Ever since some initial discussions with Raph Levien (author of
font-rs) at RustConf last
September, I've been thinking about ways to improve vector graphics rendering using modern graphics
hardware, specifically for fonts. These ideas began to come together in December, and over the past
couple of months I've been working on actually putting them into a real, usable library. They've
proved promising, and now I have some results to show.
Today I'm pleased to announce Pathfinder, a Rust library for OpenType font rendering. The goal is nothing less than to be the fastest vector graphics renderer in existence, and the results so far are extremely encouraging. Not only is it very fast according to the traditional metric of raw rasterization performance, it's practical, featuring very low setup time (end-to-end time superior to the best CPU rasterizers), best-in-class rasterization performance even at small glyph sizes, minimal memory consumption (both on CPU and GPU), compatibility with existing font formats, portability to most graphics hardware manufactured in the past five years (DirectX 10 level), and security/safety.
To illustrate what it means to be both practical and fast, consider these two graphs:
(Click each graph for a larger version.)
The first graph is a comparison of Pathfinder with other rasterization algorithms with all vectors already prepared for rendering (and uploaded to the GPU, in the case of the GPU algorithms). The second graph is the total time taken to prepare and rasterize a glyph at a typical size, measured from the point right after loading the OTF file in memory to the completion of rasterization. Lower numbers are better. All times were measured on a Haswell Intel Iris Pro (mid-2015 MacBook Pro).
From these graphs, we can see two major problems with existing GPU-based approaches:
Many algorithms aren't that fast, especially at small sizes. Algorithms aren't fast just because they run on the GPU! In general, we want rendering on the GPU to be faster than rendering on the CPU; that's often easier said than done, because modern CPUs are surprisingly speedy. (Note that, even if the GPU is somewhat slower at a task than the CPU, it may be a win for CPU-bound apps to offload some work; however, this makes the use of the algorithm highly situational.) It's much better to have an algorithm that actually beats the CPU.
Long setup times can easily eliminate the speedup of algorithms in practice. This is known as the "end-to-end" time, and real-world applications must carefully pay attention to it. One of the most common use cases for a font rasterizer is to open a font file, rasterize a character set from it (Latin-1, say) at one pixel size for later use, and throw away the file. With Web fonts now commonplace, this use case becomes even more important, because Web fonts are frequently rasterized once and then thrown away as the user navigates to a new page. Long setup times, whether the result of tessellation or more exotic approaches, are real problems for these scenarios, since what the user cares about is the document appearing quickly. Faster rasterization doesn't help if it regresses that metric.
(Of the two problems mentioned above, the second is often totally ignored in the copious literature on GPU-based vector rasterization. I'd like to see researchers start to pay attention to it. In most scenarios, we don't have the luxury of inventing our own GPU-friendly vector format. We're not going to get the world to move away from OpenType and SVG.)
Vector drawing basics
In order to understand the details of the algorithm, some basic knowledge of vector graphics is necessary. Feel free to skip this section if you're already familiar with Bézier curves and fill rules.
OpenType fonts are defined in terms of resolution-independent Bézier curves. TrueType outlines contain lines and quadratic Béziers only, while OpenType outlines can contain lines, quadratic Béziers, and cubic Béziers. (Right now, Pathfinder only supports quadratic Béziers, but extending the algorithm to support cubic Béziers should be straightforward.)
In order to fill vector paths, we need a fill rule. A fill rule is essentially a test that determines, for every point, whether that point is inside or outside the curve (and therefore whether it should be filled in). OpenType's fill rule is the winding rule, which can be expressed as follows:
Pick a point that we want to determine the color of. Call it P.
Choose any point outside the curve. (This is easy to determine since any point outside the bounding box of the curve is trivially outside the curve.) Call it Q.
Let the winding number be 0.
Trace a straight line from Q to P. Every time we cross a curve going clockwise, add 1 to the winding number. Every time we cross a curve going counterclockwise, subtract 1 from the winding number.
The point is inside the curve (and so should be filled) if and only if the winding number is not zero.
How it works, conceptually
The basic algorithm that Pathfinder uses is the by-now-standard trapezoidal pixel coverage
algorithm pioneered by Raph Levien's libart (to the best of my knowledge). Variations of it are
used in FreeType,
stb_truetype version 2.0 and up, and
font-rs. These implementations
differ as to whether they use sparse or dense representations for the coverage buffer. Following
font-rs, and unlike FreeType and
stb_truetype, Pathfinder uses a dense representation for
coverage. As a result, Raph's description of the algorithm applies fairly well to Pathfinder as
There are two phases to the algorithm: drawing and accumulation. During the draw phase, Pathfinder computes coverage deltas for every pixel touching (or immediately below) each curve. During the accumulation phase, the algorithm sweeps down each column of pixels, computing winding numbers (fractional winding numbers, since we're antialiasing) and filling pixels appropriately.
The most important concept to understand is that of the coverage delta. When drawing high-quality
antialiased curves, we care not only about whether each pixel is inside or outside the curve but
also how much of the pixel is inside or outside the curve. We treat each pixel that a curve
passes through as a small square and compute how much of the square the curve occupies. Because we
break down curves into small lines before rasterizing them, these coverage areas are always
trapezoids or triangles, and so and so we can use trapezoidal area expressions to calculate them.
The exact formulas involved are somewhat messy and involve several special cases; see Sean
Barrett's description of the
stb_truetype algorithm for the details.
Rasterizers that calculate coverage in this way differ in whether they calculate winding numbers
and fill at the same time they calculate coverage or whether they fill in a separate step after
coverage calculation. Sparse implementations like FreeType and
stb_truetype usually fill as they
go, while dense implementations like
font-rs and Pathfinder fill in a separate step. Filling in a
separate step is attractive because it can be simplified to a prefix sum over each pixel column
if we store the coverage for each pixel as the difference between the coverage of the pixel and
the coverage of the pixel above it. In other words, instead of determining the area of each pixel
that a curve covers, for each pixel we determine how much additional area the curve covers,
relative to the coverage area of the immediately preceding pixel.
This modification has the very attractive property that all coverage deltas both inside and outside the curve are zero, since points completely inside a curve contribute no additional area (except for the first pixel completely inside the curve). This property is key to Pathfinder's performance relative to most vector texture algorithms. Calculating exact area coverage is slow, but calculating coverage deltas instead of absolute coverage essentially allows us to limit the expensive calculations to the edges of the curve, reducing the amount of work the GPU has to do to a fraction of what it would have to do otherwise.
In order to fill the outline and generate the final glyph image, we simply have to sweep down each
column of pixels, calculating the running total of area coverage and writing pixels according to
the winding rule. The formula to determine the color of each pixel is simple and fast:
min(|coverage total so far|, 1.0) (where 0.0 is a fully transparent pixel, 1.0 is a fully opaque
pixel, and values in between are different shades). Importantly, all columns are completely
independent and can be calculated in parallel.
With the advanced features in OpenGL 4.3, this algorithm can be straightforwardly adapted to the GPU.
As an initialization step, we create a coverage buffer to hold delta coverage values. This coverage buffer is a single-channel floating-point framebuffer. We always draw to the framebuffer with blending enabled (
GL_FUNC_ADD, both source and destination factors set to
We expand the TrueType outlines from the variable-length compressed
glyfformat inside the font to a fixed-length, but still compact, representation. This is necessary to be able to operate on vertices in parallel, since variable-length formats are inherently sequential. These outlines are then uploaded to the GPU.
Next, we draw each curve as a patch. In a tessellation-enabled drawing pipeline like the one that Pathfinder uses, rather than submitting triangles directly to the GPU, we submit abstract patches which are converted into triangles in hardware. We use indexed drawing (
glDrawElements) to take advantage of the GPU's post-transform cache, since most vertices belong to two curves.
For each path segment that represents a Bézier curve, we tessellate the Bézier curve into a series of small lines on the GPU. Then we expand all lines out to screen-aligned quads encompassing their bounding boxes. (There is a complication here when these quads overlap; we may have to generate extra one-pixel-wide quads here and strip them out with backface culling. See the comments inside the tessellation control shader for details.)
In the fragment shader, we calculate trapezoidal coverage area for each pixel and write it to the coverage buffer. This completes the draw step.
To perform the accumulation step, we attach the coverage buffer and the destination texture to images. We then dispatch a simple compute shader with one invocation per pixel column. For each row, the shader reads from the coverage buffer and writes the total coverage so far to the destination texture. The
min(|coverage total so far, 1.0)expression above need not be computed explicitly, because our unsigned normalized atlas texture stores colors in this way automatically.
The performance characteristics of this approach are excellent. No CPU preprocessing is needed other than the conversion of the variable-length TrueType outline to a fixed-length format. The number of draw calls is minimal—any number of glyphs can be rasterized in one draw call, even from different fonts—and the depth and stencil buffers remain unused. Because the tessellation is performed on the fly instead of on the CPU, the amount of data uploaded to the GPU is minimal. Area coverage is essentially only calculated for pixels on the edges of the outlines, avoiding expensive fragment shader invocations for all the pixels inside each glyph. The final accumulation step has ideal characteristics for GPU compute, since branch divergence is nonexistent and cache locality is maximized. All pixels in the final buffer are only painted at most once, regardless of the number of curves present.
For any GPU code designed to be shipping to consumers, especially OpenGL 3.0 and up, compatibility and portability are always concerns. As Pathfinder is designed for OpenGL 4.3, released in 2012, it is no exception. Fortunately, the algorithm can be adapted in various ways depending on the available functionality.
When compute shaders are not available (OpenGL 4.2 or lower), Pathfinder uses OpenCL 1.2 instead. This is the case on the Mac, since Apple has not implemented any OpenGL features newer than OpenGL 4.2 (2011). The compute-shader crate abstracts over the subset of OpenGL and OpenCL necessary to access GPU compute functionality.
When tessellation shaders are not available (OpenGL 3.3 or lower), Pathfinder uses geometry shaders, available in OpenGL 3.2 and up.
(Note that it should be possible to avoid both geometry shaders and tessellation shaders, at the cost of performing that work on the CPU. This turns out to be quite fast. However, since image load/store is a hard requirement, this seems pointless: both image load/store and geometry shaders were introduced in DirectX 10-level hardware.)
Although these system requirements may seem high at first, the integrated graphics found in any Intel Sandy Bridge (2011) CPU or later meet them.
The immediate next step for Pathfinder is to integrate into WebRender as an optional accelerated path for applicable fonts on supported GPUs. Beyond that, there are several features that could be added to extend Pathfinder itself.
Support vector graphics outside the font setting. As Pathfinder is a generic vector graphics rasterizer, it would be interesting to expose an API allowing it to be used as the backend for e.g. an SVG renderer. Rendering the entire SVG specification is outside of the scope of Pathfinder itself, but it could certainly be the path rendering component of a full SVG renderer.
Support CFF and CFF2 outlines. These have been seen more and more over time, e.g. in Apple's new San Francisco font. Adding this support involves both parsing and extracting the CFF2 format and adding support for cubic Bézier curves to Pathfinder.
Support WOFF and WOFF2. In the case of WOFF2, this involves writing a parser for the transformed
Support subpixel antialiasing. This should be straightforward.
Support emoji. The Microsoft
sbixextensions are straightforward, but the Google
SVGtable allows arbitrary SVGs to be embedded into a font. Full support for SVG is probably out of scope of Pathfinder, but perhaps the subset used in practice is small enough to support.
Optimize overlapping paths. It would be desirable to avoid antialiasing edges that are covered by other paths. The fill rule makes this trickier than it initially sounds.
Support hinting. This is low-priority since it's effectively obsolete with high-quality antialiasing, subpixel AA, and high-density displays, but it might be useful to match the system rendering on Windows.
Pathfinder is available on GitHub and should be easily buildable using the stable version of Rust and Cargo. Please feel free to check it out, build it, and report bugs! I'm especially interested in reports of poor performance, crashes, or rendering problems on a variety of hardware. As Pathfinder does use DirectX 10-level hardware features, some amount of driver pain is unavoidable. I'd like to shake these problems out as soon as possible.
Finally, I'd like to extend a special thanks to Raph Levien for many fruitful discussions and ideas. This project wouldn't have been possible without his insight and expertise.