X-Git-Url: https://www.flypig.org.uk/git/?a=blobdiff_plain;f=OpenCL%2Fm01720_a3.cl;h=9b74a9a254eea11ef652c16888e6d9c8e01d0c19;hb=8702d0e3e155e6a61ca1b0dcdc1ec924d92b8436;hp=2a363d4adc85423e866ebee655f93eccdf97a220;hpb=0bf4e3c34a6a799ccc34f403bed70119574ca9c8;p=hashcat.git

diff --git a/OpenCL/m01720_a3.cl b/OpenCL/m01720_a3.cl
index 2a363d4..9b74a9a 100644
--- a/OpenCL/m01720_a3.cl
+++ b/OpenCL/m01720_a3.cl
@@ -5,20 +5,14 @@
 
 #define _SHA512_
 
-#include "include/constants.h"
-#include "include/kernel_vendor.h"
+#define NEW_SIMD_CODE
 
-#define DGST_R0 14
-#define DGST_R1 15
-#define DGST_R2 6
-#define DGST_R3 7
-
-#include "include/kernel_functions.c"
-#include "types_ocl.c"
-#include "common.c"
-
-#define COMPARE_S "check_single_comp4.c"
-#define COMPARE_M "check_multi_comp4.c"
+#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"
 
 __constant u64 k_sha512[80] =
 {
@@ -44,33 +38,33 @@ __constant u64 k_sha512[80] =
   SHA512C4c, SHA512C4d, SHA512C4e, SHA512C4f,
 };
 
-static void sha512_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u64 digest[8])
+void sha512_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u64x digest[8])
 {
-  u64 w0_t = hl32_to_64 (w0[0], w0[1]);
-  u64 w1_t = hl32_to_64 (w0[2], w0[3]);
-  u64 w2_t = hl32_to_64 (w1[0], w1[1]);
-  u64 w3_t = hl32_to_64 (w1[2], w1[3]);
-  u64 w4_t = hl32_to_64 (w2[0], w2[1]);
-  u64 w5_t = hl32_to_64 (w2[2], w2[3]);
-  u64 w6_t = hl32_to_64 (w3[0], w3[1]);
-  u64 w7_t = 0;
-  u64 w8_t = 0;
-  u64 w9_t = 0;
-  u64 wa_t = 0;
-  u64 wb_t = 0;
-  u64 wc_t = 0;
-  u64 wd_t = 0;
-  u64 we_t = 0;
-  u64 wf_t = hl32_to_64 (w3[2], w3[3]);
-
-  u64 a = digest[0];
-  u64 b = digest[1];
-  u64 c = digest[2];
-  u64 d = digest[3];
-  u64 e = digest[4];
-  u64 f = digest[5];
-  u64 g = digest[6];
-  u64 h = digest[7];
+  u64x w0_t = hl32_to_64 (w0[0], w0[1]);
+  u64x w1_t = hl32_to_64 (w0[2], w0[3]);
+  u64x w2_t = hl32_to_64 (w1[0], w1[1]);
+  u64x w3_t = hl32_to_64 (w1[2], w1[3]);
+  u64x w4_t = hl32_to_64 (w2[0], w2[1]);
+  u64x w5_t = hl32_to_64 (w2[2], w2[3]);
+  u64x w6_t = hl32_to_64 (w3[0], w3[1]);
+  u64x w7_t = 0;
+  u64x w8_t = 0;
+  u64x w9_t = 0;
+  u64x wa_t = 0;
+  u64x wb_t = 0;
+  u64x wc_t = 0;
+  u64x wd_t = 0;
+  u64x we_t = 0;
+  u64x wf_t = hl32_to_64 (w3[2], w3[3]);
+
+  u64x a = digest[0];
+  u64x b = digest[1];
+  u64x c = digest[2];
+  u64x d = digest[3];
+  u64x e = digest[4];
+  u64x f = digest[5];
+  u64x g = digest[6];
+  u64x h = digest[7];
 
   #define ROUND_EXPAND()                            \
   {                                                 \
@@ -114,7 +108,9 @@ static void sha512_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4],
 
   ROUND_STEP (0);
 
+  #ifdef _unroll
   #pragma unroll
+  #endif
   for (int i = 16; i < 80; i += 16)
   {
     ROUND_EXPAND (); ROUND_STEP (i);
@@ -141,7 +137,7 @@ static void sha512_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4],
   digest[7] = h;
 }
 
-static void m01720m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset)
+void m01720m (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
@@ -155,116 +151,117 @@ static void m01720m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
    */
 
   u32 salt_buf0[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_buf0[0] = swap32_S (salt_bufs[salt_pos].salt_buf[ 0]);
+  salt_buf0[1] = swap32_S (salt_bufs[salt_pos].salt_buf[ 1]);
+  salt_buf0[2] = swap32_S (salt_bufs[salt_pos].salt_buf[ 2]);
+  salt_buf0[3] = swap32_S (salt_bufs[salt_pos].salt_buf[ 3]);
+  salt_buf1[0] = swap32_S (salt_bufs[salt_pos].salt_buf[ 4]);
+  salt_buf1[1] = swap32_S (salt_bufs[salt_pos].salt_buf[ 5]);
+  salt_buf1[2] = swap32_S (salt_bufs[salt_pos].salt_buf[ 6]);
+  salt_buf1[3] = swap32_S (salt_bufs[salt_pos].salt_buf[ 7]);
+  salt_buf2[0] = swap32_S (salt_bufs[salt_pos].salt_buf[ 8]);
+  salt_buf2[1] = swap32_S (salt_bufs[salt_pos].salt_buf[ 9]);
+  salt_buf2[2] = swap32_S (salt_bufs[salt_pos].salt_buf[10]);
+  salt_buf2[3] = swap32_S (salt_bufs[salt_pos].salt_buf[11]);
+  salt_buf3[0] = swap32_S (salt_bufs[salt_pos].salt_buf[12]);
+  salt_buf3[1] = swap32_S (salt_bufs[salt_pos].salt_buf[13]);
+  salt_buf3[2] = swap32_S (salt_bufs[salt_pos].salt_buf[14]);
+  salt_buf3[3] = swap32_S (salt_bufs[salt_pos].salt_buf[15]);
 
   const u32 salt_len = salt_bufs[salt_pos].salt_len;
 
   const u32 pw_salt_len = pw_len + salt_len;
 
   /**
-   * loop
+   * prepend salt
    */
 
-  u32 w0l = w0[0];
+  const u32 w0l = w0[0];
+
+  switch_buffer_by_offset_be_S (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];
+
+  u32x t0[4];
+  u32x t1[4];
+  u32x t2[4];
+  u32x t3[4];
+
+  t0[0] = w0[0];
+  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];
 
-  for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos++)
-  {
-    const u32 w0r = bfs_buf[il_pos].i;
+  /**
+   * loop
+   */
 
-    w0[0] = w0l | w0r;
+  for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
+  {
+    const u32x w0r = ix_create_bft (bfs_buf, il_pos);
 
-    /**
-     * prepend salt
-     */
+    const u32x w0lr = w0l | w0r;
 
-    u32 w0_t[4];
-    u32 w1_t[4];
-    u32 w2_t[4];
-    u32 w3_t[4];
-
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
-    w2_t[0] = swap_workaround (w2[0]);
-    w2_t[1] = swap_workaround (w2[1]);
-    w2_t[2] = swap_workaround (w2[2]);
-    w2_t[3] = swap_workaround (w2[3]);
-    w3_t[0] = swap_workaround (w3[0]);
-    w3_t[1] = swap_workaround (w3[1]);
-    w3_t[2] = swap_workaround (w3[2]);
-    w3_t[3] = swap_workaround (w3[3]);
-
-    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]  = 0;
-    w3_t[3]  = pw_salt_len * 8;
+    overwrite_at_be_4x4 (t0, t1, t2, t3, w0lr, salt_len);
 
     /**
      * sha512
      */
 
-    w0_t[0] = swap_workaround (w0_t[0]);
-    w0_t[1] = swap_workaround (w0_t[1]);
-    w0_t[2] = swap_workaround (w0_t[2]);
-    w0_t[3] = swap_workaround (w0_t[3]);
-    w1_t[0] = swap_workaround (w1_t[0]);
-    w1_t[1] = swap_workaround (w1_t[1]);
-    w1_t[2] = swap_workaround (w1_t[2]);
-    w1_t[3] = swap_workaround (w1_t[3]);
-    w2_t[0] = swap_workaround (w2_t[0]);
-    w2_t[1] = swap_workaround (w2_t[1]);
-    w2_t[2] = swap_workaround (w2_t[2]);
-    w2_t[3] = swap_workaround (w2_t[3]);
-    w3_t[0] = swap_workaround (w3_t[0]);
-    w3_t[1] = swap_workaround (w3_t[1]);
-    //w3_t[2] = swap_workaround (w3_t[2]);
-    //w3_t[3] = swap_workaround (w3_t[3]);
-
-    u64 digest[8];
+    u32x w0_t[4];
+    u32x w1_t[4];
+    u32x w2_t[4];
+    u32x w3_t[4];
+
+    w0_t[0] = t0[0];
+    w0_t[1] = t0[1];
+    w0_t[2] = t0[2];
+    w0_t[3] = t0[3];
+    w1_t[0] = t1[0];
+    w1_t[1] = t1[1];
+    w1_t[2] = t1[2];
+    w1_t[3] = t1[3];
+    w2_t[0] = t2[0];
+    w2_t[1] = t2[1];
+    w2_t[2] = t2[2];
+    w2_t[3] = t2[3];
+    w3_t[0] = t3[0];
+    w3_t[1] = t3[1];
+    w3_t[2] = 0;
+    w3_t[3] = pw_salt_len * 8;
+
+    u64x digest[8];
 
     digest[0] = SHA512M_A;
     digest[1] = SHA512M_B;
@@ -277,17 +274,16 @@ static void m01720m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
 
     sha512_transform (w0_t, w1_t, w2_t, w3_t, digest);
 
+    const u32x r0 = l32_from_64 (digest[7]);
+    const u32x r1 = h32_from_64 (digest[7]);
+    const u32x r2 = l32_from_64 (digest[3]);
+    const u32x r3 = h32_from_64 (digest[3]);
 
-    const u32 r0 = l32_from_64 (digest[7]);
-    const u32 r1 = h32_from_64 (digest[7]);
-    const u32 r2 = l32_from_64 (digest[3]);
-    const u32 r3 = h32_from_64 (digest[3]);
-
-    #include COMPARE_M
+    COMPARE_M_SIMD (r0, r1, r2, r3);
   }
 }
 
-static void m01720s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset)
+void m01720s (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
@@ -313,116 +309,117 @@ static void m01720s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
    */
 
   u32 salt_buf0[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_buf0[0] = swap32_S (salt_bufs[salt_pos].salt_buf[ 0]);
+  salt_buf0[1] = swap32_S (salt_bufs[salt_pos].salt_buf[ 1]);
+  salt_buf0[2] = swap32_S (salt_bufs[salt_pos].salt_buf[ 2]);
+  salt_buf0[3] = swap32_S (salt_bufs[salt_pos].salt_buf[ 3]);
+  salt_buf1[0] = swap32_S (salt_bufs[salt_pos].salt_buf[ 4]);
+  salt_buf1[1] = swap32_S (salt_bufs[salt_pos].salt_buf[ 5]);
+  salt_buf1[2] = swap32_S (salt_bufs[salt_pos].salt_buf[ 6]);
+  salt_buf1[3] = swap32_S (salt_bufs[salt_pos].salt_buf[ 7]);
+  salt_buf2[0] = swap32_S (salt_bufs[salt_pos].salt_buf[ 8]);
+  salt_buf2[1] = swap32_S (salt_bufs[salt_pos].salt_buf[ 9]);
+  salt_buf2[2] = swap32_S (salt_bufs[salt_pos].salt_buf[10]);
+  salt_buf2[3] = swap32_S (salt_bufs[salt_pos].salt_buf[11]);
+  salt_buf3[0] = swap32_S (salt_bufs[salt_pos].salt_buf[12]);
+  salt_buf3[1] = swap32_S (salt_bufs[salt_pos].salt_buf[13]);
+  salt_buf3[2] = swap32_S (salt_bufs[salt_pos].salt_buf[14]);
+  salt_buf3[3] = swap32_S (salt_bufs[salt_pos].salt_buf[15]);
 
   const u32 salt_len = salt_bufs[salt_pos].salt_len;
 
   const u32 pw_salt_len = pw_len + salt_len;
 
   /**
-   * loop
+   * prepend salt
    */
 
-  u32 w0l = w0[0];
+  const u32 w0l = w0[0];
+
+  switch_buffer_by_offset_be_S (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];
+
+  u32x t0[4];
+  u32x t1[4];
+  u32x t2[4];
+  u32x t3[4];
+
+  t0[0] = w0[0];
+  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];
 
-  for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos++)
-  {
-    const u32 w0r = bfs_buf[il_pos].i;
+  /**
+   * loop
+   */
 
-    w0[0] = w0l | w0r;
+  for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
+  {
+    const u32x w0r = ix_create_bft (bfs_buf, il_pos);
 
-    /**
-     * prepend salt
-     */
+    const u32x w0lr = w0l | w0r;
 
-    u32 w0_t[4];
-    u32 w1_t[4];
-    u32 w2_t[4];
-    u32 w3_t[4];
-
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
-    w2_t[0] = swap_workaround (w2[0]);
-    w2_t[1] = swap_workaround (w2[1]);
-    w2_t[2] = swap_workaround (w2[2]);
-    w2_t[3] = swap_workaround (w2[3]);
-    w3_t[0] = swap_workaround (w3[0]);
-    w3_t[1] = swap_workaround (w3[1]);
-    w3_t[2] = swap_workaround (w3[2]);
-    w3_t[3] = swap_workaround (w3[3]);
-
-    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]  = 0;
-    w3_t[3]  = pw_salt_len * 8;
+    overwrite_at_be_4x4 (t0, t1, t2, t3, w0lr, salt_len);
 
     /**
      * sha512
      */
 
-    w0_t[0] = swap_workaround (w0_t[0]);
-    w0_t[1] = swap_workaround (w0_t[1]);
-    w0_t[2] = swap_workaround (w0_t[2]);
-    w0_t[3] = swap_workaround (w0_t[3]);
-    w1_t[0] = swap_workaround (w1_t[0]);
-    w1_t[1] = swap_workaround (w1_t[1]);
-    w1_t[2] = swap_workaround (w1_t[2]);
-    w1_t[3] = swap_workaround (w1_t[3]);
-    w2_t[0] = swap_workaround (w2_t[0]);
-    w2_t[1] = swap_workaround (w2_t[1]);
-    w2_t[2] = swap_workaround (w2_t[2]);
-    w2_t[3] = swap_workaround (w2_t[3]);
-    w3_t[0] = swap_workaround (w3_t[0]);
-    w3_t[1] = swap_workaround (w3_t[1]);
-    //w3_t[2] = swap_workaround (w3_t[2]);
-    //w3_t[3] = swap_workaround (w3_t[3]);
-
-    u64 digest[8];
+    u32x w0_t[4];
+    u32x w1_t[4];
+    u32x w2_t[4];
+    u32x w3_t[4];
+
+    w0_t[0] = t0[0];
+    w0_t[1] = t0[1];
+    w0_t[2] = t0[2];
+    w0_t[3] = t0[3];
+    w1_t[0] = t1[0];
+    w1_t[1] = t1[1];
+    w1_t[2] = t1[2];
+    w1_t[3] = t1[3];
+    w2_t[0] = t2[0];
+    w2_t[1] = t2[1];
+    w2_t[2] = t2[2];
+    w2_t[3] = t2[3];
+    w3_t[0] = t3[0];
+    w3_t[1] = t3[1];
+    w3_t[2] = 0;
+    w3_t[3] = pw_salt_len * 8;
+
+    u64x digest[8];
 
     digest[0] = SHA512M_A;
     digest[1] = SHA512M_B;
@@ -435,17 +432,16 @@ static void m01720s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
 
     sha512_transform (w0_t, w1_t, w2_t, w3_t, digest);
 
+    const u32x r0 = l32_from_64 (digest[7]);
+    const u32x r1 = h32_from_64 (digest[7]);
+    const u32x r2 = l32_from_64 (digest[3]);
+    const u32x r3 = h32_from_64 (digest[3]);
 
-    const u32 r0 = l32_from_64 (digest[7]);
-    const u32 r1 = h32_from_64 (digest[7]);
-    const u32 r2 = l32_from_64 (digest[3]);
-    const u32 r3 = h32_from_64 (digest[3]);
-
-    #include COMPARE_S
+    COMPARE_S_SIMD (r0, r1, r2, r3);
   }
 }
 
-__kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01720_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
+__kernel void m01720_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
@@ -489,10 +485,10 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01720_m04 (__glo
    * main
    */
 
-  m01720m (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);
+  m01720m (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 __attribute__((reqd_work_group_size (64, 1, 1))) m01720_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
+__kernel void m01720_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
@@ -536,10 +532,10 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01720_m08 (__glo
    * main
    */
 
-  m01720m (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);
+  m01720m (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 __attribute__((reqd_work_group_size (64, 1, 1))) m01720_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
+__kernel void m01720_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
@@ -583,10 +579,10 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01720_m16 (__glo
    * main
    */
 
-  m01720m (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);
+  m01720m (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 __attribute__((reqd_work_group_size (64, 1, 1))) m01720_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
+__kernel void m01720_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
@@ -630,10 +626,10 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01720_s04 (__glo
    * main
    */
 
-  m01720s (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);
+  m01720s (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 __attribute__((reqd_work_group_size (64, 1, 1))) m01720_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
+__kernel void m01720_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
@@ -677,10 +673,10 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01720_s08 (__glo
    * main
    */
 
-  m01720s (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);
+  m01720s (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 __attribute__((reqd_work_group_size (64, 1, 1))) m01720_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
+__kernel void m01720_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
@@ -724,5 +720,5 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01720_s16 (__glo
    * main
    */
 
-  m01720s (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);
+  m01720s (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);
 }