diff --git a/OpenCL/inc_common.cl b/OpenCL/inc_common.cl
index ff41c41df..abe27347a 100644
--- a/OpenCL/inc_common.cl
+++ b/OpenCL/inc_common.cl
@@ -6947,6 +6947,1619 @@ inline void switch_buffer_by_offset_le_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w
   #endif
 }
 
+inline void switch_buffer_by_offset_8x4_le_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 w4[4], u32 w5[4], u32 w6[4], u32 w7[4], const u32 offset)
+{
+  #if defined IS_AMD || defined IS_GENERIC
+  const int offset_mod_4 = offset & 3;
+
+  const int offset_minus_4 = 4 - offset;
+
+  switch (offset / 4)
+  {
+    case 0:
+      w7[3] = amd_bytealign_S (w7[3], w7[2], offset_minus_4);
+      w7[2] = amd_bytealign_S (w7[2], w7[1], offset_minus_4);
+      w7[1] = amd_bytealign_S (w7[1], w7[0], offset_minus_4);
+      w7[0] = amd_bytealign_S (w7[0], w6[3], offset_minus_4);
+      w6[3] = amd_bytealign_S (w6[3], w6[2], offset_minus_4);
+      w6[2] = amd_bytealign_S (w6[2], w6[1], offset_minus_4);
+      w6[1] = amd_bytealign_S (w6[1], w6[0], offset_minus_4);
+      w6[0] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
+      w5[3] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
+      w5[2] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
+      w5[1] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
+      w5[0] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
+      w4[3] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
+      w4[2] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
+      w4[1] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
+      w4[0] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
+      w3[3] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
+      w3[2] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
+      w3[1] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
+      w3[0] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
+      w2[3] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
+      w2[2] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
+      w2[1] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
+      w2[0] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
+      w1[3] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
+      w1[2] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
+      w1[1] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
+      w1[0] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
+      w0[3] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
+      w0[2] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
+      w0[1] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
+      w0[0] = amd_bytealign_S (w0[0],     0, offset_minus_4);
+
+      if (offset_mod_4 == 0)
+      {
+        w0[0] = w0[1];
+        w0[1] = w0[2];
+        w0[2] = w0[3];
+        w0[3] = w1[0];
+        w1[0] = w1[1];
+        w1[1] = w1[2];
+        w1[2] = w1[3];
+        w1[3] = w2[0];
+        w2[0] = w2[1];
+        w2[1] = w2[2];
+        w2[2] = w2[3];
+        w2[3] = w3[0];
+        w3[0] = w3[1];
+        w3[1] = w3[2];
+        w3[2] = w3[3];
+        w3[3] = w4[0];
+        w4[0] = w4[1];
+        w4[1] = w4[2];
+        w4[2] = w4[3];
+        w4[3] = w5[0];
+        w5[0] = w5[1];
+        w5[1] = w5[2];
+        w5[2] = w5[3];
+        w5[3] = w6[0];
+        w6[0] = w6[1];
+        w6[1] = w6[2];
+        w6[2] = w6[3];
+        w6[3] = w7[0];
+        w7[0] = w7[1];
+        w7[1] = w7[2];
+        w7[2] = w7[3];
+        w7[3] = 0;
+      }
+
+      break;
+
+    case 1:
+      w7[3] = amd_bytealign_S (w7[2], w7[1], offset_minus_4);
+      w7[2] = amd_bytealign_S (w7[1], w7[0], offset_minus_4);
+      w7[1] = amd_bytealign_S (w7[0], w6[3], offset_minus_4);
+      w7[0] = amd_bytealign_S (w6[3], w6[2], offset_minus_4);
+      w6[3] = amd_bytealign_S (w6[2], w6[1], offset_minus_4);
+      w6[2] = amd_bytealign_S (w6[1], w6[0], offset_minus_4);
+      w6[1] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
+      w6[0] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
+      w5[3] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
+      w5[2] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
+      w5[1] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
+      w5[0] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
+      w4[3] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
+      w4[2] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
+      w4[1] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
+      w4[0] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
+      w3[3] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
+      w3[2] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
+      w3[1] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
+      w3[0] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
+      w2[3] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
+      w2[2] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
+      w2[1] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
+      w2[0] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
+      w1[3] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
+      w1[2] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
+      w1[1] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
+      w1[0] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
+      w0[3] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
+      w0[2] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
+      w0[1] = amd_bytealign_S (w0[0],     0, offset_minus_4);
+      w0[0] = 0;
+
+      if (offset_mod_4 == 0)
+      {
+        w0[1] = w0[2];
+        w0[2] = w0[3];
+        w0[3] = w1[0];
+        w1[0] = w1[1];
+        w1[1] = w1[2];
+        w1[2] = w1[3];
+        w1[3] = w2[0];
+        w2[0] = w2[1];
+        w2[1] = w2[2];
+        w2[2] = w2[3];
+        w2[3] = w3[0];
+        w3[0] = w3[1];
+        w3[1] = w3[2];
+        w3[2] = w3[3];
+        w3[3] = w4[0];
+        w4[0] = w4[1];
+        w4[1] = w4[2];
+        w4[2] = w4[3];
+        w4[3] = w5[0];
+        w5[0] = w5[1];
+        w5[1] = w5[2];
+        w5[2] = w5[3];
+        w5[3] = w6[0];
+        w6[0] = w6[1];
+        w6[1] = w6[2];
+        w6[2] = w6[3];
+        w6[3] = w7[0];
+        w7[0] = w7[1];
+        w7[1] = w7[2];
+        w7[2] = w7[3];
+        w7[3] = 0;
+      }
+
+      break;
+
+    case 2:
+      w7[3] = amd_bytealign_S (w7[1], w7[0], offset_minus_4);
+      w7[2] = amd_bytealign_S (w7[0], w6[3], offset_minus_4);
+      w7[1] = amd_bytealign_S (w6[3], w6[2], offset_minus_4);
+      w7[0] = amd_bytealign_S (w6[2], w6[1], offset_minus_4);
+      w6[3] = amd_bytealign_S (w6[1], w6[0], offset_minus_4);
+      w6[2] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
+      w6[1] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
+      w6[0] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
+      w5[3] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
+      w5[2] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
+      w5[1] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
+      w5[0] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
+      w4[3] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
+      w4[2] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
+      w4[1] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
+      w4[0] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
+      w3[3] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
+      w3[2] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
+      w3[1] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
+      w3[0] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
+      w2[3] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
+      w2[2] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
+      w2[1] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
+      w2[0] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
+      w1[3] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
+      w1[2] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
+      w1[1] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
+      w1[0] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
+      w0[3] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
+      w0[2] = amd_bytealign_S (w0[0],     0, offset_minus_4);
+      w0[1] = 0;
+      w0[0] = 0;
+
+      if (offset_mod_4 == 0)
+      {
+        w0[2] = w0[3];
+        w0[3] = w1[0];
+        w1[0] = w1[1];
+        w1[1] = w1[2];
+        w1[2] = w1[3];
+        w1[3] = w2[0];
+        w2[0] = w2[1];
+        w2[1] = w2[2];
+        w2[2] = w2[3];
+        w2[3] = w3[0];
+        w3[0] = w3[1];
+        w3[1] = w3[2];
+        w3[2] = w3[3];
+        w3[3] = w4[0];
+        w4[0] = w4[1];
+        w4[1] = w4[2];
+        w4[2] = w4[3];
+        w4[3] = w5[0];
+        w5[0] = w5[1];
+        w5[1] = w5[2];
+        w5[2] = w5[3];
+        w5[3] = w6[0];
+        w6[0] = w6[1];
+        w6[1] = w6[2];
+        w6[2] = w6[3];
+        w6[3] = w7[0];
+        w7[0] = w7[1];
+        w7[1] = w7[2];
+        w7[2] = w7[3];
+        w7[3] = 0;
+      }
+
+      break;
+
+    case 3:
+      w7[3] = amd_bytealign_S (w7[0], w6[3], offset_minus_4);
+      w7[2] = amd_bytealign_S (w6[3], w6[2], offset_minus_4);
+      w7[1] = amd_bytealign_S (w6[2], w6[1], offset_minus_4);
+      w7[0] = amd_bytealign_S (w6[1], w6[0], offset_minus_4);
+      w6[3] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
+      w6[2] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
+      w6[1] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
+      w6[0] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
+      w5[3] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
+      w5[2] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
+      w5[1] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
+      w5[0] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
+      w4[3] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
+      w4[2] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
+      w4[1] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
+      w4[0] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
+      w3[3] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
+      w3[2] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
+      w3[1] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
+      w3[0] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
+      w2[3] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
+      w2[2] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
+      w2[1] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
+      w2[0] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
+      w1[3] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
+      w1[2] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
+      w1[1] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
+      w1[0] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
+      w0[3] = amd_bytealign_S (w0[0],     0, offset_minus_4);
+      w0[2] = 0;
+      w0[1] = 0;
+      w0[0] = 0;
+
+      if (offset_mod_4 == 0)
+      {
+        w0[3] = w1[0];
+        w1[0] = w1[1];
+        w1[1] = w1[2];
+        w1[2] = w1[3];
+        w1[3] = w2[0];
+        w2[0] = w2[1];
+        w2[1] = w2[2];
+        w2[2] = w2[3];
+        w2[3] = w3[0];
+        w3[0] = w3[1];
+        w3[1] = w3[2];
+        w3[2] = w3[3];
+        w3[3] = w4[0];
+        w4[0] = w4[1];
+        w4[1] = w4[2];
+        w4[2] = w4[3];
+        w4[3] = w5[0];
+        w5[0] = w5[1];
+        w5[1] = w5[2];
+        w5[2] = w5[3];
+        w5[3] = w6[0];
+        w6[0] = w6[1];
+        w6[1] = w6[2];
+        w6[2] = w6[3];
+        w6[3] = w7[0];
+        w7[0] = w7[1];
+        w7[1] = w7[2];
+        w7[2] = w7[3];
+        w7[3] = 0;
+      }
+
+      break;
+
+    case 4:
+      w7[3] = amd_bytealign_S (w6[3], w6[2], offset_minus_4);
+      w7[2] = amd_bytealign_S (w6[2], w6[1], offset_minus_4);
+      w7[1] = amd_bytealign_S (w6[1], w6[0], offset_minus_4);
+      w7[0] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
+      w6[3] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
+      w6[2] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
+      w6[1] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
+      w6[0] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
+      w5[3] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
+      w5[2] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
+      w5[1] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
+      w5[0] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
+      w4[3] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
+      w4[2] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
+      w4[1] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
+      w4[0] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
+      w3[3] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
+      w3[2] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
+      w3[1] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
+      w3[0] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
+      w2[3] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
+      w2[2] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
+      w2[1] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
+      w2[0] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
+      w1[3] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
+      w1[2] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
+      w1[1] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
+      w1[0] = amd_bytealign_S (w0[0],     0, offset_minus_4);
+      w0[3] = 0;
+      w0[2] = 0;
+      w0[1] = 0;
+      w0[0] = 0;
+
+      if (offset_mod_4 == 0)
+      {
+        w1[0] = w1[1];
+        w1[1] = w1[2];
+        w1[2] = w1[3];
+        w1[3] = w2[0];
+        w2[0] = w2[1];
+        w2[1] = w2[2];
+        w2[2] = w2[3];
+        w2[3] = w3[0];
+        w3[0] = w3[1];
+        w3[1] = w3[2];
+        w3[2] = w3[3];
+        w3[3] = w4[0];
+        w4[0] = w4[1];
+        w4[1] = w4[2];
+        w4[2] = w4[3];
+        w4[3] = w5[0];
+        w5[0] = w5[1];
+        w5[1] = w5[2];
+        w5[2] = w5[3];
+        w5[3] = w6[0];
+        w6[0] = w6[1];
+        w6[1] = w6[2];
+        w6[2] = w6[3];
+        w6[3] = w7[0];
+        w7[0] = w7[1];
+        w7[1] = w7[2];
+        w7[2] = w7[3];
+        w7[3] = 0;
+      }
+
+      break;
+
+    case 5:
+      w7[3] = amd_bytealign_S (w6[2], w6[1], offset_minus_4);
+      w7[2] = amd_bytealign_S (w6[1], w6[0], offset_minus_4);
+      w7[1] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
+      w7[0] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
+      w6[3] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
+      w6[2] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
+      w6[1] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
+      w6[0] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
+      w5[3] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
+      w5[2] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
+      w5[1] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
+      w5[0] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
+      w4[3] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
+      w4[2] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
+      w4[1] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
+      w4[0] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
+      w3[3] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
+      w3[2] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
+      w3[1] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
+      w3[0] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
+      w2[3] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
+      w2[2] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
+      w2[1] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
+      w2[0] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
+      w1[3] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
+      w1[2] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
+      w1[1] = amd_bytealign_S (w0[0],     0, offset_minus_4);
+      w1[0] = 0;
+      w0[3] = 0;
+      w0[2] = 0;
+      w0[1] = 0;
+      w0[0] = 0;
+
+      if (offset_mod_4 == 0)
+      {
+        w1[1] = w1[2];
+        w1[2] = w1[3];
+        w1[3] = w2[0];
+        w2[0] = w2[1];
+        w2[1] = w2[2];
+        w2[2] = w2[3];
+        w2[3] = w3[0];
+        w3[0] = w3[1];
+        w3[1] = w3[2];
+        w3[2] = w3[3];
+        w3[3] = w4[0];
+        w4[0] = w4[1];
+        w4[1] = w4[2];
+        w4[2] = w4[3];
+        w4[3] = w5[0];
+        w5[0] = w5[1];
+        w5[1] = w5[2];
+        w5[2] = w5[3];
+        w5[3] = w6[0];
+        w6[0] = w6[1];
+        w6[1] = w6[2];
+        w6[2] = w6[3];
+        w6[3] = w7[0];
+        w7[0] = w7[1];
+        w7[1] = w7[2];
+        w7[2] = w7[3];
+        w7[3] = 0;
+      }
+
+      break;
+
+    case 6:
+      w7[3] = amd_bytealign_S (w6[1], w6[0], offset_minus_4);
+      w7[2] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
+      w7[1] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
+      w7[0] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
+      w6[3] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
+      w6[2] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
+      w6[1] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
+      w6[0] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
+      w5[3] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
+      w5[2] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
+      w5[1] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
+      w5[0] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
+      w4[3] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
+      w4[2] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
+      w4[1] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
+      w4[0] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
+      w3[3] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
+      w3[2] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
+      w3[1] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
+      w3[0] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
+      w2[3] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
+      w2[2] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
+      w2[1] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
+      w2[0] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
+      w1[3] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
+      w1[2] = amd_bytealign_S (w0[0],     0, offset_minus_4);
+      w1[1] = 0;
+      w1[0] = 0;
+      w0[3] = 0;
+      w0[2] = 0;
+      w0[1] = 0;
+      w0[0] = 0;
+
+      if (offset_mod_4 == 0)
+      {
+        w1[2] = w1[3];
+        w1[3] = w2[0];
+        w2[0] = w2[1];
+        w2[1] = w2[2];
+        w2[2] = w2[3];
+        w2[3] = w3[0];
+        w3[0] = w3[1];
+        w3[1] = w3[2];
+        w3[2] = w3[3];
+        w3[3] = w4[0];
+        w4[0] = w4[1];
+        w4[1] = w4[2];
+        w4[2] = w4[3];
+        w4[3] = w5[0];
+        w5[0] = w5[1];
+        w5[1] = w5[2];
+        w5[2] = w5[3];
+        w5[3] = w6[0];
+        w6[0] = w6[1];
+        w6[1] = w6[2];
+        w6[2] = w6[3];
+        w6[3] = w7[0];
+        w7[0] = w7[1];
+        w7[1] = w7[2];
+        w7[2] = w7[3];
+        w7[3] = 0;
+      }
+
+      break;
+
+    case 7:
+      w7[3] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
+      w7[2] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
+      w7[1] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
+      w7[0] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
+      w6[3] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
+      w6[2] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
+      w6[1] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
+      w6[0] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
+      w5[3] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
+      w5[2] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
+      w5[1] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
+      w5[0] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
+      w4[3] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
+      w4[2] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
+      w4[1] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
+      w4[0] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
+      w3[3] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
+      w3[2] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
+      w3[1] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
+      w3[0] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
+      w2[3] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
+      w2[2] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
+      w2[1] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
+      w2[0] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
+      w1[3] = amd_bytealign_S (w0[0],     0, offset_minus_4);
+      w1[2] = 0;
+      w1[1] = 0;
+      w1[0] = 0;
+      w0[3] = 0;
+      w0[2] = 0;
+      w0[1] = 0;
+      w0[0] = 0;
+
+      if (offset_mod_4 == 0)
+      {
+        w1[3] = w2[0];
+        w2[0] = w2[1];
+        w2[1] = w2[2];
+        w2[2] = w2[3];
+        w2[3] = w3[0];
+        w3[0] = w3[1];
+        w3[1] = w3[2];
+        w3[2] = w3[3];
+        w3[3] = w4[0];
+        w4[0] = w4[1];
+        w4[1] = w4[2];
+        w4[2] = w4[3];
+        w4[3] = w5[0];
+        w5[0] = w5[1];
+        w5[1] = w5[2];
+        w5[2] = w5[3];
+        w5[3] = w6[0];
+        w6[0] = w6[1];
+        w6[1] = w6[2];
+        w6[2] = w6[3];
+        w6[3] = w7[0];
+        w7[0] = w7[1];
+        w7[1] = w7[2];
+        w7[2] = w7[3];
+        w7[3] = 0;
+      }
+
+      break;
+
+    case 8:
+      w7[3] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
+      w7[2] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
+      w7[1] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
+      w7[0] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
+      w6[3] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
+      w6[2] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
+      w6[1] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
+      w6[0] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
+      w5[3] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
+      w5[2] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
+      w5[1] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
+      w5[0] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
+      w4[3] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
+      w4[2] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
+      w4[1] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
+      w4[0] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
+      w3[3] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
+      w3[2] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
+      w3[1] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
+      w3[0] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
+      w2[3] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
+      w2[2] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
+      w2[1] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
+      w2[0] = amd_bytealign_S (w0[0],     0, offset_minus_4);
+      w1[3] = 0;
+      w1[2] = 0;
+      w1[1] = 0;
+      w1[0] = 0;
+      w0[3] = 0;
+      w0[2] = 0;
+      w0[1] = 0;
+      w0[0] = 0;
+
+      if (offset_mod_4 == 0)
+      {
+        w2[0] = w2[1];
+        w2[1] = w2[2];
+        w2[2] = w2[3];
+        w2[3] = w3[0];
+        w3[0] = w3[1];
+        w3[1] = w3[2];
+        w3[2] = w3[3];
+        w3[3] = w4[0];
+        w4[0] = w4[1];
+        w4[1] = w4[2];
+        w4[2] = w4[3];
+        w4[3] = w5[0];
+        w5[0] = w5[1];
+        w5[1] = w5[2];
+        w5[2] = w5[3];
+        w5[3] = w6[0];
+        w6[0] = w6[1];
+        w6[1] = w6[2];
+        w6[2] = w6[3];
+        w6[3] = w7[0];
+        w7[0] = w7[1];
+        w7[1] = w7[2];
+        w7[2] = w7[3];
+        w7[3] = 0;
+      }
+
+      break;
+
+    case 9:
+      w7[3] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
+      w7[2] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
+      w7[1] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
+      w7[0] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
+      w6[3] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
+      w6[2] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
+      w6[1] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
+      w6[0] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
+      w5[3] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
+      w5[2] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
+      w5[1] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
+      w5[0] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
+      w4[3] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
+      w4[2] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
+      w4[1] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
+      w4[0] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
+      w3[3] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
+      w3[2] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
+      w3[1] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
+      w3[0] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
+      w2[3] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
+      w2[2] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
+      w2[1] = amd_bytealign_S (w0[0],     0, offset_minus_4);
+      w2[0] = 0;
+      w1[3] = 0;
+      w1[2] = 0;
+      w1[1] = 0;
+      w1[0] = 0;
+      w0[3] = 0;
+      w0[2] = 0;
+      w0[1] = 0;
+      w0[0] = 0;
+
+      if (offset_mod_4 == 0)
+      {
+        w2[1] = w2[2];
+        w2[2] = w2[3];
+        w2[3] = w3[0];
+        w3[0] = w3[1];
+        w3[1] = w3[2];
+        w3[2] = w3[3];
+        w3[3] = w4[0];
+        w4[0] = w4[1];
+        w4[1] = w4[2];
+        w4[2] = w4[3];
+        w4[3] = w5[0];
+        w5[0] = w5[1];
+        w5[1] = w5[2];
+        w5[2] = w5[3];
+        w5[3] = w6[0];
+        w6[0] = w6[1];
+        w6[1] = w6[2];
+        w6[2] = w6[3];
+        w6[3] = w7[0];
+        w7[0] = w7[1];
+        w7[1] = w7[2];
+        w7[2] = w7[3];
+        w7[3] = 0;
+      }
+
+      break;
+
+    case 10:
+      w7[3] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
+      w7[2] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
+      w7[1] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
+      w7[0] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
+      w6[3] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
+      w6[2] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
+      w6[1] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
+      w6[0] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
+      w5[3] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
+      w5[2] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
+      w5[1] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
+      w5[0] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
+      w4[3] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
+      w4[2] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
+      w4[1] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
+      w4[0] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
+      w3[3] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
+      w3[2] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
+      w3[1] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
+      w3[0] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
+      w2[3] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
+      w2[2] = amd_bytealign_S (w0[0],     0, offset_minus_4);
+      w2[1] = 0;
+      w2[0] = 0;
+      w1[3] = 0;
+      w1[2] = 0;
+      w1[1] = 0;
+      w1[0] = 0;
+      w0[3] = 0;
+      w0[2] = 0;
+      w0[1] = 0;
+      w0[0] = 0;
+
+      if (offset_mod_4 == 0)
+      {
+        w2[2] = w2[3];
+        w2[3] = w3[0];
+        w3[0] = w3[1];
+        w3[1] = w3[2];
+        w3[2] = w3[3];
+        w3[3] = w4[0];
+        w4[0] = w4[1];
+        w4[1] = w4[2];
+        w4[2] = w4[3];
+        w4[3] = w5[0];
+        w5[0] = w5[1];
+        w5[1] = w5[2];
+        w5[2] = w5[3];
+        w5[3] = w6[0];
+        w6[0] = w6[1];
+        w6[1] = w6[2];
+        w6[2] = w6[3];
+        w6[3] = w7[0];
+        w7[0] = w7[1];
+        w7[1] = w7[2];
+        w7[2] = w7[3];
+        w7[3] = 0;
+      }
+
+      break;
+
+    case 11:
+      w7[3] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
+      w7[2] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
+      w7[1] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
+      w7[0] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
+      w6[3] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
+      w6[2] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
+      w6[1] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
+      w6[0] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
+      w5[3] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
+      w5[2] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
+      w5[1] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
+      w5[0] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
+      w4[3] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
+      w4[2] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
+      w4[1] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
+      w4[0] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
+      w3[3] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
+      w3[2] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
+      w3[1] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
+      w3[0] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
+      w2[3] = amd_bytealign_S (w0[0],     0, offset_minus_4);
+      w2[2] = 0;
+      w2[1] = 0;
+      w2[0] = 0;
+      w1[3] = 0;
+      w1[2] = 0;
+      w1[1] = 0;
+      w1[0] = 0;
+      w0[3] = 0;
+      w0[2] = 0;
+      w0[1] = 0;
+      w0[0] = 0;
+
+      if (offset_mod_4 == 0)
+      {
+        w2[3] = w3[0];
+        w3[0] = w3[1];
+        w3[1] = w3[2];
+        w3[2] = w3[3];
+        w3[3] = w4[0];
+        w4[0] = w4[1];
+        w4[1] = w4[2];
+        w4[2] = w4[3];
+        w4[3] = w5[0];
+        w5[0] = w5[1];
+        w5[1] = w5[2];
+        w5[2] = w5[3];
+        w5[3] = w6[0];
+        w6[0] = w6[1];
+        w6[1] = w6[2];
+        w6[2] = w6[3];
+        w6[3] = w7[0];
+        w7[0] = w7[1];
+        w7[1] = w7[2];
+        w7[2] = w7[3];
+        w7[3] = 0;
+      }
+
+      break;
+
+    case 12:
+      w7[3] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
+      w7[2] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
+      w7[1] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
+      w7[0] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
+      w6[3] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
+      w6[2] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
+      w6[1] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
+      w6[0] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
+      w5[3] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
+      w5[2] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
+      w5[1] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
+      w5[0] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
+      w4[3] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
+      w4[2] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
+      w4[1] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
+      w4[0] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
+      w3[3] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
+      w3[2] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
+      w3[1] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
+      w3[0] = amd_bytealign_S (w0[0],     0, offset_minus_4);
+      w2[3] = 0;
+      w2[2] = 0;
+      w2[1] = 0;
+      w2[0] = 0;
+      w1[3] = 0;
+      w1[2] = 0;
+      w1[1] = 0;
+      w1[0] = 0;
+      w0[3] = 0;
+      w0[2] = 0;
+      w0[1] = 0;
+      w0[0] = 0;
+
+      if (offset_mod_4 == 0)
+      {
+        w3[0] = w3[1];
+        w3[1] = w3[2];
+        w3[2] = w3[3];
+        w3[3] = w4[0];
+        w4[0] = w4[1];
+        w4[1] = w4[2];
+        w4[2] = w4[3];
+        w4[3] = w5[0];
+        w5[0] = w5[1];
+        w5[1] = w5[2];
+        w5[2] = w5[3];
+        w5[3] = w6[0];
+        w6[0] = w6[1];
+        w6[1] = w6[2];
+        w6[2] = w6[3];
+        w6[3] = w7[0];
+        w7[0] = w7[1];
+        w7[1] = w7[2];
+        w7[2] = w7[3];
+        w7[3] = 0;
+      }
+
+      break;
+
+    case 13:
+      w7[3] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
+      w7[2] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
+      w7[1] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
+      w7[0] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
+      w6[3] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
+      w6[2] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
+      w6[1] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
+      w6[0] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
+      w5[3] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
+      w5[2] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
+      w5[1] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
+      w5[0] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
+      w4[3] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
+      w4[2] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
+      w4[1] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
+      w4[0] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
+      w3[3] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
+      w3[2] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
+      w3[1] = amd_bytealign_S (w0[0],     0, offset_minus_4);
+      w3[0] = 0;
+      w2[3] = 0;
+      w2[2] = 0;
+      w2[1] = 0;
+      w2[0] = 0;
+      w1[3] = 0;
+      w1[2] = 0;
+      w1[1] = 0;
+      w1[0] = 0;
+      w0[3] = 0;
+      w0[2] = 0;
+      w0[1] = 0;
+      w0[0] = 0;
+
+      if (offset_mod_4 == 0)
+      {
+        w3[1] = w3[2];
+        w3[2] = w3[3];
+        w3[3] = w4[0];
+        w4[0] = w4[1];
+        w4[1] = w4[2];
+        w4[2] = w4[3];
+        w4[3] = w5[0];
+        w5[0] = w5[1];
+        w5[1] = w5[2];
+        w5[2] = w5[3];
+        w5[3] = w6[0];
+        w6[0] = w6[1];
+        w6[1] = w6[2];
+        w6[2] = w6[3];
+        w6[3] = w7[0];
+        w7[0] = w7[1];
+        w7[1] = w7[2];
+        w7[2] = w7[3];
+        w7[3] = 0;
+      }
+
+      break;
+
+    case 14:
+      w7[3] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
+      w7[2] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
+      w7[1] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
+      w7[0] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
+      w6[3] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
+      w6[2] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
+      w6[1] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
+      w6[0] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
+      w5[3] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
+      w5[2] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
+      w5[1] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
+      w5[0] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
+      w4[3] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
+      w4[2] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
+      w4[1] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
+      w4[0] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
+      w3[3] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
+      w3[2] = amd_bytealign_S (w0[0],     0, offset_minus_4);
+      w3[1] = 0;
+      w3[0] = 0;
+      w2[3] = 0;
+      w2[2] = 0;
+      w2[1] = 0;
+      w2[0] = 0;
+      w1[3] = 0;
+      w1[2] = 0;
+      w1[1] = 0;
+      w1[0] = 0;
+      w0[3] = 0;
+      w0[2] = 0;
+      w0[1] = 0;
+      w0[0] = 0;
+
+      if (offset_mod_4 == 0)
+      {
+        w3[2] = w3[3];
+        w3[3] = w4[0];
+        w4[0] = w4[1];
+        w4[1] = w4[2];
+        w4[2] = w4[3];
+        w4[3] = w5[0];
+        w5[0] = w5[1];
+        w5[1] = w5[2];
+        w5[2] = w5[3];
+        w5[3] = w6[0];
+        w6[0] = w6[1];
+        w6[1] = w6[2];
+        w6[2] = w6[3];
+        w6[3] = w7[0];
+        w7[0] = w7[1];
+        w7[1] = w7[2];
+        w7[2] = w7[3];
+        w7[3] = 0;
+      }
+
+      break;
+
+    case 15:
+      w7[3] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
+      w7[2] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
+      w7[1] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
+      w7[0] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
+      w6[3] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
+      w6[2] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
+      w6[1] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
+      w6[0] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
+      w5[3] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
+      w5[2] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
+      w5[1] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
+      w5[0] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
+      w4[3] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
+      w4[2] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
+      w4[1] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
+      w4[0] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
+      w3[3] = amd_bytealign_S (w0[0],     0, offset_minus_4);
+      w3[2] = 0;
+      w3[1] = 0;
+      w3[0] = 0;
+      w2[3] = 0;
+      w2[2] = 0;
+      w2[1] = 0;
+      w2[0] = 0;
+      w1[3] = 0;
+      w1[2] = 0;
+      w1[1] = 0;
+      w1[0] = 0;
+      w0[3] = 0;
+      w0[2] = 0;
+      w0[1] = 0;
+      w0[0] = 0;
+
+      if (offset_mod_4 == 0)
+      {
+        w3[3] = w4[0];
+        w4[0] = w4[1];
+        w4[1] = w4[2];
+        w4[2] = w4[3];
+        w4[3] = w5[0];
+        w5[0] = w5[1];
+        w5[1] = w5[2];
+        w5[2] = w5[3];
+        w5[3] = w6[0];
+        w6[0] = w6[1];
+        w6[1] = w6[2];
+        w6[2] = w6[3];
+        w6[3] = w7[0];
+        w7[0] = w7[1];
+        w7[1] = w7[2];
+        w7[2] = w7[3];
+        w7[3] = 0;
+      }
+
+      break;
+  }
+  #endif
+
+  #ifdef IS_NV
+  const int offset_minus_4 = 4 - (offset % 4);
+
+  const int selector = (0x76543210 >> (offset_minus_4 * 4)) & 0xffff;
+
+  switch (offset / 4)
+  {
+    case 0:
+      w7[3] = __byte_perm_S (w7[2], w7[3], selector);
+      w7[2] = __byte_perm_S (w7[1], w7[2], selector);
+      w7[1] = __byte_perm_S (w7[0], w7[1], selector);
+      w7[0] = __byte_perm_S (w6[3], w7[0], selector);
+      w6[3] = __byte_perm_S (w6[2], w6[3], selector);
+      w6[2] = __byte_perm_S (w6[1], w6[2], selector);
+      w6[1] = __byte_perm_S (w6[0], w6[1], selector);
+      w6[0] = __byte_perm_S (w5[3], w6[0], selector);
+      w5[3] = __byte_perm_S (w5[2], w5[3], selector);
+      w5[2] = __byte_perm_S (w5[1], w5[2], selector);
+      w5[1] = __byte_perm_S (w5[0], w5[1], selector);
+      w5[0] = __byte_perm_S (w4[3], w5[0], selector);
+      w4[3] = __byte_perm_S (w4[2], w4[3], selector);
+      w4[2] = __byte_perm_S (w4[1], w4[2], selector);
+      w4[1] = __byte_perm_S (w4[0], w4[1], selector);
+      w4[0] = __byte_perm_S (w3[3], w4[0], selector);
+      w3[3] = __byte_perm_S (w3[2], w3[3], selector);
+      w3[2] = __byte_perm_S (w3[1], w3[2], selector);
+      w3[1] = __byte_perm_S (w3[0], w3[1], selector);
+      w3[0] = __byte_perm_S (w2[3], w3[0], selector);
+      w2[3] = __byte_perm_S (w2[2], w2[3], selector);
+      w2[2] = __byte_perm_S (w2[1], w2[2], selector);
+      w2[1] = __byte_perm_S (w2[0], w2[1], selector);
+      w2[0] = __byte_perm_S (w1[3], w2[0], selector);
+      w1[3] = __byte_perm_S (w1[2], w1[3], selector);
+      w1[2] = __byte_perm_S (w1[1], w1[2], selector);
+      w1[1] = __byte_perm_S (w1[0], w1[1], selector);
+      w1[0] = __byte_perm_S (w0[3], w1[0], selector);
+      w0[3] = __byte_perm_S (w0[2], w0[3], selector);
+      w0[2] = __byte_perm_S (w0[1], w0[2], selector);
+      w0[1] = __byte_perm_S (w0[0], w0[1], selector);
+      w0[0] = __byte_perm_S (    0, w0[0], selector);
+      break;
+
+    case 1:
+      w7[3] = __byte_perm_S (w7[1], w7[2], selector);
+      w7[2] = __byte_perm_S (w7[0], w7[1], selector);
+      w7[1] = __byte_perm_S (w6[3], w7[0], selector);
+      w7[0] = __byte_perm_S (w6[2], w6[3], selector);
+      w6[3] = __byte_perm_S (w6[1], w6[2], selector);
+      w6[2] = __byte_perm_S (w6[0], w6[1], selector);
+      w6[1] = __byte_perm_S (w5[3], w6[0], selector);
+      w6[0] = __byte_perm_S (w5[2], w5[3], selector);
+      w5[3] = __byte_perm_S (w5[1], w5[2], selector);
+      w5[2] = __byte_perm_S (w5[0], w5[1], selector);
+      w5[1] = __byte_perm_S (w4[3], w5[0], selector);
+      w5[0] = __byte_perm_S (w4[2], w4[3], selector);
+      w4[3] = __byte_perm_S (w4[1], w4[2], selector);
+      w4[2] = __byte_perm_S (w4[0], w4[1], selector);
+      w4[1] = __byte_perm_S (w3[3], w4[0], selector);
+      w4[0] = __byte_perm_S (w3[2], w3[3], selector);
+      w3[3] = __byte_perm_S (w3[1], w3[2], selector);
+      w3[2] = __byte_perm_S (w3[0], w3[1], selector);
+      w3[1] = __byte_perm_S (w2[3], w3[0], selector);
+      w3[0] = __byte_perm_S (w2[2], w2[3], selector);
+      w2[3] = __byte_perm_S (w2[1], w2[2], selector);
+      w2[2] = __byte_perm_S (w2[0], w2[1], selector);
+      w2[1] = __byte_perm_S (w1[3], w2[0], selector);
+      w2[0] = __byte_perm_S (w1[2], w1[3], selector);
+      w1[3] = __byte_perm_S (w1[1], w1[2], selector);
+      w1[2] = __byte_perm_S (w1[0], w1[1], selector);
+      w1[1] = __byte_perm_S (w0[3], w1[0], selector);
+      w1[0] = __byte_perm_S (w0[2], w0[3], selector);
+      w0[3] = __byte_perm_S (w0[1], w0[2], selector);
+      w0[2] = __byte_perm_S (w0[0], w0[1], selector);
+      w0[1] = __byte_perm_S (    0, w0[0], selector);
+      w0[0] = 0;
+      break;
+
+    case 2:
+      w7[3] = __byte_perm_S (w7[0], w7[1], selector);
+      w7[2] = __byte_perm_S (w6[3], w7[0], selector);
+      w7[1] = __byte_perm_S (w6[2], w6[3], selector);
+      w7[0] = __byte_perm_S (w6[1], w6[2], selector);
+      w6[3] = __byte_perm_S (w6[0], w6[1], selector);
+      w6[2] = __byte_perm_S (w5[3], w6[0], selector);
+      w6[1] = __byte_perm_S (w5[2], w5[3], selector);
+      w6[0] = __byte_perm_S (w5[1], w5[2], selector);
+      w5[3] = __byte_perm_S (w5[0], w5[1], selector);
+      w5[2] = __byte_perm_S (w4[3], w5[0], selector);
+      w5[1] = __byte_perm_S (w4[2], w4[3], selector);
+      w5[0] = __byte_perm_S (w4[1], w4[2], selector);
+      w4[3] = __byte_perm_S (w4[0], w4[1], selector);
+      w4[2] = __byte_perm_S (w3[3], w4[0], selector);
+      w4[1] = __byte_perm_S (w3[2], w3[3], selector);
+      w4[0] = __byte_perm_S (w3[1], w3[2], selector);
+      w3[3] = __byte_perm_S (w3[0], w3[1], selector);
+      w3[2] = __byte_perm_S (w2[3], w3[0], selector);
+      w3[1] = __byte_perm_S (w2[2], w2[3], selector);
+      w3[0] = __byte_perm_S (w2[1], w2[2], selector);
+      w2[3] = __byte_perm_S (w2[0], w2[1], selector);
+      w2[2] = __byte_perm_S (w1[3], w2[0], selector);
+      w2[1] = __byte_perm_S (w1[2], w1[3], selector);
+      w2[0] = __byte_perm_S (w1[1], w1[2], selector);
+      w1[3] = __byte_perm_S (w1[0], w1[1], selector);
+      w1[2] = __byte_perm_S (w0[3], w1[0], selector);
+      w1[1] = __byte_perm_S (w0[2], w0[3], selector);
+      w1[0] = __byte_perm_S (w0[1], w0[2], selector);
+      w0[3] = __byte_perm_S (w0[0], w0[1], selector);
+      w0[2] = __byte_perm_S (    0, w0[0], selector);
+      w0[1] = 0;
+      w0[0] = 0;
+      break;
+
+    case 3:
+      w7[3] = __byte_perm_S (w6[3], w7[0], selector);
+      w7[2] = __byte_perm_S (w6[2], w6[3], selector);
+      w7[1] = __byte_perm_S (w6[1], w6[2], selector);
+      w7[0] = __byte_perm_S (w6[0], w6[1], selector);
+      w6[3] = __byte_perm_S (w5[3], w6[0], selector);
+      w6[2] = __byte_perm_S (w5[2], w5[3], selector);
+      w6[1] = __byte_perm_S (w5[1], w5[2], selector);
+      w6[0] = __byte_perm_S (w5[0], w5[1], selector);
+      w5[3] = __byte_perm_S (w4[3], w5[0], selector);
+      w5[2] = __byte_perm_S (w4[2], w4[3], selector);
+      w5[1] = __byte_perm_S (w4[1], w4[2], selector);
+      w5[0] = __byte_perm_S (w4[0], w4[1], selector);
+      w4[3] = __byte_perm_S (w3[3], w4[0], selector);
+      w4[2] = __byte_perm_S (w3[2], w3[3], selector);
+      w4[1] = __byte_perm_S (w3[1], w3[2], selector);
+      w4[0] = __byte_perm_S (w3[0], w3[1], selector);
+      w3[3] = __byte_perm_S (w2[3], w3[0], selector);
+      w3[2] = __byte_perm_S (w2[2], w2[3], selector);
+      w3[1] = __byte_perm_S (w2[1], w2[2], selector);
+      w3[0] = __byte_perm_S (w2[0], w2[1], selector);
+      w2[3] = __byte_perm_S (w1[3], w2[0], selector);
+      w2[2] = __byte_perm_S (w1[2], w1[3], selector);
+      w2[1] = __byte_perm_S (w1[1], w1[2], selector);
+      w2[0] = __byte_perm_S (w1[0], w1[1], selector);
+      w1[3] = __byte_perm_S (w0[3], w1[0], selector);
+      w1[2] = __byte_perm_S (w0[2], w0[3], selector);
+      w1[1] = __byte_perm_S (w0[1], w0[2], selector);
+      w1[0] = __byte_perm_S (w0[0], w0[1], selector);
+      w0[3] = __byte_perm_S (    0, w0[0], selector);
+      w0[2] = 0;
+      w0[1] = 0;
+      w0[0] = 0;
+      break;
+
+    case 4:
+      w7[3] = __byte_perm_S (w6[2], w6[3], selector);
+      w7[2] = __byte_perm_S (w6[1], w6[2], selector);
+      w7[1] = __byte_perm_S (w6[0], w6[1], selector);
+      w7[0] = __byte_perm_S (w5[3], w6[0], selector);
+      w6[3] = __byte_perm_S (w5[2], w5[3], selector);
+      w6[2] = __byte_perm_S (w5[1], w5[2], selector);
+      w6[1] = __byte_perm_S (w5[0], w5[1], selector);
+      w6[0] = __byte_perm_S (w4[3], w5[0], selector);
+      w5[3] = __byte_perm_S (w4[2], w4[3], selector);
+      w5[2] = __byte_perm_S (w4[1], w4[2], selector);
+      w5[1] = __byte_perm_S (w4[0], w4[1], selector);
+      w5[0] = __byte_perm_S (w3[3], w4[0], selector);
+      w4[3] = __byte_perm_S (w3[2], w3[3], selector);
+      w4[2] = __byte_perm_S (w3[1], w3[2], selector);
+      w4[1] = __byte_perm_S (w3[0], w3[1], selector);
+      w4[0] = __byte_perm_S (w2[3], w3[0], selector);
+      w3[3] = __byte_perm_S (w2[2], w2[3], selector);
+      w3[2] = __byte_perm_S (w2[1], w2[2], selector);
+      w3[1] = __byte_perm_S (w2[0], w2[1], selector);
+      w3[0] = __byte_perm_S (w1[3], w2[0], selector);
+      w2[3] = __byte_perm_S (w1[2], w1[3], selector);
+      w2[2] = __byte_perm_S (w1[1], w1[2], selector);
+      w2[1] = __byte_perm_S (w1[0], w1[1], selector);
+      w2[0] = __byte_perm_S (w0[3], w1[0], selector);
+      w1[3] = __byte_perm_S (w0[2], w0[3], selector);
+      w1[2] = __byte_perm_S (w0[1], w0[2], selector);
+      w1[1] = __byte_perm_S (w0[0], w0[1], selector);
+      w1[0] = __byte_perm_S (    0, w0[0], selector);
+      w0[3] = 0;
+      w0[2] = 0;
+      w0[1] = 0;
+      w0[0] = 0;
+      break;
+
+    case 5:
+      w7[3] = __byte_perm_S (w6[1], w6[2], selector);
+      w7[2] = __byte_perm_S (w6[0], w6[1], selector);
+      w7[1] = __byte_perm_S (w5[3], w6[0], selector);
+      w7[0] = __byte_perm_S (w5[2], w5[3], selector);
+      w6[3] = __byte_perm_S (w5[1], w5[2], selector);
+      w6[2] = __byte_perm_S (w5[0], w5[1], selector);
+      w6[1] = __byte_perm_S (w4[3], w5[0], selector);
+      w6[0] = __byte_perm_S (w4[2], w4[3], selector);
+      w5[3] = __byte_perm_S (w4[1], w4[2], selector);
+      w5[2] = __byte_perm_S (w4[0], w4[1], selector);
+      w5[1] = __byte_perm_S (w3[3], w4[0], selector);
+      w5[0] = __byte_perm_S (w3[2], w3[3], selector);
+      w4[3] = __byte_perm_S (w3[1], w3[2], selector);
+      w4[2] = __byte_perm_S (w3[0], w3[1], selector);
+      w4[1] = __byte_perm_S (w2[3], w3[0], selector);
+      w4[0] = __byte_perm_S (w2[2], w2[3], selector);
+      w3[3] = __byte_perm_S (w2[1], w2[2], selector);
+      w3[2] = __byte_perm_S (w2[0], w2[1], selector);
+      w3[1] = __byte_perm_S (w1[3], w2[0], selector);
+      w3[0] = __byte_perm_S (w1[2], w1[3], selector);
+      w2[3] = __byte_perm_S (w1[1], w1[2], selector);
+      w2[2] = __byte_perm_S (w1[0], w1[1], selector);
+      w2[1] = __byte_perm_S (w0[3], w1[0], selector);
+      w2[0] = __byte_perm_S (w0[2], w0[3], selector);
+      w1[3] = __byte_perm_S (w0[1], w0[2], selector);
+      w1[2] = __byte_perm_S (w0[0], w0[1], selector);
+      w1[1] = __byte_perm_S (    0, w0[0], selector);
+      w1[0] = 0;
+      w0[3] = 0;
+      w0[2] = 0;
+      w0[1] = 0;
+      w0[0] = 0;
+      break;
+
+    case 6:
+      w7[3] = __byte_perm_S (w6[0], w6[1], selector);
+      w7[2] = __byte_perm_S (w5[3], w6[0], selector);
+      w7[1] = __byte_perm_S (w5[2], w5[3], selector);
+      w7[0] = __byte_perm_S (w5[1], w5[2], selector);
+      w6[3] = __byte_perm_S (w5[0], w5[1], selector);
+      w6[2] = __byte_perm_S (w4[3], w5[0], selector);
+      w6[1] = __byte_perm_S (w4[2], w4[3], selector);
+      w6[0] = __byte_perm_S (w4[1], w4[2], selector);
+      w5[3] = __byte_perm_S (w4[0], w4[1], selector);
+      w5[2] = __byte_perm_S (w3[3], w4[0], selector);
+      w5[1] = __byte_perm_S (w3[2], w3[3], selector);
+      w5[0] = __byte_perm_S (w3[1], w3[2], selector);
+      w4[3] = __byte_perm_S (w3[0], w3[1], selector);
+      w4[2] = __byte_perm_S (w2[3], w3[0], selector);
+      w4[1] = __byte_perm_S (w2[2], w2[3], selector);
+      w4[0] = __byte_perm_S (w2[1], w2[2], selector);
+      w3[3] = __byte_perm_S (w2[0], w2[1], selector);
+      w3[2] = __byte_perm_S (w1[3], w2[0], selector);
+      w3[1] = __byte_perm_S (w1[2], w1[3], selector);
+      w3[0] = __byte_perm_S (w1[1], w1[2], selector);
+      w2[3] = __byte_perm_S (w1[0], w1[1], selector);
+      w2[2] = __byte_perm_S (w0[3], w1[0], selector);
+      w2[1] = __byte_perm_S (w0[2], w0[3], selector);
+      w2[0] = __byte_perm_S (w0[1], w0[2], selector);
+      w1[3] = __byte_perm_S (w0[0], w0[1], selector);
+      w1[2] = __byte_perm_S (    0, w0[0], selector);
+      w1[1] = 0;
+      w1[0] = 0;
+      w0[3] = 0;
+      w0[2] = 0;
+      w0[1] = 0;
+      w0[0] = 0;
+      break;
+
+    case 7:
+      w7[3] = __byte_perm_S (w5[3], w6[0], selector);
+      w7[2] = __byte_perm_S (w5[2], w5[3], selector);
+      w7[1] = __byte_perm_S (w5[1], w5[2], selector);
+      w7[0] = __byte_perm_S (w5[0], w5[1], selector);
+      w6[3] = __byte_perm_S (w4[3], w5[0], selector);
+      w6[2] = __byte_perm_S (w4[2], w4[3], selector);
+      w6[1] = __byte_perm_S (w4[1], w4[2], selector);
+      w6[0] = __byte_perm_S (w4[0], w4[1], selector);
+      w5[3] = __byte_perm_S (w3[3], w4[0], selector);
+      w5[2] = __byte_perm_S (w3[2], w3[3], selector);
+      w5[1] = __byte_perm_S (w3[1], w3[2], selector);
+      w5[0] = __byte_perm_S (w3[0], w3[1], selector);
+      w4[3] = __byte_perm_S (w2[3], w3[0], selector);
+      w4[2] = __byte_perm_S (w2[2], w2[3], selector);
+      w4[1] = __byte_perm_S (w2[1], w2[2], selector);
+      w4[0] = __byte_perm_S (w2[0], w2[1], selector);
+      w3[3] = __byte_perm_S (w1[3], w2[0], selector);
+      w3[2] = __byte_perm_S (w1[2], w1[3], selector);
+      w3[1] = __byte_perm_S (w1[1], w1[2], selector);
+      w3[0] = __byte_perm_S (w1[0], w1[1], selector);
+      w2[3] = __byte_perm_S (w0[3], w1[0], selector);
+      w2[2] = __byte_perm_S (w0[2], w0[3], selector);
+      w2[1] = __byte_perm_S (w0[1], w0[2], selector);
+      w2[0] = __byte_perm_S (w0[0], w0[1], selector);
+      w1[3] = __byte_perm_S (    0, w0[0], selector);
+      w1[2] = 0;
+      w1[1] = 0;
+      w1[0] = 0;
+      w0[3] = 0;
+      w0[2] = 0;
+      w0[1] = 0;
+      w0[0] = 0;
+      break;
+
+    case 8:
+      w7[3] = __byte_perm_S (w5[2], w5[3], selector);
+      w7[2] = __byte_perm_S (w5[1], w5[2], selector);
+      w7[1] = __byte_perm_S (w5[0], w5[1], selector);
+      w7[0] = __byte_perm_S (w4[3], w5[0], selector);
+      w6[3] = __byte_perm_S (w4[2], w4[3], selector);
+      w6[2] = __byte_perm_S (w4[1], w4[2], selector);
+      w6[1] = __byte_perm_S (w4[0], w4[1], selector);
+      w6[0] = __byte_perm_S (w3[3], w4[0], selector);
+      w5[3] = __byte_perm_S (w3[2], w3[3], selector);
+      w5[2] = __byte_perm_S (w3[1], w3[2], selector);
+      w5[1] = __byte_perm_S (w3[0], w3[1], selector);
+      w5[0] = __byte_perm_S (w2[3], w3[0], selector);
+      w4[3] = __byte_perm_S (w2[2], w2[3], selector);
+      w4[2] = __byte_perm_S (w2[1], w2[2], selector);
+      w4[1] = __byte_perm_S (w2[0], w2[1], selector);
+      w4[0] = __byte_perm_S (w1[3], w2[0], selector);
+      w3[3] = __byte_perm_S (w1[2], w1[3], selector);
+      w3[2] = __byte_perm_S (w1[1], w1[2], selector);
+      w3[1] = __byte_perm_S (w1[0], w1[1], selector);
+      w3[0] = __byte_perm_S (w0[3], w1[0], selector);
+      w2[3] = __byte_perm_S (w0[2], w0[3], selector);
+      w2[2] = __byte_perm_S (w0[1], w0[2], selector);
+      w2[1] = __byte_perm_S (w0[0], w0[1], selector);
+      w2[0] = __byte_perm_S (    0, w0[0], selector);
+      w1[3] = 0;
+      w1[2] = 0;
+      w1[1] = 0;
+      w1[0] = 0;
+      w0[3] = 0;
+      w0[2] = 0;
+      w0[1] = 0;
+      w0[0] = 0;
+      break;
+
+    case 9:
+      w7[3] = __byte_perm_S (w5[1], w5[2], selector);
+      w7[2] = __byte_perm_S (w5[0], w5[1], selector);
+      w7[1] = __byte_perm_S (w4[3], w5[0], selector);
+      w7[0] = __byte_perm_S (w4[2], w4[3], selector);
+      w6[3] = __byte_perm_S (w4[1], w4[2], selector);
+      w6[2] = __byte_perm_S (w4[0], w4[1], selector);
+      w6[1] = __byte_perm_S (w3[3], w4[0], selector);
+      w6[0] = __byte_perm_S (w3[2], w3[3], selector);
+      w5[3] = __byte_perm_S (w3[1], w3[2], selector);
+      w5[2] = __byte_perm_S (w3[0], w3[1], selector);
+      w5[1] = __byte_perm_S (w2[3], w3[0], selector);
+      w5[0] = __byte_perm_S (w2[2], w2[3], selector);
+      w4[3] = __byte_perm_S (w2[1], w2[2], selector);
+      w4[2] = __byte_perm_S (w2[0], w2[1], selector);
+      w4[1] = __byte_perm_S (w1[3], w2[0], selector);
+      w4[0] = __byte_perm_S (w1[2], w1[3], selector);
+      w3[3] = __byte_perm_S (w1[1], w1[2], selector);
+      w3[2] = __byte_perm_S (w1[0], w1[1], selector);
+      w3[1] = __byte_perm_S (w0[3], w1[0], selector);
+      w3[0] = __byte_perm_S (w0[2], w0[3], selector);
+      w2[3] = __byte_perm_S (w0[1], w0[2], selector);
+      w2[2] = __byte_perm_S (w0[0], w0[1], selector);
+      w2[1] = __byte_perm_S (    0, w0[0], selector);
+      w2[0] = 0;
+      w1[3] = 0;
+      w1[2] = 0;
+      w1[1] = 0;
+      w1[0] = 0;
+      w0[3] = 0;
+      w0[2] = 0;
+      w0[1] = 0;
+      w0[0] = 0;
+      break;
+
+    case 10:
+      w7[3] = __byte_perm_S (w5[0], w5[1], selector);
+      w7[2] = __byte_perm_S (w4[3], w5[0], selector);
+      w7[1] = __byte_perm_S (w4[2], w4[3], selector);
+      w7[0] = __byte_perm_S (w4[1], w4[2], selector);
+      w6[3] = __byte_perm_S (w4[0], w4[1], selector);
+      w6[2] = __byte_perm_S (w3[3], w4[0], selector);
+      w6[1] = __byte_perm_S (w3[2], w3[3], selector);
+      w6[0] = __byte_perm_S (w3[1], w3[2], selector);
+      w5[3] = __byte_perm_S (w3[0], w3[1], selector);
+      w5[2] = __byte_perm_S (w2[3], w3[0], selector);
+      w5[1] = __byte_perm_S (w2[2], w2[3], selector);
+      w5[0] = __byte_perm_S (w2[1], w2[2], selector);
+      w4[3] = __byte_perm_S (w2[0], w2[1], selector);
+      w4[2] = __byte_perm_S (w1[3], w2[0], selector);
+      w4[1] = __byte_perm_S (w1[2], w1[3], selector);
+      w4[0] = __byte_perm_S (w1[1], w1[2], selector);
+      w3[3] = __byte_perm_S (w1[0], w1[1], selector);
+      w3[2] = __byte_perm_S (w0[3], w1[0], selector);
+      w3[1] = __byte_perm_S (w0[2], w0[3], selector);
+      w3[0] = __byte_perm_S (w0[1], w0[2], selector);
+      w2[3] = __byte_perm_S (w0[0], w0[1], selector);
+      w2[2] = __byte_perm_S (    0, w0[0], selector);
+      w2[1] = 0;
+      w2[0] = 0;
+      w1[3] = 0;
+      w1[2] = 0;
+      w1[1] = 0;
+      w1[0] = 0;
+      w0[3] = 0;
+      w0[2] = 0;
+      w0[1] = 0;
+      w0[0] = 0;
+      break;
+
+    case 11:
+      w7[3] = __byte_perm_S (w4[3], w5[0], selector);
+      w7[2] = __byte_perm_S (w4[2], w4[3], selector);
+      w7[1] = __byte_perm_S (w4[1], w4[2], selector);
+      w7[0] = __byte_perm_S (w4[0], w4[1], selector);
+      w6[3] = __byte_perm_S (w3[3], w4[0], selector);
+      w6[2] = __byte_perm_S (w3[2], w3[3], selector);
+      w6[1] = __byte_perm_S (w3[1], w3[2], selector);
+      w6[0] = __byte_perm_S (w3[0], w3[1], selector);
+      w5[3] = __byte_perm_S (w2[3], w3[0], selector);
+      w5[2] = __byte_perm_S (w2[2], w2[3], selector);
+      w5[1] = __byte_perm_S (w2[1], w2[2], selector);
+      w5[0] = __byte_perm_S (w2[0], w2[1], selector);
+      w4[3] = __byte_perm_S (w1[3], w2[0], selector);
+      w4[2] = __byte_perm_S (w1[2], w1[3], selector);
+      w4[1] = __byte_perm_S (w1[1], w1[2], selector);
+      w4[0] = __byte_perm_S (w1[0], w1[1], selector);
+      w3[3] = __byte_perm_S (w0[3], w1[0], selector);
+      w3[2] = __byte_perm_S (w0[2], w0[3], selector);
+      w3[1] = __byte_perm_S (w0[1], w0[2], selector);
+      w3[0] = __byte_perm_S (w0[0], w0[1], selector);
+      w2[3] = __byte_perm_S (    0, w0[0], selector);
+      w2[2] = 0;
+      w2[1] = 0;
+      w2[0] = 0;
+      w1[3] = 0;
+      w1[2] = 0;
+      w1[1] = 0;
+      w1[0] = 0;
+      w0[3] = 0;
+      w0[2] = 0;
+      w0[1] = 0;
+      w0[0] = 0;
+      break;
+
+    case 12:
+      w7[3] = __byte_perm_S (w4[2], w4[3], selector);
+      w7[2] = __byte_perm_S (w4[1], w4[2], selector);
+      w7[1] = __byte_perm_S (w4[0], w4[1], selector);
+      w7[0] = __byte_perm_S (w3[3], w4[0], selector);
+      w6[3] = __byte_perm_S (w3[2], w3[3], selector);
+      w6[2] = __byte_perm_S (w3[1], w3[2], selector);
+      w6[1] = __byte_perm_S (w3[0], w3[1], selector);
+      w6[0] = __byte_perm_S (w2[3], w3[0], selector);
+      w5[3] = __byte_perm_S (w2[2], w2[3], selector);
+      w5[2] = __byte_perm_S (w2[1], w2[2], selector);
+      w5[1] = __byte_perm_S (w2[0], w2[1], selector);
+      w5[0] = __byte_perm_S (w1[3], w2[0], selector);
+      w4[3] = __byte_perm_S (w1[2], w1[3], selector);
+      w4[2] = __byte_perm_S (w1[1], w1[2], selector);
+      w4[1] = __byte_perm_S (w1[0], w1[1], selector);
+      w4[0] = __byte_perm_S (w0[3], w1[0], selector);
+      w3[3] = __byte_perm_S (w0[2], w0[3], selector);
+      w3[2] = __byte_perm_S (w0[1], w0[2], selector);
+      w3[1] = __byte_perm_S (w0[0], w0[1], selector);
+      w3[0] = __byte_perm_S (    0, w0[0], selector);
+      w2[3] = 0;
+      w2[2] = 0;
+      w2[1] = 0;
+      w2[0] = 0;
+      w1[3] = 0;
+      w1[2] = 0;
+      w1[1] = 0;
+      w1[0] = 0;
+      w0[3] = 0;
+      w0[2] = 0;
+      w0[1] = 0;
+      w0[0] = 0;
+      break;
+
+    case 13:
+      w7[3] = __byte_perm_S (w4[1], w4[2], selector);
+      w7[2] = __byte_perm_S (w4[0], w4[1], selector);
+      w7[1] = __byte_perm_S (w3[3], w4[0], selector);
+      w7[0] = __byte_perm_S (w3[2], w3[3], selector);
+      w6[3] = __byte_perm_S (w3[1], w3[2], selector);
+      w6[2] = __byte_perm_S (w3[0], w3[1], selector);
+      w6[1] = __byte_perm_S (w2[3], w3[0], selector);
+      w6[0] = __byte_perm_S (w2[2], w2[3], selector);
+      w5[3] = __byte_perm_S (w2[1], w2[2], selector);
+      w5[2] = __byte_perm_S (w2[0], w2[1], selector);
+      w5[1] = __byte_perm_S (w1[3], w2[0], selector);
+      w5[0] = __byte_perm_S (w1[2], w1[3], selector);
+      w4[3] = __byte_perm_S (w1[1], w1[2], selector);
+      w4[2] = __byte_perm_S (w1[0], w1[1], selector);
+      w4[1] = __byte_perm_S (w0[3], w1[0], selector);
+      w4[0] = __byte_perm_S (w0[2], w0[3], selector);
+      w3[3] = __byte_perm_S (w0[1], w0[2], selector);
+      w3[2] = __byte_perm_S (w0[0], w0[1], selector);
+      w3[1] = __byte_perm_S (    0, w0[0], selector);
+      w3[0] = 0;
+      w2[3] = 0;
+      w2[2] = 0;
+      w2[1] = 0;
+      w2[0] = 0;
+      w1[3] = 0;
+      w1[2] = 0;
+      w1[1] = 0;
+      w1[0] = 0;
+      w0[3] = 0;
+      w0[2] = 0;
+      w0[1] = 0;
+      w0[0] = 0;
+      break;
+
+    case 14:
+      w7[3] = __byte_perm_S (w4[0], w4[1], selector);
+      w7[2] = __byte_perm_S (w3[3], w4[0], selector);
+      w7[1] = __byte_perm_S (w3[2], w3[3], selector);
+      w7[0] = __byte_perm_S (w3[1], w3[2], selector);
+      w6[3] = __byte_perm_S (w3[0], w3[1], selector);
+      w6[2] = __byte_perm_S (w2[3], w3[0], selector);
+      w6[1] = __byte_perm_S (w2[2], w2[3], selector);
+      w6[0] = __byte_perm_S (w2[1], w2[2], selector);
+      w5[3] = __byte_perm_S (w2[0], w2[1], selector);
+      w5[2] = __byte_perm_S (w1[3], w2[0], selector);
+      w5[1] = __byte_perm_S (w1[2], w1[3], selector);
+      w5[0] = __byte_perm_S (w1[1], w1[2], selector);
+      w4[3] = __byte_perm_S (w1[0], w1[1], selector);
+      w4[2] = __byte_perm_S (w0[3], w1[0], selector);
+      w4[1] = __byte_perm_S (w0[2], w0[3], selector);
+      w4[0] = __byte_perm_S (w0[1], w0[2], selector);
+      w3[3] = __byte_perm_S (w0[0], w0[1], selector);
+      w3[2] = __byte_perm_S (    0, w0[0], selector);
+      w3[1] = 0;
+      w3[0] = 0;
+      w2[3] = 0;
+      w2[2] = 0;
+      w2[1] = 0;
+      w2[0] = 0;
+      w1[3] = 0;
+      w1[2] = 0;
+      w1[1] = 0;
+      w1[0] = 0;
+      w0[3] = 0;
+      w0[2] = 0;
+      w0[1] = 0;
+      w0[0] = 0;
+      break;
+
+    case 15:
+      w7[3] = __byte_perm_S (w3[3], w4[0], selector);
+      w7[2] = __byte_perm_S (w3[2], w3[3], selector);
+      w7[1] = __byte_perm_S (w3[1], w3[2], selector);
+      w7[0] = __byte_perm_S (w3[0], w3[1], selector);
+      w6[3] = __byte_perm_S (w2[3], w3[0], selector);
+      w6[2] = __byte_perm_S (w2[2], w2[3], selector);
+      w6[1] = __byte_perm_S (w2[1], w2[2], selector);
+      w6[0] = __byte_perm_S (w2[0], w2[1], selector);
+      w5[3] = __byte_perm_S (w1[3], w2[0], selector);
+      w5[2] = __byte_perm_S (w1[2], w1[3], selector);
+      w5[1] = __byte_perm_S (w1[1], w1[2], selector);
+      w5[0] = __byte_perm_S (w1[0], w1[1], selector);
+      w4[3] = __byte_perm_S (w0[3], w1[0], selector);
+      w4[2] = __byte_perm_S (w0[2], w0[3], selector);
+      w4[1] = __byte_perm_S (w0[1], w0[2], selector);
+      w4[0] = __byte_perm_S (w0[0], w0[1], selector);
+      w3[3] = __byte_perm_S (    0, w0[0], selector);
+      w3[2] = 0;
+      w3[1] = 0;
+      w3[0] = 0;
+      w2[3] = 0;
+      w2[2] = 0;
+      w2[1] = 0;
+      w2[0] = 0;
+      w1[3] = 0;
+      w1[2] = 0;
+      w1[1] = 0;
+      w1[0] = 0;
+      w0[3] = 0;
+      w0[2] = 0;
+      w0[1] = 0;
+      w0[0] = 0;
+      break;
+  }
+  #endif
+}
+
 inline void switch_buffer_by_offset_carry_le_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 c0[4], u32 c1[4], u32 c2[4], u32 c3[4], const u32 offset)
 {
   const int offset_mod_4 = offset & 3;
@@ -8862,6 +10475,26 @@ inline void switch_buffer_by_offset_carry_be_S (u32 w0[4], u32 w1[4], u32 w2[4],
   PACKSV4 (s2, v2, e);                      \
   PACKSV4 (s3, v3, e);
 
+#define PACKVS84(s0,s1,s2,s3,s4,s5,s6,s7,v0,v1,v2,v3,v4,v5,v6,v7,e) \
+  PACKVS4 (s0, v0, e);                                              \
+  PACKVS4 (s1, v1, e);                                              \
+  PACKVS4 (s2, v2, e);                                              \
+  PACKVS4 (s3, v3, e);                                              \
+  PACKVS4 (s4, v4, e);                                              \
+  PACKVS4 (s5, v5, e);                                              \
+  PACKVS4 (s6, v6, e);                                              \
+  PACKVS4 (s7, v7, e);
+
+#define PACKSV84(s0,s1,s2,s3,s4,s5,s6,s7,v0,v1,v2,v3,v4,v5,v6,v7,e) \
+  PACKSV4 (s0, v0, e);                                              \
+  PACKSV4 (s1, v1, e);                                              \
+  PACKSV4 (s2, v2, e);                                              \
+  PACKSV4 (s3, v3, e);                                              \
+  PACKSV4 (s4, v4, e);                                              \
+  PACKSV4 (s5, v5, e);                                              \
+  PACKSV4 (s6, v6, e);                                              \
+  PACKSV4 (s7, v7, e);
+
 inline void switch_buffer_by_offset_le_VV (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const u32x offset)
 {
   #if VECT_SIZE == 1
@@ -8922,6 +10555,186 @@ inline void switch_buffer_by_offset_le_VV (u32x w0[4], u32x w1[4], u32x w2[4], u
   #endif
 }
 
+inline void switch_buffer_by_offset_8x4_le_VV (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], const u32x offset)
+{
+  #if VECT_SIZE == 1
+
+  switch_buffer_by_offset_8x4_le_S (w0, w1, w2, w3, w4, w5, w6, w7, offset);
+
+  #else
+
+  u32 t0[4];
+  u32 t1[4];
+  u32 t2[4];
+  u32 t3[4];
+  u32 t4[4];
+  u32 t5[4];
+  u32 t6[4];
+  u32 t7[4];
+
+  #endif
+
+  #if   VECT_SIZE == 2
+
+  // 1
+  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 0);
+  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s0);
+  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 0);
+
+  // 2
+  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 1);
+  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s1);
+  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 1);
+
+  #elif VECT_SIZE == 4
+
+  // 1
+  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 0);
+  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s0);
+  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 0);
+
+  // 2
+  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 1);
+  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s1);
+  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 1);
+
+  // 3
+  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 2);
+  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s2);
+  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 2);
+
+  // 4
+  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 3);
+  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s3);
+  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 3);
+
+  #elif VECT_SIZE == 8
+
+  // 1
+  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 0);
+  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s0);
+  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 0);
+
+  // 2
+  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 1);
+  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s1);
+  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 1);
+
+  // 3
+  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 2);
+  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s2);
+  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 2);
+
+  // 4
+  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 3);
+  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s3);
+  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 3);
+
+  // 5
+  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 4);
+  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s4);
+  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 4);
+
+  // 6
+  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 5);
+  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s5);
+  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 5);
+
+  // 7
+  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 6);
+  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s6);
+  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 6);
+
+  // 8
+  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 7);
+  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s7);
+  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 7);
+
+  #elif VECT_SIZE == 16
+
+  // 1
+  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 0);
+  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s0);
+  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 0);
+
+  // 2
+  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 1);
+  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s1);
+  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 1);
+
+  // 3
+  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 2);
+  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s2);
+  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 2);
+
+  // 4
+  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 3);
+  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s3);
+  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 3);
+
+  // 5
+  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 4);
+  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s4);
+  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 4);
+
+  // 6
+  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 5);
+  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s5);
+  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 5);
+
+  // 7
+  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 6);
+  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s6);
+  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 6);
+
+  // 8
+  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 7);
+  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s7);
+  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 7);
+
+  // 9
+  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 8);
+  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s8);
+  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 8);
+
+  // 10
+  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 9);
+  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s9);
+  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 9);
+
+  // 11
+  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, a);
+  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.sa);
+  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, a);
+
+  // 12
+  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, b);
+  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.sb);
+  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, b);
+
+  // 13
+  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, c);
+  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.sc);
+  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, c);
+
+  // 14
+  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, d);
+  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.sd);
+  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, d);
+
+  // 15
+  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, e);
+  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.se);
+  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, e);
+
+  // 16
+  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, f);
+  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.sf);
+  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, f);
+
+  #endif
+}
+
 inline void append_0x01_2x4_VV (u32x w0[4], u32x w1[4], const u32x offset)
 {
   #if VECT_SIZE == 1
diff --git a/OpenCL/m15000_a0.cl b/OpenCL/m15000_a0.cl
index 25b75e40a..35a95f9ec 100644
--- a/OpenCL/m15000_a0.cl
+++ b/OpenCL/m15000_a0.cl
@@ -16,1819 +16,6 @@
 #include "inc_rp.cl"
 #include "inc_simd.cl"
 
-inline void switch_buffer_by_offset_8x4_le_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 w4[4], u32 w5[4], u32 w6[4], u32 w7[4], const u32 offset)
-{
-  #if defined IS_AMD || defined IS_GENERIC
-  const int offset_mod_4 = offset & 3;
-
-  const int offset_minus_4 = 4 - offset;
-
-  switch (offset / 4)
-  {
-    case 0:
-      w7[3] = amd_bytealign_S (w7[3], w7[2], offset_minus_4);
-      w7[2] = amd_bytealign_S (w7[2], w7[1], offset_minus_4);
-      w7[1] = amd_bytealign_S (w7[1], w7[0], offset_minus_4);
-      w7[0] = amd_bytealign_S (w7[0], w6[3], offset_minus_4);
-      w6[3] = amd_bytealign_S (w6[3], w6[2], offset_minus_4);
-      w6[2] = amd_bytealign_S (w6[2], w6[1], offset_minus_4);
-      w6[1] = amd_bytealign_S (w6[1], w6[0], offset_minus_4);
-      w6[0] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
-      w5[3] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
-      w5[2] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
-      w5[1] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w5[0] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w4[3] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w4[2] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w4[1] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w4[0] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w3[3] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w3[2] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w3[1] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w3[0] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w2[3] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w2[2] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w2[1] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w2[0] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w1[3] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w1[2] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w1[1] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w1[0] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w0[3] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w0[2] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w0[1] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w0[0] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-
-      if (offset_mod_4 == 0)
-      {
-        w0[0] = w0[1];
-        w0[1] = w0[2];
-        w0[2] = w0[3];
-        w0[3] = w1[0];
-        w1[0] = w1[1];
-        w1[1] = w1[2];
-        w1[2] = w1[3];
-        w1[3] = w2[0];
-        w2[0] = w2[1];
-        w2[1] = w2[2];
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 1:
-      w7[3] = amd_bytealign_S (w7[2], w7[1], offset_minus_4);
-      w7[2] = amd_bytealign_S (w7[1], w7[0], offset_minus_4);
-      w7[1] = amd_bytealign_S (w7[0], w6[3], offset_minus_4);
-      w7[0] = amd_bytealign_S (w6[3], w6[2], offset_minus_4);
-      w6[3] = amd_bytealign_S (w6[2], w6[1], offset_minus_4);
-      w6[2] = amd_bytealign_S (w6[1], w6[0], offset_minus_4);
-      w6[1] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
-      w6[0] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
-      w5[3] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
-      w5[2] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w5[1] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w5[0] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w4[3] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w4[2] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w4[1] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w4[0] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w3[3] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w3[2] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w3[1] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w3[0] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w2[3] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w2[2] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w2[1] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w2[0] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w1[3] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w1[2] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w1[1] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w1[0] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w0[3] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w0[2] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w0[1] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w0[1] = w0[2];
-        w0[2] = w0[3];
-        w0[3] = w1[0];
-        w1[0] = w1[1];
-        w1[1] = w1[2];
-        w1[2] = w1[3];
-        w1[3] = w2[0];
-        w2[0] = w2[1];
-        w2[1] = w2[2];
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 2:
-      w7[3] = amd_bytealign_S (w7[1], w7[0], offset_minus_4);
-      w7[2] = amd_bytealign_S (w7[0], w6[3], offset_minus_4);
-      w7[1] = amd_bytealign_S (w6[3], w6[2], offset_minus_4);
-      w7[0] = amd_bytealign_S (w6[2], w6[1], offset_minus_4);
-      w6[3] = amd_bytealign_S (w6[1], w6[0], offset_minus_4);
-      w6[2] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
-      w6[1] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
-      w6[0] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
-      w5[3] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w5[2] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w5[1] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w5[0] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w4[3] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w4[2] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w4[1] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w4[0] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w3[3] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w3[2] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w3[1] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w3[0] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w2[3] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w2[2] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w2[1] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w2[0] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w1[3] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w1[2] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w1[1] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w1[0] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w0[3] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w0[2] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w0[2] = w0[3];
-        w0[3] = w1[0];
-        w1[0] = w1[1];
-        w1[1] = w1[2];
-        w1[2] = w1[3];
-        w1[3] = w2[0];
-        w2[0] = w2[1];
-        w2[1] = w2[2];
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 3:
-      w7[3] = amd_bytealign_S (w7[0], w6[3], offset_minus_4);
-      w7[2] = amd_bytealign_S (w6[3], w6[2], offset_minus_4);
-      w7[1] = amd_bytealign_S (w6[2], w6[1], offset_minus_4);
-      w7[0] = amd_bytealign_S (w6[1], w6[0], offset_minus_4);
-      w6[3] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
-      w6[2] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
-      w6[1] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
-      w6[0] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w5[3] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w5[2] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w5[1] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w5[0] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w4[3] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w4[2] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w4[1] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w4[0] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w3[3] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w3[2] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w3[1] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w3[0] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w2[3] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w2[2] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w2[1] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w2[0] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w1[3] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w1[2] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w1[1] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w1[0] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w0[3] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w0[3] = w1[0];
-        w1[0] = w1[1];
-        w1[1] = w1[2];
-        w1[2] = w1[3];
-        w1[3] = w2[0];
-        w2[0] = w2[1];
-        w2[1] = w2[2];
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 4:
-      w7[3] = amd_bytealign_S (w6[3], w6[2], offset_minus_4);
-      w7[2] = amd_bytealign_S (w6[2], w6[1], offset_minus_4);
-      w7[1] = amd_bytealign_S (w6[1], w6[0], offset_minus_4);
-      w7[0] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
-      w6[3] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
-      w6[2] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
-      w6[1] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w6[0] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w5[3] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w5[2] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w5[1] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w5[0] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w4[3] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w4[2] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w4[1] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w4[0] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w3[3] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w3[2] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w3[1] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w3[0] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w2[3] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w2[2] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w2[1] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w2[0] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w1[3] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w1[2] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w1[1] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w1[0] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w1[0] = w1[1];
-        w1[1] = w1[2];
-        w1[2] = w1[3];
-        w1[3] = w2[0];
-        w2[0] = w2[1];
-        w2[1] = w2[2];
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 5:
-      w7[3] = amd_bytealign_S (w6[2], w6[1], offset_minus_4);
-      w7[2] = amd_bytealign_S (w6[1], w6[0], offset_minus_4);
-      w7[1] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
-      w7[0] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
-      w6[3] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
-      w6[2] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w6[1] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w6[0] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w5[3] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w5[2] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w5[1] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w5[0] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w4[3] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w4[2] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w4[1] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w4[0] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w3[3] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w3[2] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w3[1] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w3[0] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w2[3] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w2[2] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w2[1] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w2[0] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w1[3] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w1[2] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w1[1] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w1[1] = w1[2];
-        w1[2] = w1[3];
-        w1[3] = w2[0];
-        w2[0] = w2[1];
-        w2[1] = w2[2];
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 6:
-      w7[3] = amd_bytealign_S (w6[1], w6[0], offset_minus_4);
-      w7[2] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
-      w7[1] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
-      w7[0] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
-      w6[3] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w6[2] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w6[1] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w6[0] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w5[3] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w5[2] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w5[1] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w5[0] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w4[3] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w4[2] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w4[1] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w4[0] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w3[3] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w3[2] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w3[1] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w3[0] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w2[3] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w2[2] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w2[1] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w2[0] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w1[3] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w1[2] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w1[2] = w1[3];
-        w1[3] = w2[0];
-        w2[0] = w2[1];
-        w2[1] = w2[2];
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 7:
-      w7[3] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
-      w7[2] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
-      w7[1] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
-      w7[0] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w6[3] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w6[2] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w6[1] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w6[0] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w5[3] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w5[2] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w5[1] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w5[0] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w4[3] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w4[2] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w4[1] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w4[0] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w3[3] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w3[2] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w3[1] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w3[0] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w2[3] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w2[2] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w2[1] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w2[0] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w1[3] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w1[3] = w2[0];
-        w2[0] = w2[1];
-        w2[1] = w2[2];
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 8:
-      w7[3] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
-      w7[2] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
-      w7[1] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w7[0] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w6[3] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w6[2] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w6[1] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w6[0] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w5[3] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w5[2] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w5[1] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w5[0] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w4[3] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w4[2] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w4[1] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w4[0] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w3[3] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w3[2] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w3[1] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w3[0] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w2[3] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w2[2] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w2[1] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w2[0] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w2[0] = w2[1];
-        w2[1] = w2[2];
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 9:
-      w7[3] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
-      w7[2] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w7[1] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w7[0] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w6[3] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w6[2] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w6[1] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w6[0] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w5[3] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w5[2] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w5[1] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w5[0] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w4[3] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w4[2] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w4[1] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w4[0] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w3[3] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w3[2] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w3[1] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w3[0] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w2[3] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w2[2] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w2[1] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w2[1] = w2[2];
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 10:
-      w7[3] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w7[2] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w7[1] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w7[0] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w6[3] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w6[2] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w6[1] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w6[0] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w5[3] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w5[2] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w5[1] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w5[0] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w4[3] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w4[2] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w4[1] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w4[0] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w3[3] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w3[2] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w3[1] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w3[0] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w2[3] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w2[2] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 11:
-      w7[3] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w7[2] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w7[1] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w7[0] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w6[3] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w6[2] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w6[1] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w6[0] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w5[3] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w5[2] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w5[1] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w5[0] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w4[3] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w4[2] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w4[1] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w4[0] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w3[3] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w3[2] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w3[1] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w3[0] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w2[3] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w2[2] = 0;
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 12:
-      w7[3] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w7[2] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w7[1] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w7[0] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w6[3] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w6[2] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w6[1] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w6[0] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w5[3] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w5[2] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w5[1] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w5[0] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w4[3] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w4[2] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w4[1] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w4[0] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w3[3] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w3[2] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w3[1] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w3[0] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w2[3] = 0;
-      w2[2] = 0;
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 13:
-      w7[3] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w7[2] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w7[1] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w7[0] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w6[3] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w6[2] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w6[1] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w6[0] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w5[3] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w5[2] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w5[1] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w5[0] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w4[3] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w4[2] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w4[1] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w4[0] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w3[3] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w3[2] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w3[1] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w3[0] = 0;
-      w2[3] = 0;
-      w2[2] = 0;
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 14:
-      w7[3] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w7[2] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w7[1] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w7[0] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w6[3] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w6[2] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w6[1] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w6[0] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w5[3] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w5[2] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w5[1] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w5[0] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w4[3] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w4[2] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w4[1] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w4[0] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w3[3] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w3[2] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w3[1] = 0;
-      w3[0] = 0;
-      w2[3] = 0;
-      w2[2] = 0;
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 15:
-      w7[3] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w7[2] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w7[1] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w7[0] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w6[3] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w6[2] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w6[1] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w6[0] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w5[3] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w5[2] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w5[1] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w5[0] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w4[3] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w4[2] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w4[1] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w4[0] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w3[3] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w3[2] = 0;
-      w3[1] = 0;
-      w3[0] = 0;
-      w2[3] = 0;
-      w2[2] = 0;
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-  }
-  #endif
-
-  #ifdef IS_NV
-  const int offset_minus_4 = 4 - (offset % 4);
-
-  const int selector = (0x76543210 >> (offset_minus_4 * 4)) & 0xffff;
-
-  switch (offset / 4)
-  {
-    case 0:
-      w7[3] = __byte_perm_S (w7[2], w7[3], selector);
-      w7[2] = __byte_perm_S (w7[1], w7[2], selector);
-      w7[1] = __byte_perm_S (w7[0], w7[1], selector);
-      w7[0] = __byte_perm_S (w6[3], w7[0], selector);
-      w6[3] = __byte_perm_S (w6[2], w6[3], selector);
-      w6[2] = __byte_perm_S (w6[1], w6[2], selector);
-      w6[1] = __byte_perm_S (w6[0], w6[1], selector);
-      w6[0] = __byte_perm_S (w5[3], w6[0], selector);
-      w5[3] = __byte_perm_S (w5[2], w5[3], selector);
-      w5[2] = __byte_perm_S (w5[1], w5[2], selector);
-      w5[1] = __byte_perm_S (w5[0], w5[1], selector);
-      w5[0] = __byte_perm_S (w4[3], w5[0], selector);
-      w4[3] = __byte_perm_S (w4[2], w4[3], selector);
-      w4[2] = __byte_perm_S (w4[1], w4[2], selector);
-      w4[1] = __byte_perm_S (w4[0], w4[1], selector);
-      w4[0] = __byte_perm_S (w3[3], w4[0], selector);
-      w3[3] = __byte_perm_S (w3[2], w3[3], selector);
-      w3[2] = __byte_perm_S (w3[1], w3[2], selector);
-      w3[1] = __byte_perm_S (w3[0], w3[1], selector);
-      w3[0] = __byte_perm_S (w2[3], w3[0], selector);
-      w2[3] = __byte_perm_S (w2[2], w2[3], selector);
-      w2[2] = __byte_perm_S (w2[1], w2[2], selector);
-      w2[1] = __byte_perm_S (w2[0], w2[1], selector);
-      w2[0] = __byte_perm_S (w1[3], w2[0], selector);
-      w1[3] = __byte_perm_S (w1[2], w1[3], selector);
-      w1[2] = __byte_perm_S (w1[1], w1[2], selector);
-      w1[1] = __byte_perm_S (w1[0], w1[1], selector);
-      w1[0] = __byte_perm_S (w0[3], w1[0], selector);
-      w0[3] = __byte_perm_S (w0[2], w0[3], selector);
-      w0[2] = __byte_perm_S (w0[1], w0[2], selector);
-      w0[1] = __byte_perm_S (w0[0], w0[1], selector);
-      w0[0] = __byte_perm_S (    0, w0[0], selector);
-      break;
-
-    case 1:
-      w7[3] = __byte_perm_S (w7[1], w7[2], selector);
-      w7[2] = __byte_perm_S (w7[0], w7[1], selector);
-      w7[1] = __byte_perm_S (w6[3], w7[0], selector);
-      w7[0] = __byte_perm_S (w6[2], w6[3], selector);
-      w6[3] = __byte_perm_S (w6[1], w6[2], selector);
-      w6[2] = __byte_perm_S (w6[0], w6[1], selector);
-      w6[1] = __byte_perm_S (w5[3], w6[0], selector);
-      w6[0] = __byte_perm_S (w5[2], w5[3], selector);
-      w5[3] = __byte_perm_S (w5[1], w5[2], selector);
-      w5[2] = __byte_perm_S (w5[0], w5[1], selector);
-      w5[1] = __byte_perm_S (w4[3], w5[0], selector);
-      w5[0] = __byte_perm_S (w4[2], w4[3], selector);
-      w4[3] = __byte_perm_S (w4[1], w4[2], selector);
-      w4[2] = __byte_perm_S (w4[0], w4[1], selector);
-      w4[1] = __byte_perm_S (w3[3], w4[0], selector);
-      w4[0] = __byte_perm_S (w3[2], w3[3], selector);
-      w3[3] = __byte_perm_S (w3[1], w3[2], selector);
-      w3[2] = __byte_perm_S (w3[0], w3[1], selector);
-      w3[1] = __byte_perm_S (w2[3], w3[0], selector);
-      w3[0] = __byte_perm_S (w2[2], w2[3], selector);
-      w2[3] = __byte_perm_S (w2[1], w2[2], selector);
-      w2[2] = __byte_perm_S (w2[0], w2[1], selector);
-      w2[1] = __byte_perm_S (w1[3], w2[0], selector);
-      w2[0] = __byte_perm_S (w1[2], w1[3], selector);
-      w1[3] = __byte_perm_S (w1[1], w1[2], selector);
-      w1[2] = __byte_perm_S (w1[0], w1[1], selector);
-      w1[1] = __byte_perm_S (w0[3], w1[0], selector);
-      w1[0] = __byte_perm_S (w0[2], w0[3], selector);
-      w0[3] = __byte_perm_S (w0[1], w0[2], selector);
-      w0[2] = __byte_perm_S (w0[0], w0[1], selector);
-      w0[1] = __byte_perm_S (    0, w0[0], selector);
-      w0[0] = 0;
-      break;
-
-    case 2:
-      w7[3] = __byte_perm_S (w7[0], w7[1], selector);
-      w7[2] = __byte_perm_S (w6[3], w7[0], selector);
-      w7[1] = __byte_perm_S (w6[2], w6[3], selector);
-      w7[0] = __byte_perm_S (w6[1], w6[2], selector);
-      w6[3] = __byte_perm_S (w6[0], w6[1], selector);
-      w6[2] = __byte_perm_S (w5[3], w6[0], selector);
-      w6[1] = __byte_perm_S (w5[2], w5[3], selector);
-      w6[0] = __byte_perm_S (w5[1], w5[2], selector);
-      w5[3] = __byte_perm_S (w5[0], w5[1], selector);
-      w5[2] = __byte_perm_S (w4[3], w5[0], selector);
-      w5[1] = __byte_perm_S (w4[2], w4[3], selector);
-      w5[0] = __byte_perm_S (w4[1], w4[2], selector);
-      w4[3] = __byte_perm_S (w4[0], w4[1], selector);
-      w4[2] = __byte_perm_S (w3[3], w4[0], selector);
-      w4[1] = __byte_perm_S (w3[2], w3[3], selector);
-      w4[0] = __byte_perm_S (w3[1], w3[2], selector);
-      w3[3] = __byte_perm_S (w3[0], w3[1], selector);
-      w3[2] = __byte_perm_S (w2[3], w3[0], selector);
-      w3[1] = __byte_perm_S (w2[2], w2[3], selector);
-      w3[0] = __byte_perm_S (w2[1], w2[2], selector);
-      w2[3] = __byte_perm_S (w2[0], w2[1], selector);
-      w2[2] = __byte_perm_S (w1[3], w2[0], selector);
-      w2[1] = __byte_perm_S (w1[2], w1[3], selector);
-      w2[0] = __byte_perm_S (w1[1], w1[2], selector);
-      w1[3] = __byte_perm_S (w1[0], w1[1], selector);
-      w1[2] = __byte_perm_S (w0[3], w1[0], selector);
-      w1[1] = __byte_perm_S (w0[2], w0[3], selector);
-      w1[0] = __byte_perm_S (w0[1], w0[2], selector);
-      w0[3] = __byte_perm_S (w0[0], w0[1], selector);
-      w0[2] = __byte_perm_S (    0, w0[0], selector);
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 3:
-      w7[3] = __byte_perm_S (w6[3], w7[0], selector);
-      w7[2] = __byte_perm_S (w6[2], w6[3], selector);
-      w7[1] = __byte_perm_S (w6[1], w6[2], selector);
-      w7[0] = __byte_perm_S (w6[0], w6[1], selector);
-      w6[3] = __byte_perm_S (w5[3], w6[0], selector);
-      w6[2] = __byte_perm_S (w5[2], w5[3], selector);
-      w6[1] = __byte_perm_S (w5[1], w5[2], selector);
-      w6[0] = __byte_perm_S (w5[0], w5[1], selector);
-      w5[3] = __byte_perm_S (w4[3], w5[0], selector);
-      w5[2] = __byte_perm_S (w4[2], w4[3], selector);
-      w5[1] = __byte_perm_S (w4[1], w4[2], selector);
-      w5[0] = __byte_perm_S (w4[0], w4[1], selector);
-      w4[3] = __byte_perm_S (w3[3], w4[0], selector);
-      w4[2] = __byte_perm_S (w3[2], w3[3], selector);
-      w4[1] = __byte_perm_S (w3[1], w3[2], selector);
-      w4[0] = __byte_perm_S (w3[0], w3[1], selector);
-      w3[3] = __byte_perm_S (w2[3], w3[0], selector);
-      w3[2] = __byte_perm_S (w2[2], w2[3], selector);
-      w3[1] = __byte_perm_S (w2[1], w2[2], selector);
-      w3[0] = __byte_perm_S (w2[0], w2[1], selector);
-      w2[3] = __byte_perm_S (w1[3], w2[0], selector);
-      w2[2] = __byte_perm_S (w1[2], w1[3], selector);
-      w2[1] = __byte_perm_S (w1[1], w1[2], selector);
-      w2[0] = __byte_perm_S (w1[0], w1[1], selector);
-      w1[3] = __byte_perm_S (w0[3], w1[0], selector);
-      w1[2] = __byte_perm_S (w0[2], w0[3], selector);
-      w1[1] = __byte_perm_S (w0[1], w0[2], selector);
-      w1[0] = __byte_perm_S (w0[0], w0[1], selector);
-      w0[3] = __byte_perm_S (    0, w0[0], selector);
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 4:
-      w7[3] = __byte_perm_S (w6[2], w6[3], selector);
-      w7[2] = __byte_perm_S (w6[1], w6[2], selector);
-      w7[1] = __byte_perm_S (w6[0], w6[1], selector);
-      w7[0] = __byte_perm_S (w5[3], w6[0], selector);
-      w6[3] = __byte_perm_S (w5[2], w5[3], selector);
-      w6[2] = __byte_perm_S (w5[1], w5[2], selector);
-      w6[1] = __byte_perm_S (w5[0], w5[1], selector);
-      w6[0] = __byte_perm_S (w4[3], w5[0], selector);
-      w5[3] = __byte_perm_S (w4[2], w4[3], selector);
-      w5[2] = __byte_perm_S (w4[1], w4[2], selector);
-      w5[1] = __byte_perm_S (w4[0], w4[1], selector);
-      w5[0] = __byte_perm_S (w3[3], w4[0], selector);
-      w4[3] = __byte_perm_S (w3[2], w3[3], selector);
-      w4[2] = __byte_perm_S (w3[1], w3[2], selector);
-      w4[1] = __byte_perm_S (w3[0], w3[1], selector);
-      w4[0] = __byte_perm_S (w2[3], w3[0], selector);
-      w3[3] = __byte_perm_S (w2[2], w2[3], selector);
-      w3[2] = __byte_perm_S (w2[1], w2[2], selector);
-      w3[1] = __byte_perm_S (w2[0], w2[1], selector);
-      w3[0] = __byte_perm_S (w1[3], w2[0], selector);
-      w2[3] = __byte_perm_S (w1[2], w1[3], selector);
-      w2[2] = __byte_perm_S (w1[1], w1[2], selector);
-      w2[1] = __byte_perm_S (w1[0], w1[1], selector);
-      w2[0] = __byte_perm_S (w0[3], w1[0], selector);
-      w1[3] = __byte_perm_S (w0[2], w0[3], selector);
-      w1[2] = __byte_perm_S (w0[1], w0[2], selector);
-      w1[1] = __byte_perm_S (w0[0], w0[1], selector);
-      w1[0] = __byte_perm_S (    0, w0[0], selector);
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 5:
-      w7[3] = __byte_perm_S (w6[1], w6[2], selector);
-      w7[2] = __byte_perm_S (w6[0], w6[1], selector);
-      w7[1] = __byte_perm_S (w5[3], w6[0], selector);
-      w7[0] = __byte_perm_S (w5[2], w5[3], selector);
-      w6[3] = __byte_perm_S (w5[1], w5[2], selector);
-      w6[2] = __byte_perm_S (w5[0], w5[1], selector);
-      w6[1] = __byte_perm_S (w4[3], w5[0], selector);
-      w6[0] = __byte_perm_S (w4[2], w4[3], selector);
-      w5[3] = __byte_perm_S (w4[1], w4[2], selector);
-      w5[2] = __byte_perm_S (w4[0], w4[1], selector);
-      w5[1] = __byte_perm_S (w3[3], w4[0], selector);
-      w5[0] = __byte_perm_S (w3[2], w3[3], selector);
-      w4[3] = __byte_perm_S (w3[1], w3[2], selector);
-      w4[2] = __byte_perm_S (w3[0], w3[1], selector);
-      w4[1] = __byte_perm_S (w2[3], w3[0], selector);
-      w4[0] = __byte_perm_S (w2[2], w2[3], selector);
-      w3[3] = __byte_perm_S (w2[1], w2[2], selector);
-      w3[2] = __byte_perm_S (w2[0], w2[1], selector);
-      w3[1] = __byte_perm_S (w1[3], w2[0], selector);
-      w3[0] = __byte_perm_S (w1[2], w1[3], selector);
-      w2[3] = __byte_perm_S (w1[1], w1[2], selector);
-      w2[2] = __byte_perm_S (w1[0], w1[1], selector);
-      w2[1] = __byte_perm_S (w0[3], w1[0], selector);
-      w2[0] = __byte_perm_S (w0[2], w0[3], selector);
-      w1[3] = __byte_perm_S (w0[1], w0[2], selector);
-      w1[2] = __byte_perm_S (w0[0], w0[1], selector);
-      w1[1] = __byte_perm_S (    0, w0[0], selector);
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 6:
-      w7[3] = __byte_perm_S (w6[0], w6[1], selector);
-      w7[2] = __byte_perm_S (w5[3], w6[0], selector);
-      w7[1] = __byte_perm_S (w5[2], w5[3], selector);
-      w7[0] = __byte_perm_S (w5[1], w5[2], selector);
-      w6[3] = __byte_perm_S (w5[0], w5[1], selector);
-      w6[2] = __byte_perm_S (w4[3], w5[0], selector);
-      w6[1] = __byte_perm_S (w4[2], w4[3], selector);
-      w6[0] = __byte_perm_S (w4[1], w4[2], selector);
-      w5[3] = __byte_perm_S (w4[0], w4[1], selector);
-      w5[2] = __byte_perm_S (w3[3], w4[0], selector);
-      w5[1] = __byte_perm_S (w3[2], w3[3], selector);
-      w5[0] = __byte_perm_S (w3[1], w3[2], selector);
-      w4[3] = __byte_perm_S (w3[0], w3[1], selector);
-      w4[2] = __byte_perm_S (w2[3], w3[0], selector);
-      w4[1] = __byte_perm_S (w2[2], w2[3], selector);
-      w4[0] = __byte_perm_S (w2[1], w2[2], selector);
-      w3[3] = __byte_perm_S (w2[0], w2[1], selector);
-      w3[2] = __byte_perm_S (w1[3], w2[0], selector);
-      w3[1] = __byte_perm_S (w1[2], w1[3], selector);
-      w3[0] = __byte_perm_S (w1[1], w1[2], selector);
-      w2[3] = __byte_perm_S (w1[0], w1[1], selector);
-      w2[2] = __byte_perm_S (w0[3], w1[0], selector);
-      w2[1] = __byte_perm_S (w0[2], w0[3], selector);
-      w2[0] = __byte_perm_S (w0[1], w0[2], selector);
-      w1[3] = __byte_perm_S (w0[0], w0[1], selector);
-      w1[2] = __byte_perm_S (    0, w0[0], selector);
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 7:
-      w7[3] = __byte_perm_S (w5[3], w6[0], selector);
-      w7[2] = __byte_perm_S (w5[2], w5[3], selector);
-      w7[1] = __byte_perm_S (w5[1], w5[2], selector);
-      w7[0] = __byte_perm_S (w5[0], w5[1], selector);
-      w6[3] = __byte_perm_S (w4[3], w5[0], selector);
-      w6[2] = __byte_perm_S (w4[2], w4[3], selector);
-      w6[1] = __byte_perm_S (w4[1], w4[2], selector);
-      w6[0] = __byte_perm_S (w4[0], w4[1], selector);
-      w5[3] = __byte_perm_S (w3[3], w4[0], selector);
-      w5[2] = __byte_perm_S (w3[2], w3[3], selector);
-      w5[1] = __byte_perm_S (w3[1], w3[2], selector);
-      w5[0] = __byte_perm_S (w3[0], w3[1], selector);
-      w4[3] = __byte_perm_S (w2[3], w3[0], selector);
-      w4[2] = __byte_perm_S (w2[2], w2[3], selector);
-      w4[1] = __byte_perm_S (w2[1], w2[2], selector);
-      w4[0] = __byte_perm_S (w2[0], w2[1], selector);
-      w3[3] = __byte_perm_S (w1[3], w2[0], selector);
-      w3[2] = __byte_perm_S (w1[2], w1[3], selector);
-      w3[1] = __byte_perm_S (w1[1], w1[2], selector);
-      w3[0] = __byte_perm_S (w1[0], w1[1], selector);
-      w2[3] = __byte_perm_S (w0[3], w1[0], selector);
-      w2[2] = __byte_perm_S (w0[2], w0[3], selector);
-      w2[1] = __byte_perm_S (w0[1], w0[2], selector);
-      w2[0] = __byte_perm_S (w0[0], w0[1], selector);
-      w1[3] = __byte_perm_S (    0, w0[0], selector);
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 8:
-      w7[3] = __byte_perm_S (w5[2], w5[3], selector);
-      w7[2] = __byte_perm_S (w5[1], w5[2], selector);
-      w7[1] = __byte_perm_S (w5[0], w5[1], selector);
-      w7[0] = __byte_perm_S (w4[3], w5[0], selector);
-      w6[3] = __byte_perm_S (w4[2], w4[3], selector);
-      w6[2] = __byte_perm_S (w4[1], w4[2], selector);
-      w6[1] = __byte_perm_S (w4[0], w4[1], selector);
-      w6[0] = __byte_perm_S (w3[3], w4[0], selector);
-      w5[3] = __byte_perm_S (w3[2], w3[3], selector);
-      w5[2] = __byte_perm_S (w3[1], w3[2], selector);
-      w5[1] = __byte_perm_S (w3[0], w3[1], selector);
-      w5[0] = __byte_perm_S (w2[3], w3[0], selector);
-      w4[3] = __byte_perm_S (w2[2], w2[3], selector);
-      w4[2] = __byte_perm_S (w2[1], w2[2], selector);
-      w4[1] = __byte_perm_S (w2[0], w2[1], selector);
-      w4[0] = __byte_perm_S (w1[3], w2[0], selector);
-      w3[3] = __byte_perm_S (w1[2], w1[3], selector);
-      w3[2] = __byte_perm_S (w1[1], w1[2], selector);
-      w3[1] = __byte_perm_S (w1[0], w1[1], selector);
-      w3[0] = __byte_perm_S (w0[3], w1[0], selector);
-      w2[3] = __byte_perm_S (w0[2], w0[3], selector);
-      w2[2] = __byte_perm_S (w0[1], w0[2], selector);
-      w2[1] = __byte_perm_S (w0[0], w0[1], selector);
-      w2[0] = __byte_perm_S (    0, w0[0], selector);
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 9:
-      w7[3] = __byte_perm_S (w5[1], w5[2], selector);
-      w7[2] = __byte_perm_S (w5[0], w5[1], selector);
-      w7[1] = __byte_perm_S (w4[3], w5[0], selector);
-      w7[0] = __byte_perm_S (w4[2], w4[3], selector);
-      w6[3] = __byte_perm_S (w4[1], w4[2], selector);
-      w6[2] = __byte_perm_S (w4[0], w4[1], selector);
-      w6[1] = __byte_perm_S (w3[3], w4[0], selector);
-      w6[0] = __byte_perm_S (w3[2], w3[3], selector);
-      w5[3] = __byte_perm_S (w3[1], w3[2], selector);
-      w5[2] = __byte_perm_S (w3[0], w3[1], selector);
-      w5[1] = __byte_perm_S (w2[3], w3[0], selector);
-      w5[0] = __byte_perm_S (w2[2], w2[3], selector);
-      w4[3] = __byte_perm_S (w2[1], w2[2], selector);
-      w4[2] = __byte_perm_S (w2[0], w2[1], selector);
-      w4[1] = __byte_perm_S (w1[3], w2[0], selector);
-      w4[0] = __byte_perm_S (w1[2], w1[3], selector);
-      w3[3] = __byte_perm_S (w1[1], w1[2], selector);
-      w3[2] = __byte_perm_S (w1[0], w1[1], selector);
-      w3[1] = __byte_perm_S (w0[3], w1[0], selector);
-      w3[0] = __byte_perm_S (w0[2], w0[3], selector);
-      w2[3] = __byte_perm_S (w0[1], w0[2], selector);
-      w2[2] = __byte_perm_S (w0[0], w0[1], selector);
-      w2[1] = __byte_perm_S (    0, w0[0], selector);
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 10:
-      w7[3] = __byte_perm_S (w5[0], w5[1], selector);
-      w7[2] = __byte_perm_S (w4[3], w5[0], selector);
-      w7[1] = __byte_perm_S (w4[2], w4[3], selector);
-      w7[0] = __byte_perm_S (w4[1], w4[2], selector);
-      w6[3] = __byte_perm_S (w4[0], w4[1], selector);
-      w6[2] = __byte_perm_S (w3[3], w4[0], selector);
-      w6[1] = __byte_perm_S (w3[2], w3[3], selector);
-      w6[0] = __byte_perm_S (w3[1], w3[2], selector);
-      w5[3] = __byte_perm_S (w3[0], w3[1], selector);
-      w5[2] = __byte_perm_S (w2[3], w3[0], selector);
-      w5[1] = __byte_perm_S (w2[2], w2[3], selector);
-      w5[0] = __byte_perm_S (w2[1], w2[2], selector);
-      w4[3] = __byte_perm_S (w2[0], w2[1], selector);
-      w4[2] = __byte_perm_S (w1[3], w2[0], selector);
-      w4[1] = __byte_perm_S (w1[2], w1[3], selector);
-      w4[0] = __byte_perm_S (w1[1], w1[2], selector);
-      w3[3] = __byte_perm_S (w1[0], w1[1], selector);
-      w3[2] = __byte_perm_S (w0[3], w1[0], selector);
-      w3[1] = __byte_perm_S (w0[2], w0[3], selector);
-      w3[0] = __byte_perm_S (w0[1], w0[2], selector);
-      w2[3] = __byte_perm_S (w0[0], w0[1], selector);
-      w2[2] = __byte_perm_S (    0, w0[0], selector);
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 11:
-      w7[3] = __byte_perm_S (w4[3], w5[0], selector);
-      w7[2] = __byte_perm_S (w4[2], w4[3], selector);
-      w7[1] = __byte_perm_S (w4[1], w4[2], selector);
-      w7[0] = __byte_perm_S (w4[0], w4[1], selector);
-      w6[3] = __byte_perm_S (w3[3], w4[0], selector);
-      w6[2] = __byte_perm_S (w3[2], w3[3], selector);
-      w6[1] = __byte_perm_S (w3[1], w3[2], selector);
-      w6[0] = __byte_perm_S (w3[0], w3[1], selector);
-      w5[3] = __byte_perm_S (w2[3], w3[0], selector);
-      w5[2] = __byte_perm_S (w2[2], w2[3], selector);
-      w5[1] = __byte_perm_S (w2[1], w2[2], selector);
-      w5[0] = __byte_perm_S (w2[0], w2[1], selector);
-      w4[3] = __byte_perm_S (w1[3], w2[0], selector);
-      w4[2] = __byte_perm_S (w1[2], w1[3], selector);
-      w4[1] = __byte_perm_S (w1[1], w1[2], selector);
-      w4[0] = __byte_perm_S (w1[0], w1[1], selector);
-      w3[3] = __byte_perm_S (w0[3], w1[0], selector);
-      w3[2] = __byte_perm_S (w0[2], w0[3], selector);
-      w3[1] = __byte_perm_S (w0[1], w0[2], selector);
-      w3[0] = __byte_perm_S (w0[0], w0[1], selector);
-      w2[3] = __byte_perm_S (    0, w0[0], selector);
-      w2[2] = 0;
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 12:
-      w7[3] = __byte_perm_S (w4[2], w4[3], selector);
-      w7[2] = __byte_perm_S (w4[1], w4[2], selector);
-      w7[1] = __byte_perm_S (w4[0], w4[1], selector);
-      w7[0] = __byte_perm_S (w3[3], w4[0], selector);
-      w6[3] = __byte_perm_S (w3[2], w3[3], selector);
-      w6[2] = __byte_perm_S (w3[1], w3[2], selector);
-      w6[1] = __byte_perm_S (w3[0], w3[1], selector);
-      w6[0] = __byte_perm_S (w2[3], w3[0], selector);
-      w5[3] = __byte_perm_S (w2[2], w2[3], selector);
-      w5[2] = __byte_perm_S (w2[1], w2[2], selector);
-      w5[1] = __byte_perm_S (w2[0], w2[1], selector);
-      w5[0] = __byte_perm_S (w1[3], w2[0], selector);
-      w4[3] = __byte_perm_S (w1[2], w1[3], selector);
-      w4[2] = __byte_perm_S (w1[1], w1[2], selector);
-      w4[1] = __byte_perm_S (w1[0], w1[1], selector);
-      w4[0] = __byte_perm_S (w0[3], w1[0], selector);
-      w3[3] = __byte_perm_S (w0[2], w0[3], selector);
-      w3[2] = __byte_perm_S (w0[1], w0[2], selector);
-      w3[1] = __byte_perm_S (w0[0], w0[1], selector);
-      w3[0] = __byte_perm_S (    0, w0[0], selector);
-      w2[3] = 0;
-      w2[2] = 0;
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 13:
-      w7[3] = __byte_perm_S (w4[1], w4[2], selector);
-      w7[2] = __byte_perm_S (w4[0], w4[1], selector);
-      w7[1] = __byte_perm_S (w3[3], w4[0], selector);
-      w7[0] = __byte_perm_S (w3[2], w3[3], selector);
-      w6[3] = __byte_perm_S (w3[1], w3[2], selector);
-      w6[2] = __byte_perm_S (w3[0], w3[1], selector);
-      w6[1] = __byte_perm_S (w2[3], w3[0], selector);
-      w6[0] = __byte_perm_S (w2[2], w2[3], selector);
-      w5[3] = __byte_perm_S (w2[1], w2[2], selector);
-      w5[2] = __byte_perm_S (w2[0], w2[1], selector);
-      w5[1] = __byte_perm_S (w1[3], w2[0], selector);
-      w5[0] = __byte_perm_S (w1[2], w1[3], selector);
-      w4[3] = __byte_perm_S (w1[1], w1[2], selector);
-      w4[2] = __byte_perm_S (w1[0], w1[1], selector);
-      w4[1] = __byte_perm_S (w0[3], w1[0], selector);
-      w4[0] = __byte_perm_S (w0[2], w0[3], selector);
-      w3[3] = __byte_perm_S (w0[1], w0[2], selector);
-      w3[2] = __byte_perm_S (w0[0], w0[1], selector);
-      w3[1] = __byte_perm_S (    0, w0[0], selector);
-      w3[0] = 0;
-      w2[3] = 0;
-      w2[2] = 0;
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 14:
-      w7[3] = __byte_perm_S (w4[0], w4[1], selector);
-      w7[2] = __byte_perm_S (w3[3], w4[0], selector);
-      w7[1] = __byte_perm_S (w3[2], w3[3], selector);
-      w7[0] = __byte_perm_S (w3[1], w3[2], selector);
-      w6[3] = __byte_perm_S (w3[0], w3[1], selector);
-      w6[2] = __byte_perm_S (w2[3], w3[0], selector);
-      w6[1] = __byte_perm_S (w2[2], w2[3], selector);
-      w6[0] = __byte_perm_S (w2[1], w2[2], selector);
-      w5[3] = __byte_perm_S (w2[0], w2[1], selector);
-      w5[2] = __byte_perm_S (w1[3], w2[0], selector);
-      w5[1] = __byte_perm_S (w1[2], w1[3], selector);
-      w5[0] = __byte_perm_S (w1[1], w1[2], selector);
-      w4[3] = __byte_perm_S (w1[0], w1[1], selector);
-      w4[2] = __byte_perm_S (w0[3], w1[0], selector);
-      w4[1] = __byte_perm_S (w0[2], w0[3], selector);
-      w4[0] = __byte_perm_S (w0[1], w0[2], selector);
-      w3[3] = __byte_perm_S (w0[0], w0[1], selector);
-      w3[2] = __byte_perm_S (    0, w0[0], selector);
-      w3[1] = 0;
-      w3[0] = 0;
-      w2[3] = 0;
-      w2[2] = 0;
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 15:
-      w7[3] = __byte_perm_S (w3[3], w4[0], selector);
-      w7[2] = __byte_perm_S (w3[2], w3[3], selector);
-      w7[1] = __byte_perm_S (w3[1], w3[2], selector);
-      w7[0] = __byte_perm_S (w3[0], w3[1], selector);
-      w6[3] = __byte_perm_S (w2[3], w3[0], selector);
-      w6[2] = __byte_perm_S (w2[2], w2[3], selector);
-      w6[1] = __byte_perm_S (w2[1], w2[2], selector);
-      w6[0] = __byte_perm_S (w2[0], w2[1], selector);
-      w5[3] = __byte_perm_S (w1[3], w2[0], selector);
-      w5[2] = __byte_perm_S (w1[2], w1[3], selector);
-      w5[1] = __byte_perm_S (w1[1], w1[2], selector);
-      w5[0] = __byte_perm_S (w1[0], w1[1], selector);
-      w4[3] = __byte_perm_S (w0[3], w1[0], selector);
-      w4[2] = __byte_perm_S (w0[2], w0[3], selector);
-      w4[1] = __byte_perm_S (w0[1], w0[2], selector);
-      w4[0] = __byte_perm_S (w0[0], w0[1], selector);
-      w3[3] = __byte_perm_S (    0, w0[0], selector);
-      w3[2] = 0;
-      w3[1] = 0;
-      w3[0] = 0;
-      w2[3] = 0;
-      w2[2] = 0;
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-  }
-  #endif
-}
-
-#define PACKVS84(s0,s1,s2,s3,s4,s5,s6,s7,v0,v1,v2,v3,v4,v5,v6,v7,e) \
-  PACKVS4 (s0, v0, e);                                              \
-  PACKVS4 (s1, v1, e);                                              \
-  PACKVS4 (s2, v2, e);                                              \
-  PACKVS4 (s3, v3, e);                                              \
-  PACKVS4 (s4, v4, e);                                              \
-  PACKVS4 (s5, v5, e);                                              \
-  PACKVS4 (s6, v6, e);                                              \
-  PACKVS4 (s7, v7, e);
-
-#define PACKSV84(s0,s1,s2,s3,s4,s5,s6,s7,v0,v1,v2,v3,v4,v5,v6,v7,e) \
-  PACKSV4 (s0, v0, e);                                              \
-  PACKSV4 (s1, v1, e);                                              \
-  PACKSV4 (s2, v2, e);                                              \
-  PACKSV4 (s3, v3, e);                                              \
-  PACKSV4 (s4, v4, e);                                              \
-  PACKSV4 (s5, v5, e);                                              \
-  PACKSV4 (s6, v6, e);                                              \
-  PACKSV4 (s7, v7, e);
-
-inline void switch_buffer_by_offset_8x4_le_VV (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], const u32x offset)
-{
-  #if VECT_SIZE == 1
-
-  switch_buffer_by_offset_8x4_le_S (w0, w1, w2, w3, w4, w5, w6, w7, offset);
-
-  #else
-
-  u32 t0[4];
-  u32 t1[4];
-  u32 t2[4];
-  u32 t3[4];
-  u32 t4[4];
-  u32 t5[4];
-  u32 t6[4];
-  u32 t7[4];
-
-  #endif
-
-  #if   VECT_SIZE == 2
-
-  // 1
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 0);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s0);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 0);
-
-  // 2
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 1);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s1);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 1);
-
-  #elif VECT_SIZE == 4
-
-  // 1
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 0);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s0);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 0);
-
-  // 2
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 1);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s1);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 1);
-
-  // 3
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 2);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s2);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 2);
-
-  // 4
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 3);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s3);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 3);
-
-  #elif VECT_SIZE == 8
-
-  // 1
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 0);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s0);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 0);
-
-  // 2
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 1);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s1);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 1);
-
-  // 3
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 2);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s2);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 2);
-
-  // 4
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 3);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s3);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 3);
-
-  // 5
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 4);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s4);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 4);
-
-  // 6
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 5);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s5);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 5);
-
-  // 7
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 6);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s6);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 6);
-
-  // 8
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 7);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s7);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 7);
-
-  #elif VECT_SIZE == 16
-
-  // 1
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 0);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s0);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 0);
-
-  // 2
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 1);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s1);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 1);
-
-  // 3
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 2);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s2);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 2);
-
-  // 4
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 3);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s3);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 3);
-
-  // 5
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 4);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s4);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 4);
-
-  // 6
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 5);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s5);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 5);
-
-  // 7
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 6);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s6);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 6);
-
-  // 8
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 7);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s7);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 7);
-
-  // 9
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 8);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s8);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 8);
-
-  // 10
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 9);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s9);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 9);
-
-  // 11
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, a);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.sa);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, a);
-
-  // 12
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, b);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.sb);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, b);
-
-  // 13
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, c);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.sc);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, c);
-
-  // 14
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, d);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.sd);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, d);
-
-  // 15
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, e);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.se);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, e);
-
-  // 16
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, f);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.sf);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, f);
-
-  #endif
-}
-
 __constant u64 k_sha512[80] =
 {
   SHA512C00, SHA512C01, SHA512C02, SHA512C03,
diff --git a/OpenCL/m15000_a1.cl b/OpenCL/m15000_a1.cl
index 86274b263..86f3270eb 100644
--- a/OpenCL/m15000_a1.cl
+++ b/OpenCL/m15000_a1.cl
@@ -14,1819 +14,6 @@
 #include "inc_common.cl"
 #include "inc_simd.cl"
 
-inline void switch_buffer_by_offset_8x4_le_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 w4[4], u32 w5[4], u32 w6[4], u32 w7[4], const u32 offset)
-{
-  #if defined IS_AMD || defined IS_GENERIC
-  const int offset_mod_4 = offset & 3;
-
-  const int offset_minus_4 = 4 - offset;
-
-  switch (offset / 4)
-  {
-    case 0:
-      w7[3] = amd_bytealign_S (w7[3], w7[2], offset_minus_4);
-      w7[2] = amd_bytealign_S (w7[2], w7[1], offset_minus_4);
-      w7[1] = amd_bytealign_S (w7[1], w7[0], offset_minus_4);
-      w7[0] = amd_bytealign_S (w7[0], w6[3], offset_minus_4);
-      w6[3] = amd_bytealign_S (w6[3], w6[2], offset_minus_4);
-      w6[2] = amd_bytealign_S (w6[2], w6[1], offset_minus_4);
-      w6[1] = amd_bytealign_S (w6[1], w6[0], offset_minus_4);
-      w6[0] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
-      w5[3] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
-      w5[2] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
-      w5[1] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w5[0] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w4[3] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w4[2] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w4[1] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w4[0] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w3[3] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w3[2] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w3[1] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w3[0] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w2[3] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w2[2] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w2[1] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w2[0] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w1[3] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w1[2] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w1[1] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w1[0] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w0[3] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w0[2] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w0[1] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w0[0] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-
-      if (offset_mod_4 == 0)
-      {
-        w0[0] = w0[1];
-        w0[1] = w0[2];
-        w0[2] = w0[3];
-        w0[3] = w1[0];
-        w1[0] = w1[1];
-        w1[1] = w1[2];
-        w1[2] = w1[3];
-        w1[3] = w2[0];
-        w2[0] = w2[1];
-        w2[1] = w2[2];
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 1:
-      w7[3] = amd_bytealign_S (w7[2], w7[1], offset_minus_4);
-      w7[2] = amd_bytealign_S (w7[1], w7[0], offset_minus_4);
-      w7[1] = amd_bytealign_S (w7[0], w6[3], offset_minus_4);
-      w7[0] = amd_bytealign_S (w6[3], w6[2], offset_minus_4);
-      w6[3] = amd_bytealign_S (w6[2], w6[1], offset_minus_4);
-      w6[2] = amd_bytealign_S (w6[1], w6[0], offset_minus_4);
-      w6[1] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
-      w6[0] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
-      w5[3] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
-      w5[2] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w5[1] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w5[0] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w4[3] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w4[2] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w4[1] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w4[0] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w3[3] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w3[2] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w3[1] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w3[0] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w2[3] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w2[2] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w2[1] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w2[0] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w1[3] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w1[2] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w1[1] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w1[0] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w0[3] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w0[2] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w0[1] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w0[1] = w0[2];
-        w0[2] = w0[3];
-        w0[3] = w1[0];
-        w1[0] = w1[1];
-        w1[1] = w1[2];
-        w1[2] = w1[3];
-        w1[3] = w2[0];
-        w2[0] = w2[1];
-        w2[1] = w2[2];
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 2:
-      w7[3] = amd_bytealign_S (w7[1], w7[0], offset_minus_4);
-      w7[2] = amd_bytealign_S (w7[0], w6[3], offset_minus_4);
-      w7[1] = amd_bytealign_S (w6[3], w6[2], offset_minus_4);
-      w7[0] = amd_bytealign_S (w6[2], w6[1], offset_minus_4);
-      w6[3] = amd_bytealign_S (w6[1], w6[0], offset_minus_4);
-      w6[2] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
-      w6[1] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
-      w6[0] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
-      w5[3] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w5[2] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w5[1] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w5[0] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w4[3] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w4[2] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w4[1] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w4[0] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w3[3] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w3[2] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w3[1] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w3[0] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w2[3] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w2[2] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w2[1] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w2[0] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w1[3] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w1[2] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w1[1] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w1[0] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w0[3] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w0[2] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w0[2] = w0[3];
-        w0[3] = w1[0];
-        w1[0] = w1[1];
-        w1[1] = w1[2];
-        w1[2] = w1[3];
-        w1[3] = w2[0];
-        w2[0] = w2[1];
-        w2[1] = w2[2];
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 3:
-      w7[3] = amd_bytealign_S (w7[0], w6[3], offset_minus_4);
-      w7[2] = amd_bytealign_S (w6[3], w6[2], offset_minus_4);
-      w7[1] = amd_bytealign_S (w6[2], w6[1], offset_minus_4);
-      w7[0] = amd_bytealign_S (w6[1], w6[0], offset_minus_4);
-      w6[3] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
-      w6[2] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
-      w6[1] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
-      w6[0] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w5[3] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w5[2] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w5[1] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w5[0] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w4[3] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w4[2] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w4[1] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w4[0] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w3[3] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w3[2] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w3[1] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w3[0] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w2[3] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w2[2] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w2[1] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w2[0] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w1[3] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w1[2] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w1[1] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w1[0] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w0[3] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w0[3] = w1[0];
-        w1[0] = w1[1];
-        w1[1] = w1[2];
-        w1[2] = w1[3];
-        w1[3] = w2[0];
-        w2[0] = w2[1];
-        w2[1] = w2[2];
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 4:
-      w7[3] = amd_bytealign_S (w6[3], w6[2], offset_minus_4);
-      w7[2] = amd_bytealign_S (w6[2], w6[1], offset_minus_4);
-      w7[1] = amd_bytealign_S (w6[1], w6[0], offset_minus_4);
-      w7[0] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
-      w6[3] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
-      w6[2] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
-      w6[1] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w6[0] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w5[3] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w5[2] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w5[1] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w5[0] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w4[3] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w4[2] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w4[1] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w4[0] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w3[3] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w3[2] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w3[1] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w3[0] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w2[3] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w2[2] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w2[1] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w2[0] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w1[3] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w1[2] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w1[1] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w1[0] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w1[0] = w1[1];
-        w1[1] = w1[2];
-        w1[2] = w1[3];
-        w1[3] = w2[0];
-        w2[0] = w2[1];
-        w2[1] = w2[2];
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 5:
-      w7[3] = amd_bytealign_S (w6[2], w6[1], offset_minus_4);
-      w7[2] = amd_bytealign_S (w6[1], w6[0], offset_minus_4);
-      w7[1] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
-      w7[0] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
-      w6[3] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
-      w6[2] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w6[1] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w6[0] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w5[3] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w5[2] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w5[1] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w5[0] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w4[3] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w4[2] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w4[1] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w4[0] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w3[3] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w3[2] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w3[1] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w3[0] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w2[3] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w2[2] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w2[1] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w2[0] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w1[3] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w1[2] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w1[1] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w1[1] = w1[2];
-        w1[2] = w1[3];
-        w1[3] = w2[0];
-        w2[0] = w2[1];
-        w2[1] = w2[2];
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 6:
-      w7[3] = amd_bytealign_S (w6[1], w6[0], offset_minus_4);
-      w7[2] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
-      w7[1] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
-      w7[0] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
-      w6[3] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w6[2] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w6[1] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w6[0] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w5[3] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w5[2] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w5[1] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w5[0] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w4[3] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w4[2] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w4[1] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w4[0] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w3[3] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w3[2] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w3[1] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w3[0] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w2[3] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w2[2] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w2[1] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w2[0] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w1[3] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w1[2] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w1[2] = w1[3];
-        w1[3] = w2[0];
-        w2[0] = w2[1];
-        w2[1] = w2[2];
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 7:
-      w7[3] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
-      w7[2] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
-      w7[1] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
-      w7[0] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w6[3] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w6[2] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w6[1] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w6[0] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w5[3] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w5[2] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w5[1] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w5[0] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w4[3] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w4[2] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w4[1] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w4[0] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w3[3] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w3[2] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w3[1] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w3[0] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w2[3] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w2[2] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w2[1] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w2[0] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w1[3] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w1[3] = w2[0];
-        w2[0] = w2[1];
-        w2[1] = w2[2];
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 8:
-      w7[3] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
-      w7[2] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
-      w7[1] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w7[0] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w6[3] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w6[2] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w6[1] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w6[0] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w5[3] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w5[2] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w5[1] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w5[0] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w4[3] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w4[2] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w4[1] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w4[0] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w3[3] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w3[2] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w3[1] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w3[0] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w2[3] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w2[2] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w2[1] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w2[0] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w2[0] = w2[1];
-        w2[1] = w2[2];
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 9:
-      w7[3] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
-      w7[2] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w7[1] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w7[0] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w6[3] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w6[2] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w6[1] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w6[0] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w5[3] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w5[2] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w5[1] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w5[0] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w4[3] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w4[2] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w4[1] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w4[0] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w3[3] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w3[2] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w3[1] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w3[0] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w2[3] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w2[2] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w2[1] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w2[1] = w2[2];
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 10:
-      w7[3] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w7[2] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w7[1] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w7[0] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w6[3] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w6[2] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w6[1] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w6[0] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w5[3] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w5[2] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w5[1] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w5[0] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w4[3] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w4[2] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w4[1] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w4[0] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w3[3] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w3[2] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w3[1] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w3[0] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w2[3] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w2[2] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 11:
-      w7[3] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w7[2] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w7[1] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w7[0] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w6[3] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w6[2] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w6[1] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w6[0] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w5[3] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w5[2] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w5[1] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w5[0] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w4[3] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w4[2] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w4[1] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w4[0] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w3[3] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w3[2] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w3[1] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w3[0] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w2[3] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w2[2] = 0;
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 12:
-      w7[3] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w7[2] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w7[1] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w7[0] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w6[3] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w6[2] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w6[1] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w6[0] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w5[3] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w5[2] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w5[1] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w5[0] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w4[3] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w4[2] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w4[1] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w4[0] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w3[3] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w3[2] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w3[1] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w3[0] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w2[3] = 0;
-      w2[2] = 0;
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 13:
-      w7[3] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w7[2] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w7[1] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w7[0] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w6[3] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w6[2] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w6[1] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w6[0] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w5[3] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w5[2] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w5[1] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w5[0] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w4[3] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w4[2] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w4[1] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w4[0] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w3[3] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w3[2] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w3[1] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w3[0] = 0;
-      w2[3] = 0;
-      w2[2] = 0;
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 14:
-      w7[3] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w7[2] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w7[1] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w7[0] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w6[3] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w6[2] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w6[1] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w6[0] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w5[3] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w5[2] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w5[1] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w5[0] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w4[3] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w4[2] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w4[1] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w4[0] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w3[3] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w3[2] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w3[1] = 0;
-      w3[0] = 0;
-      w2[3] = 0;
-      w2[2] = 0;
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 15:
-      w7[3] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w7[2] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w7[1] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w7[0] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w6[3] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w6[2] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w6[1] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w6[0] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w5[3] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w5[2] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w5[1] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w5[0] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w4[3] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w4[2] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w4[1] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w4[0] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w3[3] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w3[2] = 0;
-      w3[1] = 0;
-      w3[0] = 0;
-      w2[3] = 0;
-      w2[2] = 0;
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-  }
-  #endif
-
-  #ifdef IS_NV
-  const int offset_minus_4 = 4 - (offset % 4);
-
-  const int selector = (0x76543210 >> (offset_minus_4 * 4)) & 0xffff;
-
-  switch (offset / 4)
-  {
-    case 0:
-      w7[3] = __byte_perm_S (w7[2], w7[3], selector);
-      w7[2] = __byte_perm_S (w7[1], w7[2], selector);
-      w7[1] = __byte_perm_S (w7[0], w7[1], selector);
-      w7[0] = __byte_perm_S (w6[3], w7[0], selector);
-      w6[3] = __byte_perm_S (w6[2], w6[3], selector);
-      w6[2] = __byte_perm_S (w6[1], w6[2], selector);
-      w6[1] = __byte_perm_S (w6[0], w6[1], selector);
-      w6[0] = __byte_perm_S (w5[3], w6[0], selector);
-      w5[3] = __byte_perm_S (w5[2], w5[3], selector);
-      w5[2] = __byte_perm_S (w5[1], w5[2], selector);
-      w5[1] = __byte_perm_S (w5[0], w5[1], selector);
-      w5[0] = __byte_perm_S (w4[3], w5[0], selector);
-      w4[3] = __byte_perm_S (w4[2], w4[3], selector);
-      w4[2] = __byte_perm_S (w4[1], w4[2], selector);
-      w4[1] = __byte_perm_S (w4[0], w4[1], selector);
-      w4[0] = __byte_perm_S (w3[3], w4[0], selector);
-      w3[3] = __byte_perm_S (w3[2], w3[3], selector);
-      w3[2] = __byte_perm_S (w3[1], w3[2], selector);
-      w3[1] = __byte_perm_S (w3[0], w3[1], selector);
-      w3[0] = __byte_perm_S (w2[3], w3[0], selector);
-      w2[3] = __byte_perm_S (w2[2], w2[3], selector);
-      w2[2] = __byte_perm_S (w2[1], w2[2], selector);
-      w2[1] = __byte_perm_S (w2[0], w2[1], selector);
-      w2[0] = __byte_perm_S (w1[3], w2[0], selector);
-      w1[3] = __byte_perm_S (w1[2], w1[3], selector);
-      w1[2] = __byte_perm_S (w1[1], w1[2], selector);
-      w1[1] = __byte_perm_S (w1[0], w1[1], selector);
-      w1[0] = __byte_perm_S (w0[3], w1[0], selector);
-      w0[3] = __byte_perm_S (w0[2], w0[3], selector);
-      w0[2] = __byte_perm_S (w0[1], w0[2], selector);
-      w0[1] = __byte_perm_S (w0[0], w0[1], selector);
-      w0[0] = __byte_perm_S (    0, w0[0], selector);
-      break;
-
-    case 1:
-      w7[3] = __byte_perm_S (w7[1], w7[2], selector);
-      w7[2] = __byte_perm_S (w7[0], w7[1], selector);
-      w7[1] = __byte_perm_S (w6[3], w7[0], selector);
-      w7[0] = __byte_perm_S (w6[2], w6[3], selector);
-      w6[3] = __byte_perm_S (w6[1], w6[2], selector);
-      w6[2] = __byte_perm_S (w6[0], w6[1], selector);
-      w6[1] = __byte_perm_S (w5[3], w6[0], selector);
-      w6[0] = __byte_perm_S (w5[2], w5[3], selector);
-      w5[3] = __byte_perm_S (w5[1], w5[2], selector);
-      w5[2] = __byte_perm_S (w5[0], w5[1], selector);
-      w5[1] = __byte_perm_S (w4[3], w5[0], selector);
-      w5[0] = __byte_perm_S (w4[2], w4[3], selector);
-      w4[3] = __byte_perm_S (w4[1], w4[2], selector);
-      w4[2] = __byte_perm_S (w4[0], w4[1], selector);
-      w4[1] = __byte_perm_S (w3[3], w4[0], selector);
-      w4[0] = __byte_perm_S (w3[2], w3[3], selector);
-      w3[3] = __byte_perm_S (w3[1], w3[2], selector);
-      w3[2] = __byte_perm_S (w3[0], w3[1], selector);
-      w3[1] = __byte_perm_S (w2[3], w3[0], selector);
-      w3[0] = __byte_perm_S (w2[2], w2[3], selector);
-      w2[3] = __byte_perm_S (w2[1], w2[2], selector);
-      w2[2] = __byte_perm_S (w2[0], w2[1], selector);
-      w2[1] = __byte_perm_S (w1[3], w2[0], selector);
-      w2[0] = __byte_perm_S (w1[2], w1[3], selector);
-      w1[3] = __byte_perm_S (w1[1], w1[2], selector);
-      w1[2] = __byte_perm_S (w1[0], w1[1], selector);
-      w1[1] = __byte_perm_S (w0[3], w1[0], selector);
-      w1[0] = __byte_perm_S (w0[2], w0[3], selector);
-      w0[3] = __byte_perm_S (w0[1], w0[2], selector);
-      w0[2] = __byte_perm_S (w0[0], w0[1], selector);
-      w0[1] = __byte_perm_S (    0, w0[0], selector);
-      w0[0] = 0;
-      break;
-
-    case 2:
-      w7[3] = __byte_perm_S (w7[0], w7[1], selector);
-      w7[2] = __byte_perm_S (w6[3], w7[0], selector);
-      w7[1] = __byte_perm_S (w6[2], w6[3], selector);
-      w7[0] = __byte_perm_S (w6[1], w6[2], selector);
-      w6[3] = __byte_perm_S (w6[0], w6[1], selector);
-      w6[2] = __byte_perm_S (w5[3], w6[0], selector);
-      w6[1] = __byte_perm_S (w5[2], w5[3], selector);
-      w6[0] = __byte_perm_S (w5[1], w5[2], selector);
-      w5[3] = __byte_perm_S (w5[0], w5[1], selector);
-      w5[2] = __byte_perm_S (w4[3], w5[0], selector);
-      w5[1] = __byte_perm_S (w4[2], w4[3], selector);
-      w5[0] = __byte_perm_S (w4[1], w4[2], selector);
-      w4[3] = __byte_perm_S (w4[0], w4[1], selector);
-      w4[2] = __byte_perm_S (w3[3], w4[0], selector);
-      w4[1] = __byte_perm_S (w3[2], w3[3], selector);
-      w4[0] = __byte_perm_S (w3[1], w3[2], selector);
-      w3[3] = __byte_perm_S (w3[0], w3[1], selector);
-      w3[2] = __byte_perm_S (w2[3], w3[0], selector);
-      w3[1] = __byte_perm_S (w2[2], w2[3], selector);
-      w3[0] = __byte_perm_S (w2[1], w2[2], selector);
-      w2[3] = __byte_perm_S (w2[0], w2[1], selector);
-      w2[2] = __byte_perm_S (w1[3], w2[0], selector);
-      w2[1] = __byte_perm_S (w1[2], w1[3], selector);
-      w2[0] = __byte_perm_S (w1[1], w1[2], selector);
-      w1[3] = __byte_perm_S (w1[0], w1[1], selector);
-      w1[2] = __byte_perm_S (w0[3], w1[0], selector);
-      w1[1] = __byte_perm_S (w0[2], w0[3], selector);
-      w1[0] = __byte_perm_S (w0[1], w0[2], selector);
-      w0[3] = __byte_perm_S (w0[0], w0[1], selector);
-      w0[2] = __byte_perm_S (    0, w0[0], selector);
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 3:
-      w7[3] = __byte_perm_S (w6[3], w7[0], selector);
-      w7[2] = __byte_perm_S (w6[2], w6[3], selector);
-      w7[1] = __byte_perm_S (w6[1], w6[2], selector);
-      w7[0] = __byte_perm_S (w6[0], w6[1], selector);
-      w6[3] = __byte_perm_S (w5[3], w6[0], selector);
-      w6[2] = __byte_perm_S (w5[2], w5[3], selector);
-      w6[1] = __byte_perm_S (w5[1], w5[2], selector);
-      w6[0] = __byte_perm_S (w5[0], w5[1], selector);
-      w5[3] = __byte_perm_S (w4[3], w5[0], selector);
-      w5[2] = __byte_perm_S (w4[2], w4[3], selector);
-      w5[1] = __byte_perm_S (w4[1], w4[2], selector);
-      w5[0] = __byte_perm_S (w4[0], w4[1], selector);
-      w4[3] = __byte_perm_S (w3[3], w4[0], selector);
-      w4[2] = __byte_perm_S (w3[2], w3[3], selector);
-      w4[1] = __byte_perm_S (w3[1], w3[2], selector);
-      w4[0] = __byte_perm_S (w3[0], w3[1], selector);
-      w3[3] = __byte_perm_S (w2[3], w3[0], selector);
-      w3[2] = __byte_perm_S (w2[2], w2[3], selector);
-      w3[1] = __byte_perm_S (w2[1], w2[2], selector);
-      w3[0] = __byte_perm_S (w2[0], w2[1], selector);
-      w2[3] = __byte_perm_S (w1[3], w2[0], selector);
-      w2[2] = __byte_perm_S (w1[2], w1[3], selector);
-      w2[1] = __byte_perm_S (w1[1], w1[2], selector);
-      w2[0] = __byte_perm_S (w1[0], w1[1], selector);
-      w1[3] = __byte_perm_S (w0[3], w1[0], selector);
-      w1[2] = __byte_perm_S (w0[2], w0[3], selector);
-      w1[1] = __byte_perm_S (w0[1], w0[2], selector);
-      w1[0] = __byte_perm_S (w0[0], w0[1], selector);
-      w0[3] = __byte_perm_S (    0, w0[0], selector);
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 4:
-      w7[3] = __byte_perm_S (w6[2], w6[3], selector);
-      w7[2] = __byte_perm_S (w6[1], w6[2], selector);
-      w7[1] = __byte_perm_S (w6[0], w6[1], selector);
-      w7[0] = __byte_perm_S (w5[3], w6[0], selector);
-      w6[3] = __byte_perm_S (w5[2], w5[3], selector);
-      w6[2] = __byte_perm_S (w5[1], w5[2], selector);
-      w6[1] = __byte_perm_S (w5[0], w5[1], selector);
-      w6[0] = __byte_perm_S (w4[3], w5[0], selector);
-      w5[3] = __byte_perm_S (w4[2], w4[3], selector);
-      w5[2] = __byte_perm_S (w4[1], w4[2], selector);
-      w5[1] = __byte_perm_S (w4[0], w4[1], selector);
-      w5[0] = __byte_perm_S (w3[3], w4[0], selector);
-      w4[3] = __byte_perm_S (w3[2], w3[3], selector);
-      w4[2] = __byte_perm_S (w3[1], w3[2], selector);
-      w4[1] = __byte_perm_S (w3[0], w3[1], selector);
-      w4[0] = __byte_perm_S (w2[3], w3[0], selector);
-      w3[3] = __byte_perm_S (w2[2], w2[3], selector);
-      w3[2] = __byte_perm_S (w2[1], w2[2], selector);
-      w3[1] = __byte_perm_S (w2[0], w2[1], selector);
-      w3[0] = __byte_perm_S (w1[3], w2[0], selector);
-      w2[3] = __byte_perm_S (w1[2], w1[3], selector);
-      w2[2] = __byte_perm_S (w1[1], w1[2], selector);
-      w2[1] = __byte_perm_S (w1[0], w1[1], selector);
-      w2[0] = __byte_perm_S (w0[3], w1[0], selector);
-      w1[3] = __byte_perm_S (w0[2], w0[3], selector);
-      w1[2] = __byte_perm_S (w0[1], w0[2], selector);
-      w1[1] = __byte_perm_S (w0[0], w0[1], selector);
-      w1[0] = __byte_perm_S (    0, w0[0], selector);
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 5:
-      w7[3] = __byte_perm_S (w6[1], w6[2], selector);
-      w7[2] = __byte_perm_S (w6[0], w6[1], selector);
-      w7[1] = __byte_perm_S (w5[3], w6[0], selector);
-      w7[0] = __byte_perm_S (w5[2], w5[3], selector);
-      w6[3] = __byte_perm_S (w5[1], w5[2], selector);
-      w6[2] = __byte_perm_S (w5[0], w5[1], selector);
-      w6[1] = __byte_perm_S (w4[3], w5[0], selector);
-      w6[0] = __byte_perm_S (w4[2], w4[3], selector);
-      w5[3] = __byte_perm_S (w4[1], w4[2], selector);
-      w5[2] = __byte_perm_S (w4[0], w4[1], selector);
-      w5[1] = __byte_perm_S (w3[3], w4[0], selector);
-      w5[0] = __byte_perm_S (w3[2], w3[3], selector);
-      w4[3] = __byte_perm_S (w3[1], w3[2], selector);
-      w4[2] = __byte_perm_S (w3[0], w3[1], selector);
-      w4[1] = __byte_perm_S (w2[3], w3[0], selector);
-      w4[0] = __byte_perm_S (w2[2], w2[3], selector);
-      w3[3] = __byte_perm_S (w2[1], w2[2], selector);
-      w3[2] = __byte_perm_S (w2[0], w2[1], selector);
-      w3[1] = __byte_perm_S (w1[3], w2[0], selector);
-      w3[0] = __byte_perm_S (w1[2], w1[3], selector);
-      w2[3] = __byte_perm_S (w1[1], w1[2], selector);
-      w2[2] = __byte_perm_S (w1[0], w1[1], selector);
-      w2[1] = __byte_perm_S (w0[3], w1[0], selector);
-      w2[0] = __byte_perm_S (w0[2], w0[3], selector);
-      w1[3] = __byte_perm_S (w0[1], w0[2], selector);
-      w1[2] = __byte_perm_S (w0[0], w0[1], selector);
-      w1[1] = __byte_perm_S (    0, w0[0], selector);
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 6:
-      w7[3] = __byte_perm_S (w6[0], w6[1], selector);
-      w7[2] = __byte_perm_S (w5[3], w6[0], selector);
-      w7[1] = __byte_perm_S (w5[2], w5[3], selector);
-      w7[0] = __byte_perm_S (w5[1], w5[2], selector);
-      w6[3] = __byte_perm_S (w5[0], w5[1], selector);
-      w6[2] = __byte_perm_S (w4[3], w5[0], selector);
-      w6[1] = __byte_perm_S (w4[2], w4[3], selector);
-      w6[0] = __byte_perm_S (w4[1], w4[2], selector);
-      w5[3] = __byte_perm_S (w4[0], w4[1], selector);
-      w5[2] = __byte_perm_S (w3[3], w4[0], selector);
-      w5[1] = __byte_perm_S (w3[2], w3[3], selector);
-      w5[0] = __byte_perm_S (w3[1], w3[2], selector);
-      w4[3] = __byte_perm_S (w3[0], w3[1], selector);
-      w4[2] = __byte_perm_S (w2[3], w3[0], selector);
-      w4[1] = __byte_perm_S (w2[2], w2[3], selector);
-      w4[0] = __byte_perm_S (w2[1], w2[2], selector);
-      w3[3] = __byte_perm_S (w2[0], w2[1], selector);
-      w3[2] = __byte_perm_S (w1[3], w2[0], selector);
-      w3[1] = __byte_perm_S (w1[2], w1[3], selector);
-      w3[0] = __byte_perm_S (w1[1], w1[2], selector);
-      w2[3] = __byte_perm_S (w1[0], w1[1], selector);
-      w2[2] = __byte_perm_S (w0[3], w1[0], selector);
-      w2[1] = __byte_perm_S (w0[2], w0[3], selector);
-      w2[0] = __byte_perm_S (w0[1], w0[2], selector);
-      w1[3] = __byte_perm_S (w0[0], w0[1], selector);
-      w1[2] = __byte_perm_S (    0, w0[0], selector);
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 7:
-      w7[3] = __byte_perm_S (w5[3], w6[0], selector);
-      w7[2] = __byte_perm_S (w5[2], w5[3], selector);
-      w7[1] = __byte_perm_S (w5[1], w5[2], selector);
-      w7[0] = __byte_perm_S (w5[0], w5[1], selector);
-      w6[3] = __byte_perm_S (w4[3], w5[0], selector);
-      w6[2] = __byte_perm_S (w4[2], w4[3], selector);
-      w6[1] = __byte_perm_S (w4[1], w4[2], selector);
-      w6[0] = __byte_perm_S (w4[0], w4[1], selector);
-      w5[3] = __byte_perm_S (w3[3], w4[0], selector);
-      w5[2] = __byte_perm_S (w3[2], w3[3], selector);
-      w5[1] = __byte_perm_S (w3[1], w3[2], selector);
-      w5[0] = __byte_perm_S (w3[0], w3[1], selector);
-      w4[3] = __byte_perm_S (w2[3], w3[0], selector);
-      w4[2] = __byte_perm_S (w2[2], w2[3], selector);
-      w4[1] = __byte_perm_S (w2[1], w2[2], selector);
-      w4[0] = __byte_perm_S (w2[0], w2[1], selector);
-      w3[3] = __byte_perm_S (w1[3], w2[0], selector);
-      w3[2] = __byte_perm_S (w1[2], w1[3], selector);
-      w3[1] = __byte_perm_S (w1[1], w1[2], selector);
-      w3[0] = __byte_perm_S (w1[0], w1[1], selector);
-      w2[3] = __byte_perm_S (w0[3], w1[0], selector);
-      w2[2] = __byte_perm_S (w0[2], w0[3], selector);
-      w2[1] = __byte_perm_S (w0[1], w0[2], selector);
-      w2[0] = __byte_perm_S (w0[0], w0[1], selector);
-      w1[3] = __byte_perm_S (    0, w0[0], selector);
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 8:
-      w7[3] = __byte_perm_S (w5[2], w5[3], selector);
-      w7[2] = __byte_perm_S (w5[1], w5[2], selector);
-      w7[1] = __byte_perm_S (w5[0], w5[1], selector);
-      w7[0] = __byte_perm_S (w4[3], w5[0], selector);
-      w6[3] = __byte_perm_S (w4[2], w4[3], selector);
-      w6[2] = __byte_perm_S (w4[1], w4[2], selector);
-      w6[1] = __byte_perm_S (w4[0], w4[1], selector);
-      w6[0] = __byte_perm_S (w3[3], w4[0], selector);
-      w5[3] = __byte_perm_S (w3[2], w3[3], selector);
-      w5[2] = __byte_perm_S (w3[1], w3[2], selector);
-      w5[1] = __byte_perm_S (w3[0], w3[1], selector);
-      w5[0] = __byte_perm_S (w2[3], w3[0], selector);
-      w4[3] = __byte_perm_S (w2[2], w2[3], selector);
-      w4[2] = __byte_perm_S (w2[1], w2[2], selector);
-      w4[1] = __byte_perm_S (w2[0], w2[1], selector);
-      w4[0] = __byte_perm_S (w1[3], w2[0], selector);
-      w3[3] = __byte_perm_S (w1[2], w1[3], selector);
-      w3[2] = __byte_perm_S (w1[1], w1[2], selector);
-      w3[1] = __byte_perm_S (w1[0], w1[1], selector);
-      w3[0] = __byte_perm_S (w0[3], w1[0], selector);
-      w2[3] = __byte_perm_S (w0[2], w0[3], selector);
-      w2[2] = __byte_perm_S (w0[1], w0[2], selector);
-      w2[1] = __byte_perm_S (w0[0], w0[1], selector);
-      w2[0] = __byte_perm_S (    0, w0[0], selector);
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 9:
-      w7[3] = __byte_perm_S (w5[1], w5[2], selector);
-      w7[2] = __byte_perm_S (w5[0], w5[1], selector);
-      w7[1] = __byte_perm_S (w4[3], w5[0], selector);
-      w7[0] = __byte_perm_S (w4[2], w4[3], selector);
-      w6[3] = __byte_perm_S (w4[1], w4[2], selector);
-      w6[2] = __byte_perm_S (w4[0], w4[1], selector);
-      w6[1] = __byte_perm_S (w3[3], w4[0], selector);
-      w6[0] = __byte_perm_S (w3[2], w3[3], selector);
-      w5[3] = __byte_perm_S (w3[1], w3[2], selector);
-      w5[2] = __byte_perm_S (w3[0], w3[1], selector);
-      w5[1] = __byte_perm_S (w2[3], w3[0], selector);
-      w5[0] = __byte_perm_S (w2[2], w2[3], selector);
-      w4[3] = __byte_perm_S (w2[1], w2[2], selector);
-      w4[2] = __byte_perm_S (w2[0], w2[1], selector);
-      w4[1] = __byte_perm_S (w1[3], w2[0], selector);
-      w4[0] = __byte_perm_S (w1[2], w1[3], selector);
-      w3[3] = __byte_perm_S (w1[1], w1[2], selector);
-      w3[2] = __byte_perm_S (w1[0], w1[1], selector);
-      w3[1] = __byte_perm_S (w0[3], w1[0], selector);
-      w3[0] = __byte_perm_S (w0[2], w0[3], selector);
-      w2[3] = __byte_perm_S (w0[1], w0[2], selector);
-      w2[2] = __byte_perm_S (w0[0], w0[1], selector);
-      w2[1] = __byte_perm_S (    0, w0[0], selector);
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 10:
-      w7[3] = __byte_perm_S (w5[0], w5[1], selector);
-      w7[2] = __byte_perm_S (w4[3], w5[0], selector);
-      w7[1] = __byte_perm_S (w4[2], w4[3], selector);
-      w7[0] = __byte_perm_S (w4[1], w4[2], selector);
-      w6[3] = __byte_perm_S (w4[0], w4[1], selector);
-      w6[2] = __byte_perm_S (w3[3], w4[0], selector);
-      w6[1] = __byte_perm_S (w3[2], w3[3], selector);
-      w6[0] = __byte_perm_S (w3[1], w3[2], selector);
-      w5[3] = __byte_perm_S (w3[0], w3[1], selector);
-      w5[2] = __byte_perm_S (w2[3], w3[0], selector);
-      w5[1] = __byte_perm_S (w2[2], w2[3], selector);
-      w5[0] = __byte_perm_S (w2[1], w2[2], selector);
-      w4[3] = __byte_perm_S (w2[0], w2[1], selector);
-      w4[2] = __byte_perm_S (w1[3], w2[0], selector);
-      w4[1] = __byte_perm_S (w1[2], w1[3], selector);
-      w4[0] = __byte_perm_S (w1[1], w1[2], selector);
-      w3[3] = __byte_perm_S (w1[0], w1[1], selector);
-      w3[2] = __byte_perm_S (w0[3], w1[0], selector);
-      w3[1] = __byte_perm_S (w0[2], w0[3], selector);
-      w3[0] = __byte_perm_S (w0[1], w0[2], selector);
-      w2[3] = __byte_perm_S (w0[0], w0[1], selector);
-      w2[2] = __byte_perm_S (    0, w0[0], selector);
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 11:
-      w7[3] = __byte_perm_S (w4[3], w5[0], selector);
-      w7[2] = __byte_perm_S (w4[2], w4[3], selector);
-      w7[1] = __byte_perm_S (w4[1], w4[2], selector);
-      w7[0] = __byte_perm_S (w4[0], w4[1], selector);
-      w6[3] = __byte_perm_S (w3[3], w4[0], selector);
-      w6[2] = __byte_perm_S (w3[2], w3[3], selector);
-      w6[1] = __byte_perm_S (w3[1], w3[2], selector);
-      w6[0] = __byte_perm_S (w3[0], w3[1], selector);
-      w5[3] = __byte_perm_S (w2[3], w3[0], selector);
-      w5[2] = __byte_perm_S (w2[2], w2[3], selector);
-      w5[1] = __byte_perm_S (w2[1], w2[2], selector);
-      w5[0] = __byte_perm_S (w2[0], w2[1], selector);
-      w4[3] = __byte_perm_S (w1[3], w2[0], selector);
-      w4[2] = __byte_perm_S (w1[2], w1[3], selector);
-      w4[1] = __byte_perm_S (w1[1], w1[2], selector);
-      w4[0] = __byte_perm_S (w1[0], w1[1], selector);
-      w3[3] = __byte_perm_S (w0[3], w1[0], selector);
-      w3[2] = __byte_perm_S (w0[2], w0[3], selector);
-      w3[1] = __byte_perm_S (w0[1], w0[2], selector);
-      w3[0] = __byte_perm_S (w0[0], w0[1], selector);
-      w2[3] = __byte_perm_S (    0, w0[0], selector);
-      w2[2] = 0;
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 12:
-      w7[3] = __byte_perm_S (w4[2], w4[3], selector);
-      w7[2] = __byte_perm_S (w4[1], w4[2], selector);
-      w7[1] = __byte_perm_S (w4[0], w4[1], selector);
-      w7[0] = __byte_perm_S (w3[3], w4[0], selector);
-      w6[3] = __byte_perm_S (w3[2], w3[3], selector);
-      w6[2] = __byte_perm_S (w3[1], w3[2], selector);
-      w6[1] = __byte_perm_S (w3[0], w3[1], selector);
-      w6[0] = __byte_perm_S (w2[3], w3[0], selector);
-      w5[3] = __byte_perm_S (w2[2], w2[3], selector);
-      w5[2] = __byte_perm_S (w2[1], w2[2], selector);
-      w5[1] = __byte_perm_S (w2[0], w2[1], selector);
-      w5[0] = __byte_perm_S (w1[3], w2[0], selector);
-      w4[3] = __byte_perm_S (w1[2], w1[3], selector);
-      w4[2] = __byte_perm_S (w1[1], w1[2], selector);
-      w4[1] = __byte_perm_S (w1[0], w1[1], selector);
-      w4[0] = __byte_perm_S (w0[3], w1[0], selector);
-      w3[3] = __byte_perm_S (w0[2], w0[3], selector);
-      w3[2] = __byte_perm_S (w0[1], w0[2], selector);
-      w3[1] = __byte_perm_S (w0[0], w0[1], selector);
-      w3[0] = __byte_perm_S (    0, w0[0], selector);
-      w2[3] = 0;
-      w2[2] = 0;
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 13:
-      w7[3] = __byte_perm_S (w4[1], w4[2], selector);
-      w7[2] = __byte_perm_S (w4[0], w4[1], selector);
-      w7[1] = __byte_perm_S (w3[3], w4[0], selector);
-      w7[0] = __byte_perm_S (w3[2], w3[3], selector);
-      w6[3] = __byte_perm_S (w3[1], w3[2], selector);
-      w6[2] = __byte_perm_S (w3[0], w3[1], selector);
-      w6[1] = __byte_perm_S (w2[3], w3[0], selector);
-      w6[0] = __byte_perm_S (w2[2], w2[3], selector);
-      w5[3] = __byte_perm_S (w2[1], w2[2], selector);
-      w5[2] = __byte_perm_S (w2[0], w2[1], selector);
-      w5[1] = __byte_perm_S (w1[3], w2[0], selector);
-      w5[0] = __byte_perm_S (w1[2], w1[3], selector);
-      w4[3] = __byte_perm_S (w1[1], w1[2], selector);
-      w4[2] = __byte_perm_S (w1[0], w1[1], selector);
-      w4[1] = __byte_perm_S (w0[3], w1[0], selector);
-      w4[0] = __byte_perm_S (w0[2], w0[3], selector);
-      w3[3] = __byte_perm_S (w0[1], w0[2], selector);
-      w3[2] = __byte_perm_S (w0[0], w0[1], selector);
-      w3[1] = __byte_perm_S (    0, w0[0], selector);
-      w3[0] = 0;
-      w2[3] = 0;
-      w2[2] = 0;
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 14:
-      w7[3] = __byte_perm_S (w4[0], w4[1], selector);
-      w7[2] = __byte_perm_S (w3[3], w4[0], selector);
-      w7[1] = __byte_perm_S (w3[2], w3[3], selector);
-      w7[0] = __byte_perm_S (w3[1], w3[2], selector);
-      w6[3] = __byte_perm_S (w3[0], w3[1], selector);
-      w6[2] = __byte_perm_S (w2[3], w3[0], selector);
-      w6[1] = __byte_perm_S (w2[2], w2[3], selector);
-      w6[0] = __byte_perm_S (w2[1], w2[2], selector);
-      w5[3] = __byte_perm_S (w2[0], w2[1], selector);
-      w5[2] = __byte_perm_S (w1[3], w2[0], selector);
-      w5[1] = __byte_perm_S (w1[2], w1[3], selector);
-      w5[0] = __byte_perm_S (w1[1], w1[2], selector);
-      w4[3] = __byte_perm_S (w1[0], w1[1], selector);
-      w4[2] = __byte_perm_S (w0[3], w1[0], selector);
-      w4[1] = __byte_perm_S (w0[2], w0[3], selector);
-      w4[0] = __byte_perm_S (w0[1], w0[2], selector);
-      w3[3] = __byte_perm_S (w0[0], w0[1], selector);
-      w3[2] = __byte_perm_S (    0, w0[0], selector);
-      w3[1] = 0;
-      w3[0] = 0;
-      w2[3] = 0;
-      w2[2] = 0;
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 15:
-      w7[3] = __byte_perm_S (w3[3], w4[0], selector);
-      w7[2] = __byte_perm_S (w3[2], w3[3], selector);
-      w7[1] = __byte_perm_S (w3[1], w3[2], selector);
-      w7[0] = __byte_perm_S (w3[0], w3[1], selector);
-      w6[3] = __byte_perm_S (w2[3], w3[0], selector);
-      w6[2] = __byte_perm_S (w2[2], w2[3], selector);
-      w6[1] = __byte_perm_S (w2[1], w2[2], selector);
-      w6[0] = __byte_perm_S (w2[0], w2[1], selector);
-      w5[3] = __byte_perm_S (w1[3], w2[0], selector);
-      w5[2] = __byte_perm_S (w1[2], w1[3], selector);
-      w5[1] = __byte_perm_S (w1[1], w1[2], selector);
-      w5[0] = __byte_perm_S (w1[0], w1[1], selector);
-      w4[3] = __byte_perm_S (w0[3], w1[0], selector);
-      w4[2] = __byte_perm_S (w0[2], w0[3], selector);
-      w4[1] = __byte_perm_S (w0[1], w0[2], selector);
-      w4[0] = __byte_perm_S (w0[0], w0[1], selector);
-      w3[3] = __byte_perm_S (    0, w0[0], selector);
-      w3[2] = 0;
-      w3[1] = 0;
-      w3[0] = 0;
-      w2[3] = 0;
-      w2[2] = 0;
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-  }
-  #endif
-}
-
-#define PACKVS84(s0,s1,s2,s3,s4,s5,s6,s7,v0,v1,v2,v3,v4,v5,v6,v7,e) \
-  PACKVS4 (s0, v0, e);                                              \
-  PACKVS4 (s1, v1, e);                                              \
-  PACKVS4 (s2, v2, e);                                              \
-  PACKVS4 (s3, v3, e);                                              \
-  PACKVS4 (s4, v4, e);                                              \
-  PACKVS4 (s5, v5, e);                                              \
-  PACKVS4 (s6, v6, e);                                              \
-  PACKVS4 (s7, v7, e);
-
-#define PACKSV84(s0,s1,s2,s3,s4,s5,s6,s7,v0,v1,v2,v3,v4,v5,v6,v7,e) \
-  PACKSV4 (s0, v0, e);                                              \
-  PACKSV4 (s1, v1, e);                                              \
-  PACKSV4 (s2, v2, e);                                              \
-  PACKSV4 (s3, v3, e);                                              \
-  PACKSV4 (s4, v4, e);                                              \
-  PACKSV4 (s5, v5, e);                                              \
-  PACKSV4 (s6, v6, e);                                              \
-  PACKSV4 (s7, v7, e);
-
-inline void switch_buffer_by_offset_8x4_le_VV (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x w4[4], u32x w5[4], u32x w6[4], u32x w7[4], const u32x offset)
-{
-  #if VECT_SIZE == 1
-
-  switch_buffer_by_offset_8x4_le_S (w0, w1, w2, w3, w4, w5, w6, w7, offset);
-
-  #else
-
-  u32 t0[4];
-  u32 t1[4];
-  u32 t2[4];
-  u32 t3[4];
-  u32 t4[4];
-  u32 t5[4];
-  u32 t6[4];
-  u32 t7[4];
-
-  #endif
-
-  #if   VECT_SIZE == 2
-
-  // 1
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 0);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s0);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 0);
-
-  // 2
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 1);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s1);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 1);
-
-  #elif VECT_SIZE == 4
-
-  // 1
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 0);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s0);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 0);
-
-  // 2
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 1);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s1);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 1);
-
-  // 3
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 2);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s2);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 2);
-
-  // 4
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 3);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s3);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 3);
-
-  #elif VECT_SIZE == 8
-
-  // 1
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 0);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s0);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 0);
-
-  // 2
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 1);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s1);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 1);
-
-  // 3
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 2);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s2);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 2);
-
-  // 4
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 3);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s3);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 3);
-
-  // 5
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 4);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s4);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 4);
-
-  // 6
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 5);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s5);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 5);
-
-  // 7
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 6);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s6);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 6);
-
-  // 8
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 7);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s7);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 7);
-
-  #elif VECT_SIZE == 16
-
-  // 1
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 0);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s0);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 0);
-
-  // 2
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 1);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s1);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 1);
-
-  // 3
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 2);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s2);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 2);
-
-  // 4
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 3);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s3);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 3);
-
-  // 5
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 4);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s4);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 4);
-
-  // 6
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 5);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s5);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 5);
-
-  // 7
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 6);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s6);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 6);
-
-  // 8
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 7);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s7);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 7);
-
-  // 9
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 8);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s8);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 8);
-
-  // 10
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 9);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.s9);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, 9);
-
-  // 11
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, a);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.sa);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, a);
-
-  // 12
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, b);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.sb);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, b);
-
-  // 13
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, c);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.sc);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, c);
-
-  // 14
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, d);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.sd);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, d);
-
-  // 15
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, e);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.se);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, e);
-
-  // 16
-  PACKVS84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, f);
-  switch_buffer_by_offset_8x4_le_S (t0, t1, t2, t3, t4, t5, t6, t7, offset.sf);
-  PACKSV84 (t0, t1, t2, t3, t4, t5, t6, t7, w0, w1, w2, w3, w4, w5, w6, w7, f);
-
-  #endif
-}
-
 __constant u64 k_sha512[80] =
 {
   SHA512C00, SHA512C01, SHA512C02, SHA512C03,
diff --git a/OpenCL/m15000_a3.cl b/OpenCL/m15000_a3.cl
index a3db83c55..5451e88be 100644
--- a/OpenCL/m15000_a3.cl
+++ b/OpenCL/m15000_a3.cl
@@ -14,1619 +14,6 @@
 #include "inc_common.cl"
 #include "inc_simd.cl"
 
-inline void switch_buffer_by_offset_8x4_le_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 w4[4], u32 w5[4], u32 w6[4], u32 w7[4], const u32 offset)
-{
-  #if defined IS_AMD || defined IS_GENERIC
-  const int offset_mod_4 = offset & 3;
-
-  const int offset_minus_4 = 4 - offset;
-
-  switch (offset / 4)
-  {
-    case 0:
-      w7[3] = amd_bytealign_S (w7[3], w7[2], offset_minus_4);
-      w7[2] = amd_bytealign_S (w7[2], w7[1], offset_minus_4);
-      w7[1] = amd_bytealign_S (w7[1], w7[0], offset_minus_4);
-      w7[0] = amd_bytealign_S (w7[0], w6[3], offset_minus_4);
-      w6[3] = amd_bytealign_S (w6[3], w6[2], offset_minus_4);
-      w6[2] = amd_bytealign_S (w6[2], w6[1], offset_minus_4);
-      w6[1] = amd_bytealign_S (w6[1], w6[0], offset_minus_4);
-      w6[0] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
-      w5[3] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
-      w5[2] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
-      w5[1] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w5[0] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w4[3] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w4[2] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w4[1] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w4[0] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w3[3] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w3[2] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w3[1] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w3[0] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w2[3] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w2[2] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w2[1] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w2[0] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w1[3] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w1[2] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w1[1] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w1[0] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w0[3] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w0[2] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w0[1] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w0[0] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-
-      if (offset_mod_4 == 0)
-      {
-        w0[0] = w0[1];
-        w0[1] = w0[2];
-        w0[2] = w0[3];
-        w0[3] = w1[0];
-        w1[0] = w1[1];
-        w1[1] = w1[2];
-        w1[2] = w1[3];
-        w1[3] = w2[0];
-        w2[0] = w2[1];
-        w2[1] = w2[2];
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 1:
-      w7[3] = amd_bytealign_S (w7[2], w7[1], offset_minus_4);
-      w7[2] = amd_bytealign_S (w7[1], w7[0], offset_minus_4);
-      w7[1] = amd_bytealign_S (w7[0], w6[3], offset_minus_4);
-      w7[0] = amd_bytealign_S (w6[3], w6[2], offset_minus_4);
-      w6[3] = amd_bytealign_S (w6[2], w6[1], offset_minus_4);
-      w6[2] = amd_bytealign_S (w6[1], w6[0], offset_minus_4);
-      w6[1] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
-      w6[0] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
-      w5[3] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
-      w5[2] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w5[1] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w5[0] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w4[3] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w4[2] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w4[1] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w4[0] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w3[3] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w3[2] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w3[1] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w3[0] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w2[3] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w2[2] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w2[1] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w2[0] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w1[3] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w1[2] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w1[1] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w1[0] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w0[3] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w0[2] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w0[1] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w0[1] = w0[2];
-        w0[2] = w0[3];
-        w0[3] = w1[0];
-        w1[0] = w1[1];
-        w1[1] = w1[2];
-        w1[2] = w1[3];
-        w1[3] = w2[0];
-        w2[0] = w2[1];
-        w2[1] = w2[2];
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 2:
-      w7[3] = amd_bytealign_S (w7[1], w7[0], offset_minus_4);
-      w7[2] = amd_bytealign_S (w7[0], w6[3], offset_minus_4);
-      w7[1] = amd_bytealign_S (w6[3], w6[2], offset_minus_4);
-      w7[0] = amd_bytealign_S (w6[2], w6[1], offset_minus_4);
-      w6[3] = amd_bytealign_S (w6[1], w6[0], offset_minus_4);
-      w6[2] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
-      w6[1] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
-      w6[0] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
-      w5[3] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w5[2] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w5[1] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w5[0] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w4[3] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w4[2] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w4[1] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w4[0] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w3[3] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w3[2] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w3[1] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w3[0] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w2[3] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w2[2] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w2[1] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w2[0] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w1[3] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w1[2] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w1[1] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w1[0] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w0[3] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w0[2] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w0[2] = w0[3];
-        w0[3] = w1[0];
-        w1[0] = w1[1];
-        w1[1] = w1[2];
-        w1[2] = w1[3];
-        w1[3] = w2[0];
-        w2[0] = w2[1];
-        w2[1] = w2[2];
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 3:
-      w7[3] = amd_bytealign_S (w7[0], w6[3], offset_minus_4);
-      w7[2] = amd_bytealign_S (w6[3], w6[2], offset_minus_4);
-      w7[1] = amd_bytealign_S (w6[2], w6[1], offset_minus_4);
-      w7[0] = amd_bytealign_S (w6[1], w6[0], offset_minus_4);
-      w6[3] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
-      w6[2] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
-      w6[1] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
-      w6[0] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w5[3] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w5[2] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w5[1] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w5[0] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w4[3] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w4[2] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w4[1] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w4[0] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w3[3] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w3[2] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w3[1] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w3[0] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w2[3] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w2[2] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w2[1] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w2[0] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w1[3] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w1[2] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w1[1] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w1[0] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w0[3] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w0[3] = w1[0];
-        w1[0] = w1[1];
-        w1[1] = w1[2];
-        w1[2] = w1[3];
-        w1[3] = w2[0];
-        w2[0] = w2[1];
-        w2[1] = w2[2];
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 4:
-      w7[3] = amd_bytealign_S (w6[3], w6[2], offset_minus_4);
-      w7[2] = amd_bytealign_S (w6[2], w6[1], offset_minus_4);
-      w7[1] = amd_bytealign_S (w6[1], w6[0], offset_minus_4);
-      w7[0] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
-      w6[3] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
-      w6[2] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
-      w6[1] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w6[0] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w5[3] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w5[2] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w5[1] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w5[0] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w4[3] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w4[2] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w4[1] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w4[0] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w3[3] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w3[2] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w3[1] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w3[0] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w2[3] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w2[2] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w2[1] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w2[0] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w1[3] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w1[2] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w1[1] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w1[0] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w1[0] = w1[1];
-        w1[1] = w1[2];
-        w1[2] = w1[3];
-        w1[3] = w2[0];
-        w2[0] = w2[1];
-        w2[1] = w2[2];
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 5:
-      w7[3] = amd_bytealign_S (w6[2], w6[1], offset_minus_4);
-      w7[2] = amd_bytealign_S (w6[1], w6[0], offset_minus_4);
-      w7[1] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
-      w7[0] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
-      w6[3] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
-      w6[2] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w6[1] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w6[0] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w5[3] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w5[2] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w5[1] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w5[0] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w4[3] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w4[2] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w4[1] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w4[0] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w3[3] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w3[2] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w3[1] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w3[0] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w2[3] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w2[2] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w2[1] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w2[0] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w1[3] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w1[2] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w1[1] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w1[1] = w1[2];
-        w1[2] = w1[3];
-        w1[3] = w2[0];
-        w2[0] = w2[1];
-        w2[1] = w2[2];
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 6:
-      w7[3] = amd_bytealign_S (w6[1], w6[0], offset_minus_4);
-      w7[2] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
-      w7[1] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
-      w7[0] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
-      w6[3] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w6[2] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w6[1] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w6[0] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w5[3] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w5[2] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w5[1] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w5[0] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w4[3] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w4[2] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w4[1] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w4[0] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w3[3] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w3[2] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w3[1] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w3[0] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w2[3] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w2[2] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w2[1] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w2[0] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w1[3] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w1[2] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w1[2] = w1[3];
-        w1[3] = w2[0];
-        w2[0] = w2[1];
-        w2[1] = w2[2];
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 7:
-      w7[3] = amd_bytealign_S (w6[0], w5[3], offset_minus_4);
-      w7[2] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
-      w7[1] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
-      w7[0] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w6[3] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w6[2] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w6[1] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w6[0] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w5[3] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w5[2] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w5[1] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w5[0] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w4[3] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w4[2] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w4[1] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w4[0] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w3[3] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w3[2] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w3[1] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w3[0] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w2[3] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w2[2] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w2[1] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w2[0] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w1[3] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w1[3] = w2[0];
-        w2[0] = w2[1];
-        w2[1] = w2[2];
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 8:
-      w7[3] = amd_bytealign_S (w5[3], w5[2], offset_minus_4);
-      w7[2] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
-      w7[1] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w7[0] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w6[3] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w6[2] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w6[1] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w6[0] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w5[3] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w5[2] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w5[1] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w5[0] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w4[3] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w4[2] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w4[1] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w4[0] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w3[3] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w3[2] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w3[1] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w3[0] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w2[3] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w2[2] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w2[1] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w2[0] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w2[0] = w2[1];
-        w2[1] = w2[2];
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 9:
-      w7[3] = amd_bytealign_S (w5[2], w5[1], offset_minus_4);
-      w7[2] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w7[1] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w7[0] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w6[3] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w6[2] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w6[1] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w6[0] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w5[3] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w5[2] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w5[1] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w5[0] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w4[3] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w4[2] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w4[1] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w4[0] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w3[3] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w3[2] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w3[1] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w3[0] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w2[3] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w2[2] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w2[1] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w2[1] = w2[2];
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 10:
-      w7[3] = amd_bytealign_S (w5[1], w5[0], offset_minus_4);
-      w7[2] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w7[1] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w7[0] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w6[3] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w6[2] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w6[1] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w6[0] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w5[3] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w5[2] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w5[1] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w5[0] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w4[3] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w4[2] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w4[1] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w4[0] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w3[3] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w3[2] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w3[1] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w3[0] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w2[3] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w2[2] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w2[2] = w2[3];
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 11:
-      w7[3] = amd_bytealign_S (w5[0], w5[3], offset_minus_4);
-      w7[2] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w7[1] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w7[0] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w6[3] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w6[2] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w6[1] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w6[0] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w5[3] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w5[2] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w5[1] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w5[0] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w4[3] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w4[2] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w4[1] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w4[0] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w3[3] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w3[2] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w3[1] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w3[0] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w2[3] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w2[2] = 0;
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w2[3] = w3[0];
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 12:
-      w7[3] = amd_bytealign_S (w4[3], w4[2], offset_minus_4);
-      w7[2] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w7[1] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w7[0] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w6[3] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w6[2] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w6[1] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w6[0] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w5[3] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w5[2] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w5[1] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w5[0] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w4[3] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w4[2] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w4[1] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w4[0] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w3[3] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w3[2] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w3[1] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w3[0] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w2[3] = 0;
-      w2[2] = 0;
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w3[0] = w3[1];
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 13:
-      w7[3] = amd_bytealign_S (w4[2], w4[1], offset_minus_4);
-      w7[2] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w7[1] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w7[0] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w6[3] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w6[2] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w6[1] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w6[0] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w5[3] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w5[2] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w5[1] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w5[0] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w4[3] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w4[2] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w4[1] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w4[0] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w3[3] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w3[2] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w3[1] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w3[0] = 0;
-      w2[3] = 0;
-      w2[2] = 0;
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w3[1] = w3[2];
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 14:
-      w7[3] = amd_bytealign_S (w4[1], w4[0], offset_minus_4);
-      w7[2] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w7[1] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w7[0] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w6[3] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w6[2] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w6[1] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w6[0] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w5[3] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w5[2] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w5[1] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w5[0] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w4[3] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w4[2] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w4[1] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w4[0] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w3[3] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w3[2] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w3[1] = 0;
-      w3[0] = 0;
-      w2[3] = 0;
-      w2[2] = 0;
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w3[2] = w3[3];
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-
-    case 15:
-      w7[3] = amd_bytealign_S (w4[0], w3[3], offset_minus_4);
-      w7[2] = amd_bytealign_S (w3[3], w3[2], offset_minus_4);
-      w7[1] = amd_bytealign_S (w3[2], w3[1], offset_minus_4);
-      w7[0] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
-      w6[3] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
-      w6[2] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
-      w6[1] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
-      w6[0] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
-      w5[3] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
-      w5[2] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
-      w5[1] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
-      w5[0] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
-      w4[3] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
-      w4[2] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
-      w4[1] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
-      w4[0] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
-      w3[3] = amd_bytealign_S (w0[0],     0, offset_minus_4);
-      w3[2] = 0;
-      w3[1] = 0;
-      w3[0] = 0;
-      w2[3] = 0;
-      w2[2] = 0;
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-
-      if (offset_mod_4 == 0)
-      {
-        w3[3] = w4[0];
-        w4[0] = w4[1];
-        w4[1] = w4[2];
-        w4[2] = w4[3];
-        w4[3] = w5[0];
-        w5[0] = w5[1];
-        w5[1] = w5[2];
-        w5[2] = w5[3];
-        w5[3] = w6[0];
-        w6[0] = w6[1];
-        w6[1] = w6[2];
-        w6[2] = w6[3];
-        w6[3] = w7[0];
-        w7[0] = w7[1];
-        w7[1] = w7[2];
-        w7[2] = w7[3];
-        w7[3] = 0;
-      }
-
-      break;
-  }
-  #endif
-
-  #ifdef IS_NV
-  const int offset_minus_4 = 4 - (offset % 4);
-
-  const int selector = (0x76543210 >> (offset_minus_4 * 4)) & 0xffff;
-
-  switch (offset / 4)
-  {
-    case 0:
-      w7[3] = __byte_perm_S (w7[2], w7[3], selector);
-      w7[2] = __byte_perm_S (w7[1], w7[2], selector);
-      w7[1] = __byte_perm_S (w7[0], w7[1], selector);
-      w7[0] = __byte_perm_S (w6[3], w7[0], selector);
-      w6[3] = __byte_perm_S (w6[2], w6[3], selector);
-      w6[2] = __byte_perm_S (w6[1], w6[2], selector);
-      w6[1] = __byte_perm_S (w6[0], w6[1], selector);
-      w6[0] = __byte_perm_S (w5[3], w6[0], selector);
-      w5[3] = __byte_perm_S (w5[2], w5[3], selector);
-      w5[2] = __byte_perm_S (w5[1], w5[2], selector);
-      w5[1] = __byte_perm_S (w5[0], w5[1], selector);
-      w5[0] = __byte_perm_S (w4[3], w5[0], selector);
-      w4[3] = __byte_perm_S (w4[2], w4[3], selector);
-      w4[2] = __byte_perm_S (w4[1], w4[2], selector);
-      w4[1] = __byte_perm_S (w4[0], w4[1], selector);
-      w4[0] = __byte_perm_S (w3[3], w4[0], selector);
-      w3[3] = __byte_perm_S (w3[2], w3[3], selector);
-      w3[2] = __byte_perm_S (w3[1], w3[2], selector);
-      w3[1] = __byte_perm_S (w3[0], w3[1], selector);
-      w3[0] = __byte_perm_S (w2[3], w3[0], selector);
-      w2[3] = __byte_perm_S (w2[2], w2[3], selector);
-      w2[2] = __byte_perm_S (w2[1], w2[2], selector);
-      w2[1] = __byte_perm_S (w2[0], w2[1], selector);
-      w2[0] = __byte_perm_S (w1[3], w2[0], selector);
-      w1[3] = __byte_perm_S (w1[2], w1[3], selector);
-      w1[2] = __byte_perm_S (w1[1], w1[2], selector);
-      w1[1] = __byte_perm_S (w1[0], w1[1], selector);
-      w1[0] = __byte_perm_S (w0[3], w1[0], selector);
-      w0[3] = __byte_perm_S (w0[2], w0[3], selector);
-      w0[2] = __byte_perm_S (w0[1], w0[2], selector);
-      w0[1] = __byte_perm_S (w0[0], w0[1], selector);
-      w0[0] = __byte_perm_S (    0, w0[0], selector);
-      break;
-
-    case 1:
-      w7[3] = __byte_perm_S (w7[1], w7[2], selector);
-      w7[2] = __byte_perm_S (w7[0], w7[1], selector);
-      w7[1] = __byte_perm_S (w6[3], w7[0], selector);
-      w7[0] = __byte_perm_S (w6[2], w6[3], selector);
-      w6[3] = __byte_perm_S (w6[1], w6[2], selector);
-      w6[2] = __byte_perm_S (w6[0], w6[1], selector);
-      w6[1] = __byte_perm_S (w5[3], w6[0], selector);
-      w6[0] = __byte_perm_S (w5[2], w5[3], selector);
-      w5[3] = __byte_perm_S (w5[1], w5[2], selector);
-      w5[2] = __byte_perm_S (w5[0], w5[1], selector);
-      w5[1] = __byte_perm_S (w4[3], w5[0], selector);
-      w5[0] = __byte_perm_S (w4[2], w4[3], selector);
-      w4[3] = __byte_perm_S (w4[1], w4[2], selector);
-      w4[2] = __byte_perm_S (w4[0], w4[1], selector);
-      w4[1] = __byte_perm_S (w3[3], w4[0], selector);
-      w4[0] = __byte_perm_S (w3[2], w3[3], selector);
-      w3[3] = __byte_perm_S (w3[1], w3[2], selector);
-      w3[2] = __byte_perm_S (w3[0], w3[1], selector);
-      w3[1] = __byte_perm_S (w2[3], w3[0], selector);
-      w3[0] = __byte_perm_S (w2[2], w2[3], selector);
-      w2[3] = __byte_perm_S (w2[1], w2[2], selector);
-      w2[2] = __byte_perm_S (w2[0], w2[1], selector);
-      w2[1] = __byte_perm_S (w1[3], w2[0], selector);
-      w2[0] = __byte_perm_S (w1[2], w1[3], selector);
-      w1[3] = __byte_perm_S (w1[1], w1[2], selector);
-      w1[2] = __byte_perm_S (w1[0], w1[1], selector);
-      w1[1] = __byte_perm_S (w0[3], w1[0], selector);
-      w1[0] = __byte_perm_S (w0[2], w0[3], selector);
-      w0[3] = __byte_perm_S (w0[1], w0[2], selector);
-      w0[2] = __byte_perm_S (w0[0], w0[1], selector);
-      w0[1] = __byte_perm_S (    0, w0[0], selector);
-      w0[0] = 0;
-      break;
-
-    case 2:
-      w7[3] = __byte_perm_S (w7[0], w7[1], selector);
-      w7[2] = __byte_perm_S (w6[3], w7[0], selector);
-      w7[1] = __byte_perm_S (w6[2], w6[3], selector);
-      w7[0] = __byte_perm_S (w6[1], w6[2], selector);
-      w6[3] = __byte_perm_S (w6[0], w6[1], selector);
-      w6[2] = __byte_perm_S (w5[3], w6[0], selector);
-      w6[1] = __byte_perm_S (w5[2], w5[3], selector);
-      w6[0] = __byte_perm_S (w5[1], w5[2], selector);
-      w5[3] = __byte_perm_S (w5[0], w5[1], selector);
-      w5[2] = __byte_perm_S (w4[3], w5[0], selector);
-      w5[1] = __byte_perm_S (w4[2], w4[3], selector);
-      w5[0] = __byte_perm_S (w4[1], w4[2], selector);
-      w4[3] = __byte_perm_S (w4[0], w4[1], selector);
-      w4[2] = __byte_perm_S (w3[3], w4[0], selector);
-      w4[1] = __byte_perm_S (w3[2], w3[3], selector);
-      w4[0] = __byte_perm_S (w3[1], w3[2], selector);
-      w3[3] = __byte_perm_S (w3[0], w3[1], selector);
-      w3[2] = __byte_perm_S (w2[3], w3[0], selector);
-      w3[1] = __byte_perm_S (w2[2], w2[3], selector);
-      w3[0] = __byte_perm_S (w2[1], w2[2], selector);
-      w2[3] = __byte_perm_S (w2[0], w2[1], selector);
-      w2[2] = __byte_perm_S (w1[3], w2[0], selector);
-      w2[1] = __byte_perm_S (w1[2], w1[3], selector);
-      w2[0] = __byte_perm_S (w1[1], w1[2], selector);
-      w1[3] = __byte_perm_S (w1[0], w1[1], selector);
-      w1[2] = __byte_perm_S (w0[3], w1[0], selector);
-      w1[1] = __byte_perm_S (w0[2], w0[3], selector);
-      w1[0] = __byte_perm_S (w0[1], w0[2], selector);
-      w0[3] = __byte_perm_S (w0[0], w0[1], selector);
-      w0[2] = __byte_perm_S (    0, w0[0], selector);
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 3:
-      w7[3] = __byte_perm_S (w6[3], w7[0], selector);
-      w7[2] = __byte_perm_S (w6[2], w6[3], selector);
-      w7[1] = __byte_perm_S (w6[1], w6[2], selector);
-      w7[0] = __byte_perm_S (w6[0], w6[1], selector);
-      w6[3] = __byte_perm_S (w5[3], w6[0], selector);
-      w6[2] = __byte_perm_S (w5[2], w5[3], selector);
-      w6[1] = __byte_perm_S (w5[1], w5[2], selector);
-      w6[0] = __byte_perm_S (w5[0], w5[1], selector);
-      w5[3] = __byte_perm_S (w4[3], w5[0], selector);
-      w5[2] = __byte_perm_S (w4[2], w4[3], selector);
-      w5[1] = __byte_perm_S (w4[1], w4[2], selector);
-      w5[0] = __byte_perm_S (w4[0], w4[1], selector);
-      w4[3] = __byte_perm_S (w3[3], w4[0], selector);
-      w4[2] = __byte_perm_S (w3[2], w3[3], selector);
-      w4[1] = __byte_perm_S (w3[1], w3[2], selector);
-      w4[0] = __byte_perm_S (w3[0], w3[1], selector);
-      w3[3] = __byte_perm_S (w2[3], w3[0], selector);
-      w3[2] = __byte_perm_S (w2[2], w2[3], selector);
-      w3[1] = __byte_perm_S (w2[1], w2[2], selector);
-      w3[0] = __byte_perm_S (w2[0], w2[1], selector);
-      w2[3] = __byte_perm_S (w1[3], w2[0], selector);
-      w2[2] = __byte_perm_S (w1[2], w1[3], selector);
-      w2[1] = __byte_perm_S (w1[1], w1[2], selector);
-      w2[0] = __byte_perm_S (w1[0], w1[1], selector);
-      w1[3] = __byte_perm_S (w0[3], w1[0], selector);
-      w1[2] = __byte_perm_S (w0[2], w0[3], selector);
-      w1[1] = __byte_perm_S (w0[1], w0[2], selector);
-      w1[0] = __byte_perm_S (w0[0], w0[1], selector);
-      w0[3] = __byte_perm_S (    0, w0[0], selector);
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 4:
-      w7[3] = __byte_perm_S (w6[2], w6[3], selector);
-      w7[2] = __byte_perm_S (w6[1], w6[2], selector);
-      w7[1] = __byte_perm_S (w6[0], w6[1], selector);
-      w7[0] = __byte_perm_S (w5[3], w6[0], selector);
-      w6[3] = __byte_perm_S (w5[2], w5[3], selector);
-      w6[2] = __byte_perm_S (w5[1], w5[2], selector);
-      w6[1] = __byte_perm_S (w5[0], w5[1], selector);
-      w6[0] = __byte_perm_S (w4[3], w5[0], selector);
-      w5[3] = __byte_perm_S (w4[2], w4[3], selector);
-      w5[2] = __byte_perm_S (w4[1], w4[2], selector);
-      w5[1] = __byte_perm_S (w4[0], w4[1], selector);
-      w5[0] = __byte_perm_S (w3[3], w4[0], selector);
-      w4[3] = __byte_perm_S (w3[2], w3[3], selector);
-      w4[2] = __byte_perm_S (w3[1], w3[2], selector);
-      w4[1] = __byte_perm_S (w3[0], w3[1], selector);
-      w4[0] = __byte_perm_S (w2[3], w3[0], selector);
-      w3[3] = __byte_perm_S (w2[2], w2[3], selector);
-      w3[2] = __byte_perm_S (w2[1], w2[2], selector);
-      w3[1] = __byte_perm_S (w2[0], w2[1], selector);
-      w3[0] = __byte_perm_S (w1[3], w2[0], selector);
-      w2[3] = __byte_perm_S (w1[2], w1[3], selector);
-      w2[2] = __byte_perm_S (w1[1], w1[2], selector);
-      w2[1] = __byte_perm_S (w1[0], w1[1], selector);
-      w2[0] = __byte_perm_S (w0[3], w1[0], selector);
-      w1[3] = __byte_perm_S (w0[2], w0[3], selector);
-      w1[2] = __byte_perm_S (w0[1], w0[2], selector);
-      w1[1] = __byte_perm_S (w0[0], w0[1], selector);
-      w1[0] = __byte_perm_S (    0, w0[0], selector);
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 5:
-      w7[3] = __byte_perm_S (w6[1], w6[2], selector);
-      w7[2] = __byte_perm_S (w6[0], w6[1], selector);
-      w7[1] = __byte_perm_S (w5[3], w6[0], selector);
-      w7[0] = __byte_perm_S (w5[2], w5[3], selector);
-      w6[3] = __byte_perm_S (w5[1], w5[2], selector);
-      w6[2] = __byte_perm_S (w5[0], w5[1], selector);
-      w6[1] = __byte_perm_S (w4[3], w5[0], selector);
-      w6[0] = __byte_perm_S (w4[2], w4[3], selector);
-      w5[3] = __byte_perm_S (w4[1], w4[2], selector);
-      w5[2] = __byte_perm_S (w4[0], w4[1], selector);
-      w5[1] = __byte_perm_S (w3[3], w4[0], selector);
-      w5[0] = __byte_perm_S (w3[2], w3[3], selector);
-      w4[3] = __byte_perm_S (w3[1], w3[2], selector);
-      w4[2] = __byte_perm_S (w3[0], w3[1], selector);
-      w4[1] = __byte_perm_S (w2[3], w3[0], selector);
-      w4[0] = __byte_perm_S (w2[2], w2[3], selector);
-      w3[3] = __byte_perm_S (w2[1], w2[2], selector);
-      w3[2] = __byte_perm_S (w2[0], w2[1], selector);
-      w3[1] = __byte_perm_S (w1[3], w2[0], selector);
-      w3[0] = __byte_perm_S (w1[2], w1[3], selector);
-      w2[3] = __byte_perm_S (w1[1], w1[2], selector);
-      w2[2] = __byte_perm_S (w1[0], w1[1], selector);
-      w2[1] = __byte_perm_S (w0[3], w1[0], selector);
-      w2[0] = __byte_perm_S (w0[2], w0[3], selector);
-      w1[3] = __byte_perm_S (w0[1], w0[2], selector);
-      w1[2] = __byte_perm_S (w0[0], w0[1], selector);
-      w1[1] = __byte_perm_S (    0, w0[0], selector);
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 6:
-      w7[3] = __byte_perm_S (w6[0], w6[1], selector);
-      w7[2] = __byte_perm_S (w5[3], w6[0], selector);
-      w7[1] = __byte_perm_S (w5[2], w5[3], selector);
-      w7[0] = __byte_perm_S (w5[1], w5[2], selector);
-      w6[3] = __byte_perm_S (w5[0], w5[1], selector);
-      w6[2] = __byte_perm_S (w4[3], w5[0], selector);
-      w6[1] = __byte_perm_S (w4[2], w4[3], selector);
-      w6[0] = __byte_perm_S (w4[1], w4[2], selector);
-      w5[3] = __byte_perm_S (w4[0], w4[1], selector);
-      w5[2] = __byte_perm_S (w3[3], w4[0], selector);
-      w5[1] = __byte_perm_S (w3[2], w3[3], selector);
-      w5[0] = __byte_perm_S (w3[1], w3[2], selector);
-      w4[3] = __byte_perm_S (w3[0], w3[1], selector);
-      w4[2] = __byte_perm_S (w2[3], w3[0], selector);
-      w4[1] = __byte_perm_S (w2[2], w2[3], selector);
-      w4[0] = __byte_perm_S (w2[1], w2[2], selector);
-      w3[3] = __byte_perm_S (w2[0], w2[1], selector);
-      w3[2] = __byte_perm_S (w1[3], w2[0], selector);
-      w3[1] = __byte_perm_S (w1[2], w1[3], selector);
-      w3[0] = __byte_perm_S (w1[1], w1[2], selector);
-      w2[3] = __byte_perm_S (w1[0], w1[1], selector);
-      w2[2] = __byte_perm_S (w0[3], w1[0], selector);
-      w2[1] = __byte_perm_S (w0[2], w0[3], selector);
-      w2[0] = __byte_perm_S (w0[1], w0[2], selector);
-      w1[3] = __byte_perm_S (w0[0], w0[1], selector);
-      w1[2] = __byte_perm_S (    0, w0[0], selector);
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 7:
-      w7[3] = __byte_perm_S (w5[3], w6[0], selector);
-      w7[2] = __byte_perm_S (w5[2], w5[3], selector);
-      w7[1] = __byte_perm_S (w5[1], w5[2], selector);
-      w7[0] = __byte_perm_S (w5[0], w5[1], selector);
-      w6[3] = __byte_perm_S (w4[3], w5[0], selector);
-      w6[2] = __byte_perm_S (w4[2], w4[3], selector);
-      w6[1] = __byte_perm_S (w4[1], w4[2], selector);
-      w6[0] = __byte_perm_S (w4[0], w4[1], selector);
-      w5[3] = __byte_perm_S (w3[3], w4[0], selector);
-      w5[2] = __byte_perm_S (w3[2], w3[3], selector);
-      w5[1] = __byte_perm_S (w3[1], w3[2], selector);
-      w5[0] = __byte_perm_S (w3[0], w3[1], selector);
-      w4[3] = __byte_perm_S (w2[3], w3[0], selector);
-      w4[2] = __byte_perm_S (w2[2], w2[3], selector);
-      w4[1] = __byte_perm_S (w2[1], w2[2], selector);
-      w4[0] = __byte_perm_S (w2[0], w2[1], selector);
-      w3[3] = __byte_perm_S (w1[3], w2[0], selector);
-      w3[2] = __byte_perm_S (w1[2], w1[3], selector);
-      w3[1] = __byte_perm_S (w1[1], w1[2], selector);
-      w3[0] = __byte_perm_S (w1[0], w1[1], selector);
-      w2[3] = __byte_perm_S (w0[3], w1[0], selector);
-      w2[2] = __byte_perm_S (w0[2], w0[3], selector);
-      w2[1] = __byte_perm_S (w0[1], w0[2], selector);
-      w2[0] = __byte_perm_S (w0[0], w0[1], selector);
-      w1[3] = __byte_perm_S (    0, w0[0], selector);
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 8:
-      w7[3] = __byte_perm_S (w5[2], w5[3], selector);
-      w7[2] = __byte_perm_S (w5[1], w5[2], selector);
-      w7[1] = __byte_perm_S (w5[0], w5[1], selector);
-      w7[0] = __byte_perm_S (w4[3], w5[0], selector);
-      w6[3] = __byte_perm_S (w4[2], w4[3], selector);
-      w6[2] = __byte_perm_S (w4[1], w4[2], selector);
-      w6[1] = __byte_perm_S (w4[0], w4[1], selector);
-      w6[0] = __byte_perm_S (w3[3], w4[0], selector);
-      w5[3] = __byte_perm_S (w3[2], w3[3], selector);
-      w5[2] = __byte_perm_S (w3[1], w3[2], selector);
-      w5[1] = __byte_perm_S (w3[0], w3[1], selector);
-      w5[0] = __byte_perm_S (w2[3], w3[0], selector);
-      w4[3] = __byte_perm_S (w2[2], w2[3], selector);
-      w4[2] = __byte_perm_S (w2[1], w2[2], selector);
-      w4[1] = __byte_perm_S (w2[0], w2[1], selector);
-      w4[0] = __byte_perm_S (w1[3], w2[0], selector);
-      w3[3] = __byte_perm_S (w1[2], w1[3], selector);
-      w3[2] = __byte_perm_S (w1[1], w1[2], selector);
-      w3[1] = __byte_perm_S (w1[0], w1[1], selector);
-      w3[0] = __byte_perm_S (w0[3], w1[0], selector);
-      w2[3] = __byte_perm_S (w0[2], w0[3], selector);
-      w2[2] = __byte_perm_S (w0[1], w0[2], selector);
-      w2[1] = __byte_perm_S (w0[0], w0[1], selector);
-      w2[0] = __byte_perm_S (    0, w0[0], selector);
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 9:
-      w7[3] = __byte_perm_S (w5[1], w5[2], selector);
-      w7[2] = __byte_perm_S (w5[0], w5[1], selector);
-      w7[1] = __byte_perm_S (w4[3], w5[0], selector);
-      w7[0] = __byte_perm_S (w4[2], w4[3], selector);
-      w6[3] = __byte_perm_S (w4[1], w4[2], selector);
-      w6[2] = __byte_perm_S (w4[0], w4[1], selector);
-      w6[1] = __byte_perm_S (w3[3], w4[0], selector);
-      w6[0] = __byte_perm_S (w3[2], w3[3], selector);
-      w5[3] = __byte_perm_S (w3[1], w3[2], selector);
-      w5[2] = __byte_perm_S (w3[0], w3[1], selector);
-      w5[1] = __byte_perm_S (w2[3], w3[0], selector);
-      w5[0] = __byte_perm_S (w2[2], w2[3], selector);
-      w4[3] = __byte_perm_S (w2[1], w2[2], selector);
-      w4[2] = __byte_perm_S (w2[0], w2[1], selector);
-      w4[1] = __byte_perm_S (w1[3], w2[0], selector);
-      w4[0] = __byte_perm_S (w1[2], w1[3], selector);
-      w3[3] = __byte_perm_S (w1[1], w1[2], selector);
-      w3[2] = __byte_perm_S (w1[0], w1[1], selector);
-      w3[1] = __byte_perm_S (w0[3], w1[0], selector);
-      w3[0] = __byte_perm_S (w0[2], w0[3], selector);
-      w2[3] = __byte_perm_S (w0[1], w0[2], selector);
-      w2[2] = __byte_perm_S (w0[0], w0[1], selector);
-      w2[1] = __byte_perm_S (    0, w0[0], selector);
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 10:
-      w7[3] = __byte_perm_S (w5[0], w5[1], selector);
-      w7[2] = __byte_perm_S (w4[3], w5[0], selector);
-      w7[1] = __byte_perm_S (w4[2], w4[3], selector);
-      w7[0] = __byte_perm_S (w4[1], w4[2], selector);
-      w6[3] = __byte_perm_S (w4[0], w4[1], selector);
-      w6[2] = __byte_perm_S (w3[3], w4[0], selector);
-      w6[1] = __byte_perm_S (w3[2], w3[3], selector);
-      w6[0] = __byte_perm_S (w3[1], w3[2], selector);
-      w5[3] = __byte_perm_S (w3[0], w3[1], selector);
-      w5[2] = __byte_perm_S (w2[3], w3[0], selector);
-      w5[1] = __byte_perm_S (w2[2], w2[3], selector);
-      w5[0] = __byte_perm_S (w2[1], w2[2], selector);
-      w4[3] = __byte_perm_S (w2[0], w2[1], selector);
-      w4[2] = __byte_perm_S (w1[3], w2[0], selector);
-      w4[1] = __byte_perm_S (w1[2], w1[3], selector);
-      w4[0] = __byte_perm_S (w1[1], w1[2], selector);
-      w3[3] = __byte_perm_S (w1[0], w1[1], selector);
-      w3[2] = __byte_perm_S (w0[3], w1[0], selector);
-      w3[1] = __byte_perm_S (w0[2], w0[3], selector);
-      w3[0] = __byte_perm_S (w0[1], w0[2], selector);
-      w2[3] = __byte_perm_S (w0[0], w0[1], selector);
-      w2[2] = __byte_perm_S (    0, w0[0], selector);
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 11:
-      w7[3] = __byte_perm_S (w4[3], w5[0], selector);
-      w7[2] = __byte_perm_S (w4[2], w4[3], selector);
-      w7[1] = __byte_perm_S (w4[1], w4[2], selector);
-      w7[0] = __byte_perm_S (w4[0], w4[1], selector);
-      w6[3] = __byte_perm_S (w3[3], w4[0], selector);
-      w6[2] = __byte_perm_S (w3[2], w3[3], selector);
-      w6[1] = __byte_perm_S (w3[1], w3[2], selector);
-      w6[0] = __byte_perm_S (w3[0], w3[1], selector);
-      w5[3] = __byte_perm_S (w2[3], w3[0], selector);
-      w5[2] = __byte_perm_S (w2[2], w2[3], selector);
-      w5[1] = __byte_perm_S (w2[1], w2[2], selector);
-      w5[0] = __byte_perm_S (w2[0], w2[1], selector);
-      w4[3] = __byte_perm_S (w1[3], w2[0], selector);
-      w4[2] = __byte_perm_S (w1[2], w1[3], selector);
-      w4[1] = __byte_perm_S (w1[1], w1[2], selector);
-      w4[0] = __byte_perm_S (w1[0], w1[1], selector);
-      w3[3] = __byte_perm_S (w0[3], w1[0], selector);
-      w3[2] = __byte_perm_S (w0[2], w0[3], selector);
-      w3[1] = __byte_perm_S (w0[1], w0[2], selector);
-      w3[0] = __byte_perm_S (w0[0], w0[1], selector);
-      w2[3] = __byte_perm_S (    0, w0[0], selector);
-      w2[2] = 0;
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 12:
-      w7[3] = __byte_perm_S (w4[2], w4[3], selector);
-      w7[2] = __byte_perm_S (w4[1], w4[2], selector);
-      w7[1] = __byte_perm_S (w4[0], w4[1], selector);
-      w7[0] = __byte_perm_S (w3[3], w4[0], selector);
-      w6[3] = __byte_perm_S (w3[2], w3[3], selector);
-      w6[2] = __byte_perm_S (w3[1], w3[2], selector);
-      w6[1] = __byte_perm_S (w3[0], w3[1], selector);
-      w6[0] = __byte_perm_S (w2[3], w3[0], selector);
-      w5[3] = __byte_perm_S (w2[2], w2[3], selector);
-      w5[2] = __byte_perm_S (w2[1], w2[2], selector);
-      w5[1] = __byte_perm_S (w2[0], w2[1], selector);
-      w5[0] = __byte_perm_S (w1[3], w2[0], selector);
-      w4[3] = __byte_perm_S (w1[2], w1[3], selector);
-      w4[2] = __byte_perm_S (w1[1], w1[2], selector);
-      w4[1] = __byte_perm_S (w1[0], w1[1], selector);
-      w4[0] = __byte_perm_S (w0[3], w1[0], selector);
-      w3[3] = __byte_perm_S (w0[2], w0[3], selector);
-      w3[2] = __byte_perm_S (w0[1], w0[2], selector);
-      w3[1] = __byte_perm_S (w0[0], w0[1], selector);
-      w3[0] = __byte_perm_S (    0, w0[0], selector);
-      w2[3] = 0;
-      w2[2] = 0;
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 13:
-      w7[3] = __byte_perm_S (w4[1], w4[2], selector);
-      w7[2] = __byte_perm_S (w4[0], w4[1], selector);
-      w7[1] = __byte_perm_S (w3[3], w4[0], selector);
-      w7[0] = __byte_perm_S (w3[2], w3[3], selector);
-      w6[3] = __byte_perm_S (w3[1], w3[2], selector);
-      w6[2] = __byte_perm_S (w3[0], w3[1], selector);
-      w6[1] = __byte_perm_S (w2[3], w3[0], selector);
-      w6[0] = __byte_perm_S (w2[2], w2[3], selector);
-      w5[3] = __byte_perm_S (w2[1], w2[2], selector);
-      w5[2] = __byte_perm_S (w2[0], w2[1], selector);
-      w5[1] = __byte_perm_S (w1[3], w2[0], selector);
-      w5[0] = __byte_perm_S (w1[2], w1[3], selector);
-      w4[3] = __byte_perm_S (w1[1], w1[2], selector);
-      w4[2] = __byte_perm_S (w1[0], w1[1], selector);
-      w4[1] = __byte_perm_S (w0[3], w1[0], selector);
-      w4[0] = __byte_perm_S (w0[2], w0[3], selector);
-      w3[3] = __byte_perm_S (w0[1], w0[2], selector);
-      w3[2] = __byte_perm_S (w0[0], w0[1], selector);
-      w3[1] = __byte_perm_S (    0, w0[0], selector);
-      w3[0] = 0;
-      w2[3] = 0;
-      w2[2] = 0;
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 14:
-      w7[3] = __byte_perm_S (w4[0], w4[1], selector);
-      w7[2] = __byte_perm_S (w3[3], w4[0], selector);
-      w7[1] = __byte_perm_S (w3[2], w3[3], selector);
-      w7[0] = __byte_perm_S (w3[1], w3[2], selector);
-      w6[3] = __byte_perm_S (w3[0], w3[1], selector);
-      w6[2] = __byte_perm_S (w2[3], w3[0], selector);
-      w6[1] = __byte_perm_S (w2[2], w2[3], selector);
-      w6[0] = __byte_perm_S (w2[1], w2[2], selector);
-      w5[3] = __byte_perm_S (w2[0], w2[1], selector);
-      w5[2] = __byte_perm_S (w1[3], w2[0], selector);
-      w5[1] = __byte_perm_S (w1[2], w1[3], selector);
-      w5[0] = __byte_perm_S (w1[1], w1[2], selector);
-      w4[3] = __byte_perm_S (w1[0], w1[1], selector);
-      w4[2] = __byte_perm_S (w0[3], w1[0], selector);
-      w4[1] = __byte_perm_S (w0[2], w0[3], selector);
-      w4[0] = __byte_perm_S (w0[1], w0[2], selector);
-      w3[3] = __byte_perm_S (w0[0], w0[1], selector);
-      w3[2] = __byte_perm_S (    0, w0[0], selector);
-      w3[1] = 0;
-      w3[0] = 0;
-      w2[3] = 0;
-      w2[2] = 0;
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-
-    case 15:
-      w7[3] = __byte_perm_S (w3[3], w4[0], selector);
-      w7[2] = __byte_perm_S (w3[2], w3[3], selector);
-      w7[1] = __byte_perm_S (w3[1], w3[2], selector);
-      w7[0] = __byte_perm_S (w3[0], w3[1], selector);
-      w6[3] = __byte_perm_S (w2[3], w3[0], selector);
-      w6[2] = __byte_perm_S (w2[2], w2[3], selector);
-      w6[1] = __byte_perm_S (w2[1], w2[2], selector);
-      w6[0] = __byte_perm_S (w2[0], w2[1], selector);
-      w5[3] = __byte_perm_S (w1[3], w2[0], selector);
-      w5[2] = __byte_perm_S (w1[2], w1[3], selector);
-      w5[1] = __byte_perm_S (w1[1], w1[2], selector);
-      w5[0] = __byte_perm_S (w1[0], w1[1], selector);
-      w4[3] = __byte_perm_S (w0[3], w1[0], selector);
-      w4[2] = __byte_perm_S (w0[2], w0[3], selector);
-      w4[1] = __byte_perm_S (w0[1], w0[2], selector);
-      w4[0] = __byte_perm_S (w0[0], w0[1], selector);
-      w3[3] = __byte_perm_S (    0, w0[0], selector);
-      w3[2] = 0;
-      w3[1] = 0;
-      w3[0] = 0;
-      w2[3] = 0;
-      w2[2] = 0;
-      w2[1] = 0;
-      w2[0] = 0;
-      w1[3] = 0;
-      w1[2] = 0;
-      w1[1] = 0;
-      w1[0] = 0;
-      w0[3] = 0;
-      w0[2] = 0;
-      w0[1] = 0;
-      w0[0] = 0;
-      break;
-  }
-  #endif
-}
-
 __constant u64 k_sha512[80] =
 {
   SHA512C00, SHA512C01, SHA512C02, SHA512C03,