Cleanup -m 110xx kernels to latest standard

OpenCL/m11000_a0.cl

@@ -39,18 +39,16 @@ __kernel void m11000_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_buf0[0] = pws[gid].i[0];
-  pw_buf1[1] = pws[gid].i[ 5];
+  pw_buf0[1] = pws[gid].i[1];
-  pw_buf1[2] = pws[gid].i[ 6];
+  pw_buf0[2] = pws[gid].i[2];
-  pw_buf1[3] = pws[gid].i[ 7];
+  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;
 
@@ -58,32 +56,27 @@ __kernel void m11000_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    * salt
    */
 
-
   u32 salt_buf0[4];
+  u32 salt_buf1[4];
+  u32 salt_buf2[4];
+  u32 salt_buf3[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] = 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];
-
-  u32 salt_buf3[2];
-
   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;
 
@@ -109,6 +102,7 @@ __kernel void m11000_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
      */
 
     // first step fixed 56 bytes of salt
+    // after 56 byte salt, we have beginning of the password
 
     u32x w0_t[4];
     u32x w1_t[4];
@@ -129,9 +123,6 @@ __kernel void m11000_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     w2_t[3] = salt_buf2[3];
     w3_t[0] = salt_buf3[0];
     w3_t[1] = salt_buf3[1];
-
-    // after 56 byte salt, we have beginning of the password
-
     w3_t[2] = w0[0];
     w3_t[3] = w0[1];
 
@@ -139,8 +130,6 @@ __kernel void m11000_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
      * md5
      */
 
-    // first transform
-
     u32x a = MD5M_A;
     u32x b = MD5M_B;
     u32x c = MD5M_C;
@@ -232,14 +221,14 @@ __kernel void m11000_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     w0_t[3] = w1[1];
     w1_t[0] = w1[2];
     w1_t[1] = w1[3];
-    w1_t[2] = w2[0];
+    w1_t[2] = 0;
-    w1_t[3] = w2[1];
+    w1_t[3] = 0;
-    w2_t[0] = w2[2];
+    w2_t[0] = 0;
-    w2_t[1] = w2[3];
+    w2_t[1] = 0;
-    w2_t[2] = w3[0];
+    w2_t[2] = 0;
-    w2_t[3] = w3[1];
+    w2_t[3] = 0;
-    w3_t[0] = w3[2];
+    w3_t[0] = 0;
-    w3_t[1] = w3[3];
+    w3_t[1] = 0;
     w3_t[2] = pw_salt_len * 8;
     w3_t[3] = 0;
 
@@ -345,18 +334,16 @@ __kernel void m11000_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_buf0[0] = pws[gid].i[0];
-  pw_buf1[1] = pws[gid].i[ 5];
+  pw_buf0[1] = pws[gid].i[1];
-  pw_buf1[2] = pws[gid].i[ 6];
+  pw_buf0[2] = pws[gid].i[2];
-  pw_buf1[3] = pws[gid].i[ 7];
+  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;
 
@@ -364,32 +351,27 @@ __kernel void m11000_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    * salt
    */
 
-
   u32 salt_buf0[4];
+  u32 salt_buf1[4];
+  u32 salt_buf2[4];
+  u32 salt_buf3[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] = 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];
-
-  u32 salt_buf3[2];
-
   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;
 
@@ -427,6 +409,7 @@ __kernel void m11000_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
      */
 
     // first step fixed 56 bytes of salt
+    // after 56 byte salt, we have beginning of the password
 
     u32x w0_t[4];
     u32x w1_t[4];
@@ -447,9 +430,6 @@ __kernel void m11000_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     w2_t[3] = salt_buf2[3];
     w3_t[0] = salt_buf3[0];
     w3_t[1] = salt_buf3[1];
-
-    // after 56 byte salt, we have beginning of the password
-
     w3_t[2] = w0[0];
     w3_t[3] = w0[1];
 
@@ -457,8 +437,6 @@ __kernel void m11000_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
      * md5
      */
 
-    // first transform
-
     u32x a = MD5M_A;
     u32x b = MD5M_B;
     u32x c = MD5M_C;
@@ -550,14 +528,14 @@ __kernel void m11000_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     w0_t[3] = w1[1];
     w1_t[0] = w1[2];
     w1_t[1] = w1[3];
-    w1_t[2] = w2[0];
+    w1_t[2] = 0;
-    w1_t[3] = w2[1];
+    w1_t[3] = 0;
-    w2_t[0] = w2[2];
+    w2_t[0] = 0;
-    w2_t[1] = w2[3];
+    w2_t[1] = 0;
-    w2_t[2] = w3[0];
+    w2_t[2] = 0;
-    w2_t[3] = w3[1];
+    w2_t[3] = 0;
-    w3_t[0] = w3[2];
+    w3_t[0] = 0;
-    w3_t[1] = w3[3];
+    w3_t[1] = 0;
     w3_t[2] = pw_salt_len * 8;
     w3_t[3] = 0;
 

OpenCL/m11000_a1.cl

@@ -5,6 +5,8 @@
 
 #define _MD5_
 
+#define NEW_SIMD_CODE
+
 #include "include/constants.h"
 #include "include/kernel_vendor.h"
 
@@ -16,9 +18,7 @@
 #include "include/kernel_functions.c"
 #include "OpenCL/types_ocl.c"
 #include "OpenCL/common.c"
-
+#include "OpenCL/simd.c"
-#define COMPARE_S "OpenCL/check_single_comp4.c"
-#define COMPARE_M "OpenCL/check_multi_comp4.c"
 
 __kernel void m11000_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)
 {
@@ -36,70 +36,45 @@ __kernel void m11000_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];
+  pw_buf0[0] = pws[gid].i[0];
-  wordl0[1] = pws[gid].i[ 1];
+  pw_buf0[1] = pws[gid].i[1];
-  wordl0[2] = pws[gid].i[ 2];
+  pw_buf0[2] = pws[gid].i[2];
-  wordl0[3] = pws[gid].i[ 3];
+  pw_buf0[3] = pws[gid].i[3];
-
+  pw_buf1[0] = pws[gid].i[4];
-  u32 wordl1[4];
+  pw_buf1[1] = pws[gid].i[5];
-
+  pw_buf1[2] = pws[gid].i[6];
-  wordl1[0] = pws[gid].i[ 4];
+  pw_buf1[3] = pws[gid].i[7];
-  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;
 
   const u32 pw_l_len = pws[gid].pw_len;
 
-  if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
-  {
-    switch_buffer_by_offset_le (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
-  }
-
   /**
    * salt
    */
 
   u32 salt_buf0[4];
+  u32 salt_buf1[4];
+  u32 salt_buf2[4];
+  u32 salt_buf3[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] = 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];
-
-  u32 salt_buf3[2];
-
   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;
 
@@ -107,43 +82,57 @@ __kernel void m11000_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];
+     * concat password candidate
-    u32 wordr2[4];
+     */
-    u32 wordr3[4];
 
-    wordr0[0] = combs_buf[il_pos].i[0];
+    u32x wordl0[4] = { 0 };
-    wordr0[1] = combs_buf[il_pos].i[1];
+    u32x wordl1[4] = { 0 };
-    wordr0[2] = combs_buf[il_pos].i[2];
+    u32x wordl2[4] = { 0 };
-    wordr0[3] = combs_buf[il_pos].i[3];
+    u32x wordl3[4] = { 0 };
-    wordr1[0] = combs_buf[il_pos].i[4];
+
-    wordr1[1] = combs_buf[il_pos].i[5];
+    wordl0[0] = pw_buf0[0];
-    wordr1[2] = combs_buf[il_pos].i[6];
+    wordl0[1] = pw_buf0[1];
-    wordr1[3] = combs_buf[il_pos].i[7];
+    wordl0[2] = pw_buf0[2];
-    wordr2[0] = 0;
+    wordl0[3] = pw_buf0[3];
-    wordr2[1] = 0;
+    wordl1[0] = pw_buf1[0];
-    wordr2[2] = 0;
+    wordl1[1] = pw_buf1[1];
-    wordr2[3] = 0;
+    wordl1[2] = pw_buf1[2];
-    wordr3[0] = 0;
+    wordl1[3] = pw_buf1[3];
-    wordr3[1] = 0;
+
-    wordr3[2] = 0;
+    u32x wordr0[4] = { 0 };
-    wordr3[3] = 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];
-    u32 w1[4];
+    u32x w1[4];
-    u32 w2[4];
+    u32x w2[4];
-    u32 w3[4];
+    u32x w3[4];
 
     w0[0] = wordl0[0] | wordr0[0];
     w0[1] = wordl0[1] | wordr0[1];
@@ -162,20 +151,19 @@ __kernel void m11000_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     w3[2] = wordl3[2] | wordr3[2];
     w3[3] = wordl3[3] | wordr3[3];
 
-    append_0x80_4x4 (w0, w1, w2, w3, pw_len);
+    const u32x pw_salt_len = salt_len + pw_len;
-
-    const u32 pw_salt_len = pw_len + salt_len;
 
     /**
      * prepend salt
      */
 
     // first step fixed 56 bytes of salt
+    // after 56 byte salt, we have beginning of the password
 
-    u32 w0_t[4];
+    u32x w0_t[4];
-    u32 w1_t[4];
+    u32x w1_t[4];
-    u32 w2_t[4];
+    u32x w2_t[4];
-    u32 w3_t[4];
+    u32x w3_t[4];
 
     w0_t[0] = salt_buf0[0];
     w0_t[1] = salt_buf0[1];
@@ -191,9 +179,6 @@ __kernel void m11000_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     w2_t[3] = salt_buf2[3];
     w3_t[0] = salt_buf3[0];
     w3_t[1] = salt_buf3[1];
-
-    // after 56 byte salt, we have beginning of the password
-
     w3_t[2] = w0[0];
     w3_t[3] = w0[1];
 
@@ -201,12 +186,10 @@ __kernel void m11000_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
      * md5
      */
 
-    // first transform
+    u32x a = MD5M_A;
-
+    u32x b = MD5M_B;
-    u32 a = MD5M_A;
+    u32x c = MD5M_C;
-    u32 b = MD5M_B;
+    u32x d = MD5M_D;
-    u32 c = MD5M_C;
-    u32 d = MD5M_D;
 
     MD5_STEP (MD5_Fo, a, b, c, d, w0_t[0], MD5C00, MD5S00);
     MD5_STEP (MD5_Fo, d, a, b, c, w0_t[1], MD5C01, MD5S01);
@@ -281,10 +264,10 @@ __kernel void m11000_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     c += MD5M_C;
     d += MD5M_D;
 
-    u32 r_a = a;
+    u32x r_a = a;
-    u32 r_b = b;
+    u32x r_b = b;
-    u32 r_c = c;
+    u32x r_c = c;
-    u32 r_d = d;
+    u32x r_d = d;
 
     // 2nd transform
 
@@ -378,12 +361,7 @@ __kernel void m11000_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     c += r_c;
     d += r_d;
 
-    const u32 r0 = a;
+    COMPARE_M_SIMD (a, d, c, b);
-    const u32 r1 = d;
-    const u32 r2 = c;
-    const u32 r3 = b;
-
-    #include COMPARE_M
   }
 }
 
@@ -411,70 +389,45 @@ __kernel void m11000_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];
+  pw_buf0[0] = pws[gid].i[0];
-  wordl0[1] = pws[gid].i[ 1];
+  pw_buf0[1] = pws[gid].i[1];
-  wordl0[2] = pws[gid].i[ 2];
+  pw_buf0[2] = pws[gid].i[2];
-  wordl0[3] = pws[gid].i[ 3];
+  pw_buf0[3] = pws[gid].i[3];
-
+  pw_buf1[0] = pws[gid].i[4];
-  u32 wordl1[4];
+  pw_buf1[1] = pws[gid].i[5];
-
+  pw_buf1[2] = pws[gid].i[6];
-  wordl1[0] = pws[gid].i[ 4];
+  pw_buf1[3] = pws[gid].i[7];
-  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;
 
   const u32 pw_l_len = pws[gid].pw_len;
 
-  if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
-  {
-    switch_buffer_by_offset_le (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
-  }
-
   /**
    * salt
    */
 
   u32 salt_buf0[4];
+  u32 salt_buf1[4];
+  u32 salt_buf2[4];
+  u32 salt_buf3[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] = 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];
-
-  u32 salt_buf3[2];
-
   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;
 
@@ -494,43 +447,57 @@ __kernel void m11000_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];
+     * concat password candidate
-    u32 wordr2[4];
+     */
-    u32 wordr3[4];
 
-    wordr0[0] = combs_buf[il_pos].i[0];
+    u32x wordl0[4] = { 0 };
-    wordr0[1] = combs_buf[il_pos].i[1];
+    u32x wordl1[4] = { 0 };
-    wordr0[2] = combs_buf[il_pos].i[2];
+    u32x wordl2[4] = { 0 };
-    wordr0[3] = combs_buf[il_pos].i[3];
+    u32x wordl3[4] = { 0 };
-    wordr1[0] = combs_buf[il_pos].i[4];
+
-    wordr1[1] = combs_buf[il_pos].i[5];
+    wordl0[0] = pw_buf0[0];
-    wordr1[2] = combs_buf[il_pos].i[6];
+    wordl0[1] = pw_buf0[1];
-    wordr1[3] = combs_buf[il_pos].i[7];
+    wordl0[2] = pw_buf0[2];
-    wordr2[0] = 0;
+    wordl0[3] = pw_buf0[3];
-    wordr2[1] = 0;
+    wordl1[0] = pw_buf1[0];
-    wordr2[2] = 0;
+    wordl1[1] = pw_buf1[1];
-    wordr2[3] = 0;
+    wordl1[2] = pw_buf1[2];
-    wordr3[0] = 0;
+    wordl1[3] = pw_buf1[3];
-    wordr3[1] = 0;
+
-    wordr3[2] = 0;
+    u32x wordr0[4] = { 0 };
-    wordr3[3] = 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];
-    u32 w1[4];
+    u32x w1[4];
-    u32 w2[4];
+    u32x w2[4];
-    u32 w3[4];
+    u32x w3[4];
 
     w0[0] = wordl0[0] | wordr0[0];
     w0[1] = wordl0[1] | wordr0[1];
@@ -549,20 +516,19 @@ __kernel void m11000_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     w3[2] = wordl3[2] | wordr3[2];
     w3[3] = wordl3[3] | wordr3[3];
 
-    append_0x80_4x4 (w0, w1, w2, w3, pw_len);
+    const u32x pw_salt_len = salt_len + pw_len;
-
-    const u32 pw_salt_len = pw_len + salt_len;
 
     /**
      * prepend salt
      */
 
     // first step fixed 56 bytes of salt
+    // after 56 byte salt, we have beginning of the password
 
-    u32 w0_t[4];
+    u32x w0_t[4];
-    u32 w1_t[4];
+    u32x w1_t[4];
-    u32 w2_t[4];
+    u32x w2_t[4];
-    u32 w3_t[4];
+    u32x w3_t[4];
 
     w0_t[0] = salt_buf0[0];
     w0_t[1] = salt_buf0[1];
@@ -578,9 +544,6 @@ __kernel void m11000_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     w2_t[3] = salt_buf2[3];
     w3_t[0] = salt_buf3[0];
     w3_t[1] = salt_buf3[1];
-
-    // after 56 byte salt, we have beginning of the password
-
     w3_t[2] = w0[0];
     w3_t[3] = w0[1];
 
@@ -588,12 +551,10 @@ __kernel void m11000_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
      * md5
      */
 
-    // first transform
+    u32x a = MD5M_A;
-
+    u32x b = MD5M_B;
-    u32 a = MD5M_A;
+    u32x c = MD5M_C;
-    u32 b = MD5M_B;
+    u32x d = MD5M_D;
-    u32 c = MD5M_C;
-    u32 d = MD5M_D;
 
     MD5_STEP (MD5_Fo, a, b, c, d, w0_t[0], MD5C00, MD5S00);
     MD5_STEP (MD5_Fo, d, a, b, c, w0_t[1], MD5C01, MD5S01);
@@ -668,10 +629,10 @@ __kernel void m11000_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     c += MD5M_C;
     d += MD5M_D;
 
-    u32 r_a = a;
+    u32x r_a = a;
-    u32 r_b = b;
+    u32x r_b = b;
-    u32 r_c = c;
+    u32x r_c = c;
-    u32 r_d = d;
+    u32x r_d = d;
 
     // 2nd transform
 
@@ -765,12 +726,7 @@ __kernel void m11000_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     c += r_c;
     d += r_d;
 
-    const u32 r0 = a;
+    COMPARE_S_SIMD (a, d, c, b);
-    const u32 r1 = d;
-    const u32 r2 = c;
-    const u32 r3 = b;
-
-    #include COMPARE_S
   }
 }
 

OpenCL/m11000_a3.cl

@@ -34,30 +34,26 @@ static void m11000m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
    */
 
   u32 salt_buf0[4];
+  u32 salt_buf1[4];
+  u32 salt_buf2[4];
+  u32 salt_buf3[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] = 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];
-
-  u32 salt_buf3[2];
-
   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;
 
@@ -80,6 +76,7 @@ static void m11000m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
      */
 
     // first step fixed 56 bytes of salt
+    // after 56 byte salt, we have beginning of the password
 
     u32x w0_t[4];
     u32x w1_t[4];
@@ -100,9 +97,6 @@ static void m11000m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
     w2_t[3] = salt_buf2[3];
     w3_t[0] = salt_buf3[0];
     w3_t[1] = salt_buf3[1];
-
-    // after 56 byte salt, we have beginning of the password
-
     w3_t[2] = w0lr;
     w3_t[3] = w0[1];
 
@@ -110,8 +104,6 @@ static void m11000m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
      * md5
      */
 
-    // first transform
-
     u32x a = MD5M_A;
     u32x b = MD5M_B;
     u32x c = MD5M_C;
@@ -300,6 +292,36 @@ static void m11000s (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
+   */
+
+  u32 salt_buf0[4];
+  u32 salt_buf1[4];
+  u32 salt_buf2[4];
+  u32 salt_buf3[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];
+  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;
+
+  const u32 pw_salt_len = pw_len + salt_len;
+
   /**
    * digest
    */
@@ -312,40 +334,6 @@ static void m11000s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
     digests_buf[digests_offset].digest_buf[DGST_R3]
   };
 
-  /**
-   * 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];
-
-  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] = 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];
-
-  u32 salt_buf3[2];
-
-  salt_buf3[0] = salt_bufs[salt_pos].salt_buf[12];
-  salt_buf3[1] = salt_bufs[salt_pos].salt_buf[13];
-
-  const u32 salt_len = salt_bufs[salt_pos].salt_len;
-
-  const u32 pw_salt_len = pw_len + salt_len;
-
   /**
    * loop
    */
@@ -363,6 +351,7 @@ static void m11000s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
      */
 
     // first step fixed 56 bytes of salt
+    // after 56 byte salt, we have beginning of the password
 
     u32x w0_t[4];
     u32x w1_t[4];
@@ -383,9 +372,6 @@ static void m11000s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
     w2_t[3] = salt_buf2[3];
     w3_t[0] = salt_buf3[0];
     w3_t[1] = salt_buf3[1];
-
-    // after 56 byte salt, we have beginning of the password
-
     w3_t[2] = w0lr;
     w3_t[3] = w0[1];
 
@@ -393,8 +379,6 @@ static void m11000s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
      * md5
      */
 
-    // first transform
-
     u32x a = MD5M_A;
     u32x b = MD5M_B;
     u32x c = MD5M_C;