Developer Guide and Reference

Contents

Using the -fp-model (/fp) Option

The
-fp-model
(Linux*
and
macOS*
) or
/fp
(Windows*) option allows you to control the optimizations on floating-point data. You can use this option to tune the performance, level of accuracy, or result consistency for floating-point applications across platforms and optimization levels.
For applications that do not require support for
denormalized
numbers, the
-fp-model
or
/fp
option can be combined with the
[Q]ftz
option to flush
denormalized
results to zero in order to obtain improved runtime performance on processors based on all Intel® architectures.
You can use keywords to specify the semantics to be used. The keywords specified for this option may influence the choice of math routines that are invoked. Many routines in the
libirc
,
libm
, and
libsvml
libraries are more highly optimized for Intel microprocessors than for non-Intel microprocessors. Possible values of the keywords are as follows:
Keyword
Description
precise
Enables value-safe optimizations on floating-point data.
fast[=1|2]
Enables more aggressive optimizations on floating-point data.
consistent
Enables consistent, reproducible results for different optimization levels or between different processors of the same architecture. This setting is equivalent to the use of the following options:
Windows*:
/fp:precise /Qfma- /Qimf-arch-consistency:true
Linux*
and
macOS*
:
-fp-model precise -no-fma -fimf-arch-consistency=true
strict
Enables
precise
and
except
, disables contractions,
and enables pragma
stdc fenv_access
.
source
Rounds intermediate results to source-defined precision and enables value-safe optimizations.
double
Rounds intermediate results to 53-bit (double) precision and enables value-safe optimizations.
extended
Rounds intermediate results to 64-bit (extended) precision and enables value-safe optimizations.
[no-]except
(Linux*
and
macOS*
) or
except[-]
(Windows*)
Determines whether strict floating-point exception semantics are used.
The default value of the option is
-fp-model fast=1
or
/fp:fast=1
, which means that the compiler uses more aggressive optimizations on floating-point calculations.
Using the default option keyword
-fp-model fast
or
/fp:fast
, you may get significant differences in your result depending on whether the compiler uses x87 or SSE/AVX instructions to implement floating-point operations. Results are more consistent when the other option keywords are used.
Several examples are provided to illustrate the usage of the keywords. These examples show:
  • A small example of source code.
    The same source code is considered in all the included examples.
  • The semantics that are used to interpret floating-point calculations in the source code.
  • One or more possible ways the compiler may interpret the source code.
    There are several ways the compiler may interpret the code; we show just some of these possibilities.

-fp-model fast
or
/fp:fast

Example source code:
Example
float t0, t1, t2; ... t0 = 4.0f + 0.1f + t1 + t2;
When this option is specified, the compiler applies the following semantics:
  • Additions may be performed in any order.
  • Intermediate expressions may use
    single, double,
    or
    extended double
    precision.
  • The constant addition may be pre-computed, assuming the default rounding mode.
Using these semantics, some possible ways the compiler may interpret the original code are given below:
Example
float t0, t1, t2; ... t0 = (float)((double)t1 + (double)t2) + 4.1f;
float t0, t1, t2; ... t0 = (t1 + t2) + 4.1f;
float t0, t1, t2; ... t0 = (t1 + 4.1f) + t2;

-fp-model extended
or
/fp:extended

This setting is equivalent to
-fp-model precise
on Linux* operating systems based on the IA-32 architecture.
Example source code:
float t0, t1, t2; ... t0 = 4.0f + 0.1f + t1 + t2;
When this option is specified, the compiler applies the following semantics:
  • Additions are performed in program order
  • Intermediate expressions use extended double precision
  • The constant addition may be pre-computed, assuming the default rounding mode
Using these semantics, a possible way the compiler may interpret the original code is shown below:
float t0, t1, t2; ... t0 = (float)(((long double)4.1 + (long double)t1) + (long double)t2);

-fp-model source
or
/fp:source

This setting is equivalent to
-fp-model precise
or
/fp:precise
on systems based on the Intel® 64 architecture.
Source code example