Texture atlas by Maldela · Pull Request #154 · iced-rs/iced

Maldela · 2020-01-12T02:08:58Z

I finally came around to implementing the texture atlas for images. I'm using guillotiere to manage the atlas space. In the current implementation the atlas only grows if necessary. It never shrinks.
I also wrote a small build script that compiles shaders at compile time using shaderc. Manually compiling shaders during development was super annoying.
Let me know what you think.

hecrj

Looks like a decent start! Thank you for giving it a shot :)

There are some things that need more work before we can merge, though! Currently, we are not getting any benefits from using an atlas.

We should also have in mind that a texture cannot keep growing indefinitely. Different hardware has different limits. To be on the safe side, we should assume atlases cannot have a dimension bigger than 8192.

Therefore, we will need to leverage multiple atlases (maybe texture arrays or arrays of textures could help us!) and even split huge images in multiple smaller ones. I am not saying we should implement this on a first iteration, but it's something we should keep in mind when designing the current data model.

Remember you can reach out to me on Zulip to discuss design ideas and ask any questions. When it comes to involved features like this one, I think this kind of interaction is crucial to help me understand your approach, get on the same page, avoid wasting efforts, and finally get your code merged. Code should be the last step. It's the easy part!

Maldela · 2020-01-17T19:53:49Z

Sorry it took me so long to update the PR. I think the new one addresses all your requests.

hecrj · 2020-01-18T10:04:29Z

You actually went ahead and implemented a texture array approach! And it seems to even split big images in multiple allocations! Awesome! 🎉

There is a lot to unpack here, so it may take me a while to go over all the changes. I'd appreciate it if you could share some details about the general implementation and any challenges, tradeoffs, future improvements or anything in particular we should be aware of.

How have you tested this? It would be great to create an example application to try these changes and, at the same time, showcase how the library can be used to create an actual application. Maybe some kind of image viewer? It may be a good idea to write some unit tests for the allocation logic.

I just took a quick glance and I think I can take it to the finish line from here, if you don't mind of course. It should be mostly a matter of moving and renaming some stuff.

Thanks for the amazing work! 🥂

Maldela · 2020-02-10T22:50:57Z

Just rebased on current master. Go ahead and rename or move anything you like.

…line.

…as empty

…ations

- Split into multiple modules - Rename some concepts - Change API details

- Update texture view on grow - Fix atlas texture coordinates - Fix fragmented uploads

hecrj

I have finally gone through the changes and refactored some stuff! Sorry for the delay here 😅

I mostly changed names, moved code around, and simplified a bunch of things. So far, I am quite happy with the result. Let me know what you think!

We should be quite close to merging this 🎉

hecrj · 2020-02-26T19:59:29Z

+                            },
+                            wgpu::VertexAttributeDescriptor {
+                                shader_location: 5,
+                                format: wgpu::VertexFormat::Uint,


I changed this to Uint to make passing non-sensical values as the layer index impossible.

hecrj · 2020-02-26T20:01:24Z

+                });
+
+            self.texture_version = texture_version;
+        }


Not sure how expensive creating a bind group is, but now we are only re-creating it when necessary instead of every frame.

hecrj · 2020-02-26T20:04:44Z

+                &self.instances,
+                0,
+                (amount * std::mem::size_of::<Instance>()) as u64,
+            );


Recreating a vertex buffer every frame is expensive. Instead, we create one with a fixed size and reuse it every frame, uploading data from a staging buffer using offsets when needed.

hecrj · 2020-02-26T20:06:35Z

+                if let Memory::Device(entry) = memory {
+                    atlas.remove(entry);
+                }
+            }


I reverted the smart deallocation on drop, as it added a bunch of Rc and RefCell complexity. I think we should keep deallocation centralized and easy to reason about.

Wasn't sure about that one either. Automatic deallocation just seemed kind of "rusty".

hecrj · 2020-02-26T20:11:39Z

+        Atlas {
+            texture,
+            texture_view,
+            layers: vec![Layer::Empty, Layer::Empty],


Apparently, there is no way to create a texture array with an array_layer_count of 1. When we do this, we get a bunch of Vulkan validation errors because the created texture is a simple 2D one and the sampler is not compatible.

I am not sure if this is a limitation of the modern graphics backends or a wgpu design decision. We should probably bring it up in the wgpu matrix channel.

In any case, we need to start with at least 2 layers for now.

Maybe we should decrease the atlas size then or make it configurable somehow. 4096 squared pixels times two is already 128 MB of data if I calculated that right.

It's 2048 currently, we could maybe lower it more.

In the long run, wgpu should offer a way to check available VRAM so we can choose a proper value at runtime.

2048 is fine, I think.

hecrj · 2020-02-26T20:12:26Z

+            Entry::Fragmented { fragments, .. } => {
+                for fragment in fragments {
+                    let (x, y) = fragment.position;
+                    let offset = (y * width + x) as usize * 4;


I have simplified this a bunch. It turns out you can use copy_buffer_to_texture to upload a sub-region of an image quite easily.

hecrj · 2020-02-26T20:12:59Z

+        self.texture = new_texture;
+        self.texture_view = self.texture.create_default_view();
+    }
+}


Overall, I really like the simplicity of this module.

Maldela · 2020-02-27T14:24:54Z

If you're happy with it, then I'm happy with it as well.

hecrj · 2020-02-28T14:17:06Z

I have changed the renderer so it doesn't create a useless image::Pipeline (and therefore, an empty atlas) when both image and svg flags are disabled.

I think this ship is ready to sail. Thank you for all the work here! 🎉

hecrj requested changes Jan 13, 2020

View reviewed changes

Comment thread wgpu/Cargo.toml Outdated

Comment thread wgpu/src/image.rs Outdated

Comment thread wgpu/src/image.rs Outdated

Comment thread wgpu/src/image.rs Outdated

Comment thread wgpu/src/image/raster.rs Outdated

hecrj added the improvement An internal improvement label Jan 13, 2020

hecrj mentioned this pull request Jan 15, 2020

InfiniteList widget #160

Open

Maldela force-pushed the atlas branch from 02f803d to b9e14c1 Compare January 17, 2020 19:49

hecrj mentioned this pull request Jan 18, 2020

Cull images and free VRAM #163

Closed

hecrj mentioned this pull request Jan 18, 2020

Make svg_fmt dependency optional nical/guillotiere#12

Merged

Maldela force-pushed the atlas branch from 6bbeebd to 74bfb0c Compare February 10, 2020 22:46

hecrj added this to the 0.1.0 milestone Feb 19, 2020

Maldela and others added 11 commits February 25, 2020 13:26

Implemented a texture atlas for images and svgs.

1bcfc9a

Merged image and svg texture atlases into one owned by the image pipe…

743637e

…line.

Some small debug changes

82e0675

Fixed texture bleeding

8562a4c

Use array of atlases instead of one growing indefinitely.

2f77a6b

Implement allocating large images across multiple texture array layers.

3f38835

Batch image draw calls into one with multiple instances

c099692

Updated shaders and removed debug_stub_derive dependency

2f695ef

When deallocating the last allocation in an allocator mark its layer …

8f9f44b

…as empty

Implemented automatic deallocation of texture space for dropped alloc…

4617da2

…ations

Recompile image shaders

82f0a49

hecrj force-pushed the atlas branch from 2c066a6 to 82f0a49 Compare February 25, 2020 12:28

hecrj added 7 commits February 26, 2020 12:34

Refactor texture atlas

59d45a5

- Split into multiple modules - Rename some concepts - Change API details

Avoid creating a vertex buffer every frame

c58d94f

Fix multiple issues from the refactoring

48d7028

- Update texture view on grow - Fix atlas texture coordinates - Fix fragmented uploads

Deallocate atlas entries and remove padding

d06d06e

Add env_logger to svg example

883a9f2

Remove unused code warnings in iced_wgpu::image

6cb7fb6

Make new texture module private for now

deedf6e

hecrj added 3 commits February 26, 2020 20:47

Derive Debug for raster::Memory

271725f

Move Debug implementation for vector::Svg

bb397cc

Move Atlas::deallocate after allocate

fc55e3a

hecrj approved these changes Feb 26, 2020

View reviewed changes

hecrj added 2 commits February 28, 2020 14:38

Stop creating image pipeline when unnecessary

4e7159c

Remove unnecessary pub(crate) use

88d4cd0

hecrj merged commit bab7dbc into iced-rs:master Feb 28, 2020

Maldela deleted the atlas branch February 28, 2020 17:06

hecrj mentioned this pull request Apr 2, 2020

Release 0.1 - Custom styling, event subscriptions, additional widgets, new examples, and more! #253

Merged

Uh oh!

Conversation

Maldela commented Jan 12, 2020

Uh oh!

hecrj left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Maldela commented Jan 17, 2020

Uh oh!

hecrj commented Jan 18, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Maldela commented Feb 10, 2020

Uh oh!

hecrj left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Maldela Feb 27, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

hecrj Feb 27, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

hecrj Feb 26, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Maldela commented Feb 27, 2020

Uh oh!

hecrj commented Feb 28, 2020

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

hecrj left a comment •

edited

Loading

hecrj commented Jan 18, 2020 •

edited

Loading

hecrj left a comment •

edited

Loading

Maldela Feb 27, 2020 •

edited

Loading

hecrj Feb 27, 2020 •

edited

Loading

hecrj Feb 26, 2020 •

edited

Loading