[dpdk-dev] [PATCH v1 1/2] rte_memcmp functions using Intel AVX and SSE intrinsics
Ferruh Yigit
ferruh.yigit at intel.com
Fri Dec 21 00:30:10 CET 2018
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/
More information about the dev
mailing list