X-Git-Url: https://www.flypig.org.uk/git/?a=blobdiff_plain;f=OpenCL%2Fm03800_a0.cl;h=3ca0d7a48cacfd157f82d2e0d1e4411fc3e5d51a;hb=161a6eb4bc643d8e636e96eda613f5137d30da59;hp=c58da6510cc49a4396eeb1aa755a353f69a5f609;hpb=6893a9ad857a1d4fe63f1a92365802f5d9d3b39e;p=hashcat.git

diff --git a/OpenCL/m03800_a0.cl b/OpenCL/m03800_a0.cl
index c58da65..3ca0d7a 100644
--- a/OpenCL/m03800_a0.cl
+++ b/OpenCL/m03800_a0.cl
@@ -5,24 +5,18 @@
 
 #define _MD5_
 
-#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 "types_ocl.c"
-#include "common.c"
-#include "include/rp_gpu.h"
-#include "rp.c"
-
-#define COMPARE_S "check_single_comp4.c"
-#define COMPARE_M "check_multi_comp4.c"
-
-__kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m03800_m04 (__global pw_t *pws, __global gpu_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)
+#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.h"
+#include "inc_rp.cl"
+#include "inc_simd.cl"
+
+__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)
 {
   /**
    * modifier
@@ -36,19 +30,19 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m03800_m04 (__glo
 
   const u32 gid = get_global_id (0);
 
-  u32 pw_buf0[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];
+  if (gid >= gid_max) return;
 
+  u32 pw_buf0[4];
   u32 pw_buf1[4];
 
-  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];
+  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;
 
@@ -57,311 +51,209 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m03800_m04 (__glo
    */
 
   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];
-
-  u32 salt_buf2[4];
-
-  salt_buf2[0] = 0;
-  salt_buf2[1] = 0;
-  salt_buf2[2] = 0;
-  salt_buf2[3] = 0;
-
-  u32 salt_buf3[4];
-
-  salt_buf3[0] = 0;
-  salt_buf3[1] = 0;
-  salt_buf3[2] = 0;
-  salt_buf3[3] = 0;
+  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;
 
-  /**
-   * bin2asc table
-   */
-
-  __local u32 l_bin2asc[256];
-
-  const u32 lid4 = lid * 4;
-
-  const u32 lid40 = lid4 + 0;
-  const u32 lid41 = lid4 + 1;
-  const u32 lid42 = lid4 + 2;
-  const u32 lid43 = lid4 + 3;
-
-  const u32 v400 = (lid40 >> 0) & 15;
-  const u32 v401 = (lid40 >> 4) & 15;
-  const u32 v410 = (lid41 >> 0) & 15;
-  const u32 v411 = (lid41 >> 4) & 15;
-  const u32 v420 = (lid42 >> 0) & 15;
-  const u32 v421 = (lid42 >> 4) & 15;
-  const u32 v430 = (lid43 >> 0) & 15;
-  const u32 v431 = (lid43 >> 4) & 15;
-
-  l_bin2asc[lid40] = ((v400 < 10) ? '0' + v400 : 'a' - 10 + v400) << 8
-                   | ((v401 < 10) ? '0' + v401 : 'a' - 10 + v401) << 0;
-  l_bin2asc[lid41] = ((v410 < 10) ? '0' + v410 : 'a' - 10 + v410) << 8
-                   | ((v411 < 10) ? '0' + v411 : 'a' - 10 + v411) << 0;
-  l_bin2asc[lid42] = ((v420 < 10) ? '0' + v420 : 'a' - 10 + v420) << 8
-                   | ((v421 < 10) ? '0' + v421 : 'a' - 10 + v421) << 0;
-  l_bin2asc[lid43] = ((v430 < 10) ? '0' + v430 : 'a' - 10 + v430) << 8
-                   | ((v431 < 10) ? '0' + v431 : 'a' - 10 + v431) << 0;
-
-  barrier (CLK_LOCAL_MEM_FENCE);
-
-  if (gid >= gid_max) return;
-
   /**
    * loop
    */
 
-  for (u32 il_pos = 0; il_pos < rules_cnt; il_pos++)
+  for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
   {
-    u32 w0[4];
-
-    w0[0] = pw_buf0[0];
-    w0[1] = pw_buf0[1];
-    w0[2] = pw_buf0[2];
-    w0[3] = pw_buf0[3];
-
-    u32 w1[4];
-
-    w1[0] = pw_buf1[0];
-    w1[1] = pw_buf1[1];
-    w1[2] = pw_buf1[2];
-    w1[3] = pw_buf1[3];
-
-    u32 w2[4];
-
-    w2[0] = 0;
-    w2[1] = 0;
-    w2[2] = 0;
-    w2[3] = 0;
-
-    u32 w3[4];
-
-    w3[0] = 0;
-    w3[1] = 0;
-    w3[2] = 0;
-    w3[3] = 0;
-
-    const u32 out_len = apply_rules (rules_buf[il_pos].cmds, w0, w1, pw_len);
-
-    u32 w0_t[4];
-
-    w0_t[0] = w0[0];
-    w0_t[1] = w0[1];
-    w0_t[2] = w0[2];
-    w0_t[3] = w0[3];
-
-    u32 w1_t[4];
-
-    w1_t[0] = w1[0];
-    w1_t[1] = w1[1];
-    w1_t[2] = w1[2];
-    w1_t[3] = w1[3];
-
-    u32 w2_t[4];
-
-    w2_t[0] = w2[0];
-    w2_t[1] = w2[1];
-    w2_t[2] = w2[2];
-    w2_t[3] = w2[3];
-
-    u32 w3_t[4];
-
-    w3_t[0] = w3[0];
-    w3_t[1] = w3[1];
-    w3_t[2] = w3[2];
-    w3_t[3] = w3[3];
+    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);
 
     /**
      * prepend salt
      */
 
-    switch_buffer_by_offset (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];
+    const u32x out_salt_len = out_len + salt_len;
+
+    switch_buffer_by_offset_le_VV (w0, w1, w2, w3, salt_len);
+
+    w0[0] |= salt_buf0[0];
+    w0[1] |= salt_buf0[1];
+    w0[2] |= salt_buf0[2];
+    w0[3] |= salt_buf0[3];
+    w1[0] |= salt_buf1[0];
+    w1[1] |= salt_buf1[1];
+    w1[2] |= salt_buf1[2];
+    w1[3] |= salt_buf1[3];
+    w2[0] |= salt_buf2[0];
+    w2[1] |= salt_buf2[1];
+    w2[2] |= salt_buf2[2];
+    w2[3] |= salt_buf2[3];
+    w3[0] |= salt_buf3[0];
+    w3[1] |= salt_buf3[1];
+    w3[2] |= salt_buf3[2];
+    w3[3] |= salt_buf3[3];
 
     /**
      * append salt
      */
 
-    u32 s0[4];
+    u32x s0[4];
+    u32x s1[4];
+    u32x s2[4];
+    u32x s3[4];
 
     s0[0] = salt_buf0[0];
     s0[1] = salt_buf0[1];
     s0[2] = salt_buf0[2];
     s0[3] = salt_buf0[3];
-
-    u32 s1[4];
-
     s1[0] = salt_buf1[0];
     s1[1] = salt_buf1[1];
     s1[2] = salt_buf1[2];
     s1[3] = salt_buf1[3];
-
-    u32 s2[4];
-
-    s2[0] = 0;
-    s2[1] = 0;
-    s2[2] = 0;
-    s2[3] = 0;
-
-    u32 s3[4];
-
-    s3[0] = 0;
-    s3[1] = 0;
-    s3[2] = 0;
-    s3[3] = 0;
-
-    switch_buffer_by_offset (s0, s1, s2, s3, salt_len + out_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];
-
-    const u32 pw_salt_len = salt_len + out_len + salt_len;
-
-    append_0x80_4x4 (w0_t, w1_t, w2_t, w3_t, pw_salt_len);
-
-    w3_t[2] = pw_salt_len * 8;
+    s2[0] = salt_buf2[0];
+    s2[1] = salt_buf2[1];
+    s2[2] = salt_buf2[2];
+    s2[3] = salt_buf2[3];
+    s3[0] = salt_buf3[0];
+    s3[1] = salt_buf3[1];
+    s3[2] = salt_buf3[2];
+    s3[3] = salt_buf3[3];
+
+    switch_buffer_by_offset_le_VV (s0, s1, s2, s3, out_salt_len);
+
+    w0[0] |= s0[0];
+    w0[1] |= s0[1];
+    w0[2] |= s0[2];
+    w0[3] |= s0[3];
+    w1[0] |= s1[0];
+    w1[1] |= s1[1];
+    w1[2] |= s1[2];
+    w1[3] |= s1[3];
+    w2[0] |= s2[0];
+    w2[1] |= s2[1];
+    w2[2] |= s2[2];
+    w2[3] |= s2[3];
+    w3[0] |= s3[0];
+    w3[1] |= s3[1];
+    w3[2] |= s3[2];
+    w3[3] |= s3[3];
+
+    const u32x salt_out_salt_len = salt_len + out_len + salt_len;
+
+    append_0x80_4x4_VV (w0, w1, w2, w3, salt_out_salt_len);
+
+    w3[2] = salt_out_salt_len * 8;
+    w3[3] = 0;
 
     /**
      * md5
      */
 
-    u32 a = MD5M_A;
-    u32 b = MD5M_B;
-    u32 c = MD5M_C;
-    u32 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);
-
-    const u32 r0 = a;
-    const u32 r1 = d;
-    const u32 r2 = c;
-    const u32 r3 = b;
-
-    #include COMPARE_M
+    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 __attribute__((reqd_work_group_size (64, 1, 1))) m03800_m08 (__global pw_t *pws, __global gpu_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 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)
 {
 }
 
-__kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m03800_m16 (__global pw_t *pws, __global gpu_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 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)
 {
 }
 
-__kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m03800_s04 (__global pw_t *pws, __global gpu_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)
+__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)
 {
   /**
    * modifier
@@ -375,19 +267,19 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m03800_s04 (__glo
 
   const u32 gid = get_global_id (0);
 
-  u32 pw_buf0[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];
+  if (gid >= gid_max) return;
 
+  u32 pw_buf0[4];
   u32 pw_buf1[4];
 
-  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];
+  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;
 
@@ -396,32 +288,26 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m03800_s04 (__glo
    */
 
   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];
-
-  u32 salt_buf2[4];
-
-  salt_buf2[0] = 0;
-  salt_buf2[1] = 0;
-  salt_buf2[2] = 0;
-  salt_buf2[3] = 0;
-
-  u32 salt_buf3[4];
-
-  salt_buf3[0] = 0;
-  salt_buf3[1] = 0;
-  salt_buf3[2] = 0;
-  salt_buf3[3] = 0;
+  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;
 
@@ -437,276 +323,184 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m03800_s04 (__glo
     digests_buf[digests_offset].digest_buf[DGST_R3]
   };
 
-  /**
-   * bin2asc table
-   */
-
-  __local u32 l_bin2asc[256];
-
-  const u32 lid4 = lid * 4;
-
-  const u32 lid40 = lid4 + 0;
-  const u32 lid41 = lid4 + 1;
-  const u32 lid42 = lid4 + 2;
-  const u32 lid43 = lid4 + 3;
-
-  const u32 v400 = (lid40 >> 0) & 15;
-  const u32 v401 = (lid40 >> 4) & 15;
-  const u32 v410 = (lid41 >> 0) & 15;
-  const u32 v411 = (lid41 >> 4) & 15;
-  const u32 v420 = (lid42 >> 0) & 15;
-  const u32 v421 = (lid42 >> 4) & 15;
-  const u32 v430 = (lid43 >> 0) & 15;
-  const u32 v431 = (lid43 >> 4) & 15;
-
-  l_bin2asc[lid40] = ((v400 < 10) ? '0' + v400 : 'a' - 10 + v400) << 8
-                   | ((v401 < 10) ? '0' + v401 : 'a' - 10 + v401) << 0;
-  l_bin2asc[lid41] = ((v410 < 10) ? '0' + v410 : 'a' - 10 + v410) << 8
-                   | ((v411 < 10) ? '0' + v411 : 'a' - 10 + v411) << 0;
-  l_bin2asc[lid42] = ((v420 < 10) ? '0' + v420 : 'a' - 10 + v420) << 8
-                   | ((v421 < 10) ? '0' + v421 : 'a' - 10 + v421) << 0;
-  l_bin2asc[lid43] = ((v430 < 10) ? '0' + v430 : 'a' - 10 + v430) << 8
-                   | ((v431 < 10) ? '0' + v431 : 'a' - 10 + v431) << 0;
-
-  barrier (CLK_LOCAL_MEM_FENCE);
-
-  if (gid >= gid_max) return;
-
   /**
    * loop
    */
 
-  for (u32 il_pos = 0; il_pos < rules_cnt; il_pos++)
+  for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
   {
-    u32 w0[4];
-
-    w0[0] = pw_buf0[0];
-    w0[1] = pw_buf0[1];
-    w0[2] = pw_buf0[2];
-    w0[3] = pw_buf0[3];
-
-    u32 w1[4];
-
-    w1[0] = pw_buf1[0];
-    w1[1] = pw_buf1[1];
-    w1[2] = pw_buf1[2];
-    w1[3] = pw_buf1[3];
-
-    u32 w2[4];
-
-    w2[0] = 0;
-    w2[1] = 0;
-    w2[2] = 0;
-    w2[3] = 0;
-
-    u32 w3[4];
-
-    w3[0] = 0;
-    w3[1] = 0;
-    w3[2] = 0;
-    w3[3] = 0;
-
-    const u32 out_len = apply_rules (rules_buf[il_pos].cmds, w0, w1, pw_len);
-
-    u32 w0_t[4];
-
-    w0_t[0] = w0[0];
-    w0_t[1] = w0[1];
-    w0_t[2] = w0[2];
-    w0_t[3] = w0[3];
-
-    u32 w1_t[4];
-
-    w1_t[0] = w1[0];
-    w1_t[1] = w1[1];
-    w1_t[2] = w1[2];
-    w1_t[3] = w1[3];
-
-    u32 w2_t[4];
-
-    w2_t[0] = w2[0];
-    w2_t[1] = w2[1];
-    w2_t[2] = w2[2];
-    w2_t[3] = w2[3];
-
-    u32 w3_t[4];
-
-    w3_t[0] = w3[0];
-    w3_t[1] = w3[1];
-    w3_t[2] = w3[2];
-    w3_t[3] = w3[3];
+    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);
 
     /**
      * prepend salt
      */
 
-    switch_buffer_by_offset (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];
+    const u32x out_salt_len = out_len + salt_len;
+
+    switch_buffer_by_offset_le_VV (w0, w1, w2, w3, salt_len);
+
+    w0[0] |= salt_buf0[0];
+    w0[1] |= salt_buf0[1];
+    w0[2] |= salt_buf0[2];
+    w0[3] |= salt_buf0[3];
+    w1[0] |= salt_buf1[0];
+    w1[1] |= salt_buf1[1];
+    w1[2] |= salt_buf1[2];
+    w1[3] |= salt_buf1[3];
+    w2[0] |= salt_buf2[0];
+    w2[1] |= salt_buf2[1];
+    w2[2] |= salt_buf2[2];
+    w2[3] |= salt_buf2[3];
+    w3[0] |= salt_buf3[0];
+    w3[1] |= salt_buf3[1];
+    w3[2] |= salt_buf3[2];
+    w3[3] |= salt_buf3[3];
 
     /**
      * append salt
      */
 
-    u32 s0[4];
+    u32x s0[4];
+    u32x s1[4];
+    u32x s2[4];
+    u32x s3[4];
 
     s0[0] = salt_buf0[0];
     s0[1] = salt_buf0[1];
     s0[2] = salt_buf0[2];
     s0[3] = salt_buf0[3];
-
-    u32 s1[4];
-
     s1[0] = salt_buf1[0];
     s1[1] = salt_buf1[1];
     s1[2] = salt_buf1[2];
     s1[3] = salt_buf1[3];
-
-    u32 s2[4];
-
-    s2[0] = 0;
-    s2[1] = 0;
-    s2[2] = 0;
-    s2[3] = 0;
-
-    u32 s3[4];
-
-    s3[0] = 0;
-    s3[1] = 0;
-    s3[2] = 0;
-    s3[3] = 0;
-
-    switch_buffer_by_offset (s0, s1, s2, s3, salt_len + out_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];
-
-    const u32 pw_salt_len = salt_len + out_len + salt_len;
-
-    append_0x80_4x4 (w0_t, w1_t, w2_t, w3_t, pw_salt_len);
-
-    w3_t[2] = pw_salt_len * 8;
+    s2[0] = salt_buf2[0];
+    s2[1] = salt_buf2[1];
+    s2[2] = salt_buf2[2];
+    s2[3] = salt_buf2[3];
+    s3[0] = salt_buf3[0];
+    s3[1] = salt_buf3[1];
+    s3[2] = salt_buf3[2];
+    s3[3] = salt_buf3[3];
+
+    switch_buffer_by_offset_le_VV (s0, s1, s2, s3, out_salt_len);
+
+    w0[0] |= s0[0];
+    w0[1] |= s0[1];
+    w0[2] |= s0[2];
+    w0[3] |= s0[3];
+    w1[0] |= s1[0];
+    w1[1] |= s1[1];
+    w1[2] |= s1[2];
+    w1[3] |= s1[3];
+    w2[0] |= s2[0];
+    w2[1] |= s2[1];
+    w2[2] |= s2[2];
+    w2[3] |= s2[3];
+    w3[0] |= s3[0];
+    w3[1] |= s3[1];
+    w3[2] |= s3[2];
+    w3[3] |= s3[3];
+
+    const u32x salt_out_salt_len = salt_len + out_len + salt_len;
+
+    append_0x80_4x4_VV (w0, w1, w2, w3, salt_out_salt_len);
+
+    w3[2] = salt_out_salt_len * 8;
+    w3[3] = 0;
 
     /**
      * md5
      */
 
-    u32 a = MD5M_A;
-    u32 b = MD5M_B;
-    u32 c = MD5M_C;
-    u32 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);
-
-    const u32 r0 = a;
-    const u32 r1 = d;
-    const u32 r2 = c;
-    const u32 r3 = b;
-
-    #include COMPARE_S
+    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 __attribute__((reqd_work_group_size (64, 1, 1))) m03800_s08 (__global pw_t *pws, __global gpu_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 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)
 {
 }
 
-__kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m03800_s16 (__global pw_t *pws, __global gpu_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 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)
 {
 }