X-Git-Url: https://www.flypig.org.uk/git/?a=blobdiff_plain;f=OpenCL%2Fm03800_a3.cl;h=5c3a8922632895006434e39b421028c461e1575e;hb=161a6eb4bc643d8e636e96eda613f5137d30da59;hp=502295c967281758b85e4e1ee71a5b6ae03b5b7d;hpb=1d3795a3ab79613b42a4acb4066782aa2c42c58c;p=hashcat.git

diff --git a/OpenCL/m03800_a3.cl b/OpenCL/m03800_a3.cl
index 502295c..5c3a892 100644
--- a/OpenCL/m03800_a3.cl
+++ b/OpenCL/m03800_a3.cl
@@ -7,20 +7,14 @@
 
 #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 "OpenCL/simd.c"
-
-static void m03800m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset)
+#include "inc_vendor.cl"
+#include "inc_hash_constants.h"
+#include "inc_hash_functions.cl"
+#include "inc_types.cl"
+#include "inc_common.cl"
+#include "inc_simd.cl"
+
+void m03800m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __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_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_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 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -29,45 +23,80 @@ static void m03800m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
   const u32 gid = get_global_id (0);
   const u32 lid = get_local_id (0);
 
-  /**
-   * digest
-   */
-
   /**
    * 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];
-
-  u32 salt_buf1[4];
-
   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];
+
+  u32 salt_buf0_t[4];
+  u32 salt_buf1_t[4];
+  u32 salt_buf2_t[4];
+  u32 salt_buf3_t[4];
+
+  salt_buf0_t[0] = salt_bufs[salt_pos].salt_buf[ 0];
+  salt_buf0_t[1] = salt_bufs[salt_pos].salt_buf[ 1];
+  salt_buf0_t[2] = salt_bufs[salt_pos].salt_buf[ 2];
+  salt_buf0_t[3] = salt_bufs[salt_pos].salt_buf[ 3];
+  salt_buf1_t[0] = salt_bufs[salt_pos].salt_buf[ 4];
+  salt_buf1_t[1] = salt_bufs[salt_pos].salt_buf[ 5];
+  salt_buf1_t[2] = salt_bufs[salt_pos].salt_buf[ 6];
+  salt_buf1_t[3] = salt_bufs[salt_pos].salt_buf[ 7];
+  salt_buf2_t[0] = salt_bufs[salt_pos].salt_buf[ 8];
+  salt_buf2_t[1] = salt_bufs[salt_pos].salt_buf[ 9];
+  salt_buf2_t[2] = salt_bufs[salt_pos].salt_buf[10];
+  salt_buf2_t[3] = salt_bufs[salt_pos].salt_buf[11];
+  salt_buf3_t[0] = salt_bufs[salt_pos].salt_buf[12];
+  salt_buf3_t[1] = salt_bufs[salt_pos].salt_buf[13];
+  salt_buf3_t[2] = salt_bufs[salt_pos].salt_buf[14];
+  salt_buf3_t[3] = salt_bufs[salt_pos].salt_buf[15];
 
-  u32 salt_buf2[4];
+  const u32 salt_len = salt_bufs[salt_pos].salt_len;
 
-  salt_buf2[0] = 0;
-  salt_buf2[1] = 0;
-  salt_buf2[2] = 0;
-  salt_buf2[3] = 0;
+  const u32 pw_salt_len = pw_len + salt_len;
 
-  u32 salt_buf3[4];
+  const u32 salt_pw_salt_len = salt_len + pw_len + salt_len;
 
-  salt_buf3[0] = 0;
-  salt_buf3[1] = 0;
-  salt_buf3[2] = 0;
-  salt_buf3[3] = 0;
+  switch_buffer_by_offset_le_S (salt_buf0_t, salt_buf1_t, salt_buf2_t, salt_buf3_t, pw_salt_len);
 
-  const u32 salt_len = salt_bufs[salt_pos].salt_len;
+  salt_buf0[0] |= salt_buf0_t[0];
+  salt_buf0[1] |= salt_buf0_t[1];
+  salt_buf0[2] |= salt_buf0_t[2];
+  salt_buf0[3] |= salt_buf0_t[3];
+  salt_buf1[0] |= salt_buf1_t[0];
+  salt_buf1[1] |= salt_buf1_t[1];
+  salt_buf1[2] |= salt_buf1_t[2];
+  salt_buf1[3] |= salt_buf1_t[3];
+  salt_buf2[0] |= salt_buf2_t[0];
+  salt_buf2[1] |= salt_buf2_t[1];
+  salt_buf2[2] |= salt_buf2_t[2];
+  salt_buf2[3] |= salt_buf2_t[3];
+  salt_buf3[0] |= salt_buf3_t[0];
+  salt_buf3[1] |= salt_buf3_t[1];
+  salt_buf3[2] |= salt_buf3_t[2];
+  salt_buf3[3] |= salt_buf3_t[3];
 
-  const u32 pw_salt_len = salt_len + pw_len + salt_len;
+  append_0x80_4x4_S (salt_buf0, salt_buf1, salt_buf2, salt_buf3, salt_pw_salt_len);
 
   /**
    * loop
@@ -75,118 +104,56 @@ static void m03800m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
 
   u32 w0l = w0[0];
 
-  for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos += VECT_SIZE)
+  for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
   {
-    const u32x w0r = w0r_create_bft (bfs_buf, il_pos);
+    const u32x w0r = ix_create_bft (bfs_buf, il_pos);
 
     const u32x w0lr = w0l | w0r;
 
-    u32x w0_t[4];
-
-    w0_t[0] = w0lr;
-    w0_t[1] = w0[1];
-    w0_t[2] = w0[2];
-    w0_t[3] = w0[3];
-
-    u32x w1_t[4];
-
-    w1_t[0] = w1[0];
-    w1_t[1] = w1[1];
-    w1_t[2] = w1[2];
-    w1_t[3] = w1[3];
-
-    u32x w2_t[4];
-
-    w2_t[0] = w2[0];
-    w2_t[1] = w2[1];
-    w2_t[2] = w2[2];
-    w2_t[3] = w2[3];
-
-    u32x w3_t[4];
-
-    w3_t[0] = w3[0];
-    w3_t[1] = w3[1];
-    w3_t[2] = w3[2];
-    w3_t[3] = w3[3];
-
+    u32x t0[4];
+    u32x t1[4];
+    u32x t2[4];
+    u32x t3[4];
+
+    t0[0] = w0lr;
+    t0[1] = w0[1];
+    t0[2] = w0[2];
+    t0[3] = w0[3];
+    t1[0] = w1[0];
+    t1[1] = w1[1];
+    t1[2] = w1[2];
+    t1[3] = w1[3];
+    t2[0] = w2[0];
+    t2[1] = w2[1];
+    t2[2] = w2[2];
+    t2[3] = w2[3];
+    t3[0] = w3[0];
+    t3[1] = w3[1];
+    t3[2] = w3[2];
+    t3[3] = w3[3];
 
     /**
-     * prepend salt
+     * put the password after the first salt but before the second salt
      */
 
-    switch_buffer_by_offset_le (w0_t, w1_t, w2_t, w3_t, salt_len);
-
-    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] |= salt_buf3[2];
-    w3_t[3] |= salt_buf3[3];
-
-    /**
-     * append salt
-     */
-
-    u32x s0[4];
-
-    s0[0] = salt_buf0[0];
-    s0[1] = salt_buf0[1];
-    s0[2] = salt_buf0[2];
-    s0[3] = salt_buf0[3];
-
-    u32x s1[4];
-
-    s1[0] = salt_buf1[0];
-    s1[1] = salt_buf1[1];
-    s1[2] = salt_buf1[2];
-    s1[3] = salt_buf1[3];
-
-    u32x s2[4];
-
-    s2[0] = 0;
-    s2[1] = 0;
-    s2[2] = 0;
-    s2[3] = 0;
-
-    u32x s3[4];
-
-    s3[0] = 0;
-    s3[1] = 0;
-    s3[2] = 0;
-    s3[3] = 0;
-
-    switch_buffer_by_offset_le (s0, s1, s2, s3, salt_len + pw_len);
-
-    w0_t[0] |= s0[0];
-    w0_t[1] |= s0[1];
-    w0_t[2] |= s0[2];
-    w0_t[3] |= s0[3];
-    w1_t[0] |= s1[0];
-    w1_t[1] |= s1[1];
-    w1_t[2] |= s1[2];
-    w1_t[3] |= s1[3];
-    w2_t[0] |= s2[0];
-    w2_t[1] |= s2[1];
-    w2_t[2] |= s2[2];
-    w2_t[3] |= s2[3];
-    w3_t[0] |= s3[0];
-    w3_t[1] |= s3[1];
-    w3_t[2] |= s3[2];
-    w3_t[3] |= s3[3];
-
-    append_0x80_4x4 (w0_t, w1_t, w2_t, w3_t, pw_salt_len);
-
-    w3_t[2] = pw_salt_len * 8;
+    switch_buffer_by_offset_le (t0, t1, t2, t3, salt_len);
+
+    t0[0] |= salt_buf0[0];
+    t0[1] |= salt_buf0[1];
+    t0[2] |= salt_buf0[2];
+    t0[3] |= salt_buf0[3];
+    t1[0] |= salt_buf1[0];
+    t1[1] |= salt_buf1[1];
+    t1[2] |= salt_buf1[2];
+    t1[3] |= salt_buf1[3];
+    t2[0] |= salt_buf2[0];
+    t2[1] |= salt_buf2[1];
+    t2[2] |= salt_buf2[2];
+    t2[3] |= salt_buf2[3];
+    t3[0] |= salt_buf3[0];
+    t3[1] |= salt_buf3[1];
+    t3[2]  = salt_pw_salt_len * 8;
+    t3[3]  = 0;
 
     /**
      * md5
@@ -197,80 +164,79 @@ static void m03800m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
     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);
-
-    MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
-    MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
-    MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
-    MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
-    MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
-    MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
-    MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
-    MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
-    MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
-    MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
-    MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
-    MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
-    MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
-    MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
-    MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
-    MD5_STEP (MD5_H , 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);
+    MD5_STEP (MD5_Fo, a, b, c, d, t0[0], MD5C00, MD5S00);
+    MD5_STEP (MD5_Fo, d, a, b, c, t0[1], MD5C01, MD5S01);
+    MD5_STEP (MD5_Fo, c, d, a, b, t0[2], MD5C02, MD5S02);
+    MD5_STEP (MD5_Fo, b, c, d, a, t0[3], MD5C03, MD5S03);
+    MD5_STEP (MD5_Fo, a, b, c, d, t1[0], MD5C04, MD5S00);
+    MD5_STEP (MD5_Fo, d, a, b, c, t1[1], MD5C05, MD5S01);
+    MD5_STEP (MD5_Fo, c, d, a, b, t1[2], MD5C06, MD5S02);
+    MD5_STEP (MD5_Fo, b, c, d, a, t1[3], MD5C07, MD5S03);
+    MD5_STEP (MD5_Fo, a, b, c, d, t2[0], MD5C08, MD5S00);
+    MD5_STEP (MD5_Fo, d, a, b, c, t2[1], MD5C09, MD5S01);
+    MD5_STEP (MD5_Fo, c, d, a, b, t2[2], MD5C0a, MD5S02);
+    MD5_STEP (MD5_Fo, b, c, d, a, t2[3], MD5C0b, MD5S03);
+    MD5_STEP (MD5_Fo, a, b, c, d, t3[0], MD5C0c, MD5S00);
+    MD5_STEP (MD5_Fo, d, a, b, c, t3[1], MD5C0d, MD5S01);
+    MD5_STEP (MD5_Fo, c, d, a, b, t3[2], MD5C0e, MD5S02);
+    MD5_STEP (MD5_Fo, b, c, d, a, t3[3], MD5C0f, MD5S03);
+
+    MD5_STEP (MD5_Go, a, b, c, d, t0[1], MD5C10, MD5S10);
+    MD5_STEP (MD5_Go, d, a, b, c, t1[2], MD5C11, MD5S11);
+    MD5_STEP (MD5_Go, c, d, a, b, t2[3], MD5C12, MD5S12);
+    MD5_STEP (MD5_Go, b, c, d, a, t0[0], MD5C13, MD5S13);
+    MD5_STEP (MD5_Go, a, b, c, d, t1[1], MD5C14, MD5S10);
+    MD5_STEP (MD5_Go, d, a, b, c, t2[2], MD5C15, MD5S11);
+    MD5_STEP (MD5_Go, c, d, a, b, t3[3], MD5C16, MD5S12);
+    MD5_STEP (MD5_Go, b, c, d, a, t1[0], MD5C17, MD5S13);
+    MD5_STEP (MD5_Go, a, b, c, d, t2[1], MD5C18, MD5S10);
+    MD5_STEP (MD5_Go, d, a, b, c, t3[2], MD5C19, MD5S11);
+    MD5_STEP (MD5_Go, c, d, a, b, t0[3], MD5C1a, MD5S12);
+    MD5_STEP (MD5_Go, b, c, d, a, t2[0], MD5C1b, MD5S13);
+    MD5_STEP (MD5_Go, a, b, c, d, t3[1], MD5C1c, MD5S10);
+    MD5_STEP (MD5_Go, d, a, b, c, t0[2], MD5C1d, MD5S11);
+    MD5_STEP (MD5_Go, c, d, a, b, t1[3], MD5C1e, MD5S12);
+    MD5_STEP (MD5_Go, b, c, d, a, t3[0], MD5C1f, MD5S13);
+
+    MD5_STEP (MD5_H , a, b, c, d, t1[1], MD5C20, MD5S20);
+    MD5_STEP (MD5_H , d, a, b, c, t2[0], MD5C21, MD5S21);
+    MD5_STEP (MD5_H , c, d, a, b, t2[3], MD5C22, MD5S22);
+    MD5_STEP (MD5_H , b, c, d, a, t3[2], MD5C23, MD5S23);
+    MD5_STEP (MD5_H , a, b, c, d, t0[1], MD5C24, MD5S20);
+    MD5_STEP (MD5_H , d, a, b, c, t1[0], MD5C25, MD5S21);
+    MD5_STEP (MD5_H , c, d, a, b, t1[3], MD5C26, MD5S22);
+    MD5_STEP (MD5_H , b, c, d, a, t2[2], MD5C27, MD5S23);
+    MD5_STEP (MD5_H , a, b, c, d, t3[1], MD5C28, MD5S20);
+    MD5_STEP (MD5_H , d, a, b, c, t0[0], MD5C29, MD5S21);
+    MD5_STEP (MD5_H , c, d, a, b, t0[3], MD5C2a, MD5S22);
+    MD5_STEP (MD5_H , b, c, d, a, t1[2], MD5C2b, MD5S23);
+    MD5_STEP (MD5_H , a, b, c, d, t2[1], MD5C2c, MD5S20);
+    MD5_STEP (MD5_H , d, a, b, c, t3[0], MD5C2d, MD5S21);
+    MD5_STEP (MD5_H , c, d, a, b, t3[3], MD5C2e, MD5S22);
+    MD5_STEP (MD5_H , b, c, d, a, t0[2], MD5C2f, MD5S23);
+
+    MD5_STEP (MD5_I , a, b, c, d, t0[0], MD5C30, MD5S30);
+    MD5_STEP (MD5_I , d, a, b, c, t1[3], MD5C31, MD5S31);
+    MD5_STEP (MD5_I , c, d, a, b, t3[2], MD5C32, MD5S32);
+    MD5_STEP (MD5_I , b, c, d, a, t1[1], MD5C33, MD5S33);
+    MD5_STEP (MD5_I , a, b, c, d, t3[0], MD5C34, MD5S30);
+    MD5_STEP (MD5_I , d, a, b, c, t0[3], MD5C35, MD5S31);
+    MD5_STEP (MD5_I , c, d, a, b, t2[2], MD5C36, MD5S32);
+    MD5_STEP (MD5_I , b, c, d, a, t0[1], MD5C37, MD5S33);
+    MD5_STEP (MD5_I , a, b, c, d, t2[0], MD5C38, MD5S30);
+    MD5_STEP (MD5_I , d, a, b, c, t3[3], MD5C39, MD5S31);
+    MD5_STEP (MD5_I , c, d, a, b, t1[2], MD5C3a, MD5S32);
+    MD5_STEP (MD5_I , b, c, d, a, t3[1], MD5C3b, MD5S33);
+    MD5_STEP (MD5_I , a, b, c, d, t1[0], MD5C3c, MD5S30);
+    MD5_STEP (MD5_I , d, a, b, c, t2[3], MD5C3d, MD5S31);
+    MD5_STEP (MD5_I , c, d, a, b, t0[2], MD5C3e, MD5S32);
+    MD5_STEP (MD5_I , b, c, d, a, t2[1], MD5C3f, MD5S33);
 
     COMPARE_M_SIMD (a, d, c, b);
   }
 }
 
-static void m03800s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset)
+void m03800s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __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_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_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 il_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -279,53 +245,92 @@ static void m03800s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
   const u32 gid = get_global_id (0);
   const u32 lid = get_local_id (0);
 
-  /**
-   * 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]
-  };
-
   /**
    * 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];
-
-  u32 salt_buf1[4];
-
   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];
+
+  u32 salt_buf0_t[4];
+  u32 salt_buf1_t[4];
+  u32 salt_buf2_t[4];
+  u32 salt_buf3_t[4];
+
+  salt_buf0_t[0] = salt_bufs[salt_pos].salt_buf[ 0];
+  salt_buf0_t[1] = salt_bufs[salt_pos].salt_buf[ 1];
+  salt_buf0_t[2] = salt_bufs[salt_pos].salt_buf[ 2];
+  salt_buf0_t[3] = salt_bufs[salt_pos].salt_buf[ 3];
+  salt_buf1_t[0] = salt_bufs[salt_pos].salt_buf[ 4];
+  salt_buf1_t[1] = salt_bufs[salt_pos].salt_buf[ 5];
+  salt_buf1_t[2] = salt_bufs[salt_pos].salt_buf[ 6];
+  salt_buf1_t[3] = salt_bufs[salt_pos].salt_buf[ 7];
+  salt_buf2_t[0] = salt_bufs[salt_pos].salt_buf[ 8];
+  salt_buf2_t[1] = salt_bufs[salt_pos].salt_buf[ 9];
+  salt_buf2_t[2] = salt_bufs[salt_pos].salt_buf[10];
+  salt_buf2_t[3] = salt_bufs[salt_pos].salt_buf[11];
+  salt_buf3_t[0] = salt_bufs[salt_pos].salt_buf[12];
+  salt_buf3_t[1] = salt_bufs[salt_pos].salt_buf[13];
+  salt_buf3_t[2] = salt_bufs[salt_pos].salt_buf[14];
+  salt_buf3_t[3] = salt_bufs[salt_pos].salt_buf[15];
 
-  u32 salt_buf2[4];
+  const u32 salt_len = salt_bufs[salt_pos].salt_len;
 
-  salt_buf2[0] = 0;
-  salt_buf2[1] = 0;
-  salt_buf2[2] = 0;
-  salt_buf2[3] = 0;
+  const u32 pw_salt_len = pw_len + salt_len;
 
-  u32 salt_buf3[4];
+  const u32 salt_pw_salt_len = salt_len + pw_len + salt_len;
 
-  salt_buf3[0] = 0;
-  salt_buf3[1] = 0;
-  salt_buf3[2] = 0;
-  salt_buf3[3] = 0;
+  switch_buffer_by_offset_le_S (salt_buf0_t, salt_buf1_t, salt_buf2_t, salt_buf3_t, pw_salt_len);
 
-  const u32 salt_len = salt_bufs[salt_pos].salt_len;
+  salt_buf0[0] |= salt_buf0_t[0];
+  salt_buf0[1] |= salt_buf0_t[1];
+  salt_buf0[2] |= salt_buf0_t[2];
+  salt_buf0[3] |= salt_buf0_t[3];
+  salt_buf1[0] |= salt_buf1_t[0];
+  salt_buf1[1] |= salt_buf1_t[1];
+  salt_buf1[2] |= salt_buf1_t[2];
+  salt_buf1[3] |= salt_buf1_t[3];
+  salt_buf2[0] |= salt_buf2_t[0];
+  salt_buf2[1] |= salt_buf2_t[1];
+  salt_buf2[2] |= salt_buf2_t[2];
+  salt_buf2[3] |= salt_buf2_t[3];
+  salt_buf3[0] |= salt_buf3_t[0];
+  salt_buf3[1] |= salt_buf3_t[1];
+  salt_buf3[2] |= salt_buf3_t[2];
+  salt_buf3[3] |= salt_buf3_t[3];
+
+  append_0x80_4x4_S (salt_buf0, salt_buf1, salt_buf2, salt_buf3, salt_pw_salt_len);
+
+  /**
+   * digest
+   */
 
-  const u32 pw_salt_len = salt_len + pw_len + salt_len;
+  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
@@ -333,118 +338,56 @@ static void m03800s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
 
   u32 w0l = w0[0];
 
-  for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos += VECT_SIZE)
+  for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
   {
-    const u32x w0r = w0r_create_bft (bfs_buf, il_pos);
+    const u32x w0r = ix_create_bft (bfs_buf, il_pos);
 
     const u32x w0lr = w0l | w0r;
 
-    u32x w0_t[4];
-
-    w0_t[0] = w0lr;
-    w0_t[1] = w0[1];
-    w0_t[2] = w0[2];
-    w0_t[3] = w0[3];
-
-    u32x w1_t[4];
-
-    w1_t[0] = w1[0];
-    w1_t[1] = w1[1];
-    w1_t[2] = w1[2];
-    w1_t[3] = w1[3];
-
-    u32x w2_t[4];
-
-    w2_t[0] = w2[0];
-    w2_t[1] = w2[1];
-    w2_t[2] = w2[2];
-    w2_t[3] = w2[3];
-
-    u32x w3_t[4];
-
-    w3_t[0] = w3[0];
-    w3_t[1] = w3[1];
-    w3_t[2] = w3[2];
-    w3_t[3] = w3[3];
-
-
-    /**
-     * prepend salt
-     */
-
-    switch_buffer_by_offset_le (w0_t, w1_t, w2_t, w3_t, salt_len);
-
-    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] |= salt_buf3[2];
-    w3_t[3] |= salt_buf3[3];
+    u32x t0[4];
+    u32x t1[4];
+    u32x t2[4];
+    u32x t3[4];
+
+    t0[0] = w0lr;
+    t0[1] = w0[1];
+    t0[2] = w0[2];
+    t0[3] = w0[3];
+    t1[0] = w1[0];
+    t1[1] = w1[1];
+    t1[2] = w1[2];
+    t1[3] = w1[3];
+    t2[0] = w2[0];
+    t2[1] = w2[1];
+    t2[2] = w2[2];
+    t2[3] = w2[3];
+    t3[0] = w3[0];
+    t3[1] = w3[1];
+    t3[2] = w3[2];
+    t3[3] = w3[3];
 
     /**
-     * append salt
+     * put the password after the first salt but before the second salt
      */
 
-    u32x s0[4];
-
-    s0[0] = salt_buf0[0];
-    s0[1] = salt_buf0[1];
-    s0[2] = salt_buf0[2];
-    s0[3] = salt_buf0[3];
-
-    u32x s1[4];
-
-    s1[0] = salt_buf1[0];
-    s1[1] = salt_buf1[1];
-    s1[2] = salt_buf1[2];
-    s1[3] = salt_buf1[3];
-
-    u32x s2[4];
-
-    s2[0] = 0;
-    s2[1] = 0;
-    s2[2] = 0;
-    s2[3] = 0;
-
-    u32x s3[4];
-
-    s3[0] = 0;
-    s3[1] = 0;
-    s3[2] = 0;
-    s3[3] = 0;
-
-    switch_buffer_by_offset_le (s0, s1, s2, s3, salt_len + pw_len);
-
-    w0_t[0] |= s0[0];
-    w0_t[1] |= s0[1];
-    w0_t[2] |= s0[2];
-    w0_t[3] |= s0[3];
-    w1_t[0] |= s1[0];
-    w1_t[1] |= s1[1];
-    w1_t[2] |= s1[2];
-    w1_t[3] |= s1[3];
-    w2_t[0] |= s2[0];
-    w2_t[1] |= s2[1];
-    w2_t[2] |= s2[2];
-    w2_t[3] |= s2[3];
-    w3_t[0] |= s3[0];
-    w3_t[1] |= s3[1];
-    w3_t[2] |= s3[2];
-    w3_t[3] |= s3[3];
-
-    append_0x80_4x4 (w0_t, w1_t, w2_t, w3_t, pw_salt_len);
-
-    w3_t[2] = pw_salt_len * 8;
+    switch_buffer_by_offset_le (t0, t1, t2, t3, salt_len);
+
+    t0[0] |= salt_buf0[0];
+    t0[1] |= salt_buf0[1];
+    t0[2] |= salt_buf0[2];
+    t0[3] |= salt_buf0[3];
+    t1[0] |= salt_buf1[0];
+    t1[1] |= salt_buf1[1];
+    t1[2] |= salt_buf1[2];
+    t1[3] |= salt_buf1[3];
+    t2[0] |= salt_buf2[0];
+    t2[1] |= salt_buf2[1];
+    t2[2] |= salt_buf2[2];
+    t2[3] |= salt_buf2[3];
+    t3[0] |= salt_buf3[0];
+    t3[1] |= salt_buf3[1];
+    t3[2]  = salt_pw_salt_len * 8;
+    t3[3]  = 0;
 
     /**
      * md5
@@ -455,79 +398,82 @@ static void m03800s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
     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);
-
-    MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
-    MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
-    MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
-    MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
-    MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
-    MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
-    MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
-    MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
-    MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
-    MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
-    MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
-    MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
-    MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
-    MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
-    MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
-    MD5_STEP (MD5_H , 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);
+    MD5_STEP (MD5_Fo, a, b, c, d, t0[0], MD5C00, MD5S00);
+    MD5_STEP (MD5_Fo, d, a, b, c, t0[1], MD5C01, MD5S01);
+    MD5_STEP (MD5_Fo, c, d, a, b, t0[2], MD5C02, MD5S02);
+    MD5_STEP (MD5_Fo, b, c, d, a, t0[3], MD5C03, MD5S03);
+    MD5_STEP (MD5_Fo, a, b, c, d, t1[0], MD5C04, MD5S00);
+    MD5_STEP (MD5_Fo, d, a, b, c, t1[1], MD5C05, MD5S01);
+    MD5_STEP (MD5_Fo, c, d, a, b, t1[2], MD5C06, MD5S02);
+    MD5_STEP (MD5_Fo, b, c, d, a, t1[3], MD5C07, MD5S03);
+    MD5_STEP (MD5_Fo, a, b, c, d, t2[0], MD5C08, MD5S00);
+    MD5_STEP (MD5_Fo, d, a, b, c, t2[1], MD5C09, MD5S01);
+    MD5_STEP (MD5_Fo, c, d, a, b, t2[2], MD5C0a, MD5S02);
+    MD5_STEP (MD5_Fo, b, c, d, a, t2[3], MD5C0b, MD5S03);
+    MD5_STEP (MD5_Fo, a, b, c, d, t3[0], MD5C0c, MD5S00);
+    MD5_STEP (MD5_Fo, d, a, b, c, t3[1], MD5C0d, MD5S01);
+    MD5_STEP (MD5_Fo, c, d, a, b, t3[2], MD5C0e, MD5S02);
+    MD5_STEP (MD5_Fo, b, c, d, a, t3[3], MD5C0f, MD5S03);
+
+    MD5_STEP (MD5_Go, a, b, c, d, t0[1], MD5C10, MD5S10);
+    MD5_STEP (MD5_Go, d, a, b, c, t1[2], MD5C11, MD5S11);
+    MD5_STEP (MD5_Go, c, d, a, b, t2[3], MD5C12, MD5S12);
+    MD5_STEP (MD5_Go, b, c, d, a, t0[0], MD5C13, MD5S13);
+    MD5_STEP (MD5_Go, a, b, c, d, t1[1], MD5C14, MD5S10);
+    MD5_STEP (MD5_Go, d, a, b, c, t2[2], MD5C15, MD5S11);
+    MD5_STEP (MD5_Go, c, d, a, b, t3[3], MD5C16, MD5S12);
+    MD5_STEP (MD5_Go, b, c, d, a, t1[0], MD5C17, MD5S13);
+    MD5_STEP (MD5_Go, a, b, c, d, t2[1], MD5C18, MD5S10);
+    MD5_STEP (MD5_Go, d, a, b, c, t3[2], MD5C19, MD5S11);
+    MD5_STEP (MD5_Go, c, d, a, b, t0[3], MD5C1a, MD5S12);
+    MD5_STEP (MD5_Go, b, c, d, a, t2[0], MD5C1b, MD5S13);
+    MD5_STEP (MD5_Go, a, b, c, d, t3[1], MD5C1c, MD5S10);
+    MD5_STEP (MD5_Go, d, a, b, c, t0[2], MD5C1d, MD5S11);
+    MD5_STEP (MD5_Go, c, d, a, b, t1[3], MD5C1e, MD5S12);
+    MD5_STEP (MD5_Go, b, c, d, a, t3[0], MD5C1f, MD5S13);
+
+    MD5_STEP (MD5_H , a, b, c, d, t1[1], MD5C20, MD5S20);
+    MD5_STEP (MD5_H , d, a, b, c, t2[0], MD5C21, MD5S21);
+    MD5_STEP (MD5_H , c, d, a, b, t2[3], MD5C22, MD5S22);
+    MD5_STEP (MD5_H , b, c, d, a, t3[2], MD5C23, MD5S23);
+    MD5_STEP (MD5_H , a, b, c, d, t0[1], MD5C24, MD5S20);
+    MD5_STEP (MD5_H , d, a, b, c, t1[0], MD5C25, MD5S21);
+    MD5_STEP (MD5_H , c, d, a, b, t1[3], MD5C26, MD5S22);
+    MD5_STEP (MD5_H , b, c, d, a, t2[2], MD5C27, MD5S23);
+    MD5_STEP (MD5_H , a, b, c, d, t3[1], MD5C28, MD5S20);
+    MD5_STEP (MD5_H , d, a, b, c, t0[0], MD5C29, MD5S21);
+    MD5_STEP (MD5_H , c, d, a, b, t0[3], MD5C2a, MD5S22);
+    MD5_STEP (MD5_H , b, c, d, a, t1[2], MD5C2b, MD5S23);
+    MD5_STEP (MD5_H , a, b, c, d, t2[1], MD5C2c, MD5S20);
+    MD5_STEP (MD5_H , d, a, b, c, t3[0], MD5C2d, MD5S21);
+    MD5_STEP (MD5_H , c, d, a, b, t3[3], MD5C2e, MD5S22);
+    MD5_STEP (MD5_H , b, c, d, a, t0[2], MD5C2f, MD5S23);
+
+    MD5_STEP (MD5_I , a, b, c, d, t0[0], MD5C30, MD5S30);
+    MD5_STEP (MD5_I , d, a, b, c, t1[3], MD5C31, MD5S31);
+    MD5_STEP (MD5_I , c, d, a, b, t3[2], MD5C32, MD5S32);
+    MD5_STEP (MD5_I , b, c, d, a, t1[1], MD5C33, MD5S33);
+    MD5_STEP (MD5_I , a, b, c, d, t3[0], MD5C34, MD5S30);
+    MD5_STEP (MD5_I , d, a, b, c, t0[3], MD5C35, MD5S31);
+    MD5_STEP (MD5_I , c, d, a, b, t2[2], MD5C36, MD5S32);
+    MD5_STEP (MD5_I , b, c, d, a, t0[1], MD5C37, MD5S33);
+    MD5_STEP (MD5_I , a, b, c, d, t2[0], MD5C38, MD5S30);
+    MD5_STEP (MD5_I , d, a, b, c, t3[3], MD5C39, MD5S31);
+    MD5_STEP (MD5_I , c, d, a, b, t1[2], MD5C3a, MD5S32);
+    MD5_STEP (MD5_I , b, c, d, a, t3[1], MD5C3b, MD5S33);
+    MD5_STEP (MD5_I , a, b, c, d, t1[0], MD5C3c, MD5S30);
+
+    if (MATCHES_NONE_VS (a, search[0])) continue;
+
+    MD5_STEP (MD5_I , d, a, b, c, t2[3], MD5C3d, MD5S31);
+    MD5_STEP (MD5_I , c, d, a, b, t0[2], MD5C3e, MD5S32);
+    MD5_STEP (MD5_I , b, c, d, a, t2[1], MD5C3f, MD5S33);
 
     COMPARE_S_SIMD (a, d, c, b);
   }
 }
 
-__kernel void m03800_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
+__kernel void m03800_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_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_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 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
 {
   /**
    * base
@@ -577,10 +523,10 @@ __kernel void m03800_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    * main
    */
 
-  m03800m (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset);
+  m03800m (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV0_buf, d_scryptV1_buf, d_scryptV2_buf, d_scryptV3_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset);
 }
 
-__kernel void m03800_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
+__kernel void m03800_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_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_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 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
 {
   /**
    * base
@@ -630,10 +576,10 @@ __kernel void m03800_m08 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    * main
    */
 
-  m03800m (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset);
+  m03800m (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV0_buf, d_scryptV1_buf, d_scryptV2_buf, d_scryptV3_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset);
 }
 
-__kernel void m03800_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
+__kernel void m03800_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_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_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 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
 {
   /**
    * base
@@ -683,10 +629,10 @@ __kernel void m03800_m16 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    * main
    */
 
-  m03800m (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset);
+  m03800m (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV0_buf, d_scryptV1_buf, d_scryptV2_buf, d_scryptV3_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset);
 }
 
-__kernel void m03800_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
+__kernel void m03800_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_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_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 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
 {
   /**
    * base
@@ -736,10 +682,10 @@ __kernel void m03800_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    * main
    */
 
-  m03800s (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset);
+  m03800s (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV0_buf, d_scryptV1_buf, d_scryptV2_buf, d_scryptV3_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset);
 }
 
-__kernel void m03800_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
+__kernel void m03800_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_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_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 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
 {
   /**
    * base
@@ -789,10 +735,10 @@ __kernel void m03800_s08 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    * main
    */
 
-  m03800s (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset);
+  m03800s (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV0_buf, d_scryptV1_buf, d_scryptV2_buf, d_scryptV3_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset);
 }
 
-__kernel void m03800_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
+__kernel void m03800_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_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_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 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
 {
   /**
    * base
@@ -842,5 +788,5 @@ __kernel void m03800_s16 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    * main
    */
 
-  m03800s (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset);
+  m03800s (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV0_buf, d_scryptV1_buf, d_scryptV2_buf, d_scryptV3_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset);
 }