aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorDamjan Marion <damjan.marion@gmail.com>2020-02-16 13:07:13 +0100
committerNeale Ranns <nranns@cisco.com>2020-02-17 10:40:01 +0000
commit627fb6a16d8e7430e84aa664cb2b8f89a5688fab (patch)
tree76a05a6c79302f3ab0601425ada1d56663209a79
parent4fbb9daa90b53e0abaa060cf6db7762e708ce5b6 (diff)
crypto-native: calculate ghash using vpclmulqdq instructions
vpclmulqdq is introduced on intel icelake architecture and allows computing 4 carry-less multiplications in paralled by using 512-bit SIMD registers Type: feature Change-Id: Idb09d6f51ba6f116bba11649b2d99f649356d449 Signed-off-by: Damjan Marion <damjan.marion@gmail.com>
-rw-r--r--src/plugins/crypto_native/ghash.h126
-rw-r--r--src/vppinfra/vector_avx512.h15
2 files changed, 141 insertions, 0 deletions
diff --git a/src/plugins/crypto_native/ghash.h b/src/plugins/crypto_native/ghash.h
index a2886a468e9..abbfd2856f0 100644
--- a/src/plugins/crypto_native/ghash.h
+++ b/src/plugins/crypto_native/ghash.h
@@ -257,6 +257,132 @@ ghash_mul (u8x16 a, u8x16 b)
return ghash_final (gd);
}
+#ifdef __VPCLMULQDQ__
+
+static const u8x64 ghash4_poly2 = {
+ 0x00, 0x00, 0x00, 0xc2, 0x01, 0x00, 0x00, 0x00,
+ 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xc2,
+ 0x00, 0x00, 0x00, 0xc2, 0x01, 0x00, 0x00, 0x00,
+ 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xc2,
+ 0x00, 0x00, 0x00, 0xc2, 0x01, 0x00, 0x00, 0x00,
+ 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xc2,
+ 0x00, 0x00, 0x00, 0xc2, 0x01, 0x00, 0x00, 0x00,
+ 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xc2,
+};
+
+typedef struct
+{
+ u8x64 hi, lo, mid, tmp_lo, tmp_hi;
+ int pending;
+} ghash4_data_t;
+
+static_always_inline u8x64
+gmul4_lo_lo (u8x64 a, u8x64 b)
+{
+ return (u8x64) _mm512_clmulepi64_epi128 ((__m512i) a, (__m512i) b, 0x00);
+}
+
+static_always_inline u8x64
+gmul4_hi_lo (u8x64 a, u8x64 b)
+{
+ return (u8x64) _mm512_clmulepi64_epi128 ((__m512i) a, (__m512i) b, 0x01);
+}
+
+static_always_inline u8x64
+gmul4_lo_hi (u8x64 a, u8x64 b)
+{
+ return (u8x64) _mm512_clmulepi64_epi128 ((__m512i) a, (__m512i) b, 0x10);
+}
+
+static_always_inline u8x64
+gmul4_hi_hi (u8x64 a, u8x64 b)
+{
+ return (u8x64) _mm512_clmulepi64_epi128 ((__m512i) a, (__m512i) b, 0x11);
+}
+
+
+static_always_inline void
+ghash4_mul_first (ghash4_data_t * gd, u8x64 a, u8x64 b)
+{
+ gd->hi = gmul4_hi_hi (a, b);
+ gd->lo = gmul4_lo_lo (a, b);
+ gd->mid = (gmul4_hi_lo (a, b) ^ gmul4_lo_hi (a, b));
+ gd->pending = 0;
+}
+
+static_always_inline void
+ghash4_mul_next (ghash4_data_t * gd, u8x64 a, u8x64 b)
+{
+ u8x64 hi = gmul4_hi_hi (a, b);
+ u8x64 lo = gmul4_lo_lo (a, b);
+
+ if (gd->pending)
+ {
+ /* there is peding data from previous invocation so we can XOR */
+ gd->hi = u8x64_xor3 (gd->hi, gd->tmp_hi, hi);
+ gd->lo = u8x64_xor3 (gd->lo, gd->tmp_lo, lo);
+ gd->pending = 0;
+ }
+ else
+ {
+ /* there is no peding data from previous invocation so we postpone XOR */
+ gd->tmp_hi = hi;
+ gd->tmp_lo = lo;
+ gd->pending = 1;
+ }
+ gd->mid = u8x64_xor3 (gd->mid, gmul4_hi_lo (a, b), gmul4_lo_hi (a, b));
+}
+
+static_always_inline void
+ghash4_reduce (ghash4_data_t * gd)
+{
+ u8x64 r;
+
+ /* Final combination:
+ gd->lo ^= gd->mid << 64
+ gd->hi ^= gd->mid >> 64 */
+
+ u8x64 midl = u8x64_word_shift_left (gd->mid, 8);
+ u8x64 midr = u8x64_word_shift_right (gd->mid, 8);
+
+ if (gd->pending)
+ {
+ gd->lo = u8x64_xor3 (gd->lo, gd->tmp_lo, midl);
+ gd->hi = u8x64_xor3 (gd->hi, gd->tmp_hi, midr);
+ }
+ else
+ {
+ gd->lo ^= midl;
+ gd->hi ^= midr;
+ }
+
+ r = gmul4_hi_lo (ghash4_poly2, gd->lo);
+ gd->lo ^= u8x64_word_shift_left (r, 8);
+
+}
+
+static_always_inline void
+ghash4_reduce2 (ghash4_data_t * gd)
+{
+ gd->tmp_lo = gmul4_lo_lo (ghash4_poly2, gd->lo);
+ gd->tmp_hi = gmul4_lo_hi (ghash4_poly2, gd->lo);
+}
+
+static_always_inline u8x16
+ghash4_final (ghash4_data_t * gd)
+{
+ u8x64 r;
+ u8x32 t;
+
+ r = u8x64_xor3 (gd->hi, u8x64_word_shift_right (gd->tmp_lo, 4),
+ u8x64_word_shift_left (gd->tmp_hi, 4));
+
+ /* horizontal XOR of 4 128-bit lanes */
+ t = u8x64_extract_lo (r) ^ u8x64_extract_hi (r);
+ return u8x32_extract_hi (t) ^ u8x32_extract_lo (t);
+}
+#endif
+
static_always_inline void
ghash_precompute (u8x16 H, u8x16 * Hi, int count)
{
diff --git a/src/vppinfra/vector_avx512.h b/src/vppinfra/vector_avx512.h
index e27c7c041e4..efdc78bdc02 100644
--- a/src/vppinfra/vector_avx512.h
+++ b/src/vppinfra/vector_avx512.h
@@ -113,6 +113,18 @@ u32x16_extract_hi (u32x16 v)
return (u32x8) _mm512_extracti64x4_epi64 ((__m512i) v, 1);
}
+static_always_inline u8x32
+u8x64_extract_lo (u8x64 v)
+{
+ return (u8x32) _mm512_extracti64x4_epi64 ((__m512i) v, 0);
+}
+
+static_always_inline u8x32
+u8x64_extract_hi (u8x64 v)
+{
+ return (u8x32) _mm512_extracti64x4_epi64 ((__m512i) v, 1);
+}
+
static_always_inline u32
u32x16_min_scalar (u32x16 v)
{
@@ -149,6 +161,9 @@ u64x8_permute (u64x8 a, u64x8 b, u64x8 mask)
#define u8x64_extract_u8x16(a, n) \
(u8x16) _mm512_extracti64x2_epi64 ((__m512i) (a), n)
+#define u8x64_word_shift_left(a,n) (u8x64) _mm512_bslli_epi128((__m512i) a, n)
+#define u8x64_word_shift_right(a,n) (u8x64) _mm512_bsrli_epi128((__m512i) a, n)
+
static_always_inline u8x64
u8x64_xor3 (u8x64 a, u8x64 b, u8x64 c)
{