From 132086f62008b61f9d6d2e6a1cfcdfe8dd040400 Mon Sep 17 00:00:00 2001
From: Jens Steube <jens.steube@gmail.com>
Date: Fri, 4 Mar 2016 16:54:12 +0100
Subject: [PATCH] Cleanup -m 300 kernels to latest standard

---
 OpenCL/m00300_a0.cl |  36 ++---
 OpenCL/m00300_a1.cl | 360 +++++++++++++++++++-------------------------
 OpenCL/m00300_a3.cl | 111 +++++---------
 3 files changed, 208 insertions(+), 299 deletions(-)

diff --git a/OpenCL/m00300_a0.cl b/OpenCL/m00300_a0.cl
index 8985f51af..c083b0267 100644
--- a/OpenCL/m00300_a0.cl
+++ b/OpenCL/m00300_a0.cl
@@ -39,18 +39,16 @@ __kernel void m00300_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;
 
@@ -346,18 +344,16 @@ __kernel void m00300_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;
 
diff --git a/OpenCL/m00300_a1.cl b/OpenCL/m00300_a1.cl
index 2195ab71e..2af68fee4 100644
--- a/OpenCL/m00300_a1.cl
+++ b/OpenCL/m00300_a1.cl
@@ -5,6 +5,8 @@
 
 #define _SHA1_
 
+#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"
 
 __kernel void m00300_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
 {
@@ -36,142 +36,119 @@ __kernel void m00300_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)
-  {
-    append_0x80_2x4 (wordl0, wordl1, pw_l_len);
-
-    switch_buffer_by_offset_le (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
-  }
-
   /**
    * 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];
+    /**
+     * 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];
+    u32x wordl0[4] = { 0 };
+    u32x wordl1[4] = { 0 };
+    u32x wordl2[4] = { 0 };
+    u32x wordl3[4] = { 0 };
 
-    u32 wordr1[4];
+    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];
 
-    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];
+    u32x wordr0[4] = { 0 };
+    u32x wordr1[4] = { 0 };
+    u32x wordr2[4] = { 0 };
+    u32x wordr3[4] = { 0 };
 
-    u32 wordr2[4];
-
-    wordr2[0] = 0;
-    wordr2[1] = 0;
-    wordr2[2] = 0;
-    wordr2[3] = 0;
-
-    u32 wordr3[4];
-
-    wordr3[0] = 0;
-    wordr3[1] = 0;
-    wordr3[2] = 0;
-    wordr3[3] = 0;
+    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)
     {
-      append_0x80_2x4 (wordr0, wordr1, pw_r_len);
-
-      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] = 0;
-    w3[3] = 0;
+    w3[2] = wordl3[2] | wordr3[2];
+    w3[3] = wordl3[3] | wordr3[3];
 
     /**
      * sha1
      */
 
-    u32 w0_t = swap32 (w0[0]);
-    u32 w1_t = swap32 (w0[1]);
-    u32 w2_t = swap32 (w0[2]);
-    u32 w3_t = swap32 (w0[3]);
-    u32 w4_t = swap32 (w1[0]);
-    u32 w5_t = swap32 (w1[1]);
-    u32 w6_t = swap32 (w1[2]);
-    u32 w7_t = swap32 (w1[3]);
-    u32 w8_t = swap32 (w2[0]);
-    u32 w9_t = swap32 (w2[1]);
-    u32 wa_t = swap32 (w2[2]);
-    u32 wb_t = swap32 (w2[3]);
-    u32 wc_t = swap32 (w3[0]);
-    u32 wd_t = swap32 (w3[1]);
-    u32 we_t = 0;
-    u32 wf_t = pw_len * 8;
+    u32x w0_t = swap32 (w0[0]);
+    u32x w1_t = swap32 (w0[1]);
+    u32x w2_t = swap32 (w0[2]);
+    u32x w3_t = swap32 (w0[3]);
+    u32x w4_t = swap32 (w1[0]);
+    u32x w5_t = swap32 (w1[1]);
+    u32x w6_t = swap32 (w1[2]);
+    u32x w7_t = swap32 (w1[3]);
+    u32x w8_t = swap32 (w2[0]);
+    u32x w9_t = swap32 (w2[1]);
+    u32x wa_t = swap32 (w2[2]);
+    u32x wb_t = swap32 (w2[3]);
+    u32x wc_t = swap32 (w3[0]);
+    u32x wd_t = swap32 (w3[1]);
+    u32x we_t = 0;
+    u32x wf_t = pw_len * 8;
 
-    u32 a = SHA1M_A;
-    u32 b = SHA1M_B;
-    u32 c = SHA1M_C;
-    u32 d = SHA1M_D;
-    u32 e = SHA1M_E;
+    u32x a = SHA1M_A;
+    u32x b = SHA1M_B;
+    u32x c = SHA1M_C;
+    u32x d = SHA1M_D;
+    u32x e = SHA1M_E;
 
     #undef K
     #define K SHA1C00
@@ -394,13 +371,7 @@ __kernel void m00300_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, we_t);
     wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, wf_t);
 
-
-    const u32 r0 = d;
-    const u32 r1 = e;
-    const u32 r2 = c;
-    const u32 r3 = b;
-
-    #include COMPARE_M
+    COMPARE_M_SIMD (d, e, c, b);
   }
 }
 
@@ -428,43 +399,20 @@ __kernel void m00300_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)
-  {
-    append_0x80_2x4 (wordl0, wordl1, pw_l_len);
-
-    switch_buffer_by_offset_le (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
-  }
-
   /**
    * digest
    */
@@ -481,107 +429,107 @@ __kernel void m00300_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    * reverse
    */
 
-  const u32 e_rev = rotl32 (search[1], 2u);
+  const u32 e_rev = rotl32_S (search[1], 2u);
 
   /**
    * 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];
+    /**
+     * 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];
+    u32x wordl0[4] = { 0 };
+    u32x wordl1[4] = { 0 };
+    u32x wordl2[4] = { 0 };
+    u32x wordl3[4] = { 0 };
 
-    u32 wordr1[4];
+    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];
 
-    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];
+    u32x wordr0[4] = { 0 };
+    u32x wordr1[4] = { 0 };
+    u32x wordr2[4] = { 0 };
+    u32x wordr3[4] = { 0 };
 
-    u32 wordr2[4];
-
-    wordr2[0] = 0;
-    wordr2[1] = 0;
-    wordr2[2] = 0;
-    wordr2[3] = 0;
-
-    u32 wordr3[4];
-
-    wordr3[0] = 0;
-    wordr3[1] = 0;
-    wordr3[2] = 0;
-    wordr3[3] = 0;
+    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)
     {
-      append_0x80_2x4 (wordr0, wordr1, pw_r_len);
-
-      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] = 0;
-    w3[3] = 0;
+    w3[2] = wordl3[2] | wordr3[2];
+    w3[3] = wordl3[3] | wordr3[3];
 
     /**
      * sha1
      */
 
-    u32 w0_t = swap32 (w0[0]);
-    u32 w1_t = swap32 (w0[1]);
-    u32 w2_t = swap32 (w0[2]);
-    u32 w3_t = swap32 (w0[3]);
-    u32 w4_t = swap32 (w1[0]);
-    u32 w5_t = swap32 (w1[1]);
-    u32 w6_t = swap32 (w1[2]);
-    u32 w7_t = swap32 (w1[3]);
-    u32 w8_t = swap32 (w2[0]);
-    u32 w9_t = swap32 (w2[1]);
-    u32 wa_t = swap32 (w2[2]);
-    u32 wb_t = swap32 (w2[3]);
-    u32 wc_t = swap32 (w3[0]);
-    u32 wd_t = swap32 (w3[1]);
-    u32 we_t = 0;
-    u32 wf_t = pw_len * 8;
+    u32x w0_t = swap32 (w0[0]);
+    u32x w1_t = swap32 (w0[1]);
+    u32x w2_t = swap32 (w0[2]);
+    u32x w3_t = swap32 (w0[3]);
+    u32x w4_t = swap32 (w1[0]);
+    u32x w5_t = swap32 (w1[1]);
+    u32x w6_t = swap32 (w1[2]);
+    u32x w7_t = swap32 (w1[3]);
+    u32x w8_t = swap32 (w2[0]);
+    u32x w9_t = swap32 (w2[1]);
+    u32x wa_t = swap32 (w2[2]);
+    u32x wb_t = swap32 (w2[3]);
+    u32x wc_t = swap32 (w3[0]);
+    u32x wd_t = swap32 (w3[1]);
+    u32x we_t = 0;
+    u32x wf_t = pw_len * 8;
 
-    u32 a = SHA1M_A;
-    u32 b = SHA1M_B;
-    u32 c = SHA1M_C;
-    u32 d = SHA1M_D;
-    u32 e = SHA1M_E;
+    u32x a = SHA1M_A;
+    u32x b = SHA1M_B;
+    u32x c = SHA1M_C;
+    u32x d = SHA1M_D;
+    u32x e = SHA1M_E;
 
     #undef K
     #define K SHA1C00
@@ -800,20 +748,14 @@ __kernel void m00300_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, wa_t);
     wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, wb_t);
 
-    if (allx (e != e_rev)) continue;
+    if (MATCHES_NONE_VS (e, e_rev)) continue;
 
     wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wc_t);
     wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, wd_t);
     we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, we_t);
     wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, wf_t);
 
-
-    const u32 r0 = d;
-    const u32 r1 = e;
-    const u32 r2 = c;
-    const u32 r3 = b;
-
-    #include COMPARE_S
+    COMPARE_S_SIMD (d, e, c, b);
   }
 }
 
diff --git a/OpenCL/m00300_a3.cl b/OpenCL/m00300_a3.cl
index 162c8a603..19953e2ee 100644
--- a/OpenCL/m00300_a3.cl
+++ b/OpenCL/m00300_a3.cl
@@ -93,10 +93,6 @@ static void m00300m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global k
   const u32 c_73s = rotl32_S ((c_70s ^ c_65s ^ c_59s ^ c_57s), 1u);
   const u32 c_74s = rotl32_S ((c_71s ^ c_66s ^ c_60s ^ c_58s), 1u);
   const u32 c_75s = rotl32_S ((c_72s ^ c_67s ^ c_61s ^ c_59s), 1u);
-  const u32 c_76s = rotl32_S ((c_73s ^ c_68s ^ c_62s ^ c_60s), 1u);
-  const u32 c_77s = rotl32_S ((c_74s ^ c_69s ^ c_63s ^ c_61s), 1u);
-  const u32 c_78s = rotl32_S ((c_75s ^ c_70s ^ c_64s ^ c_62s), 1u);
-  const u32 c_79s = rotl32_S ((c_76s ^ c_71s ^ c_65s ^ c_63s), 1u);
 
   const u32 c_17sK = c_17s + SHA1C00;
   const u32 c_18sK = c_18s + SHA1C00;
@@ -114,23 +110,7 @@ static void m00300m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global k
   const u32 c_65sK = c_65s + SHA1C03;
   const u32 c_69sK = c_69s + SHA1C03;
 
-  const u32 w1 = w[ 1] + SHA1C00;
-  const u32 w2 = w[ 2] + SHA1C00;
-  const u32 w3 = w[ 3] + SHA1C00;
-  const u32 w4 = w[ 4] + SHA1C00;
-  const u32 w5 = w[ 5] + SHA1C00;
-  const u32 w6 = w[ 6] + SHA1C00;
-  const u32 w7 = w[ 7] + SHA1C00;
-  const u32 w8 = w[ 8] + SHA1C00;
-  const u32 w9 = w[ 9] + SHA1C00;
-  const u32 wa = w[10] + SHA1C00;
-  const u32 wb = w[11] + SHA1C00;
-  const u32 wc = w[12] + SHA1C00;
-  const u32 wd = w[13] + SHA1C00;
-  const u32 we = w[14] + SHA1C00;
-  const u32 wf = w[15] + SHA1C00;
-
- /**
+  /**
    * loop
    */
 
@@ -162,8 +142,6 @@ static void m00300m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global k
     const u32x w0s18 = rotl32 (w0, 18u);
     const u32x w0s19 = rotl32 (w0, 19u);
     const u32x w0s20 = rotl32 (w0, 20u);
-    const u32x w0s21 = rotl32 (w0, 21u);
-    const u32x w0s22 = rotl32 (w0, 22U);
 
     const u32x w0s04___w0s06 = w0s04 ^ w0s06;
     const u32x w0s04___w0s08 = w0s04 ^ w0s08;
@@ -180,21 +158,21 @@ static void m00300m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global k
     #define K SHA1C00
 
     SHA1_STEP (SHA1_F0o, a, b, c, d, e, w0);
-    SHA1_STEPX(SHA1_F0o, e, a, b, c, d, w1);
-    SHA1_STEPX(SHA1_F0o, d, e, a, b, c, w2);
-    SHA1_STEPX(SHA1_F0o, c, d, e, a, b, w3);
-    SHA1_STEPX(SHA1_F0o, b, c, d, e, a, w4);
-    SHA1_STEPX(SHA1_F0o, a, b, c, d, e, w5);
-    SHA1_STEPX(SHA1_F0o, e, a, b, c, d, w6);
-    SHA1_STEPX(SHA1_F0o, d, e, a, b, c, w7);
-    SHA1_STEPX(SHA1_F0o, c, d, e, a, b, w8);
-    SHA1_STEPX(SHA1_F0o, b, c, d, e, a, w9);
-    SHA1_STEPX(SHA1_F0o, a, b, c, d, e, wa);
-    SHA1_STEPX(SHA1_F0o, e, a, b, c, d, wb);
-    SHA1_STEPX(SHA1_F0o, d, e, a, b, c, wc);
-    SHA1_STEPX(SHA1_F0o, c, d, e, a, b, wd);
-    SHA1_STEPX(SHA1_F0o, b, c, d, e, a, we);
-    SHA1_STEPX(SHA1_F0o, a, b, c, d, e, wf);
+    SHA1_STEP (SHA1_F0o, e, a, b, c, d, w[ 1]);
+    SHA1_STEP (SHA1_F0o, d, e, a, b, c, w[ 2]);
+    SHA1_STEP (SHA1_F0o, c, d, e, a, b, w[ 3]);
+    SHA1_STEP (SHA1_F0o, b, c, d, e, a, w[ 4]);
+    SHA1_STEP (SHA1_F0o, a, b, c, d, e, w[ 5]);
+    SHA1_STEP (SHA1_F0o, e, a, b, c, d, w[ 6]);
+    SHA1_STEP (SHA1_F0o, d, e, a, b, c, w[ 7]);
+    SHA1_STEP (SHA1_F0o, c, d, e, a, b, w[ 8]);
+    SHA1_STEP (SHA1_F0o, b, c, d, e, a, w[ 9]);
+    SHA1_STEP (SHA1_F0o, a, b, c, d, e, w[10]);
+    SHA1_STEP (SHA1_F0o, e, a, b, c, d, w[11]);
+    SHA1_STEP (SHA1_F0o, d, e, a, b, c, w[12]);
+    SHA1_STEP (SHA1_F0o, c, d, e, a, b, w[13]);
+    SHA1_STEP (SHA1_F0o, b, c, d, e, a, w[14]);
+    SHA1_STEP (SHA1_F0o, a, b, c, d, e, w[15]);
 
     SHA1_STEP (SHA1_F0o, e, a, b, c, d, (c_16s ^ w0s01));
     SHA1_STEPX(SHA1_F0o, d, e, a, b, c, (c_17sK));
@@ -268,6 +246,15 @@ static void m00300m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global k
     SHA1_STEP (SHA1_F1 , c, d, e, a, b, (c_73s ^ w0s20));
     SHA1_STEP (SHA1_F1 , b, c, d, e, a, (c_74s ^ w0s08 ^ w0s16));
     SHA1_STEP (SHA1_F1 , a, b, c, d, e, (c_75s ^ w0s06 ^ w0s12 ^ w0s14));
+
+    const u32x c_76s = rotl32 ((c_73s ^ c_68s ^ c_62s ^ c_60s), 1u);
+    const u32x c_77s = rotl32 ((c_74s ^ c_69s ^ c_63s ^ c_61s), 1u);
+    const u32x c_78s = rotl32 ((c_75s ^ c_70s ^ c_64s ^ c_62s), 1u);
+    const u32x c_79s = rotl32 ((c_76s ^ c_71s ^ c_65s ^ c_63s), 1u);
+
+    const u32x w0s21 = rotl32 (w0, 21u);
+    const u32x w0s22 = rotl32 (w0, 22U);
+
     SHA1_STEP (SHA1_F1 , e, a, b, c, d, (c_76s ^ w0s07 ^ w0s08___w0s12 ^ w0s16 ^ w0s21));
     SHA1_STEP (SHA1_F1 , d, e, a, b, c, (c_77s));
     SHA1_STEP (SHA1_F1 , c, d, e, a, b, (c_78s ^ w0s07 ^ w0s08 ^ w0s15 ^ w0s18 ^ w0s20));
@@ -492,22 +479,6 @@ static void m00300s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global k
   const u32 c_65sK = c_65s + SHA1C03;
   const u32 c_69sK = c_69s + SHA1C03;
 
-  const u32 w1 = w[ 1] + SHA1C00;
-  const u32 w2 = w[ 2] + SHA1C00;
-  const u32 w3 = w[ 3] + SHA1C00;
-  const u32 w4 = w[ 4] + SHA1C00;
-  const u32 w5 = w[ 5] + SHA1C00;
-  const u32 w6 = w[ 6] + SHA1C00;
-  const u32 w7 = w[ 7] + SHA1C00;
-  const u32 w8 = w[ 8] + SHA1C00;
-  const u32 w9 = w[ 9] + SHA1C00;
-  const u32 wa = w[10] + SHA1C00;
-  const u32 wb = w[11] + SHA1C00;
-  const u32 wc = w[12] + SHA1C00;
-  const u32 wd = w[13] + SHA1C00;
-  const u32 we = w[14] + SHA1C00;
-  const u32 wf = w[15] + SHA1C00;
-
   /**
    * digest
    */
@@ -526,7 +497,7 @@ static void m00300s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global k
 
   const u32 e_rev = rotl32_S (search[1], 2u);
 
- /**
+  /**
    * loop
    */
 
@@ -574,21 +545,21 @@ static void m00300s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global k
     #define K SHA1C00
 
     SHA1_STEP (SHA1_F0o, a, b, c, d, e, w0);
-    SHA1_STEPX(SHA1_F0o, e, a, b, c, d, w1);
-    SHA1_STEPX(SHA1_F0o, d, e, a, b, c, w2);
-    SHA1_STEPX(SHA1_F0o, c, d, e, a, b, w3);
-    SHA1_STEPX(SHA1_F0o, b, c, d, e, a, w4);
-    SHA1_STEPX(SHA1_F0o, a, b, c, d, e, w5);
-    SHA1_STEPX(SHA1_F0o, e, a, b, c, d, w6);
-    SHA1_STEPX(SHA1_F0o, d, e, a, b, c, w7);
-    SHA1_STEPX(SHA1_F0o, c, d, e, a, b, w8);
-    SHA1_STEPX(SHA1_F0o, b, c, d, e, a, w9);
-    SHA1_STEPX(SHA1_F0o, a, b, c, d, e, wa);
-    SHA1_STEPX(SHA1_F0o, e, a, b, c, d, wb);
-    SHA1_STEPX(SHA1_F0o, d, e, a, b, c, wc);
-    SHA1_STEPX(SHA1_F0o, c, d, e, a, b, wd);
-    SHA1_STEPX(SHA1_F0o, b, c, d, e, a, we);
-    SHA1_STEPX(SHA1_F0o, a, b, c, d, e, wf);
+    SHA1_STEP (SHA1_F0o, e, a, b, c, d, w[ 1]);
+    SHA1_STEP (SHA1_F0o, d, e, a, b, c, w[ 2]);
+    SHA1_STEP (SHA1_F0o, c, d, e, a, b, w[ 3]);
+    SHA1_STEP (SHA1_F0o, b, c, d, e, a, w[ 4]);
+    SHA1_STEP (SHA1_F0o, a, b, c, d, e, w[ 5]);
+    SHA1_STEP (SHA1_F0o, e, a, b, c, d, w[ 6]);
+    SHA1_STEP (SHA1_F0o, d, e, a, b, c, w[ 7]);
+    SHA1_STEP (SHA1_F0o, c, d, e, a, b, w[ 8]);
+    SHA1_STEP (SHA1_F0o, b, c, d, e, a, w[ 9]);
+    SHA1_STEP (SHA1_F0o, a, b, c, d, e, w[10]);
+    SHA1_STEP (SHA1_F0o, e, a, b, c, d, w[11]);
+    SHA1_STEP (SHA1_F0o, d, e, a, b, c, w[12]);
+    SHA1_STEP (SHA1_F0o, c, d, e, a, b, w[13]);
+    SHA1_STEP (SHA1_F0o, b, c, d, e, a, w[14]);
+    SHA1_STEP (SHA1_F0o, a, b, c, d, e, w[15]);
 
     SHA1_STEP (SHA1_F0o, e, a, b, c, d, (c_16s ^ w0s01));
     SHA1_STEPX(SHA1_F0o, d, e, a, b, c, (c_17sK));