moa/docs/discussion.txt
transistor 708e7706f3 Fixed many of the undocumented tests
Now it's up to 99% tests passing, 15362 are still failing, but many
of those are the unimplemented IN/OUT instructions, which the Genesis
doesn't seem to use
2023-05-22 21:12:05 -07:00

109 lines
4.8 KiB
Plaintext

2021/10/21:
Frontend/Backend Interface
--------------------------
- need a way for the frontend window to be updated with graphics data from a backend device
- also need access to input key and joystick presses; other interfaces can be used for audio, etc
- it would be nice if it was possible to have multiple video output devices in a system
outputting to different windows (ie. one window per video device)
- either the frontend calls a function in the backend to update the window, or the backend
calls an indirect function on the frontend to update the window
- Frontend calling backend:
- Opt 1 - supply a callback and object separately, pass object to callback (no closure)
- older way but should work
- object needs to be Arc in order to share, but doesn't need to be wrapped in a tuple struct
- con: if object is device itself, would still need to use tuple struct wrapper for Addressable
- Opt 2 - supply a trait object with an update method
- frontend would store a Box<dyn Window>
- device would have to make a wrapper: `struct WindowWrapper(Arc<ActualDevice>)` and
impl Window on WindowWrapper, and `struct DeviceWrapper(Arc<ActualDevice>)` and impl
Addressable/Steppable on DeviceWrapper
- pro: in-sync on-demand rendering
- con: lots of complications and indirection
- Opt 3 - supply a common object that devices can update, and that then updates the window
- backend would define a Frame type object which would contain a rendered frame
- system thread would render to the Frame, ui thread would then copy the Frame to window
- con: lots of copying of pixel data
- con: out of sync rendering
- pro: ui can handle any scaling
- Backend calling frontend:
- Opt 4 - host can produce a generic window object that satisfies Window trait
- the device struct which needs the window would have a type parameter for the window
object, and the system thread (step) would call an update function on the generic
window to copy the rendered frame to a ui buffer, which then copies again on ui thread
- window object can only have a buffer, but it can use a native format rather than internal format
- con: out of sync rendering
- Opt 5 - host can produce a Window trait object
- the device would just store the trait object so no need for a type param
- the frontend can't put the native window in the window object so it would need to be a buffer
- con: would need a wrapper if the frontend needs internal access to the common device
- this is
- Opt 3, 4, and 5 would all involve an intermediate buffer, but with Opt 4 or 5, that buffer can be native-compatible
- Opt 3, 4, and 5 are out of sync updating
- Opt 1 and 2 can be in-sync updating if the device object is supplied, but not if it uses an intermediate buffer
- it seems like Opt 4 isn't working because you can't make an existential generic (can't return MiniWindow as W: Window)
- it seems Opt 5 works, but is trying to update the screen waaay too much, and causing the sim to almost never move.
Even with a simple count limit, it seems to pause when it tries updating the screen, probably due to lock contention
- Opt 1 is misleading because you still need a shared object, and it can't be a generic, so it's either a backend-specific
struct or a dyn trait object, so really it's the same as Opt 2, or Opt 3
- it might be possible to have a common data struct that contains 2 frames, one that can only be updated by the sim, and one
that can only be read by in-sync update function, and they are swapped on update (if the writable one isn't locked), so
that the updating is still in sync in the ui thread, but the rendering is happening in another thread
2021/12/07:
Signals, Etc.
-------------
* think more about what kinds of signals are used:
- one setter with one or more passive listeners (bank_register, updated by writing to and used whenever a value is read)
- one or more setters and one listeners (reset, bus_request for the CPUs)
- one one-shot setter (no reset) with one active listener that resets the signal (certain interrupts, including a vsync interrupt)
- what about interrupt controller?
2023/05/22:
Refactoring Execution of m68k
-----------------------------
* would it be possible to construct something that behaves more like the microcode executor using steps in chained form,
with side effects and quirks and flags (and exceptions) handled as such
```
// ADD
executor.start()
.fetch_src(src)
.fetch_dest(dest)
.add()
.write_dest(dest)
.set_flags("XCNVS")
.finish()?;
// BRA
executor.start()
.branch_offset(offset)
.finish()?;
// JMP
executor.start()
.fetch_src(target)
.branch_absolute()
.finish()?;
```