gputools - OpenCL accelerated volume processing in Python

GPU accelerated 2D/3D image processing, e.g. for microscopy, via OpenCL (i.e. works on both NVIDIA and non-NVIDIA devices).

• convolutions
• denoising
• synthetic noise
• ffts (simple wrapper around reikna)
• affine transforms

Built on the excellent pyopencl bindings.

Some examples of processing tasks and their respective runtime (tests/benchmark/benchmark.py):

Task	Image Size/type	CPU[1]	GPU[2]	GPU (w/o transfer)
Mean filter 7x7x7	(128, 1024, 1024) uint8	1708 ms	22 ms	5 ms
Median filter 3x3x3	(128, 1024, 1024) uint8	37591 ms	49 ms	31 ms
Gaussian filter 5x5x5	(128, 1024, 1024) float32	4536 ms	86 ms	19 ms
Zoom/Scale 2x2x2	(128, 1024, 1024) uint8	18433 ms	133 ms	-
NLM denoising	(64, 256, 256) float32	21109 ms	105 ms	-
FFT (pow2)	(128, 1024, 1024) complex64	3418 ms	156 ms	21 ms

[1] AMD Ryzen Threadripper PRO 9965WX 
[2] NVIDIA GeForce RTX 5090 

Requirements

• Python 3.9+
• a working OpenCL environment (check with clinfo).

Installation

pip install gputools

Or the development version:

pip install git+https://github.com/maweigert/gputools

Check if basic stuff is working:

python -m gputools

Troubleshooting

If you experience installation issues, this might be due to pyopencl not being properly installed. The easiest way to get a working pyopencl is via conda:

conda install -c conda-forge pyopencl

Usage

Docs are still to be done ;)

Most of the methods work on both numpy arrays or GPU memory objects (gputools.OCLArrays/OCLImage). The latter saving the memory transfer (which e.g. for simple convolutions accounts for the main run time)

Convolutions

• 2D-3D convolutions
• separable convolutions
• fft based convolution
• spatially varying convolutions

import gputools

d = np.zeros((128,128), np.float32)
d[64,64] = 1.
h = np.ones((17,17))
res = gputools.convolve(d,h)

d = np.zeros((128,128,128), np.float32)
d[64,64,64] = 1.
hx,hy,hz = np.ones(7),np.ones(9),np.ones(11)
res = gputools.convolve_sep3(d,hx,hy,hz)

Denoising

bilateral filter, non local means

...
d = np.zeros((128,128,128), np.float32)
d[50:78,50:78,50:78:2] = 4.
d = d+np.random.normal(0,1,d.shape)
res_nlm = gputools.denoise.nlm3(d,2.,2,3)
res_bilat = gputools.denoise.bilateral3(d,3,4.)

Perlin noise

fast 2d and 3d perlin noise calculations

gputools.perlin3(size = (256,256,256), scale = (10.,10.,10.))

Transforms

scaling, translate, rotate, affine...

gputools.transforms.scale(d,.2)
gputools.transforms.rotate(d,axis = (64,64,64),angle = .3)
gputools.transforms.shift(d,(10,20,30))
...

fft

wraps around reikna

gputools.fft(d)
gputools.fft(d, inverse = True)

Configuration

Some configuration data (e.g. the default OpenCL platform and devic) can be changed in the config file "~/.gputools" (create it if necessary)

#~/.gputools

id_platform = 0
id_device = 1

See

gputools.config.defaults

for available keys and their defaults.

Alternatively, the used OpenCL Device can be set via the environment variables gputools_id_device, gputools_id_platform, and gputools_use_gpu (variables present in the config file will take precendence, however).