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.

acl_op_compiler.h 4.2 kB

4 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115
  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_EXTERNAL_ACL_ACL_OP_COMPILER_H_
  17. #define INC_EXTERNAL_ACL_ACL_OP_COMPILER_H_
  18. #include "acl_base.h"
  19. #include "acl_op.h"
  20. #ifdef __cplusplus
  21. extern "C" {
  22. #endif
  23. typedef enum aclCompileType {
  24. ACL_COMPILE_SYS,
  25. ACL_COMPILE_UNREGISTERED
  26. } aclopCompileType;
  27. typedef enum {
  28. ACL_PRECISION_MODE,
  29. ACL_AICORE_NUM,
  30. ACL_AUTO_TUNE_MODE,
  31. ACL_OP_SELECT_IMPL_MODE,
  32. ACL_OPTYPELIST_FOR_IMPLMODE,
  33. ACL_OP_DEBUG_LEVEL,
  34. ACL_DEBUG_DIR,
  35. ACL_OP_COMPILER_CACHE_MODE,
  36. ACL_OP_COMPILER_CACHE_DIR
  37. } aclCompileOpt;
  38. /**
  39. * @ingroup AscendCL
  40. * @brief compile op
  41. *
  42. * @param opType [IN] op type
  43. * @param numInputs [IN] number of inputs
  44. * @param inputDesc [IN] pointer to array of input tensor descriptions
  45. * @param numOutputs [IN] number of outputs
  46. * @param outputDesc [IN] pointer to array of output tensor descriptions
  47. * @param attr [IN] pointer to instance of aclopAttr.
  48. * may pass nullptr if the op has no attribute
  49. * @param engineType [IN] engine type
  50. * @param compileFlag [IN] compile flag
  51. * @param opPath [IN] path of op
  52. *
  53. * @retval ACL_SUCCESS The function is successfully executed.
  54. * @retval OtherValues Failure
  55. */
  56. ACL_FUNC_VISIBILITY aclError aclopCompile(const char *opType,
  57. int numInputs,
  58. const aclTensorDesc *const inputDesc[],
  59. int numOutputs,
  60. const aclTensorDesc *const outputDesc[],
  61. const aclopAttr *attr,
  62. aclopEngineType engineType,
  63. aclopCompileType compileFlag,
  64. const char *opPath);
  65. /**
  66. * @ingroup AscendCL
  67. * @brief compile and execute op
  68. *
  69. * @param opType [IN] op type
  70. * @param numInputs [IN] number of inputs
  71. * @param inputDesc [IN] pointer to array of input tensor descriptions
  72. * @param inputs [IN] pointer to array of input buffers
  73. * @param numOutputs [IN] number of outputs
  74. * @param outputDesc [IN] pointer to array of output tensor descriptions
  75. * @param outputs [IN] pointer to array of outputs buffers
  76. * @param attr [IN] pointer to instance of aclopAttr.
  77. * may pass nullptr if the op has no attribute
  78. * @param engineType [IN] engine type
  79. * @param compileFlag [IN] compile flag
  80. * @param opPath [IN] path of op
  81. * @param stream [IN] stream handle
  82. *
  83. * @retval ACL_SUCCESS The function is successfully executed.
  84. * @retval OtherValues Failure
  85. */
  86. ACL_FUNC_VISIBILITY aclError aclopCompileAndExecute(const char *opType,
  87. int numInputs, const aclTensorDesc *const inputDesc[], const aclDataBuffer *const inputs[],
  88. int numOutputs, const aclTensorDesc *const outputDesc[], aclDataBuffer *const outputs[],
  89. const aclopAttr *attr, aclopEngineType engineType, aclopCompileType compileFlag,
  90. const char *opPath, aclrtStream stream);
  91. /**
  92. * @ingroup AscendCL
  93. * @brief set compile option
  94. *
  95. * @param aclCompileOpt [IN] compile option
  96. * @param value [IN] pointer for the option value
  97. *
  98. * @retval ACL_SUCCESS The function is successfully executed.
  99. * @retval OtherValues Failure
  100. */
  101. ACL_FUNC_VISIBILITY aclError aclSetCompileopt(aclCompileOpt opt, const char *value);
  102. #ifdef __cplusplus
  103. }
  104. #endif
  105. #endif // INC_EXTERNAL_ACL_ACL_OP_COMPILER_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