From cd0e287827bf02e5743b698d4962f1eec8da83a6 Mon Sep 17 00:00:00 2001
From: Jens Steube <jens.steube@gmail.com>
Date: Fri, 4 Mar 2016 14:49:44 +0100
Subject: [PATCH] Cleanup -m 1740 kernels to latest standard

---
 OpenCL/m01740_a0.cl | 252 ++++++++++++----------
 OpenCL/m01740_a1.cl | 512 +++++++++++++++++++++++---------------------
 OpenCL/m01740_a3.cl | 371 +++++++++++++-------------------
 3 files changed, 549 insertions(+), 586 deletions(-)

diff --git a/OpenCL/m01740_a0.cl b/OpenCL/m01740_a0.cl
index b41e625a8..9a6b68504 100644
--- a/OpenCL/m01740_a0.cl
+++ b/OpenCL/m01740_a0.cl
@@ -160,18 +160,16 @@ __kernel void m01740_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
   if (gid >= gid_max) return;
 
   u32 pw_buf0[4];
-
-  pw_buf0[0] = pws[gid].i[ 0];
-  pw_buf0[1] = pws[gid].i[ 1];
-  pw_buf0[2] = pws[gid].i[ 2];
-  pw_buf0[3] = pws[gid].i[ 3];
-
   u32 pw_buf1[4];
 
-  pw_buf1[0] = pws[gid].i[ 4];
-  pw_buf1[1] = pws[gid].i[ 5];
-  pw_buf1[2] = pws[gid].i[ 6];
-  pw_buf1[3] = pws[gid].i[ 7];
+  pw_buf0[0] = pws[gid].i[0];
+  pw_buf0[1] = pws[gid].i[1];
+  pw_buf0[2] = pws[gid].i[2];
+  pw_buf0[3] = pws[gid].i[3];
+  pw_buf1[0] = pws[gid].i[4];
+  pw_buf1[1] = pws[gid].i[5];
+  pw_buf1[2] = pws[gid].i[6];
+  pw_buf1[3] = pws[gid].i[7];
 
   const u32 pw_len = pws[gid].pw_len;
 
@@ -180,18 +178,26 @@ __kernel void m01740_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    */
 
   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];
+  u32 salt_buf2[4];
+  u32 salt_buf3[4];
 
-  salt_buf1[0] = salt_bufs[salt_pos].salt_buf[4];
-  salt_buf1[1] = salt_bufs[salt_pos].salt_buf[5];
-  salt_buf1[2] = salt_bufs[salt_pos].salt_buf[6];
-  salt_buf1[3] = salt_bufs[salt_pos].salt_buf[7];
+  salt_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];
+  salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
+  salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
+  salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
+  salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
+  salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8];
+  salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9];
+  salt_buf2[2] = salt_bufs[salt_pos].salt_buf[10];
+  salt_buf2[3] = salt_bufs[salt_pos].salt_buf[11];
+  salt_buf3[0] = salt_bufs[salt_pos].salt_buf[12];
+  salt_buf3[1] = salt_bufs[salt_pos].salt_buf[13];
+  salt_buf3[2] = salt_bufs[salt_pos].salt_buf[14];
+  salt_buf3[3] = salt_bufs[salt_pos].salt_buf[15];
 
   const u32 salt_len = salt_bufs[salt_pos].salt_len;
 
@@ -208,51 +214,61 @@ __kernel void m01740_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 
     const u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
 
+    make_unicode (w1, w2, w3);
+    make_unicode (w0, w0, w1);
+
+    const u32x out_len2 = out_len * 2;
+
     /**
      * prepend salt
      */
 
-    const u32x out_salt_len = (out_len * 2) + salt_len;
+    const u32x out_salt_len = out_len2 + salt_len;
+
+    switch_buffer_by_offset_le_VV (w0, w1, w2, w3, salt_len);
+
+    w0[0] |= salt_buf0[0];
+    w0[1] |= salt_buf0[1];
+    w0[2] |= salt_buf0[2];
+    w0[3] |= salt_buf0[3];
+    w1[0] |= salt_buf1[0];
+    w1[1] |= salt_buf1[1];
+    w1[2] |= salt_buf1[2];
+    w1[3] |= salt_buf1[3];
+    w2[0] |= salt_buf2[0];
+    w2[1] |= salt_buf2[1];
+    w2[2] |= salt_buf2[2];
+    w2[3] |= salt_buf2[3];
+    w3[0] |= salt_buf3[0];
+    w3[1] |= salt_buf3[1];
+    w3[2] |= salt_buf3[2];
+    w3[3] |= salt_buf3[3];
+
+    append_0x80_4x4_VV (w0, w1, w2, w3, out_salt_len);
+
+    /**
+     * sha512
+     */
 
     u32x w0_t[4];
     u32x w1_t[4];
     u32x w2_t[4];
     u32x w3_t[4];
 
-    make_unicode (w0, w0_t, w1_t);
-    make_unicode (w1, w2_t, w3_t);
-
-    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_VV (w0_t, w1_t, w2_t, w3_t, out_salt_len);
-
-    /**
-     * sha512
-     */
-
-    w0_t[0] = swap32 (w0_t[0]);
-    w0_t[1] = swap32 (w0_t[1]);
-    w0_t[2] = swap32 (w0_t[2]);
-    w0_t[3] = swap32 (w0_t[3]);
-    w1_t[0] = swap32 (w1_t[0]);
-    w1_t[1] = swap32 (w1_t[1]);
-    w1_t[2] = swap32 (w1_t[2]);
-    w1_t[3] = swap32 (w1_t[3]);
-    w2_t[0] = swap32 (w2_t[0]);
-    w2_t[1] = swap32 (w2_t[1]);
-    w2_t[2] = swap32 (w2_t[2]);
-    w2_t[3] = swap32 (w2_t[3]);
-    w3_t[0] = swap32 (w3_t[0]);
-    w3_t[1] = swap32 (w3_t[1]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
+    w2_t[0] = swap32 (w2[0]);
+    w2_t[1] = swap32 (w2[1]);
+    w2_t[2] = swap32 (w2[2]);
+    w2_t[3] = swap32 (w2[3]);
+    w3_t[0] = swap32 (w3[0]);
+    w3_t[1] = swap32 (w3[1]);
     w3_t[2] = 0;
     w3_t[3] = out_salt_len * 8;
 
@@ -303,18 +319,16 @@ __kernel void m01740_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
   if (gid >= gid_max) return;
 
   u32 pw_buf0[4];
-
-  pw_buf0[0] = pws[gid].i[ 0];
-  pw_buf0[1] = pws[gid].i[ 1];
-  pw_buf0[2] = pws[gid].i[ 2];
-  pw_buf0[3] = pws[gid].i[ 3];
-
   u32 pw_buf1[4];
 
-  pw_buf1[0] = pws[gid].i[ 4];
-  pw_buf1[1] = pws[gid].i[ 5];
-  pw_buf1[2] = pws[gid].i[ 6];
-  pw_buf1[3] = pws[gid].i[ 7];
+  pw_buf0[0] = pws[gid].i[0];
+  pw_buf0[1] = pws[gid].i[1];
+  pw_buf0[2] = pws[gid].i[2];
+  pw_buf0[3] = pws[gid].i[3];
+  pw_buf1[0] = pws[gid].i[4];
+  pw_buf1[1] = pws[gid].i[5];
+  pw_buf1[2] = pws[gid].i[6];
+  pw_buf1[3] = pws[gid].i[7];
 
   const u32 pw_len = pws[gid].pw_len;
 
@@ -323,18 +337,26 @@ __kernel void m01740_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    */
 
   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];
+  u32 salt_buf2[4];
+  u32 salt_buf3[4];
 
-  salt_buf1[0] = salt_bufs[salt_pos].salt_buf[4];
-  salt_buf1[1] = salt_bufs[salt_pos].salt_buf[5];
-  salt_buf1[2] = salt_bufs[salt_pos].salt_buf[6];
-  salt_buf1[3] = salt_bufs[salt_pos].salt_buf[7];
+  salt_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];
+  salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
+  salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
+  salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
+  salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
+  salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8];
+  salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9];
+  salt_buf2[2] = salt_bufs[salt_pos].salt_buf[10];
+  salt_buf2[3] = salt_bufs[salt_pos].salt_buf[11];
+  salt_buf3[0] = salt_bufs[salt_pos].salt_buf[12];
+  salt_buf3[1] = salt_bufs[salt_pos].salt_buf[13];
+  salt_buf3[2] = salt_bufs[salt_pos].salt_buf[14];
+  salt_buf3[3] = salt_bufs[salt_pos].salt_buf[15];
 
   const u32 salt_len = salt_bufs[salt_pos].salt_len;
 
@@ -363,51 +385,61 @@ __kernel void m01740_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 
     const u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
 
+    make_unicode (w1, w2, w3);
+    make_unicode (w0, w0, w1);
+
+    const u32x out_len2 = out_len * 2;
+
     /**
      * prepend salt
      */
 
-    const u32x out_salt_len = (out_len * 2) + salt_len;
+    const u32x out_salt_len = out_len2 + salt_len;
+
+    switch_buffer_by_offset_le_VV (w0, w1, w2, w3, salt_len);
+
+    w0[0] |= salt_buf0[0];
+    w0[1] |= salt_buf0[1];
+    w0[2] |= salt_buf0[2];
+    w0[3] |= salt_buf0[3];
+    w1[0] |= salt_buf1[0];
+    w1[1] |= salt_buf1[1];
+    w1[2] |= salt_buf1[2];
+    w1[3] |= salt_buf1[3];
+    w2[0] |= salt_buf2[0];
+    w2[1] |= salt_buf2[1];
+    w2[2] |= salt_buf2[2];
+    w2[3] |= salt_buf2[3];
+    w3[0] |= salt_buf3[0];
+    w3[1] |= salt_buf3[1];
+    w3[2] |= salt_buf3[2];
+    w3[3] |= salt_buf3[3];
+
+    append_0x80_4x4_VV (w0, w1, w2, w3, out_salt_len);
+
+    /**
+     * sha512
+     */
 
     u32x w0_t[4];
     u32x w1_t[4];
     u32x w2_t[4];
     u32x w3_t[4];
 
-    make_unicode (w0, w0_t, w1_t);
-    make_unicode (w1, w2_t, w3_t);
-
-    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_VV (w0_t, w1_t, w2_t, w3_t, out_salt_len);
-
-    /**
-     * sha512
-     */
-
-    w0_t[0] = swap32 (w0_t[0]);
-    w0_t[1] = swap32 (w0_t[1]);
-    w0_t[2] = swap32 (w0_t[2]);
-    w0_t[3] = swap32 (w0_t[3]);
-    w1_t[0] = swap32 (w1_t[0]);
-    w1_t[1] = swap32 (w1_t[1]);
-    w1_t[2] = swap32 (w1_t[2]);
-    w1_t[3] = swap32 (w1_t[3]);
-    w2_t[0] = swap32 (w2_t[0]);
-    w2_t[1] = swap32 (w2_t[1]);
-    w2_t[2] = swap32 (w2_t[2]);
-    w2_t[3] = swap32 (w2_t[3]);
-    w3_t[0] = swap32 (w3_t[0]);
-    w3_t[1] = swap32 (w3_t[1]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
+    w2_t[0] = swap32 (w2[0]);
+    w2_t[1] = swap32 (w2[1]);
+    w2_t[2] = swap32 (w2[2]);
+    w2_t[3] = swap32 (w2[3]);
+    w3_t[0] = swap32 (w3[0]);
+    w3_t[1] = swap32 (w3[1]);
     w3_t[2] = 0;
     w3_t[3] = out_salt_len * 8;
 
diff --git a/OpenCL/m01740_a1.cl b/OpenCL/m01740_a1.cl
index c1b980935..2e8b6a972 100644
--- a/OpenCL/m01740_a1.cl
+++ b/OpenCL/m01740_a1.cl
@@ -5,6 +5,8 @@
 
 #define _SHA512_
 
+#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"
 
 __constant u64 k_sha512[80] =
 {
@@ -44,33 +44,33 @@ __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])
+static 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]);
+  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]);
 
-  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 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()                            \
   {                                                 \
@@ -157,58 +157,45 @@ __kernel void m01740_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 
   if (gid >= gid_max) return;
 
-  u32 wordl0[4];
+  u32 pw_buf0[4];
+  u32 pw_buf1[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];
-
-  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_le (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
-  }
-
   /**
    * salt
    */
 
   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];
+  u32 salt_buf2[4];
+  u32 salt_buf3[4];
 
-  salt_buf1[0] = salt_bufs[salt_pos].salt_buf[4];
-  salt_buf1[1] = salt_bufs[salt_pos].salt_buf[5];
-  salt_buf1[2] = salt_bufs[salt_pos].salt_buf[6];
-  salt_buf1[3] = salt_bufs[salt_pos].salt_buf[7];
+  salt_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];
+  salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
+  salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
+  salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
+  salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
+  salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8];
+  salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9];
+  salt_buf2[2] = salt_bufs[salt_pos].salt_buf[10];
+  salt_buf2[3] = salt_bufs[salt_pos].salt_buf[11];
+  salt_buf3[0] = salt_bufs[salt_pos].salt_buf[12];
+  salt_buf3[1] = salt_bufs[salt_pos].salt_buf[13];
+  salt_buf3[2] = salt_bufs[salt_pos].salt_buf[14];
+  salt_buf3[3] = salt_bufs[salt_pos].salt_buf[15];
 
   const u32 salt_len = salt_bufs[salt_pos].salt_len;
 
@@ -216,43 +203,57 @@ __kernel void m01740_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    * loop
    */
 
-  for (u32 il_pos = 0; il_pos < il_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 u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos);
 
-    const u32 pw_len = pw_l_len + pw_r_len;
+    const u32x pw_len = pw_l_len + pw_r_len;
 
-    u32 wordr0[4];
-    u32 wordr1[4];
-    u32 wordr2[4];
-    u32 wordr3[4];
+    /**
+     * concat password candidate
+     */
 
-    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];
-    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];
-    wordr2[0] = 0;
-    wordr2[1] = 0;
-    wordr2[2] = 0;
-    wordr2[3] = 0;
-    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] = 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_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);
     }
 
-    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];
@@ -271,55 +272,63 @@ __kernel void m01740_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     w3[2] = wordl3[2] | wordr3[2];
     w3[3] = wordl3[3] | wordr3[3];
 
+    make_unicode (w1, w2, w3);
+    make_unicode (w0, w0, w1);
+
+    const u32x pw_len2 = pw_len * 2;
+
     /**
      * prepend salt
      */
 
-    const u32 pw_salt_len = (pw_len * 2) + salt_len;
+    switch_buffer_by_offset_le (w0, w1, w2, w3, salt_len);
 
-    u32 w0_t[4];
-    u32 w1_t[4];
-    u32 w2_t[4];
-    u32 w3_t[4];
+    const u32x pw_salt_len = pw_len2 + salt_len;
 
-    make_unicode (w0, w0_t, w1_t);
-    make_unicode (w1, w2_t, w3_t);
-
-    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);
+    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];
 
     /**
      * sha512
      */
 
-    w0_t[0] = swap32 (w0_t[0]);
-    w0_t[1] = swap32 (w0_t[1]);
-    w0_t[2] = swap32 (w0_t[2]);
-    w0_t[3] = swap32 (w0_t[3]);
-    w1_t[0] = swap32 (w1_t[0]);
-    w1_t[1] = swap32 (w1_t[1]);
-    w1_t[2] = swap32 (w1_t[2]);
-    w1_t[3] = swap32 (w1_t[3]);
-    w2_t[0] = swap32 (w2_t[0]);
-    w2_t[1] = swap32 (w2_t[1]);
-    w2_t[2] = swap32 (w2_t[2]);
-    w2_t[3] = swap32 (w2_t[3]);
-    w3_t[0] = swap32 (w3_t[0]);
-    w3_t[1] = swap32 (w3_t[1]);
+    u32x w0_t[4];
+    u32x w1_t[4];
+    u32x w2_t[4];
+    u32x w3_t[4];
+
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
+    w2_t[0] = swap32 (w2[0]);
+    w2_t[1] = swap32 (w2[1]);
+    w2_t[2] = swap32 (w2[2]);
+    w2_t[3] = swap32 (w2[3]);
+    w3_t[0] = swap32 (w3[0]);
+    w3_t[1] = swap32 (w3[1]);
     w3_t[2] = 0;
     w3_t[3] = pw_salt_len * 8;
 
-    u64 digest[8];
+    u64x digest[8];
 
     digest[0] = SHA512M_A;
     digest[1] = SHA512M_B;
@@ -332,13 +341,12 @@ __kernel void m01740_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 
     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);
   }
 }
 
@@ -366,58 +374,45 @@ __kernel void m01740_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 
   if (gid >= gid_max) return;
 
-  u32 wordl0[4];
+  u32 pw_buf0[4];
+  u32 pw_buf1[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];
-
-  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_le (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
-  }
-
   /**
    * salt
    */
 
   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];
+  u32 salt_buf2[4];
+  u32 salt_buf3[4];
 
-  salt_buf1[0] = salt_bufs[salt_pos].salt_buf[4];
-  salt_buf1[1] = salt_bufs[salt_pos].salt_buf[5];
-  salt_buf1[2] = salt_bufs[salt_pos].salt_buf[6];
-  salt_buf1[3] = salt_bufs[salt_pos].salt_buf[7];
+  salt_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];
+  salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
+  salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
+  salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
+  salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
+  salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8];
+  salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9];
+  salt_buf2[2] = salt_bufs[salt_pos].salt_buf[10];
+  salt_buf2[3] = salt_bufs[salt_pos].salt_buf[11];
+  salt_buf3[0] = salt_bufs[salt_pos].salt_buf[12];
+  salt_buf3[1] = salt_bufs[salt_pos].salt_buf[13];
+  salt_buf3[2] = salt_bufs[salt_pos].salt_buf[14];
+  salt_buf3[3] = salt_bufs[salt_pos].salt_buf[15];
 
   const u32 salt_len = salt_bufs[salt_pos].salt_len;
 
@@ -437,43 +432,57 @@ __kernel void m01740_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    * loop
    */
 
-  for (u32 il_pos = 0; il_pos < il_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 u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos);
 
-    const u32 pw_len = pw_l_len + pw_r_len;
+    const u32x pw_len = pw_l_len + pw_r_len;
 
-    u32 wordr0[4];
-    u32 wordr1[4];
-    u32 wordr2[4];
-    u32 wordr3[4];
+    /**
+     * concat password candidate
+     */
 
-    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];
-    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];
-    wordr2[0] = 0;
-    wordr2[1] = 0;
-    wordr2[2] = 0;
-    wordr2[3] = 0;
-    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] = 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_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);
     }
 
-    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];
@@ -492,55 +501,63 @@ __kernel void m01740_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     w3[2] = wordl3[2] | wordr3[2];
     w3[3] = wordl3[3] | wordr3[3];
 
+    make_unicode (w1, w2, w3);
+    make_unicode (w0, w0, w1);
+
+    const u32x pw_len2 = pw_len * 2;
+
     /**
      * prepend salt
      */
 
-    const u32 pw_salt_len = (pw_len * 2) + salt_len;
+    switch_buffer_by_offset_le (w0, w1, w2, w3, salt_len);
 
-    u32 w0_t[4];
-    u32 w1_t[4];
-    u32 w2_t[4];
-    u32 w3_t[4];
+    const u32x pw_salt_len = pw_len2 + salt_len;
 
-    make_unicode (w0, w0_t, w1_t);
-    make_unicode (w1, w2_t, w3_t);
-
-    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);
+    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];
 
     /**
      * sha512
      */
 
-    w0_t[0] = swap32 (w0_t[0]);
-    w0_t[1] = swap32 (w0_t[1]);
-    w0_t[2] = swap32 (w0_t[2]);
-    w0_t[3] = swap32 (w0_t[3]);
-    w1_t[0] = swap32 (w1_t[0]);
-    w1_t[1] = swap32 (w1_t[1]);
-    w1_t[2] = swap32 (w1_t[2]);
-    w1_t[3] = swap32 (w1_t[3]);
-    w2_t[0] = swap32 (w2_t[0]);
-    w2_t[1] = swap32 (w2_t[1]);
-    w2_t[2] = swap32 (w2_t[2]);
-    w2_t[3] = swap32 (w2_t[3]);
-    w3_t[0] = swap32 (w3_t[0]);
-    w3_t[1] = swap32 (w3_t[1]);
+    u32x w0_t[4];
+    u32x w1_t[4];
+    u32x w2_t[4];
+    u32x w3_t[4];
+
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
+    w2_t[0] = swap32 (w2[0]);
+    w2_t[1] = swap32 (w2[1]);
+    w2_t[2] = swap32 (w2[2]);
+    w2_t[3] = swap32 (w2[3]);
+    w3_t[0] = swap32 (w3[0]);
+    w3_t[1] = swap32 (w3[1]);
     w3_t[2] = 0;
     w3_t[3] = pw_salt_len * 8;
 
-    u64 digest[8];
+    u64x digest[8];
 
     digest[0] = SHA512M_A;
     digest[1] = SHA512M_B;
@@ -553,13 +570,12 @@ __kernel void m01740_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 
     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);
   }
 }
 
diff --git a/OpenCL/m01740_a3.cl b/OpenCL/m01740_a3.cl
index 0d17a7a45..005099eae 100644
--- a/OpenCL/m01740_a3.cl
+++ b/OpenCL/m01740_a3.cl
@@ -155,159 +155,116 @@ static void m01740m (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;
 
-
   /**
    * prepend salt
    */
 
-  u32 w0_t[4];
-  u32 w1_t[4];
-  u32 w2_t[4];
-  u32 w3_t[4];
+  const u32 w0l = w0[0];
 
-  w0_t[0] = swap32_S (w0[0]);
-  w0_t[1] = swap32_S (w0[1]);
-  w0_t[2] = swap32_S (w0[2]);
-  w0_t[3] = swap32_S (w0[3]);
-  w1_t[0] = swap32_S (w1[0]);
-  w1_t[1] = swap32_S (w1[1]);
-  w1_t[2] = swap32_S (w1[2]);
-  w1_t[3] = swap32_S (w1[3]);
-  w2_t[0] = swap32_S (w2[0]);
-  w2_t[1] = swap32_S (w2[1]);
-  w2_t[2] = swap32_S (w2[2]);
-  w2_t[3] = swap32_S (w2[3]);
-  w3_t[0] = swap32_S (w3[0]);
-  w3_t[1] = swap32_S (w3[1]);
-  w3_t[2] = swap32_S (w3[2]);
-  w3_t[3] = swap32_S (w3[3]);
+  switch_buffer_by_offset_be_S (w0, w1, w2, w3, salt_len);
 
-  switch_buffer_by_offset_le_S (w0_t, w1_t, w2_t, w3_t, 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];
 
-  w0_t[0] |= salt_buf0[0];
-  w0_t[1] |= salt_buf0[1];
-  w0_t[2] |= salt_buf0[2];
-  w0_t[3] |= salt_buf0[3];
-  w1_t[0] |= salt_buf1[0];
-  w1_t[1] |= salt_buf1[1];
-  w1_t[2] |= salt_buf1[2];
-  w1_t[3] |= salt_buf1[3];
-  w2_t[0] |= salt_buf2[0];
-  w2_t[1] |= salt_buf2[1];
-  w2_t[2] |= salt_buf2[2];
-  w2_t[3] |= salt_buf2[3];
-  w3_t[0] |= salt_buf3[0];
-  w3_t[1] |= salt_buf3[1];
-  w3_t[2] |= salt_buf3[2];
-  w3_t[3] |= salt_buf3[3];
+  u32x t0[4];
+  u32x t1[4];
+  u32x t2[4];
+  u32x t3[4];
 
-  w0_t[0] = swap32_S (w0_t[0]);
-  w0_t[1] = swap32_S (w0_t[1]);
-  w0_t[2] = swap32_S (w0_t[2]);
-  w0_t[3] = swap32_S (w0_t[3]);
-  w1_t[0] = swap32_S (w1_t[0]);
-  w1_t[1] = swap32_S (w1_t[1]);
-  w1_t[2] = swap32_S (w1_t[2]);
-  w1_t[3] = swap32_S (w1_t[3]);
-  w2_t[0] = swap32_S (w2_t[0]);
-  w2_t[1] = swap32_S (w2_t[1]);
-  w2_t[2] = swap32_S (w2_t[2]);
-  w2_t[3] = swap32_S (w2_t[3]);
-  w3_t[0] = swap32_S (w3_t[0]);
-  w3_t[1] = swap32_S (w3_t[1]);
-  w3_t[2] = swap32_S (w3_t[2]);
-  w3_t[3] = swap32_S (w3_t[3]);
+  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];
 
   /**
    * loop
    */
 
-  u32 w0l = w0[0];
-
   for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
   {
     const u32x w0r = ix_create_bft (bfs_buf, il_pos);
 
     const u32x w0lr = w0l | w0r;
 
-    u32x wx[16];
+    overwrite_at_be_4x4 (t0, t1, t2, t3, w0lr, salt_len);
 
-    wx[ 0] = w0_t[0];
-    wx[ 1] = w0_t[1];
-    wx[ 2] = w0_t[2];
-    wx[ 3] = w0_t[3];
-    wx[ 4] = w1_t[0];
-    wx[ 5] = w1_t[1];
-    wx[ 6] = w1_t[2];
-    wx[ 7] = w1_t[3];
-    wx[ 8] = w2_t[0];
-    wx[ 9] = w2_t[1];
-    wx[10] = w2_t[2];
-    wx[11] = w2_t[3];
-    wx[12] = w3_t[0];
-    wx[13] = w3_t[1];
-    wx[14] = w3_t[2];
-    wx[15] = w3_t[3];
-
-    overwrite_at_be (wx, w0lr, salt_len);
+    /**
+     * sha512
+     */
 
     u32x w0_t[4];
     u32x w1_t[4];
     u32x w2_t[4];
     u32x w3_t[4];
 
-    w0_t[0] = wx[ 0];
-    w0_t[1] = wx[ 1];
-    w0_t[2] = wx[ 2];
-    w0_t[3] = wx[ 3];
-    w1_t[0] = wx[ 4];
-    w1_t[1] = wx[ 5];
-    w1_t[2] = wx[ 6];
-    w1_t[3] = wx[ 7];
-    w2_t[0] = wx[ 8];
-    w2_t[1] = wx[ 9];
-    w2_t[2] = wx[10];
-    w2_t[3] = wx[11];
-    w3_t[0] = wx[12];
-    w3_t[1] = wx[13];
+    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;
 
-    /**
-     * sha512
-     */
-
     u64x digest[8];
 
     digest[0] = SHA512M_A;
@@ -356,32 +313,26 @@ static void m01740s (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;
 
@@ -391,120 +342,84 @@ static void m01740s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
    * prepend salt
    */
 
-  u32 w0_t[4];
-  u32 w1_t[4];
-  u32 w2_t[4];
-  u32 w3_t[4];
+  const u32 w0l = w0[0];
 
-  w0_t[0] = swap32_S (w0[0]);
-  w0_t[1] = swap32_S (w0[1]);
-  w0_t[2] = swap32_S (w0[2]);
-  w0_t[3] = swap32_S (w0[3]);
-  w1_t[0] = swap32_S (w1[0]);
-  w1_t[1] = swap32_S (w1[1]);
-  w1_t[2] = swap32_S (w1[2]);
-  w1_t[3] = swap32_S (w1[3]);
-  w2_t[0] = swap32_S (w2[0]);
-  w2_t[1] = swap32_S (w2[1]);
-  w2_t[2] = swap32_S (w2[2]);
-  w2_t[3] = swap32_S (w2[3]);
-  w3_t[0] = swap32_S (w3[0]);
-  w3_t[1] = swap32_S (w3[1]);
-  w3_t[2] = swap32_S (w3[2]);
-  w3_t[3] = swap32_S (w3[3]);
+  switch_buffer_by_offset_be_S (w0, w1, w2, w3, salt_len);
 
-  switch_buffer_by_offset_le_S (w0_t, w1_t, w2_t, w3_t, 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];
 
-  w0_t[0] |= salt_buf0[0];
-  w0_t[1] |= salt_buf0[1];
-  w0_t[2] |= salt_buf0[2];
-  w0_t[3] |= salt_buf0[3];
-  w1_t[0] |= salt_buf1[0];
-  w1_t[1] |= salt_buf1[1];
-  w1_t[2] |= salt_buf1[2];
-  w1_t[3] |= salt_buf1[3];
-  w2_t[0] |= salt_buf2[0];
-  w2_t[1] |= salt_buf2[1];
-  w2_t[2] |= salt_buf2[2];
-  w2_t[3] |= salt_buf2[3];
-  w3_t[0] |= salt_buf3[0];
-  w3_t[1] |= salt_buf3[1];
-  w3_t[2] |= salt_buf3[2];
-  w3_t[3] |= salt_buf3[3];
+  u32x t0[4];
+  u32x t1[4];
+  u32x t2[4];
+  u32x t3[4];
 
-  w0_t[0] = swap32_S (w0_t[0]);
-  w0_t[1] = swap32_S (w0_t[1]);
-  w0_t[2] = swap32_S (w0_t[2]);
-  w0_t[3] = swap32_S (w0_t[3]);
-  w1_t[0] = swap32_S (w1_t[0]);
-  w1_t[1] = swap32_S (w1_t[1]);
-  w1_t[2] = swap32_S (w1_t[2]);
-  w1_t[3] = swap32_S (w1_t[3]);
-  w2_t[0] = swap32_S (w2_t[0]);
-  w2_t[1] = swap32_S (w2_t[1]);
-  w2_t[2] = swap32_S (w2_t[2]);
-  w2_t[3] = swap32_S (w2_t[3]);
-  w3_t[0] = swap32_S (w3_t[0]);
-  w3_t[1] = swap32_S (w3_t[1]);
-  w3_t[2] = swap32_S (w3_t[2]);
-  w3_t[3] = swap32_S (w3_t[3]);
+  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];
 
   /**
    * loop
    */
 
-  u32 w0l = w0[0];
-
   for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
   {
     const u32x w0r = ix_create_bft (bfs_buf, il_pos);
 
     const u32x w0lr = w0l | w0r;
 
+    overwrite_at_be_4x4 (t0, t1, t2, t3, w0lr, salt_len);
+
     /**
-     * prepend salt
+     * sha512
      */
 
-    u32x wx[16];
-
-    wx[ 0] = w0_t[0];
-    wx[ 1] = w0_t[1];
-    wx[ 2] = w0_t[2];
-    wx[ 3] = w0_t[3];
-    wx[ 4] = w1_t[0];
-    wx[ 5] = w1_t[1];
-    wx[ 6] = w1_t[2];
-    wx[ 7] = w1_t[3];
-    wx[ 8] = w2_t[0];
-    wx[ 9] = w2_t[1];
-    wx[10] = w2_t[2];
-    wx[11] = w2_t[3];
-    wx[12] = w3_t[0];
-    wx[13] = w3_t[1];
-    wx[14] = w3_t[2];
-    wx[15] = w3_t[3];
-
-    overwrite_at_be (wx, w0lr, salt_len);
-
     u32x w0_t[4];
     u32x w1_t[4];
     u32x w2_t[4];
     u32x w3_t[4];
 
-    w0_t[0] = wx[ 0];
-    w0_t[1] = wx[ 1];
-    w0_t[2] = wx[ 2];
-    w0_t[3] = wx[ 3];
-    w1_t[0] = wx[ 4];
-    w1_t[1] = wx[ 5];
-    w1_t[2] = wx[ 6];
-    w1_t[3] = wx[ 7];
-    w2_t[0] = wx[ 8];
-    w2_t[1] = wx[ 9];
-    w2_t[2] = wx[10];
-    w2_t[3] = wx[11];
-    w3_t[0] = wx[12];
-    w3_t[1] = wx[13];
+    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;