Two subjects: 64-bit OS2/eCs, Innotek Products

Two subjects: 64-bit OS2/eCs, Innotek Products

Post by Jonathan de Boyne Pollar » Thu, 22 Jan 2004 01:12:49



PM> Several people in this thread have mentioned that wide registers
PM> are needed for big number-crunching applications.  True, but the
PM> registers in that case are the floating point registers, which in
PM> the 80x86 architecture are separate from the main processor
PM> registers.  The floating point registers in these processors have
PM> always been 80 bits wide, and probably always will be.

AS> IIRC, the IEEE standard on floating point maths has something to
AS> say about the required precision and may therefore have a bearing
AS> on this.

I don't have a copy of IEEE 754 to hand, but people have stated that 80 bits
is simply the _minimum_ size for extended precision floating point, single
precision and double precision being more strictly defined.

 
 
 

Two subjects: 64-bit OS2/eCs, Innotek Products

Post by Edvin Catovi » Sat, 24 Jan 2004 08:10:32



Quote:> PM> Several people in this thread have mentioned that wide registers
> PM> are needed for big number-crunching applications.  True, but the
> PM> registers in that case are the floating point registers, which in
> PM> the 80x86 architecture are separate from the main processor
> PM> registers.  The floating point registers in these processors have
> PM> always been 80 bits wide, and probably always will be.

> AS> IIRC, the IEEE standard on floating point maths has something to
> AS> say about the required precision and may therefore have a bearing
> AS> on this.

> I don't have a copy of IEEE 754 to hand, but people have stated that 80 bits
> is simply the _minimum_ size for extended precision floating point, single
> precision and double precision being more strictly defined.

The minimum width of double-extended precision format as defined in the
IEEE-754 standard is 79 bits. An extra bit in the x86 80-bit format is an
explicit leading significand bit (single and double precision formats use
an implicit (hidden) leading significand bit).