Cleanup -m 1450 kernels to latest standard

diff --git a/OpenCL/m01450_a0.cl b/OpenCL/m01450_a0.cl
index 8b0741c..012f039 100644 (file)
--- a/OpenCL/m01450_a0.cl
+++ b/OpenCL/m01450_a0.cl
@@ -246,18 +246,16 @@ __kernel void m01450_m04 (__global pw_t *pws, __global kernel_rule_t *  rules_bu
   if (gid >= gid_max) return;
 
   u32 pw_buf0[4];
-
-  pw_buf0[0] = pws[gid].i[ 0];
-  pw_buf0[1] = pws[gid].i[ 1];
-  pw_buf0[2] = pws[gid].i[ 2];
-  pw_buf0[3] = pws[gid].i[ 3];
-
   u32 pw_buf1[4];
 
-  pw_buf1[0] = pws[gid].i[ 4];
-  pw_buf1[1] = pws[gid].i[ 5];
-  pw_buf1[2] = pws[gid].i[ 6];
-  pw_buf1[3] = pws[gid].i[ 7];
+  pw_buf0[0] = pws[gid].i[0];
+  pw_buf0[1] = pws[gid].i[1];
+  pw_buf0[2] = pws[gid].i[2];
+  pw_buf0[3] = pws[gid].i[3];
+  pw_buf1[0] = pws[gid].i[4];
+  pw_buf1[1] = pws[gid].i[5];
+  pw_buf1[2] = pws[gid].i[6];
+  pw_buf1[3] = pws[gid].i[7];
 
   const u32 pw_len = pws[gid].pw_len;
 
@@ -266,18 +264,26 @@ __kernel void m01450_m04 (__global pw_t *pws, __global kernel_rule_t *  rules_bu
    */
 
   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] = swap32_S (salt_bufs[salt_pos].salt_buf[ 0]);
+  salt_buf0[1] = swap32_S (salt_bufs[salt_pos].salt_buf[ 1]);
+  salt_buf0[2] = swap32_S (salt_bufs[salt_pos].salt_buf[ 2]);
+  salt_buf0[3] = swap32_S (salt_bufs[salt_pos].salt_buf[ 3]);
+  salt_buf1[0] = swap32_S (salt_bufs[salt_pos].salt_buf[ 4]);
+  salt_buf1[1] = swap32_S (salt_bufs[salt_pos].salt_buf[ 5]);
+  salt_buf1[2] = swap32_S (salt_bufs[salt_pos].salt_buf[ 6]);
+  salt_buf1[3] = swap32_S (salt_bufs[salt_pos].salt_buf[ 7]);
+  salt_buf2[0] = swap32_S (salt_bufs[salt_pos].salt_buf[ 8]);
+  salt_buf2[1] = swap32_S (salt_bufs[salt_pos].salt_buf[ 9]);
+  salt_buf2[2] = swap32_S (salt_bufs[salt_pos].salt_buf[10]);
+  salt_buf2[3] = swap32_S (salt_bufs[salt_pos].salt_buf[11]);
+  salt_buf3[0] = swap32_S (salt_bufs[salt_pos].salt_buf[12]);
+  salt_buf3[1] = swap32_S (salt_bufs[salt_pos].salt_buf[13]);
+  salt_buf3[2] = swap32_S (salt_bufs[salt_pos].salt_buf[14]);
+  salt_buf3[3] = swap32_S (salt_bufs[salt_pos].salt_buf[15]);
 
   const u32 salt_len = salt_bufs[salt_pos].salt_len;
 
@@ -294,63 +300,44 @@ __kernel void m01450_m04 (__global pw_t *pws, __global kernel_rule_t *  rules_bu
 
     apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
 
+    w0[0] = swap32 (w0[0]);
+    w0[1] = swap32 (w0[1]);
+    w0[2] = swap32 (w0[2]);
+    w0[3] = swap32 (w0[3]);
+    w1[0] = swap32 (w1[0]);
+    w1[1] = swap32 (w1[1]);
+    w1[2] = swap32 (w1[2]);
+    w1[3] = swap32 (w1[3]);
+
     /**
      * pads
      */
 
-    u32x w0_t[4];
-
-    w0_t[0] = swap32 (w0[0]);
-    w0_t[1] = swap32 (w0[1]);
-    w0_t[2] = swap32 (w0[2]);
-    w0_t[3] = swap32 (w0[3]);
-
-    u32x w1_t[4];
-
-    w1_t[0] = swap32 (w1[0]);
-    w1_t[1] = swap32 (w1[1]);
-    w1_t[2] = swap32 (w1[2]);
-    w1_t[3] = swap32 (w1[3]);
-
-    u32x w2_t[4];
-
-    w2_t[0] = 0;
-    w2_t[1] = 0;
-    w2_t[2] = 0;
-    w2_t[3] = 0;
-
-    u32x w3_t[4];
-
-    w3_t[0] = 0;
-    w3_t[1] = 0;
-    w3_t[2] = 0;
-    w3_t[3] = 0;
-
     u32x ipad[8];
     u32x opad[8];
 
-    hmac_sha256_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
-
-    w0_t[0] = swap32 (salt_buf0[0]);
-    w0_t[1] = swap32 (salt_buf0[1]);
-    w0_t[2] = swap32 (salt_buf0[2]);
-    w0_t[3] = swap32 (salt_buf0[3]);
-    w1_t[0] = swap32 (salt_buf1[0]);
-    w1_t[1] = swap32 (salt_buf1[1]);
-    w1_t[2] = swap32 (salt_buf1[2]);
-    w1_t[3] = swap32 (salt_buf1[3]);
-    w2_t[0] = 0;
-    w2_t[1] = 0;
-    w2_t[2] = 0;
-    w2_t[3] = 0;
-    w3_t[0] = 0;
-    w3_t[1] = 0;
-    w3_t[2] = 0;
-    w3_t[3] = (64 + salt_len) * 8;
+    hmac_sha256_pad (w0, w1, w2, w3, ipad, opad);
+
+    w0[0] = salt_buf0[0];
+    w0[1] = salt_buf0[1];
+    w0[2] = salt_buf0[2];
+    w0[3] = salt_buf0[3];
+    w1[0] = salt_buf1[0];
+    w1[1] = salt_buf1[1];
+    w1[2] = salt_buf1[2];
+    w1[3] = salt_buf1[3];
+    w2[0] = salt_buf2[0];
+    w2[1] = salt_buf2[1];
+    w2[2] = salt_buf2[2];
+    w2[3] = salt_buf2[3];
+    w3[0] = salt_buf3[0];
+    w3[1] = salt_buf3[1];
+    w3[2] = 0;
+    w3[3] = (64 + salt_len) * 8;
 
     u32x digest[8];
 
-    hmac_sha256_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
+    hmac_sha256_run (w0, w1, w2, w3, ipad, opad, digest);
 
     COMPARE_M_SIMD (digest[3], digest[7], digest[2], digest[6]);
   }
@@ -381,18 +368,16 @@ __kernel void m01450_s04 (__global pw_t *pws, __global kernel_rule_t *  rules_bu
   if (gid >= gid_max) return;
 
   u32 pw_buf0[4];
-
-  pw_buf0[0] = pws[gid].i[ 0];
-  pw_buf0[1] = pws[gid].i[ 1];
-  pw_buf0[2] = pws[gid].i[ 2];
-  pw_buf0[3] = pws[gid].i[ 3];
-
   u32 pw_buf1[4];
 
-  pw_buf1[0] = pws[gid].i[ 4];
-  pw_buf1[1] = pws[gid].i[ 5];
-  pw_buf1[2] = pws[gid].i[ 6];
-  pw_buf1[3] = pws[gid].i[ 7];
+  pw_buf0[0] = pws[gid].i[0];
+  pw_buf0[1] = pws[gid].i[1];
+  pw_buf0[2] = pws[gid].i[2];
+  pw_buf0[3] = pws[gid].i[3];
+  pw_buf1[0] = pws[gid].i[4];
+  pw_buf1[1] = pws[gid].i[5];
+  pw_buf1[2] = pws[gid].i[6];
+  pw_buf1[3] = pws[gid].i[7];
 
   const u32 pw_len = pws[gid].pw_len;
 
@@ -401,18 +386,26 @@ __kernel void m01450_s04 (__global pw_t *pws, __global kernel_rule_t *  rules_bu
    */
 
   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] = swap32_S (salt_bufs[salt_pos].salt_buf[ 0]);
+  salt_buf0[1] = swap32_S (salt_bufs[salt_pos].salt_buf[ 1]);
+  salt_buf0[2] = swap32_S (salt_bufs[salt_pos].salt_buf[ 2]);
+  salt_buf0[3] = swap32_S (salt_bufs[salt_pos].salt_buf[ 3]);
+  salt_buf1[0] = swap32_S (salt_bufs[salt_pos].salt_buf[ 4]);
+  salt_buf1[1] = swap32_S (salt_bufs[salt_pos].salt_buf[ 5]);
+  salt_buf1[2] = swap32_S (salt_bufs[salt_pos].salt_buf[ 6]);
+  salt_buf1[3] = swap32_S (salt_bufs[salt_pos].salt_buf[ 7]);
+  salt_buf2[0] = swap32_S (salt_bufs[salt_pos].salt_buf[ 8]);
+  salt_buf2[1] = swap32_S (salt_bufs[salt_pos].salt_buf[ 9]);
+  salt_buf2[2] = swap32_S (salt_bufs[salt_pos].salt_buf[10]);
+  salt_buf2[3] = swap32_S (salt_bufs[salt_pos].salt_buf[11]);
+  salt_buf3[0] = swap32_S (salt_bufs[salt_pos].salt_buf[12]);
+  salt_buf3[1] = swap32_S (salt_bufs[salt_pos].salt_buf[13]);
+  salt_buf3[2] = swap32_S (salt_bufs[salt_pos].salt_buf[14]);
+  salt_buf3[3] = swap32_S (salt_bufs[salt_pos].salt_buf[15]);
 
   const u32 salt_len = salt_bufs[salt_pos].salt_len;
 
@@ -441,63 +434,44 @@ __kernel void m01450_s04 (__global pw_t *pws, __global kernel_rule_t *  rules_bu
 
     apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
 
+    w0[0] = swap32 (w0[0]);
+    w0[1] = swap32 (w0[1]);
+    w0[2] = swap32 (w0[2]);
+    w0[3] = swap32 (w0[3]);
+    w1[0] = swap32 (w1[0]);
+    w1[1] = swap32 (w1[1]);
+    w1[2] = swap32 (w1[2]);
+    w1[3] = swap32 (w1[3]);
+
     /**
      * pads
      */
 
-    u32x w0_t[4];
-
-    w0_t[0] = swap32 (w0[0]);
-    w0_t[1] = swap32 (w0[1]);
-    w0_t[2] = swap32 (w0[2]);
-    w0_t[3] = swap32 (w0[3]);
-
-    u32x w1_t[4];
-
-    w1_t[0] = swap32 (w1[0]);
-    w1_t[1] = swap32 (w1[1]);
-    w1_t[2] = swap32 (w1[2]);
-    w1_t[3] = swap32 (w1[3]);
-
-    u32x w2_t[4];
-
-    w2_t[0] = 0;
-    w2_t[1] = 0;
-    w2_t[2] = 0;
-    w2_t[3] = 0;
-
-    u32x w3_t[4];
-
-    w3_t[0] = 0;
-    w3_t[1] = 0;
-    w3_t[2] = 0;
-    w3_t[3] = 0;
-
     u32x ipad[8];
     u32x opad[8];
 
-    hmac_sha256_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
-
-    w0_t[0] = swap32 (salt_buf0[0]);
-    w0_t[1] = swap32 (salt_buf0[1]);
-    w0_t[2] = swap32 (salt_buf0[2]);
-    w0_t[3] = swap32 (salt_buf0[3]);
-    w1_t[0] = swap32 (salt_buf1[0]);
-    w1_t[1] = swap32 (salt_buf1[1]);
-    w1_t[2] = swap32 (salt_buf1[2]);
-    w1_t[3] = swap32 (salt_buf1[3]);
-    w2_t[0] = 0;
-    w2_t[1] = 0;
-    w2_t[2] = 0;
-    w2_t[3] = 0;
-    w3_t[0] = 0;
-    w3_t[1] = 0;
-    w3_t[2] = 0;
-    w3_t[3] = (64 + salt_len) * 8;
+    hmac_sha256_pad (w0, w1, w2, w3, ipad, opad);
+
+    w0[0] = salt_buf0[0];
+    w0[1] = salt_buf0[1];
+    w0[2] = salt_buf0[2];
+    w0[3] = salt_buf0[3];
+    w1[0] = salt_buf1[0];
+    w1[1] = salt_buf1[1];
+    w1[2] = salt_buf1[2];
+    w1[3] = salt_buf1[3];
+    w2[0] = salt_buf2[0];
+    w2[1] = salt_buf2[1];
+    w2[2] = salt_buf2[2];
+    w2[3] = salt_buf2[3];
+    w3[0] = salt_buf3[0];
+    w3[1] = salt_buf3[1];
+    w3[2] = 0;
+    w3[3] = (64 + salt_len) * 8;
 
     u32x digest[8];
 
-    hmac_sha256_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
+    hmac_sha256_run (w0, w1, w2, w3, ipad, opad, digest);
 
     COMPARE_S_SIMD (digest[3], digest[7], digest[2], digest[6]);
   }

diff --git a/OpenCL/m01450_a3.cl b/OpenCL/m01450_a3.cl
index 3eebb2c..40ece67 100644 (file)
--- a/OpenCL/m01450_a3.cl
+++ b/OpenCL/m01450_a3.cl
@@ -241,18 +241,26 @@ static void m01450m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
    */
 
   u32 salt_buf0[4];
-
-  salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0];
-  salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1];
-  salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2];
-  salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3];
-
   u32 salt_buf1[4];
-
-  salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
-  salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
-  salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
-  salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
+  u32 salt_buf2[4];
+  u32 salt_buf3[4];
+
+  salt_buf0[0] = swap32_S (salt_bufs[salt_pos].salt_buf[ 0]);
+  salt_buf0[1] = swap32_S (salt_bufs[salt_pos].salt_buf[ 1]);
+  salt_buf0[2] = swap32_S (salt_bufs[salt_pos].salt_buf[ 2]);
+  salt_buf0[3] = swap32_S (salt_bufs[salt_pos].salt_buf[ 3]);
+  salt_buf1[0] = swap32_S (salt_bufs[salt_pos].salt_buf[ 4]);
+  salt_buf1[1] = swap32_S (salt_bufs[salt_pos].salt_buf[ 5]);
+  salt_buf1[2] = swap32_S (salt_bufs[salt_pos].salt_buf[ 6]);
+  salt_buf1[3] = swap32_S (salt_bufs[salt_pos].salt_buf[ 7]);
+  salt_buf2[0] = swap32_S (salt_bufs[salt_pos].salt_buf[ 8]);
+  salt_buf2[1] = swap32_S (salt_bufs[salt_pos].salt_buf[ 9]);
+  salt_buf2[2] = swap32_S (salt_bufs[salt_pos].salt_buf[10]);
+  salt_buf2[3] = swap32_S (salt_bufs[salt_pos].salt_buf[11]);
+  salt_buf3[0] = swap32_S (salt_bufs[salt_pos].salt_buf[12]);
+  salt_buf3[1] = swap32_S (salt_bufs[salt_pos].salt_buf[13]);
+  salt_buf3[2] = swap32_S (salt_bufs[salt_pos].salt_buf[14]);
+  salt_buf3[3] = swap32_S (salt_bufs[salt_pos].salt_buf[15]);
 
   const u32 salt_len = salt_bufs[salt_pos].salt_len;
 
@@ -273,52 +281,46 @@ static void m01450m (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] = w0lr;
     w0_t[1] = w0[1];
     w0_t[2] = w0[2];
     w0_t[3] = w0[3];
-
-    u32x w1_t[4];
-
     w1_t[0] = w1[0];
     w1_t[1] = w1[1];
     w1_t[2] = w1[2];
     w1_t[3] = w1[3];
-
-    u32x w2_t[4];
-
     w2_t[0] = w2[0];
     w2_t[1] = w2[1];
     w2_t[2] = w2[2];
     w2_t[3] = w2[3];
-
-    u32x w3_t[4];
-
     w3_t[0] = w3[0];
     w3_t[1] = w3[1];
-    w3_t[2] = 0;
-    w3_t[3] = 0;
+    w3_t[2] = w3[2];
+    w3_t[3] = w3[3];
 
     u32x ipad[8];
     u32x opad[8];
 
     hmac_sha256_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
 
-    w0_t[0] = swap32 (salt_buf0[0]);
-    w0_t[1] = swap32 (salt_buf0[1]);
-    w0_t[2] = swap32 (salt_buf0[2]);
-    w0_t[3] = swap32 (salt_buf0[3]);
-    w1_t[0] = swap32 (salt_buf1[0]);
-    w1_t[1] = swap32 (salt_buf1[1]);
-    w1_t[2] = swap32 (salt_buf1[2]);
-    w1_t[3] = swap32 (salt_buf1[3]);
-    w2_t[0] = 0;
-    w2_t[1] = 0;
-    w2_t[2] = 0;
-    w2_t[3] = 0;
-    w3_t[0] = 0;
-    w3_t[1] = 0;
+    w0_t[0] = salt_buf0[0];
+    w0_t[1] = salt_buf0[1];
+    w0_t[2] = salt_buf0[2];
+    w0_t[3] = salt_buf0[3];
+    w1_t[0] = salt_buf1[0];
+    w1_t[1] = salt_buf1[1];
+    w1_t[2] = salt_buf1[2];
+    w1_t[3] = salt_buf1[3];
+    w2_t[0] = salt_buf2[0];
+    w2_t[1] = salt_buf2[1];
+    w2_t[2] = salt_buf2[2];
+    w2_t[3] = salt_buf2[3];
+    w3_t[0] = salt_buf3[0];
+    w3_t[1] = salt_buf3[1];
     w3_t[2] = 0;
     w3_t[3] = (64 + salt_len) * 8;
 
@@ -344,18 +346,26 @@ static void m01450s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
    */
 
   u32 salt_buf0[4];
-
-  salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0];
-  salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1];
-  salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2];
-  salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3];
-
   u32 salt_buf1[4];
-
-  salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
-  salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
-  salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
-  salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
+  u32 salt_buf2[4];
+  u32 salt_buf3[4];
+
+  salt_buf0[0] = swap32_S (salt_bufs[salt_pos].salt_buf[ 0]);
+  salt_buf0[1] = swap32_S (salt_bufs[salt_pos].salt_buf[ 1]);
+  salt_buf0[2] = swap32_S (salt_bufs[salt_pos].salt_buf[ 2]);
+  salt_buf0[3] = swap32_S (salt_bufs[salt_pos].salt_buf[ 3]);
+  salt_buf1[0] = swap32_S (salt_bufs[salt_pos].salt_buf[ 4]);
+  salt_buf1[1] = swap32_S (salt_bufs[salt_pos].salt_buf[ 5]);
+  salt_buf1[2] = swap32_S (salt_bufs[salt_pos].salt_buf[ 6]);
+  salt_buf1[3] = swap32_S (salt_bufs[salt_pos].salt_buf[ 7]);
+  salt_buf2[0] = swap32_S (salt_bufs[salt_pos].salt_buf[ 8]);
+  salt_buf2[1] = swap32_S (salt_bufs[salt_pos].salt_buf[ 9]);
+  salt_buf2[2] = swap32_S (salt_bufs[salt_pos].salt_buf[10]);
+  salt_buf2[3] = swap32_S (salt_bufs[salt_pos].salt_buf[11]);
+  salt_buf3[0] = swap32_S (salt_bufs[salt_pos].salt_buf[12]);
+  salt_buf3[1] = swap32_S (salt_bufs[salt_pos].salt_buf[13]);
+  salt_buf3[2] = swap32_S (salt_bufs[salt_pos].salt_buf[14]);
+  salt_buf3[3] = swap32_S (salt_bufs[salt_pos].salt_buf[15]);
 
   const u32 salt_len = salt_bufs[salt_pos].salt_len;
 
@@ -388,52 +398,46 @@ static void m01450s (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] = w0lr;
     w0_t[1] = w0[1];
     w0_t[2] = w0[2];
     w0_t[3] = w0[3];
-
-    u32x w1_t[4];
-
     w1_t[0] = w1[0];
     w1_t[1] = w1[1];
     w1_t[2] = w1[2];
     w1_t[3] = w1[3];
-
-    u32x w2_t[4];
-
     w2_t[0] = w2[0];
     w2_t[1] = w2[1];
     w2_t[2] = w2[2];
     w2_t[3] = w2[3];
-
-    u32x w3_t[4];
-
     w3_t[0] = w3[0];
     w3_t[1] = w3[1];
-    w3_t[2] = 0;
-    w3_t[3] = 0;
+    w3_t[2] = w3[2];
+    w3_t[3] = w3[3];
 
     u32x ipad[8];
     u32x opad[8];
 
     hmac_sha256_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
 
-    w0_t[0] = swap32 (salt_buf0[0]);
-    w0_t[1] = swap32 (salt_buf0[1]);
-    w0_t[2] = swap32 (salt_buf0[2]);
-    w0_t[3] = swap32 (salt_buf0[3]);
-    w1_t[0] = swap32 (salt_buf1[0]);
-    w1_t[1] = swap32 (salt_buf1[1]);
-    w1_t[2] = swap32 (salt_buf1[2]);
-    w1_t[3] = swap32 (salt_buf1[3]);
-    w2_t[0] = 0;
-    w2_t[1] = 0;
-    w2_t[2] = 0;
-    w2_t[3] = 0;
-    w3_t[0] = 0;
-    w3_t[1] = 0;
+    w0_t[0] = salt_buf0[0];
+    w0_t[1] = salt_buf0[1];
+    w0_t[2] = salt_buf0[2];
+    w0_t[3] = salt_buf0[3];
+    w1_t[0] = salt_buf1[0];
+    w1_t[1] = salt_buf1[1];
+    w1_t[2] = salt_buf1[2];
+    w1_t[3] = salt_buf1[3];
+    w2_t[0] = salt_buf2[0];
+    w2_t[1] = salt_buf2[1];
+    w2_t[2] = salt_buf2[2];
+    w2_t[3] = salt_buf2[3];
+    w3_t[0] = salt_buf3[0];
+    w3_t[1] = salt_buf3[1];
     w3_t[2] = 0;
     w3_t[3] = (64 + salt_len) * 8;