symengine/symengine

Fast symbolic manipulation library, written in C++

Richard Otis

@richardotis

(calphadpy3) [rotis@kaufman pycalphad]$ python -c 'import symengine ; import pycalphad'
Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "/home/rotis/git/pycalphad/pycalphad/__init__.py", line 37, in <module>
    from pycalphad.core.calculate import calculate
  File "/home/rotis/git/pycalphad/pycalphad/core/calculate.py", line 7, in <module>
    from pycalphad.codegen.callables import build_phase_records
  File "/home/rotis/git/pycalphad/pycalphad/codegen/callables.py", line 5, in <module>
    from pycalphad.core.phase_rec import PhaseRecord
ImportError: /home/rotis/git/pycalphad/pycalphad/core/phase_rec.cpython-37m-x86_64-linux-gnu.so: undefined symbol: _ZN9SymEngine17LLVMDoubleVisitor5loadsERKNSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEE

Isuru Fernando

@isuruf

what does ldd /path/to/python/lib/python3.7/site-packages/symengine/lib/symengine_wrapper.cpython-37m-x86_64-linux-gnu.so give you?

Richard Otis

@richardotis

(calphadpy3) [rotis@kaufman pycalphad]$ ldd /home/rotis/anaconda3/envs/calphadpy3/lib/python3.7/site-packages/symengine/lib/symengine_wrapper.cpython-37m-x86_64-linux-gnu.so
        linux-vdso.so.1 =>  (0x00007ffc20ffa000)
        libsymengine.so.0.4 => /home/rotis/anaconda3/envs/calphadpy3/lib/python3.7/site-packages/symengine/lib/../../../../libsymengine.so.0.4 (0x00007f1867949000)
        libflint.so.13 => /home/rotis/anaconda3/envs/calphadpy3/lib/python3.7/site-packages/symengine/lib/../../../../libflint.so.13 (0x00007f18672fd000)
        libmpc.so.3 => /home/rotis/anaconda3/envs/calphadpy3/lib/python3.7/site-packages/symengine/lib/../../../../libmpc.so.3 (0x00007f18672e2000)
        libmpfr.so.6 => /home/rotis/anaconda3/envs/calphadpy3/lib/python3.7/site-packages/symengine/lib/../../../../libmpfr.so.6 (0x00007f1867263000)
        libstdc++.so.6 => /home/rotis/anaconda3/envs/calphadpy3/lib/python3.7/site-packages/symengine/lib/../../../../libstdc++.so.6 (0x00007f1867121000)
        libm.so.6 => /lib64/libm.so.6 (0x00007f1866df6000)
        libgcc_s.so.1 => /home/rotis/anaconda3/envs/calphadpy3/lib/python3.7/site-packages/symengine/lib/../../../../libgcc_s.so.1 (0x00007f1866de1000)
        libc.so.6 => /lib64/libc.so.6 (0x00007f1866a1d000)
        libgmp.so.10 => /home/rotis/anaconda3/envs/calphadpy3/lib/python3.7/site-packages/symengine/lib/../../../.././libgmp.so.10 (0x00007f1866789000)
        libdl.so.2 => /lib64/libdl.so.2 (0x00007f1866584000)
        libpthread.so.0 => /lib64/libpthread.so.0 (0x00007f1866368000)
        /lib64/ld-linux-x86-64.so.2 (0x00007f18696ba000)
        libntl.so.33 => /home/rotis/anaconda3/envs/calphadpy3/lib/python3.7/site-packages/symengine/lib/../../../.././libntl.so.33 (0x00007f1865f3a000)
        libgf2x.so.1 => /home/rotis/anaconda3/envs/calphadpy3/lib/python3.7/site-packages/symengine/lib/../../../../././libgf2x.so.1 (0x00007f1865d2d000)

Isuru Fernando

@isuruf

What's the output of,

nm -g /home/rotis/anaconda3/envs/calphadpy3/lib/python3.7/site-packages/symengine/lib/../../../../libsymengine.so.0.4 | grep _ZN9SymEngine17LLVMDoubleVisitor5loadsERKNSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEE

Richard Otis

@richardotis

(calphadpy3) [rotis@kaufman pycalphad]$ nm -g /home/rotis/anaconda3/envs/calphadpy3/lib/python3.7/site-packages/symengine/lib/../../../../libsymengine.so.0.4 | grep _ZN9SymEngine17LLVMDoubleVisitor5loadsERKNSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEE00000000003e0e80 T _ZN9SymEngine17LLVMDoubleVisitor5loadsERKNSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEE

00000000003e0e80 T _ZN9SymEngine17LLVMDoubleVisitor5loadsERKNSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEE

Isuru Fernando

@isuruf

Hmm, you'll have to link to libsymengine.so. I see no way around that

Richard Otis

@richardotis

Would adding it as a libraries argument in the Extension do the trick

Isuru Fernando

@isuruf

Yes, libraries = ['symengine'] should work

Richard Otis

@richardotis

Looks like that did the trick! Thanks again

Brandon Bocklund

@bocklund

Just wanted to check in again - we've transitioned all the codegen of pycalphad from SymPy+Cython to symengine. Our test suite has gone from taking almost 30 minutes on my local dev machine to 3m 30s! 30s of that is just calling sympify, which we hope to remove in the future (I think pickle support is a blocker). The codegen and differentiation are about 10x faster and it looks like calling our objective is somewhere around 2x faster. The codegen boost is really impactful for our users and startup times will go from 1-5 minutes to just a few seconds. Thank you for this effort!

Isuru Fernando

@isuruf

Nice. We would like to have serialization, but unfortunately none of us have experience in doing that in C++. symengine/symengine#1394

Brandon Bocklund

@bocklund

Yeah, we saw that. I don't have any experience either, unfortunately.

Any idea what some use cases of symengine would be where you need guarantees of compatibility as mentioned in the last comment?

Their 2 main downsides are that 1) they serialize into some custom format, which is not much usable as a data exchange format with other libraries/programs 2) they offer few guarantees of forward compatibility (e.g., if you serialize an object with Boost 1.66 it's not guaranteed to be deserializable with Boost 1.65).

Isuru Fernando

@isuruf

@bocklund, a use case is where you pickle an object and I unpickle it

Brandon Bocklund

@bocklund

IIRC pickles are guaranteed to be Python version independent, but not platform independent

Isuru Fernando

@isuruf

pickles are supposed to be backward compatible

and independent of platform

if you pickle in python 2 on osx, you can unpickle in python 3 on linux

you can't pickle in python3 on osx and unpickle in python 2 on linux

so, pickle is not forward compatible, but if 2 people used the same python minor version, they'll be compatible. If we use boost, there's an additional constraint on boost version

Richard Otis

@richardotis

One use case where you're less sensitive to environment compatibility is using dask or multiprocessing, where serialization is basically just IPC

Though I very much understand that you don't want to implement serialization with a footgun that users might not notice

Richard Otis

@richardotis

Are the Windows symengine packages on conda-forge built with LLVM enabled?

Isuru Fernando

@isuruf

Yes. And wheels too

Richard Otis

@richardotis

And it's built with MSVC?

Richard Otis

@richardotis

I am trying to link to symengine.lib using the g++ that comes with the mingw-w64 package. It finds it, but I'm getting all undefined references during the compile step.

Isuru Fernando

@isuruf

Yes, they are built with MSVC

Richard Otis

@richardotis

Okay, looks like I'll need to try converting my toolchain over. Thanks.

Isuru Fernando

@isuruf

To use with mingw-w64, you'll have to use the C interface instead of C++

Richard Otis

@richardotis

Do I have any control over that from the Cython side?

Isuru Fernando

@isuruf

Sure. You can use the functions in cwrapper.h instead of the C++ classes

Richard Otis

@richardotis

That would mean recreating and vendoring the portion of symengine_wrapper I'm using?

Isuru Fernando

@isuruf

Yes, mostly the LLVMDoubleVisitor related functions

Isuru Fernando

@isuruf

@richardotis, @bocklund I looked at the pycalphad PR. One thing I saw was that you were building lots of LLVMDoubleVisitors. Are they called with the same input? If so, you might be missing out on cse optimizations.

Richard Otis

@richardotis

They usually are, but we can't guarantee that due to the structure of our optimization problem.

Richard Otis

@richardotis

We do some grouping of our constraint functions, which does get the CSE benefit

We are SymEngine and LLVM newbies though, if there are any new superpowers we may have gained, please let us know. ;)

Isuru Fernando

@isuruf

Can you post the symbolic expressions for something like mass_hess?

Richard Otis

@richardotis

We could probably pickle the sympy versions of the objects

Isuru Fernando

@isuruf

yeah, that works

python >3.4

Richard Otis

@richardotis

@bocklund probably has a good one for a complex system. With the latest update we're hoping to push beyond the limits of what the commercial codes can do in terms of function complexity

Isuru Fernando

@isuruf

For something like mass_hess, it might be beneficial to have a loop with the data in column major format to get SIMD benefits

Brandon Bocklund

@bocklund

Pickling and unpickling the LLVMDouble does not correctly reset the LLVMDoubleVisitor that is supposed to be in llvm_double[0].

Reproducing the bug

from symengine import symbols
x, y, z = symbols('x y z')
ex = x**2 + y**2 + z**2
from symengine import lambdify
l = lambdify([x, y, z], [ex], backend='llvm')
import pickle
print('pickle.loads(pickle.dumps(l)) == l  :: ', pickle.loads(pickle.dumps(l)) == l)
print(l.__reduce__())
print(pickle.loads(pickle.dumps(l)).__reduce__())

gives

pickle.loads(pickle.dumps(l)) == l  ::  False
(<built-in function llvm_loading_func>, (3, 1, [()], True, 1, 'C', [0, 1], <class 'numpy.float64'>, b"\xcf\xfa\xed\xfe\x07\x00\x00\x01\x03\x00\x00\x00\x01\x00\x00\x00\x04\x00\x00\x00`\x01\x00\x00\x00 \x00\x00\x00\x00\x00\x00\x19\x00\x00\x00\xe8\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00`\x00\x00\x00\x00\x00\x00\x00\x80\x01\x00\x00\x00\x00\x00\x00`\x00\x00\x00\x00\x00\x00\x00\x07\x00\x00\x00\x07\x00\x00\x00\x02\x00\x00\x00\x00\x00\x00\x00__text\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00__TEXT\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00'\x00\x00\x00\x00\x00\x00\x00\x80\x01\x00\x00\x04\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x04\x00\x80\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00__eh_frame\x00\x00\x00\x00\x00\x00__TEXT\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00(\x00\x00\x00\x00\x00\x00\x008\x00\x00\x00\x00\x00\x00\x00\xa8\x01\x00\x00\x03\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x0b\x00\x00h\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00$\x00\x00\x00\x10\x00\x00\x00\x00\x0e\n\x00\x00\x00\x00\x00\x02\x00\x00\x00\x18\x00\x00\x00\xe0\x01\x00\x00\x01\x00\x00\x00\xf0\x01\x00\x00\x10\x00\x00\x00\x0b\x00\x00\x00P\x00\x00\x00\x00\x00\x00\x00\x01\x00\x00\x00\x01\x00\x00\x00\x00\x00\x00\x00\x01\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\xf2\x0f\x10\x07\xf2\x0f\x10O\x08\xf2\x0f\x10W\x10\xf2\x0fY\xd2\xf2\x0fY\xc0\xf2\x0fX\xc2\xf2\x0fY\xc9\xf2\x0fX\xc8\xf2\x0f\x11\x0e\xc3\x00\x14\x00\x00\x00\x00\x00\x00\x00\x01zR\x00\x01x\x10\x01\x10\x0c\x07\x08\x90\x01\x00\x00\x1c\x00\x00\x00\x1c\x00\x00\x00\xb8\xff\xff\xff\xff\xff\xff\xff'\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01\x00\x00\x00\x0e\x01\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00___unnamed_1\x00\x00\x00"))
(<built-in function llvm_loading_func>, (3, 1, [()], True, 1, 'C', [0, 1], <class 'numpy.float64'>, b''))

You can see that the pickled/unpickled version gives an empty byte string for the dumped LLVMDoubleVisitor bytes. @richardotis and I were tracking this down and the issue looks to be that the membuffer is not correctly reset in LLVMDoubleVisitor.loads here when called in self.llvm_double[0].loads as part of the LLVMDouble._load method here. In that case, the root fix would be to set the membuffer in LLVMDoubleVisitor.loads in the same way as LLVMDoubleVisitor.__init__.

A test for this might be to do pickle.loads(pickle.dumps(pickle.loads(pickle.dumps(l)))), which currently will kill the kernel with a LLVM ERROR: The file was not recognized as a valid object file

Isuru Fernando

@isuruf

Want to send a PR?

Brandon Bocklund

@bocklund

I'm not sure what the correct fix for the LLVMDoubleVisitor.loads method looks like

I'm not very fluent in C++

Isuru Fernando

@isuruf

copying the string to membuffer

membuffer = s at the beginning of loads should fix it

Where communities thrive

People

Repo info

Activity

Reproducing the bug