I fixed a super-annoying case of "ImportError: DLL load failed while importing" in a way that generally applies to Windows, so let me share it with the group. Here is the question:
I installed FINUFFT via pip install finufft. When I import finufft, I get this error:
ImportError: DLL load failed while importing _finufft: The specified module could not be found.
How do I fix it?
Read to the end before doing anything.
The error means that a DLL cannot find another DLL that it was linked with. But which other DLL?
Download Dependencies.
Locate your problematic DLL. In this specific case: Locate the folder ...\Lib\site-packages\finufft\ of the FINUFFT installation that you want to fix. ...\ is the path of your standard python installation or of your python virtual environment.
Start DependenciesGui.exe and use it to open the problematic DLL, e.g. ...\finufft\_finufft.cp38-win_amd64.pyd. (A .pyd is a regular DLL with some specific entry points for python.)
On the left, you will see a complete list of the problematic DLL's direct dependencies, whose dependencies you can in turn unfold by mouse click. Apart from typical Windows-DLLs, like kernel32.dll and MSVCRT.dll, and apart from the FFTW-DLLs, which should already be in the FINUFFT-folder, there will also be some - possibly missing - Linux-DLLs. For me, it was libgcc_s_seh-1.dll, libgomp-1.dll and libstdc++-6.dll. By checking their direct dependencies, I also discovered libwinpthread-1.dll as missing.
[See EDIT below!!!] I found those DLLs in Anaconda (...\Anaconda3\Library\mingw-w64\bin\), but you can probably also get them from cygwin (...\cygwin64\bin\), git (...\Git\mingw64\bin\) or anything else that downloads mingw64 and its packages on Windows.
To solve the problem, copy the respective DLLs into ...\Lib\site-packages\finufft\ and give them the exact filenames that the FINUFFT-DLL is expecting according to Dependencies. This works because Windows and because of the Windows DLL search order.
Now, import finufft should work in the specific python environment whose FINUFFT installation you fixed. Clearly, this method can be applied anytime DLL dependencies are missing.
EDIT - correction of my answer by #CristiFati: If possible, DLLs and similar things should always be built with the same toolchain. So if you don't compile them yourself, get them from as few different places as possible, i.e. don't mix regular python, Anaconda, cygwin, etc. - if possible. Of course, Windows DLLs will have a different origin from Linux DLLs.
Related
I'm currently having interop problems between a proprietry library and pyembree.
Problem
I have two mostly identical conda enviroments. Both contain this in-house library. The only difference is that I installed pyembree via these instructions in one environment. Now in this environment, the in-house library cannot load a DLL which it has no problem loading in the other. The absolute (correct) path is even supplied to the ctypes.WinDLL object:
On module level (in a module of the proprietatry library):
kernel32 = ctypes.WinDLL('kernel32', use_last_error=True)
kernel32.AddDllDirectory(os.path.abspath(os.path.dirname(__file__)))
On import this is called (where name is the correct absolute path and mode is 4096):
handle = kernel32.LoadLibraryExW(name, None, mode)
Which raises a ctypes.WinError with the message 'OSError: [WinError 127] The specified procedure could not be found.'.
Tracking overwritten DLLs
I tried working with that error message based on https://stackoverflow.com/a/2603386/5409315 . That answer says that the error message means that a DLL is loaded, but in the wrong version. Initially, that made a lot of sense - the install of pyembree could have overwritten a dependency of the proprietry lib. (Although I don't understand why that message would be triggered on load as compared to when the procedure is supposed to be used.)
However, what I found with Process Monitor was that in both environments two dozen DLLs are searched but not found (e.g. OPENGL32.dll) and that the only DLL that is successfully loaded in both environments (vcomp140.dll) does not differ between the two environments.
I cross-checked this with Dependency Walker, where I found a lot more depended DLLs but had the same ultimate result: In both environments, the depended libraries are either literally identical or bytewise identical.
The PATH environment variable is also equivalent (that is, where a path contains the environment directory, it differs in exactly that regard between the environments).
Minimal reproducible example
I can't provide a minimally reproducible example because the proprietary library is not publicly available.
Another approach
I'd like to dump all global variables just before the attempt to load the DLL in both environments, diff the output and thus find the crucial difference. Is this possible in VS Code? Is this possible in some other context (pdb, ...)?
I have a C++ application (X-Plane) for which there is a plugin which permits the use of python scripts (XPPython3 plugin). The plugin is written in C, using python CAPI, and works great, allowing me to write python scripts which get executed within the C++ application.
On Windows 10, I want to extend my python features by importing imgui. I have a python cython-built pyd file (_imgui.cp39-win_amd64.pyd).
If I place the pyd file in C\Program Files\Python39\DLLs, it works as expected: C++ application calls CAPI to python, which loads script which imports and executes imgui code.
If I place the pyd file anywhere else, embedded python either reports "module not found" -- if the pyd isn't on sys.path(), or if it is on sys.path():
ImportError: DLL load failed while importing _imgui: The parameter is incorrect.'
Changes using: os.add_dll_directory(r'D:\somewhere\else')
Does not effect whether the module is found or not, nor does it change the 'parameter incorrect' error. (see https://bugs.python.org/issue36085 for details on this change. -- my guess is add_dll_directory changes lookup for DLLs, but not for pyd?) sys.path appears to be used for locating pyd.
Yes, the pyd is compiled with python3.9: I've compiled it both with mingw and with visual studio toolchains, in case that might be a difference.
For fun, I moved python-standard _zoneinfo.pyd from Python39\DLLs and it fails in the same way in embedded python: "The parameter is incorrect". So, that would appear to rule out my specific pyd file.
The key question is/are:
Other than placing a pyd file under PythonXX\DLLs, is there a way to load a PYD in an embedded python implementation? (I want to avoid having to tell users to move my pyd file into the Python39\DLLs directory... because they'll forget.)
Note that using IDLE or python.exe, I can load pyds without error -- anywhere on sys.path -- so they don't have to be under Python39\DLLs. It's only when trying to load from embedded python that the "Parameter is incorrect" appears. And of course, this works flawlessly on Mac.
(Bonus question: what parameter? It appears to be python passing through a windows error.)
There seems to be a simple answer, though I suspect it's better characterized as a python bug.
There is nothing magical about Python39\DLLs directory.
The problem is using absolute vs relative paths in sys.path.
Python can find modules using absolute or relative paths. So if zippy.py is in folder foobar,
sys.path.append('foobar')
import zippy
# Success
Python and find, BUT NOT LOAD pyd files using relative paths. For example, move _zoneinfo.pyd from PythonXX\LDDs to foobar
sys.path.append('foobar')
import _zoneinfo
# ImportError: DLL load failed while importing _zoneinfo: The parameter is incorrect.'
Instead, use absolute path, and it will find and load PYD:
sys.path.append(r'c:\MyTest\foobar')
import _zoneinfo
# Success
So, there is actually a way to do this—that is, ship your application with the desired libraries. The solution is to use an embedded distribution and ship this with your application. You can find the correct distribution on the official Python download page corresponding to your desired version (here's the link to the lastest 3.9 release which seems to be what you're using: https://www.python.org/downloads/release/python-392/). Look for the Windows Embeddable Package.
You can then simply drop in your .pyd file alongside the standard library files (note that if your third-party library is dependent on any other libraries, you will have to include them, as well). Shipping your application with an embeddable distribution should not only solve your current issue, but will also mean that your application will work regardless of which version of Python a user has installed (or without having Python installed at all).
Pyahk documentation says:
A copy of the ANSI 32-bit dll must be provided either in the system location of your version of Windows, or in the same folder as the ahk.py file. (The required dll is not provided as part of this distribution, see the AutoHotKey.dll site for download instructions (alternate link AutoHotKey_H).)
Autohotkey.dll site does not provide any download instructions. The alternate link does not work.
I found some dlls here. Pyahk is not able to import the 32bit dll versions. I'm able to import ahk in python without the missing dll warning with the 64bit dll, but upon a call the library complains about some chinese characters.
Possibly I need to compile autohotkey.dll myself. If so I would prefer a method which does not require me to install a ton of compilation tools.
You can find the source code here: https://github.com/HotKeyIt/ahkdll
And the releases can be found here: https://github.com/HotKeyIt/ahkdll-v1-release
You'll find a couple of different versions (unicode, ansi, 64 bit, 32 bit ...).
When I import a module I built, I get this boost-python related error:
Traceback (most recent call last):
File "<string>", line 1, in <module>
ImportError: dlopen(./myMod.so, 2): Symbol not found: __ZN5boost6python7objects15function_objectERKNS1_11py_functionERKSt4pairIPKNS0_6detail7keywordES9_E
Referenced from: ./myMod.so
Expected in: flat namespace
in ./myMod.so
What does this actually mean? Why was this error raised?
Description
The problem was caused by mixing objects that compiled with libc++ and object that compiled with libstdc++.
In our case, the library myMod.so (compiled with libstdc++) need boost-python that compiled with libstdc++ (boost-python-libstdc++ from now). When boost-python is boost-python-libstdc++, it will work fine. Otherwise - on computer that its boost-python has compiled with libc++ (or another c++ library), it will have a problem loading and running it.
In our case, it happens because that libc++ developers intentionally changed the name of all of their symbols to prevent you (and save you) from mixing code from their library and code from a different one: myMod.so need a function that take an argument from the type. In libc++, this type's name is std::__1::pair. Therefore, this symbol was not found.
To understand why mixing two version of the same API is bad, consider this situation: There are two libraries: Foo and Bar. They both have a function that takes a std::string and uses it for something but they use a different c++ library. When a std::string that has been created by Foo will be passed to Bar, Bar will think that this is an instance of its c++ library's std::string and then bad things can happen (they are a completely different objects).
Note: In some cases, there would be no problem with two or more different versions of the same API in a completely different parts of a program. There will be a problem if they will pass this API's objects between them. However, checking that can be very hard, especially if they pass the API object only as a member of another object. Also, a library's initialization function can do things that should not happen twice. Another version may do these things again.
How to solve that?
You can always recompile your libraries and make them match each other.
You can link boost-python to your library as a static library. Then, it will work on almost every computer (even one that doesn't has boost-python installed). See more about that here.
Summary
myMod.so need another version of boost-python, one that compiled with a specific c++ library. Therefore, It would not work with any another version.
In my case I was receiving:
ImportError: dlopen(/Library/Frameworks/Python.framework/Versions/3.8/lib/python3.8/site-packages/xmlsec.cpython-38-darwin.so, 0x0002): symbol not found in flat namespace '_xmlSecDSigNs'
BACKGROUND:
M1 MacBook Pro with Montery
I was working with a python virtualenv (using pyenv) to use an earlier version of python3.8 (3.8.2), while my system had 3.8.10 installed natively.
While I was in the activated 3.8.2 virtualenv I noticed the path in dlopen() was pointing to the package in the native python install NOT the virtualenv install.
SOLUTION:
In my case, I did not need the native 3.8 version at all so I simply removed it and this solved the problem.
I encounter the same problem.
Expected in: flat namespace
Add the linker flag fixes the problem
-lboost_python37
change the dynamic library name to the one installed on the os.
By the way, my os is macOS High Sierra and I use brew to install boost_python3.
Symbol not found means the definition of the declared function or variable was not found. When a header file of a shared object is compiled with your program, linker adds symbols of declared functions and objects to your compiled program. When your program is loaded by the OS's loader, the symbols are resolved so that their definition will be loaded. It is only at this time where if the implementation is missing, loader complains it couldn't find the definition due to may be failing to resolve the actual path to the library or the library itself wasn't compiled with the implementation/source file where the definition of the function or object resides. There is a good article on this on the linux journal http://www.linuxjournal.com/article/6463.
In my case I was just failing to import all the required sources (c++ files) when compiling with Cython.
From the string after "Symbol not found" you can understand which library you are missing.
One of the solutions I found was to uninstall and reinstall it using the no-binary flag, which forces pip to compile the module from source instead of installing from precompiled wheel.
pip install --no-binary :all: <name-of-module>
Found this solution here
Here's what I've learned (osx):
If this is supposed to work (i.e. it works on another computer), you may be experiencing clang/gcc issues. To debug this, use otool -l on the .so file which is raising the error, or a suspect library (in my example it's a boost-python dylib file) and examine the contents. Anything in the /System/ folder is built with clang, and should be installed somewhere else with the gcc compiler. Never delete anything in the /System folder.
.so files are dynamic libraries (so = shared object). On Windows they are called .dll (dynamic-link library). They contain compiled code which contains functions available for usage to any executable which links them.
What is important to notice here is that those .so are not Python files. They were probably compiled from C or C++ code and contain public functions which can be used from Python code (see documentation on Extending Python with C or C++).
On your case, well, you have a corrupt .so. Try reinstalling the affected libraries, or Python, or both.
Problem
I had this same issue when running puma as part of Rails app
LoadError:
dlopen(/Users/alucard/.rbenv/versions/2.7.6/lib/ruby/gems/2.7.0/gems/puma-5.6.4/lib/puma/puma_http11.bundle, 0x0009): symbol not found in flat namespace '_ERR_load_crypto_strings'
/Users/alucard/.rbenv/versions/2.7.6/lib/ruby/gems/2.7.0/gems/puma-5.6.4/lib/puma/puma_http11.bundle
Solution
It was solved just by installing puma gem again gem install puma
I recently updated from Python 2.5 to 2.7 (I tried 2.6 during my hassles) and while everything works fine from the command line or in the Django runserver, mod_wsgi cannot load any module that contains DLLs (pyd) built with MSVC.
For example, if I build my own versions of pycrypto or lxml then I will get the following error only from mod_wsgi:
ImportError at /
DLL load failed: The specified module could not be found.
Even the official PIL binaries will fail to import the _imaging C module in mod_wsgi but that may be another problem.
However, if I use a version of pycrypto built with MinGW from somewhere like http://www.voidspace.org.uk/python/modules.shtml#pycrypto then it will import fine even in mod_wsgi. I don't find this solution satisfactory though since the whole reason I updated Python was to avoid needing to hunt for prebuilt binaries and I can't build them myself because MinGW fails >50% of the time for me.
EDIT2:
I noticed this in Python27/Lib/distutils/msvc9compiler.py on lines 680-705:
try:
# Remove references to the Visual C runtime, so they will
# fall through to the Visual C dependency of Python.exe.
# This way, when installed for a restricted user (e.g.
# runtimes are not in WinSxS folder, but in Python's own
# folder), the runtimes do not need to be in every folder
# with .pyd's.
manifest_f = open(manifest_file)
try:
manifest_buf = manifest_f.read()
finally:
manifest_f.close()
pattern = re.compile(
r"""<assemblyIdentity.*?name=("|')Microsoft\."""\
r"""VC\d{2}\.CRT("|').*?(/>|</assemblyIdentity>)""",
re.DOTALL)
manifest_buf = re.sub(pattern, "", manifest_buf)
pattern = "<dependentAssembly>\s*</dependentAssembly>"
manifest_buf = re.sub(pattern, "", manifest_buf)
manifest_f = open(manifest_file, 'w')
try:
manifest_f.write(manifest_buf)
finally:
manifest_f.close()
except IOError:
pass
This probably explains why everything works from the command line but not in mod_wsgi. Commenting all this out seems to fix the problem but doesn't feel like the proper fix. The question now is where to put msvcr90.dll so that Apache can use it? I notice that Apache's bin folder contains msvcr70.dll and msvcr80.dll but putting 90 in there doesn't work.
I've had a similar issue and eventually found a solution here: download/update your apache server with one from http://www.apachelounge.com/download/.
While I don't know anything about mod_wsgi, I dare to guess that the most probable reason is missing runtime dependencies. You may want to inspect your MSVC-build with Dependency Walker that ships with MSVC (e.g., in MSVC 2005, it's located at \Common7\Tools\Bin\Depends.Exe). It will show you which DLLs are required by a binary.
As another workaround, it should be possible to build your modules with statically linked runtime (see Project Properties -> C/C++ -> Code Generation -> Runtime -- choose "Multithreaded" (not "Multithreaded DLL"); or, if building from command line, make sure /MT is used instead of /MD). However there may be problems if runtime-dependent things (e.g. FILE* objects) cross module boundary.
UPD If you have correct VC redist installed, the reason may be a problem with SxS configuration (i.e. manifest of .pyd itself is wrong or missing, or conflicts with manifest of the application that loads the .pyd). You can use sxstrace utility to see what exactly is going on. See Diagnosing SideBySide failures.
Also, did you try static linking of the runtime? Or, better yet, check what are requirements of your host process.
I was getting this error with zmq. The solution was to include the python27.dll manifest in the libzmq.pyd file (and it'll most likely work for other pyd/dll's). Make sure you use all 64-bit or all 32-bit.
"C:\Program Files (x86)\Windows Kits\8.0\bin\x64\mt.exe" -inputresource:C:\windows\system32\python27.dll;#2 -outputresource:libzmq.pyd;#2
See https://code.google.com/p/pyodbc/issues/detail?id=214