sha256.cpp

   1 // Copyright (c) 2010 Nils Schneider
   2 // Distributed under the MIT/X11 software license, see the accompanying
   3 // file license.txt or http://www.opensource.org/licenses/mit-license.php.
   4
   5 // 4-way 128-bit SSE2 SHA-256
   6
   7 #ifdef FOURWAYSSE2
   8
   9 #include <string.h>
  10 #include <assert.h>
  11
  12 #include <xmmintrin.h>
  13 #include <stdint.h>
  14 #include <stdio.h>
  15
  16 #define NPAR 32
  17
  18 extern void DoubleBlockSHA256(const void* pin, void* pout, const void* pinit, unsigned int hash[8][NPAR], const void* init2);
  19
  20 static const unsigned int sha256_consts[] = {
  21     0x428a2f98, 0x71374491, 0xb5c0fbcf, 0xe9b5dba5, /*  0 */
  22     0x3956c25b, 0x59f111f1, 0x923f82a4, 0xab1c5ed5,
  23     0xd807aa98, 0x12835b01, 0x243185be, 0x550c7dc3, /*  8 */
  24     0x72be5d74, 0x80deb1fe, 0x9bdc06a7, 0xc19bf174,
  25     0xe49b69c1, 0xefbe4786, 0x0fc19dc6, 0x240ca1cc, /* 16 */
  26     0x2de92c6f, 0x4a7484aa, 0x5cb0a9dc, 0x76f988da,
  27     0x983e5152, 0xa831c66d, 0xb00327c8, 0xbf597fc7, /* 24 */
  28     0xc6e00bf3, 0xd5a79147, 0x06ca6351, 0x14292967,
  29     0x27b70a85, 0x2e1b2138, 0x4d2c6dfc, 0x53380d13, /* 32 */
  30     0x650a7354, 0x766a0abb, 0x81c2c92e, 0x92722c85,
  31     0xa2bfe8a1, 0xa81a664b, 0xc24b8b70, 0xc76c51a3, /* 40 */
  32     0xd192e819, 0xd6990624, 0xf40e3585, 0x106aa070,
  33     0x19a4c116, 0x1e376c08, 0x2748774c, 0x34b0bcb5, /* 48 */
  34     0x391c0cb3, 0x4ed8aa4a, 0x5b9cca4f, 0x682e6ff3,
  35     0x748f82ee, 0x78a5636f, 0x84c87814, 0x8cc70208, /* 56 */
  36     0x90befffa, 0xa4506ceb, 0xbef9a3f7, 0xc67178f2
  37 };
  38
  39
  40 static inline __m128i Ch(const __m128i b, const __m128i c, const __m128i d) {
  41     return (b & c) ^ (~b & d);
  42 }
  43
  44 static inline __m128i Maj(const __m128i b, const __m128i c, const __m128i d) {
  45     return (b & c) ^ (b & d) ^ (c & d);
  46 }
  47
  48 static inline __m128i ROTR(__m128i x, const int n) {
  49     return _mm_srli_epi32(x, n) | _mm_slli_epi32(x, 32 - n);
  50 }
  51
  52 static inline __m128i SHR(__m128i x, const int n) {
  53     return _mm_srli_epi32(x, n);
  54 }
  55
  56 /* SHA256 Functions */
  57 #define BIGSIGMA0_256(x)    (ROTR((x), 2) ^ ROTR((x), 13) ^ ROTR((x), 22))
  58 #define BIGSIGMA1_256(x)    (ROTR((x), 6) ^ ROTR((x), 11) ^ ROTR((x), 25))
  59 #define SIGMA0_256(x)       (ROTR((x), 7) ^ ROTR((x), 18) ^ SHR((x), 3))
  60 #define SIGMA1_256(x)       (ROTR((x), 17) ^ ROTR((x), 19) ^ SHR((x), 10))
  61
  62 static inline unsigned int store32(const __m128i x, int i) {
  63     union { unsigned int ret[4]; __m128i x; } box;
  64     box.x = x;
  65     return box.ret[i];
  66 }
  67
  68 static inline void store_epi32(const __m128i x, unsigned int *x0, unsigned int *x1, unsigned int *x2, unsigned int *x3) {
  69     union { unsigned int ret[4]; __m128i x; } box;
  70     box.x = x;
  71     *x0 = box.ret[3]; *x1 = box.ret[2]; *x2 = box.ret[1]; *x3 = box.ret[0];
  72 }
  73
  74 #define add4(x0, x1, x2, x3) _mm_add_epi32(_mm_add_epi32(_mm_add_epi32(x0, x1), x2), x3)
  75 #define add5(x0, x1, x2, x3, x4) _mm_add_epi32(add4(x0, x1, x2, x3), x4)
  76
  77 #define SHA256ROUND(a, b, c, d, e, f, g, h, i, w)                       \
  78     T1 = add5(h, BIGSIGMA1_256(e), Ch(e, f, g), _mm_set1_epi32(sha256_consts[i]), w);   \
  79 d = _mm_add_epi32(d, T1);                                           \
  80 h = _mm_add_epi32(T1, _mm_add_epi32(BIGSIGMA0_256(a), Maj(a, b, c)));
  81
  82 static inline void dumpreg(__m128i x, char *msg) {
  83     union { unsigned int ret[4]; __m128i x; } box;
  84     box.x = x ;
  85     printf("%s %08x %08x %08x %08x\n", msg, box.ret[0], box.ret[1], box.ret[2], box.ret[3]);
  86 }
  87
  88 #if 1
  89 #define dumpstate(i) printf("%s: %08x %08x %08x %08x %08x %08x %08x %08x %08x\n", \
  90         __func__, store32(w0, i), store32(a, i), store32(b, i), store32(c, i), store32(d, i), store32(e, i), store32(f, i), store32(g, i), store32(h, i));
  91 #else
  92 #define dumpstate()
  93 #endif
  94
  95 // Align by increasing pointer, must have extra space at end of buffer
  96 template <size_t nBytes, typename T>
  97 T* alignup(T* p)
  98 {
  99     union
 100     {
 101         T* ptr;
 102         size_t n;
 103     } u;
 104     u.ptr = p;
 105     u.n = (u.n + (nBytes-1)) & ~(nBytes-1);
 106     return u.ptr;
 107 }
 108
 109 static const unsigned int pSHA256InitState[8] =
 110 {0x6a09e667, 0xbb67ae85, 0x3c6ef372, 0xa54ff53a, 0x510e527f, 0x9b05688c, 0x1f83d9ab, 0x5be0cd19};
 111
 112
 113 unsigned int ScanHash_4WaySSE2(char* pmidstate, char* pdata, char* phash1, char* phash, unsigned int& nHashesDone)
 114 {
 115     unsigned int& nNonce = *(unsigned int*)(pdata + 12);
 116     for (;;)
 117     {
 118         nNonce += NPAR;
 119         unsigned int thashbuf[9][NPAR];
 120         unsigned int (&thash)[9][NPAR] = *alignup<16>(&thashbuf);
 121         DoubleBlockSHA256(pdata, phash1, pmidstate, thash, pSHA256InitState);
 122
 123         for (int j = 0; j < NPAR; j++)
 124         {
 125             if (thash[7][j] == 0)
 126             {
 127                 for (int i = 0; i < 32/4; i++)
 128                     ((unsigned int*)phash)[i] = thash[i][j];
 129                 return nNonce + j;
 130             }
 131         }
 132
 133         if ((nNonce & 0xffff) == 0)
 134         {
 135             nHashesDone = 0xffff+1;
 136             return -1;
 137         }
 138     }
 139 }
 140
 141
 142 void DoubleBlockSHA256(const void* pin, void* pad, const void *pre, unsigned int thash[9][NPAR], const void *init)
 143 {
 144     unsigned int* In = (unsigned int*)pin;
 145     unsigned int* Pad = (unsigned int*)pad;
 146     unsigned int* hPre = (unsigned int*)pre;
 147     unsigned int* hInit = (unsigned int*)init;
 148     unsigned int i, j, k;
 149
 150     /* vectors used in calculation */
 151     __m128i w0, w1, w2, w3, w4, w5, w6, w7;
 152     __m128i w8, w9, w10, w11, w12, w13, w14, w15;
 153     __m128i T1;
 154     __m128i a, b, c, d, e, f, g, h;
 155     __m128i nonce;
 156
 157     /* nonce offset for vector */
 158     __m128i offset = _mm_set_epi32(0x00000003, 0x00000002, 0x00000001, 0x00000000);
 159
 160
 161     for(k = 0; k<NPAR; k+=4) {
 162         w0 = _mm_set1_epi32(In[0]);
 163         w1 = _mm_set1_epi32(In[1]);
 164         w2 = _mm_set1_epi32(In[2]);
 165         //w3 = _mm_set1_epi32(In[3]); nonce will be later hacked into the hash
 166         w4 = _mm_set1_epi32(In[4]);
 167         w5 = _mm_set1_epi32(In[5]);
 168         w6 = _mm_set1_epi32(In[6]);
 169         w7 = _mm_set1_epi32(In[7]);
 170         w8 = _mm_set1_epi32(In[8]);
 171         w9 = _mm_set1_epi32(In[9]);
 172         w10 = _mm_set1_epi32(In[10]);
 173         w11 = _mm_set1_epi32(In[11]);
 174         w12 = _mm_set1_epi32(In[12]);
 175         w13 = _mm_set1_epi32(In[13]);
 176         w14 = _mm_set1_epi32(In[14]);
 177         w15 = _mm_set1_epi32(In[15]);
 178
 179         /* hack nonce into lowest byte of w3 */
 180         nonce = _mm_set1_epi32(In[3]);
 181         nonce = _mm_add_epi32(nonce, offset);
 182         nonce = _mm_add_epi32(nonce, _mm_set1_epi32(k));
 183         w3 = nonce;
 184
 185         a = _mm_set1_epi32(hPre[0]);
 186         b = _mm_set1_epi32(hPre[1]);
 187         c = _mm_set1_epi32(hPre[2]);
 188         d = _mm_set1_epi32(hPre[3]);
 189         e = _mm_set1_epi32(hPre[4]);
 190         f = _mm_set1_epi32(hPre[5]);
 191         g = _mm_set1_epi32(hPre[6]);
 192         h = _mm_set1_epi32(hPre[7]);
 193
 194         SHA256ROUND(a, b, c, d, e, f, g, h, 0, w0);
 195         SHA256ROUND(h, a, b, c, d, e, f, g, 1, w1);
 196         SHA256ROUND(g, h, a, b, c, d, e, f, 2, w2);
 197         SHA256ROUND(f, g, h, a, b, c, d, e, 3, w3);
 198         SHA256ROUND(e, f, g, h, a, b, c, d, 4, w4);
 199         SHA256ROUND(d, e, f, g, h, a, b, c, 5, w5);
 200         SHA256ROUND(c, d, e, f, g, h, a, b, 6, w6);
 201         SHA256ROUND(b, c, d, e, f, g, h, a, 7, w7);
 202         SHA256ROUND(a, b, c, d, e, f, g, h, 8, w8);
 203         SHA256ROUND(h, a, b, c, d, e, f, g, 9, w9);
 204         SHA256ROUND(g, h, a, b, c, d, e, f, 10, w10);
 205         SHA256ROUND(f, g, h, a, b, c, d, e, 11, w11);
 206         SHA256ROUND(e, f, g, h, a, b, c, d, 12, w12);
 207         SHA256ROUND(d, e, f, g, h, a, b, c, 13, w13);
 208         SHA256ROUND(c, d, e, f, g, h, a, b, 14, w14);
 209         SHA256ROUND(b, c, d, e, f, g, h, a, 15, w15);
 210
 211         w0 = add4(SIGMA1_256(w14), w9, SIGMA0_256(w1), w0);
 212         SHA256ROUND(a, b, c, d, e, f, g, h, 16, w0);
 213         w1 = add4(SIGMA1_256(w15), w10, SIGMA0_256(w2), w1);
 214         SHA256ROUND(h, a, b, c, d, e, f, g, 17, w1);
 215         w2 = add4(SIGMA1_256(w0), w11, SIGMA0_256(w3), w2);
 216         SHA256ROUND(g, h, a, b, c, d, e, f, 18, w2);
 217         w3 = add4(SIGMA1_256(w1), w12, SIGMA0_256(w4), w3);
 218         SHA256ROUND(f, g, h, a, b, c, d, e, 19, w3);
 219         w4 = add4(SIGMA1_256(w2), w13, SIGMA0_256(w5), w4);
 220         SHA256ROUND(e, f, g, h, a, b, c, d, 20, w4);
 221         w5 = add4(SIGMA1_256(w3), w14, SIGMA0_256(w6), w5);
 222         SHA256ROUND(d, e, f, g, h, a, b, c, 21, w5);
 223         w6 = add4(SIGMA1_256(w4), w15, SIGMA0_256(w7), w6);
 224         SHA256ROUND(c, d, e, f, g, h, a, b, 22, w6);
 225         w7 = add4(SIGMA1_256(w5), w0, SIGMA0_256(w8), w7);
 226         SHA256ROUND(b, c, d, e, f, g, h, a, 23, w7);
 227         w8 = add4(SIGMA1_256(w6), w1, SIGMA0_256(w9), w8);
 228         SHA256ROUND(a, b, c, d, e, f, g, h, 24, w8);
 229         w9 = add4(SIGMA1_256(w7), w2, SIGMA0_256(w10), w9);
 230         SHA256ROUND(h, a, b, c, d, e, f, g, 25, w9);
 231         w10 = add4(SIGMA1_256(w8), w3, SIGMA0_256(w11), w10);
 232         SHA256ROUND(g, h, a, b, c, d, e, f, 26, w10);
 233         w11 = add4(SIGMA1_256(w9), w4, SIGMA0_256(w12), w11);
 234         SHA256ROUND(f, g, h, a, b, c, d, e, 27, w11);
 235         w12 = add4(SIGMA1_256(w10), w5, SIGMA0_256(w13), w12);
 236         SHA256ROUND(e, f, g, h, a, b, c, d, 28, w12);
 237         w13 = add4(SIGMA1_256(w11), w6, SIGMA0_256(w14), w13);
 238         SHA256ROUND(d, e, f, g, h, a, b, c, 29, w13);
 239         w14 = add4(SIGMA1_256(w12), w7, SIGMA0_256(w15), w14);
 240         SHA256ROUND(c, d, e, f, g, h, a, b, 30, w14);
 241         w15 = add4(SIGMA1_256(w13), w8, SIGMA0_256(w0), w15);
 242         SHA256ROUND(b, c, d, e, f, g, h, a, 31, w15);
 243
 244         w0 = add4(SIGMA1_256(w14), w9, SIGMA0_256(w1), w0);
 245         SHA256ROUND(a, b, c, d, e, f, g, h, 32, w0);
 246         w1 = add4(SIGMA1_256(w15), w10, SIGMA0_256(w2), w1);
 247         SHA256ROUND(h, a, b, c, d, e, f, g, 33, w1);
 248         w2 = add4(SIGMA1_256(w0), w11, SIGMA0_256(w3), w2);
 249         SHA256ROUND(g, h, a, b, c, d, e, f, 34, w2);
 250         w3 = add4(SIGMA1_256(w1), w12, SIGMA0_256(w4), w3);
 251         SHA256ROUND(f, g, h, a, b, c, d, e, 35, w3);
 252         w4 = add4(SIGMA1_256(w2), w13, SIGMA0_256(w5), w4);
 253         SHA256ROUND(e, f, g, h, a, b, c, d, 36, w4);
 254         w5 = add4(SIGMA1_256(w3), w14, SIGMA0_256(w6), w5);
 255         SHA256ROUND(d, e, f, g, h, a, b, c, 37, w5);
 256         w6 = add4(SIGMA1_256(w4), w15, SIGMA0_256(w7), w6);
 257         SHA256ROUND(c, d, e, f, g, h, a, b, 38, w6);
 258         w7 = add4(SIGMA1_256(w5), w0, SIGMA0_256(w8), w7);
 259         SHA256ROUND(b, c, d, e, f, g, h, a, 39, w7);
 260         w8 = add4(SIGMA1_256(w6), w1, SIGMA0_256(w9), w8);
 261         SHA256ROUND(a, b, c, d, e, f, g, h, 40, w8);
 262         w9 = add4(SIGMA1_256(w7), w2, SIGMA0_256(w10), w9);
 263         SHA256ROUND(h, a, b, c, d, e, f, g, 41, w9);
 264         w10 = add4(SIGMA1_256(w8), w3, SIGMA0_256(w11), w10);
 265         SHA256ROUND(g, h, a, b, c, d, e, f, 42, w10);
 266         w11 = add4(SIGMA1_256(w9), w4, SIGMA0_256(w12), w11);
 267         SHA256ROUND(f, g, h, a, b, c, d, e, 43, w11);
 268         w12 = add4(SIGMA1_256(w10), w5, SIGMA0_256(w13), w12);
 269         SHA256ROUND(e, f, g, h, a, b, c, d, 44, w12);
 270         w13 = add4(SIGMA1_256(w11), w6, SIGMA0_256(w14), w13);
 271         SHA256ROUND(d, e, f, g, h, a, b, c, 45, w13);
 272         w14 = add4(SIGMA1_256(w12), w7, SIGMA0_256(w15), w14);
 273         SHA256ROUND(c, d, e, f, g, h, a, b, 46, w14);
 274         w15 = add4(SIGMA1_256(w13), w8, SIGMA0_256(w0), w15);
 275         SHA256ROUND(b, c, d, e, f, g, h, a, 47, w15);
 276
 277         w0 = add4(SIGMA1_256(w14), w9, SIGMA0_256(w1), w0);
 278         SHA256ROUND(a, b, c, d, e, f, g, h, 48, w0);
 279         w1 = add4(SIGMA1_256(w15), w10, SIGMA0_256(w2), w1);
 280         SHA256ROUND(h, a, b, c, d, e, f, g, 49, w1);
 281         w2 = add4(SIGMA1_256(w0), w11, SIGMA0_256(w3), w2);
 282         SHA256ROUND(g, h, a, b, c, d, e, f, 50, w2);
 283         w3 = add4(SIGMA1_256(w1), w12, SIGMA0_256(w4), w3);
 284         SHA256ROUND(f, g, h, a, b, c, d, e, 51, w3);
 285         w4 = add4(SIGMA1_256(w2), w13, SIGMA0_256(w5), w4);
 286         SHA256ROUND(e, f, g, h, a, b, c, d, 52, w4);
 287         w5 = add4(SIGMA1_256(w3), w14, SIGMA0_256(w6), w5);
 288         SHA256ROUND(d, e, f, g, h, a, b, c, 53, w5);
 289         w6 = add4(SIGMA1_256(w4), w15, SIGMA0_256(w7), w6);
 290         SHA256ROUND(c, d, e, f, g, h, a, b, 54, w6);
 291         w7 = add4(SIGMA1_256(w5), w0, SIGMA0_256(w8), w7);
 292         SHA256ROUND(b, c, d, e, f, g, h, a, 55, w7);
 293         w8 = add4(SIGMA1_256(w6), w1, SIGMA0_256(w9), w8);
 294         SHA256ROUND(a, b, c, d, e, f, g, h, 56, w8);
 295         w9 = add4(SIGMA1_256(w7), w2, SIGMA0_256(w10), w9);
 296         SHA256ROUND(h, a, b, c, d, e, f, g, 57, w9);
 297         w10 = add4(SIGMA1_256(w8), w3, SIGMA0_256(w11), w10);
 298         SHA256ROUND(g, h, a, b, c, d, e, f, 58, w10);
 299         w11 = add4(SIGMA1_256(w9), w4, SIGMA0_256(w12), w11);
 300         SHA256ROUND(f, g, h, a, b, c, d, e, 59, w11);
 301         w12 = add4(SIGMA1_256(w10), w5, SIGMA0_256(w13), w12);
 302         SHA256ROUND(e, f, g, h, a, b, c, d, 60, w12);
 303         w13 = add4(SIGMA1_256(w11), w6, SIGMA0_256(w14), w13);
 304         SHA256ROUND(d, e, f, g, h, a, b, c, 61, w13);
 305         w14 = add4(SIGMA1_256(w12), w7, SIGMA0_256(w15), w14);
 306         SHA256ROUND(c, d, e, f, g, h, a, b, 62, w14);
 307         w15 = add4(SIGMA1_256(w13), w8, SIGMA0_256(w0), w15);
 308         SHA256ROUND(b, c, d, e, f, g, h, a, 63, w15);
 309
 310 #define store_load(x, i, dest) \
 311         T1 = _mm_set1_epi32((hPre)[i]); \
 312         dest = _mm_add_epi32(T1, x);
 313
 314         store_load(a, 0, w0);
 315         store_load(b, 1, w1);
 316         store_load(c, 2, w2);
 317         store_load(d, 3, w3);
 318         store_load(e, 4, w4);
 319         store_load(f, 5, w5);
 320         store_load(g, 6, w6);
 321         store_load(h, 7, w7);
 322
 323         w8 = _mm_set1_epi32(Pad[8]);
 324         w9 = _mm_set1_epi32(Pad[9]);
 325         w10 = _mm_set1_epi32(Pad[10]);
 326         w11 = _mm_set1_epi32(Pad[11]);
 327         w12 = _mm_set1_epi32(Pad[12]);
 328         w13 = _mm_set1_epi32(Pad[13]);
 329         w14 = _mm_set1_epi32(Pad[14]);
 330         w15 = _mm_set1_epi32(Pad[15]);
 331
 332         a = _mm_set1_epi32(hInit[0]);
 333         b = _mm_set1_epi32(hInit[1]);
 334         c = _mm_set1_epi32(hInit[2]);
 335         d = _mm_set1_epi32(hInit[3]);
 336         e = _mm_set1_epi32(hInit[4]);
 337         f = _mm_set1_epi32(hInit[5]);
 338         g = _mm_set1_epi32(hInit[6]);
 339         h = _mm_set1_epi32(hInit[7]);
 340
 341         SHA256ROUND(a, b, c, d, e, f, g, h, 0, w0);
 342         SHA256ROUND(h, a, b, c, d, e, f, g, 1, w1);
 343         SHA256ROUND(g, h, a, b, c, d, e, f, 2, w2);
 344         SHA256ROUND(f, g, h, a, b, c, d, e, 3, w3);
 345         SHA256ROUND(e, f, g, h, a, b, c, d, 4, w4);
 346         SHA256ROUND(d, e, f, g, h, a, b, c, 5, w5);
 347         SHA256ROUND(c, d, e, f, g, h, a, b, 6, w6);
 348         SHA256ROUND(b, c, d, e, f, g, h, a, 7, w7);
 349         SHA256ROUND(a, b, c, d, e, f, g, h, 8, w8);
 350         SHA256ROUND(h, a, b, c, d, e, f, g, 9, w9);
 351         SHA256ROUND(g, h, a, b, c, d, e, f, 10, w10);
 352         SHA256ROUND(f, g, h, a, b, c, d, e, 11, w11);
 353         SHA256ROUND(e, f, g, h, a, b, c, d, 12, w12);
 354         SHA256ROUND(d, e, f, g, h, a, b, c, 13, w13);
 355         SHA256ROUND(c, d, e, f, g, h, a, b, 14, w14);
 356         SHA256ROUND(b, c, d, e, f, g, h, a, 15, w15);
 357
 358         w0 = add4(SIGMA1_256(w14), w9, SIGMA0_256(w1), w0);
 359         SHA256ROUND(a, b, c, d, e, f, g, h, 16, w0);
 360         w1 = add4(SIGMA1_256(w15), w10, SIGMA0_256(w2), w1);
 361         SHA256ROUND(h, a, b, c, d, e, f, g, 17, w1);
 362         w2 = add4(SIGMA1_256(w0), w11, SIGMA0_256(w3), w2);
 363         SHA256ROUND(g, h, a, b, c, d, e, f, 18, w2);
 364         w3 = add4(SIGMA1_256(w1), w12, SIGMA0_256(w4), w3);
 365         SHA256ROUND(f, g, h, a, b, c, d, e, 19, w3);
 366         w4 = add4(SIGMA1_256(w2), w13, SIGMA0_256(w5), w4);
 367         SHA256ROUND(e, f, g, h, a, b, c, d, 20, w4);
 368         w5 = add4(SIGMA1_256(w3), w14, SIGMA0_256(w6), w5);
 369         SHA256ROUND(d, e, f, g, h, a, b, c, 21, w5);
 370         w6 = add4(SIGMA1_256(w4), w15, SIGMA0_256(w7), w6);
 371         SHA256ROUND(c, d, e, f, g, h, a, b, 22, w6);
 372         w7 = add4(SIGMA1_256(w5), w0, SIGMA0_256(w8), w7);
 373         SHA256ROUND(b, c, d, e, f, g, h, a, 23, w7);
 374         w8 = add4(SIGMA1_256(w6), w1, SIGMA0_256(w9), w8);
 375         SHA256ROUND(a, b, c, d, e, f, g, h, 24, w8);
 376         w9 = add4(SIGMA1_256(w7), w2, SIGMA0_256(w10), w9);
 377         SHA256ROUND(h, a, b, c, d, e, f, g, 25, w9);
 378         w10 = add4(SIGMA1_256(w8), w3, SIGMA0_256(w11), w10);
 379         SHA256ROUND(g, h, a, b, c, d, e, f, 26, w10);
 380         w11 = add4(SIGMA1_256(w9), w4, SIGMA0_256(w12), w11);
 381         SHA256ROUND(f, g, h, a, b, c, d, e, 27, w11);
 382         w12 = add4(SIGMA1_256(w10), w5, SIGMA0_256(w13), w12);
 383         SHA256ROUND(e, f, g, h, a, b, c, d, 28, w12);
 384         w13 = add4(SIGMA1_256(w11), w6, SIGMA0_256(w14), w13);
 385         SHA256ROUND(d, e, f, g, h, a, b, c, 29, w13);
 386         w14 = add4(SIGMA1_256(w12), w7, SIGMA0_256(w15), w14);
 387         SHA256ROUND(c, d, e, f, g, h, a, b, 30, w14);
 388         w15 = add4(SIGMA1_256(w13), w8, SIGMA0_256(w0), w15);
 389         SHA256ROUND(b, c, d, e, f, g, h, a, 31, w15);
 390
 391         w0 = add4(SIGMA1_256(w14), w9, SIGMA0_256(w1), w0);
 392         SHA256ROUND(a, b, c, d, e, f, g, h, 32, w0);
 393         w1 = add4(SIGMA1_256(w15), w10, SIGMA0_256(w2), w1);
 394         SHA256ROUND(h, a, b, c, d, e, f, g, 33, w1);
 395         w2 = add4(SIGMA1_256(w0), w11, SIGMA0_256(w3), w2);
 396         SHA256ROUND(g, h, a, b, c, d, e, f, 34, w2);
 397         w3 = add4(SIGMA1_256(w1), w12, SIGMA0_256(w4), w3);
 398         SHA256ROUND(f, g, h, a, b, c, d, e, 35, w3);
 399         w4 = add4(SIGMA1_256(w2), w13, SIGMA0_256(w5), w4);
 400         SHA256ROUND(e, f, g, h, a, b, c, d, 36, w4);
 401         w5 = add4(SIGMA1_256(w3), w14, SIGMA0_256(w6), w5);
 402         SHA256ROUND(d, e, f, g, h, a, b, c, 37, w5);
 403         w6 = add4(SIGMA1_256(w4), w15, SIGMA0_256(w7), w6);
 404         SHA256ROUND(c, d, e, f, g, h, a, b, 38, w6);
 405         w7 = add4(SIGMA1_256(w5), w0, SIGMA0_256(w8), w7);
 406         SHA256ROUND(b, c, d, e, f, g, h, a, 39, w7);
 407         w8 = add4(SIGMA1_256(w6), w1, SIGMA0_256(w9), w8);
 408         SHA256ROUND(a, b, c, d, e, f, g, h, 40, w8);
 409         w9 = add4(SIGMA1_256(w7), w2, SIGMA0_256(w10), w9);
 410         SHA256ROUND(h, a, b, c, d, e, f, g, 41, w9);
 411         w10 = add4(SIGMA1_256(w8), w3, SIGMA0_256(w11), w10);
 412         SHA256ROUND(g, h, a, b, c, d, e, f, 42, w10);
 413         w11 = add4(SIGMA1_256(w9), w4, SIGMA0_256(w12), w11);
 414         SHA256ROUND(f, g, h, a, b, c, d, e, 43, w11);
 415         w12 = add4(SIGMA1_256(w10), w5, SIGMA0_256(w13), w12);
 416         SHA256ROUND(e, f, g, h, a, b, c, d, 44, w12);
 417         w13 = add4(SIGMA1_256(w11), w6, SIGMA0_256(w14), w13);
 418         SHA256ROUND(d, e, f, g, h, a, b, c, 45, w13);
 419         w14 = add4(SIGMA1_256(w12), w7, SIGMA0_256(w15), w14);
 420         SHA256ROUND(c, d, e, f, g, h, a, b, 46, w14);
 421         w15 = add4(SIGMA1_256(w13), w8, SIGMA0_256(w0), w15);
 422         SHA256ROUND(b, c, d, e, f, g, h, a, 47, w15);
 423
 424         w0 = add4(SIGMA1_256(w14), w9, SIGMA0_256(w1), w0);
 425         SHA256ROUND(a, b, c, d, e, f, g, h, 48, w0);
 426         w1 = add4(SIGMA1_256(w15), w10, SIGMA0_256(w2), w1);
 427         SHA256ROUND(h, a, b, c, d, e, f, g, 49, w1);
 428         w2 = add4(SIGMA1_256(w0), w11, SIGMA0_256(w3), w2);
 429         SHA256ROUND(g, h, a, b, c, d, e, f, 50, w2);
 430         w3 = add4(SIGMA1_256(w1), w12, SIGMA0_256(w4), w3);
 431         SHA256ROUND(f, g, h, a, b, c, d, e, 51, w3);
 432         w4 = add4(SIGMA1_256(w2), w13, SIGMA0_256(w5), w4);
 433         SHA256ROUND(e, f, g, h, a, b, c, d, 52, w4);
 434         w5 = add4(SIGMA1_256(w3), w14, SIGMA0_256(w6), w5);
 435         SHA256ROUND(d, e, f, g, h, a, b, c, 53, w5);
 436         w6 = add4(SIGMA1_256(w4), w15, SIGMA0_256(w7), w6);
 437         SHA256ROUND(c, d, e, f, g, h, a, b, 54, w6);
 438         w7 = add4(SIGMA1_256(w5), w0, SIGMA0_256(w8), w7);
 439         SHA256ROUND(b, c, d, e, f, g, h, a, 55, w7);
 440         w8 = add4(SIGMA1_256(w6), w1, SIGMA0_256(w9), w8);
 441         SHA256ROUND(a, b, c, d, e, f, g, h, 56, w8);
 442         w9 = add4(SIGMA1_256(w7), w2, SIGMA0_256(w10), w9);
 443         SHA256ROUND(h, a, b, c, d, e, f, g, 57, w9);
 444         w10 = add4(SIGMA1_256(w8), w3, SIGMA0_256(w11), w10);
 445         SHA256ROUND(g, h, a, b, c, d, e, f, 58, w10);
 446         w11 = add4(SIGMA1_256(w9), w4, SIGMA0_256(w12), w11);
 447         SHA256ROUND(f, g, h, a, b, c, d, e, 59, w11);
 448         w12 = add4(SIGMA1_256(w10), w5, SIGMA0_256(w13), w12);
 449         SHA256ROUND(e, f, g, h, a, b, c, d, 60, w12);
 450         w13 = add4(SIGMA1_256(w11), w6, SIGMA0_256(w14), w13);
 451         SHA256ROUND(d, e, f, g, h, a, b, c, 61, w13);
 452         w14 = add4(SIGMA1_256(w12), w7, SIGMA0_256(w15), w14);
 453         SHA256ROUND(c, d, e, f, g, h, a, b, 62, w14);
 454         w15 = add4(SIGMA1_256(w13), w8, SIGMA0_256(w0), w15);
 455         SHA256ROUND(b, c, d, e, f, g, h, a, 63, w15);
 456
 457         /* store resulsts directly in thash */
 458 #define store_2(x,i)  \
 459         w0 = _mm_set1_epi32(hInit[i]); \
 460         *(__m128i *)&(thash)[i][0+k] = _mm_add_epi32(w0, x);
 461
 462         store_2(a, 0);
 463         store_2(b, 1);
 464         store_2(c, 2);
 465         store_2(d, 3);
 466         store_2(e, 4);
 467         store_2(f, 5);
 468         store_2(g, 6);
 469         store_2(h, 7);
 470         *(__m128i *)&(thash)[8][0+k] = nonce;
 471     }
 472
 473 }
 474
 475 #endif // FOURWAYSSE2