dpdk: improve perf of buffer indices calc in the input node

Change-Id: I16557189aa4a763ec496cb4a45f6e12f2d46971f
Signed-off-by: Damjan Marion <damarion@cisco.com>

1 files changed, 53 insertions, 34 deletions

diff --git a/src/plugins/dpdk/device/node.c b/src/plugins/dpdk/device/node.c
index b8fe834755c..9b59830af33 100644
--- a/src/plugins/dpdk/device/node.c
+++ b/src/plugins/dpdk/device/node.c
@@ -214,37 +214,62 @@ poll_rate_limit (dpdk_main_t * dm)
 */
 
 static_always_inline void
-dpdk_mbuf_to_buffer_index_x4 (vlib_main_t * vm, struct rte_mbuf **mb,
-			      u32 * buffers)
+dpdk_mbufs_to_buffer_indices (vlib_main_t * vm, struct rte_mbuf **mb,
+			      u32 * bi, uword n_left)
 {
 #ifdef CLIB_HAVE_VEC256
-  vlib_buffer_main_t *bm = &buffer_main;
-  u64x4 v = *(u64x4 *) mb;
-  u32x8 v2, mask = { 0, 2, 4, 6, 1, 3, 5, 7 };
-
-  /* load 4 pointers into 256-bit register */
-  v = u64x4_load_unaligned (mb);
-
-  /* vlib_buffer_t is straight after rte_mbuf so advance all 4
-     pointers for size of rte_mbuf */
-  v += u64x4_splat (sizeof (struct rte_mbuf));
-
-  /* calculate 4 buffer indices in paralled */
-  v = (v - u64x4_splat (bm->buffer_mem_start)) >> CLIB_LOG2_CACHE_LINE_BYTES;
-
-  /* permute 256-bit register so lower u32s of each buffer index are
-   * placed into lower 128-bits */
-  v2 = u32x8_permute ((u32x8) v, mask);
+  u32x8 mask = { 0, 2, 4, 6, 1, 3, 5, 7 };
+  u64x4 off4 = u64x4_splat (buffer_main.buffer_mem_start -
+			    sizeof (struct rte_mbuf));
+#endif
 
-  /* extract lower 128-bits and save them to the array of buffer indices */
-  u32x4_store_unaligned (u32x8_extract_lo (v2), buffers);
+  while (n_left >= 8)
+    {
+#ifdef CLIB_HAVE_VEC256
+      /* load 4 pointers into 256-bit register */
+      u64x4 v0 = u64x4_load_unaligned (mb);
+      u64x4 v1 = u64x4_load_unaligned (mb + 4);
+      u32x8 v2, v3;
+
+      /* calculate 4 buffer indices in parallel
+         vlib_buffer_t is straight after rte_mbuf so advance all 4
+         pointers for size of rte_mbuf */
+      v0 -= off4;
+      v1 -= off4;
+
+      v0 >>= CLIB_LOG2_CACHE_LINE_BYTES;
+      v1 >>= CLIB_LOG2_CACHE_LINE_BYTES;
+
+      /* permute 256-bit register so lower u32s of each buffer index are
+       * placed into lower 128-bits */
+      v2 = u32x8_permute ((u32x8) v0, mask);
+      v3 = u32x8_permute ((u32x8) v1, mask);
+
+      /* extract lower 128-bits and save them to the array of buffer indices */
+      u32x4_store_unaligned (u32x8_extract_lo (v2), bi);
+      u32x4_store_unaligned (u32x8_extract_lo (v3), bi + 4);
 #else
-  /* equivalent non-nector implementation */
-  buffers[0] = vlib_get_buffer_index (vm, vlib_buffer_from_rte_mbuf (mb[0]));
-  buffers[1] = vlib_get_buffer_index (vm, vlib_buffer_from_rte_mbuf (mb[1]));
-  buffers[2] = vlib_get_buffer_index (vm, vlib_buffer_from_rte_mbuf (mb[2]));
-  buffers[3] = vlib_get_buffer_index (vm, vlib_buffer_from_rte_mbuf (mb[3]));
+      /* equivalent non-nector implementation */
+      bi[0] = vlib_get_buffer_index (vm, vlib_buffer_from_rte_mbuf (mb[0]));
+      bi[1] = vlib_get_buffer_index (vm, vlib_buffer_from_rte_mbuf (mb[1]));
+      bi[2] = vlib_get_buffer_index (vm, vlib_buffer_from_rte_mbuf (mb[2]));
+      bi[3] = vlib_get_buffer_index (vm, vlib_buffer_from_rte_mbuf (mb[3]));
+      bi[4] = vlib_get_buffer_index (vm, vlib_buffer_from_rte_mbuf (mb[4]));
+      bi[5] = vlib_get_buffer_index (vm, vlib_buffer_from_rte_mbuf (mb[4]));
+      bi[6] = vlib_get_buffer_index (vm, vlib_buffer_from_rte_mbuf (mb[6]));
+      bi[7] = vlib_get_buffer_index (vm, vlib_buffer_from_rte_mbuf (mb[7]));
 #endif
+      bi += 8;
+      mb += 8;
+      n_left -= 8;
+    }
+  while (n_left)
+    {
+      bi[0] = vlib_get_buffer_index (vm, vlib_buffer_from_rte_mbuf (mb[0]));
+      bi += 1;
+      mb += 1;
+      n_left -= 1;
+    }
 }
 
 static_always_inline u8
@@ -539,6 +564,7 @@ dpdk_device_input (vlib_main_t * vm, dpdk_main_t * dm, dpdk_device_t * xd,
       }
 
   /* enqueue buffers to the next node */
+  dpdk_mbufs_to_buffer_indices (vm, ptd->mbufs, ptd->buffers, n_rx_packets);
   n_left = n_rx_packets;
   next = ptd->next;
   buffers = ptd->buffers;
@@ -554,10 +580,6 @@ dpdk_device_input (vlib_main_t * vm, dpdk_main_t * dm, dpdk_device_t * xd,
 	  u16x16 next16 = u16x16_load_unaligned (next);
 	  if (u16x16_is_all_equal (next16, next_index))
 	    {
-	      dpdk_mbuf_to_buffer_index_x4 (vm, mb, buffers);
-	      dpdk_mbuf_to_buffer_index_x4 (vm, mb + 4, buffers + 4);
-	      dpdk_mbuf_to_buffer_index_x4 (vm, mb + 8, buffers + 8);
-	      dpdk_mbuf_to_buffer_index_x4 (vm, mb + 12, buffers + 12);
 	      clib_memcpy (to_next, buffers, 16 * sizeof (u32));
 	      to_next += 16;
 	      n_left_to_next -= 16;
@@ -568,7 +590,6 @@ dpdk_device_input (vlib_main_t * vm, dpdk_main_t * dm, dpdk_device_t * xd,
 	    }
 	  else
 	    {
-	      dpdk_mbuf_to_buffer_index_x4 (vm, mb, buffers);
 	      clib_memcpy (to_next, buffers, 4 * sizeof (u32));
 	      to_next += 4;
 	      n_left_to_next -= 4;
@@ -588,7 +609,6 @@ dpdk_device_input (vlib_main_t * vm, dpdk_main_t * dm, dpdk_device_t * xd,
 #endif
       while (n_left >= 4 && n_left_to_next >= 4)
 	{
-	  dpdk_mbuf_to_buffer_index_x4 (vm, mb, buffers);
 	  clib_memcpy (to_next, buffers, 4 * sizeof (u32));
 	  to_next += 4;
 	  n_left_to_next -= 4;
@@ -606,8 +626,7 @@ dpdk_device_input (vlib_main_t * vm, dpdk_main_t * dm, dpdk_device_t * xd,
 	}
       while (n_left && n_left_to_next)
 	{
-	  to_next[0] = buffers[0] =
-	    vlib_get_buffer_index (vm, vlib_buffer_from_rte_mbuf (mb[0]));
+	  clib_memcpy (to_next, buffers, 1 * sizeof (u32));
 	  to_next += 1;
 	  n_left_to_next -= 1;
 	  vlib_validate_buffer_enqueue_x1 (vm, node, next_index, to_next,