From 1a353b54ffd1e74cd55c3813d52083161c88f1d0 Mon Sep 17 00:00:00 2001
From: Jens Steube <jens.steube@gmail.com>
Date: Wed, 13 Apr 2016 20:58:10 +0200
Subject: [PATCH] Cleanup -m 73xx kernels to latest standard

---
 OpenCL/m07300_a0.cl | 258 +++++++++-----------
 OpenCL/m07300_a1.cl | 578 +++++++++++++++++++-------------------------
 OpenCL/m07300_a3.cl |  56 ++---
 3 files changed, 388 insertions(+), 504 deletions(-)

diff --git a/OpenCL/m07300_a0.cl b/OpenCL/m07300_a0.cl
index 670a24178..b0ef0f9be 100644
--- a/OpenCL/m07300_a0.cl
+++ b/OpenCL/m07300_a0.cl
@@ -256,18 +256,16 @@ __kernel void m07300_m04 (__global pw_t *pws, __global kernel_rule_t *  rules_bu
   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;
 
@@ -290,42 +288,23 @@ __kernel void m07300_m04 (__global pw_t *pws, __global kernel_rule_t *  rules_bu
 
     apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
 
+    w0[0] = swap32 (w0[0]);
+    w0[1] = swap32 (w0[1]);
+    w0[2] = swap32 (w0[2]);
+    w0[3] = swap32 (w0[3]);
+    w1[0] = swap32 (w1[0]);
+    w1[1] = swap32 (w1[1]);
+    w1[2] = swap32 (w1[2]);
+    w1[3] = swap32 (w1[3]);
+
     /**
-     * pads
+     * RAKP
      */
 
-    u32x w0_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]);
-
-    u32x w1_t[4];
-
-    w1_t[0] = swap32 (w1[0]);
-    w1_t[1] = swap32 (w1[1]);
-    w1_t[2] = swap32 (w1[2]);
-    w1_t[3] = swap32 (w1[3]);
-
-    u32x w2_t[4];
-
-    w2_t[0] = 0;
-    w2_t[1] = 0;
-    w2_t[2] = 0;
-    w2_t[3] = 0;
-
-    u32x w3_t[4];
-
-    w3_t[0] = 0;
-    w3_t[1] = 0;
-    w3_t[2] = 0;
-    w3_t[3] = 0;
-
     u32x ipad[5];
     u32x opad[5];
 
-    hmac_sha1_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
+    hmac_sha1_pad (w0, w1, w2, w3, ipad, opad);
 
     int esalt_size = esalt_len;
 
@@ -334,46 +313,46 @@ __kernel void m07300_m04 (__global pw_t *pws, __global kernel_rule_t *  rules_bu
 
     for (esalt_left = esalt_size, esalt_off = 0; esalt_left >= 56; esalt_left -= 64, esalt_off += 16)
     {
-      w0_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  0];
-      w0_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  1];
-      w0_t[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  2];
-      w0_t[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  3];
-      w1_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  4];
-      w1_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  5];
-      w1_t[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  6];
-      w1_t[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  7];
-      w2_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  8];
-      w2_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  9];
-      w2_t[2] = rakp_bufs[salt_pos].salt_buf[esalt_off + 10];
-      w2_t[3] = rakp_bufs[salt_pos].salt_buf[esalt_off + 11];
-      w3_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off + 12];
-      w3_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off + 13];
-      w3_t[2] = rakp_bufs[salt_pos].salt_buf[esalt_off + 14];
-      w3_t[3] = rakp_bufs[salt_pos].salt_buf[esalt_off + 15];
-
-      sha1_transform (w0_t, w1_t, w2_t, w3_t, ipad);
+      w0[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  0];
+      w0[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  1];
+      w0[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  2];
+      w0[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  3];
+      w1[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  4];
+      w1[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  5];
+      w1[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  6];
+      w1[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  7];
+      w2[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  8];
+      w2[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  9];
+      w2[2] = rakp_bufs[salt_pos].salt_buf[esalt_off + 10];
+      w2[3] = rakp_bufs[salt_pos].salt_buf[esalt_off + 11];
+      w3[0] = rakp_bufs[salt_pos].salt_buf[esalt_off + 12];
+      w3[1] = rakp_bufs[salt_pos].salt_buf[esalt_off + 13];
+      w3[2] = rakp_bufs[salt_pos].salt_buf[esalt_off + 14];
+      w3[3] = rakp_bufs[salt_pos].salt_buf[esalt_off + 15];
+
+      sha1_transform (w0, w1, w2, w3, ipad);
     }
 
-    w0_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  0];
-    w0_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  1];
-    w0_t[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  2];
-    w0_t[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  3];
-    w1_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  4];
-    w1_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  5];
-    w1_t[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  6];
-    w1_t[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  7];
-    w2_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  8];
-    w2_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  9];
-    w2_t[2] = rakp_bufs[salt_pos].salt_buf[esalt_off + 10];
-    w2_t[3] = rakp_bufs[salt_pos].salt_buf[esalt_off + 11];
-    w3_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off + 12];
-    w3_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off + 13];
-    w3_t[2] = 0;
-    w3_t[3] = (64 + esalt_size) * 8;
+    w0[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  0];
+    w0[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  1];
+    w0[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  2];
+    w0[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  3];
+    w1[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  4];
+    w1[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  5];
+    w1[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  6];
+    w1[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  7];
+    w2[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  8];
+    w2[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  9];
+    w2[2] = rakp_bufs[salt_pos].salt_buf[esalt_off + 10];
+    w2[3] = rakp_bufs[salt_pos].salt_buf[esalt_off + 11];
+    w3[0] = rakp_bufs[salt_pos].salt_buf[esalt_off + 12];
+    w3[1] = rakp_bufs[salt_pos].salt_buf[esalt_off + 13];
+    w3[2] = 0;
+    w3[3] = (64 + esalt_size) * 8;
 
     u32x digest[5];
 
-    hmac_sha1_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
+    hmac_sha1_run (w0, w1, w2, w3, ipad, opad, digest);
 
     COMPARE_M_SIMD (digest[3], digest[4], digest[2], digest[1]);
   }
@@ -404,18 +383,16 @@ __kernel void m07300_s04 (__global pw_t *pws, __global kernel_rule_t *  rules_bu
   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;
 
@@ -450,42 +427,23 @@ __kernel void m07300_s04 (__global pw_t *pws, __global kernel_rule_t *  rules_bu
 
     apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
 
+    w0[0] = swap32 (w0[0]);
+    w0[1] = swap32 (w0[1]);
+    w0[2] = swap32 (w0[2]);
+    w0[3] = swap32 (w0[3]);
+    w1[0] = swap32 (w1[0]);
+    w1[1] = swap32 (w1[1]);
+    w1[2] = swap32 (w1[2]);
+    w1[3] = swap32 (w1[3]);
+
     /**
-     * pads
+     * RAKP
      */
 
-    u32x w0_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]);
-
-    u32x w1_t[4];
-
-    w1_t[0] = swap32 (w1[0]);
-    w1_t[1] = swap32 (w1[1]);
-    w1_t[2] = swap32 (w1[2]);
-    w1_t[3] = swap32 (w1[3]);
-
-    u32x w2_t[4];
-
-    w2_t[0] = 0;
-    w2_t[1] = 0;
-    w2_t[2] = 0;
-    w2_t[3] = 0;
-
-    u32x w3_t[4];
-
-    w3_t[0] = 0;
-    w3_t[1] = 0;
-    w3_t[2] = 0;
-    w3_t[3] = 0;
-
     u32x ipad[5];
     u32x opad[5];
 
-    hmac_sha1_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
+    hmac_sha1_pad (w0, w1, w2, w3, ipad, opad);
 
     int esalt_size = esalt_len;
 
@@ -494,46 +452,46 @@ __kernel void m07300_s04 (__global pw_t *pws, __global kernel_rule_t *  rules_bu
 
     for (esalt_left = esalt_size, esalt_off = 0; esalt_left >= 56; esalt_left -= 64, esalt_off += 16)
     {
-      w0_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  0];
-      w0_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  1];
-      w0_t[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  2];
-      w0_t[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  3];
-      w1_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  4];
-      w1_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  5];
-      w1_t[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  6];
-      w1_t[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  7];
-      w2_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  8];
-      w2_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  9];
-      w2_t[2] = rakp_bufs[salt_pos].salt_buf[esalt_off + 10];
-      w2_t[3] = rakp_bufs[salt_pos].salt_buf[esalt_off + 11];
-      w3_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off + 12];
-      w3_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off + 13];
-      w3_t[2] = rakp_bufs[salt_pos].salt_buf[esalt_off + 14];
-      w3_t[3] = rakp_bufs[salt_pos].salt_buf[esalt_off + 15];
-
-      sha1_transform (w0_t, w1_t, w2_t, w3_t, ipad);
+      w0[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  0];
+      w0[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  1];
+      w0[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  2];
+      w0[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  3];
+      w1[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  4];
+      w1[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  5];
+      w1[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  6];
+      w1[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  7];
+      w2[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  8];
+      w2[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  9];
+      w2[2] = rakp_bufs[salt_pos].salt_buf[esalt_off + 10];
+      w2[3] = rakp_bufs[salt_pos].salt_buf[esalt_off + 11];
+      w3[0] = rakp_bufs[salt_pos].salt_buf[esalt_off + 12];
+      w3[1] = rakp_bufs[salt_pos].salt_buf[esalt_off + 13];
+      w3[2] = rakp_bufs[salt_pos].salt_buf[esalt_off + 14];
+      w3[3] = rakp_bufs[salt_pos].salt_buf[esalt_off + 15];
+
+      sha1_transform (w0, w1, w2, w3, ipad);
     }
 
-    w0_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  0];
-    w0_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  1];
-    w0_t[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  2];
-    w0_t[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  3];
-    w1_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  4];
-    w1_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  5];
-    w1_t[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  6];
-    w1_t[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  7];
-    w2_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  8];
-    w2_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  9];
-    w2_t[2] = rakp_bufs[salt_pos].salt_buf[esalt_off + 10];
-    w2_t[3] = rakp_bufs[salt_pos].salt_buf[esalt_off + 11];
-    w3_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off + 12];
-    w3_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off + 13];
-    w3_t[2] = 0;
-    w3_t[3] = (64 + esalt_size) * 8;
+    w0[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  0];
+    w0[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  1];
+    w0[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  2];
+    w0[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  3];
+    w1[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  4];
+    w1[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  5];
+    w1[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  6];
+    w1[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  7];
+    w2[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  8];
+    w2[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  9];
+    w2[2] = rakp_bufs[salt_pos].salt_buf[esalt_off + 10];
+    w2[3] = rakp_bufs[salt_pos].salt_buf[esalt_off + 11];
+    w3[0] = rakp_bufs[salt_pos].salt_buf[esalt_off + 12];
+    w3[1] = rakp_bufs[salt_pos].salt_buf[esalt_off + 13];
+    w3[2] = 0;
+    w3[3] = (64 + esalt_size) * 8;
 
     u32x digest[5];
 
-    hmac_sha1_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
+    hmac_sha1_run (w0, w1, w2, w3, ipad, opad, digest);
 
     COMPARE_S_SIMD (digest[3], digest[4], digest[2], digest[1]);
   }
diff --git a/OpenCL/m07300_a1.cl b/OpenCL/m07300_a1.cl
index fd1890a11..b8a9ce365 100644
--- a/OpenCL/m07300_a1.cl
+++ b/OpenCL/m07300_a1.cl
@@ -5,6 +5,8 @@
 
 #define _SHA1_
 
+#define NEW_SIMD_CODE
+
 #include "include/constants.h"
 #include "include/kernel_vendor.h"
 
@@ -16,34 +18,32 @@
 #include "include/kernel_functions.c"
 #include "OpenCL/types_ocl.c"
 #include "OpenCL/common.c"
+#include "OpenCL/simd.c"
 
-#define COMPARE_S "OpenCL/check_single_comp4.c"
-#define COMPARE_M "OpenCL/check_multi_comp4.c"
-
-static void sha1_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[5])
+static void sha1_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[5])
 {
-  u32 A = digest[0];
-  u32 B = digest[1];
-  u32 C = digest[2];
-  u32 D = digest[3];
-  u32 E = digest[4];
-
-  u32 w0_t = w0[0];
-  u32 w1_t = w0[1];
-  u32 w2_t = w0[2];
-  u32 w3_t = w0[3];
-  u32 w4_t = w1[0];
-  u32 w5_t = w1[1];
-  u32 w6_t = w1[2];
-  u32 w7_t = w1[3];
-  u32 w8_t = w2[0];
-  u32 w9_t = w2[1];
-  u32 wa_t = w2[2];
-  u32 wb_t = w2[3];
-  u32 wc_t = w3[0];
-  u32 wd_t = w3[1];
-  u32 we_t = w3[2];
-  u32 wf_t = w3[3];
+  u32x A = digest[0];
+  u32x B = digest[1];
+  u32x C = digest[2];
+  u32x D = digest[3];
+  u32x E = digest[4];
+
+  u32x w0_t = w0[0];
+  u32x w1_t = w0[1];
+  u32x w2_t = w0[2];
+  u32x w3_t = w0[3];
+  u32x w4_t = w1[0];
+  u32x w5_t = w1[1];
+  u32x w6_t = w1[2];
+  u32x w7_t = w1[3];
+  u32x w8_t = w2[0];
+  u32x w9_t = w2[1];
+  u32x wa_t = w2[2];
+  u32x wb_t = w2[3];
+  u32x wc_t = w3[0];
+  u32x wd_t = w3[1];
+  u32x we_t = w3[2];
+  u32x wf_t = w3[3];
 
   #undef K
   #define K SHA1C00
@@ -148,7 +148,7 @@ static void sha1_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], c
   digest[4] += E;
 }
 
-static void hmac_sha1_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[5], u32 opad[5])
+static void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
 {
   w0[0] = w0[0] ^ 0x36363636;
   w0[1] = w0[1] ^ 0x36363636;
@@ -201,7 +201,7 @@ static void hmac_sha1_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[
   sha1_transform (w0, w1, w2, w3, opad);
 }
 
-static void hmac_sha1_run (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[5], u32 opad[5], u32 digest[5])
+static void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
 {
   digest[0] = ipad[0];
   digest[1] = ipad[1];
@@ -253,41 +253,20 @@ __kernel void m07300_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 pw_buf0[4];
+  u32 pw_buf1[4];
 
-  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
    */
@@ -298,109 +277,100 @@ __kernel void m07300_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 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] = 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];
+    u32x w0[4];
+    u32x w1[4];
+    u32x w2[4];
+    u32x w3[4];
 
     w0[0] = wordl0[0] | wordr0[0];
     w0[1] = wordl0[1] | wordr0[1];
     w0[2] = wordl0[2] | wordr0[2];
     w0[3] = wordl0[3] | wordr0[3];
-
-    u32 w1[4];
-
     w1[0] = wordl1[0] | wordr1[0];
     w1[1] = wordl1[1] | wordr1[1];
     w1[2] = wordl1[2] | wordr1[2];
     w1[3] = wordl1[3] | wordr1[3];
-
-    u32 w2[4];
-
     w2[0] = wordl2[0] | wordr2[0];
     w2[1] = wordl2[1] | wordr2[1];
     w2[2] = wordl2[2] | wordr2[2];
     w2[3] = wordl2[3] | wordr2[3];
-
-    u32 w3[4];
-
     w3[0] = wordl3[0] | wordr3[0];
     w3[1] = wordl3[1] | wordr3[1];
     w3[2] = wordl3[2] | wordr3[2];
     w3[3] = wordl3[3] | wordr3[3];
 
+    w0[0] = swap32 (w0[0]);
+    w0[1] = swap32 (w0[1]);
+    w0[2] = swap32 (w0[2]);
+    w0[3] = swap32 (w0[3]);
+    w1[0] = swap32 (w1[0]);
+    w1[1] = swap32 (w1[1]);
+    w1[2] = swap32 (w1[2]);
+    w1[3] = swap32 (w1[3]);
+    w2[0] = swap32 (w2[0]);
+    w2[1] = swap32 (w2[1]);
+    w2[2] = swap32 (w2[2]);
+    w2[3] = swap32 (w2[3]);
+    w3[0] = swap32 (w3[0]);
+    w3[1] = swap32 (w3[1]);
+    w3[2] = swap32 (w3[2]);
+    w3[3] = swap32 (w3[3]);
+
     /**
-     * pads
+     * RAKP
      */
 
-    u32 w0_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]);
-
-    u32 w1_t[4];
-
-    w1_t[0] = swap32 (w1[0]);
-    w1_t[1] = swap32 (w1[1]);
-    w1_t[2] = swap32 (w1[2]);
-    w1_t[3] = swap32 (w1[3]);
-
-    u32 w2_t[4];
-
-    w2_t[0] = 0;
-    w2_t[1] = 0;
-    w2_t[2] = 0;
-    w2_t[3] = 0;
-
-    u32 w3_t[4];
-
-    w3_t[0] = 0;
-    w3_t[1] = 0;
-    w3_t[2] = 0;
-    w3_t[3] = 0;
+    u32x ipad[5];
+    u32x opad[5];
 
-    u32 ipad[5];
-    u32 opad[5];
-
-    hmac_sha1_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
+    hmac_sha1_pad (w0, w1, w2, w3, ipad, opad);
 
     int esalt_size = esalt_len;
 
@@ -409,53 +379,48 @@ __kernel void m07300_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 
     for (esalt_left = esalt_size, esalt_off = 0; esalt_left >= 56; esalt_left -= 64, esalt_off += 16)
     {
-      w0_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  0];
-      w0_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  1];
-      w0_t[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  2];
-      w0_t[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  3];
-      w1_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  4];
-      w1_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  5];
-      w1_t[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  6];
-      w1_t[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  7];
-      w2_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  8];
-      w2_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  9];
-      w2_t[2] = rakp_bufs[salt_pos].salt_buf[esalt_off + 10];
-      w2_t[3] = rakp_bufs[salt_pos].salt_buf[esalt_off + 11];
-      w3_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off + 12];
-      w3_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off + 13];
-      w3_t[2] = rakp_bufs[salt_pos].salt_buf[esalt_off + 14];
-      w3_t[3] = rakp_bufs[salt_pos].salt_buf[esalt_off + 15];
-
-      sha1_transform (w0_t, w1_t, w2_t, w3_t, ipad);
+      w0[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  0];
+      w0[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  1];
+      w0[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  2];
+      w0[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  3];
+      w1[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  4];
+      w1[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  5];
+      w1[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  6];
+      w1[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  7];
+      w2[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  8];
+      w2[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  9];
+      w2[2] = rakp_bufs[salt_pos].salt_buf[esalt_off + 10];
+      w2[3] = rakp_bufs[salt_pos].salt_buf[esalt_off + 11];
+      w3[0] = rakp_bufs[salt_pos].salt_buf[esalt_off + 12];
+      w3[1] = rakp_bufs[salt_pos].salt_buf[esalt_off + 13];
+      w3[2] = rakp_bufs[salt_pos].salt_buf[esalt_off + 14];
+      w3[3] = rakp_bufs[salt_pos].salt_buf[esalt_off + 15];
+
+      sha1_transform (w0, w1, w2, w3, ipad);
     }
 
-    w0_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  0];
-    w0_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  1];
-    w0_t[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  2];
-    w0_t[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  3];
-    w1_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  4];
-    w1_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  5];
-    w1_t[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  6];
-    w1_t[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  7];
-    w2_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  8];
-    w2_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  9];
-    w2_t[2] = rakp_bufs[salt_pos].salt_buf[esalt_off + 10];
-    w2_t[3] = rakp_bufs[salt_pos].salt_buf[esalt_off + 11];
-    w3_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off + 12];
-    w3_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off + 13];
-    w3_t[2] = 0;
-    w3_t[3] = (64 + esalt_size) * 8;
-
-    u32 digest[5];
-
-    hmac_sha1_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
-
-    const u32 r0 = digest[3];
-    const u32 r1 = digest[4];
-    const u32 r2 = digest[2];
-    const u32 r3 = digest[1];
-
-    #include COMPARE_M
+    w0[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  0];
+    w0[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  1];
+    w0[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  2];
+    w0[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  3];
+    w1[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  4];
+    w1[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  5];
+    w1[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  6];
+    w1[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  7];
+    w2[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  8];
+    w2[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  9];
+    w2[2] = rakp_bufs[salt_pos].salt_buf[esalt_off + 10];
+    w2[3] = rakp_bufs[salt_pos].salt_buf[esalt_off + 11];
+    w3[0] = rakp_bufs[salt_pos].salt_buf[esalt_off + 12];
+    w3[1] = rakp_bufs[salt_pos].salt_buf[esalt_off + 13];
+    w3[2] = 0;
+    w3[3] = (64 + esalt_size) * 8;
+
+    u32x digest[5];
+
+    hmac_sha1_run (w0, w1, w2, w3, ipad, opad, digest);
+
+    COMPARE_M_SIMD (digest[3], digest[4], digest[2], digest[1]);
   }
 }
 
@@ -483,41 +448,20 @@ __kernel void m07300_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_le (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
-  }
-
   /**
    * salt
    */
@@ -540,109 +484,100 @@ __kernel void m07300_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 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];
+    const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos);
 
-    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];
+    const u32x pw_len = pw_l_len + pw_r_len;
 
-    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] = 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];
+    u32x w0[4];
+    u32x w1[4];
+    u32x w2[4];
+    u32x w3[4];
 
     w0[0] = wordl0[0] | wordr0[0];
     w0[1] = wordl0[1] | wordr0[1];
     w0[2] = wordl0[2] | wordr0[2];
     w0[3] = wordl0[3] | wordr0[3];
-
-    u32 w1[4];
-
     w1[0] = wordl1[0] | wordr1[0];
     w1[1] = wordl1[1] | wordr1[1];
     w1[2] = wordl1[2] | wordr1[2];
     w1[3] = wordl1[3] | wordr1[3];
-
-    u32 w2[4];
-
     w2[0] = wordl2[0] | wordr2[0];
     w2[1] = wordl2[1] | wordr2[1];
     w2[2] = wordl2[2] | wordr2[2];
     w2[3] = wordl2[3] | wordr2[3];
-
-    u32 w3[4];
-
     w3[0] = wordl3[0] | wordr3[0];
     w3[1] = wordl3[1] | wordr3[1];
     w3[2] = wordl3[2] | wordr3[2];
     w3[3] = wordl3[3] | wordr3[3];
 
+    w0[0] = swap32 (w0[0]);
+    w0[1] = swap32 (w0[1]);
+    w0[2] = swap32 (w0[2]);
+    w0[3] = swap32 (w0[3]);
+    w1[0] = swap32 (w1[0]);
+    w1[1] = swap32 (w1[1]);
+    w1[2] = swap32 (w1[2]);
+    w1[3] = swap32 (w1[3]);
+    w2[0] = swap32 (w2[0]);
+    w2[1] = swap32 (w2[1]);
+    w2[2] = swap32 (w2[2]);
+    w2[3] = swap32 (w2[3]);
+    w3[0] = swap32 (w3[0]);
+    w3[1] = swap32 (w3[1]);
+    w3[2] = swap32 (w3[2]);
+    w3[3] = swap32 (w3[3]);
+
     /**
-     * pads
+     * RAKP
      */
 
-    u32 w0_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]);
-
-    u32 w1_t[4];
-
-    w1_t[0] = swap32 (w1[0]);
-    w1_t[1] = swap32 (w1[1]);
-    w1_t[2] = swap32 (w1[2]);
-    w1_t[3] = swap32 (w1[3]);
-
-    u32 w2_t[4];
-
-    w2_t[0] = 0;
-    w2_t[1] = 0;
-    w2_t[2] = 0;
-    w2_t[3] = 0;
-
-    u32 w3_t[4];
-
-    w3_t[0] = 0;
-    w3_t[1] = 0;
-    w3_t[2] = 0;
-    w3_t[3] = 0;
-
-    u32 ipad[5];
-    u32 opad[5];
+    u32x ipad[5];
+    u32x opad[5];
 
-    hmac_sha1_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
+    hmac_sha1_pad (w0, w1, w2, w3, ipad, opad);
 
     int esalt_size = esalt_len;
 
@@ -651,53 +586,48 @@ __kernel void m07300_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 
     for (esalt_left = esalt_size, esalt_off = 0; esalt_left >= 56; esalt_left -= 64, esalt_off += 16)
     {
-      w0_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  0];
-      w0_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  1];
-      w0_t[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  2];
-      w0_t[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  3];
-      w1_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  4];
-      w1_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  5];
-      w1_t[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  6];
-      w1_t[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  7];
-      w2_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  8];
-      w2_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  9];
-      w2_t[2] = rakp_bufs[salt_pos].salt_buf[esalt_off + 10];
-      w2_t[3] = rakp_bufs[salt_pos].salt_buf[esalt_off + 11];
-      w3_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off + 12];
-      w3_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off + 13];
-      w3_t[2] = rakp_bufs[salt_pos].salt_buf[esalt_off + 14];
-      w3_t[3] = rakp_bufs[salt_pos].salt_buf[esalt_off + 15];
-
-      sha1_transform (w0_t, w1_t, w2_t, w3_t, ipad);
+      w0[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  0];
+      w0[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  1];
+      w0[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  2];
+      w0[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  3];
+      w1[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  4];
+      w1[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  5];
+      w1[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  6];
+      w1[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  7];
+      w2[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  8];
+      w2[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  9];
+      w2[2] = rakp_bufs[salt_pos].salt_buf[esalt_off + 10];
+      w2[3] = rakp_bufs[salt_pos].salt_buf[esalt_off + 11];
+      w3[0] = rakp_bufs[salt_pos].salt_buf[esalt_off + 12];
+      w3[1] = rakp_bufs[salt_pos].salt_buf[esalt_off + 13];
+      w3[2] = rakp_bufs[salt_pos].salt_buf[esalt_off + 14];
+      w3[3] = rakp_bufs[salt_pos].salt_buf[esalt_off + 15];
+
+      sha1_transform (w0, w1, w2, w3, ipad);
     }
 
-    w0_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  0];
-    w0_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  1];
-    w0_t[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  2];
-    w0_t[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  3];
-    w1_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  4];
-    w1_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  5];
-    w1_t[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  6];
-    w1_t[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  7];
-    w2_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  8];
-    w2_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  9];
-    w2_t[2] = rakp_bufs[salt_pos].salt_buf[esalt_off + 10];
-    w2_t[3] = rakp_bufs[salt_pos].salt_buf[esalt_off + 11];
-    w3_t[0] = rakp_bufs[salt_pos].salt_buf[esalt_off + 12];
-    w3_t[1] = rakp_bufs[salt_pos].salt_buf[esalt_off + 13];
-    w3_t[2] = 0;
-    w3_t[3] = (64 + esalt_size) * 8;
-
-    u32 digest[5];
-
-    hmac_sha1_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
-
-    const u32 r0 = digest[3];
-    const u32 r1 = digest[4];
-    const u32 r2 = digest[2];
-    const u32 r3 = digest[1];
-
-    #include COMPARE_S
+    w0[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  0];
+    w0[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  1];
+    w0[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  2];
+    w0[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  3];
+    w1[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  4];
+    w1[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  5];
+    w1[2] = rakp_bufs[salt_pos].salt_buf[esalt_off +  6];
+    w1[3] = rakp_bufs[salt_pos].salt_buf[esalt_off +  7];
+    w2[0] = rakp_bufs[salt_pos].salt_buf[esalt_off +  8];
+    w2[1] = rakp_bufs[salt_pos].salt_buf[esalt_off +  9];
+    w2[2] = rakp_bufs[salt_pos].salt_buf[esalt_off + 10];
+    w2[3] = rakp_bufs[salt_pos].salt_buf[esalt_off + 11];
+    w3[0] = rakp_bufs[salt_pos].salt_buf[esalt_off + 12];
+    w3[1] = rakp_bufs[salt_pos].salt_buf[esalt_off + 13];
+    w3[2] = 0;
+    w3[3] = (64 + esalt_size) * 8;
+
+    u32x digest[5];
+
+    hmac_sha1_run (w0, w1, w2, w3, ipad, opad, digest);
+
+    COMPARE_S_SIMD (digest[3], digest[4], digest[2], digest[1]);
   }
 }
 
diff --git a/OpenCL/m07300_a3.cl b/OpenCL/m07300_a3.cl
index 08459e77f..e778cfdb0 100644
--- a/OpenCL/m07300_a3.cl
+++ b/OpenCL/m07300_a3.cl
@@ -246,6 +246,10 @@ static void m07300m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
   const u32 gid = get_global_id (0);
   const u32 lid = get_local_id (0);
 
+  /**
+   * salt
+   */
+
   const u32 esalt_len = rakp_bufs[salt_pos].salt_len;
 
   /**
@@ -260,37 +264,31 @@ static void m07300m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
 
     const u32x w0lr = w0l | w0r;
 
-    /**
-     * pads
-     */
-
     u32x w0_t[4];
+    u32x w1_t[4];
+    u32x w2_t[4];
+    u32x w3_t[4];
 
     w0_t[0] = w0lr;
     w0_t[1] = w0[1];
     w0_t[2] = w0[2];
     w0_t[3] = w0[3];
-
-    u32x w1_t[4];
-
     w1_t[0] = w1[0];
     w1_t[1] = w1[1];
     w1_t[2] = w1[2];
     w1_t[3] = w1[3];
-
-    u32x w2_t[4];
-
     w2_t[0] = w2[0];
     w2_t[1] = w2[1];
     w2_t[2] = w2[2];
     w2_t[3] = w2[3];
-
-    u32x w3_t[4];
-
     w3_t[0] = w3[0];
     w3_t[1] = w3[1];
-    w3_t[2] = 0;
-    w3_t[3] = 0;
+    w3_t[2] = w3[2];
+    w3_t[3] = w3[3];
+
+    /**
+     * RAKP
+     */
 
     u32x ipad[5];
     u32x opad[5];
@@ -358,6 +356,10 @@ static void m07300s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
   const u32 gid = get_global_id (0);
   const u32 lid = get_local_id (0);
 
+  /**
+   * salt
+   */
+
   const u32 esalt_len = rakp_bufs[salt_pos].salt_len;
 
   /**
@@ -384,37 +386,31 @@ static void m07300s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
 
     const u32x w0lr = w0l | w0r;
 
-    /**
-     * pads
-     */
-
     u32x w0_t[4];
+    u32x w1_t[4];
+    u32x w2_t[4];
+    u32x w3_t[4];
 
     w0_t[0] = w0lr;
     w0_t[1] = w0[1];
     w0_t[2] = w0[2];
     w0_t[3] = w0[3];
-
-    u32x w1_t[4];
-
     w1_t[0] = w1[0];
     w1_t[1] = w1[1];
     w1_t[2] = w1[2];
     w1_t[3] = w1[3];
-
-    u32x w2_t[4];
-
     w2_t[0] = w2[0];
     w2_t[1] = w2[1];
     w2_t[2] = w2[2];
     w2_t[3] = w2[3];
-
-    u32x w3_t[4];
-
     w3_t[0] = w3[0];
     w3_t[1] = w3[1];
-    w3_t[2] = 0;
-    w3_t[3] = 0;
+    w3_t[2] = w3[2];
+    w3_t[3] = w3[3];
+
+    /**
+     * RAKP
+     */
 
     u32x ipad[5];
     u32x opad[5];