Matrix multiplication using Python3.8, CUDA, PyCUDA

• Implemented the fastest way to multiply matrix ( using numpy.dot() );

• Compared a time running of two ways on gpu: 1) with numpy.dot(); 2) with C++ kernel in PyCUDA

• Results:

  N	CPU time, ms	GPU time, ms	Speedup
  128	0.112	0.416	0.27
  256	0.604	0.607	1.00
  512	4.458	2.122	2.10
  1024	33.154	9.679	3.43
  2048	256.891	61.235	4.20

• Conclusion:

  • On small matrix sizes CPU calculate better than GPU, but, with the matrix size growing, GPU shows better results, in compare with CPU.
  • Most reasonable for multiplying high-sizes matrix is to use GPU, instead of CPU.

Harris Corner Detector using Python3.8, CUDA, PyCUDA

• Results:

  • For Image (384 by 384 pixels):

    CPU time, ms	GPU time, ms	Speedup
    3.216	0.263	12.22

Original Image:	Processed Image:

• Conclusion:
  • GPU shows better results, in compare with CPU.

Salt and Pepper noise filtering using Python3.8, CUDA, PyCUDA

• To obtain an array characterizing the color of pixels, the Pillow library was used

• Each element of the output image was calculated (on the GPU) by a separate thread

• To speed up calculations on the GPU inside each image block, copying of elements from global memory to shared memory was implemented, which reduces the number of calls to global memory

• Results for filter 3х3 (median)

  File	CPU time, ms	GPU time, ms	Speedup
  256.bmp	524.320	0.289	1813.15
  512.bmp	2273.571	0.492	4619.92
  1024.bmp	9105.398	1.484	6135.44

• Results for filter 5х5 (median)

  File	CPU time, ms	GPU time, ms	Speedup
  256.bmp	1210.220	0.925	1308.35
  512.bmp	5102.574	2.189	2331.01
  1024.bmp	19996.483	6.873	2909.43

  Original Image:	CPU Processed Image:	GPU Processed Image:

• Conclusion:

  • Using the GPU in the framework of the median filtering task gives great increase in speed.