Re: [dpdk-dev] [PATCH v1 1/2] rte_memcmp functions using Intel AVX and SSE intrinsics

[Ferruh Yigit <ferruh.yigit@intel.com>]

On 5/26/2016 9:57 AM, zhihong.wang at intel.com (Wang, Zhihong) wrote:
>> -----Original Message-----
>> From: dev [mailto:dev-bounces at dpdk.org] On Behalf Of Ravi Kerur
>> Sent: Tuesday, March 8, 2016 7:01 AM
>> To: dev at dpdk.org
>> Subject: [dpdk-dev] [PATCH v1 1/2] rte_memcmp functions using Intel AVX and
>> SSE intrinsics
>>
>> v1:
>>         This patch adds memcmp functionality using AVX and SSE
>>         intrinsics provided by Intel. For other architectures
>>         supported by DPDK regular memcmp function is used.
>>
>>         Compiled and tested on Ubuntu 14.04(non-NUMA) and 15.10(NUMA)
>>         systems.
>>
> [...]
> 
>> +	if (unlikely(!_mm_testz_si128(xmm2, xmm2))) {
>> +		__m128i idx =
>> +			_mm_setr_epi8(15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0);
> 
> line over 80 characters ;)
> 
>> +
>> +		/*
>> +		 * Reverse byte order
>> +		 */
>> +		xmm0 = _mm_shuffle_epi8(xmm0, idx);
>> +		xmm1 = _mm_shuffle_epi8(xmm1, idx);
>> +
>> +		/*
>> +		* Compare unsigned bytes with instructions for signed bytes
>> +		*/
>> +		xmm0 = _mm_xor_si128(xmm0, _mm_set1_epi8(0x80));
>> +		xmm1 = _mm_xor_si128(xmm1, _mm_set1_epi8(0x80));
>> +
>> +		return _mm_movemask_epi8(xmm0 > xmm1) -
>> _mm_movemask_epi8(xmm1 > xmm0);
>> +	}
>> +
>> +	return 0;
>> +}
> 
> [...]
> 
>> +static inline int
>> +rte_memcmp(const void *_src_1, const void *_src_2, size_t n)
>> +{
>> +	const uint8_t *src_1 = (const uint8_t *)_src_1;
>> +	const uint8_t *src_2 = (const uint8_t *)_src_2;
>> +	int ret = 0;
>> +
>> +	if (n < 16)
>> +		return rte_memcmp_regular(src_1, src_2, n);
> [...]
>> +
>> +	while (n > 512) {
>> +		ret = rte_cmp256(src_1 + 0 * 256, src_2 + 0 * 256);
> 
> Thanks for the great work!
> 
> Seems to me there's a big improvement area before going into detailed
> instruction layout tuning that -- No unalignment handling here for large
> size memcmp.
> 
> So almost without a doubt the performance will be low in micro-architectures
> like Sandy Bridge if the start address is unaligned, which might be a
> common case.

Patch is waiting for comment for a long time, since 2016 May. Updating patch
status as rejected.

Anyone planning to work on vectorized version of rte_memcmp() can benefit from
this patch:
https://patches.dpdk.org/patch/11156/
https://patches.dpdk.org/patch/11157/