Cython build issues in combination with generated sources #31

rgommers · 2021-07-04T13:57:02Z

There's multiple places where we can't express dependencies correctly, or we can but they go missing. I filed one issue upstream: mesonbuild/meson#8961.

Extensions that have problems:

`scipy.special._ufuncs`

[1/539] Compiling Cython source scipy/special/_ufuncs.pyx
FAILED: scipy/special/_ufuncs.cpython-39-x86_64-linux-gnu.so.p/scipy/special/_ufuncs.pyx.c 
cython --fast-fail -3 -o scipy/special/_ufuncs.cpython-39-x86_64-linux-gnu.so.p/scipy/special/_ufuncs.pyx.c scipy/special/_ufuncs.pyx

Error compiling Cython file:
------------------------------------------------------------
...
# This file is automatically generated by _generate_pyx.py.
# Do not edit manually!

include "_ufuncs_extra_code_common.pxi"
^

`scipy.linalg._decomp_update`

If we build without expressing the dependency on cython_blas.pxd:

TypeError: C function scipy.linalg.cython_blas.daxpy has wrong signature (expected void (int *, __pyx_t_5scipy_6linalg_11cython_blas_d *, __pyx_t_5scipy_6linalg_11cython_blas_d *, int *, __pyx_t_5scipy_6linalg_11cython_blas_d *, int *), got void (int *, __pyx_t_11cython_blas_d *, __pyx_t_11cython_blas_d *, int *, __pyx_t_11cython_blas_d *, int *))

And if we add it, that doesn't seem supported:

Found ninja-1.10.2 at /home/rgommers/anaconda3/envs/scipy-meson/bin/ninja
Traceback (most recent call last):                                                                  
  File "/home/rgommers/anaconda3/envs/scipy-meson/lib/python3.9/site-packages/mesonbuild/mesonmain.py", line 223, in run
    return options.run_func(options)
  File "/home/rgommers/anaconda3/envs/scipy-meson/lib/python3.9/site-packages/mesonbuild/msetup.py", line 281, in run
    app.generate()
  File "/home/rgommers/anaconda3/envs/scipy-meson/lib/python3.9/site-packages/mesonbuild/msetup.py", line 184, in generate
    self._generate(env)
  File "/home/rgommers/anaconda3/envs/scipy-meson/lib/python3.9/site-packages/mesonbuild/msetup.py", line 246, in _generate
    intr.backend.generate()
  File "/home/rgommers/anaconda3/envs/scipy-meson/lib/python3.9/site-packages/mesonbuild/backend/ninjabackend.py", line 544, in generate
    self.generate_target(t)
  File "/home/rgommers/anaconda3/envs/scipy-meson/lib/python3.9/site-packages/mesonbuild/backend/ninjabackend.py", line 753, in generate_target
    transpiled_sources = self.generate_cython_transpile(target)
  File "/home/rgommers/anaconda3/envs/scipy-meson/lib/python3.9/site-packages/mesonbuild/backend/ninjabackend.py", line 1601, in generate_cython_transpile
    generated_sources[ssrc] = mesonlib.File.from_built_file(gen.subdir, ssrc)
AttributeError: 'CustomTargetIndex' object has no attribute 'subdir'
FAILED: build.ninja

If we add all of _generate_cy as a dependency, then we rebuild cython_blas.pyx and cython_lapack.pyx which isn't right.

`optimize._lsq.givens_elimination`

Not built yet, but also depends on cython_lapack.pxd

The text was updated successfully, but these errors were encountered:

rgommers · 2021-07-04T14:22:09Z

The problem in both cases is missing files in the build/ directory. A workaround is to copy them over manually, then re-run the build. Once the build works once, it will keep working until you do git clean -xdf.

So:

cp scipy/*.pxd build/scipy/

cp scipy/special/*.pxi build/scipy/special/
cp scipy/special/*.pxd build/scipy/special/
cp scipy/special/__init__.py build/scipy/special/

python scipy/linalg/_generate_pyx.py --outdir build
cp scipy/linalg/__init__.py build/scipy/linalg/

rgommers · 2021-07-04T14:50:50Z

@Smit-create maybe one of these was your problem? If so, the workaround should help.

Smit-create · 2021-07-04T17:07:04Z

I had the following error:

$ meson setup build --prefix=$PWD/installdir
The Meson build system
Version: 0.58.1
Source dir: /home/smit/Smitlunagariya/scipy
Build dir: /home/smit/Smitlunagariya/scipy/build
Build type: native build
Project name: SciPy
Project version: 1.7.0.dev0+ababababab
Fortran compiler for the host machine: /home/smit/anaconda3/envs/scipy-dev/bin/x86_64-conda-linux-gnu-

gfortran (gcc 9.3.0 "GNU Fortran (crosstool-NG 1.24.0.133_b0863d8_dirty) 9.3.0")
Fortran linker for the host machine: /home/smit/anaconda3/envs/scipy-dev/bin/x86_64-conda-linux-gnu-gf
ortran ld.bfd 2.36.1
C compiler for the host machine: /home/smit/anaconda3/envs/scipy-dev/bin/x86_64-conda-linux-gnu-cc (gc
c 9.3.0 "x86_64-conda-linux-gnu-cc (crosstool-NG 1.24.0.133_b0863d8_dirty) 9.3.0")
C linker for the host machine: /home/smit/anaconda3/envs/scipy-dev/bin/x86_64-conda-linux-gnu-cc ld.bf
d 2.36.1
C++ compiler for the host machine: /home/smit/anaconda3/envs/scipy-dev/bin/x86_64-conda-linux-gnu-c++
(gcc 9.3.0 "x86_64-conda-linux-gnu-c++ (crosstool-NG 1.24.0.133_b0863d8_dirty) 9.3.0")
C++ linker for the host machine: /home/smit/anaconda3/envs/scipy-dev/bin/x86_64-conda-linux-gnu-c++ ld
.bfd 2.36.1

meson.build:1:0: ERROR: Tried to use unknown language "cython".

When, I removed cython as the 1st language just to check whether it works without it, I get:

$ meson setup build --prefix=$PWD/installdir
The Meson build system
Version: 0.58.1
Source dir: /home/smit/Smitlunagariya/scipy
Build dir: /home/smit/Smitlunagariya/scipy/build
Build type: native build
Project name: SciPy
Project version: 1.7.0.dev0+ababababab
Fortran compiler for the host machine: /home/smit/anaconda3/envs/scipy-dev/bin/x86_64-conda-linux-gnu-
gfortran (gcc 9.3.0 "GNU Fortran (crosstool-NG 1.24.0.133_b0863d8_dirty) 9.3.0")
Fortran linker for the host machine: /home/smit/anaconda3/envs/scipy-dev/bin/x86_64-conda-linux-gnu-gf
ortran ld.bfd 2.36.1

C compiler for the host machine: /home/smit/anaconda3/envs/scipy-dev/bin/x86_64-conda-linux-gnu-cc (gc
c 9.3.0 "x86_64-conda-linux-gnu-cc (crosstool-NG 1.24.0.133_b0863d8_dirty) 9.3.0")
C linker for the host machine: /home/smit/anaconda3/envs/scipy-dev/bin/x86_64-conda-linux-gnu-cc ld.bf
d 2.36.1
C++ compiler for the host machine: /home/smit/anaconda3/envs/scipy-dev/bin/x86_64-conda-linux-gnu-c++
(gcc 9.3.0 "x86_64-conda-linux-gnu-c++ (crosstool-NG 1.24.0.133_b0863d8_dirty) 9.3.0")
C++ linker for the host machine: /home/smit/anaconda3/envs/scipy-dev/bin/x86_64-conda-linux-gnu-c++ ld
.bfd 2.36.1
Host machine cpu family: x86_64
Host machine cpu: x86_64
Program cython found: YES (/home/smit/anaconda3/envs/scipy-dev/bin/cython)
Program pythran found: YES (/home/smit/anaconda3/envs/scipy-dev/bin/pythran)
Program python3 found: YES (/home/smit/anaconda3/envs/scipy-dev/bin/python)
Found pkg-config: /home/smit/anaconda3/envs/scipy-dev/bin/pkg-config (0.29.2)
Dependency python found: YES (pkgconfig)
Library npymath found: YES
Library npyrandom found: YES
Run-time dependency openblas found: YES 0.3.15
Dependency openblas found: YES 0.3.15 (cached)
Program cp found: YES (/bin/cp)
Program cp found: YES (/bin/cp)
Program cp found: YES (/bin/cp)
Run-time dependency threads found: YES
Program cp found: YES (/bin/cp)
Checking for size of "void*" : 8
Dependency threads found: YES unknown (cached)

Checking for function "open_memstream" : NO
Configuring messagestream_config.h using configuration
Compiler for C++ supports arguments -fvisibility=hidden: YES
Program cp found: YES (/bin/cp)
Program cp found: YES (/bin/cp)
Program cp found: YES (/bin/cp)
Compiler for C++ supports arguments -fvisibility=hidden: YES (cached)
Program cp found: YES (/bin/cp)
Dependency openblas found: YES 0.3.15 (cached)
Program cp found: YES (/bin/cp)
Build targets in project: 145

Found ninja-1.8.2 at /usr/bin/ninja

ERROR: No specified compiler can handle file scipy/sparse/csgraph/_flow.pyx

rgommers · 2021-07-04T19:00:15Z

Ah, that does mean you have an old version, it says you're using meson 0.58.1. If you had installed from the master branch, you'd be at 0.58.999 (it is what I have locally, and it is the version number at https://github.com/mesonbuild/meson/blob/master/mesonbuild/coredata.py#L46). You can either use:

python -m pip install git+https://github.com/mesonbuild/meson.git@master

or you can just clone the repo and locally run pip install . in the root of the repo.

rgommers · 2021-10-06T10:27:18Z

Some more failures in gh-76. Cause should be mesonbuild/meson#8961.

rgommers · 2021-12-27T10:17:04Z

This is starting to be urgent to fix. It continues to fail on the macOS CI job, and on Windows it needs -j 128 to work around it.

Here is a simple explanation of the issue: scipy#14847 (comment)

And a potential solution in Cython that I hadn't considered before: scipy#14847 (comment).

rgommers · 2021-12-29T15:58:43Z

Cython finds the module name from the path in a package here: https://github.com/cython/cython/blob/f372c5ab67c73a4c28e68e2aa8bee16a2ec0e4e1/Cython/Compiler/Main.py#L381

full_module_name is passed through a whole chain of functions, because it may be auto-determined by extract_module_name, or for files outside a package it's the filename, or it can be specified by the user in the Extension module (setuptools-specific). Related: exporting module names can be skipped with CYTHON_NO_PYINIT_EXPORT. Where the chain starts:

calling cython .... somefile.pyx on the command line calls main()
main() basically does options, sources = parse_command_line(args) followed by compile(sources, options)
compile() has a full_module_name=None parameter, but this cannot be used from the command line, only from Extension or another Python interface.

Generated C code will end up in a build dir, like at: build/scipy/special/_ufuncs.cpython-39-x86_64-linux-gnu.so.p/scipy/special/_ufuncs.pyx.c

The generated C code will contain the module name, like so:

#define __Pyx_MODULE_NAME "scipy.special._ufuncs"

If something goes wrong related to this issue, that name changes:

#define __Pyx_MODULE_NAME "_ufuncs"

A decent design would be to allow using cython ... --name scipy.special._ufuncs _ufuncs.pyx. This doesn't look too difficult to implement, and would get rid of all the .py file copying. And it's just bringing feature parity between the command line interface and Extension.

matthew-brett · 2021-12-29T16:39:07Z

@rgommers - I'm happy to have a go at that - let me know?

rgommers · 2021-12-29T17:25:50Z

That would be great, thanks @matthew-brett. I had just started looking into it, but I'm really short on time so would be great if you wanted to drive this one.

Here's the full extent of my WIP patch, to save you about 5 minutes:

$ git diff
diff --git a/Cython/Compiler/CmdLine.py b/Cython/Compiler/CmdLine.py
index ffff6a61c..6126ea3e2 100644
--- a/Cython/Compiler/CmdLine.py
+++ b/Cython/Compiler/CmdLine.py
@@ -145,6 +145,10 @@ def create_cython_argparser():
                       dest='compile_time_env', type=str,
                       action=ParseCompileTimeEnvAction,
                       help='Provides compile time env like DEF would do.')
+    parser.add_argument("--name", dest='name', type=str, nargs=1, action='store',
+                      help='Fully qualified module name. If not given, is '
+                           'deduced from the import path if source file is in '
+                           'a package, or equals the filename otherwise.')
     parser.add_argument('sources', nargs='*', default=[])
 
     # TODO: add help

matthew-brett · 2021-12-29T17:27:29Z

Thanks - will have a go at this tomorrow.

eli-schwartz · 2021-12-29T19:50:30Z

That looks neat. Since meson knows what the fully qualified module name is supposed to be -- using py_installation.extension_module('_ufuncs', sources, subdir: 'scipy/special') it knows it will install scipy/special/_ufuncs.cpython-310-x86_64-linux-gnu.so which is also the above mentioned module name -- it should be easy to automatically handle in meson without requiring any options (just like setuptools does by taking advantage of the same logic and using cython as a python module instead of a command line program).

I will look forward to the opportunity to plug that prospective support into meson.

It can be useful to specify the module name for the output file directly, rather than working it out from the enclosing file tree - particularly for out of tree build systems, like Meson. See: rgommers/scipy#31 (comment) for background.

matthew-brett · 2021-12-30T23:35:39Z

Sketch at cython/cython#4548

It can be useful to specify the module name for the output file directly, rather than working it out from the enclosing file tree - particularly for out of tree build systems, like Meson. See: rgommers/scipy#31 (comment) for background.

matthew-brett · 2022-01-10T15:58:37Z

I think this allows us to do something neat like https://github.com/matthew-brett/cyext/blob/main/meson.build .

rgommers · 2022-01-10T16:04:40Z

I think this allows us to do something neat like https://github.com/matthew-brett/cyext/blob/main/meson.build .

That'll be a very temporary thing right? We want support in Meson, and then cython will be invoked by ninja. Or maybe I'm missing what is neat here?

matthew-brett · 2022-01-10T23:55:42Z

Hmm - I hadn't registered how complex the Cython build was. Just for my reference, there are two issues for the source / build tree and Cython:

Finding the correct module name (can be solved by the --module-name argument);
Making sure needed generated ~~.pyx~~ .pxd files are cimportable at build time (not solved by new argument).

Here's a first sketch at respecifying a dependency. I believe it also correctly adds a cimport dependency (on ~~cython_blas.pyx~~ cython_blas.pyx):

matthew-brett@7b33fce

rgommers · 2022-01-11T08:52:06Z

Making sure needed generated .pyx files are cimportable at build time (not solved by new argument).

I don't think I understand this. You cannot cimport a .pyx file normally, this is what .pxd files are for. It looks like Cython has an option to allow cimporting .pyx file, but it's disabled by default and I have never seen it used: https://cython.readthedocs.io/en/latest/src/userguide/source_files_and_compilation.html#Cython.Compiler.Options.cimport_from_pyx

I suspect what you're seeing is that a .pxd file that is still needed in the build directory got copied there too late.

matthew-brett · 2022-01-11T09:54:11Z

Sorry - typing too quickly, and too late - I meant .pxd files - I'll edit in the original comment. In particular, I was working (above) on build for a file that cimports the generated cython_blas (cython_blas.pxd). The point I was reminding myself of was that our current need for copies of the import tree during Cython compilation comes both from the need for correct module name discovery (removed by the cython --module-name arg), and the need for cimport to work when referencing generated .pxd files.

It is somewhat unrelated to that point, but yes, in testing, it turned out that the _bglu_dense extension did not declare its dependence on cython_blas.pxd, so it was possible for the build to fail, if it got out of order.

rgommers · 2022-01-11T10:12:47Z

and the need for cimport to work when referencing generated .pxd files.

Ugh. Yes, you're right. We still need a fix for mesonbuild/meson#8961.

matthew-brett · 2022-01-11T19:20:47Z

What about splitting out the .pyx to .c compilation steps, until we do have a good solution to the problem as defined in mesonbuild/meson#8961 ? It's a bit of work, but probably only a few hours, it wouldn't take long to reverse when we do have a solution.

rgommers · 2022-01-11T20:08:11Z

That should be fine. It won't solve everything, but it will solve this particular issue I suspect.

matthew-brett · 2022-01-13T23:36:10Z

I've done the split in scipy#15407 . I don't get any out-of-order errors now. I think I caught all the dependencies, but please do review.

As discussed in rgommers#31 and Meson issue gh-8961, the current implementation for Cython (and maybe other) transpile steps means that we cannot specify dependencies that apply to the transpile (pyx -> c) step; dependencies only apply to the c building step. This in turn means it is possible to start a pyx -> c transpile without e.g. pxd or `__init__py` files in place, that Cython needs for module name discovery or `cimports`. This PR drops 'cython' as a project compiler, and splits all the Cython transpile steps out into separate steps prior to extension building. We can therefore add the transpile dependencies explicitly. I have tested a `ninja` `-j 1` compile, and several default compiles, without getting any of the Cython transpile errors we were getting before, without very high `-j` numbers.

rgommers · 2022-01-18T06:57:18Z

Since this should now be fixed, I'll close it. I will add the upstream Meson issue to our tracking issue (gh-22), because we do want to get that fixed at some point, so we can simplify our meson.build files.

It can be useful to specify the module name for the output file directly, rather than working it out from the enclosing file tree - particularly for out of tree build systems, like Meson. See background in rgommers/scipy#31 (comment)

It can be useful to specify the module name for the output file directly, rather than working it out from the enclosing file tree - particularly for out of tree build systems, like Meson. See background in rgommers/scipy#31 (comment) Backported-By: H. Vetinari <h.vetinari@gmx.com>

Backport of #4548 It can be useful to specify the module name for the output file directly, rather than working it out from the enclosing file tree - particularly for out of tree build systems, like Meson. See background in rgommers/scipy#31 (comment)

rgommers self-assigned this Jul 6, 2021

rgommers mentioned this issue Jul 7, 2021

BLD: Added interpolate to meson #34

Merged

rgommers mentioned this issue Oct 6, 2021

Failure on cluster._vq build with Meson #76

Closed

rgommers mentioned this issue Oct 16, 2021

ENH: BLD: enable building SciPy with Meson scipy/scipy#14847

Merged

rgommers mentioned this issue Nov 30, 2021

MAINT: Add dev.py for building and installing scipy with meson #94

Merged

rgommers mentioned this issue Dec 30, 2021

Tracking issue for Meson support of submodules #22

Closed

41 tasks

matthew-brett mentioned this issue Dec 30, 2021

MRG: add --module-name argument to cython command cython/cython#4548

Merged

matthew-brett mentioned this issue Jan 13, 2022

BLD: meson: split pyx->c and Python extension build scipy/scipy#15407

Merged

rgommers closed this as completed Jan 18, 2022

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Cython build issues in combination with generated sources #31

Cython build issues in combination with generated sources #31

rgommers commented Jul 4, 2021 •

edited

Loading

rgommers commented Jul 4, 2021 •

edited

Loading

rgommers commented Jul 4, 2021

Smit-create commented Jul 4, 2021

rgommers commented Jul 4, 2021

rgommers commented Oct 6, 2021

rgommers commented Dec 27, 2021

rgommers commented Dec 29, 2021

matthew-brett commented Dec 29, 2021

rgommers commented Dec 29, 2021

matthew-brett commented Dec 29, 2021 via email

eli-schwartz commented Dec 29, 2021

matthew-brett commented Dec 30, 2021

matthew-brett commented Jan 10, 2022

rgommers commented Jan 10, 2022

matthew-brett commented Jan 10, 2022 •

edited

Loading

rgommers commented Jan 11, 2022

matthew-brett commented Jan 11, 2022

rgommers commented Jan 11, 2022

matthew-brett commented Jan 11, 2022

rgommers commented Jan 11, 2022

matthew-brett commented Jan 13, 2022

rgommers commented Jan 18, 2022

Cython build issues in combination with generated sources #31

Cython build issues in combination with generated sources #31

Comments

rgommers commented Jul 4, 2021 • edited Loading

scipy.special._ufuncs

scipy.linalg._decomp_update

optimize._lsq.givens_elimination

rgommers commented Jul 4, 2021 • edited Loading

rgommers commented Jul 4, 2021

Smit-create commented Jul 4, 2021

rgommers commented Jul 4, 2021

rgommers commented Oct 6, 2021

rgommers commented Dec 27, 2021

rgommers commented Dec 29, 2021

matthew-brett commented Dec 29, 2021

rgommers commented Dec 29, 2021

matthew-brett commented Dec 29, 2021 via email

eli-schwartz commented Dec 29, 2021

matthew-brett commented Dec 30, 2021

matthew-brett commented Jan 10, 2022

rgommers commented Jan 10, 2022

matthew-brett commented Jan 10, 2022 • edited Loading

rgommers commented Jan 11, 2022

matthew-brett commented Jan 11, 2022

rgommers commented Jan 11, 2022

matthew-brett commented Jan 11, 2022

rgommers commented Jan 11, 2022

matthew-brett commented Jan 13, 2022

rgommers commented Jan 18, 2022

rgommers commented Jul 4, 2021 •

edited

Loading

`scipy.special._ufuncs`

`scipy.linalg._decomp_update`

`optimize._lsq.givens_elimination`

rgommers commented Jul 4, 2021 •

edited

Loading

matthew-brett commented Jan 10, 2022 •

edited

Loading