summaryrefslogtreecommitdiffstats
path: root/src/vnet/ip/ip_packet.h
diff options
context:
space:
mode:
authorDamjan Marion <damarion@cisco.com>2018-06-18 22:21:40 +0200
committerDamjan Marion <dmarion@me.com>2018-06-28 10:00:05 +0000
commit08bca80c998bdae87060306350124714606a1604 (patch)
treeb207c71fc42030e9beb26addf5e4e4056a39e73d /src/vnet/ip/ip_packet.h
parent0d427d8d8a48af404dfc6535ae72d334b790f809 (diff)
ip: vectorized ip checksum
Change-Id: Ida678e6f31daa8decb18189da712a350336326e2 Signed-off-by: Damjan Marion <damarion@cisco.com>
Diffstat (limited to 'src/vnet/ip/ip_packet.h')
-rw-r--r--src/vnet/ip/ip_packet.h91
1 files changed, 91 insertions, 0 deletions
diff --git a/src/vnet/ip/ip_packet.h b/src/vnet/ip/ip_packet.h
index 6c86e3e046e..c4990976188 100644
--- a/src/vnet/ip/ip_packet.h
+++ b/src/vnet/ip/ip_packet.h
@@ -86,6 +86,97 @@ typedef enum
/* 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 += u16x8_extend_to_u32x8 (u16x16_extract_lo (v1));
+ sum8 += u16x8_extend_to_u32x8 (u16x16_extract_hi (v1));
+ sum8 += u16x8_extend_to_u32x8 (u16x16_extract_lo (v2));
+ sum8 += u16x8_extend_to_u32x8 (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
+ v1 = u16x16_byte_swap (u16x16_load_unaligned (data));
+ sum8 += u16x8_extend_to_u32x8 (u16x16_extract_lo (v1));
+ sum8 += u16x8_extend_to_u32x8 (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 += u16x8_extend_to_u32x8 (u16x16_extract_lo (v1));
+ sum8 += u16x8_extend_to_u32x8 (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;