Avoid allocations in in-place image resizing

[pxl-th]

This PR avoids allocations for the BSpline types of interpolations when doing in-place resizing. If I'm not mistaking anything, BSpline interpolations do not modify the original array?

The benefits are considerable. Also maybe we should expose the API?

• With allocations:

  @btime ImageTransformations.imresize!(o, x; method=Linear());
  15.253 ms (2 allocations: 32.00 MiB)

• Without allocations:

  @btime ImageTransformations.imresize!(o, x; method=Linear());
  405.207 μs (0 allocations: 0 bytes)

Code to benchmark:

x = rand(Float64, 2048, 2048);
o = rand(Float64, 256, 256);
@btime ImageTransformations.imresize!(o, x; method=Linear());

[johnnychen94]

Also maybe we should expose the API?

I'm okay to this. Previously we were concerned about the naming issues because the name imresize happens to be identical to MATLAB. But since the Google vs Oracle java API lawsuit ended I have good reason to believe this won't be an issue at all.

[codecov[bot]]

Codecov Report

Merging #150 (c3eb09e) into master (284bb1e) will increase coverage by 0.13%. The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master     #150      +/-   ##
==========================================
+ Coverage   90.14%   90.27%   +0.13%     
==========================================
  Files           8        8              
  Lines         213      216       +3     
==========================================
+ Hits          192      195       +3     
  Misses         21       21              

Impacted Files	Coverage Δ
src/ImageTransformations.jl	83.33% <ø> (ø)
src/resizing.jl	100.00% <100.00%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more Δ = absolute <relative> (impact), ø = not affected, ? = missing data Powered by Codecov. Last update 284bb1e...c3eb09e. Read the comment docs.

[pxl-th]

Great! Can you add some test to ensure that it's working equivalently to imresize?

Added

[johnnychen94]

Just sent you an invitation in case you want to commit more changes

[pxl-th]

Just sent you an invitation in case you want to commit more changes

Thanks!

[timholy]

If I'm not mistaking anything, BSpline interpolations do not modify the original array?

Not for Constant and Linear, but they do for Quadratic and Cubic. I don't think you can use interpolate! with Cubic, but you can for quadratic, search https://juliamath.github.io/Interpolations.jl/latest/control/ for "destroys".

[timholy]

We might need a fix for the quadratic case pronto.

[pxl-th]

I'll make a PR asap and add tests to ensure the original image is not changed (should've done this in this PR).