Add vlib_buffer_enqueue_to_next inline function

Change-Id: I1042c0fe179b57a00ce99c8d62cb1bdbe24d9184
Signed-off-by: Damjan Marion <damarion@cisco.com>

5 files changed, 140 insertions, 76 deletions

diff --git a/src/plugins/dpdk/device/node.c b/src/plugins/dpdk/device/node.c
index 3311ac4830d..a1acc1f0d1e 100644
--- a/src/plugins/dpdk/device/node.c
+++ b/src/plugins/dpdk/device/node.c
@@ -548,81 +548,9 @@ dpdk_device_input (vlib_main_t * vm, dpdk_main_t * dm, dpdk_device_t * xd,
   vlib_get_buffer_indices_with_offset (vm, (void **) ptd->mbufs, ptd->buffers,
 				       n_rx_packets,
 				       sizeof (struct rte_mbuf));
-  n_left = n_rx_packets;
-  next = ptd->next;
-  buffers = ptd->buffers;
-  mb = ptd->mbufs;
-  while (n_left)
-    {
-      u32 n_left_to_next;
-      u32 *to_next;
-      vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next);
-#ifdef CLIB_HAVE_VEC256
-      while (n_left >= 16 && n_left_to_next >= 16)
-	{
-	  u16x16 next16 = u16x16_load_unaligned (next);
-	  if (u16x16_is_all_equal (next16, next_index))
-	    {
-	      clib_memcpy (to_next, buffers, 16 * sizeof (u32));
-	      to_next += 16;
-	      n_left_to_next -= 16;
-	      buffers += 16;
-	      n_left -= 16;
-	      next += 16;
-	      mb += 16;
-	    }
-	  else
-	    {
-	      clib_memcpy (to_next, buffers, 4 * sizeof (u32));
-	      to_next += 4;
-	      n_left_to_next -= 4;
-
-	      vlib_validate_buffer_enqueue_x4 (vm, node, next_index, to_next,
-					       n_left_to_next, buffers[0],
-					       buffers[1], buffers[2],
-					       buffers[3], next[0], next[1],
-					       next[2], next[3]);
-	      /* next */
-	      buffers += 4;
-	      n_left -= 4;
-	      next += 4;
-	      mb += 4;
-	    }
-	}
-#endif
-      while (n_left >= 4 && n_left_to_next >= 4)
-	{
-	  clib_memcpy (to_next, buffers, 4 * sizeof (u32));
-	  to_next += 4;
-	  n_left_to_next -= 4;
-
-	  vlib_validate_buffer_enqueue_x4 (vm, node, next_index, to_next,
-					   n_left_to_next, buffers[0],
-					   buffers[1], buffers[2], buffers[3],
-					   next[0], next[1], next[2],
-					   next[3]);
-	  /* next */
-	  buffers += 4;
-	  n_left -= 4;
-	  next += 4;
-	  mb += 4;
-	}
-      while (n_left && n_left_to_next)
-	{
-	  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,
-					   n_left_to_next, buffers[0],
-					   next[0]);
-	  /* next */
-	  buffers += 1;
-	  n_left -= 1;
-	  next += 1;
-	  mb += 1;
-	}
-      vlib_put_next_frame (vm, node, next_index, n_left_to_next);
-    }
+
+  vlib_buffer_enqueue_to_next (vm, node, ptd->buffers, ptd->next,
+			       n_rx_packets);
 
   /* packet trace if enabled */
   if ((n_trace = vlib_get_trace_count (vm, node)))
diff --git a/src/vlib/buffer_node.h b/src/vlib/buffer_node.h
index f9e8b3fa4fe..1c4f4e7eea7 100644
--- a/src/vlib/buffer_node.h
+++ b/src/vlib/buffer_node.h
@@ -328,6 +328,121 @@ generic_buffer_node_inline (vlib_main_t * vm,
   return frame->n_vectors;
 }
 
+static_always_inline void
+vlib_buffer_enqueue_to_next (vlib_main_t * vm, vlib_node_runtime_t * node,
+			     u32 * buffers, u16 * nexts, uword count)
+{
+  u32 *to_next, n_left_to_next, max;
+  u16 next_index;
+
+  next_index = nexts[0];
+  vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next);
+  max = clib_min (n_left_to_next, count);
+
+  while (count)
+    {
+      u32 n_enqueued;
+      if ((nexts[0] != next_index) || n_left_to_next == 0)
+	{
+	  vlib_put_next_frame (vm, node, next_index, n_left_to_next);
+	  next_index = nexts[0];
+	  vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next);
+	  max = clib_min (n_left_to_next, count);
+	}
+#if defined(CLIB_HAVE_VEC512)
+      u16x32 next32 = u16x32_load_unaligned (nexts);
+      next32 = (next32 == u16x32_splat (next32[0]));
+      u64 bitmap = u16x32_msb_mask (next32);
+      n_enqueued = count_trailing_zeros (~bitmap);
+#elif defined(CLIB_HAVE_VEC256)
+      u16x16 next16 = u16x16_load_unaligned (nexts);
+      next16 = (next16 == u16x16_splat (next16[0]));
+      u64 bitmap = u8x32_msb_mask ((u8x32) next16);
+      n_enqueued = count_trailing_zeros (~bitmap) / 2;
+#elif defined(CLIB_HAVE_VEC128)
+      u16x8 next8 = u16x8_load_unaligned (nexts);
+      next8 = (next8 == u16x8_splat (next8[0]));
+      u64 bitmap = u8x16_msb_mask ((u8x16) next8);
+      n_enqueued = count_trailing_zeros (~bitmap) / 2;
+#else
+      u16 x = 0;
+      x |= next_index ^ nexts[1];
+      x |= next_index ^ nexts[2];
+      x |= next_index ^ nexts[3];
+      n_enqueued = (x == 0) ? 4 : 1;
+#endif
+
+      if (PREDICT_FALSE (n_enqueued > max))
+	n_enqueued = max;
+
+#ifdef CLIB_HAVE_VEC512
+      if (n_enqueued >= 32)
+	{
+	  clib_memcpy (to_next, buffers, 32 * sizeof (u32));
+	  nexts += 32;
+	  to_next += 32;
+	  buffers += 32;
+	  n_left_to_next -= 32;
+	  count -= 32;
+	  max -= 32;
+	  continue;
+	}
+#endif
+
+#ifdef CLIB_HAVE_VEC256
+      if (n_enqueued >= 16)
+	{
+	  clib_memcpy (to_next, buffers, 16 * sizeof (u32));
+	  nexts += 16;
+	  to_next += 16;
+	  buffers += 16;
+	  n_left_to_next -= 16;
+	  count -= 16;
+	  max -= 16;
+	  continue;
+	}
+#endif
+
+#ifdef CLIB_HAVE_VEC128
+      if (n_enqueued >= 8)
+	{
+	  clib_memcpy (to_next, buffers, 8 * sizeof (u32));
+	  nexts += 8;
+	  to_next += 8;
+	  buffers += 8;
+	  n_left_to_next -= 8;
+	  count -= 8;
+	  max -= 8;
+	  continue;
+	}
+#endif
+
+      if (n_enqueued >= 4)
+	{
+	  clib_memcpy (to_next, buffers, 4 * sizeof (u32));
+	  nexts += 4;
+	  to_next += 4;
+	  buffers += 4;
+	  n_left_to_next -= 4;
+	  count -= 4;
+	  max -= 4;
+	  continue;
+	}
+
+      /* copy */
+      to_next[0] = buffers[0];
+
+      /* next */
+      nexts += 1;
+      to_next += 1;
+      buffers += 1;
+      n_left_to_next -= 1;
+      count -= 1;
+      max -= 1;
+    }
+  vlib_put_next_frame (vm, node, next_index, n_left_to_next);
+}
+
 #endif /* included_vlib_buffer_node_h */
 
 /*
diff --git a/src/vppinfra/vector_avx2.h b/src/vppinfra/vector_avx2.h
index f651392559b..3f0b397b828 100644
--- a/src/vppinfra/vector_avx2.h
+++ b/src/vppinfra/vector_avx2.h
@@ -81,6 +81,12 @@ u32x8_insert_hi (u32x8 v1, u32x4 v2)
   return (u32x8) _mm256_inserti128_si256 ((__m256i) v1, (__m128i) v2, 1);
 }
 
+static_always_inline u32
+u8x32_msb_mask (u8x32 v)
+{
+  return _mm256_movemask_epi8 ((__m256i) v);
+}
+
 /* _extend_to_ */
 /* *INDENT-OFF* */
 #define _(f,t,i) \
diff --git a/src/vppinfra/vector_avx512.h b/src/vppinfra/vector_avx512.h
index ac4c09b8375..c1b7c42a260 100644
--- a/src/vppinfra/vector_avx512.h
+++ b/src/vppinfra/vector_avx512.h
@@ -27,6 +27,7 @@
   _(f,32,8,ps) _(f,64,4,pd)
 
 /* splat, load_unaligned, store_unaligned */
+/* *INDENT-OFF* */
 #define _(t, s, c, i) \
 static_always_inline t##s##x##c						\
 t##s##x##c##_splat (t##s x)						\
@@ -43,7 +44,15 @@ t##s##x##c##_store_unaligned (t##s##x##c v, void *p)			\
 
 foreach_avx512_vec512i foreach_avx512_vec512u
 #undef _
-#endif				/* included_vector_avx512_h */
+/* *INDENT-ON* */
+
+static_always_inline u32
+u16x32_msb_mask (u16x32 v)
+{
+  return (u32) _mm512_movepi16_mask ((__m512i) v);
+}
+
+#endif /* included_vector_avx512_h */
 /*
  * fd.io coding-style-patch-verification: ON
  *
diff --git a/src/vppinfra/vector_sse42.h b/src/vppinfra/vector_sse42.h
index cf7f158b873..50aa662a925 100644
--- a/src/vppinfra/vector_sse42.h
+++ b/src/vppinfra/vector_sse42.h
@@ -574,6 +574,12 @@ i16x8_min_scalar (i16x8 x)
   return _mm_extract_epi16 ((__m128i) x, 0);
 }
 
+static_always_inline u16
+u8x16_msb_mask (u8x16 v)
+{
+  return _mm_movemask_epi8 ((__m128i) v);
+}
+
 #undef _signed_binop
 
 #endif /* included_vector_sse2_h */