[dpdk-dev] [PATCH v4 2/2] net/mlx5: add non temporal store for WQE fields

[Aman Kumar <aman.kumar@vvdntech.in>]

add non temporal store for few WQE fields to optimize
data path. Define RTE_LIBRTE_MLX5_NT_STORE in build
configurations to enable this optimization.

Signed-off-by: Aman Kumar <aman.kumar@vvdntech.in>
---
 drivers/net/mlx5/meson.build     |   1 +
 drivers/net/mlx5/mlx5.c          |  17 ++
 drivers/net/mlx5/mlx5.h          |   4 +
 drivers/net/mlx5/mlx5_rxq.c      |   3 +
 drivers/net/mlx5/mlx5_rxtx.c     | 321 ++++++++++++++++++++++++++++++-
 drivers/net/mlx5/mlx5_rxtx.h     |   6 +
 drivers/net/mlx5/mlx5_rxtx_vec.h |  29 ++-
 drivers/net/mlx5/mlx5_txq.c      |   3 +
 meson_options.txt                |   2 +
 9 files changed, 377 insertions(+), 9 deletions(-)

diff --git a/drivers/net/mlx5/meson.build b/drivers/net/mlx5/meson.build
index 38e93fdc1..347ca6527 100644
--- a/drivers/net/mlx5/meson.build
+++ b/drivers/net/mlx5/meson.build
@@ -48,6 +48,7 @@ foreach option:cflags_options
 	endif
 endforeach
 dpdk_conf.set('RTE_LIBRTE_MLX5_NTLOAD_TSTORE_ALIGN_COPY', get_option('mlx5_ntload_tstore'))
+dpdk_conf.set('RTE_LIBRTE_MLX5_NT_STORE', get_option('mlx5_ntstore'))
 if get_option('buildtype').contains('debug')
 	cflags += [ '-pedantic', '-DPEDANTIC' ]
 else
diff --git a/drivers/net/mlx5/mlx5.c b/drivers/net/mlx5/mlx5.c
index a0913e161..04c8a7c08 100644
--- a/drivers/net/mlx5/mlx5.c
+++ b/drivers/net/mlx5/mlx5.c
@@ -164,6 +164,13 @@
 /* mprq_tstore_memcpy */
 #define MLX5_MPRQ_TSTORE_MEMCPY "mprq_tstore_memcpy"
 #endif
+#ifdef RTE_LIBRTE_MLX5_NT_STORE
+/* tx_wqe_field_ntstore */
+#define MLX5_TX_WQE_FIELD_NTSTORE "tx_wqe_field_ntstore"
+
+/* vec_rx_wqe_field_ntstore */
+#define MLX5_VEC_RX_WQE_FIELD_NTSTORE "vec_rx_wqe_field_ntstore"
+#endif
 
 /*
  * Device parameter to configure the total data buffer size for a single
@@ -1663,6 +1670,12 @@ mlx5_args_check(const char *key, const char *val, void *opaque)
 #ifdef RTE_LIBRTE_MLX5_NTLOAD_TSTORE_ALIGN_COPY
 	} else if (strcmp(MLX5_MPRQ_TSTORE_MEMCPY, key) == 0) {
 		config->mprq_tstore_memcpy = tmp;
+#endif
+#ifdef RTE_LIBRTE_MLX5_NT_STORE
+	} else if (strcmp(MLX5_TX_WQE_FIELD_NTSTORE, key) == 0) {
+		config->tx_wqe_field_ntstore = tmp;
+	} else if (strcmp(MLX5_VEC_RX_WQE_FIELD_NTSTORE, key) == 0) {
+		config->vec_rx_wqe_field_ntstore = tmp;
 #endif
 	} else {
 		DRV_LOG(WARNING, "%s: unknown parameter", key);
@@ -1726,6 +1739,10 @@ mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
 		MLX5_DECAP_EN,
 #ifdef RTE_LIBRTE_MLX5_NTLOAD_TSTORE_ALIGN_COPY
 		MLX5_MPRQ_TSTORE_MEMCPY,
+#endif
+#ifdef RTE_LIBRTE_MLX5_NT_STORE
+		MLX5_TX_WQE_FIELD_NTSTORE,
+		MLX5_VEC_RX_WQE_FIELD_NTSTORE,
 #endif
 		NULL,
 	};
diff --git a/drivers/net/mlx5/mlx5.h b/drivers/net/mlx5/mlx5.h
index 42934f6ca..978a9d697 100644
--- a/drivers/net/mlx5/mlx5.h
+++ b/drivers/net/mlx5/mlx5.h
@@ -239,6 +239,10 @@ struct mlx5_dev_config {
 #ifdef RTE_LIBRTE_MLX5_NTLOAD_TSTORE_ALIGN_COPY
 	unsigned int mprq_tstore_memcpy:1;
 #endif
+#ifdef RTE_LIBRTE_MLX5_NT_STORE
+	unsigned int tx_wqe_field_ntstore:1;
+	unsigned int vec_rx_wqe_field_ntstore:1;
+#endif
 };
 
 
diff --git a/drivers/net/mlx5/mlx5_rxq.c b/drivers/net/mlx5/mlx5_rxq.c
index 59b635e0b..4a8f301e1 100644
--- a/drivers/net/mlx5/mlx5_rxq.c
+++ b/drivers/net/mlx5/mlx5_rxq.c
@@ -1382,6 +1382,9 @@ mlx5_rxq_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
 		tmpl->irq = 1;
 #ifdef RTE_LIBRTE_MLX5_NTLOAD_TSTORE_ALIGN_COPY
 	tmpl->rxq.mprq_tstore_memcpy = config->mprq_tstore_memcpy;
+#endif
+#ifdef RTE_LIBRTE_MLX5_NT_STORE
+	tmpl->rxq.vec_rx_wqe_field_ntstore = config->vec_rx_wqe_field_ntstore;
 #endif
 	mprq_stride_nums = config->mprq.stride_num_n ?
 		config->mprq.stride_num_n : MLX5_MPRQ_STRIDE_NUM_N;
diff --git a/drivers/net/mlx5/mlx5_rxtx.c b/drivers/net/mlx5/mlx5_rxtx.c
index 761dc88f3..84f6fea9e 100644
--- a/drivers/net/mlx5/mlx5_rxtx.c
+++ b/drivers/net/mlx5/mlx5_rxtx.c
@@ -212,6 +212,300 @@ static void *memcpy_aligned_rx_tstore_16B(void *dst, void *src, int len)
 }
 #endif
 
+#if defined(RTE_ARCH_X86_64) && defined(RTE_LIBRTE_MLX5_NT_STORE)
+static void *amd_memcpy(void *dest, const void *src, size_t size)
+{
+	asm goto("movq	%0, %%rsi\n\t"
+	"movq	%1, %%rdi\n\t"
+	"movq	%2, %%rdx\n\t"
+	"movq    %%rdi, %%rax\n\t"
+	"cmp     $32, %%rdx\n\t"
+	"jb      less_vec\n\t"
+	"cmp     $(32 * 2), %%rdx\n\t"
+	"ja      more_2x_vec\n\t"
+	"vmovdqu   (%%rsi), %%ymm0\n\t"
+	"vmovdqu   -32(%%rsi,%%rdx), %%ymm1\n\t"
+	"vmovdqu   %%ymm0, (%%rdi)\n\t"
+	"vmovdqu   %%ymm1, -32(%%rdi,%%rdx)\n\t"
+	"vzeroupper\n\t"
+	"jmp %l[done]\n\t"
+	"less_vec:\n\t"
+	/* Less than 1 VEC.  */
+	"cmpb    $32, %%dl\n\t"
+	"jae     between_32_63\n\t"
+	"cmpb    $16, %%dl\n\t"
+	"jae     between_16_31\n\t"
+	"cmpb    $8, %%dl\n\t"
+	"jae     between_8_15\n\t"
+	"cmpb    $4, %%dl\n\t"
+	"jae     between_4_7\n\t"
+	"cmpb    $1, %%dl\n\t"
+	"ja      between_2_3\n\t"
+	"jb      1f\n\t"
+	"movzbl  (%%rsi), %%ecx\n\t"
+	"movb    %%cl, (%%rdi)\n\t"
+	"1:\n\t"
+	"jmp %l[done]\n\t"
+	"between_32_63:\n\t"
+	/* From 32 to 63.  No branch when size == 32.  */
+	"vmovdqu (%%rsi), %%ymm0\n\t"
+	"vmovdqu -32(%%rsi,%%rdx), %%ymm1\n\t"
+	"vmovdqu %%ymm0, (%%rdi)\n\t"
+	"vmovdqu %%ymm1, -32(%%rdi,%%rdx)\n\t"
+	"vzeroupper\n\t"
+	"jmp %l[done]\n\t"
+	/* From 16 to 31.  No branch when size == 16.  */
+	"between_16_31:\n\t"
+	"vmovdqu (%%rsi), %%xmm0\n\t"
+	"vmovdqu -16(%%rsi,%%rdx), %%xmm1\n\t"
+	"vmovdqu %%xmm0, (%%rdi)\n\t"
+	"vmovdqu %%xmm1, -16(%%rdi,%%rdx)\n\t"
+	"jmp %l[done]\n\t"
+	"between_8_15:\n\t"
+	/* From 8 to 15.  No branch when size == 8.  */
+	"movq    -8(%%rsi,%%rdx), %%rcx\n\t"
+	"movq    (%%rsi), %%rsi\n\t"
+	"movq    %%rcx, -8(%%rdi,%%rdx)\n\t"
+	"movq    %%rsi, (%%rdi)\n\t"
+	"jmp %l[done]\n\t"
+	"between_4_7:\n\t"
+	/* From 4 to 7.  No branch when size == 4.  */
+	"movl    -4(%%rsi,%%rdx), %%ecx\n\t"
+	"movl    (%%rsi), %%esi\n\t"
+	"movl    %%ecx, -4(%%rdi,%%rdx)\n\t"
+	"movl    %%esi, (%%rdi)\n\t"
+	"jmp %l[done]\n\t"
+	"between_2_3:\n\t"
+	/* From 2 to 3.  No branch when size == 2.  */
+	"movzwl  -2(%%rsi,%%rdx), %%ecx\n\t"
+	"movzwl  (%%rsi), %%esi\n\t"
+	"movw    %%cx, -2(%%rdi,%%rdx)\n\t"
+	"movw    %%si, (%%rdi)\n\t"
+	"jmp %l[done]\n\t"
+	"more_2x_vec:\n\t"
+	/* More than 2 * VEC and there may be overlap between destination */
+	/* and source.  */
+	"cmpq    $(32 * 8), %%rdx\n\t"
+	"ja      more_8x_vec\n\t"
+	"cmpq    $(32 * 4), %%rdx\n\t"
+	"jb      last_4x_vec\n\t"
+	/* Copy from 4 * VEC to 8 * VEC, inclusively. */
+	"vmovdqu   (%%rsi), %%ymm0\n\t"
+	"vmovdqu   32(%%rsi), %%ymm1\n\t"
+	"vmovdqu   (32 * 2)(%%rsi), %%ymm2\n\t"
+	"vmovdqu   (32 * 3)(%%rsi), %%ymm3\n\t"
+	"vmovdqu   -32(%%rsi,%%rdx), %%ymm4\n\t"
+	"vmovdqu   -(32 * 2)(%%rsi,%%rdx), %%ymm5\n\t"
+	"vmovdqu   -(32 * 3)(%%rsi,%%rdx), %%ymm6\n\t"
+	"vmovdqu   -(32 * 4)(%%rsi,%%rdx), %%ymm7\n\t"
+	"vmovdqu   %%ymm0, (%%rdi)\n\t"
+	"vmovdqu   %%ymm1, 32(%%rdi)\n\t"
+	"vmovdqu   %%ymm2, (32 * 2)(%%rdi)\n\t"
+	"vmovdqu   %%ymm3, (32 * 3)(%%rdi)\n\t"
+	"vmovdqu   %%ymm4, -32(%%rdi,%%rdx)\n\t"
+	"vmovdqu   %%ymm5, -(32 * 2)(%%rdi,%%rdx)\n\t"
+	"vmovdqu   %%ymm6, -(32 * 3)(%%rdi,%%rdx)\n\t"
+	"vmovdqu   %%ymm7, -(32 * 4)(%%rdi,%%rdx)\n\t"
+	"vzeroupper\n\t"
+	"jmp %l[done]\n\t"
+	"last_4x_vec:\n\t"
+	/* Copy from 2 * VEC to 4 * VEC. */
+	"vmovdqu   (%%rsi), %%ymm0\n\t"
+	"vmovdqu   32(%%rsi), %%ymm1\n\t"
+	"vmovdqu   -32(%%rsi,%%rdx), %%ymm2\n\t"
+	"vmovdqu   -(32 * 2)(%%rsi,%%rdx), %%ymm3\n\t"
+	"vmovdqu   %%ymm0, (%%rdi)\n\t"
+	"vmovdqu   %%ymm1, 32(%%rdi)\n\t"
+	"vmovdqu   %%ymm2, -32(%%rdi,%%rdx)\n\t"
+	"vmovdqu   %%ymm3, -(32 * 2)(%%rdi,%%rdx)\n\t"
+	"vzeroupper\n\t"
+	"nop:\n\t"
+	"jmp %l[done]\n\t"
+	"more_8x_vec:\n\t"
+	"cmpq    %%rsi, %%rdi\n\t"
+	"ja      more_8x_vec_backward\n\t"
+	/* Source == destination is less common.  */
+	"je      nop\n\t"
+	/* Load the first VEC and last 4 * VEC to support overlapping addresses.  */
+	"vmovdqu   (%%rsi), %%ymm4\n\t"
+	"vmovdqu   -32(%%rsi, %%rdx), %%ymm5\n\t"
+	"vmovdqu   -(32 * 2)(%%rsi, %%rdx), %%ymm6\n\t"
+	"vmovdqu   -(32 * 3)(%%rsi, %%rdx), %%ymm7\n\t"
+	"vmovdqu   -(32 * 4)(%%rsi, %%rdx), %%ymm8\n\t"
+	/* Save start and stop of the destination buffer.  */
+	"movq    %%rdi, %%r11\n\t"
+	"leaq    -32(%%rdi, %%rdx), %%rcx\n\t"
+	/* Align destination for aligned stores in the loop.  Compute */
+	/* how much destination is misaligned.  */
+	"movq    %%rdi, %%r8\n\t"
+	"andq    $(32 - 1), %%r8\n\t"
+	/* Get the negative of offset for alignment.  */
+	"subq    $32, %%r8\n\t"
+	/* Adjust source.  */
+	"subq    %%r8, %%rsi\n\t"
+	/* Adjust destination which should be aligned now.  */
+	"subq    %%r8, %%rdi\n\t"
+	/* Adjust length.  */
+	"addq    %%r8, %%rdx\n\t"
+	/* Check non-temporal store threshold.  */
+	"cmpq	 $(1024*1024), %%rdx\n\t"
+	"ja      large_forward\n\t"
+	"loop_4x_vec_forward:\n\t"
+	/* Copy 4 * VEC a time forward.  */
+	"vmovdqu   (%%rsi), %%ymm0\n\t"
+	"vmovdqu   32(%%rsi), %%ymm1\n\t"
+	"vmovdqu   (32 * 2)(%%rsi), %%ymm2\n\t"
+	"vmovdqu   (32 * 3)(%%rsi), %%ymm3\n\t"
+	"addq    $(32 * 4), %%rsi\n\t"
+	"subq    $(32 * 4), %%rdx\n\t"
+	"vmovdqa   %%ymm0, (%%rdi)\n\t"
+	"vmovdqa   %%ymm1, 32(%%rdi)\n\t"
+	"vmovdqa   %%ymm2, (32 * 2)(%%rdi)\n\t"
+	"vmovdqa   %%ymm3, (32 * 3)(%%rdi)\n\t"
+	"addq    $(32 * 4), %%rdi\n\t"
+	"cmpq    $(32 * 4), %%rdx\n\t"
+	"ja      loop_4x_vec_forward\n\t"
+	/* Store the last 4 * VEC.  */
+	"vmovdqu   %%ymm5, (%%rcx)\n\t"
+	"vmovdqu   %%ymm6, -32(%%rcx)\n\t"
+	"vmovdqu   %%ymm7, -(32 * 2)(%%rcx)\n\t"
+	"vmovdqu   %%ymm8, -(32 * 3)(%%rcx)\n\t"
+	/* Store the first VEC.  */
+	"vmovdqu   %%ymm4, (%%r11)\n\t"
+	"vzeroupper\n\t"
+	"jmp %l[done]\n\t"
+	"more_8x_vec_backward:\n\t"
+	/* Load the first 4*VEC and last VEC to support overlapping addresses.*/
+	"vmovdqu   (%%rsi), %%ymm4\n\t"
+	"vmovdqu   32(%%rsi), %%ymm5\n\t"
+	"vmovdqu   (32 * 2)(%%rsi), %%ymm6\n\t"
+	"vmovdqu   (32 * 3)(%%rsi), %%ymm7\n\t"
+	"vmovdqu   -32(%%rsi,%%rdx), %%ymm8\n\t"
+	/* Save stop of the destination buffer.  */
+	"leaq    -32(%%rdi, %%rdx), %%r11\n\t"
+	/* Align destination end for aligned stores in the loop.  Compute */
+	/* how much destination end is misaligned.  */
+	"leaq    -32(%%rsi, %%rdx), %%rcx\n\t"
+	"movq    %%r11, %%r9\n\t"
+	"movq    %%r11, %%r8\n\t"
+	"andq    $(32 - 1), %%r8\n\t"
+	/* Adjust source.  */
+	"subq    %%r8, %%rcx\n\t"
+	/* Adjust the end of destination which should be aligned now.  */
+	"subq    %%r8, %%r9\n\t"
+	/* Adjust length.  */
+	"subq    %%r8, %%rdx\n\t"
+	 /* Check non-temporal store threshold.  */
+	"cmpq	 $(1024*1024), %%rdx\n\t"
+	"ja      large_backward\n\t"
+	"loop_4x_vec_backward:\n\t"
+	/* Copy 4 * VEC a time backward.  */
+	"vmovdqu   (%%rcx), %%ymm0\n\t"
+	"vmovdqu   -32(%%rcx), %%ymm1\n\t"
+	"vmovdqu   -(32 * 2)(%%rcx), %%ymm2\n\t"
+	"vmovdqu   -(32 * 3)(%%rcx), %%ymm3\n\t"
+	"subq    $(32 * 4), %%rcx\n\t"
+	"subq    $(32 * 4), %%rdx\n\t"
+	"vmovdqa   %%ymm0, (%%r9)\n\t"
+	"vmovdqa   %%ymm1, -32(%%r9)\n\t"
+	"vmovdqa   %%ymm2, -(32 * 2)(%%r9)\n\t"
+	"vmovdqa   %%ymm3, -(32 * 3)(%%r9)\n\t"
+	"subq    $(32 * 4), %%r9\n\t"
+	"cmpq    $(32 * 4), %%rdx\n\t"
+	"ja      loop_4x_vec_backward\n\t"
+	/* Store the first 4 * VEC. */
+	"vmovdqu   %%ymm4, (%%rdi)\n\t"
+	"vmovdqu   %%ymm5, 32(%%rdi)\n\t"
+	"vmovdqu   %%ymm6, (32 * 2)(%%rdi)\n\t"
+	"vmovdqu   %%ymm7, (32 * 3)(%%rdi)\n\t"
+	/* Store the last VEC. */
+	"vmovdqu   %%ymm8, (%%r11)\n\t"
+	"vzeroupper\n\t"
+	"jmp %l[done]\n\t"
+
+	"large_forward:\n\t"
+	/* Don't use non-temporal store if there is overlap between */
+	/* destination and source since destination may be in cache */
+	/* when source is loaded. */
+	"leaq    (%%rdi, %%rdx), %%r10\n\t"
+	"cmpq    %%r10, %%rsi\n\t"
+	"jb      loop_4x_vec_forward\n\t"
+	"loop_large_forward:\n\t"
+	/* Copy 4 * VEC a time forward with non-temporal stores.  */
+	"prefetcht0 (32*4*2)(%%rsi)\n\t"
+	"prefetcht0 (32*4*2 + 64)(%%rsi)\n\t"
+	"prefetcht0 (32*4*3)(%%rsi)\n\t"
+	"prefetcht0 (32*4*3 + 64)(%%rsi)\n\t"
+	"vmovdqu   (%%rsi), %%ymm0\n\t"
+	"vmovdqu   32(%%rsi), %%ymm1\n\t"
+	"vmovdqu   (32 * 2)(%%rsi), %%ymm2\n\t"
+	"vmovdqu   (32 * 3)(%%rsi), %%ymm3\n\t"
+	"addq    $(32*4), %%rsi\n\t"
+	"subq    $(32*4), %%rdx\n\t"
+	"vmovntdq  %%ymm0, (%%rdi)\n\t"
+	"vmovntdq  %%ymm1, 32(%%rdi)\n\t"
+	"vmovntdq  %%ymm2, (32 * 2)(%%rdi)\n\t"
+	"vmovntdq  %%ymm3, (32 * 3)(%%rdi)\n\t"
+	"addq    $(32*4), %%rdi\n\t"
+	"cmpq    $(32*4), %%rdx\n\t"
+	"ja      loop_large_forward\n\t"
+	"sfence\n\t"
+	/* Store the last 4 * VEC.  */
+	"vmovdqu   %%ymm5, (%%rcx)\n\t"
+	"vmovdqu   %%ymm6, -32(%%rcx)\n\t"
+	"vmovdqu   %%ymm7, -(32 * 2)(%%rcx)\n\t"
+	"vmovdqu   %%ymm8, -(32 * 3)(%%rcx)\n\t"
+	/* Store the first VEC.  */
+	"vmovdqu   %%ymm4, (%%r11)\n\t"
+	"vzeroupper\n\t"
+	"jmp %l[done]\n\t"
+	"large_backward:\n\t"
+	/* Don't use non-temporal store if there is overlap between */
+	/* destination and source since destination may be in cache */
+	/* when source is loaded.  */
+	"leaq    (%%rcx, %%rdx), %%r10\n\t"
+	"cmpq    %%r10, %%r9\n\t"
+	"jb      loop_4x_vec_backward\n\t"
+	"loop_large_backward:\n\t"
+	/* Copy 4 * VEC a time backward with non-temporal stores. */
+	"prefetcht0 (-32 * 4 * 2)(%%rcx)\n\t"
+	"prefetcht0 (-32 * 4 * 2 - 64)(%%rcx)\n\t"
+	"prefetcht0 (-32 * 4 * 3)(%%rcx)\n\t"
+	"prefetcht0 (-32 * 4 * 3 - 64)(%%rcx)\n\t"
+	"vmovdqu   (%%rcx), %%ymm0\n\t"
+	"vmovdqu   -32(%%rcx), %%ymm1\n\t"
+	"vmovdqu   -(32 * 2)(%%rcx), %%ymm2\n\t"
+	"vmovdqu   -(32 * 3)(%%rcx), %%ymm3\n\t"
+	"subq    $(32*4), %%rcx\n\t"
+	"subq    $(32*4), %%rdx\n\t"
+	"vmovntdq  %%ymm0, (%%r9)\n\t"
+	"vmovntdq  %%ymm1, -32(%%r9)\n\t"
+	"vmovntdq  %%ymm2, -(32 * 2)(%%r9)\n\t"
+	"vmovntdq  %%ymm3, -(32 * 3)(%%r9)\n\t"
+	"subq    $(32 * 4), %%r9\n\t"
+	"cmpq    $(32 * 4), %%rdx\n\t"
+	"ja      loop_large_backward\n\t"
+	"sfence\n\t"
+	/* Store the first 4 * VEC.  */
+	"vmovdqu   %%ymm4, (%%rdi)\n\t"
+	"vmovdqu   %%ymm5, 32(%%rdi)\n\t"
+	"vmovdqu   %%ymm6, (32 * 2)(%%rdi)\n\t"
+	"vmovdqu   %%ymm7, (32 * 3)(%%rdi)\n\t"
+	/* Store the last VEC.  */
+	"vmovdqu   %%ymm8, (%%r11)\n\t"
+	"vzeroupper\n\t"
+	"jmp %l[done]"
+	:
+	: "r"(src), "r"(dest), "r"(size)
+	: "rax", "rcx", "rdx", "rdi", "rsi", "r8", "r9", "r10", "r11", "r12", "ymm0",
+	"ymm1", "ymm2", "ymm3", "ymm4", "ymm5", "ymm6", "ymm7", "ymm8", "memory"
+	: done
+	);
+done:
+	return dest;
+}
+#endif
+
 /**
  * Build a table to translate Rx completion flags to packet type.
  *
@@ -2323,6 +2617,9 @@ mlx5_tx_request_completion(struct mlx5_txq_data *__rte_restrict txq,
 {
 	uint16_t head = txq->elts_head;
 	unsigned int part;
+#ifdef RTE_LIBRTE_MLX5_NT_STORE
+	register uint32_t flags;
+#endif
 
 	part = MLX5_TXOFF_CONFIG(INLINE) ?
 	       0 : loc->pkts_sent - loc->pkts_copy;
@@ -2336,9 +2633,20 @@ mlx5_tx_request_completion(struct mlx5_txq_data *__rte_restrict txq,
 		txq->elts_comp = head;
 		if (MLX5_TXOFF_CONFIG(INLINE))
 			txq->wqe_comp = txq->wqe_ci;
-		/* Request unconditional completion on last WQE. */
-		last->cseg.flags = RTE_BE32(MLX5_COMP_ALWAYS <<
-					    MLX5_COMP_MODE_OFFSET);
+#ifdef RTE_LIBRTE_MLX5_NT_STORE
+		if (txq->tx_wqe_field_ntstore) {
+			flags = RTE_BE32(MLX5_COMP_ALWAYS <<
+					MLX5_COMP_MODE_OFFSET);
+			_mm_stream_si32(((void *)(uintptr_t)&last->cseg.flags),
+					flags);
+		} else {
+#endif
+			/* Request unconditional completion on last WQE. */
+			last->cseg.flags = RTE_BE32(MLX5_COMP_ALWAYS <<
+					MLX5_COMP_MODE_OFFSET);
+#ifdef RTE_LIBRTE_MLX5_NT_STORE
+		}
+#endif
 		/* Save elts_head in dedicated free on completion queue. */
 #ifdef RTE_LIBRTE_MLX5_DEBUG
 		txq->fcqs[txq->cq_pi++ & txq->cqe_m] = head |
@@ -3147,7 +3455,12 @@ mlx5_tx_dseg_vlan(struct mlx5_txq_data *__rte_restrict txq,
 	part = (uint8_t *)txq->wqes_end - pdst;
 	part = RTE_MIN(part, len);
 	do {
-		rte_memcpy(pdst, buf, part);
+#if defined(RTE_ARCH_X86_64) && defined(RTE_LIBRTE_MLX5_NT_STORE)
+		if (likely(txq->tx_wqe_field_ntstore))
+			amd_memcpy(pdst, buf, part);
+		else
+#endif
+			rte_memcpy(pdst, buf, part);
 		len -= part;
 		if (likely(!len)) {
 			pdst += part;
diff --git a/drivers/net/mlx5/mlx5_rxtx.h b/drivers/net/mlx5/mlx5_rxtx.h
index 750371014..11fd0918b 100644
--- a/drivers/net/mlx5/mlx5_rxtx.h
+++ b/drivers/net/mlx5/mlx5_rxtx.h
@@ -156,6 +156,9 @@ struct mlx5_rxq_data {
 #ifdef RTE_LIBRTE_MLX5_NTLOAD_TSTORE_ALIGN_COPY
 	unsigned int mprq_tstore_memcpy:1;
 #endif
+#ifdef RTE_LIBRTE_MLX5_NT_STORE
+	unsigned int vec_rx_wqe_field_ntstore:1;
+#endif
 } __rte_cache_aligned;
 
 enum mlx5_rxq_type {
@@ -256,6 +259,9 @@ struct mlx5_txq_data {
 	int32_t ts_offset; /* Timestamp field dynamic offset. */
 	struct mlx5_dev_ctx_shared *sh; /* Shared context. */
 	struct mlx5_txq_stats stats; /* TX queue counters. */
+#ifdef RTE_LIBRTE_MLX5_NT_STORE
+	unsigned int tx_wqe_field_ntstore:1;
+#endif
 #ifndef RTE_ARCH_64
 	rte_spinlock_t *uar_lock;
 	/* UAR access lock required for 32bit implementations */
diff --git a/drivers/net/mlx5/mlx5_rxtx_vec.h b/drivers/net/mlx5/mlx5_rxtx_vec.h
index a8d6c4f41..413f863ba 100644
--- a/drivers/net/mlx5/mlx5_rxtx_vec.h
+++ b/drivers/net/mlx5/mlx5_rxtx_vec.h
@@ -86,6 +86,10 @@ mlx5_rx_replenish_bulk_mbuf(struct mlx5_rxq_data *rxq, uint16_t n)
 	volatile struct mlx5_wqe_data_seg *wq =
 		&((volatile struct mlx5_wqe_data_seg *)rxq->wqes)[elts_idx];
 	unsigned int i;
+#ifdef RTE_LIBRTE_MLX5_NT_STORE
+	register uint64_t buf_addr2;
+	register uint32_t lkey_t;
+#endif
 
 	MLX5_ASSERT(n >= MLX5_VPMD_RXQ_RPLNSH_THRESH(q_n));
 	MLX5_ASSERT(n <= (uint16_t)(q_n - (rxq->rq_ci - rxq->rq_pi)));
@@ -107,11 +111,26 @@ mlx5_rx_replenish_bulk_mbuf(struct mlx5_rxq_data *rxq, uint16_t n)
 		 * impact the performance.
 		 */
 		buf_addr = elts[i]->buf_addr;
-		wq[i].addr = rte_cpu_to_be_64((uintptr_t)buf_addr +
-					      RTE_PKTMBUF_HEADROOM);
-		/* If there's only one MR, no need to replace LKey in WQE. */
-		if (unlikely(mlx5_mr_btree_len(&rxq->mr_ctrl.cache_bh) > 1))
-			wq[i].lkey = mlx5_rx_mb2mr(rxq, elts[i]);
+#ifdef RTE_LIBRTE_MLX5_NT_STORE
+		if (rxq->vec_rx_wqe_field_ntstore) {
+			buf_addr2 = (uint64_t)rte_cpu_to_be_64((uintptr_t)buf_addr +
+							       RTE_PKTMBUF_HEADROOM);
+			_mm_stream_si64(((void *)(uintptr_t)&wq[i].addr), buf_addr2);
+			/* If there's only one MR, no need to replace LKey in WQE. */
+			if (unlikely(mlx5_mr_btree_len(&rxq->mr_ctrl.cache_bh) > 1)) {
+				lkey_t = (uint32_t)mlx5_rx_mb2mr(rxq, elts[i]);
+				_mm_stream_si32(((void *)(uintptr_t)&wq[i].lkey), lkey_t);
+			}
+		} else {
+#endif
+			wq[i].addr = rte_cpu_to_be_64((uintptr_t)buf_addr +
+					RTE_PKTMBUF_HEADROOM);
+			/* If there's only one MR, no need to replace LKey in WQE. */
+			if (unlikely(mlx5_mr_btree_len(&rxq->mr_ctrl.cache_bh) > 1))
+				wq[i].lkey = mlx5_rx_mb2mr(rxq, elts[i]);
+#ifdef RTE_LIBRTE_MLX5_NT_STORE
+		}
+#endif
 	}
 	rxq->rq_ci += n;
 	/* Prevent overflowing into consumed mbufs. */
diff --git a/drivers/net/mlx5/mlx5_txq.c b/drivers/net/mlx5/mlx5_txq.c
index af84f5f72..f9160f551 100644
--- a/drivers/net/mlx5/mlx5_txq.c
+++ b/drivers/net/mlx5/mlx5_txq.c
@@ -779,6 +779,9 @@ txq_set_params(struct mlx5_txq_ctrl *txq_ctrl)
 					    DEV_TX_OFFLOAD_UDP_TNL_TSO);
 	bool vlan_inline;
 	unsigned int temp;
+#ifdef RTE_LIBRTE_MLX5_NT_STORE
+	txq_ctrl->txq.tx_wqe_field_ntstore = config->tx_wqe_field_ntstore;
+#endif
 
 	if (config->txqs_inline == MLX5_ARG_UNSET)
 		txqs_inline =
diff --git a/meson_options.txt b/meson_options.txt
index a4bc565d2..21c31d57b 100644
--- a/meson_options.txt
+++ b/meson_options.txt
@@ -32,6 +32,8 @@ option('max_numa_nodes', type: 'integer', value: 4,
 	description: 'maximum number of NUMA nodes supported by EAL')
 option('mlx5_ntload_tstore', type: 'boolean', value: false,
 	description: 'to enable optimized MPRQ in RX datapath')
+option('mlx5_ntstore', type: 'boolean', value: false,
+	description: 'to enable optimized MLX5 TX datapath')
 option('enable_trace_fp', type: 'boolean', value: false,
 	description: 'enable fast path trace points.')
 option('tests', type: 'boolean', value: true,
-- 
2.25.1