[dpdk-dev] [PATCH v5 1/8] eal/x86: introduce AVX 512-bit type

Medvedkin, Vladimir vladimir.medvedkin at intel.com
Mon Jul 13 12:23:27 CEST 2020


Hi Thomas,

On 10/07/2020 22:49, Thomas Monjalon wrote:
> Please Cc those who participated in the review previously.
> Adding Ray, Jerin, David.
> 
> 10/07/2020 16:46, Vladimir Medvedkin:
>> New data type to manipulate 512 bit AVX values.
> [...]
>> +#ifdef __AVX512F__
>> +
>> +#define	RTE_X86_ZMM_SIZE	(sizeof(__m512i))
>> +#define	RTE_X86_ZMM_MASK	(ZMM_SIZE - 1)
> 
> Why do you use tabs?

Will resend v6

> 
>> +
>> +typedef union __rte_x86_zmm  {
> 
> Double space

Will fix in v6

> 
>> +	__m512i	 z;
>> +	ymm_t    y[RTE_X86_ZMM_SIZE / sizeof(ymm_t)];
>> +	xmm_t    x[RTE_X86_ZMM_SIZE / sizeof(xmm_t)];
>> +	uint8_t  u8[RTE_X86_ZMM_SIZE / sizeof(uint8_t)];
>> +	uint16_t u16[RTE_X86_ZMM_SIZE / sizeof(uint16_t)];
>> +	uint32_t u32[RTE_X86_ZMM_SIZE / sizeof(uint32_t)];
>> +	uint64_t u64[RTE_X86_ZMM_SIZE / sizeof(uint64_t)];
>> +	double   pd[RTE_X86_ZMM_SIZE / sizeof(double)];
>> +} __rte_aligned(RTE_X86_ZMM_SIZE) __rte_x86_zmm_t;
>> +
>> +#endif /* __AVX512F__ */
> 
> You were supposed to undef the macros above.

It was intentional. It could be used later by other libs, like XMM_SIZE:
git grep -lw XMM_SIZE
lib/librte_acl/acl_gen.c
lib/librte_acl/acl_run.h
lib/librte_acl/rte_acl.h
lib/librte_eal/arm/include/rte_vect.h
lib/librte_eal/ppc/include/rte_vect.h
lib/librte_eal/x86/include/rte_vect.h
lib/librte_hash/rte_thash.h

> 
> Vladimir, after your recent contributions,
> it seems you are not interested in details.
> Please understand we have to maintain a project with consistency
> and good doc. Please pay attention, thanks.
> 
> 

-- 
Regards,
Vladimir


More information about the dev mailing list