This commit does the following:
* factor out class AccessTransientLEDMode to an individual file
* factor out class LEDModeInterface to individual files
* fix the file header comment in LEDMode.h
Signed-off-by: Florian Fleissner <florian.fleissner@inpartik.de>
This PR introduces the concept of dynamic LED modes. Those are LED modes whose class instances
have a restricted lifetime that lasts only as long as a LED mode is active. By this means
it is possible to support a greater amount of LED modes - especially RAM-hungry ones - in the same firmware build. The amount of RAM used to store dynamic LED modes is now bounded
by the maximum size (`sizeof(...)`) of the largest dynamic LED mode.
Old-style LED modes are furtheron called _static_ in the terminology of this PR. They are still supported and blend in nicely with the newly introduced dynamic LED modes.
All changes are entirely backward compatible. No user sketches or existing user plugins require changes.
The greatest benefit of this change is that it drastically reduces the consumption of RAM
when multiple complex LED modes are used. Currently the most complex stock LED mode is
the wavepool effect. Its plugin requires around 140 bytes of RAM that are statically allocated and cannot be shared with any other features.
With this change it becomes possible to have a large number of such resource-hungry LED modes in parallel without a significant gain in RAM consumption.
For the stock firmware this change means a small (~30 byte) growth in terms of PROGMEM. On the other hand it reduces the amount of statically consumed RAM by ~90 bytes. As the current atmel architectures come with around ten times as much PROGMEM as RAM, this means a great improvement as RAM is the more critical resource.
If the wavepool effect, a especially RAM-hungry LED mode is added to the stock firmware,
the saving of RAM increases to 160 bytes which is almost 8% of RAM of the Keyboardio Model01.
A new interface class `LEDModeInterface` was introduced that those plugins
that export dynamic LED modes inherit from. To remain backward compatible, the `LEDMode` class that all pre-existing LED mode plugins inherited from is also derived from `LEDModeInterface`.
The new interface class currently lives in header `LEDMode.h` (see information about this new header below). This is because `LEDMode` and `LEDModeInterface` will
always be used together by dynamic LED modes. Thus, an extra header for `LEDModeInterface` would only mean extra include work for users writing plugins.
Those plugins that export dynamic LED modes must furtheron provide a exported type `DynamicLEDMode`.
This can either be done by defining a nested class of that name or by typedef-ing a class that is defined at global scope to `DynamicLEDMode`. See the modified stock LED modes for examples.
Some of those plugins that export dynamic led modes require access to their particular
dynamic LED mode. By adding the macro `ACCESS_THIS_LED_MODE` to the plugin class definition,
additional data and methods (an integer `led_mode_id_) are synthesized, that enable the plugin class to gain access to their particular dynamic LED mode instance (as long as it is active).
The synthesized integer member `led_mode_id_` can be used to query if the currently active LED mode is the oned handled by the plugin class instance (note that there might be more than one plugin instance of the same class and thus also several dynamic LED modes, see e.g. the solid color LED mode).
A query in the plugin's event handler e.g. looks as follows.
```cpp
if (::LEDControl.get_mode_index() != led_mode_id_)
return EventHandlerResult::OK;
```
All stock LED modes have been adapted to export dynamic LED modes (if possible).
This does not apply to all of them as for some the transition would have provided no gain.
It would even have meant a deterioration of resource consumption for those few pre-existing stock LED mode plugins that hardly have no (static) data-members at all (like e.g. `LEDOff`).
To reduce the amount of compile unit and header interdependencies, the class `LEDMode` has been moved to a header/implementation file of its own.
The `LEDControl` class now does not have a static array anymore to store LED mode pointers.
Instead, it delegates the core LED mode handling to a newly introduced `LEDModeManager` class
that lives in internal namespace. The `LEDModeManager` class is there to restrict access
to LED modes but also to wrap up core LED mode handling. If this functionality would
have been added to class `LEDControl`, far too much of the internals of LED mode handling would have been exposed to users through header `LEDControl.h`.
The new internal header `array_like_storage.h` contains a template class that is used to generate
array-like storages. Here array-like means that the contained pieces of information
are stored contiguously in memory in the same way as they would be when defining
language intrinsic (C-style) arrays. This type of storage is especially useful to generate array-like data struktures
in PROGMEM at compile time based on a list of global objects or POD data. By casting the array-like storage's address
to the content's pointer type, an array-like indexed access is possible.
In this PR an array-like data structure is used to generate a PROGMEM
array of LED mode factories. Array-like data structures could also become useful in other places and for future applications.
The most complex part of the implementation of the new LED mode handling is wrapped up in
`LEDModeManager.h` and `LEDModeManager.cpp` to hide it from users' site.
There, recursive template classes are used to setup an array-like data structure of `LEDModeFactory` instances in PROGMEM. Each of the stored `LEDModeFactory`s are associated with one LED mode-plugin as specified in the sketch. The template mechanism filters out any other plugins unrelated to LED modes. `LEDModeFactory`s thereby handle both static and dynamic LED modes.
Class `LEDModeManager` provides access to the LED mode factories and LED modes in general. It exports methods to query the number of LED modes and to activate a LED mode by its mode-ID. Most of this is only available to `LEDControl` that represents the actual user interface.
When a dynamic LED mode is activated, a dedicated `LEDModeFactory` generates an instance of the dynamic LED mode class in the
LED mode buffer. This buffer is shared by all dynamic LED modes. Its size has been determined at compile time by examining all exported dynamic LED mode types and determining the maximum necessary amount of RAM to store any of those.
All LED mode handling related data structures are generated at compile time, based on
the list of plugins that are passed to `KALEIDOSCOPE_INIT_PLUGINS(...)`. This function macro invokes a new function macro `_INIT_LED_MODE_MANAGER` from `LEDModeManager.h` that handles the LED mode related stuff.
Signed-off-by: Florian Fleissner <florian.fleissner@inpartik.de>
A traits class that is used to check if a class implements a
method with a given signature was already used by
the eventhandler signature check. It has now been moved to
its own traits header (macro name is DEFINE_HAS_METHOD_TRAITS).
Signed-off-by: Florian Fleissner <florian.fleissner@inpartik.de>
This is a short-term solution for the problem of qukeys rollover from
mod-flagged keys. When flushing its queue, Qukeys stashes a copy of the
current keyboard HID report, and bases the new (usually mid-cycle) report on
the previous one, because it's complete, unlike the partial current report. This
prevents bugs where it would cause undesired repeats of keys scanned later in
the cycle, but results in a bypassing of the system in the HIDAdaptor that masks
modifier flags when rolling over to a key without those flags.
By first storing a copy of the previous key report's modifiers byte and then
restoring it after the report for a released qukey is sent, we remove any
modifier flags it added to that report, so the next key in the queue (if any)
won't base it's report on those modifiers (which would normally be masked by the
HIDAdaptor).
This is a temporary fix. The real fix will probably involve changes to both
KeyboardioHID and the HIDAdaptor, and will allow Qukeys to stop accessing the
HID reports, which really ought to be private member variables of the Keyboard
class.
Fixes#619.
Signed-off-by: Michael Richters <gedankenexperimenter@gmail.com>
Recently a proposed change to the firmware required a change to one of
the event handlers'/hooks' signatures.
Up to now, we only supported a single overload/implementation of a
event handler of a specific type. This means that changing handler signatures
was only possible by adding a handler method with a similar name and
a different signature like
EventHandlerResult
onKeyswitchEvent(Key &mappedKey, byte row, byte col, uint8_t keyState);
and
EventHandlerResult
onKeyswitchEvent2(Key &mappedKey, KeyAddr key_addr, uint8_t keyState);
As this was regarded as unacceptable confusion of the users of our
programming interface, this commit introduces some changes to the
event handler management.
Important changes:
* handlers can now be overloaded,
* handler signatures can be versioned,
* individual versions can be declared as deprecated,
* we now abort the compile if a handler reimplementation with a bad signature
is found,
* or if two or more handlers with correct signatures are detected that
are furtheron considered as ambiguous.
Impact on resources:
The proposed changes only affect the way things are handled at compile time.
No changes in terms of PROGMEM or RAM are to be expected.
Signed-off-by: Florian Fleissner <fleissner@inpartik.de>
Previously, if a Topsy key was pressed, then a non-Topsy key was
pressed, then the Topsy key was released, we'd return the event consumed
before toggling off the active state, leaving Shift keys consumed until
a Topsy key got pressed again.
Toggle the active state before the early return for Topsy key events.
Signed-off-by: Lisa Ugray <lisa.ugray@gmail.com>
Introduce an MCU driver, with a few convenience methods to disable JTAG or clock
division. Both of these were re-implemented by various hardware plugins on their
own, this collapses them into a common implementation.
Signed-off-by: Gergely Nagy <algernon@keyboard.io>
If we found no modifiers on the current layer, force a re-scan. This way we'll
scan the keymap without having to change layers first. We can't do the same
thing at `onSetup()` time, because that's too early, so we do a check in
`beforeReportingState()`.
Fixes#608.
Signed-off-by: Gergely Nagy <algernon@keyboard.io>
OneShot keys were failing to activate on the first press because of how the timeout was
implemented. Most of the time, at the end of a cycle, `should_cancel_` was being set to
`true` because the current time was being compared to the last time a OneShot key was
activated, regardless of whether or not a OneShot key was active. As a result, OneShot
keys that were pressed for the first time in a while would fail to register as OneShots.
This change prevents checking the timeout unless there is an active OneShot key.
Fixes#603.
Signed-off-by: Michael Richters <gedankenexperimenter@gmail.com>
This should stop layers from getting stuck on when a release delay is in use. Instead of
recording the keycode of the key that will be released, I now have Qukeys just use the
key's coordinates, and have it do a lookup in `Layer.live_composite_keymap_` the usual
way.
I also stopped overloading the `key_queue[x].start_time` and instead record the release
delay time in a separate variable. This greatly simplifies the timeout math, and saves
another good-sized chunk of PROGMEM, as well as avoiding timeout-conflict bugs.
Signed-off-by: Michael Richters <gedankenexperimenter@gmail.com>
Now, when we flush a key from the queue, we just insert whatever its final `Key` value
should be into `Layer.live_composite_keymap_`, obviating the need to independently track
whether a flushed qukey is in its primary or alternate state. We can also get rid of all
the code for correcting `mapped_key`, because once a key has been flushed, it won't show
up as a qukey (or DualUse key) any longer.
This substantially reduces PROGMEM usage (~300 bytes) and reduces RAM usage by 8
bytes (for the Model01 hardware).
Signed-off-by: Michael Richters <gedankenexperimenter@gmail.com>
With qukey defined with an alternate `Key` value of `ShiftToLayer(N)`, we need a way to
send the toggle-off event in order to get the layer to deactivate. Since the physical key
was already released, we need to keep track of the delayed-release qukey, and send the
appropriate release event after the rest of the queue has been flushed.
This change records the delayed key at the time of the delay, and sends the release event
after the ensuing key has been flushed from the queue.
Signed-off-by: Michael Richters <gedankenexperimenter@gmail.com>
With the changes to the updating of `Layer.live_composite_keymap_`, the code that changes
the `mapped_key` values for keys that have been flushed from the queue is unnecessary.
Signed-off-by: Michael Richters <gedankenexperimenter@gmail.com>
SpaceCadet was sending a keyswitch event with the row & column coordinates of the wrong
key. The coordinates were coming from the keypress, but the injected event was for a
previously-pressed (different) key that's still held. With the mutable
`live_composite_keymap_[]` change, this meant that pressing (and holding) `shift` then
another key would result in only a single character, rather than a repeating
character. Regardless of the minor bug, using the row & col for the event was still
logically incorrect here.
Signed-off-by: Michael Richters <gedankenexperimenter@gmail.com>
An `EPHEMERAL` keyswitch event is one that gets ignored by updates to
`live_composite_keymap_[]`, and thus doesn't change the keymap at all. This is useful for
a plugin that needs to retain its own key type in the keymap in order to do something
particular when it toggles off, but needs to inject an event with a different `Key` value,
with real key coordinates.
Signed-off-by: Michael Richters <gedankenexperimenter@gmail.com>
If we're sending injected events with `UNKNOWN_KEYSWITCH_LOCATION` instead of
the (row,col) of the OneShot key, there's no need for the function that does this
conversion.
Signed-off-by: Michael Richters <gedankenexperimenter@gmail.com>
With the changes to live_composite_keymap_, OneShot keys were no longer working. OneShot
modifiers would get stuck on, because the entry for the `OSM` key would get changed to the
modifier key value, and when the key was released, it wouldn't be processed by the OneShot
plugin, which would continue to re-apply the modifier indefinitely. At the same time,
OneShot layers wouldn't work at all, because the key would get changed to the layer shift
key, which would then turn the layer off when it was released.
This change fixes the problem by injecting keyswitch events from OneShot without physical
key coordinates, so nothing in `live_composite_keymap_` can get updated from it. OneShot
wants the `Key` entries there to remain as OneShot keys, not as whatever they're changed
to, so this is the most appropriate way I can think of to fix the issue.
Signed-off-by: Michael Richters <gedankenexperimenter@gmail.com>
If `handleKeyswitchEvent()` is called with a `mappedKey` other than `Key_NoKey`, update
`live_composite_keymap_` with the specified `mappedKey` instead of doing a keymap lookup
on active layers. This allows a plugin to call `handleKeyswitchEvent()` and specify what
the `Key` value should be, and not have to separately track what value to change it to
every cycle.
This is especially important when there's a layer change. In particular, this is the
simplest way to allow Qukeys to use `ShiftToLayer()` keys that work properly.
Fixes#501.
Signed-off-by: Michael Richters <gedankenexperimenter@gmail.com>
This adds a new version of `updateLiveCompositeKeymap()` which takes three parameters:
row, column, and a `Key` value. Instead of looking up the value to update in the key map,
it just updates `live_composite_keymap_` with the specified `Key` value.
Signed-off-by: Michael Richters <gedankenexperimenter@gmail.com>
Instead of trying to be clever by checking if a mask applies to a layer index,
just check if the layer is higher or equal (or lower, as appropriate) than our
`IGNORE_HARDCODED_LAYER` value.
This addresses keyboardio/Chrysalis#341.
Signed-off-by: Gergely Nagy <algernon@keyboard.io>
Since LED modes only send updates to the hardware at `syncDelay` intervals, calling
`update()` on the active mode every cycle doesn't make animations any smoother.
Signed-off-by: Michael Richters <gedankenexperimenter@gmail.com>
This decreases the rate at which LEDs get updated from ~60Hz to ~30Hz, which should be
fast enough that human eyes won't be able to tell the difference.
Signed-off-by: Michael Richters <gedankenexperimenter@gmail.com>
The LED-Stalker plugin was looping through its table of `step` values (one for each key)
on every call to `update()`, and calling `compute()` on every value there, but those table
entries were only updated once every 50ms (by default), resulting in a lot of repeated
computation. This resulted in the mean idle cycle time increasing from 565µs to 1317µs on
the Model01 with an (almost) unmodified standard sketch.
This change moves the check for the timeout before the loop through the `step` values, and
aborts processing there if not enough time has elapsed. The resulting mean idle cycle time
becomes 577µs -- an almost negligible increase. It also reduces the PROGMEM footprint by
44 bytes.
Signed-off-by: Michael Richters <gedankenexperimenter@gmail.com>
This commit silences a warning due to an undefined variable.
A first attemt to locally supress the warning using compiler pragmas was
unsuccessfull due to a compiler bug in all gcc versions. This bug
in all gcc versions below 6.1 makes local diagnostics suppression useless.
As future versions of Arduino will ship with later versions of gcc
from a certain point on, the supplied warning suppression macros will become useful
to avoid the necessity to set global compiler flags like -Wno... and to
enable more precise warning suppression.
The new header file comes with an example that explains how to use
the suppression mechanism.
Signed-off-by: Florian Fleissner <florian.fleissner@inpartik.de>
Various old methods provided by `Layer` have been deprecated for a while, and
were scheduled to be removed by February 14, 2019. We're past that, so lets
remove them.
Fixes#577.
Signed-off-by: Gergely Nagy <algernon@keyboard.io>
On AVR8, 16-bit bit-operations are expensive. Switch from using 3 16-bit
bitfields to using a 16-element array of a carefully constructed struct.
This saves us almost 300 bytes of PROGMEM at the cost of 10 bytes of memory.
Signed-off-by: Gergely Nagy <algernon@keyboard.io>
I tried to be smart and use `sizeof()` instead of hard-coding the array size,
but I used it wrong, and we iterated to 32 instead of 16, overflowing the array
and looking at parts of the memory we had no business looking at. This resulted
in `isPressed()` and `isSticky()` always being true.
Instead of trying to be clever, use `OneShot::ONESHOT_KEY_COUNT` throughout,
which is a constexpr defined at 16, the number of oneshot keys we have.
Fixes#572.
Signed-off-by: Gergely Nagy <algernon@keyboard.io>
In `beforeReportingState()` and `afterEachCycle()` we used to return early if no
OneShots were active. This was easy to do when we were using a bitfield, we'd
just tested for non-zero. Since we're using an array now, this check is more
expensive, and the extra work negates any benefit of returning early.
For this reason, these early returns are removed.
Signed-off-by: Gergely Nagy <algernon@keyboard.io>
Now that we're using a struct with descriptive member names, drop the helper
macros that made using bitfield manipulation readable. The code's readable
without them by now.
Signed-off-by: Gergely Nagy <algernon@keyboard.io>
On AVR8, 16-bit bit-operations are expensive. Switch from using 4(!) 16-bit
bitfields to using a 16-element array of a carefully constructed struct.
This saves us about 242 PROGMEM at the cost of 8 bytes of memory, and a tiny
performance hit.
Signed-off-by: Gergely Nagy <algernon@keyboard.io>
We want to be able to notice when the layout of the EEPROM *settings*
changed (which the CRC does not cover). For this reason, we're repurposing the
existing version setting, which wasn't widely used: it is now internal.
We use the version to determine whether the EEPROM has been written to yet, or
if it is uninitialized. This helps us make sure we're starting up with sensible
defaults.
Fixes#559, and fixes#558.
Signed-off-by: Gergely Nagy <algernon@keyboard.io>
Previously we requested a slice in `onSetup()`, but this had a major,
problematic implication: if we went from a sketch with `EEPROM-Keymap`, but
withot `LED-Palette-Theme` enabled to one with it, the EEPROM layout was not
compatible, no matter what the order of plugins in `KALEIDOSCOPE_INIT_PLUGINS`
were. This happened because `EEPROM-Keymap` requested space *after* `onSetup()`
already ran.
To fix this issue, request a slice in `.reserveThemes`, only if we didn't
request a slice yet. This way the EEPROM layout is decided by the order of
initialization in `setup()`, and we do not sneakily steal a slice in
`onSetup()`.
This is required to address keyboardio/Chrysalis#270.
Signed-off-by: Gergely Nagy <algernon@keyboard.io>
Instead of having a single `keymap.map` Focus command that tries to be smart
about what layers it presents, we should have two separate commands: one to
query the defaults (PROGMEM), and one to get/set the custom keymap (EEPROM).
This makes `keymap.map` and `keymap.roLayers` obsolete, and they're now removed.
Furthermore, having to specify whether the EEPROM keymap extends the built-in
one or not proved to be unflexible. So we re-purposed the highest bit of the
first EEPROM byte, to signal whether we should use EEPROM layers only or not.
Signed-off-by: Gergely Nagy <algernon@keyboard.io>
Florian Fleissner
Florian Fleissner
Jesse Vincent
Jesse Vincent
Michael Richters
Gergely Nagy
Lisa Ugray
Gergely Nagy
Jordan Galby
Scott Perry