Working on GPUs

By default, the MinKit package only works in CPU backends. However, the usage of GPUs is supported through the Reikna, which is itself an API for PyCUDA and PyOpenCL. In order to enable MinKit to work on GPUs, it is necessary to install at least two of the previous packages (or both). Please visit Notes for GPU compatibility to know about the dependencies of the package and the way to install them.

There are two (compatible) ways of specifying the backend in MinKit. The environmental variable MINKIT_BACKEND controls the default backend that is used in the package. In case a script is created with no explicit declaration of a backend, the following lines will execute it in CPU, CUDA and OpenCL backends, sequentially:

MINKIT_BACKEND=CPU python script.py
MINKIT_BACKEND=CUDA python script.py
MINKIT_BACKEND=OPENCL python script.py

The name of the backend is not case sensitive, so it is also possible to specify them as cpu, cuda and opencl, respectively.

In a single system it is possible to control several GPUs. The environmental variable that controls the device number is MINKIT_DEVICE, so

MINKIT_BACKEND=CUDA MINKIT_DEVICE=0 python script.py
MINKIT_BACKEND=CUDA MINKIT_DEVICE=1 python script.py

will execute the script in the first device and then in the second.

The backend can also be specified in python scripts, where different backends can coexist. It is possible to configure the script so MinKit asks the user what device to use:

import minkit
bk1 = minkit.Backend('cuda', device=0)
bk2 = minkit.Backend('cuda', interactive=True)
bk3 = minkit.Backend('cuda', device=2, interactive=True)

In the previous example, the first backend will work on the first device, whilst in the second it will depend on the input from the user. In the third case, the device is set initially to 2, however, if this device does not exist, MinKit will ask the user to introduce a new backend. For the last case, if interactive is not provided the first device will be used and a warning will be displayed.

The only objects that depend on backends are PDFs and data sets, and they can be directly built using the backend instance.

import minkit

bk = minkit.Backend('cuda', device=0)

m = minkit.Parameter('m', bounds=(10, 20))
c = minkit.Parameter('c', 15)
s = minkit.Parameter('s', 2)

ga = bk.Gaussian('g', m, c, s)
gb = minkit.Gaussian('g', m, c, s, backend=bk)

In the last example, the two last lines are equivalent.