add support for planar images

[chrstphrbrns]

Fixes #125

[tlnagy]

What's the rational for the explicit SIMD calls and the dependence on SIMD.jl?

[chrstphrbrns]

Better performance

[tlnagy]

How significant of an impact does it make? Do you have any comparisons with and without?

I prefer to avoid adding dependencies if the difference is relatively minor.

[chrstphrbrns]

For one particular 5120x2880 image, it's roughly 130ms vs 80ms (results are similar for compressed and uncompressed versions of the image)

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (bda25ff) 92.20% compared to head (a920e3e) 92.76%.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #136      +/-   ##
==========================================
+ Coverage   92.20%   92.76%   +0.56%     
==========================================
  Files          13       13              
  Lines         924      982      +58     
==========================================
+ Hits          852      911      +59     
+ Misses         72       71       -1     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[tlnagy]

I hacked up a quick script to measure the timings:

Script

module TestMain
    println(@__MODULE__)
    using Pkg
    Pkg.activate(; temp = true)
    Pkg.add("TiffImages")

    a = @elapsed Base.compilecache(Base.identify_package("TiffImages"))

    b = @elapsed using TiffImages

    open("timings.txt", "w") do f
        write(f, join([string(@__MODULE__), a, b], "\t"), "\n")
    end
end

module TestPR136
    println(@__MODULE__)

    using Pkg
    Pkg.activate(; temp = true)
    Pkg.develop("TiffImages")

    a = @elapsed Base.compilecache(Base.identify_package("TiffImages"))

    b = @elapsed using TiffImages

    open("timings.txt", "a") do f
        write(f, join([string(@__MODULE__), a, b], "\t"), "\n")
    end
end

It doesn't quite work for proper using measurements so I switched the module orders to get accurate first timings for the main branch vs this PR.

Which gave the following:

Condition	Precompilation (s)	Using (s)	Using Delta (%)
Main.TestMain	38.981615158	0.869952215	1.0
Main.TestPR136	39.184539139	1.031776947	1.19

So this PR results in approximately the same precompilation time, but almost 20% reduction in speed when loading. I'm not sure if the slight reduction in planar image loading speed is worth a 20% reduction in speed for all users.

EDIT: this is with commit a920e3e

[chrstphrbrns]

For me, a 40% per-image win easily justifies an imperceptible (or barely perceptible) increase in one-time load cost

FWIW, the PR for #58 also has a SIMD path, which gives somewhat better improvements for "unusual" bit depths

Twice as fast is nothing to sneeze at

[tlnagy]

The SIMD path would only accelerate the speed of loading planar images, no? The 20% increase in speed hits everyone, while the increase in speed would only benefit folks loading the planar images (which is likely a small subset of folks). I'm not totally against including SIMD as a dependency, I just want to make sure that it's justified.

[chrstphrbrns]

The additional delay is fixed and imperceptible, so I round it down to zero, whereas the optimization is per-image and the benefits therefore unbounded. One might argue that few will process enough images to appreciate the benefits, but then, again, the cost is ~zero

Do you prefer to keep the code simple for now and defer optimization until later?

[tlnagy]

Sorry for the delay with this, but I decided that the delay (particularly with the advances in Julia v1.10 on that front) should be fine. I think it likely makes sense to let this stew on main for a bit and then tag v0.9 to release it.