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

#define NEW_SIMD_CODE

#define XSTR(x) #x
#define STR(x) XSTR(x)

#ifdef KERNEL_STATIC
#include STR(INCLUDE_PATH/inc_vendor.h)
#include STR(INCLUDE_PATH/inc_types.h)
#include STR(INCLUDE_PATH/inc_platform.cl)
#include STR(INCLUDE_PATH/inc_common.cl)
#include STR(INCLUDE_PATH/inc_simd.cl)
#include STR(INCLUDE_PATH/inc_hash_md5.cl)
#include STR(INCLUDE_PATH/inc_hash_sha1.cl)
#endif

#define COMPARE_S STR(INCLUDE_PATH/inc_comp_single.cl)
#define COMPARE_M STR(INCLUDE_PATH/inc_comp_multi.cl)

typedef struct mongodb_sha1_tmp
{
  u32  ipad[5];
  u32  opad[5];

  u32  dgst[5];
  u32  out[5];

} mongodb_sha1_tmp_t;

typedef struct mongodb_sha1
{
  u32 salt[16];
  u32 user[16];

  u32 user_len;

} mongodb_sha1_t;

DECLSPEC void hmac_sha1_run_V (PRIVATE_AS u32x *w0, PRIVATE_AS u32x *w1, PRIVATE_AS u32x *w2, PRIVATE_AS u32x *w3, PRIVATE_AS u32x *ipad, PRIVATE_AS u32x *opad, PRIVATE_AS u32x *digest)
{
  digest[0] = ipad[0];
  digest[1] = ipad[1];
  digest[2] = ipad[2];
  digest[3] = ipad[3];
  digest[4] = ipad[4];

  sha1_transform_vector (w0, w1, w2, w3, digest);

  w0[0] = digest[0];
  w0[1] = digest[1];
  w0[2] = digest[2];
  w0[3] = digest[3];
  w1[0] = digest[4];
  w1[1] = 0x80000000;
  w1[2] = 0;
  w1[3] = 0;
  w2[0] = 0;
  w2[1] = 0;
  w2[2] = 0;
  w2[3] = 0;
  w3[0] = 0;
  w3[1] = 0;
  w3[2] = 0;
  w3[3] = (64 + 20) * 8;

  digest[0] = opad[0];
  digest[1] = opad[1];
  digest[2] = opad[2];
  digest[3] = opad[3];
  digest[4] = opad[4];

  sha1_transform_vector (w0, w1, w2, w3, digest);
}

KERNEL_FQ void m24100_init (KERN_ATTR_TMPS_ESALT (mongodb_sha1_tmp_t, mongodb_sha1_t))
{
  /**
   * modifier
   */

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

  /**
   * bin2asc table
   */

  LOCAL_VK u32 l_bin2asc[256];

  for (u32 i = lid; i < 256; i += lsz)
  {
    const u32 i0 = (i >> 0) & 15;
    const u32 i1 = (i >> 4) & 15;

    l_bin2asc[i] = ((i0 < 10) ? '0' + i0 : 'a' - 10 + i0) << 0
                 | ((i1 < 10) ? '0' + i1 : 'a' - 10 + i1) << 8;
  }

  SYNC_THREADS ();

  if (gid >= GID_CNT) return;

  md5_ctx_t md5_ctx;

  md5_init (&md5_ctx);

  md5_update_global (&md5_ctx, esalt_bufs[DIGESTS_OFFSET_HOST].user, esalt_bufs[DIGESTS_OFFSET_HOST].user_len);
  md5_update_global (&md5_ctx, pws[gid].i, pws[gid].pw_len);

  md5_final (&md5_ctx);

  u32 a = md5_ctx.h[0];
  u32 b = md5_ctx.h[1];
  u32 c = md5_ctx.h[2];
  u32 d = md5_ctx.h[3];

  #define uint_to_hex_lower8(i) l_bin2asc[(i)]

  u32 hex[16] = { 0 };

  hex[0] = uint_to_hex_lower8 ((a >>  8) & 255) <<  0
         | uint_to_hex_lower8 ((a >>  0) & 255) << 16;
  hex[1] = uint_to_hex_lower8 ((a >> 24) & 255) <<  0
         | uint_to_hex_lower8 ((a >> 16) & 255) << 16;
  hex[2] = uint_to_hex_lower8 ((b >>  8) & 255) <<  0
         | uint_to_hex_lower8 ((b >>  0) & 255) << 16;
  hex[3] = uint_to_hex_lower8 ((b >> 24) & 255) <<  0
         | uint_to_hex_lower8 ((b >> 16) & 255) << 16;
  hex[4] = uint_to_hex_lower8 ((c >>  8) & 255) <<  0
         | uint_to_hex_lower8 ((c >>  0) & 255) << 16;
  hex[5] = uint_to_hex_lower8 ((c >> 24) & 255) <<  0
         | uint_to_hex_lower8 ((c >> 16) & 255) << 16;
  hex[6] = uint_to_hex_lower8 ((d >>  8) & 255) <<  0
         | uint_to_hex_lower8 ((d >>  0) & 255) << 16;
  hex[7] = uint_to_hex_lower8 ((d >> 24) & 255) <<  0
         | uint_to_hex_lower8 ((d >> 16) & 255) << 16;

  sha1_hmac_ctx_t sha1_hmac_ctx;

  sha1_hmac_init (&sha1_hmac_ctx, hex, 32);

  tmps[gid].ipad[0] = sha1_hmac_ctx.ipad.h[0];
  tmps[gid].ipad[1] = sha1_hmac_ctx.ipad.h[1];
  tmps[gid].ipad[2] = sha1_hmac_ctx.ipad.h[2];
  tmps[gid].ipad[3] = sha1_hmac_ctx.ipad.h[3];
  tmps[gid].ipad[4] = sha1_hmac_ctx.ipad.h[4];

  tmps[gid].opad[0] = sha1_hmac_ctx.opad.h[0];
  tmps[gid].opad[1] = sha1_hmac_ctx.opad.h[1];
  tmps[gid].opad[2] = sha1_hmac_ctx.opad.h[2];
  tmps[gid].opad[3] = sha1_hmac_ctx.opad.h[3];
  tmps[gid].opad[4] = sha1_hmac_ctx.opad.h[4];

  sha1_hmac_update_global (&sha1_hmac_ctx, esalt_bufs[DIGESTS_OFFSET_HOST].salt, 16);

  for (u32 i = 0, j = 1; i < 4; i += 5, j += 1)
  {
    sha1_hmac_ctx_t sha1_hmac_ctx2 = sha1_hmac_ctx;

    u32 w0[4];
    u32 w1[4];
    u32 w2[4];
    u32 w3[4];

    w0[0] = j;
    w0[1] = 0;
    w0[2] = 0;
    w0[3] = 0;
    w1[0] = 0;
    w1[1] = 0;
    w1[2] = 0;
    w1[3] = 0;
    w2[0] = 0;
    w2[1] = 0;
    w2[2] = 0;
    w2[3] = 0;
    w3[0] = 0;
    w3[1] = 0;
    w3[2] = 0;
    w3[3] = 0;

    sha1_hmac_update_64 (&sha1_hmac_ctx2, w0, w1, w2, w3, 4);

    sha1_hmac_final (&sha1_hmac_ctx2);

    tmps[gid].dgst[i + 0] = sha1_hmac_ctx2.opad.h[0];
    tmps[gid].dgst[i + 1] = sha1_hmac_ctx2.opad.h[1];
    tmps[gid].dgst[i + 2] = sha1_hmac_ctx2.opad.h[2];
    tmps[gid].dgst[i + 3] = sha1_hmac_ctx2.opad.h[3];
    tmps[gid].dgst[i + 4] = sha1_hmac_ctx2.opad.h[4];

    tmps[gid].out[i + 0] = tmps[gid].dgst[i + 0];
    tmps[gid].out[i + 1] = tmps[gid].dgst[i + 1];
    tmps[gid].out[i + 2] = tmps[gid].dgst[i + 2];
    tmps[gid].out[i + 3] = tmps[gid].dgst[i + 3];
    tmps[gid].out[i + 4] = tmps[gid].dgst[i + 4];
  }
}

KERNEL_FQ void m24100_loop (KERN_ATTR_TMPS_ESALT (mongodb_sha1_tmp_t, mongodb_sha1_t))
{
  const u64 gid = get_global_id (0);

  if ((gid * VECT_SIZE) >= GID_CNT) return;

  u32x ipad[5];
  u32x opad[5];

  ipad[0] = packv (tmps, ipad, gid, 0);
  ipad[1] = packv (tmps, ipad, gid, 1);
  ipad[2] = packv (tmps, ipad, gid, 2);
  ipad[3] = packv (tmps, ipad, gid, 3);
  ipad[4] = packv (tmps, ipad, gid, 4);

  opad[0] = packv (tmps, opad, gid, 0);
  opad[1] = packv (tmps, opad, gid, 1);
  opad[2] = packv (tmps, opad, gid, 2);
  opad[3] = packv (tmps, opad, gid, 3);
  opad[4] = packv (tmps, opad, gid, 4);

  for (u32 i = 0; i < 4; i += 5)
  {
    u32x dgst[5];
    u32x out[5];

    dgst[0] = packv (tmps, dgst, gid, i + 0);
    dgst[1] = packv (tmps, dgst, gid, i + 1);
    dgst[2] = packv (tmps, dgst, gid, i + 2);
    dgst[3] = packv (tmps, dgst, gid, i + 3);
    dgst[4] = packv (tmps, dgst, gid, i + 4);

    out[0] = packv (tmps, out, gid, i + 0);
    out[1] = packv (tmps, out, gid, i + 1);
    out[2] = packv (tmps, out, gid, i + 2);
    out[3] = packv (tmps, out, gid, i + 3);
    out[4] = packv (tmps, out, gid, i + 4);

    for (u32 j = 0; j < LOOP_CNT; j++)
    {
      u32x w0[4];
      u32x w1[4];
      u32x w2[4];
      u32x w3[4];

      w0[0] = dgst[0];
      w0[1] = dgst[1];
      w0[2] = dgst[2];
      w0[3] = dgst[3];
      w1[0] = dgst[4];
      w1[1] = 0x80000000;
      w1[2] = 0;
      w1[3] = 0;
      w2[0] = 0;
      w2[1] = 0;
      w2[2] = 0;
      w2[3] = 0;
      w3[0] = 0;
      w3[1] = 0;
      w3[2] = 0;
      w3[3] = (64 + 20) * 8;

      hmac_sha1_run_V (w0, w1, w2, w3, ipad, opad, dgst);

      out[0] ^= dgst[0];
      out[1] ^= dgst[1];
      out[2] ^= dgst[2];
      out[3] ^= dgst[3];
      out[4] ^= dgst[4];
    }

    unpackv (tmps, dgst, gid, i + 0, dgst[0]);
    unpackv (tmps, dgst, gid, i + 1, dgst[1]);
    unpackv (tmps, dgst, gid, i + 2, dgst[2]);
    unpackv (tmps, dgst, gid, i + 3, dgst[3]);
    unpackv (tmps, dgst, gid, i + 4, dgst[4]);

    unpackv (tmps, out, gid, i + 0, out[0]);
    unpackv (tmps, out, gid, i + 1, out[1]);
    unpackv (tmps, out, gid, i + 2, out[2]);
    unpackv (tmps, out, gid, i + 3, out[3]);
    unpackv (tmps, out, gid, i + 4, out[4]);
  }
}

KERNEL_FQ void m24100_comp (KERN_ATTR_TMPS_ESALT (mongodb_sha1_tmp_t, mongodb_sha1_t))
{
  /**
   * base
   */

  const u64 gid = get_global_id (0);

  if (gid >= GID_CNT) return;

  const u64 lid = get_local_id (0);

  u32 out[5];

  out[0] = tmps[gid].out[0];
  out[1] = tmps[gid].out[1];
  out[2] = tmps[gid].out[2];
  out[3] = tmps[gid].out[3];
  out[4] = tmps[gid].out[4];

  // HMAC-SHA1 with "Server Key" salt:

  u32 w0[4];
  u32 w1[4];
  u32 w2[4];
  u32 w3[4];

  w0[0] = out[0];
  w0[1] = out[1];
  w0[2] = out[2];
  w0[3] = out[3];
  w1[0] = out[4];
  w1[1] = 0;
  w1[2] = 0;
  w1[3] = 0;
  w2[0] = 0;
  w2[1] = 0;
  w2[2] = 0;
  w2[3] = 0;
  w3[0] = 0;
  w3[1] = 0;
  w3[2] = 0;
  w3[3] = 0;

  sha1_hmac_ctx_t sha1_hmac_ctx;

  sha1_hmac_init_64 (&sha1_hmac_ctx, w0, w1, w2, w3);

  w0[0] = 0x53657276; // Serv
  w0[1] = 0x6572204b; // er K
  w0[2] = 0x65790000; // ey
  w0[3] = 0;
  w1[0] = 0;
  w1[1] = 0;
  w1[2] = 0;
  w1[3] = 0;
  w2[0] = 0;
  w2[1] = 0;
  w2[2] = 0;
  w2[3] = 0;
  w3[0] = 0;
  w3[1] = 0;
  w3[2] = 0;
  w3[3] = 0;

  sha1_hmac_update_64 (&sha1_hmac_ctx, w0, w1, w2, w3, 10);

  sha1_hmac_final (&sha1_hmac_ctx);

  const u32 r0 = sha1_hmac_ctx.opad.h[DGST_R0];
  const u32 r1 = sha1_hmac_ctx.opad.h[DGST_R1];
  const u32 r2 = sha1_hmac_ctx.opad.h[DGST_R2];
  const u32 r3 = sha1_hmac_ctx.opad.h[DGST_R3];

  #define il_pos 0

  #ifdef KERNEL_STATIC
  #include COMPARE_M
  #endif
}