DPDK patches and discussions
 help / color / mirror / Atom feed
From: Herbert Guan <herbert.guan@arm.com>
To: dev@dpdk.org, jerin.jacob@caviumnetworks.com
Cc: Herbert Guan <herbert.guan@arm.com>
Subject: [dpdk-dev] [PATCH v4] arch/arm: optimization for memcpy on AArch64
Date: Thu, 21 Dec 2017 13:33:47 +0800	[thread overview]
Message-ID: <1513834427-12635-1-git-send-email-herbert.guan@arm.com> (raw)
In-Reply-To: <1513565664-19509-1-git-send-email-herbert.guan@arm.com>

This patch provides an option to do rte_memcpy() using 'restrict'
qualifier, which can induce GCC to do optimizations by using more
efficient instructions, providing some performance gain over memcpy()
on some AArch64 platforms/enviroments.

The memory copy performance differs between different AArch64
platforms. And a more recent glibc (e.g. 2.23 or later)
can provide a better memcpy() performance compared to old glibc
versions. It's always suggested to use a more recent glibc if
possible, from which the entire system can get benefit. If for some
reason an old glibc has to be used, this patch is provided for an
alternative.

This implementation can improve memory copy on some AArch64
platforms, when an old glibc (e.g. 2.19, 2.17...) is being used.
It is disabled by default and needs "RTE_ARCH_ARM64_MEMCPY"
defined to activate. It's not always proving better performance
than memcpy() so users need to run DPDK unit test
"memcpy_perf_autotest" and customize parameters in "customization
section" in rte_memcpy_64.h for best performance.

Compiler version will also impact the rte_memcpy() performance.
It's observed on some platforms and with the same code, GCC 7.2.0
compiled binary can provide better performance than GCC 4.8.5. It's
suggested to use GCC 5.4.0 or later.

Signed-off-by: Herbert Guan <herbert.guan@arm.com>
---
 config/common_armv8a_linuxapp                      |   6 +
 .../common/include/arch/arm/rte_memcpy_64.h        | 287 +++++++++++++++++++++
 2 files changed, 293 insertions(+)

diff --git a/config/common_armv8a_linuxapp b/config/common_armv8a_linuxapp
index 6732d1e..8f0cbed 100644
--- a/config/common_armv8a_linuxapp
+++ b/config/common_armv8a_linuxapp
@@ -44,6 +44,12 @@ CONFIG_RTE_FORCE_INTRINSICS=y
 # to address minimum DMA alignment across all arm64 implementations.
 CONFIG_RTE_CACHE_LINE_SIZE=128
 
+# Accelarate rte_memcpy.  Be sure to run unit test to determine the
+# best threshold in code.  Refer to notes in source file
+# (lib/librte_eal/common/include/arch/arm/rte_memcpy_64.h) for more
+# info.
+CONFIG_RTE_ARCH_ARM64_MEMCPY=n
+
 CONFIG_RTE_LIBRTE_FM10K_PMD=n
 CONFIG_RTE_LIBRTE_SFC_EFX_PMD=n
 CONFIG_RTE_LIBRTE_AVP_PMD=n
diff --git a/lib/librte_eal/common/include/arch/arm/rte_memcpy_64.h b/lib/librte_eal/common/include/arch/arm/rte_memcpy_64.h
index b80d8ba..b269f34 100644
--- a/lib/librte_eal/common/include/arch/arm/rte_memcpy_64.h
+++ b/lib/librte_eal/common/include/arch/arm/rte_memcpy_64.h
@@ -42,6 +42,291 @@
 
 #include "generic/rte_memcpy.h"
 
+#ifdef RTE_ARCH_ARM64_MEMCPY
+#include <rte_common.h>
+#include <rte_branch_prediction.h>
+
+/*
+ * The memory copy performance differs on different AArch64 micro-architectures.
+ * And the most recent glibc (e.g. 2.23 or later) can provide a better memcpy()
+ * performance compared to old glibc versions. It's always suggested to use a
+ * more recent glibc if possible, from which the entire system can get benefit.
+ *
+ * This implementation improves memory copy on some aarch64 micro-architectures,
+ * when an old glibc (e.g. 2.19, 2.17...) is being used. It is disabled by
+ * default and needs "RTE_ARCH_ARM64_MEMCPY" defined to activate. It's not
+ * always providing better performance than memcpy() so users need to run unit
+ * test "memcpy_perf_autotest" and customize parameters in customization section
+ * below for best performance.
+ *
+ * Compiler version will also impact the rte_memcpy() performance. It's observed
+ * on some platforms and with the same code, GCC 7.2.0 compiled binaries can
+ * provide better performance than GCC 4.8.5 compiled binaries.
+ */
+
+/**************************************
+ * Beginning of customization section
+ **************************************/
+#define RTE_ARM64_MEMCPY_ALIGN_MASK 0x0F
+#ifndef RTE_ARCH_ARM64_MEMCPY_STRICT_ALIGN
+/* Only src unalignment will be treaed as unaligned copy */
+#define IS_UNALIGNED_COPY(dst, src) \
+	((uintptr_t)(dst) & RTE_ARM64_MEMCPY_ALIGN_MASK)
+#else
+/* Both dst and src unalignment will be treated as unaligned copy */
+#define IS_UNALIGNED_COPY(dst, src) \
+	(((uintptr_t)(dst) | (uintptr_t)(src)) & RTE_ARM64_MEMCPY_ALIGN_MASK)
+#endif
+
+
+/*
+ * If copy size is larger than threshold, memcpy() will be used.
+ * Run "memcpy_perf_autotest" to determine the proper threshold.
+ */
+#define RTE_ARM64_MEMCPY_ALIGNED_THRESHOLD       ((size_t)(0xffffffff))
+#define RTE_ARM64_MEMCPY_UNALIGNED_THRESHOLD     ((size_t)(0xffffffff))
+
+/*
+ * The logic of USE_RTE_MEMCPY() can also be modified to best fit platform.
+ */
+#define USE_RTE_MEMCPY(dst, src, n) \
+((!IS_UNALIGNED_COPY(dst, src) && n <= RTE_ARM64_MEMCPY_ALIGNED_THRESHOLD) \
+|| (IS_UNALIGNED_COPY(dst, src) && n <= RTE_ARM64_MEMCPY_UNALIGNED_THRESHOLD))
+
+
+/**************************************
+ * End of customization section
+ **************************************/
+#if defined(RTE_TOOLCHAIN_GCC) && !defined(RTE_AARCH64_SKIP_GCC_VERSION_CHECK)
+#if (GCC_VERSION < 50400)
+#warning "The GCC version is quite old, which may result in sub-optimal \
+performance of the compiled code. It is suggested that at least GCC 5.4.0 \
+be used."
+#endif
+#endif
+
+static __rte_always_inline void rte_mov16(uint8_t *dst, const uint8_t *src)
+{
+	__uint128_t *dst128 = (__uint128_t *)dst;
+	const __uint128_t *src128 = (const __uint128_t *)src;
+	*dst128 = *src128;
+}
+
+static __rte_always_inline void rte_mov32(uint8_t *dst, const uint8_t *src)
+{
+	__uint128_t *dst128 = (__uint128_t *)dst;
+	const __uint128_t *src128 = (const __uint128_t *)src;
+	const __uint128_t x0 = src128[0], x1 = src128[1];
+	dst128[0] = x0;
+	dst128[1] = x1;
+}
+
+static __rte_always_inline void rte_mov48(uint8_t *dst, const uint8_t *src)
+{
+	__uint128_t *dst128 = (__uint128_t *)dst;
+	const __uint128_t *src128 = (const __uint128_t *)src;
+	const __uint128_t x0 = src128[0], x1 = src128[1], x2 = src128[2];
+	dst128[0] = x0;
+	dst128[1] = x1;
+	dst128[2] = x2;
+}
+
+static __rte_always_inline void rte_mov64(uint8_t *dst, const uint8_t *src)
+{
+	__uint128_t *dst128 = (__uint128_t *)dst;
+	const __uint128_t *src128 = (const __uint128_t *)src;
+	const __uint128_t
+		x0 = src128[0], x1 = src128[1], x2 = src128[2], x3 = src128[3];
+	dst128[0] = x0;
+	dst128[1] = x1;
+	dst128[2] = x2;
+	dst128[3] = x3;
+}
+
+static __rte_always_inline void rte_mov128(uint8_t *dst, const uint8_t *src)
+{
+	__uint128_t *dst128 = (__uint128_t *)dst;
+	const __uint128_t *src128 = (const __uint128_t *)src;
+	/* Keep below declaration & copy sequence for optimized instructions */
+	const __uint128_t
+		x0 = src128[0], x1 = src128[1], x2 = src128[2], x3 = src128[3];
+	dst128[0] = x0;
+	__uint128_t x4 = src128[4];
+	dst128[1] = x1;
+	__uint128_t x5 = src128[5];
+	dst128[2] = x2;
+	__uint128_t x6 = src128[6];
+	dst128[3] = x3;
+	__uint128_t x7 = src128[7];
+	dst128[4] = x4;
+	dst128[5] = x5;
+	dst128[6] = x6;
+	dst128[7] = x7;
+}
+
+static __rte_always_inline void rte_mov256(uint8_t *dst, const uint8_t *src)
+{
+	rte_mov128(dst, src);
+	rte_mov128(dst + 128, src + 128);
+}
+
+static __rte_always_inline void
+rte_memcpy_lt16(uint8_t *dst, const uint8_t *src, size_t n)
+{
+	if (n & 0x08) {
+		/* copy 8 ~ 15 bytes */
+		*(uint64_t *)dst = *(const uint64_t *)src;
+		*(uint64_t *)(dst - 8 + n) = *(const uint64_t *)(src - 8 + n);
+	} else if (n & 0x04) {
+		/* copy 4 ~ 7 bytes */
+		*(uint32_t *)dst = *(const uint32_t *)src;
+		*(uint32_t *)(dst - 4 + n) = *(const uint32_t *)(src - 4 + n);
+	} else if (n & 0x02) {
+		/* copy 2 ~ 3 bytes */
+		*(uint16_t *)dst = *(const uint16_t *)src;
+		*(uint16_t *)(dst - 2 + n) = *(const uint16_t *)(src - 2 + n);
+	} else if (n & 0x01) {
+		/* copy 1 byte */
+		*dst = *src;
+	}
+}
+
+static __rte_always_inline void
+rte_memcpy_ge16_lt128(uint8_t *dst, const uint8_t *src, size_t n)
+{
+	if (n < 64) {
+		if (n == 16) {
+			rte_mov16(dst, src);
+		} else if (n <= 32) {
+			rte_mov16(dst, src);
+			rte_mov16(dst - 16 + n, src - 16 + n);
+		} else if (n <= 48) {
+			rte_mov32(dst, src);
+			rte_mov16(dst - 16 + n, src - 16 + n);
+		} else {
+			rte_mov48(dst, src);
+			rte_mov16(dst - 16 + n, src - 16 + n);
+		}
+	} else {
+		rte_mov64((uint8_t *)dst, (const uint8_t *)src);
+		if (n > 48 + 64)
+			rte_mov64(dst - 64 + n, src - 64 + n);
+		else if (n > 32 + 64)
+			rte_mov48(dst - 48 + n, src - 48 + n);
+		else if (n > 16 + 64)
+			rte_mov32(dst - 32 + n, src - 32 + n);
+		else if (n > 64)
+			rte_mov16(dst - 16 + n, src - 16 + n);
+	}
+}
+
+static __rte_always_inline void
+rte_memcpy_ge128(uint8_t *dst, const uint8_t *src, size_t n)
+{
+	do {
+		rte_mov128(dst, src);
+		src += 128;
+		dst += 128;
+		n -= 128;
+	} while (likely(n >= 128));
+
+	if (likely(n)) {
+		if (n <= 16)
+			rte_mov16(dst - 16 + n, src - 16 + n);
+		else if (n <= 32)
+			rte_mov32(dst - 32 + n, src - 32 + n);
+		else if (n <= 48)
+			rte_mov48(dst - 48 + n, src - 48 + n);
+		else if (n <= 64)
+			rte_mov64(dst - 64 + n, src - 64 + n);
+		else
+			rte_memcpy_ge16_lt128(dst, src, n);
+	}
+}
+
+static __rte_always_inline void
+rte_memcpy_ge16_lt64(uint8_t *dst, const uint8_t *src, size_t n)
+{
+	if (n == 16) {
+		rte_mov16(dst, src);
+	} else if (n <= 32) {
+		rte_mov16(dst, src);
+		rte_mov16(dst - 16 + n, src - 16 + n);
+	} else if (n <= 48) {
+		rte_mov32(dst, src);
+		rte_mov16(dst - 16 + n, src - 16 + n);
+	} else {
+		rte_mov48(dst, src);
+		rte_mov16(dst - 16 + n, src - 16 + n);
+	}
+}
+
+static __rte_always_inline void
+rte_memcpy_ge64(uint8_t *dst, const uint8_t *src, size_t n)
+{
+	do {
+		rte_mov64(dst, src);
+		src += 64;
+		dst += 64;
+		n -= 64;
+	} while (likely(n >= 64));
+
+	if (likely(n)) {
+		if (n <= 16)
+			rte_mov16(dst - 16 + n, src - 16 + n);
+		else if (n <= 32)
+			rte_mov32(dst - 32 + n, src - 32 + n);
+		else if (n <= 48)
+			rte_mov48(dst - 48 + n, src - 48 + n);
+		else
+			rte_mov64(dst - 64 + n, src - 64 + n);
+	}
+}
+
+#if RTE_CACHE_LINE_SIZE >= 128
+static __rte_always_inline void *
+rte_memcpy(void *dst, const void *src, size_t n)
+{
+	if (n < 16) {
+		rte_memcpy_lt16((uint8_t *)dst, (const uint8_t *)src, n);
+		return dst;
+	}
+	if (n < 128) {
+		rte_memcpy_ge16_lt128((uint8_t *)dst, (const uint8_t *)src, n);
+		return dst;
+	}
+	__builtin_prefetch(src, 0, 0);
+	__builtin_prefetch(dst, 1, 0);
+	if (likely(USE_RTE_MEMCPY(dst, src, n))) {
+		rte_memcpy_ge128((uint8_t *)dst, (const uint8_t *)src, n);
+		return dst;
+	} else
+		return memcpy(dst, src, n);
+}
+
+#else
+static __rte_always_inline void *
+rte_memcpy(void *dst, const void *src, size_t n)
+{
+	if (n < 16) {
+		rte_memcpy_lt16((uint8_t *)dst, (const uint8_t *)src, n);
+		return dst;
+	}
+	if (n < 64) {
+		rte_memcpy_ge16_lt64((uint8_t *)dst, (const uint8_t *)src, n);
+		return dst;
+	}
+	__builtin_prefetch(src, 0, 0);
+	__builtin_prefetch(dst, 1, 0);
+	if (likely(USE_RTE_MEMCPY(dst, src, n))) {
+		rte_memcpy_ge64((uint8_t *)dst, (const uint8_t *)src, n);
+		return dst;
+	} else
+		return memcpy(dst, src, n);
+}
+#endif /* RTE_CACHE_LINE_SIZE >= 128 */
+
+#else
 static inline void
 rte_mov16(uint8_t *dst, const uint8_t *src)
 {
@@ -80,6 +365,8 @@
 
 #define rte_memcpy(d, s, n)	memcpy((d), (s), (n))
 
+#endif /* RTE_ARCH_ARM64_MEMCPY */
+
 #ifdef __cplusplus
 }
 #endif
-- 
1.8.3.1

  parent reply	other threads:[~2017-12-21  5:33 UTC|newest]

Thread overview: 26+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2017-11-27  7:49 [dpdk-dev] [PATCH] " Herbert Guan
2017-11-29 12:31 ` Jerin Jacob
2017-12-03 12:37   ` Herbert Guan
2017-12-15  4:06     ` Jerin Jacob
2017-12-18  2:51       ` Herbert Guan
2017-12-18  4:17         ` Jerin Jacob
2017-12-02  7:33 ` Pavan Nikhilesh Bhagavatula
2017-12-03 12:38   ` Herbert Guan
2017-12-03 14:20     ` Pavan Nikhilesh Bhagavatula
2017-12-04  7:14       ` Herbert Guan
2017-12-05  6:02 ` [dpdk-dev] [PATCH v2] " Herbert Guan
2017-12-18  2:54 ` [dpdk-dev] [PATCH v3] " Herbert Guan
2017-12-18  7:43   ` Jerin Jacob
2017-12-19  5:33     ` Herbert Guan
2017-12-19  7:24       ` Jerin Jacob
2017-12-21  5:33   ` Herbert Guan [this message]
2018-01-03 13:35     ` [dpdk-dev] [PATCH v4] " Jerin Jacob
2018-01-04 10:23       ` Herbert Guan
2018-01-04 10:20 ` [dpdk-dev] [PATCH v5] " Herbert Guan
2018-01-12 17:03   ` Thomas Monjalon
2018-01-15 10:57     ` Herbert Guan
2018-01-15 11:37       ` Thomas Monjalon
2018-01-18 23:54         ` Thomas Monjalon
2018-01-19  6:16           ` [dpdk-dev] 答复: " Herbert Guan
2018-01-19  6:10   ` [dpdk-dev] [PATCH v6] arch/arm: optimization for memcpy on ARM64 Herbert Guan
2018-01-20 16:21     ` Thomas Monjalon

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=1513834427-12635-1-git-send-email-herbert.guan@arm.com \
    --to=herbert.guan@arm.com \
    --cc=dev@dpdk.org \
    --cc=jerin.jacob@caviumnetworks.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).