MindSpore:图算融合报错

问题描述：

在mindspore1.1.1上做图算融合，硬件环境是Ascend 910，程序没有问题，添加运行图算融合（enable_graph_kernel=True）参数之后程序报错：

ms.context.set_context(mode=ms.context.GRAPH_MODE, device_target=args.device_target,

       save_graphs=False, max_call_depth=10000, device_id=args.device_id, enable_graph_kernel=True)

报错记录如下：

Traceback (most recent call last):

  File "/opt/python3.7.5/lib/python3.7/site-packages/mindspore/_extends/parallel_compile/akg_compiler/compiler.py", line 35, in

    run_compiler(sys.argv[1])

  File "/opt/python3.7.5/lib/python3.7/site-packages/mindspore/_extends/parallel_compile/akg_compiler/compiler.py", line 28, in run_compiler

    p = __import__("akg", globals(), locals(), ['ms'], 0)

  File "/opt/python3.7.5/lib/python3.7/site-packages/akg/__init__.py", line 84, in

    from . import autodiff

  File "/opt/python3.7.5/lib/python3.7/site-packages/akg/autodiff.py", line 19, in

    from akg.tvm._ffi.function import _init_api

  File "", line 983, in _find_and_load

  File "", line 967, in _find_and_load_unlocked

  File "", line 668, in _load_unlocked

  File "", line 638, in _load_backward_compatible

  File "/opt/python3.7.5/lib/python3.7/site-packages/akg/__init__.py", line 62, in load_module

    __import__(self.__rname, globals(), locals())

  File "/opt/python3.7.5/lib/python3.7/site-packages/akg/tvm/__init__.py", line 27, in

    from . import tensor

  File "/opt/python3.7.5/lib/python3.7/site-packages/akg/tvm/tensor.py", line 20, in

    from ._ffi.node import NodeBase, NodeGeneric, register_node, convert_to_node

  File "/opt/python3.7.5/lib/python3.7/site-packages/akg/tvm/_ffi/node.py", line 24, in

    from .object import Object, register_object, _set_class_node

  File "/opt/python3.7.5/lib/python3.7/site-packages/akg/tvm/_ffi/object.py", line 23, in

    from .base import _FFI_MODE, _RUNTIME_ONLY, check_call, _LIB, c_str

  File "/opt/python3.7.5/lib/python3.7/site-packages/akg/tvm/_ffi/base.py", line 81, in

    _LIB, _LIB_NAME = _load_lib()

  File "/opt/python3.7.5/lib/python3.7/site-packages/akg/tvm/_ffi/base.py", line 73, in _load_lib

    lib = ctypes.CDLL(lib_path[0], ctypes.RTLD_GLOBAL)

  File "/opt/python3.7.5/lib/python3.7/ctypes/__init__.py", line 364, in __init__

    self._handle = _dlopen(self._name, mode)

OSError: /opt/python3.7.5/lib/python3.7/site-packages/mindspore/lib/libakg.so: undefined symbol: pthread_atfork

Traceback (most recent call last):

Traceback (most recent call last):

  File "/opt/python3.7.5/lib/python3.7/site-packages/mindspore/_extends/parallel_compile/akg_compiler/compiler.py", line 35, in

  File "/opt/python3.7.5/lib/python3.7/site-packages/mindspore/_extends/parallel_compile/akg_compiler/compiler.py", line 35, in

    run_compiler(sys.argv[1])

    run_compiler(sys.argv[1])

  File "/opt/python3.7.5/lib/python3.7/site-packages/mindspore/_extends/parallel_compile/akg_compiler/compiler.py", line 28, in run_compiler

  File "/opt/python3.7.5/lib/python3.7/site-packages/mindspore/_extends/parallel_compile/akg_compiler/compiler.py", line 28, in run_compiler

    p = __import__("akg", globals(), locals(), ['ms'], 0)

    p = __import__("akg", globals(), locals(), ['ms'], 0)

这个报错记录太长了，超过帖子最大字数限制，我把报错记录放在附件error.txt中了

求问有哪位大神指导这个问题如何解决吗？

解决方案：

这个问题的原因在于使用了高版本的gcc编译akg时找不到pthread_atfork。

pthread_atfork是POSIX标准，在编译时要加上-pthread

-lpthread是老版本的gcc编译器用的，在新版本中应该用-pthread取代-lpthread

可以在akg 的 cmakelist下的target_link_libraries增加pthread重新编译一次