Mozilla GFX: WebRender newsletter #44

Вторник, 30 Апреля 2019 г. 19:33 + в цитатник

WebRender is a GPU based 2D rendering engine for web written in Rust, currently powering Mozilla’s research web browser servo and on its way to becoming Firefox‘s rendering engine.

WebRender on Linux in Firefox Nightly

Right after the previous newsletter was published, Andrew and Jeff enabled WebRender for Linux users on Intel integrated GPUs with Mesa 18.2 or newer on Nightly if their screen resolution is 3440x1440 or less.
We decided to start with Mesa is thanks to the quality of the drivers. Users with 4k screens will have to wait a little longer though (or enable WebRender manually) as there are a number of specific optimizations we want to do before we are comfortable getting WebRender used on these very high resolution screens. While most recent discreet GPUs can stomach about anything we throw at them, integrated GPUs operate on a much tighter budget and compete with the CPU for memory bandwidth. 4k screens are real little memory bandwidth eating monsters.

WebRender roadmap

Jessie put together a roadmap of the WebRender project and other graphics endeavors from the items discussed in the week in Toronto.
It gives a good idea of the topics that we are focusing on for the coming months.

A week in Toronto – Part deux

In the previous newsletter I went over a number of the topics that we discussed during the graphics team’s last get-together in Toronto. Let’s continue here.

WebRender on Android

We went over a number of the items in WebRender’s Android TODO-list. Getting WebRender to work at all on Android is one thing. A lot of platform-specific low level glue code which Sotaro has been steadily improving lately.

On top of that come mores questions:

Which portion of the Android user population support the OpenGL features that WebRender relies on?
Which OpenGL features we could stop relying on to cover more users
What do we do about the remaining users which have such a small OpenGL feature set available that we don’t plan to get WebRender in the foreseeable future.

Among the features that WebRender currently heavily relies on but are (surprisingly) not universally supported in this day and age:

texture arrays,
float 32 textures,
texture fetches in vertex shaders,
instancing,

We discussed various workarounds. Some of them will be easy to implement, some harder, some will come at a cost, some we are not sure will provide an acceptable user experience. As it turns out, building a modern rendering engine while also targetting devices that are everything but modern is quite a challenge, who would have thought!

Frame scheduling

Rendering a frame, from a change of layout triggered by some JavaScript to photons flying out of the screen, goes through a long pipeline. Sometimes some steps in that pipeline take longer than we would want but other parts of the pipeline sort of absorb and hide the issue and all is mostly fine. Sometimes, though, slowdowns in particular places with the wrong timing can cause a chain of bad interactions which results a back and forth between a rapid burst of a few frames followed by a couple of missed frames as parts the system oscillate between throttle themselves on and off.

I am describing this in the abstract because the technical description of how and why this can happen in Gecko is complicated. It’s a big topic that impacts the design of a lot of pieces in Firefox’s rendering engine. We talked about this and came up with some short and long term potential improvements.

Intel 4K performance

I mentioned this towards the beginning of this post. Integrated GPUs tend to be more limited in, well in most things, but most importantly in memory bandwidth, which is exacerbated by sharing RAM with the CPU. When high resolution screens don’t fit in the integrated GPU’s dedicated caches, Jeff and Markus made the observation that it can be significantly faster to split the screen into a few large regions and render them one by one. This is at the cost of batch breaks and an increased amount of draw calls, however the restricting rendering to smaller portions of the screen gives the GPU a more cache-friendly workload than rendering the entire screen in a single pass.

This approach is interestingly similar to the way tiled GPUs common on mobile devices work.
On top of that there are some optimizations that we want to investigate to reduce the amount of batch breaks caused by text on platforms that do not support dual-source blending, as well as continued investigation in progress of what is slow specifically on Intel devices.

In conclusion

It’s been a very useful and busy week. The graphics team will meet next in Whistler in June along with the rest of Mozilla. By then Firefox 67 will ship, enabling WebRender for a subset of Windows users in the release channel which is huge milestone for us.