| Re: Floating Point Bug on Intel Pentium CPUs [message #3156] |
Fri, 25 November 1994 05:51 |
nowicki
Messages: 11
Registered: May 1993
|
Junior Member |
|
|
In article <3b04ks$3c6@jhunix1.hcf.jhu.edu>, murthy@pha.jhu.edu (Jayant Murthy) writes:
> Gregory D. Nowicki (nowicki@tardis.larc.nasa.gov) wrote:
> : In article <3avsus$22n@spool.cs.wisc.edu>, Liam Gumley <liamg@ssec.wisc.edu> writes:
> : >This is not strictly IDL related, but may affect IDL users on Pentium
> : >platforms who run complicated models. Recent discussion in the
> : >comp.sys.intel newsgroup has focussed on the discovery of a bug
> : >in the floating point unit on Intel Pentium CPUs. The bug is evident
> : >during divide operations. The following code reportedly will show the
> : >bug. It should give zero (or near enough to it), but on the Pentium
> : >it returns z = 256.
> : >
> : >x = 4195835.
> : >y = 3145727.
> : >z = x - (x/y)*y
> : >
> : >Apparently some kind of chip replacement strategy is being considered
> : >by Intel. For more details, tune to comp.sys.intel
> : >
> : >Cheers,
> : >Liam.
> : >liamg@ssec.wisc.edu
> : >
>
> : Above problem only there if using DOUBLE precision math, (DOUBLE in IDL,
> : real*8 in Fortran, double in C, etc.)
>
> : -Greg
>
> : /* Greg Nowicki | Mail Stop 401A | LIDAR Applications Group */
> : /* NASA Langley Research Center | Hampton, Virginia 23681-0001 */
> : /* Voice: (804) 864-2713 | FAX: (804) 864-7790 */
> : /* nowicki@tardis.larc.nasa.gov | My opinions and mine alone . . . */
>
> Not true. I get it using plain old FLOATs on my P90.
>
> Jayant
> --
>
> Jayant Murthy murthy@pha.jhu.edu
> 410-516-7027
Sorry to repost all the above stuff, but much is lost if I cut some out.
You must have found a different bug in compiler, chip, or ??? This FDIV
bug is said to ONLY affect numbers when doing double precision floating
point math, and the precision that is affected is said to be below the
threshold of pecision using any other type of math.
The above info has been gleened from comp.sys.intel, where if this
discussion continues, I suggest we more it to.
Sorry to continue off the topic of IDL.
-Greg
/* Greg Nowicki | Mail Stop 401A | LIDAR Applications Group */
/* NASA Langley Research Center | Hampton, Virginia 23681-0001 */
/* Voice: (804) 864-2713 | FAX: (804) 864-7790 */
/* nowicki@tardis.larc.nasa.gov | My opinions and mine alone . . . */
|
|
|
|
| Re: Floating Point Bug on Intel Pentium CPUs [message #3167 is a reply to message #3156] |
Wed, 23 November 1994 11:19  |
murthy
Messages: 7
Registered: February 1994
|
Junior Member |
|
|
Gregory D. Nowicki (nowicki@tardis.larc.nasa.gov) wrote:
: In article <3avsus$22n@spool.cs.wisc.edu>, Liam Gumley <liamg@ssec.wisc.edu> writes:
: >This is not strictly IDL related, but may affect IDL users on Pentium
: >platforms who run complicated models. Recent discussion in the
: >comp.sys.intel newsgroup has focussed on the discovery of a bug
: >in the floating point unit on Intel Pentium CPUs. The bug is evident
: >during divide operations. The following code reportedly will show the
: >bug. It should give zero (or near enough to it), but on the Pentium
: >it returns z = 256.
: >
: >x = 4195835.
: >y = 3145727.
: >z = x - (x/y)*y
: >
: >Apparently some kind of chip replacement strategy is being considered
: >by Intel. For more details, tune to comp.sys.intel
: >
: >Cheers,
: >Liam.
: >liamg@ssec.wisc.edu
: >
: Above problem only there if using DOUBLE precision math, (DOUBLE in IDL,
: real*8 in Fortran, double in C, etc.)
: -Greg
: /* Greg Nowicki | Mail Stop 401A | LIDAR Applications Group */
: /* NASA Langley Research Center | Hampton, Virginia 23681-0001 */
: /* Voice: (804) 864-2713 | FAX: (804) 864-7790 */
: /* nowicki@tardis.larc.nasa.gov | My opinions and mine alone . . . */
Not true. I get it using plain old FLOATs on my P90.
Jayant
--
Jayant Murthy murthy@pha.jhu.edu
410-516-7027
|
|
|
|
| Re: Floating Point Bug on Intel Pentium CPUs [message #3169 is a reply to message #3167] |
Wed, 23 November 1994 10:43 |
nowicki
Messages: 11
Registered: May 1993
|
Junior Member |
|
|
In article <3avsus$22n@spool.cs.wisc.edu>, Liam Gumley <liamg@ssec.wisc.edu> writes:
> This is not strictly IDL related, but may affect IDL users on Pentium
> platforms who run complicated models. Recent discussion in the
> comp.sys.intel newsgroup has focussed on the discovery of a bug
> in the floating point unit on Intel Pentium CPUs. The bug is evident
> during divide operations. The following code reportedly will show the
> bug. It should give zero (or near enough to it), but on the Pentium
> it returns z = 256.
>
> x = 4195835.
> y = 3145727.
> z = x - (x/y)*y
>
> Apparently some kind of chip replacement strategy is being considered
> by Intel. For more details, tune to comp.sys.intel
>
> Cheers,
> Liam.
> liamg@ssec.wisc.edu
>
Above problem only there if using DOUBLE precision math, (DOUBLE in IDL,
real*8 in Fortran, double in C, etc.)
-Greg
/* Greg Nowicki | Mail Stop 401A | LIDAR Applications Group */
/* NASA Langley Research Center | Hampton, Virginia 23681-0001 */
/* Voice: (804) 864-2713 | FAX: (804) 864-7790 */
/* nowicki@tardis.larc.nasa.gov | My opinions and mine alone . . . */
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
