[Mesa-dev] [Bug 106232] LLVM unit tests have error in random number handling

[bugzilla-daemon at freedesktop.org]

https://bugs.freedesktop.org/show_bug.cgi?id=106232

--- Comment #4 from Roland Scheidegger <sroland at vmware.com> ---
Created attachment 139140
  --> https://bugs.freedesktop.org/attachment.cgi?id=139140&action=edit
patch to fix bogus random numbers

Ok this patch should fix it.
Non-normalized signed ints had completely bogus ref values, which I fixed as
well. (As a side note, for things like si32x4 -> si16x8 or the unsigned
equivalent, the probability that we hit something which _isn't_ clamped is very
low - for si32x4 -> si8x16 it's not even in part-per-million range.)

But we can't apply it until we fix the remaining failures. They don't actually
look too bad, all involving signed normalized numbers - all look like rounding
issues, we expect a max error of 1 but can get an error of 2. Might be due to
the snorm conversion formulas we (have to) use even as indeed it looks like we
get these errors only for negative numbers < -0.5 or so, but I'm not entirely
sure what happens.

-- 
You are receiving this mail because:
You are the QA Contact for the bug.
You are the assignee for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20180426/c804f884/attachment-0001.html>