implement custom style layers #7039

ansis · 2018-07-27T17:50:05Z

This implements #6456. It follows the approach taken in -native and keeps the scope fairly tight. It's an alternative to #6124.

Custom layers can be created by implementing CustomLayerInterface and can be added to the map with map.addLayer(customLayer). An example. Another example using ThreeJS.

The interface

type CustomRenderMethod = (gl: WebGLRenderingContext, matrix: Array<number>) => void;

type CustomLayerInterface = {
    id: string,
    type: "custom",
    renderingMode: "2d" | "3d",
    render: CustomRenderMethod, // for rendering directly into the texture
    prerender: ?CustomRenderMethod, // for rendering into a texture
    onAdd: ?(map: Map, gl: WebGLRenderingContext) => void,
    onRemove: ?(map: Map) => void
}

Custom layers trigger repaints with the newly exposed map.triggerRepaint().

The matrix

The matrix projects spherical mercator coordinates to gl coordinates. The spherical mercator coordinate [0, 0] represents the top left corner of the mercator world and [1, 1] represents the bottom right corner. The z coordinate is conformal. A box with identical x, y, and z lengths in mercator units would be rendered as a cube.

Shared depth

Custom 3D layers share depth space with our 3D layers. /debug/threejs.html contains an example of ThreeJS sharing depth space with fill extrusions:

API design questions

Currently a layer implements render(...) if it wants to render a 2D layer and render3D(...) if it wants to render in 3D. Layers can implement both. Would it be better to only have render(...) and differentiate based on something else? For example, having two types: "type": "custom-3d" and "type": "custom"

When a custom layer takes over the gl context the state is in a mostly undefined state. Should we reset all (or a subset) of the state to default values?

Things left for later

an api for projecting to/from mercator and custom coordinate spaces
convenience methods for producing matrices for custom defined coordinate spaces
exposing support or helpers for tiled rendering

Checklist

briefly describe the changes in this PR
write tests for all new functionality
document any changes to public APIs
post benchmark scores
manually test the debug page

@davidmanzanares @ibgreen does this meet your basic needs?

andrewharvey · 2018-07-28T23:34:43Z

This is huge in what it will unlock! ❤️ Thanks @ansis

As someone writing a custom layer do I need to worry about if I'm drawing off screen or not? Is that my responsibility or does GL JS worry about that for me?

Surely there's performance considerations with your THREE example if I'm making boxes which are halfway across the world and not in the current view.

an api for projecting to/from mercator and custom coordinate spaces

I think the ultimate convenience is working with WGS84 longitude, latitude. For example from your https://github.com/mapbox/mapbox-gl-js/blob/custom-layers/debug/threejs.html example many use cases would be I have the long,lat coordinates and want to use THREE to draw something here.

mourner · 2018-07-30T15:22:45Z

@ansis the API and the implementation look really great! Exploring it in custom-layers-wind, having some troubles...

No layers draw after the wind layer. E.g. if I place it before building, no layers that come after building will be drawn. So the wind layer somehow corrupts the gl state I suppose.
I still need to do gl.pixelStorei(gl.UNPACK_PREMULTIPLY_ALPHA_WEBGL, false); in onAdd. I guess we should reset the context before onAdd too, and also expose a method to reset it, e.g. if the GL context is initialized asynchronously.
The map doesn't draw properly if I don't do gl.viewport(0, 0, gl.canvas.width, gl.canvas.height) at the end of the render call — I guess viewport isn't being properly reset at the beginning of GL rendering?
The particles now exhibit some weird striping artifacts that I haven't noticed before (especially noticeable when zooming to a certain zoom), and I have no idea what might be causing them and why we haven't seen them with the previous implementation.

anandthakker

This is great!

I wonder how we can best include tests for this API. The render tests don't make sense, since custom layers can't be specified in a style document... maybe a (much simplified) analogue to those, but as a unit test using pixelmatch?

Would it be better to only have render(...) and differentiate based on something else? For example, having two types: "type": "custom-3d" and "type": "custom"

My gut feeling would be to only have render(), leave type: "custom", and differentiate with a separate property renderingMode: '2d' | '3d'. By the way: from the implementor's point of view, is the only difference that 2d must not use the depth buffer, while 3d can?

anandthakker · 2018-07-30T14:48:38Z

src/render/draw_custom.js

+            drawOffscreenTexture(painter, layer, 1);
+        }
+
+    } else if (painter.renderPass === 'offscreen' && layer.implementation.render3D) {


Nit: for consistency with fill_extrusion (and because offscreen pass happens first), would it be clearer to swap the order of the translucent/offscreen cases here?

anandthakker · 2018-07-30T15:57:41Z

src/render/painter.js

+
+    setBaseState() {
+        this.context.cullFace.set(false);
+    }


I'm surprised that this is the only property we have to set here -- why is that?

anandthakker · 2018-07-30T16:11:48Z

src/style/style_layer.js

@@ -53,38 +55,47 @@ class StyleLayer extends Evented {
                              pixelsToTileUnits: number,
                              posMatrix: Float32Array) => boolean;

-    constructor(layer: LayerSpecification, properties: {layout?: Properties<*>, paint: Properties<*>}) {
+    +onAdd: (map: Map) => void;
+    +onRemove: (map: Map) => void;


I think these should be ?(map: Map) => void; to express that they might be absent (i.e. so that Flow enforces the if (layer.onAdd), etc. checks we're doing in style.js)

anandthakker · 2018-07-30T16:15:57Z

src/style/style_layer.js

-            this.sourceLayer = layer['source-layer'];
-            this.filter = layer.filter;
-        }
+        if (layer.type !== 'custom') {


Maybe invert this by doing if (layer.type === 'custom') return;?

anandthakker · 2018-07-30T18:24:23Z

src/style/style_layer/custom_style_layer.js

+    render3D: CustomRenderMethod,
+    onAdd: (map: Map, gl: WebGLRenderingContext) => void,
+    onRemove(map: Map): void
+}


This type definition says that render and render3D are both defined. Should they both be typed ?CustomRenderMethod instead?

anandthakker · 2018-07-30T18:31:45Z

src/style/style.js

-        }
+        let layer;
+        if (layerObject.type === 'custom') {
+            layer = createStyleLayer(layerObject);


Since this isn't going through the style spec validation path, the provided implementation needs to be validated 1) for layer id uniqueness, and 2) conformance to the CustomLayerInterface. Probably 1) makes sense to do here, and maybe 2) in the CustomStyleLayer constructor?

ryanhamley · 2018-07-30T18:56:38Z

I think the ultimate convenience is working with WGS84 longitude, latitude. For example from your https://github.com/mapbox/mapbox-gl-js/blob/custom-layers/debug/threejs.html example many use cases would be I have the long,lat coordinates and want to use THREE to draw something here.

Backing up @andrewharvey's point here with a specific example, this is exactly the kind of thing that would have been a game changer at my previous company. We had all these meshes and point clouds of real-world sites that we rendered with Three.js but we weren't able to geolocate them on a map so we simply rendered them in their own separate viewer. Being able to place them on a map would have given the users a ton of added value.

markusjohnsson · 2018-08-02T12:49:17Z

This looks very promising!

It looks as if it will allow us to move our custom canvas-based layers to this API instead. I think we will need to keep some of our scene handling logic though, like rendering two world widths to allow graphics crossing the +/-180 meridian. It would be nice if that could be handled by mapbox in the future, though I don't know how it would work.

markusjohnsson · 2018-08-03T13:39:25Z

I tested this branch to feel it out, but I got problems with label layers:

It only happened after I added a texture (using THREE.TextureLoader) and subsequently used it in a shader (THREE.ShaderMaterial). Prior to using textures it worked fine.

Gist with example source:
https://gist.github.com/markusjohnsson/b6a2d8750cd15f593010c4dfa68ce2e3

andrewharvey · 2018-08-06T07:59:28Z

Backing up @andrewharvey's point here with a specific example, this is exactly the kind of thing that would have been a game changer at my previous company. We had all these meshes and point clouds of real-world sites that we rendered with Three.js but we weren't able to geolocate them on a map so we simply rendered them in their own separate viewer. Being able to place them on a map would have given the users a ton of added value.

This function works for me to transform from WGS84 long, lat into a unit vector, anchored at the top left, as needed by this PR.

 var fromLL = function (lon,lat) {
     // derived from https://gist.github.com/springmeyer/871897
     var extent = 20037508.34;
 
     var x = lon * extent / 180;
     var y = Math.log(Math.tan((90 + lat) * Math.PI / 360)) / (Math.PI / 180);
     y = y * extent / 180;
 
     return [(x + extent) / (2 * extent), 1 - ((y + extent) / (2 * extent))];
 }

Providing a helper function for this would make it much easier for people to do #3996 (custom 3d models).

Here's an example built on top of @ansis's code, but placing a model with long, lat coords.

https://bl.ocks.org/andrewharvey/7b61e9bdb4165e8832b7495c2a4f17f7

davidmanzanares · 2018-08-06T12:28:49Z

Hi,

First of all, thank you so much for pushing this, this is an amazing feature.

Regarding your question, we've only made one small PoC of integrating this with CARTO VL, but it looked quite good. We found two small issues regarding WebGL state sharing( #7080 and #7081) but apart from that, we haven't seen any other rendering issue.

However, we do have seen a performance regression compared to our previous approach. After some profiling, I saw that the actor.js:58 receive() function took much more time when using this integration: from a total time of 9% to a total time of 19%. We also saw an important increase in execution time in web_worker_transfer.js:197 deserialize(). To be honest, I don't know if this makes any sense, I don't see any WebGL related calls in those functions. I could be seeing another performance regression not related to this branch since our integration is not aligned with master. If you think this is the case, we could update our integration with master and see if the regression is there.

We also have thought of another way to share the WebGL state in a safe way without changing this new API. I think that monkey-patching the WebGL context to provide a patched version to custom layers could be a possible way of knowing when you need to mark some state as dirty. This will remove the need of calling Context.setDirty in every frame and state would be marked as dirty as needed. The number of WebGL drawing calls could be decreased that way, but I don't know if the performance advantage is worthwhile.

Thanks!

mourner · 2018-08-06T13:07:13Z

@davidmanzanares yes, the performance regression you're seeing is likely in master too — see #7011 for more context. There's also #7015 as an attempt to win some perf back, but we're not sure it's the right approach yet.

andrewharvey · 2018-08-10T05:57:08Z

I ran into an issue where a THREE.js scene wasn't showing up with an empty map style:

style: {
    version: 8,
    sources: {},
    layers: []
}

I needed to add a background layer first. It might have been a problem on my end, but I'm not sure.

TripleMap · 2018-08-10T12:28:24Z

@andrewharvey,

Your example inspired me to use models instead of circles. But the question arises. I can not find the input point in src mapbox, to work with the current source. For example, I have about 20 point (lat, lng, height). How can I get data from source? (I belive, that it is possible).

ansis · 2018-08-13T23:31:04Z

I've pushed a bunch of fixes including some api changes:

therender3D method was removed and a "renderingMode": "2d" | "3d" property was added
an optional prerender method was added so that users can separate rendering into textures from rendering into the main framebuffer. This was added to work around cases where the user changes the blend and depth settings when drawing into the texture and would then need to reset these settings when drawing onto the map

Two render tests were added.

I think the ultimate convenience is working with WGS84 longitude, latitude. For example from your https://github.com/mapbox/mapbox-gl-js/blob/custom-layers/debug/threejs.html example many use cases would be I have the long,lat coordinates and want to use THREE to draw something here.

Yep! I think we should add this but outside of this pr so that we can get the api right for that.

I think we will need to keep some of our scene handling logic though, like rendering two world widths to allow graphics crossing the +/-180 meridian. It would be nice if that could be handled by mapbox in the future, though I don't know how it would work.

Yeah, this is tricky to implement but it's outside the scope of the current custom layers work. I'm not sure how this would work on the mapbox-gl side.

I tested this branch to feel it out, but I got problems with label layers:
@markusjohnsson thanks for the test case. It should be fixed!

We also have thought of another way to share the WebGL state in a safe way without changing this new API. I think that monkey-patching the WebGL context to provide a patched version to custom layers could be a possible way of knowing when you need to mark some state as dirty.

@davidmanzanares yep, that could be a good approach! If the setDirty() approach turns out to be a performance or safety problem we should try this.

I ran into an issue where a THREE.js scene wasn't showing up with an empty map style:

@andrewharvey Thanks for catching this. It should now be fixed!

Jesus89 · 2018-08-23T13:44:09Z

Hi @ansis,

I have tested the integration of this branch with CARTO VL, adding the new interface (attributes and functions) to our carto.Layer object to match the API.

I'm a bit worried about exposing these attributes and functions in the Layer because a user can be confused using them as part of the public API.

Could you add something to differentiate this internal API from the public API, for example, adding '$' at the beginning?

type CustomLayerInterface = {
    $id: string,
    $type: "custom",
    $render: CustomRenderMethod,
    $render3D: CustomRenderMethod,
    $onAdd: (map: Map, gl: WebGLRenderingContext) => void,
    $onRemove(map: Map): void
}

Thanks for your work!

Regards.

ansis · 2018-08-23T14:32:05Z

I'm a bit worried about exposing these attributes and functions in the Layer because a user can be confused using them as part of the public API.

Could you add something to differentiate this internal API from the public API, for example, adding '$' at the beginning?

@Jesus89 from the mapbox-gl-js perspective this will be a public API. I'm not sure if I'm understanding your issue correctly, but if adding these members to carto.Layer is an issue, could you implement them on a separate object and then add that to the map, instead of the carto.Layer itself?

Also: thanks for testing it out!

markusjohnsson · 2018-08-23T15:42:51Z

@ansis I think that what @Jesus89 refers to is that if you get a reference to the layer you would be able to make a call to one of these methods e.g. map.getLayer().render(), in that sense, the methods should not be public (but rather protected, in the TypeScript/C#/Java sense).

Edit: not that I really agree.. I think the API as it is is very nice to use.

src/geo/transform.js

mourner · 2018-08-27T08:42:55Z

debug/custom3d.html

+            x + d, y + d, z,
+            x, y + d + d, 0,
+            x + d, y + d + d, 0]);
+        const indexArray = new window.Uint16Array([


Nit: why do we access those with window? If linter complains, this probably needs a fix on the lint rules side.

src/render/draw_custom.js

mourner · 2018-08-27T08:52:16Z

src/style/style.js

-            layers: this._order.map((id) => this._layers[id].serialize())
+            layers: this._order.map((id) => this._layers[id])
+                .filter((layer) => layer.type !== 'custom')
+                .map((layer) => layer.serialize())


This is duplicated with _serializeLayers. Let's DRY up. Maybe also write with a for loop / push for less short-lived allocations?

sakeenaq · 2018-08-30T10:44:01Z

@markusjohnsson wrt #7039 (comment) I was facing the same issue with the label layers, did you come across any resolution for it?

markusjohnsson · 2018-08-30T10:58:48Z

@SakeenaQureshee no, I just tested with the latest changes, and I still have the same problem (at least at certain zoom levels).

andrewharvey · 2018-08-31T04:07:25Z

@SakeenaQureshee no, I just tested with the latest changes, and I still have the same problem (at least at certain zoom levels).

I think it affects other textures too including raster sources.

mourner

This looks good to me, pending a green rebase on master. We can follow-up on improvements with separate PRs.

ansis · 2018-09-06T20:30:23Z

@Jesus89 thanks for the feedback but we've decided to leave the interface as is

@measuredweighed the current implementation should be pretty stable. The first production release it will be included in will be released October 10th.

Jesus89 · 2018-09-06T20:51:56Z

OK. Thanks for your work! 👍

tibotiber · 2018-09-26T12:56:37Z

This feature is huge, thanks for the amazing work. A quick question: if I wanted to add a lot of independent objects (e.g. threejs 3D models), each with separate locations, would adding a custom layer (with its own scene, camera and renderer) for each object pose any performance issue? Or would you recommend I create a single custom layer in which I add the different objects computing location based on camera matrix and intended position?

In the second case, I'd imagine the location computation will be a bit more complex. Any pointer on this? Probably something like gl_Position = projectionMatrix * viewMatrix * modelMatrix * vec4( position, 1.0 ); as seen in mrdoob/three.js#1188 (comment) but not sure where I'd get viewMatrix (camera.matrixWorldInverse) from?

Cc @pviejo

tibotiber · 2018-09-27T08:07:54Z

Update on my previous comment. I have gone the second way and implemented one single custom layer centered in (lat,lng)=(0,0) in which I add multiple objects for which the position is computed using the fromLL method shared by @andrewharvey. Works really well and I believe more efficient than having a separate renderer for each object. Thanks again for the great work 👍.

tibotiber · 2018-09-28T03:32:04Z

Re-update FWIW. threejs would have rendering artifacts on large scenes, so it's better locating the custom layer scene near your area of interest and offsetting objects in the scene accordingly.

wandergis · 2018-10-12T09:03:18Z

Update on my previous comment. I have gone the second way and implemented one single custom layer centered in (lat,lng)=(0,0) in which I add multiple objects for which the position is computed using the fromLL method shared by @andrewharvey. Works really well and I believe more efficient than having a separate renderer for each object. Thanks again for the great work 👍.

Does Raycaster works also works well ?

tibotiber · 2018-10-12T10:36:31Z

@wandergis no, completely out of sync 😬

rveciana · 2018-10-23T06:25:42Z

Hi, does somebody know how to calculate the positions? The initial comment says:

The spherical mercator coordinate [0, 0] represents the top left corner of the mercator world and [1, 1] represents the bottom right corner.

Taking the Three.js example, for instance:

The horizontal coordinate is easy: (lon +180)/360. In the example, would be (-79.390307 + 180) / 360 = 0.279471369, which is the position in the code
The vertical position doesn't seem to work the same way. It should be something like (90 - lat)/180, but applying it to the example, 0.257450244, very far away from the correct 0.364935
I know that the upper limit seems 85 degrees, not 90, but changing the values doesn't give a good result
0.5 is the equator, as expected, but 45 degrees is 0.33, instead of the expected 0.25. Does the position depends on the cos(lat)? I don't find the formula neither...

andrewharvey · 2018-10-23T06:28:09Z

@rveciana See the code snippet in my comment at #7039 (comment)

lileialg · 2018-10-31T06:02:21Z

For the first,this feature is very great,I can upload my shader to mapbox-gl,So I can display unlimited visualization; but I have a question, when I increase the zoom to 17,18,19,20,21,22....., the position of point is not stable, the position of points are shaking with the zoom increasing or map rotating, the attachment is my code file,I am looking for your response, thanks very much
custom3d-4.txt

measuredweighed · 2018-10-31T09:52:44Z

@lileialg This is related to a lack of floating point precision. I opened an issue about this (#7268) and also provided one possible workaround.

lileialg · 2018-10-31T10:27:59Z

@lileialg This is related to a lack of floating point precision. I opened an issue about this (#7268) and also provided one possible workaround.
Thank you very much, that's really help

lileialg · 2018-11-01T08:42:05Z

@ansis @measuredweighed I covert position to pixel in javascript instead of in shader script,It looks like working fine.
custom3d-13.txt

bhison · 2019-07-26T13:14:05Z

Re-update FWIW. threejs would have rendering artifacts on large scenes, so it's better locating the custom layer scene near your area of interest and offsetting objects in the scene accordingly.

@tibotiber Thanks for your above comments, they've really set me on the right path for rendering multiple 3D objects across my map. I've started from the Add a 3D model example, set the center of my map as the translation for the Three projection matrix and then applied an offset transform on the object to have it positioned correctly. Fairly standard stuff

My model is now in the correct position, with the right scale, a slightly off rotation (which I assume is the symptom of floating point issues), but the major issue is for some reason it now looks like the normals are inverted on my model. I was wondering if you came up against anything like this? I've tried fiddling with a lot of options but I'm perplexed, I can't see how this could be happening.

tibotiber · 2019-07-26T13:40:11Z

@bhison glad the comments could help :). And no I haven't met any issue concerning normals. Sorry, can't be of much help on this.

bhison · 2019-07-29T13:01:20Z

@bhison glad the comments could help :). And no I haven't met any issue concerning normals. Sorry, can't be of much help on this.

Knowing you haven't had the same issue is kind of helpful from a process of elimination perspective. Much appreciated.

ansis requested review from mourner and anandthakker July 27, 2018 17:50

ansis force-pushed the custom-layers branch from 7f302f3 to 663d22f Compare July 27, 2018 18:29

ryanbaumann mentioned this pull request Jul 28, 2018

Custom WebGL layer #6124

Closed

anandthakker reviewed Jul 30, 2018

View reviewed changes

ibgreen mentioned this pull request Jul 31, 2018

Mapbox Custom Layer Integration visgl/deck.gl#2132

Closed

11 tasks

anandthakker mentioned this pull request Aug 1, 2018

Custom WebGL Layers API #6456

Closed

mourner mentioned this pull request Aug 1, 2018

Eliminate opaque pass rendering optimization #7063

Closed

andrewharvey mentioned this pull request Aug 6, 2018

Add support for rendering 3d models #3996

Closed

go-d mentioned this pull request Aug 10, 2018

Performance regression: feature-state _idMap building is expensive #7110

Closed

1ec5 mentioned this pull request Aug 14, 2018

Allow annotation views to lie flat against a tilted map mapbox/mapbox-gl-native#10498

Closed

ansis force-pushed the custom-layers branch from 79aeaa3 to c893194 Compare August 14, 2018 19:33

mourner reviewed Aug 27, 2018

View reviewed changes

mourner approved these changes Sep 6, 2018

View reviewed changes

implement custom style layers

2309d63

ansis force-pushed the custom-layers branch from f51bdba to 2309d63 Compare September 6, 2018 20:08

ansis merged commit f556bee into master Sep 6, 2018

ryanhamley mentioned this pull request Sep 20, 2018

Canvas source not respecting alpha values set on canvas element #7302

Closed

corradio mentioned this pull request Oct 5, 2018

Investigate webgl-based wind layer electricitymaps/electricitymaps-contrib#944

Closed

dinglei19 mentioned this pull request Oct 25, 2018

how to calculate the value of tanslate based on longitude and latitude. #7481

Closed

ansis mentioned this pull request Oct 25, 2018

expose mercator projection for custom layers #7485

Closed

rlueder mentioned this pull request Nov 15, 2018

Question - Possibility to add 3D model or gif into the map alex3165/react-mapbox-gl#624

Closed

Wykks mentioned this pull request Dec 9, 2018

Integrate new custom layer feature if possible (at least make an example) Wykks/ngx-mapbox-gl#82

Open

Kamrad mentioned this pull request Dec 23, 2018

CustomLayer + three.js // need for camera position #7395

Closed

This was referenced May 12, 2019

SceneKit annotation view mapbox/ios-sdk-examples#111

Open

Allow custom layers to share depth buffer with fill extrusion layers mapbox/mapbox-gl-native#14641

Closed

mourner deleted the custom-layers branch August 1, 2019 11:47

implement custom style layers #7039

implement custom style layers #7039

Conversation

ansis commented Jul 27, 2018 • edited Loading

The interface

The matrix

Shared depth

API design questions

Things left for later

Checklist

andrewharvey commented Jul 28, 2018

mourner commented Jul 30, 2018

anandthakker left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

ryanhamley commented Jul 30, 2018

markusjohnsson commented Aug 2, 2018

markusjohnsson commented Aug 3, 2018 • edited Loading

andrewharvey commented Aug 6, 2018 • edited Loading

davidmanzanares commented Aug 6, 2018

mourner commented Aug 6, 2018

andrewharvey commented Aug 10, 2018

TripleMap commented Aug 10, 2018

ansis commented Aug 13, 2018

Jesus89 commented Aug 23, 2018

ansis commented Aug 23, 2018 • edited Loading

markusjohnsson commented Aug 23, 2018 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

sakeenaq commented Aug 30, 2018 • edited Loading

markusjohnsson commented Aug 30, 2018

andrewharvey commented Aug 31, 2018

mourner left a comment

Choose a reason for hiding this comment

ansis commented Sep 6, 2018

Jesus89 commented Sep 6, 2018

tibotiber commented Sep 26, 2018 • edited Loading

tibotiber commented Sep 27, 2018 • edited Loading

tibotiber commented Sep 28, 2018

wandergis commented Oct 12, 2018

tibotiber commented Oct 12, 2018

rveciana commented Oct 23, 2018

andrewharvey commented Oct 23, 2018

lileialg commented Oct 31, 2018

measuredweighed commented Oct 31, 2018

lileialg commented Oct 31, 2018

lileialg commented Nov 1, 2018 • edited Loading

bhison commented Jul 26, 2019 • edited Loading

tibotiber commented Jul 26, 2019

bhison commented Jul 29, 2019 • edited Loading

ansis commented Jul 27, 2018 •

edited

Loading

markusjohnsson commented Aug 3, 2018 •

edited

Loading

andrewharvey commented Aug 6, 2018 •

edited

Loading

ansis commented Aug 23, 2018 •

edited

Loading

markusjohnsson commented Aug 23, 2018 •

edited

Loading

sakeenaq commented Aug 30, 2018 •

edited

Loading

tibotiber commented Sep 26, 2018 •

edited

Loading

tibotiber commented Sep 27, 2018 •

edited

Loading

lileialg commented Nov 1, 2018 •

edited

Loading

bhison commented Jul 26, 2019 •

edited

Loading

bhison commented Jul 29, 2019 •

edited

Loading