[PyCUDA] CUDA Kernel for Python Code

[Subhradeep Dutta]

I am trying to write a CUDA kernel for computing a piece of Python code. I would need to sort the data inside the kernel but since there are no built in functions that would sort the data so I am not sure what would be the best way to implement this. Any suggestions on how to implement it would be appreciated. Feel free to put some sample code or point me to online resources that have a similar problem.


for i in xrange(0,img_size_x-window_size):
for j in xrange(0,img_size_y-window_size):
kernel = img[i:i+window_size,j:j+window_size]
kernel_flat = np.sort(kernel.flatten())
# Calculate the rank
rank = np.where(kernel_flat == img[i,j])[0][0]
img_mod[i,j] = int((rank * 255 )/(window_size_squared))


Thanks

[Jerome Kieffer]

On Sat, 6 May 2017 10:24:51 +0000

Post by Subhradeep Dutta
I am trying to write a CUDA kernel for computing a piece of Python code. I would need to sort the data inside the kernel but since there are no built in functions that would sort the data so I am not sure what would be the best way to implement this. Any suggestions on how to implement it would be appreciated. Feel free to put some sample code or point me to online resources that have a similar problem.
kernel = img[i:i+window_size,j:j+window_size]
kernel_flat = np.sort(kernel.flatten())
# Calculate the rank
rank = np.where(kernel_flat == img[i,j])[0][0]
img_mod[i,j] = int((rank * 255 )/(window_size_squared))

There are a couple of sort available in parallel, among them the
bitonic-sort is pretty well adapted to this kind of problems.

Your code looks like a median filter. Resources are available on the
net about this.

Cheers,