/**
 * Author......: Jens Steube <jens.steube@gmail.com>
 * License.....: MIT
 */

#define _MD5_

#define NEW_SIMD_CODE

#include "include/constants.h"
#include "include/kernel_vendor.h"

#define DGST_R0 0
#define DGST_R1 3
#define DGST_R2 2
#define DGST_R3 1

#include "include/kernel_functions.c"
#include "OpenCL/types_ocl.c"
#include "OpenCL/common.c"
#include "include/rp_kernel.h"
#include "OpenCL/rp.c"
#include "OpenCL/simd.c"

__kernel void m00000_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
{
  /**
   * modifier
   */

  const u32 lid = get_local_id (0);

  /**
   * base
   */

  const u32 gid = get_global_id (0);

  if (gid >= gid_max) 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_len = pws[gid].pw_len;

  /**
   * loop
   */

  for (u32 il_pos = 0; il_pos < rules_cnt; il_pos += VECT_SIZE)
  {
    u32x w0[4] = { 0 };
    u32x w1[4] = { 0 };
    u32x w2[4] = { 0 };
    u32x w3[4] = { 0 };

    const u32 out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);

    append_0x80_2x4 (w0, w1, out_len);

    w3[2] = out_len * 8;
    w3[3] = 0;

    u32x a = MD5M_A;
    u32x b = MD5M_B;
    u32x c = MD5M_C;
    u32x d = MD5M_D;

    MD5_STEP (MD5_Fo, a, b, c, d, w0[0], MD5C00, MD5S00);
    MD5_STEP (MD5_Fo, d, a, b, c, w0[1], MD5C01, MD5S01);
    MD5_STEP (MD5_Fo, c, d, a, b, w0[2], MD5C02, MD5S02);
    MD5_STEP (MD5_Fo, b, c, d, a, w0[3], MD5C03, MD5S03);
    MD5_STEP (MD5_Fo, a, b, c, d, w1[0], MD5C04, MD5S00);
    MD5_STEP (MD5_Fo, d, a, b, c, w1[1], MD5C05, MD5S01);
    MD5_STEP (MD5_Fo, c, d, a, b, w1[2], MD5C06, MD5S02);
    MD5_STEP (MD5_Fo, b, c, d, a, w1[3], MD5C07, MD5S03);
    MD5_STEP (MD5_Fo, a, b, c, d, w2[0], MD5C08, MD5S00);
    MD5_STEP (MD5_Fo, d, a, b, c, w2[1], MD5C09, MD5S01);
    MD5_STEP (MD5_Fo, c, d, a, b, w2[2], MD5C0a, MD5S02);
    MD5_STEP (MD5_Fo, b, c, d, a, w2[3], MD5C0b, MD5S03);
    MD5_STEP (MD5_Fo, a, b, c, d, w3[0], MD5C0c, MD5S00);
    MD5_STEP (MD5_Fo, d, a, b, c, w3[1], MD5C0d, MD5S01);
    MD5_STEP (MD5_Fo, c, d, a, b, w3[2], MD5C0e, MD5S02);
    MD5_STEP (MD5_Fo, b, c, d, a, w3[3], MD5C0f, MD5S03);

    MD5_STEP (MD5_Go, a, b, c, d, w0[1], MD5C10, MD5S10);
    MD5_STEP (MD5_Go, d, a, b, c, w1[2], MD5C11, MD5S11);
    MD5_STEP (MD5_Go, c, d, a, b, w2[3], MD5C12, MD5S12);
    MD5_STEP (MD5_Go, b, c, d, a, w0[0], MD5C13, MD5S13);
    MD5_STEP (MD5_Go, a, b, c, d, w1[1], MD5C14, MD5S10);
    MD5_STEP (MD5_Go, d, a, b, c, w2[2], MD5C15, MD5S11);
    MD5_STEP (MD5_Go, c, d, a, b, w3[3], MD5C16, MD5S12);
    MD5_STEP (MD5_Go, b, c, d, a, w1[0], MD5C17, MD5S13);
    MD5_STEP (MD5_Go, a, b, c, d, w2[1], MD5C18, MD5S10);
    MD5_STEP (MD5_Go, d, a, b, c, w3[2], MD5C19, MD5S11);
    MD5_STEP (MD5_Go, c, d, a, b, w0[3], MD5C1a, MD5S12);
    MD5_STEP (MD5_Go, b, c, d, a, w2[0], MD5C1b, MD5S13);
    MD5_STEP (MD5_Go, a, b, c, d, w3[1], MD5C1c, MD5S10);
    MD5_STEP (MD5_Go, d, a, b, c, w0[2], MD5C1d, MD5S11);
    MD5_STEP (MD5_Go, c, d, a, b, w1[3], MD5C1e, MD5S12);
    MD5_STEP (MD5_Go, b, c, d, a, w3[0], MD5C1f, MD5S13);

    MD5_STEP (MD5_H , a, b, c, d, w1[1], MD5C20, MD5S20);
    MD5_STEP (MD5_H , d, a, b, c, w2[0], MD5C21, MD5S21);
    MD5_STEP (MD5_H , c, d, a, b, w2[3], MD5C22, MD5S22);
    MD5_STEP (MD5_H , b, c, d, a, w3[2], MD5C23, MD5S23);
    MD5_STEP (MD5_H , a, b, c, d, w0[1], MD5C24, MD5S20);
    MD5_STEP (MD5_H , d, a, b, c, w1[0], MD5C25, MD5S21);
    MD5_STEP (MD5_H , c, d, a, b, w1[3], MD5C26, MD5S22);
    MD5_STEP (MD5_H , b, c, d, a, w2[2], MD5C27, MD5S23);
    MD5_STEP (MD5_H , a, b, c, d, w3[1], MD5C28, MD5S20);
    MD5_STEP (MD5_H , d, a, b, c, w0[0], MD5C29, MD5S21);
    MD5_STEP (MD5_H , c, d, a, b, w0[3], MD5C2a, MD5S22);
    MD5_STEP (MD5_H , b, c, d, a, w1[2], MD5C2b, MD5S23);
    MD5_STEP (MD5_H , a, b, c, d, w2[1], MD5C2c, MD5S20);
    MD5_STEP (MD5_H , d, a, b, c, w3[0], MD5C2d, MD5S21);
    MD5_STEP (MD5_H , c, d, a, b, w3[3], MD5C2e, MD5S22);
    MD5_STEP (MD5_H , b, c, d, a, w0[2], MD5C2f, MD5S23);

    MD5_STEP (MD5_I , a, b, c, d, w0[0], MD5C30, MD5S30);
    MD5_STEP (MD5_I , d, a, b, c, w1[3], MD5C31, MD5S31);
    MD5_STEP (MD5_I , c, d, a, b, w3[2], MD5C32, MD5S32);
    MD5_STEP (MD5_I , b, c, d, a, w1[1], MD5C33, MD5S33);
    MD5_STEP (MD5_I , a, b, c, d, w3[0], MD5C34, MD5S30);
    MD5_STEP (MD5_I , d, a, b, c, w0[3], MD5C35, MD5S31);
    MD5_STEP (MD5_I , c, d, a, b, w2[2], MD5C36, MD5S32);
    MD5_STEP (MD5_I , b, c, d, a, w0[1], MD5C37, MD5S33);
    MD5_STEP (MD5_I , a, b, c, d, w2[0], MD5C38, MD5S30);
    MD5_STEP (MD5_I , d, a, b, c, w3[3], MD5C39, MD5S31);
    MD5_STEP (MD5_I , c, d, a, b, w1[2], MD5C3a, MD5S32);
    MD5_STEP (MD5_I , b, c, d, a, w3[1], MD5C3b, MD5S33);
    MD5_STEP (MD5_I , a, b, c, d, w1[0], MD5C3c, MD5S30);
    MD5_STEP (MD5_I , d, a, b, c, w2[3], MD5C3d, MD5S31);
    MD5_STEP (MD5_I , c, d, a, b, w0[2], MD5C3e, MD5S32);
    MD5_STEP (MD5_I , b, c, d, a, w2[1], MD5C3f, MD5S33);

    COMPARE_M_SIMD (a, d, c, b);
  }
}

__kernel void m00000_m08 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
{
}

__kernel void m00000_m16 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
{
}

__kernel void m00000_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
{
  /**
   * modifier
   */

  const u32 lid = get_local_id (0);

  /**
   * base
   */

  const u32 gid = get_global_id (0);

  if (gid >= gid_max) 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_len = pws[gid].pw_len;

  /**
   * digest
   */

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

  /**
   * loop
   */

  for (u32 il_pos = 0; il_pos < rules_cnt; il_pos += VECT_SIZE)
  {
    u32x w0[4] = { 0 };
    u32x w1[4] = { 0 };
    u32x w2[4] = { 0 };
    u32x w3[4] = { 0 };

    const u32 out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);

    append_0x80_2x4 (w0, w1, out_len);

    w3[2] = out_len * 8;
    w3[3] = 0;

    u32x a = MD5M_A;
    u32x b = MD5M_B;
    u32x c = MD5M_C;
    u32x d = MD5M_D;

    MD5_STEP (MD5_Fo, a, b, c, d, w0[0], MD5C00, MD5S00);
    MD5_STEP (MD5_Fo, d, a, b, c, w0[1], MD5C01, MD5S01);
    MD5_STEP (MD5_Fo, c, d, a, b, w0[2], MD5C02, MD5S02);
    MD5_STEP (MD5_Fo, b, c, d, a, w0[3], MD5C03, MD5S03);
    MD5_STEP (MD5_Fo, a, b, c, d, w1[0], MD5C04, MD5S00);
    MD5_STEP (MD5_Fo, d, a, b, c, w1[1], MD5C05, MD5S01);
    MD5_STEP (MD5_Fo, c, d, a, b, w1[2], MD5C06, MD5S02);
    MD5_STEP (MD5_Fo, b, c, d, a, w1[3], MD5C07, MD5S03);
    MD5_STEP (MD5_Fo, a, b, c, d, w2[0], MD5C08, MD5S00);
    MD5_STEP (MD5_Fo, d, a, b, c, w2[1], MD5C09, MD5S01);
    MD5_STEP (MD5_Fo, c, d, a, b, w2[2], MD5C0a, MD5S02);
    MD5_STEP (MD5_Fo, b, c, d, a, w2[3], MD5C0b, MD5S03);
    MD5_STEP (MD5_Fo, a, b, c, d, w3[0], MD5C0c, MD5S00);
    MD5_STEP (MD5_Fo, d, a, b, c, w3[1], MD5C0d, MD5S01);
    MD5_STEP (MD5_Fo, c, d, a, b, w3[2], MD5C0e, MD5S02);
    MD5_STEP (MD5_Fo, b, c, d, a, w3[3], MD5C0f, MD5S03);

    MD5_STEP (MD5_Go, a, b, c, d, w0[1], MD5C10, MD5S10);
    MD5_STEP (MD5_Go, d, a, b, c, w1[2], MD5C11, MD5S11);
    MD5_STEP (MD5_Go, c, d, a, b, w2[3], MD5C12, MD5S12);
    MD5_STEP (MD5_Go, b, c, d, a, w0[0], MD5C13, MD5S13);
    MD5_STEP (MD5_Go, a, b, c, d, w1[1], MD5C14, MD5S10);
    MD5_STEP (MD5_Go, d, a, b, c, w2[2], MD5C15, MD5S11);
    MD5_STEP (MD5_Go, c, d, a, b, w3[3], MD5C16, MD5S12);
    MD5_STEP (MD5_Go, b, c, d, a, w1[0], MD5C17, MD5S13);
    MD5_STEP (MD5_Go, a, b, c, d, w2[1], MD5C18, MD5S10);
    MD5_STEP (MD5_Go, d, a, b, c, w3[2], MD5C19, MD5S11);
    MD5_STEP (MD5_Go, c, d, a, b, w0[3], MD5C1a, MD5S12);
    MD5_STEP (MD5_Go, b, c, d, a, w2[0], MD5C1b, MD5S13);
    MD5_STEP (MD5_Go, a, b, c, d, w3[1], MD5C1c, MD5S10);
    MD5_STEP (MD5_Go, d, a, b, c, w0[2], MD5C1d, MD5S11);
    MD5_STEP (MD5_Go, c, d, a, b, w1[3], MD5C1e, MD5S12);
    MD5_STEP (MD5_Go, b, c, d, a, w3[0], MD5C1f, MD5S13);

    MD5_STEP (MD5_H , a, b, c, d, w1[1], MD5C20, MD5S20);
    MD5_STEP (MD5_H , d, a, b, c, w2[0], MD5C21, MD5S21);
    MD5_STEP (MD5_H , c, d, a, b, w2[3], MD5C22, MD5S22);
    MD5_STEP (MD5_H , b, c, d, a, w3[2], MD5C23, MD5S23);
    MD5_STEP (MD5_H , a, b, c, d, w0[1], MD5C24, MD5S20);
    MD5_STEP (MD5_H , d, a, b, c, w1[0], MD5C25, MD5S21);
    MD5_STEP (MD5_H , c, d, a, b, w1[3], MD5C26, MD5S22);
    MD5_STEP (MD5_H , b, c, d, a, w2[2], MD5C27, MD5S23);
    MD5_STEP (MD5_H , a, b, c, d, w3[1], MD5C28, MD5S20);
    MD5_STEP (MD5_H , d, a, b, c, w0[0], MD5C29, MD5S21);
    MD5_STEP (MD5_H , c, d, a, b, w0[3], MD5C2a, MD5S22);
    MD5_STEP (MD5_H , b, c, d, a, w1[2], MD5C2b, MD5S23);
    MD5_STEP (MD5_H , a, b, c, d, w2[1], MD5C2c, MD5S20);
    MD5_STEP (MD5_H , d, a, b, c, w3[0], MD5C2d, MD5S21);
    MD5_STEP (MD5_H , c, d, a, b, w3[3], MD5C2e, MD5S22);
    MD5_STEP (MD5_H , b, c, d, a, w0[2], MD5C2f, MD5S23);

    MD5_STEP (MD5_I , a, b, c, d, w0[0], MD5C30, MD5S30);
    MD5_STEP (MD5_I , d, a, b, c, w1[3], MD5C31, MD5S31);
    MD5_STEP (MD5_I , c, d, a, b, w3[2], MD5C32, MD5S32);
    MD5_STEP (MD5_I , b, c, d, a, w1[1], MD5C33, MD5S33);
    MD5_STEP (MD5_I , a, b, c, d, w3[0], MD5C34, MD5S30);
    MD5_STEP (MD5_I , d, a, b, c, w0[3], MD5C35, MD5S31);
    MD5_STEP (MD5_I , c, d, a, b, w2[2], MD5C36, MD5S32);
    MD5_STEP (MD5_I , b, c, d, a, w0[1], MD5C37, MD5S33);
    MD5_STEP (MD5_I , a, b, c, d, w2[0], MD5C38, MD5S30);
    MD5_STEP (MD5_I , d, a, b, c, w3[3], MD5C39, MD5S31);
    MD5_STEP (MD5_I , c, d, a, b, w1[2], MD5C3a, MD5S32);
    MD5_STEP (MD5_I , b, c, d, a, w3[1], MD5C3b, MD5S33);
    MD5_STEP (MD5_I , a, b, c, d, w1[0], MD5C3c, MD5S30);

    if (MATCHES_NONE_VS (a, search[0])) continue;

    MD5_STEP (MD5_I , d, a, b, c, w2[3], MD5C3d, MD5S31);
    MD5_STEP (MD5_I , c, d, a, b, w0[2], MD5C3e, MD5S32);
    MD5_STEP (MD5_I , b, c, d, a, w2[1], MD5C3f, MD5S33);

    COMPARE_S_SIMD (a, d, c, b);
  }
}

__kernel void m00000_s08 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
{
}

__kernel void m00000_s16 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
{
}