From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mga04.intel.com (mga04.intel.com [192.55.52.120]) by dpdk.org (Postfix) with ESMTP id 595C88E93 for ; Thu, 14 Jan 2016 14:16:48 +0100 (CET) Received: from fmsmga001.fm.intel.com ([10.253.24.23]) by fmsmga104.fm.intel.com with ESMTP; 14 Jan 2016 05:16:48 -0800 X-ExtLoop1: 1 X-IronPort-AV: E=Sophos;i="5.22,294,1449561600"; d="scan'208";a="881268072" Received: from unknown (HELO dpdk5.sh.intel.com) ([10.239.129.244]) by fmsmga001.fm.intel.com with ESMTP; 14 Jan 2016 05:16:46 -0800 From: Zhihong Wang To: dev@dpdk.org Date: Thu, 14 Jan 2016 01:13:21 -0500 Message-Id: <1452752002-107586-4-git-send-email-zhihong.wang@intel.com> X-Mailer: git-send-email 2.5.0 In-Reply-To: <1452752002-107586-1-git-send-email-zhihong.wang@intel.com> References: <1452752002-107586-1-git-send-email-zhihong.wang@intel.com> Subject: [dpdk-dev] [PATCH 3/4] lib/librte_eal: Optimize memcpy for AVX512 platforms X-BeenThere: dev@dpdk.org X-Mailman-Version: 2.1.15 Precedence: list List-Id: patches and discussions about DPDK List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Thu, 14 Jan 2016 13:16:48 -0000 Implement AVX512 memcpy and choose the right implementation based on predefined macros, to make full utilization of hardware resources and deliver high performance. In current DPDK, memcpy holds a large proportion of execution time in libs like Vhost, especially for large packets, and this patch can bring considerable benefits for AVX512 platforms. The implementation is based on the current DPDK memcpy framework, some background introduction can be found in these threads: http://dpdk.org/ml/archives/dev/2014-November/008158.html http://dpdk.org/ml/archives/dev/2015-January/011800.html Signed-off-by: Zhihong Wang --- .../common/include/arch/x86/rte_memcpy.h | 247 ++++++++++++++++++++- 1 file changed, 243 insertions(+), 4 deletions(-) diff --git a/lib/librte_eal/common/include/arch/x86/rte_memcpy.h b/lib/librte_eal/common/include/arch/x86/rte_memcpy.h index 6a57426..fee954a 100644 --- a/lib/librte_eal/common/include/arch/x86/rte_memcpy.h +++ b/lib/librte_eal/common/include/arch/x86/rte_memcpy.h @@ -37,7 +37,7 @@ /** * @file * - * Functions for SSE/AVX/AVX2 implementation of memcpy(). + * Functions for SSE/AVX/AVX2/AVX512 implementation of memcpy(). */ #include @@ -67,7 +67,246 @@ extern "C" { static inline void * rte_memcpy(void *dst, const void *src, size_t n) __attribute__((always_inline)); -#ifdef RTE_MACHINE_CPUFLAG_AVX2 +#ifdef RTE_MACHINE_CPUFLAG_AVX512F + +/** + * AVX512 implementation below + */ + +/** + * Copy 16 bytes from one location to another, + * locations should not overlap. + */ +static inline void +rte_mov16(uint8_t *dst, const uint8_t *src) +{ + __m128i xmm0; + + xmm0 = _mm_loadu_si128((const __m128i *)src); + _mm_storeu_si128((__m128i *)dst, xmm0); +} + +/** + * Copy 32 bytes from one location to another, + * locations should not overlap. + */ +static inline void +rte_mov32(uint8_t *dst, const uint8_t *src) +{ + __m256i ymm0; + + ymm0 = _mm256_loadu_si256((const __m256i *)src); + _mm256_storeu_si256((__m256i *)dst, ymm0); +} + +/** + * Copy 64 bytes from one location to another, + * locations should not overlap. + */ +static inline void +rte_mov64(uint8_t *dst, const uint8_t *src) +{ + __m512i zmm0; + + zmm0 = _mm512_loadu_si512((const void *)src); + _mm512_storeu_si512((void *)dst, zmm0); +} + +/** + * Copy 128 bytes from one location to another, + * locations should not overlap. + */ +static inline void +rte_mov128(uint8_t *dst, const uint8_t *src) +{ + rte_mov64(dst + 0 * 64, src + 0 * 64); + rte_mov64(dst + 1 * 64, src + 1 * 64); +} + +/** + * Copy 256 bytes from one location to another, + * locations should not overlap. + */ +static inline void +rte_mov256(uint8_t *dst, const uint8_t *src) +{ + rte_mov64(dst + 0 * 64, src + 0 * 64); + rte_mov64(dst + 1 * 64, src + 1 * 64); + rte_mov64(dst + 2 * 64, src + 2 * 64); + rte_mov64(dst + 3 * 64, src + 3 * 64); +} + +/** + * Copy 128-byte blocks from one location to another, + * locations should not overlap. + */ +static inline void +rte_mov128blocks(uint8_t *dst, const uint8_t *src, size_t n) +{ + __m512i zmm0, zmm1; + + while (n >= 128) { + zmm0 = _mm512_loadu_si512((const void *)(src + 0 * 64)); + n -= 128; + zmm1 = _mm512_loadu_si512((const void *)(src + 1 * 64)); + src = src + 128; + _mm512_storeu_si512((void *)(dst + 0 * 64), zmm0); + _mm512_storeu_si512((void *)(dst + 1 * 64), zmm1); + dst = dst + 128; + } +} + +/** + * Copy 512-byte blocks from one location to another, + * locations should not overlap. + */ +static inline void +rte_mov512blocks(uint8_t *dst, const uint8_t *src, size_t n) +{ + __m512i zmm0, zmm1, zmm2, zmm3, zmm4, zmm5, zmm6, zmm7; + + while (n >= 512) { + zmm0 = _mm512_loadu_si512((const void *)(src + 0 * 64)); + n -= 512; + zmm1 = _mm512_loadu_si512((const void *)(src + 1 * 64)); + zmm2 = _mm512_loadu_si512((const void *)(src + 2 * 64)); + zmm3 = _mm512_loadu_si512((const void *)(src + 3 * 64)); + zmm4 = _mm512_loadu_si512((const void *)(src + 4 * 64)); + zmm5 = _mm512_loadu_si512((const void *)(src + 5 * 64)); + zmm6 = _mm512_loadu_si512((const void *)(src + 6 * 64)); + zmm7 = _mm512_loadu_si512((const void *)(src + 7 * 64)); + src = src + 512; + _mm512_storeu_si512((void *)(dst + 0 * 64), zmm0); + _mm512_storeu_si512((void *)(dst + 1 * 64), zmm1); + _mm512_storeu_si512((void *)(dst + 2 * 64), zmm2); + _mm512_storeu_si512((void *)(dst + 3 * 64), zmm3); + _mm512_storeu_si512((void *)(dst + 4 * 64), zmm4); + _mm512_storeu_si512((void *)(dst + 5 * 64), zmm5); + _mm512_storeu_si512((void *)(dst + 6 * 64), zmm6); + _mm512_storeu_si512((void *)(dst + 7 * 64), zmm7); + dst = dst + 512; + } +} + +static inline void * +rte_memcpy(void *dst, const void *src, size_t n) +{ + uintptr_t dstu = (uintptr_t)dst; + uintptr_t srcu = (uintptr_t)src; + void *ret = dst; + size_t dstofss; + size_t bits; + + /** + * Copy less than 16 bytes + */ + if (n < 16) { + if (n & 0x01) { + *(uint8_t *)dstu = *(const uint8_t *)srcu; + srcu = (uintptr_t)((const uint8_t *)srcu + 1); + dstu = (uintptr_t)((uint8_t *)dstu + 1); + } + if (n & 0x02) { + *(uint16_t *)dstu = *(const uint16_t *)srcu; + srcu = (uintptr_t)((const uint16_t *)srcu + 1); + dstu = (uintptr_t)((uint16_t *)dstu + 1); + } + if (n & 0x04) { + *(uint32_t *)dstu = *(const uint32_t *)srcu; + srcu = (uintptr_t)((const uint32_t *)srcu + 1); + dstu = (uintptr_t)((uint32_t *)dstu + 1); + } + if (n & 0x08) + *(uint64_t *)dstu = *(const uint64_t *)srcu; + return ret; + } + + /** + * Fast way when copy size doesn't exceed 512 bytes + */ + if (n <= 32) { + rte_mov16((uint8_t *)dst, (const uint8_t *)src); + rte_mov16((uint8_t *)dst - 16 + n, + (const uint8_t *)src - 16 + n); + return ret; + } + if (n <= 64) { + rte_mov32((uint8_t *)dst, (const uint8_t *)src); + rte_mov32((uint8_t *)dst - 32 + n, + (const uint8_t *)src - 32 + n); + return ret; + } + if (n <= 512) { + if (n >= 256) { + n -= 256; + rte_mov256((uint8_t *)dst, (const uint8_t *)src); + src = (const uint8_t *)src + 256; + dst = (uint8_t *)dst + 256; + } + if (n >= 128) { + n -= 128; + rte_mov128((uint8_t *)dst, (const uint8_t *)src); + src = (const uint8_t *)src + 128; + dst = (uint8_t *)dst + 128; + } +COPY_BLOCK_128_BACK63: + if (n > 64) { + rte_mov64((uint8_t *)dst, (const uint8_t *)src); + rte_mov64((uint8_t *)dst - 64 + n, + (const uint8_t *)src - 64 + n); + return ret; + } + if (n > 0) + rte_mov64((uint8_t *)dst - 64 + n, + (const uint8_t *)src - 64 + n); + return ret; + } + + /** + * Make store aligned when copy size exceeds 512 bytes + */ + dstofss = ((uintptr_t)dst & 0x3F); + if (dstofss > 0) { + dstofss = 64 - dstofss; + n -= dstofss; + rte_mov64((uint8_t *)dst, (const uint8_t *)src); + src = (const uint8_t *)src + dstofss; + dst = (uint8_t *)dst + dstofss; + } + + /** + * Copy 512-byte blocks. + * Use copy block function for better instruction order control, + * which is important when load is unaligned. + */ + rte_mov512blocks((uint8_t *)dst, (const uint8_t *)src, n); + bits = n; + n = n & 511; + bits -= n; + src = (const uint8_t *)src + bits; + dst = (uint8_t *)dst + bits; + + /** + * Copy 128-byte blocks. + * Use copy block function for better instruction order control, + * which is important when load is unaligned. + */ + if (n >= 128) { + rte_mov128blocks((uint8_t *)dst, (const uint8_t *)src, n); + bits = n; + n = n & 127; + bits -= n; + src = (const uint8_t *)src + bits; + dst = (uint8_t *)dst + bits; + } + + /** + * Copy whatever left + */ + goto COPY_BLOCK_128_BACK63; +} + +#elif RTE_MACHINE_CPUFLAG_AVX2 /** * AVX2 implementation below @@ -311,7 +550,7 @@ COPY_BLOCK_64_BACK31: goto COPY_BLOCK_64_BACK31; } -#else /* RTE_MACHINE_CPUFLAG_AVX2 */ +#else /* RTE_MACHINE_CPUFLAG */ /** * SSE & AVX implementation below @@ -630,7 +869,7 @@ COPY_BLOCK_64_BACK15: goto COPY_BLOCK_64_BACK15; } -#endif /* RTE_MACHINE_CPUFLAG_AVX2 */ +#endif /* RTE_MACHINE_CPUFLAG */ #ifdef __cplusplus } -- 2.5.0