20th Jan 2017
Here at Blender Institute we challenge ourselves to make industry quality films while improving and developing Blender and open source pipeline tools.
One of the artistic goals for Agent 327 Barbershop is to have high quality motion blur. Rendering with motion blur is known to be a technical challenge, and as result render times are usually very high. This is because we work with:
As soon as animation was ready for one of the shots, Andy prepared it for rendering with the final settings (full resolution, full amount of Branched Path Tracing samples, etc.) and shipped it to the render farm (IT4Innovations, VSB – Technical University of Ostrava).
Here are the specs of some of the machines used.
gcc_node_1- Blender compiled with gcc 5.3, with_cpu_sse=on, qbvh=on (24 cores = 2x Intel Xeon E5-2680v3, 2.5GHz)
intel_cpu24_2xMIC- compiled with intel 2016.03 with patch https://developer.blender.org/D2396 (+ loop over samples moved to loop over pixels), with_cpu_sse=off, qbvh=off (using Xeon Phi coprocessor)
Rendering took a while, and when all frames where completed we produced the following clip:
The graph is made by collecting the render time on the 2 different system configurations (yellow is
gcc_node_1 and green is
intel_cpu24_2xMIC) and overlaying it on the animation, with a time marker matching the current frame.
The problem: some frames were taking over 100 hours to render. The solution: fix Cycles!
The first step was to reduce the overall render time of the scene, in order to do more tests and collect measurements more quickly.
Here is a chart comparing the render times of the same animation with and without motion blur. Notice that the y axis uses a logarithmic scale. In some cases a frame would be 100x slower with motion blur.
This appeared to be was mostly due to motion blur, especially hair. Who knew!
After a few days of investigation, Sergey improved the layout of hair bounding boxes for BVH structure. What does this mean? A more in-depth explanation is coming soon.
This led to a dramatic performance improvement, with render times going down 10x.
After that, he applied the same optimization to triangles (for actual character geometry), and this led to further performance improvements, speeding up renders up to 8x on the full frames.
Notice that the graph uses a logarithmic scale on the y axis. To get a better idea of the performance impact, check out the following chart, with a linear scale.
This optimization affects memory usage, but in a marginal way (only 15% increase in the memory usage for this shot).
The production file is available for everyone to test and verify the results.
A similar challenge was solved during the Gooseberry project. In that case, the focus was on memory optimization.
This production challenge was solved once again thanks to the Blender Cloud Subscribers, who are supporting content-driven development: the best way to improve 3D software.
Not a Blender Cloud subscriber? Consider becoming one!