Speeding up Rigging Code
So our game NBA 2K18 comes out Friday so the art team has been pretty relaxed for the past couple of weeks. With this down time I’ve finally had a moment to handle our largest complaint from the character team, Speed.
The main complaints about speed were that our rigging code took too long and our tools felt unresponsive .
I started by running cProfile on our rigging code to find any bottlenecks. The starting profile showed that it was talking about 96 seconds to rig my test character. It also showed that we were spending a lot of time doing file I/O and making blendshapes.
I noticed that we were reading in a bunch of files, moving them all by some amount, then saving them out to use later. I saved about 10 seconds by saving out the transform values in a text file then applying it right before we need to use each object in the rig. There was some cleanup code that was run on each of those intermediate files that need to be modified to run at import time instead of on a clean file, but that was fairly straight forward.
We make tons of blendshapes by wrap deforming things to the main geo. I saved 5 seconds by not making blendshapes were we didn’t need to. So no right side shapes on meshes that are only affected by the left side.
I tried checking about 10 percent of the verts on the meshes to see if anything moved before making a blendshape. Unfortunately vert look up is pretty slow in pymel so it didn’t really save any time.
I also switched from duplicating the wrapped mesh and adding all the blendshapes at once to repeatedly adding one blendshape from the wrapped mesh to the final mesh, breaking the connections between them, and renaming the alias by hand. It didn’t save much time either but I felt more confident that the blend shape names wouldn’t be renamed by maya.
I also wrote my own code to import objs for blend shapes. Once I have one of the blendshapes loaded I duplicate it then I only need to read the vert positions of every other shape and set the point on the duplicate. It was just barely faster but I think I might be able to improve it with multithreading.
We also have some generic blendshapes that most characters use. I have a modified version of the blendshape import code that calculates the character specific versions of those files using numpy arrays. Using arrays of vert positions you can get the final blend shape’s vert positions with code like this. We calculate the characters delta first so we can reuse it for all the generic shapes.
character_shape_delta = character_base_shape - generic_baseshape for generic_pose in generic_poses: generated_blendshape = characer_shape_delta + generic_pose
Finally I slimmed down some of the partial rigs we use to generate things, saving another 10 seconds or so there.
In the end I ended up saving 30 second with my best test time of 66 seconds. I was hoping to get under a minute, but I hit a brick wall there and I was ready to move onto speeding up our tools which I will talk a little bit about later.
Most of our rigging code goes back at least 4 years and when you are releasing a yearly game a lot of things get added and removed from it every time, leaving a lot of cruf sitting around. I think every time I tried to answer a “TODO: why are we doing things this way” I speed things up a little bit, or at least I could provide an actual answer to the question.