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

#define NEW_SIMD_CODE

#include "inc_vendor.cl"
#include "inc_hash_constants.h"
#include "inc_hash_functions.cl"
#include "inc_types.cl"
#include "inc_common.cl"
#include "inc_rp_optimized.h"
#include "inc_rp_optimized.cl"
#include "inc_simd.cl"

__kernel void m11000_m04 (KERN_ATTR_RULES)
{
  /**
   * modifier
   */

  const u64 lid = get_local_id (0);

  /**
   * base
   */

  const u64 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;

  /**
   * salt
   */

  u32 salt_buf0[4];
  u32 salt_buf1[4];
  u32 salt_buf2[4];
  u32 salt_buf3[4];

  salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0];
  salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1];
  salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2];
  salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3];
  salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
  salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
  salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
  salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
  salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8];
  salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9];
  salt_buf2[2] = salt_bufs[salt_pos].salt_buf[10];
  salt_buf2[3] = salt_bufs[salt_pos].salt_buf[11];
  salt_buf3[0] = salt_bufs[salt_pos].salt_buf[12];
  salt_buf3[1] = salt_bufs[salt_pos].salt_buf[13];
  salt_buf3[2] = salt_bufs[salt_pos].salt_buf[14];
  salt_buf3[3] = salt_bufs[salt_pos].salt_buf[15];

  const u32 salt_len = salt_bufs[salt_pos].salt_len;

  /**
   * loop
   */

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

    const u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);

    append_0x80_2x4_VV (w0, w1, out_len);

    const u32x pw_salt_len = salt_len + out_len;

    /**
     * prepend salt
     */

    // first step fixed 56 bytes of salt
    // after 56 byte salt, we have beginning of the password

    u32x w0_t[4];
    u32x w1_t[4];
    u32x w2_t[4];
    u32x w3_t[4];

    w0_t[0] = salt_buf0[0];
    w0_t[1] = salt_buf0[1];
    w0_t[2] = salt_buf0[2];
    w0_t[3] = salt_buf0[3];
    w1_t[0] = salt_buf1[0];
    w1_t[1] = salt_buf1[1];
    w1_t[2] = salt_buf1[2];
    w1_t[3] = salt_buf1[3];
    w2_t[0] = salt_buf2[0];
    w2_t[1] = salt_buf2[1];
    w2_t[2] = salt_buf2[2];
    w2_t[3] = salt_buf2[3];
    w3_t[0] = salt_buf3[0];
    w3_t[1] = salt_buf3[1];
    w3_t[2] = w0[0];
    w3_t[3] = w0[1];

    /**
     * md5
     */

    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_t[0], MD5C00, MD5S00);
    MD5_STEP (MD5_Fo, d, a, b, c, w0_t[1], MD5C01, MD5S01);
    MD5_STEP (MD5_Fo, c, d, a, b, w0_t[2], MD5C02, MD5S02);
    MD5_STEP (MD5_Fo, b, c, d, a, w0_t[3], MD5C03, MD5S03);
    MD5_STEP (MD5_Fo, a, b, c, d, w1_t[0], MD5C04, MD5S00);
    MD5_STEP (MD5_Fo, d, a, b, c, w1_t[1], MD5C05, MD5S01);
    MD5_STEP (MD5_Fo, c, d, a, b, w1_t[2], MD5C06, MD5S02);
    MD5_STEP (MD5_Fo, b, c, d, a, w1_t[3], MD5C07, MD5S03);
    MD5_STEP (MD5_Fo, a, b, c, d, w2_t[0], MD5C08, MD5S00);
    MD5_STEP (MD5_Fo, d, a, b, c, w2_t[1], MD5C09, MD5S01);
    MD5_STEP (MD5_Fo, c, d, a, b, w2_t[2], MD5C0a, MD5S02);
    MD5_STEP (MD5_Fo, b, c, d, a, w2_t[3], MD5C0b, MD5S03);
    MD5_STEP (MD5_Fo, a, b, c, d, w3_t[0], MD5C0c, MD5S00);
    MD5_STEP (MD5_Fo, d, a, b, c, w3_t[1], MD5C0d, MD5S01);
    MD5_STEP (MD5_Fo, c, d, a, b, w3_t[2], MD5C0e, MD5S02);
    MD5_STEP (MD5_Fo, b, c, d, a, w3_t[3], MD5C0f, MD5S03);

    MD5_STEP (MD5_Go, a, b, c, d, w0_t[1], MD5C10, MD5S10);
    MD5_STEP (MD5_Go, d, a, b, c, w1_t[2], MD5C11, MD5S11);
    MD5_STEP (MD5_Go, c, d, a, b, w2_t[3], MD5C12, MD5S12);
    MD5_STEP (MD5_Go, b, c, d, a, w0_t[0], MD5C13, MD5S13);
    MD5_STEP (MD5_Go, a, b, c, d, w1_t[1], MD5C14, MD5S10);
    MD5_STEP (MD5_Go, d, a, b, c, w2_t[2], MD5C15, MD5S11);
    MD5_STEP (MD5_Go, c, d, a, b, w3_t[3], MD5C16, MD5S12);
    MD5_STEP (MD5_Go, b, c, d, a, w1_t[0], MD5C17, MD5S13);
    MD5_STEP (MD5_Go, a, b, c, d, w2_t[1], MD5C18, MD5S10);
    MD5_STEP (MD5_Go, d, a, b, c, w3_t[2], MD5C19, MD5S11);
    MD5_STEP (MD5_Go, c, d, a, b, w0_t[3], MD5C1a, MD5S12);
    MD5_STEP (MD5_Go, b, c, d, a, w2_t[0], MD5C1b, MD5S13);
    MD5_STEP (MD5_Go, a, b, c, d, w3_t[1], MD5C1c, MD5S10);
    MD5_STEP (MD5_Go, d, a, b, c, w0_t[2], MD5C1d, MD5S11);
    MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
    MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);

    u32x t;

    MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
    MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
    MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
    MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
    MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
    MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
    MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
    MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
    MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
    MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
    MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
    MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
    MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
    MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
    MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
    MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);

    MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
    MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
    MD5_STEP (MD5_I , c, d, a, b, w3_t[2], MD5C32, MD5S32);
    MD5_STEP (MD5_I , b, c, d, a, w1_t[1], MD5C33, MD5S33);
    MD5_STEP (MD5_I , a, b, c, d, w3_t[0], MD5C34, MD5S30);
    MD5_STEP (MD5_I , d, a, b, c, w0_t[3], MD5C35, MD5S31);
    MD5_STEP (MD5_I , c, d, a, b, w2_t[2], MD5C36, MD5S32);
    MD5_STEP (MD5_I , b, c, d, a, w0_t[1], MD5C37, MD5S33);
    MD5_STEP (MD5_I , a, b, c, d, w2_t[0], MD5C38, MD5S30);
    MD5_STEP (MD5_I , d, a, b, c, w3_t[3], MD5C39, MD5S31);
    MD5_STEP (MD5_I , c, d, a, b, w1_t[2], MD5C3a, MD5S32);
    MD5_STEP (MD5_I , b, c, d, a, w3_t[1], MD5C3b, MD5S33);
    MD5_STEP (MD5_I , a, b, c, d, w1_t[0], MD5C3c, MD5S30);
    MD5_STEP (MD5_I , d, a, b, c, w2_t[3], MD5C3d, MD5S31);
    MD5_STEP (MD5_I , c, d, a, b, w0_t[2], MD5C3e, MD5S32);
    MD5_STEP (MD5_I , b, c, d, a, w2_t[1], MD5C3f, MD5S33);

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

    u32x r_a = a;
    u32x r_b = b;
    u32x r_c = c;
    u32x r_d = d;

    // 2nd transform

    w0_t[0] = w0[2];
    w0_t[1] = w0[3];
    w0_t[2] = w1[0];
    w0_t[3] = w1[1];
    w1_t[0] = w1[2];
    w1_t[1] = w1[3];
    w1_t[2] = 0;
    w1_t[3] = 0;
    w2_t[0] = 0;
    w2_t[1] = 0;
    w2_t[2] = 0;
    w2_t[3] = 0;
    w3_t[0] = 0;
    w3_t[1] = 0;
    w3_t[2] = pw_salt_len * 8;
    w3_t[3] = 0;

    MD5_STEP (MD5_Fo, a, b, c, d, w0_t[0], MD5C00, MD5S00);
    MD5_STEP (MD5_Fo, d, a, b, c, w0_t[1], MD5C01, MD5S01);
    MD5_STEP (MD5_Fo, c, d, a, b, w0_t[2], MD5C02, MD5S02);
    MD5_STEP (MD5_Fo, b, c, d, a, w0_t[3], MD5C03, MD5S03);
    MD5_STEP (MD5_Fo, a, b, c, d, w1_t[0], MD5C04, MD5S00);
    MD5_STEP (MD5_Fo, d, a, b, c, w1_t[1], MD5C05, MD5S01);
    MD5_STEP (MD5_Fo, c, d, a, b, w1_t[2], MD5C06, MD5S02);
    MD5_STEP (MD5_Fo, b, c, d, a, w1_t[3], MD5C07, MD5S03);
    MD5_STEP (MD5_Fo, a, b, c, d, w2_t[0], MD5C08, MD5S00);
    MD5_STEP (MD5_Fo, d, a, b, c, w2_t[1], MD5C09, MD5S01);
    MD5_STEP (MD5_Fo, c, d, a, b, w2_t[2], MD5C0a, MD5S02);
    MD5_STEP (MD5_Fo, b, c, d, a, w2_t[3], MD5C0b, MD5S03);
    MD5_STEP (MD5_Fo, a, b, c, d, w3_t[0], MD5C0c, MD5S00);
    MD5_STEP (MD5_Fo, d, a, b, c, w3_t[1], MD5C0d, MD5S01);
    MD5_STEP (MD5_Fo, c, d, a, b, w3_t[2], MD5C0e, MD5S02);
    MD5_STEP (MD5_Fo, b, c, d, a, w3_t[3], MD5C0f, MD5S03);

    MD5_STEP (MD5_Go, a, b, c, d, w0_t[1], MD5C10, MD5S10);
    MD5_STEP (MD5_Go, d, a, b, c, w1_t[2], MD5C11, MD5S11);
    MD5_STEP (MD5_Go, c, d, a, b, w2_t[3], MD5C12, MD5S12);
    MD5_STEP (MD5_Go, b, c, d, a, w0_t[0], MD5C13, MD5S13);
    MD5_STEP (MD5_Go, a, b, c, d, w1_t[1], MD5C14, MD5S10);
    MD5_STEP (MD5_Go, d, a, b, c, w2_t[2], MD5C15, MD5S11);
    MD5_STEP (MD5_Go, c, d, a, b, w3_t[3], MD5C16, MD5S12);
    MD5_STEP (MD5_Go, b, c, d, a, w1_t[0], MD5C17, MD5S13);
    MD5_STEP (MD5_Go, a, b, c, d, w2_t[1], MD5C18, MD5S10);
    MD5_STEP (MD5_Go, d, a, b, c, w3_t[2], MD5C19, MD5S11);
    MD5_STEP (MD5_Go, c, d, a, b, w0_t[3], MD5C1a, MD5S12);
    MD5_STEP (MD5_Go, b, c, d, a, w2_t[0], MD5C1b, MD5S13);
    MD5_STEP (MD5_Go, a, b, c, d, w3_t[1], MD5C1c, MD5S10);
    MD5_STEP (MD5_Go, d, a, b, c, w0_t[2], MD5C1d, MD5S11);
    MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
    MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);

    MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
    MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
    MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
    MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
    MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
    MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
    MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
    MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
    MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
    MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
    MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
    MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
    MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
    MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
    MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
    MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);

    MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
    MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
    MD5_STEP (MD5_I , c, d, a, b, w3_t[2], MD5C32, MD5S32);
    MD5_STEP (MD5_I , b, c, d, a, w1_t[1], MD5C33, MD5S33);
    MD5_STEP (MD5_I , a, b, c, d, w3_t[0], MD5C34, MD5S30);
    MD5_STEP (MD5_I , d, a, b, c, w0_t[3], MD5C35, MD5S31);
    MD5_STEP (MD5_I , c, d, a, b, w2_t[2], MD5C36, MD5S32);
    MD5_STEP (MD5_I , b, c, d, a, w0_t[1], MD5C37, MD5S33);
    MD5_STEP (MD5_I , a, b, c, d, w2_t[0], MD5C38, MD5S30);
    MD5_STEP (MD5_I , d, a, b, c, w3_t[3], MD5C39, MD5S31);
    MD5_STEP (MD5_I , c, d, a, b, w1_t[2], MD5C3a, MD5S32);
    MD5_STEP (MD5_I , b, c, d, a, w3_t[1], MD5C3b, MD5S33);
    MD5_STEP (MD5_I , a, b, c, d, w1_t[0], MD5C3c, MD5S30);
    MD5_STEP (MD5_I , d, a, b, c, w2_t[3], MD5C3d, MD5S31);
    MD5_STEP (MD5_I , c, d, a, b, w0_t[2], MD5C3e, MD5S32);
    MD5_STEP (MD5_I , b, c, d, a, w2_t[1], MD5C3f, MD5S33);

    a += r_a;
    b += r_b;
    c += r_c;
    d += r_d;

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

__kernel void m11000_m08 (KERN_ATTR_RULES)
{
}

__kernel void m11000_m16 (KERN_ATTR_RULES)
{
}

__kernel void m11000_s04 (KERN_ATTR_RULES)
{
  /**
   * modifier
   */

  const u64 lid = get_local_id (0);

  /**
   * base
   */

  const u64 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;

  /**
   * salt
   */

  u32 salt_buf0[4];
  u32 salt_buf1[4];
  u32 salt_buf2[4];
  u32 salt_buf3[4];

  salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0];
  salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1];
  salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2];
  salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3];
  salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
  salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
  salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
  salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
  salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8];
  salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9];
  salt_buf2[2] = salt_bufs[salt_pos].salt_buf[10];
  salt_buf2[3] = salt_bufs[salt_pos].salt_buf[11];
  salt_buf3[0] = salt_bufs[salt_pos].salt_buf[12];
  salt_buf3[1] = salt_bufs[salt_pos].salt_buf[13];
  salt_buf3[2] = salt_bufs[salt_pos].salt_buf[14];
  salt_buf3[3] = salt_bufs[salt_pos].salt_buf[15];

  const u32 salt_len = salt_bufs[salt_pos].salt_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 < il_cnt; il_pos += VECT_SIZE)
  {
    u32x w0[4] = { 0 };
    u32x w1[4] = { 0 };
    u32x w2[4] = { 0 };
    u32x w3[4] = { 0 };

    const u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);

    append_0x80_2x4_VV (w0, w1, out_len);

    const u32x pw_salt_len = salt_len + out_len;

    /**
     * prepend salt
     */

    // first step fixed 56 bytes of salt
    // after 56 byte salt, we have beginning of the password

    u32x w0_t[4];
    u32x w1_t[4];
    u32x w2_t[4];
    u32x w3_t[4];

    w0_t[0] = salt_buf0[0];
    w0_t[1] = salt_buf0[1];
    w0_t[2] = salt_buf0[2];
    w0_t[3] = salt_buf0[3];
    w1_t[0] = salt_buf1[0];
    w1_t[1] = salt_buf1[1];
    w1_t[2] = salt_buf1[2];
    w1_t[3] = salt_buf1[3];
    w2_t[0] = salt_buf2[0];
    w2_t[1] = salt_buf2[1];
    w2_t[2] = salt_buf2[2];
    w2_t[3] = salt_buf2[3];
    w3_t[0] = salt_buf3[0];
    w3_t[1] = salt_buf3[1];
    w3_t[2] = w0[0];
    w3_t[3] = w0[1];

    /**
     * md5
     */

    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_t[0], MD5C00, MD5S00);
    MD5_STEP (MD5_Fo, d, a, b, c, w0_t[1], MD5C01, MD5S01);
    MD5_STEP (MD5_Fo, c, d, a, b, w0_t[2], MD5C02, MD5S02);
    MD5_STEP (MD5_Fo, b, c, d, a, w0_t[3], MD5C03, MD5S03);
    MD5_STEP (MD5_Fo, a, b, c, d, w1_t[0], MD5C04, MD5S00);
    MD5_STEP (MD5_Fo, d, a, b, c, w1_t[1], MD5C05, MD5S01);
    MD5_STEP (MD5_Fo, c, d, a, b, w1_t[2], MD5C06, MD5S02);
    MD5_STEP (MD5_Fo, b, c, d, a, w1_t[3], MD5C07, MD5S03);
    MD5_STEP (MD5_Fo, a, b, c, d, w2_t[0], MD5C08, MD5S00);
    MD5_STEP (MD5_Fo, d, a, b, c, w2_t[1], MD5C09, MD5S01);
    MD5_STEP (MD5_Fo, c, d, a, b, w2_t[2], MD5C0a, MD5S02);
    MD5_STEP (MD5_Fo, b, c, d, a, w2_t[3], MD5C0b, MD5S03);
    MD5_STEP (MD5_Fo, a, b, c, d, w3_t[0], MD5C0c, MD5S00);
    MD5_STEP (MD5_Fo, d, a, b, c, w3_t[1], MD5C0d, MD5S01);
    MD5_STEP (MD5_Fo, c, d, a, b, w3_t[2], MD5C0e, MD5S02);
    MD5_STEP (MD5_Fo, b, c, d, a, w3_t[3], MD5C0f, MD5S03);

    MD5_STEP (MD5_Go, a, b, c, d, w0_t[1], MD5C10, MD5S10);
    MD5_STEP (MD5_Go, d, a, b, c, w1_t[2], MD5C11, MD5S11);
    MD5_STEP (MD5_Go, c, d, a, b, w2_t[3], MD5C12, MD5S12);
    MD5_STEP (MD5_Go, b, c, d, a, w0_t[0], MD5C13, MD5S13);
    MD5_STEP (MD5_Go, a, b, c, d, w1_t[1], MD5C14, MD5S10);
    MD5_STEP (MD5_Go, d, a, b, c, w2_t[2], MD5C15, MD5S11);
    MD5_STEP (MD5_Go, c, d, a, b, w3_t[3], MD5C16, MD5S12);
    MD5_STEP (MD5_Go, b, c, d, a, w1_t[0], MD5C17, MD5S13);
    MD5_STEP (MD5_Go, a, b, c, d, w2_t[1], MD5C18, MD5S10);
    MD5_STEP (MD5_Go, d, a, b, c, w3_t[2], MD5C19, MD5S11);
    MD5_STEP (MD5_Go, c, d, a, b, w0_t[3], MD5C1a, MD5S12);
    MD5_STEP (MD5_Go, b, c, d, a, w2_t[0], MD5C1b, MD5S13);
    MD5_STEP (MD5_Go, a, b, c, d, w3_t[1], MD5C1c, MD5S10);
    MD5_STEP (MD5_Go, d, a, b, c, w0_t[2], MD5C1d, MD5S11);
    MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
    MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);

    u32x t;

    MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
    MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
    MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
    MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
    MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
    MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
    MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
    MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
    MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
    MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
    MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
    MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
    MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
    MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
    MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
    MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);

    MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
    MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
    MD5_STEP (MD5_I , c, d, a, b, w3_t[2], MD5C32, MD5S32);
    MD5_STEP (MD5_I , b, c, d, a, w1_t[1], MD5C33, MD5S33);
    MD5_STEP (MD5_I , a, b, c, d, w3_t[0], MD5C34, MD5S30);
    MD5_STEP (MD5_I , d, a, b, c, w0_t[3], MD5C35, MD5S31);
    MD5_STEP (MD5_I , c, d, a, b, w2_t[2], MD5C36, MD5S32);
    MD5_STEP (MD5_I , b, c, d, a, w0_t[1], MD5C37, MD5S33);
    MD5_STEP (MD5_I , a, b, c, d, w2_t[0], MD5C38, MD5S30);
    MD5_STEP (MD5_I , d, a, b, c, w3_t[3], MD5C39, MD5S31);
    MD5_STEP (MD5_I , c, d, a, b, w1_t[2], MD5C3a, MD5S32);
    MD5_STEP (MD5_I , b, c, d, a, w3_t[1], MD5C3b, MD5S33);
    MD5_STEP (MD5_I , a, b, c, d, w1_t[0], MD5C3c, MD5S30);
    MD5_STEP (MD5_I , d, a, b, c, w2_t[3], MD5C3d, MD5S31);
    MD5_STEP (MD5_I , c, d, a, b, w0_t[2], MD5C3e, MD5S32);
    MD5_STEP (MD5_I , b, c, d, a, w2_t[1], MD5C3f, MD5S33);

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

    u32x r_a = a;
    u32x r_b = b;
    u32x r_c = c;
    u32x r_d = d;

    // 2nd transform

    w0_t[0] = w0[2];
    w0_t[1] = w0[3];
    w0_t[2] = w1[0];
    w0_t[3] = w1[1];
    w1_t[0] = w1[2];
    w1_t[1] = w1[3];
    w1_t[2] = 0;
    w1_t[3] = 0;
    w2_t[0] = 0;
    w2_t[1] = 0;
    w2_t[2] = 0;
    w2_t[3] = 0;
    w3_t[0] = 0;
    w3_t[1] = 0;
    w3_t[2] = pw_salt_len * 8;
    w3_t[3] = 0;

    MD5_STEP (MD5_Fo, a, b, c, d, w0_t[0], MD5C00, MD5S00);
    MD5_STEP (MD5_Fo, d, a, b, c, w0_t[1], MD5C01, MD5S01);
    MD5_STEP (MD5_Fo, c, d, a, b, w0_t[2], MD5C02, MD5S02);
    MD5_STEP (MD5_Fo, b, c, d, a, w0_t[3], MD5C03, MD5S03);
    MD5_STEP (MD5_Fo, a, b, c, d, w1_t[0], MD5C04, MD5S00);
    MD5_STEP (MD5_Fo, d, a, b, c, w1_t[1], MD5C05, MD5S01);
    MD5_STEP (MD5_Fo, c, d, a, b, w1_t[2], MD5C06, MD5S02);
    MD5_STEP (MD5_Fo, b, c, d, a, w1_t[3], MD5C07, MD5S03);
    MD5_STEP (MD5_Fo, a, b, c, d, w2_t[0], MD5C08, MD5S00);
    MD5_STEP (MD5_Fo, d, a, b, c, w2_t[1], MD5C09, MD5S01);
    MD5_STEP (MD5_Fo, c, d, a, b, w2_t[2], MD5C0a, MD5S02);
    MD5_STEP (MD5_Fo, b, c, d, a, w2_t[3], MD5C0b, MD5S03);
    MD5_STEP (MD5_Fo, a, b, c, d, w3_t[0], MD5C0c, MD5S00);
    MD5_STEP (MD5_Fo, d, a, b, c, w3_t[1], MD5C0d, MD5S01);
    MD5_STEP (MD5_Fo, c, d, a, b, w3_t[2], MD5C0e, MD5S02);
    MD5_STEP (MD5_Fo, b, c, d, a, w3_t[3], MD5C0f, MD5S03);

    MD5_STEP (MD5_Go, a, b, c, d, w0_t[1], MD5C10, MD5S10);
    MD5_STEP (MD5_Go, d, a, b, c, w1_t[2], MD5C11, MD5S11);
    MD5_STEP (MD5_Go, c, d, a, b, w2_t[3], MD5C12, MD5S12);
    MD5_STEP (MD5_Go, b, c, d, a, w0_t[0], MD5C13, MD5S13);
    MD5_STEP (MD5_Go, a, b, c, d, w1_t[1], MD5C14, MD5S10);
    MD5_STEP (MD5_Go, d, a, b, c, w2_t[2], MD5C15, MD5S11);
    MD5_STEP (MD5_Go, c, d, a, b, w3_t[3], MD5C16, MD5S12);
    MD5_STEP (MD5_Go, b, c, d, a, w1_t[0], MD5C17, MD5S13);
    MD5_STEP (MD5_Go, a, b, c, d, w2_t[1], MD5C18, MD5S10);
    MD5_STEP (MD5_Go, d, a, b, c, w3_t[2], MD5C19, MD5S11);
    MD5_STEP (MD5_Go, c, d, a, b, w0_t[3], MD5C1a, MD5S12);
    MD5_STEP (MD5_Go, b, c, d, a, w2_t[0], MD5C1b, MD5S13);
    MD5_STEP (MD5_Go, a, b, c, d, w3_t[1], MD5C1c, MD5S10);
    MD5_STEP (MD5_Go, d, a, b, c, w0_t[2], MD5C1d, MD5S11);
    MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
    MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);

    MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
    MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
    MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
    MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
    MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
    MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
    MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
    MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
    MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
    MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
    MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
    MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
    MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
    MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
    MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
    MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);

    MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
    MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
    MD5_STEP (MD5_I , c, d, a, b, w3_t[2], MD5C32, MD5S32);
    MD5_STEP (MD5_I , b, c, d, a, w1_t[1], MD5C33, MD5S33);
    MD5_STEP (MD5_I , a, b, c, d, w3_t[0], MD5C34, MD5S30);
    MD5_STEP (MD5_I , d, a, b, c, w0_t[3], MD5C35, MD5S31);
    MD5_STEP (MD5_I , c, d, a, b, w2_t[2], MD5C36, MD5S32);
    MD5_STEP (MD5_I , b, c, d, a, w0_t[1], MD5C37, MD5S33);
    MD5_STEP (MD5_I , a, b, c, d, w2_t[0], MD5C38, MD5S30);
    MD5_STEP (MD5_I , d, a, b, c, w3_t[3], MD5C39, MD5S31);
    MD5_STEP (MD5_I , c, d, a, b, w1_t[2], MD5C3a, MD5S32);
    MD5_STEP (MD5_I , b, c, d, a, w3_t[1], MD5C3b, MD5S33);
    MD5_STEP (MD5_I , a, b, c, d, w1_t[0], MD5C3c, MD5S30);
    MD5_STEP (MD5_I , d, a, b, c, w2_t[3], MD5C3d, MD5S31);
    MD5_STEP (MD5_I , c, d, a, b, w0_t[2], MD5C3e, MD5S32);
    MD5_STEP (MD5_I , b, c, d, a, w2_t[1], MD5C3f, MD5S33);

    a += r_a;
    b += r_b;
    c += r_c;
    d += r_d;

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

__kernel void m11000_s08 (KERN_ATTR_RULES)
{
}

__kernel void m11000_s16 (KERN_ATTR_RULES)
{
}