APM - Mini CUDA project: Image processing

Build

Dependencies

• A C++ compiler supporting ISO C++11 ;
• NVIDIA's CUDA compiler nvcc and NVIDIA's CUDA runtime libraries supporting ISO C++ 11 ;
• The FreeImage C++ library.

The provided Makefile assumes that the FreeImage library has been installed at ./ext/FreeImage/build from the root of the project. If this is not the case, please update the relevant parts of the Makefile as follows:

CXXFLAGS := -std=c++11 -Wall -Wextra -fopenmp -g -I $(INCDIR)/ \
            -I path/to/FreeImage/include/                      \
            -L path/to/FreeImage/lib/
NVCXXFLAGS := -std=c++11 -O3 -use_fast_math -extra-device-vectorization -Xcompiler -fopenmp -g -I $(INCDIR)/ \
              -I path/to/FreeImage/include/                                                                  \
              -L path/to/FreeImage/lib/

Do not forget to re-export the correct location of the library with LD_LIBRARY_PATH:

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:path/to/FreeImage/lib/

The user must specify the default Compute Capability for the CUDA using the COMPUTE_CAPABILITY variable. Once these prerequisites are matched the project can be built using the provided Makefile, e.g. for NVIDIA A100 GPUs:

COMPUTE_CAPABILITY=80 make

Features

The project provides multiple runtime options and flags that are listed in the help command:

target/cuda-image-transform [-h|--help]

Arguments in between angle brackets < > are required, those in between square brackets [ ] are optional.

Note: Because of getopt/getopt_long special treatment of optional arguments (see their manpage), flags that take optional arguments are required to have no space between the flag and the argument, e.g. when specifying a threshold for the Sobel filter:

target/cuda-image-transform --sobel 100 # BAD: argument `100` is ignored
target/cuda-image-transform --sobel100  # OK!: Threshold is set to `100`

The implemented filters are:

• Blur and iterative blur ;
• Diapositive ;
• Grayscale ;
• Horizontal flip ;
• Pop-Art ;
• Saturation (red, green and blue) ;
• Sobel filter.

Apart from the pop-art filter which is GPU-only, all filters are implemented for both CPU and GPU. The target can be set from the command line when launching the program using the -t or --target option (default target is the GPU device).

Moreover, the filters can be chained together with no memory transfers by specifying them in-order from the command line, e.g.:

# Flip the image, apply iterative blur 50 times, saturate green and saturate blue:
target/cuda-image-transform --horizontal-flip --blur50 --saturategreen --saturateblue
# Or with the shorthand options:
target/cuda-image-transform -f -b50 -sg -sb

Note: some filters may override each other.