vppinfra: new vectorized ip checksum functions incl. csum_and_copy

Type: improvement
Change-Id: Id5810b7f4a6d6e4ce16b73c235b50db5d475ebf7
Signed-off-by: Damjan Marion <damarion@cisco.com>

3 files changed, 11 insertions, 98 deletions

diff --git a/src/vnet/ip/ip4_input.h b/src/vnet/ip/ip4_input.h
index 383ef31758c..53948d60266 100644
--- a/src/vnet/ip/ip4_input.h
+++ b/src/vnet/ip/ip4_input.h
@@ -42,6 +42,7 @@
 
 #include <vnet/ip/ip.h>
 #include <vnet/ethernet/ethernet.h>
+#include <vppinfra/vector/ip_csum.h>
 
 typedef enum
 {
@@ -63,15 +64,16 @@ check_ver_opt_csum (ip4_header_t * ip, u8 * error, int verify_checksum)
       if ((ip->ip_version_and_header_length & 0xf) != 5)
 	{
 	  *error = IP4_ERROR_OPTIONS;
-	  if (verify_checksum && ip_csum (ip, ip4_header_bytes (ip)) != 0)
+	  if (verify_checksum &&
+	      clib_ip_csum ((u8 *) ip, ip4_header_bytes (ip)) != 0)
 	    *error = IP4_ERROR_BAD_CHECKSUM;
 	}
       else
 	*error = IP4_ERROR_VERSION;
     }
-  else
-    if (PREDICT_FALSE (verify_checksum &&
-		       ip_csum (ip, sizeof (ip4_header_t)) != 0))
+  else if (PREDICT_FALSE (verify_checksum &&
+			  clib_ip_csum ((u8 *) ip, sizeof (ip4_header_t)) !=
+			    0))
     *error = IP4_ERROR_BAD_CHECKSUM;
 }
 
diff --git a/src/vnet/ip/ip_packet.h b/src/vnet/ip/ip_packet.h
index 837b3df8563..04cf9f11d70 100755
--- a/src/vnet/ip/ip_packet.h
+++ b/src/vnet/ip/ip_packet.h
@@ -149,98 +149,6 @@ STATIC_ASSERT_SIZEOF (ip_ecn_t, 1);
 
 extern u8 *format_ip_ecn (u8 * s, va_list * va);
 
-/* IP checksum support. */
-
-static_always_inline u16
-ip_csum (void *data, u16 n_left)
-{
-  u32 sum;
-#ifdef CLIB_HAVE_VEC256
-  u16x16 v1, v2;
-  u32x8 zero = { 0 };
-  u32x8 sum8 = { 0 };
-  u32x4 sum4;
-#endif
-
-  /* if there is odd number of bytes, pad by zero and store in sum */
-  sum = (n_left & 1) ? ((u8 *) data)[n_left - 1] << 8 : 0;
-
-  /* we deal with words */
-  n_left >>= 1;
-
-#ifdef CLIB_HAVE_VEC256
-  while (n_left >= 32)
-    {
-      v1 = u16x16_load_unaligned (data);
-      v2 = u16x16_load_unaligned (data + 32);
-
-#ifdef CLIB_ARCH_IS_LITTLE_ENDIAN
-      v1 = u16x16_byte_swap (v1);
-      v2 = u16x16_byte_swap (v2);
-#endif
-      sum8 += u32x8_from_u16x8 (u16x16_extract_lo (v1));
-      sum8 += u32x8_from_u16x8 (u16x16_extract_hi (v1));
-      sum8 += u32x8_from_u16x8 (u16x16_extract_lo (v2));
-      sum8 += u32x8_from_u16x8 (u16x16_extract_hi (v2));
-      n_left -= 32;
-      data += 64;
-    }
-
-  if (n_left >= 16)
-    {
-      v1 = u16x16_load_unaligned (data);
-#ifdef CLIB_ARCH_IS_LITTLE_ENDIAN
-      v1 = u16x16_byte_swap (v1);
-#endif
-      sum8 += u32x8_from_u16x8 (u16x16_extract_lo (v1));
-      sum8 += u32x8_from_u16x8 (u16x16_extract_hi (v1));
-      n_left -= 16;
-      data += 32;
-    }
-
-  if (n_left)
-    {
-      v1 = u16x16_load_unaligned (data);
-#ifdef CLIB_ARCH_IS_LITTLE_ENDIAN
-      v1 = u16x16_byte_swap (v1);
-#endif
-      v1 = u16x16_mask_last (v1, 16 - n_left);
-      sum8 += u32x8_from_u16x8 (u16x16_extract_lo (v1));
-      sum8 += u32x8_from_u16x8 (u16x16_extract_hi (v1));
-    }
-
-  sum8 = u32x8_hadd (sum8, zero);
-  sum4 = u32x8_extract_lo (sum8) + u32x8_extract_hi (sum8);
-  sum += sum4[0] + sum4[1];
-
-#else
-  /* scalar version */
-  while (n_left >= 8)
-    {
-      sum += clib_net_to_host_u16 (*((u16 *) data + 0));
-      sum += clib_net_to_host_u16 (*((u16 *) data + 1));
-      sum += clib_net_to_host_u16 (*((u16 *) data + 2));
-      sum += clib_net_to_host_u16 (*((u16 *) data + 3));
-      sum += clib_net_to_host_u16 (*((u16 *) data + 4));
-      sum += clib_net_to_host_u16 (*((u16 *) data + 5));
-      sum += clib_net_to_host_u16 (*((u16 *) data + 6));
-      sum += clib_net_to_host_u16 (*((u16 *) data + 7));
-      n_left -= 8;
-      data += 16;
-    }
-  while (n_left)
-    {
-      sum += clib_net_to_host_u16 (*(u16 *) data);
-      n_left -= 1;
-      data += 2;
-    }
-#endif
-
-  sum = (sum & 0xffff) + (sum >> 16);
-  sum = (sum & 0xffff) + (sum >> 16);
-  return ~((u16) sum);
-}
-
 /* Incremental checksum update. */
 typedef uword ip_csum_t;
 
diff --git a/src/vnet/ip/ip_psh_cksum.h b/src/vnet/ip/ip_psh_cksum.h
index eaac401f223..8723749865f 100644
--- a/src/vnet/ip/ip_psh_cksum.h
+++ b/src/vnet/ip/ip_psh_cksum.h
@@ -7,6 +7,7 @@
 #define included_ip_psh_cksum_h
 
 #include <vnet/ip/ip.h>
+#include <vppinfra/vector/ip_csum.h>
 
 typedef struct _ip4_psh
 {
@@ -37,7 +38,8 @@ ip4_pseudo_header_cksum (ip4_header_t *ip4)
   psh.proto = ip4->protocol;
   psh.l4len = clib_host_to_net_u16 (clib_net_to_host_u16 (ip4->length) -
 				    sizeof (ip4_header_t));
-  return ~clib_net_to_host_u16 (ip_csum (&psh, sizeof (ip4_psh_t)));
+  return ~clib_net_to_host_u16 (
+    clib_ip_csum ((u8 *) &psh, sizeof (ip4_psh_t)));
 }
 
 static_always_inline u16
@@ -48,7 +50,8 @@ ip6_pseudo_header_cksum (ip6_header_t *ip6)
   psh.dst = ip6->dst_address;
   psh.l4len = ip6->payload_length;
   psh.proto = clib_host_to_net_u32 ((u32) ip6->protocol);
-  return ~clib_net_to_host_u16 (ip_csum (&psh, sizeof (ip6_psh_t)));
+  return ~clib_net_to_host_u16 (
+    clib_ip_csum ((u8 *) &psh, sizeof (ip6_psh_t)));
 }
 
 #endif /* included_ip_psh_cksum_h */