deprecate invwarpedview in favor of InvWarpedView

[johnnychen94]

Should be the last PR before v0.9 release.

Now that InvWarpedView accepts method and fillvalue keywords, and also commits to the parent(wv) === img assumption. (This is also an improvement to #24 and thus makes invwarpedview not really necessary.)

This also introduces the same set of already deprecated API to keep consistent with warpedview, e.g., InvWarpedView(img, tinv, inds, method, fillvalue) is newly added as a deprecated method.

Other unrelated changes:

• propagate @inbounds for getindex method for both WarpedView and InvWarpedView
• simplify Base.size methods and remove IndexStyle; it's sufficient to use the fallback implementation in Base.

TODO:

• bring the SubArray trick back appropriately (This trick is directly removed)

[codecov]

Codecov Report

Merging #138 (36e9ea8) into master (eeee0bf) will decrease coverage by 1.05%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master     #138      +/-   ##
==========================================
- Coverage   92.15%   91.09%   -1.06%     
==========================================
  Files           8        8              
  Lines         204      191      -13     
==========================================
- Hits          188      174      -14     
- Misses         16       17       +1     

Impacted Files	Coverage Δ
src/compat.jl	80.00% <ø> (ø)
src/warp.jl	100.00% <ø> (ø)
src/invwarpedview.jl	84.00% <100.00%> (-13.15%)	⬇️
src/warpedview.jl	94.11% <100.00%> (+9.11%)	⬆️

---

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 eeee0bf...36e9ea8. Read the comment docs.

[timholy]

This seems reasonable. One distinction between InvWarpedView and invwarpedview is that the first should be a constructor and always return an InvWarpedView. The second would have license to, e.g., strip a WarpedView wrapper that happens to be inverted from the one being applied. That's probably a corner case that would only arise by design rather than in "real" usage.

[johnnychen94]

The more concerning issue I find with this trick with this "license" to merge two or more lazy warp views through SubArray is that out-of-domain values are not guaranteed to be valid. This can be a deeply buried root of potential bugs so I choose to remove this to make things easier to intuit.

img_camera = testimage("cameraman")
img = similar(img_camera) # uninitialized array
v = view(img, 75:195, 245:370)
v .= img_camera[v.indices...]

tfm = recenter(RotMatrix(-pi/8), center(img_camera))
wrong_result = collect(invwarpedview(v, tfm)) # there's no way to check if `parent(v)` are filled with real useful data.
correct_result = collect(InvWarpedView(v, tfm))
mosaic(wrong_result, correct_result; nrow=1)

wrong_result = collect(invwarpedview(v, tfm, axes(v)))
correct_result = collect(InvWarpedView(v, tfm, axes(v)))
mosaic(wrong_result, correct_result; nrow=1)

It should be the caller's responsibility to check whether this optimization is valid and not the callee invwarpedview/warpedview.

[timholy]

I like this way of thinking! 100% agreed.