Re: GSL on OS/2

[Henry Sobotka <sobotka at axess dot com>]

Brian Gough wrote:
> 
> The errors picked up there are large compared to exact agreement but ok in
> those cases, in terms of the underlying test.  Useful to check anyway.

I'm curious to see what effect varying optimization levels and other
pertinent compiler flags might have on the margins of error.

> We should probably start running our own tests with fdlibm in order to
> take account of variation between math libraries, since fdlibm is
> supposed to be strictly IEEE compliant.  I ran a quick check but did
> not get any failures on specfunc (or complex) with fdlibm so it may
> depend on other factors as well.

One factor at this end is linkage. Our libm is automatically included
when we link dynamically with C-runtime. This puts two versions of math
functions within calling range when also linking with fdlibm. Moreover,
specifying -lm may be pulling in a third set of static copies. Strictly
static linkage should eliminate this potential source of havoc.

Another factor may be the precision discrepancy between IEEE 754 and our
default 64-bit mantissa.

 > Let us know when your OS/2 IEEE routine is ready and I will add it.

Attached.

thx,

h~

gsldiffs.os2.gz

fp-os2emx.c.gz