From 492f9f2290e716583082ca68dee79ddc96f92a6f Mon Sep 17 00:00:00 2001
From: jsteube <jens.steube@gmail.com>
Date: Mon, 31 Oct 2016 14:22:00 +0100
Subject: [PATCH] For some easy actions, maxwell GPU prefer not to use switch()

---
 OpenCL/inc_common.cl | 1852 +++++++-----------------------------------
 1 file changed, 282 insertions(+), 1570 deletions(-)

diff --git a/OpenCL/inc_common.cl b/OpenCL/inc_common.cl
index 75c4d8bf5..ac11e9f0d 100644
--- a/OpenCL/inc_common.cl
+++ b/OpenCL/inc_common.cl
@@ -6069,1662 +6069,374 @@ inline void overwrite_at_be_4x4 (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4],
 
 inline void append_0x01_1x4_S (u32 w0[4], const u32 offset)
 {
-  switch (offset)
-  {
-    case 0:
-      w0[0]  = 0x01;
-      break;
+  u32 w[4];
 
-    case 1:
-      w0[0] = w0[0] | 0x0100;
-      break;
+  w[0] = w0[0];
+  w[1] = w0[1];
+  w[2] = w0[2];
+  w[3] = w0[3];
 
-    case 2:
-      w0[0] = w0[0] | 0x010000;
-      break;
+  u8 *w_ptr = (u8 *) w;
 
-    case 3:
-      w0[0] = w0[0] | 0x01000000;
-      break;
+  w_ptr[offset] = 0x01;
 
-    case 4:
-      w0[1] = 0x01;
-      break;
-
-    case 5:
-      w0[1] = w0[1] | 0x0100;
-      break;
-
-    case 6:
-      w0[1] = w0[1] | 0x010000;
-      break;
-
-    case 7:
-      w0[1] = w0[1] | 0x01000000;
-      break;
-
-    case 8:
-      w0[2] = 0x01;
-      break;
-
-    case 9:
-      w0[2] = w0[2] | 0x0100;
-      break;
-
-    case 10:
-      w0[2] = w0[2] | 0x010000;
-      break;
-
-    case 11:
-      w0[2] = w0[2] | 0x01000000;
-      break;
-
-    case 12:
-      w0[3] = 0x01;
-      break;
-
-    case 13:
-      w0[3] = w0[3] | 0x0100;
-      break;
-
-    case 14:
-      w0[3] = w0[3] | 0x010000;
-      break;
-
-    case 15:
-      w0[3] = w0[3] | 0x01000000;
-      break;
-  }
+  w0[0] = w[0];
+  w0[1] = w[1];
+  w0[2] = w[2];
+  w0[3] = w[3];
 }
 
 inline void append_0x01_2x4_S (u32 w0[4], u32 w1[4], const u32 offset)
 {
-  switch (offset)
-  {
-    case 0:
-      w0[0] = 0x01;
-      break;
+  u32 w[8];
 
-    case 1:
-      w0[0] = w0[0] | 0x0100;
-      break;
+  w[0] = w0[0];
+  w[1] = w0[1];
+  w[2] = w0[2];
+  w[3] = w0[3];
+  w[4] = w1[0];
+  w[5] = w1[1];
+  w[6] = w1[2];
+  w[7] = w1[3];
 
-    case 2:
-      w0[0] = w0[0] | 0x010000;
-      break;
+  u8 *w_ptr = (u8 *) w;
 
-    case 3:
-      w0[0] = w0[0] | 0x01000000;
-      break;
+  w_ptr[offset] = 0x01;
 
-    case 4:
-      w0[1] = 0x01;
-      break;
-
-    case 5:
-      w0[1] = w0[1] | 0x0100;
-      break;
-
-    case 6:
-      w0[1] = w0[1] | 0x010000;
-      break;
-
-    case 7:
-      w0[1] = w0[1] | 0x01000000;
-      break;
-
-    case 8:
-      w0[2] = 0x01;
-      break;
-
-    case 9:
-      w0[2] = w0[2] | 0x0100;
-      break;
-
-    case 10:
-      w0[2] = w0[2] | 0x010000;
-      break;
-
-    case 11:
-      w0[2] = w0[2] | 0x01000000;
-      break;
-
-    case 12:
-      w0[3] = 0x01;
-      break;
-
-    case 13:
-      w0[3] = w0[3] | 0x0100;
-      break;
-
-    case 14:
-      w0[3] = w0[3] | 0x010000;
-      break;
-
-    case 15:
-      w0[3] = w0[3] | 0x01000000;
-      break;
-
-    case 16:
-      w1[0] = 0x01;
-      break;
-
-    case 17:
-      w1[0] = w1[0] | 0x0100;
-      break;
-
-    case 18:
-      w1[0] = w1[0] | 0x010000;
-      break;
-
-    case 19:
-      w1[0] = w1[0] | 0x01000000;
-      break;
-
-    case 20:
-      w1[1] = 0x01;
-      break;
-
-    case 21:
-      w1[1] = w1[1] | 0x0100;
-      break;
-
-    case 22:
-      w1[1] = w1[1] | 0x010000;
-      break;
-
-    case 23:
-      w1[1] = w1[1] | 0x01000000;
-      break;
-
-    case 24:
-      w1[2] = 0x01;
-      break;
-
-    case 25:
-      w1[2] = w1[2] | 0x0100;
-      break;
-
-    case 26:
-      w1[2] = w1[2] | 0x010000;
-      break;
-
-    case 27:
-      w1[2] = w1[2] | 0x01000000;
-      break;
-
-    case 28:
-      w1[3] = 0x01;
-      break;
-
-    case 29:
-      w1[3] = w1[3] | 0x0100;
-      break;
-
-    case 30:
-      w1[3] = w1[3] | 0x010000;
-      break;
-
-    case 31:
-      w1[3] = w1[3] | 0x01000000;
-      break;
-  }
+  w0[0] = w[0];
+  w0[1] = w[1];
+  w0[2] = w[2];
+  w0[3] = w[3];
+  w1[0] = w[4];
+  w1[1] = w[5];
+  w1[2] = w[6];
+  w1[3] = w[7];
 }
 
 inline void append_0x01_3x4_S (u32 w0[4], u32 w1[4], u32 w2[4], const u32 offset)
 {
-  switch (offset)
-  {
-    case 0:
-      w0[0] = 0x01;
-      break;
+  u32 w[12];
 
-    case 1:
-      w0[0] = w0[0] | 0x0100;
-      break;
+  w[ 0] = w0[0];
+  w[ 1] = w0[1];
+  w[ 2] = w0[2];
+  w[ 3] = w0[3];
+  w[ 4] = w1[0];
+  w[ 5] = w1[1];
+  w[ 6] = w1[2];
+  w[ 7] = w1[3];
+  w[ 8] = w2[0];
+  w[ 9] = w2[1];
+  w[10] = w2[2];
+  w[11] = w2[3];
 
-    case 2:
-      w0[0] = w0[0] | 0x010000;
-      break;
+  u8 *w_ptr = (u8 *) w;
 
-    case 3:
-      w0[0] = w0[0] | 0x01000000;
-      break;
+  w_ptr[offset] = 0x01;
 
-    case 4:
-      w0[1] = 0x01;
-      break;
-
-    case 5:
-      w0[1] = w0[1] | 0x0100;
-      break;
-
-    case 6:
-      w0[1] = w0[1] | 0x010000;
-      break;
-
-    case 7:
-      w0[1] = w0[1] | 0x01000000;
-      break;
-
-    case 8:
-      w0[2] = 0x01;
-      break;
-
-    case 9:
-      w0[2] = w0[2] | 0x0100;
-      break;
-
-    case 10:
-      w0[2] = w0[2] | 0x010000;
-      break;
-
-    case 11:
-      w0[2] = w0[2] | 0x01000000;
-      break;
-
-    case 12:
-      w0[3] = 0x01;
-      break;
-
-    case 13:
-      w0[3] = w0[3] | 0x0100;
-      break;
-
-    case 14:
-      w0[3] = w0[3] | 0x010000;
-      break;
-
-    case 15:
-      w0[3] = w0[3] | 0x01000000;
-      break;
-
-    case 16:
-      w1[0] = 0x01;
-      break;
-
-    case 17:
-      w1[0] = w1[0] | 0x0100;
-      break;
-
-    case 18:
-      w1[0] = w1[0] | 0x010000;
-      break;
-
-    case 19:
-      w1[0] = w1[0] | 0x01000000;
-      break;
-
-    case 20:
-      w1[1] = 0x01;
-      break;
-
-    case 21:
-      w1[1] = w1[1] | 0x0100;
-      break;
-
-    case 22:
-      w1[1] = w1[1] | 0x010000;
-      break;
-
-    case 23:
-      w1[1] = w1[1] | 0x01000000;
-      break;
-
-    case 24:
-      w1[2] = 0x01;
-      break;
-
-    case 25:
-      w1[2] = w1[2] | 0x0100;
-      break;
-
-    case 26:
-      w1[2] = w1[2] | 0x010000;
-      break;
-
-    case 27:
-      w1[2] = w1[2] | 0x01000000;
-      break;
-
-    case 28:
-      w1[3] = 0x01;
-      break;
-
-    case 29:
-      w1[3] = w1[3] | 0x0100;
-      break;
-
-    case 30:
-      w1[3] = w1[3] | 0x010000;
-      break;
-
-    case 31:
-      w1[3] = w1[3] | 0x01000000;
-      break;
-
-    case 32:
-      w2[0] = 0x01;
-      break;
-
-    case 33:
-      w2[0] = w2[0] | 0x0100;
-      break;
-
-    case 34:
-      w2[0] = w2[0] | 0x010000;
-      break;
-
-    case 35:
-      w2[0] = w2[0] | 0x01000000;
-      break;
-
-    case 36:
-      w2[1] = 0x01;
-      break;
-
-    case 37:
-      w2[1] = w2[1] | 0x0100;
-      break;
-
-    case 38:
-      w2[1] = w2[1] | 0x010000;
-      break;
-
-    case 39:
-      w2[1] = w2[1] | 0x01000000;
-      break;
-
-    case 40:
-      w2[2] = 0x01;
-      break;
-
-    case 41:
-      w2[2] = w2[2] | 0x0100;
-      break;
-
-    case 42:
-      w2[2] = w2[2] | 0x010000;
-      break;
-
-    case 43:
-      w2[2] = w2[2] | 0x01000000;
-      break;
-
-    case 44:
-      w2[3] = 0x01;
-      break;
-
-    case 45:
-      w2[3] = w2[3] | 0x0100;
-      break;
-
-    case 46:
-      w2[3] = w2[3] | 0x010000;
-      break;
-
-    case 47:
-      w2[3] = w2[3] | 0x01000000;
-      break;
-  }
+  w0[0] = w[ 0];
+  w0[1] = w[ 1];
+  w0[2] = w[ 2];
+  w0[3] = w[ 3];
+  w1[0] = w[ 4];
+  w1[1] = w[ 5];
+  w1[2] = w[ 6];
+  w1[3] = w[ 7];
+  w2[0] = w[ 8];
+  w2[1] = w[ 9];
+  w2[2] = w[10];
+  w2[3] = w[11];
 }
 
 inline void append_0x01_4x4_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 offset)
 {
-  switch (offset)
-  {
-    case 0:
-      w0[0] = 0x01;
-      break;
+  u32 w[16];
 
-    case 1:
-      w0[0] = w0[0] | 0x0100;
-      break;
+  w[ 0] = w0[0];
+  w[ 1] = w0[1];
+  w[ 2] = w0[2];
+  w[ 3] = w0[3];
+  w[ 4] = w1[0];
+  w[ 5] = w1[1];
+  w[ 6] = w1[2];
+  w[ 7] = w1[3];
+  w[ 8] = w2[0];
+  w[ 9] = w2[1];
+  w[10] = w2[2];
+  w[11] = w2[3];
+  w[12] = w3[0];
+  w[13] = w3[1];
+  w[14] = w3[2];
+  w[15] = w3[3];
 
-    case 2:
-      w0[0] = w0[0] | 0x010000;
-      break;
+  u8 *w_ptr = (u8 *) w;
 
-    case 3:
-      w0[0] = w0[0] | 0x01000000;
-      break;
+  w_ptr[offset] = 0x01;
 
-    case 4:
-      w0[1] = 0x01;
-      break;
+  w0[0] = w[ 0];
+  w0[1] = w[ 1];
+  w0[2] = w[ 2];
+  w0[3] = w[ 3];
+  w1[0] = w[ 4];
+  w1[1] = w[ 5];
+  w1[2] = w[ 6];
+  w1[3] = w[ 7];
+  w2[0] = w[ 8];
+  w2[1] = w[ 9];
+  w2[2] = w[10];
+  w2[3] = w[11];
+  w3[0] = w[12];
+  w3[1] = w[13];
+  w3[2] = w[14];
+  w3[3] = w[15];
+}
 
-    case 5:
-      w0[1] = w0[1] | 0x0100;
-      break;
+inline void append_0x02_1x4_S (u32 w0[4], const u32 offset)
+{
+  u32 w[4];
 
-    case 6:
-      w0[1] = w0[1] | 0x010000;
-      break;
+  w[0] = w0[0];
+  w[1] = w0[1];
+  w[2] = w0[2];
+  w[3] = w0[3];
 
-    case 7:
-      w0[1] = w0[1] | 0x01000000;
-      break;
+  u8 *w_ptr = (u8 *) w;
 
-    case 8:
-      w0[2] = 0x01;
-      break;
+  w_ptr[offset] = 0x02;
 
-    case 9:
-      w0[2] = w0[2] | 0x0100;
-      break;
-
-    case 10:
-      w0[2] = w0[2] | 0x010000;
-      break;
-
-    case 11:
-      w0[2] = w0[2] | 0x01000000;
-      break;
-
-    case 12:
-      w0[3] = 0x01;
-      break;
-
-    case 13:
-      w0[3] = w0[3] | 0x0100;
-      break;
-
-    case 14:
-      w0[3] = w0[3] | 0x010000;
-      break;
-
-    case 15:
-      w0[3] = w0[3] | 0x01000000;
-      break;
-
-    case 16:
-      w1[0] = 0x01;
-      break;
-
-    case 17:
-      w1[0] = w1[0] | 0x0100;
-      break;
-
-    case 18:
-      w1[0] = w1[0] | 0x010000;
-      break;
-
-    case 19:
-      w1[0] = w1[0] | 0x01000000;
-      break;
-
-    case 20:
-      w1[1] = 0x01;
-      break;
-
-    case 21:
-      w1[1] = w1[1] | 0x0100;
-      break;
-
-    case 22:
-      w1[1] = w1[1] | 0x010000;
-      break;
-
-    case 23:
-      w1[1] = w1[1] | 0x01000000;
-      break;
-
-    case 24:
-      w1[2] = 0x01;
-      break;
-
-    case 25:
-      w1[2] = w1[2] | 0x0100;
-      break;
-
-    case 26:
-      w1[2] = w1[2] | 0x010000;
-      break;
-
-    case 27:
-      w1[2] = w1[2] | 0x01000000;
-      break;
-
-    case 28:
-      w1[3] = 0x01;
-      break;
-
-    case 29:
-      w1[3] = w1[3] | 0x0100;
-      break;
-
-    case 30:
-      w1[3] = w1[3] | 0x010000;
-      break;
-
-    case 31:
-      w1[3] = w1[3] | 0x01000000;
-      break;
-
-    case 32:
-      w2[0] = 0x01;
-      break;
-
-    case 33:
-      w2[0] = w2[0] | 0x0100;
-      break;
-
-    case 34:
-      w2[0] = w2[0] | 0x010000;
-      break;
-
-    case 35:
-      w2[0] = w2[0] | 0x01000000;
-      break;
-
-    case 36:
-      w2[1] = 0x01;
-      break;
-
-    case 37:
-      w2[1] = w2[1] | 0x0100;
-      break;
-
-    case 38:
-      w2[1] = w2[1] | 0x010000;
-      break;
-
-    case 39:
-      w2[1] = w2[1] | 0x01000000;
-      break;
-
-    case 40:
-      w2[2] = 0x01;
-      break;
-
-    case 41:
-      w2[2] = w2[2] | 0x0100;
-      break;
-
-    case 42:
-      w2[2] = w2[2] | 0x010000;
-      break;
-
-    case 43:
-      w2[2] = w2[2] | 0x01000000;
-      break;
-
-    case 44:
-      w2[3] = 0x01;
-      break;
-
-    case 45:
-      w2[3] = w2[3] | 0x0100;
-      break;
-
-    case 46:
-      w2[3] = w2[3] | 0x010000;
-      break;
-
-    case 47:
-      w2[3] = w2[3] | 0x01000000;
-      break;
-
-    case 48:
-      w3[0] = 0x01;
-      break;
-
-    case 49:
-      w3[0] = w3[0] | 0x0100;
-      break;
-
-    case 50:
-      w3[0] = w3[0] | 0x010000;
-      break;
-
-    case 51:
-      w3[0] = w3[0] | 0x01000000;
-      break;
-
-    case 52:
-      w3[1] = 0x01;
-      break;
-
-    case 53:
-      w3[1] = w3[1] | 0x0100;
-      break;
-
-    case 54:
-      w3[1] = w3[1] | 0x010000;
-      break;
-
-    case 55:
-      w3[1] = w3[1] | 0x01000000;
-      break;
-
-    case 56:
-      w3[2] = 0x01;
-      break;
-
-    case 57:
-      w3[2] = w3[2] | 0x0100;
-      break;
-
-    case 58:
-      w3[2] = w3[2] | 0x010000;
-      break;
-
-    case 59:
-      w3[2] = w3[2] | 0x01000000;
-      break;
-
-    case 60:
-      w3[3] = 0x01;
-      break;
-
-    case 61:
-      w3[3] = w3[3] | 0x0100;
-      break;
-
-    case 62:
-      w3[3] = w3[3] | 0x010000;
-      break;
-
-    case 63:
-      w3[3] = w3[3] | 0x01000000;
-      break;
-  }
+  w0[0] = w[0];
+  w0[1] = w[1];
+  w0[2] = w[2];
+  w0[3] = w[3];
 }
 
 inline void append_0x02_2x4_S (u32 w0[4], u32 w1[4], const u32 offset)
 {
-  switch (offset)
-  {
-    case 0:
-      w0[0] = 0x02;
-      break;
+  u32 w[8];
 
-    case 1:
-      w0[0] = w0[0] | 0x0200;
-      break;
+  w[0] = w0[0];
+  w[1] = w0[1];
+  w[2] = w0[2];
+  w[3] = w0[3];
+  w[4] = w1[0];
+  w[5] = w1[1];
+  w[6] = w1[2];
+  w[7] = w1[3];
 
-    case 2:
-      w0[0] = w0[0] | 0x020000;
-      break;
+  u8 *w_ptr = (u8 *) w;
 
-    case 3:
-      w0[0] = w0[0] | 0x02000000;
-      break;
+  w_ptr[offset] = 0x02;
 
-    case 4:
-      w0[1] = 0x02;
-      break;
-
-    case 5:
-      w0[1] = w0[1] | 0x0200;
-      break;
-
-    case 6:
-      w0[1] = w0[1] | 0x020000;
-      break;
-
-    case 7:
-      w0[1] = w0[1] | 0x02000000;
-      break;
-
-    case 8:
-      w0[2] = 0x02;
-      break;
-
-    case 9:
-      w0[2] = w0[2] | 0x0200;
-      break;
-
-    case 10:
-      w0[2] = w0[2] | 0x020000;
-      break;
-
-    case 11:
-      w0[2] = w0[2] | 0x02000000;
-      break;
-
-    case 12:
-      w0[3] = 0x02;
-      break;
-
-    case 13:
-      w0[3] = w0[3] | 0x0200;
-      break;
-
-    case 14:
-      w0[3] = w0[3] | 0x020000;
-      break;
-
-    case 15:
-      w0[3] = w0[3] | 0x02000000;
-      break;
-
-    case 16:
-      w1[0] = 0x02;
-      break;
-
-    case 17:
-      w1[0] = w1[0] | 0x0200;
-      break;
-
-    case 18:
-      w1[0] = w1[0] | 0x020000;
-      break;
-
-    case 19:
-      w1[0] = w1[0] | 0x02000000;
-      break;
-
-    case 20:
-      w1[1] = 0x02;
-      break;
-
-    case 21:
-      w1[1] = w1[1] | 0x0200;
-      break;
-
-    case 22:
-      w1[1] = w1[1] | 0x020000;
-      break;
-
-    case 23:
-      w1[1] = w1[1] | 0x02000000;
-      break;
-
-    case 24:
-      w1[2] = 0x02;
-      break;
-
-    case 25:
-      w1[2] = w1[2] | 0x0200;
-      break;
-
-    case 26:
-      w1[2] = w1[2] | 0x020000;
-      break;
-
-    case 27:
-      w1[2] = w1[2] | 0x02000000;
-      break;
-
-    case 28:
-      w1[3] = 0x02;
-      break;
-
-    case 29:
-      w1[3] = w1[3] | 0x0200;
-      break;
-
-    case 30:
-      w1[3] = w1[3] | 0x020000;
-      break;
-
-    case 31:
-      w1[3] = w1[3] | 0x02000000;
-      break;
-  }
+  w0[0] = w[0];
+  w0[1] = w[1];
+  w0[2] = w[2];
+  w0[3] = w[3];
+  w1[0] = w[4];
+  w1[1] = w[5];
+  w1[2] = w[6];
+  w1[3] = w[7];
 }
 
 inline void append_0x02_3x4_S (u32 w0[4], u32 w1[4], u32 w2[4], const u32 offset)
 {
-  switch (offset)
-  {
-    case 0:
-      w0[0] = 0x02;
-      break;
+  u32 w[12];
 
-    case 1:
-      w0[0] = w0[0] | 0x0200;
-      break;
+  w[ 0] = w0[0];
+  w[ 1] = w0[1];
+  w[ 2] = w0[2];
+  w[ 3] = w0[3];
+  w[ 4] = w1[0];
+  w[ 5] = w1[1];
+  w[ 6] = w1[2];
+  w[ 7] = w1[3];
+  w[ 8] = w2[0];
+  w[ 9] = w2[1];
+  w[10] = w2[2];
+  w[11] = w2[3];
 
-    case 2:
-      w0[0] = w0[0] | 0x020000;
-      break;
+  u8 *w_ptr = (u8 *) w;
 
-    case 3:
-      w0[0] = w0[0] | 0x02000000;
-      break;
+  w_ptr[offset] = 0x02;
 
-    case 4:
-      w0[1] = 0x02;
-      break;
+  w0[0] = w[ 0];
+  w0[1] = w[ 1];
+  w0[2] = w[ 2];
+  w0[3] = w[ 3];
+  w1[0] = w[ 4];
+  w1[1] = w[ 5];
+  w1[2] = w[ 6];
+  w1[3] = w[ 7];
+  w2[0] = w[ 8];
+  w2[1] = w[ 9];
+  w2[2] = w[10];
+  w2[3] = w[11];
+}
 
-    case 5:
-      w0[1] = w0[1] | 0x0200;
-      break;
+inline void append_0x02_4x4_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 offset)
+{
+  u32 w[16];
 
-    case 6:
-      w0[1] = w0[1] | 0x020000;
-      break;
+  w[ 0] = w0[0];
+  w[ 1] = w0[1];
+  w[ 2] = w0[2];
+  w[ 3] = w0[3];
+  w[ 4] = w1[0];
+  w[ 5] = w1[1];
+  w[ 6] = w1[2];
+  w[ 7] = w1[3];
+  w[ 8] = w2[0];
+  w[ 9] = w2[1];
+  w[10] = w2[2];
+  w[11] = w2[3];
+  w[12] = w3[0];
+  w[13] = w3[1];
+  w[14] = w3[2];
+  w[15] = w3[3];
 
-    case 7:
-      w0[1] = w0[1] | 0x02000000;
-      break;
+  u8 *w_ptr = (u8 *) w;
 
-    case 8:
-      w0[2] = 0x02;
-      break;
+  w_ptr[offset] = 0x02;
 
-    case 9:
-      w0[2] = w0[2] | 0x0200;
-      break;
-
-    case 10:
-      w0[2] = w0[2] | 0x020000;
-      break;
-
-    case 11:
-      w0[2] = w0[2] | 0x02000000;
-      break;
-
-    case 12:
-      w0[3] = 0x02;
-      break;
-
-    case 13:
-      w0[3] = w0[3] | 0x0200;
-      break;
-
-    case 14:
-      w0[3] = w0[3] | 0x020000;
-      break;
-
-    case 15:
-      w0[3] = w0[3] | 0x02000000;
-      break;
-
-    case 16:
-      w1[0] = 0x02;
-      break;
-
-    case 17:
-      w1[0] = w1[0] | 0x0200;
-      break;
-
-    case 18:
-      w1[0] = w1[0] | 0x020000;
-      break;
-
-    case 19:
-      w1[0] = w1[0] | 0x02000000;
-      break;
-
-    case 20:
-      w1[1] = 0x02;
-      break;
-
-    case 21:
-      w1[1] = w1[1] | 0x0200;
-      break;
-
-    case 22:
-      w1[1] = w1[1] | 0x020000;
-      break;
-
-    case 23:
-      w1[1] = w1[1] | 0x02000000;
-      break;
-
-    case 24:
-      w1[2] = 0x02;
-      break;
-
-    case 25:
-      w1[2] = w1[2] | 0x0200;
-      break;
-
-    case 26:
-      w1[2] = w1[2] | 0x020000;
-      break;
-
-    case 27:
-      w1[2] = w1[2] | 0x02000000;
-      break;
-
-    case 28:
-      w1[3] = 0x02;
-      break;
-
-    case 29:
-      w1[3] = w1[3] | 0x0200;
-      break;
-
-    case 30:
-      w1[3] = w1[3] | 0x020000;
-      break;
-
-    case 31:
-      w1[3] = w1[3] | 0x02000000;
-      break;
-
-    case 32:
-      w2[0] = 0x02;
-      break;
-
-    case 33:
-      w2[0] = w2[0] | 0x0200;
-      break;
-
-    case 34:
-      w2[0] = w2[0] | 0x020000;
-      break;
-
-    case 35:
-      w2[0] = w2[0] | 0x02000000;
-      break;
-
-    case 36:
-      w2[1] = 0x02;
-      break;
-
-    case 37:
-      w2[1] = w2[1] | 0x0200;
-      break;
-
-    case 38:
-      w2[1] = w2[1] | 0x020000;
-      break;
-
-    case 39:
-      w2[1] = w2[1] | 0x02000000;
-      break;
-
-    case 40:
-      w2[2] = 0x02;
-      break;
-
-    case 41:
-      w2[2] = w2[2] | 0x0200;
-      break;
-
-    case 42:
-      w2[2] = w2[2] | 0x020000;
-      break;
-
-    case 43:
-      w2[2] = w2[2] | 0x02000000;
-      break;
-
-    case 44:
-      w2[3] = 0x02;
-      break;
-
-    case 45:
-      w2[3] = w2[3] | 0x0200;
-      break;
-
-    case 46:
-      w2[3] = w2[3] | 0x020000;
-      break;
-
-    case 47:
-      w2[3] = w2[3] | 0x02000000;
-      break;
-  }
+  w0[0] = w[ 0];
+  w0[1] = w[ 1];
+  w0[2] = w[ 2];
+  w0[3] = w[ 3];
+  w1[0] = w[ 4];
+  w1[1] = w[ 5];
+  w1[2] = w[ 6];
+  w1[3] = w[ 7];
+  w2[0] = w[ 8];
+  w2[1] = w[ 9];
+  w2[2] = w[10];
+  w2[3] = w[11];
+  w3[0] = w[12];
+  w3[1] = w[13];
+  w3[2] = w[14];
+  w3[3] = w[15];
 }
 
 inline void append_0x80_1x4_S (u32 w0[4], const u32 offset)
 {
-  switch (offset)
-  {
-    case 0:
-      w0[0]  = 0x80;
-      break;
+  u32 w[4];
 
-    case 1:
-      w0[0] = w0[0] | 0x8000;
-      break;
+  w[0] = w0[0];
+  w[1] = w0[1];
+  w[2] = w0[2];
+  w[3] = w0[3];
 
-    case 2:
-      w0[0] = w0[0] | 0x800000;
-      break;
+  u8 *w_ptr = (u8 *) w;
 
-    case 3:
-      w0[0] = w0[0] | 0x80000000;
-      break;
+  w_ptr[offset] = 0x80;
 
-    case 4:
-      w0[1] = 0x80;
-      break;
-
-    case 5:
-      w0[1] = w0[1] | 0x8000;
-      break;
-
-    case 6:
-      w0[1] = w0[1] | 0x800000;
-      break;
-
-    case 7:
-      w0[1] = w0[1] | 0x80000000;
-      break;
-
-    case 8:
-      w0[2] = 0x80;
-      break;
-
-    case 9:
-      w0[2] = w0[2] | 0x8000;
-      break;
-
-    case 10:
-      w0[2] = w0[2] | 0x800000;
-      break;
-
-    case 11:
-      w0[2] = w0[2] | 0x80000000;
-      break;
-
-    case 12:
-      w0[3] = 0x80;
-      break;
-
-    case 13:
-      w0[3] = w0[3] | 0x8000;
-      break;
-
-    case 14:
-      w0[3] = w0[3] | 0x800000;
-      break;
-
-    case 15:
-      w0[3] = w0[3] | 0x80000000;
-      break;
-  }
+  w0[0] = w[0];
+  w0[1] = w[1];
+  w0[2] = w[2];
+  w0[3] = w[3];
 }
 
 inline void append_0x80_2x4_S (u32 w0[4], u32 w1[4], const u32 offset)
 {
-  switch (offset)
-  {
-    case 0:
-      w0[0] = 0x80;
-      break;
+  u32 w[8];
 
-    case 1:
-      w0[0] = w0[0] | 0x8000;
-      break;
+  w[0] = w0[0];
+  w[1] = w0[1];
+  w[2] = w0[2];
+  w[3] = w0[3];
+  w[4] = w1[0];
+  w[5] = w1[1];
+  w[6] = w1[2];
+  w[7] = w1[3];
 
-    case 2:
-      w0[0] = w0[0] | 0x800000;
-      break;
+  u8 *w_ptr = (u8 *) w;
 
-    case 3:
-      w0[0] = w0[0] | 0x80000000;
-      break;
+  w_ptr[offset] = 0x80;
 
-    case 4:
-      w0[1] = 0x80;
-      break;
-
-    case 5:
-      w0[1] = w0[1] | 0x8000;
-      break;
-
-    case 6:
-      w0[1] = w0[1] | 0x800000;
-      break;
-
-    case 7:
-      w0[1] = w0[1] | 0x80000000;
-      break;
-
-    case 8:
-      w0[2] = 0x80;
-      break;
-
-    case 9:
-      w0[2] = w0[2] | 0x8000;
-      break;
-
-    case 10:
-      w0[2] = w0[2] | 0x800000;
-      break;
-
-    case 11:
-      w0[2] = w0[2] | 0x80000000;
-      break;
-
-    case 12:
-      w0[3] = 0x80;
-      break;
-
-    case 13:
-      w0[3] = w0[3] | 0x8000;
-      break;
-
-    case 14:
-      w0[3] = w0[3] | 0x800000;
-      break;
-
-    case 15:
-      w0[3] = w0[3] | 0x80000000;
-      break;
-
-    case 16:
-      w1[0] = 0x80;
-      break;
-
-    case 17:
-      w1[0] = w1[0] | 0x8000;
-      break;
-
-    case 18:
-      w1[0] = w1[0] | 0x800000;
-      break;
-
-    case 19:
-      w1[0] = w1[0] | 0x80000000;
-      break;
-
-    case 20:
-      w1[1] = 0x80;
-      break;
-
-    case 21:
-      w1[1] = w1[1] | 0x8000;
-      break;
-
-    case 22:
-      w1[1] = w1[1] | 0x800000;
-      break;
-
-    case 23:
-      w1[1] = w1[1] | 0x80000000;
-      break;
-
-    case 24:
-      w1[2] = 0x80;
-      break;
-
-    case 25:
-      w1[2] = w1[2] | 0x8000;
-      break;
-
-    case 26:
-      w1[2] = w1[2] | 0x800000;
-      break;
-
-    case 27:
-      w1[2] = w1[2] | 0x80000000;
-      break;
-
-    case 28:
-      w1[3] = 0x80;
-      break;
-
-    case 29:
-      w1[3] = w1[3] | 0x8000;
-      break;
-
-    case 30:
-      w1[3] = w1[3] | 0x800000;
-      break;
-
-    case 31:
-      w1[3] = w1[3] | 0x80000000;
-      break;
-  }
+  w0[0] = w[0];
+  w0[1] = w[1];
+  w0[2] = w[2];
+  w0[3] = w[3];
+  w1[0] = w[4];
+  w1[1] = w[5];
+  w1[2] = w[6];
+  w1[3] = w[7];
 }
 
 inline void append_0x80_3x4_S (u32 w0[4], u32 w1[4], u32 w2[4], const u32 offset)
 {
-  switch (offset)
-  {
-    case 0:
-      w0[0] = 0x80;
-      break;
+  u32 w[12];
 
-    case 1:
-      w0[0] = w0[0] | 0x8000;
-      break;
+  w[ 0] = w0[0];
+  w[ 1] = w0[1];
+  w[ 2] = w0[2];
+  w[ 3] = w0[3];
+  w[ 4] = w1[0];
+  w[ 5] = w1[1];
+  w[ 6] = w1[2];
+  w[ 7] = w1[3];
+  w[ 8] = w2[0];
+  w[ 9] = w2[1];
+  w[10] = w2[2];
+  w[11] = w2[3];
 
-    case 2:
-      w0[0] = w0[0] | 0x800000;
-      break;
+  u8 *w_ptr = (u8 *) w;
 
-    case 3:
-      w0[0] = w0[0] | 0x80000000;
-      break;
+  w_ptr[offset] = 0x80;
 
-    case 4:
-      w0[1] = 0x80;
-      break;
-
-    case 5:
-      w0[1] = w0[1] | 0x8000;
-      break;
-
-    case 6:
-      w0[1] = w0[1] | 0x800000;
-      break;
-
-    case 7:
-      w0[1] = w0[1] | 0x80000000;
-      break;
-
-    case 8:
-      w0[2] = 0x80;
-      break;
-
-    case 9:
-      w0[2] = w0[2] | 0x8000;
-      break;
-
-    case 10:
-      w0[2] = w0[2] | 0x800000;
-      break;
-
-    case 11:
-      w0[2] = w0[2] | 0x80000000;
-      break;
-
-    case 12:
-      w0[3] = 0x80;
-      break;
-
-    case 13:
-      w0[3] = w0[3] | 0x8000;
-      break;
-
-    case 14:
-      w0[3] = w0[3] | 0x800000;
-      break;
-
-    case 15:
-      w0[3] = w0[3] | 0x80000000;
-      break;
-
-    case 16:
-      w1[0] = 0x80;
-      break;
-
-    case 17:
-      w1[0] = w1[0] | 0x8000;
-      break;
-
-    case 18:
-      w1[0] = w1[0] | 0x800000;
-      break;
-
-    case 19:
-      w1[0] = w1[0] | 0x80000000;
-      break;
-
-    case 20:
-      w1[1] = 0x80;
-      break;
-
-    case 21:
-      w1[1] = w1[1] | 0x8000;
-      break;
-
-    case 22:
-      w1[1] = w1[1] | 0x800000;
-      break;
-
-    case 23:
-      w1[1] = w1[1] | 0x80000000;
-      break;
-
-    case 24:
-      w1[2] = 0x80;
-      break;
-
-    case 25:
-      w1[2] = w1[2] | 0x8000;
-      break;
-
-    case 26:
-      w1[2] = w1[2] | 0x800000;
-      break;
-
-    case 27:
-      w1[2] = w1[2] | 0x80000000;
-      break;
-
-    case 28:
-      w1[3] = 0x80;
-      break;
-
-    case 29:
-      w1[3] = w1[3] | 0x8000;
-      break;
-
-    case 30:
-      w1[3] = w1[3] | 0x800000;
-      break;
-
-    case 31:
-      w1[3] = w1[3] | 0x80000000;
-      break;
-
-    case 32:
-      w2[0] = 0x80;
-      break;
-
-    case 33:
-      w2[0] = w2[0] | 0x8000;
-      break;
-
-    case 34:
-      w2[0] = w2[0] | 0x800000;
-      break;
-
-    case 35:
-      w2[0] = w2[0] | 0x80000000;
-      break;
-
-    case 36:
-      w2[1] = 0x80;
-      break;
-
-    case 37:
-      w2[1] = w2[1] | 0x8000;
-      break;
-
-    case 38:
-      w2[1] = w2[1] | 0x800000;
-      break;
-
-    case 39:
-      w2[1] = w2[1] | 0x80000000;
-      break;
-
-    case 40:
-      w2[2] = 0x80;
-      break;
-
-    case 41:
-      w2[2] = w2[2] | 0x8000;
-      break;
-
-    case 42:
-      w2[2] = w2[2] | 0x800000;
-      break;
-
-    case 43:
-      w2[2] = w2[2] | 0x80000000;
-      break;
-
-    case 44:
-      w2[3] = 0x80;
-      break;
-
-    case 45:
-      w2[3] = w2[3] | 0x8000;
-      break;
-
-    case 46:
-      w2[3] = w2[3] | 0x800000;
-      break;
-
-    case 47:
-      w2[3] = w2[3] | 0x80000000;
-      break;
-  }
+  w0[0] = w[ 0];
+  w0[1] = w[ 1];
+  w0[2] = w[ 2];
+  w0[3] = w[ 3];
+  w1[0] = w[ 4];
+  w1[1] = w[ 5];
+  w1[2] = w[ 6];
+  w1[3] = w[ 7];
+  w2[0] = w[ 8];
+  w2[1] = w[ 9];
+  w2[2] = w[10];
+  w2[3] = w[11];
 }
 
 inline void append_0x80_4x4_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 offset)
 {
-  switch (offset)
-  {
-    case 0:
-      w0[0] = 0x80;
-      break;
+  u32 w[16];
 
-    case 1:
-      w0[0] = w0[0] | 0x8000;
-      break;
+  w[ 0] = w0[0];
+  w[ 1] = w0[1];
+  w[ 2] = w0[2];
+  w[ 3] = w0[3];
+  w[ 4] = w1[0];
+  w[ 5] = w1[1];
+  w[ 6] = w1[2];
+  w[ 7] = w1[3];
+  w[ 8] = w2[0];
+  w[ 9] = w2[1];
+  w[10] = w2[2];
+  w[11] = w2[3];
+  w[12] = w3[0];
+  w[13] = w3[1];
+  w[14] = w3[2];
+  w[15] = w3[3];
 
-    case 2:
-      w0[0] = w0[0] | 0x800000;
-      break;
+  u8 *w_ptr = (u8 *) w;
 
-    case 3:
-      w0[0] = w0[0] | 0x80000000;
-      break;
+  w_ptr[offset] = 0x80;
 
-    case 4:
-      w0[1] = 0x80;
-      break;
-
-    case 5:
-      w0[1] = w0[1] | 0x8000;
-      break;
-
-    case 6:
-      w0[1] = w0[1] | 0x800000;
-      break;
-
-    case 7:
-      w0[1] = w0[1] | 0x80000000;
-      break;
-
-    case 8:
-      w0[2] = 0x80;
-      break;
-
-    case 9:
-      w0[2] = w0[2] | 0x8000;
-      break;
-
-    case 10:
-      w0[2] = w0[2] | 0x800000;
-      break;
-
-    case 11:
-      w0[2] = w0[2] | 0x80000000;
-      break;
-
-    case 12:
-      w0[3] = 0x80;
-      break;
-
-    case 13:
-      w0[3] = w0[3] | 0x8000;
-      break;
-
-    case 14:
-      w0[3] = w0[3] | 0x800000;
-      break;
-
-    case 15:
-      w0[3] = w0[3] | 0x80000000;
-      break;
-
-    case 16:
-      w1[0] = 0x80;
-      break;
-
-    case 17:
-      w1[0] = w1[0] | 0x8000;
-      break;
-
-    case 18:
-      w1[0] = w1[0] | 0x800000;
-      break;
-
-    case 19:
-      w1[0] = w1[0] | 0x80000000;
-      break;
-
-    case 20:
-      w1[1] = 0x80;
-      break;
-
-    case 21:
-      w1[1] = w1[1] | 0x8000;
-      break;
-
-    case 22:
-      w1[1] = w1[1] | 0x800000;
-      break;
-
-    case 23:
-      w1[1] = w1[1] | 0x80000000;
-      break;
-
-    case 24:
-      w1[2] = 0x80;
-      break;
-
-    case 25:
-      w1[2] = w1[2] | 0x8000;
-      break;
-
-    case 26:
-      w1[2] = w1[2] | 0x800000;
-      break;
-
-    case 27:
-      w1[2] = w1[2] | 0x80000000;
-      break;
-
-    case 28:
-      w1[3] = 0x80;
-      break;
-
-    case 29:
-      w1[3] = w1[3] | 0x8000;
-      break;
-
-    case 30:
-      w1[3] = w1[3] | 0x800000;
-      break;
-
-    case 31:
-      w1[3] = w1[3] | 0x80000000;
-      break;
-
-    case 32:
-      w2[0] = 0x80;
-      break;
-
-    case 33:
-      w2[0] = w2[0] | 0x8000;
-      break;
-
-    case 34:
-      w2[0] = w2[0] | 0x800000;
-      break;
-
-    case 35:
-      w2[0] = w2[0] | 0x80000000;
-      break;
-
-    case 36:
-      w2[1] = 0x80;
-      break;
-
-    case 37:
-      w2[1] = w2[1] | 0x8000;
-      break;
-
-    case 38:
-      w2[1] = w2[1] | 0x800000;
-      break;
-
-    case 39:
-      w2[1] = w2[1] | 0x80000000;
-      break;
-
-    case 40:
-      w2[2] = 0x80;
-      break;
-
-    case 41:
-      w2[2] = w2[2] | 0x8000;
-      break;
-
-    case 42:
-      w2[2] = w2[2] | 0x800000;
-      break;
-
-    case 43:
-      w2[2] = w2[2] | 0x80000000;
-      break;
-
-    case 44:
-      w2[3] = 0x80;
-      break;
-
-    case 45:
-      w2[3] = w2[3] | 0x8000;
-      break;
-
-    case 46:
-      w2[3] = w2[3] | 0x800000;
-      break;
-
-    case 47:
-      w2[3] = w2[3] | 0x80000000;
-      break;
-
-    case 48:
-      w3[0] = 0x80;
-      break;
-
-    case 49:
-      w3[0] = w3[0] | 0x8000;
-      break;
-
-    case 50:
-      w3[0] = w3[0] | 0x800000;
-      break;
-
-    case 51:
-      w3[0] = w3[0] | 0x80000000;
-      break;
-
-    case 52:
-      w3[1] = 0x80;
-      break;
-
-    case 53:
-      w3[1] = w3[1] | 0x8000;
-      break;
-
-    case 54:
-      w3[1] = w3[1] | 0x800000;
-      break;
-
-    case 55:
-      w3[1] = w3[1] | 0x80000000;
-      break;
-
-    case 56:
-      w3[2] = 0x80;
-      break;
-
-    case 57:
-      w3[2] = w3[2] | 0x8000;
-      break;
-
-    case 58:
-      w3[2] = w3[2] | 0x800000;
-      break;
-
-    case 59:
-      w3[2] = w3[2] | 0x80000000;
-      break;
-
-    case 60:
-      w3[3] = 0x80;
-      break;
-
-    case 61:
-      w3[3] = w3[3] | 0x8000;
-      break;
-
-    case 62:
-      w3[3] = w3[3] | 0x800000;
-      break;
-
-    case 63:
-      w3[3] = w3[3] | 0x80000000;
-      break;
-  }
+  w0[0] = w[ 0];
+  w0[1] = w[ 1];
+  w0[2] = w[ 2];
+  w0[3] = w[ 3];
+  w1[0] = w[ 4];
+  w1[1] = w[ 5];
+  w1[2] = w[ 6];
+  w1[3] = w[ 7];
+  w2[0] = w[ 8];
+  w2[1] = w[ 9];
+  w2[2] = w[10];
+  w2[3] = w[11];
+  w3[0] = w[12];
+  w3[1] = w[13];
+  w3[2] = w[14];
+  w3[3] = w[15];
 }
 
 inline void truncate_block_S (u32 w[4], const u32 len)