How to convert this vectorized code into GPU code for MAXIMUM speedup ?

[Jan]

Please post code as text and not as a screenshot. Then it can be used by copy&paste to create an answer.

Answer 1

[Joss Knight]

I was able to get a marginal speedup with additional vectorization of the mask:

   x = sum(I < cat(3, 120, 155, 160), 3) == true;

but otherwise you've done pretty well. You've got to wonder why you need to replicate the output on every channel however. Why not discard the colour channels if you're using grayscale?

[Tanmay Virnodkar]

Actually i want to show speedup or lets say difference between normal cpu time and gpu time..... so if i first convert the picture into grayscale and use imtool then cpu time is also very less so i am not able to show speed up ...hence i decided not to discard R G B channels.

[Joss Knight]

Right, but then you're including the cost of replicating data in GPU memory and doing indexing, which is memory-bound and doesn't necessarily show the GPU in a great light.

Answer 2

[Jan]

The bottlenecks of the code are the darn clear all and the disk access using imwrite. Moving this to the GPU will not help.