[Numpy-discussion] ufunc for user-defined type

[Neal Becker]

Travis E. Oliphant wrote:

> Neal Becker wrote:
>> Now that I have my user-defined type, I want to add some funcs.
>> According to the numpy book, I need to use:
>>
>> PyUFunc_RegisterLoopForType
>>
>> The book says I first need a ufunc.  The only way I see to create one is
>>
>> PyUFunc_FromFuncAndData.
>>
>> Is the the correct procedure?
>>   
> You'll have to do some digging because the API allowing low-level loops
> for user-defined types is not well tested or vetted.   Also, don't be
> surprised if you uncover some bugs.
> 
> The concept is that you create the UFunc with the "built-in" types using
> PyUFunc_FromFuncAndData and then you add loops for the user-defined type
> using RegisterLoopForType.
> 
> So, you have the basic procedure down.
> 
> -Travis

It looks like this works.  Basically, the only thing used from 
the ufunc obj is: name, nin, nout, nargs, and ntypes.

// Make a generic new ufunc obj
static PyUFuncObject * new_ufunc (char* name) {
  PyUFuncObject * u = PyObject_New (PyUFuncObject, &PyUFunc_Type);
  u->nin = 0;
  u->nout = 0;
  u->nargs = 0;
  u->identity = 0;
  u->functions = 0;
  u->data = 0;
  u->types = 0;
  u->ntypes = 0;
  u->check_return = 0;
  u->ptr = NULL;
  u->obj = NULL;
  u->userloops=NULL;

  u->name = name;

  u->doc = "NULL";
  return u;
}

//fill in required fields
  PyUFuncObject * u = new_ufunc ("magsqr");
  u->nin = 1;
  u->nout = 1;
  u->nargs = 2;
  u->ntypes = 1;

  {
    int types[] = {d1->type_num, NPY_INT32};
    register_unary_loop_func (u, d1->type_num, types, mag_sqr_impl<cmplx_int32_t>());
  }

where 'register_unary_loop_func' will basically call:
  PyUFunc_RegisterLoopForType (u, typenum, &loop1d_unary<in_t,out_t,unary_func>, types, 0);