/gpu-programming-research

Primary LanguageCuda

K-Means Image Segmentation in CUDA


K-Means example (K=3, Iterations=1000)

Requirements:

Clone repository:

$ git clone https://github.com/dmadariaga/gpu-programming-research.git

Run GPU algorithm:

$ nvcc -o kmeans_gpu kmeans_gpu.cu
$ kmeans_gpu input.ppm k numIter output.ppm

Where:

  • input.ppm: Image to apply segmentation. Must be a raw PPM file (file type = P6).
  • k: Number of clusters
  • numIter: Number of iterations
  • output.ppm: Name of the image file created to save the results.

Run CPU algorithm:

$ gcc -std=gnu99 -o kmeans_cpu kmeans_cpu.c -fcilkplus -lcilkrts -lm
$ kmeans_cpu input.ppm k numIter output.ppm

Params are the same that GPU version

Run GPU tests:

$ ./test_gpu.sh >> gpu_results.csv

gpu_results.csv format is:

unit,size,k,iter,t
gpu,160000,3,100,0.001152
gpu,160000,3,1000,0.627432
gpu,160000,3,5000,3.854647
...

Where:

  • unit: Unit of execution (GPU or CPU)
  • size: Number of pixels in image
  • k: Number of clusters
  • iter: Number of iterations
  • t: Elapsed time int seconds

Run CPU tests:

$ ./test_cpu.sh >> cpu_results.csv

cpu_results.csv format is:

unit,size,k,iter,t
cpu,160000,3,100,0.570475
cpu,160000,3,1000,5.699365
cpu,160000,3,5000,3.793827
...

Column headers are the same that GPU version