This is part 2 of a pair of posts describing my work to convert Mozilla's SVG implementation to directly use Moz2D. Part 1 provided some background information and details about the process. This post will discuss the performance benefits of the conversion of the SVG code and future work.
For the most part the performance improvements from the conversion to Moz2D were gradual; as code was incrementally converted, little by little gfxContext overhead was avoided. On doing an audit of our open SVG performance bugs it seems that painting performance is no longer one of the reasons that we perform poorly, except for when we us cairo backed DrawTargets (Linux, Windows XP and other Windows versions with blacklisted drivers), and with the exception of one bug that needs further investigation. (See below for the issues that still do causes SVG performance problems.)
Besides the incremental improvements, there have been a couple of interesting perf bumps that are worth mentioning.
The biggest perf bump by far came when I converted the code that does the actual filling and stroking of SVG geometry to directly use a DrawTarget. The time taken to render this map then dropped from about 15s to 1-2s on my Mac. On the same machine Chrome Canary shows a blank window for about 5s, and then takes a further 20s to render. Now, to be honest, this improvement will be down to something pathological that has been removed rather than being down to avoiding Thebes overhead. (I haven't got to the bottom of exactly what that was yet.) The DrawTarget object being drawn to is ultimately the same object, and Thebes overhead isn't likely to be more than a few percent of any time spent in this code. Nevertheless, it's still a welcome win.
Another perf bump that came from the Moz2D conversion was that it enabled us to cache path objects. When using Thebes, paths are built up using gfxContext API calls and the consumer never gets to touch the resulting path. This prevents the consumer from keeping hold of the path and reusing it in future. This can be a disadvantage when the path is reused frequently, especially when D2D is being used where path creation is relatively expensive. Converting to Moz2D has allowed the SVG code to hold on to the path objects that it creates and reuse them. (For example, in addition to their obvious use during rasterization, paths might be reused for bounds calculations (think invalidation areas, objectBoundingBox content, getBBox() calls) and hit-testing.) Caching paths made us noticeably more responsive on this cool data visualization (temporarily mirrored here while the site is down) when mousing over the table rows, and gave us a +25% boost on this NYT article, for example.
For those of you that are interested in Talos, I did take a look at the SVG test data, but the unfortunately frequent up-and-down of unrelated regressions and wins makes it impossible to use that to show any overall impact of Moz2D conversion on the Talos tests. (Since the beginning of the year the times on Windows have improved slightly while on Mac they have regressed slightly.) The incremental nature of most of the work also unfortunately meant that the impact of individual patches couldn't usually be distinguished from the noise in Talos' results. One notable exception was the change to make SVG geometry use a path object directly which resulted in an improvement in the region of 6% for the svg_opacity suite on Windows 7 and 8.
Other than the performance benefits, several parts of the SVG implementation that were pretty messy and hard to get into and debug have become a lot more approachable. This has already allowed me to fix various SVG bugs that would otherwise have taken a lot longer to work out, and I hope it makes the code easier to approach for devs who aren't so familiar with it.
One final note on performance for any of you that will do your own testing to compare build - note that the enabling of e10s and tiled layers has caused significant changes in performance characteristics. You might want to turn those off.
As I noted above there are still SVG performance issues unrelated to graphics speed. There are three sources of significant SVG performance issues that can make Mozilla perform poorly on SVG relative to other implementations. There is our lack of hardware acceleration of SVG filters; there's the issue of display list overhead dwarfing painting on SVGs that contain huge numbers of elements (display lists being an implementation detail, and one that gave us very large wins in many other cases); and there are a whole bunch of "strange" bugs that I expect are related to our layers infrastructure that are causing us to over invalidate (and thus do work painting when we shouldn't need to).
Currently these three issues are not on a schedule, but as other higher priority Mozilla work gets ticked of I expect we'll add them.
The performance benefits from the Moz2D conversion on the SVG code do seem to have been positive enough that I expect that we will continue converting the rest of layout in the future. As usual, it will all depend on relative priorities though.
One thing that we should do is audit all the code that creates DrawTargets to check for backend type compatibility. Mixing hardware and software backed DrawTargets when we don't need to can cause us to unwittingly be taking big performance hits due to readback from and/or upload to the GPU. I fixed several instances of mismatch that I happened to notice during the conversion work, and in one case accidentally introduced one which fortunately was caught because it caused a 10-25% regression in a specific Talos test. We know that we still have outstanding bugs on this (such as bug 944571) and I'm sure there are a bunch of cases that we're unaware of.
I mentioned above that painting performance is still a significant issue on machines that fall back to using cairo backed DrawTargets. I believe that the Graphics team's plan to solve this is to finish the Skia backend for Moz2D and use that on the platforms that don't support D2D.
There are a few things that need to be added to Moz2D before we can completely get rid of gfxContext. The main thing we're missing is push-group API on DrawTarget. This is the main reason that gfxContexts actually wraps a stack of DrawTargets, which has all sorts of irritating fallout. Most annoying it makes it hazardous to set clip paths or transforms directly on DrawTargets that may be accessed via a wrapping gfxContext before the DrawTarget's clip stack and transform has been restored, and why I had to continue passing gfxContexts to a lot of code that now only paints directly via the DrawTarget.
The only Moz2D design decision that I've found myself to be somewhat unhappy with is the decision to make patterns relative to user-space. This is what most other hardware accelerated libraries do, but I don't think it's a good fit for 2D browser rendering. Typically crisp rendering is very important to web content, so we render patterns assuming a specific user-space to device-space transform and device space pixel alignment. To maintain crisp rendering we have to make sure that patterns are used with the device-space transform that they were created for, and having to do this manually can be irksome. Anyway, it's a small detail, but something I'll be discussing with the Graphics guys when I see them face-to-face in a couple of weeks.
Modulo the two issues above (and all the changes that I and others had made to it over the last year) I've found the Moz2D API to be a pleasure to work with and I feel the SVG code is better performing and a lot cleaner for converting to it. Well done Graphics team!
Tags: Mozilla, Moz2D, SVG
One of my main work items this year was the conversion of the graphics portions of Mozilla's SVG implementation to directly use Moz2D APIs instead of using the old gfxContext/gfxASurface Thebes APIs. This pair of posts will provide some information on that work. This post will give some background and information on the conversion process, while part 2 will provide some discussion about the benefits of the work and what steps we might want to carry out next.
When Moz2D was first being put together it was initially developed and tested as an alternative rendering backend for Mozilla's implementation of HTML
<canvas>. Canvas was chosen as the initial testbed because its drawing is largely self contained, it requires a relatively small number of features from any rendering backend, and because we knew from profiling that it was being particularly impacted by Thebes/cairo overhead.
As Moz2D started to become more stable, Thebes' gfxContext class was extended to allow it to wrap a Moz2D DrawTarget (prior to that it was backed only by an instance of a Thebes gfxASurface subclass, in turn backed by a cairo_surface_t). This might seem a bit strange since, after all, Moz2D is supposed to replace Thebes, not be wrapped by it adding yet another layer of abstraction and overhead. However, it was an important step to allow the Graphics team to start testing Moz2D on Mozilla's more complicated, non-canvas, rendering scenarios. It allowed many classes of Moz2D bugs and missing Moz2D features to be worked on/out before beginning a larger effort to convert the masses of non-canvas rendering code to Moz2D.
In order to switch any of the large number of instances of gfxContext to be backed by a DrawTarget, any code that might encounter that gfxContext and try to get a gfxASurface from it had to be updated to handle DrawTargets too. For example, lots of forks in the code had to be added to BasicLayerManager, and gfxFont required a new GlyphBufferAzure class to be written. As this work progressed some instances of Thebes gfxContexts were permanently flipped to being backed by a Moz2D DrawTarget, helping keep working Moz2D code paths from regressing.
Towards the end of 2013 it was felt that Moz2D was sufficiently ready to start thinking about converting Mozilla's layout code to use Moz2D directly and eliminate its use of gfxContext API. (The layout code being the code that decides where and how most things are placed on the screen, and by far the biggest consumer of the graphics code.) Before committing a lot of engineering time and resources to a large scale conversion, Jet wanted to convert a specific part of the layout code to ensure that Moz2D could meet its needs and determine what performance benefits it could provide to layout. The SVG code was chosen for this purpose since it was considered to be the most complicated to convert (if Moz2D could work for SVG, it could work for the rest of layout).
After drawing up a rough list of the work to convert the SVG code to Moz2D I got stuck in. The initial plan was to add code paths to the SVG code to check for and extract DrawTargets from gfxContexts that were passed in (if the gfxContext was backed by one) and operate directly on the DrawTarget in that case. (At some future point the Thebes forks could then be removed.) It soon became apparent that these forks were often not how we would want the code to be structured on completion of Moz2D conversion though. To leverage Moz2D more effectively I frequently found myself wanting to refactor the code quite substantially, and in ways that were not compatible with the existing Thebes code paths. Rather than spending months writing suboptimal Moz2D code paths only to have to rewrite things again when we got rid of the Thebes paths I decided to save time in the long run and first make sure that any gfxContexts that were passed into SVG code would be wrapping a DrawTarget. That way maintaining Thebes forks would be unnecessary.
It wasn't trivial to determine which gfxContexts might end up being passed to SVG code. The complexity of the code paths and the virtually limitless permutations in which Web content can be combined meant that I only identified about a dozen gfxContexts that could not end up in SVG code. As a result I ended up working to convert all gfxContexts in the Mozilla code. (The small amount of additional work to convert the instances that couldn't end up in SVG code allowed us to reduce a whole bunch of code complexity (and remove a lot of then dead code) and simplified things for other devs working with Thebes/Moz2D.)
Ensuring that all the gfxContexts that might be passed to SVG code would be backed by a DrawTarget turned out to be quite a task. I started this work when relatively few gfxContexts had been converted to wrap a DrawTarget so unsurprisingly things were a bit rough. I tripped over several Moz2D bugs at this point. Mostly though the headaches were caused by the amount of code that assumed gfxContexts wrapped and could provide them with a gfxASurface/cairo_surface_t/platform library object, possibly getting or then passing those objects from/to seemingly far corners of the Mozilla code. Particularly challenging was converting the image code where the sources and destinations of gfxASurfaces turned out to be particularly far reaching requiring the code to be converted incrementally in 34 separate bugs. Doing this without temporary performance regressions was tricky.
Besides preparing the ground for the SVG conversion, this work resulted in a decent number of performance improvements in its own right.
Converting the SVG code to Moz2D was a lot more than a simple case of switching calls from one graphics API to another. The stateful context provided by a retained mode API like Thebes or cairo allows consumer code to set context state (for example, fill pattern, or anti-alias mode) in points of the code that can seem far removed from other code that takes an action (for example, filling a path) that relies on that state having been set. The SVG code made use of this a lot since in many cases (for example, when passing things through for callbacks) it simplified the code to only pass a context rather than a context and some state to set.
This wouldn't have been all that bad if it wasn't for another fundamental difference between Thebes/cairo and Moz2D -- in Moz2D paths and patterns are relative to user-space, whereas in Thebes/cairo they are relative to device-space. Whereas with Thebes we could set a path/pattern and then change the transform before drawing (perhaps, say, to apply a clip in a different space) and the position of the path/pattern would be unaffected, with Moz2D such a transform change would change (and thus break) the rendering. This, incidentally, was why the SVG code was expected to be the hardest area to switch to Moz2D. Partly for historic reasons, and partly because some of the features that SVG supports lead it to, the SVG code did a lot of setting state, changing transforms, setting some more state and then drawing. Often the complexity of the code made it difficult to figure out which code could be setting relevant state before a transform change, requiring more involved refactoring. On the plus side, sorting this out has made parts of the code significantly easier to understand, and has been something I've wanted to find the time to do for years.
To continue reading about the performance benefits of the conversion of the SVG code and some possible next steps continue to part 2.
Tags: Mozilla, Moz2D, SVG