- Graphics Principles
- Graphics Elements
- Under the Hood
- Where Are We?
This chapter is from the book
This chapter is from the book
This chapter is from the book
Under the Hood
Previous sections have discussed the graphics principles and the graphics API elements. This section goes "under the hood" to describe how Silverlight draws XAML content and displays it in the browser window. Understanding this process will help you understand the Silverlight runtime performance characteristics. Furthermore, you will understand the problems solved by the runtime and the problems your application must solve.
In particular, this section discusses the following:
- The draw loop process that takes changes to the graph of objects and draws it to an off screen back buffer
- The rasterization process that converts vector graphics primitives to pixels in an offscreen back buffer
- Performance optimizations such as incremental redraw, occlusion culling, and multi-core
- How the offscreen back buffer gets displayed in the browser window
Draw Loop
Silverlight draws at a regular timer interval set by the MaxFrameRate property. On each tick of the timer, Silverlight does the following:
- Checks for any changes to the properties of our graph of Canvas and Shape elements. If no changes exist, Silverlight does no further work for this timer tick.
- Performs any pending layout operations. Chapter 7, "Layout," will discuss these layout operations further.
- Gathers rendering changes and prepares to rasterize them.
- Incrementally rasterizes the changes for the current timer tick. The graphics state at the current timer tick is the current frame.
- Notifies the browser that a frame (or an incremental delta to an existing frame) is complete for display.
Rasterization
After the draw loop has identified which elements need to be redrawn, Silverlight converts those elements to a set of pixels in our offscreen back buffer. The previous discussion of shapes described how to specify path outlines and a method of specifying the inside and the outside of the shape. However, the geometry describes an abstract infinite resolution outline of a shape and a screen has a finite number of pixels to color. Rasterization is the process of converting from a path outline to discrete pixels. This section describes how rasterization is accomplished.
The simplest method to convert geometry to pixels is a process called sampling. The sampling process uses a discrete number of sample points to convert from the infinite shape description to pixels. For example, consider the simple sample pattern consisting of a uniform grid of sample points with one sample point per pixel. If the sample point is contained within the geometry, light up the pixel. If the sample point is not contained within the geometry, do not light the pixel. For example, the circle specified by the following XAML would light the pixels shown in Figure 3.33.
)
Figure 3.33 Sampling a circle
<Canvas xmlns="http://schemas.microsoft.com/client/2007">
<Ellipse
Fill="Black"
Width="15"
Height="15"
/>
</Canvas>
You may have noticed that the integer coordinates were located at the top left of the pixel and the sample points were in the center of a pixel. This convention enables a symmetrical curved surface specified on integer coordinates to produce a symmetrical rasterization. If the sample points were on integer coordinates instead, the ellipse would lose symmetry as shown in Figure 3.34.
)
Figure 3.34 Sampling a circle with integer sample point coordinates
The rasterization shown in Figure 3.33 appears to have jagged edges. This jagged appearance is the consequence of aliasing. Aliasing is the loss of information that results from converting from a continuous curve to a discrete set of samples. Anti-aliasing is a term that refers to a technique that attempts to minimize aliasing artifacts.
The Silverlight anti-aliasing technique consists of sampling multiple times per pixel and applying a box filter to produce the final pixel color. Silverlight conceptually samples 64 times per pixel as shown in Figure 3.35. The box filter averages the contribution of all samples within a rectangle bordering the pixel to produce a final pixel color. If some partial number of samples is in the box, Silverlight applies transparency to blend smoothly with what is underneath the geometry as shown in Figure 3.36. This anti-aliasing technique produces a smooth transition from inside the shape to outside the shape along edges.
)
Figure 3.35 Anti-aliasing sampling pattern
)
Figure 3.36 Anti-aliased rasterization
One artifact of anti-aliasing is a visible seam that sometimes results from drawing two adjacent shapes. For example, the following two rectangles that meet in the middle of a pixel generate a seam:
<Canvas xmlns="http://schemas.microsoft.com/client/2007">
<Rectangle
Fill="Black"
Width="100.5"
Height="100.5"
/>
<Rectangle
Fill="Black"
Canvas.Left="100.5"
Width="100.5"
Height="100.5"
/>
</Canvas>
The previous XAML results in the rasterization shown in Figure 3.37. Notice the gap between the two rectangles. The rectangles joined perfectly in the input XAML, so why is there a seam in the rendered result?
)
Figure 3.37 Anti-aliasing seam example
These seams are a result of the rasterization rules described in this section. Consider the rasterization process applied to a light gray pixel X shown in Figure 3.37. Rectangle A is covering exactly half the samples for pixel X. Silverlight consequently draws that pixel of Rectangle A with 0.5 anti-aliasing alpha. Alpha is a term that refers to the transparency used to blend colors with a formula such as
Color_destination = alpha * Color_source + (1 – alpha) * Color_destination
In our example, alpha = 0.5, Color_source = Black, and Color_destination = White. Blending along the edge of rectangle A results in a destination color of
0.5 * Black + (1 – 0.5) * White = 0.5 * White
Rectangle B also covers half its sample points. Silverlight also blends pixel X of rectangle B with alpha = 0.5 to a background color of 0.5 * White. Consequently, the resulting color is
0.5 * Black + (1 – 0.5) * (0.5 White) = 0.25 White.
The final pixel color has one quarter of the background color showing through as a visible seam.
To avoid these seams, you should snap edges to pixel boundaries as shown in Figure 3.38. Snapping also produces a sharper edge between the two shapes. However, pixel snapping only removes seams if you align the shapes edges with the x-axis or the y-axis. For the rotated edges shown in Figure 3.39, snapping does not remove the artifact. For rotated edges, the common technique to avoid this seam is to overlap the edges so that the background is no longer visible.
)
Figure 3.38 Pixel snapped rasterization
)
Figure 3.39 Seams with a rotated edge
Bilinear Filtering
The previous section discussed how Silverlight converts an arbitrary geometry to a set of pixels to fill. Silverlight then colors the filled pixels based on the brush specified. This process is straightforward for solid color brushes and gradient brushes. However, with image brushes, Silverlight must map from the destination pixels to the original image data, which may be at a different resolution. This section describes the mapping function used to achieve the image data stretch shown in Figure 3.40.
)
Figure 3.40 Image with bilinear filtering
Nearest neighbor is a simple image scaling function that transforms the destination pixel to an image bitmap position and picks the nearest pixel color. Nearest neighbor sampling generates ugly aliasing artifacts when the image is displayed with a scale or rotation as shown in Figure 3.41. You will notice jagged lines if you look at the picture closely.
)
Figure 3.41 Image with nearest neighbor
Silverlight uses nearest neighbor sampling in the special case where the brush image data maps exactly onto centers of pixels. For rotated, scaled, or non-integer translated images, bilinear filtering is used to produce the result shown in Figure 3.40.
Bilinear filtering maps the screen position to a position (u,v) in image space. The bilinear filtering process then interpolates a color from pixels (floor(u), floor(v)), (floor(u) + 1, floor(v)), (floor(u), floor(v) + 1), and (floor(u) + 1, floor(v)+1). Figure 3.42 illustrates this process. Bilinear filtering generates good results for scales that are within a factor of two of the original image size. Figure 3.43 demonstrates the results of scaling an image in two sizes within reasonable limits.
)
Figure 3.42 The bilinear filtering process
)
Figure 3.43 Image scaling within good limits
With bilinear filtering, if you scale up an image significantly, it becomes blurry. Conversely, if you scale down an image significantly, it looks aliased. Figure 3.44 shows examples of both these artifacts.
)
Figure 3.44 Image scaling extremes
Incremental Redraw
In addition to drawing static objects for a single frame, Silverlight must constantly redraw objects as they are changing. If an object moves from one position to another, it would be wasteful to redraw all the pixels on the screen. Instead, Silverlight marks the old position as needing a redraw and marks the new position as also needing a redraw. To simplify this marking algorithm, Silverlight uses the bounding box of a shape instead of the tight shape bounds.
For example, suppose the shape shown in the following XAML moves from position 0,0 to position 100,100. Figure 3.45 shows the area that is redrawn.
)
Figure 3.45 Incremental redraw regions
<Canvas xmlns="http://schemas.microsoft.com/client/2007">
<Ellipse
Fill="Black"
Width="100"
Height="100"
/>
</Canvas>
To view a visualization of these incremental redraw regions in an application, use the following JavaScript:
function loadHandler()
{
plugin = document.getElementById("MySilverlightPlugin");
plugin.settings.EnableRedrawRegions = true;
}
This visualization blends a transparent color on top of any content drawn and cycles to a different color each frame. Consequently, any content that is flashing represents content that Silverlight is constantly redrawing. Any content that stabilizes on a single color has not changed for several frames.
Occlusion Culling
The most expensive operation in the draw loop is the rasterization process, which writes each of the destination pixels. For example, a full screen animation can consist of processing several hundred million pixels per second. Each of these pixels applies at least one brush operation. If there are overlapping brushes, the computational requirements can multiply by a factor of 3 to 10.
As the graph of elements gets more complicated, it may no longer render at the desired frame rate. To optimize the rasterization process, Silverlight avoids brush operations for completely occluded brush pixels. For example, if you draw a full screen background and an almost full screen image, Silverlight computes all the image pixels and only those background pixels not covered by the image itself. For complicated graphs of elements, this optimization can produce a 3–10x speedup.
Multi-Core Rendering
Silverlight takes advantage of multiple CPU cores to produce faster rendering throughput. In particular, Silverlight subdivides a frame into a set of horizontal bands and distributes the rasterization of those bands across CPU cores as shown in Figure 3.46. Currently, only the frame rasterization step, effects, mip-map generation, and media operations run in parallel across CPU cores. Systems such as layout, control templating, application user code, and animation all run on a single thread. Consequently, you can determine if your application is rasterization bound by simply setting your framerate to 10000 frames per second and measuring your CPU usage percentage. If you achieve almost 100% CPU usage on a dual core machine, you are almost entirely rasterization bound. If you achieve 70% CPU usage on a dual core machine (at 10000 frames per second), that means 30% of the work is not running in parallel.
)
Figure 3.46 Dividing a scene for multi-core rendering
How Content Gets to the Screen
As previously discussed, the draw loop first draws a frame to an offscreen back buffer and then it notifies the browser that the frame is ready for display. With windowless=false mode, Silverlight content goes to the screen without browser intervention on most operating systems. With windowless=true, the browser copies the offscreen frame to its display area. This extra step is both slow and can result in visual tearing effects in a number of browsers. The worst mode of operation is when windowless=true is specified with a transparent color for the background of the control. The transparent color causes the Web browser to redraw the content underneath the control each time any control content has changed.