Use Float32 for the matrix coefficients for low precision colors

[kimikage]

This also changes the definition of the sRGB transform matrix slightly. (cf. #355 (comment))

cf. #477

[codecov]

Codecov Report

Merging #482 (dcd3281) into master (9f27da0) will decrease coverage by 0.02%.
The diff coverage is 96.77%.

@@            Coverage Diff             @@
##           master     #482      +/-   ##
==========================================
- Coverage   93.42%   93.40%   -0.03%     
==========================================
  Files           9        9              
  Lines        1050     1061      +11     
==========================================
+ Hits          981      991      +10     
- Misses         69       70       +1     

Impacted Files	Coverage Δ
src/conversions.jl	99.38% <92.30%> (-0.32%)	⬇️
src/algorithms.jl	60.95% <100.00%> (-0.37%)	⬇️
src/utilities.jl	98.87% <100.00%> (+0.11%)	⬆️

---

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 9f27da0...dcd3281. Read the comment docs.

[johnnychen94]

Feel free to resolve these comments if they don't make sense.

[kimikage]

I will submit some additional PRs to mitigate the precompilation problem.

[kimikage]

Benchmark

using Colors, FixedPointNumbers, BenchmarkTools
N = 1000;
rgb_f64 = rand(RGB{Float64}, N, N);
rgb_f32 = rand(RGB{Float32}, N, N);
rgb_n0f8 = rand(RGB{N0f8}, N, N);
xyz_f64 = XYZ.(rgb_f64);
xyz_f32 = XYZ.(rgb_f32);
yiq_f64 = YIQ.(rgb_f64);
yiq_f32 = YIQ.(rgb_f32);

@btime convert.(XYZ, $rgb_f64);
@btime convert.(XYZ, $rgb_f32);
@btime convert.(XYZ, $rgb_n0f8);
@btime convert.(RGB, $xyz_f64);
@btime convert.(RGB, $xyz_f32);
@btime convert.(RGB{N0f8}, $xyz_f32);

@btime convert.(YIQ, $rgb_f64);
@btime convert.(YIQ, $rgb_f32);
@btime convert.(YIQ, $rgb_n0f8);
@btime convert.(RGB, $yiq_f64);
@btime convert.(RGB, $yiq_f32);
@btime convert.(RGB{N0f8}, $yiq_f32);

	(v0.12.8)	before	after
RGB{Float64}-->XYZ{Float64}	56.330 ms	30.209 ms	30.318 ms
RGB{Float32}-->XYZ{Float32}	25.483 ms	25.528 ms	23.325 ms
RGB{N0f8} -->XYZ{Float32}	3.547 ms	3.506 ms	2.396 ms
XYZ{Float64}-->RGB{Float64}	39.305 ms	36.156 ms	34.093 ms
XYZ{Float32}-->RGB{Float32}	55.132 ms	47.960 ms	*36.715 ms
XYZ{Float32}-->RGB{N0f8}	55.942 ms	47.224 ms	*42.136 ms

*: see also PR #493

	before	after
RGB{Float64}-->YIQ{Float64}	5.379 ms	5.245 ms
RGB{Float32}-->YIQ{Float32}	3.076 ms	2.373 ms
RGB{N0f8} -->YIQ{Float32}	9.060 ms	2.006 ms
YIQ{Float64}-->RGB{Float64}	5.709 ms	5.755 ms
YIQ{Float32}-->RGB{Float32}	3.398 ms	2.546 ms
YIQ{Float32}-->RGB{N0f8}	18.193 ms	16.755 ms

[kimikage]

Now that the RGB<-->XYZ conversions avoids the precompilation problem, I think we are ready to merge this.