I will try, I tried with the 2020-09-10 (https://buildbot.pypy.org/nightly/py3.6/pypy-c-jit-latest-linux64.tar.bz2) but your changes might have not made it into that one. The errors I get is:

https://github.com/apache/arrow/blob/apache-arrow-1.0.1/cpp/src/arrow/python/common.h#L256

In file included from /arrow/cpp/src/arrow/python/common.cc:18:
/arrow/cpp/src/arrow/python/common.h: In member function ‘arrow::Status arrow::py::PyBytesView::FromBinary(PyObject*, const char*)’:
/arrow/cpp/src/arrow/python/common.h:256:31: error: ‘PyMemoryView_GetContiguous’ was not declared in this scope
       PyObject* contig_view = PyMemoryView_GetContiguous(obj, PyBUF_READ, 'C');
                               ^~~~~~~~~~~~~~~~~~~~~~~~~~
/arrow/cpp/src/arrow/python/common.h:256:31: note: suggested alternative: ‘PyMemoryView_FromMemory’
       PyObject* contig_view = PyMemoryView_GetContiguous(obj, PyBUF_READ, 'C');
                               ^~~~~~~~~~~~~~~~~~~~~~~~~~
                               PyMemoryView_FromMemory

Or I might need to patch something :)

Odd thing is that running it on the latest code I don’t get any errors on date time, only on above. That said, it might simply be that the errors are “below” the above errors

I did a new branch for latest: https://github.com/bivald/pyarrow-on-pypy3/blob/feature/latest-pypy-latest-pyarrow/Dockerfile

On 10 Sep 2020, at 11:30, Matti Picus <matti.picus@gmail.com> wrote:

I implemented the easy part (without allocating and copying non-contiguous data to a new contiguous buffer) of PyMemoryView_GetContiguous, which will make it into the upcoming release. You can try it out from tonight's nightlies
Matti

On 9/9/20 3:08 PM, Niklas B wrote:
Thank you, I managed to get it built using an older version (instructions and whl file for those who are interested is available at https://github.com/bivald/pyarrow-on-pypy3)

I have setup a monthly 50USD recurring donation (under “Enplore”)

On 9 Sep 2020, at 00:06, Matti Picus <matti.picus@gmail.com> wrote:

Have you tried the documented interface https://docs.python.org/3.6/c-api/datetime.html

which is to dispense with all that code and use


#import <datetime.h>

|PyDateTime_IMPORT|


which does the right thing on each implementation (CPython - does the PyCapsule_Import; PyPy - calls _PyDateTime_Import() )


A donation to https://opencollective.com/pypy is always appreciated.

Matti


On 9/9/20 12:12 AM, Niklas B wrote:
Hi,

I’ve been trying to build data science library pyarrow (the arrow library, for parquet files mainly in my case) for PyPy. I’ve gotten it working for pypy2 a few years ago, and is now trying for PyPy3. Overall I get it to build and produce a pyarrow wheel file by following the arrow instructions. So far so good. I expect a massive part of pyarrow not to work, but for my case I really only need `pandas.read_parquet`. However I am stuck trying to figure out how to use the pyppy cpyext for Datetime.

The code I’m trying to build is:

https://github.com/apache/arrow/blob/maint-0.15.x/cpp/src/arrow/python/datetime.cc#L37 (older branch, better luck in building it)

Which is basically:

PyDateTime_CAPI* datetime_api = nullptr;



void InitDatetime() {

PyAcquireGIL lock;

datetime_api =

reinterpret_cast<PyDateTime_CAPI*>(PyCapsule_Import(PyDateTime_CAPSULE_NAME, 0));

if (datetime_api == nullptr) {

Py_FatalError("Could not import datetime C API");

}

}


I’ve tried about a million different ways, but I’m way outside my comfort zone :) I can get it to build by doing:

datetime_api = PyDateTimeAPI;

And also:

datetime_api = reinterpret_cast<PyDateTime_CAPI*>(PyCapsule_Import("datetime", 0));

And:

datetime_api = reinterpret_cast<PyDateTime_CAPI*>(PyCapsule_Import("datetime.datetime_CAPI", 0));

But both of these trigger the fatal error in the code after (“could not import date time C API” or "PyCapsule_Import "datetime" is not valid” or module 'datetime' has no attribute 'datetime_CAPI')

I will be posting reproducible builds once I get them working.

I am more than happy to pay 300USD to anyone (or to PyPy) who can help me getting this to run:

Import pandas
d = pandas.read_parquet(‘file.parq’)

Obviously that’s not enough money to cover things but at least it’s something :) obviously all results and builds will be public

Regards,
Niklas

_______________________________________________
pypy-dev mailing list
pypy-dev@python.org
https://mail.python.org/mailman/listinfo/pypy-dev
_______________________________________________
pypy-dev mailing list
pypy-dev@python.org
https://mail.python.org/mailman/listinfo/pypy-dev