Write optimization via caching

[IanButterworth]

Gives a slight increase in write speed. Benchmarks over in JuliaIO/FileIO.jl#290

[tlnagy]

Huh. I guess eager evaluation is a negative here.

[IanButterworth]

Turns out a lot of the slowness was julia's write handler unfolding the reinterpret array type sent to write.
Doing a collect after the views greatly improves speed.

This PR now takes a 1000x1000 image save time from 70ms to 2ms (results here)

[IanButterworth]

@tlnagy Cache now set up, and performs well JuliaIO/FileIO.jl#290 (comment)

[tlnagy]

Hi @IanButterworth, thanks for implementing the cache idea! I think it's unnecessary to build the write-cache on construction of the DenseTaggedImage type because many TIFFs will only be read and never written. I think it's better to construct this on write.

[codecov-io]

Codecov Report

Merging #33 (2005b07) into master (c974bf9) will decrease coverage by 15.30%.
The diff coverage is 2.54%.

@@             Coverage Diff             @@
##           master      #33       +/-   ##
===========================================
- Coverage   87.80%   72.49%   -15.31%     
===========================================
  Files          12       13        +1     
  Lines         541      658      +117     
===========================================
+ Hits          475      477        +2     
- Misses         66      181      +115     

Impacted Files	Coverage Δ
src/TiffImages.jl	100.00% <ø> (ø)
src/precompile.jl	0.00% <0.00%> (ø)
src/types/dense.jl	97.05% <100.00%> (+0.08%)	⬆️

---

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 c974bf9...2005b07. Read the comment docs.

[IanButterworth]

Good point! Moved the cache to write time.
Also set up the SnoopCompile tooling to auto-generate and install precomp statements. Results again over in JuliaIO/FileIO.jl#290 (comment)

[tlnagy]

I think this looks good to be merged from my end!

I'm not sure if the snooping makes a difference in its current form, but perhaps that's because of latency higher up in the stack.

[tlnagy]

Could we move the Snoopcompile stuff into a separate PR and I can go ahead and merge the cache-collect portion of this?

[IanButterworth]

Could we move the Snoopcompile stuff into a separate PR

👍🏻 #35

[tlnagy]

Awesome, thanks for figuring this out!