[PATCH v2 1/5] examples/l3fwd: fix port group mask generation

[David Christensen]

On 9/2/22 2:18 AM, pbhagavatula at marvell.com wrote:
> From: Pavan Nikhilesh <pbhagavatula at marvell.com>
> 
> Fix port group mask generation in altivec, vec_any_eq returns
> 0 or 1 while port_groupx4 expects comparison mask result.
> 
> Fixes: 2193b7467f7a ("examples/l3fwd: optimize packet processing on powerpc")
> Cc: stable at dpdk.org
> 
> Signed-off-by: Pavan Nikhilesh <pbhagavatula at marvell.com>
> ---
>   v2 Changes:
>   - Fix PPC, RISC-V, aarch32 compilation.
> 
>   examples/common/altivec/port_group.h | 11 +++++++++--
>   1 file changed, 9 insertions(+), 2 deletions(-)
> 
> diff --git a/examples/common/altivec/port_group.h b/examples/common/altivec/port_group.h
> index 5e209b02fa..592ef80b7f 100644
> --- a/examples/common/altivec/port_group.h
> +++ b/examples/common/altivec/port_group.h
> @@ -26,12 +26,19 @@ port_groupx4(uint16_t pn[FWDSTEP + 1], uint16_t *lp,
>   		uint16_t u16[FWDSTEP + 1];
>   		uint64_t u64;
>   	} *pnum = (void *)pn;
> +	union u_vec {
> +		__vector unsigned short v_us;
> +		unsigned short s[8];
> +	};
> 
> +	union u_vec res;
>   	int32_t v;
> 
> -	v = vec_any_eq(dp1, dp2);
> -
> +	dp1 = (__vector unsigned short)vec_cmpeq(dp1, dp2);

Altivec vec_cmpeq() is similar to Intel _mm_cmpeq_*(), so this looks 
right to me.

> +	res.v_us = dp1;
> 
> +	v = (res.s[0] & 0x1) | (res.s[1] & 0x2) | (res.s[2] & 0x4) |
> +	    (res.s[3] & 0x8);

This can be vectorized too.  The Intel _mm_unpacklo_epi16() intrinsic 
can be replaced with the following Altivec code:

extern __inline __m128i __attribute__((__gnu_inline__, 
__always_inline__, __artificial__))
_mm_unpacklo_epi16 (__m128i __A, __m128i __B)
{
   return (__m128i) vec_mergeh ((__v8hi)__A, (__v8hi)__B);
}

The Intel _mm_movemask_ps() intrinsic can be replaced with the following 
Altivec implementation:

/* Creates a 4-bit mask from the most significant bits of the SPFP 
values.  */
extern __inline int __attribute__((__gnu_inline__, __always_inline__, 
__artificial__))
_mm_movemask_ps (__m128  __A)
{
   __vector unsigned long long result;
   static const __vector unsigned int perm_mask =
     {
#ifdef __LITTLE_ENDIAN__
         0x00204060, 0x80808080, 0x80808080, 0x80808080
#else
       0x80808080, 0x80808080, 0x80808080, 0x00204060
#endif
     };

   result = ((__vector unsigned long long)
             vec_vbpermq ((__vector unsigned char) __A,
                          (__vector unsigned char) perm_mask));

#ifdef __LITTLE_ENDIAN__
   return result[1];
#else
   return result[0];
#endif
}

Dave