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.

module.mk 8.0 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216
  1. LOCAL_PATH := $(call my-dir)
  2. local_ge_executor_src_files := \
  3. ge_executor.cc \
  4. ../common/profiling/profiling_manager.cc \
  5. ../common/dump/dump_properties.cc \
  6. ../common/dump/dump_manager.cc \
  7. ../common/dump/dump_op.cc \
  8. ../common/ge/plugin_manager.cc \
  9. ../common/ge/op_tiling_manager.cc \
  10. ../graph/load/graph_loader.cc \
  11. ../graph/execute/graph_execute.cc \
  12. ../omm/csa_interact.cc \
  13. ../graph/manager/graph_manager_utils.cc \
  14. ../graph/manager/graph_var_manager.cc \
  15. ../graph/manager/rdma_pool_allocator.cc \
  16. ../graph/manager/graph_mem_allocator.cc \
  17. ../graph/manager/graph_caching_allocator.cc \
  18. ../graph/manager/trans_var_data_utils.cc \
  19. ../graph/manager/util/debug.cc \
  20. ../model/ge_model.cc \
  21. ../model/ge_root_model.cc \
  22. ../graph/load/new_model_manager/davinci_model.cc \
  23. ../graph/load/new_model_manager/davinci_model_parser.cc \
  24. ../graph/load/new_model_manager/model_manager.cc \
  25. ../graph/load/new_model_manager/tbe_handle_store.cc \
  26. ../graph/load/new_model_manager/cpu_queue_schedule.cc \
  27. ../graph/load/new_model_manager/model_utils.cc \
  28. ../graph/load/new_model_manager/aipp_utils.cc \
  29. ../graph/load/new_model_manager/data_inputer.cc \
  30. ../graph/load/new_model_manager/data_dumper.cc \
  31. ../graph/load/new_model_manager/zero_copy_task.cc \
  32. ../graph/load/new_model_manager/zero_copy_offset.cc \
  33. ../graph/load/new_model_manager/task_info/task_info.cc \
  34. ../graph/load/new_model_manager/task_info/event_record_task_info.cc \
  35. ../graph/load/new_model_manager/task_info/event_wait_task_info.cc \
  36. ../graph/load/new_model_manager/task_info/fusion_start_task_info.cc \
  37. ../graph/load/new_model_manager/task_info/fusion_stop_task_info.cc \
  38. ../graph/load/new_model_manager/task_info/kernel_ex_task_info.cc \
  39. ../graph/load/new_model_manager/task_info/kernel_task_info.cc \
  40. ../graph/load/new_model_manager/task_info/label_set_task_info.cc \
  41. ../graph/load/new_model_manager/task_info/label_switch_by_index_task_info.cc \
  42. ../graph/load/new_model_manager/task_info/label_goto_ex_task_info.cc \
  43. ../graph/load/new_model_manager/task_info/memcpy_async_task_info.cc \
  44. ../graph/load/new_model_manager/task_info/memcpy_addr_async_task_info.cc \
  45. ../graph/load/new_model_manager/task_info/profiler_trace_task_info.cc \
  46. ../graph/load/new_model_manager/task_info/stream_active_task_info.cc \
  47. ../graph/load/new_model_manager/task_info/stream_switch_task_info.cc \
  48. ../graph/load/new_model_manager/task_info/stream_switchn_task_info.cc \
  49. ../graph/load/new_model_manager/task_info/end_graph_task_info.cc \
  50. ../graph/load/new_model_manager/task_info/super_kernel/super_kernel_factory.cc \
  51. ../graph/load/new_model_manager/task_info/super_kernel/super_kernel.cc \
  52. ../single_op/single_op_manager.cc \
  53. ../single_op/single_op_model.cc \
  54. ../single_op/single_op.cc \
  55. ../single_op/stream_resource.cc \
  56. ../single_op/task/op_task.cc \
  57. ../single_op/task/build_task_utils.cc \
  58. ../single_op/task/tbe_task_builder.cc \
  59. ../single_op/task/aicpu_task_builder.cc \
  60. ../single_op/task/aicpu_kernel_task_builder.cc \
  61. ../hybrid/hybrid_davinci_model_stub.cc\
  62. ../hybrid/node_executor/aicpu/aicpu_ext_info.cc \
  63. local_ge_executor_c_include := \
  64. proto/insert_op.proto \
  65. proto/op_mapping_info.proto \
  66. proto/dump_task.proto \
  67. proto/ge_ir.proto \
  68. proto/task.proto \
  69. proto/om.proto \
  70. $(TOPDIR)inc/external \
  71. $(TOPDIR)inc/external/graph \
  72. $(TOPDIR)inc/framework \
  73. $(TOPDIR)inc \
  74. $(LOCAL_PATH)/../ \
  75. $(TOPDIR)libc_sec/include \
  76. third_party/protobuf/include \
  77. third_party/json/include \
  78. local_ge_executor_shared_library := \
  79. libprotobuf \
  80. libc_sec \
  81. libge_common \
  82. libruntime \
  83. libslog \
  84. libmmpa \
  85. libgraph \
  86. libregister \
  87. libmsprof \
  88. liberror_manager \
  89. local_ge_executor_ldflags := -lrt -ldl \
  90. #compile arm device dynamic lib
  91. include $(CLEAR_VARS)
  92. LOCAL_MODULE := libge_executor
  93. LOCAL_CFLAGS += -Werror
  94. LOCAL_CFLAGS += -DPROTOBUF_INLINE_NOT_IN_HEADERS=0 -O2 -DDAVINCI_SUPPORT_PROFILING
  95. LOCAL_SRC_FILES := $(local_ge_executor_src_files)
  96. LOCAL_C_INCLUDES := $(local_ge_executor_c_include)
  97. LOCAL_SHARED_LIBRARIES := $(local_ge_executor_shared_library)
  98. ifeq ($(device_os),android)
  99. LOCAL_LDFLAGS += -ldl
  100. LOCAL_LDLIBS += -L$(PWD)/prebuilts/clang/linux-x86/aarch64/android-ndk-r21/sysroot/usr/lib/aarch64-linux-android/29 -llog
  101. else
  102. LOCAL_LDFLAGS += $(local_ge_executor_ldflags)
  103. endif
  104. include $(BUILD_SHARED_LIBRARY)
  105. #compile x86 host dynamic lib
  106. include $(CLEAR_VARS)
  107. LOCAL_MODULE := libge_executor
  108. LOCAL_CFLAGS += -Werror
  109. LOCAL_CFLAGS += -DPROTOBUF_INLINE_NOT_IN_HEADERS=0 -DDAVINCI_SUPPORT_PROFILING
  110. ifeq ($(DEBUG), 1)
  111. LOCAL_CFLAGS += -g -O0
  112. else
  113. LOCAL_CFLAGS += -O2
  114. endif
  115. LOCAL_SRC_FILES := $(local_ge_executor_src_files)
  116. LOCAL_C_INCLUDES := $(local_ge_executor_c_include)
  117. LOCAL_SHARED_LIBRARIES := \
  118. libprotobuf \
  119. libc_sec \
  120. libge_common \
  121. libruntime \
  122. libslog \
  123. libmmpa \
  124. libgraph \
  125. libregister \
  126. libmsprof \
  127. liberror_manager \
  128. LOCAL_LDFLAGS += $(local_ge_executor_ldflags)
  129. include $(BUILD_HOST_SHARED_LIBRARY)
  130. #compile for host static lib
  131. include $(CLEAR_VARS)
  132. LOCAL_MODULE := libge_executor
  133. LOCAL_CFLAGS += -Werror
  134. LOCAL_CFLAGS += -DPROTOBUF_INLINE_NOT_IN_HEADERS=0 -DDAVINCI_SUPPORT_PROFILING
  135. ifeq ($(DEBUG), 1)
  136. LOCAL_CFLAGS += -g -O0
  137. else
  138. LOCAL_CFLAGS += -O2
  139. endif
  140. LOCAL_SRC_FILES := $(local_ge_executor_src_files)
  141. LOCAL_C_INCLUDES := $(local_ge_executor_c_include)
  142. LOCAL_STATIC_LIBRARIES := \
  143. libge_common \
  144. libgraph \
  145. libregister \
  146. libprotobuf \
  147. LOCAL_SHARED_LIBRARIES := \
  148. libc_sec \
  149. libruntime \
  150. libslog \
  151. libmmpa \
  152. libmsprof \
  153. LOCAL_LDFLAGS += $(local_ge_executor_ldflags)
  154. include $(BUILD_HOST_STATIC_LIBRARY)
  155. #compile for device static lib
  156. include $(CLEAR_VARS)
  157. LOCAL_MODULE := libge_executor
  158. LOCAL_CFLAGS += -Werror
  159. LOCAL_CFLAGS += -DPROTOBUF_INLINE_NOT_IN_HEADERS=0 -DDAVINCI_SUPPORT_PROFILING
  160. ifeq ($(DEBUG), 1)
  161. LOCAL_CFLAGS += -g -O0
  162. else
  163. LOCAL_CFLAGS += -O2
  164. endif
  165. LOCAL_SRC_FILES := $(local_ge_executor_src_files)
  166. LOCAL_C_INCLUDES := $(local_ge_executor_c_include)
  167. LOCAL_STATIC_LIBRARIES := \
  168. libge_common \
  169. libgraph \
  170. libregister \
  171. libprotobuf \
  172. LOCAL_SHARED_LIBRARIES := \
  173. libc_sec \
  174. libruntime \
  175. libslog \
  176. libmmpa \
  177. libmsprof \
  178. ifeq ($(device_os),android)
  179. LOCAL_LDFLAGS += -ldl
  180. LOCAL_LDLIBS += -L$(PWD)/prebuilts/clang/linux-x86/aarch64/android-ndk-r21/sysroot/usr/lib/aarch64-linux-android/29 -llog
  181. else
  182. LOCAL_LDFLAGS += $(local_ge_executor_ldflags)
  183. endif
  184. include $(BUILD_STATIC_LIBRARY)

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