hashcat/OpenCL/m33502_a3-optimized.cl

/**
 * Author......: See docs/credits.txt
 * License.....: MIT
 */

//#define NEW_SIMD_CODE

#ifdef KERNEL_STATIC
#include M2S(INCLUDE_PATH/inc_vendor.h)
#include M2S(INCLUDE_PATH/inc_types.h)
#include M2S(INCLUDE_PATH/inc_platform.cl)
#include M2S(INCLUDE_PATH/inc_common.cl)
#include M2S(INCLUDE_PATH/inc_simd.cl)
#include M2S(INCLUDE_PATH/inc_cipher_rc4.cl)
#endif

typedef struct rc4
{
  u32 dropN;
  u32 ct_len;
  u32 pt_len;
  u32 pt_off;

  u32 pt[2];
  u32 ct[16];

} rc4_t;

CONSTANT_VK u32 pt_masks[16] =
{
  0x00000000,
  0x000000FF,
  0x0000FFFF,
  0x00FFFFFF,
  0xFFFFFFFF,
  0x000000FF,
  0x0000FFFF,
  0x00FFFFFF,
  0xFFFFFFFF,
  0x000000FF,
  0x0000FFFF,
  0x00FFFFFF,
  0xFFFFFFFF,
  0x000000FF,
  0,
  0
};

DECLSPEC void m33502m (LOCAL_AS u32 *S, PRIVATE_AS u32 *w0, PRIVATE_AS u32 *w1, PRIVATE_AS u32 *w2, PRIVATE_AS u32 *w3, const u32 pw_len, KERN_ATTR_FUNC_ESALT (rc4_t))
{
  /**
   * modifiers are taken from args
   */

  /**
   * loop
   */

  const u32 dropN  = esalt_bufs[DIGESTS_OFFSET_HOST].dropN;
  const u32 pt_len = esalt_bufs[DIGESTS_OFFSET_HOST].pt_len;

  const u32 ct[4] =
  {
    esalt_bufs[DIGESTS_OFFSET_HOST].ct[0],
    esalt_bufs[DIGESTS_OFFSET_HOST].ct[1],
    esalt_bufs[DIGESTS_OFFSET_HOST].ct[2],
    esalt_bufs[DIGESTS_OFFSET_HOST].ct[3]
  };

  u32 w0l = w0[0];

  for (u32 il_pos = 0; il_pos < IL_CNT; il_pos += VECT_SIZE)
  {
    const u32 w0r = ix_create_bft (bfs_buf, il_pos);

    const u32 w0lr = w0l | w0r;

    w0[0] = w0lr;

    rc4_init_104 (S, w0, lid);

    u32 out[4];

    u8 i = 0;
    u8 j = 0;

    if (dropN > 0)
    {
      rc4_dropN (S, &i, &j, dropN, lid);
    }

    rc4_next_16 (S, i, j, ct, out, lid);

    if (pt_len == 13)
    {
      out[3] &= pt_masks[1];
    }
    else
    {
      out[3] = 0;

      if (pt_len < 9)
      {
        out[2] = 0;

        if (pt_len < 5)
        {
          out[1] = 0;

          if (pt_len >= 1 && pt_len <= 3)
          {
            out[0] &= pt_masks[pt_len];
          }
        }
        else if (pt_len <= 7)
        {
          out[1] &= pt_masks[pt_len];
        }
      }
      else if (pt_len <= 11)
      {
        out[2] &= pt_masks[pt_len];
      }
    }

    COMPARE_M_SIMD (out[0], out[1], out[2], out[3]);
  }
}

DECLSPEC void m33502s (LOCAL_AS u32 *S, PRIVATE_AS u32 *w0, PRIVATE_AS u32 *w1, PRIVATE_AS u32 *w2, PRIVATE_AS u32 *w3, const u32 pw_len, KERN_ATTR_FUNC_ESALT (rc4_t))
{
  /**
   * modifiers are taken from args
   */

  /**
   * digest
   */

  const u32 search[4] =
  {
    digests_buf[DIGESTS_OFFSET_HOST].digest_buf[DGST_R0],
    digests_buf[DIGESTS_OFFSET_HOST].digest_buf[DGST_R1],
    digests_buf[DIGESTS_OFFSET_HOST].digest_buf[DGST_R2],
    digests_buf[DIGESTS_OFFSET_HOST].digest_buf[DGST_R3]
  };

  /**
   * loop
   */

  const u32 dropN  = esalt_bufs[DIGESTS_OFFSET_HOST].dropN;
  const u32 pt_len = esalt_bufs[DIGESTS_OFFSET_HOST].pt_len;

  const u32 ct[4] =
  {
    esalt_bufs[DIGESTS_OFFSET_HOST].ct[0],
    esalt_bufs[DIGESTS_OFFSET_HOST].ct[1],
    esalt_bufs[DIGESTS_OFFSET_HOST].ct[2],
    esalt_bufs[DIGESTS_OFFSET_HOST].ct[3]
  };

  u32 w0l = w0[0];

  for (u32 il_pos = 0; il_pos < IL_CNT; il_pos += VECT_SIZE)
  {
    const u32 w0r = ix_create_bft (bfs_buf, il_pos);

    const u32 w0lr = w0l | w0r;

    w0[0] = w0lr;

    rc4_init_104 (S, w0, lid);

    u32 out[4];

    u8 i = 0;
    u8 j = 0;

    if (dropN > 0)
    {
      rc4_dropN (S, &i, &j, dropN, lid);
    }

    rc4_next_16 (S, i, j, ct, out, lid);

    if (pt_len == 13)
    {
      out[3] &= pt_masks[1];
    }
    else
    {
      out[3] = 0;

      if (pt_len < 9)
      {
        out[2] = 0;

        if (pt_len < 5)
        {
          out[1] = 0;

          if (pt_len >= 1 && pt_len <= 3)
          {
            out[0] &= pt_masks[pt_len];
          }
        }
        else if (pt_len <= 7)
        {
          out[1] &= pt_masks[pt_len];
        }
      }
      else if (pt_len <= 11)
      {
        out[2] &= pt_masks[pt_len];
      }
    }

    COMPARE_S_SIMD (out[0], out[1], out[2], out[3]);
  }
}

KERNEL_FQ void FIXED_THREAD_COUNT(FIXED_LOCAL_SIZE) m33502_m04 (KERN_ATTR_ESALT (rc4_t))
{
  /**
   * base
   */

  const u64 lid = get_local_id (0);
  const u64 gid = get_global_id (0);
  const u64 lsz = get_local_size (0);

  if (gid >= GID_CNT) return;

  u32 w0[4];

  w0[0] = pws[gid].i[ 0];
  w0[1] = pws[gid].i[ 1];
  w0[2] = pws[gid].i[ 2];
  w0[3] = pws[gid].i[ 3];

  u32 w1[4];

  w1[0] = 0;
  w1[1] = 0;
  w1[2] = 0;
  w1[3] = 0;

  u32 w2[4];

  w2[0] = 0;
  w2[1] = 0;
  w2[2] = 0;
  w2[3] = 0;

  u32 w3[4];

  w3[0] = 0;
  w3[1] = 0;
  w3[2] = 0;
  w3[3] = 0;

  const u32 pw_len = pws[gid].pw_len & 63;

  /**
   * main
   */

  LOCAL_VK u32 S[64 * FIXED_LOCAL_SIZE];

  m33502m (S, w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_extra0_buf, d_extra1_buf, d_extra2_buf, d_extra3_buf, kernel_param, gid, lid, lsz);
}

KERNEL_FQ void FIXED_THREAD_COUNT(FIXED_LOCAL_SIZE) m33502_m08 (KERN_ATTR_ESALT (rc4_t))
{
  /**
   * base
   */

  const u64 lid = get_local_id (0);
  const u64 gid = get_global_id (0);
  const u64 lsz = get_local_size (0);

  if (gid >= GID_CNT) return;

  u32 w0[4];

  w0[0] = pws[gid].i[ 0];
  w0[1] = pws[gid].i[ 1];
  w0[2] = pws[gid].i[ 2];
  w0[3] = pws[gid].i[ 3];

  u32 w1[4];

  w1[0] = pws[gid].i[ 4];
  w1[1] = pws[gid].i[ 5];
  w1[2] = pws[gid].i[ 6];
  w1[3] = pws[gid].i[ 7];

  u32 w2[4];

  w2[0] = 0;
  w2[1] = 0;
  w2[2] = 0;
  w2[3] = 0;

  u32 w3[4];

  w3[0] = 0;
  w3[1] = 0;
  w3[2] = 0;
  w3[3] = 0;

  const u32 pw_len = pws[gid].pw_len & 63;

  /**
   * main
   */

  LOCAL_VK u32 S[64 * FIXED_LOCAL_SIZE];

  m33502m (S, w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_extra0_buf, d_extra1_buf, d_extra2_buf, d_extra3_buf, kernel_param, gid, lid, lsz);
}

KERNEL_FQ void FIXED_THREAD_COUNT(FIXED_LOCAL_SIZE) m33502_m16 (KERN_ATTR_ESALT (rc4_t))
{
  /**
   * base
   */

  const u64 lid = get_local_id (0);
  const u64 gid = get_global_id (0);
  const u64 lsz = get_local_size (0);

  if (gid >= GID_CNT) return;

  u32 w0[4];

  w0[0] = pws[gid].i[ 0];
  w0[1] = pws[gid].i[ 1];
  w0[2] = pws[gid].i[ 2];
  w0[3] = pws[gid].i[ 3];

  u32 w1[4];

  w1[0] = pws[gid].i[ 4];
  w1[1] = pws[gid].i[ 5];
  w1[2] = pws[gid].i[ 6];
  w1[3] = pws[gid].i[ 7];

  u32 w2[4];

  w2[0] = pws[gid].i[ 8];
  w2[1] = pws[gid].i[ 9];
  w2[2] = pws[gid].i[10];
  w2[3] = pws[gid].i[11];

  u32 w3[4];

  w3[0] = pws[gid].i[12];
  w3[1] = pws[gid].i[13];
  w3[2] = 0;
  w3[3] = 0;

  const u32 pw_len = pws[gid].pw_len & 63;

  /**
   * main
   */

  LOCAL_VK u32 S[64 * FIXED_LOCAL_SIZE];

  m33502m (S, w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_extra0_buf, d_extra1_buf, d_extra2_buf, d_extra3_buf, kernel_param, gid, lid, lsz);
}

KERNEL_FQ void FIXED_THREAD_COUNT(FIXED_LOCAL_SIZE) m33502_s04 (KERN_ATTR_ESALT (rc4_t))
{
  /**
   * base
   */

  const u64 lid = get_local_id (0);
  const u64 gid = get_global_id (0);
  const u64 lsz = get_local_size (0);

  if (gid >= GID_CNT) return;

  u32 w0[4];

  w0[0] = pws[gid].i[ 0];
  w0[1] = pws[gid].i[ 1];
  w0[2] = pws[gid].i[ 2];
  w0[3] = pws[gid].i[ 3];

  u32 w1[4];

  w1[0] = 0;
  w1[1] = 0;
  w1[2] = 0;
  w1[3] = 0;

  u32 w2[4];

  w2[0] = 0;
  w2[1] = 0;
  w2[2] = 0;
  w2[3] = 0;

  u32 w3[4];

  w3[0] = 0;
  w3[1] = 0;
  w3[2] = 0;
  w3[3] = 0;

  const u32 pw_len = pws[gid].pw_len & 63;

  /**
   * main
   */

  LOCAL_VK u32 S[64 * FIXED_LOCAL_SIZE];

  m33502s (S, w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_extra0_buf, d_extra1_buf, d_extra2_buf, d_extra3_buf, kernel_param, gid, lid, lsz);
}

KERNEL_FQ void FIXED_THREAD_COUNT(FIXED_LOCAL_SIZE) m33502_s08 (KERN_ATTR_ESALT (rc4_t))
{
  /**
   * base
   */

  const u64 lid = get_local_id (0);
  const u64 gid = get_global_id (0);
  const u64 lsz = get_local_size (0);

  if (gid >= GID_CNT) return;

  u32 w0[4];

  w0[0] = pws[gid].i[ 0];
  w0[1] = pws[gid].i[ 1];
  w0[2] = pws[gid].i[ 2];
  w0[3] = pws[gid].i[ 3];

  u32 w1[4];

  w1[0] = pws[gid].i[ 4];
  w1[1] = pws[gid].i[ 5];
  w1[2] = pws[gid].i[ 6];
  w1[3] = pws[gid].i[ 7];

  u32 w2[4];

  w2[0] = 0;
  w2[1] = 0;
  w2[2] = 0;
  w2[3] = 0;

  u32 w3[4];

  w3[0] = 0;
  w3[1] = 0;
  w3[2] = 0;
  w3[3] = 0;

  const u32 pw_len = pws[gid].pw_len & 63;

  /**
   * main
   */

  LOCAL_VK u32 S[64 * FIXED_LOCAL_SIZE];

  m33502s (S, w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_extra0_buf, d_extra1_buf, d_extra2_buf, d_extra3_buf, kernel_param, gid, lid, lsz);
}

KERNEL_FQ void FIXED_THREAD_COUNT(FIXED_LOCAL_SIZE) m33502_s16 (KERN_ATTR_ESALT (rc4_t))
{
  /**
   * base
   */

  const u64 lid = get_local_id (0);
  const u64 gid = get_global_id (0);
  const u64 lsz = get_local_size (0);

  if (gid >= GID_CNT) return;

  u32 w0[4];

  w0[0] = pws[gid].i[ 0];
  w0[1] = pws[gid].i[ 1];
  w0[2] = pws[gid].i[ 2];
  w0[3] = pws[gid].i[ 3];

  u32 w1[4];

  w1[0] = pws[gid].i[ 4];
  w1[1] = pws[gid].i[ 5];
  w1[2] = pws[gid].i[ 6];
  w1[3] = pws[gid].i[ 7];

  u32 w2[4];

  w2[0] = pws[gid].i[ 8];
  w2[1] = pws[gid].i[ 9];
  w2[2] = pws[gid].i[10];
  w2[3] = pws[gid].i[11];

  u32 w3[4];

  w3[0] = pws[gid].i[12];
  w3[1] = pws[gid].i[13];
  w3[2] = 0;
  w3[3] = 0;

  const u32 pw_len = pws[gid].pw_len & 63;

  /**
   * main
   */

  LOCAL_VK u32 S[64 * FIXED_LOCAL_SIZE];

  m33502s (S, w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_extra0_buf, d_extra1_buf, d_extra2_buf, d_extra3_buf, kernel_param, gid, lid, lsz);
}