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.

CMakeLists.txt 1.5 kB

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768
  1. ############ libge_runtime.so ############
  2. set(GE_SRC_LIST
  3. "model_runner.cc"
  4. "runtime_model.cc"
  5. "output.cc"
  6. "task/aicpu_task.cc"
  7. "task/cce_task.cc"
  8. "task/tbe_task.cc"
  9. "task/event_record_task.cc"
  10. "task/event_wait_task.cc"
  11. "task/stream_active_task.cc"
  12. "task/stream_switch_task.cc"
  13. "task/hccl_task.cc"
  14. "task/memcpy_async_task.cc"
  15. "task/profiler_task.cc"
  16. )
  17. add_library(ge_runtime SHARED ${GE_SRC_LIST})
  18. target_compile_options(ge_runtime PRIVATE
  19. -Werror
  20. -O2
  21. )
  22. target_compile_definitions(ge_runtime PRIVATE
  23. PROTOBUF_INLINE_NOT_IN_HEADERS=0
  24. )
  25. target_include_directories(ge_runtime PRIVATE
  26. ${CMAKE_CURRENT_LIST_DIR}
  27. ${GE_CODE_DIR}
  28. ${GE_CODE_DIR}/ge
  29. ${GE_CODE_DIR}/inc
  30. ${GE_CODE_DIR}/inc/graph
  31. ${GE_CODE_DIR}/inc/external
  32. ${GE_CODE_DIR}/inc/framework
  33. ${GE_CODE_DIR}/inc/framework/common
  34. ${GE_CODE_DIR}/inc/framework/ge_runtime
  35. ${GE_CODE_DIR}/inc/cce
  36. ${GE_CODE_DIR}/third_party/fwkacllib/inc
  37. ${METADEF_DIR}
  38. ${METADEF_DIR}/inc
  39. ${METADEF_DIR}/inc/external/graph
  40. ${METADEF_DIR}/inc/external
  41. ${METADEF_DIR}/inc/graph
  42. ${CMAKE_BINARY_DIR}
  43. ${CMAKE_BINARY_DIR}/proto/ge
  44. )
  45. target_link_libraries(ge_runtime PRIVATE
  46. $<BUILD_INTERFACE:intf_pub>
  47. -Wl,--no-as-needed
  48. slog
  49. runtime
  50. c_sec
  51. graph
  52. -Wl,--as-needed
  53. -lrt
  54. -ldl
  55. )
  56. ############ install ############
  57. set(INSTALL_BASE_DIR "")
  58. set(INSTALL_LIBRARY_DIR lib)
  59. install(TARGETS ge_runtime OPTIONAL
  60. LIBRARY DESTINATION ${INSTALL_LIBRARY_DIR}
  61. )

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