pycudwt-multitarget 1.0.5

pycudwt-multitarget

pycudwt-multitarget is a python module for parallel Discrete Wavelet Transform. This is a fork of the wrapper of PDWT.

Note: this project is much the same as pycudwt, but it has the ability to compile for multiple different GPUs to obviate the need to have separate containers (Docker or enroot) for different instance types having different GPUs (ie: one image for both A100s and H100s).

When these changes are merged back into pycudwt, I will no point there and make a note in the new description.

Installation

Requirements

You need cython and nvcc (the Nvidia CUDA compiler, available in the NVIDIA CUDA Toolkit).

For the tests, you need pywavelets. python-pywt is packaged for Debian-like distributions, more recent changes are available on the new repository.

Stable version (from pypi)

pip install pycudwt-multitarget

From conda recipe

Conda build for a specific cudatoolkit version that matches one in your conda environment, e.g.:

export CUDA_VERSION="10.1.243"
conda build conda-recipe/

Development version (from github)

git clone https://github.com/pierrepaleo/pypwt
cd pypwt
pip install .

You can specify the compute capability when building the library:

PYCUDWT_CC=86 pip install .

# or to target multiple specific GPUs
PYCUDWT_CC=80,90

# or to let nvcc target your current GPU(s)
PYCUDWT_CC=all

Learn more here.

Testing

If pywavelet is available, you can check if pycudwt gives consistent results :

cd test
python test_all.py

the results are stored in results.log.

Getting started

Computing a Wavelet Transform wity pycudwt is simple. In ipython:

from pycudwt import Wavelets
from scipy.misc import lena
l = lena()
W = Wavelets(l, "db2", 3)
W
------------- Wavelet transform infos ------------
Wavelet name : db2
Number of levels : 3
Stationary WT : no
Cycle spinning : no
Separable transform : yes
Estimated memory footprint : 5.2 MB
Running on device : GeForce GTX TITAN X
--------------------------------------------------
W.forward()
W.soft_threshold(10)
W.inverse()
imshow(W.image)