From: Jens Steube <jens.steube@gmail.com>
Date: Mon, 22 Feb 2016 09:01:53 +0000 (+0100)
Subject: New SIMD code for -a 1 -m 40
X-Git-Tag: v3.00-beta~252
X-Git-Url: https://www.flypig.org.uk/git/?a=commitdiff_plain;h=d6c6af040b0c2572c3dad6268f92c3c7b632dcb7;p=hashcat.git

New SIMD code for -a 1 -m 40
---

diff --git a/OpenCL/m00040_a1.cl b/OpenCL/m00040_a1.cl
index 4812eae..f704c55 100644
--- a/OpenCL/m00040_a1.cl
+++ b/OpenCL/m00040_a1.cl
@@ -5,6 +5,8 @@
 
 #define _MD5_
 
+#define NEW_SIMD_CODE
+
 #include "include/constants.h"
 #include "include/kernel_vendor.h"
 
@@ -16,9 +18,7 @@
 #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 "OpenCL/simd.c"
 
 // no unicode yet
 
@@ -38,54 +38,31 @@ __kernel void m00040_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 pws0[4] = { 0 };
+  u32 pws1[4] = { 0 };
 
-  u32 wordl2[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;
+  pws0[0] = pws[gid].i[0];
+  pws0[1] = pws[gid].i[1];
+  pws0[2] = pws[gid].i[2];
+  pws0[3] = pws[gid].i[3];
+  pws1[0] = pws[gid].i[4];
+  pws1[1] = pws[gid].i[5];
+  pws1[2] = pws[gid].i[6];
+  pws1[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_le (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
-  }
-
   /**
    * salt
    */
 
   u32 salt_buf0[4];
+  u32 salt_buf1[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];
@@ -97,55 +74,57 @@ __kernel void m00040_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 < combs_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];
+    const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos);
 
-    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];
+    const u32x pw_len = pw_l_len + pw_r_len;
 
-    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];
+    /**
+     * concat password candidate
+     */
 
-    wordr3[0] = 0;
-    wordr3[1] = 0;
-    wordr3[2] = 0;
-    wordr3[3] = 0;
+    u32x wordl0[4] = { 0 };
+    u32x wordl1[4] = { 0 };
+    u32x wordl2[4] = { 0 };
+    u32x wordl3[4] = { 0 };
+
+    wordl0[0] = pws0[0];
+    wordl0[1] = pws0[1];
+    wordl0[2] = pws0[2];
+    wordl0[3] = pws0[3];
+    wordl1[0] = pws1[0];
+    wordl1[1] = pws1[1];
+    wordl1[2] = pws1[2];
+    wordl1[3] = pws1[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_le (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);
     }
 
-    /**
-     * prepend salt
-     */
-
-    const u32 pw_salt_len = (pw_len * 2) + salt_len;
-
-    u32 w0[4];
-    u32 w1[4];
-    u32 w2[4];
-    u32 w3[4];
+    u32x w0[4];
+    u32x w1[4];
+    u32x w2[4];
+    u32x w3[4];
 
     w0[0] = wordl0[0] | wordr0[0];
     w0[1] = wordl0[1] | wordr0[1];
@@ -154,7 +133,7 @@ __kernel void m00040_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     w1[0] = wordl1[0] | wordr1[0];
     w1[1] = wordl1[1] | wordr1[1];
     w1[2] = wordl1[2] | wordr1[2];
-    w1[3] = wordl1[3] | wordr1[3];
+    w1[3] = 0;
     w2[0] = 0;
     w2[1] = 0;
     w2[2] = 0;
@@ -164,113 +143,114 @@ __kernel void m00040_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     w3[2] = 0;
     w3[3] = 0;
 
-    u32 w0_t[4];
-    u32 w1_t[4];
-    u32 w2_t[4];
-    u32 w3_t[4];
+    make_unicode (w1, w2, w3);
+    make_unicode (w0, w0, w1);
 
-    make_unicode (w0, w0_t, w1_t);
-    make_unicode (w1, w2_t, w3_t);
+    const u32x pw_len2 = pw_len * 2;
 
-    switch_buffer_by_offset_le (w0_t, w1_t, w2_t, w3_t, salt_len);
-
-    w0_t[0] |= salt_buf0[0];
-    w0_t[1] |= salt_buf0[1];
-    w0_t[2] |= salt_buf0[2];
-    w0_t[3] |= salt_buf0[3];
-    w1_t[0] |= salt_buf1[0];
-    w1_t[1] |= salt_buf1[1];
-    w1_t[2] |= salt_buf1[2];
-    w1_t[3] |= salt_buf1[3];
-
-    append_0x80_4x4 (w0_t, w1_t, w2_t, w3_t, pw_salt_len);
+    /**
+     * prepend salt
+     */
 
-    w3_t[2] = pw_salt_len * 8;
+    switch_buffer_by_offset_le (w0, w1, w2, w3, salt_len);
+
+    const u32x pw_salt_len = pw_len2 + 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] |= 0;
+    w2[1] |= 0;
+    w2[2] |= 0;
+    w2[3] |= 0;
+    w3[0] |= 0;
+    w3[1] |= 0;
+    w3[2]  = pw_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);
   }
 }
 
@@ -298,54 +278,31 @@ __kernel void m00040_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 pws0[4] = { 0 };
+  u32 pws1[4] = { 0 };
 
-  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;
+  pws0[0] = pws[gid].i[0];
+  pws0[1] = pws[gid].i[1];
+  pws0[2] = pws[gid].i[2];
+  pws0[3] = pws[gid].i[3];
+  pws1[0] = pws[gid].i[4];
+  pws1[1] = pws[gid].i[5];
+  pws1[2] = pws[gid].i[6];
+  pws1[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_le (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
-  }
-
   /**
    * salt
    */
 
   u32 salt_buf0[4];
+  u32 salt_buf1[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];
@@ -369,55 +326,57 @@ __kernel void m00040_s04 (__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 < combs_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];
+    const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos);
 
-    u32 wordr1[4];
+    const u32x pw_len = pw_l_len + pw_r_len;
 
-    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];
+    /**
+     * concat password candidate
+     */
 
-    wordr3[0] = 0;
-    wordr3[1] = 0;
-    wordr3[2] = 0;
-    wordr3[3] = 0;
+    u32x wordl0[4] = { 0 };
+    u32x wordl1[4] = { 0 };
+    u32x wordl2[4] = { 0 };
+    u32x wordl3[4] = { 0 };
+
+    wordl0[0] = pws0[0];
+    wordl0[1] = pws0[1];
+    wordl0[2] = pws0[2];
+    wordl0[3] = pws0[3];
+    wordl1[0] = pws1[0];
+    wordl1[1] = pws1[1];
+    wordl1[2] = pws1[2];
+    wordl1[3] = pws1[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_le (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);
     }
 
-    /**
-     * prepend salt
-     */
-
-    const u32 pw_salt_len = (pw_len * 2) + salt_len;
-
-    u32 w0[4];
-    u32 w1[4];
-    u32 w2[4];
-    u32 w3[4];
+    u32x w0[4];
+    u32x w1[4];
+    u32x w2[4];
+    u32x w3[4];
 
     w0[0] = wordl0[0] | wordr0[0];
     w0[1] = wordl0[1] | wordr0[1];
@@ -426,7 +385,7 @@ __kernel void m00040_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     w1[0] = wordl1[0] | wordr1[0];
     w1[1] = wordl1[1] | wordr1[1];
     w1[2] = wordl1[2] | wordr1[2];
-    w1[3] = wordl1[3] | wordr1[3];
+    w1[3] = 0;
     w2[0] = 0;
     w2[1] = 0;
     w2[2] = 0;
@@ -436,118 +395,117 @@ __kernel void m00040_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     w3[2] = 0;
     w3[3] = 0;
 
-    u32 w0_t[4];
-    u32 w1_t[4];
-    u32 w2_t[4];
-    u32 w3_t[4];
-
-    make_unicode (w0, w0_t, w1_t);
-    make_unicode (w1, w2_t, w3_t);
+    make_unicode (w1, w2, w3);
+    make_unicode (w0, w0, w1);
 
-    switch_buffer_by_offset_le (w0_t, w1_t, w2_t, w3_t, salt_len);
+    const u32x pw_len2 = pw_len * 2;
 
-    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];
-
-    append_0x80_4x4 (w0_t, w1_t, w2_t, w3_t, pw_salt_len);
+    /**
+     * prepend salt
+     */
 
-    w3_t[2] = pw_salt_len * 8;
+    switch_buffer_by_offset_le (w0, w1, w2, w3, salt_len);
+
+    const u32x pw_salt_len = pw_len2 + 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] |= 0;
+    w2[1] |= 0;
+    w2[2] |= 0;
+    w2[3] |= 0;
+    w3[0] |= 0;
+    w3[1] |= 0;
+    w3[2]  = pw_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);
-
-    bool q_cond = allx (search[0] != a);
-
-    if (q_cond) continue;
-
-    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);
   }
 }