CODEWITHSUNDEEP
gccfloating-pointclangiccfast-math
marcin
marcin

Reputation: 3571

What is GCC/Clang equivalent of -fp-model fast=1 in ICC

As I read on Intel's website:

Intel compiler uses /fp-model fast=1 as defaults. This optimization favors speed over standards compliance. You may use compiler option -mieee-fp to get compliant code.

My understanding of the fp-model option in ICC is that (correct me if I'm wrong):

What options in GCC or Clang would make floating point math most similar to Intel's default -fp-model fast=1?

Upvotes: 5

Views: 3762

Answers (2)

Mingye Wang
Mingye Wang

Reputation: 1374

Let's read what Intel says:

First, Intel disables exceptions by default, so -fno-trapping-math there. But that applies even to "precise" mode, so not quite an answer.

There is a distinction between "precise" and "consistent", but it is not explained.

There is no explanation of re-association and other assumptions like GCC has. A Intel manual (1) says limited reassoc happens at fast=1, partly to enable SIMD reduction. An Intel presentation (2) mentions GCC's funsafe-math-optimizations and lists some assocs, but it is not clarified if those reassocs are the "less risky" ones enabled by fast=1. I would speculate that they are somewhat similar, considering "fun & safe" does have the reputation of being the safer of the bunch.

1 also mentions that:

  • "fast" turns on FMA contract, while "precise" turns it off. -ffp-contract=fast in GCC.
  • fast=1 does not have access to the FPU enviornment, so no FTZ/DAZ.

Behavior that GCC doesn't have

We do have this behavior on choosing what libm implementation to use, which has no equivalent in GCC as far as I know:

The -fp-model and /fp options determine the setting for the maximum allowable relative error for math library function results (max-error) if none of the following options are specified (the following options are only available for ifort):

  • -fimf-accuracy-bits (Linux* and macOS) or /Qimf-accuracy-bits (Windows*)
  • -fimf-max-error (Linux and macOS) or /Qimf-max-error (Windows)
  • -fimf-precision (Linux and macOS) or /Qimf-precision (Windows)
  • [Q]fast-transcendentals

Option -fp-model=fast (and /fp:fast) sets option -fimf-precision=medium (/Qimf-precision:medium) and option -fp-model=precise (and /fp:precise) implies -fimf-precision=high (and /Qimf-precision:high). Option -fp-model=fast=2 (and /fp:fast2) sets option -fimf-precision=medium (and /Qimf-precision:medium) and option -fimf-domain-exclusion=15 (and /Qimf-domain-exclusion=15).

Here -fimf-precision is described as equivalent to -fimf-max-error=4, though the documentation does not tell us what 1 and 4 means (what's a relative error? ULP?).

-fimf-domain-exclusion=15 is a little easier to decipher: it means that math functions assumes that no "extreme normal number inputs", nans, infinites, nor denormals will be passed. That roughly corresponds to, for libm calls only, -fcx-limited-range -ffinite-math-only, plus the denormal handling (x86-specific) -mdaz-ftz.

Upvotes: 2

Konstantin Vladimirov
Konstantin Vladimirov

Reputation: 6999

As it follows from GCC's set_fast_math_flags function, ffast-math option (at least in GCC 5.2) is equivalent to

(1) unsafe opts group:

-fno-trapping-math
-fassociative_math
-fno-signed-zeros
-freciprocal-math

(2) other guys:

-ffinite-math-only
-fno-errno-math
-fno-signaling-nans
-fno-rounding-math
-fcx-limited-range

First group is abbreviated with the option -funsafe-math-optimizations.

You should figure out what comes in ICC and try to combine those flags to make desired effect.

Upvotes: 3

Related Questions