Search All of the Math Forum:
Views expressed in these public forums are not endorsed by
Drexel University or The Math Forum.
|
|
|
|
Re: forcing intel floating registers to use 64-bit arithmetic
Posted:
Feb 3, 2013 10:48 PM
|
|
On 2013-01-11 12:46, Tom Stockfisch wrote:
> I need to force all double precision calculations to proceed strictly in 64-bit -- no 80-bit
> intel register arithmetic. Can someone tell me the current function calls or
> command line settings to achieve this on both MacOS and linux?
>
> I need this to be able to track down platform-dependent differences in numerical code.
>
You have two options:
- Use compiler switches to force intermediates to memory, look for
switches called strict or something like that.
This will affect the computation of expressions like r = a*b+c.
Here the result of a*b will be stored in a 64 bit memory location
and reloaded prior to adding c.
Of course these additional stores and loads are not efficient and
this mode is precisely for the type of inter platform comparisons
which you plan on doing.
- Use single SSE registers but you will probably have to use
Assembler.
Note that the "precision setting" controls the original Intel FPU
only and, contrary to what many posters claim, affects only the
precision of floating point division and square root which are
computed using iterative algorithms.
The remaining arithmetic operations [+,-,*] are always done in
extended precision as are all transcendentals other than square
root such as [cos,sin,log,2^x] and a few others.
Unless you must do billions of divisions or square roots and the
precision really does not matter take my advice and leave the FPU
in extended precision.
If your code contains no transcendentals you may get identical
results between different architectures using these two approaches.
Considering transcendentals keep in mind that the Intel FPU
supports only the minimum required set of transcendentals and that
all transcendentals which you call from your program will be library
functions which are built on top of the former or which are simple
wrappers which do range and error checking.
I am not sure whether "strict" compiler switches imply the use of
less precise transcendentals, you should check.
Good luck consolidating transcendentals between machines.
As a previous poster remarked, small differences may occur between
compilers because the language standard leaves the order of
evaluation of expressions unspecified. So you use parentheses when
needed to remove such ambiguities.
I could say a lot more but have not much time so I will leave it at
this. Hope this helps!
|
|
|
|
