Month: June 2015

Now that we have a quite complete implementation of CSS Grid Layout specification it’s time to take care of performance analysis and optimizations. In this essay, which is the first of a series of posts about performance, I’ll first introduce briefly how to use Blink (Chrome) and WebKit (Safari) performance analysis tools, some of the most interesting cases I’ve seen during my work on the implementation of this spec and, finally, a basic case to compare Flexbox and Grid layout models, which I’d like to evolve and analyze further in the coming months.

Performance analysis tools

Both WebKit and Blink projects provide several useful and easy to use scrips (python) to run a set of test cases and take different measurements and early analysis. They were written before the fork, that’s why related documentation can be found at WebKit’s track, but both engines still uses them, for the time being.

Tools/Scripts/run-perf-tests
Tools/Scripts/webkitpy/performance_tests/

There are a wide set of performance tests under PerformanceTest folder, at Blink’s/WebKit’s root directory, but even though both engines share a substantial number of tests, there are some differences.

(blink’s root directory) $ ls PerformanceTests/
Bindings BlinkGC Canvas CSS DOM Dromaeo Events inspector Layout Mutation OWNERS Parser resources ShadowDOM Skipped SunSpider SVG XMLHttpRequest XSSAuditor

Chromium project has introduced a new performance tool, called Telemetry, which in addition of running the above mentioned tests, it’s designed to execute more complex cases like running specific PageSets or doing benchmarking to compare results with a preset recording (WebPageRelay). It’s also possible to send patches to performance try bots, directly from gclient or git (depot_tools) command line. There are quite much information available in the following links:

• Telemetry
• Bisecting performance regressions
• Performance try-bots

Regarding profiling tools, it’s possible both in Webkit and Blink to use the –profiler option when running the performance tests so we can collect profiling data. However, while WebKit recommends perf for linux, Google’s Blink engine provides some alternatives.

CSS Grid Layout performance tests and current status

While implementing a new browser feature is not easy to measure performance while code evolves so much and quickly and, what it’s worst, be aware of regressions introduced by new logic. When the feature’s syntax changes or there are missing or incomplete functionality, it’s not always possible to establish a well defined baseline for performance. It’s also a though decision to determine which use cases we might care about; obviously the faster the better, but adding performance optimizations usually complicates code, it may affect its robustness and it could lead to unexpected, and even worst, hard to find bugs.

At the time of this writing, we had 3 basic performance tests:

• auto-grid-lots-of-data.html – 100 x 20 auto-sized grid (no stretch).
• fixed-grid-lots-of-data.html – 100 x 20 fixed-sized grid (no stretch).
• fixed-grid-lots-of-stretched-data.html – 100 x 20 fixed-sized grid (stretch)

Why we have selected those uses cases to measure and keep track of performance regression ? First of all, note that auto-sizing one of the most expensive branches inside the grid track sizing algorithm, so we are really interested on both, improving it and keeping track of regressions on this code path.

body {
    display: grid;
    grid-template-rows: repeat(100, auto);
    grid-template-columns: repeat(20, auto);
}
.gridItem {
    height: 200px;
    width: 200px;
}

On the other hand, fixed-sized is the easiest/fastest path of the algorithm, so besides the importance of avoiding regressions (when possible), it’s also a good case to compare with auto-sized.

body {
    display: grid;
    grid-template-rows: repeat(100, 200px);
    grid-template-columns: repeat(20, 200px);
}
.gridItem {
    height: 200px;
    width: 200px;
}

Finally, a stretching use cases was added because it’s the default alignment value for grid items and the two test cases already described use fixed size items, hence no stretch (even though items fill the whole grid cell area). Given that I implemented CSS Box Alignment support for grid I was conscious of how expensive the stretching logic is, so I considered it an important use case to analyze and optimize as much as possible. Actually, I’ve already introduced several optimizations because the early implementation was quite slow, around 40% slower than using any other basic alignment (start, end, center). We will talk more about this later when we analyze a case to compare Flexbox and Grid performance in layout.

body {
    display: grid;
    grid-template-rows: repeat(100, 200px);
    grid-template-columns: repeat(20, 200px);
}
.gridItem {
    height: auto;
    width: auto;
}

The basic HTML body of these 3 tests is quite simple because we want to analyze performance of very specific parts of the Grid Layout logic, in order to detect regressions in sensible code paths. We’d like to have eventually some real use cases to analyze and create many more performance tests, but chrome performance platform it’s definitively not the place to do so. The following graphs show performance evolution during 2015 for the 3 tests we have defined so far.

Note that yellow trace shows data taken from a reference build, so we can discount temporary glitches on the machine running the performance tests of target build, which are shown in the blue trace; this reference trace is also useful to detect invalid regression alerts.

Why performance is so different for these cases ?

The 3 tests we have for Grid Layout use runs/second values as a way to measure performance; this is the preferred method for both WebKit and Blink engines because we can detect regressions with relatively small tests. It’s possible, though, to do other kind of measurements. Looking at the graphs above we can extract the following data:

• auto-sized grid: around 650 runs/sec
• fixed-sized grid: around 1400 runs/sec
• fixed-sized stretched grid: around 1250 runs/sec

Before analyzing possible causes of performance drop for each case, I’ve defined some additional tests to stress even more these 3 cases, so we can realize how grid size affect to the obtained results. I defined 20 tests for these cases, each one with different grid items; from 10×10 up to 200×200 grids. I run those tests in my own laptop, so let’s take the absolute numbers of each case with a grain of salt; although differences between each of these 3 scenarios should be coherent. The table below shows some numeric results of this experiment.

First of all, recall that these 3 tests produce the same web visualization, consisting of grids with NxN items of 100px each one. The only difference is the grid layout strategy used to produce such result: auto-sizing, fixed-sizing and stretching. So now, focusing on previous table’s data we can evaluate the cost, in terms of layout performance, of using auto-sized tracks for defining the grid (which may be the only solution for certain cases). Performance drop is even growing with the number of grid items, but we can conclude that it’s stabilized around 60%. On the other hand stretching is also slower but, unlike auto-sized, in this case performance drop does not show a high dependency of grid size, more or less constant around 15%.

Impact of auto-sized tracks in layout performance

Basically, the track sizing algorithm can be described in the following 4 steps:

• 1- Initialize per Grid track variables.
• 2- Resolve content-based TrackSizingFunctions.
• 3- Grow all Grid tracks in GridTracks from their baseSize up to their growthLimit value until freeSpace is exhausted.
• 4- Grow all Grid tracks having a fraction as the MaxTrackSizingFunction.

These steps will be executed twice, first cycle for determining column tracks’s size and another cycle to set row tracks’s size which it may depend on grid’s width. When using just fixed-sized tracks in the very simple case we are testing, the only computation required to determine grid’s size is completing step 1 and determining free available space based on the specified fixed-size values of each track.

// 1. Initialize per Grid track variables.
for (size_t i = 0; i < tracks.size(); ++i) {
    GridTrack& track = tracks[i];
    GridTrackSize trackSize = gridTrackSize(direction, i);
    const GridLength& minTrackBreadth = trackSize.minTrackBreadth();
    const GridLength& maxTrackBreadth = trackSize.maxTrackBreadth();
 
    track.setBaseSize(computeUsedBreadthOfMinLength(direction, minTrackBreadth));
    track.setGrowthLimit(computeUsedBreadthOfMaxLength(direction, maxTrackBreadth, track.baseSize()));
 
    if (trackSize.isContentSized())
        sizingData.contentSizedTracksIndex.append(i);
    if (trackSize.maxTrackBreadth().isFlex())
        flexibleSizedTracksIndex.append(i);
}
for (const auto& track: tracks) {
    freeSpace -= track.baseSize();
}

Focusing now on the auto-sized scenario, we will have the overhead of resolving content-sized functions for all the grid items.

// 2. Resolve content-based TrackSizingFunctions.
if (!sizingData.contentSizedTracksIndex.isEmpty())
    resolveContentBasedTrackSizingFunctions(direction, sizingData);

I didn’t add source code of resolveContentBasedTrackSizingFunctions because it’s quite complex, but basically it implies a cost proportional to the number of grid tracks (minimum of 2x), in order to determine minContent and maxContent values for each grid item. It might imply additional computation overhead when using spanning items; it would require to sort them based on their spanning value and iterate over them again to resolve their content-sized functions.

Some issues may be interesting to analyze in the future:

• How much each content-sized track costs ?
• What is the impact on performance of using flexible-sized tracks ? Would it be the worst case scenario ? Considering it will require to follow the four steps of track sizing algorithm, it likely will.
• Which are the performance implications of using spanning items ?

Why stretching is so performance drain ?

This is an interesting issue, given that stretch is the default value for both Grid and Flexbox items. Actually, it’s the root cause of why Grid beats Flexbox in terms of layout performance for the cases when stretch alignment is used. As I’ll explain later, Flexbox doesn’t have the optimizations I’ve implemented for Grid Layout.

Stretching logic takes place during the grid container layout operations, after all tracks have their size precisely determined and we have properly computed all grid track’s positions relatively to the grid container. It happens before the alignment logic is executed because stretching may imply changing some grid item’s size, hence they will be marked for layout (if they wasn’t already).

Obviously, stretching only takes place when the corresponding Self Alignment properties (align-self, justify-self) have either auto or stretch as value, but there are other conditions that must be fulfilled to trigger this operation:

• box’s computed width/height (as appropriate to the axis) is auto.
• neither of its margins (in the appropriate axis) are auto
• still respecting the constraints imposed by min-height/min-width/max-height/max-width

In that scenario, stretching logic implies the following operations:

LayoutUnit stretchedLogicalHeight = availableAlignmentSpaceForChildBeforeStretching(gridAreaBreadthForChild, child);
LayoutUnit desiredLogicalHeight = child.constrainLogicalHeightByMinMax(stretchedLogicalHeight, -1);
 
bool childNeedsRelayout = desiredLogicalHeight != child.logicalHeight();
if (childNeedsRelayout || !child.hasOverrideLogicalContentHeight())
    child.setOverrideLogicalContentHeight(desiredLogicalHeight - child.borderAndPaddingLogicalHeight());
if (childNeedsRelayout) {
    child.setLogicalHeight(0);
    child.setNeedsLayout();
}
 
LayoutUnit LayoutGrid::availableAlignmentSpaceForChildBeforeStretching(LayoutUnit gridAreaBreadthForChild, const LayoutBox& child) const
{
    LayoutUnit childMarginLogicalHeight = marginLogicalHeightForChild(child);
 
    // Because we want to avoid multiple layouts, stretching logic might be performed before
    // children are laid out, so we can't use the child cached values. Hence, we need to
    // compute margins in order to determine the available height before stretching.
    if (childMarginLogicalHeight == 0)
        childMarginLogicalHeight = computeMarginLogicalHeightForChild(child);
 
    return gridAreaBreadthForChild - childMarginLogicalHeight;
}

In addition to the extra layout required for changing grid item’s size, computing the available space for stretching adds an additional overhead, overall if we have to compute grid item’s margins because some layout operations are still incomplete.

Given that grid container relies on generic block’s layout operations to determine the stretched width, this specific logic is only executed for determining the stretched height. Hence performance drop is alleviated, compared with the auto-sized tracks scenario.

Grid VS Flexbox layout performance

One of the main goals of CSS Grid Layout specification is to complement Flexbox layout model for 2 dimensions. It’s expectable that creating grid designs with Flexbox will be more inefficient than using a layout model specifically designed for these cases, not only regarding CSS syntax, but also regarding layout performance.

However, I think it’s interesting to measure Grid Layout performance in 1-dimensional cases, usually managed using Flexbox, so we can have comparable scenarios to evaluate both models. In this post I’ll start with such cases, using a very simple one in this occasion. I’d like to get more complex examples in future posts, the ones more usual in Flexbox based designs.

So, let’s consider the following simple test case:

<div class="className">
   <div class="i1">Item 1</div> 
   <div class="i2">Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.</div>
   <div class="i3">Item 3 longer</div>
</div>

I evaluated the simple HTML example above with both Flexbox and Grid layouts to measure performance. I used a CPU profiler to figure out where the bottlenecks are for each model, trying to explain where differences came from. So, I defined 2 CSS classes for each layout model, as follows:

.flex {
    background-color: silver;
    display: flex;
    height: 100px;
    align-items: start;
}
.grid {
    background-color: silver;
    display: grid;
    grid-template-columns: 100px 1fr auto;
    grid-template-rows: 100px;
    align-items: start;
    justify-items: start;
}
.i1 { 
    background-color: cyan;
    flex-basis: 100px; 
}
.i2 { 
    background-color: magenta;
    flex: 1; 
}
.i3 { 
    background-color: yellow; 
}

Given that there is not concept of row in Flexbox, I evaluated performance of 100 up to 2000 grid or flex containers, creating 20 tests to be run inside the chrome performance framework, described at the beginning of this post. You can check out resources and a script to generate them at our github examples repo.

When comparing both layout models targeting layout times, we see clearly that Grid Layout beats Flexbox using the default values for CSS properties controlling layout itself and alignment, which is stretch for these containers. As it was explained before, the stretching logic adds an important computation overhead, which as we can see now in the numeric table above, has more weight for Flexbox than Grid.

Looking at the plot about differences in layout time, we see that for the default case, Grid performance improvement is stabilized around 7%. However, when we avoid the stretching logic, for instance by using any other alignment value, layout performance it’s considerable worse than Flexbox, for this test case, around 15% slower. This is something sensible, as this test case is the idea for Flexbox, while a bit artificial for Grid; using a single Grid with N rows improves performance considerably, getting much better numbers than Flexbox, but we will see these cases in future analysis.

Grid layout better results for the default case (stretch) are explained because I implemented several optimizations for Grid. Probably Flexbox should do the same, as it’s the default value and it could affect many sites using this layout model in their designs.

Thanks to Bloomberg for sponsoring this work, as part of the efforts that Igalia has been doing all these years pursuing a better and more open web.

In my last post I introduced the concept of Content Distribution alignment and how it affects Grid Layout implementation. At that time, it was possible to use all the <content-position> values to select grid tracks position inside a grid container, moving them across the available space. However, it wasn’t until recently that users can distribute grid tracks along such available space, literally adding gaps in between or even stretching them.

In this post I’ll describe how each <content-distribution> value affect tracks in a Grid Layout, their position and size, using different grid structures (eg. number of tracks, span).

Let’s start analyzing the new Content Distribution alignment syntax defined in the CSS Box Alignment specification:

auto | <baseline-position> | <content-distribution> || [ <overflow-position>? && <content-position> ]

In case of a <content-distribution> value can’t be applied, its associated fallback <content-distribution> value should be used instead. However, this CSS syntax allow users to specify a preferred fallback value:

If both a <content-distribution> and <content-position> are given, the <content-position> provides an explicit fallback alignment.

Before going into each value, I think it’s a good idea to refresh the concepts of alignment container and alignment subject and how they apply in the context of Grid Layout:

The alignment container is the grid container’s content box. The alignment subjects are the grid tracks.

The different <content-distribution> values that can be used for align-content and justify-content CSS properties are defined as follows:

• space-between – The alignment subjects are evenly distributed in the alignment container. Default fallback: start.
• space-around – The alignment subjects are evenly distributed in the alignment container, with a half-size space on either end. Default fallback: center.
• space-evenly – The alignment subjects are evenly distributed in the alignment container, with a full-size space on either end. Default fallback: center.
• stretch – Any auto-sized alignment subjects have their size increased equally (not proportionally) so that the combined size exactly fills the alignment container. Default fallback: start.

Picture below describes how these values would behave depending on the number of grid tracks; for simplicity I only use justify-content property, so tracks are distributed along the inline (row) axis. In next examples we will see how both properties work together using more complex grid definitions.

Previous examples were defined with grid items filling grid areas of just 1×1 tracks, which makes distribution pretty simple and easier to predict. But thanks to the flexibility of Grid Layout syntax we can define irregular grids, for instance, using the grid-template-areas property like in the next example.

Since Content Distribution alignment considers grid tracks as the alignment subject, distributing tracks along the available space may have the consequence of modifying the dimensions of grid-areas defined by more than one track. The following picture shows the result of the code above and provides and excellent example of how powerful is the Content Alignment effect on a Grid Layout.

These use cases can be obtained from Igalia’s Grid Layout examples repository, so anybody can play with different grid designs and alignment values combinations. They are also available at our codepen repository.

Grid Layout behind the scene

Now I’d like to explain a bit what I had to implement in the browser’s webcore to get these new features done; just some small pieces of source code, the ones I considered better, to get an idea of what implementing new behavior in browsers implies.

As you might already know because of my previous posts, CSS Box Alignment specification was born to generalize Flexbox’s alignment behavior so that it can be used for grid and even regular blocks. Several new properties were added, like justify-items and justify-self, and CSS syntax has changed considerably. Specially noteworthy how Content Distribution alignment properties have changed from their initial Flexbox definition. They now support complex values like ‘space-between true’, ‘space-around start’, or even ‘stretch center safe’. This makes possible to express more info than using the previous simple keyword form, although it requires a new CSS parsing logic in Browsers.

More complex CSS parsing

Since both align-content and justify-content properties accept multiple optional keywords I needed to re-implement completely their parsing logic. I’m happy to announce that it recently landed WebKit’s trunk too, so now both web engines support the new CSS syntax.

Due to the complex values defined for theses CSS properties, a new CSSValue derived class was defined to hold all the Content Alignment data, named as CSSContentDistributionValue. This data is then converted to something meaningful for the style logic using the StyleBuilderConverter class. This is the preferred method in both WebKit and Blink engines, which it just needs to be declared in the CSSPropertyNames.in and CSSProperties.in template files, respectively.

align-content initial=initialContentAlignment, converter=convertContentAlignmentData
justify-content initial=initialContentAlignment, converter=convertContentAlignmentData

The StyleBuildConverter logic is pretty simple thanks to these 2 new data structures, as it can be appreciated in the following excerpt of source code:

StyleContentAlignmentData StyleBuilderConverter::convertContentAlignmentData(StyleResolverState&, CSSValue* value)
{
    StyleContentAlignmentData alignmentData = ComputedStyle::initialContentAlignment();
    CSSContentDistributionValue* contentValue = toCSSContentDistributionValue(value);
    if (contentValue->distribution()->getValueID() != CSSValueInvalid)
        alignmentData.setDistribution(*contentValue->;distribution());
    if (contentValue->position()->getValueID() != CSSValueInvalid)
        alignmentData.setPosition(*contentValue->;position());
    if (contentValue->overflow()->getValueID() != CSSValueInvalid)
        alignmentData.setOverflow(*contentValue->overflow());
    return alignmentData;
}

The StyleContentAlignmentData class was defined to simplify how we manage these complex values, so that we can handle properties as they had an atomic value. This approach allows a more efficient and robust way of detecting and managing style changes in these properties.

New Layout operations

Once this new CSS syntax is correctly parsed and a LayoutStyle instance generated according to user defined CSS style rules, I needed to modify Flexbox’s layout code for adapting it to the new data structures, ensuring browser backward compatibility and passing all the Layout and Unit tests. I implemented from scratch this logic for Grid Layout so I had the opportunity to introduce several performance optimizations to avoid unnecessary layouts and repaints. This area is pretty interesting and I’ll talk about it soon in a new post.

One interesting aspect of Content Distribution alignment is that it might take part in the track sizing algorithm. As it was explained in my previous post about Self Alignment, stretch value increases alignment subject’s size to fill its alignment container’s available space. This is also the case of Content Alignment, but considering tracks as alignment subject. However, there is another case not so obvious where <content-distribution> values may influence in track sizing resolution, or perhaps better said, grid area sizing.

Let’s consider this example of grid where there are certain areas using more than one track:

grid-template-areas: "a a b"
                     "c d b"
grid-auto-columns: 20px;
grid-auto-rows: 40px;
width: 150px;
height: 300px;

The example above defines a grid with 3 column tracks of 20px and 2 row tracks of 40px, which would be laid out as it’s shown in the following diagram:

This fact has interesting implementation implications due to the fact that in certain cases, in order to determine grid item’s logical height we need its logical width to be resolved first. Track sizing algorithm uses children grid area size to determine grid cell’s logical height, hence given that alignment logic needs track sizes have been already resolved, it may imply a re-layout of the grid items which size could be affected by the used content-distribution value. The following source code shows how I handle this scenario:

LayoutUnit LayoutGrid::gridAreaBreadthForChild(const LayoutBox& child, GridTrackSizingDirection direction, const Vector& tracks) const
{
    const GridCoordinate& coordinate = cachedGridCoordinate(child);
    const GridSpan& span = (direction == ForColumns) ? coordinate.columns : coordinate.rows;
    const Vector& trackPositions = (direction == ForColumns) ? m_columnPositions : m_rowPositions;
    if (span.resolvedFinalPosition.toInt() < trackPositions.size()) {
        LayoutUnit startOftrack = trackPositions[span.resolvedInitialPosition.toInt()];
        LayoutUnit endOfTrack = trackPositions[span.resolvedFinalPosition.toInt()];
        return endOfTrack - startOftrack + tracks[span.resolvedFinalPosition.toInt()].baseSize();
    }
    LayoutUnit gridAreaBreadth = 0;
    for (GridSpan::iterator trackPosition = span.begin(); trackPosition != span.end(); ++trackPosition)
        gridAreaBreadth += tracks[trackPosition.toInt()].baseSize();
 
    return gridAreaBreadth;

The code above will return different results, in the cases mentioned before, depending on whether it’s run during track sizing alignment or after applying the alignment logic. This will likely make needed a new layout of the whole grid, or at least, the affected grid items, which it likely has a negative impact on performance.

Current status and next steps

I’d like to finish this post with a snapshot of current situation and challenges for the next months, as I’ve been regularly doing in my last posts.

Unlike last reports, this time I’ve got good news regarding reduction of implementation gaps between the two web engines we are focusing our efforts on, WebKit and Blink. The following table describes current situation:

The table above indicates that several milestones were reached since the last report, although there are still some pending issues:

• I’ve completed the implementation in WebKit of the parsing logic for the new Box Alignment properties: align-items and align-self.
• As a side effect, I’ve also upgraded the ones already present because of Flexbox to the latest CSS3 Box Alignment specification.
• WebKit has now full support for Default and Self Alignment fro Grid Layout, including also overflow handling
• Blink has now full support for Content Distribution alignment, which missed <content-distrbuton> values.
• WebKit’s Grid Layout implementation still misses support for Content Distribution alignment.
• Baseline Alignment is still missing in both web engines

In addition to the above mentioned pending issues, our roadmap include the following tasks as part of my todo list for the next months:

• Even though there s support for different writing-modes and flow directions, there are still some issues with orthogonal flows. I’ve got already some promising patches but they still have to be reviewed by Blink and WebKit engineers.
• Optimizations of style and repaint invalidations triggered by changes on the alignment properties. As commented before, this is a very interesting topic involving, which I’ll elaborate further in next posts.
• Performance analysis of relevant Grid Layout use cases, which hopefully will lead to optimizations proposals.

All this work and many other contributions to Grid Layout for WebKit and Blink web engines are the result of the collaboration between Bloomberg and Igalia to implement this W3C specification.