[1/3] net/mlx5: optimize mprq memcpy
diff mbox series

Message ID 20200907193214.38426-1-aman.kumar@vvdntech.in
State Superseded
Delegated to: Raslan Darawsheh
Headers show
Series
  • [1/3] net/mlx5: optimize mprq memcpy
Related show

Checks

Context Check Description
ci/checkpatch warning coding style issues

Commit Message

Aman Kumar Sept. 7, 2020, 7:32 p.m. UTC
add non temporal load and temporal store for mprq memcpy.
Enabling CONFIG_RTE_LIBRTE_MLX5_NTLOAD_TSTORE_ALIGN_COPY=y in
dpdk config will enable this optimization.

Signed-off-by: Aman Kumar <aman.kumar@vvdntech.in>
---
 config/common_base                      |  1 +
 drivers/net/mlx5/mlx5.c                 | 12 ++++
 drivers/net/mlx5/mlx5.h                 |  3 +
 drivers/net/mlx5/mlx5_rxq.c             |  3 +
 drivers/net/mlx5/mlx5_rxtx.c            | 17 ++++-
 drivers/net/mlx5/mlx5_rxtx.h            |  3 +
 lib/librte_eal/x86/include/rte_memcpy.h | 92 +++++++++++++++++++++++++
 7 files changed, 129 insertions(+), 2 deletions(-)

Patch
diff mbox series

diff --git a/config/common_base b/config/common_base
index fbf0ee70c..1476cf334 100644
--- a/config/common_base
+++ b/config/common_base
@@ -371,6 +371,7 @@  CONFIG_RTE_LIBRTE_MLX4_DEBUG=n
 # ConnectX-6 & BlueField (MLX5) PMD
 #
 CONFIG_RTE_LIBRTE_MLX5_PMD=n
+CONFIG_RTE_LIBRTE_MLX5_NTLOAD_TSTORE_ALIGN_COPY=n
 CONFIG_RTE_LIBRTE_MLX5_DEBUG=n
 
 #
diff --git a/drivers/net/mlx5/mlx5.c b/drivers/net/mlx5/mlx5.c
index 1e4c695f8..6eb85dfac 100644
--- a/drivers/net/mlx5/mlx5.c
+++ b/drivers/net/mlx5/mlx5.c
@@ -163,6 +163,11 @@ 
 /* Configure timeout of LRO session (in microseconds). */
 #define MLX5_LRO_TIMEOUT_USEC "lro_timeout_usec"
 
+#ifdef RTE_LIBRTE_MLX5_NTLOAD_TSTORE_ALIGN_COPY
+/* mprq_tstore_memcpy */
+#define MLX5_MPRQ_TSTORE_MEMCPY "mprq_tstore_memcpy"
+#endif
+
 /*
  * Device parameter to configure the total data buffer size for a single
  * hairpin queue (logarithm value).
@@ -1621,6 +1626,10 @@  mlx5_args_check(const char *key, const char *val, void *opaque)
 		config->sys_mem_en = !!tmp;
 	} else if (strcmp(MLX5_DECAP_EN, key) == 0) {
 		config->decap_en = !!tmp;
+#ifdef RTE_LIBRTE_MLX5_NTLOAD_TSTORE_ALIGN_COPY
+	} else if (strcmp(MLX5_MPRQ_TSTORE_MEMCPY, key) == 0) {
+		config->mprq_tstore_memcpy = tmp;
+#endif
 	} else {
 		DRV_LOG(WARNING, "%s: unknown parameter", key);
 		rte_errno = EINVAL;
@@ -1681,6 +1690,9 @@  mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
 		MLX5_RECLAIM_MEM,
 		MLX5_SYS_MEM_EN,
 		MLX5_DECAP_EN,
+#ifdef RTE_LIBRTE_MLX5_NTLOAD_TSTORE_ALIGN_COPY
+		MLX5_MPRQ_TSTORE_MEMCPY,
+#endif
 		NULL,
 	};
 	struct rte_kvargs *kvlist;
diff --git a/drivers/net/mlx5/mlx5.h b/drivers/net/mlx5/mlx5.h
index 78d6eb728..09dc90953 100644
--- a/drivers/net/mlx5/mlx5.h
+++ b/drivers/net/mlx5/mlx5.h
@@ -236,6 +236,9 @@  struct mlx5_dev_config {
 	int tx_skew; /* Tx scheduling skew between WQE and data on wire. */
 	struct mlx5_hca_attr hca_attr; /* HCA attributes. */
 	struct mlx5_lro_config lro; /* LRO configuration. */
+#ifdef RTE_LIBRTE_MLX5_NTLOAD_TSTORE_ALIGN_COPY
+	unsigned int mprq_tstore_memcpy:1;
+#endif
 };
 
 
diff --git a/drivers/net/mlx5/mlx5_rxq.c b/drivers/net/mlx5/mlx5_rxq.c
index 79eb8f8d7..bee5c03bc 100644
--- a/drivers/net/mlx5/mlx5_rxq.c
+++ b/drivers/net/mlx5/mlx5_rxq.c
@@ -2302,6 +2302,9 @@  mlx5_rxq_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
 	tmpl->socket = socket;
 	if (dev->data->dev_conf.intr_conf.rxq)
 		tmpl->irq = 1;
+#ifdef RTE_LIBRTE_MLX5_NTLOAD_TSTORE_ALIGN_COPY
+	tmpl->rxq.mprq_tstore_memcpy = config->mprq_tstore_memcpy;
+#endif
 	mprq_stride_nums = config->mprq.stride_num_n ?
 		config->mprq.stride_num_n : MLX5_MPRQ_STRIDE_NUM_N;
 	mprq_stride_size = non_scatter_min_mbuf_size <=
diff --git a/drivers/net/mlx5/mlx5_rxtx.c b/drivers/net/mlx5/mlx5_rxtx.c
index 1b71e9422..62ade3775 100644
--- a/drivers/net/mlx5/mlx5_rxtx.c
+++ b/drivers/net/mlx5/mlx5_rxtx.c
@@ -1774,8 +1774,21 @@  mlx5_rx_burst_mprq(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
 		    rxq->mprq_repl == NULL ||
 		    (hdrm_overlap > 0 && !rxq->strd_scatter_en)) {
 			if (likely(rte_pktmbuf_tailroom(pkt) >= len)) {
-				rte_memcpy(rte_pktmbuf_mtod(pkt, void *),
-					   addr, len);
+#ifdef RTE_LIBRTE_MLX5_NTLOAD_TSTORE_ALIGN_COPY
+				if ((rxq->mprq_tstore_memcpy) &&
+				    (!(((uintptr_t)(rte_pktmbuf_mtod(pkt,
+								     void *)) |
+					(uintptr_t)addr) & ALIGNMENT_MASK))) {
+					memcpy_aligned_rx_tstore_16B(
+						rte_pktmbuf_mtod(pkt, void *),
+						addr, len);
+				} else {
+#endif
+					rte_memcpy(rte_pktmbuf_mtod(pkt, void *),
+							addr, len);
+#ifdef RTE_LIBRTE_MLX5_NTLOAD_TSTORE_ALIGN_COPY
+				}
+#endif
 				DATA_LEN(pkt) = len;
 			} else if (rxq->strd_scatter_en) {
 				struct rte_mbuf *prev = pkt;
diff --git a/drivers/net/mlx5/mlx5_rxtx.h b/drivers/net/mlx5/mlx5_rxtx.h
index c02a007c8..72763962f 100644
--- a/drivers/net/mlx5/mlx5_rxtx.h
+++ b/drivers/net/mlx5/mlx5_rxtx.h
@@ -153,6 +153,9 @@  struct mlx5_rxq_data {
 	uint32_t tunnel; /* Tunnel information. */
 	uint64_t flow_meta_mask;
 	int32_t flow_meta_offset;
+#ifdef RTE_LIBRTE_MLX5_NTLOAD_TSTORE_ALIGN_COPY
+	unsigned int mprq_tstore_memcpy:1;
+#endif
 } __rte_cache_aligned;
 
 enum mlx5_rxq_obj_type {
diff --git a/lib/librte_eal/x86/include/rte_memcpy.h b/lib/librte_eal/x86/include/rte_memcpy.h
index 9c67232df..6345572a7 100644
--- a/lib/librte_eal/x86/include/rte_memcpy.h
+++ b/lib/librte_eal/x86/include/rte_memcpy.h
@@ -874,6 +874,98 @@  rte_memcpy(void *dst, const void *src, size_t n)
 		return rte_memcpy_generic(dst, src, n);
 }
 
+#ifdef RTE_LIBRTE_MLX5_NTLOAD_TSTORE_ALIGN_COPY
+static __rte_always_inline
+void copy16B_ts(void *dst, void *src)
+{
+	__m128i var128;
+
+	var128 = _mm_stream_load_si128((__m128i *)src);
+	_mm_storeu_si128((__m128i *)dst, var128);
+}
+
+static __rte_always_inline
+void copy32B_ts(void *dst, void *src)
+{
+	__m256i ymm0;
+
+	ymm0 = _mm256_stream_load_si256((const __m256i *)src);
+	_mm256_storeu_si256((__m256i *)dst, ymm0);
+}
+
+static __rte_always_inline
+void copy64B_ts(void *dst, void *src)
+{
+	__m256i ymm0, ymm1;
+
+	ymm0 = _mm256_stream_load_si256((const __m256i *)src);
+	ymm1 = _mm256_stream_load_si256((const __m256i *)((uint8_t *)src + 32));
+	_mm256_storeu_si256((__m256i *)dst, ymm0);
+	_mm256_storeu_si256((__m256i *)((uint8_t *)dst + 32), ymm1);
+}
+
+static __rte_always_inline
+void copy128B_ts(void *dst, void *src)
+{
+	__m256i ymm0, ymm1, ymm2, ymm3;
+
+	ymm0 = _mm256_stream_load_si256((const __m256i *)src);
+	ymm1 = _mm256_stream_load_si256((const __m256i *)((uint8_t *)src + 32));
+	ymm2 = _mm256_stream_load_si256((const __m256i *)((uint8_t *)src + 64));
+	ymm3 = _mm256_stream_load_si256((const __m256i *)((uint8_t *)src + 96));
+	_mm256_storeu_si256((__m256i *)dst, ymm0);
+	_mm256_storeu_si256((__m256i *)((uint8_t *)dst + 32), ymm1);
+	_mm256_storeu_si256((__m256i *)((uint8_t *)dst + 64), ymm2);
+	_mm256_storeu_si256((__m256i *)((uint8_t *)dst + 96), ymm3);
+}
+
+static __rte_always_inline
+void memcpy_aligned_rx_tstore_16B(void *dst, void *src, int len)
+{
+	while (len >= 128) {
+		copy128B_ts(dst, src);
+		dst = (uint8_t *)dst + 128;
+		src = (uint8_t *)src + 128;
+		len -= 128;
+	}
+	while (len >= 64) {
+		copy64B_ts(dst, src);
+		dst = (uint8_t *)dst + 64;
+		src = (uint8_t *)src + 64;
+		len -= 64;
+	}
+	while (len >= 32) {
+		copy32B_ts(dst, src);
+		dst = (uint8_t *)dst + 32;
+		src = (uint8_t *)src + 32;
+		len -= 32;
+	}
+	if (len >= 16) {
+		copy16B_ts(dst, src);
+		dst = (uint8_t *)dst + 16;
+		src = (uint8_t *)src + 16;
+		len -= 16;
+	}
+	if (len >= 8) {
+		*(uint64_t *)dst = *(const uint64_t *)src;
+		dst = (uint8_t *)dst + 8;
+		src = (uint8_t *)src + 8;
+		len -= 8;
+	}
+	if (len >= 4) {
+		*(uint32_t *)dst = *(const uint32_t *)src;
+		dst = (uint8_t *)dst + 4;
+		src = (uint8_t *)src + 4;
+		len -= 4;
+	}
+	if (len != 0) {
+		dst = (uint8_t *)dst - (4 - len);
+		src = (uint8_t *)src - (4 - len);
+		*(uint32_t *)dst = *(const uint32_t *)src;
+	}
+}
+#endif
+
 #if defined(RTE_TOOLCHAIN_GCC) && (GCC_VERSION >= 100000)
 #pragma GCC diagnostic pop
 #endif