80 bit variables for intense math functions?

[vibrantcascade]

Is it possible in ifort to somehow retain 80 bit precision in variables to prevent rounding losses? The performance hit going from double to quad precision is huge, and if I could get an extra 4 or 5 digits of precision without a performance hit due to the math being done in 80 bit anyhow it would greatly help my program.

I know some fortran compilers support real*10 or real(kind=10) but I can't seem to find the equivalent in ifort.
I've tried -double-size 80 as well to no avail.

Is there by chance a hidden feature which allows me to make all my double precision variables 80 bits or am I out of luck?

[TimP]

ifort doesn't have kind=10. With -mia32 and -pc80 (32-bit mode only) you can get some expression evaluation in that mode, but not for stored results.

[jimdempseyatthecove]

You could create a user defined type of 80-bits...
and define operator functions for this type which calls C functions that perform the FPU operations.
Actual implementation may be harder. E.g. trig functions will balk at using your type, but you can write your own functions which call out to your C routines. Though execution is via function call it is likely much faster than emulation.

Jim Dempsey