Fix _unsafe_trunc to reduce the likelihood of arbitrary values

[kimikage]

This fixes (mitigates) #288.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 96.93%. Comparing base (2604b5a) to head (c7081b5).

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #291      +/-   ##
==========================================
- Coverage   97.05%   96.93%   -0.13%     
==========================================
  Files           7        7              
  Lines         782      784       +2     
==========================================
+ Hits          759      760       +1     
- Misses         23       24       +1     

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

[kimikage]

Benchmark (x86-64 Windows)

As for x86-64, there is no major performance regression.
There was about a 2x regression between v1.6.7 and v1.10.2 for 32-bit values, but this does not seem to be related to this change.

using FixedPointNumbers
using BenchmarkTools

mat_f64 = rand(Float64, 1000, 1000) .* 2 .- 1;
mat_f32 = rand(Float32, 1000, 1000) .* 2 .- 1;

@btime $mat_f64 .% N0f8;
@btime $mat_f64 .% N0f16;
@btime $mat_f64 .% N16f16;
@btime $mat_f32 .% N0f8;
@btime $mat_f32 .% N0f16;
@btime $mat_f32 .% N16f16;
@btime $mat_f64 .% Q0f7;
@btime $mat_f64 .% Q0f15;
@btime $mat_f64 .% Q15f16;
@btime $mat_f32 .% Q0f7;
@btime $mat_f32 .% Q0f15;
@btime $mat_f32 .% Q15f16;

julia> versioninfo()
Julia Version 1.10.2
Commit bd47eca2c8 (2024-03-01 10:14 UTC)
Build Info:
  Official https://julialang.org/ release
Platform Info:
  OS: Windows (x86_64-w64-mingw32)
  CPU: 8 × 11th Gen Intel(R) Core(TM) i7-1165G7 @ 2.80GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-15.0.7 (ORCJIT, tigerlake)
Threads: 1 default, 0 interactive, 1 GC (on 8 virtual cores)

before

julia> @btime $mat_f64 .% N0f8;
  214.200 μs (2 allocations: 976.67 KiB)

julia> @btime $mat_f64 .% N0f16;
  428.900 μs (2 allocations: 1.91 MiB)

julia> @btime $mat_f64 .% N16f16;
  1.118 ms (2 allocations: 3.81 MiB)

julia> @btime $mat_f32 .% N0f8;
  119.000 μs (2 allocations: 976.67 KiB)

julia> @btime $mat_f32 .% N0f16;
  327.400 μs (2 allocations: 1.91 MiB)

julia> @btime $mat_f32 .% N16f16;
  1.095 ms (2 allocations: 3.81 MiB)

julia> @btime $mat_f64 .% Q0f7;
  216.300 μs (2 allocations: 976.67 KiB)

julia> @btime $mat_f64 .% Q0f15;
  436.200 μs (2 allocations: 1.91 MiB)

julia> @btime $mat_f64 .% Q15f16;
  1.123 ms (2 allocations: 3.81 MiB)

julia> @btime $mat_f32 .% Q0f7;
  116.900 μs (2 allocations: 976.67 KiB)

julia> @btime $mat_f32 .% Q0f15;
  288.300 μs (2 allocations: 1.91 MiB)

julia> @btime $mat_f32 .% Q15f16;
  1.122 ms (2 allocations: 3.81 MiB)

after

julia> @btime $mat_f64 .% N0f8;
  202.500 μs (2 allocations: 976.67 KiB)

julia> @btime $mat_f64 .% N0f16;
  415.000 μs (2 allocations: 1.91 MiB)

julia> @btime $mat_f64 .% N16f16;
  1.403 ms (2 allocations: 3.81 MiB)

julia> @btime $mat_f32 .% N0f8;
  115.600 μs (2 allocations: 976.67 KiB)

julia> @btime $mat_f32 .% N0f16;
  322.200 μs (2 allocations: 1.91 MiB)

julia> @btime $mat_f32 .% N16f16;
  1.383 ms (2 allocations: 3.81 MiB)

julia> @btime $mat_f64 .% Q0f7;
  205.500 μs (2 allocations: 976.67 KiB)

julia> @btime $mat_f64 .% Q0f15;
  417.700 μs (2 allocations: 1.91 MiB)

julia> @btime $mat_f64 .% Q15f16;
  1.120 ms (2 allocations: 3.81 MiB)

julia> @btime $mat_f32 .% Q0f7;
  116.200 μs (2 allocations: 976.67 KiB)

julia> @btime $mat_f32 .% Q0f15;
  329.200 μs (2 allocations: 1.91 MiB)

julia> @btime $mat_f32 .% Q15f16;
  1.175 ms (2 allocations: 3.81 MiB)

[kimikage]

We are having problems testing ColorTypes.jl. So, I am going to backport this to v0.8.
However, since v0.8.4, the files have been rewritten a lot. Any thoughts?

[kimikage]

In any case, this PR is a bug fix, so I will merge it first.
Theoretically, there should be no major performance regression, and even if there is, it should be tackled in another PR.