[PATCH] serpent-avx512-x86: use depth-optimized vpternlogd S-boxes

Jussi Kivilinna jussi.kivilinna at iki.fi
Sat Nov 14 16:22:53 CET 2026


* cipher/serpent-avx512-x86.c (vpternlogd): New.
(SBOX0, SBOX0_INVERSE, SBOX1, SBOX1_INVERSE, SBOX2, SBOX2_INVERSE)
(SBOX3, SBOX3_INVERSE, SBOX4, SBOX4_INVERSE, SBOX5, SBOX5_INVERSE)
(SBOX6, SBOX6_INVERSE, SBOX7, SBOX7_INVERSE): Replace with ternary
logic circuits.
--

S-box circuits were found with SAT solver search over 3-input gates,
minimizing critical path through S-box, key mixing and linear
transformation.

Benchmark on AMD Ryzen 9 9950X3D (zen5):

 Before:
 SERPENT128     |  nanosecs/byte   mebibytes/sec   cycles/byte  auto Mhz
        ECB enc |     0.487 ns/B      1958 MiB/s      2.10 c/B      4320
        ECB dec |     0.409 ns/B      2329 MiB/s      1.77 c/B      4317±1

 After (~27% faster encryption, ~14% faster decryption):
 SERPENT128     |  nanosecs/byte   mebibytes/sec   cycles/byte  auto Mhz
        ECB enc |     0.382 ns/B      2495 MiB/s      1.65 c/B      4314
        ECB dec |     0.359 ns/B      2658 MiB/s      1.55 c/B      4312

Benchmark on AMD Ryzen 9 7900X (zen4):

 Before:
 SERPENT128     |  nanosecs/byte   mebibytes/sec   cycles/byte  auto Mhz
        ECB enc |     0.572 ns/B      1668 MiB/s      2.66 c/B      4650
        ECB dec |     0.549 ns/B      1737 MiB/s      2.55 c/B      4650

 After (~19% faster encryption, ~11% faster decryption):
 SERPENT128     |  nanosecs/byte   mebibytes/sec   cycles/byte  auto Mhz
        ECB enc |     0.479 ns/B      1989 MiB/s      2.23 c/B      4650
        ECB dec |     0.495 ns/B      1926 MiB/s      2.30 c/B      4650

Benchmark on Intel Core i3-1115G4 (tigerlake):

 Before:
 SERPENT128     |  nanosecs/byte   mebibytes/sec   cycles/byte  auto Mhz
        ECB enc |     0.495 ns/B      1927 MiB/s      2.03 c/B      4097±4
        ECB dec |     0.511 ns/B      1867 MiB/s      2.09 c/B      4097±3

 After (~14% faster encryption, ~11% faster decryption):
 SERPENT128     |  nanosecs/byte   mebibytes/sec   cycles/byte  auto Mhz
        ECB enc |     0.433 ns/B      2205 MiB/s      1.77 c/B      4104
        ECB dec |     0.461 ns/B      2071 MiB/s      1.89 c/B      4097±3

Signed-off-by: Jussi Kivilinna <jussi.kivilinna at iki.fi>
---
 cipher/serpent-avx512-x86.c | 438 ++++++++++++++++--------------------
 1 file changed, 197 insertions(+), 241 deletions(-)

diff --git a/cipher/serpent-avx512-x86.c b/cipher/serpent-avx512-x86.c
index dfbe2432..626206cd 100644
--- a/cipher/serpent-avx512-x86.c
+++ b/cipher/serpent-avx512-x86.c
@@ -59,6 +59,9 @@ typedef unsigned int serpent_subkeys_t[ROUNDS + 1][4];
 #define vror(x, s) _mm512_ror_epi32((x), (s))
 #define vshl(x, s) _mm512_slli_epi32((x), (s))
 
+#define vpternlogd(a, b, c, imm) \
+  _mm512_ternarylogic_epi32((a), (b), (c), (imm))
+
 /* 4x4 32-bit integer matrix transpose */
 #define transpose_4x4(x0, x1, x2, x3, t1, t2, t3) \
 	vpunpckhdq(x1, x0, t2); \
@@ -73,310 +76,263 @@ typedef unsigned int serpent_subkeys_t[ROUNDS + 1][4];
 	vpunpckhqdq(x2, t2, x3); \
 	vpunpcklqdq(x2, t2, x2);
 
-/*
- * These are the S-Boxes of Serpent from following research paper.
- *
- *  D. A. Osvik, “Speeding up Serpent,” in Third AES Candidate Conference,
- *   (New York, New York, USA), p. 317–329, National Institute of Standards and
- *   Technology, 2000.
- *
- * Paper is also available at: http://www.ii.uib.no/~osvik/pub/aes3.pdf
- *
- * --
- *
- * Following logic gets heavily optimized by compiler to use AVX512F
- * 'vpternlogq' instruction. This gives higher performance increase than
- * would be expected from simple wideing of vectors from AVX2/256bit to
- * AVX512/512bit.
- *
- */
+/* Serpent S-boxes as minimal-gate-count AVX512 ternary logic circuits.  */
 
 #define SBOX0(r0, r1, r2, r3, w, x, y, z) \
   { \
-    __m512i r4; \
-    \
-    r3 ^= r0; r4 =  r1; \
-    r1 &= r3; r4 ^= r2; \
-    r1 ^= r0; r0 |= r3; \
-    r0 ^= r4; r4 ^= r3; \
-    r3 ^= r2; r2 |= r1; \
-    r2 ^= r4; r4 = ~r4; \
-    r4 |= r1; r1 ^= r3; \
-    r1 ^= r4; r3 |= r0; \
-    r1 ^= r3; r4 ^= r3; \
-    \
-    w = r1; x = r4; y = r2; z = r0; \
+    __m512i t4, t5, t6, t7, t8, t9, t10, t11; \
+    \
+    t4 = vpternlogd(r0, r1, r3, 0xb8); \
+    t5 = vpternlogd(r0, r1, r3, 0x36); \
+    t6 = vpternlogd(r0, r1, r3, 0x6c); \
+    t7 = vpternlogd(r0, r2, r3, 0x96); \
+    t8 = r2 ^ t5; \
+    t9 = vpternlogd(t4, t5, t7, 0x6c); \
+    t10 = vpternlogd(t4, t7, t8, 0x1b); \
+    t11 = vpternlogd(t4, t6, t7, 0x4d); \
+    \
+    w = t11; x = t10; y = t9; z = t8; \
   }
 
 #define SBOX0_INVERSE(r0, r1, r2, r3, w, x, y, z) \
   { \
-    __m512i r4; \
-    \
-    r2 = ~r2; r4 =  r1; \
-    r1 |= r0; r4 = ~r4; \
-    r1 ^= r2; r2 |= r4; \
-    r1 ^= r3; r0 ^= r4; \
-    r2 ^= r0; r0 &= r3; \
-    r4 ^= r0; r0 |= r1; \
-    r0 ^= r2; r3 ^= r4; \
-    r2 ^= r1; r3 ^= r0; \
-    r3 ^= r1; \
-    r2 &= r3; \
-    r4 ^= r2; \
-    \
-    w = r0; x = r4; y = r1; z = r3; \
+    __m512i t4, t5, t6, t7, t8, t9, t10, t11; \
+    \
+    t4 = vpternlogd(r0, r1, r3, 0x4e); \
+    t5 = vpternlogd(r0, r1, r2, 0xa9); \
+    t6 = r3 ^ t5; \
+    t7 = vpternlogd(r0, r1, r2, 0x63); \
+    t8 = vpternlogd(r0, r1, r2, 0x81); \
+    t9 = vpternlogd(r3, t7, t8, 0xc9); \
+    t10 = vpternlogd(t4, t5, t9, 0x27); \
+    t11 = vpternlogd(t4, t5, t9, 0x69); \
+    \
+    w = t9; x = t10; y = t6; z = t11; \
   }
 
 #define SBOX1(r0, r1, r2, r3, w, x, y, z) \
   { \
-    __m512i r4; \
-    \
-    r0 = ~r0; r2 = ~r2; \
-    r4 =  r0; r0 &= r1; \
-    r2 ^= r0; r0 |= r3; \
-    r3 ^= r2; r1 ^= r0; \
-    r0 ^= r4; r4 |= r1; \
-    r1 ^= r3; r2 |= r0; \
-    r2 &= r4; r0 ^= r1; \
-    r1 &= r2; \
-    r1 ^= r0; r0 &= r2; \
-    r0 ^= r4; \
-    \
-    w = r2; x = r0; y = r3; z = r1; \
+    __m512i t4, t5, t6, t7, t8, t9, t10; \
+    \
+    t4 = vpternlogd(r0, r1, r3, 0xd8); \
+    t5 = vpternlogd(r0, r1, r3, 0x96); \
+    t6 = vpternlogd(r0, r1, r2, 0x59); \
+    t7 = vpternlogd(t4, t5, t6, 0x36); \
+    t8 = vpternlogd(t4, t5, t6, 0x93); \
+    t9 = r3 ^ t6; \
+    t10 = vpternlogd(t4, t5, t6, 0x2b); \
+    \
+    w = t10; x = t7; y = t9; z = t8; \
   }
 
 #define SBOX1_INVERSE(r0, r1, r2, r3, w, x, y, z) \
   { \
-    __m512i r4; \
-    \
-    r4 =  r1; r1 ^= r3; \
-    r3 &= r1; r4 ^= r2; \
-    r3 ^= r0; r0 |= r1; \
-    r2 ^= r3; r0 ^= r4; \
-    r0 |= r2; r1 ^= r3; \
-    r0 ^= r1; r1 |= r3; \
-    r1 ^= r0; r4 = ~r4; \
-    r4 ^= r1; r1 |= r0; \
-    r1 ^= r0; \
-    r1 |= r4; \
-    r3 ^= r1; \
-    \
-    w = r4; x = r0; y = r3; z = r2; \
+    __m512i t4, t5, t6, t7, t8, t9, t10, t11; \
+    \
+    t4 = vpternlogd(r0, r1, r2, 0x6a); \
+    t5 = vpternlogd(r0, r1, r2, 0xe4); \
+    t6 = vpternlogd(r0, r1, r2, 0x7c); \
+    t7 = vpternlogd(r3, t4, t6, 0x69); \
+    t8 = vpternlogd(r0, t5, t7, 0x8d); \
+    t9 = vpternlogd(r1, t4, t7, 0xc5); \
+    t10 = vpternlogd(r3, t5, t6, 0x95); \
+    t11 = vpternlogd(r2, t7, t10, 0x6c); \
+    \
+    w = t10; x = t8; y = t11; z = t9; \
   }
 
 #define SBOX2(r0, r1, r2, r3, w, x, y, z) \
   { \
-    __m512i r4; \
-    \
-    r4 =  r0; r0 &= r2; \
-    r0 ^= r3; r2 ^= r1; \
-    r2 ^= r0; r3 |= r4; \
-    r3 ^= r1; r4 ^= r2; \
-    r1 =  r3; r3 |= r4; \
-    r3 ^= r0; r0 &= r1; \
-    r4 ^= r0; r1 ^= r3; \
-    r1 ^= r4; r4 = ~r4; \
-    \
-    w = r2; x = r3; y = r1; z = r4; \
+    __m512i t4, t5, t6, t7, t8, t9, t10; \
+    \
+    t4 = vpternlogd(r0, r2, r3, 0x6a); \
+    t5 = vpternlogd(r0, r1, r2, 0x96); \
+    t6 = vpternlogd(r0, r2, r3, 0xc6); \
+    t7 = vpternlogd(r0, t4, t5, 0x96); \
+    t8 = vpternlogd(t4, t5, t6, 0x8e); \
+    t9 = vpternlogd(t4, t5, t6, 0xb4); \
+    t10 = vpternlogd(t5, t6, t9, 0x8d); \
+    \
+    w = t7; x = t8; y = t9; z = t10; \
   }
 
 #define SBOX2_INVERSE(r0, r1, r2, r3, w, x, y, z) \
   { \
-    __m512i r4; \
-    \
-    r2 ^= r3; r3 ^= r0; \
-    r4 =  r3; r3 &= r2; \
-    r3 ^= r1; r1 |= r2; \
-    r1 ^= r4; r4 &= r3; \
-    r2 ^= r3; r4 &= r0; \
-    r4 ^= r2; r2 &= r1; \
-    r2 |= r0; r3 = ~r3; \
-    r2 ^= r3; r0 ^= r3; \
-    r0 &= r1; r3 ^= r4; \
-    r3 ^= r0; \
-    \
-    w = r1; x = r4; y = r2; z = r3; \
+    __m512i t4, t5, t6, t7, t8, t9, t10, t11; \
+    \
+    t4 = vpternlogd(r0, r1, r2, 0x1e); \
+    t5 = vpternlogd(r1, r3, t4, 0x6a); \
+    t6 = vpternlogd(r1, r2, r3, 0xca); \
+    t7 = vpternlogd(r1, r2, r3, 0x8d); \
+    t8 = vpternlogd(r1, r2, r3, 0x49); \
+    t9 = vpternlogd(t4, t7, t8, 0x56); \
+    t10 = vpternlogd(t5, t6, t7, 0x93); \
+    t11 = vpternlogd(t5, t6, t7, 0x3a); \
+    \
+    w = t5; x = t9; y = t10; z = t11; \
   }
 
 #define SBOX3(r0, r1, r2, r3, w, x, y, z) \
   { \
-    __m512i r4; \
-    \
-    r4 =  r0; r0 |= r3; \
-    r3 ^= r1; r1 &= r4; \
-    r4 ^= r2; r2 ^= r3; \
-    r3 &= r0; r4 |= r1; \
-    r3 ^= r4; r0 ^= r1; \
-    r4 &= r0; r1 ^= r3; \
-    r4 ^= r2; r1 |= r0; \
-    r1 ^= r2; r0 ^= r3; \
-    r2 =  r1; r1 |= r3; \
-    r1 ^= r0; \
-    \
-    w = r1; x = r2; y = r3; z = r4; \
+    __m512i t4, t5, t6, t7, t8, t9, t10, t11; \
+    \
+    t4 = vpternlogd(r0, r1, r3, 0x1c); \
+    t5 = vpternlogd(r0, r2, r3, 0x62); \
+    t6 = vpternlogd(r0, r1, r2, 0x1a); \
+    t7 = vpternlogd(r1, r2, r3, 0xb0); \
+    t8 = vpternlogd(r0, t5, t7, 0x56); \
+    t9 = vpternlogd(r1, t5, t6, 0x16); \
+    t10 = vpternlogd(r3, t4, t6, 0x9a); \
+    t11 = vpternlogd(r3, t4, t5, 0x96); \
+    \
+    w = t8; x = t11; y = t10; z = t9; \
   }
 
 #define SBOX3_INVERSE(r0, r1, r2, r3, w, x, y, z) \
   { \
-    __m512i r4; \
-    \
-    r4 =  r2; r2 ^= r1; \
-    r0 ^= r2; r4 &= r2; \
-    r4 ^= r0; r0 &= r1; \
-    r1 ^= r3; r3 |= r4; \
-    r2 ^= r3; r0 ^= r3; \
-    r1 ^= r4; r3 &= r2; \
-    r3 ^= r1; r1 ^= r0; \
-    r1 |= r2; r0 ^= r3; \
-    r1 ^= r4; \
-    r0 ^= r1; \
-    \
-    w = r2; x = r1; y = r3; z = r0; \
+    __m512i t4, t5, t6, t7, t8, t9, t10, t11; \
+    \
+    t4 = r1 ^ r2; \
+    t5 = vpternlogd(r0, r1, r2, 0xb4); \
+    t6 = vpternlogd(r3, t4, t5, 0x36); \
+    t7 = vpternlogd(r3, t4, t5, 0xe3); \
+    t8 = vpternlogd(r1, t5, t7, 0x75); \
+    t9 = vpternlogd(r1, r3, t5, 0x96); \
+    t10 = vpternlogd(r2, t7, t9, 0x79); \
+    t11 = vpternlogd(t4, t6, t9, 0xa6); \
+    \
+    w = t6; x = t8; y = t11; z = t10; \
   }
 
 #define SBOX4(r0, r1, r2, r3, w, x, y, z) \
   { \
-    __m512i r4; \
-    \
-    r1 ^= r3; r3 = ~r3; \
-    r2 ^= r3; r3 ^= r0; \
-    r4 =  r1; r1 &= r3; \
-    r1 ^= r2; r4 ^= r3; \
-    r0 ^= r4; r2 &= r4; \
-    r2 ^= r0; r0 &= r1; \
-    r3 ^= r0; r4 |= r1; \
-    r4 ^= r0; r0 |= r3; \
-    r0 ^= r2; r2 &= r3; \
-    r0 = ~r0; r4 ^= r2; \
-    \
-    w = r1; x = r4; y = r0; z = r3; \
+    __m512i t4, t5, t6, t7, t8, t9, t10, t11; \
+    \
+    t4 = vpternlogd(r0, r2, r3, 0x36); \
+    t5 = vpternlogd(r0, r1, r3, 0x96); \
+    t6 = vpternlogd(r0, r2, r3, 0x93); \
+    t7 = vpternlogd(r0, r1, r3, 0x2e); \
+    t8 = vpternlogd(t5, t6, t7, 0x4b); \
+    t9 = t6 ^ t7; \
+    t10 = vpternlogd(r1, t5, t9, 0xc9); \
+    t11 = vpternlogd(t4, t5, t8, 0xd2); \
+    \
+    w = t9; x = t11; y = t8; z = t10; \
   }
 
 #define SBOX4_INVERSE(r0, r1, r2, r3, w, x, y, z) \
   { \
-    __m512i r4; \
-    \
-    r4 =  r2; r2 &= r3; \
-    r2 ^= r1; r1 |= r3; \
-    r1 &= r0; r4 ^= r2; \
-    r4 ^= r1; r1 &= r2; \
-    r0 = ~r0; r3 ^= r4; \
-    r1 ^= r3; r3 &= r0; \
-    r3 ^= r2; r0 ^= r1; \
-    r2 &= r0; r3 ^= r0; \
-    r2 ^= r4; \
-    r2 |= r3; r3 ^= r0; \
-    r2 ^= r1; \
-    \
-    w = r0; x = r3; y = r2; z = r4; \
+    __m512i t4, t5, t6, t7, t8, t9, t10, t11; \
+    \
+    t4 = vpternlogd(r0, r2, r3, 0x26); \
+    t5 = vpternlogd(r0, r1, r3, 0x6e); \
+    t6 = vpternlogd(r0, r1, r2, 0x3a); \
+    t7 = vpternlogd(r0, t4, t5, 0x6c); \
+    t8 = vpternlogd(r0, r1, r2, 0x96); \
+    t9 = vpternlogd(r2, t5, t6, 0xc9); \
+    t10 = vpternlogd(r3, t6, t8, 0x65); \
+    t11 = vpternlogd(t4, t6, t10, 0x69); \
+    \
+    w = t10; x = t7; y = t9; z = t11; \
   }
 
 #define SBOX5(r0, r1, r2, r3, w, x, y, z) \
   { \
-    __m512i r4; \
-    \
-    r0 ^= r1; r1 ^= r3; \
-    r3 = ~r3; r4 =  r1; \
-    r1 &= r0; r2 ^= r3; \
-    r1 ^= r2; r2 |= r4; \
-    r4 ^= r3; r3 &= r1; \
-    r3 ^= r0; r4 ^= r1; \
-    r4 ^= r2; r2 ^= r0; \
-    r0 &= r3; r2 = ~r2; \
-    r0 ^= r4; r4 |= r3; \
-    r2 ^= r4; \
-    \
-    w = r1; x = r3; y = r0; z = r2; \
+    __m512i t4, t5, t6, t7, t8, t9, t10, t11; \
+    \
+    t4 = vpternlogd(r0, r1, r2, 0x6e); \
+    t5 = vpternlogd(r0, r1, r3, 0x38); \
+    t6 = vpternlogd(r0, r1, r3, 0x7e); \
+    t7 = vpternlogd(r0, r2, t6, 0x69); \
+    t8 = vpternlogd(r3, t4, t6, 0x87); \
+    t9 = vpternlogd(r3, t5, t8, 0x4d); \
+    t10 = vpternlogd(r1, t7, t9, 0x96); \
+    t11 = vpternlogd(r2, r3, t5, 0xa9); \
+    \
+    w = t7; x = t11; y = t8; z = t10; \
   }
 
 #define SBOX5_INVERSE(r0, r1, r2, r3, w, x, y, z) \
   { \
-    __m512i r4; \
-    \
-    r1 = ~r1; r4 =  r3; \
-    r2 ^= r1; r3 |= r0; \
-    r3 ^= r2; r2 |= r1; \
-    r2 &= r0; r4 ^= r3; \
-    r2 ^= r4; r4 |= r0; \
-    r4 ^= r1; r1 &= r2; \
-    r1 ^= r3; r4 ^= r2; \
-    r3 &= r4; r4 ^= r1; \
-    r3 ^= r4; r4 = ~r4; \
-    r3 ^= r0; \
-    \
-    w = r1; x = r4; y = r3; z = r2; \
+    __m512i t4, t5, t6, t7, t8, t9, t10, t11; \
+    \
+    t4 = vpternlogd(r0, r1, r2, 0x5b); \
+    t5 = vpternlogd(r0, r1, r3, 0x38); \
+    t6 = vpternlogd(r0, r3, t4, 0xc5); \
+    t7 = vpternlogd(r1, r3, t4, 0xac); \
+    t8 = vpternlogd(r1, r2, t6, 0x59); \
+    t9 = vpternlogd(r2, t5, t6, 0x9c); \
+    t10 = vpternlogd(r3, t4, t5, 0xc9); \
+    t11 = t5 ^ t7; \
+    \
+    w = t11; x = t10; y = t9; z = t8; \
   }
 
 #define SBOX6(r0, r1, r2, r3, w, x, y, z) \
   { \
-    __m512i r4; \
-    \
-    r2 = ~r2; r4 =  r3; \
-    r3 &= r0; r0 ^= r4; \
-    r3 ^= r2; r2 |= r4; \
-    r1 ^= r3; r2 ^= r0; \
-    r0 |= r1; r2 ^= r1; \
-    r4 ^= r0; r0 |= r3; \
-    r0 ^= r2; r4 ^= r3; \
-    r4 ^= r0; r3 = ~r3; \
-    r2 &= r4; \
-    r2 ^= r3; \
-    \
-    w = r0; x = r1; y = r4; z = r2; \
+    __m512i t4, t5, t6, t7, t8, t9, t10, t11; \
+    \
+    t4 = vpternlogd(r0, r1, r3, 0x16); \
+    t5 = vpternlogd(r0, r1, r3, 0x62); \
+    t6 = vpternlogd(r0, r1, r3, 0x2e); \
+    t7 = vpternlogd(r0, r2, r3, 0x93); \
+    t8 = vpternlogd(r2, t4, t5, 0xa3); \
+    t9 = vpternlogd(r2, t4, t6, 0xc9); \
+    t10 = r1 ^ t7; \
+    t11 = vpternlogd(t7, t8, t9, 0x2d); \
+    \
+    w = t8; x = t10; y = t9; z = t11; \
   }
 
 #define SBOX6_INVERSE(r0, r1, r2, r3, w, x, y, z) \
   { \
-    __m512i r4; \
-    \
-    r0 ^= r2; r4 =  r2; \
-    r2 &= r0; r4 ^= r3; \
-    r2 = ~r2; r3 ^= r1; \
-    r2 ^= r3; r4 |= r0; \
-    r0 ^= r2; r3 ^= r4; \
-    r4 ^= r1; r1 &= r3; \
-    r1 ^= r0; r0 ^= r3; \
-    r0 |= r2; r3 ^= r1; \
-    r4 ^= r0; \
-    \
-    w = r1; x = r2; y = r4; z = r3; \
+    __m512i t4, t5, t6, t7, t8, t9, t10, t11, t12; \
+    \
+    t4 = vpternlogd(r0, r2, r3, 0xd8); \
+    t5 = vpternlogd(r0, r1, r2, 0xd0); \
+    t6 = vpternlogd(r0, r2, r3, 0x78); \
+    t7 = vpternlogd(r0, r1, r2, 0x48); \
+    t8 = vpternlogd(r0, r1, r2, 0x39); \
+    t9 = r3 ^ t8; \
+    t10 = vpternlogd(r3, t5, t7, 0xc9); \
+    t11 = vpternlogd(r1, t6, t10, 0x93); \
+    t12 = vpternlogd(t4, t8, t10, 0x69); \
+    \
+    w = t10; x = t9; y = t11; z = t12; \
   }
 
 #define SBOX7(r0, r1, r2, r3, w, x, y, z) \
   { \
-    __m512i r4; \
-    \
-    r4 =  r1; r1 |= r2; \
-    r1 ^= r3; r4 ^= r2; \
-    r2 ^= r1; r3 |= r4; \
-    r3 &= r0; r4 ^= r2; \
-    r3 ^= r1; r1 |= r4; \
-    r1 ^= r0; r0 |= r4; \
-    r0 ^= r2; r1 ^= r4; \
-    r2 ^= r1; r1 &= r0; \
-    r1 ^= r4; r2 = ~r2; \
-    r2 |= r0; \
-    r4 ^= r2; \
-    \
-    w = r4; x = r3; y = r1; z = r0; \
+    __m512i t4, t5, t6, t7, t8, t9, t10, t11, t12; \
+    \
+    t4 = vpternlogd(r0, r2, r3, 0x64); \
+    t5 = vpternlogd(r0, r1, r2, 0x62); \
+    t6 = vpternlogd(r0, r2, r3, 0xd2); \
+    t7 = vpternlogd(r0, r1, r2, 0x16); \
+    t8 = vpternlogd(r0, t4, t7, 0x6a); \
+    t9 = vpternlogd(r1, t4, t6, 0x93); \
+    t10 = vpternlogd(r3, t5, t7, 0x3a); \
+    t11 = vpternlogd(r3, t5, t7, 0x26); \
+    t12 = vpternlogd(t6, t9, t11, 0x59); \
+    \
+    w = t9; x = t12; y = t10; z = t8; \
   }
 
 #define SBOX7_INVERSE(r0, r1, r2, r3, w, x, y, z) \
   { \
-    __m512i r4; \
-    \
-    r4 =  r2; r2 ^= r0; \
-    r0 &= r3; r4 |= r3; \
-    r2 = ~r2; r3 ^= r1; \
-    r1 |= r0; r0 ^= r2; \
-    r2 &= r4; r3 &= r4; \
-    r1 ^= r2; r2 ^= r0; \
-    r0 |= r2; r4 ^= r1; \
-    r0 ^= r3; r3 ^= r4; \
-    r4 |= r0; r3 ^= r2; \
-    r4 ^= r2; \
-    \
-    w = r3; x = r0; y = r1; z = r4; \
+    __m512i t4, t5, t6, t7, t8, t9, t10, t11, t12; \
+    \
+    t4 = vpternlogd(r0, r1, r2, 0x87); \
+    t5 = vpternlogd(r0, r1, r2, 0x26); \
+    t6 = vpternlogd(r1, r3, t5, 0x76); \
+    t7 = vpternlogd(r1, r3, t4, 0x34); \
+    t8 = t5 ^ t7; \
+    t9 = vpternlogd(r0, t6, t7, 0x49); \
+    t10 = vpternlogd(r2, r3, t4, 0x2e); \
+    t11 = vpternlogd(r2, t4, t6, 0x2a); \
+    t12 = t7 ^ t10; \
+    \
+    w = t12; x = t9; y = t11; z = t8; \
   }
 
 /* XOR BLOCK1 into BLOCK0.  */
-- 
2.53.0



More information about the Gcrypt-devel mailing list