/*                                                                *
 * This is an OpenCL implementation of the encryption algorithm:  *
 *                                                                *
 *   GOST R 34.12-2015 Kuznyechik by A.S.Kuzmin and A.A.Nechaev   *
 *                                                                *
 * Author of the original C implementation:                       *
 *                                                                *
 *   Markku-Juhani O. Saarinen <mjos@iki.fi>                      *
 *   https://github.com/mjosaarinen/kuznechik                     *
 *                                                                *
 * Adapted for GPU use with hashcat by Ruslan Yushaev.            *
 *                                                                *
 */

#include "inc_vendor.h"
#include "inc_types.h"
#include "inc_platform.h"
#include "inc_common.h"
#include "inc_cipher_kuznyechik.h"

CONSTANT_AS u32a k_sbox[256] =
{
  0xfc, 0xee, 0xdd, 0x11, 0xcf, 0x6e, 0x31, 0x16,
  0xfb, 0xc4, 0xfa, 0xda, 0x23, 0xc5, 0x04, 0x4d,
  0xe9, 0x77, 0xf0, 0xdb, 0x93, 0x2e, 0x99, 0xba,
  0x17, 0x36, 0xf1, 0xbb, 0x14, 0xcd, 0x5f, 0xc1,
  0xf9, 0x18, 0x65, 0x5a, 0xe2, 0x5c, 0xef, 0x21,
  0x81, 0x1c, 0x3c, 0x42, 0x8b, 0x01, 0x8e, 0x4f,
  0x05, 0x84, 0x02, 0xae, 0xe3, 0x6a, 0x8f, 0xa0,
  0x06, 0x0b, 0xed, 0x98, 0x7f, 0xd4, 0xd3, 0x1f,
  0xeb, 0x34, 0x2c, 0x51, 0xea, 0xc8, 0x48, 0xab,
  0xf2, 0x2a, 0x68, 0xa2, 0xfd, 0x3a, 0xce, 0xcc,
  0xb5, 0x70, 0x0e, 0x56, 0x08, 0x0c, 0x76, 0x12,
  0xbf, 0x72, 0x13, 0x47, 0x9c, 0xb7, 0x5d, 0x87,
  0x15, 0xa1, 0x96, 0x29, 0x10, 0x7b, 0x9a, 0xc7,
  0xf3, 0x91, 0x78, 0x6f, 0x9d, 0x9e, 0xb2, 0xb1,
  0x32, 0x75, 0x19, 0x3d, 0xff, 0x35, 0x8a, 0x7e,
  0x6d, 0x54, 0xc6, 0x80, 0xc3, 0xbd, 0x0d, 0x57,
  0xdf, 0xf5, 0x24, 0xa9, 0x3e, 0xa8, 0x43, 0xc9,
  0xd7, 0x79, 0xd6, 0xf6, 0x7c, 0x22, 0xb9, 0x03,
  0xe0, 0x0f, 0xec, 0xde, 0x7a, 0x94, 0xb0, 0xbc,
  0xdc, 0xe8, 0x28, 0x50, 0x4e, 0x33, 0x0a, 0x4a,
  0xa7, 0x97, 0x60, 0x73, 0x1e, 0x00, 0x62, 0x44,
  0x1a, 0xb8, 0x38, 0x82, 0x64, 0x9f, 0x26, 0x41,
  0xad, 0x45, 0x46, 0x92, 0x27, 0x5e, 0x55, 0x2f,
  0x8c, 0xa3, 0xa5, 0x7d, 0x69, 0xd5, 0x95, 0x3b,
  0x07, 0x58, 0xb3, 0x40, 0x86, 0xac, 0x1d, 0xf7,
  0x30, 0x37, 0x6b, 0xe4, 0x88, 0xd9, 0xe7, 0x89,
  0xe1, 0x1b, 0x83, 0x49, 0x4c, 0x3f, 0xf8, 0xfe,
  0x8d, 0x53, 0xaa, 0x90, 0xca, 0xd8, 0x85, 0x61,
  0x20, 0x71, 0x67, 0xa4, 0x2d, 0x2b, 0x09, 0x5b,
  0xcb, 0x9b, 0x25, 0xd0, 0xbe, 0xe5, 0x6c, 0x52,
  0x59, 0xa6, 0x74, 0xd2, 0xe6, 0xf4, 0xb4, 0xc0,
  0xd1, 0x66, 0xaf, 0xc2, 0x39, 0x4b, 0x63, 0xb6
};

CONSTANT_AS u32a k_sbox_inv[256] =
{
  0xa5, 0x2d, 0x32, 0x8f, 0x0e, 0x30, 0x38, 0xc0,
  0x54, 0xe6, 0x9e, 0x39, 0x55, 0x7e, 0x52, 0x91,
  0x64, 0x03, 0x57, 0x5a, 0x1c, 0x60, 0x07, 0x18,
  0x21, 0x72, 0xa8, 0xd1, 0x29, 0xc6, 0xa4, 0x3f,
  0xe0, 0x27, 0x8d, 0x0c, 0x82, 0xea, 0xae, 0xb4,
  0x9a, 0x63, 0x49, 0xe5, 0x42, 0xe4, 0x15, 0xb7,
  0xc8, 0x06, 0x70, 0x9d, 0x41, 0x75, 0x19, 0xc9,
  0xaa, 0xfc, 0x4d, 0xbf, 0x2a, 0x73, 0x84, 0xd5,
  0xc3, 0xaf, 0x2b, 0x86, 0xa7, 0xb1, 0xb2, 0x5b,
  0x46, 0xd3, 0x9f, 0xfd, 0xd4, 0x0f, 0x9c, 0x2f,
  0x9b, 0x43, 0xef, 0xd9, 0x79, 0xb6, 0x53, 0x7f,
  0xc1, 0xf0, 0x23, 0xe7, 0x25, 0x5e, 0xb5, 0x1e,
  0xa2, 0xdf, 0xa6, 0xfe, 0xac, 0x22, 0xf9, 0xe2,
  0x4a, 0xbc, 0x35, 0xca, 0xee, 0x78, 0x05, 0x6b,
  0x51, 0xe1, 0x59, 0xa3, 0xf2, 0x71, 0x56, 0x11,
  0x6a, 0x89, 0x94, 0x65, 0x8c, 0xbb, 0x77, 0x3c,
  0x7b, 0x28, 0xab, 0xd2, 0x31, 0xde, 0xc4, 0x5f,
  0xcc, 0xcf, 0x76, 0x2c, 0xb8, 0xd8, 0x2e, 0x36,
  0xdb, 0x69, 0xb3, 0x14, 0x95, 0xbe, 0x62, 0xa1,
  0x3b, 0x16, 0x66, 0xe9, 0x5c, 0x6c, 0x6d, 0xad,
  0x37, 0x61, 0x4b, 0xb9, 0xe3, 0xba, 0xf1, 0xa0,
  0x85, 0x83, 0xda, 0x47, 0xc5, 0xb0, 0x33, 0xfa,
  0x96, 0x6f, 0x6e, 0xc2, 0xf6, 0x50, 0xff, 0x5d,
  0xa9, 0x8e, 0x17, 0x1b, 0x97, 0x7d, 0xec, 0x58,
  0xf7, 0x1f, 0xfb, 0x7c, 0x09, 0x0d, 0x7a, 0x67,
  0x45, 0x87, 0xdc, 0xe8, 0x4f, 0x1d, 0x4e, 0x04,
  0xeb, 0xf8, 0xf3, 0x3e, 0x3d, 0xbd, 0x8a, 0x88,
  0xdd, 0xcd, 0x0b, 0x13, 0x98, 0x02, 0x93, 0x80,
  0x90, 0xd0, 0x24, 0x34, 0xcb, 0xed, 0xf4, 0xce,
  0x99, 0x10, 0x44, 0x40, 0x92, 0x3a, 0x01, 0x26,
  0x12, 0x1a, 0x48, 0x68, 0xf5, 0x81, 0x8b, 0xc7,
  0xd6, 0x20, 0x0a, 0x08, 0x00, 0x4c, 0xd7, 0x74
};

#define extract_byte(x,n) (((x) >> (8 * (n))) & 0xff)

#define k_lookup(w,sbox)                      \
  for (int i = 0; i < 4; i++)                 \
    w[i] = sbox[extract_byte (w[i], 0)] <<  0 \
         | sbox[extract_byte (w[i], 1)] <<  8 \
         | sbox[extract_byte (w[i], 2)] << 16 \
         | sbox[extract_byte (w[i], 3)] << 24

#define k_xor(n)                      \
  for (int i = (n); i > 0; i /= 2)    \
  {                                   \
    z ^= x * (i % 2);                 \
    x = (x << 1) ^ ((x >> 7) * 0xc3); \
    x &= 0xff;                        \
  }

DECLSPEC void kuznyechik_linear (u32 *w)
{
  // used in k_xor macro
  u32 x;
  u32 z;

  for (int i = 0; i < 16; i++)
  {
    z = 0;

    // k_xor (1) yields the same result as a simple xor
    x = extract_byte (w[3], 3); z ^= x;
    x = extract_byte (w[3], 2); k_xor (148);
    x = extract_byte (w[3], 1); k_xor (32);
    x = extract_byte (w[3], 0); k_xor (133);
    x = extract_byte (w[2], 3); k_xor (16);
    x = extract_byte (w[2], 2); k_xor (194);
    x = extract_byte (w[2], 1); k_xor (192);
    x = extract_byte (w[2], 0); z ^= x;
    x = extract_byte (w[1], 3); k_xor (251);
    x = extract_byte (w[1], 2); z ^= x;
    x = extract_byte (w[1], 1); k_xor (192);
    x = extract_byte (w[1], 0); k_xor (194);
    x = extract_byte (w[0], 3); k_xor (16);
    x = extract_byte (w[0], 2); k_xor (133);
    x = extract_byte (w[0], 1); k_xor (32);
    x = extract_byte (w[0], 0); k_xor (148);

    // right-shift data block, prepend calculated byte
    w[3] = (w[3] << 8) | (w[2] >> 24);
    w[2] = (w[2] << 8) | (w[1] >> 24);
    w[1] = (w[1] << 8) | (w[0] >> 24);
    w[0] = (w[0] << 8) | z;
  }
}

DECLSPEC void kuznyechik_linear_inv (u32 *w)
{
  // used in k_xor macro
  u32 x;
  u32 z;

  for (int i = 0; i < 16; i++)
  {
    z = extract_byte (w[0], 0);

    //left-shift data block
    w[0] = (w[0] >> 8) | (w[1] << 24);
    w[1] = (w[1] >> 8) | (w[2] << 24);
    w[2] = (w[2] >> 8) | (w[3] << 24);
    w[3] = (w[3] >> 8);

    x = extract_byte (w[0], 0); k_xor (148);
    x = extract_byte (w[0], 1); k_xor (32);
    x = extract_byte (w[0], 2); k_xor (133);
    x = extract_byte (w[0], 3); k_xor (16);
    x = extract_byte (w[1], 0); k_xor (194);
    x = extract_byte (w[1], 1); k_xor (192);
    x = extract_byte (w[1], 2); z ^= x;
    x = extract_byte (w[1], 3); k_xor (251);
    x = extract_byte (w[2], 0); z ^= x;
    x = extract_byte (w[2], 1); k_xor (192);
    x = extract_byte (w[2], 2); k_xor (194);
    x = extract_byte (w[2], 3); k_xor (16);
    x = extract_byte (w[3], 0); k_xor (133);
    x = extract_byte (w[3], 1); k_xor (32);
    x = extract_byte (w[3], 2); k_xor (148);

    //append calculated byte
    w[3] |= (z << 24);
  }
}

DECLSPEC void kuznyechik_set_key (u32 *ks, const u32 *ukey)
{
  u32 counter[4];
  u32 x[4];
  u32 y[4];
  u32 z[4];

  x[0] = ukey[0];
  x[1] = ukey[1];
  x[2] = ukey[2];
  x[3] = ukey[3];

  y[0] = ukey[4];
  y[1] = ukey[5];
  y[2] = ukey[6];
  y[3] = ukey[7];

  ks[0] = ukey[0];
  ks[1] = ukey[1];
  ks[2] = ukey[2];
  ks[3] = ukey[3];
  ks[4] = ukey[4];
  ks[5] = ukey[5];
  ks[6] = ukey[6];
  ks[7] = ukey[7];

  for (int i = 1; i <= 32; i++)
  {
    counter[0] = 0;
    counter[1] = 0;
    counter[2] = 0;
    counter[3] = (i << 24);

    kuznyechik_linear (counter);

    z[0] = x[0] ^ counter[0];
    z[1] = x[1] ^ counter[1];
    z[2] = x[2] ^ counter[2];
    z[3] = x[3] ^ counter[3];

    k_lookup (z, k_sbox);

    kuznyechik_linear (z);

    z[0] ^= y[0];
    z[1] ^= y[1];
    z[2] ^= y[2];
    z[3] ^= y[3];

    y[0] = x[0];
    y[1] = x[1];
    y[2] = x[2];
    y[3] = x[3];

    x[0] = z[0];
    x[1] = z[1];
    x[2] = z[2];
    x[3] = z[3];

    if ((i & 7) == 0)
    {
      ks[i + 0] = x[0];
      ks[i + 1] = x[1];
      ks[i + 2] = x[2];
      ks[i + 3] = x[3];

      ks[i + 4] = y[0];
      ks[i + 5] = y[1];
      ks[i + 6] = y[2];
      ks[i + 7] = y[3];
    }
  }
}

DECLSPEC void kuznyechik_encrypt (const u32 *ks, const u32 *in, u32 *out)
{
  out[0] = in[0];
  out[1] = in[1];
  out[2] = in[2];
  out[3] = in[3];

  for (int i = 0; i < 9; i++)
  {
    out[0] ^= ks[4 * i + 0];
    out[1] ^= ks[4 * i + 1];
    out[2] ^= ks[4 * i + 2];
    out[3] ^= ks[4 * i + 3];

    k_lookup (out, k_sbox);

    kuznyechik_linear (out);
  }

  out[0] ^= ks[4 * 9 + 0];
  out[1] ^= ks[4 * 9 + 1];
  out[2] ^= ks[4 * 9 + 2];
  out[3] ^= ks[4 * 9 + 3];
}

DECLSPEC void kuznyechik_decrypt (const u32 *ks, const u32 *in, u32 *out)
{
  out[0] = in[0];
  out[1] = in[1];
  out[2] = in[2];
  out[3] = in[3];

  out[0] ^= ks[4 * 9 + 0];
  out[1] ^= ks[4 * 9 + 1];
  out[2] ^= ks[4 * 9 + 2];
  out[3] ^= ks[4 * 9 + 3];

  for (int i = 8; i >= 0; i--)
  {
    kuznyechik_linear_inv (out);

    k_lookup (out, k_sbox_inv);

    out[0] ^= ks[4 * i + 0];
    out[1] ^= ks[4 * i + 1];
    out[2] ^= ks[4 * i + 2];
    out[3] ^= ks[4 * i + 3];
  }
}

#undef k_xor
#undef k_lookup
#undef extract_byte