Now we should use material passes instead of context and defines to
select a needed shader. It allows us to reduce the number of
shader combinations we use and reduce memory consumption.
All shaders were selected by context (set of defines) which increases
the number of needed permutations. For example: we don't need to know
about shadows when rendering silhouettes.
We were unnecessary passing all combinations of our scene shader context
even if we don't need it as for shadows and silhouettes for terrain
rendering. Since we only need to draw a solid color.
When the returned value of the compare function is `0` the order is
preserved. Before the sort, the order between joining clients and
non-joining clients differ.
Additionally sorting by name should make the order consintent.
The AI only noticed for changes to `isIdleConfirmed`. That is done to
avoid recursions and don't send to many messages. But when the AI is
deserialized it get's `isIdle` from the simulation. Leading to an OOS.
When the returned value of the compare function is `0` the order is
preserved. Before the sort, the order between joining clients and
non-joining clients differ.
Now everything important (which is used later on) is included in the
comparison. If still `0` is returned it's not importent which element is
taken.
Another solution would be to have one array per category.
The metadata where written every the turn of deserialization. When on
the non rejoining client it was only written when that AI makes a turn
and handles the events.
On a rejoined client the data isn't the same anymore. But it's equal.
Likely this is because the data isn't aquired at the same turn.
This could also be solved more generally by copying all data which is
pushed to the AI. Or by using the same serializer for AI as for the rest
of the simulation.
Clients which didn't late-join only get a change to `uptadeTerritories`
when the ai playes a turn. When joining clients do it at the turn when
they deserialize the ai they might be ahead of the others. That would
lead to an out of sync.
`entityObj` is not available on clients which just have rejoined. As not
all clients have access to it the data shouldn't be used for anything
elaborate.
The information where a unit is and which resource it provides wasn't
available on joining clients. Now the `resourceMap` is reset when a
resource is destroide.
It would be performance intensive when the `resourceMap` would be reset
when any unit would be destroied. To prevent this the `SharedScript`
remembers which entities are resourcesources.
When destroying a builder it might not be known anymore what it was able
to build. Now the entityIDs of the builder is stores. So that it can
easily be looked up and removed.
On newly deserialized clients there is no property for the destroied
entities. Because of that the metadata of the entity wasn't remembered
in the event and wasn't destroied.
`GetFullRepresentation` doesn't reset `this.changes` to `null`. So when
a message arrives the ai-interface will not be informed.
The ai-interface calls `GetFullRepresentation` and removed the dirty
flag from the ai interface.
This two behaviors lead to the error that the ai never receives messages
from entities which exist at game start.
The assignement has to be done for new dropsite but must not always be
done as it sends a command.
Serialized data shouldn't be too complicated. Because of that the `ent`
property is removed.
The progressive load timeout was hardcoded to 10ms, causing the CPU
to be mostly idle during map loading. This commit replaces the fixed
value with a dynamic budget derived from the display refresh rate
when VSync is enabled, or from the configured adaptivefps.menu limit
otherwise.
Fixes: #8774
Commit 6a8366b99a added additional
libraries to the report but didn't consider gloox be optional. Address
this false assumption allowing to build pyrogenesis without lobby
support.
Signed-off-by: Ralph Sennhauser <ralph.sennhauser@gmail.com>
