Representation of size-0 arrays in __cuda_array_interface__

I have a design question about __cuda_array_interface__. How (if at all) should size-0 arrays be represented? In cupy, the data field currently uses 0 as the int_ptr.

In [7]: import cupy

In [8]: arr = cupy.random.random(0)

In [9]: arr.__cuda_array_interface__
Out[9]:
{'shape': (0,),
 'typestr': '<f8',
 'descr': [('', '<f8')],
 'data': (0, False),
 'version': 0}

Compare with numba, which uses None

In [10]: import numba.cuda

In [11]: numba.cuda.device_array(0).__cuda_array_interface__
Out[11]:
{'shape': (0,),
 'strides': (8,),
 'data': (None, False),
 'typestr': '<f8',
 'version': 0}

Currently, cupy doesn’t round-trip a size-0 array

In [12]: class Dummy:
    ...:     def __init__(self, descr):
    ...:         self.__cuda_array_interface__ = descr
    ...:

In [13]: obj = Dummy(arr.__cuda_array_interface__)

In [14]: cupy.asarray(obj)

---------------------------------------------------------------------------
CUDARuntimeError                          Traceback (most recent call last)
<ipython-input-14-306044f096f0> in <module>
----> 1 cupy.asarray(obj)

~/miniconda3/lib/python3.6/site-packages/cupy/creation/from_data.py in asarray(a, dtype, order)
     84     """
     85     if not isinstance(a, ndarray) and hasattr(a, '__cuda_array_interface__'):
---> 86         return _convert_object_with_cuda_array_interface(a)
     87     return core.array(a, dtype, False, order)
     88

~/miniconda3/lib/python3.6/site-packages/cupy/creation/from_data.py in _convert_object_with_cuda_array_interface(a)
     56         strides = None
     57         nbytes = numpy.prod(shape) * dtype.itemsize
---> 58     mem = memory.UnownedMemory(desc['data'][0], nbytes, a)
     59     memptr = memory.MemoryPointer(mem, 0)
     60     return ndarray(shape, dtype=dtype, memptr=memptr, strides=strides)

cupy/cuda/memory.pyx in cupy.cuda.memory.UnownedMemory.__init__()

cupy/cuda/runtime.pyx in cupy.cuda.runtime.pointerGetAttributes()

cupy/cuda/runtime.pyx in cupy.cuda.runtime.check_status()

CUDARuntimeError: cudaErrorInvalidValue: invalid argument

Info:

In [6]:  cupy.show_config()
CuPy Version          : 6.0.0b3
CUDA Root             : /usr/local/cuda-10.0
CUDA Build Version    : 9020
CUDA Driver Version   : 9020
CUDA Runtime Version  : 9020
cuDNN Build Version   : 7402
cuDNN Version         : 7402
NCCL Build Version    : 2307
NCCL Runtime Version  : 2307

cc @sklam @seibert for the numba side of things.

[leofang]

This issue is resolved by #2491. The new v2 protocol specifies that the pointer to 0-size arrays should be 0.

[leofang]

Just trying to follow up: as I raised in numba/numba#4175, Numba’s way of representing a 0-size array with __cuda_array_interface__['data'][0] = None is not complying with the standard they proposed, which requires an int. This should be considered a bug on the Numba side, not CuPy.

cc: @seibert