Tracker: Performance - Githubissues

ibgreen commented 7 years ago

[x] Provide a target array for attribute updaters to write into, avoiding generation of millions of temporary objects during certain attribute generation.
[x] Support generic vertex attributes: Setting an accessor to a constant would cause a generic vertex attribute to be set, completely bypassing attribute generation (requires luma.gl fix).
[x] Create a shared instancePickingColors attribute (resized to a high water mark) that can be shared between all layers, instead of generating one for each layer (could give a 20-25% boost in attribute update time for many layers, in addition to obvious memory savings).
[ ] Use single interleaved WebGL buffers with offsets and strides separating attributes (may only give <5% improvement on modern GPUs)
[x] Pack colors and picking colors as Uint8s instead of Float32s
[x] Allow apps to pre generate and pass in WebGL Buffers (not just typed arrays) as props.
[ ] Investigate inactive tabs and GPU memory optimizations #1775

[ ] Eliminate high cost of geojson splitting for Paths and Polygons
- [ ] Make PathLayer support multipaths, and empty paths
- [ ] Make SolidPolygonLayer support multipolygons, and empty polygons
- [ ] Make GeoJson pass through unmodified FeatureCollection to PathLayer and SolidPolygonLayer
[ ] Improve Tesselation support to ensure custom geometries are built in most efficient way.

[ ] Separate changeFlag dirtying from layer updates
[ ] Disable attribute updates for non-visible layers (requires previous)
[x] Disable CompositeLayer.renderLayers on shouldUpdate returns false (i.e. reuse layers from last render)
[x] Allow reuse of layers and optimize Layer matching accordingly (see #257)
[x] Layer matching - Use a map to match layer ids vs. repeated linear searches O(N) vs O(N2)

[ ] Allow the app to control what type of picking a layer supports (e.g. hover vs. click vs. none) - as hovering over a very large layer can cause noticeable CPU activity, while clicks are much less frequent.
[ ] Under WebGL2, use WebGLBuffer for readPixels to reduce GPU roundtrip costs.
[x] Mode that disables picking during pan and zoom operations?
[ ] Document how to override default color based picking, for layers that can provide more efficient mathematical implementations (e.g. regular tilings like grid and hexagons layers).
[x] Allow multiple layers to render picking colors into the same framebuffer, so that we only need one (or at least fewer) expensive calls to gl.getPixels
[x] Reuse framebuffer between picking operations.

[ ] Provide another appropriately resized framebuffer or texture as common scratch pad area, to avoid each effect allocating its own?
[ ] Refactor life cycle management so that effects can also use it?

[ ] Animation mode - allow some uniforms to be updated between WebGL frames without requiring passing new layers to deck.gl.
[ ] Expose AttributeManager to apps - so that apps can precalculate attributes, and reuse buffers that are shared between layers.

k commented 7 years ago

Are there any plans for when these performance improvements will be done?

More specifically reusing layers.

Pessimistress commented 7 years ago

@k Reusing layers should be available in 4.1.0-beta. It is still experimental, so testing is welcome.

k commented 7 years ago

@Pessimistress great thanks. I'll check it out soon.

visgl / deck.gl