | virtual void SetMemoryGroup(const std::shared_ptr<arm_compute::IMemoryGroup>& memoryGroup) override |
| 282 | { |
| 283 | IgnoreUnused(memoryGroup); |
| 284 | } |
| 285 | |
| 286 | TensorShape GetStrides() const override |
| 287 | { |
| 288 | return armcomputetensorutils::GetStrides(m_Tensor.info()->strides_in_bytes()); |
| 289 | } |
| 290 | |
| 291 | TensorShape GetShape() const override |
| 292 | { |
| 293 | return armcomputetensorutils::GetShape(m_Tensor.info()->tensor_shape()); |
| 294 | } |
| 295 | |
| 296 | private: |
| 297 | // Only used for testing |
| 298 | void CopyOutTo(void* memory) const override |
| 299 | { |
| 300 | const_cast<ClImportSubTensorHandle*>(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::F16: |
| 313 | armcomputetensorutils::CopyArmComputeITensorData(this->GetTensor(), |
| 314 | static_cast<armnn::Half*>(memory)); |
| 315 | break; |
| 316 | case arm_compute::DataType::QSYMM8_PER_CHANNEL: |
| 317 | case arm_compute::DataType::QASYMM8_SIGNED: |
| 318 | armcomputetensorutils::CopyArmComputeITensorData(this->GetTensor(), |
| 319 | static_cast<int8_t*>(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<ClImportSubTensorHandle*>(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::QSYMM8_PER_CHANNEL: |
| 358 | case arm_compute::DataType::QASYMM8_SIGNED: |
| 359 | armcomputetensorutils::CopyArmComputeITensorData(static_cast<const int8_t*>(memory), |
| 360 | this->GetTensor()); |
| 361 | break; |
| 362 | case arm_compute::DataType::S16: |
| 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 | mutable arm_compute::CLSubTensor m_Tensor; |
| 380 | ITensorHandle* parentHandle = nullptr; |
| 381 | }; |
| 382 | |
| 383 | } // namespace armnn |