diff --git a/OpenCL/inc_cipher_aes.cl b/OpenCL/inc_cipher_aes.cl
index 12d928c03..6a706cf70 100644
--- a/OpenCL/inc_cipher_aes.cl
+++ b/OpenCL/inc_cipher_aes.cl
@@ -700,9 +700,6 @@ DECLSPEC void aes128_ExpandKey (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, S
   ks[2] = ukey[2];
   ks[3] = ukey[3];
 
-  #ifdef _unroll
-  #pragma unroll
-  #endif
   for (int i = 0, j = 0; i < 10; i += 1, j += 4)
   {
     u32 temp = ks[j + 3];
@@ -748,9 +745,6 @@ DECLSPEC void aes128_InvertKey (u32 *ks, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te
   temp = ks[18]; ks[18] = ks[26]; ks[26] = temp;
   temp = ks[19]; ks[19] = ks[27]; ks[27] = temp;
 
-  #ifdef _unroll
-  #pragma unroll
-  #endif
   for (int i = 1, j = 4; i < 10; i += 1, j += 4)
   {
     const u32 x0s0 = (ks[j + 0] >>  0) & 0xff;
@@ -827,83 +821,82 @@ DECLSPEC void aes128_set_decrypt_key (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_
 DECLSPEC void aes128_encrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4);
 DECLSPEC void aes128_encrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
 {
-  const u32 in_s0 = swap32_S (in[0]);
-  const u32 in_s1 = swap32_S (in[1]);
-  const u32 in_s2 = swap32_S (in[2]);
-  const u32 in_s3 = swap32_S (in[3]);
+  u32 in_s[4];
 
-  u32 t0 = in_s0 ^ ks[0];
-  u32 t1 = in_s1 ^ ks[1];
-  u32 t2 = in_s2 ^ ks[2];
-  u32 t3 = in_s3 ^ ks[3];
+  in_s[0] = swap32_S (in[0]);
+  in_s[1] = swap32_S (in[1]);
+  in_s[2] = swap32_S (in[2]);
+  in_s[3] = swap32_S (in[3]);
 
-  #ifdef _unroll
-  #pragma unroll
-  #endif
-  for (int i = 4; i < 40; i += 4)
-  {
-    const u32 x0s0 = (t0 >>  0) & 0xff;
-    const u32 x0s1 = (t0 >>  8) & 0xff;
-    const u32 x0s2 = (t0 >> 16) & 0xff;
-    const u32 x0s3 = (t0 >> 24) & 0xff;
-    const u32 x1s0 = (t1 >>  0) & 0xff;
-    const u32 x1s1 = (t1 >>  8) & 0xff;
-    const u32 x1s2 = (t1 >> 16) & 0xff;
-    const u32 x1s3 = (t1 >> 24) & 0xff;
-    const u32 x2s0 = (t2 >>  0) & 0xff;
-    const u32 x2s1 = (t2 >>  8) & 0xff;
-    const u32 x2s2 = (t2 >> 16) & 0xff;
-    const u32 x2s3 = (t2 >> 24) & 0xff;
-    const u32 x3s0 = (t3 >>  0) & 0xff;
-    const u32 x3s1 = (t3 >>  8) & 0xff;
-    const u32 x3s2 = (t3 >> 16) & 0xff;
-    const u32 x3s3 = (t3 >> 24) & 0xff;
+  u32 s0 = in_s[0] ^ ks[0];
+  u32 s1 = in_s[1] ^ ks[1];
+  u32 s2 = in_s[2] ^ ks[2];
+  u32 s3 = in_s[3] ^ ks[3];
 
-    t0 = s_te0[x0s3] ^ s_te1[x1s2] ^ s_te2[x2s1] ^ s_te3[x3s0] ^ ks[i + 0];
-    t1 = s_te0[x1s3] ^ s_te1[x2s2] ^ s_te2[x3s1] ^ s_te3[x0s0] ^ ks[i + 1];
-    t2 = s_te0[x2s3] ^ s_te1[x3s2] ^ s_te2[x0s1] ^ s_te3[x1s0] ^ ks[i + 2];
-    t3 = s_te0[x3s3] ^ s_te1[x0s2] ^ s_te2[x1s1] ^ s_te3[x2s0] ^ ks[i + 3];
-  }
+  u32 t0;
+  u32 t1;
+  u32 t2;
+  u32 t3;
 
-  const u32 x0s0 = (t0 >>  0) & 0xff;
-  const u32 x0s1 = (t0 >>  8) & 0xff;
-  const u32 x0s2 = (t0 >> 16) & 0xff;
-  const u32 x0s3 = (t0 >> 24) & 0xff;
-  const u32 x1s0 = (t1 >>  0) & 0xff;
-  const u32 x1s1 = (t1 >>  8) & 0xff;
-  const u32 x1s2 = (t1 >> 16) & 0xff;
-  const u32 x1s3 = (t1 >> 24) & 0xff;
-  const u32 x2s0 = (t2 >>  0) & 0xff;
-  const u32 x2s1 = (t2 >>  8) & 0xff;
-  const u32 x2s2 = (t2 >> 16) & 0xff;
-  const u32 x2s3 = (t2 >> 24) & 0xff;
-  const u32 x3s0 = (t3 >>  0) & 0xff;
-  const u32 x3s1 = (t3 >>  8) & 0xff;
-  const u32 x3s2 = (t3 >> 16) & 0xff;
-  const u32 x3s3 = (t3 >> 24) & 0xff;
+  t0 = s_te0[s0 >> 24] ^ s_te1[(s1 >> 16) & 0xff] ^ s_te2[(s2 >>  8) & 0xff] ^ s_te3[s3 & 0xff] ^ ks[ 4];
+  t1 = s_te0[s1 >> 24] ^ s_te1[(s2 >> 16) & 0xff] ^ s_te2[(s3 >>  8) & 0xff] ^ s_te3[s0 & 0xff] ^ ks[ 5];
+  t2 = s_te0[s2 >> 24] ^ s_te1[(s3 >> 16) & 0xff] ^ s_te2[(s0 >>  8) & 0xff] ^ s_te3[s1 & 0xff] ^ ks[ 6];
+  t3 = s_te0[s3 >> 24] ^ s_te1[(s0 >> 16) & 0xff] ^ s_te2[(s1 >>  8) & 0xff] ^ s_te3[s2 & 0xff] ^ ks[ 7];
+  s0 = s_te0[t0 >> 24] ^ s_te1[(t1 >> 16) & 0xff] ^ s_te2[(t2 >>  8) & 0xff] ^ s_te3[t3 & 0xff] ^ ks[ 8];
+  s1 = s_te0[t1 >> 24] ^ s_te1[(t2 >> 16) & 0xff] ^ s_te2[(t3 >>  8) & 0xff] ^ s_te3[t0 & 0xff] ^ ks[ 9];
+  s2 = s_te0[t2 >> 24] ^ s_te1[(t3 >> 16) & 0xff] ^ s_te2[(t0 >>  8) & 0xff] ^ s_te3[t1 & 0xff] ^ ks[10];
+  s3 = s_te0[t3 >> 24] ^ s_te1[(t0 >> 16) & 0xff] ^ s_te2[(t1 >>  8) & 0xff] ^ s_te3[t2 & 0xff] ^ ks[11];
+  t0 = s_te0[s0 >> 24] ^ s_te1[(s1 >> 16) & 0xff] ^ s_te2[(s2 >>  8) & 0xff] ^ s_te3[s3 & 0xff] ^ ks[12];
+  t1 = s_te0[s1 >> 24] ^ s_te1[(s2 >> 16) & 0xff] ^ s_te2[(s3 >>  8) & 0xff] ^ s_te3[s0 & 0xff] ^ ks[13];
+  t2 = s_te0[s2 >> 24] ^ s_te1[(s3 >> 16) & 0xff] ^ s_te2[(s0 >>  8) & 0xff] ^ s_te3[s1 & 0xff] ^ ks[14];
+  t3 = s_te0[s3 >> 24] ^ s_te1[(s0 >> 16) & 0xff] ^ s_te2[(s1 >>  8) & 0xff] ^ s_te3[s2 & 0xff] ^ ks[15];
+  s0 = s_te0[t0 >> 24] ^ s_te1[(t1 >> 16) & 0xff] ^ s_te2[(t2 >>  8) & 0xff] ^ s_te3[t3 & 0xff] ^ ks[16];
+  s1 = s_te0[t1 >> 24] ^ s_te1[(t2 >> 16) & 0xff] ^ s_te2[(t3 >>  8) & 0xff] ^ s_te3[t0 & 0xff] ^ ks[17];
+  s2 = s_te0[t2 >> 24] ^ s_te1[(t3 >> 16) & 0xff] ^ s_te2[(t0 >>  8) & 0xff] ^ s_te3[t1 & 0xff] ^ ks[18];
+  s3 = s_te0[t3 >> 24] ^ s_te1[(t0 >> 16) & 0xff] ^ s_te2[(t1 >>  8) & 0xff] ^ s_te3[t2 & 0xff] ^ ks[19];
+  t0 = s_te0[s0 >> 24] ^ s_te1[(s1 >> 16) & 0xff] ^ s_te2[(s2 >>  8) & 0xff] ^ s_te3[s3 & 0xff] ^ ks[20];
+  t1 = s_te0[s1 >> 24] ^ s_te1[(s2 >> 16) & 0xff] ^ s_te2[(s3 >>  8) & 0xff] ^ s_te3[s0 & 0xff] ^ ks[21];
+  t2 = s_te0[s2 >> 24] ^ s_te1[(s3 >> 16) & 0xff] ^ s_te2[(s0 >>  8) & 0xff] ^ s_te3[s1 & 0xff] ^ ks[22];
+  t3 = s_te0[s3 >> 24] ^ s_te1[(s0 >> 16) & 0xff] ^ s_te2[(s1 >>  8) & 0xff] ^ s_te3[s2 & 0xff] ^ ks[23];
+  s0 = s_te0[t0 >> 24] ^ s_te1[(t1 >> 16) & 0xff] ^ s_te2[(t2 >>  8) & 0xff] ^ s_te3[t3 & 0xff] ^ ks[24];
+  s1 = s_te0[t1 >> 24] ^ s_te1[(t2 >> 16) & 0xff] ^ s_te2[(t3 >>  8) & 0xff] ^ s_te3[t0 & 0xff] ^ ks[25];
+  s2 = s_te0[t2 >> 24] ^ s_te1[(t3 >> 16) & 0xff] ^ s_te2[(t0 >>  8) & 0xff] ^ s_te3[t1 & 0xff] ^ ks[26];
+  s3 = s_te0[t3 >> 24] ^ s_te1[(t0 >> 16) & 0xff] ^ s_te2[(t1 >>  8) & 0xff] ^ s_te3[t2 & 0xff] ^ ks[27];
+  t0 = s_te0[s0 >> 24] ^ s_te1[(s1 >> 16) & 0xff] ^ s_te2[(s2 >>  8) & 0xff] ^ s_te3[s3 & 0xff] ^ ks[28];
+  t1 = s_te0[s1 >> 24] ^ s_te1[(s2 >> 16) & 0xff] ^ s_te2[(s3 >>  8) & 0xff] ^ s_te3[s0 & 0xff] ^ ks[29];
+  t2 = s_te0[s2 >> 24] ^ s_te1[(s3 >> 16) & 0xff] ^ s_te2[(s0 >>  8) & 0xff] ^ s_te3[s1 & 0xff] ^ ks[30];
+  t3 = s_te0[s3 >> 24] ^ s_te1[(s0 >> 16) & 0xff] ^ s_te2[(s1 >>  8) & 0xff] ^ s_te3[s2 & 0xff] ^ ks[31];
+  s0 = s_te0[t0 >> 24] ^ s_te1[(t1 >> 16) & 0xff] ^ s_te2[(t2 >>  8) & 0xff] ^ s_te3[t3 & 0xff] ^ ks[32];
+  s1 = s_te0[t1 >> 24] ^ s_te1[(t2 >> 16) & 0xff] ^ s_te2[(t3 >>  8) & 0xff] ^ s_te3[t0 & 0xff] ^ ks[33];
+  s2 = s_te0[t2 >> 24] ^ s_te1[(t3 >> 16) & 0xff] ^ s_te2[(t0 >>  8) & 0xff] ^ s_te3[t1 & 0xff] ^ ks[34];
+  s3 = s_te0[t3 >> 24] ^ s_te1[(t0 >> 16) & 0xff] ^ s_te2[(t1 >>  8) & 0xff] ^ s_te3[t2 & 0xff] ^ ks[35];
+  t0 = s_te0[s0 >> 24] ^ s_te1[(s1 >> 16) & 0xff] ^ s_te2[(s2 >>  8) & 0xff] ^ s_te3[s3 & 0xff] ^ ks[36];
+  t1 = s_te0[s1 >> 24] ^ s_te1[(s2 >> 16) & 0xff] ^ s_te2[(s3 >>  8) & 0xff] ^ s_te3[s0 & 0xff] ^ ks[37];
+  t2 = s_te0[s2 >> 24] ^ s_te1[(s3 >> 16) & 0xff] ^ s_te2[(s0 >>  8) & 0xff] ^ s_te3[s1 & 0xff] ^ ks[38];
+  t3 = s_te0[s3 >> 24] ^ s_te1[(s0 >> 16) & 0xff] ^ s_te2[(s1 >>  8) & 0xff] ^ s_te3[s2 & 0xff] ^ ks[39];
 
-  out[0] = (s_te4[x0s3] & 0xff000000)
-         ^ (s_te4[x1s2] & 0x00ff0000)
-         ^ (s_te4[x2s1] & 0x0000ff00)
-         ^ (s_te4[x3s0] & 0x000000ff)
+  out[0] = (s_te4[(t0 >> 24) & 0xff] & 0xff000000)
+         ^ (s_te4[(t1 >> 16) & 0xff] & 0x00ff0000)
+         ^ (s_te4[(t2 >>  8) & 0xff] & 0x0000ff00)
+         ^ (s_te4[(t3 >>  0) & 0xff] & 0x000000ff)
          ^ ks[40];
 
-  out[1] = (s_te4[x1s3] & 0xff000000)
-         ^ (s_te4[x2s2] & 0x00ff0000)
-         ^ (s_te4[x3s1] & 0x0000ff00)
-         ^ (s_te4[x0s0] & 0x000000ff)
+  out[1] = (s_te4[(t1 >> 24) & 0xff] & 0xff000000)
+         ^ (s_te4[(t2 >> 16) & 0xff] & 0x00ff0000)
+         ^ (s_te4[(t3 >>  8) & 0xff] & 0x0000ff00)
+         ^ (s_te4[(t0 >>  0) & 0xff] & 0x000000ff)
          ^ ks[41];
 
-  out[2] = (s_te4[x2s3] & 0xff000000)
-         ^ (s_te4[x3s2] & 0x00ff0000)
-         ^ (s_te4[x0s1] & 0x0000ff00)
-         ^ (s_te4[x1s0] & 0x000000ff)
+  out[2] = (s_te4[(t2 >> 24) & 0xff] & 0xff000000)
+         ^ (s_te4[(t3 >> 16) & 0xff] & 0x00ff0000)
+         ^ (s_te4[(t0 >>  8) & 0xff] & 0x0000ff00)
+         ^ (s_te4[(t1 >>  0) & 0xff] & 0x000000ff)
          ^ ks[42];
 
-  out[3] = (s_te4[x3s3] & 0xff000000)
-         ^ (s_te4[x0s2] & 0x00ff0000)
-         ^ (s_te4[x1s1] & 0x0000ff00)
-         ^ (s_te4[x2s0] & 0x000000ff)
+  out[3] = (s_te4[(t3 >> 24) & 0xff] & 0xff000000)
+         ^ (s_te4[(t0 >> 16) & 0xff] & 0x00ff0000)
+         ^ (s_te4[(t1 >>  8) & 0xff] & 0x0000ff00)
+         ^ (s_te4[(t2 >>  0) & 0xff] & 0x000000ff)
          ^ ks[43];
 
   out[0] = swap32_S (out[0]);
@@ -915,83 +908,82 @@ DECLSPEC void aes128_encrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u
 DECLSPEC void aes128_decrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4);
 DECLSPEC void aes128_decrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
-  const u32 in_s0 = swap32_S (in[0]);
-  const u32 in_s1 = swap32_S (in[1]);
-  const u32 in_s2 = swap32_S (in[2]);
-  const u32 in_s3 = swap32_S (in[3]);
+  u32 in_s[4];
 
-  u32 t0 = in_s0 ^ ks[0];
-  u32 t1 = in_s1 ^ ks[1];
-  u32 t2 = in_s2 ^ ks[2];
-  u32 t3 = in_s3 ^ ks[3];
+  in_s[0] = swap32_S (in[0]);
+  in_s[1] = swap32_S (in[1]);
+  in_s[2] = swap32_S (in[2]);
+  in_s[3] = swap32_S (in[3]);
 
-  #ifdef _unroll
-  #pragma unroll
-  #endif
-  for (int i = 4; i < 40; i += 4)
-  {
-    const u32 x0s0 = (t0 >>  0) & 0xff;
-    const u32 x0s1 = (t0 >>  8) & 0xff;
-    const u32 x0s2 = (t0 >> 16) & 0xff;
-    const u32 x0s3 = (t0 >> 24) & 0xff;
-    const u32 x1s0 = (t1 >>  0) & 0xff;
-    const u32 x1s1 = (t1 >>  8) & 0xff;
-    const u32 x1s2 = (t1 >> 16) & 0xff;
-    const u32 x1s3 = (t1 >> 24) & 0xff;
-    const u32 x2s0 = (t2 >>  0) & 0xff;
-    const u32 x2s1 = (t2 >>  8) & 0xff;
-    const u32 x2s2 = (t2 >> 16) & 0xff;
-    const u32 x2s3 = (t2 >> 24) & 0xff;
-    const u32 x3s0 = (t3 >>  0) & 0xff;
-    const u32 x3s1 = (t3 >>  8) & 0xff;
-    const u32 x3s2 = (t3 >> 16) & 0xff;
-    const u32 x3s3 = (t3 >> 24) & 0xff;
+  u32 s0 = in_s[0] ^ ks[0];
+  u32 s1 = in_s[1] ^ ks[1];
+  u32 s2 = in_s[2] ^ ks[2];
+  u32 s3 = in_s[3] ^ ks[3];
 
-    t0 = s_td0[x0s3] ^ s_td1[x3s2] ^ s_td2[x2s1] ^ s_td3[x1s0] ^ ks[i + 0];
-    t1 = s_td0[x1s3] ^ s_td1[x0s2] ^ s_td2[x3s1] ^ s_td3[x2s0] ^ ks[i + 1];
-    t2 = s_td0[x2s3] ^ s_td1[x1s2] ^ s_td2[x0s1] ^ s_td3[x3s0] ^ ks[i + 2];
-    t3 = s_td0[x3s3] ^ s_td1[x2s2] ^ s_td2[x1s1] ^ s_td3[x0s0] ^ ks[i + 3];
-  }
+  u32 t0;
+  u32 t1;
+  u32 t2;
+  u32 t3;
 
-  const u32 x0s0 = (t0 >>  0) & 0xff;
-  const u32 x0s1 = (t0 >>  8) & 0xff;
-  const u32 x0s2 = (t0 >> 16) & 0xff;
-  const u32 x0s3 = (t0 >> 24) & 0xff;
-  const u32 x1s0 = (t1 >>  0) & 0xff;
-  const u32 x1s1 = (t1 >>  8) & 0xff;
-  const u32 x1s2 = (t1 >> 16) & 0xff;
-  const u32 x1s3 = (t1 >> 24) & 0xff;
-  const u32 x2s0 = (t2 >>  0) & 0xff;
-  const u32 x2s1 = (t2 >>  8) & 0xff;
-  const u32 x2s2 = (t2 >> 16) & 0xff;
-  const u32 x2s3 = (t2 >> 24) & 0xff;
-  const u32 x3s0 = (t3 >>  0) & 0xff;
-  const u32 x3s1 = (t3 >>  8) & 0xff;
-  const u32 x3s2 = (t3 >> 16) & 0xff;
-  const u32 x3s3 = (t3 >> 24) & 0xff;
+  t0 = s_td0[s0 >> 24] ^ s_td1[(s3 >> 16) & 0xff] ^ s_td2[(s2 >>  8) & 0xff] ^ s_td3[s1 & 0xff] ^ ks[ 4];
+  t1 = s_td0[s1 >> 24] ^ s_td1[(s0 >> 16) & 0xff] ^ s_td2[(s3 >>  8) & 0xff] ^ s_td3[s2 & 0xff] ^ ks[ 5];
+  t2 = s_td0[s2 >> 24] ^ s_td1[(s1 >> 16) & 0xff] ^ s_td2[(s0 >>  8) & 0xff] ^ s_td3[s3 & 0xff] ^ ks[ 6];
+  t3 = s_td0[s3 >> 24] ^ s_td1[(s2 >> 16) & 0xff] ^ s_td2[(s1 >>  8) & 0xff] ^ s_td3[s0 & 0xff] ^ ks[ 7];
+  s0 = s_td0[t0 >> 24] ^ s_td1[(t3 >> 16) & 0xff] ^ s_td2[(t2 >>  8) & 0xff] ^ s_td3[t1 & 0xff] ^ ks[ 8];
+  s1 = s_td0[t1 >> 24] ^ s_td1[(t0 >> 16) & 0xff] ^ s_td2[(t3 >>  8) & 0xff] ^ s_td3[t2 & 0xff] ^ ks[ 9];
+  s2 = s_td0[t2 >> 24] ^ s_td1[(t1 >> 16) & 0xff] ^ s_td2[(t0 >>  8) & 0xff] ^ s_td3[t3 & 0xff] ^ ks[10];
+  s3 = s_td0[t3 >> 24] ^ s_td1[(t2 >> 16) & 0xff] ^ s_td2[(t1 >>  8) & 0xff] ^ s_td3[t0 & 0xff] ^ ks[11];
+  t0 = s_td0[s0 >> 24] ^ s_td1[(s3 >> 16) & 0xff] ^ s_td2[(s2 >>  8) & 0xff] ^ s_td3[s1 & 0xff] ^ ks[12];
+  t1 = s_td0[s1 >> 24] ^ s_td1[(s0 >> 16) & 0xff] ^ s_td2[(s3 >>  8) & 0xff] ^ s_td3[s2 & 0xff] ^ ks[13];
+  t2 = s_td0[s2 >> 24] ^ s_td1[(s1 >> 16) & 0xff] ^ s_td2[(s0 >>  8) & 0xff] ^ s_td3[s3 & 0xff] ^ ks[14];
+  t3 = s_td0[s3 >> 24] ^ s_td1[(s2 >> 16) & 0xff] ^ s_td2[(s1 >>  8) & 0xff] ^ s_td3[s0 & 0xff] ^ ks[15];
+  s0 = s_td0[t0 >> 24] ^ s_td1[(t3 >> 16) & 0xff] ^ s_td2[(t2 >>  8) & 0xff] ^ s_td3[t1 & 0xff] ^ ks[16];
+  s1 = s_td0[t1 >> 24] ^ s_td1[(t0 >> 16) & 0xff] ^ s_td2[(t3 >>  8) & 0xff] ^ s_td3[t2 & 0xff] ^ ks[17];
+  s2 = s_td0[t2 >> 24] ^ s_td1[(t1 >> 16) & 0xff] ^ s_td2[(t0 >>  8) & 0xff] ^ s_td3[t3 & 0xff] ^ ks[18];
+  s3 = s_td0[t3 >> 24] ^ s_td1[(t2 >> 16) & 0xff] ^ s_td2[(t1 >>  8) & 0xff] ^ s_td3[t0 & 0xff] ^ ks[19];
+  t0 = s_td0[s0 >> 24] ^ s_td1[(s3 >> 16) & 0xff] ^ s_td2[(s2 >>  8) & 0xff] ^ s_td3[s1 & 0xff] ^ ks[20];
+  t1 = s_td0[s1 >> 24] ^ s_td1[(s0 >> 16) & 0xff] ^ s_td2[(s3 >>  8) & 0xff] ^ s_td3[s2 & 0xff] ^ ks[21];
+  t2 = s_td0[s2 >> 24] ^ s_td1[(s1 >> 16) & 0xff] ^ s_td2[(s0 >>  8) & 0xff] ^ s_td3[s3 & 0xff] ^ ks[22];
+  t3 = s_td0[s3 >> 24] ^ s_td1[(s2 >> 16) & 0xff] ^ s_td2[(s1 >>  8) & 0xff] ^ s_td3[s0 & 0xff] ^ ks[23];
+  s0 = s_td0[t0 >> 24] ^ s_td1[(t3 >> 16) & 0xff] ^ s_td2[(t2 >>  8) & 0xff] ^ s_td3[t1 & 0xff] ^ ks[24];
+  s1 = s_td0[t1 >> 24] ^ s_td1[(t0 >> 16) & 0xff] ^ s_td2[(t3 >>  8) & 0xff] ^ s_td3[t2 & 0xff] ^ ks[25];
+  s2 = s_td0[t2 >> 24] ^ s_td1[(t1 >> 16) & 0xff] ^ s_td2[(t0 >>  8) & 0xff] ^ s_td3[t3 & 0xff] ^ ks[26];
+  s3 = s_td0[t3 >> 24] ^ s_td1[(t2 >> 16) & 0xff] ^ s_td2[(t1 >>  8) & 0xff] ^ s_td3[t0 & 0xff] ^ ks[27];
+  t0 = s_td0[s0 >> 24] ^ s_td1[(s3 >> 16) & 0xff] ^ s_td2[(s2 >>  8) & 0xff] ^ s_td3[s1 & 0xff] ^ ks[28];
+  t1 = s_td0[s1 >> 24] ^ s_td1[(s0 >> 16) & 0xff] ^ s_td2[(s3 >>  8) & 0xff] ^ s_td3[s2 & 0xff] ^ ks[29];
+  t2 = s_td0[s2 >> 24] ^ s_td1[(s1 >> 16) & 0xff] ^ s_td2[(s0 >>  8) & 0xff] ^ s_td3[s3 & 0xff] ^ ks[30];
+  t3 = s_td0[s3 >> 24] ^ s_td1[(s2 >> 16) & 0xff] ^ s_td2[(s1 >>  8) & 0xff] ^ s_td3[s0 & 0xff] ^ ks[31];
+  s0 = s_td0[t0 >> 24] ^ s_td1[(t3 >> 16) & 0xff] ^ s_td2[(t2 >>  8) & 0xff] ^ s_td3[t1 & 0xff] ^ ks[32];
+  s1 = s_td0[t1 >> 24] ^ s_td1[(t0 >> 16) & 0xff] ^ s_td2[(t3 >>  8) & 0xff] ^ s_td3[t2 & 0xff] ^ ks[33];
+  s2 = s_td0[t2 >> 24] ^ s_td1[(t1 >> 16) & 0xff] ^ s_td2[(t0 >>  8) & 0xff] ^ s_td3[t3 & 0xff] ^ ks[34];
+  s3 = s_td0[t3 >> 24] ^ s_td1[(t2 >> 16) & 0xff] ^ s_td2[(t1 >>  8) & 0xff] ^ s_td3[t0 & 0xff] ^ ks[35];
+  t0 = s_td0[s0 >> 24] ^ s_td1[(s3 >> 16) & 0xff] ^ s_td2[(s2 >>  8) & 0xff] ^ s_td3[s1 & 0xff] ^ ks[36];
+  t1 = s_td0[s1 >> 24] ^ s_td1[(s0 >> 16) & 0xff] ^ s_td2[(s3 >>  8) & 0xff] ^ s_td3[s2 & 0xff] ^ ks[37];
+  t2 = s_td0[s2 >> 24] ^ s_td1[(s1 >> 16) & 0xff] ^ s_td2[(s0 >>  8) & 0xff] ^ s_td3[s3 & 0xff] ^ ks[38];
+  t3 = s_td0[s3 >> 24] ^ s_td1[(s2 >> 16) & 0xff] ^ s_td2[(s1 >>  8) & 0xff] ^ s_td3[s0 & 0xff] ^ ks[39];
 
-  out[0] = (s_td4[x0s3] & 0xff000000)
-         ^ (s_td4[x3s2] & 0x00ff0000)
-         ^ (s_td4[x2s1] & 0x0000ff00)
-         ^ (s_td4[x1s0] & 0x000000ff)
+  out[0] = (s_td4[(t0 >> 24) & 0xff] & 0xff000000)
+         ^ (s_td4[(t3 >> 16) & 0xff] & 0x00ff0000)
+         ^ (s_td4[(t2 >>  8) & 0xff] & 0x0000ff00)
+         ^ (s_td4[(t1 >>  0) & 0xff] & 0x000000ff)
          ^ ks[40];
 
-  out[1] = (s_td4[x1s3] & 0xff000000)
-         ^ (s_td4[x0s2] & 0x00ff0000)
-         ^ (s_td4[x3s1] & 0x0000ff00)
-         ^ (s_td4[x2s0] & 0x000000ff)
+  out[1] = (s_td4[(t1 >> 24) & 0xff] & 0xff000000)
+         ^ (s_td4[(t0 >> 16) & 0xff] & 0x00ff0000)
+         ^ (s_td4[(t3 >>  8) & 0xff] & 0x0000ff00)
+         ^ (s_td4[(t2 >>  0) & 0xff] & 0x000000ff)
          ^ ks[41];
 
-  out[2] = (s_td4[x2s3] & 0xff000000)
-         ^ (s_td4[x1s2] & 0x00ff0000)
-         ^ (s_td4[x0s1] & 0x0000ff00)
-         ^ (s_td4[x3s0] & 0x000000ff)
+  out[2] = (s_td4[(t2 >> 24) & 0xff] & 0xff000000)
+         ^ (s_td4[(t1 >> 16) & 0xff] & 0x00ff0000)
+         ^ (s_td4[(t0 >>  8) & 0xff] & 0x0000ff00)
+         ^ (s_td4[(t3 >>  0) & 0xff] & 0x000000ff)
          ^ ks[42];
 
-  out[3] = (s_td4[x3s3] & 0xff000000)
-         ^ (s_td4[x2s2] & 0x00ff0000)
-         ^ (s_td4[x1s1] & 0x0000ff00)
-         ^ (s_td4[x0s0] & 0x000000ff)
+  out[3] = (s_td4[(t3 >> 24) & 0xff] & 0xff000000)
+         ^ (s_td4[(t2 >> 16) & 0xff] & 0x00ff0000)
+         ^ (s_td4[(t1 >>  8) & 0xff] & 0x0000ff00)
+         ^ (s_td4[(t0 >>  0) & 0xff] & 0x000000ff)
          ^ ks[43];
 
   out[0] = swap32_S (out[0]);
@@ -1017,9 +1009,6 @@ DECLSPEC void aes256_ExpandKey (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_te0, S
   int i;
   int j;
 
-  #ifdef _unroll
-  #pragma unroll
-  #endif
   for (int i = 0, j = 0; i < 7; i += 1, j += 8)
   {
     const u32 temp1 = ks[j + 7];
@@ -1085,9 +1074,6 @@ DECLSPEC void aes256_InvertKey (u32 *ks, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te
   temp = ks[26]; ks[26] = ks[34]; ks[34] = temp;
   temp = ks[27]; ks[27] = ks[35]; ks[35] = temp;
 
-  #ifdef _unroll
-  #pragma unroll
-  #endif
   for (int i = 1, j = 4; i < 14; i += 1, j += 4)
   {
     const u32 x0s0 = (ks[j + 0] >>  0) & 0xff;
@@ -1172,83 +1158,98 @@ DECLSPEC void aes256_set_decrypt_key (u32 *ks, const u32 *ukey, SHM_TYPE u32 *s_
 DECLSPEC void aes256_encrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4);
 DECLSPEC void aes256_encrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
 {
-  const u32 in_s0 = swap32_S (in[0]);
-  const u32 in_s1 = swap32_S (in[1]);
-  const u32 in_s2 = swap32_S (in[2]);
-  const u32 in_s3 = swap32_S (in[3]);
+  u32 in_s[4];
 
-  u32 t0 = in_s0 ^ ks[0];
-  u32 t1 = in_s1 ^ ks[1];
-  u32 t2 = in_s2 ^ ks[2];
-  u32 t3 = in_s3 ^ ks[3];
+  in_s[0] = swap32_S (in[0]);
+  in_s[1] = swap32_S (in[1]);
+  in_s[2] = swap32_S (in[2]);
+  in_s[3] = swap32_S (in[3]);
 
-  #ifdef _unroll
-  #pragma unroll
-  #endif
-  for (int i = 4; i < 56; i += 4)
-  {
-    const u32 x0s0 = (t0 >>  0) & 0xff;
-    const u32 x0s1 = (t0 >>  8) & 0xff;
-    const u32 x0s2 = (t0 >> 16) & 0xff;
-    const u32 x0s3 = (t0 >> 24) & 0xff;
-    const u32 x1s0 = (t1 >>  0) & 0xff;
-    const u32 x1s1 = (t1 >>  8) & 0xff;
-    const u32 x1s2 = (t1 >> 16) & 0xff;
-    const u32 x1s3 = (t1 >> 24) & 0xff;
-    const u32 x2s0 = (t2 >>  0) & 0xff;
-    const u32 x2s1 = (t2 >>  8) & 0xff;
-    const u32 x2s2 = (t2 >> 16) & 0xff;
-    const u32 x2s3 = (t2 >> 24) & 0xff;
-    const u32 x3s0 = (t3 >>  0) & 0xff;
-    const u32 x3s1 = (t3 >>  8) & 0xff;
-    const u32 x3s2 = (t3 >> 16) & 0xff;
-    const u32 x3s3 = (t3 >> 24) & 0xff;
+  u32 s0 = in_s[0] ^ ks[0];
+  u32 s1 = in_s[1] ^ ks[1];
+  u32 s2 = in_s[2] ^ ks[2];
+  u32 s3 = in_s[3] ^ ks[3];
 
-    t0 = s_te0[x0s3] ^ s_te1[x1s2] ^ s_te2[x2s1] ^ s_te3[x3s0] ^ ks[i + 0];
-    t1 = s_te0[x1s3] ^ s_te1[x2s2] ^ s_te2[x3s1] ^ s_te3[x0s0] ^ ks[i + 1];
-    t2 = s_te0[x2s3] ^ s_te1[x3s2] ^ s_te2[x0s1] ^ s_te3[x1s0] ^ ks[i + 2];
-    t3 = s_te0[x3s3] ^ s_te1[x0s2] ^ s_te2[x1s1] ^ s_te3[x2s0] ^ ks[i + 3];
-  }
+  u32 t0;
+  u32 t1;
+  u32 t2;
+  u32 t3;
 
-  const u32 x0s0 = (t0 >>  0) & 0xff;
-  const u32 x0s1 = (t0 >>  8) & 0xff;
-  const u32 x0s2 = (t0 >> 16) & 0xff;
-  const u32 x0s3 = (t0 >> 24) & 0xff;
-  const u32 x1s0 = (t1 >>  0) & 0xff;
-  const u32 x1s1 = (t1 >>  8) & 0xff;
-  const u32 x1s2 = (t1 >> 16) & 0xff;
-  const u32 x1s3 = (t1 >> 24) & 0xff;
-  const u32 x2s0 = (t2 >>  0) & 0xff;
-  const u32 x2s1 = (t2 >>  8) & 0xff;
-  const u32 x2s2 = (t2 >> 16) & 0xff;
-  const u32 x2s3 = (t2 >> 24) & 0xff;
-  const u32 x3s0 = (t3 >>  0) & 0xff;
-  const u32 x3s1 = (t3 >>  8) & 0xff;
-  const u32 x3s2 = (t3 >> 16) & 0xff;
-  const u32 x3s3 = (t3 >> 24) & 0xff;
+  t0 = s_te0[s0 >> 24] ^ s_te1[(s1 >> 16) & 0xff] ^ s_te2[(s2 >>  8) & 0xff] ^ s_te3[s3 & 0xff] ^ ks[ 4];
+  t1 = s_te0[s1 >> 24] ^ s_te1[(s2 >> 16) & 0xff] ^ s_te2[(s3 >>  8) & 0xff] ^ s_te3[s0 & 0xff] ^ ks[ 5];
+  t2 = s_te0[s2 >> 24] ^ s_te1[(s3 >> 16) & 0xff] ^ s_te2[(s0 >>  8) & 0xff] ^ s_te3[s1 & 0xff] ^ ks[ 6];
+  t3 = s_te0[s3 >> 24] ^ s_te1[(s0 >> 16) & 0xff] ^ s_te2[(s1 >>  8) & 0xff] ^ s_te3[s2 & 0xff] ^ ks[ 7];
+  s0 = s_te0[t0 >> 24] ^ s_te1[(t1 >> 16) & 0xff] ^ s_te2[(t2 >>  8) & 0xff] ^ s_te3[t3 & 0xff] ^ ks[ 8];
+  s1 = s_te0[t1 >> 24] ^ s_te1[(t2 >> 16) & 0xff] ^ s_te2[(t3 >>  8) & 0xff] ^ s_te3[t0 & 0xff] ^ ks[ 9];
+  s2 = s_te0[t2 >> 24] ^ s_te1[(t3 >> 16) & 0xff] ^ s_te2[(t0 >>  8) & 0xff] ^ s_te3[t1 & 0xff] ^ ks[10];
+  s3 = s_te0[t3 >> 24] ^ s_te1[(t0 >> 16) & 0xff] ^ s_te2[(t1 >>  8) & 0xff] ^ s_te3[t2 & 0xff] ^ ks[11];
+  t0 = s_te0[s0 >> 24] ^ s_te1[(s1 >> 16) & 0xff] ^ s_te2[(s2 >>  8) & 0xff] ^ s_te3[s3 & 0xff] ^ ks[12];
+  t1 = s_te0[s1 >> 24] ^ s_te1[(s2 >> 16) & 0xff] ^ s_te2[(s3 >>  8) & 0xff] ^ s_te3[s0 & 0xff] ^ ks[13];
+  t2 = s_te0[s2 >> 24] ^ s_te1[(s3 >> 16) & 0xff] ^ s_te2[(s0 >>  8) & 0xff] ^ s_te3[s1 & 0xff] ^ ks[14];
+  t3 = s_te0[s3 >> 24] ^ s_te1[(s0 >> 16) & 0xff] ^ s_te2[(s1 >>  8) & 0xff] ^ s_te3[s2 & 0xff] ^ ks[15];
+  s0 = s_te0[t0 >> 24] ^ s_te1[(t1 >> 16) & 0xff] ^ s_te2[(t2 >>  8) & 0xff] ^ s_te3[t3 & 0xff] ^ ks[16];
+  s1 = s_te0[t1 >> 24] ^ s_te1[(t2 >> 16) & 0xff] ^ s_te2[(t3 >>  8) & 0xff] ^ s_te3[t0 & 0xff] ^ ks[17];
+  s2 = s_te0[t2 >> 24] ^ s_te1[(t3 >> 16) & 0xff] ^ s_te2[(t0 >>  8) & 0xff] ^ s_te3[t1 & 0xff] ^ ks[18];
+  s3 = s_te0[t3 >> 24] ^ s_te1[(t0 >> 16) & 0xff] ^ s_te2[(t1 >>  8) & 0xff] ^ s_te3[t2 & 0xff] ^ ks[19];
+  t0 = s_te0[s0 >> 24] ^ s_te1[(s1 >> 16) & 0xff] ^ s_te2[(s2 >>  8) & 0xff] ^ s_te3[s3 & 0xff] ^ ks[20];
+  t1 = s_te0[s1 >> 24] ^ s_te1[(s2 >> 16) & 0xff] ^ s_te2[(s3 >>  8) & 0xff] ^ s_te3[s0 & 0xff] ^ ks[21];
+  t2 = s_te0[s2 >> 24] ^ s_te1[(s3 >> 16) & 0xff] ^ s_te2[(s0 >>  8) & 0xff] ^ s_te3[s1 & 0xff] ^ ks[22];
+  t3 = s_te0[s3 >> 24] ^ s_te1[(s0 >> 16) & 0xff] ^ s_te2[(s1 >>  8) & 0xff] ^ s_te3[s2 & 0xff] ^ ks[23];
+  s0 = s_te0[t0 >> 24] ^ s_te1[(t1 >> 16) & 0xff] ^ s_te2[(t2 >>  8) & 0xff] ^ s_te3[t3 & 0xff] ^ ks[24];
+  s1 = s_te0[t1 >> 24] ^ s_te1[(t2 >> 16) & 0xff] ^ s_te2[(t3 >>  8) & 0xff] ^ s_te3[t0 & 0xff] ^ ks[25];
+  s2 = s_te0[t2 >> 24] ^ s_te1[(t3 >> 16) & 0xff] ^ s_te2[(t0 >>  8) & 0xff] ^ s_te3[t1 & 0xff] ^ ks[26];
+  s3 = s_te0[t3 >> 24] ^ s_te1[(t0 >> 16) & 0xff] ^ s_te2[(t1 >>  8) & 0xff] ^ s_te3[t2 & 0xff] ^ ks[27];
+  t0 = s_te0[s0 >> 24] ^ s_te1[(s1 >> 16) & 0xff] ^ s_te2[(s2 >>  8) & 0xff] ^ s_te3[s3 & 0xff] ^ ks[28];
+  t1 = s_te0[s1 >> 24] ^ s_te1[(s2 >> 16) & 0xff] ^ s_te2[(s3 >>  8) & 0xff] ^ s_te3[s0 & 0xff] ^ ks[29];
+  t2 = s_te0[s2 >> 24] ^ s_te1[(s3 >> 16) & 0xff] ^ s_te2[(s0 >>  8) & 0xff] ^ s_te3[s1 & 0xff] ^ ks[30];
+  t3 = s_te0[s3 >> 24] ^ s_te1[(s0 >> 16) & 0xff] ^ s_te2[(s1 >>  8) & 0xff] ^ s_te3[s2 & 0xff] ^ ks[31];
+  s0 = s_te0[t0 >> 24] ^ s_te1[(t1 >> 16) & 0xff] ^ s_te2[(t2 >>  8) & 0xff] ^ s_te3[t3 & 0xff] ^ ks[32];
+  s1 = s_te0[t1 >> 24] ^ s_te1[(t2 >> 16) & 0xff] ^ s_te2[(t3 >>  8) & 0xff] ^ s_te3[t0 & 0xff] ^ ks[33];
+  s2 = s_te0[t2 >> 24] ^ s_te1[(t3 >> 16) & 0xff] ^ s_te2[(t0 >>  8) & 0xff] ^ s_te3[t1 & 0xff] ^ ks[34];
+  s3 = s_te0[t3 >> 24] ^ s_te1[(t0 >> 16) & 0xff] ^ s_te2[(t1 >>  8) & 0xff] ^ s_te3[t2 & 0xff] ^ ks[35];
+  t0 = s_te0[s0 >> 24] ^ s_te1[(s1 >> 16) & 0xff] ^ s_te2[(s2 >>  8) & 0xff] ^ s_te3[s3 & 0xff] ^ ks[36];
+  t1 = s_te0[s1 >> 24] ^ s_te1[(s2 >> 16) & 0xff] ^ s_te2[(s3 >>  8) & 0xff] ^ s_te3[s0 & 0xff] ^ ks[37];
+  t2 = s_te0[s2 >> 24] ^ s_te1[(s3 >> 16) & 0xff] ^ s_te2[(s0 >>  8) & 0xff] ^ s_te3[s1 & 0xff] ^ ks[38];
+  t3 = s_te0[s3 >> 24] ^ s_te1[(s0 >> 16) & 0xff] ^ s_te2[(s1 >>  8) & 0xff] ^ s_te3[s2 & 0xff] ^ ks[39];
+  s0 = s_te0[t0 >> 24] ^ s_te1[(t1 >> 16) & 0xff] ^ s_te2[(t2 >>  8) & 0xff] ^ s_te3[t3 & 0xff] ^ ks[40];
+  s1 = s_te0[t1 >> 24] ^ s_te1[(t2 >> 16) & 0xff] ^ s_te2[(t3 >>  8) & 0xff] ^ s_te3[t0 & 0xff] ^ ks[41];
+  s2 = s_te0[t2 >> 24] ^ s_te1[(t3 >> 16) & 0xff] ^ s_te2[(t0 >>  8) & 0xff] ^ s_te3[t1 & 0xff] ^ ks[42];
+  s3 = s_te0[t3 >> 24] ^ s_te1[(t0 >> 16) & 0xff] ^ s_te2[(t1 >>  8) & 0xff] ^ s_te3[t2 & 0xff] ^ ks[43];
+  t0 = s_te0[s0 >> 24] ^ s_te1[(s1 >> 16) & 0xff] ^ s_te2[(s2 >>  8) & 0xff] ^ s_te3[s3 & 0xff] ^ ks[44];
+  t1 = s_te0[s1 >> 24] ^ s_te1[(s2 >> 16) & 0xff] ^ s_te2[(s3 >>  8) & 0xff] ^ s_te3[s0 & 0xff] ^ ks[45];
+  t2 = s_te0[s2 >> 24] ^ s_te1[(s3 >> 16) & 0xff] ^ s_te2[(s0 >>  8) & 0xff] ^ s_te3[s1 & 0xff] ^ ks[46];
+  t3 = s_te0[s3 >> 24] ^ s_te1[(s0 >> 16) & 0xff] ^ s_te2[(s1 >>  8) & 0xff] ^ s_te3[s2 & 0xff] ^ ks[47];
+  s0 = s_te0[t0 >> 24] ^ s_te1[(t1 >> 16) & 0xff] ^ s_te2[(t2 >>  8) & 0xff] ^ s_te3[t3 & 0xff] ^ ks[48];
+  s1 = s_te0[t1 >> 24] ^ s_te1[(t2 >> 16) & 0xff] ^ s_te2[(t3 >>  8) & 0xff] ^ s_te3[t0 & 0xff] ^ ks[49];
+  s2 = s_te0[t2 >> 24] ^ s_te1[(t3 >> 16) & 0xff] ^ s_te2[(t0 >>  8) & 0xff] ^ s_te3[t1 & 0xff] ^ ks[50];
+  s3 = s_te0[t3 >> 24] ^ s_te1[(t0 >> 16) & 0xff] ^ s_te2[(t1 >>  8) & 0xff] ^ s_te3[t2 & 0xff] ^ ks[51];
+  t0 = s_te0[s0 >> 24] ^ s_te1[(s1 >> 16) & 0xff] ^ s_te2[(s2 >>  8) & 0xff] ^ s_te3[s3 & 0xff] ^ ks[52];
+  t1 = s_te0[s1 >> 24] ^ s_te1[(s2 >> 16) & 0xff] ^ s_te2[(s3 >>  8) & 0xff] ^ s_te3[s0 & 0xff] ^ ks[53];
+  t2 = s_te0[s2 >> 24] ^ s_te1[(s3 >> 16) & 0xff] ^ s_te2[(s0 >>  8) & 0xff] ^ s_te3[s1 & 0xff] ^ ks[54];
+  t3 = s_te0[s3 >> 24] ^ s_te1[(s0 >> 16) & 0xff] ^ s_te2[(s1 >>  8) & 0xff] ^ s_te3[s2 & 0xff] ^ ks[55];
 
-  out[0] = (s_te4[x0s3] & 0xff000000)
-         ^ (s_te4[x1s2] & 0x00ff0000)
-         ^ (s_te4[x2s1] & 0x0000ff00)
-         ^ (s_te4[x3s0] & 0x000000ff)
+  out[0] = (s_te4[(t0 >> 24) & 0xff] & 0xff000000)
+         ^ (s_te4[(t1 >> 16) & 0xff] & 0x00ff0000)
+         ^ (s_te4[(t2 >>  8) & 0xff] & 0x0000ff00)
+         ^ (s_te4[(t3 >>  0) & 0xff] & 0x000000ff)
          ^ ks[56];
 
-  out[1] = (s_te4[x1s3] & 0xff000000)
-         ^ (s_te4[x2s2] & 0x00ff0000)
-         ^ (s_te4[x3s1] & 0x0000ff00)
-         ^ (s_te4[x0s0] & 0x000000ff)
+  out[1] = (s_te4[(t1 >> 24) & 0xff] & 0xff000000)
+         ^ (s_te4[(t2 >> 16) & 0xff] & 0x00ff0000)
+         ^ (s_te4[(t3 >>  8) & 0xff] & 0x0000ff00)
+         ^ (s_te4[(t0 >>  0) & 0xff] & 0x000000ff)
          ^ ks[57];
 
-  out[2] = (s_te4[x2s3] & 0xff000000)
-         ^ (s_te4[x3s2] & 0x00ff0000)
-         ^ (s_te4[x0s1] & 0x0000ff00)
-         ^ (s_te4[x1s0] & 0x000000ff)
+  out[2] = (s_te4[(t2 >> 24) & 0xff] & 0xff000000)
+         ^ (s_te4[(t3 >> 16) & 0xff] & 0x00ff0000)
+         ^ (s_te4[(t0 >>  8) & 0xff] & 0x0000ff00)
+         ^ (s_te4[(t1 >>  0) & 0xff] & 0x000000ff)
          ^ ks[58];
 
-  out[3] = (s_te4[x3s3] & 0xff000000)
-         ^ (s_te4[x0s2] & 0x00ff0000)
-         ^ (s_te4[x1s1] & 0x0000ff00)
-         ^ (s_te4[x2s0] & 0x000000ff)
+  out[3] = (s_te4[(t3 >> 24) & 0xff] & 0xff000000)
+         ^ (s_te4[(t0 >> 16) & 0xff] & 0x00ff0000)
+         ^ (s_te4[(t1 >>  8) & 0xff] & 0x0000ff00)
+         ^ (s_te4[(t2 >>  0) & 0xff] & 0x000000ff)
          ^ ks[59];
 
   out[0] = swap32_S (out[0]);
@@ -1260,83 +1261,98 @@ DECLSPEC void aes256_encrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u
 DECLSPEC void aes256_decrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4);
 DECLSPEC void aes256_decrypt (const u32 *ks, const u32 *in, u32 *out, SHM_TYPE u32 *s_td0, SHM_TYPE u32 *s_td1, SHM_TYPE u32 *s_td2, SHM_TYPE u32 *s_td3, SHM_TYPE u32 *s_td4)
 {
-  const u32 in_s0 = swap32_S (in[0]);
-  const u32 in_s1 = swap32_S (in[1]);
-  const u32 in_s2 = swap32_S (in[2]);
-  const u32 in_s3 = swap32_S (in[3]);
+  u32 in_s[4];
 
-  u32 t0 = in_s0 ^ ks[0];
-  u32 t1 = in_s1 ^ ks[1];
-  u32 t2 = in_s2 ^ ks[2];
-  u32 t3 = in_s3 ^ ks[3];
+  in_s[0] = swap32_S (in[0]);
+  in_s[1] = swap32_S (in[1]);
+  in_s[2] = swap32_S (in[2]);
+  in_s[3] = swap32_S (in[3]);
 
-  #ifdef _unroll
-  #pragma unroll
-  #endif
-  for (int i = 4; i < 56; i += 4)
-  {
-    const u32 x0s0 = (t0 >>  0) & 0xff;
-    const u32 x0s1 = (t0 >>  8) & 0xff;
-    const u32 x0s2 = (t0 >> 16) & 0xff;
-    const u32 x0s3 = (t0 >> 24) & 0xff;
-    const u32 x1s0 = (t1 >>  0) & 0xff;
-    const u32 x1s1 = (t1 >>  8) & 0xff;
-    const u32 x1s2 = (t1 >> 16) & 0xff;
-    const u32 x1s3 = (t1 >> 24) & 0xff;
-    const u32 x2s0 = (t2 >>  0) & 0xff;
-    const u32 x2s1 = (t2 >>  8) & 0xff;
-    const u32 x2s2 = (t2 >> 16) & 0xff;
-    const u32 x2s3 = (t2 >> 24) & 0xff;
-    const u32 x3s0 = (t3 >>  0) & 0xff;
-    const u32 x3s1 = (t3 >>  8) & 0xff;
-    const u32 x3s2 = (t3 >> 16) & 0xff;
-    const u32 x3s3 = (t3 >> 24) & 0xff;
+  u32 s0 = in_s[0] ^ ks[0];
+  u32 s1 = in_s[1] ^ ks[1];
+  u32 s2 = in_s[2] ^ ks[2];
+  u32 s3 = in_s[3] ^ ks[3];
 
-    t0 = s_td0[x0s3] ^ s_td1[x3s2] ^ s_td2[x2s1] ^ s_td3[x1s0] ^ ks[i + 0];
-    t1 = s_td0[x1s3] ^ s_td1[x0s2] ^ s_td2[x3s1] ^ s_td3[x2s0] ^ ks[i + 1];
-    t2 = s_td0[x2s3] ^ s_td1[x1s2] ^ s_td2[x0s1] ^ s_td3[x3s0] ^ ks[i + 2];
-    t3 = s_td0[x3s3] ^ s_td1[x2s2] ^ s_td2[x1s1] ^ s_td3[x0s0] ^ ks[i + 3];
-  }
+  u32 t0;
+  u32 t1;
+  u32 t2;
+  u32 t3;
 
-  const u32 x0s0 = (t0 >>  0) & 0xff;
-  const u32 x0s1 = (t0 >>  8) & 0xff;
-  const u32 x0s2 = (t0 >> 16) & 0xff;
-  const u32 x0s3 = (t0 >> 24) & 0xff;
-  const u32 x1s0 = (t1 >>  0) & 0xff;
-  const u32 x1s1 = (t1 >>  8) & 0xff;
-  const u32 x1s2 = (t1 >> 16) & 0xff;
-  const u32 x1s3 = (t1 >> 24) & 0xff;
-  const u32 x2s0 = (t2 >>  0) & 0xff;
-  const u32 x2s1 = (t2 >>  8) & 0xff;
-  const u32 x2s2 = (t2 >> 16) & 0xff;
-  const u32 x2s3 = (t2 >> 24) & 0xff;
-  const u32 x3s0 = (t3 >>  0) & 0xff;
-  const u32 x3s1 = (t3 >>  8) & 0xff;
-  const u32 x3s2 = (t3 >> 16) & 0xff;
-  const u32 x3s3 = (t3 >> 24) & 0xff;
+  t0 = s_td0[s0 >> 24] ^ s_td1[(s3 >> 16) & 0xff] ^ s_td2[(s2 >>  8) & 0xff] ^ s_td3[s1 & 0xff] ^ ks[ 4];
+  t1 = s_td0[s1 >> 24] ^ s_td1[(s0 >> 16) & 0xff] ^ s_td2[(s3 >>  8) & 0xff] ^ s_td3[s2 & 0xff] ^ ks[ 5];
+  t2 = s_td0[s2 >> 24] ^ s_td1[(s1 >> 16) & 0xff] ^ s_td2[(s0 >>  8) & 0xff] ^ s_td3[s3 & 0xff] ^ ks[ 6];
+  t3 = s_td0[s3 >> 24] ^ s_td1[(s2 >> 16) & 0xff] ^ s_td2[(s1 >>  8) & 0xff] ^ s_td3[s0 & 0xff] ^ ks[ 7];
+  s0 = s_td0[t0 >> 24] ^ s_td1[(t3 >> 16) & 0xff] ^ s_td2[(t2 >>  8) & 0xff] ^ s_td3[t1 & 0xff] ^ ks[ 8];
+  s1 = s_td0[t1 >> 24] ^ s_td1[(t0 >> 16) & 0xff] ^ s_td2[(t3 >>  8) & 0xff] ^ s_td3[t2 & 0xff] ^ ks[ 9];
+  s2 = s_td0[t2 >> 24] ^ s_td1[(t1 >> 16) & 0xff] ^ s_td2[(t0 >>  8) & 0xff] ^ s_td3[t3 & 0xff] ^ ks[10];
+  s3 = s_td0[t3 >> 24] ^ s_td1[(t2 >> 16) & 0xff] ^ s_td2[(t1 >>  8) & 0xff] ^ s_td3[t0 & 0xff] ^ ks[11];
+  t0 = s_td0[s0 >> 24] ^ s_td1[(s3 >> 16) & 0xff] ^ s_td2[(s2 >>  8) & 0xff] ^ s_td3[s1 & 0xff] ^ ks[12];
+  t1 = s_td0[s1 >> 24] ^ s_td1[(s0 >> 16) & 0xff] ^ s_td2[(s3 >>  8) & 0xff] ^ s_td3[s2 & 0xff] ^ ks[13];
+  t2 = s_td0[s2 >> 24] ^ s_td1[(s1 >> 16) & 0xff] ^ s_td2[(s0 >>  8) & 0xff] ^ s_td3[s3 & 0xff] ^ ks[14];
+  t3 = s_td0[s3 >> 24] ^ s_td1[(s2 >> 16) & 0xff] ^ s_td2[(s1 >>  8) & 0xff] ^ s_td3[s0 & 0xff] ^ ks[15];
+  s0 = s_td0[t0 >> 24] ^ s_td1[(t3 >> 16) & 0xff] ^ s_td2[(t2 >>  8) & 0xff] ^ s_td3[t1 & 0xff] ^ ks[16];
+  s1 = s_td0[t1 >> 24] ^ s_td1[(t0 >> 16) & 0xff] ^ s_td2[(t3 >>  8) & 0xff] ^ s_td3[t2 & 0xff] ^ ks[17];
+  s2 = s_td0[t2 >> 24] ^ s_td1[(t1 >> 16) & 0xff] ^ s_td2[(t0 >>  8) & 0xff] ^ s_td3[t3 & 0xff] ^ ks[18];
+  s3 = s_td0[t3 >> 24] ^ s_td1[(t2 >> 16) & 0xff] ^ s_td2[(t1 >>  8) & 0xff] ^ s_td3[t0 & 0xff] ^ ks[19];
+  t0 = s_td0[s0 >> 24] ^ s_td1[(s3 >> 16) & 0xff] ^ s_td2[(s2 >>  8) & 0xff] ^ s_td3[s1 & 0xff] ^ ks[20];
+  t1 = s_td0[s1 >> 24] ^ s_td1[(s0 >> 16) & 0xff] ^ s_td2[(s3 >>  8) & 0xff] ^ s_td3[s2 & 0xff] ^ ks[21];
+  t2 = s_td0[s2 >> 24] ^ s_td1[(s1 >> 16) & 0xff] ^ s_td2[(s0 >>  8) & 0xff] ^ s_td3[s3 & 0xff] ^ ks[22];
+  t3 = s_td0[s3 >> 24] ^ s_td1[(s2 >> 16) & 0xff] ^ s_td2[(s1 >>  8) & 0xff] ^ s_td3[s0 & 0xff] ^ ks[23];
+  s0 = s_td0[t0 >> 24] ^ s_td1[(t3 >> 16) & 0xff] ^ s_td2[(t2 >>  8) & 0xff] ^ s_td3[t1 & 0xff] ^ ks[24];
+  s1 = s_td0[t1 >> 24] ^ s_td1[(t0 >> 16) & 0xff] ^ s_td2[(t3 >>  8) & 0xff] ^ s_td3[t2 & 0xff] ^ ks[25];
+  s2 = s_td0[t2 >> 24] ^ s_td1[(t1 >> 16) & 0xff] ^ s_td2[(t0 >>  8) & 0xff] ^ s_td3[t3 & 0xff] ^ ks[26];
+  s3 = s_td0[t3 >> 24] ^ s_td1[(t2 >> 16) & 0xff] ^ s_td2[(t1 >>  8) & 0xff] ^ s_td3[t0 & 0xff] ^ ks[27];
+  t0 = s_td0[s0 >> 24] ^ s_td1[(s3 >> 16) & 0xff] ^ s_td2[(s2 >>  8) & 0xff] ^ s_td3[s1 & 0xff] ^ ks[28];
+  t1 = s_td0[s1 >> 24] ^ s_td1[(s0 >> 16) & 0xff] ^ s_td2[(s3 >>  8) & 0xff] ^ s_td3[s2 & 0xff] ^ ks[29];
+  t2 = s_td0[s2 >> 24] ^ s_td1[(s1 >> 16) & 0xff] ^ s_td2[(s0 >>  8) & 0xff] ^ s_td3[s3 & 0xff] ^ ks[30];
+  t3 = s_td0[s3 >> 24] ^ s_td1[(s2 >> 16) & 0xff] ^ s_td2[(s1 >>  8) & 0xff] ^ s_td3[s0 & 0xff] ^ ks[31];
+  s0 = s_td0[t0 >> 24] ^ s_td1[(t3 >> 16) & 0xff] ^ s_td2[(t2 >>  8) & 0xff] ^ s_td3[t1 & 0xff] ^ ks[32];
+  s1 = s_td0[t1 >> 24] ^ s_td1[(t0 >> 16) & 0xff] ^ s_td2[(t3 >>  8) & 0xff] ^ s_td3[t2 & 0xff] ^ ks[33];
+  s2 = s_td0[t2 >> 24] ^ s_td1[(t1 >> 16) & 0xff] ^ s_td2[(t0 >>  8) & 0xff] ^ s_td3[t3 & 0xff] ^ ks[34];
+  s3 = s_td0[t3 >> 24] ^ s_td1[(t2 >> 16) & 0xff] ^ s_td2[(t1 >>  8) & 0xff] ^ s_td3[t0 & 0xff] ^ ks[35];
+  t0 = s_td0[s0 >> 24] ^ s_td1[(s3 >> 16) & 0xff] ^ s_td2[(s2 >>  8) & 0xff] ^ s_td3[s1 & 0xff] ^ ks[36];
+  t1 = s_td0[s1 >> 24] ^ s_td1[(s0 >> 16) & 0xff] ^ s_td2[(s3 >>  8) & 0xff] ^ s_td3[s2 & 0xff] ^ ks[37];
+  t2 = s_td0[s2 >> 24] ^ s_td1[(s1 >> 16) & 0xff] ^ s_td2[(s0 >>  8) & 0xff] ^ s_td3[s3 & 0xff] ^ ks[38];
+  t3 = s_td0[s3 >> 24] ^ s_td1[(s2 >> 16) & 0xff] ^ s_td2[(s1 >>  8) & 0xff] ^ s_td3[s0 & 0xff] ^ ks[39];
+  s0 = s_td0[t0 >> 24] ^ s_td1[(t3 >> 16) & 0xff] ^ s_td2[(t2 >>  8) & 0xff] ^ s_td3[t1 & 0xff] ^ ks[40];
+  s1 = s_td0[t1 >> 24] ^ s_td1[(t0 >> 16) & 0xff] ^ s_td2[(t3 >>  8) & 0xff] ^ s_td3[t2 & 0xff] ^ ks[41];
+  s2 = s_td0[t2 >> 24] ^ s_td1[(t1 >> 16) & 0xff] ^ s_td2[(t0 >>  8) & 0xff] ^ s_td3[t3 & 0xff] ^ ks[42];
+  s3 = s_td0[t3 >> 24] ^ s_td1[(t2 >> 16) & 0xff] ^ s_td2[(t1 >>  8) & 0xff] ^ s_td3[t0 & 0xff] ^ ks[43];
+  t0 = s_td0[s0 >> 24] ^ s_td1[(s3 >> 16) & 0xff] ^ s_td2[(s2 >>  8) & 0xff] ^ s_td3[s1 & 0xff] ^ ks[44];
+  t1 = s_td0[s1 >> 24] ^ s_td1[(s0 >> 16) & 0xff] ^ s_td2[(s3 >>  8) & 0xff] ^ s_td3[s2 & 0xff] ^ ks[45];
+  t2 = s_td0[s2 >> 24] ^ s_td1[(s1 >> 16) & 0xff] ^ s_td2[(s0 >>  8) & 0xff] ^ s_td3[s3 & 0xff] ^ ks[46];
+  t3 = s_td0[s3 >> 24] ^ s_td1[(s2 >> 16) & 0xff] ^ s_td2[(s1 >>  8) & 0xff] ^ s_td3[s0 & 0xff] ^ ks[47];
+  s0 = s_td0[t0 >> 24] ^ s_td1[(t3 >> 16) & 0xff] ^ s_td2[(t2 >>  8) & 0xff] ^ s_td3[t1 & 0xff] ^ ks[48];
+  s1 = s_td0[t1 >> 24] ^ s_td1[(t0 >> 16) & 0xff] ^ s_td2[(t3 >>  8) & 0xff] ^ s_td3[t2 & 0xff] ^ ks[49];
+  s2 = s_td0[t2 >> 24] ^ s_td1[(t1 >> 16) & 0xff] ^ s_td2[(t0 >>  8) & 0xff] ^ s_td3[t3 & 0xff] ^ ks[50];
+  s3 = s_td0[t3 >> 24] ^ s_td1[(t2 >> 16) & 0xff] ^ s_td2[(t1 >>  8) & 0xff] ^ s_td3[t0 & 0xff] ^ ks[51];
+  t0 = s_td0[s0 >> 24] ^ s_td1[(s3 >> 16) & 0xff] ^ s_td2[(s2 >>  8) & 0xff] ^ s_td3[s1 & 0xff] ^ ks[52];
+  t1 = s_td0[s1 >> 24] ^ s_td1[(s0 >> 16) & 0xff] ^ s_td2[(s3 >>  8) & 0xff] ^ s_td3[s2 & 0xff] ^ ks[53];
+  t2 = s_td0[s2 >> 24] ^ s_td1[(s1 >> 16) & 0xff] ^ s_td2[(s0 >>  8) & 0xff] ^ s_td3[s3 & 0xff] ^ ks[54];
+  t3 = s_td0[s3 >> 24] ^ s_td1[(s2 >> 16) & 0xff] ^ s_td2[(s1 >>  8) & 0xff] ^ s_td3[s0 & 0xff] ^ ks[55];
 
-  out[0] = (s_td4[x0s3] & 0xff000000)
-         ^ (s_td4[x3s2] & 0x00ff0000)
-         ^ (s_td4[x2s1] & 0x0000ff00)
-         ^ (s_td4[x1s0] & 0x000000ff)
+  out[0] = (s_td4[(t0 >> 24) & 0xff] & 0xff000000)
+         ^ (s_td4[(t3 >> 16) & 0xff] & 0x00ff0000)
+         ^ (s_td4[(t2 >>  8) & 0xff] & 0x0000ff00)
+         ^ (s_td4[(t1 >>  0) & 0xff] & 0x000000ff)
          ^ ks[56];
 
-  out[1] = (s_td4[x1s3] & 0xff000000)
-         ^ (s_td4[x0s2] & 0x00ff0000)
-         ^ (s_td4[x3s1] & 0x0000ff00)
-         ^ (s_td4[x2s0] & 0x000000ff)
+  out[1] = (s_td4[(t1 >> 24) & 0xff] & 0xff000000)
+         ^ (s_td4[(t0 >> 16) & 0xff] & 0x00ff0000)
+         ^ (s_td4[(t3 >>  8) & 0xff] & 0x0000ff00)
+         ^ (s_td4[(t2 >>  0) & 0xff] & 0x000000ff)
          ^ ks[57];
 
-  out[2] = (s_td4[x2s3] & 0xff000000)
-         ^ (s_td4[x1s2] & 0x00ff0000)
-         ^ (s_td4[x0s1] & 0x0000ff00)
-         ^ (s_td4[x3s0] & 0x000000ff)
+  out[2] = (s_td4[(t2 >> 24) & 0xff] & 0xff000000)
+         ^ (s_td4[(t1 >> 16) & 0xff] & 0x00ff0000)
+         ^ (s_td4[(t0 >>  8) & 0xff] & 0x0000ff00)
+         ^ (s_td4[(t3 >>  0) & 0xff] & 0x000000ff)
          ^ ks[58];
 
-  out[3] = (s_td4[x3s3] & 0xff000000)
-         ^ (s_td4[x2s2] & 0x00ff0000)
-         ^ (s_td4[x1s1] & 0x0000ff00)
-         ^ (s_td4[x0s0] & 0x000000ff)
+  out[3] = (s_td4[(t3 >> 24) & 0xff] & 0xff000000)
+         ^ (s_td4[(t2 >> 16) & 0xff] & 0x00ff0000)
+         ^ (s_td4[(t1 >>  8) & 0xff] & 0x0000ff00)
+         ^ (s_td4[(t0 >>  0) & 0xff] & 0x000000ff)
          ^ ks[59];
 
   out[0] = swap32_S (out[0]);
diff --git a/OpenCL/inc_cipher_des.cl b/OpenCL/inc_cipher_des.cl
index 1272b6064..503221532 100644
--- a/OpenCL/inc_cipher_des.cl
+++ b/OpenCL/inc_cipher_des.cl
@@ -366,9 +366,6 @@ DECLSPEC void _des_crypt_encrypt (u32x *iv, u32x *data, u32x *Kc, u32x *Kd, SHM_
   r = rotl32 (r, 3u);
   l = rotl32 (l, 3u);
 
-  #ifdef _unroll
-  #pragma unroll
-  #endif
   for (u32 i = 0; i < 16; i += 2)
   {
     u32x u;
@@ -419,9 +416,6 @@ DECLSPEC void _des_crypt_decrypt (u32x *iv, u32x *data, u32x *Kc, u32x *Kd, SHM_
   r = rotl32 (r, 3u);
   l = rotl32 (l, 3u);
 
-  #ifdef _unroll
-  #pragma unroll
-  #endif
   for (u32 i = 16; i > 0; i -= 2)
   {
     u32x u;
@@ -478,9 +472,6 @@ DECLSPEC void _des_crypt_keysetup (u32x c, u32x d, u32x *Kc, u32x *Kd, SHM_TYPE
 
   c = c & 0x0fffffff;
 
-  #ifdef _unroll
-  #pragma unroll
-  #endif
   for (u32 i = 0; i < 16; i++)
   {
     if ((i < 2) || (i == 8) || (i == 15))
diff --git a/OpenCL/inc_cipher_serpent.cl b/OpenCL/inc_cipher_serpent.cl
index 70f4b8639..19aadd954 100644
--- a/OpenCL/inc_cipher_serpent.cl
+++ b/OpenCL/inc_cipher_serpent.cl
@@ -406,31 +406,146 @@
 DECLSPEC void serpent128_set_key (u32 *ks, const u32 *ukey);
 DECLSPEC void serpent128_set_key (u32 *ks, const u32 *ukey)
 {
-  #ifdef _unroll
-  #pragma unroll
-  #endif
-  for (int i = 0; i < 4; i++)
-  {
-    ks[i] = ukey[i];
-  }
-
-  #ifdef _unroll
-  #pragma unroll
-  #endif
-  for (int i = 4; i < 8; i++)
-  {
-    ks[i] = 0;
-  }
-
-  ks[4] = 1;
-
-  #ifdef _unroll
-  #pragma unroll
-  #endif
-  for (int i = 0; i < 132; i++)
-  {
-    ks[i + 8] = rotl32_S (ks[i + 7] ^ ks[i + 5] ^ ks[i + 3] ^ ks[i + 0] ^ 0x9e3779b9 ^ i, 11);
-  }
+  ks[  0] = ukey[0];
+  ks[  1] = ukey[1];
+  ks[  2] = ukey[2];
+  ks[  3] = ukey[3];
+  ks[  4] = 1;
+  ks[  5] = 0;
+  ks[  6] = 0;
+  ks[  7] = 0;
+  ks[  8] = rotl32_S ((ks[  7] ^ ks[  5] ^ ks[  3] ^ ks[  0] ^ 0x9e3779b9 ^   0), 11);
+  ks[  9] = rotl32_S ((ks[  8] ^ ks[  6] ^ ks[  4] ^ ks[  1] ^ 0x9e3779b9 ^   1), 11);
+  ks[ 10] = rotl32_S ((ks[  9] ^ ks[  7] ^ ks[  5] ^ ks[  2] ^ 0x9e3779b9 ^   2), 11);
+  ks[ 11] = rotl32_S ((ks[ 10] ^ ks[  8] ^ ks[  6] ^ ks[  3] ^ 0x9e3779b9 ^   3), 11);
+  ks[ 12] = rotl32_S ((ks[ 11] ^ ks[  9] ^ ks[  7] ^ ks[  4] ^ 0x9e3779b9 ^   4), 11);
+  ks[ 13] = rotl32_S ((ks[ 12] ^ ks[ 10] ^ ks[  8] ^ ks[  5] ^ 0x9e3779b9 ^   5), 11);
+  ks[ 14] = rotl32_S ((ks[ 13] ^ ks[ 11] ^ ks[  9] ^ ks[  6] ^ 0x9e3779b9 ^   6), 11);
+  ks[ 15] = rotl32_S ((ks[ 14] ^ ks[ 12] ^ ks[ 10] ^ ks[  7] ^ 0x9e3779b9 ^   7), 11);
+  ks[ 16] = rotl32_S ((ks[ 15] ^ ks[ 13] ^ ks[ 11] ^ ks[  8] ^ 0x9e3779b9 ^   8), 11);
+  ks[ 17] = rotl32_S ((ks[ 16] ^ ks[ 14] ^ ks[ 12] ^ ks[  9] ^ 0x9e3779b9 ^   9), 11);
+  ks[ 18] = rotl32_S ((ks[ 17] ^ ks[ 15] ^ ks[ 13] ^ ks[ 10] ^ 0x9e3779b9 ^  10), 11);
+  ks[ 19] = rotl32_S ((ks[ 18] ^ ks[ 16] ^ ks[ 14] ^ ks[ 11] ^ 0x9e3779b9 ^  11), 11);
+  ks[ 20] = rotl32_S ((ks[ 19] ^ ks[ 17] ^ ks[ 15] ^ ks[ 12] ^ 0x9e3779b9 ^  12), 11);
+  ks[ 21] = rotl32_S ((ks[ 20] ^ ks[ 18] ^ ks[ 16] ^ ks[ 13] ^ 0x9e3779b9 ^  13), 11);
+  ks[ 22] = rotl32_S ((ks[ 21] ^ ks[ 19] ^ ks[ 17] ^ ks[ 14] ^ 0x9e3779b9 ^  14), 11);
+  ks[ 23] = rotl32_S ((ks[ 22] ^ ks[ 20] ^ ks[ 18] ^ ks[ 15] ^ 0x9e3779b9 ^  15), 11);
+  ks[ 24] = rotl32_S ((ks[ 23] ^ ks[ 21] ^ ks[ 19] ^ ks[ 16] ^ 0x9e3779b9 ^  16), 11);
+  ks[ 25] = rotl32_S ((ks[ 24] ^ ks[ 22] ^ ks[ 20] ^ ks[ 17] ^ 0x9e3779b9 ^  17), 11);
+  ks[ 26] = rotl32_S ((ks[ 25] ^ ks[ 23] ^ ks[ 21] ^ ks[ 18] ^ 0x9e3779b9 ^  18), 11);
+  ks[ 27] = rotl32_S ((ks[ 26] ^ ks[ 24] ^ ks[ 22] ^ ks[ 19] ^ 0x9e3779b9 ^  19), 11);
+  ks[ 28] = rotl32_S ((ks[ 27] ^ ks[ 25] ^ ks[ 23] ^ ks[ 20] ^ 0x9e3779b9 ^  20), 11);
+  ks[ 29] = rotl32_S ((ks[ 28] ^ ks[ 26] ^ ks[ 24] ^ ks[ 21] ^ 0x9e3779b9 ^  21), 11);
+  ks[ 30] = rotl32_S ((ks[ 29] ^ ks[ 27] ^ ks[ 25] ^ ks[ 22] ^ 0x9e3779b9 ^  22), 11);
+  ks[ 31] = rotl32_S ((ks[ 30] ^ ks[ 28] ^ ks[ 26] ^ ks[ 23] ^ 0x9e3779b9 ^  23), 11);
+  ks[ 32] = rotl32_S ((ks[ 31] ^ ks[ 29] ^ ks[ 27] ^ ks[ 24] ^ 0x9e3779b9 ^  24), 11);
+  ks[ 33] = rotl32_S ((ks[ 32] ^ ks[ 30] ^ ks[ 28] ^ ks[ 25] ^ 0x9e3779b9 ^  25), 11);
+  ks[ 34] = rotl32_S ((ks[ 33] ^ ks[ 31] ^ ks[ 29] ^ ks[ 26] ^ 0x9e3779b9 ^  26), 11);
+  ks[ 35] = rotl32_S ((ks[ 34] ^ ks[ 32] ^ ks[ 30] ^ ks[ 27] ^ 0x9e3779b9 ^  27), 11);
+  ks[ 36] = rotl32_S ((ks[ 35] ^ ks[ 33] ^ ks[ 31] ^ ks[ 28] ^ 0x9e3779b9 ^  28), 11);
+  ks[ 37] = rotl32_S ((ks[ 36] ^ ks[ 34] ^ ks[ 32] ^ ks[ 29] ^ 0x9e3779b9 ^  29), 11);
+  ks[ 38] = rotl32_S ((ks[ 37] ^ ks[ 35] ^ ks[ 33] ^ ks[ 30] ^ 0x9e3779b9 ^  30), 11);
+  ks[ 39] = rotl32_S ((ks[ 38] ^ ks[ 36] ^ ks[ 34] ^ ks[ 31] ^ 0x9e3779b9 ^  31), 11);
+  ks[ 40] = rotl32_S ((ks[ 39] ^ ks[ 37] ^ ks[ 35] ^ ks[ 32] ^ 0x9e3779b9 ^  32), 11);
+  ks[ 41] = rotl32_S ((ks[ 40] ^ ks[ 38] ^ ks[ 36] ^ ks[ 33] ^ 0x9e3779b9 ^  33), 11);
+  ks[ 42] = rotl32_S ((ks[ 41] ^ ks[ 39] ^ ks[ 37] ^ ks[ 34] ^ 0x9e3779b9 ^  34), 11);
+  ks[ 43] = rotl32_S ((ks[ 42] ^ ks[ 40] ^ ks[ 38] ^ ks[ 35] ^ 0x9e3779b9 ^  35), 11);
+  ks[ 44] = rotl32_S ((ks[ 43] ^ ks[ 41] ^ ks[ 39] ^ ks[ 36] ^ 0x9e3779b9 ^  36), 11);
+  ks[ 45] = rotl32_S ((ks[ 44] ^ ks[ 42] ^ ks[ 40] ^ ks[ 37] ^ 0x9e3779b9 ^  37), 11);
+  ks[ 46] = rotl32_S ((ks[ 45] ^ ks[ 43] ^ ks[ 41] ^ ks[ 38] ^ 0x9e3779b9 ^  38), 11);
+  ks[ 47] = rotl32_S ((ks[ 46] ^ ks[ 44] ^ ks[ 42] ^ ks[ 39] ^ 0x9e3779b9 ^  39), 11);
+  ks[ 48] = rotl32_S ((ks[ 47] ^ ks[ 45] ^ ks[ 43] ^ ks[ 40] ^ 0x9e3779b9 ^  40), 11);
+  ks[ 49] = rotl32_S ((ks[ 48] ^ ks[ 46] ^ ks[ 44] ^ ks[ 41] ^ 0x9e3779b9 ^  41), 11);
+  ks[ 50] = rotl32_S ((ks[ 49] ^ ks[ 47] ^ ks[ 45] ^ ks[ 42] ^ 0x9e3779b9 ^  42), 11);
+  ks[ 51] = rotl32_S ((ks[ 50] ^ ks[ 48] ^ ks[ 46] ^ ks[ 43] ^ 0x9e3779b9 ^  43), 11);
+  ks[ 52] = rotl32_S ((ks[ 51] ^ ks[ 49] ^ ks[ 47] ^ ks[ 44] ^ 0x9e3779b9 ^  44), 11);
+  ks[ 53] = rotl32_S ((ks[ 52] ^ ks[ 50] ^ ks[ 48] ^ ks[ 45] ^ 0x9e3779b9 ^  45), 11);
+  ks[ 54] = rotl32_S ((ks[ 53] ^ ks[ 51] ^ ks[ 49] ^ ks[ 46] ^ 0x9e3779b9 ^  46), 11);
+  ks[ 55] = rotl32_S ((ks[ 54] ^ ks[ 52] ^ ks[ 50] ^ ks[ 47] ^ 0x9e3779b9 ^  47), 11);
+  ks[ 56] = rotl32_S ((ks[ 55] ^ ks[ 53] ^ ks[ 51] ^ ks[ 48] ^ 0x9e3779b9 ^  48), 11);
+  ks[ 57] = rotl32_S ((ks[ 56] ^ ks[ 54] ^ ks[ 52] ^ ks[ 49] ^ 0x9e3779b9 ^  49), 11);
+  ks[ 58] = rotl32_S ((ks[ 57] ^ ks[ 55] ^ ks[ 53] ^ ks[ 50] ^ 0x9e3779b9 ^  50), 11);
+  ks[ 59] = rotl32_S ((ks[ 58] ^ ks[ 56] ^ ks[ 54] ^ ks[ 51] ^ 0x9e3779b9 ^  51), 11);
+  ks[ 60] = rotl32_S ((ks[ 59] ^ ks[ 57] ^ ks[ 55] ^ ks[ 52] ^ 0x9e3779b9 ^  52), 11);
+  ks[ 61] = rotl32_S ((ks[ 60] ^ ks[ 58] ^ ks[ 56] ^ ks[ 53] ^ 0x9e3779b9 ^  53), 11);
+  ks[ 62] = rotl32_S ((ks[ 61] ^ ks[ 59] ^ ks[ 57] ^ ks[ 54] ^ 0x9e3779b9 ^  54), 11);
+  ks[ 63] = rotl32_S ((ks[ 62] ^ ks[ 60] ^ ks[ 58] ^ ks[ 55] ^ 0x9e3779b9 ^  55), 11);
+  ks[ 64] = rotl32_S ((ks[ 63] ^ ks[ 61] ^ ks[ 59] ^ ks[ 56] ^ 0x9e3779b9 ^  56), 11);
+  ks[ 65] = rotl32_S ((ks[ 64] ^ ks[ 62] ^ ks[ 60] ^ ks[ 57] ^ 0x9e3779b9 ^  57), 11);
+  ks[ 66] = rotl32_S ((ks[ 65] ^ ks[ 63] ^ ks[ 61] ^ ks[ 58] ^ 0x9e3779b9 ^  58), 11);
+  ks[ 67] = rotl32_S ((ks[ 66] ^ ks[ 64] ^ ks[ 62] ^ ks[ 59] ^ 0x9e3779b9 ^  59), 11);
+  ks[ 68] = rotl32_S ((ks[ 67] ^ ks[ 65] ^ ks[ 63] ^ ks[ 60] ^ 0x9e3779b9 ^  60), 11);
+  ks[ 69] = rotl32_S ((ks[ 68] ^ ks[ 66] ^ ks[ 64] ^ ks[ 61] ^ 0x9e3779b9 ^  61), 11);
+  ks[ 70] = rotl32_S ((ks[ 69] ^ ks[ 67] ^ ks[ 65] ^ ks[ 62] ^ 0x9e3779b9 ^  62), 11);
+  ks[ 71] = rotl32_S ((ks[ 70] ^ ks[ 68] ^ ks[ 66] ^ ks[ 63] ^ 0x9e3779b9 ^  63), 11);
+  ks[ 72] = rotl32_S ((ks[ 71] ^ ks[ 69] ^ ks[ 67] ^ ks[ 64] ^ 0x9e3779b9 ^  64), 11);
+  ks[ 73] = rotl32_S ((ks[ 72] ^ ks[ 70] ^ ks[ 68] ^ ks[ 65] ^ 0x9e3779b9 ^  65), 11);
+  ks[ 74] = rotl32_S ((ks[ 73] ^ ks[ 71] ^ ks[ 69] ^ ks[ 66] ^ 0x9e3779b9 ^  66), 11);
+  ks[ 75] = rotl32_S ((ks[ 74] ^ ks[ 72] ^ ks[ 70] ^ ks[ 67] ^ 0x9e3779b9 ^  67), 11);
+  ks[ 76] = rotl32_S ((ks[ 75] ^ ks[ 73] ^ ks[ 71] ^ ks[ 68] ^ 0x9e3779b9 ^  68), 11);
+  ks[ 77] = rotl32_S ((ks[ 76] ^ ks[ 74] ^ ks[ 72] ^ ks[ 69] ^ 0x9e3779b9 ^  69), 11);
+  ks[ 78] = rotl32_S ((ks[ 77] ^ ks[ 75] ^ ks[ 73] ^ ks[ 70] ^ 0x9e3779b9 ^  70), 11);
+  ks[ 79] = rotl32_S ((ks[ 78] ^ ks[ 76] ^ ks[ 74] ^ ks[ 71] ^ 0x9e3779b9 ^  71), 11);
+  ks[ 80] = rotl32_S ((ks[ 79] ^ ks[ 77] ^ ks[ 75] ^ ks[ 72] ^ 0x9e3779b9 ^  72), 11);
+  ks[ 81] = rotl32_S ((ks[ 80] ^ ks[ 78] ^ ks[ 76] ^ ks[ 73] ^ 0x9e3779b9 ^  73), 11);
+  ks[ 82] = rotl32_S ((ks[ 81] ^ ks[ 79] ^ ks[ 77] ^ ks[ 74] ^ 0x9e3779b9 ^  74), 11);
+  ks[ 83] = rotl32_S ((ks[ 82] ^ ks[ 80] ^ ks[ 78] ^ ks[ 75] ^ 0x9e3779b9 ^  75), 11);
+  ks[ 84] = rotl32_S ((ks[ 83] ^ ks[ 81] ^ ks[ 79] ^ ks[ 76] ^ 0x9e3779b9 ^  76), 11);
+  ks[ 85] = rotl32_S ((ks[ 84] ^ ks[ 82] ^ ks[ 80] ^ ks[ 77] ^ 0x9e3779b9 ^  77), 11);
+  ks[ 86] = rotl32_S ((ks[ 85] ^ ks[ 83] ^ ks[ 81] ^ ks[ 78] ^ 0x9e3779b9 ^  78), 11);
+  ks[ 87] = rotl32_S ((ks[ 86] ^ ks[ 84] ^ ks[ 82] ^ ks[ 79] ^ 0x9e3779b9 ^  79), 11);
+  ks[ 88] = rotl32_S ((ks[ 87] ^ ks[ 85] ^ ks[ 83] ^ ks[ 80] ^ 0x9e3779b9 ^  80), 11);
+  ks[ 89] = rotl32_S ((ks[ 88] ^ ks[ 86] ^ ks[ 84] ^ ks[ 81] ^ 0x9e3779b9 ^  81), 11);
+  ks[ 90] = rotl32_S ((ks[ 89] ^ ks[ 87] ^ ks[ 85] ^ ks[ 82] ^ 0x9e3779b9 ^  82), 11);
+  ks[ 91] = rotl32_S ((ks[ 90] ^ ks[ 88] ^ ks[ 86] ^ ks[ 83] ^ 0x9e3779b9 ^  83), 11);
+  ks[ 92] = rotl32_S ((ks[ 91] ^ ks[ 89] ^ ks[ 87] ^ ks[ 84] ^ 0x9e3779b9 ^  84), 11);
+  ks[ 93] = rotl32_S ((ks[ 92] ^ ks[ 90] ^ ks[ 88] ^ ks[ 85] ^ 0x9e3779b9 ^  85), 11);
+  ks[ 94] = rotl32_S ((ks[ 93] ^ ks[ 91] ^ ks[ 89] ^ ks[ 86] ^ 0x9e3779b9 ^  86), 11);
+  ks[ 95] = rotl32_S ((ks[ 94] ^ ks[ 92] ^ ks[ 90] ^ ks[ 87] ^ 0x9e3779b9 ^  87), 11);
+  ks[ 96] = rotl32_S ((ks[ 95] ^ ks[ 93] ^ ks[ 91] ^ ks[ 88] ^ 0x9e3779b9 ^  88), 11);
+  ks[ 97] = rotl32_S ((ks[ 96] ^ ks[ 94] ^ ks[ 92] ^ ks[ 89] ^ 0x9e3779b9 ^  89), 11);
+  ks[ 98] = rotl32_S ((ks[ 97] ^ ks[ 95] ^ ks[ 93] ^ ks[ 90] ^ 0x9e3779b9 ^  90), 11);
+  ks[ 99] = rotl32_S ((ks[ 98] ^ ks[ 96] ^ ks[ 94] ^ ks[ 91] ^ 0x9e3779b9 ^  91), 11);
+  ks[100] = rotl32_S ((ks[ 99] ^ ks[ 97] ^ ks[ 95] ^ ks[ 92] ^ 0x9e3779b9 ^  92), 11);
+  ks[101] = rotl32_S ((ks[100] ^ ks[ 98] ^ ks[ 96] ^ ks[ 93] ^ 0x9e3779b9 ^  93), 11);
+  ks[102] = rotl32_S ((ks[101] ^ ks[ 99] ^ ks[ 97] ^ ks[ 94] ^ 0x9e3779b9 ^  94), 11);
+  ks[103] = rotl32_S ((ks[102] ^ ks[100] ^ ks[ 98] ^ ks[ 95] ^ 0x9e3779b9 ^  95), 11);
+  ks[104] = rotl32_S ((ks[103] ^ ks[101] ^ ks[ 99] ^ ks[ 96] ^ 0x9e3779b9 ^  96), 11);
+  ks[105] = rotl32_S ((ks[104] ^ ks[102] ^ ks[100] ^ ks[ 97] ^ 0x9e3779b9 ^  97), 11);
+  ks[106] = rotl32_S ((ks[105] ^ ks[103] ^ ks[101] ^ ks[ 98] ^ 0x9e3779b9 ^  98), 11);
+  ks[107] = rotl32_S ((ks[106] ^ ks[104] ^ ks[102] ^ ks[ 99] ^ 0x9e3779b9 ^  99), 11);
+  ks[108] = rotl32_S ((ks[107] ^ ks[105] ^ ks[103] ^ ks[100] ^ 0x9e3779b9 ^ 100), 11);
+  ks[109] = rotl32_S ((ks[108] ^ ks[106] ^ ks[104] ^ ks[101] ^ 0x9e3779b9 ^ 101), 11);
+  ks[110] = rotl32_S ((ks[109] ^ ks[107] ^ ks[105] ^ ks[102] ^ 0x9e3779b9 ^ 102), 11);
+  ks[111] = rotl32_S ((ks[110] ^ ks[108] ^ ks[106] ^ ks[103] ^ 0x9e3779b9 ^ 103), 11);
+  ks[112] = rotl32_S ((ks[111] ^ ks[109] ^ ks[107] ^ ks[104] ^ 0x9e3779b9 ^ 104), 11);
+  ks[113] = rotl32_S ((ks[112] ^ ks[110] ^ ks[108] ^ ks[105] ^ 0x9e3779b9 ^ 105), 11);
+  ks[114] = rotl32_S ((ks[113] ^ ks[111] ^ ks[109] ^ ks[106] ^ 0x9e3779b9 ^ 106), 11);
+  ks[115] = rotl32_S ((ks[114] ^ ks[112] ^ ks[110] ^ ks[107] ^ 0x9e3779b9 ^ 107), 11);
+  ks[116] = rotl32_S ((ks[115] ^ ks[113] ^ ks[111] ^ ks[108] ^ 0x9e3779b9 ^ 108), 11);
+  ks[117] = rotl32_S ((ks[116] ^ ks[114] ^ ks[112] ^ ks[109] ^ 0x9e3779b9 ^ 109), 11);
+  ks[118] = rotl32_S ((ks[117] ^ ks[115] ^ ks[113] ^ ks[110] ^ 0x9e3779b9 ^ 110), 11);
+  ks[119] = rotl32_S ((ks[118] ^ ks[116] ^ ks[114] ^ ks[111] ^ 0x9e3779b9 ^ 111), 11);
+  ks[120] = rotl32_S ((ks[119] ^ ks[117] ^ ks[115] ^ ks[112] ^ 0x9e3779b9 ^ 112), 11);
+  ks[121] = rotl32_S ((ks[120] ^ ks[118] ^ ks[116] ^ ks[113] ^ 0x9e3779b9 ^ 113), 11);
+  ks[122] = rotl32_S ((ks[121] ^ ks[119] ^ ks[117] ^ ks[114] ^ 0x9e3779b9 ^ 114), 11);
+  ks[123] = rotl32_S ((ks[122] ^ ks[120] ^ ks[118] ^ ks[115] ^ 0x9e3779b9 ^ 115), 11);
+  ks[124] = rotl32_S ((ks[123] ^ ks[121] ^ ks[119] ^ ks[116] ^ 0x9e3779b9 ^ 116), 11);
+  ks[125] = rotl32_S ((ks[124] ^ ks[122] ^ ks[120] ^ ks[117] ^ 0x9e3779b9 ^ 117), 11);
+  ks[126] = rotl32_S ((ks[125] ^ ks[123] ^ ks[121] ^ ks[118] ^ 0x9e3779b9 ^ 118), 11);
+  ks[127] = rotl32_S ((ks[126] ^ ks[124] ^ ks[122] ^ ks[119] ^ 0x9e3779b9 ^ 119), 11);
+  ks[128] = rotl32_S ((ks[127] ^ ks[125] ^ ks[123] ^ ks[120] ^ 0x9e3779b9 ^ 120), 11);
+  ks[129] = rotl32_S ((ks[128] ^ ks[126] ^ ks[124] ^ ks[121] ^ 0x9e3779b9 ^ 121), 11);
+  ks[130] = rotl32_S ((ks[129] ^ ks[127] ^ ks[125] ^ ks[122] ^ 0x9e3779b9 ^ 122), 11);
+  ks[131] = rotl32_S ((ks[130] ^ ks[128] ^ ks[126] ^ ks[123] ^ 0x9e3779b9 ^ 123), 11);
+  ks[132] = rotl32_S ((ks[131] ^ ks[129] ^ ks[127] ^ ks[124] ^ 0x9e3779b9 ^ 124), 11);
+  ks[133] = rotl32_S ((ks[132] ^ ks[130] ^ ks[128] ^ ks[125] ^ 0x9e3779b9 ^ 125), 11);
+  ks[134] = rotl32_S ((ks[133] ^ ks[131] ^ ks[129] ^ ks[126] ^ 0x9e3779b9 ^ 126), 11);
+  ks[135] = rotl32_S ((ks[134] ^ ks[132] ^ ks[130] ^ ks[127] ^ 0x9e3779b9 ^ 127), 11);
+  ks[136] = rotl32_S ((ks[135] ^ ks[133] ^ ks[131] ^ ks[128] ^ 0x9e3779b9 ^ 128), 11);
+  ks[137] = rotl32_S ((ks[136] ^ ks[134] ^ ks[132] ^ ks[129] ^ 0x9e3779b9 ^ 129), 11);
+  ks[138] = rotl32_S ((ks[137] ^ ks[135] ^ ks[133] ^ ks[130] ^ 0x9e3779b9 ^ 130), 11);
+  ks[139] = rotl32_S ((ks[138] ^ ks[136] ^ ks[134] ^ ks[131] ^ 0x9e3779b9 ^ 131), 11);
 
   u32  a,b,c,d,e,f,g,h;
   u32  t1,t2,t3,t4,t5,t6,t7,t8,t9,t10,t11,t12,t13,t14,t15,t16;
@@ -577,21 +692,146 @@ DECLSPEC void serpent128_decrypt (const u32 *ks, const u32 *in, u32 *out)
 DECLSPEC void serpent256_set_key (u32 *ks, const u32 *ukey);
 DECLSPEC void serpent256_set_key (u32 *ks, const u32 *ukey)
 {
-  #ifdef _unroll
-  #pragma unroll
-  #endif
-  for (int i = 0; i < 8; i++)
-  {
-    ks[i] = ukey[i];
-  }
-
-  #ifdef _unroll
-  #pragma unroll
-  #endif
-  for (int i = 0; i < 132; i++)
-  {
-    ks[i + 8] = rotl32_S (ks[i + 7] ^ ks[i + 5] ^ ks[i + 3] ^ ks[i + 0] ^ 0x9e3779b9 ^ i, 11);
-  }
+  ks[  0] = ukey[0];
+  ks[  1] = ukey[1];
+  ks[  2] = ukey[2];
+  ks[  3] = ukey[3];
+  ks[  4] = ukey[4];
+  ks[  5] = ukey[5];
+  ks[  6] = ukey[6];
+  ks[  7] = ukey[7];
+  ks[  8] = rotl32_S ((ks[  7] ^ ks[  5] ^ ks[  3] ^ ks[  0] ^ 0x9e3779b9 ^   0), 11);
+  ks[  9] = rotl32_S ((ks[  8] ^ ks[  6] ^ ks[  4] ^ ks[  1] ^ 0x9e3779b9 ^   1), 11);
+  ks[ 10] = rotl32_S ((ks[  9] ^ ks[  7] ^ ks[  5] ^ ks[  2] ^ 0x9e3779b9 ^   2), 11);
+  ks[ 11] = rotl32_S ((ks[ 10] ^ ks[  8] ^ ks[  6] ^ ks[  3] ^ 0x9e3779b9 ^   3), 11);
+  ks[ 12] = rotl32_S ((ks[ 11] ^ ks[  9] ^ ks[  7] ^ ks[  4] ^ 0x9e3779b9 ^   4), 11);
+  ks[ 13] = rotl32_S ((ks[ 12] ^ ks[ 10] ^ ks[  8] ^ ks[  5] ^ 0x9e3779b9 ^   5), 11);
+  ks[ 14] = rotl32_S ((ks[ 13] ^ ks[ 11] ^ ks[  9] ^ ks[  6] ^ 0x9e3779b9 ^   6), 11);
+  ks[ 15] = rotl32_S ((ks[ 14] ^ ks[ 12] ^ ks[ 10] ^ ks[  7] ^ 0x9e3779b9 ^   7), 11);
+  ks[ 16] = rotl32_S ((ks[ 15] ^ ks[ 13] ^ ks[ 11] ^ ks[  8] ^ 0x9e3779b9 ^   8), 11);
+  ks[ 17] = rotl32_S ((ks[ 16] ^ ks[ 14] ^ ks[ 12] ^ ks[  9] ^ 0x9e3779b9 ^   9), 11);
+  ks[ 18] = rotl32_S ((ks[ 17] ^ ks[ 15] ^ ks[ 13] ^ ks[ 10] ^ 0x9e3779b9 ^  10), 11);
+  ks[ 19] = rotl32_S ((ks[ 18] ^ ks[ 16] ^ ks[ 14] ^ ks[ 11] ^ 0x9e3779b9 ^  11), 11);
+  ks[ 20] = rotl32_S ((ks[ 19] ^ ks[ 17] ^ ks[ 15] ^ ks[ 12] ^ 0x9e3779b9 ^  12), 11);
+  ks[ 21] = rotl32_S ((ks[ 20] ^ ks[ 18] ^ ks[ 16] ^ ks[ 13] ^ 0x9e3779b9 ^  13), 11);
+  ks[ 22] = rotl32_S ((ks[ 21] ^ ks[ 19] ^ ks[ 17] ^ ks[ 14] ^ 0x9e3779b9 ^  14), 11);
+  ks[ 23] = rotl32_S ((ks[ 22] ^ ks[ 20] ^ ks[ 18] ^ ks[ 15] ^ 0x9e3779b9 ^  15), 11);
+  ks[ 24] = rotl32_S ((ks[ 23] ^ ks[ 21] ^ ks[ 19] ^ ks[ 16] ^ 0x9e3779b9 ^  16), 11);
+  ks[ 25] = rotl32_S ((ks[ 24] ^ ks[ 22] ^ ks[ 20] ^ ks[ 17] ^ 0x9e3779b9 ^  17), 11);
+  ks[ 26] = rotl32_S ((ks[ 25] ^ ks[ 23] ^ ks[ 21] ^ ks[ 18] ^ 0x9e3779b9 ^  18), 11);
+  ks[ 27] = rotl32_S ((ks[ 26] ^ ks[ 24] ^ ks[ 22] ^ ks[ 19] ^ 0x9e3779b9 ^  19), 11);
+  ks[ 28] = rotl32_S ((ks[ 27] ^ ks[ 25] ^ ks[ 23] ^ ks[ 20] ^ 0x9e3779b9 ^  20), 11);
+  ks[ 29] = rotl32_S ((ks[ 28] ^ ks[ 26] ^ ks[ 24] ^ ks[ 21] ^ 0x9e3779b9 ^  21), 11);
+  ks[ 30] = rotl32_S ((ks[ 29] ^ ks[ 27] ^ ks[ 25] ^ ks[ 22] ^ 0x9e3779b9 ^  22), 11);
+  ks[ 31] = rotl32_S ((ks[ 30] ^ ks[ 28] ^ ks[ 26] ^ ks[ 23] ^ 0x9e3779b9 ^  23), 11);
+  ks[ 32] = rotl32_S ((ks[ 31] ^ ks[ 29] ^ ks[ 27] ^ ks[ 24] ^ 0x9e3779b9 ^  24), 11);
+  ks[ 33] = rotl32_S ((ks[ 32] ^ ks[ 30] ^ ks[ 28] ^ ks[ 25] ^ 0x9e3779b9 ^  25), 11);
+  ks[ 34] = rotl32_S ((ks[ 33] ^ ks[ 31] ^ ks[ 29] ^ ks[ 26] ^ 0x9e3779b9 ^  26), 11);
+  ks[ 35] = rotl32_S ((ks[ 34] ^ ks[ 32] ^ ks[ 30] ^ ks[ 27] ^ 0x9e3779b9 ^  27), 11);
+  ks[ 36] = rotl32_S ((ks[ 35] ^ ks[ 33] ^ ks[ 31] ^ ks[ 28] ^ 0x9e3779b9 ^  28), 11);
+  ks[ 37] = rotl32_S ((ks[ 36] ^ ks[ 34] ^ ks[ 32] ^ ks[ 29] ^ 0x9e3779b9 ^  29), 11);
+  ks[ 38] = rotl32_S ((ks[ 37] ^ ks[ 35] ^ ks[ 33] ^ ks[ 30] ^ 0x9e3779b9 ^  30), 11);
+  ks[ 39] = rotl32_S ((ks[ 38] ^ ks[ 36] ^ ks[ 34] ^ ks[ 31] ^ 0x9e3779b9 ^  31), 11);
+  ks[ 40] = rotl32_S ((ks[ 39] ^ ks[ 37] ^ ks[ 35] ^ ks[ 32] ^ 0x9e3779b9 ^  32), 11);
+  ks[ 41] = rotl32_S ((ks[ 40] ^ ks[ 38] ^ ks[ 36] ^ ks[ 33] ^ 0x9e3779b9 ^  33), 11);
+  ks[ 42] = rotl32_S ((ks[ 41] ^ ks[ 39] ^ ks[ 37] ^ ks[ 34] ^ 0x9e3779b9 ^  34), 11);
+  ks[ 43] = rotl32_S ((ks[ 42] ^ ks[ 40] ^ ks[ 38] ^ ks[ 35] ^ 0x9e3779b9 ^  35), 11);
+  ks[ 44] = rotl32_S ((ks[ 43] ^ ks[ 41] ^ ks[ 39] ^ ks[ 36] ^ 0x9e3779b9 ^  36), 11);
+  ks[ 45] = rotl32_S ((ks[ 44] ^ ks[ 42] ^ ks[ 40] ^ ks[ 37] ^ 0x9e3779b9 ^  37), 11);
+  ks[ 46] = rotl32_S ((ks[ 45] ^ ks[ 43] ^ ks[ 41] ^ ks[ 38] ^ 0x9e3779b9 ^  38), 11);
+  ks[ 47] = rotl32_S ((ks[ 46] ^ ks[ 44] ^ ks[ 42] ^ ks[ 39] ^ 0x9e3779b9 ^  39), 11);
+  ks[ 48] = rotl32_S ((ks[ 47] ^ ks[ 45] ^ ks[ 43] ^ ks[ 40] ^ 0x9e3779b9 ^  40), 11);
+  ks[ 49] = rotl32_S ((ks[ 48] ^ ks[ 46] ^ ks[ 44] ^ ks[ 41] ^ 0x9e3779b9 ^  41), 11);
+  ks[ 50] = rotl32_S ((ks[ 49] ^ ks[ 47] ^ ks[ 45] ^ ks[ 42] ^ 0x9e3779b9 ^  42), 11);
+  ks[ 51] = rotl32_S ((ks[ 50] ^ ks[ 48] ^ ks[ 46] ^ ks[ 43] ^ 0x9e3779b9 ^  43), 11);
+  ks[ 52] = rotl32_S ((ks[ 51] ^ ks[ 49] ^ ks[ 47] ^ ks[ 44] ^ 0x9e3779b9 ^  44), 11);
+  ks[ 53] = rotl32_S ((ks[ 52] ^ ks[ 50] ^ ks[ 48] ^ ks[ 45] ^ 0x9e3779b9 ^  45), 11);
+  ks[ 54] = rotl32_S ((ks[ 53] ^ ks[ 51] ^ ks[ 49] ^ ks[ 46] ^ 0x9e3779b9 ^  46), 11);
+  ks[ 55] = rotl32_S ((ks[ 54] ^ ks[ 52] ^ ks[ 50] ^ ks[ 47] ^ 0x9e3779b9 ^  47), 11);
+  ks[ 56] = rotl32_S ((ks[ 55] ^ ks[ 53] ^ ks[ 51] ^ ks[ 48] ^ 0x9e3779b9 ^  48), 11);
+  ks[ 57] = rotl32_S ((ks[ 56] ^ ks[ 54] ^ ks[ 52] ^ ks[ 49] ^ 0x9e3779b9 ^  49), 11);
+  ks[ 58] = rotl32_S ((ks[ 57] ^ ks[ 55] ^ ks[ 53] ^ ks[ 50] ^ 0x9e3779b9 ^  50), 11);
+  ks[ 59] = rotl32_S ((ks[ 58] ^ ks[ 56] ^ ks[ 54] ^ ks[ 51] ^ 0x9e3779b9 ^  51), 11);
+  ks[ 60] = rotl32_S ((ks[ 59] ^ ks[ 57] ^ ks[ 55] ^ ks[ 52] ^ 0x9e3779b9 ^  52), 11);
+  ks[ 61] = rotl32_S ((ks[ 60] ^ ks[ 58] ^ ks[ 56] ^ ks[ 53] ^ 0x9e3779b9 ^  53), 11);
+  ks[ 62] = rotl32_S ((ks[ 61] ^ ks[ 59] ^ ks[ 57] ^ ks[ 54] ^ 0x9e3779b9 ^  54), 11);
+  ks[ 63] = rotl32_S ((ks[ 62] ^ ks[ 60] ^ ks[ 58] ^ ks[ 55] ^ 0x9e3779b9 ^  55), 11);
+  ks[ 64] = rotl32_S ((ks[ 63] ^ ks[ 61] ^ ks[ 59] ^ ks[ 56] ^ 0x9e3779b9 ^  56), 11);
+  ks[ 65] = rotl32_S ((ks[ 64] ^ ks[ 62] ^ ks[ 60] ^ ks[ 57] ^ 0x9e3779b9 ^  57), 11);
+  ks[ 66] = rotl32_S ((ks[ 65] ^ ks[ 63] ^ ks[ 61] ^ ks[ 58] ^ 0x9e3779b9 ^  58), 11);
+  ks[ 67] = rotl32_S ((ks[ 66] ^ ks[ 64] ^ ks[ 62] ^ ks[ 59] ^ 0x9e3779b9 ^  59), 11);
+  ks[ 68] = rotl32_S ((ks[ 67] ^ ks[ 65] ^ ks[ 63] ^ ks[ 60] ^ 0x9e3779b9 ^  60), 11);
+  ks[ 69] = rotl32_S ((ks[ 68] ^ ks[ 66] ^ ks[ 64] ^ ks[ 61] ^ 0x9e3779b9 ^  61), 11);
+  ks[ 70] = rotl32_S ((ks[ 69] ^ ks[ 67] ^ ks[ 65] ^ ks[ 62] ^ 0x9e3779b9 ^  62), 11);
+  ks[ 71] = rotl32_S ((ks[ 70] ^ ks[ 68] ^ ks[ 66] ^ ks[ 63] ^ 0x9e3779b9 ^  63), 11);
+  ks[ 72] = rotl32_S ((ks[ 71] ^ ks[ 69] ^ ks[ 67] ^ ks[ 64] ^ 0x9e3779b9 ^  64), 11);
+  ks[ 73] = rotl32_S ((ks[ 72] ^ ks[ 70] ^ ks[ 68] ^ ks[ 65] ^ 0x9e3779b9 ^  65), 11);
+  ks[ 74] = rotl32_S ((ks[ 73] ^ ks[ 71] ^ ks[ 69] ^ ks[ 66] ^ 0x9e3779b9 ^  66), 11);
+  ks[ 75] = rotl32_S ((ks[ 74] ^ ks[ 72] ^ ks[ 70] ^ ks[ 67] ^ 0x9e3779b9 ^  67), 11);
+  ks[ 76] = rotl32_S ((ks[ 75] ^ ks[ 73] ^ ks[ 71] ^ ks[ 68] ^ 0x9e3779b9 ^  68), 11);
+  ks[ 77] = rotl32_S ((ks[ 76] ^ ks[ 74] ^ ks[ 72] ^ ks[ 69] ^ 0x9e3779b9 ^  69), 11);
+  ks[ 78] = rotl32_S ((ks[ 77] ^ ks[ 75] ^ ks[ 73] ^ ks[ 70] ^ 0x9e3779b9 ^  70), 11);
+  ks[ 79] = rotl32_S ((ks[ 78] ^ ks[ 76] ^ ks[ 74] ^ ks[ 71] ^ 0x9e3779b9 ^  71), 11);
+  ks[ 80] = rotl32_S ((ks[ 79] ^ ks[ 77] ^ ks[ 75] ^ ks[ 72] ^ 0x9e3779b9 ^  72), 11);
+  ks[ 81] = rotl32_S ((ks[ 80] ^ ks[ 78] ^ ks[ 76] ^ ks[ 73] ^ 0x9e3779b9 ^  73), 11);
+  ks[ 82] = rotl32_S ((ks[ 81] ^ ks[ 79] ^ ks[ 77] ^ ks[ 74] ^ 0x9e3779b9 ^  74), 11);
+  ks[ 83] = rotl32_S ((ks[ 82] ^ ks[ 80] ^ ks[ 78] ^ ks[ 75] ^ 0x9e3779b9 ^  75), 11);
+  ks[ 84] = rotl32_S ((ks[ 83] ^ ks[ 81] ^ ks[ 79] ^ ks[ 76] ^ 0x9e3779b9 ^  76), 11);
+  ks[ 85] = rotl32_S ((ks[ 84] ^ ks[ 82] ^ ks[ 80] ^ ks[ 77] ^ 0x9e3779b9 ^  77), 11);
+  ks[ 86] = rotl32_S ((ks[ 85] ^ ks[ 83] ^ ks[ 81] ^ ks[ 78] ^ 0x9e3779b9 ^  78), 11);
+  ks[ 87] = rotl32_S ((ks[ 86] ^ ks[ 84] ^ ks[ 82] ^ ks[ 79] ^ 0x9e3779b9 ^  79), 11);
+  ks[ 88] = rotl32_S ((ks[ 87] ^ ks[ 85] ^ ks[ 83] ^ ks[ 80] ^ 0x9e3779b9 ^  80), 11);
+  ks[ 89] = rotl32_S ((ks[ 88] ^ ks[ 86] ^ ks[ 84] ^ ks[ 81] ^ 0x9e3779b9 ^  81), 11);
+  ks[ 90] = rotl32_S ((ks[ 89] ^ ks[ 87] ^ ks[ 85] ^ ks[ 82] ^ 0x9e3779b9 ^  82), 11);
+  ks[ 91] = rotl32_S ((ks[ 90] ^ ks[ 88] ^ ks[ 86] ^ ks[ 83] ^ 0x9e3779b9 ^  83), 11);
+  ks[ 92] = rotl32_S ((ks[ 91] ^ ks[ 89] ^ ks[ 87] ^ ks[ 84] ^ 0x9e3779b9 ^  84), 11);
+  ks[ 93] = rotl32_S ((ks[ 92] ^ ks[ 90] ^ ks[ 88] ^ ks[ 85] ^ 0x9e3779b9 ^  85), 11);
+  ks[ 94] = rotl32_S ((ks[ 93] ^ ks[ 91] ^ ks[ 89] ^ ks[ 86] ^ 0x9e3779b9 ^  86), 11);
+  ks[ 95] = rotl32_S ((ks[ 94] ^ ks[ 92] ^ ks[ 90] ^ ks[ 87] ^ 0x9e3779b9 ^  87), 11);
+  ks[ 96] = rotl32_S ((ks[ 95] ^ ks[ 93] ^ ks[ 91] ^ ks[ 88] ^ 0x9e3779b9 ^  88), 11);
+  ks[ 97] = rotl32_S ((ks[ 96] ^ ks[ 94] ^ ks[ 92] ^ ks[ 89] ^ 0x9e3779b9 ^  89), 11);
+  ks[ 98] = rotl32_S ((ks[ 97] ^ ks[ 95] ^ ks[ 93] ^ ks[ 90] ^ 0x9e3779b9 ^  90), 11);
+  ks[ 99] = rotl32_S ((ks[ 98] ^ ks[ 96] ^ ks[ 94] ^ ks[ 91] ^ 0x9e3779b9 ^  91), 11);
+  ks[100] = rotl32_S ((ks[ 99] ^ ks[ 97] ^ ks[ 95] ^ ks[ 92] ^ 0x9e3779b9 ^  92), 11);
+  ks[101] = rotl32_S ((ks[100] ^ ks[ 98] ^ ks[ 96] ^ ks[ 93] ^ 0x9e3779b9 ^  93), 11);
+  ks[102] = rotl32_S ((ks[101] ^ ks[ 99] ^ ks[ 97] ^ ks[ 94] ^ 0x9e3779b9 ^  94), 11);
+  ks[103] = rotl32_S ((ks[102] ^ ks[100] ^ ks[ 98] ^ ks[ 95] ^ 0x9e3779b9 ^  95), 11);
+  ks[104] = rotl32_S ((ks[103] ^ ks[101] ^ ks[ 99] ^ ks[ 96] ^ 0x9e3779b9 ^  96), 11);
+  ks[105] = rotl32_S ((ks[104] ^ ks[102] ^ ks[100] ^ ks[ 97] ^ 0x9e3779b9 ^  97), 11);
+  ks[106] = rotl32_S ((ks[105] ^ ks[103] ^ ks[101] ^ ks[ 98] ^ 0x9e3779b9 ^  98), 11);
+  ks[107] = rotl32_S ((ks[106] ^ ks[104] ^ ks[102] ^ ks[ 99] ^ 0x9e3779b9 ^  99), 11);
+  ks[108] = rotl32_S ((ks[107] ^ ks[105] ^ ks[103] ^ ks[100] ^ 0x9e3779b9 ^ 100), 11);
+  ks[109] = rotl32_S ((ks[108] ^ ks[106] ^ ks[104] ^ ks[101] ^ 0x9e3779b9 ^ 101), 11);
+  ks[110] = rotl32_S ((ks[109] ^ ks[107] ^ ks[105] ^ ks[102] ^ 0x9e3779b9 ^ 102), 11);
+  ks[111] = rotl32_S ((ks[110] ^ ks[108] ^ ks[106] ^ ks[103] ^ 0x9e3779b9 ^ 103), 11);
+  ks[112] = rotl32_S ((ks[111] ^ ks[109] ^ ks[107] ^ ks[104] ^ 0x9e3779b9 ^ 104), 11);
+  ks[113] = rotl32_S ((ks[112] ^ ks[110] ^ ks[108] ^ ks[105] ^ 0x9e3779b9 ^ 105), 11);
+  ks[114] = rotl32_S ((ks[113] ^ ks[111] ^ ks[109] ^ ks[106] ^ 0x9e3779b9 ^ 106), 11);
+  ks[115] = rotl32_S ((ks[114] ^ ks[112] ^ ks[110] ^ ks[107] ^ 0x9e3779b9 ^ 107), 11);
+  ks[116] = rotl32_S ((ks[115] ^ ks[113] ^ ks[111] ^ ks[108] ^ 0x9e3779b9 ^ 108), 11);
+  ks[117] = rotl32_S ((ks[116] ^ ks[114] ^ ks[112] ^ ks[109] ^ 0x9e3779b9 ^ 109), 11);
+  ks[118] = rotl32_S ((ks[117] ^ ks[115] ^ ks[113] ^ ks[110] ^ 0x9e3779b9 ^ 110), 11);
+  ks[119] = rotl32_S ((ks[118] ^ ks[116] ^ ks[114] ^ ks[111] ^ 0x9e3779b9 ^ 111), 11);
+  ks[120] = rotl32_S ((ks[119] ^ ks[117] ^ ks[115] ^ ks[112] ^ 0x9e3779b9 ^ 112), 11);
+  ks[121] = rotl32_S ((ks[120] ^ ks[118] ^ ks[116] ^ ks[113] ^ 0x9e3779b9 ^ 113), 11);
+  ks[122] = rotl32_S ((ks[121] ^ ks[119] ^ ks[117] ^ ks[114] ^ 0x9e3779b9 ^ 114), 11);
+  ks[123] = rotl32_S ((ks[122] ^ ks[120] ^ ks[118] ^ ks[115] ^ 0x9e3779b9 ^ 115), 11);
+  ks[124] = rotl32_S ((ks[123] ^ ks[121] ^ ks[119] ^ ks[116] ^ 0x9e3779b9 ^ 116), 11);
+  ks[125] = rotl32_S ((ks[124] ^ ks[122] ^ ks[120] ^ ks[117] ^ 0x9e3779b9 ^ 117), 11);
+  ks[126] = rotl32_S ((ks[125] ^ ks[123] ^ ks[121] ^ ks[118] ^ 0x9e3779b9 ^ 118), 11);
+  ks[127] = rotl32_S ((ks[126] ^ ks[124] ^ ks[122] ^ ks[119] ^ 0x9e3779b9 ^ 119), 11);
+  ks[128] = rotl32_S ((ks[127] ^ ks[125] ^ ks[123] ^ ks[120] ^ 0x9e3779b9 ^ 120), 11);
+  ks[129] = rotl32_S ((ks[128] ^ ks[126] ^ ks[124] ^ ks[121] ^ 0x9e3779b9 ^ 121), 11);
+  ks[130] = rotl32_S ((ks[129] ^ ks[127] ^ ks[125] ^ ks[122] ^ 0x9e3779b9 ^ 122), 11);
+  ks[131] = rotl32_S ((ks[130] ^ ks[128] ^ ks[126] ^ ks[123] ^ 0x9e3779b9 ^ 123), 11);
+  ks[132] = rotl32_S ((ks[131] ^ ks[129] ^ ks[127] ^ ks[124] ^ 0x9e3779b9 ^ 124), 11);
+  ks[133] = rotl32_S ((ks[132] ^ ks[130] ^ ks[128] ^ ks[125] ^ 0x9e3779b9 ^ 125), 11);
+  ks[134] = rotl32_S ((ks[133] ^ ks[131] ^ ks[129] ^ ks[126] ^ 0x9e3779b9 ^ 126), 11);
+  ks[135] = rotl32_S ((ks[134] ^ ks[132] ^ ks[130] ^ ks[127] ^ 0x9e3779b9 ^ 127), 11);
+  ks[136] = rotl32_S ((ks[135] ^ ks[133] ^ ks[131] ^ ks[128] ^ 0x9e3779b9 ^ 128), 11);
+  ks[137] = rotl32_S ((ks[136] ^ ks[134] ^ ks[132] ^ ks[129] ^ 0x9e3779b9 ^ 129), 11);
+  ks[138] = rotl32_S ((ks[137] ^ ks[135] ^ ks[133] ^ ks[130] ^ 0x9e3779b9 ^ 130), 11);
+  ks[139] = rotl32_S ((ks[138] ^ ks[136] ^ ks[134] ^ ks[131] ^ 0x9e3779b9 ^ 131), 11);
 
   u32  a,b,c,d,e,f,g,h;
   u32  t1,t2,t3,t4,t5,t6,t7,t8,t9,t10,t11,t12,t13,t14,t15,t16;