X-Git-Url: https://www.flypig.org.uk/git/?a=blobdiff_plain;f=OpenCL%2Fm09710_a1.cl;h=ef6bdf17b38353cf80948683e8bf1a345da08f89;hb=161a6eb4bc643d8e636e96eda613f5137d30da59;hp=36a79b8c0c9c057964af74c22b2c9384d657af7b;hpb=5c07a412ec247a9e2b32c4d79996c4db5768560c;p=hashcat.git

diff --git a/OpenCL/m09710_a1.cl b/OpenCL/m09710_a1.cl
index 36a79b8..ef6bdf1 100644
--- a/OpenCL/m09710_a1.cl
+++ b/OpenCL/m09710_a1.cl
@@ -5,20 +5,15 @@
 
 #define _OLDOFFICE01_
 
-#include "include/constants.h"
-#include "include/kernel_vendor.h"
+//too much register pressure
+//#define NEW_SIMD_CODE
 
-#define DGST_R0 0
-#define DGST_R1 1
-#define DGST_R2 2
-#define DGST_R3 3
-
-#include "include/kernel_functions.c"
-#include "OpenCL/types_ocl.c"
-#include "OpenCL/common.c"
-
-#define COMPARE_S "OpenCL/check_single_comp4.c"
-#define COMPARE_M "OpenCL/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"
 
 typedef struct
 {
@@ -28,7 +23,7 @@ typedef struct
 
 } RC4_KEY;
 
-static void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
+void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
 {
   u8 tmp;
 
@@ -37,14 +32,16 @@ static void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
   rc4_key->S[j] = tmp;
 }
 
-static void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
+void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
 {
   u32 v = 0x03020100;
   u32 a = 0x04040404;
 
   __local u32 *ptr = (__local u32 *) rc4_key->S;
 
+  #ifdef _unroll
   #pragma unroll
+  #endif
   for (u32 i = 0; i < 64; i++)
   {
     *ptr++ = v; v += a;
@@ -88,9 +85,11 @@ static void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
   }
 }
 
-static u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
+u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
 {
+  #ifdef _unroll
   #pragma unroll
+  #endif
   for (u32 k = 0; k < 4; k++)
   {
     u32 xor4 = 0;
@@ -139,7 +138,7 @@ static u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u3
   return j;
 }
 
-static void md5_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4])
+void md5_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4])
 {
   u32 a = digest[0];
   u32 b = digest[1];
@@ -237,7 +236,7 @@ static void md5_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], co
   digest[3] += d;
 }
 
-__kernel void m09710_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 oldoffice01_t *oldoffice01_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 m09710_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 oldoffice01_t *oldoffice01_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
@@ -245,10 +244,6 @@ __kernel void m09710_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 
   const u32 lid = get_local_id (0);
 
-  __local RC4_KEY rc4_keys[64];
-
-  __local RC4_KEY *rc4_key = &rc4_keys[lid];
-
   /**
    * base
    */
@@ -257,40 +252,27 @@ __kernel void m09710_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 
   if (gid >= gid_max) return;
 
-  u32 wordl0[4];
-
-  wordl0[0] = pws[gid].i[ 0];
-  wordl0[1] = pws[gid].i[ 1];
-  wordl0[2] = pws[gid].i[ 2];
-  wordl0[3] = pws[gid].i[ 3];
-
-  u32 wordl1[4];
-
-  wordl1[0] = pws[gid].i[ 4];
-  wordl1[1] = pws[gid].i[ 5];
-  wordl1[2] = pws[gid].i[ 6];
-  wordl1[3] = pws[gid].i[ 7];
-
-  u32 wordl2[4];
+  u32 pw_buf0[4];
+  u32 pw_buf1[4];
 
-  wordl2[0] = 0;
-  wordl2[1] = 0;
-  wordl2[2] = 0;
-  wordl2[3] = 0;
+  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];
 
-  u32 wordl3[4];
+  const u32 pw_l_len = pws[gid].pw_len;
 
-  wordl3[0] = 0;
-  wordl3[1] = 0;
-  wordl3[2] = 0;
-  wordl3[3] = 0;
+  /**
+   * shared
+   */
 
-  const u32 pw_l_len = pws[gid].pw_len;
+  __local RC4_KEY rc4_keys[64];
 
-  if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
-  {
-    switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
-  }
+  __local RC4_KEY *rc4_key = &rc4_keys[lid];
 
   /**
    * esalt
@@ -309,68 +291,81 @@ __kernel void m09710_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    * loop
    */
 
-  for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
+  for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
   {
-    const u32 pw_r_len = combs_buf[il_pos].pw_len;
-
-    const u32 pw_len = pw_l_len + pw_r_len;
-
-    u32 wordr0[4];
-
-    wordr0[0] = combs_buf[il_pos].i[0];
-    wordr0[1] = combs_buf[il_pos].i[1];
-    wordr0[2] = combs_buf[il_pos].i[2];
-    wordr0[3] = combs_buf[il_pos].i[3];
-
-    u32 wordr1[4];
-
-    wordr1[0] = combs_buf[il_pos].i[4];
-    wordr1[1] = combs_buf[il_pos].i[5];
-    wordr1[2] = combs_buf[il_pos].i[6];
-    wordr1[3] = combs_buf[il_pos].i[7];
-
-    u32 wordr2[4];
-
-    wordr2[0] = 0;
-    wordr2[1] = 0;
-    wordr2[2] = 0;
-    wordr2[3] = 0;
-
-    u32 wordr3[4];
-
-    wordr3[0] = 0;
-    wordr3[1] = 0;
-    wordr3[2] = 0;
-    wordr3[3] = 0;
+    const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos);
+
+    const u32x pw_len = pw_l_len + pw_r_len;
+
+    /**
+     * concat password candidate
+     */
+
+    u32x wordl0[4] = { 0 };
+    u32x wordl1[4] = { 0 };
+    u32x wordl2[4] = { 0 };
+    u32x wordl3[4] = { 0 };
+
+    wordl0[0] = pw_buf0[0];
+    wordl0[1] = pw_buf0[1];
+    wordl0[2] = pw_buf0[2];
+    wordl0[3] = pw_buf0[3];
+    wordl1[0] = pw_buf1[0];
+    wordl1[1] = pw_buf1[1];
+    wordl1[2] = pw_buf1[2];
+    wordl1[3] = pw_buf1[3];
+
+    u32x wordr0[4] = { 0 };
+    u32x wordr1[4] = { 0 };
+    u32x wordr2[4] = { 0 };
+    u32x wordr3[4] = { 0 };
+
+    wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
+    wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
+    wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
+    wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
+    wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
+    wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
+    wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
+    wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);
 
     if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
     {
-      switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
+      switch_buffer_by_offset_le_VV (wordr0, wordr1, wordr2, wordr3, pw_l_len);
+    }
+    else
+    {
+      switch_buffer_by_offset_le_VV (wordl0, wordl1, wordl2, wordl3, pw_r_len);
     }
 
-    // first md5 to generate RC4 128 bit key
-
-    u32 w0_t[4];
-    u32 w1_t[4];
-    u32 w2_t[4];
-    u32 w3_t[4];
-
-    w0_t[0]  =  wordl0[0] | wordr0[0];
-    w0_t[1]  = (wordl0[1] | wordr0[1]) & 0xff;
-    w0_t[2]  = 0x8000;
-    w0_t[3]  = 0;
-    w1_t[0]  = 0;
-    w1_t[1]  = 0;
-    w1_t[2]  = 0;
-    w1_t[3]  = 0;
-    w2_t[0]  = 0;
-    w2_t[1]  = 0;
-    w2_t[2]  = 0;
-    w2_t[3]  = 0;
-    w3_t[0]  = 0;
-    w3_t[1]  = 0;
-    w3_t[2]  = 9 * 8;
-    w3_t[3]  = 0;
+    u32x w0[4];
+    u32x w1[4];
+    u32x w2[4];
+    u32x w3[4];
+
+    w0[0] = wordl0[0] | wordr0[0];
+    w0[1] = wordl0[1] | wordr0[1];
+
+    /**
+     * md5
+     */
+
+    w0[0]  = w0[0];
+    w0[1]  = w0[1] & 0xff;
+    w0[2]  = 0x8000;
+    w0[3]  = 0;
+    w1[0]  = 0;
+    w1[1]  = 0;
+    w1[2]  = 0;
+    w1[3]  = 0;
+    w2[0]  = 0;
+    w2[1]  = 0;
+    w2[2]  = 0;
+    w2[3]  = 0;
+    w3[0]  = 0;
+    w3[1]  = 0;
+    w3[2]  = 9 * 8;
+    w3[3]  = 0;
 
     u32 digest[4];
 
@@ -379,67 +374,55 @@ __kernel void m09710_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     digest[2] = MD5M_C;
     digest[3] = MD5M_D;
 
-    md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
+    md5_transform (w0, w1, w2, w3, digest);
 
     // now the RC4 part
 
-    u32 key[4];
-
-    key[0] = digest[0];
-    key[1] = digest[1];
-    key[2] = digest[2];
-    key[3] = digest[3];
-
-    rc4_init_16 (rc4_key, key);
+    rc4_init_16 (rc4_key, digest);
 
     u32 out[4];
 
     u8 j = rc4_next_16 (rc4_key, 0, 0, encryptedVerifier, out);
 
-    w0_t[0] = out[0];
-    w0_t[1] = out[1];
-    w0_t[2] = out[2];
-    w0_t[3] = out[3];
-    w1_t[0] = 0x80;
-    w1_t[1] = 0;
-    w1_t[2] = 0;
-    w1_t[3] = 0;
-    w2_t[0] = 0;
-    w2_t[1] = 0;
-    w2_t[2] = 0;
-    w2_t[3] = 0;
-    w3_t[0] = 0;
-    w3_t[1] = 0;
-    w3_t[2] = 16 * 8;
-    w3_t[3] = 0;
+    w0[0] = out[0];
+    w0[1] = out[1];
+    w0[2] = out[2];
+    w0[3] = out[3];
+    w1[0] = 0x80;
+    w1[1] = 0;
+    w1[2] = 0;
+    w1[3] = 0;
+    w2[0] = 0;
+    w2[1] = 0;
+    w2[2] = 0;
+    w2[3] = 0;
+    w3[0] = 0;
+    w3[1] = 0;
+    w3[2] = 16 * 8;
+    w3[3] = 0;
 
     digest[0] = MD5M_A;
     digest[1] = MD5M_B;
     digest[2] = MD5M_C;
     digest[3] = MD5M_D;
 
-    md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
+    md5_transform (w0, w1, w2, w3, digest);
 
     rc4_next_16 (rc4_key, 16, j, digest, out);
 
-    const u32 r0 = out[0];
-    const u32 r1 = out[1];
-    const u32 r2 = out[2];
-    const u32 r3 = out[3];
-
-    #include COMPARE_M
+    COMPARE_M_SIMD (out[0], out[1], out[2], out[3]);
   }
 }
 
-__kernel void m09710_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 oldoffice01_t *oldoffice01_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 m09710_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 oldoffice01_t *oldoffice01_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 m09710_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 oldoffice01_t *oldoffice01_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 m09710_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 oldoffice01_t *oldoffice01_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 m09710_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 oldoffice01_t *oldoffice01_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 m09710_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 oldoffice01_t *oldoffice01_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
@@ -447,10 +430,6 @@ __kernel void m09710_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 
   const u32 lid = get_local_id (0);
 
-  __local RC4_KEY rc4_keys[64];
-
-  __local RC4_KEY *rc4_key = &rc4_keys[lid];
-
   /**
    * base
    */
@@ -459,52 +438,27 @@ __kernel void m09710_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 
   if (gid >= gid_max) return;
 
-  u32 wordl0[4];
-
-  wordl0[0] = pws[gid].i[ 0];
-  wordl0[1] = pws[gid].i[ 1];
-  wordl0[2] = pws[gid].i[ 2];
-  wordl0[3] = pws[gid].i[ 3];
-
-  u32 wordl1[4];
-
-  wordl1[0] = pws[gid].i[ 4];
-  wordl1[1] = pws[gid].i[ 5];
-  wordl1[2] = pws[gid].i[ 6];
-  wordl1[3] = pws[gid].i[ 7];
-
-  u32 wordl2[4];
+  u32 pw_buf0[4];
+  u32 pw_buf1[4];
 
-  wordl2[0] = 0;
-  wordl2[1] = 0;
-  wordl2[2] = 0;
-  wordl2[3] = 0;
-
-  u32 wordl3[4];
-
-  wordl3[0] = 0;
-  wordl3[1] = 0;
-  wordl3[2] = 0;
-  wordl3[3] = 0;
+  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_l_len = pws[gid].pw_len;
 
-  if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
-  {
-    switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
-  }
-
   /**
-   * digest
+   * shared
    */
 
-  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]
-  };
+  __local RC4_KEY rc4_keys[64];
+
+  __local RC4_KEY *rc4_key = &rc4_keys[lid];
 
   /**
    * esalt
@@ -520,71 +474,96 @@ __kernel void m09710_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
   encryptedVerifier[3] = oldoffice01_bufs[salt_pos].encryptedVerifier[3];
 
   /**
-   * loop
+   * digest
    */
 
-  for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
+  const u32 search[4] =
   {
-    const u32 pw_r_len = combs_buf[il_pos].pw_len;
-
-    const u32 pw_len = pw_l_len + pw_r_len;
-
-    u32 wordr0[4];
-
-    wordr0[0] = combs_buf[il_pos].i[0];
-    wordr0[1] = combs_buf[il_pos].i[1];
-    wordr0[2] = combs_buf[il_pos].i[2];
-    wordr0[3] = combs_buf[il_pos].i[3];
-
-    u32 wordr1[4];
-
-    wordr1[0] = combs_buf[il_pos].i[4];
-    wordr1[1] = combs_buf[il_pos].i[5];
-    wordr1[2] = combs_buf[il_pos].i[6];
-    wordr1[3] = combs_buf[il_pos].i[7];
-
-    u32 wordr2[4];
-
-    wordr2[0] = 0;
-    wordr2[1] = 0;
-    wordr2[2] = 0;
-    wordr2[3] = 0;
+    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]
+  };
 
-    u32 wordr3[4];
+  /**
+   * loop
+   */
 
-    wordr3[0] = 0;
-    wordr3[1] = 0;
-    wordr3[2] = 0;
-    wordr3[3] = 0;
+  for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
+  {
+    const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos);
+
+    const u32x pw_len = pw_l_len + pw_r_len;
+
+    /**
+     * concat password candidate
+     */
+
+    u32x wordl0[4] = { 0 };
+    u32x wordl1[4] = { 0 };
+    u32x wordl2[4] = { 0 };
+    u32x wordl3[4] = { 0 };
+
+    wordl0[0] = pw_buf0[0];
+    wordl0[1] = pw_buf0[1];
+    wordl0[2] = pw_buf0[2];
+    wordl0[3] = pw_buf0[3];
+    wordl1[0] = pw_buf1[0];
+    wordl1[1] = pw_buf1[1];
+    wordl1[2] = pw_buf1[2];
+    wordl1[3] = pw_buf1[3];
+
+    u32x wordr0[4] = { 0 };
+    u32x wordr1[4] = { 0 };
+    u32x wordr2[4] = { 0 };
+    u32x wordr3[4] = { 0 };
+
+    wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
+    wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
+    wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
+    wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
+    wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
+    wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
+    wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
+    wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);
 
     if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
     {
-      switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
+      switch_buffer_by_offset_le_VV (wordr0, wordr1, wordr2, wordr3, pw_l_len);
+    }
+    else
+    {
+      switch_buffer_by_offset_le_VV (wordl0, wordl1, wordl2, wordl3, pw_r_len);
     }
 
-    // first md5 to generate RC4 128 bit key
-
-    u32 w0_t[4];
-    u32 w1_t[4];
-    u32 w2_t[4];
-    u32 w3_t[4];
-
-    w0_t[0]  =  wordl0[0] | wordr0[0];
-    w0_t[1]  = (wordl0[1] | wordr0[1]) & 0xff;
-    w0_t[2]  = 0x8000;
-    w0_t[3]  = 0;
-    w1_t[0]  = 0;
-    w1_t[1]  = 0;
-    w1_t[2]  = 0;
-    w1_t[3]  = 0;
-    w2_t[0]  = 0;
-    w2_t[1]  = 0;
-    w2_t[2]  = 0;
-    w2_t[3]  = 0;
-    w3_t[0]  = 0;
-    w3_t[1]  = 0;
-    w3_t[2]  = 9 * 8;
-    w3_t[3]  = 0;
+    u32x w0[4];
+    u32x w1[4];
+    u32x w2[4];
+    u32x w3[4];
+
+    w0[0] = wordl0[0] | wordr0[0];
+    w0[1] = wordl0[1] | wordr0[1];
+
+    /**
+     * md5
+     */
+
+    w0[0]  = w0[0];
+    w0[1]  = w0[1] & 0xff;
+    w0[2]  = 0x8000;
+    w0[3]  = 0;
+    w1[0]  = 0;
+    w1[1]  = 0;
+    w1[2]  = 0;
+    w1[3]  = 0;
+    w2[0]  = 0;
+    w2[1]  = 0;
+    w2[2]  = 0;
+    w2[3]  = 0;
+    w3[0]  = 0;
+    w3[1]  = 0;
+    w3[2]  = 9 * 8;
+    w3[3]  = 0;
 
     u32 digest[4];
 
@@ -593,62 +572,50 @@ __kernel void m09710_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     digest[2] = MD5M_C;
     digest[3] = MD5M_D;
 
-    md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
+    md5_transform (w0, w1, w2, w3, digest);
 
     // now the RC4 part
 
-    u32 key[4];
-
-    key[0] = digest[0];
-    key[1] = digest[1];
-    key[2] = digest[2];
-    key[3] = digest[3];
-
-    rc4_init_16 (rc4_key, key);
+    rc4_init_16 (rc4_key, digest);
 
     u32 out[4];
 
     u8 j = rc4_next_16 (rc4_key, 0, 0, encryptedVerifier, out);
 
-    w0_t[0] = out[0];
-    w0_t[1] = out[1];
-    w0_t[2] = out[2];
-    w0_t[3] = out[3];
-    w1_t[0] = 0x80;
-    w1_t[1] = 0;
-    w1_t[2] = 0;
-    w1_t[3] = 0;
-    w2_t[0] = 0;
-    w2_t[1] = 0;
-    w2_t[2] = 0;
-    w2_t[3] = 0;
-    w3_t[0] = 0;
-    w3_t[1] = 0;
-    w3_t[2] = 16 * 8;
-    w3_t[3] = 0;
+    w0[0] = out[0];
+    w0[1] = out[1];
+    w0[2] = out[2];
+    w0[3] = out[3];
+    w1[0] = 0x80;
+    w1[1] = 0;
+    w1[2] = 0;
+    w1[3] = 0;
+    w2[0] = 0;
+    w2[1] = 0;
+    w2[2] = 0;
+    w2[3] = 0;
+    w3[0] = 0;
+    w3[1] = 0;
+    w3[2] = 16 * 8;
+    w3[3] = 0;
 
     digest[0] = MD5M_A;
     digest[1] = MD5M_B;
     digest[2] = MD5M_C;
     digest[3] = MD5M_D;
 
-    md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
+    md5_transform (w0, w1, w2, w3, digest);
 
     rc4_next_16 (rc4_key, 16, j, digest, out);
 
-    const u32 r0 = out[0];
-    const u32 r1 = out[1];
-    const u32 r2 = out[2];
-    const u32 r3 = out[3];
-
-    #include COMPARE_S
+    COMPARE_S_SIMD (out[0], out[1], out[2], out[3]);
   }
 }
 
-__kernel void m09710_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 oldoffice01_t *oldoffice01_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 m09710_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 oldoffice01_t *oldoffice01_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 m09710_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 oldoffice01_t *oldoffice01_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 m09710_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 oldoffice01_t *oldoffice01_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)
 {
 }