Quad feature #528
Quad feature #528
Conversation
…nt image. Move getBuffer from various gl features to geo.util.getGeomBuffer so as to avoid duplicating code. Requires a change to vgl to support adjusting the modelViewOriginUniform. If multiple quads use the same texture, that texture is only created once. Preview images are supported for tiles, but only if those preview images are loaded -- if they are still loading, they won't be used. It would require quite a bit more logic to try to load both the preview and the desired image, and properly only show the desired image if it loaded first. Optionally cache information about quads to reduce computation. Cache vgl textures inside images to reduce recomputing them. Support solid-colored quads, including previewColor. Switch the tile layer to use the quad feature. If the tiles haven't changed, don't requeue them. Before, on all updates, we would calculate which tiles are needed and then make sure they were queued.
This is not really sufficient, as reprojecting tiles cause the tiles to have a non-uniform density, which means our calculation of the appropriate tiles to show at different zoom levels is incorrect. If a projection cannot wrap and wrapping is turned on, we may try to ask for a functionally infinite set of tiles. Do a little less work while checking the tile queue.
Add some protection against too many tiles.
The mockVGLRenderer functions now count how often GL functions are invoked. Note that with the change for the tile layer from gl.planeFeature to gl.quadFeature, the gl.planeFeature source module is no longer covered.
Using other example's thumbnails seems too ugly. I've swapped them out for flower pictures.
Add some more core to improve the behavior with reprojection. If the tile layer had a minimum tile level other than zero and keepLower was true, zooming out would make all tiles vanish.
|
This addresses issues #517 (use a single shader program), issue #228 (update planeFeature -- though, strictly, planeFeature hasn't been altered), issue #374 (non-square tiles -- additional work will be needed to truly generalize this). This helps issue #473 (optimize tileLayer._update -- it does much less work if the tile set isn't changing) and issue #474 (reduce memory allocations). |
|
I added the qlQuad screenshot. Is the new parallelProjection failure expected? |
|
No. I was expecting everything to pass. |
|
When I load the parallelProjection test locally, it looks fine. |
|
I like the new test images 😄 lily is my favorite. Will review it but it won't be quick since it has reasonable changes |
The resize event was fired when the camera viewport had been resized but its bounds had not been updated. This meant that during a resize event the displayToGcs function would return invalid results.
|
Here are some timing results, using the transitions example, starting with the browser having cached the tile images, but geojs not having cached the tiles except those visible in the default view, and running through the five transitions in the order shown in the example a total of three times. In all cases I had the developer console open on the browser, which probably impacted the results some. The frame times are the amount of time spent in
The tests were run on my desktop and my laptop. The desktop has an AMD FirePro W5000. Short summary: the quads are vastly faster than the planes. Except for Firefox on my laptop, the speed is very good, with very few dropped frames. IE reported very few slow frames, but only animated half as many frames as expected. This means that it spent a lot of time outside of |
…h low resolution z-buffers. Specifically, on Chrome on Android (current version), using too small a z-value results in the layers flickering.
This is awesome @manthey How did you capture these times? |
| reset_minimum_zoom(); | ||
| var newZoom = fix_zoom(m_zoom); | ||
| if (newZoom !== m_zoom) { | ||
| m_this.zoom(newZoom); | ||
| } | ||
| m_this.camera().viewport = {width: w, height: h}; |
There was a problem hiding this comment.
@manthey why was this necessary?
|
@manthey looks great! I only have few questions. I didn't review the test code line by line but did look into the API and GL implementation and it looks great ! |
|
@manthey I locally merged your branch into master did npm install and grunt and got other examples running but the quad one is failing: http://localhost:8082/data/grid.jpg Failed to load resource: the server responded with a status of 404 (Not Found) |
|
Also, in wms example, I don't see wms layer unless I zoom in or pan. Not sure if this branch caused this may be this was the case before. (http://localhost:8082/examples/wms/) |
|
I captured the times by instrumenting the |
|
Regarding http://localhost:8082/data/grid.jpg -- I think this file is only pulled down from Girder if you also build the selenium tests. |
|
I see the issue with the WMS example, too. I'll look into it. |
|
@aashish24 The issue with the WMS example is that we change the url after the wms tile layer is created, and I was failing to clear a cache. It is good that it showed up this error, but from an efficiency standpoint, the WMS example could just start with the correct url, so it doesn't bother to load a handful of different tiles before loading the tiles we want. |
|
I don't know if @aashish24 wants to chime in again, but this LGTM. |
|
LGTM Ship it! |
Adding a quad feature which allows multiple quads each with a different image.
Moved getBuffer from various gl features to geo.util.getGeomBuffer so as to avoid duplicating code.
Updates to the latest vgl as it needs to be able to set the origin of the modelViewOrigin uniform.
If multiple quads use the same texture, that texture is only created once. This doesn't get leveraged with an osm layer using a single image, as the tile layer creates a separate Image element per tile, but does get the benefit with general quads.
Preview images are supported for tiles, but only if those preview images are loaded -- if they are still loading, they won't be used.
Optionally cache information about quads to reduce computation. Cache vgl textures inside images to reduce recomputing them.
Support solid-colored quads, including previewColor.
Switched the tile layer to use the quad feature.
If the tiles haven't changed, don't requeue them. Before, on all updates, we would calculate which tiles are needed and then make sure they were queued.
Added some support for altering tile projections. This is not really sufficient, as reprojecting tiles cause the tiles to have a non-uniform density, which means our calculation of the appropriate tiles to show at different zoom levels is incorrect. If a projection cannot wrap and wrapping is turned on, we may try to ask for a functionally infinite set of tiles.
Do a little less work while checking the tile queue. This greatly reduces the cpu load when no new tiles are needed or old tiles go away.
The mockVGLRenderer functions now count how often GL functions are invoked. This is used in some of the tests.
Note that with the change for the tile layer from gl.planeFeature to gl.quadFeature, the gl.planeFeature source module is no longer covered.
Fixed a bug that could occur if a tile layer didn't include tiles for zoom level 0.