Hacker News Reader

Comment by adamnemecek | Hacker News Reader

2d graphics are kinda hard. The problem is that gpus don't support them out of the box so you have to do a Lot yourself.

View on HN

I assume the point is that you have to implement a pure software rasterizer. It's not actually that difficult, but it feels like a lot for an undergrad who's not familiar with the space. I assume this is done as part of a class that gives you context.

For an example of a rasterizer, you can take a look at my pure JS implementation of canvas (which is roughly the same imaging model as SVG). All lines and shapes are flattened into a pure line polygon, then drawn with a scanline rasterizer. Getting the basics working is fairly easy. Handling the endless edge cases, however.... :)

https://github.com/joshmarinacci/node-pureimage

It's get somewhat tricky when you want things like gpu support, text rendering etc.

You have to do things like stencil texture to speed things up etc etc.

GPUs aren't really designed for 2D graphics out of the box. I've written about this before [0].

Text rendering can be tricky, but not that much trickier -- it's just the same curves at smaller scales.

Not sure what you mean by stencil textures. Are you talking about the NV_path_rendering approach where you stencil out the path? Yeah, that's not really a thing that people do these days.

[0] https://blog.mecheye.net/2019/05/why-is-2d-graphics-is-harde...

Yeah I'm talking about that. What do people use then?

Fast, efficient 2D graphics using the PostScript (i.e. SVG) rendering model is still an open research problem. Most shipping implementations these days do full tessellation to triangles on CPU.

> it's just the same curves at smaller scales.

It's not that simple unfortunately. Hinting for TrueType fonts is a thing and without properly aligning resulting pixels at their own scale fonts tend to look very ugly when following pure curve definitions. Oversampling at e.g. 8x size is one possible option but then there are other problems that pop up once downscaling to display size.

This is a beginner CS course, it's a lot simpler than that.

Yeah, writing a scanline rasterizer from scratch is fun. For beginners, I advocate the sample-based approach I've written about, as functional "is inside shape" can be easier to reason about for simple cases. It's harder for non-convex polygons though :)

https://magcius.github.io/xplain/article/rast1.html

The edge cases are similar -- precision in computing the path intersections and winding curves.

Making a 2D rasterizer is fairly trivial, making a rasterizer that looks good is insanely difficult. Handling things like aliasing without overblowing and making it work in all possible cases is tricky and one spends 99% of time addressing those tiny little details that "make one fall in love with software".

Of course GPUs support 2D. It is just 3D with one less dimension.

What do you mean GPUs don't support them? Most opengl tutorials I've read have people building triangles in 2d to learn fragment shaders and vertices before going 3d. Most of them also support parametric curves, so even hardware accelerated bezier curves should be possible.

Wait, why do you say GPUs support parametric curves? GPUs are based on triangles mostly.

GPUs can take in way more than just triangles as input. There are particle simulation and even Ray tracing implemented on GPUs nowadays. Support for parametric curves was one of the more recent additions.

That's a very... naive view of how a GPU works. Particle simulation is done in compute, and ray tracing, as implemented in RTX/DXR, is done on a soup of triangles. The core of rasterization is still done on triangles, and can't easily be done in compute. Have any references to parametric curves on GPUs? All the approaches I know of, like the recent mesh shader work, still output triangle meshes.

"In compute" just means that there aren't inputs and outputs related to the current output frame in the calculation job - the computational possibilities are the same.

Triangles aren't necessarily involved in in rendering either, see eg how the stuff on shadertoy.com works.

This is from 2007: https://developer.nvidia.com/gpugems/gpugems3/part-iv-image-...

Again, the vast majority of use cases for GPUs is 3d vertex graphics. But they're capable of more than that. Modern GPUs are very different from early GPUs that only worked with triangles. Some of the early ones were actually ASICs, and couldn't even load different shader programs.

GPU aren't based on anything. They are turing complete, you can run Linux on it if you wanted to.

What you are saying is outdated by maybe a couple of decades.

A Turing Machine is turing-complete, but it would be inefficient to run Linux on it. We're not talking about raw computability here, but feasibility. And still, I am not aware of anything "running Linux on a GPU", their scheduling engines are not designed for those sorts of workloads.

See eg https://github.com/mattdesl/parametric-curves/ and https://developer.nvidia.com/gpugems/gpugems3/part-iv-image-...

You can builds all kinds on top of pixels and triangles.

That's like saying "GPUs support Master Chief". You can model Master Chief with triangles, and you can model parametric curves with triangles. But I wouldn't call it "supporting parametric curves", you're still rasterizing triangles, they're just morphed into the shape of a curve. And most practical, shipping versions of this technique would do adaptive triangulation on the CPU, since otherwise you don't have an idea of your mesh density and are either over-submitting or under-submitting triangles.

Loop-Blinn, similarly, is mostly a CPU-side approach and has a lot of drawbacks, but at it's core it's using the pixel shader to define a curve profile.

The parametric curves don't get transformed into triangles. This isn't tessellation or similar techniques. You aren't feeding the GPU any triangles - you're only feeding it the function that defines the parametric curve. Then, using that function for the parametric curve the GPU can calculate pixel output. Again, modern GPUs (really, most GPUs in the past decade) can support more than just triangles. These more exotic techniques don't get as much attention, since most graphics assets are still implemented with triangle meshes.

The first triangle is still in 3d space but the z coordinates are set to 0

Or you can just not define a z coordinate in the vertex buffer object (at least, in opengl). I think the vertex shader might still need to output a 3rd coordinate. But you can always just discard it.

Reply on news.ycombinator.com