add non temporal load and temporal store for mprq memcpy.
This utilizes AMD EPYC2 optimized rte_memcpy* routines and
only enabled if config/x86/x86_amd_epyc_linux_gcc is build.
Signed-off-by: Aman Kumar <aman.ku...@vvdntech.in>
---
drivers/net/mlx5/mlx5_rx.h | 19 +++++++++++++++++++
1 file changed, 19 insertions(+)
diff --git a/drivers/net/mlx5/mlx5_rx.h b/drivers/net/mlx5/mlx5_rx.h
index 2b7ad3e48b..cda6aa02f2 100644
--- a/drivers/net/mlx5/mlx5_rx.h
+++ b/drivers/net/mlx5/mlx5_rx.h
@@ -422,6 +422,14 @@ mprq_buf_to_pkt(struct mlx5_rxq_data *rxq, struct rte_mbuf
*pkt, uint32_t len,
const uint32_t offset = strd_idx * strd_sz + strd_shift;
void *addr = RTE_PTR_ADD(mlx5_mprq_buf_addr(buf, strd_n), offset);
+#ifdef RTE_MEMCPY_AMDEPYC
+ if (len <= rxq->mprq_max_memcpy_len) {
+ rte_prefetch1(addr);
+ if (len > RTE_CACHE_LINE_SIZE)
+ rte_prefetch2((void *)((uintptr_t)addr +
+ RTE_CACHE_LINE_SIZE));
+ }
+#endif
/*
* Memcpy packets to the target mbuf if:
* - The size of packet is smaller than mprq_max_memcpy_len.
@@ -433,8 +441,19 @@ mprq_buf_to_pkt(struct mlx5_rxq_data *rxq, struct rte_mbuf
*pkt, uint32_t len,
(hdrm_overlap > 0 && !rxq->strd_scatter_en)) {
if (likely(len <=
(uint32_t)(pkt->buf_len - RTE_PKTMBUF_HEADROOM))) {
+#ifdef RTE_MEMCPY_AMDEPYC
+ uintptr_t data_addr;
+
+ data_addr = (uintptr_t)rte_pktmbuf_mtod(pkt, void *);
+ if (!((data_addr | (uintptr_t)addr) & ALIGNMENT_MASK))
+ rte_memcpy_aligned_tstore16((void *)data_addr,
+ addr, len);
+ else
+ rte_memcpy((void *)data_addr, addr, len);
+#else
rte_memcpy(rte_pktmbuf_mtod(pkt, void *),
addr, len);
+#endif
DATA_LEN(pkt) = len;
} else if (rxq->strd_scatter_en) {
struct rte_mbuf *prev = pkt;
--
2.25.1
