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 4.5 kB

3 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121
  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 { ACL_COMPILE_SYS, ACL_COMPILE_UNREGISTERED } aclopCompileType;
  24. typedef enum {
  25. ACL_PRECISION_MODE,
  26. ACL_AICORE_NUM,
  27. ACL_AUTO_TUNE_MODE,
  28. ACL_OP_SELECT_IMPL_MODE,
  29. ACL_OPTYPELIST_FOR_IMPLMODE,
  30. ACL_OP_DEBUG_LEVEL,
  31. ACL_DEBUG_DIR,
  32. ACL_OP_COMPILER_CACHE_MODE,
  33. ACL_OP_COMPILER_CACHE_DIR,
  34. ACL_OP_PERFORMANCE_MODE
  35. } aclCompileOpt;
  36. typedef enum aclCompileFlag { ACL_OP_COMPILE_DEFAULT, ACL_OP_COMPILE_FUZZ } aclOpCompileFlag;
  37. /**
  38. * @ingroup AscendCL
  39. * @brief compile op
  40. *
  41. * @param opType [IN] op type
  42. * @param numInputs [IN] number of inputs
  43. * @param inputDesc [IN] pointer to array of input tensor descriptions
  44. * @param numOutputs [IN] number of outputs
  45. * @param outputDesc [IN] pointer to array of output tensor descriptions
  46. * @param attr [IN] pointer to instance of aclopAttr.
  47. * may pass nullptr if the op has no attribute
  48. * @param engineType [IN] engine type
  49. * @param compileFlag [IN] compile flag
  50. * @param opPath [IN] path of op
  51. *
  52. * @retval ACL_SUCCESS The function is successfully executed.
  53. * @retval OtherValues Failure
  54. */
  55. ACL_FUNC_VISIBILITY aclError aclopCompile(const char *opType, int numInputs, const aclTensorDesc *const inputDesc[],
  56. int numOutputs, const aclTensorDesc *const outputDesc[],
  57. const aclopAttr *attr, aclopEngineType engineType,
  58. aclopCompileType compileFlag, const char *opPath);
  59. /**
  60. * @ingroup AscendCL
  61. * @brief compile and execute op
  62. *
  63. * @param opType [IN] op type
  64. * @param numInputs [IN] number of inputs
  65. * @param inputDesc [IN] pointer to array of input tensor descriptions
  66. * @param inputs [IN] pointer to array of input buffers
  67. * @param numOutputs [IN] number of outputs
  68. * @param outputDesc [IN] pointer to array of output tensor descriptions
  69. * @param outputs [IN] pointer to array of outputs buffers
  70. * @param attr [IN] pointer to instance of aclopAttr.
  71. * may pass nullptr if the op has no attribute
  72. * @param engineType [IN] engine type
  73. * @param compileFlag [IN] compile flag
  74. * @param opPath [IN] path of op
  75. * @param stream [IN] stream handle
  76. *
  77. * @retval ACL_SUCCESS The function is successfully executed.
  78. * @retval OtherValues Failure
  79. */
  80. ACL_FUNC_VISIBILITY aclError aclopCompileAndExecute(
  81. const char *opType, int numInputs, const aclTensorDesc *const inputDesc[], const aclDataBuffer *const inputs[],
  82. int numOutputs, const aclTensorDesc *const outputDesc[], aclDataBuffer *const outputs[], const aclopAttr *attr,
  83. aclopEngineType engineType, aclopCompileType compileFlag, const char *opPath, aclrtStream stream);
  84. /**
  85. * @ingroup AscendCL
  86. * @brief set compile option
  87. *
  88. * @param aclCompileOpt [IN] compile option
  89. * @param value [IN] pointer for the option value
  90. *
  91. * @retval ACL_SUCCESS The function is successfully executed.
  92. * @retval OtherValues Failure
  93. */
  94. ACL_FUNC_VISIBILITY aclError aclSetCompileopt(aclCompileOpt opt, const char *value);
  95. /**
  96. * @ingroup AscendCL
  97. * @brief set compile flag
  98. *
  99. * @param flag [IN] compile flag, ACL_OP_COMPILE_DEFAULT means compile with default mode
  100. * ACL_OP_COMPILE_FUZZ means compile with fuzz mode
  101. *
  102. * @retval ACL_SUCCESS The function is successfully executed.
  103. * @retval OtherValues Failure
  104. */
  105. ACL_FUNC_VISIBILITY aclError aclopSetCompileFlag(aclOpCompileFlag flag);
  106. #ifdef __cplusplus
  107. }
  108. #endif
  109. #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