hpyproject / numpy-hpy Goto Github PK

What is the status of the hpy, graal-team/hpy and graal-team/hpy-0.0.5 branches here?

hpyproject/hpy HEAD now has more support for the interfaces needed here, and NumPy HEAD has progressed. It would be nice to update this fork and add a blurb to the README so others can try it out.

There is a missing tp_traverse on the New_PyArrayDescr_spec (which derives from the New_PyArrayDescr_spec_prototype), since PyArray_Descr has HPyFields. PyPy requires these, and will not let HPyField_Store work if the tp_traverse is missing. Since HPyField_Store has a void return value, PyPy will exit. I will submit a PR to add the missing tp_traverse and also raise a HPy issue to change the return value so an error can properly be reported.

When building, I get lots of warnings about using forbidden API. I think this is because the code builds with HPY_ABI_UNIVERSAL. Shouldn't it use HPY_ABI_HYBRID until the migration is complete?

array_iter_impl -> HPySeqIter_New -> look up __iter__ and call it -> array_iter_impl.

You can hit this by doing np.range(10).__iter__(). This was added in 3a7a0d2.

For some reason calling src_cycle = cycle(self.pool) in setting up the random.bit_generator is hitting this in PyPy but not in CPython. Maybe cycle is implemented slightly differently in CPython, and avoids calling self.pool.__iter__().

I currently try to get rid of HPy_SetType in branch graal-team/hpy . We used it (as a replacement for Py_SET_TYPE) in function dtypemeta_wrap_legacy_descriptor. I was able to refactor the creation of (_hpy)_builtin_descrs (see https://github.com/hpyproject/numpy-hpy/blob/graal-team/hpy/numpy/core/src/multiarray/arraytypes.c.src#L4842). I did so by first creating the new dtype and then instantiating the singleton. However, it is not that simple for Numpy API (H)PyArray_RegisterDataType because the descriptor to register is provided as argument (e.g. sometimes created with (H)PyArray_DescrNew).

My question is: are (H)PyArray_DescrNew and (H)PyArray_RegisterDataType legacy APIs that have some replacement and could we just drop them in Numpy/HPy?
Or is there some other way to solve this?