/**
 * 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_hash_blake2s.cl)
#endif

KERNEL_FQ void m31000_m04 (KERN_ATTR_BASIC ())
{
  /**
   * base
   */

  const u64 gid = get_global_id (0);

  if (gid >= GID_CNT) return;

  u32 pw_buf0[4];
  u32 pw_buf1[4];

  pw_buf0[0] = pws[gid].i[0];
  pw_buf0[1] = pws[gid].i[1];
  pw_buf0[2] = pws[gid].i[2];
  pw_buf0[3] = pws[gid].i[3];
  pw_buf1[0] = pws[gid].i[4];
  pw_buf1[1] = pws[gid].i[5];
  pw_buf1[2] = pws[gid].i[6];
  pw_buf1[3] = pws[gid].i[7];

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

  /**
   * loop
   */

  for (u32 il_pos = 0; il_pos < IL_CNT; il_pos += VECT_SIZE)
  {
    const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos) & 63;

    const u32x out_len = pw_l_len + pw_r_len;

    /**
     * concat password candidate
     */

    u32x wordl0[4] = { 0 };
    u32x wordl1[4] = { 0 };
    u32x wordl2[4] = { 0 };
    u32x wordl3[4] = { 0 };

    wordl0[0] = pw_buf0[0];
    wordl0[1] = pw_buf0[1];
    wordl0[2] = pw_buf0[2];
    wordl0[3] = pw_buf0[3];
    wordl1[0] = pw_buf1[0];
    wordl1[1] = pw_buf1[1];
    wordl1[2] = pw_buf1[2];
    wordl1[3] = pw_buf1[3];

    u32x wordr0[4] = { 0 };
    u32x wordr1[4] = { 0 };
    u32x wordr2[4] = { 0 };
    u32x wordr3[4] = { 0 };

    wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
    wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
    wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
    wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
    wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
    wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
    wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
    wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);

    if (COMBS_MODE == COMBINATOR_MODE_BASE_LEFT)
    {
      switch_buffer_by_offset_le_VV (wordr0, wordr1, wordr2, wordr3, pw_l_len);
    }
    else
    {
      switch_buffer_by_offset_le_VV (wordl0, wordl1, wordl2, wordl3, pw_r_len);
    }

    u32x w0[4];
    u32x w1[4];

    w0[0] = wordl0[0] | wordr0[0];
    w0[1] = wordl0[1] | wordr0[1];
    w0[2] = wordl0[2] | wordr0[2];
    w0[3] = wordl0[3] | wordr0[3];
    w1[0] = wordl1[0] | wordr1[0];
    w1[1] = wordl1[1] | wordr1[1];
    w1[2] = wordl1[2] | wordr1[2];
    w1[3] = wordl1[3] | wordr1[3];

    u32x m[16];

    m[ 0] = w0[0];
    m[ 1] = w0[1];
    m[ 2] = w0[2];
    m[ 3] = w0[3];
    m[ 4] = w1[0];
    m[ 5] = w1[1];
    m[ 6] = w1[2];
    m[ 7] = w1[3];
    m[ 8] = 0;
    m[ 9] = 0;
    m[10] = 0;
    m[11] = 0;
    m[12] = 0;
    m[13] = 0;
    m[14] = 0;
    m[15] = 0;

    u32x h[8];

    h[0] = BLAKE2S_IV_00 ^ 0x01010020;
    h[1] = BLAKE2S_IV_01;
    h[2] = BLAKE2S_IV_02;
    h[3] = BLAKE2S_IV_03;
    h[4] = BLAKE2S_IV_04;
    h[5] = BLAKE2S_IV_05;
    h[6] = BLAKE2S_IV_06;
    h[7] = BLAKE2S_IV_07;

    blake2s_transform_vector (h, m, out_len, BLAKE2S_FINAL);

    const u32x r0 = h[DGST_R0];
    const u32x r1 = h[DGST_R1];
    const u32x r2 = h[DGST_R2];
    const u32x r3 = h[DGST_R3];

    COMPARE_M_SIMD (r0, r1, r2, r3);
  }
}

KERNEL_FQ void m31000_m08 (KERN_ATTR_BASIC ())
{
}

KERNEL_FQ void m31000_m16 (KERN_ATTR_BASIC ())
{
}

KERNEL_FQ void m31000_s04 (KERN_ATTR_BASIC ())
{
  /**
   * base
   */

  const u64 gid = get_global_id (0);

  if (gid >= GID_CNT) return;

  u32 pw_buf0[4];
  u32 pw_buf1[4];

  pw_buf0[0] = pws[gid].i[0];
  pw_buf0[1] = pws[gid].i[1];
  pw_buf0[2] = pws[gid].i[2];
  pw_buf0[3] = pws[gid].i[3];
  pw_buf1[0] = pws[gid].i[4];
  pw_buf1[1] = pws[gid].i[5];
  pw_buf1[2] = pws[gid].i[6];
  pw_buf1[3] = pws[gid].i[7];

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

  /**
   * 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
   */

  for (u32 il_pos = 0; il_pos < IL_CNT; il_pos += VECT_SIZE)
  {
    const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos) & 63;

    const u32x out_len = pw_l_len + pw_r_len;

    /**
     * concat password candidate
     */

    u32x wordl0[4] = { 0 };
    u32x wordl1[4] = { 0 };
    u32x wordl2[4] = { 0 };
    u32x wordl3[4] = { 0 };

    wordl0[0] = pw_buf0[0];
    wordl0[1] = pw_buf0[1];
    wordl0[2] = pw_buf0[2];
    wordl0[3] = pw_buf0[3];
    wordl1[0] = pw_buf1[0];
    wordl1[1] = pw_buf1[1];
    wordl1[2] = pw_buf1[2];
    wordl1[3] = pw_buf1[3];

    u32x wordr0[4] = { 0 };
    u32x wordr1[4] = { 0 };
    u32x wordr2[4] = { 0 };
    u32x wordr3[4] = { 0 };

    wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
    wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
    wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
    wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
    wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
    wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
    wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
    wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);

    if (COMBS_MODE == COMBINATOR_MODE_BASE_LEFT)
    {
      switch_buffer_by_offset_le_VV (wordr0, wordr1, wordr2, wordr3, pw_l_len);
    }
    else
    {
      switch_buffer_by_offset_le_VV (wordl0, wordl1, wordl2, wordl3, pw_r_len);
    }

    u32x w0[4];
    u32x w1[4];

    w0[0] = wordl0[0] | wordr0[0];
    w0[1] = wordl0[1] | wordr0[1];
    w0[2] = wordl0[2] | wordr0[2];
    w0[3] = wordl0[3] | wordr0[3];
    w1[0] = wordl1[0] | wordr1[0];
    w1[1] = wordl1[1] | wordr1[1];
    w1[2] = wordl1[2] | wordr1[2];
    w1[3] = wordl1[3] | wordr1[3];

    u32x m[16];

    m[ 0] = w0[0];
    m[ 1] = w0[1];
    m[ 2] = w0[2];
    m[ 3] = w0[3];
    m[ 4] = w1[0];
    m[ 5] = w1[1];
    m[ 6] = w1[2];
    m[ 7] = w1[3];
    m[ 8] = 0;
    m[ 9] = 0;
    m[10] = 0;
    m[11] = 0;
    m[12] = 0;
    m[13] = 0;
    m[14] = 0;
    m[15] = 0;

    u32x h[8];

    h[0] = BLAKE2S_IV_00 ^ 0x01010020;
    h[1] = BLAKE2S_IV_01;
    h[2] = BLAKE2S_IV_02;
    h[3] = BLAKE2S_IV_03;
    h[4] = BLAKE2S_IV_04;
    h[5] = BLAKE2S_IV_05;
    h[6] = BLAKE2S_IV_06;
    h[7] = BLAKE2S_IV_07;

    blake2s_transform_vector (h, m, out_len, BLAKE2S_FINAL);

    const u32x r0 = h[DGST_R0];
    const u32x r1 = h[DGST_R1];
    const u32x r2 = h[DGST_R2];
    const u32x r3 = h[DGST_R3];

    COMPARE_S_SIMD (r0, r1, r2, r3);
  }
}

KERNEL_FQ void m31000_s08 (KERN_ATTR_BASIC ())
{
}

KERNEL_FQ void m31000_s16 (KERN_ATTR_BASIC ())
{
}