1 #ifndef SHA256AVX_H 2 #define SHA256AVX_H 3 4 #include <immintrin.h> 5 #include <stdint.h> 6 7 static const unsigned int RC[] = { 8 0x428a2f98, 0x71374491, 0xb5c0fbcf, 0xe9b5dba5, 9 0x3956c25b, 0x59f111f1, 0x923f82a4, 0xab1c5ed5, 10 0xd807aa98, 0x12835b01, 0x243185be, 0x550c7dc3, 11 0x72be5d74, 0x80deb1fe, 0x9bdc06a7, 0xc19bf174, 12 0xe49b69c1, 0xefbe4786, 0x0fc19dc6, 0x240ca1cc, 13 0x2de92c6f, 0x4a7484aa, 0x5cb0a9dc, 0x76f988da, 14 0x983e5152, 0xa831c66d, 0xb00327c8, 0xbf597fc7, 15 0xc6e00bf3, 0xd5a79147, 0x06ca6351, 0x14292967, 16 0x27b70a85, 0x2e1b2138, 0x4d2c6dfc, 0x53380d13, 17 0x650a7354, 0x766a0abb, 0x81c2c92e, 0x92722c85, 18 0xa2bfe8a1, 0xa81a664b, 0xc24b8b70, 0xc76c51a3, 19 0xd192e819, 0xd6990624, 0xf40e3585, 0x106aa070, 20 0x19a4c116, 0x1e376c08, 0x2748774c, 0x34b0bcb5, 21 0x391c0cb3, 0x4ed8aa4a, 0x5b9cca4f, 0x682e6ff3, 22 0x748f82ee, 0x78a5636f, 0x84c87814, 0x8cc70208, 23 0x90befffa, 0xa4506ceb, 0xbef9a3f7, 0xc67178f2 24 }; 25 26 #define u32 uint32_t 27 #define u256 __m256i 28 29 #define XOR _mm256_xor_si256 30 #define OR _mm256_or_si256 31 #define AND _mm256_and_si256 32 #define ADD32 _mm256_add_epi32 33 #define NOT(x) _mm256_xor_si256(x, _mm256_set_epi32(-1, -1, -1, -1, -1, -1, -1, -1)) 34 35 #define LOAD(src) _mm256_loadu_si256((__m256i *)(src)) 36 #define STORE(dest,src) _mm256_storeu_si256((__m256i *)(dest),src) 37 38 #define BYTESWAP(x) _mm256_shuffle_epi8(x, _mm256_set_epi8(0xc,0xd,0xe,0xf,0x8,0x9,0xa,0xb,0x4,0x5,0x6,0x7,0x0,0x1,0x2,0x3,0xc,0xd,0xe,0xf,0x8,0x9,0xa,0xb,0x4,0x5,0x6,0x7,0x0,0x1,0x2,0x3)) 39 40 #define SHIFTR32(x, y) _mm256_srli_epi32(x, y) 41 #define SHIFTL32(x, y) _mm256_slli_epi32(x, y) 42 43 #define ROTR32(x, y) OR(SHIFTR32(x, y), SHIFTL32(x, 32 - (y))) 44 #define ROTL32(x, y) OR(SHIFTL32(x, y), SHIFTR32(x, 32 - (y))) 45 46 #define XOR3(a, b, c) XOR(XOR(a, b), c) 47 48 #define ADD3_32(a, b, c) ADD32(ADD32(a, b), c) 49 #define ADD4_32(a, b, c, d) ADD32(ADD32(ADD32(a, b), c), d) 50 #define ADD5_32(a, b, c, d, e) ADD32(ADD32(ADD32(ADD32(a, b), c), d), e) 51 52 #define MAJ_AVX(a, b, c) XOR3(AND(a, b), AND(a, c), AND(b, c)) 53 #define CH_AVX(a, b, c) XOR(AND(a, b), AND(NOT(a), c)) 54 55 #define SIGMA1_AVX(x) XOR3(ROTR32(x, 6), ROTR32(x, 11), ROTR32(x, 25)) 56 #define SIGMA0_AVX(x) XOR3(ROTR32(x, 2), ROTR32(x, 13), ROTR32(x, 22)) 57 58 #define WSIGMA1_AVX(x) XOR3(ROTR32(x, 17), ROTR32(x, 19), SHIFTR32(x, 10)) 59 #define WSIGMA0_AVX(x) XOR3(ROTR32(x, 7), ROTR32(x, 18), SHIFTR32(x, 3)) 60 61 #define SHA256ROUND_AVX(a, b, c, d, e, f, g, h, rc, w) \ 62 T0 = ADD5_32(h, SIGMA1_AVX(e), CH_AVX(e, f, g), _mm256_set1_epi32((int)RC[rc]), w); \ 63 (d) = ADD32(d, T0); \ 64 T1 = ADD32(SIGMA0_AVX(a), MAJ_AVX(a, b, c)); \ 65 (h) = ADD32(T0, T1); 66 67 typedef struct SHA256state { 68 u256 s[8]; 69 uint8_t msgblocks[8 * 64]; 70 unsigned int datalen; 71 uint64_t msglen; 72 } sha256ctxx8; 73 74 75 void PQCLEAN_SPHINCSSHA256192FROBUST_AVX2_transpose(u256 s[8]); 76 void PQCLEAN_SPHINCSSHA256192FROBUST_AVX2_sha256_init_frombytes_x8(sha256ctxx8 *ctx, const uint8_t *s, unsigned long long msglen); 77 void PQCLEAN_SPHINCSSHA256192FROBUST_AVX2_sha256_init8x(sha256ctxx8 *ctx); 78 void PQCLEAN_SPHINCSSHA256192FROBUST_AVX2_sha256_update8x(sha256ctxx8 *ctx, 79 const unsigned char *d0, 80 const unsigned char *d1, 81 const unsigned char *d2, 82 const unsigned char *d3, 83 const unsigned char *d4, 84 const unsigned char *d5, 85 const unsigned char *d6, 86 const unsigned char *d7, 87 unsigned long long len); 88 void PQCLEAN_SPHINCSSHA256192FROBUST_AVX2_sha256_final8x(sha256ctxx8 *ctx, 89 unsigned char *out0, 90 unsigned char *out1, 91 unsigned char *out2, 92 unsigned char *out3, 93 unsigned char *out4, 94 unsigned char *out5, 95 unsigned char *out6, 96 unsigned char *out7); 97 98 void PQCLEAN_SPHINCSSHA256192FROBUST_AVX2_sha256_transform8x(sha256ctxx8 *ctx, const unsigned char *data); 99 100 void PQCLEAN_SPHINCSSHA256192FROBUST_AVX2_sha256_clone_statex8(sha256ctxx8 *outctx, const sha256ctxx8 *inctx); 101 102 103 #endif 104