ytsedan
commented
7 years ago Let me share some notes and observations on performance, as these kind of questions pop up here regularly.
Display list are still a thing if your performance problems are CPU bound. They will help if you render the same shape over multiple frames and you want to only apply affine transforms to them (Note: heavy scaling will be noticeable, as edges become blurry due to scaling up the aa fringe, and you might see the tessellation error). However, if you have lots of dynamic shapes that change each frame, display lists will not help at all as you still need to tesselate the shape again and again. On a side note, if you want to render the same static shape multiple time in the same frame (as you often do in 2d tile-based games), the shape will be transferred multiple times to the GPU which is costly, too. For this use case it’s a good idea to render the shape to a bitmap and then draw the bitmap multiple times. You can check out my attempt on display lists in my fork of this project.
Regarding lower performance devices, I did some optimizations to improve rendering performance on older iOS devices (iPad 3 generation). These devices especially have poor pixel shader performance; not sure what the problem is with raspberries, though. Here are some things you might try to improve GPU performance
1) Spit the Übershader into multiple smaller shaders. nanovg uses a single shader which has multiple branches for different tasks, splitting that into separated shaders and switch shaders between draw calls helped a lot for me.
2) Scissoring. nanovg does scissoring right in the pixel shader, the advantage is, that scissor rects do not have to be axis-aligned. However, removing this code from the shader and just using glScissor did help performance in my use-case (drawback: you have to get along with axis-aligned clipping, which was fine for me).
3) In my app I’m drawing mainly axis-aligned rectangle (for a UI system), for this the whole nanovg tessellation is kind of overkill. So I added support for drawing simple rectangles, that will just boil down to two triangles (no anti-aliasing) and a solid color (or textured) pixel shader. However, nanovg will still create a draw call for each of these rectangles - so there is room for improvement here.
A new project on GitHub appeared, that addresses these issues by basically re-implementing most of nanovg. It’s not finished yet, but looks very promising and you might want to keep an eye on it. https://github.com/jdryg/vg-renderer
natinusala
olliwang
lieff
SpinyOwl
Walther
memononen
mgood7123
I've been trying to run nanovg on a Raspberry Pi 3 B to see how it performs (GLES2). Well, it's very slow (<10fps). I made a profile of the example : https://pastebin.com/rY3LfLAF
You can easily see that almost 50% of all the calls are from drawing and calculation functions. I already implemented the AFD tesselation function from issue #328 with no great improvement.
This leads me to the display lists, I saw that multiple issues were talking about them, is it still a thing ? I believe it would fix the performance issues of embedded devices such as Android or the Rpi.