* Re: musl libmath (implied algorithm issue)
[not found] <CAJDAfTDD+x3zS0kbbNU=5XzGqmNTene1HmZiG44feDXpkb=s2A@mail.gmail.com>
@ 2018-04-01 20:42 ` Szabolcs Nagy
0 siblings, 0 replies; only message in thread
From: Szabolcs Nagy @ 2018-04-01 20:42 UTC (permalink / raw)
To: Alba Pompeo; +Cc: musl
[-- Attachment #1: Type: text/plain, Size: 497 bytes --]
* Alba Pompeo <albapompeo@gmail.com> [2018-04-01 10:54:54 -0300]:
> I found a commit elsewhere that implies musl's fma algorithm is wrong.
> You may be interested in investigating.
> https://git.eigenstate.org/npnth/mc.git/commit/?h=libmath&id=debd13efe12eaca072229e9859aaa6df8708a09f
thanks, it is a wrong result bug in fmaf, copied from freebsd.
near half way cases may get incorrectly rounded under some
circumstances, i didnt try to understand their fix, but
used a minimal change for now.
[-- Attachment #2: 0001-fix-fmaf-wrong-result.patch --]
[-- Type: text/x-diff, Size: 1418 bytes --]
From 2a38f5d89f7a69d03861a47a7f981136ee60630f Mon Sep 17 00:00:00 2001
From: Szabolcs Nagy <nsz@port70.net>
Date: Sun, 1 Apr 2018 20:02:01 +0000
Subject: [PATCH] fix fmaf wrong result
if double precision r=x*y+z is not a half way case between two single
precision floats or it is an exact result then fmaf returns (float)r.
however the exactness check was wrong when |x*y| < |z| and could cause
incorrectly rounded result in nearest rounding mode when r is a half
way case.
fmaf(-0x1.26524ep-54, -0x1.cb7868p+11, 0x1.d10f5ep-29)
was incorrectly rounded up to 0x1.d117ap-29 instead of 0x1.d1179ep-29.
(exact result is 0x1.d1179efffffffecp-29, r is 0x1.d1179fp-29)
this example was reported in
https://git.eigenstate.org/npnth/mc.git/commit/?h=libmath&id=debd13efe12eaca072229e9859aaa6df8708a09f
---
src/math/fmaf.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/src/math/fmaf.c b/src/math/fmaf.c
index aa57feb6..80f5cd8a 100644
--- a/src/math/fmaf.c
+++ b/src/math/fmaf.c
@@ -50,7 +50,7 @@ float fmaf(float x, float y, float z)
/* Common case: The double precision result is fine. */
if ((u.i & 0x1fffffff) != 0x10000000 || /* not a halfway case */
e == 0x7ff || /* NaN */
- result - xy == z || /* exact */
+ (result - xy == z && result - z == xy) || /* exact */
fegetround() != FE_TONEAREST) /* not round-to-nearest */
{
/*
--
2.16.3
^ permalink raw reply [flat|nested] only message in thread