Installing¶

pip install vidmag

That is the whole thing on any machine. It always gives you a working library on the processor cores; whether it also builds graphics acceleration depends on the hardware. The two sections below say how to get each kind.

NVIDIA graphics processor¶

The fastest path is hand-written CUDA, compiled during installation. It needs the CUDA toolkit — specifically nvcc — already on the machine.

nvcc --version          # prints a version => the extension will be built
pip install vidmag

If nvcc is absent the install still succeeds and prints that the extension was skipped — a missing compiler should not stop you installing a library that works without it. Two environment variables change the defaults:

Variable	Effect
`VIDMAG_CUDA_REQUIRE=1`	Turn a missing `nvcc` into a build error instead of a skip
`VIDMAG_CUDA_ARCHS=all`	Build for a range of NVIDIA cards, not just the one compiling

Check what you got:

import vidmag.cuda

print(vidmag.cuda.have_cuda)

Apple, AMD or Intel graphics processor¶

These are reached through OpenCL, which needs one extra package:

pip install "vidmag[opencl]"

The driver comes from your operating system or graphics vendor, not from this project. macOS ships one; on Linux it comes from the vendor's driver package. Installing a Python package and installing a driver are different jobs, so ask which half is missing:

from vidmag.opencl import runtime

print(runtime.unavailable_reason() or f"ready: {runtime.device_name()}")

That prints the device name if it works, and otherwise says what is missing. Metal and Vulkan are alternatives on the same hardware — see backends and hardware.

Which one gets used¶

By default, the fastest that will run: NVIDIA, then Metal, Vulkan, OpenCL, PyTorch, then the processor cores. The choice is reported through the vidmag logger, and you can ask before running:

from vidmag import backend

name, _ = backend.select("auto")
print(name)

Naming one explicitly is honoured exactly and fails loudly if it cannot run — it is never quietly swapped for a slower one.

From a checkout¶

git clone https://github.com/iamkucuk/eulerian-video-magnification-cuda
cd eulerian-video-magnification-cuda
pip install -e ".[dev]"
python -m pytest tests/ -q

The suite reports skips as well as passes. On a machine with no graphics processor the whole hardware-comparison suite skips, so a green run there says nothing about that hardware — read the skip count, not just the pass count.