Re: [PATCH v2 0/8] centralize AVX-512 feature detection

[Bruce Richardson <bruce.richardson@intel.com>]

On Tue, Oct 08, 2024 at 10:49:39AM +0200, David Marchand wrote:
> On Tue, Oct 1, 2024 at 1:19 PM Bruce Richardson
> <bruce.richardson@intel.com> wrote:
> >
> > The meson code to detect CPU and compiler support for AVX512 was duplicated
> > across multiple drivers. Do all detection in just a single place to simplify
> > the code.
> >
> > v2: ensure that target_has_avx512 is always defined on x86 to fix build errors
> >
> > Bruce Richardson (8):
> >   config/x86: add global defines for checking AVX-512
> >   event/dlb2: use global AVX-512 variables
> >   common/idpf: use global AVX-512 variables
> >   net/cpfl: use global AVX-512 variables
> >   net/i40e: use global AVX-512 variables
> >   net/iavf: use global AVX-512 variables
> >   net/ice: use global AVX-512 variables
> >   net/idpf: use global AVX-512 variables
> >
> >  config/x86/meson.build          | 19 +++++++++++----
> >  drivers/common/idpf/meson.build | 17 ++-----------
> >  drivers/event/dlb2/meson.build  | 42 +++++++--------------------------
> >  drivers/net/cpfl/meson.build    | 19 ++-------------
> >  drivers/net/i40e/meson.build    | 13 ++--------
> >  drivers/net/iavf/meson.build    | 13 ++--------
> >  drivers/net/ice/meson.build     | 15 ++----------
> >  drivers/net/idpf/meson.build    | 19 ++-------------
> >  8 files changed, 36 insertions(+), 121 deletions(-)
> 
> Thanks for this cleanup, I have two comments.
> 
> - Some drivers were going into great lenghts to check that individiual
> avx512 features were available.
> With this series, we end up requiring support for all features to
> announce avx512 availability.
> Are we perhaps disabling AVX512 support with some toolchains, out
> there, supporting only part of the set?
> 

The various AVX-512 feature sets checked for (F, BW, VL, DQ) were all
introduced in the same hardware generation - all are available in gcc when
using -march=skylake-avx512 or later, or -march=znver4. On the toolchain
side, gcc introduced all these flags simultaneously in gcc-6 [1]. For
clang/llvm, testing with godbolt for compiler errors/warnings indicates
that all these 4 avx512 flags are available from clang 3.6 - the minimum we
support in DPDK [2]

[1] https://gcc.gnu.org/gcc-6/changes.html
[2] https://doc.dpdk.org/guides/linux_gsg/sys_reqs.html#compilation-of-the-dpdk

> - Some drivers were checking for presence of -mno-avx512f in
> machine_args as a way to disable building any AVX512 stuff.
> This gets discarded with this series.
> 

Yes, because it should no longer be necessary. The places in the build
system where we set the no-avx512f flag are reworked so that we don't have
cc_has_avx512 set.

Regards,
/Bruce