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

#define NEW_SIMD_CODE

#include "inc_vendor.h"
#include "inc_types.h"
#include "inc_common.cl"
#include "inc_simd.cl"
#include "inc_hash_md4.cl"
#include "inc_hash_sha1.cl"

#define COMPARE_S "inc_comp_single.cl"
#define COMPARE_M "inc_comp_multi.cl"

typedef struct dcc2_tmp
{
  u32 ipad[5];
  u32 opad[5];

  u32 dgst[5];
  u32 out[4];

} dcc2_tmp_t;

DECLSPEC void hmac_sha1_run_V (u32x *w0, u32x *w1, u32x *w2, u32x *w3, u32x *ipad, u32x *opad, 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 void m02100_init (KERN_ATTR_TMPS (dcc2_tmp_t))
{
  /**
   * base
   */

  const u64 gid = get_global_id (0);

  if (gid >= gid_max) return;

  // dcc1

  md4_ctx_t md4_ctx1;

  md4_init (&md4_ctx1);

  md4_update_global_utf16le (&md4_ctx1, pws[gid].i, pws[gid].pw_len);

  md4_final (&md4_ctx1);

  md4_ctx_t md4_ctx2;

  md4_init (&md4_ctx2);

  md4_ctx2.w0[0] = md4_ctx1.h[0];
  md4_ctx2.w0[1] = md4_ctx1.h[1];
  md4_ctx2.w0[2] = md4_ctx1.h[2];
  md4_ctx2.w0[3] = md4_ctx1.h[3];

  md4_ctx2.len = 16;

  md4_update_global_utf16le (&md4_ctx2, salt_bufs[salt_pos].salt_buf, salt_bufs[salt_pos].salt_len);

  md4_final (&md4_ctx2);

  md4_ctx2.h[0] = swap32_S (md4_ctx2.h[0]);
  md4_ctx2.h[1] = swap32_S (md4_ctx2.h[1]);
  md4_ctx2.h[2] = swap32_S (md4_ctx2.h[2]);
  md4_ctx2.h[3] = swap32_S (md4_ctx2.h[3]);

  // dcc2

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

  w0[0] = md4_ctx2.h[0];
  w0[1] = md4_ctx2.h[1];
  w0[2] = md4_ctx2.h[2];
  w0[3] = md4_ctx2.h[3];
  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_ctx_t sha1_hmac_ctx;

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

  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_utf16le_swap (&sha1_hmac_ctx, salt_bufs[salt_pos].salt_buf, salt_bufs[salt_pos].salt_len);

  w0[0] = 1;
  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_ctx, w0, w1, w2, w3, 4);

  sha1_hmac_final (&sha1_hmac_ctx);

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

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

__kernel void m02100_loop (KERN_ATTR_TMPS (dcc2_tmp_t))
{
  /**
   * base
   */

  const u64 gid = get_global_id (0);

  if ((gid * VECT_SIZE) >= gid_max) 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);

  /**
   * iter1
   */

  u32x dgst[5];
  u32x out[4];

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

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

  for (u32 i = 0; i < loop_cnt; i++)
  {
    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];
  }

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

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

__kernel void m02100_comp (KERN_ATTR_TMPS (dcc2_tmp_t))
{
  /**
   * base
   */

  const u64 gid = get_global_id (0);

  if (gid >= gid_max) return;

  const u64 lid = get_local_id (0);

  const u32 r0 = tmps[gid].out[DGST_R0];
  const u32 r1 = tmps[gid].out[DGST_R1];
  const u32 r2 = tmps[gid].out[DGST_R2];
  const u32 r3 = tmps[gid].out[DGST_R3];

  #define il_pos 0

  #include COMPARE_M
}