Activity

1. leofang commented on Jun 7, 2021

   

   leofang

   on Jun 7, 2021

   Collaborator

   More actions

   We should list a few items that would be great to do if we plan to break ABI.
2. hpkfft commented on Sep 11, 2026

   

   hpkfft

   on Sep 11, 2026

   More actions

   🌱 Let me share the basic outline of an idea, which is to add a flag DLPACK_FLAG_BITMASK_COMPLEX_COMPONENTS.
   This means that the data is complex, but the bits member of DLDataType is the size of the real component. The imaginary component is understood to be the same size and to follow the real component contiguously in memory.
   The shape array gives the shape of the complex tensor.
   The strides array gives the strides measured in number of real-valued-sized elements.

   Example A:

   >>> a = np.array([[1, 2, 3], [4, 5, 6]], dtype=np.complex128)
   >>> a
   array([[1.+0.j, 2.+0.j, 3.+0.j],
          [4.+0.j, 5.+0.j, 6.+0.j]])
   >>> a.shape
   (2, 3)
   >>> a.strides
   (48, 16)
   

   For DLPack, code=kDLFloat and bits=64 since components are IEEE binary64, shape={2,3}, and strides={6,2}.
   NumPy measures strides in bytes; DLPack measures strides in 8B units because bits is 8B.
   Of course, the flag bit DLPACK_FLAG_BITMASK_COMPLEX_COMPONENTS is set.

   Example B:

   >>> a = np.array([[1, 2, 3, 4, 0], [5, 6, 7, 8, 0]], dtype=np.float64)
   >>> s = a[:, 0:4]  # slice
   >>> b = s.view(np.complex128)
   >>> b
   array([[1.+2.j, 3.+4.j],
          [5.+6.j, 7.+8.j]])
   >>> b.shape
   (2, 2)
   >>> b.strides
   (40, 16)
   

   For DLPack, code=kDLFloat and bits=64 since components are IEEE binary64, shape={2,2}, and strides={5,2}.
   Note that, today, b cannot be dlpacked:

   >>> b.__dlpack__()
   Traceback (most recent call last):
     File "<python-input-38>", line 1, in <module>
       b.__dlpack__()
       ~~~~~~~~~~~~^^
   BufferError: DLPack only supports strides which are a multiple of itemsize.
   

   but with DLPACK_FLAG_BITMASK_COMPLEX_COMPONENTS it can be.
   And it ought to be dlpackable, since alignof(std::complex<double>) == 8.

   Example C:

   >>> c = np.array([1,2,3], dtype=ComplexQuadPrecDType())  # Or whatever it's called
   

   For DLPack, code=kDLFloat and bits=128 since components are IEEE binary128, shape={3}, and strides={2}.
   With DLPACK_FLAG_BITMASK_COMPLEX_COMPONENTS set, uint8_t bits is wide enough.
3. seberg commented on Sep 11, 2026

   

   seberg

   on Sep 11, 2026

   CollaboratorAuthor

   More actions

   Can't say this strikes me as clearly nice. Sacrificing a bit on the dtype enum to mean "complex version" would make more sense to me. But even then, things like striding suddenly being per component feels at least a bit awkward, if maybe not so awkward to be terrible if there is a need.