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.

op_tiling_registry.h 4.5 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153
  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 INC_REGISTER_OP_TILING_REGISTRY_H_
  17. #define INC_REGISTER_OP_TILING_REGISTRY_H_
  18. #include <functional>
  19. #include <map>
  20. #include <nlohmann/json.hpp>
  21. #include <sstream>
  22. #include <string>
  23. #include <vector>
  24. #include "external/register/register_types.h"
  25. #include "external/graph/tensor.h"
  26. #define REGISTER_OP_TILING_FUNC(optype, opfunc) \
  27. REGISTER_OP_TILING_FUNC_UNIQ_HELPER(optype, opfunc, __COUNTER__)
  28. #define REGISTER_OP_TILING_FUNC_UNIQ_HELPER(optype, opfunc, counter) \
  29. REGISTER_OP_TILING_FUNC_UNIQ(optype, opfunc, counter)
  30. #define REGISTER_OP_TILING_FUNC_UNIQ(optype, opfunc, counter) \
  31. static OpTilingInterf g_##optype##TilingInterf##counter(#optype, opfunc)
  32. #define REGISTER_OP_TILING_FUNC_NEW(optype, opfunc) \
  33. REGISTER_OP_TILING_UNIQ_HELPER(optype, opfunc, __COUNTER__)
  34. #define REGISTER_OP_TILING(optype, opfunc) \
  35. REGISTER_OP_TILING_UNIQ_HELPER(optype, opfunc, __COUNTER__)
  36. #define REGISTER_OP_TILING_UNIQ_HELPER(optype, opfunc, counter) \
  37. REGISTER_OP_TILING_UNIQ(optype, opfunc, counter)
  38. #define REGISTER_OP_TILING_UNIQ(optype, opfunc, counter) \
  39. static OpTilingRegistryInterf g_##optype##TilingRegistryInterf##counter(#optype, opfunc)
  40. namespace optiling {
  41. enum TensorArgType {
  42. TA_NONE,
  43. TA_SINGLE,
  44. TA_LIST,
  45. };
  46. using ByteBuffer = std::stringstream;
  47. struct TeOpTensor {
  48. std::vector<int64_t> shape;
  49. std::vector<int64_t> ori_shape;
  50. std::string format;
  51. std::string ori_format;
  52. std::string dtype;
  53. std::map<std::string, std::string> attrs;
  54. };
  55. struct TeOpTensorArg {
  56. TensorArgType arg_type;
  57. std::vector<TeOpTensor> tensor;
  58. };
  59. struct OpRunInfo {
  60. uint32_t block_dim;
  61. std::vector<int64_t> workspaces;
  62. ByteBuffer tiling_data;
  63. bool clear_atomic;
  64. };
  65. using TeOpAttrArgs = std::vector<std::string>;
  66. using TeConstTensorData = std::tuple<const uint8_t*, size_t, ge::Tensor>;
  67. struct TeOpParas {
  68. std::vector<TeOpTensorArg> inputs;
  69. std::vector<TeOpTensorArg> outputs;
  70. std::map<std::string, TeConstTensorData> const_inputs;
  71. TeOpAttrArgs attrs;
  72. std::string op_type;
  73. };
  74. using OpTilingFunc = std::function<bool(const std::string&, const TeOpParas&,
  75. const nlohmann::json& , OpRunInfo&)>;
  76. using OpTilingFuncPtr = bool(*)(const std::string&, const TeOpParas&, const nlohmann::json& , OpRunInfo&);
  77. class FMK_FUNC_HOST_VISIBILITY OpTilingInterf
  78. {
  79. public:
  80. OpTilingInterf(std::string op_type, OpTilingFunc func);
  81. ~OpTilingInterf() = default;
  82. static std::map<std::string, OpTilingFunc> &RegisteredOpInterf();
  83. static std::string OpTilingUuid;
  84. };
  85. struct OpCompileInfo {
  86. std::string str;
  87. std::string key;
  88. };
  89. using OpTilingFuncNew = std::function<bool(const TeOpParas&, const OpCompileInfo& , OpRunInfo&)>;
  90. using OpTilingFuncPtrNew = bool(*)(const TeOpParas&, const OpCompileInfo& , OpRunInfo&);
  91. class FMK_FUNC_HOST_VISIBILITY OpTilingRegistryInterf {
  92. public:
  93. OpTilingRegistryInterf(std::string op_type, OpTilingFuncNew func);
  94. ~OpTilingRegistryInterf() = default;
  95. static std::map<std::string, OpTilingFuncNew> &RegisteredOpInterf();
  96. };
  97. template <class T>
  98. ByteBuffer& ByteBufferPut(ByteBuffer &buf, const T &value)
  99. {
  100. buf.write(reinterpret_cast<const char*>(&value), sizeof(value));
  101. buf.flush();
  102. return buf;
  103. }
  104. template <class T>
  105. ByteBuffer& ByteBufferGet(ByteBuffer &buf, T &value)
  106. {
  107. buf.read(reinterpret_cast<char*>(&value), sizeof(value));
  108. return buf;
  109. }
  110. inline size_t ByteBufferGetAll(ByteBuffer &buf, char *dest, size_t dest_len)
  111. {
  112. size_t nread = 0;
  113. size_t rn = 0;
  114. do {
  115. rn = buf.readsome(dest + nread, dest_len - nread);
  116. nread += rn;
  117. } while (rn > 0 && dest_len > nread);
  118. return nread;
  119. }
  120. }
  121. #endif // INC_REGISTER_OP_TILING_REGISTRY_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