Clojure wrapper & highlevel processing pipeline ops for JOCL/OpenCL
SimpleCL intends to enable a somewhat declarative interop approach between Clojure and OpenCL compute kernels. It provides thin wrappers for basic OpenCL data structures (contexts, devices, buffers, programs, kernels), as well as an highlevel API to define & configure multi-kernel & multi-program workflows and transforming data between Clojure’s data structures and their native OpenCL representation as byte buffers.
SimpleCL is still a young project (although it’s was originally created in 2012, it has been dormant for a while), yet has been sucessfully used in production in several projects and is fairly well documented.
[thi.ng/simplecl "0.2.2"]Please see the extensive doc strings in the source and the examples below:
A trivial “Hello world” style example is located in this namespace. It defines a kernel to factor two float arrays and shows the required steps to run this on either CPU or GPU devices (by default the most performant is chosen automatically):
;; chooses max performance device lein trampoline run -m thi.ng.simplecl.test.hello ;; or force usage of given device type lein trampoline run -m thi.ng.simplecl.test.hello gpu lein trampoline run -m thi.ng.simplecl.test.hello cpu
Running the command below will run a basic 2d cloth simulation with
40,000 particles & 80,000 springs and generate an image sequence of
120 PNGs in the /export sub-directory.
lein trampoline run -m thi.ng.simplecl.test.verlet
The image sequence can then be converted into an MP4 using this `ffmpeg` command:
ffmpeg -r 30 -f image2 -start_number 5 -i "export/verlet-%04d.png" \
-vcodec libx264 -preset slow -crf 23 -pix_fmt yuv420p \
-f mp4 -threads 0 verlet.mp4 -y
A video of three different variations is here: verlet-cloth-sim.mp4
This demo is also showcasing usage & integration of the structgen partner library to ease the task of aligning nested Clojure data structures in an OpenCL compatible way (mapping to OpenCL structs).
See thi.ng/raymarchcl for full details.
A sneak peak of an upcoming example kernel pipeline to:
- compute a 2d strange attractor
- apply a 2D distortion filter to all points
- map the result points onto the surface of a sphere and
- apply a 3D distortion filter to all points
- do a standard 3d > 2d camera-screen projection
A short video fragment: attractor.mp4 (10M points/frame, ~780-840ms/frame)
More examples & documentation will be added ASAP.
Copyright © 2012 - 2015 Karsten Schmidt
Distributed under the Apache Software License 2.0