The VHDL standard's own float package returns the wrong answer for sqrt(2)

The VHDL standard's float package gets sqrt(2) wrong

An audit of ieee.float_pkg, the floating-point package shipped in IEEE 1076-2008 and 1076-2019 and bundled unchanged in GHDL and NVC, compares every operation against exact arithmetic. Plain calls already fail: sqrt(2.0) in float64 lands one unit off, mac(-6.0, -1.0, -5.0) returns -15.0, and some resizes crash the simulator. Across 10,000-call sweeps, mac is wrong roughly a quarter of the time under round_nearest, and sqrt misses 735 of 10,000 float32 cases. The author traces each fault to specific lines and offers a 33-hunk patch.

Of the integers 1 to 4,096, sqrt returns a float32 one unit off for 256 and a float64 one unit off for 264.
  1. bee_rider

    The first link in the post doesn’t seem to be valid (on my iPhone at least). Maybe it is an issue of mobile safari not wanting to open… a vhdl blob, though? Maybe double check that, a link to their main page could be more useful.

    [edit, it seemed to work the second time, I dunno, maybe it was just a temporary issue, sorry!]

    Also, it looks like in the last six minutes (as of me writing this post) the README was updated with an intro that starts thing off with a better high-level description of what’s going on. I think the write-up could still use some work. But it isn’t entirely useless, really, this sort of library should be fully correct, down to the ULP, it is quite foundational after all.

  2. rustybolt

    My god this is utterly unreadable for humans.

  3. IshKebab

    ... by one ULP. Anyway I don't think anyone is just feeding a raw sqrt into their synthesiser.

More from this day

2026-11-14