Adds Range and NdRange as supported types in numba_dpex.dpjit. #1148
Conversation
diptorupd
force-pushed
the
feature/range_ndrange_type_support
branch
from
66d6bd3 to
b0a2d1e
Compare
diptorupd
force-pushed
the
feature/range_ndrange_type_support
branch
2 times, most recently
from
69330d8 to
f3b7d26
Compare
| with cgutils.early_exit_if_null(c.builder, stack, dim2_obj): | ||
| c.builder.store(fail_obj, ret_ptr) | ||
|
|
||
| class_obj = c.pyapi.unserialize(c.pyapi.serialize_object(Range)) |
There was a problem hiding this comment.
Sorry, lack of knowledge, could you explain why do we do this?
There was a problem hiding this comment.
First a mea culpa, I have no idea! Purely "monkey see, monkey do" from the Numba extension guide.
But, I dug up what it does. Here is the paper trail:
https://docs.python.org/3/c-api/bytes.html#c.PyBytes_FromStringAndSize
My superficial reading of what is happening here is that for every registered type Numba is maintaining a cache of how to serialize and unserialize an object of that type using the Python pickle module. It is a functionality I believe was designed mainly for on-disk caching of compiled functions and that involves being able to box and unbox argument types. So, we are using that API to call the CPython PyBytes_FromStringAndSize to get the PyObject structure that represents a Range object in this case. Once we have that we are next doing a PyObject_CallFunctionObjArgs to call the constructor for that object.
We might need a better CPython expert like @oleksandr-pavlyk for further explanation.
| """ | ||
| Convert a native range structure to a Range object. | ||
| """ | ||
| ret_ptr = cgutils.alloca_once(c.builder, c.pyapi.pyobj) |
There was a problem hiding this comment.
Am I right that this one creates reference on the created object, that's why we don't have to increase reference manually?
There was a problem hiding this comment.
ret_ptr is the PyObject* into which the res (Range) object to be stored is stored. Since it is a new object that gets returned from Numba to CPython, a incref is not needed.
There was a problem hiding this comment.
So, how does it getting cleaned then?
There was a problem hiding this comment.
When a PyObject is created the ref count should be 1. Once Numba returns the object to CPython, then CPython will track the lifetime of the object, and decref once the object is out of scope and dead. The garbage collector will free it up at that point.
- The numba_dpex.core.kernel_interface.indexers.Range and
NdRange classes can now be used inside a numba_dpex.dpjit
decorated function, as arguments to a dpjit decorated function,
and returned from a dpjit decorated function.
- Defines a datamodel, type class, overloads for the constructors,
box, and unbox functions for both classes.
diptorupd
force-pushed
the
feature/range_ndrange_type_support
branch
from
38ee434 to
af0208d
Compare
- A static variable Range.UNDEFINED_DIMENSION is now used
instead of the magic -1 value to indicate that the dimension
extent is undefined.
- An ndim_obj need not be created when boxing a Range.
diptorupd
force-pushed
the
feature/range_ndrange_type_support
branch
from
af0208d to
7e44902
Compare
https://enccs.github.io/sycl-workshop/expressing-parallelism-nd-range/ |
…support Adds Range and NdRange as supported types in numba_dpex.dpjit. cd5332f
NdRange classes can now be used inside a numba_dpex.dpjit
decorated function, as arguments to a dpjit decorated function,
and returned from a dpjit decorated function.
box, and unbox functions for both classes.
The PR should be merged post 0.21.3 tag