[dpdk-dev] [PATCH v1 1/2] rte_memcmp functions using Intel AVX and SSE intrinsics
Wang, Zhihong
zhihong.wang at intel.com
Thu May 26 10:57:30 CEST 2016
> -----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.
> + if (unlikely(ret != 0))
> + return ret;
> +
> + ret = rte_cmp256(src_1 + 1 * 256, src_2 + 1 * 256);
> + if (unlikely(ret != 0))
> + return ret;
> +
> + src_1 = src_1 + 512;
> + src_2 = src_2 + 512;
> + n -= 512;
> + }
> + goto CMP_BLOCK_LESS_THAN_512;
> +}
> +
> +#else /* RTE_MACHINE_CPUFLAG_AVX2 */
More information about the dev
mailing list