[dpdk-dev] [PATCH] net/mlx5: prefetch CQEs for a faster decompression

[Alexander Kozyrev]

Invalidation of consumed CQEs incurs a performance penalty
due to many cache misses caused by a non-sequential CQEs access.
Prefetch CQEs to get a better data locality and speed up the
decompression of CQEs. Prefetching reduces CPI rate of the
rxq_cq_decompress_v() function from 1 to 0.85 in my environment,
resulting in 2% boost in mpps for 64B frames single core test.

Signed-off-by: Alexander Kozyrev <akozyrev at mellanox.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo at mellanox.com>
---
 drivers/net/mlx5/mlx5_rxtx_vec_altivec.h | 5 +++--
 drivers/net/mlx5/mlx5_rxtx_vec_neon.h    | 6 +++---
 drivers/net/mlx5/mlx5_rxtx_vec_sse.h     | 6 ++++--
 3 files changed, 10 insertions(+), 7 deletions(-)

diff --git a/drivers/net/mlx5/mlx5_rxtx_vec_altivec.h b/drivers/net/mlx5/mlx5_rxtx_vec_altivec.h
index aa43cab084..90548ea22d 100644
--- a/drivers/net/mlx5/mlx5_rxtx_vec_altivec.h
+++ b/drivers/net/mlx5/mlx5_rxtx_vec_altivec.h
@@ -155,8 +155,9 @@ rxq_cq_decompress_v(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cq,
 		const vector unsigned long shmax = {64, 64};
 #endif
 
-		if (!(pos & 0x7) && pos + 8 < mcqe_n)
-			rte_prefetch0((void *)(cq + pos + 8));
+		for (i = 0; i < MLX5_VPMD_DESCS_PER_LOOP; ++i)
+			if (likely(pos + i < mcqe_n))
+				rte_prefetch0((void *)(cq + pos + i));
 
 		/* A.1 load mCQEs into a 128bit register. */
 		mcqe1 = (vector unsigned char)vec_vsx_ld(0,
diff --git a/drivers/net/mlx5/mlx5_rxtx_vec_neon.h b/drivers/net/mlx5/mlx5_rxtx_vec_neon.h
index 6d952df787..44f662e1c1 100644
--- a/drivers/net/mlx5/mlx5_rxtx_vec_neon.h
+++ b/drivers/net/mlx5/mlx5_rxtx_vec_neon.h
@@ -145,9 +145,9 @@ rxq_cq_decompress_v(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cq,
 				    -1UL << ((mcqe_n - pos) *
 					     sizeof(uint16_t) * 8) : 0);
 #endif
-
-		if (!(pos & 0x7) && pos + 8 < mcqe_n)
-			rte_prefetch0((void *)(cq + pos + 8));
+		for (i = 0; i < MLX5_VPMD_DESCS_PER_LOOP; ++i)
+			if (likely(pos + i < mcqe_n))
+				rte_prefetch0((void *)(cq + pos + i));
 		__asm__ volatile (
 		/* A.1 load mCQEs into a 128bit register. */
 		"ld1 {v16.16b - v17.16b}, [%[mcq]] \n\t"
diff --git a/drivers/net/mlx5/mlx5_rxtx_vec_sse.h b/drivers/net/mlx5/mlx5_rxtx_vec_sse.h
index 406f23f595..9db9003acd 100644
--- a/drivers/net/mlx5/mlx5_rxtx_vec_sse.h
+++ b/drivers/net/mlx5/mlx5_rxtx_vec_sse.h
@@ -133,8 +133,10 @@ rxq_cq_decompress_v(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cq,
 		__m128i byte_cnt, invalid_mask;
 #endif
 
-		if (!(pos & 0x7) && pos + 8 < mcqe_n)
-			rte_prefetch0((void *)(cq + pos + 8));
+		for (i = 0; i < MLX5_VPMD_DESCS_PER_LOOP; ++i)
+			if (likely(pos + i < mcqe_n))
+				rte_prefetch0((void *)(cq + pos + i));
+
 		/* A.1 load mCQEs into a 128bit register. */
 		mcqe1 = _mm_loadu_si128((__m128i *)&mcq[pos % 8]);
 		mcqe2 = _mm_loadu_si128((__m128i *)&mcq[pos % 8 + 2]);
-- 
2.18.2