[PATCH v2 1/1] mldev: split bfloat16 routines to separate files

[Srikanth Yalavarthi]

> -----Original Message-----
> From: Srikanth Yalavarthi <syalavarthi at marvell.com>
> Sent: 15 March 2023 16:12
> To: Ruifeng Wang <Ruifeng.Wang at arm.com>
> Cc: dev at dpdk.org; Shivah Shankar Shankar Narayan Rao
> <sshankarnara at marvell.com>; david.marchand at redhat.com; nd
> <nd at arm.com>; Srikanth Yalavarthi <syalavarthi at marvell.com>; Srikanth
> Yalavarthi <syalavarthi at marvell.com>
> Subject: RE: [PATCH v2 1/1] mldev: split bfloat16 routines to separate files
> 
> > -----Original Message-----
> > From: Ruifeng Wang <Ruifeng.Wang at arm.com>
> > Sent: 15 March 2023 15:32
> > To: Srikanth Yalavarthi <syalavarthi at marvell.com>
> > Cc: dev at dpdk.org; Shivah Shankar Shankar Narayan Rao
> > <sshankarnara at marvell.com>; david.marchand at redhat.com; nd
> <nd at arm.com>
> > Subject: [EXT] RE: [PATCH v2 1/1] mldev: split bfloat16 routines to
> > separate files
> >
> > External Email
> >
> > ----------------------------------------------------------------------
> > > -----Original Message-----
> > > From: Srikanth Yalavarthi <syalavarthi at marvell.com>
> > > Sent: Monday, March 13, 2023 8:03 PM
> > > To: Srikanth Yalavarthi <syalavarthi at marvell.com>; Ruifeng Wang
> > > <Ruifeng.Wang at arm.com>
> > > Cc: dev at dpdk.org; sshankarnara at marvell.com;
> > david.marchand at redhat.com
> > > Subject: [PATCH v2 1/1] mldev: split bfloat16 routines to separate
> > > files
> > >
> > > Since bfloat16 intrinsics are not supported on all ARM platforms
> > > that support NEON,
> > > bfloat16 routines are moved to separate files.
> > > This would enable using scalar implementation for bfloat16 on
> > > unsupported
> > ARM platforms.
> > >
> > > Bugzilla ID: 1179
> > > Fixes: fc54766b1612 ("mldev: add Arm NEON type conversion")
> > >
> > > Signed-off-by: Srikanth Yalavarthi <syalavarthi at marvell.com>
> > > ---
> > > Depends-on: patch-120653 ("mldev: remove weak symbols use in type
> > > conversions")
> > > Depends-on: patch-125035 ("mldev: fix identical code in conditional
> > > branches")
> > >
> > >  lib/mldev/meson.build                   |  11 +-
> > >  lib/mldev/mldev_utils_neon.c            | 142 +------------
> > >  lib/mldev/mldev_utils_neon_bfloat16.c   | 154 ++++++++++++++
> > >  lib/mldev/mldev_utils_scalar.c          | 262 +-----------------------
> > >  lib/mldev/mldev_utils_scalar.h          |  80 ++++++++
> > >  lib/mldev/mldev_utils_scalar_bfloat16.c | 197 ++++++++++++++++++
> > >  6 files changed, 445 insertions(+), 401 deletions(-)  create mode
> > > 100644 lib/mldev/mldev_utils_neon_bfloat16.c
> > >  create mode 100644 lib/mldev/mldev_utils_scalar.h  create mode
> > > 100644 lib/mldev/mldev_utils_scalar_bfloat16.c
> > >
> > > diff --git a/lib/mldev/meson.build b/lib/mldev/meson.build index
> > > c9db42257b..5769b0640a
> > > 100644
> > > --- a/lib/mldev/meson.build
> > > +++ b/lib/mldev/meson.build
> > > @@ -7,12 +7,21 @@ sources = files(
> > >          'mldev_utils.c',
> > >  )
> > >
> > > -if dpdk_conf.has('RTE_ARCH_ARM64')
> > > +if (dpdk_conf.has('RTE_ARCH_ARM64') and
> > > +    cc.get_define('__ARM_NEON', args: machine_args) != '')
> >
> > I found in ACLE document that "__ARM_NEON" is always set to 1 for
> > AArch64".
> > So this line of check is redundant?
> 
> Checking for __ARM_NEON should be enough.
> We can drop the  dpdk_conf.has('RTE_ARCH_ARM64') check.
> I will test the builds and submit a revised patch.
> 

Correction. Ideally checking for RTE_ARCH_ARM64 is enough.

But, __ARM_NEON check is required when building with gcc-4.8.x I have tested this on CentOS-7 with GCC-4.8.5
Refer https://bugs.dpdk.org/show_bug.cgi?id=1179

Below errors, are reported with GCC-4.8, when __ARM_NEON check is not used

../lib/mldev/mldev_utils_neon.c:220:2: warning: nested extern declaration of 'vcvtas_u32_f32' [-Wnested-externs]
../lib/mldev/mldev_utils_neon.c: In function '__uint8_to_float32_neon_f32x1':
../lib/mldev/mldev_utils_neon.c:297:2: warning: implicit declaration of function 'vcvts_f32_u32' [-Wimplicit-function-declaration]
  *output = scale * vcvts_f32_u32((uint32_t)*input);
  ^
../lib/mldev/mldev_utils_neon.c:297:2: warning: nested extern declaration of 'vcvts_f32_u32' [-Wnested-externs]
../lib/mldev/mldev_utils_neon.c: At top level:
../lib/mldev/mldev_utils_neon.c:604:51: error: unknown type name 'float16_t'
 __float32_to_float16_neon_f16x4(float32_t *input, float16_t *output)
                                                   ^

So, we will need both checks.


> >
> > >      sources += files('mldev_utils_neon.c')  else
> > >      sources += files('mldev_utils_scalar.c')  endif
> > >
> > > +if (dpdk_conf.has('RTE_ARCH_ARM64') and
> > > +    cc.get_define('__ARM_NEON', args: machine_args) != '' and
> >
> > Same here.
> >
> > > +    cc.get_define('__ARM_FEATURE_BF16', args: machine_args) != '')
> > > +    sources += files('mldev_utils_neon_bfloat16.c')
> > > +else
> > > +    sources += files('mldev_utils_scalar_bfloat16.c')
> > > +endif
> > > +
> > >  headers = files(
> > >          'rte_mldev.h',
> > >  )
> > <snip>