Some timings of rendering a short zoom video with and :

kf rendering 77 keyframes at 12288x1360:
14mins OpenCL AMD RX 580 GPU
30mins OpenCL AMD 2700x CPU
72mins regular AMD 2700x CPU

exrsubsample 77 keyframes to 1536x170 for interactive use in zoomasm:
13mins

reprojecting 12288x1360 keyframes to 1920x1080p60 zoom video (duration 1min) in zoomasm (includes encoding with ffmpeg):
20mins

All times wall-clock, on a desktop with other light usage happening at the same time. Formula was a hybrid of cubic Mandelbrot and cubic Burning Ship, maximum iteration count at the end was about 24000.

OpenCL on GPU is about 5.1 times faster than the regular CPU implementation, and 2.1 times faster than the OpenCL on CPU. OpenCL on CPU is about 2.4x faster than regular CPU, presumably because the branches in inner loops are resolved at kernel compile time instead of run time.

As exrsubsample is probably IO bound, and I used a slow USB2 external hard drive for the data, it's possible that the timings could be improved using a RAM disk. Required space for this project is 20GB.

Follow

Using a RAM disk for all the intermediary data, and enabling OpenEXR multithreading in exrsubsample and zoomasm, and the times are reduced quite a bit

13m46 KF rendering (GPU, OpenCL)
2m05 exrsubsample
11m18 zoomasm

It is a bit more risky, as a system crash or power loss will mean you need to restart from scratch...

· · Web · 0 · 0 · 0
Sign in to participate in the conversation
post.lurk.org

Welcome to post.lurk.org, an instance for discussions around cultural freedom, experimental, new media art, net and computational culture, and things like that.