Cool Story Bro - SingleBlendMesh and my interest for code optimization
by Luca Di Sera
In this post I’d like to talk about the hows of how I got interested in code optimization.
It all started somewhat randomly about a week ago….
The SingleBlendMesh deformer
Just out of school I was looking for some project to develop my skills. Being interested in working in the VFX/Film industry the more reasonable choice was to continue deepening my knowledge of Maya API. Deformers are one of the main work a C++ developer has to do in a production environment when working with Maya so the choice of what to concentrate on was simple. There were many deformers I could choose from : Smoothing deformers, Blendshape deformers, esoteric or funny deformers and so on. Being interested in morphing algorithm, but wanting to keep it simple as a start, a BlendShape deformer was an obvious choice. It presented a lot of interesting challenges, from deepening my array attributes knowledge to Node-architecture to confront some of the challenges that may arise like changing defined blendshapes or offering a good attribute interface with Maya templates.
A few hours of work and the base to work on was done. A simple BlendShape deformer that accepted only one mesh. Looking to improve it I was getting ready to make it a multiple BlendShape deformer…
…But then I got lost…
One lingering interest I had for a while was working with Maya Profiler. I always loved profilers. They can give you an insight on your code workings and provide you with data to back up your development timeline choices. I could not let such an opportunity pass and decided to use this project to start working with Maya Profiler. Oh boy what I had done. From there it was a descent to the hell of computation time obsessions.
The first profiling and a few optimization passes
Working with Maya Profiler is really easy. There isn’t much you have to do to set it up and it offers you a decent profiling interface while compromising a bit about accuracy and versatility. Half-an-hour of work and I had my hands on some numbers to crunch. Here you can get a look at some of them:
| Test Scene | Node Evaluation Total | Node Evaluation Average | Deform Evaluation Total | Deform Evaluation Average | Deform Evaluation Min | Deform Evaluation Max : |
|---|---|---|---|---|---|---|
| 4-sphere | 14000ms | 90ms | 9000ms | ~70ms | ~70ms | ~91ms |
| 1-sphere | 7580ms | 57.3ms | 5894ms | 49.17ms | 48.2ms | 51ms |
This was ugly to me. I wasn’t dissatisfied with the deformer code, for how simple it was there was still a beauty to it. But looking at this I felt a ache in my heart. It just seemed so FUCKING SLOW!
I could never let it pass
It was a detour sure, but a didactical one at that. I started a first optimization pass. A bit of data caching and restructuring of code. Not much work, the whole of this story actually spans not more than a few days of work, but pretty effective. With my heart full of hope I decided to do a new profiling.
"Yet what was 'fore my eyes
But if not the brightest light."
The following data was spread before me:
| Test Scene | Node Evaluation Total | Node Evaluation Average | Deform Evaluation Total | Deform Evaluation Average | Deform Evaluation Min | Deform Evaluation Max : |
|---|---|---|---|---|---|---|
| 4-sphere | 15000ms | 100ms | 10000ms | ~80ms | ~80ms | ~94ms |
| 1-sphere | 6526ms | 47.56ms | 4751ms | 39.60ms | 30ms | 42ms |
It wasn’t much, and I actually somehow slowed the 4-sphere test scene. But my heart was filled with joy. Like a baby moving its first step into a world full of wonder I was intoxicated by a sense of discovery. I could never stop here. What was initially a detour was now the main objective of my work. I dipped my toes into whatever optimization source I could find. In the end I got to multithread my work. I read many theoretical book about multithreading but never tried it practically. It worked, beautifully. It was able to cut almost 50% of the computation times. This filled me with pride.
A friendly hand
I decided to make a post about this work on facebook. To share my joy. To my surprise someone commented. I could not believe it when I saw that it was Marco Giordano. I admire him a lot. That one of my idol would respond to me was a great surprise. To further my joy he offered a friendly hand. He asked me for the code so he would review it for me. I got the opportunity as soon as possible. It was incredible. He could teach me so much in so little. He showed me some detail that I could never see before and my curiosity was out of this world. It was a friendly hand indeed, but one that brought me to the pit of obsession.
Everything I was learning filled me with joy. Every millisecond I could tear down furthered my sense of pride. When I dipped into AVX I was completely captured. I wanted to work in the High-performance computing and code-optimizations world.
And so the story ends…
This was the story of why I’m studying this type of programming and the reason this blog exists. It was short but it holds a great meaning to myself. The SingleBlendMesh deformer would be a perfect fit for a Case-Study and I would have all the code ready. I decided to not use it as such since I’m starting a new project which will contain all of the things I learned on this project but gives me more space to experiment and learn.
Well this was all, I hope I could get you engaged and that I could let you know a bit about how I’m made.
"He who fights with monsters should be careful lest he thereby become a monster.
And if thou gaze long into an abyss, the abyss will also gaze into thee."