3rd-party/catch2
1
0
Fork 0
mirror of https://github.com/catchorg/Catch2.git synced 2024-11-22 05:16:10 +01:00
Code Issues Packages Projects Releases Wiki Activity

Mention x87 and ffast-math breaking FP random gen reproducibility

Browse Source
This commit is contained in:
Martin Hořeňovský 2024-02-17 11:01:43 +01:00
parent 2e1b02a0e2
commit 28c2f0b0c2
No known key found for this signature in database
GPG Key ID: DE48307B8B0D381A
1 changed files with 20 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

26
docs/generators.md
View File

@ -206,12 +206,26 @@ or OO ranges.
Unlike `std::uniform_int_distribution`, Catch2's generators also support Unlike `std::uniform_int_distribution`, Catch2's generators also support
various single-byte integral types, such as `char` or `bool`. various single-byte integral types, such as `char` or `bool`.
Given the same seed, the output from the integral generators is
reproducible across different platforms. For floating point generators, #### Reproducibility
we only promise reproducibility on platforms that obey the IEEE 754
standard, and where `float` is 4 bytes and `double` is 8 bytes. We provide Given the same seed, the output from the integral generators is fully
no guarantees for `long double`, as the internals of `long double` can reproducible across different platforms.
vary wildly across different platforms.
For floating point generators, the situation is much more complex.
Generally Catch2 only promises reproducibility (or even just correctness!)
on platforms that obey the IEEE-754 standard. Furthermore, reproducibility
only applies between binaries that perform floating point math in the
same way, e.g. if you compile a binary targetting the x87 FPU and another
one targetting SSE2 for floating point math, their results will vary.
Similarly, binaries compiled with compiler flags that relax the IEEE-754
adherence, e.g. `-ffast-math`, might provide different results than those
compiled for strict IEEE-754 adherence.
Finally, we provide zero guarantees on the reproducibility of generating
`long double`s, as the internals of `long double` varies across different
platforms.
## Generator interface ## Generator interface