This is the mail archive of the mailing list for the glibc project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH 9/N v2] [x86_64] Vectorized math functions

On Thu, 11 Dec 2014, Andrew Senkevich wrote:

> 2014-12-11 19:48 GMT+03:00 Joseph Myers <>:
> > On Mon, 8 Dec 2014, Andrew Senkevich wrote:
> >
> >> +vlen2-arch-ext-cflags = -msse4
> >> +vlen4-arch-ext-cflags = -mavx
> >
> > I think these variables should have names starting "double-" to reflect
> > that they are for tests for double (there's no particular reason, if
> > vector functions for float are added in future, for them to use the same
> > flags for a given vector length as double tests; it's quite plausible
> > float-vlen4 and double-vlen2 would use the same flags, being for the same
> > bit-length of vectors).
> Yes, I just added cosf and catch it also. May be simply
> arch-ext1-cflags and arch-ext2-cflags to avoid such duplication of the
> same flags?

I'm not clear what abstraction you are proposing here - what properties 
"test X and test Y use the same flags" you think will be the same between 
architectures.  It's not necessarily the case that the same options are 
applicable to 128-bit vectors of 4 x float and 128-bit vectors of 2 x 
double, for example (I don't think x86 is unique in having float vectors - 
SSE - come before double vectors - SSE2 - and it's up to the people adding 
vector functions to work out what versions are useful).

Without something clear shared between architectures, I think you need 
double-vlen2-arch-ext-cflags and double-vlen4-arch-ext-cflags and further 
such variables for float tests.

Joseph S. Myers

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]