Pixelify

Image processor that applies filters to images using CUDA.

Installation steps:

On Linux:

Required dependencies:

• git
• GoogleTest
• OpenGL
• GLEW
• glfw3
• exiv2
• gcc
• CUDA

Optional dependencies:

• docker
• nvidia-container-toolkit

Install all these dependencies using your package manager, depending on which distro you have.

git clone the repository:

git clone https://github.com/robertmotr/pixelify.git

Inside your locally cloned repository, update all the external libraries:

git submodule update --init --recursive --remote

Using Docker:

If you decide to use docker (assuming you installed the optional dependencies), then start a docker daemon using dockerd, and then run the following build scripts:

sh base-build.sh && sh full-build.sh

You may need to grant the Docker container access to your host's X server. In that case, you need to run the following command: xhost +local:docker

Finally run sh run-dockerized.sh to launch Pixelify through Docker. If you decide not to use Docker, then see the next section.

Building and running natively:

Simply run the following command: mkdir build && cd build && cmake -{nproc} -DCBUILD_TYPE=Release .. Once the build finishes, launch the application by executing ./pixelify.

On Windows:

The latest executables for Windows are released under https://github.com/robertmotr/pixelify/releases/. (TODO, this has not been done yet).

TODO:

• add analytics tab and keep track of rendering times, maybe have a naive CPU comparison too
• keyboard shortcuts for tabbing between original/preview
• parse filter file -> get filter array of pixels
• support anti-aliasing methods + other applications aside from filters?
• cuda <--> amd, work on any GPU, maybe use HIP or some transpiling method

In progress:

• redo all tests to fix bugs and have safety
• realtime rendering if fast enough (maybe?)

Done:

• add zoom in/zoom out/pan tool + texture map
• profile texture/shared/constant memory for filter and use whichever one is faster
• move allocating image data and stuff to opening input file to make rendering seem faster,
• speed up kernel by using cached constant memory for both the filter_dimension and filter_data
• get rid of device_filter allocations because its useless now with constant memory
• get other_kernel functionality working again
• fix segfault from kernel formulas
• fix kernel formulas not producing results that "look right"
• add fits/xmpp data to image description
• add multiple filter passes option
• add pixel inspector
• add filter size slider, i.e 2x2 -> 9x9
• set filter size slider to be dependent on filter selected, i.e identity can only have 3x3
• optimize kernels to get ~200ms runtime preferably even lower (update: getting consistently sub 20ms which is fantastic)
• imgui implementation