On Fri, Jun 10, 2022 at 03:41:00PM +0100, McDaniel, Timothy wrote:
>
>
> > -----Original Message-----
> > From: Richardson, Bruce <bruce.richard...@intel.com>
> > Sent: Friday, June 10, 2022 8:12 AM
> > To: McDaniel, Timothy <timothy.mcdan...@intel.com>
> > Cc: jer...@marvell.com; dev@dpdk.org; Wires, Kent <kent.wi...@intel.com>
> > Subject: Re: [PATCH v5] event/dlb2: add support for single 512B write of 4
> > QEs
> >
> > On Fri, Jun 10, 2022 at 07:35:44AM -0500, Timothy McDaniel wrote:
> > > On Xeon, 512b accesses are available, so movdir64 instruction is able to
> > > perform 512b read and write to DLB producer port. In order for movdir64
> > > to be able to pull its data from store buffers (store-buffer-forwarding)
> > > (before actual write), data should be in single 512b write format.
> > > This commit add change when code is built for Xeon with 512b AVX support
> > > to make single 512b write of all 4 QEs instead of 4x64b writes.
> > >
> > > Signed-off-by: Timothy McDaniel <timothy.mcdan...@intel.com>
> > > Acked-by: Kent Wires <kent.wi...@intel.com>
> > > ===
> > >
> > > Changes since V4:
> > > 1) Add build-time control for avx512 support to meson.buildi, based
> > > on implementation found in lib/acl/meson.build
> > > 2) Add rte_vect_get_max_simd_bitwidth runtime check before using
> > > avx512 instructions
> > >
> >
> > Thanks, these changes look better for runtime support. Some further more
> > minor comments inline below.
> >
> > /Bruce
> >
> > > Changes since V3:
> > > 1) Renamed dlb2_noavx512.c to dlb2_sve.c, and fixed up meson.build
> > > for new file name.
> > >
> > > Changes since V1:
> > > 1) Split out dlb2_event_build_hcws into two implementations, one
> > > that uses AVX512 instructions, and one that does not. Each implementation
> > > is in its own source file in order to avoid build errors if the compiler
> > > does not support the newer AVX512 instructions.
> > > 2) Update meson.build to and pull in appropriate source file based on
> > > whether the compiler supports AVX512VL
> > > 3) Check if target supports AVX512VL, and use appropriate implementation
> > > based on this runtime check.
> > > ---
> > > drivers/event/dlb2/dlb2.c | 208 +-----------------------
> > > drivers/event/dlb2/dlb2_avx512.c | 267
> > +++++++++++++++++++++++++++++++
> > > drivers/event/dlb2/dlb2_priv.h | 10 ++
> > > drivers/event/dlb2/dlb2_sve.c | 219 +++++++++++++++++++++++++
> > > drivers/event/dlb2/meson.build | 53 ++++++
> > > 5 files changed, 556 insertions(+), 201 deletions(-)
> > > create mode 100644 drivers/event/dlb2/dlb2_avx512.c
> > > create mode 100644 drivers/event/dlb2/dlb2_sve.c
> > >
> > <snip>
> >
> > > diff --git a/drivers/event/dlb2/meson.build
> > > b/drivers/event/dlb2/meson.build
> > > index f963589fd3..58146e8aef 100644
> > > --- a/drivers/event/dlb2/meson.build
> > > +++ b/drivers/event/dlb2/meson.build
> > > @@ -19,6 +19,59 @@ sources = files(
> > > 'dlb2_selftest.c',
> > > )
> > >
> > > +# compile AVX512 version if:
> > > +# we are building 64-bit binary (checked above) AND binutils
> > > +# can generate proper code
> > > +
> > > +if binutils_ok
> > > +
> > > + # compile AVX512 version if either:
> > > + # a. we have AVX512 supported in minimum instruction set
> > > + # baseline
> > > + # b. it's not minimum instruction set, but supported by
> > > + # compiler
> > > + #
> > > + # in former case, just add avx512 C file to files list
> > > + # in latter case, compile c file to static lib, using correct
> > > + # compiler flags, and then have the .o file from static lib
> > > + # linked into main lib.
> > > +
> > > + # check if all required flags already enabled (variant a).
> > > + dlb2_avx512_flags = ['__AVX512F__', '__AVX512VL__',
> > > + '__AVX512CD__', '__AVX512BW__']
> >
> > Minor nit: are all 4 of these really necessary? I see the runtime portion
> > only seems to check for VL?
> >
>
> I will update to check for just VL
>
> > > +
> > > + dlb2_avx512_on = true
> > > + foreach f:dlb2_avx512_flags
> > > +
> > > + if cc.get_define(f, args: machine_args) == ''
> > > + dlb2_avx512_on = false
> > > + endif
> > > + endforeach
> > > +
> > > + if dlb2_avx512_on == true
> > > +
> > > + sources += files('dlb2_avx512.c')
> > > + cflags += '-DCC_AVX512_SUPPORT'
> > > +
> > > + elif cc.has_multi_arguments('-mavx512f', '-mavx512vl',
> > > + '-mavx512cd', '-mavx512bw')
> > > +
> > > + cflags += '-DCC_AVX512_SUPPORT'
> > > + avx512_tmplib = static_library('avx512_tmp',
> > > + 'dlb2_avx512.c',
> > > + dependencies: [static_rte_eal,
> > > + static_rte_eventdev],
> > > + c_args: cflags +
> > > + ['-mavx512f', '-mavx512vl',
> > > + '-mavx512cd', '-mavx512bw'])
> > > + objs += avx512_tmplib.extract_objects('dlb2_avx512.c')
> > > + else
> > > + sources += files('dlb2_sve.c')
> > > + endif
> > > +else
> > > + sources += files('dlb2_sve.c')
> >
> > Since this is x86 only, do you mean SSE rather than SVE?
> >
> > Also, rather than adding this in the "else" legs, does the SSE version not
> > need to always be compiled in? If the build takes the second leg, i.e.
> > build is not mandating AVX-512, but supports it if not available, is the
> > SSE code path not necessary for the case where the runtime machine does not
> > support AVX-512?
> >
>
> I'll update the name, but it's an "either or" situation. They cannot both be
> built
> as currently coded.
>
If only the AVX-512 path is built, what happens when the runtime check for
AVX-512 fails? Is there a scalar path that is used as fallback?
/Bruce
