options “--enable-vector-intrinsics=type” in arm

23 views
Skip to first unread message

hindsight

unread,
May 7, 2020, 2:13:33 AM5/7/20
to libflame-discuss
Dear libflame discussion group,

I have some problems with the compilation option “--enable-vector-intrinsics = type”.

In the option description, this option only supports sse and none. 

In the arm architecture, there is neon, but it is not supported yet.In the future, will the neon option be added?

And what is the specific role of this compilation option?

Looking forward to your reply.

Thanks,

steven.

Field G. Van Zee

unread,
May 8, 2020, 8:41:21 AM5/8/20
to libflame...@googlegroups.com
Hi Steven,

Thanks for your questions.

Unfortunately, we haven't had the time or resources to extend libflame's
underlying intrinsics kernels to support anything beyond SSE. This
includes more recent Intel-based vector extensions (AVX[2]) as well as
those for other architectures (ARM neon). This may happen someday, but
for now most of our attention has been on BLIS.

As for your second question, this option currently controls the
implementation of kernels that apply Givens rotations. I believe that is
the only kernel we have in libflame. It's a bit of an oddball,
admittedly. The kernel belongs in BLIS, but BLIS didn't exist when those
kernels were written, so we had no where else to put them. If they were
to be updated, what will most likely happen is that libflame's operation
for applying Givens rotations will be added to BLIS, where it is more
easily optimized, and then libflame will be changed to call that
operation in BLIS. At that time, the kernels in libflame will be removed
altogether, along with the --enable-vector-intrinsics configure option.

If you have any additional questions or issues you'd like to raise with
libflame developers, I invite you to use the Issue feature on GitHub [1].

Regards,
Field

[1] https://github.com/flame/libflame/issues
> --
> You received this message because you are subscribed to the Google
> Groups "libflame-discuss" group.
> To unsubscribe from this group and stop receiving emails from it, send
> an email to libflame-discu...@googlegroups.com
> <mailto:libflame-discu...@googlegroups.com>.
> To view this discussion on the web visit
> https://groups.google.com/d/msgid/libflame-discuss/7f03007d-b5ae-4dd6-8e4f-f9827d12c782%40googlegroups.com
> <https://groups.google.com/d/msgid/libflame-discuss/7f03007d-b5ae-4dd6-8e4f-f9827d12c782%40googlegroups.com?utm_medium=email&utm_source=footer>.

hindsight

unread,
May 8, 2020, 9:39:42 PM5/8/20
to libflame-discuss
Thank you Field, it is clear to me now.
Regards,
Steven

在 2020年5月8日星期五 UTC+8下午8:41:21,Field Van Zee写道:
Reply all
Reply to author
Forward
0 new messages