You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.

tensor_value.h 3.0 kB

5 years ago
5 years ago
4 years ago
4 years ago
4 years ago
5 years ago
5 years ago
4 years ago
4 years ago
5 years ago
5 years ago
4 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124
  1. /**
  2. * Copyright 2019-2020 Huawei Technologies Co., Ltd
  3. *
  4. * Licensed under the Apache License, Version 2.0 (the "License");
  5. * you may not use this file except in compliance with the License.
  6. * You may obtain a copy of the License at
  7. *
  8. * http://www.apache.org/licenses/LICENSE-2.0
  9. *
  10. * Unless required by applicable law or agreed to in writing, software
  11. * distributed under the License is distributed on an "AS IS" BASIS,
  12. * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  13. * See the License for the specific language governing permissions and
  14. * limitations under the License.
  15. */
  16. #ifndef GE_HYBRID_COMMON_TENSOR_VALUE_H_
  17. #define GE_HYBRID_COMMON_TENSOR_VALUE_H_
  18. #include <atomic>
  19. #include <cstddef>
  20. #include <memory>
  21. #include "memory/memory_api.h"
  22. #include "framework/common/util.h"
  23. namespace ge {
  24. namespace hybrid {
  25. class NpuMemoryAllocator;
  26. class AllocationAttr;
  27. class TensorBuffer {
  28. public:
  29. static std::unique_ptr<TensorBuffer> Create(NpuMemoryAllocator *allocator,
  30. size_t size,
  31. AllocationAttr *attr = nullptr);
  32. static std::unique_ptr<TensorBuffer> Create(void *buffer, size_t size);
  33. TensorBuffer(const TensorBuffer &) = delete;
  34. TensorBuffer &operator = (const TensorBuffer &) = delete;
  35. ~TensorBuffer();
  36. void* Release() {
  37. GELOGI("===========buffer is %p",buffer_);
  38. auto ret = buffer_;
  39. buffer_ = nullptr;
  40. GELOGI("============ret is %p", ret);
  41. return ret;
  42. }
  43. void *GetData() {
  44. return buffer_;
  45. }
  46. size_t GetSize() const {
  47. return size_;
  48. }
  49. MemStorageType GetMemType() const {
  50. return mem_type_;
  51. }
  52. private:
  53. TensorBuffer(NpuMemoryAllocator *allocator, void *buffer, size_t size, MemStorageType mem_type = HBM);
  54. NpuMemoryAllocator *allocator_ = nullptr;
  55. void *buffer_ = nullptr;
  56. size_t size_ = 0;
  57. MemStorageType mem_type_;
  58. };
  59. class TensorValue {
  60. public:
  61. TensorValue() = default;
  62. explicit TensorValue(std::shared_ptr<TensorBuffer> buffer);
  63. TensorValue(void *buffer, size_t size);
  64. ~TensorValue();
  65. void Destroy();
  66. void *Release() {
  67. return buffer_->Release();
  68. }
  69. bool IsEmpty() {
  70. return ref_buffer_ == nullptr && buffer_ == nullptr;
  71. }
  72. const void *GetData() const;
  73. std::string DebugString() const;
  74. void SetName(const std::string &name) {
  75. name_ = name;
  76. }
  77. MemStorageType GetMemType() const {
  78. return buffer_->GetMemType();
  79. }
  80. void *MutableData();
  81. size_t GetSize() const;
  82. template<typename T>
  83. Status CopyScalarValueToHost(T &value) const {
  84. GE_CHECK_GE(this->GetSize(), sizeof(value));
  85. return rtMemcpy(&value, sizeof(value), this->GetData(), sizeof(value), RT_MEMCPY_DEVICE_TO_HOST);
  86. }
  87. private:
  88. std::shared_ptr<TensorBuffer> buffer_;
  89. std::string name_;
  90. // for weights and variables
  91. void *ref_buffer_ = nullptr;
  92. size_t ref_size_ = 0;
  93. // shape
  94. };
  95. } // namespace hybrid
  96. } // namespace ge
  97. #endif // GE_HYBRID_COMMON_TENSOR_VALUE_H_

图引擎模块(GE)是MindSpore的一个子模块,其代码由C++实现,位于前端模块ME和底层硬件之间,起到承接作用。图引擎模块以ME下发的图作为输入,然后进行一系列的深度图优化操作,最后输出一张可以在底层硬件上高效运行的图。GE针对昇腾AI处理器的硬件结构特点,做了特定的优化工作,以此来充分发挥出昇腾AI处理器的强大算力。在进行模型训练/推理时,GE会被自动调用而用户并不感知。GE主要由GE API和GE Core两部分组成,详细的架构图如下所示