[dpdk-dev] [PATCH v4 1/8] eal: introduce zmm type for AVX 512-bit

[David Marchand]

On Thu, Jul 9, 2020 at 4:52 PM Medvedkin, Vladimir
<vladimir.medvedkin at intel.com> wrote:
> >> +
> >> +#define        ZMM_SIZE        (sizeof(__x86_zmm_t))
> >> +#define        ZMM_MASK        (ZMM_SIZE - 1)
> >
> > Macros in a public header need a RTE_ prefix + this is x86 specific,
> > then RTE_X86_.
> >
> > Looking at the whole applied series:
> > $ git grep -lw ZMM_SIZE
> > lib/librte_eal/x86/include/rte_vect.h
> > $ git grep -lw ZMM_MASK
> > lib/librte_eal/x86/include/rte_vect.h
> >
> > So I wonder if we need to export it or we can instead just #undef
> > after the struct definition.
>
> I think it's better to undef it

Even if you undef the macro, please still prefix it.
This is to avoid conflicts with macros defined before including this
rte_vect.h header.


>
> >
> >
> >> +
> >> +typedef union __rte_x86_zmm  {
> >> +       __x86_zmm_t      z;
> >> +       ymm_t    y[ZMM_SIZE / sizeof(ymm_t)];
> >> +       xmm_t    x[ZMM_SIZE / sizeof(xmm_t)];
> >> +       uint8_t  u8[ZMM_SIZE / sizeof(uint8_t)];
> >> +       uint16_t u16[ZMM_SIZE / sizeof(uint16_t)];
> >> +       uint32_t u32[ZMM_SIZE / sizeof(uint32_t)];
> >> +       uint64_t u64[ZMM_SIZE / sizeof(uint64_t)];
> >> +       double   pd[ZMM_SIZE / sizeof(double)];
> >> +} __rte_aligned(ZMM_SIZE) __rte_x86_zmm_t;
> >
> > I don't understand this forced alignment statement.
> > Would not natural alignment be enough, since all fields in this union
> > have the same size?
> >
>
> Some compilers won't align this union
> https://mails.dpdk.org/archives/dev/2020-March/159591.html

Ok, interesting, I will try to keep in mind.


-- 
David Marchand