Increase salt length for raw-md5 based algorithms

diff --git a/OpenCL/m00010_a0.cl b/OpenCL/m00010_a0.cl
index 2c8cebf..85260a9 100644 (file)
--- a/OpenCL/m00010_a0.cl
+++ b/OpenCL/m00010_a0.cl
@@ -59,18 +59,26 @@ __kernel void m00010_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];
-
-  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];
+  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;
 
@@ -91,33 +99,27 @@ __kernel void m00010_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
      * append salt
      */
 
-    u32x s0[4];
+    u32x s0[4] = { 0 };
+    u32x s1[4] = { 0 };
+    u32x s2[4] = { 0 };
+    u32x s3[4] = { 0 };
 
     s0[0] = salt_buf0[0];
     s0[1] = salt_buf0[1];
     s0[2] = salt_buf0[2];
     s0[3] = salt_buf0[3];
-
-    u32x s1[4];
-
     s1[0] = salt_buf1[0];
     s1[1] = salt_buf1[1];
     s1[2] = salt_buf1[2];
     s1[3] = salt_buf1[3];
-
-    u32x s2[4];
-
-    s2[0] = 0;
-    s2[1] = 0;
-    s2[2] = 0;
-    s2[3] = 0;
-
-    u32x s3[4];
-
-    s3[0] = 0;
-    s3[1] = 0;
-    s3[2] = 0;
-    s3[3] = 0;
+    s2[0] = salt_buf2[0];
+    s2[1] = salt_buf2[1];
+    s2[2] = salt_buf2[2];
+    s2[3] = salt_buf2[3];
+    s3[0] = salt_buf3[0];
+    s3[1] = salt_buf3[1];
+    s3[2] = salt_buf3[2];
+    s3[3] = salt_buf3[3];
 
     switch_buffer_by_offset_le (s0, s1, s2, s3, out_len);
 
@@ -127,24 +129,19 @@ __kernel void m00010_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     w0[1] |= s0[1];
     w0[2] |= s0[2];
     w0[3] |= s0[3];
-
     w1[0] |= s1[0];
     w1[1] |= s1[1];
     w1[2] |= s1[2];
     w1[3] |= s1[3];
-
     w2[0] |= s2[0];
     w2[1] |= s2[1];
     w2[2] |= s2[2];
     w2[3] |= s2[3];
-
     w3[0] |= s3[0];
     w3[1] |= s3[1];
     w3[2]  = pw_salt_len * 8;
     w3[3]  = 0;
 
-    append_0x80_4x4 (w0, w1, w2, w3, pw_salt_len);
-
     /**
      * md5
      */
@@ -271,18 +268,26 @@ __kernel void m00010_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];
-
-  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];
+  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;
 
@@ -315,33 +320,27 @@ __kernel void m00010_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
      * append salt
      */
 
-    u32x s0[4];
+    u32x s0[4] = { 0 };
+    u32x s1[4] = { 0 };
+    u32x s2[4] = { 0 };
+    u32x s3[4] = { 0 };
 
     s0[0] = salt_buf0[0];
     s0[1] = salt_buf0[1];
     s0[2] = salt_buf0[2];
     s0[3] = salt_buf0[3];
-
-    u32x s1[4];
-
     s1[0] = salt_buf1[0];
     s1[1] = salt_buf1[1];
     s1[2] = salt_buf1[2];
     s1[3] = salt_buf1[3];
-
-    u32x s2[4];
-
-    s2[0] = 0;
-    s2[1] = 0;
-    s2[2] = 0;
-    s2[3] = 0;
-
-    u32x s3[4];
-
-    s3[0] = 0;
-    s3[1] = 0;
-    s3[2] = 0;
-    s3[3] = 0;
+    s2[0] = salt_buf2[0];
+    s2[1] = salt_buf2[1];
+    s2[2] = salt_buf2[2];
+    s2[3] = salt_buf2[3];
+    s3[0] = salt_buf3[0];
+    s3[1] = salt_buf3[1];
+    s3[2] = salt_buf3[2];
+    s3[3] = salt_buf3[3];
 
     switch_buffer_by_offset_le (s0, s1, s2, s3, out_len);
 
@@ -351,24 +350,19 @@ __kernel void m00010_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     w0[1] |= s0[1];
     w0[2] |= s0[2];
     w0[3] |= s0[3];
-
     w1[0] |= s1[0];
     w1[1] |= s1[1];
     w1[2] |= s1[2];
     w1[3] |= s1[3];
-
     w2[0] |= s2[0];
     w2[1] |= s2[1];
     w2[2] |= s2[2];
     w2[3] |= s2[3];
-
     w3[0] |= s3[0];
     w3[1] |= s3[1];
     w3[2]  = pw_salt_len * 8;
     w3[3]  = 0;
 
-    append_0x80_4x4 (w0, w1, w2, w3, pw_salt_len);
-
     /**
      * md5
      */

diff --git a/OpenCL/m00010_a1.cl b/OpenCL/m00010_a1.cl
index adcf7d5..3ce73f1 100644 (file)
--- a/OpenCL/m00010_a1.cl
+++ b/OpenCL/m00010_a1.cl
@@ -56,15 +56,25 @@ __kernel void m00010_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 
   u32 salt_buf0[4];
   u32 salt_buf1[4];
-
-  salt_buf0[0] = salt_bufs[salt_pos].salt_buf[0];
-  salt_buf0[1] = salt_bufs[salt_pos].salt_buf[1];
-  salt_buf0[2] = salt_bufs[salt_pos].salt_buf[2];
-  salt_buf0[3] = salt_bufs[salt_pos].salt_buf[3];
-  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];
+  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;
 
@@ -138,8 +148,8 @@ __kernel void m00010_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     w2[3] = wordl2[3] | wordr2[3];
     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];
 
     /**
      * append salt
@@ -158,6 +168,14 @@ __kernel void m00010_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     s1[1] = salt_buf1[1];
     s1[2] = salt_buf1[2];
     s1[3] = salt_buf1[3];
+    s2[0] = salt_buf2[0];
+    s2[1] = salt_buf2[1];
+    s2[2] = salt_buf2[2];
+    s2[3] = salt_buf2[3];
+    s3[0] = salt_buf3[0];
+    s3[1] = salt_buf3[1];
+    s3[2] = salt_buf3[2];
+    s3[3] = salt_buf3[3];
 
     switch_buffer_by_offset_le_VV (s0, s1, s2, s3, pw_len);
 
@@ -305,15 +323,25 @@ __kernel void m00010_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 
   u32 salt_buf0[4];
   u32 salt_buf1[4];
-
-  salt_buf0[0] = salt_bufs[salt_pos].salt_buf[0];
-  salt_buf0[1] = salt_bufs[salt_pos].salt_buf[1];
-  salt_buf0[2] = salt_bufs[salt_pos].salt_buf[2];
-  salt_buf0[3] = salt_bufs[salt_pos].salt_buf[3];
-  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];
+  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;
 
@@ -399,8 +427,8 @@ __kernel void m00010_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     w2[3] = wordl2[3] | wordr2[3];
     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];
 
     /**
      * append salt
@@ -419,6 +447,14 @@ __kernel void m00010_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     s1[1] = salt_buf1[1];
     s1[2] = salt_buf1[2];
     s1[3] = salt_buf1[3];
+    s2[0] = salt_buf2[0];
+    s2[1] = salt_buf2[1];
+    s2[2] = salt_buf2[2];
+    s2[3] = salt_buf2[3];
+    s3[0] = salt_buf3[0];
+    s3[1] = salt_buf3[1];
+    s3[2] = salt_buf3[2];
+    s3[3] = salt_buf3[3];
 
     switch_buffer_by_offset_le_VV (s0, s1, s2, s3, pw_len);
 

diff --git a/OpenCL/m00010_a3.cl b/OpenCL/m00010_a3.cl
index ae1a83f..f2cc12a 100644 (file)
--- a/OpenCL/m00010_a3.cl
+++ b/OpenCL/m00010_a3.cl
@@ -51,32 +51,28 @@ static void m00010m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global k
    */
 
   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] = 0;
-  salt_buf2[2] = 0;
-  salt_buf2[3] = 0;
-
-  u32 salt_buf3[4];
+  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];
 
-  salt_buf3[0] = 0;
-  salt_buf3[1] = 0;
-  salt_buf3[2] = 0;
-  salt_buf3[3] = 0;
+  const u32 salt_len = salt_bufs[salt_pos].salt_len;
 
   switch_buffer_by_offset_le_S (salt_buf0, salt_buf1, salt_buf2, salt_buf3, pw_len);
 
@@ -97,11 +93,10 @@ static void m00010m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global k
   w[14] |= salt_buf3[2];
   w[15] |= salt_buf3[3];
 
-  const u32 salt_len = salt_bufs[salt_pos].salt_len;
-
   const u32 pw_salt_len = pw_len + salt_len;
 
   w[14] = pw_salt_len * 8;
+  w[15] = 0;
 
   /**
    * base

diff --git a/OpenCL/m00020_a0.cl b/OpenCL/m00020_a0.cl
index 2518b03..c10513c 100644 (file)
--- a/OpenCL/m00020_a0.cl
+++ b/OpenCL/m00020_a0.cl
@@ -59,18 +59,26 @@ __kernel void m00020_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];
-
-  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];
+  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;
 
@@ -125,10 +133,19 @@ __kernel void m00020_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     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];
 
     append_0x80_4x4 (w0_t, w1_t, w2_t, w3_t, out_salt_len);
 
     w3_t[2] = out_salt_len * 8;
+    w3_t[3] = 0;
 
     /**
      * md5
@@ -256,18 +273,26 @@ __kernel void m00020_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];
-
-  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];
+  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;
 
@@ -334,10 +359,19 @@ __kernel void m00020_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     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];
 
     append_0x80_4x4 (w0_t, w1_t, w2_t, w3_t, out_salt_len);
 
     w3_t[2] = out_salt_len * 8;
+    w3_t[3] = 0;
 
     /**
      * md5

diff --git a/OpenCL/m00020_a1.cl b/OpenCL/m00020_a1.cl
index 6afd120..86ca18e 100644 (file)
--- a/OpenCL/m00020_a1.cl
+++ b/OpenCL/m00020_a1.cl
@@ -56,15 +56,25 @@ __kernel void m00020_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 
   u32 salt_buf0[4];
   u32 salt_buf1[4];
-
-  salt_buf0[0] = salt_bufs[salt_pos].salt_buf[0];
-  salt_buf0[1] = salt_bufs[salt_pos].salt_buf[1];
-  salt_buf0[2] = salt_bufs[salt_pos].salt_buf[2];
-  salt_buf0[3] = salt_bufs[salt_pos].salt_buf[3];
-  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];
+  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;
 
@@ -157,12 +167,12 @@ __kernel void m00020_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     w1[1] |= salt_buf1[1];
     w1[2] |= salt_buf1[2];
     w1[3] |= salt_buf1[3];
-    w2[0] |= 0;
-    w2[1] |= 0;
-    w2[2] |= 0;
-    w2[3] |= 0;
-    w3[0] |= 0;
-    w3[1] |= 0;
+    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]  = pw_salt_len * 8;
     w3[3]  = 0;
 
@@ -246,6 +256,7 @@ __kernel void m00020_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     COMPARE_M_SIMD (a, d, c, b);
   }
 }
+
 __kernel void m00020_m08 (__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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
 {
 }
@@ -290,15 +301,25 @@ __kernel void m00020_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 
   u32 salt_buf0[4];
   u32 salt_buf1[4];
-
-  salt_buf0[0] = salt_bufs[salt_pos].salt_buf[0];
-  salt_buf0[1] = salt_bufs[salt_pos].salt_buf[1];
-  salt_buf0[2] = salt_bufs[salt_pos].salt_buf[2];
-  salt_buf0[3] = salt_bufs[salt_pos].salt_buf[3];
-  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];
+  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;
 
@@ -403,12 +424,12 @@ __kernel void m00020_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     w1[1] |= salt_buf1[1];
     w1[2] |= salt_buf1[2];
     w1[3] |= salt_buf1[3];
-    w2[0] |= 0;
-    w2[1] |= 0;
-    w2[2] |= 0;
-    w2[3] |= 0;
-    w3[0] |= 0;
-    w3[1] |= 0;
+    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]  = pw_salt_len * 8;
     w3[3]  = 0;
 

diff --git a/OpenCL/m00020_a3.cl b/OpenCL/m00020_a3.cl
index d74fbd7..1c6d8b1 100644 (file)
--- a/OpenCL/m00020_a3.cl
+++ b/OpenCL/m00020_a3.cl
@@ -20,7 +20,7 @@
 #include "OpenCL/common.c"
 #include "OpenCL/simd.c"
 
-static void m00020m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset)
+static void m00020m (u32 t0[4], u32 t1[4], u32 t2[4], u32 t3[4], const u32 pw_len, __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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset)
 {
   /**
    * modifier
@@ -34,32 +34,26 @@ static void m00020m (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] = 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_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;
 
@@ -74,22 +68,22 @@ static void m00020m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
   u32 w2_t[4];
   u32 w3_t[4];
 
-  w0_t[0] = w0[0];
-  w0_t[1] = w0[1];
-  w0_t[2] = w0[2];
-  w0_t[3] = w0[3];
-  w1_t[0] = w1[0];
-  w1_t[1] = w1[1];
-  w1_t[2] = w1[2];
-  w1_t[3] = w1[3];
-  w2_t[0] = w2[0];
-  w2_t[1] = w2[1];
-  w2_t[2] = w2[2];
-  w2_t[3] = w2[3];
-  w3_t[0] = w3[0];
-  w3_t[1] = w3[1];
-  w3_t[2] = w3[2];
-  w3_t[3] = w3[3];
+  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] = t3[2];
+  w3_t[3] = t3[3];
 
   switch_buffer_by_offset_le_S (w0_t, w1_t, w2_t, w3_t, salt_len);
 
@@ -114,7 +108,7 @@ static void m00020m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
    * loop
    */
 
-  u32 w0l = w0[0];
+  u32 w0l = t0[0];
 
   for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos += VECT_SIZE)
   {
@@ -143,27 +137,27 @@ static void m00020m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
 
     overwrite_at_le (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];
-    w3_t[2] = pw_salt_len * 8;
-    w3_t[3] = 0;
+    u32x w0[4];
+    u32x w1[4];
+    u32x w2[4];
+    u32x w3[4];
+
+    w0[0] = wx[ 0];
+    w0[1] = wx[ 1];
+    w0[2] = wx[ 2];
+    w0[3] = wx[ 3];
+    w1[0] = wx[ 4];
+    w1[1] = wx[ 5];
+    w1[2] = wx[ 6];
+    w1[3] = wx[ 7];
+    w2[0] = wx[ 8];
+    w2[1] = wx[ 9];
+    w2[2] = wx[10];
+    w2[3] = wx[11];
+    w3[0] = wx[12];
+    w3[1] = wx[13];
+    w3[2] = pw_salt_len * 8;
+    w3[3] = 0;
 
     /**
      * md5
@@ -174,73 +168,73 @@ static void m00020m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
     u32x c = MD5M_C;
     u32x d = MD5M_D;
 
-    MD5_STEP (MD5_Fo, a, b, c, d, w0_t[0], MD5C00, MD5S00);
-    MD5_STEP (MD5_Fo, d, a, b, c, w0_t[1], MD5C01, MD5S01);
-    MD5_STEP (MD5_Fo, c, d, a, b, w0_t[2], MD5C02, MD5S02);
-    MD5_STEP (MD5_Fo, b, c, d, a, w0_t[3], MD5C03, MD5S03);
-    MD5_STEP (MD5_Fo, a, b, c, d, w1_t[0], MD5C04, MD5S00);
-    MD5_STEP (MD5_Fo, d, a, b, c, w1_t[1], MD5C05, MD5S01);
-    MD5_STEP (MD5_Fo, c, d, a, b, w1_t[2], MD5C06, MD5S02);
-    MD5_STEP (MD5_Fo, b, c, d, a, w1_t[3], MD5C07, MD5S03);
-    MD5_STEP (MD5_Fo, a, b, c, d, w2_t[0], MD5C08, MD5S00);
-    MD5_STEP (MD5_Fo, d, a, b, c, w2_t[1], MD5C09, MD5S01);
-    MD5_STEP (MD5_Fo, c, d, a, b, w2_t[2], MD5C0a, MD5S02);
-    MD5_STEP (MD5_Fo, b, c, d, a, w2_t[3], MD5C0b, MD5S03);
-    MD5_STEP (MD5_Fo, a, b, c, d, w3_t[0], MD5C0c, MD5S00);
-    MD5_STEP (MD5_Fo, d, a, b, c, w3_t[1], MD5C0d, MD5S01);
-    MD5_STEP (MD5_Fo, c, d, a, b, w3_t[2], MD5C0e, MD5S02);
-    MD5_STEP (MD5_Fo, b, c, d, a, w3_t[3], MD5C0f, MD5S03);
-
-    MD5_STEP (MD5_Go, a, b, c, d, w0_t[1], MD5C10, MD5S10);
-    MD5_STEP (MD5_Go, d, a, b, c, w1_t[2], MD5C11, MD5S11);
-    MD5_STEP (MD5_Go, c, d, a, b, w2_t[3], MD5C12, MD5S12);
-    MD5_STEP (MD5_Go, b, c, d, a, w0_t[0], MD5C13, MD5S13);
-    MD5_STEP (MD5_Go, a, b, c, d, w1_t[1], MD5C14, MD5S10);
-    MD5_STEP (MD5_Go, d, a, b, c, w2_t[2], MD5C15, MD5S11);
-    MD5_STEP (MD5_Go, c, d, a, b, w3_t[3], MD5C16, MD5S12);
-    MD5_STEP (MD5_Go, b, c, d, a, w1_t[0], MD5C17, MD5S13);
-    MD5_STEP (MD5_Go, a, b, c, d, w2_t[1], MD5C18, MD5S10);
-    MD5_STEP (MD5_Go, d, a, b, c, w3_t[2], MD5C19, MD5S11);
-    MD5_STEP (MD5_Go, c, d, a, b, w0_t[3], MD5C1a, MD5S12);
-    MD5_STEP (MD5_Go, b, c, d, a, w2_t[0], MD5C1b, MD5S13);
-    MD5_STEP (MD5_Go, a, b, c, d, w3_t[1], MD5C1c, MD5S10);
-    MD5_STEP (MD5_Go, d, a, b, c, w0_t[2], MD5C1d, MD5S11);
-    MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
-    MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
-
-    MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
-    MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
-    MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
-    MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
-    MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
-    MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
-    MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
-    MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
-    MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
-    MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
-    MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
-    MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
-    MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
-    MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
-    MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
-    MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
-
-    MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
-    MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
-    MD5_STEP (MD5_I , c, d, a, b, w3_t[2], MD5C32, MD5S32);
-    MD5_STEP (MD5_I , b, c, d, a, w1_t[1], MD5C33, MD5S33);
-    MD5_STEP (MD5_I , a, b, c, d, w3_t[0], MD5C34, MD5S30);
-    MD5_STEP (MD5_I , d, a, b, c, w0_t[3], MD5C35, MD5S31);
-    MD5_STEP (MD5_I , c, d, a, b, w2_t[2], MD5C36, MD5S32);
-    MD5_STEP (MD5_I , b, c, d, a, w0_t[1], MD5C37, MD5S33);
-    MD5_STEP (MD5_I , a, b, c, d, w2_t[0], MD5C38, MD5S30);
-    MD5_STEP (MD5_I , d, a, b, c, w3_t[3], MD5C39, MD5S31);
-    MD5_STEP (MD5_I , c, d, a, b, w1_t[2], MD5C3a, MD5S32);
-    MD5_STEP (MD5_I , b, c, d, a, w3_t[1], MD5C3b, MD5S33);
-    MD5_STEP (MD5_I , a, b, c, d, w1_t[0], MD5C3c, MD5S30);
-    MD5_STEP (MD5_I , d, a, b, c, w2_t[3], MD5C3d, MD5S31);
-    MD5_STEP (MD5_I , c, d, a, b, w0_t[2], MD5C3e, MD5S32);
-    MD5_STEP (MD5_I , b, c, d, a, w2_t[1], MD5C3f, MD5S33);
+    MD5_STEP (MD5_Fo, a, b, c, d, w0[0], MD5C00, MD5S00);
+    MD5_STEP (MD5_Fo, d, a, b, c, w0[1], MD5C01, MD5S01);
+    MD5_STEP (MD5_Fo, c, d, a, b, w0[2], MD5C02, MD5S02);
+    MD5_STEP (MD5_Fo, b, c, d, a, w0[3], MD5C03, MD5S03);
+    MD5_STEP (MD5_Fo, a, b, c, d, w1[0], MD5C04, MD5S00);
+    MD5_STEP (MD5_Fo, d, a, b, c, w1[1], MD5C05, MD5S01);
+    MD5_STEP (MD5_Fo, c, d, a, b, w1[2], MD5C06, MD5S02);
+    MD5_STEP (MD5_Fo, b, c, d, a, w1[3], MD5C07, MD5S03);
+    MD5_STEP (MD5_Fo, a, b, c, d, w2[0], MD5C08, MD5S00);
+    MD5_STEP (MD5_Fo, d, a, b, c, w2[1], MD5C09, MD5S01);
+    MD5_STEP (MD5_Fo, c, d, a, b, w2[2], MD5C0a, MD5S02);
+    MD5_STEP (MD5_Fo, b, c, d, a, w2[3], MD5C0b, MD5S03);
+    MD5_STEP (MD5_Fo, a, b, c, d, w3[0], MD5C0c, MD5S00);
+    MD5_STEP (MD5_Fo, d, a, b, c, w3[1], MD5C0d, MD5S01);
+    MD5_STEP (MD5_Fo, c, d, a, b, w3[2], MD5C0e, MD5S02);
+    MD5_STEP (MD5_Fo, b, c, d, a, w3[3], MD5C0f, MD5S03);
+
+    MD5_STEP (MD5_Go, a, b, c, d, w0[1], MD5C10, MD5S10);
+    MD5_STEP (MD5_Go, d, a, b, c, w1[2], MD5C11, MD5S11);
+    MD5_STEP (MD5_Go, c, d, a, b, w2[3], MD5C12, MD5S12);
+    MD5_STEP (MD5_Go, b, c, d, a, w0[0], MD5C13, MD5S13);
+    MD5_STEP (MD5_Go, a, b, c, d, w1[1], MD5C14, MD5S10);
+    MD5_STEP (MD5_Go, d, a, b, c, w2[2], MD5C15, MD5S11);
+    MD5_STEP (MD5_Go, c, d, a, b, w3[3], MD5C16, MD5S12);
+    MD5_STEP (MD5_Go, b, c, d, a, w1[0], MD5C17, MD5S13);
+    MD5_STEP (MD5_Go, a, b, c, d, w2[1], MD5C18, MD5S10);
+    MD5_STEP (MD5_Go, d, a, b, c, w3[2], MD5C19, MD5S11);
+    MD5_STEP (MD5_Go, c, d, a, b, w0[3], MD5C1a, MD5S12);
+    MD5_STEP (MD5_Go, b, c, d, a, w2[0], MD5C1b, MD5S13);
+    MD5_STEP (MD5_Go, a, b, c, d, w3[1], MD5C1c, MD5S10);
+    MD5_STEP (MD5_Go, d, a, b, c, w0[2], MD5C1d, MD5S11);
+    MD5_STEP (MD5_Go, c, d, a, b, w1[3], MD5C1e, MD5S12);
+    MD5_STEP (MD5_Go, b, c, d, a, w3[0], MD5C1f, MD5S13);
+
+    MD5_STEP (MD5_H , a, b, c, d, w1[1], MD5C20, MD5S20);
+    MD5_STEP (MD5_H , d, a, b, c, w2[0], MD5C21, MD5S21);
+    MD5_STEP (MD5_H , c, d, a, b, w2[3], MD5C22, MD5S22);
+    MD5_STEP (MD5_H , b, c, d, a, w3[2], MD5C23, MD5S23);
+    MD5_STEP (MD5_H , a, b, c, d, w0[1], MD5C24, MD5S20);
+    MD5_STEP (MD5_H , d, a, b, c, w1[0], MD5C25, MD5S21);
+    MD5_STEP (MD5_H , c, d, a, b, w1[3], MD5C26, MD5S22);
+    MD5_STEP (MD5_H , b, c, d, a, w2[2], MD5C27, MD5S23);
+    MD5_STEP (MD5_H , a, b, c, d, w3[1], MD5C28, MD5S20);
+    MD5_STEP (MD5_H , d, a, b, c, w0[0], MD5C29, MD5S21);
+    MD5_STEP (MD5_H , c, d, a, b, w0[3], MD5C2a, MD5S22);
+    MD5_STEP (MD5_H , b, c, d, a, w1[2], MD5C2b, MD5S23);
+    MD5_STEP (MD5_H , a, b, c, d, w2[1], MD5C2c, MD5S20);
+    MD5_STEP (MD5_H , d, a, b, c, w3[0], MD5C2d, MD5S21);
+    MD5_STEP (MD5_H , c, d, a, b, w3[3], MD5C2e, MD5S22);
+    MD5_STEP (MD5_H , b, c, d, a, w0[2], MD5C2f, MD5S23);
+
+    MD5_STEP (MD5_I , a, b, c, d, w0[0], MD5C30, MD5S30);
+    MD5_STEP (MD5_I , d, a, b, c, w1[3], MD5C31, MD5S31);
+    MD5_STEP (MD5_I , c, d, a, b, w3[2], MD5C32, MD5S32);
+    MD5_STEP (MD5_I , b, c, d, a, w1[1], MD5C33, MD5S33);
+    MD5_STEP (MD5_I , a, b, c, d, w3[0], MD5C34, MD5S30);
+    MD5_STEP (MD5_I , d, a, b, c, w0[3], MD5C35, MD5S31);
+    MD5_STEP (MD5_I , c, d, a, b, w2[2], MD5C36, MD5S32);
+    MD5_STEP (MD5_I , b, c, d, a, w0[1], MD5C37, MD5S33);
+    MD5_STEP (MD5_I , a, b, c, d, w2[0], MD5C38, MD5S30);
+    MD5_STEP (MD5_I , d, a, b, c, w3[3], MD5C39, MD5S31);
+    MD5_STEP (MD5_I , c, d, a, b, w1[2], MD5C3a, MD5S32);
+    MD5_STEP (MD5_I , b, c, d, a, w3[1], MD5C3b, MD5S33);
+    MD5_STEP (MD5_I , a, b, c, d, w1[0], MD5C3c, MD5S30);
+    MD5_STEP (MD5_I , d, a, b, c, w2[3], MD5C3d, MD5S31);
+    MD5_STEP (MD5_I , c, d, a, b, w0[2], MD5C3e, MD5S32);
+    MD5_STEP (MD5_I , b, c, d, a, w2[1], MD5C3f, MD5S33);
 
     COMPARE_M_SIMD (a, d, c, b);
   }
@@ -272,32 +266,26 @@ static void m00020s (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] = 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_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;
 

diff --git a/OpenCL/m00030_a0.cl b/OpenCL/m00030_a0.cl
index 953f933..7e043ea 100644 (file)
--- a/OpenCL/m00030_a0.cl
+++ b/OpenCL/m00030_a0.cl
@@ -61,18 +61,26 @@ __kernel void m00030_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];
-
-  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];
+  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;
 
@@ -147,12 +155,8 @@ __kernel void m00030_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     w2_t[3] |= s2[3];
     w3_t[0] |= s3[0];
     w3_t[1] |= s3[1];
-    w3_t[2] |= s3[2];
-    w3_t[3] |= s3[3];
-
-    append_0x80_4x4 (w0_t, w1_t, w2_t, w3_t, out_salt_len);
-
-    w3_t[2] = out_salt_len * 8;
+    w3_t[2]  = out_salt_len * 8;
+    w3_t[3]  = 0;
 
     /**
      * md5
@@ -280,18 +284,26 @@ __kernel void m00030_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];
-
-  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];
+  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;
 
@@ -378,12 +390,8 @@ __kernel void m00030_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     w2_t[3] |= s2[3];
     w3_t[0] |= s3[0];
     w3_t[1] |= s3[1];
-    w3_t[2] |= s3[2];
-    w3_t[3] |= s3[3];
-
-    append_0x80_4x4 (w0_t, w1_t, w2_t, w3_t, out_salt_len);
-
-    w3_t[2] = out_salt_len * 8;
+    w3_t[2]  = out_salt_len * 8;
+    w3_t[3]  = 0;
 
     /**
      * md5

diff --git a/OpenCL/m00030_a1.cl b/OpenCL/m00030_a1.cl
index 0016776..2113be1 100644 (file)
--- a/OpenCL/m00030_a1.cl
+++ b/OpenCL/m00030_a1.cl
@@ -58,15 +58,25 @@ __kernel void m00030_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 
   u32 salt_buf0[4];
   u32 salt_buf1[4];
-
-  salt_buf0[0] = salt_bufs[salt_pos].salt_buf[0];
-  salt_buf0[1] = salt_bufs[salt_pos].salt_buf[1];
-  salt_buf0[2] = salt_bufs[salt_pos].salt_buf[2];
-  salt_buf0[3] = salt_bufs[salt_pos].salt_buf[3];
-  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];
+  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;
 
@@ -312,15 +322,25 @@ __kernel void m00030_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 
   u32 salt_buf0[4];
   u32 salt_buf1[4];
-
-  salt_buf0[0] = salt_bufs[salt_pos].salt_buf[0];
-  salt_buf0[1] = salt_bufs[salt_pos].salt_buf[1];
-  salt_buf0[2] = salt_bufs[salt_pos].salt_buf[2];
-  salt_buf0[3] = salt_bufs[salt_pos].salt_buf[3];
-  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];
+  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;
 

diff --git a/OpenCL/m00030_a3.cl b/OpenCL/m00030_a3.cl
index 946c887..a79bfc3 100644 (file)
--- a/OpenCL/m00030_a3.cl
+++ b/OpenCL/m00030_a3.cl
@@ -51,32 +51,28 @@ static void m00030m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global k
    */
 
   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] = 0;
-  salt_buf2[2] = 0;
-  salt_buf2[3] = 0;
-
-  u32 salt_buf3[4];
+  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];
 
-  salt_buf3[0] = 0;
-  salt_buf3[1] = 0;
-  salt_buf3[2] = 0;
-  salt_buf3[3] = 0;
+  const u32 salt_len = salt_bufs[salt_pos].salt_len;
 
   switch_buffer_by_offset_le_S (salt_buf0, salt_buf1, salt_buf2, salt_buf3, pw_len);
 
@@ -97,11 +93,10 @@ static void m00030m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global k
   w[14] |= salt_buf3[2];
   w[15] |= salt_buf3[3];
 
-  const u32 salt_len = salt_bufs[salt_pos].salt_len;
-
   const u32 pw_salt_len = pw_len + salt_len;
 
   w[14] = pw_salt_len * 8;
+  w[15] = 0;
 
   /**
    * base

diff --git a/OpenCL/m00040_a0.cl b/OpenCL/m00040_a0.cl
index 2f6a0aa..6c41a85 100644 (file)
--- a/OpenCL/m00040_a0.cl
+++ b/OpenCL/m00040_a0.cl
@@ -59,18 +59,26 @@ __kernel void m00040_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];
-
-  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];
+  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;
 
@@ -111,10 +119,19 @@ __kernel void m00040_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     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];
 
     append_0x80_4x4 (w0_t, w1_t, w2_t, w3_t, out_salt_len);
 
     w3_t[2] = out_salt_len * 8;
+    w3_t[3] = 0;
 
     /**
      * md5
@@ -242,18 +259,26 @@ __kernel void m00040_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];
-
-  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];
+  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;
 
@@ -306,10 +331,19 @@ __kernel void m00040_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     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];
 
     append_0x80_4x4 (w0_t, w1_t, w2_t, w3_t, out_salt_len);
 
     w3_t[2] = out_salt_len * 8;
+    w3_t[3] = 0;
 
     /**
      * md5

diff --git a/OpenCL/m00040_a1.cl b/OpenCL/m00040_a1.cl
index f704c55..9fa9651 100644 (file)
--- a/OpenCL/m00040_a1.cl
+++ b/OpenCL/m00040_a1.cl
@@ -58,15 +58,25 @@ __kernel void m00040_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 
   u32 salt_buf0[4];
   u32 salt_buf1[4];
-
-  salt_buf0[0] = salt_bufs[salt_pos].salt_buf[0];
-  salt_buf0[1] = salt_bufs[salt_pos].salt_buf[1];
-  salt_buf0[2] = salt_bufs[salt_pos].salt_buf[2];
-  salt_buf0[3] = salt_bufs[salt_pos].salt_buf[3];
-  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];
+  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;
 
@@ -164,12 +174,12 @@ __kernel void m00040_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     w1[1] |= salt_buf1[1];
     w1[2] |= salt_buf1[2];
     w1[3] |= salt_buf1[3];
-    w2[0] |= 0;
-    w2[1] |= 0;
-    w2[2] |= 0;
-    w2[3] |= 0;
-    w3[0] |= 0;
-    w3[1] |= 0;
+    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]  = pw_salt_len * 8;
     w3[3]  = 0;
 
@@ -298,15 +308,25 @@ __kernel void m00040_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 
   u32 salt_buf0[4];
   u32 salt_buf1[4];
-
-  salt_buf0[0] = salt_bufs[salt_pos].salt_buf[0];
-  salt_buf0[1] = salt_bufs[salt_pos].salt_buf[1];
-  salt_buf0[2] = salt_bufs[salt_pos].salt_buf[2];
-  salt_buf0[3] = salt_bufs[salt_pos].salt_buf[3];
-  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];
+  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;
 
@@ -416,12 +436,12 @@ __kernel void m00040_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     w1[1] |= salt_buf1[1];
     w1[2] |= salt_buf1[2];
     w1[3] |= salt_buf1[3];
-    w2[0] |= 0;
-    w2[1] |= 0;
-    w2[2] |= 0;
-    w2[3] |= 0;
-    w3[0] |= 0;
-    w3[1] |= 0;
+    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]  = pw_salt_len * 8;
     w3[3]  = 0;
 

diff --git a/OpenCL/m00040_a3.cl b/OpenCL/m00040_a3.cl
index a60c6d4..fabe52b 100644 (file)
--- a/OpenCL/m00040_a3.cl
+++ b/OpenCL/m00040_a3.cl
@@ -34,32 +34,26 @@ static void m00040m (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] = 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_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;
 
@@ -272,32 +266,26 @@ static void m00040s (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] = 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_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;
 

diff --git a/OpenCL/m00050_a0.cl b/OpenCL/m00050_a0.cl
index 575b1b4..09fda93 100644 (file)
--- a/OpenCL/m00050_a0.cl
+++ b/OpenCL/m00050_a0.cl
@@ -242,28 +242,22 @@ __kernel void m00050_m04 (__global pw_t *pws, __global kernel_rule_t *  rules_bu
    */
 
   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[4];
-
   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];
@@ -391,28 +385,22 @@ __kernel void m00050_s04 (__global pw_t *pws, __global kernel_rule_t *  rules_bu
    */
 
   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[4];
-
   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];

diff --git a/OpenCL/m00050_a3.cl b/OpenCL/m00050_a3.cl
index a954bf4..54e5e67 100644 (file)
--- a/OpenCL/m00050_a3.cl
+++ b/OpenCL/m00050_a3.cl
@@ -217,28 +217,22 @@ static void m00050m (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[4];
-
   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];
@@ -334,28 +328,22 @@ static void m00050s (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[4];
-
   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];

diff --git a/OpenCL/m00060_a0.cl b/OpenCL/m00060_a0.cl
index 9515c79..558416c 100644 (file)
--- a/OpenCL/m00060_a0.cl
+++ b/OpenCL/m00060_a0.cl
@@ -242,50 +242,54 @@ __kernel void m00060_m04 (__global pw_t *pws, __global kernel_rule_t *  rules_bu
    */
 
   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];
+  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;
 
   /**
    * pads
    */
 
   u32x w0_t[4];
+  u32x w1_t[4];
+  u32x w2_t[4];
+  u32x w3_t[4];
 
   w0_t[0] = salt_buf0[0];
   w0_t[1] = salt_buf0[1];
   w0_t[2] = salt_buf0[2];
   w0_t[3] = salt_buf0[3];
-
-  u32x w1_t[4];
-
   w1_t[0] = salt_buf1[0];
   w1_t[1] = salt_buf1[1];
   w1_t[2] = salt_buf1[2];
   w1_t[3] = salt_buf1[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;
+  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 ipad[4];
   u32x opad[4];
@@ -377,50 +381,54 @@ __kernel void m00060_s04 (__global pw_t *pws, __global kernel_rule_t *  rules_bu
    */
 
   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];
+  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;
 
   /**
    * pads
    */
 
   u32x w0_t[4];
+  u32x w1_t[4];
+  u32x w2_t[4];
+  u32x w3_t[4];
 
   w0_t[0] = salt_buf0[0];
   w0_t[1] = salt_buf0[1];
   w0_t[2] = salt_buf0[2];
   w0_t[3] = salt_buf0[3];
-
-  u32x w1_t[4];
-
   w1_t[0] = salt_buf1[0];
   w1_t[1] = salt_buf1[1];
   w1_t[2] = salt_buf1[2];
   w1_t[3] = salt_buf1[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;
+  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 ipad[4];
   u32x opad[4];

diff --git a/OpenCL/m00060_a3.cl b/OpenCL/m00060_a3.cl
index 3d5419c..b2dd85e 100644 (file)
--- a/OpenCL/m00060_a3.cl
+++ b/OpenCL/m00060_a3.cl
@@ -235,32 +235,26 @@ static void m00060m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
    */
 
   u32x w0_t[4];
+  u32x w1_t[4];
+  u32x w2_t[4];
+  u32x w3_t[4];
 
   w0_t[0] = salt_buf0[0];
   w0_t[1] = salt_buf0[1];
   w0_t[2] = salt_buf0[2];
   w0_t[3] = salt_buf0[3];
-
-  u32x w1_t[4];
-
   w1_t[0] = salt_buf1[0];
   w1_t[1] = salt_buf1[1];
   w1_t[2] = salt_buf1[2];
   w1_t[3] = salt_buf1[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;
+  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 ipad[4];
   u32x opad[4];
@@ -320,50 +314,54 @@ static void m00060s (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];
+  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;
 
   /**
    * pads
    */
 
   u32x w0_t[4];
+  u32x w1_t[4];
+  u32x w2_t[4];
+  u32x w3_t[4];
 
   w0_t[0] = salt_buf0[0];
   w0_t[1] = salt_buf0[1];
   w0_t[2] = salt_buf0[2];
   w0_t[3] = salt_buf0[3];
-
-  u32x w1_t[4];
-
   w1_t[0] = salt_buf1[0];
   w1_t[1] = salt_buf1[1];
   w1_t[2] = salt_buf1[2];
   w1_t[3] = salt_buf1[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;
+  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 ipad[4];
   u32x opad[4];