Re: Improving performance of 2D convolution

From: Olzhas Zhumabek (anonymous.from.applecity_at_[hidden])
Date: 2021-06-08 17:45:26

After battling with CMake configuration steps, we finally were able to build
benchmark branch of the Boost.GIL repo.

Prathamesh, please do the following for now:

1. Find an image you could use without attribution or royalties (it should
be public domain, not just free). If you're struggling to find one, just
draw
something and take a picture of it. The colorspace should be grayscale.
2. Scale it to 5 different sizes. Start at the minimum of 32 by 32,
and end on something big, 2048*2048. This is to test performance on
varying scales so that we will see performance scaling with input size.
3. Try to prepare a function to load the appropriate one and make it
easy to integrate with upcoming benchmarks (should accept a scale,
from 0 to 4, and return appropriately scaled image).
4. Try to find OpenCV equivalent to what we're doing now in GIL and
run the image at all scales to have ground truth prepared. If there is
none, try to write a script that uses OpenCV to generate something
close and modify the output to match GIL's output.

I will try to get the benchmark template ready as soon as possible,
but am at the moment overwhelmed by the rest of my life.
If you want to, you can try to write it yourself, we can change it later.

I will need to run your reflection PR first on something and check
with OpenCV, just to see if we have differences with them and what
the differences are. This might take a bit of time, so probably until
the end of the week. Your tests look good, I will run them locally
tomorrow, read through again and give approval.

Best,
Olzhas