FASSWG msg actions: Difference between revisions
(Imported from Redmine) |
(Removed: Extra empty spaces) |
||
Line 41: | Line 41: | ||
* '''Extrinsic:''' User/GRC knows there's a <code>multiply_const_x::set_k(float)</code> method, but <code>multiply_const</code> doesn't offer a message passing interface. Instead of despairing, we now can do | * '''Extrinsic:''' User/GRC knows there's a <code>multiply_const_x::set_k(float)</code> method, but <code>multiply_const</code> doesn't offer a message passing interface. Instead of despairing, we now can do | ||
<pre> | <pre>boost::function action = boost::bind(&repeat_impl::set_interpolation, this, _1); | ||
multiply_const_0->register_msg_action(pmt::mp("factor"), pmt::to_float, action); | |||
and send pairs (key, value) to the <code>pmt::mp("actions")</code> msg port of the respective block | </pre>and send pairs (key, value) to the <code>pmt::mp("actions")</code> msg port of the respective block | ||
== Discussion == | == Discussion == |
Revision as of 12:11, 20 March 2017
msg_actions PROPOSAL
Problem
Millions of setters in GNU Radio. Some safe to call from another thread whilst general_work is running, others not. No unified scheme of communicating settings.
Approach
Harnessing the power of message passing, and using a bit of old school boost magic to put functions into STL containers, so that they can be automagically called.
Theory of operation:
- Registering a "normal" class methods or functions as action to be executed on message won't work, because "normal" methods don't take
pmt_t
. - Define a combined converter + "normal" method: Input is
pmt_t
, output is void, but in between,pmt_t
gets converted to the native argument type of the method
- Define a converter function; in most case, the PMT standard converters do, e.g.
pmt::to_long
- Find a way to do something like a lambda without breaking with old Boost / C++-pre-0x
- Registry can then be of function(al)s that are
void action(pmt_t)
- Registry has
std::vector
of actions to be taken upon a specific key
Implementation so far
basic_block::register_msg_action(key,conv,action)
method (or rather four methods, including three convenience wrappers):
register_msg_action
(pmt_t
key
, function<ArbitraryType(pmt_t)>
converter
, function<void(ArbitraryType)>
action
)
registers an action by generating a function chain functor, and putting that in a std::vector in a std::map.
basic_block::trigger_actions(key, value)
triggers all actions registered for a single key
basic_block::process_action_msg(pmt::pair)
takes in pairs like pmt::cons(pmt::mp("interpolation"), pmt::from_long(100))
- the
boost::bind
magic that GR already uses when doingregister_msg_handler
gets used, so (hopefully) no new dependencies or boost strangenesses. - the
basic_block
constructor now sets up a message port to which one can send such pairs, and registersprocess_action_msg
as handler
Use Cases
- Intrinsic:
gr::blocks::repeat_impl
has a non-thread-safeset_interpolation(int)
method. It wants to keep that private and expose a message action. It does
boost::function action = boost::bind(&repeat_impl::set_interpolation, this, _1);@ register_msg_action(pmt::mp("interpolation"), pmt::to_long, action);
- Extrinsic: User/GRC knows there's a
multiply_const_x::set_k(float)
method, butmultiply_const
doesn't offer a message passing interface. Instead of despairing, we now can do
boost::function action = boost::bind(&repeat_impl::set_interpolation, this, _1); multiply_const_0->register_msg_action(pmt::mp("factor"), pmt::to_float, action);
and send pairs (key, value) to the pmt::mp("actions")
msg port of the respective block
Discussion
- How to do getters?
- Should GR have one central settings "bus", to which blocks' setters automatically are subscribed?
Yet to do
- drawing pretty pictures and expanding docs beyond the doxygen method docs
- block_gateway to enable python methods to be registered. Should be relatively simple, if it weren't for swig.
- replacing GRC's callback mechanism by something that registers a python function as action handler
- Will require that "Variable" GRC blocks are aware of this
- All blocks "observing" variable get their "actions" port subscribed to the variable's message_publisher
As you can see, not nearly complete, but already useful. I'm not quite sure the naming and architecture take things in the right direction, so I RFC before I PR.