Use K-means clustering to group Spotify data based on metrics found in genres_v2.csv.
The serial implementation implemented the K-means clustering algorithm as simply as possible with, a focus on isolating the functional parts of the program to aid parallelization later on.
The shared CPU implementation made use of OpenMP to parallelize
portions of logic that were frequently running. Typically, this
entailed the use of the parallel for directive to divide the
serial work among processors.
The shared GPU implementation condensed much of the code down to its bare essentials. Using CUDA allowed the creation of kernels that could be launched to seperate the work of what would ususally be a serial loop. Although the output does not identically match the serial implementation, the visualization is inidstiguishable from the above implementations.
Makefiles are included in each subdirectory.
Compile with make
The execution, validation and visualization steps can all be run with.
bash run-all.sh
Each file has been already compiled within its subdirectory, and can be run as follows:
./serial ../indata.csv > outdata
The output is an in-order list of each song's cluster.
Visulization can be run using matplotlib as follows:
python3 visualize.py serial shared-cpu shared-gpu
Validation was performed with GNU's diff program as follows:
diff -q serial/outdata shared-cpu/outdata
The input has been altered to only include the three visualized columns. An input file with more columns can be added by replacing the COLS definition in kmeans.h