- Separate concept of Flags from TaskExecutor.
- Instead, allow TaskExecutor to sync until, or while a given condition
is met.
- Flags directly store their state as an AtomicBoolean.
- Switch `executor.syncOn(flag)` to `executor.syncUntil(flag::isRaised)`
- Remove tests made redundant by improved interface.
- Add ModelPartConverter and remove ModelPartBuilder
- ModelPartConverter can convert ModelParts into Flywheel Meshes,
optionally using a transformation and TextureMapper. A
ModelLayerLocation and TextureAtlasSprite can be used instead to quickly
replicate entity models.
- Add ModelHolder and ModelCache to lazily initialize models and
automatically delete them on renderer reload
- Add SimpleModel and remove SimpleLazyModel
- Add flw.useSerialExecutor system property
- Rename ModelBufferingUtil to BakedModelBufferer
- Remove Instance.copy
- Remove Mesh.name
- Make FlwMemoryTracker thread-safe
- Fix potential thread-safety issues, texture issues, and memory leaks
related to previously memoized models
- Fix MinecartVisual only updating when not visible
- Fix ChestVisual locks not moving when opening the chest
- Add concept of flags to TaskExecutor.
- Can raise and lower flags from any thread.
- Add TaskExecutor#syncTo
- Behaves much like #syncPoint, but exits early as soon as it detects
that the requested flag has been raised.
- Document all methods in TaskExecutor.
- Do not discard tasks when destroying a VisualizationManagerImpl.
- Use flags in VisualizationManagerImpl to track frame plan and tick
plan completion.
- Use flags in BatchingEngine to track stage buffering completion and
flush completion.
- Synchronization is now needed in BatchedDrawTracker#draw.
- Use flags in IndirectEngine and InstancingEngine to track flush
completion.
- Add unit tests to validate flag behavior.
- Rename OnMainThreadPlan -> SyncedPlan.
- Add VisualizationManager and VisualManager API
- Simplify and fix memory leaks in model utilities
- Pass partialTick directly to visuals and remove AnimationTickHolder
- Fix LevelAttached and FlwTaskExecutor not being thread-safe
- Reorganize and rename many things
- Remove unnecessary things
- TransformCall now uses ForEachSlicePlan to reduce the number of
objects created.
- WaitGroup#await can now timeout. This allows the main thread to
contribute more work in a syncPoint.
- Don't normalize in transformNormal, things already are normalized.
- Not totally happy with this but it's functional and better than
directly supplying a GL object.
- Meshes provide an IndexSequence and the length of the sequence.
- IndexSequence can fill a buffer given a length.
- Special case QuadIndexSequence to optimize the most common case.
- All other sequences are treated as if they are unique
- Instancing uses EBOCache to manage ebos
- Indirect does it directly in the meshpool
- Perform various cleanup in batching code
- Remove ClientLevel argument from InstanceVertexTransformer#transform
- Rename Plan#maybeSimplify to simplify
- Rename some classes
- Move TickContext from impl to impl.visualization
- Move VertexListProviderRegistryImpl from impl to impl.vertex
- Add PlannedVisual
- Make ExampleEffect use PlannedVisual
- Remove One2ManyStorage
- Merge Storage, AbstractStorage, and One2OneStorage
- Storage directly provides update and tick plans
- Move work distribution logic to static methods in PlanUtil
- Rename RunOnAllPlan -> ForEachPlan
- Add ContextConsumer and ContextRunnable interfaces and remove
ContextAgnosticPlan
- Add VisualFrameContext for DynamicVisual#beginFrame
- Add VisualTickContext for TickableVisual#tick
- Move checks for update limiting to within the update calls themselves
- Provide update limiting/culling primitives within (B)E Visuals
- Remove methods from *Visual interfaces related to update limiting
- Add thenMap and andMap to Plan
- Add Plan primitive to transform context
- Used in Visual update dispatch
- Update dependencies, forge version
- Bump LICENCE year
- Use fma in MatrixUtil and VertexTransformations
- Remove some dead variables from TransformCall
- Use onSpinWait in WaitGroup#await
- Do not allow adding negative numbers to a WaitGroup
- Thin abstraction for TaskNotifier
- Clean up WorkerThread task polling
- Use builder pattern for defining shader compiler flows
- If only java had type inference extending to builders :whywheel:
- Support glsl 330 on instancing again
- Set uniform block binding at link time
- Remove associated glsl builder code
- Use explicit uint literals in material adapter switch cases
- No need to enable GL_ARB_explicit_attrib_location
- Fix crash trying to generate pretty errors in generated source
- Generate short, unique file names for shaders
- Cleanup some legacy code
- Catch exceptions while compiling shaders for each backend
- Catch errors while loading complex source components
- Allow compilers to progress when such an error occurs, but don't do
any actual compilation.
- Tweak error messages.
- Make resource locations in shaders default to flywheel namespace
- Make PipelineCompiler somehow simpler yet more verbose
- Switch #use to #include
- Stop ignoring errors from within shader compilers
- Track load errors in CompilerStats
- Move LoadError to separate sealed interface, LoadResult.Failure wraps
a LoadError
- Move SourceFile member parsing methods to respective classes
- Add tests for SourceFile loading
- Start work on tests for error messages
- Make ShaderSources return a LoadResult
- Refactor SourceFile loading to do all parsing outside the ctor
- Don't immediately swallow import errors, instead forward them
- Change Span to reference SourceLines instead of SourceFile
- Still WIP, but functional at this stage
- Switch to object oriented vertex array impl
- Expose vao api similar to separate attribute format
- For DSA, directly call methods
- For 33, defer state changes until bindForDraw is called
- Inline instanced mesh #drawInstances method in favor of more
fine-grained control over binding order
- Move buffer binding to GlStateTracker
- Remove raw bindVAO(0) from indirect engine
- Make VertexAttribute a sealed interface.
- Fix ebo restore state by always updating it and never restoring it
- Remove restore state wrapping in DrawCall
- Also misc cleanup.
- Add DSA abstraction to GlBuffer
- Move backing impls to their respective classes
- Remove versioned package, move impls to appropriate package
- Make fallback methods take no arguments
- Do VAO check/bind in GlStateTracker
- GlVertexArray actually checks the tracked info when binding attributes
- Add separate enable* and setup* methods in VertexArray
- Add support for DSA VAO with ARB instanced arrays incase there are
some particularly cursed drivers out there
- Misc. cleanup
- GlBuffer no longer has an explicit bind method or binding target
- Use READ/WRITE BUFFER for all data operations
- Make MappedBuffer simpler and more RAII
- Remove dead methods from GlBuffer
- Don't actually need to clear the tail of GPUInstancers on shrink
- Implement actual capability check for indirect
(I think I check all the extensions)
- Make GlCompat static
- Expand InstancedArrays compat to encompass all vertex array ops
- GlVertexArray calls into VertexArray compat layer
- VertexAttributes still use 2 separate interface methods
- Plans are now generic about a context object
- Move default composition impls to SimplyComposedPlan
- Remove PlanUtil
- Make Synchronizer implement Runnable
- Add ContextAgnosticPlan to preserve old behavior
- Reduce creation of plan objects
- Make BatchingEngine, BatchingStage directly implement plan
- Introduce BatchContext for BatchingStage plans
- Introduce FrameContext, TickContext for VisualManager plans
- Cache separate "recreation plan" for when origin shifts occur
- Ensure current backend is never null
- Fix StringUtil#countLines
- Use material index instead of file location when adapting material
shader function name
- Merge Includer and Compilation factory into ShaderCompiler and remove
builder
- Remove CompileUtil
- Remove Index
- Remove unnecessary code from ShaderSources
- Move source component classes to backend.compile.component
- Move core compilation classes to backend.compile.core
- Move pipeline classes to backend.compile
- Move GLSLVersion to glsl package
- Move SourceChecks to backend.compile
- Separate pipeline compiler from compute compiler
- Remove Pipeline field from PipelineProgramKey
- Share one UniformComponent across compilers
- Use AbstractCompiler to deduplicate some shared code
- Separate indirect programs from instancing programs
- Move compilation tracking information to CompilerStats
- Improve compilation result/link result reporting
- Raise exception when trying to slice outside DrawBuffer bounds
- Count vertices in BufferPlan#execute to account for instance creation
during Visual updates
- Mark DrawBuffers active by default
- Move pipeline code to backend package
- Store compiled programs in FlwPrograms instead of FlwCompiler
- Rename PipelineContext to PipelineProgramKey and remove CullingContext
- Remove PipelineContextSet and CullingContextSet
- Improve quality of dumped sources
- Rename pipeline directory to internal
- Add underscores before more internal GLSL variables
- Remove unnecessary GLSL files
- Use atomics to determine buffer position for batches
- Pass resultant vertex count to DrawBuffer once all transform calls
have run
- Refactor to use RunOnAllPlan within TransformCall
- Make all vertex transform logic to operate per-instance instead of
per-chunk
- Cull instances based on bounding sphere transformations
- Make BufferedMesh#mesh public to expose bounding sphere
- Roll batching #plan() arguments into FrameContext record
- Rename TransformSet -> BatchingStage
- Inline BatchingTransformManager into BatchingEngine
- Reuse one Plan object for each DrawBuffer used by a stage.
- Separate DrawBuffer acquisition from marking as active.
- Remove some unused methods in BatchingDrawTracker
- Rename variables in AnimationTickHolder
- Make flw.loadRenderDoc=false behave as expected.
- Extract TransformSet to full class
- Move planning code into BatchingTransformManager
- Reduce exposure BatchingTransformManager of internals
- Comment out WaitGroup log :ioa:
- Move ceilingDiv to MoreMath
- Use dynamic chunk size for TransformCalls
- Restore WaitGroup abstraction
- Simplify WaitGroup to make proper use of atomics
- Fix logic error in ParallelTaskExecutor causing task counter to go
below zero when executing main thread tasks
- Use ConcurrentHashMap in models to allow parallel access
- Reduce chunk size in RunOnAllPlan
- Only queue in InstanceManager
- Defer syncPoint within renderStage
- AbstractStorage return immutable view for tickable/dynamic instances
- Process InstanceManager queues off-thread
- Add FixNormalScalingMixin from Create
- Fix VisualizedRenderDispatcher#tryAddBlockEntity
- Fix flywheel.sodium.mixins.json using wrong plugin class name
- Make BatchingEngine use a shifting render origin
- Make frustum checks relative to the render origin
- Improve AbstractEntityVisual#isVisible to be closer to vanilla
- Move render origin implementation code to AbstractEngine
- Engines/InstanceManagers now provide Plans for frames/ticks
- TODO: batching is broken
- TODO: plan caching
- Add Plan interface which can be used to compose complex systems
of parallel execution
- Add Synchronizer to facilitate barriering
- Remove WorkGroup
- Summary of different primitives below, names need work :ioa:
- BarrierPlan: executes one plan then another
- NestedPlan: executes many plans at once
- OnMainThreadPlan: waits for a syncPoint to execute
- RunOnAllPlan: dispatches work over a Supplier<List>
- SimplePlan: executes many Runnables at once
- UnitPlan: does nothing
