[FEA]: Research stateful operators for cuda.parallel

[gevtushenko]

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this request and that I agree to the Code of Conduct

Area

Not sure

Is your feature request related to a problem? Please describe.

At the moment, stateful operators lead to a lot of boilerplate code:

cccl/python/cuda_cooperative/tests/test_block_scan.py

Lines 64 to 112 in 87ef1d5

	class BlockPrefixCallbackOp:
	def __init__(self, running_total):
	self.running_total = running_total
	def __call__(self_ptr, block_aggregate):
	old_prefix = self_ptr[0].running_total
	self_ptr[0] = BlockPrefixCallbackOp(old_prefix + block_aggregate)
	return old_prefix
	class BlockPrefixCallbackOpType(types.Type):
	def __init__(self):
	super().__init__(name='BlockPrefixCallbackOp')
	block_prefix_callback_op_type = BlockPrefixCallbackOpType()
	@typeof_impl.register(BlockPrefixCallbackOp)
	def typeof_block_prefix_callback_op(val, c):
	return block_prefix_callback_op_type
	@type_callable(BlockPrefixCallbackOp)
	def type__block_prefix_callback_op(context):
	def typer(running_total):
	if isinstance(running_total, types.Integer):
	return block_prefix_callback_op_type
	return typer
	@register_model(BlockPrefixCallbackOpType)
	class BlockPrefixCallbackOpModel(models.StructModel):
	def __init__(self, dmm, fe_type):
	members = [('running_total', types.int64)]
	models.StructModel.__init__(self, dmm, fe_type, members)
	make_attribute_wrapper(BlockPrefixCallbackOpType,
	'running_total', 'running_total')
	@lower_builtin(BlockPrefixCallbackOp, types.Integer)
	def impl_block_prefix_callback_op(context, builder, sig, args):
	typ = sig.return_type
	[running_total] = args
	state = cgutils.create_struct_proxy(typ)(context, builder)
	state.running_total = running_total
	return state._getvalue()

Describe the solution you'd like

Ideally, I want users to write something along the lines of:

d_input = cp.array([8, 6, 7, 5, 3, 0, 9], dtype=dtype)
cudax.bulk(3, lambda i,d_input=d_input: print(d_input[i]))

That'd involve packing as if capture list was passed to a Numba kernel. Numba is already doing that on kernel launches, we just need a state pointer to be passed to C++ foreign function call instead of cuLaunchKernel

Describe alternatives you've considered

No response

Additional context

No response

[gevtushenko]

@gmarkall for visibility