Loader.onLoadComplete is dispatched *before* the Loader is reset. If you have a `create` method in your State please note that the Loader will have been reset before this method is called. This allows you to immediately re-use the Loader without having to first reset it manually.
Correct comments:
-Phaser.Camera: checkWorldBounds =>checkBounds
-Phaser.RetroFont: Set correct @name for name and smoothed
-Phaser.DOM: inViewport => inLayoutViewport
This change implements the original suggestion of using `updateTransform`,
but applies so globally instead of within a particular postUpdate
function.
Now the game loop calls `updateTransform` after each `updateLogic` call
unconditionally; it is updates that change the world that are accounted
for, not the rendering. This removes some previous checks that were
preventing correct behavior with the previous patch.
This makes the assumption that game objects (eg. Sprites) are only
modified within callbacks triggered before the completion of the
`postUpdate` walking of the scene graph.
- User code that runs outside of the "game update", such as a `setTimeout`
timer, will need to explicitly update transformations so that the world
is synced by the next `preUpdate`: but this is not the expected case and
is already outside the Phaser update model.
- If this assumption does not hold or is too weak, the transformations
could also be applied once at the start of every game update loop
(before any render or update). This change would at most double the time
spent on apply the transformations.
The constant application of `updateTransform` passes all reported failing
cases and resolves#1424 just as the original proposal of having the
change performed in the Sprite postUpdate but will work more consistently
across all scene-bound game objects.
On a desktop Chrome browser the inclusion also has minimal relative impact
as shown by the summarized results. The percentages given are the summed
CPU time of relevant required operations along with that of the
updateTransform itself:
- 10,000 non-collision particles:
- 12% pre/post update, 2.4% updateTransform
- 100 colliding particles:
- 2% pre/post update & collision, 0.3% updateTransform
- 1000 colliding particles:
- 40% pre/post update & collision, 1% updateTransform
With this patch the updateTransform time does creep up _slightly_ (vs just
in `Sprite.postUpdate`) but it is still dominated by required game
updates, and more so, by any actual work like Physics.
- Clarified proper usage of `pageAlignVertically` and add note about 2.2
change and how to obtain 2.1 behavior.
- Removed the `@readonly` status of the `parentIsWindow` and `parentNode`;
these can be updated in a controlled manner.
- Added intra-hyperlinks
- Updated some ancillary documentation
The click trampoline added for IE prevents Chrome for Android from being
able to launch Full Screen mode with the default parameters for
`ScaleManger#startFullScreen`. (The desktop version of Chrome is not
affected.)
This fix adds an additional compatibility settings (`clickTrampoline`)
that can be used to configure when such is used. By default the
'when-not-mouse' mode is only enabled for Desktop browsers, where the
primary input is ubquitously a mouse.
There are no known breaking compatibility changes - the Full Screen should
be initiatable in Chrome for Android as it was in 2.1.x. The default
Android browser does not support Full Screen.
As pointed out, `newChild.parent` could be accessed after it was set to
undefined. This fix unifies the code from the various `destroy` methods so
the previou issue does not occur.
There are a bunch of signals added for Sprites; more when input is
enabled. However, very few of these signals are ever actually used. While
the previous performance update related to Signals addressed the size of
each Signal object, this update is to reduce the number of Signal objects
as used by the Events type.
As a comparison the "Particle: Random Sprite" demo creates 3200+ Signals;
with this change there less than 70 signals created when running the same
demo. (Each Event creates at 8 signals by default, and there is an Event
for each of the 400 particles.) While this is an idealized scenario, a
huge amount (of albeit small) object reduction should be expected.
It does this by creating a signal proxy property getter and a signal
dispatch proxy. When the event property (eg. `onEvent`) is accessed a new
Signal object is created (and cached in `_onEvent`) as required. This
ensures that no user code has to perform an existance-check on the event
property first: it just continues to use the signal property as normal.
When the Phaser game code needs to dispatch the event it uses
`event.onEvent$dispath(..)` instead of `event.onEvent.dispatch(..)`. This
special auto-generated method automatically takes care of checking for if
the Signal has been created and only dispatches the event if this is the
case. (If the game code used the `onEvent` property itself the event
deferal approach would be defeated.)
This approach is designed to require minimal changes, not negatively
affect performance, and reduce the number of Signal objects and
corresponding Signal/Event resource usage.
The only known user-code change is that code can add to signal (eg.
onInput) events even when input is not enabled - this will allow some
previously invalid code run without throwing an exception.
- Updated `readOnly` doclet to `readonly`
- `array` refined to `type[]`, where such information was immediately
determinable.
- Updated {Any}/{*} to {any}; {...*} is standard exception
- Udated {Object} to {object}
The substraction of `physicsElapsedMS` needs to be done for all individual
updates. (When current FPS ~ target FPS this is a 1-1 mapping, but catchup
updates can throw off the calculations.)
Also renamed `Game#updateNumber` (a poor initial name on my part) to
`currentUpdateID`. This matches the naming of
`Stage#currentRenderOrderID`.
- Changed `count` from 0d9678e512 to
`updateNumber` and expanded documentation; also moved primary usage back
to local variable.
- Added `updatesThisFrame` which allows (logic) code to detect if it is
the last update, or if there are pending updates the same frame. While
it could be adventageous in certain cases it will be problematic if such
update logic relies in the supplied delta time, as such should change if
fixed-timing is deviated from or extended updates are done.
- Formatting and documentation.
There are a fair amount of Signal objects created. In the current
implementation these are somewhat "fat" objects for two reasons:
- A closure / ad-hoc `dispatch` is created for each new Signal - this
increases the retained size by 138+ bytes/Signal in Chrome.
- There are a number of instance variables that never change from their
default value, and the bindings array is always created, even if never
used.
This change "lightens" the Signals such that there is significantly less
penalty for having many (unusued) signals.
As an example of how much this _does_ play a role, in the "Random Sprite"
demo ~2000 Signals are created, with only 12 of these signals being
subscribed to. This results in a shallow size of ~300K and a retained size
of ~600K (which is almost as much memory as required by the
Sprites/Sprites themselves!) for .. nothing.
With these changes the shallow and retained sizes are less than 50K each -
and this is with only ~200 sprites!
This change addresses these issues by
- Changing it so there is _no_ `dispatch` closure created. This is a
_breaking change_ (although there is no usage of such in core where it
breaks); the code referenced "#24", but no such issue could be found on
github.
In the rare case that code needs to obtain a dispatch-closure, the
`boundDispatch` property can be used to trivially obtain a (cached)
closure.
- The properties and default values are moved into the prototype; and the
`_bindings` array creation is deferred. This change, coupled with the
removal of the automatic closure, results in a very lightweight
~24bytes/object (in Chrome) for unbound signals.
This is minor breaking change, as per the removal of the automatic
closure; but no such need/usage was found in the core. (There may be cases
in the examples, which were not checked.)
If merely opting for the array creation delay and keeping the default
properties in the prototype the shallow size is still halved; less
significant but still a consideration.
This de-optimization occurred between 2.0.7 and 2.1.0 and is currently
present through dev.
`Group.forEach`, which is used by QuadTree, had an extreme de-optimization
in assigning to `arguments` - _CPU profiling showed as much as 50% of the
time was used by Group.forEach_ (after the correction it is not
registered) due to this de-optimization making the "When Particles
Collide" demo run with an unsatisfactory performance, even on a Desktop.
The fix uses a separate array and push (which is optimizable; the previous
implementation was not optimizable in Chrome, FF, or IE!).
This also fixes usages of `slice(arguments,..); ushift` elsewhere in
Group, using the same convention. It applies the same update for `iterate`
as does https://github.com/photonstorm/phaser/pull/1353 so it can also
accept null/undefined for `args` from the invoking functions.
