/Qfast-transcendentals with SVML?

/Qfast-transcendentals with SVML?

Hello.

Will the "Q/fast-transcendentals" compiler switch have an effect on the speed/accuracy of the transcendentalintrinsics in the Short Vector Math Library, or is this switch just for scalar code? Thank you.

-L

5 posts / 0 new
Last post
For more complete information about compiler optimizations, see our Optimization Notice.

Quoting - ltkeene
Hello.

Will the "Q/fast-transcendentals" compiler switch have an effect on the speed/accuracy of the transcendentalintrinsics in the Short Vector Math Library, or is this switch just for scalar code? Thank you.

-L

The fast-transcendentals switch simply enables vectorization by svml calls. It should make no difference in comparison to default vectorization, but it may enable svml vectorization when /fp:source or similar options are set. Up to now, there is only one svml function version for each function for each CPU architecture (not a very precise statement, sorry).

Quoting - tim18

The fast-transcendentals switch simply enables vectorization by svml calls. It should make no difference in comparison to default vectorization, but it may enable svml vectorization when /fp:source or similar options are set. Up to now, there is only one svml function version for each function for each CPU architecture (not a very precise statement, sorry).

Hmmm. I'm calling the SVML manually within my loops (specifically the "_mm_sin_ps()" intrinsic). Do you mean that the compiler switch only enables autovectorization of transcendentals, and therefore has no impact on manual vectorization of loops via intrinsics?

Quoting - ltkeene

Hmmm. I'm calling the SVML manually within my loops (specifically the "_mm_sin_ps()" intrinsic). Do you mean that the compiler switch only enables autovectorization of transcendentals, and therefore has no impact on manual vectorization of loops via intrinsics?

That's my understanding. You could certainly generate your own asm code with -S switch, so as to check whether fast-transcendentals has any influence.

Quoting - ltkeene
Hello.

Will the "Q/fast-transcendentals" compiler switch have an effect on the speed/accuracy of the transcendentalintrinsics in the Short Vector Math Library, or is this switch just for scalar code? Thank you.

-L

From our compiler engineer the "/Qfast-transcendentals" switch does NOT affect explicit SVML intrinsics, butonly affects the scalar calls to the standard math library routines. Our documentation will be updated to be more clear on this.

Jennifer

Login to leave a comment.