[dpdk-dev] [PATCH v4 3/3] efd: run-time dispatch over x86 EFD functions

[Ananyev, Konstantin]

> -----Original Message-----
> From: Li, Xiaoyun
> Sent: Tuesday, October 3, 2017 9:15 AM
> To: Ananyev, Konstantin <konstantin.ananyev at intel.com>; Richardson, Bruce <bruce.richardson at intel.com>
> Cc: Lu, Wenzhuo <wenzhuo.lu at intel.com>; Zhang, Helin <helin.zhang at intel.com>; dev at dpdk.org
> Subject: RE: [PATCH v4 3/3] efd: run-time dispatch over x86 EFD functions
> 
> Hi
> 
> > -----Original Message-----
> > From: Ananyev, Konstantin
> > Sent: Tuesday, October 3, 2017 00:52
> > To: Li, Xiaoyun <xiaoyun.li at intel.com>; Richardson, Bruce
> > <bruce.richardson at intel.com>
> > Cc: Lu, Wenzhuo <wenzhuo.lu at intel.com>; Zhang, Helin
> > <helin.zhang at intel.com>; dev at dpdk.org
> > Subject: RE: [PATCH v4 3/3] efd: run-time dispatch over x86 EFD functions
> >
> >
> >
> > > -----Original Message-----
> > > From: Li, Xiaoyun
> > > Sent: Monday, October 2, 2017 5:13 PM
> > > To: Ananyev, Konstantin <konstantin.ananyev at intel.com>; Richardson,
> > > Bruce <bruce.richardson at intel.com>
> > > Cc: Lu, Wenzhuo <wenzhuo.lu at intel.com>; Zhang, Helin
> > > <helin.zhang at intel.com>; dev at dpdk.org; Li, Xiaoyun
> > > <xiaoyun.li at intel.com>
> > > Subject: [PATCH v4 3/3] efd: run-time dispatch over x86 EFD functions
> > >
> > > This patch dynamically selects x86 EFD functions at run-time.
> > > This patch uses function pointer and binds it to the relative function
> > > based on CPU flags at constructor time.
> > >
> > > Signed-off-by: Xiaoyun Li <xiaoyun.li at intel.com>
> > > ---
> > >  lib/librte_efd/rte_efd_x86.h | 41
> > > ++++++++++++++++++++++++++++++++++++++---
> > >  1 file changed, 38 insertions(+), 3 deletions(-)
> > >
> > > diff --git a/lib/librte_efd/rte_efd_x86.h
> > > b/lib/librte_efd/rte_efd_x86.h index 34f37d7..93b6743 100644
> > > --- a/lib/librte_efd/rte_efd_x86.h
> > > +++ b/lib/librte_efd/rte_efd_x86.h
> > > @@ -43,12 +43,29 @@
> > >  #define EFD_LOAD_SI128(val) _mm_lddqu_si128(val)  #endif
> > >
> > > +typedef efd_value_t
> > > +(*efd_lookup_internal_avx2_t)(const efd_hashfunc_t *group_hash_idx,
> > > +		const efd_lookuptbl_t *group_lookup_table,
> > > +		const uint32_t hash_val_a, const uint32_t hash_val_b);
> > > +
> > > +static efd_lookup_internal_avx2_t efd_lookup_internal_avx2_ptr;
> > > +
> > >  static inline efd_value_t
> > >  efd_lookup_internal_avx2(const efd_hashfunc_t *group_hash_idx,
> > >  		const efd_lookuptbl_t *group_lookup_table,
> > >  		const uint32_t hash_val_a, const uint32_t hash_val_b)  { -
> > #ifdef
> > > RTE_MACHINE_CPUFLAG_AVX2
> > > +	return (*efd_lookup_internal_avx2_ptr)(group_hash_idx,
> > > +					       group_lookup_table,
> > > +					       hash_val_a, hash_val_b);
> >
> > I don't think you need all that.
> > All you need - build proper avx2 function even if current HW doesn't support
> > it.
> > The existing runtime selection here seems ok already.
> > Konstantin
> >
> 
> Sorry, not quite understand here. So don't need to change codes of efd here?
> I didn't care about the HW. CC_SUPPORT_AVX2 only means the compiler supports AVX2 since would runtime selection.
> The existing codes RTE_MACHINE_CPUFLAG_AVX2 means both the compiler and HW supports AVX2.

What I am saying - you don't need all these dances with extra function pointer.
All you need - move efd_lookup_internal_avx2() into a .c file and make sure it get compiled with -mavx2 flag.
Then at rte_efd_create() select AVX2 only when both HW and compiler supports AVX2:

...
#ifdef CC_SUPPORT_AVX2
  if (RTE_EFD_VALUE_NUM_BITS > 3 && rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2))
                table->lookup_fn = EFD_LOOKUP_AVX2;
        else
#endif
...

Konstantin

> 
> 
> Best Regards,
> Xiaoyun Li
> 
> 
> > > +}
> > > +
> > > +#ifdef CC_SUPPORT_AVX2
> > > +static inline efd_value_t
> > > +efd_lookup_internal_avx2_AVX2(const efd_hashfunc_t *group_hash_idx,
> > > +		const efd_lookuptbl_t *group_lookup_table,
> > > +		const uint32_t hash_val_a, const uint32_t hash_val_b) {
> > >  	efd_value_t value = 0;
> > >  	uint32_t i = 0;
> > >  	__m256i vhash_val_a = _mm256_set1_epi32(hash_val_a); @@ -
> > 74,13
> > > +91,31 @@ efd_lookup_internal_avx2(const efd_hashfunc_t
> > *group_hash_idx,
> > >  	}
> > >
> > >  	return value;
> > > -#else
> > > +}
> > > +#endif
> > > +
> > > +static inline efd_value_t
> > > +efd_lookup_internal_avx2_DEFAULT(const efd_hashfunc_t
> > *group_hash_idx,
> > > +		const efd_lookuptbl_t *group_lookup_table,
> > > +		const uint32_t hash_val_a, const uint32_t hash_val_b) {
> > >  	RTE_SET_USED(group_hash_idx);
> > >  	RTE_SET_USED(group_lookup_table);
> > >  	RTE_SET_USED(hash_val_a);
> > >  	RTE_SET_USED(hash_val_b);
> > >  	/* Return dummy value, only to avoid compilation breakage */
> > >  	return 0;
> > > -#endif
> > > +}
> > >
> > > +static void __attribute__((constructor))
> > > +rte_efd_x86_init(void)
> > > +{
> > > +#ifdef CC_SUPPORT_AVX2
> > > +	if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2))
> > > +		efd_lookup_internal_avx2_ptr =
> > efd_lookup_internal_avx2_AVX2;
> > > +	else
> > > +		efd_lookup_internal_avx2_ptr =
> > efd_lookup_internal_avx2_DEFAULT;
> > > +#else
> > > +	efd_lookup_internal_avx2_ptr = efd_lookup_internal_avx2_DEFAULT;
> > > +#endif
> > >  }
> > > --
> > > 2.7.4