unsupported operand type(s) for * for dataclass that does not define array_context attribute in compiled_lsrk45

[majosm]

@anderson2981 ran into an error in the wall-enabled prediction driver when using compiled_lsrk45:

  File "emirge/mirgecom/mirgecom/steppers.py", line 425, in advance_state
    _advance_state_stepper_func(
  File "emirge/mirgecom/mirgecom/steppers.py", line 159, in _advance_state_stepper_func
    state = timestepper(state=state, t=t, dt=dt, rhs=compiled_rhs)
  File "prediction.py", line 1471, in _compiled_stepper_wrapper
    return compiled_lsrk45_step(actx, state, t, dt, rhs)
  File "emirge/grudge/grudge/shortcuts.py", line 58, in compiled_lsrk45_step
    y, residual = update(y, a, b, h, rhs_val)
  File "emirge/arraycontext/arraycontext/impl/pytato/compile.py", line 336, in __call__
    output_template = self.f(
  File "emirge/grudge/grudge/shortcuts.py", line 38, in _lsrk45_update
    residual = a*residual + h*rhs_val
TypeError: unsupported operand type(s) for *: 'Placeholder' and 'WallVars'

Here WallVars is a dataclass defined in a similar manner to ConservedVars, with (at least) one distinction: it does not define an array_context method (and does not set _cls_has_array_context_attr=True) (code).

The above error doesn't occur with the regular rk4_step, and it also goes away if I define an array_context method in WallVars. Is this a bug, or is there a reason that it needs to define an array_context method in order to work?

[kaushikcfd]

For this behavior of broadcasting ArrayContext.array_types across a container, _bcast_actx_array_type needs to be set. The behavior observed here agrees with the docs (see the default values of _bcast_actx_array_type).

For this feature, we require array_context to be an attribute to figure out which array types could be legally broadcasted.