diff --git a/OpenCL/m07500_a0-optimized.cl b/OpenCL/m07500_a0-optimized.cl
index 7ecff8847..a094df5dd 100644
--- a/OpenCL/m07500_a0-optimized.cl
+++ b/OpenCL/m07500_a0-optimized.cl
@@ -16,6 +16,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_md4.cl"
 #include "inc_hash_md5.cl"
+#include "inc_cipher_rc4.cl"
 #endif
 
 typedef struct krb5pa
@@ -28,142 +29,19 @@ typedef struct krb5pa
 
 } krb5pa_t;
 
-typedef struct
+DECLSPEC int decrypt_and_check (LOCAL_AS u32 *S, u32 *data, u32 *timestamp_ct)
 {
-  u8 S[256];
-
-  u32 wtf_its_faster;
-
-} RC4_KEY;
-
-DECLSPEC void swap (LOCAL_AS RC4_KEY *rc4_key, const u8 i, const u8 j)
-{
-  u8 tmp;
-
-  tmp           = rc4_key->S[i];
-  rc4_key->S[i] = rc4_key->S[j];
-  rc4_key->S[j] = tmp;
-}
-
-DECLSPEC void rc4_init_16 (LOCAL_AS RC4_KEY *rc4_key, const u32 *data)
-{
-  u32 v = 0x03020100;
-  u32 a = 0x04040404;
-
-  LOCAL_AS u32 *ptr = (LOCAL_AS u32 *) rc4_key->S;
-
-  #ifdef _unroll
-  #pragma unroll
-  #endif
-  for (u32 i = 0; i < 64; i++)
-  {
-    *ptr++ = v; v += a;
-  }
-
-  u32 j = 0;
-
-  for (u32 i = 0; i < 16; i++)
-  {
-    u32 idx = i * 16;
-
-    u32 v;
-
-    v = data[0];
-
-    j += rc4_key->S[idx] + (v >>  0); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >>  8); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 16); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 24); swap (rc4_key, idx, j); idx++;
-
-    v = data[1];
-
-    j += rc4_key->S[idx] + (v >>  0); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >>  8); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 16); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 24); swap (rc4_key, idx, j); idx++;
-
-    v = data[2];
-
-    j += rc4_key->S[idx] + (v >>  0); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >>  8); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 16); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 24); swap (rc4_key, idx, j); idx++;
-
-    v = data[3];
-
-    j += rc4_key->S[idx] + (v >>  0); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >>  8); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 16); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 24); swap (rc4_key, idx, j); idx++;
-  }
-}
-
-DECLSPEC u8 rc4_next_16 (LOCAL_AS RC4_KEY *rc4_key, u8 i, u8 j, const u32 *in, u32 *out)
-{
-  #ifdef _unroll
-  #pragma unroll
-  #endif
-  for (u32 k = 0; k < 4; k++)
-  {
-    u32 xor4 = 0;
-
-    u8 idx;
-
-    i += 1;
-    j += rc4_key->S[i];
-
-    swap (rc4_key, i, j);
-
-    idx = rc4_key->S[i] + rc4_key->S[j];
-
-    xor4 |= rc4_key->S[idx] <<  0;
-
-    i += 1;
-    j += rc4_key->S[i];
-
-    swap (rc4_key, i, j);
-
-    idx = rc4_key->S[i] + rc4_key->S[j];
-
-    xor4 |= rc4_key->S[idx] <<  8;
-
-    i += 1;
-    j += rc4_key->S[i];
-
-    swap (rc4_key, i, j);
-
-    idx = rc4_key->S[i] + rc4_key->S[j];
-
-    xor4 |= rc4_key->S[idx] << 16;
-
-    i += 1;
-    j += rc4_key->S[i];
-
-    swap (rc4_key, i, j);
-
-    idx = rc4_key->S[i] + rc4_key->S[j];
-
-    xor4 |= rc4_key->S[idx] << 24;
-
-    out[k] = in[k] ^ xor4;
-  }
-
-  return j;
-}
-
-DECLSPEC int decrypt_and_check (LOCAL_AS RC4_KEY *rc4_key, u32 *data, u32 *timestamp_ct)
-{
-  rc4_init_16 (rc4_key, data);
+  rc4_init_128 (S, data);
 
   u32 out[4];
 
   u8 j = 0;
 
-  j = rc4_next_16 (rc4_key,  0, j, timestamp_ct + 0, out);
+  j = rc4_next_16 (S,  0, j, timestamp_ct + 0, out);
 
   if ((out[3] & 0xffff0000) != 0x30320000) return 0;
 
-  j = rc4_next_16 (rc4_key, 16, j, timestamp_ct + 4, out);
+  j = rc4_next_16 (S, 16, j, timestamp_ct + 4, out);
 
   if (((out[0] & 0xff) < '0') || ((out[0] & 0xff) > '9')) return 0; out[0] >>= 8;
   if (((out[0] & 0xff) < '0') || ((out[0] & 0xff) > '9')) return 0; out[0] >>= 8;
@@ -451,9 +329,7 @@ KERNEL_FQ void m07500_m04 (KERN_ATTR_RULES_ESALT (krb5pa_t))
    * shared
    */
 
-  LOCAL_VK RC4_KEY rc4_keys[64];
-
-  LOCAL_AS RC4_KEY *rc4_key = &rc4_keys[lid];
+  LOCAL_VK u32 S[64 * FIXED_LOCAL_SIZE];
 
   /**
    * loop
@@ -483,7 +359,7 @@ KERNEL_FQ void m07500_m04 (KERN_ATTR_RULES_ESALT (krb5pa_t))
     tmp[2] = digest[2];
     tmp[3] = digest[3];
 
-    if (decrypt_and_check (rc4_key, tmp, timestamp_ct) == 1)
+    if (decrypt_and_check (S, tmp, timestamp_ct) == 1)
     {
       if (hc_atomic_inc (&hashes_shown[DIGESTS_OFFSET]) == 0)
       {
@@ -557,9 +433,7 @@ KERNEL_FQ void m07500_s04 (KERN_ATTR_RULES_ESALT (krb5pa_t))
    * shared
    */
 
-  LOCAL_VK RC4_KEY rc4_keys[64];
-
-  LOCAL_AS RC4_KEY *rc4_key = &rc4_keys[lid];
+  LOCAL_VK u32 S[64 * FIXED_LOCAL_SIZE];
 
   /**
    * loop
@@ -589,7 +463,7 @@ KERNEL_FQ void m07500_s04 (KERN_ATTR_RULES_ESALT (krb5pa_t))
     tmp[2] = digest[2];
     tmp[3] = digest[3];
 
-    if (decrypt_and_check (rc4_key, tmp, timestamp_ct) == 1)
+    if (decrypt_and_check (S, tmp, timestamp_ct) == 1)
     {
       if (hc_atomic_inc (&hashes_shown[DIGESTS_OFFSET]) == 0)
       {
diff --git a/OpenCL/m07500_a0-pure.cl b/OpenCL/m07500_a0-pure.cl
index 4981b23bd..325e57160 100644
--- a/OpenCL/m07500_a0-pure.cl
+++ b/OpenCL/m07500_a0-pure.cl
@@ -15,6 +15,7 @@
 #include "inc_rp.cl"
 #include "inc_hash_md4.cl"
 #include "inc_hash_md5.cl"
+#include "inc_cipher_rc4.cl"
 #endif
 
 typedef struct krb5pa
@@ -27,142 +28,19 @@ typedef struct krb5pa
 
 } krb5pa_t;
 
-typedef struct
+DECLSPEC int decrypt_and_check (LOCAL_AS u32 *S, u32 *data, u32 *timestamp_ct)
 {
-  u8 S[256];
-
-  u32 wtf_its_faster;
-
-} RC4_KEY;
-
-DECLSPEC void swap (LOCAL_AS RC4_KEY *rc4_key, const u8 i, const u8 j)
-{
-  u8 tmp;
-
-  tmp           = rc4_key->S[i];
-  rc4_key->S[i] = rc4_key->S[j];
-  rc4_key->S[j] = tmp;
-}
-
-DECLSPEC void rc4_init_16 (LOCAL_AS RC4_KEY *rc4_key, const u32 *data)
-{
-  u32 v = 0x03020100;
-  u32 a = 0x04040404;
-
-  LOCAL_AS u32 *ptr = (LOCAL_AS u32 *) rc4_key->S;
-
-  #ifdef _unroll
-  #pragma unroll
-  #endif
-  for (u32 i = 0; i < 64; i++)
-  {
-    *ptr++ = v; v += a;
-  }
-
-  u32 j = 0;
-
-  for (u32 i = 0; i < 16; i++)
-  {
-    u32 idx = i * 16;
-
-    u32 v;
-
-    v = data[0];
-
-    j += rc4_key->S[idx] + (v >>  0); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >>  8); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 16); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 24); swap (rc4_key, idx, j); idx++;
-
-    v = data[1];
-
-    j += rc4_key->S[idx] + (v >>  0); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >>  8); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 16); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 24); swap (rc4_key, idx, j); idx++;
-
-    v = data[2];
-
-    j += rc4_key->S[idx] + (v >>  0); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >>  8); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 16); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 24); swap (rc4_key, idx, j); idx++;
-
-    v = data[3];
-
-    j += rc4_key->S[idx] + (v >>  0); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >>  8); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 16); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 24); swap (rc4_key, idx, j); idx++;
-  }
-}
-
-DECLSPEC u8 rc4_next_16 (LOCAL_AS RC4_KEY *rc4_key, u8 i, u8 j, const u32 *in, u32 *out)
-{
-  #ifdef _unroll
-  #pragma unroll
-  #endif
-  for (u32 k = 0; k < 4; k++)
-  {
-    u32 xor4 = 0;
-
-    u8 idx;
-
-    i += 1;
-    j += rc4_key->S[i];
-
-    swap (rc4_key, i, j);
-
-    idx = rc4_key->S[i] + rc4_key->S[j];
-
-    xor4 |= rc4_key->S[idx] <<  0;
-
-    i += 1;
-    j += rc4_key->S[i];
-
-    swap (rc4_key, i, j);
-
-    idx = rc4_key->S[i] + rc4_key->S[j];
-
-    xor4 |= rc4_key->S[idx] <<  8;
-
-    i += 1;
-    j += rc4_key->S[i];
-
-    swap (rc4_key, i, j);
-
-    idx = rc4_key->S[i] + rc4_key->S[j];
-
-    xor4 |= rc4_key->S[idx] << 16;
-
-    i += 1;
-    j += rc4_key->S[i];
-
-    swap (rc4_key, i, j);
-
-    idx = rc4_key->S[i] + rc4_key->S[j];
-
-    xor4 |= rc4_key->S[idx] << 24;
-
-    out[k] = in[k] ^ xor4;
-  }
-
-  return j;
-}
-
-DECLSPEC int decrypt_and_check (LOCAL_AS RC4_KEY *rc4_key, u32 *data, u32 *timestamp_ct)
-{
-  rc4_init_16 (rc4_key, data);
+  rc4_init_128 (S, data);
 
   u32 out[4];
 
   u8 j = 0;
 
-  j = rc4_next_16 (rc4_key,  0, j, timestamp_ct + 0, out);
+  j = rc4_next_16 (S,  0, j, timestamp_ct + 0, out);
 
   if ((out[3] & 0xffff0000) != 0x30320000) return 0;
 
-  j = rc4_next_16 (rc4_key, 16, j, timestamp_ct + 4, out);
+  j = rc4_next_16 (S, 16, j, timestamp_ct + 4, out);
 
   if (((out[0] & 0xff) < '0') || ((out[0] & 0xff) > '9')) return 0; out[0] >>= 8;
   if (((out[0] & 0xff) < '0') || ((out[0] & 0xff) > '9')) return 0; out[0] >>= 8;
@@ -296,9 +174,7 @@ KERNEL_FQ void m07500_mxx (KERN_ATTR_RULES_ESALT (krb5pa_t))
 
   COPY_PW (pws[gid]);
 
-  LOCAL_VK RC4_KEY rc4_keys[64];
-
-  LOCAL_AS RC4_KEY *rc4_key = &rc4_keys[lid];
+  LOCAL_VK u32 S[64 * FIXED_LOCAL_SIZE];
 
   u32 checksum[4];
 
@@ -340,7 +216,7 @@ KERNEL_FQ void m07500_mxx (KERN_ATTR_RULES_ESALT (krb5pa_t))
 
     kerb_prepare (ctx.h, checksum, digest);
 
-    if (decrypt_and_check (rc4_key, digest, timestamp_ct) == 1)
+    if (decrypt_and_check (S, digest, timestamp_ct) == 1)
     {
       if (hc_atomic_inc (&hashes_shown[DIGESTS_OFFSET]) == 0)
       {
@@ -367,9 +243,7 @@ KERNEL_FQ void m07500_sxx (KERN_ATTR_RULES_ESALT (krb5pa_t))
 
   COPY_PW (pws[gid]);
 
-  LOCAL_VK RC4_KEY rc4_keys[64];
-
-  LOCAL_AS RC4_KEY *rc4_key = &rc4_keys[lid];
+  LOCAL_VK u32 S[64 * FIXED_LOCAL_SIZE];
 
   u32 checksum[4];
 
@@ -411,7 +285,7 @@ KERNEL_FQ void m07500_sxx (KERN_ATTR_RULES_ESALT (krb5pa_t))
 
     kerb_prepare (ctx.h, checksum, digest);
 
-    if (decrypt_and_check (rc4_key, digest, timestamp_ct) == 1)
+    if (decrypt_and_check (S, digest, timestamp_ct) == 1)
     {
       if (hc_atomic_inc (&hashes_shown[DIGESTS_OFFSET]) == 0)
       {
diff --git a/OpenCL/m07500_a1-optimized.cl b/OpenCL/m07500_a1-optimized.cl
index 7ba0b483a..2ef9c573f 100644
--- a/OpenCL/m07500_a1-optimized.cl
+++ b/OpenCL/m07500_a1-optimized.cl
@@ -14,6 +14,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_md4.cl"
 #include "inc_hash_md5.cl"
+#include "inc_cipher_rc4.cl"
 #endif
 
 typedef struct krb5pa
@@ -26,142 +27,19 @@ typedef struct krb5pa
 
 } krb5pa_t;
 
-typedef struct
+DECLSPEC int decrypt_and_check (LOCAL_AS u32 *S, u32 *data, u32 *timestamp_ct)
 {
-  u8 S[256];
-
-  u32 wtf_its_faster;
-
-} RC4_KEY;
-
-DECLSPEC void swap (LOCAL_AS RC4_KEY *rc4_key, const u8 i, const u8 j)
-{
-  u8 tmp;
-
-  tmp           = rc4_key->S[i];
-  rc4_key->S[i] = rc4_key->S[j];
-  rc4_key->S[j] = tmp;
-}
-
-DECLSPEC void rc4_init_16 (LOCAL_AS RC4_KEY *rc4_key, const u32 *data)
-{
-  u32 v = 0x03020100;
-  u32 a = 0x04040404;
-
-  LOCAL_AS u32 *ptr = (LOCAL_AS u32 *) rc4_key->S;
-
-  #ifdef _unroll
-  #pragma unroll
-  #endif
-  for (u32 i = 0; i < 64; i++)
-  {
-    *ptr++ = v; v += a;
-  }
-
-  u32 j = 0;
-
-  for (u32 i = 0; i < 16; i++)
-  {
-    u32 idx = i * 16;
-
-    u32 v;
-
-    v = data[0];
-
-    j += rc4_key->S[idx] + (v >>  0); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >>  8); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 16); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 24); swap (rc4_key, idx, j); idx++;
-
-    v = data[1];
-
-    j += rc4_key->S[idx] + (v >>  0); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >>  8); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 16); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 24); swap (rc4_key, idx, j); idx++;
-
-    v = data[2];
-
-    j += rc4_key->S[idx] + (v >>  0); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >>  8); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 16); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 24); swap (rc4_key, idx, j); idx++;
-
-    v = data[3];
-
-    j += rc4_key->S[idx] + (v >>  0); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >>  8); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 16); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 24); swap (rc4_key, idx, j); idx++;
-  }
-}
-
-DECLSPEC u8 rc4_next_16 (LOCAL_AS RC4_KEY *rc4_key, u8 i, u8 j, const u32 *in, u32 *out)
-{
-  #ifdef _unroll
-  #pragma unroll
-  #endif
-  for (u32 k = 0; k < 4; k++)
-  {
-    u32 xor4 = 0;
-
-    u8 idx;
-
-    i += 1;
-    j += rc4_key->S[i];
-
-    swap (rc4_key, i, j);
-
-    idx = rc4_key->S[i] + rc4_key->S[j];
-
-    xor4 |= rc4_key->S[idx] <<  0;
-
-    i += 1;
-    j += rc4_key->S[i];
-
-    swap (rc4_key, i, j);
-
-    idx = rc4_key->S[i] + rc4_key->S[j];
-
-    xor4 |= rc4_key->S[idx] <<  8;
-
-    i += 1;
-    j += rc4_key->S[i];
-
-    swap (rc4_key, i, j);
-
-    idx = rc4_key->S[i] + rc4_key->S[j];
-
-    xor4 |= rc4_key->S[idx] << 16;
-
-    i += 1;
-    j += rc4_key->S[i];
-
-    swap (rc4_key, i, j);
-
-    idx = rc4_key->S[i] + rc4_key->S[j];
-
-    xor4 |= rc4_key->S[idx] << 24;
-
-    out[k] = in[k] ^ xor4;
-  }
-
-  return j;
-}
-
-DECLSPEC int decrypt_and_check (LOCAL_AS RC4_KEY *rc4_key, u32 *data, u32 *timestamp_ct)
-{
-  rc4_init_16 (rc4_key, data);
+  rc4_init_128 (S, data);
 
   u32 out[4];
 
   u8 j = 0;
 
-  j = rc4_next_16 (rc4_key,  0, j, timestamp_ct + 0, out);
+  j = rc4_next_16 (S,  0, j, timestamp_ct + 0, out);
 
   if ((out[3] & 0xffff0000) != 0x30320000) return 0;
 
-  j = rc4_next_16 (rc4_key, 16, j, timestamp_ct + 4, out);
+  j = rc4_next_16 (S, 16, j, timestamp_ct + 4, out);
 
   if (((out[0] & 0xff) < '0') || ((out[0] & 0xff) > '9')) return 0; out[0] >>= 8;
   if (((out[0] & 0xff) < '0') || ((out[0] & 0xff) > '9')) return 0; out[0] >>= 8;
@@ -449,9 +327,7 @@ KERNEL_FQ void m07500_m04 (KERN_ATTR_ESALT (krb5pa_t))
    * shared
    */
 
-  LOCAL_VK RC4_KEY rc4_keys[64];
-
-  LOCAL_AS RC4_KEY *rc4_key = &rc4_keys[lid];
+  LOCAL_VK u32 S[64 * FIXED_LOCAL_SIZE];
 
   /**
    * loop
@@ -531,7 +407,7 @@ KERNEL_FQ void m07500_m04 (KERN_ATTR_ESALT (krb5pa_t))
     tmp[2] = digest[2];
     tmp[3] = digest[3];
 
-    if (decrypt_and_check (rc4_key, tmp, timestamp_ct) == 1)
+    if (decrypt_and_check (S, tmp, timestamp_ct) == 1)
     {
       if (hc_atomic_inc (&hashes_shown[DIGESTS_OFFSET]) == 0)
       {
@@ -605,9 +481,7 @@ KERNEL_FQ void m07500_s04 (KERN_ATTR_ESALT (krb5pa_t))
    * shared
    */
 
-  LOCAL_VK RC4_KEY rc4_keys[64];
-
-  LOCAL_AS RC4_KEY *rc4_key = &rc4_keys[lid];
+  LOCAL_VK u32 S[64 * FIXED_LOCAL_SIZE];
 
   /**
    * loop
@@ -687,7 +561,7 @@ KERNEL_FQ void m07500_s04 (KERN_ATTR_ESALT (krb5pa_t))
     tmp[2] = digest[2];
     tmp[3] = digest[3];
 
-    if (decrypt_and_check (rc4_key, tmp, timestamp_ct) == 1)
+    if (decrypt_and_check (S, tmp, timestamp_ct) == 1)
     {
       if (hc_atomic_inc (&hashes_shown[DIGESTS_OFFSET]) == 0)
       {
diff --git a/OpenCL/m07500_a1-pure.cl b/OpenCL/m07500_a1-pure.cl
index 1d1873f81..bc156d284 100644
--- a/OpenCL/m07500_a1-pure.cl
+++ b/OpenCL/m07500_a1-pure.cl
@@ -13,6 +13,7 @@
 #include "inc_common.cl"
 #include "inc_hash_md4.cl"
 #include "inc_hash_md5.cl"
+#include "inc_cipher_rc4.cl"
 #endif
 
 typedef struct krb5pa
@@ -25,142 +26,19 @@ typedef struct krb5pa
 
 } krb5pa_t;
 
-typedef struct
+DECLSPEC int decrypt_and_check (LOCAL_AS u32 *S, u32 *data, u32 *timestamp_ct)
 {
-  u8 S[256];
-
-  u32 wtf_its_faster;
-
-} RC4_KEY;
-
-DECLSPEC void swap (LOCAL_AS RC4_KEY *rc4_key, const u8 i, const u8 j)
-{
-  u8 tmp;
-
-  tmp           = rc4_key->S[i];
-  rc4_key->S[i] = rc4_key->S[j];
-  rc4_key->S[j] = tmp;
-}
-
-DECLSPEC void rc4_init_16 (LOCAL_AS RC4_KEY *rc4_key, const u32 *data)
-{
-  u32 v = 0x03020100;
-  u32 a = 0x04040404;
-
-  LOCAL_AS u32 *ptr = (LOCAL_AS u32 *) rc4_key->S;
-
-  #ifdef _unroll
-  #pragma unroll
-  #endif
-  for (u32 i = 0; i < 64; i++)
-  {
-    *ptr++ = v; v += a;
-  }
-
-  u32 j = 0;
-
-  for (u32 i = 0; i < 16; i++)
-  {
-    u32 idx = i * 16;
-
-    u32 v;
-
-    v = data[0];
-
-    j += rc4_key->S[idx] + (v >>  0); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >>  8); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 16); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 24); swap (rc4_key, idx, j); idx++;
-
-    v = data[1];
-
-    j += rc4_key->S[idx] + (v >>  0); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >>  8); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 16); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 24); swap (rc4_key, idx, j); idx++;
-
-    v = data[2];
-
-    j += rc4_key->S[idx] + (v >>  0); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >>  8); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 16); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 24); swap (rc4_key, idx, j); idx++;
-
-    v = data[3];
-
-    j += rc4_key->S[idx] + (v >>  0); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >>  8); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 16); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 24); swap (rc4_key, idx, j); idx++;
-  }
-}
-
-DECLSPEC u8 rc4_next_16 (LOCAL_AS RC4_KEY *rc4_key, u8 i, u8 j, const u32 *in, u32 *out)
-{
-  #ifdef _unroll
-  #pragma unroll
-  #endif
-  for (u32 k = 0; k < 4; k++)
-  {
-    u32 xor4 = 0;
-
-    u8 idx;
-
-    i += 1;
-    j += rc4_key->S[i];
-
-    swap (rc4_key, i, j);
-
-    idx = rc4_key->S[i] + rc4_key->S[j];
-
-    xor4 |= rc4_key->S[idx] <<  0;
-
-    i += 1;
-    j += rc4_key->S[i];
-
-    swap (rc4_key, i, j);
-
-    idx = rc4_key->S[i] + rc4_key->S[j];
-
-    xor4 |= rc4_key->S[idx] <<  8;
-
-    i += 1;
-    j += rc4_key->S[i];
-
-    swap (rc4_key, i, j);
-
-    idx = rc4_key->S[i] + rc4_key->S[j];
-
-    xor4 |= rc4_key->S[idx] << 16;
-
-    i += 1;
-    j += rc4_key->S[i];
-
-    swap (rc4_key, i, j);
-
-    idx = rc4_key->S[i] + rc4_key->S[j];
-
-    xor4 |= rc4_key->S[idx] << 24;
-
-    out[k] = in[k] ^ xor4;
-  }
-
-  return j;
-}
-
-DECLSPEC int decrypt_and_check (LOCAL_AS RC4_KEY *rc4_key, u32 *data, u32 *timestamp_ct)
-{
-  rc4_init_16 (rc4_key, data);
+  rc4_init_128 (S, data);
 
   u32 out[4];
 
   u8 j = 0;
 
-  j = rc4_next_16 (rc4_key,  0, j, timestamp_ct + 0, out);
+  j = rc4_next_16 (S,  0, j, timestamp_ct + 0, out);
 
   if ((out[3] & 0xffff0000) != 0x30320000) return 0;
 
-  j = rc4_next_16 (rc4_key, 16, j, timestamp_ct + 4, out);
+  j = rc4_next_16 (S, 16, j, timestamp_ct + 4, out);
 
   if (((out[0] & 0xff) < '0') || ((out[0] & 0xff) > '9')) return 0; out[0] >>= 8;
   if (((out[0] & 0xff) < '0') || ((out[0] & 0xff) > '9')) return 0; out[0] >>= 8;
@@ -292,9 +170,7 @@ KERNEL_FQ void m07500_mxx (KERN_ATTR_ESALT (krb5pa_t))
    * base
    */
 
-  LOCAL_VK RC4_KEY rc4_keys[64];
-
-  LOCAL_AS RC4_KEY *rc4_key = &rc4_keys[lid];
+  LOCAL_VK u32 S[64 * FIXED_LOCAL_SIZE];
 
   u32 checksum[4];
 
@@ -336,7 +212,7 @@ KERNEL_FQ void m07500_mxx (KERN_ATTR_ESALT (krb5pa_t))
 
     kerb_prepare (ctx.h, checksum, digest);
 
-    if (decrypt_and_check (rc4_key, digest, timestamp_ct) == 1)
+    if (decrypt_and_check (S, digest, timestamp_ct) == 1)
     {
       if (hc_atomic_inc (&hashes_shown[DIGESTS_OFFSET]) == 0)
       {
@@ -361,9 +237,7 @@ KERNEL_FQ void m07500_sxx (KERN_ATTR_ESALT (krb5pa_t))
    * base
    */
 
-  LOCAL_VK RC4_KEY rc4_keys[64];
-
-  LOCAL_AS RC4_KEY *rc4_key = &rc4_keys[lid];
+  LOCAL_VK u32 S[64 * FIXED_LOCAL_SIZE];
 
   u32 checksum[4];
 
@@ -405,7 +279,7 @@ KERNEL_FQ void m07500_sxx (KERN_ATTR_ESALT (krb5pa_t))
 
     kerb_prepare (ctx.h, checksum, digest);
 
-    if (decrypt_and_check (rc4_key, digest, timestamp_ct) == 1)
+    if (decrypt_and_check (S, digest, timestamp_ct) == 1)
     {
       if (hc_atomic_inc (&hashes_shown[DIGESTS_OFFSET]) == 0)
       {
diff --git a/OpenCL/m07500_a3-optimized.cl b/OpenCL/m07500_a3-optimized.cl
index 7a1739825..c599eb168 100644
--- a/OpenCL/m07500_a3-optimized.cl
+++ b/OpenCL/m07500_a3-optimized.cl
@@ -14,6 +14,7 @@
 #include "inc_simd.cl"
 #include "inc_hash_md4.cl"
 #include "inc_hash_md5.cl"
+#include "inc_cipher_rc4.cl"
 #endif
 
 typedef struct krb5pa
@@ -26,142 +27,19 @@ typedef struct krb5pa
 
 } krb5pa_t;
 
-typedef struct
+DECLSPEC int decrypt_and_check (LOCAL_AS u32 *S, u32 *data, u32 *timestamp_ct)
 {
-  u8 S[256];
-
-  u32 wtf_its_faster;
-
-} RC4_KEY;
-
-DECLSPEC void swap (LOCAL_AS RC4_KEY *rc4_key, const u8 i, const u8 j)
-{
-  u8 tmp;
-
-  tmp           = rc4_key->S[i];
-  rc4_key->S[i] = rc4_key->S[j];
-  rc4_key->S[j] = tmp;
-}
-
-DECLSPEC void rc4_init_16 (LOCAL_AS RC4_KEY *rc4_key, const u32 *data)
-{
-  u32 v = 0x03020100;
-  u32 a = 0x04040404;
-
-  LOCAL_AS u32 *ptr = (LOCAL_AS u32 *) rc4_key->S;
-
-  #ifdef _unroll
-  #pragma unroll
-  #endif
-  for (u32 i = 0; i < 64; i++)
-  {
-    *ptr++ = v; v += a;
-  }
-
-  u32 j = 0;
-
-  for (u32 i = 0; i < 16; i++)
-  {
-    u32 idx = i * 16;
-
-    u32 v;
-
-    v = data[0];
-
-    j += rc4_key->S[idx] + (v >>  0); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >>  8); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 16); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 24); swap (rc4_key, idx, j); idx++;
-
-    v = data[1];
-
-    j += rc4_key->S[idx] + (v >>  0); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >>  8); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 16); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 24); swap (rc4_key, idx, j); idx++;
-
-    v = data[2];
-
-    j += rc4_key->S[idx] + (v >>  0); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >>  8); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 16); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 24); swap (rc4_key, idx, j); idx++;
-
-    v = data[3];
-
-    j += rc4_key->S[idx] + (v >>  0); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >>  8); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 16); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 24); swap (rc4_key, idx, j); idx++;
-  }
-}
-
-DECLSPEC u8 rc4_next_16 (LOCAL_AS RC4_KEY *rc4_key, u8 i, u8 j, const u32 *in, u32 *out)
-{
-  #ifdef _unroll
-  #pragma unroll
-  #endif
-  for (u32 k = 0; k < 4; k++)
-  {
-    u32 xor4 = 0;
-
-    u8 idx;
-
-    i += 1;
-    j += rc4_key->S[i];
-
-    swap (rc4_key, i, j);
-
-    idx = rc4_key->S[i] + rc4_key->S[j];
-
-    xor4 |= rc4_key->S[idx] <<  0;
-
-    i += 1;
-    j += rc4_key->S[i];
-
-    swap (rc4_key, i, j);
-
-    idx = rc4_key->S[i] + rc4_key->S[j];
-
-    xor4 |= rc4_key->S[idx] <<  8;
-
-    i += 1;
-    j += rc4_key->S[i];
-
-    swap (rc4_key, i, j);
-
-    idx = rc4_key->S[i] + rc4_key->S[j];
-
-    xor4 |= rc4_key->S[idx] << 16;
-
-    i += 1;
-    j += rc4_key->S[i];
-
-    swap (rc4_key, i, j);
-
-    idx = rc4_key->S[i] + rc4_key->S[j];
-
-    xor4 |= rc4_key->S[idx] << 24;
-
-    out[k] = in[k] ^ xor4;
-  }
-
-  return j;
-}
-
-DECLSPEC int decrypt_and_check (LOCAL_AS RC4_KEY *rc4_key, u32 *data, u32 *timestamp_ct)
-{
-  rc4_init_16 (rc4_key, data);
+  rc4_init_128 (S, data);
 
   u32 out[4];
 
   u8 j = 0;
 
-  j = rc4_next_16 (rc4_key,  0, j, timestamp_ct + 0, out);
+  j = rc4_next_16 (S,  0, j, timestamp_ct + 0, out);
 
   if ((out[3] & 0xffff0000) != 0x30320000) return 0;
 
-  j = rc4_next_16 (rc4_key, 16, j, timestamp_ct + 4, out);
+  j = rc4_next_16 (S, 16, j, timestamp_ct + 4, out);
 
   if (((out[0] & 0xff) < '0') || ((out[0] & 0xff) > '9')) return 0; out[0] >>= 8;
   if (((out[0] & 0xff) < '0') || ((out[0] & 0xff) > '9')) return 0; out[0] >>= 8;
@@ -393,7 +271,7 @@ DECLSPEC void kerb_prepare (const u32 *w0, const u32 *w1, const u32 pw_len, cons
   hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
 }
 
-DECLSPEC void m07500 (LOCAL_AS RC4_KEY *rc4_key, u32 *w0, u32 *w1, u32 *w2, u32 *w3, const u32 pw_len, KERN_ATTR_ESALT (krb5pa_t))
+DECLSPEC void m07500 (LOCAL_AS u32 *S, u32 *w0, u32 *w1, u32 *w2, u32 *w3, const u32 pw_len, KERN_ATTR_ESALT (krb5pa_t))
 {
   /**
    * modifier
@@ -473,7 +351,7 @@ DECLSPEC void m07500 (LOCAL_AS RC4_KEY *rc4_key, u32 *w0, u32 *w1, u32 *w2, u32
     tmp[2] = digest[2];
     tmp[3] = digest[3];
 
-    if (decrypt_and_check (rc4_key, tmp, timestamp_ct) == 1)
+    if (decrypt_and_check (S, tmp, timestamp_ct) == 1)
     {
       if (hc_atomic_inc (&hashes_shown[DIGESTS_OFFSET]) == 0)
       {
@@ -528,11 +406,9 @@ KERNEL_FQ void m07500_m04 (KERN_ATTR_ESALT (krb5pa_t))
    * main
    */
 
-  LOCAL_VK RC4_KEY rc4_keys[64];
-
-  LOCAL_AS RC4_KEY *rc4_key = &rc4_keys[lid];
+  LOCAL_VK u32 S[64 * FIXED_LOCAL_SIZE];
 
-  m07500 (rc4_key, w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_extra0_buf, d_extra1_buf, d_extra2_buf, d_extra3_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, SALT_POS, loop_pos, loop_cnt, il_cnt, digests_cnt, DIGESTS_OFFSET, combs_mode, salt_repeat, pws_pos, gid_max);
+  m07500 (S, w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_extra0_buf, d_extra1_buf, d_extra2_buf, d_extra3_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, SALT_POS, loop_pos, loop_cnt, il_cnt, digests_cnt, DIGESTS_OFFSET, combs_mode, salt_repeat, pws_pos, gid_max);
 }
 
 KERNEL_FQ void m07500_m08 (KERN_ATTR_ESALT (krb5pa_t))
@@ -580,11 +456,9 @@ KERNEL_FQ void m07500_m08 (KERN_ATTR_ESALT (krb5pa_t))
    * main
    */
 
-  LOCAL_VK RC4_KEY rc4_keys[64];
+  LOCAL_VK u32 S[64 * FIXED_LOCAL_SIZE];
 
-  LOCAL_AS RC4_KEY *rc4_key = &rc4_keys[lid];
-
-  m07500 (rc4_key, w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_extra0_buf, d_extra1_buf, d_extra2_buf, d_extra3_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, SALT_POS, loop_pos, loop_cnt, il_cnt, digests_cnt, DIGESTS_OFFSET, combs_mode, salt_repeat, pws_pos, gid_max);
+  m07500 (S, w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_extra0_buf, d_extra1_buf, d_extra2_buf, d_extra3_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, SALT_POS, loop_pos, loop_cnt, il_cnt, digests_cnt, DIGESTS_OFFSET, combs_mode, salt_repeat, pws_pos, gid_max);
 }
 
 KERNEL_FQ void m07500_m16 (KERN_ATTR_ESALT (krb5pa_t))
@@ -636,11 +510,9 @@ KERNEL_FQ void m07500_s04 (KERN_ATTR_ESALT (krb5pa_t))
    * main
    */
 
-  LOCAL_VK RC4_KEY rc4_keys[64];
-
-  LOCAL_AS RC4_KEY *rc4_key = &rc4_keys[lid];
+  LOCAL_VK u32 S[64 * FIXED_LOCAL_SIZE];
 
-  m07500 (rc4_key, w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_extra0_buf, d_extra1_buf, d_extra2_buf, d_extra3_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, SALT_POS, loop_pos, loop_cnt, il_cnt, digests_cnt, DIGESTS_OFFSET, combs_mode, salt_repeat, pws_pos, gid_max);
+  m07500 (S, w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_extra0_buf, d_extra1_buf, d_extra2_buf, d_extra3_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, SALT_POS, loop_pos, loop_cnt, il_cnt, digests_cnt, DIGESTS_OFFSET, combs_mode, salt_repeat, pws_pos, gid_max);
 }
 
 KERNEL_FQ void m07500_s08 (KERN_ATTR_ESALT (krb5pa_t))
@@ -688,11 +560,9 @@ KERNEL_FQ void m07500_s08 (KERN_ATTR_ESALT (krb5pa_t))
    * main
    */
 
-  LOCAL_VK RC4_KEY rc4_keys[64];
-
-  LOCAL_AS RC4_KEY *rc4_key = &rc4_keys[lid];
+  LOCAL_VK u32 S[64 * FIXED_LOCAL_SIZE];
 
-  m07500 (rc4_key, w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_extra0_buf, d_extra1_buf, d_extra2_buf, d_extra3_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, SALT_POS, loop_pos, loop_cnt, il_cnt, digests_cnt, DIGESTS_OFFSET, combs_mode, salt_repeat, pws_pos, gid_max);
+  m07500 (S, w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_extra0_buf, d_extra1_buf, d_extra2_buf, d_extra3_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, SALT_POS, loop_pos, loop_cnt, il_cnt, digests_cnt, DIGESTS_OFFSET, combs_mode, salt_repeat, pws_pos, gid_max);
 }
 
 KERNEL_FQ void m07500_s16 (KERN_ATTR_ESALT (krb5pa_t))
diff --git a/OpenCL/m07500_a3-pure.cl b/OpenCL/m07500_a3-pure.cl
index 602678806..cd8d18cda 100644
--- a/OpenCL/m07500_a3-pure.cl
+++ b/OpenCL/m07500_a3-pure.cl
@@ -13,6 +13,7 @@
 #include "inc_common.cl"
 #include "inc_hash_md4.cl"
 #include "inc_hash_md5.cl"
+#include "inc_cipher_rc4.cl"
 #endif
 
 typedef struct krb5pa
@@ -25,142 +26,19 @@ typedef struct krb5pa
 
 } krb5pa_t;
 
-typedef struct
+DECLSPEC int decrypt_and_check (LOCAL_AS u32 *S, u32 *data, u32 *timestamp_ct)
 {
-  u8 S[256];
-
-  u32 wtf_its_faster;
-
-} RC4_KEY;
-
-DECLSPEC void swap (LOCAL_AS RC4_KEY *rc4_key, const u8 i, const u8 j)
-{
-  u8 tmp;
-
-  tmp           = rc4_key->S[i];
-  rc4_key->S[i] = rc4_key->S[j];
-  rc4_key->S[j] = tmp;
-}
-
-DECLSPEC void rc4_init_16 (LOCAL_AS RC4_KEY *rc4_key, const u32 *data)
-{
-  u32 v = 0x03020100;
-  u32 a = 0x04040404;
-
-  LOCAL_AS u32 *ptr = (LOCAL_AS u32 *) rc4_key->S;
-
-  #ifdef _unroll
-  #pragma unroll
-  #endif
-  for (u32 i = 0; i < 64; i++)
-  {
-    *ptr++ = v; v += a;
-  }
-
-  u32 j = 0;
-
-  for (u32 i = 0; i < 16; i++)
-  {
-    u32 idx = i * 16;
-
-    u32 v;
-
-    v = data[0];
-
-    j += rc4_key->S[idx] + (v >>  0); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >>  8); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 16); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 24); swap (rc4_key, idx, j); idx++;
-
-    v = data[1];
-
-    j += rc4_key->S[idx] + (v >>  0); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >>  8); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 16); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 24); swap (rc4_key, idx, j); idx++;
-
-    v = data[2];
-
-    j += rc4_key->S[idx] + (v >>  0); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >>  8); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 16); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 24); swap (rc4_key, idx, j); idx++;
-
-    v = data[3];
-
-    j += rc4_key->S[idx] + (v >>  0); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >>  8); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 16); swap (rc4_key, idx, j); idx++;
-    j += rc4_key->S[idx] + (v >> 24); swap (rc4_key, idx, j); idx++;
-  }
-}
-
-DECLSPEC u8 rc4_next_16 (LOCAL_AS RC4_KEY *rc4_key, u8 i, u8 j, const u32 *in, u32 *out)
-{
-  #ifdef _unroll
-  #pragma unroll
-  #endif
-  for (u32 k = 0; k < 4; k++)
-  {
-    u32 xor4 = 0;
-
-    u8 idx;
-
-    i += 1;
-    j += rc4_key->S[i];
-
-    swap (rc4_key, i, j);
-
-    idx = rc4_key->S[i] + rc4_key->S[j];
-
-    xor4 |= rc4_key->S[idx] <<  0;
-
-    i += 1;
-    j += rc4_key->S[i];
-
-    swap (rc4_key, i, j);
-
-    idx = rc4_key->S[i] + rc4_key->S[j];
-
-    xor4 |= rc4_key->S[idx] <<  8;
-
-    i += 1;
-    j += rc4_key->S[i];
-
-    swap (rc4_key, i, j);
-
-    idx = rc4_key->S[i] + rc4_key->S[j];
-
-    xor4 |= rc4_key->S[idx] << 16;
-
-    i += 1;
-    j += rc4_key->S[i];
-
-    swap (rc4_key, i, j);
-
-    idx = rc4_key->S[i] + rc4_key->S[j];
-
-    xor4 |= rc4_key->S[idx] << 24;
-
-    out[k] = in[k] ^ xor4;
-  }
-
-  return j;
-}
-
-DECLSPEC int decrypt_and_check (LOCAL_AS RC4_KEY *rc4_key, u32 *data, u32 *timestamp_ct)
-{
-  rc4_init_16 (rc4_key, data);
+  rc4_init_128 (S, data);
 
   u32 out[4];
 
   u8 j = 0;
 
-  j = rc4_next_16 (rc4_key,  0, j, timestamp_ct + 0, out);
+  j = rc4_next_16 (S,  0, j, timestamp_ct + 0, out);
 
   if ((out[3] & 0xffff0000) != 0x30320000) return 0;
 
-  j = rc4_next_16 (rc4_key, 16, j, timestamp_ct + 4, out);
+  j = rc4_next_16 (S, 16, j, timestamp_ct + 4, out);
 
   if (((out[0] & 0xff) < '0') || ((out[0] & 0xff) > '9')) return 0; out[0] >>= 8;
   if (((out[0] & 0xff) < '0') || ((out[0] & 0xff) > '9')) return 0; out[0] >>= 8;
@@ -313,9 +191,7 @@ KERNEL_FQ void m07500_mxx (KERN_ATTR_VECTOR_ESALT (krb5pa_t))
     w[idx] = pws[gid].i[idx];
   }
 
-  LOCAL_VK RC4_KEY rc4_keys[64];
-
-  LOCAL_AS RC4_KEY *rc4_key = &rc4_keys[lid];
+  LOCAL_VK u32 S[64 * FIXED_LOCAL_SIZE];
 
   u32 checksum[4];
 
@@ -361,7 +237,7 @@ KERNEL_FQ void m07500_mxx (KERN_ATTR_VECTOR_ESALT (krb5pa_t))
 
     kerb_prepare (ctx.h, checksum, digest);
 
-    if (decrypt_and_check (rc4_key, digest, timestamp_ct) == 1)
+    if (decrypt_and_check (S, digest, timestamp_ct) == 1)
     {
       if (hc_atomic_inc (&hashes_shown[DIGESTS_OFFSET]) == 0)
       {
@@ -407,9 +283,7 @@ KERNEL_FQ void m07500_sxx (KERN_ATTR_VECTOR_ESALT (krb5pa_t))
     w[idx] = pws[gid].i[idx];
   }
 
-  LOCAL_VK RC4_KEY rc4_keys[64];
-
-  LOCAL_AS RC4_KEY *rc4_key = &rc4_keys[lid];
+  LOCAL_VK u32 S[64 * FIXED_LOCAL_SIZE];
 
   u32 checksum[4];
 
@@ -455,7 +329,7 @@ KERNEL_FQ void m07500_sxx (KERN_ATTR_VECTOR_ESALT (krb5pa_t))
 
     kerb_prepare (ctx.h, checksum, digest);
 
-    if (decrypt_and_check (rc4_key, digest, timestamp_ct) == 1)
+    if (decrypt_and_check (S, digest, timestamp_ct) == 1)
     {
       if (hc_atomic_inc (&hashes_shown[DIGESTS_OFFSET]) == 0)
       {
diff --git a/docs/changes.txt b/docs/changes.txt
index ccba2a476..f07ecf45f 100644
--- a/docs/changes.txt
+++ b/docs/changes.txt
@@ -30,7 +30,7 @@
 - AES Crypt Plugin: Reduced max password length from 256 to 128 which improved performance by 22%
 - RAR3-p (Compressed): Fix workaround in unrar library in AES constant table generation to enable multi-threading support
 - CRC32: Prevent decompression of data used in CRC32 calculation on host. This leads to false negatives with TrueCrypt/VeraCrypt keyfiles
-- RC4: Update -m 97x0 and -m 104x0 to new RC4 crypto library code, improving performance by 20% or more
+- RC4: Updated hash-mode 7500, 9710, 9720, 10400 and 10410 to new RC4 crypto library code, improving performance by 20% or more
 
 ##
 ## Technical
diff --git a/src/modules/module_07500.c b/src/modules/module_07500.c
index c7c85e1bd..931cc6b47 100644
--- a/src/modules/module_07500.c
+++ b/src/modules/module_07500.c
@@ -53,48 +53,51 @@ typedef struct krb5pa
 
 static const char *SIGNATURE_KRB5PA = "$krb5pa$23$";
 
-u64 module_esalt_size (MAYBE_UNUSED const hashconfig_t *hashconfig, MAYBE_UNUSED const user_options_t *user_options, MAYBE_UNUSED const user_options_extra_t *user_options_extra)
-{
-  const u64 esalt_size = (const u64) sizeof (krb5pa_t);
-
-  return esalt_size;
-}
-
-u32 module_kernel_threads_min (MAYBE_UNUSED const hashconfig_t *hashconfig, MAYBE_UNUSED const user_options_t *user_options, MAYBE_UNUSED const user_options_extra_t *user_options_extra)
-{
-  const u32 kernel_threads_min = 64;
-
-  return kernel_threads_min;
-}
-
-u32 module_kernel_threads_max (MAYBE_UNUSED const hashconfig_t *hashconfig, MAYBE_UNUSED const user_options_t *user_options, MAYBE_UNUSED const user_options_extra_t *user_options_extra)
-{
-  const u32 kernel_threads_max = 64;
-
-  return kernel_threads_max;
-}
-
 char *module_jit_build_options (MAYBE_UNUSED const hashconfig_t *hashconfig, MAYBE_UNUSED const user_options_t *user_options, MAYBE_UNUSED const user_options_extra_t *user_options_extra, MAYBE_UNUSED const hashes_t *hashes, MAYBE_UNUSED const hc_device_param_t *device_param)
 {
   char *jit_build_options = NULL;
 
-  // in pure -a 0 mode we reserve pws_t with 64 threads = 256 + 4 bytes = 16640.
-  // the RC4_KEY with 64 threads requires (256 + 4) 16640.
+  u32 native_threads = 0;
 
-  if ((hashconfig->opti_type & OPTI_TYPE_OPTIMIZED_KERNEL) == 0)
+  if (device_param->opencl_device_type & CL_DEVICE_TYPE_CPU)
   {
-    if (user_options_extra->attack_kern == ATTACK_KERN_STRAIGHT)
+    native_threads = 1;
+  }
+  else if (device_param->opencl_device_type & CL_DEVICE_TYPE_GPU)
+  {
+    if (device_param->opencl_device_vendor_id == VENDOR_ID_INTEL_SDK)
+    {
+      native_threads = 8;
+    }
+    else if (device_param->opencl_device_vendor_id == VENDOR_ID_AMD)
     {
       if (device_param->device_local_mem_size < 49152)
       {
-        hc_asprintf (&jit_build_options, "-D FORCE_DISABLE_SHM");
+        native_threads = 32;
+      }
+      else
+      {
+        native_threads = 64;
       }
     }
+    else
+    {
+      native_threads = 32;
+    }
   }
 
+  hc_asprintf (&jit_build_options, "-D FIXED_LOCAL_SIZE=%u -D _unroll", native_threads);
+
   return jit_build_options;
 }
 
+u64 module_esalt_size (MAYBE_UNUSED const hashconfig_t *hashconfig, MAYBE_UNUSED const user_options_t *user_options, MAYBE_UNUSED const user_options_extra_t *user_options_extra)
+{
+  const u64 esalt_size = (const u64) sizeof (krb5pa_t);
+
+  return esalt_size;
+}
+
 int module_hash_decode (MAYBE_UNUSED const hashconfig_t *hashconfig, MAYBE_UNUSED void *digest_buf, MAYBE_UNUSED salt_t *salt, MAYBE_UNUSED void *esalt_buf, MAYBE_UNUSED void *hook_salt_buf, MAYBE_UNUSED hashinfo_t *hash_info, const char *line_buf, MAYBE_UNUSED const int line_len)
 {
   u32 *digest = (u32 *) digest_buf;
@@ -293,8 +296,8 @@ void module_init (module_ctx_t *module_ctx)
   module_ctx->module_kernel_accel_min         = MODULE_DEFAULT;
   module_ctx->module_kernel_loops_max         = MODULE_DEFAULT;
   module_ctx->module_kernel_loops_min         = MODULE_DEFAULT;
-  module_ctx->module_kernel_threads_max       = module_kernel_threads_max;
-  module_ctx->module_kernel_threads_min       = module_kernel_threads_min;
+  module_ctx->module_kernel_threads_max       = MODULE_DEFAULT;
+  module_ctx->module_kernel_threads_min       = MODULE_DEFAULT;
   module_ctx->module_kern_type                = module_kern_type;
   module_ctx->module_kern_type_dynamic        = MODULE_DEFAULT;
   module_ctx->module_opti_type                = module_opti_type;