From: Jens Steube <jens.steube@gmail.com>
Date: Thu, 4 Feb 2016 14:47:52 +0000 (+0100)
Subject: Some cleanups
X-Git-Tag: v3.00-beta~404
X-Git-Url: https://www.flypig.org.uk/git/?a=commitdiff_plain;h=2381af313db6f003d0e5316158509281782e6784;p=hashcat.git

Some cleanups
---

diff --git a/OpenCL/kernel_aes256.c b/OpenCL/kernel_aes256.c
new file mode 100644
index 0000000..1127165
--- /dev/null
+++ b/OpenCL/kernel_aes256.c
@@ -0,0 +1,1048 @@
+/**
+ * Author......: Jens Steube <jens.steube@gmail.com>
+ * License.....: MIT
+ */
+
+__constant u32 te0[256] =
+{
+  0xc66363a5, 0xf87c7c84, 0xee777799, 0xf67b7b8d,
+  0xfff2f20d, 0xd66b6bbd, 0xde6f6fb1, 0x91c5c554,
+  0x60303050, 0x02010103, 0xce6767a9, 0x562b2b7d,
+  0xe7fefe19, 0xb5d7d762, 0x4dababe6, 0xec76769a,
+  0x8fcaca45, 0x1f82829d, 0x89c9c940, 0xfa7d7d87,
+  0xeffafa15, 0xb25959eb, 0x8e4747c9, 0xfbf0f00b,
+  0x41adadec, 0xb3d4d467, 0x5fa2a2fd, 0x45afafea,
+  0x239c9cbf, 0x53a4a4f7, 0xe4727296, 0x9bc0c05b,
+  0x75b7b7c2, 0xe1fdfd1c, 0x3d9393ae, 0x4c26266a,
+  0x6c36365a, 0x7e3f3f41, 0xf5f7f702, 0x83cccc4f,
+  0x6834345c, 0x51a5a5f4, 0xd1e5e534, 0xf9f1f108,
+  0xe2717193, 0xabd8d873, 0x62313153, 0x2a15153f,
+  0x0804040c, 0x95c7c752, 0x46232365, 0x9dc3c35e,
+  0x30181828, 0x379696a1, 0x0a05050f, 0x2f9a9ab5,
+  0x0e070709, 0x24121236, 0x1b80809b, 0xdfe2e23d,
+  0xcdebeb26, 0x4e272769, 0x7fb2b2cd, 0xea75759f,
+  0x1209091b, 0x1d83839e, 0x582c2c74, 0x341a1a2e,
+  0x361b1b2d, 0xdc6e6eb2, 0xb45a5aee, 0x5ba0a0fb,
+  0xa45252f6, 0x763b3b4d, 0xb7d6d661, 0x7db3b3ce,
+  0x5229297b, 0xdde3e33e, 0x5e2f2f71, 0x13848497,
+  0xa65353f5, 0xb9d1d168, 0x00000000, 0xc1eded2c,
+  0x40202060, 0xe3fcfc1f, 0x79b1b1c8, 0xb65b5bed,
+  0xd46a6abe, 0x8dcbcb46, 0x67bebed9, 0x7239394b,
+  0x944a4ade, 0x984c4cd4, 0xb05858e8, 0x85cfcf4a,
+  0xbbd0d06b, 0xc5efef2a, 0x4faaaae5, 0xedfbfb16,
+  0x864343c5, 0x9a4d4dd7, 0x66333355, 0x11858594,
+  0x8a4545cf, 0xe9f9f910, 0x04020206, 0xfe7f7f81,
+  0xa05050f0, 0x783c3c44, 0x259f9fba, 0x4ba8a8e3,
+  0xa25151f3, 0x5da3a3fe, 0x804040c0, 0x058f8f8a,
+  0x3f9292ad, 0x219d9dbc, 0x70383848, 0xf1f5f504,
+  0x63bcbcdf, 0x77b6b6c1, 0xafdada75, 0x42212163,
+  0x20101030, 0xe5ffff1a, 0xfdf3f30e, 0xbfd2d26d,
+  0x81cdcd4c, 0x180c0c14, 0x26131335, 0xc3ecec2f,
+  0xbe5f5fe1, 0x359797a2, 0x884444cc, 0x2e171739,
+  0x93c4c457, 0x55a7a7f2, 0xfc7e7e82, 0x7a3d3d47,
+  0xc86464ac, 0xba5d5de7, 0x3219192b, 0xe6737395,
+  0xc06060a0, 0x19818198, 0x9e4f4fd1, 0xa3dcdc7f,
+  0x44222266, 0x542a2a7e, 0x3b9090ab, 0x0b888883,
+  0x8c4646ca, 0xc7eeee29, 0x6bb8b8d3, 0x2814143c,
+  0xa7dede79, 0xbc5e5ee2, 0x160b0b1d, 0xaddbdb76,
+  0xdbe0e03b, 0x64323256, 0x743a3a4e, 0x140a0a1e,
+  0x924949db, 0x0c06060a, 0x4824246c, 0xb85c5ce4,
+  0x9fc2c25d, 0xbdd3d36e, 0x43acacef, 0xc46262a6,
+  0x399191a8, 0x319595a4, 0xd3e4e437, 0xf279798b,
+  0xd5e7e732, 0x8bc8c843, 0x6e373759, 0xda6d6db7,
+  0x018d8d8c, 0xb1d5d564, 0x9c4e4ed2, 0x49a9a9e0,
+  0xd86c6cb4, 0xac5656fa, 0xf3f4f407, 0xcfeaea25,
+  0xca6565af, 0xf47a7a8e, 0x47aeaee9, 0x10080818,
+  0x6fbabad5, 0xf0787888, 0x4a25256f, 0x5c2e2e72,
+  0x381c1c24, 0x57a6a6f1, 0x73b4b4c7, 0x97c6c651,
+  0xcbe8e823, 0xa1dddd7c, 0xe874749c, 0x3e1f1f21,
+  0x964b4bdd, 0x61bdbddc, 0x0d8b8b86, 0x0f8a8a85,
+  0xe0707090, 0x7c3e3e42, 0x71b5b5c4, 0xcc6666aa,
+  0x904848d8, 0x06030305, 0xf7f6f601, 0x1c0e0e12,
+  0xc26161a3, 0x6a35355f, 0xae5757f9, 0x69b9b9d0,
+  0x17868691, 0x99c1c158, 0x3a1d1d27, 0x279e9eb9,
+  0xd9e1e138, 0xebf8f813, 0x2b9898b3, 0x22111133,
+  0xd26969bb, 0xa9d9d970, 0x078e8e89, 0x339494a7,
+  0x2d9b9bb6, 0x3c1e1e22, 0x15878792, 0xc9e9e920,
+  0x87cece49, 0xaa5555ff, 0x50282878, 0xa5dfdf7a,
+  0x038c8c8f, 0x59a1a1f8, 0x09898980, 0x1a0d0d17,
+  0x65bfbfda, 0xd7e6e631, 0x844242c6, 0xd06868b8,
+  0x824141c3, 0x299999b0, 0x5a2d2d77, 0x1e0f0f11,
+  0x7bb0b0cb, 0xa85454fc, 0x6dbbbbd6, 0x2c16163a,
+};
+
+__constant u32 te1[256] =
+{
+  0xa5c66363, 0x84f87c7c, 0x99ee7777, 0x8df67b7b,
+  0x0dfff2f2, 0xbdd66b6b, 0xb1de6f6f, 0x5491c5c5,
+  0x50603030, 0x03020101, 0xa9ce6767, 0x7d562b2b,
+  0x19e7fefe, 0x62b5d7d7, 0xe64dabab, 0x9aec7676,
+  0x458fcaca, 0x9d1f8282, 0x4089c9c9, 0x87fa7d7d,
+  0x15effafa, 0xebb25959, 0xc98e4747, 0x0bfbf0f0,
+  0xec41adad, 0x67b3d4d4, 0xfd5fa2a2, 0xea45afaf,
+  0xbf239c9c, 0xf753a4a4, 0x96e47272, 0x5b9bc0c0,
+  0xc275b7b7, 0x1ce1fdfd, 0xae3d9393, 0x6a4c2626,
+  0x5a6c3636, 0x417e3f3f, 0x02f5f7f7, 0x4f83cccc,
+  0x5c683434, 0xf451a5a5, 0x34d1e5e5, 0x08f9f1f1,
+  0x93e27171, 0x73abd8d8, 0x53623131, 0x3f2a1515,
+  0x0c080404, 0x5295c7c7, 0x65462323, 0x5e9dc3c3,
+  0x28301818, 0xa1379696, 0x0f0a0505, 0xb52f9a9a,
+  0x090e0707, 0x36241212, 0x9b1b8080, 0x3ddfe2e2,
+  0x26cdebeb, 0x694e2727, 0xcd7fb2b2, 0x9fea7575,
+  0x1b120909, 0x9e1d8383, 0x74582c2c, 0x2e341a1a,
+  0x2d361b1b, 0xb2dc6e6e, 0xeeb45a5a, 0xfb5ba0a0,
+  0xf6a45252, 0x4d763b3b, 0x61b7d6d6, 0xce7db3b3,
+  0x7b522929, 0x3edde3e3, 0x715e2f2f, 0x97138484,
+  0xf5a65353, 0x68b9d1d1, 0x00000000, 0x2cc1eded,
+  0x60402020, 0x1fe3fcfc, 0xc879b1b1, 0xedb65b5b,
+  0xbed46a6a, 0x468dcbcb, 0xd967bebe, 0x4b723939,
+  0xde944a4a, 0xd4984c4c, 0xe8b05858, 0x4a85cfcf,
+  0x6bbbd0d0, 0x2ac5efef, 0xe54faaaa, 0x16edfbfb,
+  0xc5864343, 0xd79a4d4d, 0x55663333, 0x94118585,
+  0xcf8a4545, 0x10e9f9f9, 0x06040202, 0x81fe7f7f,
+  0xf0a05050, 0x44783c3c, 0xba259f9f, 0xe34ba8a8,
+  0xf3a25151, 0xfe5da3a3, 0xc0804040, 0x8a058f8f,
+  0xad3f9292, 0xbc219d9d, 0x48703838, 0x04f1f5f5,
+  0xdf63bcbc, 0xc177b6b6, 0x75afdada, 0x63422121,
+  0x30201010, 0x1ae5ffff, 0x0efdf3f3, 0x6dbfd2d2,
+  0x4c81cdcd, 0x14180c0c, 0x35261313, 0x2fc3ecec,
+  0xe1be5f5f, 0xa2359797, 0xcc884444, 0x392e1717,
+  0x5793c4c4, 0xf255a7a7, 0x82fc7e7e, 0x477a3d3d,
+  0xacc86464, 0xe7ba5d5d, 0x2b321919, 0x95e67373,
+  0xa0c06060, 0x98198181, 0xd19e4f4f, 0x7fa3dcdc,
+  0x66442222, 0x7e542a2a, 0xab3b9090, 0x830b8888,
+  0xca8c4646, 0x29c7eeee, 0xd36bb8b8, 0x3c281414,
+  0x79a7dede, 0xe2bc5e5e, 0x1d160b0b, 0x76addbdb,
+  0x3bdbe0e0, 0x56643232, 0x4e743a3a, 0x1e140a0a,
+  0xdb924949, 0x0a0c0606, 0x6c482424, 0xe4b85c5c,
+  0x5d9fc2c2, 0x6ebdd3d3, 0xef43acac, 0xa6c46262,
+  0xa8399191, 0xa4319595, 0x37d3e4e4, 0x8bf27979,
+  0x32d5e7e7, 0x438bc8c8, 0x596e3737, 0xb7da6d6d,
+  0x8c018d8d, 0x64b1d5d5, 0xd29c4e4e, 0xe049a9a9,
+  0xb4d86c6c, 0xfaac5656, 0x07f3f4f4, 0x25cfeaea,
+  0xafca6565, 0x8ef47a7a, 0xe947aeae, 0x18100808,
+  0xd56fbaba, 0x88f07878, 0x6f4a2525, 0x725c2e2e,
+  0x24381c1c, 0xf157a6a6, 0xc773b4b4, 0x5197c6c6,
+  0x23cbe8e8, 0x7ca1dddd, 0x9ce87474, 0x213e1f1f,
+  0xdd964b4b, 0xdc61bdbd, 0x860d8b8b, 0x850f8a8a,
+  0x90e07070, 0x427c3e3e, 0xc471b5b5, 0xaacc6666,
+  0xd8904848, 0x05060303, 0x01f7f6f6, 0x121c0e0e,
+  0xa3c26161, 0x5f6a3535, 0xf9ae5757, 0xd069b9b9,
+  0x91178686, 0x5899c1c1, 0x273a1d1d, 0xb9279e9e,
+  0x38d9e1e1, 0x13ebf8f8, 0xb32b9898, 0x33221111,
+  0xbbd26969, 0x70a9d9d9, 0x89078e8e, 0xa7339494,
+  0xb62d9b9b, 0x223c1e1e, 0x92158787, 0x20c9e9e9,
+  0x4987cece, 0xffaa5555, 0x78502828, 0x7aa5dfdf,
+  0x8f038c8c, 0xf859a1a1, 0x80098989, 0x171a0d0d,
+  0xda65bfbf, 0x31d7e6e6, 0xc6844242, 0xb8d06868,
+  0xc3824141, 0xb0299999, 0x775a2d2d, 0x111e0f0f,
+  0xcb7bb0b0, 0xfca85454, 0xd66dbbbb, 0x3a2c1616,
+};
+
+__constant u32 te2[256] =
+{
+  0x63a5c663, 0x7c84f87c, 0x7799ee77, 0x7b8df67b,
+  0xf20dfff2, 0x6bbdd66b, 0x6fb1de6f, 0xc55491c5,
+  0x30506030, 0x01030201, 0x67a9ce67, 0x2b7d562b,
+  0xfe19e7fe, 0xd762b5d7, 0xabe64dab, 0x769aec76,
+  0xca458fca, 0x829d1f82, 0xc94089c9, 0x7d87fa7d,
+  0xfa15effa, 0x59ebb259, 0x47c98e47, 0xf00bfbf0,
+  0xadec41ad, 0xd467b3d4, 0xa2fd5fa2, 0xafea45af,
+  0x9cbf239c, 0xa4f753a4, 0x7296e472, 0xc05b9bc0,
+  0xb7c275b7, 0xfd1ce1fd, 0x93ae3d93, 0x266a4c26,
+  0x365a6c36, 0x3f417e3f, 0xf702f5f7, 0xcc4f83cc,
+  0x345c6834, 0xa5f451a5, 0xe534d1e5, 0xf108f9f1,
+  0x7193e271, 0xd873abd8, 0x31536231, 0x153f2a15,
+  0x040c0804, 0xc75295c7, 0x23654623, 0xc35e9dc3,
+  0x18283018, 0x96a13796, 0x050f0a05, 0x9ab52f9a,
+  0x07090e07, 0x12362412, 0x809b1b80, 0xe23ddfe2,
+  0xeb26cdeb, 0x27694e27, 0xb2cd7fb2, 0x759fea75,
+  0x091b1209, 0x839e1d83, 0x2c74582c, 0x1a2e341a,
+  0x1b2d361b, 0x6eb2dc6e, 0x5aeeb45a, 0xa0fb5ba0,
+  0x52f6a452, 0x3b4d763b, 0xd661b7d6, 0xb3ce7db3,
+  0x297b5229, 0xe33edde3, 0x2f715e2f, 0x84971384,
+  0x53f5a653, 0xd168b9d1, 0x00000000, 0xed2cc1ed,
+  0x20604020, 0xfc1fe3fc, 0xb1c879b1, 0x5bedb65b,
+  0x6abed46a, 0xcb468dcb, 0xbed967be, 0x394b7239,
+  0x4ade944a, 0x4cd4984c, 0x58e8b058, 0xcf4a85cf,
+  0xd06bbbd0, 0xef2ac5ef, 0xaae54faa, 0xfb16edfb,
+  0x43c58643, 0x4dd79a4d, 0x33556633, 0x85941185,
+  0x45cf8a45, 0xf910e9f9, 0x02060402, 0x7f81fe7f,
+  0x50f0a050, 0x3c44783c, 0x9fba259f, 0xa8e34ba8,
+  0x51f3a251, 0xa3fe5da3, 0x40c08040, 0x8f8a058f,
+  0x92ad3f92, 0x9dbc219d, 0x38487038, 0xf504f1f5,
+  0xbcdf63bc, 0xb6c177b6, 0xda75afda, 0x21634221,
+  0x10302010, 0xff1ae5ff, 0xf30efdf3, 0xd26dbfd2,
+  0xcd4c81cd, 0x0c14180c, 0x13352613, 0xec2fc3ec,
+  0x5fe1be5f, 0x97a23597, 0x44cc8844, 0x17392e17,
+  0xc45793c4, 0xa7f255a7, 0x7e82fc7e, 0x3d477a3d,
+  0x64acc864, 0x5de7ba5d, 0x192b3219, 0x7395e673,
+  0x60a0c060, 0x81981981, 0x4fd19e4f, 0xdc7fa3dc,
+  0x22664422, 0x2a7e542a, 0x90ab3b90, 0x88830b88,
+  0x46ca8c46, 0xee29c7ee, 0xb8d36bb8, 0x143c2814,
+  0xde79a7de, 0x5ee2bc5e, 0x0b1d160b, 0xdb76addb,
+  0xe03bdbe0, 0x32566432, 0x3a4e743a, 0x0a1e140a,
+  0x49db9249, 0x060a0c06, 0x246c4824, 0x5ce4b85c,
+  0xc25d9fc2, 0xd36ebdd3, 0xacef43ac, 0x62a6c462,
+  0x91a83991, 0x95a43195, 0xe437d3e4, 0x798bf279,
+  0xe732d5e7, 0xc8438bc8, 0x37596e37, 0x6db7da6d,
+  0x8d8c018d, 0xd564b1d5, 0x4ed29c4e, 0xa9e049a9,
+  0x6cb4d86c, 0x56faac56, 0xf407f3f4, 0xea25cfea,
+  0x65afca65, 0x7a8ef47a, 0xaee947ae, 0x08181008,
+  0xbad56fba, 0x7888f078, 0x256f4a25, 0x2e725c2e,
+  0x1c24381c, 0xa6f157a6, 0xb4c773b4, 0xc65197c6,
+  0xe823cbe8, 0xdd7ca1dd, 0x749ce874, 0x1f213e1f,
+  0x4bdd964b, 0xbddc61bd, 0x8b860d8b, 0x8a850f8a,
+  0x7090e070, 0x3e427c3e, 0xb5c471b5, 0x66aacc66,
+  0x48d89048, 0x03050603, 0xf601f7f6, 0x0e121c0e,
+  0x61a3c261, 0x355f6a35, 0x57f9ae57, 0xb9d069b9,
+  0x86911786, 0xc15899c1, 0x1d273a1d, 0x9eb9279e,
+  0xe138d9e1, 0xf813ebf8, 0x98b32b98, 0x11332211,
+  0x69bbd269, 0xd970a9d9, 0x8e89078e, 0x94a73394,
+  0x9bb62d9b, 0x1e223c1e, 0x87921587, 0xe920c9e9,
+  0xce4987ce, 0x55ffaa55, 0x28785028, 0xdf7aa5df,
+  0x8c8f038c, 0xa1f859a1, 0x89800989, 0x0d171a0d,
+  0xbfda65bf, 0xe631d7e6, 0x42c68442, 0x68b8d068,
+  0x41c38241, 0x99b02999, 0x2d775a2d, 0x0f111e0f,
+  0xb0cb7bb0, 0x54fca854, 0xbbd66dbb, 0x163a2c16,
+};
+
+__constant u32 te3[256] =
+{
+  0x6363a5c6, 0x7c7c84f8, 0x777799ee, 0x7b7b8df6,
+  0xf2f20dff, 0x6b6bbdd6, 0x6f6fb1de, 0xc5c55491,
+  0x30305060, 0x01010302, 0x6767a9ce, 0x2b2b7d56,
+  0xfefe19e7, 0xd7d762b5, 0xababe64d, 0x76769aec,
+  0xcaca458f, 0x82829d1f, 0xc9c94089, 0x7d7d87fa,
+  0xfafa15ef, 0x5959ebb2, 0x4747c98e, 0xf0f00bfb,
+  0xadadec41, 0xd4d467b3, 0xa2a2fd5f, 0xafafea45,
+  0x9c9cbf23, 0xa4a4f753, 0x727296e4, 0xc0c05b9b,
+  0xb7b7c275, 0xfdfd1ce1, 0x9393ae3d, 0x26266a4c,
+  0x36365a6c, 0x3f3f417e, 0xf7f702f5, 0xcccc4f83,
+  0x34345c68, 0xa5a5f451, 0xe5e534d1, 0xf1f108f9,
+  0x717193e2, 0xd8d873ab, 0x31315362, 0x15153f2a,
+  0x04040c08, 0xc7c75295, 0x23236546, 0xc3c35e9d,
+  0x18182830, 0x9696a137, 0x05050f0a, 0x9a9ab52f,
+  0x0707090e, 0x12123624, 0x80809b1b, 0xe2e23ddf,
+  0xebeb26cd, 0x2727694e, 0xb2b2cd7f, 0x75759fea,
+  0x09091b12, 0x83839e1d, 0x2c2c7458, 0x1a1a2e34,
+  0x1b1b2d36, 0x6e6eb2dc, 0x5a5aeeb4, 0xa0a0fb5b,
+  0x5252f6a4, 0x3b3b4d76, 0xd6d661b7, 0xb3b3ce7d,
+  0x29297b52, 0xe3e33edd, 0x2f2f715e, 0x84849713,
+  0x5353f5a6, 0xd1d168b9, 0x00000000, 0xeded2cc1,
+  0x20206040, 0xfcfc1fe3, 0xb1b1c879, 0x5b5bedb6,
+  0x6a6abed4, 0xcbcb468d, 0xbebed967, 0x39394b72,
+  0x4a4ade94, 0x4c4cd498, 0x5858e8b0, 0xcfcf4a85,
+  0xd0d06bbb, 0xefef2ac5, 0xaaaae54f, 0xfbfb16ed,
+  0x4343c586, 0x4d4dd79a, 0x33335566, 0x85859411,
+  0x4545cf8a, 0xf9f910e9, 0x02020604, 0x7f7f81fe,
+  0x5050f0a0, 0x3c3c4478, 0x9f9fba25, 0xa8a8e34b,
+  0x5151f3a2, 0xa3a3fe5d, 0x4040c080, 0x8f8f8a05,
+  0x9292ad3f, 0x9d9dbc21, 0x38384870, 0xf5f504f1,
+  0xbcbcdf63, 0xb6b6c177, 0xdada75af, 0x21216342,
+  0x10103020, 0xffff1ae5, 0xf3f30efd, 0xd2d26dbf,
+  0xcdcd4c81, 0x0c0c1418, 0x13133526, 0xecec2fc3,
+  0x5f5fe1be, 0x9797a235, 0x4444cc88, 0x1717392e,
+  0xc4c45793, 0xa7a7f255, 0x7e7e82fc, 0x3d3d477a,
+  0x6464acc8, 0x5d5de7ba, 0x19192b32, 0x737395e6,
+  0x6060a0c0, 0x81819819, 0x4f4fd19e, 0xdcdc7fa3,
+  0x22226644, 0x2a2a7e54, 0x9090ab3b, 0x8888830b,
+  0x4646ca8c, 0xeeee29c7, 0xb8b8d36b, 0x14143c28,
+  0xdede79a7, 0x5e5ee2bc, 0x0b0b1d16, 0xdbdb76ad,
+  0xe0e03bdb, 0x32325664, 0x3a3a4e74, 0x0a0a1e14,
+  0x4949db92, 0x06060a0c, 0x24246c48, 0x5c5ce4b8,
+  0xc2c25d9f, 0xd3d36ebd, 0xacacef43, 0x6262a6c4,
+  0x9191a839, 0x9595a431, 0xe4e437d3, 0x79798bf2,
+  0xe7e732d5, 0xc8c8438b, 0x3737596e, 0x6d6db7da,
+  0x8d8d8c01, 0xd5d564b1, 0x4e4ed29c, 0xa9a9e049,
+  0x6c6cb4d8, 0x5656faac, 0xf4f407f3, 0xeaea25cf,
+  0x6565afca, 0x7a7a8ef4, 0xaeaee947, 0x08081810,
+  0xbabad56f, 0x787888f0, 0x25256f4a, 0x2e2e725c,
+  0x1c1c2438, 0xa6a6f157, 0xb4b4c773, 0xc6c65197,
+  0xe8e823cb, 0xdddd7ca1, 0x74749ce8, 0x1f1f213e,
+  0x4b4bdd96, 0xbdbddc61, 0x8b8b860d, 0x8a8a850f,
+  0x707090e0, 0x3e3e427c, 0xb5b5c471, 0x6666aacc,
+  0x4848d890, 0x03030506, 0xf6f601f7, 0x0e0e121c,
+  0x6161a3c2, 0x35355f6a, 0x5757f9ae, 0xb9b9d069,
+  0x86869117, 0xc1c15899, 0x1d1d273a, 0x9e9eb927,
+  0xe1e138d9, 0xf8f813eb, 0x9898b32b, 0x11113322,
+  0x6969bbd2, 0xd9d970a9, 0x8e8e8907, 0x9494a733,
+  0x9b9bb62d, 0x1e1e223c, 0x87879215, 0xe9e920c9,
+  0xcece4987, 0x5555ffaa, 0x28287850, 0xdfdf7aa5,
+  0x8c8c8f03, 0xa1a1f859, 0x89898009, 0x0d0d171a,
+  0xbfbfda65, 0xe6e631d7, 0x4242c684, 0x6868b8d0,
+  0x4141c382, 0x9999b029, 0x2d2d775a, 0x0f0f111e,
+  0xb0b0cb7b, 0x5454fca8, 0xbbbbd66d, 0x16163a2c,
+};
+
+__constant u32 te4[256] =
+{
+  0x63636363, 0x7c7c7c7c, 0x77777777, 0x7b7b7b7b,
+  0xf2f2f2f2, 0x6b6b6b6b, 0x6f6f6f6f, 0xc5c5c5c5,
+  0x30303030, 0x01010101, 0x67676767, 0x2b2b2b2b,
+  0xfefefefe, 0xd7d7d7d7, 0xabababab, 0x76767676,
+  0xcacacaca, 0x82828282, 0xc9c9c9c9, 0x7d7d7d7d,
+  0xfafafafa, 0x59595959, 0x47474747, 0xf0f0f0f0,
+  0xadadadad, 0xd4d4d4d4, 0xa2a2a2a2, 0xafafafaf,
+  0x9c9c9c9c, 0xa4a4a4a4, 0x72727272, 0xc0c0c0c0,
+  0xb7b7b7b7, 0xfdfdfdfd, 0x93939393, 0x26262626,
+  0x36363636, 0x3f3f3f3f, 0xf7f7f7f7, 0xcccccccc,
+  0x34343434, 0xa5a5a5a5, 0xe5e5e5e5, 0xf1f1f1f1,
+  0x71717171, 0xd8d8d8d8, 0x31313131, 0x15151515,
+  0x04040404, 0xc7c7c7c7, 0x23232323, 0xc3c3c3c3,
+  0x18181818, 0x96969696, 0x05050505, 0x9a9a9a9a,
+  0x07070707, 0x12121212, 0x80808080, 0xe2e2e2e2,
+  0xebebebeb, 0x27272727, 0xb2b2b2b2, 0x75757575,
+  0x09090909, 0x83838383, 0x2c2c2c2c, 0x1a1a1a1a,
+  0x1b1b1b1b, 0x6e6e6e6e, 0x5a5a5a5a, 0xa0a0a0a0,
+  0x52525252, 0x3b3b3b3b, 0xd6d6d6d6, 0xb3b3b3b3,
+  0x29292929, 0xe3e3e3e3, 0x2f2f2f2f, 0x84848484,
+  0x53535353, 0xd1d1d1d1, 0x00000000, 0xedededed,
+  0x20202020, 0xfcfcfcfc, 0xb1b1b1b1, 0x5b5b5b5b,
+  0x6a6a6a6a, 0xcbcbcbcb, 0xbebebebe, 0x39393939,
+  0x4a4a4a4a, 0x4c4c4c4c, 0x58585858, 0xcfcfcfcf,
+  0xd0d0d0d0, 0xefefefef, 0xaaaaaaaa, 0xfbfbfbfb,
+  0x43434343, 0x4d4d4d4d, 0x33333333, 0x85858585,
+  0x45454545, 0xf9f9f9f9, 0x02020202, 0x7f7f7f7f,
+  0x50505050, 0x3c3c3c3c, 0x9f9f9f9f, 0xa8a8a8a8,
+  0x51515151, 0xa3a3a3a3, 0x40404040, 0x8f8f8f8f,
+  0x92929292, 0x9d9d9d9d, 0x38383838, 0xf5f5f5f5,
+  0xbcbcbcbc, 0xb6b6b6b6, 0xdadadada, 0x21212121,
+  0x10101010, 0xffffffff, 0xf3f3f3f3, 0xd2d2d2d2,
+  0xcdcdcdcd, 0x0c0c0c0c, 0x13131313, 0xecececec,
+  0x5f5f5f5f, 0x97979797, 0x44444444, 0x17171717,
+  0xc4c4c4c4, 0xa7a7a7a7, 0x7e7e7e7e, 0x3d3d3d3d,
+  0x64646464, 0x5d5d5d5d, 0x19191919, 0x73737373,
+  0x60606060, 0x81818181, 0x4f4f4f4f, 0xdcdcdcdc,
+  0x22222222, 0x2a2a2a2a, 0x90909090, 0x88888888,
+  0x46464646, 0xeeeeeeee, 0xb8b8b8b8, 0x14141414,
+  0xdededede, 0x5e5e5e5e, 0x0b0b0b0b, 0xdbdbdbdb,
+  0xe0e0e0e0, 0x32323232, 0x3a3a3a3a, 0x0a0a0a0a,
+  0x49494949, 0x06060606, 0x24242424, 0x5c5c5c5c,
+  0xc2c2c2c2, 0xd3d3d3d3, 0xacacacac, 0x62626262,
+  0x91919191, 0x95959595, 0xe4e4e4e4, 0x79797979,
+  0xe7e7e7e7, 0xc8c8c8c8, 0x37373737, 0x6d6d6d6d,
+  0x8d8d8d8d, 0xd5d5d5d5, 0x4e4e4e4e, 0xa9a9a9a9,
+  0x6c6c6c6c, 0x56565656, 0xf4f4f4f4, 0xeaeaeaea,
+  0x65656565, 0x7a7a7a7a, 0xaeaeaeae, 0x08080808,
+  0xbabababa, 0x78787878, 0x25252525, 0x2e2e2e2e,
+  0x1c1c1c1c, 0xa6a6a6a6, 0xb4b4b4b4, 0xc6c6c6c6,
+  0xe8e8e8e8, 0xdddddddd, 0x74747474, 0x1f1f1f1f,
+  0x4b4b4b4b, 0xbdbdbdbd, 0x8b8b8b8b, 0x8a8a8a8a,
+  0x70707070, 0x3e3e3e3e, 0xb5b5b5b5, 0x66666666,
+  0x48484848, 0x03030303, 0xf6f6f6f6, 0x0e0e0e0e,
+  0x61616161, 0x35353535, 0x57575757, 0xb9b9b9b9,
+  0x86868686, 0xc1c1c1c1, 0x1d1d1d1d, 0x9e9e9e9e,
+  0xe1e1e1e1, 0xf8f8f8f8, 0x98989898, 0x11111111,
+  0x69696969, 0xd9d9d9d9, 0x8e8e8e8e, 0x94949494,
+  0x9b9b9b9b, 0x1e1e1e1e, 0x87878787, 0xe9e9e9e9,
+  0xcececece, 0x55555555, 0x28282828, 0xdfdfdfdf,
+  0x8c8c8c8c, 0xa1a1a1a1, 0x89898989, 0x0d0d0d0d,
+  0xbfbfbfbf, 0xe6e6e6e6, 0x42424242, 0x68686868,
+  0x41414141, 0x99999999, 0x2d2d2d2d, 0x0f0f0f0f,
+  0xb0b0b0b0, 0x54545454, 0xbbbbbbbb, 0x16161616,
+};
+
+__constant u32 td0[256] =
+{
+  0x51f4a750, 0x7e416553, 0x1a17a4c3, 0x3a275e96,
+  0x3bab6bcb, 0x1f9d45f1, 0xacfa58ab, 0x4be30393,
+  0x2030fa55, 0xad766df6, 0x88cc7691, 0xf5024c25,
+  0x4fe5d7fc, 0xc52acbd7, 0x26354480, 0xb562a38f,
+  0xdeb15a49, 0x25ba1b67, 0x45ea0e98, 0x5dfec0e1,
+  0xc32f7502, 0x814cf012, 0x8d4697a3, 0x6bd3f9c6,
+  0x038f5fe7, 0x15929c95, 0xbf6d7aeb, 0x955259da,
+  0xd4be832d, 0x587421d3, 0x49e06929, 0x8ec9c844,
+  0x75c2896a, 0xf48e7978, 0x99583e6b, 0x27b971dd,
+  0xbee14fb6, 0xf088ad17, 0xc920ac66, 0x7dce3ab4,
+  0x63df4a18, 0xe51a3182, 0x97513360, 0x62537f45,
+  0xb16477e0, 0xbb6bae84, 0xfe81a01c, 0xf9082b94,
+  0x70486858, 0x8f45fd19, 0x94de6c87, 0x527bf8b7,
+  0xab73d323, 0x724b02e2, 0xe31f8f57, 0x6655ab2a,
+  0xb2eb2807, 0x2fb5c203, 0x86c57b9a, 0xd33708a5,
+  0x302887f2, 0x23bfa5b2, 0x02036aba, 0xed16825c,
+  0x8acf1c2b, 0xa779b492, 0xf307f2f0, 0x4e69e2a1,
+  0x65daf4cd, 0x0605bed5, 0xd134621f, 0xc4a6fe8a,
+  0x342e539d, 0xa2f355a0, 0x058ae132, 0xa4f6eb75,
+  0x0b83ec39, 0x4060efaa, 0x5e719f06, 0xbd6e1051,
+  0x3e218af9, 0x96dd063d, 0xdd3e05ae, 0x4de6bd46,
+  0x91548db5, 0x71c45d05, 0x0406d46f, 0x605015ff,
+  0x1998fb24, 0xd6bde997, 0x894043cc, 0x67d99e77,
+  0xb0e842bd, 0x07898b88, 0xe7195b38, 0x79c8eedb,
+  0xa17c0a47, 0x7c420fe9, 0xf8841ec9, 0x00000000,
+  0x09808683, 0x322bed48, 0x1e1170ac, 0x6c5a724e,
+  0xfd0efffb, 0x0f853856, 0x3daed51e, 0x362d3927,
+  0x0a0fd964, 0x685ca621, 0x9b5b54d1, 0x24362e3a,
+  0x0c0a67b1, 0x9357e70f, 0xb4ee96d2, 0x1b9b919e,
+  0x80c0c54f, 0x61dc20a2, 0x5a774b69, 0x1c121a16,
+  0xe293ba0a, 0xc0a02ae5, 0x3c22e043, 0x121b171d,
+  0x0e090d0b, 0xf28bc7ad, 0x2db6a8b9, 0x141ea9c8,
+  0x57f11985, 0xaf75074c, 0xee99ddbb, 0xa37f60fd,
+  0xf701269f, 0x5c72f5bc, 0x44663bc5, 0x5bfb7e34,
+  0x8b432976, 0xcb23c6dc, 0xb6edfc68, 0xb8e4f163,
+  0xd731dcca, 0x42638510, 0x13972240, 0x84c61120,
+  0x854a247d, 0xd2bb3df8, 0xaef93211, 0xc729a16d,
+  0x1d9e2f4b, 0xdcb230f3, 0x0d8652ec, 0x77c1e3d0,
+  0x2bb3166c, 0xa970b999, 0x119448fa, 0x47e96422,
+  0xa8fc8cc4, 0xa0f03f1a, 0x567d2cd8, 0x223390ef,
+  0x87494ec7, 0xd938d1c1, 0x8ccaa2fe, 0x98d40b36,
+  0xa6f581cf, 0xa57ade28, 0xdab78e26, 0x3fadbfa4,
+  0x2c3a9de4, 0x5078920d, 0x6a5fcc9b, 0x547e4662,
+  0xf68d13c2, 0x90d8b8e8, 0x2e39f75e, 0x82c3aff5,
+  0x9f5d80be, 0x69d0937c, 0x6fd52da9, 0xcf2512b3,
+  0xc8ac993b, 0x10187da7, 0xe89c636e, 0xdb3bbb7b,
+  0xcd267809, 0x6e5918f4, 0xec9ab701, 0x834f9aa8,
+  0xe6956e65, 0xaaffe67e, 0x21bccf08, 0xef15e8e6,
+  0xbae79bd9, 0x4a6f36ce, 0xea9f09d4, 0x29b07cd6,
+  0x31a4b2af, 0x2a3f2331, 0xc6a59430, 0x35a266c0,
+  0x744ebc37, 0xfc82caa6, 0xe090d0b0, 0x33a7d815,
+  0xf104984a, 0x41ecdaf7, 0x7fcd500e, 0x1791f62f,
+  0x764dd68d, 0x43efb04d, 0xccaa4d54, 0xe49604df,
+  0x9ed1b5e3, 0x4c6a881b, 0xc12c1fb8, 0x4665517f,
+  0x9d5eea04, 0x018c355d, 0xfa877473, 0xfb0b412e,
+  0xb3671d5a, 0x92dbd252, 0xe9105633, 0x6dd64713,
+  0x9ad7618c, 0x37a10c7a, 0x59f8148e, 0xeb133c89,
+  0xcea927ee, 0xb761c935, 0xe11ce5ed, 0x7a47b13c,
+  0x9cd2df59, 0x55f2733f, 0x1814ce79, 0x73c737bf,
+  0x53f7cdea, 0x5ffdaa5b, 0xdf3d6f14, 0x7844db86,
+  0xcaaff381, 0xb968c43e, 0x3824342c, 0xc2a3405f,
+  0x161dc372, 0xbce2250c, 0x283c498b, 0xff0d9541,
+  0x39a80171, 0x080cb3de, 0xd8b4e49c, 0x6456c190,
+  0x7bcb8461, 0xd532b670, 0x486c5c74, 0xd0b85742,
+};
+
+__constant u32 td1[256] =
+{
+  0x5051f4a7, 0x537e4165, 0xc31a17a4, 0x963a275e,
+  0xcb3bab6b, 0xf11f9d45, 0xabacfa58, 0x934be303,
+  0x552030fa, 0xf6ad766d, 0x9188cc76, 0x25f5024c,
+  0xfc4fe5d7, 0xd7c52acb, 0x80263544, 0x8fb562a3,
+  0x49deb15a, 0x6725ba1b, 0x9845ea0e, 0xe15dfec0,
+  0x02c32f75, 0x12814cf0, 0xa38d4697, 0xc66bd3f9,
+  0xe7038f5f, 0x9515929c, 0xebbf6d7a, 0xda955259,
+  0x2dd4be83, 0xd3587421, 0x2949e069, 0x448ec9c8,
+  0x6a75c289, 0x78f48e79, 0x6b99583e, 0xdd27b971,
+  0xb6bee14f, 0x17f088ad, 0x66c920ac, 0xb47dce3a,
+  0x1863df4a, 0x82e51a31, 0x60975133, 0x4562537f,
+  0xe0b16477, 0x84bb6bae, 0x1cfe81a0, 0x94f9082b,
+  0x58704868, 0x198f45fd, 0x8794de6c, 0xb7527bf8,
+  0x23ab73d3, 0xe2724b02, 0x57e31f8f, 0x2a6655ab,
+  0x07b2eb28, 0x032fb5c2, 0x9a86c57b, 0xa5d33708,
+  0xf2302887, 0xb223bfa5, 0xba02036a, 0x5ced1682,
+  0x2b8acf1c, 0x92a779b4, 0xf0f307f2, 0xa14e69e2,
+  0xcd65daf4, 0xd50605be, 0x1fd13462, 0x8ac4a6fe,
+  0x9d342e53, 0xa0a2f355, 0x32058ae1, 0x75a4f6eb,
+  0x390b83ec, 0xaa4060ef, 0x065e719f, 0x51bd6e10,
+  0xf93e218a, 0x3d96dd06, 0xaedd3e05, 0x464de6bd,
+  0xb591548d, 0x0571c45d, 0x6f0406d4, 0xff605015,
+  0x241998fb, 0x97d6bde9, 0xcc894043, 0x7767d99e,
+  0xbdb0e842, 0x8807898b, 0x38e7195b, 0xdb79c8ee,
+  0x47a17c0a, 0xe97c420f, 0xc9f8841e, 0x00000000,
+  0x83098086, 0x48322bed, 0xac1e1170, 0x4e6c5a72,
+  0xfbfd0eff, 0x560f8538, 0x1e3daed5, 0x27362d39,
+  0x640a0fd9, 0x21685ca6, 0xd19b5b54, 0x3a24362e,
+  0xb10c0a67, 0x0f9357e7, 0xd2b4ee96, 0x9e1b9b91,
+  0x4f80c0c5, 0xa261dc20, 0x695a774b, 0x161c121a,
+  0x0ae293ba, 0xe5c0a02a, 0x433c22e0, 0x1d121b17,
+  0x0b0e090d, 0xadf28bc7, 0xb92db6a8, 0xc8141ea9,
+  0x8557f119, 0x4caf7507, 0xbbee99dd, 0xfda37f60,
+  0x9ff70126, 0xbc5c72f5, 0xc544663b, 0x345bfb7e,
+  0x768b4329, 0xdccb23c6, 0x68b6edfc, 0x63b8e4f1,
+  0xcad731dc, 0x10426385, 0x40139722, 0x2084c611,
+  0x7d854a24, 0xf8d2bb3d, 0x11aef932, 0x6dc729a1,
+  0x4b1d9e2f, 0xf3dcb230, 0xec0d8652, 0xd077c1e3,
+  0x6c2bb316, 0x99a970b9, 0xfa119448, 0x2247e964,
+  0xc4a8fc8c, 0x1aa0f03f, 0xd8567d2c, 0xef223390,
+  0xc787494e, 0xc1d938d1, 0xfe8ccaa2, 0x3698d40b,
+  0xcfa6f581, 0x28a57ade, 0x26dab78e, 0xa43fadbf,
+  0xe42c3a9d, 0x0d507892, 0x9b6a5fcc, 0x62547e46,
+  0xc2f68d13, 0xe890d8b8, 0x5e2e39f7, 0xf582c3af,
+  0xbe9f5d80, 0x7c69d093, 0xa96fd52d, 0xb3cf2512,
+  0x3bc8ac99, 0xa710187d, 0x6ee89c63, 0x7bdb3bbb,
+  0x09cd2678, 0xf46e5918, 0x01ec9ab7, 0xa8834f9a,
+  0x65e6956e, 0x7eaaffe6, 0x0821bccf, 0xe6ef15e8,
+  0xd9bae79b, 0xce4a6f36, 0xd4ea9f09, 0xd629b07c,
+  0xaf31a4b2, 0x312a3f23, 0x30c6a594, 0xc035a266,
+  0x37744ebc, 0xa6fc82ca, 0xb0e090d0, 0x1533a7d8,
+  0x4af10498, 0xf741ecda, 0x0e7fcd50, 0x2f1791f6,
+  0x8d764dd6, 0x4d43efb0, 0x54ccaa4d, 0xdfe49604,
+  0xe39ed1b5, 0x1b4c6a88, 0xb8c12c1f, 0x7f466551,
+  0x049d5eea, 0x5d018c35, 0x73fa8774, 0x2efb0b41,
+  0x5ab3671d, 0x5292dbd2, 0x33e91056, 0x136dd647,
+  0x8c9ad761, 0x7a37a10c, 0x8e59f814, 0x89eb133c,
+  0xeecea927, 0x35b761c9, 0xede11ce5, 0x3c7a47b1,
+  0x599cd2df, 0x3f55f273, 0x791814ce, 0xbf73c737,
+  0xea53f7cd, 0x5b5ffdaa, 0x14df3d6f, 0x867844db,
+  0x81caaff3, 0x3eb968c4, 0x2c382434, 0x5fc2a340,
+  0x72161dc3, 0x0cbce225, 0x8b283c49, 0x41ff0d95,
+  0x7139a801, 0xde080cb3, 0x9cd8b4e4, 0x906456c1,
+  0x617bcb84, 0x70d532b6, 0x74486c5c, 0x42d0b857,
+};
+
+__constant u32 td2[256] =
+{
+  0xa75051f4, 0x65537e41, 0xa4c31a17, 0x5e963a27,
+  0x6bcb3bab, 0x45f11f9d, 0x58abacfa, 0x03934be3,
+  0xfa552030, 0x6df6ad76, 0x769188cc, 0x4c25f502,
+  0xd7fc4fe5, 0xcbd7c52a, 0x44802635, 0xa38fb562,
+  0x5a49deb1, 0x1b6725ba, 0x0e9845ea, 0xc0e15dfe,
+  0x7502c32f, 0xf012814c, 0x97a38d46, 0xf9c66bd3,
+  0x5fe7038f, 0x9c951592, 0x7aebbf6d, 0x59da9552,
+  0x832dd4be, 0x21d35874, 0x692949e0, 0xc8448ec9,
+  0x896a75c2, 0x7978f48e, 0x3e6b9958, 0x71dd27b9,
+  0x4fb6bee1, 0xad17f088, 0xac66c920, 0x3ab47dce,
+  0x4a1863df, 0x3182e51a, 0x33609751, 0x7f456253,
+  0x77e0b164, 0xae84bb6b, 0xa01cfe81, 0x2b94f908,
+  0x68587048, 0xfd198f45, 0x6c8794de, 0xf8b7527b,
+  0xd323ab73, 0x02e2724b, 0x8f57e31f, 0xab2a6655,
+  0x2807b2eb, 0xc2032fb5, 0x7b9a86c5, 0x08a5d337,
+  0x87f23028, 0xa5b223bf, 0x6aba0203, 0x825ced16,
+  0x1c2b8acf, 0xb492a779, 0xf2f0f307, 0xe2a14e69,
+  0xf4cd65da, 0xbed50605, 0x621fd134, 0xfe8ac4a6,
+  0x539d342e, 0x55a0a2f3, 0xe132058a, 0xeb75a4f6,
+  0xec390b83, 0xefaa4060, 0x9f065e71, 0x1051bd6e,
+  0x8af93e21, 0x063d96dd, 0x05aedd3e, 0xbd464de6,
+  0x8db59154, 0x5d0571c4, 0xd46f0406, 0x15ff6050,
+  0xfb241998, 0xe997d6bd, 0x43cc8940, 0x9e7767d9,
+  0x42bdb0e8, 0x8b880789, 0x5b38e719, 0xeedb79c8,
+  0x0a47a17c, 0x0fe97c42, 0x1ec9f884, 0x00000000,
+  0x86830980, 0xed48322b, 0x70ac1e11, 0x724e6c5a,
+  0xfffbfd0e, 0x38560f85, 0xd51e3dae, 0x3927362d,
+  0xd9640a0f, 0xa621685c, 0x54d19b5b, 0x2e3a2436,
+  0x67b10c0a, 0xe70f9357, 0x96d2b4ee, 0x919e1b9b,
+  0xc54f80c0, 0x20a261dc, 0x4b695a77, 0x1a161c12,
+  0xba0ae293, 0x2ae5c0a0, 0xe0433c22, 0x171d121b,
+  0x0d0b0e09, 0xc7adf28b, 0xa8b92db6, 0xa9c8141e,
+  0x198557f1, 0x074caf75, 0xddbbee99, 0x60fda37f,
+  0x269ff701, 0xf5bc5c72, 0x3bc54466, 0x7e345bfb,
+  0x29768b43, 0xc6dccb23, 0xfc68b6ed, 0xf163b8e4,
+  0xdccad731, 0x85104263, 0x22401397, 0x112084c6,
+  0x247d854a, 0x3df8d2bb, 0x3211aef9, 0xa16dc729,
+  0x2f4b1d9e, 0x30f3dcb2, 0x52ec0d86, 0xe3d077c1,
+  0x166c2bb3, 0xb999a970, 0x48fa1194, 0x642247e9,
+  0x8cc4a8fc, 0x3f1aa0f0, 0x2cd8567d, 0x90ef2233,
+  0x4ec78749, 0xd1c1d938, 0xa2fe8cca, 0x0b3698d4,
+  0x81cfa6f5, 0xde28a57a, 0x8e26dab7, 0xbfa43fad,
+  0x9de42c3a, 0x920d5078, 0xcc9b6a5f, 0x4662547e,
+  0x13c2f68d, 0xb8e890d8, 0xf75e2e39, 0xaff582c3,
+  0x80be9f5d, 0x937c69d0, 0x2da96fd5, 0x12b3cf25,
+  0x993bc8ac, 0x7da71018, 0x636ee89c, 0xbb7bdb3b,
+  0x7809cd26, 0x18f46e59, 0xb701ec9a, 0x9aa8834f,
+  0x6e65e695, 0xe67eaaff, 0xcf0821bc, 0xe8e6ef15,
+  0x9bd9bae7, 0x36ce4a6f, 0x09d4ea9f, 0x7cd629b0,
+  0xb2af31a4, 0x23312a3f, 0x9430c6a5, 0x66c035a2,
+  0xbc37744e, 0xcaa6fc82, 0xd0b0e090, 0xd81533a7,
+  0x984af104, 0xdaf741ec, 0x500e7fcd, 0xf62f1791,
+  0xd68d764d, 0xb04d43ef, 0x4d54ccaa, 0x04dfe496,
+  0xb5e39ed1, 0x881b4c6a, 0x1fb8c12c, 0x517f4665,
+  0xea049d5e, 0x355d018c, 0x7473fa87, 0x412efb0b,
+  0x1d5ab367, 0xd25292db, 0x5633e910, 0x47136dd6,
+  0x618c9ad7, 0x0c7a37a1, 0x148e59f8, 0x3c89eb13,
+  0x27eecea9, 0xc935b761, 0xe5ede11c, 0xb13c7a47,
+  0xdf599cd2, 0x733f55f2, 0xce791814, 0x37bf73c7,
+  0xcdea53f7, 0xaa5b5ffd, 0x6f14df3d, 0xdb867844,
+  0xf381caaf, 0xc43eb968, 0x342c3824, 0x405fc2a3,
+  0xc372161d, 0x250cbce2, 0x498b283c, 0x9541ff0d,
+  0x017139a8, 0xb3de080c, 0xe49cd8b4, 0xc1906456,
+  0x84617bcb, 0xb670d532, 0x5c74486c, 0x5742d0b8,
+};
+
+__constant u32 td3[256] =
+{
+  0xf4a75051, 0x4165537e, 0x17a4c31a, 0x275e963a,
+  0xab6bcb3b, 0x9d45f11f, 0xfa58abac, 0xe303934b,
+  0x30fa5520, 0x766df6ad, 0xcc769188, 0x024c25f5,
+  0xe5d7fc4f, 0x2acbd7c5, 0x35448026, 0x62a38fb5,
+  0xb15a49de, 0xba1b6725, 0xea0e9845, 0xfec0e15d,
+  0x2f7502c3, 0x4cf01281, 0x4697a38d, 0xd3f9c66b,
+  0x8f5fe703, 0x929c9515, 0x6d7aebbf, 0x5259da95,
+  0xbe832dd4, 0x7421d358, 0xe0692949, 0xc9c8448e,
+  0xc2896a75, 0x8e7978f4, 0x583e6b99, 0xb971dd27,
+  0xe14fb6be, 0x88ad17f0, 0x20ac66c9, 0xce3ab47d,
+  0xdf4a1863, 0x1a3182e5, 0x51336097, 0x537f4562,
+  0x6477e0b1, 0x6bae84bb, 0x81a01cfe, 0x082b94f9,
+  0x48685870, 0x45fd198f, 0xde6c8794, 0x7bf8b752,
+  0x73d323ab, 0x4b02e272, 0x1f8f57e3, 0x55ab2a66,
+  0xeb2807b2, 0xb5c2032f, 0xc57b9a86, 0x3708a5d3,
+  0x2887f230, 0xbfa5b223, 0x036aba02, 0x16825ced,
+  0xcf1c2b8a, 0x79b492a7, 0x07f2f0f3, 0x69e2a14e,
+  0xdaf4cd65, 0x05bed506, 0x34621fd1, 0xa6fe8ac4,
+  0x2e539d34, 0xf355a0a2, 0x8ae13205, 0xf6eb75a4,
+  0x83ec390b, 0x60efaa40, 0x719f065e, 0x6e1051bd,
+  0x218af93e, 0xdd063d96, 0x3e05aedd, 0xe6bd464d,
+  0x548db591, 0xc45d0571, 0x06d46f04, 0x5015ff60,
+  0x98fb2419, 0xbde997d6, 0x4043cc89, 0xd99e7767,
+  0xe842bdb0, 0x898b8807, 0x195b38e7, 0xc8eedb79,
+  0x7c0a47a1, 0x420fe97c, 0x841ec9f8, 0x00000000,
+  0x80868309, 0x2bed4832, 0x1170ac1e, 0x5a724e6c,
+  0x0efffbfd, 0x8538560f, 0xaed51e3d, 0x2d392736,
+  0x0fd9640a, 0x5ca62168, 0x5b54d19b, 0x362e3a24,
+  0x0a67b10c, 0x57e70f93, 0xee96d2b4, 0x9b919e1b,
+  0xc0c54f80, 0xdc20a261, 0x774b695a, 0x121a161c,
+  0x93ba0ae2, 0xa02ae5c0, 0x22e0433c, 0x1b171d12,
+  0x090d0b0e, 0x8bc7adf2, 0xb6a8b92d, 0x1ea9c814,
+  0xf1198557, 0x75074caf, 0x99ddbbee, 0x7f60fda3,
+  0x01269ff7, 0x72f5bc5c, 0x663bc544, 0xfb7e345b,
+  0x4329768b, 0x23c6dccb, 0xedfc68b6, 0xe4f163b8,
+  0x31dccad7, 0x63851042, 0x97224013, 0xc6112084,
+  0x4a247d85, 0xbb3df8d2, 0xf93211ae, 0x29a16dc7,
+  0x9e2f4b1d, 0xb230f3dc, 0x8652ec0d, 0xc1e3d077,
+  0xb3166c2b, 0x70b999a9, 0x9448fa11, 0xe9642247,
+  0xfc8cc4a8, 0xf03f1aa0, 0x7d2cd856, 0x3390ef22,
+  0x494ec787, 0x38d1c1d9, 0xcaa2fe8c, 0xd40b3698,
+  0xf581cfa6, 0x7ade28a5, 0xb78e26da, 0xadbfa43f,
+  0x3a9de42c, 0x78920d50, 0x5fcc9b6a, 0x7e466254,
+  0x8d13c2f6, 0xd8b8e890, 0x39f75e2e, 0xc3aff582,
+  0x5d80be9f, 0xd0937c69, 0xd52da96f, 0x2512b3cf,
+  0xac993bc8, 0x187da710, 0x9c636ee8, 0x3bbb7bdb,
+  0x267809cd, 0x5918f46e, 0x9ab701ec, 0x4f9aa883,
+  0x956e65e6, 0xffe67eaa, 0xbccf0821, 0x15e8e6ef,
+  0xe79bd9ba, 0x6f36ce4a, 0x9f09d4ea, 0xb07cd629,
+  0xa4b2af31, 0x3f23312a, 0xa59430c6, 0xa266c035,
+  0x4ebc3774, 0x82caa6fc, 0x90d0b0e0, 0xa7d81533,
+  0x04984af1, 0xecdaf741, 0xcd500e7f, 0x91f62f17,
+  0x4dd68d76, 0xefb04d43, 0xaa4d54cc, 0x9604dfe4,
+  0xd1b5e39e, 0x6a881b4c, 0x2c1fb8c1, 0x65517f46,
+  0x5eea049d, 0x8c355d01, 0x877473fa, 0x0b412efb,
+  0x671d5ab3, 0xdbd25292, 0x105633e9, 0xd647136d,
+  0xd7618c9a, 0xa10c7a37, 0xf8148e59, 0x133c89eb,
+  0xa927eece, 0x61c935b7, 0x1ce5ede1, 0x47b13c7a,
+  0xd2df599c, 0xf2733f55, 0x14ce7918, 0xc737bf73,
+  0xf7cdea53, 0xfdaa5b5f, 0x3d6f14df, 0x44db8678,
+  0xaff381ca, 0x68c43eb9, 0x24342c38, 0xa3405fc2,
+  0x1dc37216, 0xe2250cbc, 0x3c498b28, 0x0d9541ff,
+  0xa8017139, 0x0cb3de08, 0xb4e49cd8, 0x56c19064,
+  0xcb84617b, 0x32b670d5, 0x6c5c7448, 0xb85742d0,
+};
+
+__constant u32 td4[256] =
+{
+  0x52525252, 0x09090909, 0x6a6a6a6a, 0xd5d5d5d5,
+  0x30303030, 0x36363636, 0xa5a5a5a5, 0x38383838,
+  0xbfbfbfbf, 0x40404040, 0xa3a3a3a3, 0x9e9e9e9e,
+  0x81818181, 0xf3f3f3f3, 0xd7d7d7d7, 0xfbfbfbfb,
+  0x7c7c7c7c, 0xe3e3e3e3, 0x39393939, 0x82828282,
+  0x9b9b9b9b, 0x2f2f2f2f, 0xffffffff, 0x87878787,
+  0x34343434, 0x8e8e8e8e, 0x43434343, 0x44444444,
+  0xc4c4c4c4, 0xdededede, 0xe9e9e9e9, 0xcbcbcbcb,
+  0x54545454, 0x7b7b7b7b, 0x94949494, 0x32323232,
+  0xa6a6a6a6, 0xc2c2c2c2, 0x23232323, 0x3d3d3d3d,
+  0xeeeeeeee, 0x4c4c4c4c, 0x95959595, 0x0b0b0b0b,
+  0x42424242, 0xfafafafa, 0xc3c3c3c3, 0x4e4e4e4e,
+  0x08080808, 0x2e2e2e2e, 0xa1a1a1a1, 0x66666666,
+  0x28282828, 0xd9d9d9d9, 0x24242424, 0xb2b2b2b2,
+  0x76767676, 0x5b5b5b5b, 0xa2a2a2a2, 0x49494949,
+  0x6d6d6d6d, 0x8b8b8b8b, 0xd1d1d1d1, 0x25252525,
+  0x72727272, 0xf8f8f8f8, 0xf6f6f6f6, 0x64646464,
+  0x86868686, 0x68686868, 0x98989898, 0x16161616,
+  0xd4d4d4d4, 0xa4a4a4a4, 0x5c5c5c5c, 0xcccccccc,
+  0x5d5d5d5d, 0x65656565, 0xb6b6b6b6, 0x92929292,
+  0x6c6c6c6c, 0x70707070, 0x48484848, 0x50505050,
+  0xfdfdfdfd, 0xedededed, 0xb9b9b9b9, 0xdadadada,
+  0x5e5e5e5e, 0x15151515, 0x46464646, 0x57575757,
+  0xa7a7a7a7, 0x8d8d8d8d, 0x9d9d9d9d, 0x84848484,
+  0x90909090, 0xd8d8d8d8, 0xabababab, 0x00000000,
+  0x8c8c8c8c, 0xbcbcbcbc, 0xd3d3d3d3, 0x0a0a0a0a,
+  0xf7f7f7f7, 0xe4e4e4e4, 0x58585858, 0x05050505,
+  0xb8b8b8b8, 0xb3b3b3b3, 0x45454545, 0x06060606,
+  0xd0d0d0d0, 0x2c2c2c2c, 0x1e1e1e1e, 0x8f8f8f8f,
+  0xcacacaca, 0x3f3f3f3f, 0x0f0f0f0f, 0x02020202,
+  0xc1c1c1c1, 0xafafafaf, 0xbdbdbdbd, 0x03030303,
+  0x01010101, 0x13131313, 0x8a8a8a8a, 0x6b6b6b6b,
+  0x3a3a3a3a, 0x91919191, 0x11111111, 0x41414141,
+  0x4f4f4f4f, 0x67676767, 0xdcdcdcdc, 0xeaeaeaea,
+  0x97979797, 0xf2f2f2f2, 0xcfcfcfcf, 0xcececece,
+  0xf0f0f0f0, 0xb4b4b4b4, 0xe6e6e6e6, 0x73737373,
+  0x96969696, 0xacacacac, 0x74747474, 0x22222222,
+  0xe7e7e7e7, 0xadadadad, 0x35353535, 0x85858585,
+  0xe2e2e2e2, 0xf9f9f9f9, 0x37373737, 0xe8e8e8e8,
+  0x1c1c1c1c, 0x75757575, 0xdfdfdfdf, 0x6e6e6e6e,
+  0x47474747, 0xf1f1f1f1, 0x1a1a1a1a, 0x71717171,
+  0x1d1d1d1d, 0x29292929, 0xc5c5c5c5, 0x89898989,
+  0x6f6f6f6f, 0xb7b7b7b7, 0x62626262, 0x0e0e0e0e,
+  0xaaaaaaaa, 0x18181818, 0xbebebebe, 0x1b1b1b1b,
+  0xfcfcfcfc, 0x56565656, 0x3e3e3e3e, 0x4b4b4b4b,
+  0xc6c6c6c6, 0xd2d2d2d2, 0x79797979, 0x20202020,
+  0x9a9a9a9a, 0xdbdbdbdb, 0xc0c0c0c0, 0xfefefefe,
+  0x78787878, 0xcdcdcdcd, 0x5a5a5a5a, 0xf4f4f4f4,
+  0x1f1f1f1f, 0xdddddddd, 0xa8a8a8a8, 0x33333333,
+  0x88888888, 0x07070707, 0xc7c7c7c7, 0x31313131,
+  0xb1b1b1b1, 0x12121212, 0x10101010, 0x59595959,
+  0x27272727, 0x80808080, 0xecececec, 0x5f5f5f5f,
+  0x60606060, 0x51515151, 0x7f7f7f7f, 0xa9a9a9a9,
+  0x19191919, 0xb5b5b5b5, 0x4a4a4a4a, 0x0d0d0d0d,
+  0x2d2d2d2d, 0xe5e5e5e5, 0x7a7a7a7a, 0x9f9f9f9f,
+  0x93939393, 0xc9c9c9c9, 0x9c9c9c9c, 0xefefefef,
+  0xa0a0a0a0, 0xe0e0e0e0, 0x3b3b3b3b, 0x4d4d4d4d,
+  0xaeaeaeae, 0x2a2a2a2a, 0xf5f5f5f5, 0xb0b0b0b0,
+  0xc8c8c8c8, 0xebebebeb, 0xbbbbbbbb, 0x3c3c3c3c,
+  0x83838383, 0x53535353, 0x99999999, 0x61616161,
+  0x17171717, 0x2b2b2b2b, 0x04040404, 0x7e7e7e7e,
+  0xbabababa, 0x77777777, 0xd6d6d6d6, 0x26262626,
+  0xe1e1e1e1, 0x69696969, 0x14141414, 0x63636363,
+  0x55555555, 0x21212121, 0x0c0c0c0c, 0x7d7d7d7d,
+};
+
+__constant u32 rcon[] =
+{
+  0x01000000, 0x02000000, 0x04000000, 0x08000000,
+  0x10000000, 0x20000000, 0x40000000, 0x80000000,
+  0x1b000000, 0x36000000,
+};
+
+static void aes256_ExpandKey (u32 *ks, const u32 *ukey)
+{
+  ks[0] = ukey[0];
+  ks[1] = ukey[1];
+  ks[2] = ukey[2];
+  ks[3] = ukey[3];
+  ks[4] = ukey[4];
+  ks[5] = ukey[5];
+  ks[6] = ukey[6];
+  ks[7] = ukey[7];
+
+  int i;
+  int j;
+
+  i = 0;
+  j = 0;
+
+  while (1)
+  {
+    u32 temp = ks[j +  7];
+
+    ks[j +  8] = ks[j +  0]
+           ^ (te2[(temp >> 16) & 0xff] & 0xff000000)
+           ^ (te3[(temp >>  8) & 0xff] & 0x00ff0000)
+           ^ (te0[(temp >>  0) & 0xff] & 0x0000ff00)
+           ^ (te1[(temp >> 24) & 0xff] & 0x000000ff)
+           ^ rcon[i];
+
+    ks[j +  9] = ks[j +  1] ^ ks[j +  8];
+    ks[j + 10] = ks[j +  2] ^ ks[j +  9];
+    ks[j + 11] = ks[j +  3] ^ ks[j + 10];
+
+    if (++i == 7) break;
+
+    temp = ks[j + 11];
+
+    ks[j + 12] = ks[j +  4]
+           ^ (te2[(temp >> 24) & 0xff] & 0xff000000)
+           ^ (te3[(temp >> 16) & 0xff] & 0x00ff0000)
+           ^ (te0[(temp >>  8) & 0xff] & 0x0000ff00)
+           ^ (te1[(temp >>  0) & 0xff] & 0x000000ff);
+
+    ks[j + 13] = ks[j +  5] ^ ks[j + 12];
+    ks[j + 14] = ks[j +  6] ^ ks[j + 13];
+    ks[j + 15] = ks[j +  7] ^ ks[j + 14];
+
+    j += 8;
+  }
+}
+
+static void aes256_InvertKey (u32 *ks)
+{
+  for (u32 i = 0, j = 56; i < j; i += 4, j -= 4)
+  {
+    u32 temp;
+
+    temp = ks[i + 0]; ks[i + 0] = ks[j + 0]; ks[j + 0] = temp;
+    temp = ks[i + 1]; ks[i + 1] = ks[j + 1]; ks[j + 1] = temp;
+    temp = ks[i + 2]; ks[i + 2] = ks[j + 2]; ks[j + 2] = temp;
+    temp = ks[i + 3]; ks[i + 3] = ks[j + 3]; ks[j + 3] = temp;
+  }
+
+  for (u32 i = 1, j = 4; i < 14; i += 1, j += 4)
+  {
+    ks[j + 0] =
+      td0[te1[(ks[j + 0] >> 24) & 0xff] & 0xff] ^
+      td1[te1[(ks[j + 0] >> 16) & 0xff] & 0xff] ^
+      td2[te1[(ks[j + 0] >>  8) & 0xff] & 0xff] ^
+      td3[te1[(ks[j + 0] >>  0) & 0xff] & 0xff];
+
+    ks[j + 1] =
+      td0[te1[(ks[j + 1] >> 24) & 0xff] & 0xff] ^
+      td1[te1[(ks[j + 1] >> 16) & 0xff] & 0xff] ^
+      td2[te1[(ks[j + 1] >>  8) & 0xff] & 0xff] ^
+      td3[te1[(ks[j + 1] >>  0) & 0xff] & 0xff];
+
+    ks[j + 2] =
+      td0[te1[(ks[j + 2] >> 24) & 0xff] & 0xff] ^
+      td1[te1[(ks[j + 2] >> 16) & 0xff] & 0xff] ^
+      td2[te1[(ks[j + 2] >>  8) & 0xff] & 0xff] ^
+      td3[te1[(ks[j + 2] >>  0) & 0xff] & 0xff];
+
+    ks[j + 3] =
+      td0[te1[(ks[j + 3] >> 24) & 0xff] & 0xff] ^
+      td1[te1[(ks[j + 3] >> 16) & 0xff] & 0xff] ^
+      td2[te1[(ks[j + 3] >>  8) & 0xff] & 0xff] ^
+      td3[te1[(ks[j + 3] >>  0) & 0xff] & 0xff];
+  }
+}
+
+static void aes256_set_encrypt_key (u32 *ks, const u32 *ukey)
+{
+  u32 ukey_s[8];
+
+  ukey_s[0] = swap32 (ukey[0]);
+  ukey_s[1] = swap32 (ukey[1]);
+  ukey_s[2] = swap32 (ukey[2]);
+  ukey_s[3] = swap32 (ukey[3]);
+  ukey_s[4] = swap32 (ukey[4]);
+  ukey_s[5] = swap32 (ukey[5]);
+  ukey_s[6] = swap32 (ukey[6]);
+  ukey_s[7] = swap32 (ukey[7]);
+
+  aes256_ExpandKey (ks, ukey_s);
+}
+
+static void aes256_set_decrypt_key (u32 *ks, const u32 *ukey)
+{
+  u32 ukey_s[8];
+
+  ukey_s[0] = swap32 (ukey[0]);
+  ukey_s[1] = swap32 (ukey[1]);
+  ukey_s[2] = swap32 (ukey[2]);
+  ukey_s[3] = swap32 (ukey[3]);
+  ukey_s[4] = swap32 (ukey[4]);
+  ukey_s[5] = swap32 (ukey[5]);
+  ukey_s[6] = swap32 (ukey[6]);
+  ukey_s[7] = swap32 (ukey[7]);
+
+  aes256_ExpandKey (ks, ukey_s);
+
+  aes256_InvertKey (ks);
+}
+
+static void aes256_decrypt (const u32 *ks, const u32 *in, u32 *out)
+{
+  u32 in_s[4];
+
+  in_s[0] = swap32 (in[0]);
+  in_s[1] = swap32 (in[1]);
+  in_s[2] = swap32 (in[2]);
+  in_s[3] = swap32 (in[3]);
+
+  u32 s0 = in_s[0] ^ ks[0];
+  u32 s1 = in_s[1] ^ ks[1];
+  u32 s2 = in_s[2] ^ ks[2];
+  u32 s3 = in_s[3] ^ ks[3];
+
+  u32 t0;
+  u32 t1;
+  u32 t2;
+  u32 t3;
+
+  t0 = td0[s0 >> 24] ^ td1[(s3 >> 16) & 0xff] ^ td2[(s2 >>  8) & 0xff] ^ td3[s1 & 0xff] ^ ks[ 4];
+  t1 = td0[s1 >> 24] ^ td1[(s0 >> 16) & 0xff] ^ td2[(s3 >>  8) & 0xff] ^ td3[s2 & 0xff] ^ ks[ 5];
+  t2 = td0[s2 >> 24] ^ td1[(s1 >> 16) & 0xff] ^ td2[(s0 >>  8) & 0xff] ^ td3[s3 & 0xff] ^ ks[ 6];
+  t3 = td0[s3 >> 24] ^ td1[(s2 >> 16) & 0xff] ^ td2[(s1 >>  8) & 0xff] ^ td3[s0 & 0xff] ^ ks[ 7];
+  s0 = td0[t0 >> 24] ^ td1[(t3 >> 16) & 0xff] ^ td2[(t2 >>  8) & 0xff] ^ td3[t1 & 0xff] ^ ks[ 8];
+  s1 = td0[t1 >> 24] ^ td1[(t0 >> 16) & 0xff] ^ td2[(t3 >>  8) & 0xff] ^ td3[t2 & 0xff] ^ ks[ 9];
+  s2 = td0[t2 >> 24] ^ td1[(t1 >> 16) & 0xff] ^ td2[(t0 >>  8) & 0xff] ^ td3[t3 & 0xff] ^ ks[10];
+  s3 = td0[t3 >> 24] ^ td1[(t2 >> 16) & 0xff] ^ td2[(t1 >>  8) & 0xff] ^ td3[t0 & 0xff] ^ ks[11];
+  t0 = td0[s0 >> 24] ^ td1[(s3 >> 16) & 0xff] ^ td2[(s2 >>  8) & 0xff] ^ td3[s1 & 0xff] ^ ks[12];
+  t1 = td0[s1 >> 24] ^ td1[(s0 >> 16) & 0xff] ^ td2[(s3 >>  8) & 0xff] ^ td3[s2 & 0xff] ^ ks[13];
+  t2 = td0[s2 >> 24] ^ td1[(s1 >> 16) & 0xff] ^ td2[(s0 >>  8) & 0xff] ^ td3[s3 & 0xff] ^ ks[14];
+  t3 = td0[s3 >> 24] ^ td1[(s2 >> 16) & 0xff] ^ td2[(s1 >>  8) & 0xff] ^ td3[s0 & 0xff] ^ ks[15];
+  s0 = td0[t0 >> 24] ^ td1[(t3 >> 16) & 0xff] ^ td2[(t2 >>  8) & 0xff] ^ td3[t1 & 0xff] ^ ks[16];
+  s1 = td0[t1 >> 24] ^ td1[(t0 >> 16) & 0xff] ^ td2[(t3 >>  8) & 0xff] ^ td3[t2 & 0xff] ^ ks[17];
+  s2 = td0[t2 >> 24] ^ td1[(t1 >> 16) & 0xff] ^ td2[(t0 >>  8) & 0xff] ^ td3[t3 & 0xff] ^ ks[18];
+  s3 = td0[t3 >> 24] ^ td1[(t2 >> 16) & 0xff] ^ td2[(t1 >>  8) & 0xff] ^ td3[t0 & 0xff] ^ ks[19];
+  t0 = td0[s0 >> 24] ^ td1[(s3 >> 16) & 0xff] ^ td2[(s2 >>  8) & 0xff] ^ td3[s1 & 0xff] ^ ks[20];
+  t1 = td0[s1 >> 24] ^ td1[(s0 >> 16) & 0xff] ^ td2[(s3 >>  8) & 0xff] ^ td3[s2 & 0xff] ^ ks[21];
+  t2 = td0[s2 >> 24] ^ td1[(s1 >> 16) & 0xff] ^ td2[(s0 >>  8) & 0xff] ^ td3[s3 & 0xff] ^ ks[22];
+  t3 = td0[s3 >> 24] ^ td1[(s2 >> 16) & 0xff] ^ td2[(s1 >>  8) & 0xff] ^ td3[s0 & 0xff] ^ ks[23];
+  s0 = td0[t0 >> 24] ^ td1[(t3 >> 16) & 0xff] ^ td2[(t2 >>  8) & 0xff] ^ td3[t1 & 0xff] ^ ks[24];
+  s1 = td0[t1 >> 24] ^ td1[(t0 >> 16) & 0xff] ^ td2[(t3 >>  8) & 0xff] ^ td3[t2 & 0xff] ^ ks[25];
+  s2 = td0[t2 >> 24] ^ td1[(t1 >> 16) & 0xff] ^ td2[(t0 >>  8) & 0xff] ^ td3[t3 & 0xff] ^ ks[26];
+  s3 = td0[t3 >> 24] ^ td1[(t2 >> 16) & 0xff] ^ td2[(t1 >>  8) & 0xff] ^ td3[t0 & 0xff] ^ ks[27];
+  t0 = td0[s0 >> 24] ^ td1[(s3 >> 16) & 0xff] ^ td2[(s2 >>  8) & 0xff] ^ td3[s1 & 0xff] ^ ks[28];
+  t1 = td0[s1 >> 24] ^ td1[(s0 >> 16) & 0xff] ^ td2[(s3 >>  8) & 0xff] ^ td3[s2 & 0xff] ^ ks[29];
+  t2 = td0[s2 >> 24] ^ td1[(s1 >> 16) & 0xff] ^ td2[(s0 >>  8) & 0xff] ^ td3[s3 & 0xff] ^ ks[30];
+  t3 = td0[s3 >> 24] ^ td1[(s2 >> 16) & 0xff] ^ td2[(s1 >>  8) & 0xff] ^ td3[s0 & 0xff] ^ ks[31];
+  s0 = td0[t0 >> 24] ^ td1[(t3 >> 16) & 0xff] ^ td2[(t2 >>  8) & 0xff] ^ td3[t1 & 0xff] ^ ks[32];
+  s1 = td0[t1 >> 24] ^ td1[(t0 >> 16) & 0xff] ^ td2[(t3 >>  8) & 0xff] ^ td3[t2 & 0xff] ^ ks[33];
+  s2 = td0[t2 >> 24] ^ td1[(t1 >> 16) & 0xff] ^ td2[(t0 >>  8) & 0xff] ^ td3[t3 & 0xff] ^ ks[34];
+  s3 = td0[t3 >> 24] ^ td1[(t2 >> 16) & 0xff] ^ td2[(t1 >>  8) & 0xff] ^ td3[t0 & 0xff] ^ ks[35];
+  t0 = td0[s0 >> 24] ^ td1[(s3 >> 16) & 0xff] ^ td2[(s2 >>  8) & 0xff] ^ td3[s1 & 0xff] ^ ks[36];
+  t1 = td0[s1 >> 24] ^ td1[(s0 >> 16) & 0xff] ^ td2[(s3 >>  8) & 0xff] ^ td3[s2 & 0xff] ^ ks[37];
+  t2 = td0[s2 >> 24] ^ td1[(s1 >> 16) & 0xff] ^ td2[(s0 >>  8) & 0xff] ^ td3[s3 & 0xff] ^ ks[38];
+  t3 = td0[s3 >> 24] ^ td1[(s2 >> 16) & 0xff] ^ td2[(s1 >>  8) & 0xff] ^ td3[s0 & 0xff] ^ ks[39];
+  s0 = td0[t0 >> 24] ^ td1[(t3 >> 16) & 0xff] ^ td2[(t2 >>  8) & 0xff] ^ td3[t1 & 0xff] ^ ks[40];
+  s1 = td0[t1 >> 24] ^ td1[(t0 >> 16) & 0xff] ^ td2[(t3 >>  8) & 0xff] ^ td3[t2 & 0xff] ^ ks[41];
+  s2 = td0[t2 >> 24] ^ td1[(t1 >> 16) & 0xff] ^ td2[(t0 >>  8) & 0xff] ^ td3[t3 & 0xff] ^ ks[42];
+  s3 = td0[t3 >> 24] ^ td1[(t2 >> 16) & 0xff] ^ td2[(t1 >>  8) & 0xff] ^ td3[t0 & 0xff] ^ ks[43];
+  t0 = td0[s0 >> 24] ^ td1[(s3 >> 16) & 0xff] ^ td2[(s2 >>  8) & 0xff] ^ td3[s1 & 0xff] ^ ks[44];
+  t1 = td0[s1 >> 24] ^ td1[(s0 >> 16) & 0xff] ^ td2[(s3 >>  8) & 0xff] ^ td3[s2 & 0xff] ^ ks[45];
+  t2 = td0[s2 >> 24] ^ td1[(s1 >> 16) & 0xff] ^ td2[(s0 >>  8) & 0xff] ^ td3[s3 & 0xff] ^ ks[46];
+  t3 = td0[s3 >> 24] ^ td1[(s2 >> 16) & 0xff] ^ td2[(s1 >>  8) & 0xff] ^ td3[s0 & 0xff] ^ ks[47];
+  s0 = td0[t0 >> 24] ^ td1[(t3 >> 16) & 0xff] ^ td2[(t2 >>  8) & 0xff] ^ td3[t1 & 0xff] ^ ks[48];
+  s1 = td0[t1 >> 24] ^ td1[(t0 >> 16) & 0xff] ^ td2[(t3 >>  8) & 0xff] ^ td3[t2 & 0xff] ^ ks[49];
+  s2 = td0[t2 >> 24] ^ td1[(t1 >> 16) & 0xff] ^ td2[(t0 >>  8) & 0xff] ^ td3[t3 & 0xff] ^ ks[50];
+  s3 = td0[t3 >> 24] ^ td1[(t2 >> 16) & 0xff] ^ td2[(t1 >>  8) & 0xff] ^ td3[t0 & 0xff] ^ ks[51];
+  t0 = td0[s0 >> 24] ^ td1[(s3 >> 16) & 0xff] ^ td2[(s2 >>  8) & 0xff] ^ td3[s1 & 0xff] ^ ks[52];
+  t1 = td0[s1 >> 24] ^ td1[(s0 >> 16) & 0xff] ^ td2[(s3 >>  8) & 0xff] ^ td3[s2 & 0xff] ^ ks[53];
+  t2 = td0[s2 >> 24] ^ td1[(s1 >> 16) & 0xff] ^ td2[(s0 >>  8) & 0xff] ^ td3[s3 & 0xff] ^ ks[54];
+  t3 = td0[s3 >> 24] ^ td1[(s2 >> 16) & 0xff] ^ td2[(s1 >>  8) & 0xff] ^ td3[s0 & 0xff] ^ ks[55];
+
+  out[0] = (td4[(t0 >> 24) & 0xff] & 0xff000000)
+         ^ (td4[(t3 >> 16) & 0xff] & 0x00ff0000)
+         ^ (td4[(t2 >>  8) & 0xff] & 0x0000ff00)
+         ^ (td4[(t1 >>  0) & 0xff] & 0x000000ff)
+         ^ ks[56];
+
+  out[1] = (td4[(t1 >> 24) & 0xff] & 0xff000000)
+         ^ (td4[(t0 >> 16) & 0xff] & 0x00ff0000)
+         ^ (td4[(t3 >>  8) & 0xff] & 0x0000ff00)
+         ^ (td4[(t2 >>  0) & 0xff] & 0x000000ff)
+         ^ ks[57];
+
+  out[2] = (td4[(t2 >> 24) & 0xff] & 0xff000000)
+         ^ (td4[(t1 >> 16) & 0xff] & 0x00ff0000)
+         ^ (td4[(t0 >>  8) & 0xff] & 0x0000ff00)
+         ^ (td4[(t3 >>  0) & 0xff] & 0x000000ff)
+         ^ ks[58];
+
+  out[3] = (td4[(t3 >> 24) & 0xff] & 0xff000000)
+         ^ (td4[(t2 >> 16) & 0xff] & 0x00ff0000)
+         ^ (td4[(t1 >>  8) & 0xff] & 0x0000ff00)
+         ^ (td4[(t0 >>  0) & 0xff] & 0x000000ff)
+         ^ ks[59];
+
+  out[0] = swap32 (out[0]);
+  out[1] = swap32 (out[1]);
+  out[2] = swap32 (out[2]);
+  out[3] = swap32 (out[3]);
+}
+
+static void aes256_encrypt (const u32 *ks, const u32 *in, u32 *out)
+{
+  u32 in_s[4];
+
+  in_s[0] = swap32 (in[0]);
+  in_s[1] = swap32 (in[1]);
+  in_s[2] = swap32 (in[2]);
+  in_s[3] = swap32 (in[3]);
+
+  u32 s0 = in_s[0] ^ ks[0];
+  u32 s1 = in_s[1] ^ ks[1];
+  u32 s2 = in_s[2] ^ ks[2];
+  u32 s3 = in_s[3] ^ ks[3];
+
+  u32 t0;
+  u32 t1;
+  u32 t2;
+  u32 t3;
+
+  t0 = te0[s0 >> 24] ^ te1[(s1 >> 16) & 0xff] ^ te2[(s2 >>  8) & 0xff] ^ te3[s3 & 0xff] ^ ks[ 4];
+  t1 = te0[s1 >> 24] ^ te1[(s2 >> 16) & 0xff] ^ te2[(s3 >>  8) & 0xff] ^ te3[s0 & 0xff] ^ ks[ 5];
+  t2 = te0[s2 >> 24] ^ te1[(s3 >> 16) & 0xff] ^ te2[(s0 >>  8) & 0xff] ^ te3[s1 & 0xff] ^ ks[ 6];
+  t3 = te0[s3 >> 24] ^ te1[(s0 >> 16) & 0xff] ^ te2[(s1 >>  8) & 0xff] ^ te3[s2 & 0xff] ^ ks[ 7];
+  s0 = te0[t0 >> 24] ^ te1[(t1 >> 16) & 0xff] ^ te2[(t2 >>  8) & 0xff] ^ te3[t3 & 0xff] ^ ks[ 8];
+  s1 = te0[t1 >> 24] ^ te1[(t2 >> 16) & 0xff] ^ te2[(t3 >>  8) & 0xff] ^ te3[t0 & 0xff] ^ ks[ 9];
+  s2 = te0[t2 >> 24] ^ te1[(t3 >> 16) & 0xff] ^ te2[(t0 >>  8) & 0xff] ^ te3[t1 & 0xff] ^ ks[10];
+  s3 = te0[t3 >> 24] ^ te1[(t0 >> 16) & 0xff] ^ te2[(t1 >>  8) & 0xff] ^ te3[t2 & 0xff] ^ ks[11];
+  t0 = te0[s0 >> 24] ^ te1[(s1 >> 16) & 0xff] ^ te2[(s2 >>  8) & 0xff] ^ te3[s3 & 0xff] ^ ks[12];
+  t1 = te0[s1 >> 24] ^ te1[(s2 >> 16) & 0xff] ^ te2[(s3 >>  8) & 0xff] ^ te3[s0 & 0xff] ^ ks[13];
+  t2 = te0[s2 >> 24] ^ te1[(s3 >> 16) & 0xff] ^ te2[(s0 >>  8) & 0xff] ^ te3[s1 & 0xff] ^ ks[14];
+  t3 = te0[s3 >> 24] ^ te1[(s0 >> 16) & 0xff] ^ te2[(s1 >>  8) & 0xff] ^ te3[s2 & 0xff] ^ ks[15];
+  s0 = te0[t0 >> 24] ^ te1[(t1 >> 16) & 0xff] ^ te2[(t2 >>  8) & 0xff] ^ te3[t3 & 0xff] ^ ks[16];
+  s1 = te0[t1 >> 24] ^ te1[(t2 >> 16) & 0xff] ^ te2[(t3 >>  8) & 0xff] ^ te3[t0 & 0xff] ^ ks[17];
+  s2 = te0[t2 >> 24] ^ te1[(t3 >> 16) & 0xff] ^ te2[(t0 >>  8) & 0xff] ^ te3[t1 & 0xff] ^ ks[18];
+  s3 = te0[t3 >> 24] ^ te1[(t0 >> 16) & 0xff] ^ te2[(t1 >>  8) & 0xff] ^ te3[t2 & 0xff] ^ ks[19];
+  t0 = te0[s0 >> 24] ^ te1[(s1 >> 16) & 0xff] ^ te2[(s2 >>  8) & 0xff] ^ te3[s3 & 0xff] ^ ks[20];
+  t1 = te0[s1 >> 24] ^ te1[(s2 >> 16) & 0xff] ^ te2[(s3 >>  8) & 0xff] ^ te3[s0 & 0xff] ^ ks[21];
+  t2 = te0[s2 >> 24] ^ te1[(s3 >> 16) & 0xff] ^ te2[(s0 >>  8) & 0xff] ^ te3[s1 & 0xff] ^ ks[22];
+  t3 = te0[s3 >> 24] ^ te1[(s0 >> 16) & 0xff] ^ te2[(s1 >>  8) & 0xff] ^ te3[s2 & 0xff] ^ ks[23];
+  s0 = te0[t0 >> 24] ^ te1[(t1 >> 16) & 0xff] ^ te2[(t2 >>  8) & 0xff] ^ te3[t3 & 0xff] ^ ks[24];
+  s1 = te0[t1 >> 24] ^ te1[(t2 >> 16) & 0xff] ^ te2[(t3 >>  8) & 0xff] ^ te3[t0 & 0xff] ^ ks[25];
+  s2 = te0[t2 >> 24] ^ te1[(t3 >> 16) & 0xff] ^ te2[(t0 >>  8) & 0xff] ^ te3[t1 & 0xff] ^ ks[26];
+  s3 = te0[t3 >> 24] ^ te1[(t0 >> 16) & 0xff] ^ te2[(t1 >>  8) & 0xff] ^ te3[t2 & 0xff] ^ ks[27];
+  t0 = te0[s0 >> 24] ^ te1[(s1 >> 16) & 0xff] ^ te2[(s2 >>  8) & 0xff] ^ te3[s3 & 0xff] ^ ks[28];
+  t1 = te0[s1 >> 24] ^ te1[(s2 >> 16) & 0xff] ^ te2[(s3 >>  8) & 0xff] ^ te3[s0 & 0xff] ^ ks[29];
+  t2 = te0[s2 >> 24] ^ te1[(s3 >> 16) & 0xff] ^ te2[(s0 >>  8) & 0xff] ^ te3[s1 & 0xff] ^ ks[30];
+  t3 = te0[s3 >> 24] ^ te1[(s0 >> 16) & 0xff] ^ te2[(s1 >>  8) & 0xff] ^ te3[s2 & 0xff] ^ ks[31];
+  s0 = te0[t0 >> 24] ^ te1[(t1 >> 16) & 0xff] ^ te2[(t2 >>  8) & 0xff] ^ te3[t3 & 0xff] ^ ks[32];
+  s1 = te0[t1 >> 24] ^ te1[(t2 >> 16) & 0xff] ^ te2[(t3 >>  8) & 0xff] ^ te3[t0 & 0xff] ^ ks[33];
+  s2 = te0[t2 >> 24] ^ te1[(t3 >> 16) & 0xff] ^ te2[(t0 >>  8) & 0xff] ^ te3[t1 & 0xff] ^ ks[34];
+  s3 = te0[t3 >> 24] ^ te1[(t0 >> 16) & 0xff] ^ te2[(t1 >>  8) & 0xff] ^ te3[t2 & 0xff] ^ ks[35];
+  t0 = te0[s0 >> 24] ^ te1[(s1 >> 16) & 0xff] ^ te2[(s2 >>  8) & 0xff] ^ te3[s3 & 0xff] ^ ks[36];
+  t1 = te0[s1 >> 24] ^ te1[(s2 >> 16) & 0xff] ^ te2[(s3 >>  8) & 0xff] ^ te3[s0 & 0xff] ^ ks[37];
+  t2 = te0[s2 >> 24] ^ te1[(s3 >> 16) & 0xff] ^ te2[(s0 >>  8) & 0xff] ^ te3[s1 & 0xff] ^ ks[38];
+  t3 = te0[s3 >> 24] ^ te1[(s0 >> 16) & 0xff] ^ te2[(s1 >>  8) & 0xff] ^ te3[s2 & 0xff] ^ ks[39];
+  s0 = te0[t0 >> 24] ^ te1[(t1 >> 16) & 0xff] ^ te2[(t2 >>  8) & 0xff] ^ te3[t3 & 0xff] ^ ks[40];
+  s1 = te0[t1 >> 24] ^ te1[(t2 >> 16) & 0xff] ^ te2[(t3 >>  8) & 0xff] ^ te3[t0 & 0xff] ^ ks[41];
+  s2 = te0[t2 >> 24] ^ te1[(t3 >> 16) & 0xff] ^ te2[(t0 >>  8) & 0xff] ^ te3[t1 & 0xff] ^ ks[42];
+  s3 = te0[t3 >> 24] ^ te1[(t0 >> 16) & 0xff] ^ te2[(t1 >>  8) & 0xff] ^ te3[t2 & 0xff] ^ ks[43];
+  t0 = te0[s0 >> 24] ^ te1[(s1 >> 16) & 0xff] ^ te2[(s2 >>  8) & 0xff] ^ te3[s3 & 0xff] ^ ks[44];
+  t1 = te0[s1 >> 24] ^ te1[(s2 >> 16) & 0xff] ^ te2[(s3 >>  8) & 0xff] ^ te3[s0 & 0xff] ^ ks[45];
+  t2 = te0[s2 >> 24] ^ te1[(s3 >> 16) & 0xff] ^ te2[(s0 >>  8) & 0xff] ^ te3[s1 & 0xff] ^ ks[46];
+  t3 = te0[s3 >> 24] ^ te1[(s0 >> 16) & 0xff] ^ te2[(s1 >>  8) & 0xff] ^ te3[s2 & 0xff] ^ ks[47];
+  s0 = te0[t0 >> 24] ^ te1[(t1 >> 16) & 0xff] ^ te2[(t2 >>  8) & 0xff] ^ te3[t3 & 0xff] ^ ks[48];
+  s1 = te0[t1 >> 24] ^ te1[(t2 >> 16) & 0xff] ^ te2[(t3 >>  8) & 0xff] ^ te3[t0 & 0xff] ^ ks[49];
+  s2 = te0[t2 >> 24] ^ te1[(t3 >> 16) & 0xff] ^ te2[(t0 >>  8) & 0xff] ^ te3[t1 & 0xff] ^ ks[50];
+  s3 = te0[t3 >> 24] ^ te1[(t0 >> 16) & 0xff] ^ te2[(t1 >>  8) & 0xff] ^ te3[t2 & 0xff] ^ ks[51];
+  t0 = te0[s0 >> 24] ^ te1[(s1 >> 16) & 0xff] ^ te2[(s2 >>  8) & 0xff] ^ te3[s3 & 0xff] ^ ks[52];
+  t1 = te0[s1 >> 24] ^ te1[(s2 >> 16) & 0xff] ^ te2[(s3 >>  8) & 0xff] ^ te3[s0 & 0xff] ^ ks[53];
+  t2 = te0[s2 >> 24] ^ te1[(s3 >> 16) & 0xff] ^ te2[(s0 >>  8) & 0xff] ^ te3[s1 & 0xff] ^ ks[54];
+  t3 = te0[s3 >> 24] ^ te1[(s0 >> 16) & 0xff] ^ te2[(s1 >>  8) & 0xff] ^ te3[s2 & 0xff] ^ ks[55];
+
+  out[0] = (te4[(t0 >> 24) & 0xff] & 0xff000000)
+         ^ (te4[(t1 >> 16) & 0xff] & 0x00ff0000)
+         ^ (te4[(t2 >>  8) & 0xff] & 0x0000ff00)
+         ^ (te4[(t3 >>  0) & 0xff] & 0x000000ff)
+         ^ ks[56];
+
+  out[1] = (te4[(t1 >> 24) & 0xff] & 0xff000000)
+         ^ (te4[(t2 >> 16) & 0xff] & 0x00ff0000)
+         ^ (te4[(t3 >>  8) & 0xff] & 0x0000ff00)
+         ^ (te4[(t0 >>  0) & 0xff] & 0x000000ff)
+         ^ ks[57];
+
+  out[2] = (te4[(t2 >> 24) & 0xff] & 0xff000000)
+         ^ (te4[(t3 >> 16) & 0xff] & 0x00ff0000)
+         ^ (te4[(t0 >>  8) & 0xff] & 0x0000ff00)
+         ^ (te4[(t1 >>  0) & 0xff] & 0x000000ff)
+         ^ ks[58];
+
+  out[3] = (te4[(t3 >> 24) & 0xff] & 0xff000000)
+         ^ (te4[(t0 >> 16) & 0xff] & 0x00ff0000)
+         ^ (te4[(t1 >>  8) & 0xff] & 0x0000ff00)
+         ^ (te4[(t2 >>  0) & 0xff] & 0x000000ff)
+         ^ ks[59];
+
+  out[0] = swap32 (out[0]);
+  out[1] = swap32 (out[1]);
+  out[2] = swap32 (out[2]);
+  out[3] = swap32 (out[3]);
+}
+
+static void aes256_decrypt_xts (const u32 *ukey1, const u32 *ukey2, const u32 *in, u32 *out)
+{
+  u32 T[4] = { 0 };
+  u32 Z[4] = { 0 };
+
+  out[0] = in[0];
+  out[1] = in[1];
+  out[2] = in[2];
+  out[3] = in[3];
+
+  u32 ks[60];
+
+  aes256_set_encrypt_key (ks, ukey2);
+  aes256_encrypt (ks, Z, T);
+
+  out[0] ^= T[0];
+  out[1] ^= T[1];
+  out[2] ^= T[2];
+  out[3] ^= T[3];
+
+  aes256_set_decrypt_key (ks, ukey1);
+  aes256_decrypt (ks, out, out);
+
+  out[0] ^= T[0];
+  out[1] ^= T[1];
+  out[2] ^= T[2];
+  out[3] ^= T[3];
+}
diff --git a/OpenCL/kernel_aes256_amd.c b/OpenCL/kernel_aes256_amd.c
deleted file mode 100644
index 1127165..0000000
--- a/OpenCL/kernel_aes256_amd.c
+++ /dev/null
@@ -1,1048 +0,0 @@
-/**
- * Author......: Jens Steube <jens.steube@gmail.com>
- * License.....: MIT
- */
-
-__constant u32 te0[256] =
-{
-  0xc66363a5, 0xf87c7c84, 0xee777799, 0xf67b7b8d,
-  0xfff2f20d, 0xd66b6bbd, 0xde6f6fb1, 0x91c5c554,
-  0x60303050, 0x02010103, 0xce6767a9, 0x562b2b7d,
-  0xe7fefe19, 0xb5d7d762, 0x4dababe6, 0xec76769a,
-  0x8fcaca45, 0x1f82829d, 0x89c9c940, 0xfa7d7d87,
-  0xeffafa15, 0xb25959eb, 0x8e4747c9, 0xfbf0f00b,
-  0x41adadec, 0xb3d4d467, 0x5fa2a2fd, 0x45afafea,
-  0x239c9cbf, 0x53a4a4f7, 0xe4727296, 0x9bc0c05b,
-  0x75b7b7c2, 0xe1fdfd1c, 0x3d9393ae, 0x4c26266a,
-  0x6c36365a, 0x7e3f3f41, 0xf5f7f702, 0x83cccc4f,
-  0x6834345c, 0x51a5a5f4, 0xd1e5e534, 0xf9f1f108,
-  0xe2717193, 0xabd8d873, 0x62313153, 0x2a15153f,
-  0x0804040c, 0x95c7c752, 0x46232365, 0x9dc3c35e,
-  0x30181828, 0x379696a1, 0x0a05050f, 0x2f9a9ab5,
-  0x0e070709, 0x24121236, 0x1b80809b, 0xdfe2e23d,
-  0xcdebeb26, 0x4e272769, 0x7fb2b2cd, 0xea75759f,
-  0x1209091b, 0x1d83839e, 0x582c2c74, 0x341a1a2e,
-  0x361b1b2d, 0xdc6e6eb2, 0xb45a5aee, 0x5ba0a0fb,
-  0xa45252f6, 0x763b3b4d, 0xb7d6d661, 0x7db3b3ce,
-  0x5229297b, 0xdde3e33e, 0x5e2f2f71, 0x13848497,
-  0xa65353f5, 0xb9d1d168, 0x00000000, 0xc1eded2c,
-  0x40202060, 0xe3fcfc1f, 0x79b1b1c8, 0xb65b5bed,
-  0xd46a6abe, 0x8dcbcb46, 0x67bebed9, 0x7239394b,
-  0x944a4ade, 0x984c4cd4, 0xb05858e8, 0x85cfcf4a,
-  0xbbd0d06b, 0xc5efef2a, 0x4faaaae5, 0xedfbfb16,
-  0x864343c5, 0x9a4d4dd7, 0x66333355, 0x11858594,
-  0x8a4545cf, 0xe9f9f910, 0x04020206, 0xfe7f7f81,
-  0xa05050f0, 0x783c3c44, 0x259f9fba, 0x4ba8a8e3,
-  0xa25151f3, 0x5da3a3fe, 0x804040c0, 0x058f8f8a,
-  0x3f9292ad, 0x219d9dbc, 0x70383848, 0xf1f5f504,
-  0x63bcbcdf, 0x77b6b6c1, 0xafdada75, 0x42212163,
-  0x20101030, 0xe5ffff1a, 0xfdf3f30e, 0xbfd2d26d,
-  0x81cdcd4c, 0x180c0c14, 0x26131335, 0xc3ecec2f,
-  0xbe5f5fe1, 0x359797a2, 0x884444cc, 0x2e171739,
-  0x93c4c457, 0x55a7a7f2, 0xfc7e7e82, 0x7a3d3d47,
-  0xc86464ac, 0xba5d5de7, 0x3219192b, 0xe6737395,
-  0xc06060a0, 0x19818198, 0x9e4f4fd1, 0xa3dcdc7f,
-  0x44222266, 0x542a2a7e, 0x3b9090ab, 0x0b888883,
-  0x8c4646ca, 0xc7eeee29, 0x6bb8b8d3, 0x2814143c,
-  0xa7dede79, 0xbc5e5ee2, 0x160b0b1d, 0xaddbdb76,
-  0xdbe0e03b, 0x64323256, 0x743a3a4e, 0x140a0a1e,
-  0x924949db, 0x0c06060a, 0x4824246c, 0xb85c5ce4,
-  0x9fc2c25d, 0xbdd3d36e, 0x43acacef, 0xc46262a6,
-  0x399191a8, 0x319595a4, 0xd3e4e437, 0xf279798b,
-  0xd5e7e732, 0x8bc8c843, 0x6e373759, 0xda6d6db7,
-  0x018d8d8c, 0xb1d5d564, 0x9c4e4ed2, 0x49a9a9e0,
-  0xd86c6cb4, 0xac5656fa, 0xf3f4f407, 0xcfeaea25,
-  0xca6565af, 0xf47a7a8e, 0x47aeaee9, 0x10080818,
-  0x6fbabad5, 0xf0787888, 0x4a25256f, 0x5c2e2e72,
-  0x381c1c24, 0x57a6a6f1, 0x73b4b4c7, 0x97c6c651,
-  0xcbe8e823, 0xa1dddd7c, 0xe874749c, 0x3e1f1f21,
-  0x964b4bdd, 0x61bdbddc, 0x0d8b8b86, 0x0f8a8a85,
-  0xe0707090, 0x7c3e3e42, 0x71b5b5c4, 0xcc6666aa,
-  0x904848d8, 0x06030305, 0xf7f6f601, 0x1c0e0e12,
-  0xc26161a3, 0x6a35355f, 0xae5757f9, 0x69b9b9d0,
-  0x17868691, 0x99c1c158, 0x3a1d1d27, 0x279e9eb9,
-  0xd9e1e138, 0xebf8f813, 0x2b9898b3, 0x22111133,
-  0xd26969bb, 0xa9d9d970, 0x078e8e89, 0x339494a7,
-  0x2d9b9bb6, 0x3c1e1e22, 0x15878792, 0xc9e9e920,
-  0x87cece49, 0xaa5555ff, 0x50282878, 0xa5dfdf7a,
-  0x038c8c8f, 0x59a1a1f8, 0x09898980, 0x1a0d0d17,
-  0x65bfbfda, 0xd7e6e631, 0x844242c6, 0xd06868b8,
-  0x824141c3, 0x299999b0, 0x5a2d2d77, 0x1e0f0f11,
-  0x7bb0b0cb, 0xa85454fc, 0x6dbbbbd6, 0x2c16163a,
-};
-
-__constant u32 te1[256] =
-{
-  0xa5c66363, 0x84f87c7c, 0x99ee7777, 0x8df67b7b,
-  0x0dfff2f2, 0xbdd66b6b, 0xb1de6f6f, 0x5491c5c5,
-  0x50603030, 0x03020101, 0xa9ce6767, 0x7d562b2b,
-  0x19e7fefe, 0x62b5d7d7, 0xe64dabab, 0x9aec7676,
-  0x458fcaca, 0x9d1f8282, 0x4089c9c9, 0x87fa7d7d,
-  0x15effafa, 0xebb25959, 0xc98e4747, 0x0bfbf0f0,
-  0xec41adad, 0x67b3d4d4, 0xfd5fa2a2, 0xea45afaf,
-  0xbf239c9c, 0xf753a4a4, 0x96e47272, 0x5b9bc0c0,
-  0xc275b7b7, 0x1ce1fdfd, 0xae3d9393, 0x6a4c2626,
-  0x5a6c3636, 0x417e3f3f, 0x02f5f7f7, 0x4f83cccc,
-  0x5c683434, 0xf451a5a5, 0x34d1e5e5, 0x08f9f1f1,
-  0x93e27171, 0x73abd8d8, 0x53623131, 0x3f2a1515,
-  0x0c080404, 0x5295c7c7, 0x65462323, 0x5e9dc3c3,
-  0x28301818, 0xa1379696, 0x0f0a0505, 0xb52f9a9a,
-  0x090e0707, 0x36241212, 0x9b1b8080, 0x3ddfe2e2,
-  0x26cdebeb, 0x694e2727, 0xcd7fb2b2, 0x9fea7575,
-  0x1b120909, 0x9e1d8383, 0x74582c2c, 0x2e341a1a,
-  0x2d361b1b, 0xb2dc6e6e, 0xeeb45a5a, 0xfb5ba0a0,
-  0xf6a45252, 0x4d763b3b, 0x61b7d6d6, 0xce7db3b3,
-  0x7b522929, 0x3edde3e3, 0x715e2f2f, 0x97138484,
-  0xf5a65353, 0x68b9d1d1, 0x00000000, 0x2cc1eded,
-  0x60402020, 0x1fe3fcfc, 0xc879b1b1, 0xedb65b5b,
-  0xbed46a6a, 0x468dcbcb, 0xd967bebe, 0x4b723939,
-  0xde944a4a, 0xd4984c4c, 0xe8b05858, 0x4a85cfcf,
-  0x6bbbd0d0, 0x2ac5efef, 0xe54faaaa, 0x16edfbfb,
-  0xc5864343, 0xd79a4d4d, 0x55663333, 0x94118585,
-  0xcf8a4545, 0x10e9f9f9, 0x06040202, 0x81fe7f7f,
-  0xf0a05050, 0x44783c3c, 0xba259f9f, 0xe34ba8a8,
-  0xf3a25151, 0xfe5da3a3, 0xc0804040, 0x8a058f8f,
-  0xad3f9292, 0xbc219d9d, 0x48703838, 0x04f1f5f5,
-  0xdf63bcbc, 0xc177b6b6, 0x75afdada, 0x63422121,
-  0x30201010, 0x1ae5ffff, 0x0efdf3f3, 0x6dbfd2d2,
-  0x4c81cdcd, 0x14180c0c, 0x35261313, 0x2fc3ecec,
-  0xe1be5f5f, 0xa2359797, 0xcc884444, 0x392e1717,
-  0x5793c4c4, 0xf255a7a7, 0x82fc7e7e, 0x477a3d3d,
-  0xacc86464, 0xe7ba5d5d, 0x2b321919, 0x95e67373,
-  0xa0c06060, 0x98198181, 0xd19e4f4f, 0x7fa3dcdc,
-  0x66442222, 0x7e542a2a, 0xab3b9090, 0x830b8888,
-  0xca8c4646, 0x29c7eeee, 0xd36bb8b8, 0x3c281414,
-  0x79a7dede, 0xe2bc5e5e, 0x1d160b0b, 0x76addbdb,
-  0x3bdbe0e0, 0x56643232, 0x4e743a3a, 0x1e140a0a,
-  0xdb924949, 0x0a0c0606, 0x6c482424, 0xe4b85c5c,
-  0x5d9fc2c2, 0x6ebdd3d3, 0xef43acac, 0xa6c46262,
-  0xa8399191, 0xa4319595, 0x37d3e4e4, 0x8bf27979,
-  0x32d5e7e7, 0x438bc8c8, 0x596e3737, 0xb7da6d6d,
-  0x8c018d8d, 0x64b1d5d5, 0xd29c4e4e, 0xe049a9a9,
-  0xb4d86c6c, 0xfaac5656, 0x07f3f4f4, 0x25cfeaea,
-  0xafca6565, 0x8ef47a7a, 0xe947aeae, 0x18100808,
-  0xd56fbaba, 0x88f07878, 0x6f4a2525, 0x725c2e2e,
-  0x24381c1c, 0xf157a6a6, 0xc773b4b4, 0x5197c6c6,
-  0x23cbe8e8, 0x7ca1dddd, 0x9ce87474, 0x213e1f1f,
-  0xdd964b4b, 0xdc61bdbd, 0x860d8b8b, 0x850f8a8a,
-  0x90e07070, 0x427c3e3e, 0xc471b5b5, 0xaacc6666,
-  0xd8904848, 0x05060303, 0x01f7f6f6, 0x121c0e0e,
-  0xa3c26161, 0x5f6a3535, 0xf9ae5757, 0xd069b9b9,
-  0x91178686, 0x5899c1c1, 0x273a1d1d, 0xb9279e9e,
-  0x38d9e1e1, 0x13ebf8f8, 0xb32b9898, 0x33221111,
-  0xbbd26969, 0x70a9d9d9, 0x89078e8e, 0xa7339494,
-  0xb62d9b9b, 0x223c1e1e, 0x92158787, 0x20c9e9e9,
-  0x4987cece, 0xffaa5555, 0x78502828, 0x7aa5dfdf,
-  0x8f038c8c, 0xf859a1a1, 0x80098989, 0x171a0d0d,
-  0xda65bfbf, 0x31d7e6e6, 0xc6844242, 0xb8d06868,
-  0xc3824141, 0xb0299999, 0x775a2d2d, 0x111e0f0f,
-  0xcb7bb0b0, 0xfca85454, 0xd66dbbbb, 0x3a2c1616,
-};
-
-__constant u32 te2[256] =
-{
-  0x63a5c663, 0x7c84f87c, 0x7799ee77, 0x7b8df67b,
-  0xf20dfff2, 0x6bbdd66b, 0x6fb1de6f, 0xc55491c5,
-  0x30506030, 0x01030201, 0x67a9ce67, 0x2b7d562b,
-  0xfe19e7fe, 0xd762b5d7, 0xabe64dab, 0x769aec76,
-  0xca458fca, 0x829d1f82, 0xc94089c9, 0x7d87fa7d,
-  0xfa15effa, 0x59ebb259, 0x47c98e47, 0xf00bfbf0,
-  0xadec41ad, 0xd467b3d4, 0xa2fd5fa2, 0xafea45af,
-  0x9cbf239c, 0xa4f753a4, 0x7296e472, 0xc05b9bc0,
-  0xb7c275b7, 0xfd1ce1fd, 0x93ae3d93, 0x266a4c26,
-  0x365a6c36, 0x3f417e3f, 0xf702f5f7, 0xcc4f83cc,
-  0x345c6834, 0xa5f451a5, 0xe534d1e5, 0xf108f9f1,
-  0x7193e271, 0xd873abd8, 0x31536231, 0x153f2a15,
-  0x040c0804, 0xc75295c7, 0x23654623, 0xc35e9dc3,
-  0x18283018, 0x96a13796, 0x050f0a05, 0x9ab52f9a,
-  0x07090e07, 0x12362412, 0x809b1b80, 0xe23ddfe2,
-  0xeb26cdeb, 0x27694e27, 0xb2cd7fb2, 0x759fea75,
-  0x091b1209, 0x839e1d83, 0x2c74582c, 0x1a2e341a,
-  0x1b2d361b, 0x6eb2dc6e, 0x5aeeb45a, 0xa0fb5ba0,
-  0x52f6a452, 0x3b4d763b, 0xd661b7d6, 0xb3ce7db3,
-  0x297b5229, 0xe33edde3, 0x2f715e2f, 0x84971384,
-  0x53f5a653, 0xd168b9d1, 0x00000000, 0xed2cc1ed,
-  0x20604020, 0xfc1fe3fc, 0xb1c879b1, 0x5bedb65b,
-  0x6abed46a, 0xcb468dcb, 0xbed967be, 0x394b7239,
-  0x4ade944a, 0x4cd4984c, 0x58e8b058, 0xcf4a85cf,
-  0xd06bbbd0, 0xef2ac5ef, 0xaae54faa, 0xfb16edfb,
-  0x43c58643, 0x4dd79a4d, 0x33556633, 0x85941185,
-  0x45cf8a45, 0xf910e9f9, 0x02060402, 0x7f81fe7f,
-  0x50f0a050, 0x3c44783c, 0x9fba259f, 0xa8e34ba8,
-  0x51f3a251, 0xa3fe5da3, 0x40c08040, 0x8f8a058f,
-  0x92ad3f92, 0x9dbc219d, 0x38487038, 0xf504f1f5,
-  0xbcdf63bc, 0xb6c177b6, 0xda75afda, 0x21634221,
-  0x10302010, 0xff1ae5ff, 0xf30efdf3, 0xd26dbfd2,
-  0xcd4c81cd, 0x0c14180c, 0x13352613, 0xec2fc3ec,
-  0x5fe1be5f, 0x97a23597, 0x44cc8844, 0x17392e17,
-  0xc45793c4, 0xa7f255a7, 0x7e82fc7e, 0x3d477a3d,
-  0x64acc864, 0x5de7ba5d, 0x192b3219, 0x7395e673,
-  0x60a0c060, 0x81981981, 0x4fd19e4f, 0xdc7fa3dc,
-  0x22664422, 0x2a7e542a, 0x90ab3b90, 0x88830b88,
-  0x46ca8c46, 0xee29c7ee, 0xb8d36bb8, 0x143c2814,
-  0xde79a7de, 0x5ee2bc5e, 0x0b1d160b, 0xdb76addb,
-  0xe03bdbe0, 0x32566432, 0x3a4e743a, 0x0a1e140a,
-  0x49db9249, 0x060a0c06, 0x246c4824, 0x5ce4b85c,
-  0xc25d9fc2, 0xd36ebdd3, 0xacef43ac, 0x62a6c462,
-  0x91a83991, 0x95a43195, 0xe437d3e4, 0x798bf279,
-  0xe732d5e7, 0xc8438bc8, 0x37596e37, 0x6db7da6d,
-  0x8d8c018d, 0xd564b1d5, 0x4ed29c4e, 0xa9e049a9,
-  0x6cb4d86c, 0x56faac56, 0xf407f3f4, 0xea25cfea,
-  0x65afca65, 0x7a8ef47a, 0xaee947ae, 0x08181008,
-  0xbad56fba, 0x7888f078, 0x256f4a25, 0x2e725c2e,
-  0x1c24381c, 0xa6f157a6, 0xb4c773b4, 0xc65197c6,
-  0xe823cbe8, 0xdd7ca1dd, 0x749ce874, 0x1f213e1f,
-  0x4bdd964b, 0xbddc61bd, 0x8b860d8b, 0x8a850f8a,
-  0x7090e070, 0x3e427c3e, 0xb5c471b5, 0x66aacc66,
-  0x48d89048, 0x03050603, 0xf601f7f6, 0x0e121c0e,
-  0x61a3c261, 0x355f6a35, 0x57f9ae57, 0xb9d069b9,
-  0x86911786, 0xc15899c1, 0x1d273a1d, 0x9eb9279e,
-  0xe138d9e1, 0xf813ebf8, 0x98b32b98, 0x11332211,
-  0x69bbd269, 0xd970a9d9, 0x8e89078e, 0x94a73394,
-  0x9bb62d9b, 0x1e223c1e, 0x87921587, 0xe920c9e9,
-  0xce4987ce, 0x55ffaa55, 0x28785028, 0xdf7aa5df,
-  0x8c8f038c, 0xa1f859a1, 0x89800989, 0x0d171a0d,
-  0xbfda65bf, 0xe631d7e6, 0x42c68442, 0x68b8d068,
-  0x41c38241, 0x99b02999, 0x2d775a2d, 0x0f111e0f,
-  0xb0cb7bb0, 0x54fca854, 0xbbd66dbb, 0x163a2c16,
-};
-
-__constant u32 te3[256] =
-{
-  0x6363a5c6, 0x7c7c84f8, 0x777799ee, 0x7b7b8df6,
-  0xf2f20dff, 0x6b6bbdd6, 0x6f6fb1de, 0xc5c55491,
-  0x30305060, 0x01010302, 0x6767a9ce, 0x2b2b7d56,
-  0xfefe19e7, 0xd7d762b5, 0xababe64d, 0x76769aec,
-  0xcaca458f, 0x82829d1f, 0xc9c94089, 0x7d7d87fa,
-  0xfafa15ef, 0x5959ebb2, 0x4747c98e, 0xf0f00bfb,
-  0xadadec41, 0xd4d467b3, 0xa2a2fd5f, 0xafafea45,
-  0x9c9cbf23, 0xa4a4f753, 0x727296e4, 0xc0c05b9b,
-  0xb7b7c275, 0xfdfd1ce1, 0x9393ae3d, 0x26266a4c,
-  0x36365a6c, 0x3f3f417e, 0xf7f702f5, 0xcccc4f83,
-  0x34345c68, 0xa5a5f451, 0xe5e534d1, 0xf1f108f9,
-  0x717193e2, 0xd8d873ab, 0x31315362, 0x15153f2a,
-  0x04040c08, 0xc7c75295, 0x23236546, 0xc3c35e9d,
-  0x18182830, 0x9696a137, 0x05050f0a, 0x9a9ab52f,
-  0x0707090e, 0x12123624, 0x80809b1b, 0xe2e23ddf,
-  0xebeb26cd, 0x2727694e, 0xb2b2cd7f, 0x75759fea,
-  0x09091b12, 0x83839e1d, 0x2c2c7458, 0x1a1a2e34,
-  0x1b1b2d36, 0x6e6eb2dc, 0x5a5aeeb4, 0xa0a0fb5b,
-  0x5252f6a4, 0x3b3b4d76, 0xd6d661b7, 0xb3b3ce7d,
-  0x29297b52, 0xe3e33edd, 0x2f2f715e, 0x84849713,
-  0x5353f5a6, 0xd1d168b9, 0x00000000, 0xeded2cc1,
-  0x20206040, 0xfcfc1fe3, 0xb1b1c879, 0x5b5bedb6,
-  0x6a6abed4, 0xcbcb468d, 0xbebed967, 0x39394b72,
-  0x4a4ade94, 0x4c4cd498, 0x5858e8b0, 0xcfcf4a85,
-  0xd0d06bbb, 0xefef2ac5, 0xaaaae54f, 0xfbfb16ed,
-  0x4343c586, 0x4d4dd79a, 0x33335566, 0x85859411,
-  0x4545cf8a, 0xf9f910e9, 0x02020604, 0x7f7f81fe,
-  0x5050f0a0, 0x3c3c4478, 0x9f9fba25, 0xa8a8e34b,
-  0x5151f3a2, 0xa3a3fe5d, 0x4040c080, 0x8f8f8a05,
-  0x9292ad3f, 0x9d9dbc21, 0x38384870, 0xf5f504f1,
-  0xbcbcdf63, 0xb6b6c177, 0xdada75af, 0x21216342,
-  0x10103020, 0xffff1ae5, 0xf3f30efd, 0xd2d26dbf,
-  0xcdcd4c81, 0x0c0c1418, 0x13133526, 0xecec2fc3,
-  0x5f5fe1be, 0x9797a235, 0x4444cc88, 0x1717392e,
-  0xc4c45793, 0xa7a7f255, 0x7e7e82fc, 0x3d3d477a,
-  0x6464acc8, 0x5d5de7ba, 0x19192b32, 0x737395e6,
-  0x6060a0c0, 0x81819819, 0x4f4fd19e, 0xdcdc7fa3,
-  0x22226644, 0x2a2a7e54, 0x9090ab3b, 0x8888830b,
-  0x4646ca8c, 0xeeee29c7, 0xb8b8d36b, 0x14143c28,
-  0xdede79a7, 0x5e5ee2bc, 0x0b0b1d16, 0xdbdb76ad,
-  0xe0e03bdb, 0x32325664, 0x3a3a4e74, 0x0a0a1e14,
-  0x4949db92, 0x06060a0c, 0x24246c48, 0x5c5ce4b8,
-  0xc2c25d9f, 0xd3d36ebd, 0xacacef43, 0x6262a6c4,
-  0x9191a839, 0x9595a431, 0xe4e437d3, 0x79798bf2,
-  0xe7e732d5, 0xc8c8438b, 0x3737596e, 0x6d6db7da,
-  0x8d8d8c01, 0xd5d564b1, 0x4e4ed29c, 0xa9a9e049,
-  0x6c6cb4d8, 0x5656faac, 0xf4f407f3, 0xeaea25cf,
-  0x6565afca, 0x7a7a8ef4, 0xaeaee947, 0x08081810,
-  0xbabad56f, 0x787888f0, 0x25256f4a, 0x2e2e725c,
-  0x1c1c2438, 0xa6a6f157, 0xb4b4c773, 0xc6c65197,
-  0xe8e823cb, 0xdddd7ca1, 0x74749ce8, 0x1f1f213e,
-  0x4b4bdd96, 0xbdbddc61, 0x8b8b860d, 0x8a8a850f,
-  0x707090e0, 0x3e3e427c, 0xb5b5c471, 0x6666aacc,
-  0x4848d890, 0x03030506, 0xf6f601f7, 0x0e0e121c,
-  0x6161a3c2, 0x35355f6a, 0x5757f9ae, 0xb9b9d069,
-  0x86869117, 0xc1c15899, 0x1d1d273a, 0x9e9eb927,
-  0xe1e138d9, 0xf8f813eb, 0x9898b32b, 0x11113322,
-  0x6969bbd2, 0xd9d970a9, 0x8e8e8907, 0x9494a733,
-  0x9b9bb62d, 0x1e1e223c, 0x87879215, 0xe9e920c9,
-  0xcece4987, 0x5555ffaa, 0x28287850, 0xdfdf7aa5,
-  0x8c8c8f03, 0xa1a1f859, 0x89898009, 0x0d0d171a,
-  0xbfbfda65, 0xe6e631d7, 0x4242c684, 0x6868b8d0,
-  0x4141c382, 0x9999b029, 0x2d2d775a, 0x0f0f111e,
-  0xb0b0cb7b, 0x5454fca8, 0xbbbbd66d, 0x16163a2c,
-};
-
-__constant u32 te4[256] =
-{
-  0x63636363, 0x7c7c7c7c, 0x77777777, 0x7b7b7b7b,
-  0xf2f2f2f2, 0x6b6b6b6b, 0x6f6f6f6f, 0xc5c5c5c5,
-  0x30303030, 0x01010101, 0x67676767, 0x2b2b2b2b,
-  0xfefefefe, 0xd7d7d7d7, 0xabababab, 0x76767676,
-  0xcacacaca, 0x82828282, 0xc9c9c9c9, 0x7d7d7d7d,
-  0xfafafafa, 0x59595959, 0x47474747, 0xf0f0f0f0,
-  0xadadadad, 0xd4d4d4d4, 0xa2a2a2a2, 0xafafafaf,
-  0x9c9c9c9c, 0xa4a4a4a4, 0x72727272, 0xc0c0c0c0,
-  0xb7b7b7b7, 0xfdfdfdfd, 0x93939393, 0x26262626,
-  0x36363636, 0x3f3f3f3f, 0xf7f7f7f7, 0xcccccccc,
-  0x34343434, 0xa5a5a5a5, 0xe5e5e5e5, 0xf1f1f1f1,
-  0x71717171, 0xd8d8d8d8, 0x31313131, 0x15151515,
-  0x04040404, 0xc7c7c7c7, 0x23232323, 0xc3c3c3c3,
-  0x18181818, 0x96969696, 0x05050505, 0x9a9a9a9a,
-  0x07070707, 0x12121212, 0x80808080, 0xe2e2e2e2,
-  0xebebebeb, 0x27272727, 0xb2b2b2b2, 0x75757575,
-  0x09090909, 0x83838383, 0x2c2c2c2c, 0x1a1a1a1a,
-  0x1b1b1b1b, 0x6e6e6e6e, 0x5a5a5a5a, 0xa0a0a0a0,
-  0x52525252, 0x3b3b3b3b, 0xd6d6d6d6, 0xb3b3b3b3,
-  0x29292929, 0xe3e3e3e3, 0x2f2f2f2f, 0x84848484,
-  0x53535353, 0xd1d1d1d1, 0x00000000, 0xedededed,
-  0x20202020, 0xfcfcfcfc, 0xb1b1b1b1, 0x5b5b5b5b,
-  0x6a6a6a6a, 0xcbcbcbcb, 0xbebebebe, 0x39393939,
-  0x4a4a4a4a, 0x4c4c4c4c, 0x58585858, 0xcfcfcfcf,
-  0xd0d0d0d0, 0xefefefef, 0xaaaaaaaa, 0xfbfbfbfb,
-  0x43434343, 0x4d4d4d4d, 0x33333333, 0x85858585,
-  0x45454545, 0xf9f9f9f9, 0x02020202, 0x7f7f7f7f,
-  0x50505050, 0x3c3c3c3c, 0x9f9f9f9f, 0xa8a8a8a8,
-  0x51515151, 0xa3a3a3a3, 0x40404040, 0x8f8f8f8f,
-  0x92929292, 0x9d9d9d9d, 0x38383838, 0xf5f5f5f5,
-  0xbcbcbcbc, 0xb6b6b6b6, 0xdadadada, 0x21212121,
-  0x10101010, 0xffffffff, 0xf3f3f3f3, 0xd2d2d2d2,
-  0xcdcdcdcd, 0x0c0c0c0c, 0x13131313, 0xecececec,
-  0x5f5f5f5f, 0x97979797, 0x44444444, 0x17171717,
-  0xc4c4c4c4, 0xa7a7a7a7, 0x7e7e7e7e, 0x3d3d3d3d,
-  0x64646464, 0x5d5d5d5d, 0x19191919, 0x73737373,
-  0x60606060, 0x81818181, 0x4f4f4f4f, 0xdcdcdcdc,
-  0x22222222, 0x2a2a2a2a, 0x90909090, 0x88888888,
-  0x46464646, 0xeeeeeeee, 0xb8b8b8b8, 0x14141414,
-  0xdededede, 0x5e5e5e5e, 0x0b0b0b0b, 0xdbdbdbdb,
-  0xe0e0e0e0, 0x32323232, 0x3a3a3a3a, 0x0a0a0a0a,
-  0x49494949, 0x06060606, 0x24242424, 0x5c5c5c5c,
-  0xc2c2c2c2, 0xd3d3d3d3, 0xacacacac, 0x62626262,
-  0x91919191, 0x95959595, 0xe4e4e4e4, 0x79797979,
-  0xe7e7e7e7, 0xc8c8c8c8, 0x37373737, 0x6d6d6d6d,
-  0x8d8d8d8d, 0xd5d5d5d5, 0x4e4e4e4e, 0xa9a9a9a9,
-  0x6c6c6c6c, 0x56565656, 0xf4f4f4f4, 0xeaeaeaea,
-  0x65656565, 0x7a7a7a7a, 0xaeaeaeae, 0x08080808,
-  0xbabababa, 0x78787878, 0x25252525, 0x2e2e2e2e,
-  0x1c1c1c1c, 0xa6a6a6a6, 0xb4b4b4b4, 0xc6c6c6c6,
-  0xe8e8e8e8, 0xdddddddd, 0x74747474, 0x1f1f1f1f,
-  0x4b4b4b4b, 0xbdbdbdbd, 0x8b8b8b8b, 0x8a8a8a8a,
-  0x70707070, 0x3e3e3e3e, 0xb5b5b5b5, 0x66666666,
-  0x48484848, 0x03030303, 0xf6f6f6f6, 0x0e0e0e0e,
-  0x61616161, 0x35353535, 0x57575757, 0xb9b9b9b9,
-  0x86868686, 0xc1c1c1c1, 0x1d1d1d1d, 0x9e9e9e9e,
-  0xe1e1e1e1, 0xf8f8f8f8, 0x98989898, 0x11111111,
-  0x69696969, 0xd9d9d9d9, 0x8e8e8e8e, 0x94949494,
-  0x9b9b9b9b, 0x1e1e1e1e, 0x87878787, 0xe9e9e9e9,
-  0xcececece, 0x55555555, 0x28282828, 0xdfdfdfdf,
-  0x8c8c8c8c, 0xa1a1a1a1, 0x89898989, 0x0d0d0d0d,
-  0xbfbfbfbf, 0xe6e6e6e6, 0x42424242, 0x68686868,
-  0x41414141, 0x99999999, 0x2d2d2d2d, 0x0f0f0f0f,
-  0xb0b0b0b0, 0x54545454, 0xbbbbbbbb, 0x16161616,
-};
-
-__constant u32 td0[256] =
-{
-  0x51f4a750, 0x7e416553, 0x1a17a4c3, 0x3a275e96,
-  0x3bab6bcb, 0x1f9d45f1, 0xacfa58ab, 0x4be30393,
-  0x2030fa55, 0xad766df6, 0x88cc7691, 0xf5024c25,
-  0x4fe5d7fc, 0xc52acbd7, 0x26354480, 0xb562a38f,
-  0xdeb15a49, 0x25ba1b67, 0x45ea0e98, 0x5dfec0e1,
-  0xc32f7502, 0x814cf012, 0x8d4697a3, 0x6bd3f9c6,
-  0x038f5fe7, 0x15929c95, 0xbf6d7aeb, 0x955259da,
-  0xd4be832d, 0x587421d3, 0x49e06929, 0x8ec9c844,
-  0x75c2896a, 0xf48e7978, 0x99583e6b, 0x27b971dd,
-  0xbee14fb6, 0xf088ad17, 0xc920ac66, 0x7dce3ab4,
-  0x63df4a18, 0xe51a3182, 0x97513360, 0x62537f45,
-  0xb16477e0, 0xbb6bae84, 0xfe81a01c, 0xf9082b94,
-  0x70486858, 0x8f45fd19, 0x94de6c87, 0x527bf8b7,
-  0xab73d323, 0x724b02e2, 0xe31f8f57, 0x6655ab2a,
-  0xb2eb2807, 0x2fb5c203, 0x86c57b9a, 0xd33708a5,
-  0x302887f2, 0x23bfa5b2, 0x02036aba, 0xed16825c,
-  0x8acf1c2b, 0xa779b492, 0xf307f2f0, 0x4e69e2a1,
-  0x65daf4cd, 0x0605bed5, 0xd134621f, 0xc4a6fe8a,
-  0x342e539d, 0xa2f355a0, 0x058ae132, 0xa4f6eb75,
-  0x0b83ec39, 0x4060efaa, 0x5e719f06, 0xbd6e1051,
-  0x3e218af9, 0x96dd063d, 0xdd3e05ae, 0x4de6bd46,
-  0x91548db5, 0x71c45d05, 0x0406d46f, 0x605015ff,
-  0x1998fb24, 0xd6bde997, 0x894043cc, 0x67d99e77,
-  0xb0e842bd, 0x07898b88, 0xe7195b38, 0x79c8eedb,
-  0xa17c0a47, 0x7c420fe9, 0xf8841ec9, 0x00000000,
-  0x09808683, 0x322bed48, 0x1e1170ac, 0x6c5a724e,
-  0xfd0efffb, 0x0f853856, 0x3daed51e, 0x362d3927,
-  0x0a0fd964, 0x685ca621, 0x9b5b54d1, 0x24362e3a,
-  0x0c0a67b1, 0x9357e70f, 0xb4ee96d2, 0x1b9b919e,
-  0x80c0c54f, 0x61dc20a2, 0x5a774b69, 0x1c121a16,
-  0xe293ba0a, 0xc0a02ae5, 0x3c22e043, 0x121b171d,
-  0x0e090d0b, 0xf28bc7ad, 0x2db6a8b9, 0x141ea9c8,
-  0x57f11985, 0xaf75074c, 0xee99ddbb, 0xa37f60fd,
-  0xf701269f, 0x5c72f5bc, 0x44663bc5, 0x5bfb7e34,
-  0x8b432976, 0xcb23c6dc, 0xb6edfc68, 0xb8e4f163,
-  0xd731dcca, 0x42638510, 0x13972240, 0x84c61120,
-  0x854a247d, 0xd2bb3df8, 0xaef93211, 0xc729a16d,
-  0x1d9e2f4b, 0xdcb230f3, 0x0d8652ec, 0x77c1e3d0,
-  0x2bb3166c, 0xa970b999, 0x119448fa, 0x47e96422,
-  0xa8fc8cc4, 0xa0f03f1a, 0x567d2cd8, 0x223390ef,
-  0x87494ec7, 0xd938d1c1, 0x8ccaa2fe, 0x98d40b36,
-  0xa6f581cf, 0xa57ade28, 0xdab78e26, 0x3fadbfa4,
-  0x2c3a9de4, 0x5078920d, 0x6a5fcc9b, 0x547e4662,
-  0xf68d13c2, 0x90d8b8e8, 0x2e39f75e, 0x82c3aff5,
-  0x9f5d80be, 0x69d0937c, 0x6fd52da9, 0xcf2512b3,
-  0xc8ac993b, 0x10187da7, 0xe89c636e, 0xdb3bbb7b,
-  0xcd267809, 0x6e5918f4, 0xec9ab701, 0x834f9aa8,
-  0xe6956e65, 0xaaffe67e, 0x21bccf08, 0xef15e8e6,
-  0xbae79bd9, 0x4a6f36ce, 0xea9f09d4, 0x29b07cd6,
-  0x31a4b2af, 0x2a3f2331, 0xc6a59430, 0x35a266c0,
-  0x744ebc37, 0xfc82caa6, 0xe090d0b0, 0x33a7d815,
-  0xf104984a, 0x41ecdaf7, 0x7fcd500e, 0x1791f62f,
-  0x764dd68d, 0x43efb04d, 0xccaa4d54, 0xe49604df,
-  0x9ed1b5e3, 0x4c6a881b, 0xc12c1fb8, 0x4665517f,
-  0x9d5eea04, 0x018c355d, 0xfa877473, 0xfb0b412e,
-  0xb3671d5a, 0x92dbd252, 0xe9105633, 0x6dd64713,
-  0x9ad7618c, 0x37a10c7a, 0x59f8148e, 0xeb133c89,
-  0xcea927ee, 0xb761c935, 0xe11ce5ed, 0x7a47b13c,
-  0x9cd2df59, 0x55f2733f, 0x1814ce79, 0x73c737bf,
-  0x53f7cdea, 0x5ffdaa5b, 0xdf3d6f14, 0x7844db86,
-  0xcaaff381, 0xb968c43e, 0x3824342c, 0xc2a3405f,
-  0x161dc372, 0xbce2250c, 0x283c498b, 0xff0d9541,
-  0x39a80171, 0x080cb3de, 0xd8b4e49c, 0x6456c190,
-  0x7bcb8461, 0xd532b670, 0x486c5c74, 0xd0b85742,
-};
-
-__constant u32 td1[256] =
-{
-  0x5051f4a7, 0x537e4165, 0xc31a17a4, 0x963a275e,
-  0xcb3bab6b, 0xf11f9d45, 0xabacfa58, 0x934be303,
-  0x552030fa, 0xf6ad766d, 0x9188cc76, 0x25f5024c,
-  0xfc4fe5d7, 0xd7c52acb, 0x80263544, 0x8fb562a3,
-  0x49deb15a, 0x6725ba1b, 0x9845ea0e, 0xe15dfec0,
-  0x02c32f75, 0x12814cf0, 0xa38d4697, 0xc66bd3f9,
-  0xe7038f5f, 0x9515929c, 0xebbf6d7a, 0xda955259,
-  0x2dd4be83, 0xd3587421, 0x2949e069, 0x448ec9c8,
-  0x6a75c289, 0x78f48e79, 0x6b99583e, 0xdd27b971,
-  0xb6bee14f, 0x17f088ad, 0x66c920ac, 0xb47dce3a,
-  0x1863df4a, 0x82e51a31, 0x60975133, 0x4562537f,
-  0xe0b16477, 0x84bb6bae, 0x1cfe81a0, 0x94f9082b,
-  0x58704868, 0x198f45fd, 0x8794de6c, 0xb7527bf8,
-  0x23ab73d3, 0xe2724b02, 0x57e31f8f, 0x2a6655ab,
-  0x07b2eb28, 0x032fb5c2, 0x9a86c57b, 0xa5d33708,
-  0xf2302887, 0xb223bfa5, 0xba02036a, 0x5ced1682,
-  0x2b8acf1c, 0x92a779b4, 0xf0f307f2, 0xa14e69e2,
-  0xcd65daf4, 0xd50605be, 0x1fd13462, 0x8ac4a6fe,
-  0x9d342e53, 0xa0a2f355, 0x32058ae1, 0x75a4f6eb,
-  0x390b83ec, 0xaa4060ef, 0x065e719f, 0x51bd6e10,
-  0xf93e218a, 0x3d96dd06, 0xaedd3e05, 0x464de6bd,
-  0xb591548d, 0x0571c45d, 0x6f0406d4, 0xff605015,
-  0x241998fb, 0x97d6bde9, 0xcc894043, 0x7767d99e,
-  0xbdb0e842, 0x8807898b, 0x38e7195b, 0xdb79c8ee,
-  0x47a17c0a, 0xe97c420f, 0xc9f8841e, 0x00000000,
-  0x83098086, 0x48322bed, 0xac1e1170, 0x4e6c5a72,
-  0xfbfd0eff, 0x560f8538, 0x1e3daed5, 0x27362d39,
-  0x640a0fd9, 0x21685ca6, 0xd19b5b54, 0x3a24362e,
-  0xb10c0a67, 0x0f9357e7, 0xd2b4ee96, 0x9e1b9b91,
-  0x4f80c0c5, 0xa261dc20, 0x695a774b, 0x161c121a,
-  0x0ae293ba, 0xe5c0a02a, 0x433c22e0, 0x1d121b17,
-  0x0b0e090d, 0xadf28bc7, 0xb92db6a8, 0xc8141ea9,
-  0x8557f119, 0x4caf7507, 0xbbee99dd, 0xfda37f60,
-  0x9ff70126, 0xbc5c72f5, 0xc544663b, 0x345bfb7e,
-  0x768b4329, 0xdccb23c6, 0x68b6edfc, 0x63b8e4f1,
-  0xcad731dc, 0x10426385, 0x40139722, 0x2084c611,
-  0x7d854a24, 0xf8d2bb3d, 0x11aef932, 0x6dc729a1,
-  0x4b1d9e2f, 0xf3dcb230, 0xec0d8652, 0xd077c1e3,
-  0x6c2bb316, 0x99a970b9, 0xfa119448, 0x2247e964,
-  0xc4a8fc8c, 0x1aa0f03f, 0xd8567d2c, 0xef223390,
-  0xc787494e, 0xc1d938d1, 0xfe8ccaa2, 0x3698d40b,
-  0xcfa6f581, 0x28a57ade, 0x26dab78e, 0xa43fadbf,
-  0xe42c3a9d, 0x0d507892, 0x9b6a5fcc, 0x62547e46,
-  0xc2f68d13, 0xe890d8b8, 0x5e2e39f7, 0xf582c3af,
-  0xbe9f5d80, 0x7c69d093, 0xa96fd52d, 0xb3cf2512,
-  0x3bc8ac99, 0xa710187d, 0x6ee89c63, 0x7bdb3bbb,
-  0x09cd2678, 0xf46e5918, 0x01ec9ab7, 0xa8834f9a,
-  0x65e6956e, 0x7eaaffe6, 0x0821bccf, 0xe6ef15e8,
-  0xd9bae79b, 0xce4a6f36, 0xd4ea9f09, 0xd629b07c,
-  0xaf31a4b2, 0x312a3f23, 0x30c6a594, 0xc035a266,
-  0x37744ebc, 0xa6fc82ca, 0xb0e090d0, 0x1533a7d8,
-  0x4af10498, 0xf741ecda, 0x0e7fcd50, 0x2f1791f6,
-  0x8d764dd6, 0x4d43efb0, 0x54ccaa4d, 0xdfe49604,
-  0xe39ed1b5, 0x1b4c6a88, 0xb8c12c1f, 0x7f466551,
-  0x049d5eea, 0x5d018c35, 0x73fa8774, 0x2efb0b41,
-  0x5ab3671d, 0x5292dbd2, 0x33e91056, 0x136dd647,
-  0x8c9ad761, 0x7a37a10c, 0x8e59f814, 0x89eb133c,
-  0xeecea927, 0x35b761c9, 0xede11ce5, 0x3c7a47b1,
-  0x599cd2df, 0x3f55f273, 0x791814ce, 0xbf73c737,
-  0xea53f7cd, 0x5b5ffdaa, 0x14df3d6f, 0x867844db,
-  0x81caaff3, 0x3eb968c4, 0x2c382434, 0x5fc2a340,
-  0x72161dc3, 0x0cbce225, 0x8b283c49, 0x41ff0d95,
-  0x7139a801, 0xde080cb3, 0x9cd8b4e4, 0x906456c1,
-  0x617bcb84, 0x70d532b6, 0x74486c5c, 0x42d0b857,
-};
-
-__constant u32 td2[256] =
-{
-  0xa75051f4, 0x65537e41, 0xa4c31a17, 0x5e963a27,
-  0x6bcb3bab, 0x45f11f9d, 0x58abacfa, 0x03934be3,
-  0xfa552030, 0x6df6ad76, 0x769188cc, 0x4c25f502,
-  0xd7fc4fe5, 0xcbd7c52a, 0x44802635, 0xa38fb562,
-  0x5a49deb1, 0x1b6725ba, 0x0e9845ea, 0xc0e15dfe,
-  0x7502c32f, 0xf012814c, 0x97a38d46, 0xf9c66bd3,
-  0x5fe7038f, 0x9c951592, 0x7aebbf6d, 0x59da9552,
-  0x832dd4be, 0x21d35874, 0x692949e0, 0xc8448ec9,
-  0x896a75c2, 0x7978f48e, 0x3e6b9958, 0x71dd27b9,
-  0x4fb6bee1, 0xad17f088, 0xac66c920, 0x3ab47dce,
-  0x4a1863df, 0x3182e51a, 0x33609751, 0x7f456253,
-  0x77e0b164, 0xae84bb6b, 0xa01cfe81, 0x2b94f908,
-  0x68587048, 0xfd198f45, 0x6c8794de, 0xf8b7527b,
-  0xd323ab73, 0x02e2724b, 0x8f57e31f, 0xab2a6655,
-  0x2807b2eb, 0xc2032fb5, 0x7b9a86c5, 0x08a5d337,
-  0x87f23028, 0xa5b223bf, 0x6aba0203, 0x825ced16,
-  0x1c2b8acf, 0xb492a779, 0xf2f0f307, 0xe2a14e69,
-  0xf4cd65da, 0xbed50605, 0x621fd134, 0xfe8ac4a6,
-  0x539d342e, 0x55a0a2f3, 0xe132058a, 0xeb75a4f6,
-  0xec390b83, 0xefaa4060, 0x9f065e71, 0x1051bd6e,
-  0x8af93e21, 0x063d96dd, 0x05aedd3e, 0xbd464de6,
-  0x8db59154, 0x5d0571c4, 0xd46f0406, 0x15ff6050,
-  0xfb241998, 0xe997d6bd, 0x43cc8940, 0x9e7767d9,
-  0x42bdb0e8, 0x8b880789, 0x5b38e719, 0xeedb79c8,
-  0x0a47a17c, 0x0fe97c42, 0x1ec9f884, 0x00000000,
-  0x86830980, 0xed48322b, 0x70ac1e11, 0x724e6c5a,
-  0xfffbfd0e, 0x38560f85, 0xd51e3dae, 0x3927362d,
-  0xd9640a0f, 0xa621685c, 0x54d19b5b, 0x2e3a2436,
-  0x67b10c0a, 0xe70f9357, 0x96d2b4ee, 0x919e1b9b,
-  0xc54f80c0, 0x20a261dc, 0x4b695a77, 0x1a161c12,
-  0xba0ae293, 0x2ae5c0a0, 0xe0433c22, 0x171d121b,
-  0x0d0b0e09, 0xc7adf28b, 0xa8b92db6, 0xa9c8141e,
-  0x198557f1, 0x074caf75, 0xddbbee99, 0x60fda37f,
-  0x269ff701, 0xf5bc5c72, 0x3bc54466, 0x7e345bfb,
-  0x29768b43, 0xc6dccb23, 0xfc68b6ed, 0xf163b8e4,
-  0xdccad731, 0x85104263, 0x22401397, 0x112084c6,
-  0x247d854a, 0x3df8d2bb, 0x3211aef9, 0xa16dc729,
-  0x2f4b1d9e, 0x30f3dcb2, 0x52ec0d86, 0xe3d077c1,
-  0x166c2bb3, 0xb999a970, 0x48fa1194, 0x642247e9,
-  0x8cc4a8fc, 0x3f1aa0f0, 0x2cd8567d, 0x90ef2233,
-  0x4ec78749, 0xd1c1d938, 0xa2fe8cca, 0x0b3698d4,
-  0x81cfa6f5, 0xde28a57a, 0x8e26dab7, 0xbfa43fad,
-  0x9de42c3a, 0x920d5078, 0xcc9b6a5f, 0x4662547e,
-  0x13c2f68d, 0xb8e890d8, 0xf75e2e39, 0xaff582c3,
-  0x80be9f5d, 0x937c69d0, 0x2da96fd5, 0x12b3cf25,
-  0x993bc8ac, 0x7da71018, 0x636ee89c, 0xbb7bdb3b,
-  0x7809cd26, 0x18f46e59, 0xb701ec9a, 0x9aa8834f,
-  0x6e65e695, 0xe67eaaff, 0xcf0821bc, 0xe8e6ef15,
-  0x9bd9bae7, 0x36ce4a6f, 0x09d4ea9f, 0x7cd629b0,
-  0xb2af31a4, 0x23312a3f, 0x9430c6a5, 0x66c035a2,
-  0xbc37744e, 0xcaa6fc82, 0xd0b0e090, 0xd81533a7,
-  0x984af104, 0xdaf741ec, 0x500e7fcd, 0xf62f1791,
-  0xd68d764d, 0xb04d43ef, 0x4d54ccaa, 0x04dfe496,
-  0xb5e39ed1, 0x881b4c6a, 0x1fb8c12c, 0x517f4665,
-  0xea049d5e, 0x355d018c, 0x7473fa87, 0x412efb0b,
-  0x1d5ab367, 0xd25292db, 0x5633e910, 0x47136dd6,
-  0x618c9ad7, 0x0c7a37a1, 0x148e59f8, 0x3c89eb13,
-  0x27eecea9, 0xc935b761, 0xe5ede11c, 0xb13c7a47,
-  0xdf599cd2, 0x733f55f2, 0xce791814, 0x37bf73c7,
-  0xcdea53f7, 0xaa5b5ffd, 0x6f14df3d, 0xdb867844,
-  0xf381caaf, 0xc43eb968, 0x342c3824, 0x405fc2a3,
-  0xc372161d, 0x250cbce2, 0x498b283c, 0x9541ff0d,
-  0x017139a8, 0xb3de080c, 0xe49cd8b4, 0xc1906456,
-  0x84617bcb, 0xb670d532, 0x5c74486c, 0x5742d0b8,
-};
-
-__constant u32 td3[256] =
-{
-  0xf4a75051, 0x4165537e, 0x17a4c31a, 0x275e963a,
-  0xab6bcb3b, 0x9d45f11f, 0xfa58abac, 0xe303934b,
-  0x30fa5520, 0x766df6ad, 0xcc769188, 0x024c25f5,
-  0xe5d7fc4f, 0x2acbd7c5, 0x35448026, 0x62a38fb5,
-  0xb15a49de, 0xba1b6725, 0xea0e9845, 0xfec0e15d,
-  0x2f7502c3, 0x4cf01281, 0x4697a38d, 0xd3f9c66b,
-  0x8f5fe703, 0x929c9515, 0x6d7aebbf, 0x5259da95,
-  0xbe832dd4, 0x7421d358, 0xe0692949, 0xc9c8448e,
-  0xc2896a75, 0x8e7978f4, 0x583e6b99, 0xb971dd27,
-  0xe14fb6be, 0x88ad17f0, 0x20ac66c9, 0xce3ab47d,
-  0xdf4a1863, 0x1a3182e5, 0x51336097, 0x537f4562,
-  0x6477e0b1, 0x6bae84bb, 0x81a01cfe, 0x082b94f9,
-  0x48685870, 0x45fd198f, 0xde6c8794, 0x7bf8b752,
-  0x73d323ab, 0x4b02e272, 0x1f8f57e3, 0x55ab2a66,
-  0xeb2807b2, 0xb5c2032f, 0xc57b9a86, 0x3708a5d3,
-  0x2887f230, 0xbfa5b223, 0x036aba02, 0x16825ced,
-  0xcf1c2b8a, 0x79b492a7, 0x07f2f0f3, 0x69e2a14e,
-  0xdaf4cd65, 0x05bed506, 0x34621fd1, 0xa6fe8ac4,
-  0x2e539d34, 0xf355a0a2, 0x8ae13205, 0xf6eb75a4,
-  0x83ec390b, 0x60efaa40, 0x719f065e, 0x6e1051bd,
-  0x218af93e, 0xdd063d96, 0x3e05aedd, 0xe6bd464d,
-  0x548db591, 0xc45d0571, 0x06d46f04, 0x5015ff60,
-  0x98fb2419, 0xbde997d6, 0x4043cc89, 0xd99e7767,
-  0xe842bdb0, 0x898b8807, 0x195b38e7, 0xc8eedb79,
-  0x7c0a47a1, 0x420fe97c, 0x841ec9f8, 0x00000000,
-  0x80868309, 0x2bed4832, 0x1170ac1e, 0x5a724e6c,
-  0x0efffbfd, 0x8538560f, 0xaed51e3d, 0x2d392736,
-  0x0fd9640a, 0x5ca62168, 0x5b54d19b, 0x362e3a24,
-  0x0a67b10c, 0x57e70f93, 0xee96d2b4, 0x9b919e1b,
-  0xc0c54f80, 0xdc20a261, 0x774b695a, 0x121a161c,
-  0x93ba0ae2, 0xa02ae5c0, 0x22e0433c, 0x1b171d12,
-  0x090d0b0e, 0x8bc7adf2, 0xb6a8b92d, 0x1ea9c814,
-  0xf1198557, 0x75074caf, 0x99ddbbee, 0x7f60fda3,
-  0x01269ff7, 0x72f5bc5c, 0x663bc544, 0xfb7e345b,
-  0x4329768b, 0x23c6dccb, 0xedfc68b6, 0xe4f163b8,
-  0x31dccad7, 0x63851042, 0x97224013, 0xc6112084,
-  0x4a247d85, 0xbb3df8d2, 0xf93211ae, 0x29a16dc7,
-  0x9e2f4b1d, 0xb230f3dc, 0x8652ec0d, 0xc1e3d077,
-  0xb3166c2b, 0x70b999a9, 0x9448fa11, 0xe9642247,
-  0xfc8cc4a8, 0xf03f1aa0, 0x7d2cd856, 0x3390ef22,
-  0x494ec787, 0x38d1c1d9, 0xcaa2fe8c, 0xd40b3698,
-  0xf581cfa6, 0x7ade28a5, 0xb78e26da, 0xadbfa43f,
-  0x3a9de42c, 0x78920d50, 0x5fcc9b6a, 0x7e466254,
-  0x8d13c2f6, 0xd8b8e890, 0x39f75e2e, 0xc3aff582,
-  0x5d80be9f, 0xd0937c69, 0xd52da96f, 0x2512b3cf,
-  0xac993bc8, 0x187da710, 0x9c636ee8, 0x3bbb7bdb,
-  0x267809cd, 0x5918f46e, 0x9ab701ec, 0x4f9aa883,
-  0x956e65e6, 0xffe67eaa, 0xbccf0821, 0x15e8e6ef,
-  0xe79bd9ba, 0x6f36ce4a, 0x9f09d4ea, 0xb07cd629,
-  0xa4b2af31, 0x3f23312a, 0xa59430c6, 0xa266c035,
-  0x4ebc3774, 0x82caa6fc, 0x90d0b0e0, 0xa7d81533,
-  0x04984af1, 0xecdaf741, 0xcd500e7f, 0x91f62f17,
-  0x4dd68d76, 0xefb04d43, 0xaa4d54cc, 0x9604dfe4,
-  0xd1b5e39e, 0x6a881b4c, 0x2c1fb8c1, 0x65517f46,
-  0x5eea049d, 0x8c355d01, 0x877473fa, 0x0b412efb,
-  0x671d5ab3, 0xdbd25292, 0x105633e9, 0xd647136d,
-  0xd7618c9a, 0xa10c7a37, 0xf8148e59, 0x133c89eb,
-  0xa927eece, 0x61c935b7, 0x1ce5ede1, 0x47b13c7a,
-  0xd2df599c, 0xf2733f55, 0x14ce7918, 0xc737bf73,
-  0xf7cdea53, 0xfdaa5b5f, 0x3d6f14df, 0x44db8678,
-  0xaff381ca, 0x68c43eb9, 0x24342c38, 0xa3405fc2,
-  0x1dc37216, 0xe2250cbc, 0x3c498b28, 0x0d9541ff,
-  0xa8017139, 0x0cb3de08, 0xb4e49cd8, 0x56c19064,
-  0xcb84617b, 0x32b670d5, 0x6c5c7448, 0xb85742d0,
-};
-
-__constant u32 td4[256] =
-{
-  0x52525252, 0x09090909, 0x6a6a6a6a, 0xd5d5d5d5,
-  0x30303030, 0x36363636, 0xa5a5a5a5, 0x38383838,
-  0xbfbfbfbf, 0x40404040, 0xa3a3a3a3, 0x9e9e9e9e,
-  0x81818181, 0xf3f3f3f3, 0xd7d7d7d7, 0xfbfbfbfb,
-  0x7c7c7c7c, 0xe3e3e3e3, 0x39393939, 0x82828282,
-  0x9b9b9b9b, 0x2f2f2f2f, 0xffffffff, 0x87878787,
-  0x34343434, 0x8e8e8e8e, 0x43434343, 0x44444444,
-  0xc4c4c4c4, 0xdededede, 0xe9e9e9e9, 0xcbcbcbcb,
-  0x54545454, 0x7b7b7b7b, 0x94949494, 0x32323232,
-  0xa6a6a6a6, 0xc2c2c2c2, 0x23232323, 0x3d3d3d3d,
-  0xeeeeeeee, 0x4c4c4c4c, 0x95959595, 0x0b0b0b0b,
-  0x42424242, 0xfafafafa, 0xc3c3c3c3, 0x4e4e4e4e,
-  0x08080808, 0x2e2e2e2e, 0xa1a1a1a1, 0x66666666,
-  0x28282828, 0xd9d9d9d9, 0x24242424, 0xb2b2b2b2,
-  0x76767676, 0x5b5b5b5b, 0xa2a2a2a2, 0x49494949,
-  0x6d6d6d6d, 0x8b8b8b8b, 0xd1d1d1d1, 0x25252525,
-  0x72727272, 0xf8f8f8f8, 0xf6f6f6f6, 0x64646464,
-  0x86868686, 0x68686868, 0x98989898, 0x16161616,
-  0xd4d4d4d4, 0xa4a4a4a4, 0x5c5c5c5c, 0xcccccccc,
-  0x5d5d5d5d, 0x65656565, 0xb6b6b6b6, 0x92929292,
-  0x6c6c6c6c, 0x70707070, 0x48484848, 0x50505050,
-  0xfdfdfdfd, 0xedededed, 0xb9b9b9b9, 0xdadadada,
-  0x5e5e5e5e, 0x15151515, 0x46464646, 0x57575757,
-  0xa7a7a7a7, 0x8d8d8d8d, 0x9d9d9d9d, 0x84848484,
-  0x90909090, 0xd8d8d8d8, 0xabababab, 0x00000000,
-  0x8c8c8c8c, 0xbcbcbcbc, 0xd3d3d3d3, 0x0a0a0a0a,
-  0xf7f7f7f7, 0xe4e4e4e4, 0x58585858, 0x05050505,
-  0xb8b8b8b8, 0xb3b3b3b3, 0x45454545, 0x06060606,
-  0xd0d0d0d0, 0x2c2c2c2c, 0x1e1e1e1e, 0x8f8f8f8f,
-  0xcacacaca, 0x3f3f3f3f, 0x0f0f0f0f, 0x02020202,
-  0xc1c1c1c1, 0xafafafaf, 0xbdbdbdbd, 0x03030303,
-  0x01010101, 0x13131313, 0x8a8a8a8a, 0x6b6b6b6b,
-  0x3a3a3a3a, 0x91919191, 0x11111111, 0x41414141,
-  0x4f4f4f4f, 0x67676767, 0xdcdcdcdc, 0xeaeaeaea,
-  0x97979797, 0xf2f2f2f2, 0xcfcfcfcf, 0xcececece,
-  0xf0f0f0f0, 0xb4b4b4b4, 0xe6e6e6e6, 0x73737373,
-  0x96969696, 0xacacacac, 0x74747474, 0x22222222,
-  0xe7e7e7e7, 0xadadadad, 0x35353535, 0x85858585,
-  0xe2e2e2e2, 0xf9f9f9f9, 0x37373737, 0xe8e8e8e8,
-  0x1c1c1c1c, 0x75757575, 0xdfdfdfdf, 0x6e6e6e6e,
-  0x47474747, 0xf1f1f1f1, 0x1a1a1a1a, 0x71717171,
-  0x1d1d1d1d, 0x29292929, 0xc5c5c5c5, 0x89898989,
-  0x6f6f6f6f, 0xb7b7b7b7, 0x62626262, 0x0e0e0e0e,
-  0xaaaaaaaa, 0x18181818, 0xbebebebe, 0x1b1b1b1b,
-  0xfcfcfcfc, 0x56565656, 0x3e3e3e3e, 0x4b4b4b4b,
-  0xc6c6c6c6, 0xd2d2d2d2, 0x79797979, 0x20202020,
-  0x9a9a9a9a, 0xdbdbdbdb, 0xc0c0c0c0, 0xfefefefe,
-  0x78787878, 0xcdcdcdcd, 0x5a5a5a5a, 0xf4f4f4f4,
-  0x1f1f1f1f, 0xdddddddd, 0xa8a8a8a8, 0x33333333,
-  0x88888888, 0x07070707, 0xc7c7c7c7, 0x31313131,
-  0xb1b1b1b1, 0x12121212, 0x10101010, 0x59595959,
-  0x27272727, 0x80808080, 0xecececec, 0x5f5f5f5f,
-  0x60606060, 0x51515151, 0x7f7f7f7f, 0xa9a9a9a9,
-  0x19191919, 0xb5b5b5b5, 0x4a4a4a4a, 0x0d0d0d0d,
-  0x2d2d2d2d, 0xe5e5e5e5, 0x7a7a7a7a, 0x9f9f9f9f,
-  0x93939393, 0xc9c9c9c9, 0x9c9c9c9c, 0xefefefef,
-  0xa0a0a0a0, 0xe0e0e0e0, 0x3b3b3b3b, 0x4d4d4d4d,
-  0xaeaeaeae, 0x2a2a2a2a, 0xf5f5f5f5, 0xb0b0b0b0,
-  0xc8c8c8c8, 0xebebebeb, 0xbbbbbbbb, 0x3c3c3c3c,
-  0x83838383, 0x53535353, 0x99999999, 0x61616161,
-  0x17171717, 0x2b2b2b2b, 0x04040404, 0x7e7e7e7e,
-  0xbabababa, 0x77777777, 0xd6d6d6d6, 0x26262626,
-  0xe1e1e1e1, 0x69696969, 0x14141414, 0x63636363,
-  0x55555555, 0x21212121, 0x0c0c0c0c, 0x7d7d7d7d,
-};
-
-__constant u32 rcon[] =
-{
-  0x01000000, 0x02000000, 0x04000000, 0x08000000,
-  0x10000000, 0x20000000, 0x40000000, 0x80000000,
-  0x1b000000, 0x36000000,
-};
-
-static void aes256_ExpandKey (u32 *ks, const u32 *ukey)
-{
-  ks[0] = ukey[0];
-  ks[1] = ukey[1];
-  ks[2] = ukey[2];
-  ks[3] = ukey[3];
-  ks[4] = ukey[4];
-  ks[5] = ukey[5];
-  ks[6] = ukey[6];
-  ks[7] = ukey[7];
-
-  int i;
-  int j;
-
-  i = 0;
-  j = 0;
-
-  while (1)
-  {
-    u32 temp = ks[j +  7];
-
-    ks[j +  8] = ks[j +  0]
-           ^ (te2[(temp >> 16) & 0xff] & 0xff000000)
-           ^ (te3[(temp >>  8) & 0xff] & 0x00ff0000)
-           ^ (te0[(temp >>  0) & 0xff] & 0x0000ff00)
-           ^ (te1[(temp >> 24) & 0xff] & 0x000000ff)
-           ^ rcon[i];
-
-    ks[j +  9] = ks[j +  1] ^ ks[j +  8];
-    ks[j + 10] = ks[j +  2] ^ ks[j +  9];
-    ks[j + 11] = ks[j +  3] ^ ks[j + 10];
-
-    if (++i == 7) break;
-
-    temp = ks[j + 11];
-
-    ks[j + 12] = ks[j +  4]
-           ^ (te2[(temp >> 24) & 0xff] & 0xff000000)
-           ^ (te3[(temp >> 16) & 0xff] & 0x00ff0000)
-           ^ (te0[(temp >>  8) & 0xff] & 0x0000ff00)
-           ^ (te1[(temp >>  0) & 0xff] & 0x000000ff);
-
-    ks[j + 13] = ks[j +  5] ^ ks[j + 12];
-    ks[j + 14] = ks[j +  6] ^ ks[j + 13];
-    ks[j + 15] = ks[j +  7] ^ ks[j + 14];
-
-    j += 8;
-  }
-}
-
-static void aes256_InvertKey (u32 *ks)
-{
-  for (u32 i = 0, j = 56; i < j; i += 4, j -= 4)
-  {
-    u32 temp;
-
-    temp = ks[i + 0]; ks[i + 0] = ks[j + 0]; ks[j + 0] = temp;
-    temp = ks[i + 1]; ks[i + 1] = ks[j + 1]; ks[j + 1] = temp;
-    temp = ks[i + 2]; ks[i + 2] = ks[j + 2]; ks[j + 2] = temp;
-    temp = ks[i + 3]; ks[i + 3] = ks[j + 3]; ks[j + 3] = temp;
-  }
-
-  for (u32 i = 1, j = 4; i < 14; i += 1, j += 4)
-  {
-    ks[j + 0] =
-      td0[te1[(ks[j + 0] >> 24) & 0xff] & 0xff] ^
-      td1[te1[(ks[j + 0] >> 16) & 0xff] & 0xff] ^
-      td2[te1[(ks[j + 0] >>  8) & 0xff] & 0xff] ^
-      td3[te1[(ks[j + 0] >>  0) & 0xff] & 0xff];
-
-    ks[j + 1] =
-      td0[te1[(ks[j + 1] >> 24) & 0xff] & 0xff] ^
-      td1[te1[(ks[j + 1] >> 16) & 0xff] & 0xff] ^
-      td2[te1[(ks[j + 1] >>  8) & 0xff] & 0xff] ^
-      td3[te1[(ks[j + 1] >>  0) & 0xff] & 0xff];
-
-    ks[j + 2] =
-      td0[te1[(ks[j + 2] >> 24) & 0xff] & 0xff] ^
-      td1[te1[(ks[j + 2] >> 16) & 0xff] & 0xff] ^
-      td2[te1[(ks[j + 2] >>  8) & 0xff] & 0xff] ^
-      td3[te1[(ks[j + 2] >>  0) & 0xff] & 0xff];
-
-    ks[j + 3] =
-      td0[te1[(ks[j + 3] >> 24) & 0xff] & 0xff] ^
-      td1[te1[(ks[j + 3] >> 16) & 0xff] & 0xff] ^
-      td2[te1[(ks[j + 3] >>  8) & 0xff] & 0xff] ^
-      td3[te1[(ks[j + 3] >>  0) & 0xff] & 0xff];
-  }
-}
-
-static void aes256_set_encrypt_key (u32 *ks, const u32 *ukey)
-{
-  u32 ukey_s[8];
-
-  ukey_s[0] = swap32 (ukey[0]);
-  ukey_s[1] = swap32 (ukey[1]);
-  ukey_s[2] = swap32 (ukey[2]);
-  ukey_s[3] = swap32 (ukey[3]);
-  ukey_s[4] = swap32 (ukey[4]);
-  ukey_s[5] = swap32 (ukey[5]);
-  ukey_s[6] = swap32 (ukey[6]);
-  ukey_s[7] = swap32 (ukey[7]);
-
-  aes256_ExpandKey (ks, ukey_s);
-}
-
-static void aes256_set_decrypt_key (u32 *ks, const u32 *ukey)
-{
-  u32 ukey_s[8];
-
-  ukey_s[0] = swap32 (ukey[0]);
-  ukey_s[1] = swap32 (ukey[1]);
-  ukey_s[2] = swap32 (ukey[2]);
-  ukey_s[3] = swap32 (ukey[3]);
-  ukey_s[4] = swap32 (ukey[4]);
-  ukey_s[5] = swap32 (ukey[5]);
-  ukey_s[6] = swap32 (ukey[6]);
-  ukey_s[7] = swap32 (ukey[7]);
-
-  aes256_ExpandKey (ks, ukey_s);
-
-  aes256_InvertKey (ks);
-}
-
-static void aes256_decrypt (const u32 *ks, const u32 *in, u32 *out)
-{
-  u32 in_s[4];
-
-  in_s[0] = swap32 (in[0]);
-  in_s[1] = swap32 (in[1]);
-  in_s[2] = swap32 (in[2]);
-  in_s[3] = swap32 (in[3]);
-
-  u32 s0 = in_s[0] ^ ks[0];
-  u32 s1 = in_s[1] ^ ks[1];
-  u32 s2 = in_s[2] ^ ks[2];
-  u32 s3 = in_s[3] ^ ks[3];
-
-  u32 t0;
-  u32 t1;
-  u32 t2;
-  u32 t3;
-
-  t0 = td0[s0 >> 24] ^ td1[(s3 >> 16) & 0xff] ^ td2[(s2 >>  8) & 0xff] ^ td3[s1 & 0xff] ^ ks[ 4];
-  t1 = td0[s1 >> 24] ^ td1[(s0 >> 16) & 0xff] ^ td2[(s3 >>  8) & 0xff] ^ td3[s2 & 0xff] ^ ks[ 5];
-  t2 = td0[s2 >> 24] ^ td1[(s1 >> 16) & 0xff] ^ td2[(s0 >>  8) & 0xff] ^ td3[s3 & 0xff] ^ ks[ 6];
-  t3 = td0[s3 >> 24] ^ td1[(s2 >> 16) & 0xff] ^ td2[(s1 >>  8) & 0xff] ^ td3[s0 & 0xff] ^ ks[ 7];
-  s0 = td0[t0 >> 24] ^ td1[(t3 >> 16) & 0xff] ^ td2[(t2 >>  8) & 0xff] ^ td3[t1 & 0xff] ^ ks[ 8];
-  s1 = td0[t1 >> 24] ^ td1[(t0 >> 16) & 0xff] ^ td2[(t3 >>  8) & 0xff] ^ td3[t2 & 0xff] ^ ks[ 9];
-  s2 = td0[t2 >> 24] ^ td1[(t1 >> 16) & 0xff] ^ td2[(t0 >>  8) & 0xff] ^ td3[t3 & 0xff] ^ ks[10];
-  s3 = td0[t3 >> 24] ^ td1[(t2 >> 16) & 0xff] ^ td2[(t1 >>  8) & 0xff] ^ td3[t0 & 0xff] ^ ks[11];
-  t0 = td0[s0 >> 24] ^ td1[(s3 >> 16) & 0xff] ^ td2[(s2 >>  8) & 0xff] ^ td3[s1 & 0xff] ^ ks[12];
-  t1 = td0[s1 >> 24] ^ td1[(s0 >> 16) & 0xff] ^ td2[(s3 >>  8) & 0xff] ^ td3[s2 & 0xff] ^ ks[13];
-  t2 = td0[s2 >> 24] ^ td1[(s1 >> 16) & 0xff] ^ td2[(s0 >>  8) & 0xff] ^ td3[s3 & 0xff] ^ ks[14];
-  t3 = td0[s3 >> 24] ^ td1[(s2 >> 16) & 0xff] ^ td2[(s1 >>  8) & 0xff] ^ td3[s0 & 0xff] ^ ks[15];
-  s0 = td0[t0 >> 24] ^ td1[(t3 >> 16) & 0xff] ^ td2[(t2 >>  8) & 0xff] ^ td3[t1 & 0xff] ^ ks[16];
-  s1 = td0[t1 >> 24] ^ td1[(t0 >> 16) & 0xff] ^ td2[(t3 >>  8) & 0xff] ^ td3[t2 & 0xff] ^ ks[17];
-  s2 = td0[t2 >> 24] ^ td1[(t1 >> 16) & 0xff] ^ td2[(t0 >>  8) & 0xff] ^ td3[t3 & 0xff] ^ ks[18];
-  s3 = td0[t3 >> 24] ^ td1[(t2 >> 16) & 0xff] ^ td2[(t1 >>  8) & 0xff] ^ td3[t0 & 0xff] ^ ks[19];
-  t0 = td0[s0 >> 24] ^ td1[(s3 >> 16) & 0xff] ^ td2[(s2 >>  8) & 0xff] ^ td3[s1 & 0xff] ^ ks[20];
-  t1 = td0[s1 >> 24] ^ td1[(s0 >> 16) & 0xff] ^ td2[(s3 >>  8) & 0xff] ^ td3[s2 & 0xff] ^ ks[21];
-  t2 = td0[s2 >> 24] ^ td1[(s1 >> 16) & 0xff] ^ td2[(s0 >>  8) & 0xff] ^ td3[s3 & 0xff] ^ ks[22];
-  t3 = td0[s3 >> 24] ^ td1[(s2 >> 16) & 0xff] ^ td2[(s1 >>  8) & 0xff] ^ td3[s0 & 0xff] ^ ks[23];
-  s0 = td0[t0 >> 24] ^ td1[(t3 >> 16) & 0xff] ^ td2[(t2 >>  8) & 0xff] ^ td3[t1 & 0xff] ^ ks[24];
-  s1 = td0[t1 >> 24] ^ td1[(t0 >> 16) & 0xff] ^ td2[(t3 >>  8) & 0xff] ^ td3[t2 & 0xff] ^ ks[25];
-  s2 = td0[t2 >> 24] ^ td1[(t1 >> 16) & 0xff] ^ td2[(t0 >>  8) & 0xff] ^ td3[t3 & 0xff] ^ ks[26];
-  s3 = td0[t3 >> 24] ^ td1[(t2 >> 16) & 0xff] ^ td2[(t1 >>  8) & 0xff] ^ td3[t0 & 0xff] ^ ks[27];
-  t0 = td0[s0 >> 24] ^ td1[(s3 >> 16) & 0xff] ^ td2[(s2 >>  8) & 0xff] ^ td3[s1 & 0xff] ^ ks[28];
-  t1 = td0[s1 >> 24] ^ td1[(s0 >> 16) & 0xff] ^ td2[(s3 >>  8) & 0xff] ^ td3[s2 & 0xff] ^ ks[29];
-  t2 = td0[s2 >> 24] ^ td1[(s1 >> 16) & 0xff] ^ td2[(s0 >>  8) & 0xff] ^ td3[s3 & 0xff] ^ ks[30];
-  t3 = td0[s3 >> 24] ^ td1[(s2 >> 16) & 0xff] ^ td2[(s1 >>  8) & 0xff] ^ td3[s0 & 0xff] ^ ks[31];
-  s0 = td0[t0 >> 24] ^ td1[(t3 >> 16) & 0xff] ^ td2[(t2 >>  8) & 0xff] ^ td3[t1 & 0xff] ^ ks[32];
-  s1 = td0[t1 >> 24] ^ td1[(t0 >> 16) & 0xff] ^ td2[(t3 >>  8) & 0xff] ^ td3[t2 & 0xff] ^ ks[33];
-  s2 = td0[t2 >> 24] ^ td1[(t1 >> 16) & 0xff] ^ td2[(t0 >>  8) & 0xff] ^ td3[t3 & 0xff] ^ ks[34];
-  s3 = td0[t3 >> 24] ^ td1[(t2 >> 16) & 0xff] ^ td2[(t1 >>  8) & 0xff] ^ td3[t0 & 0xff] ^ ks[35];
-  t0 = td0[s0 >> 24] ^ td1[(s3 >> 16) & 0xff] ^ td2[(s2 >>  8) & 0xff] ^ td3[s1 & 0xff] ^ ks[36];
-  t1 = td0[s1 >> 24] ^ td1[(s0 >> 16) & 0xff] ^ td2[(s3 >>  8) & 0xff] ^ td3[s2 & 0xff] ^ ks[37];
-  t2 = td0[s2 >> 24] ^ td1[(s1 >> 16) & 0xff] ^ td2[(s0 >>  8) & 0xff] ^ td3[s3 & 0xff] ^ ks[38];
-  t3 = td0[s3 >> 24] ^ td1[(s2 >> 16) & 0xff] ^ td2[(s1 >>  8) & 0xff] ^ td3[s0 & 0xff] ^ ks[39];
-  s0 = td0[t0 >> 24] ^ td1[(t3 >> 16) & 0xff] ^ td2[(t2 >>  8) & 0xff] ^ td3[t1 & 0xff] ^ ks[40];
-  s1 = td0[t1 >> 24] ^ td1[(t0 >> 16) & 0xff] ^ td2[(t3 >>  8) & 0xff] ^ td3[t2 & 0xff] ^ ks[41];
-  s2 = td0[t2 >> 24] ^ td1[(t1 >> 16) & 0xff] ^ td2[(t0 >>  8) & 0xff] ^ td3[t3 & 0xff] ^ ks[42];
-  s3 = td0[t3 >> 24] ^ td1[(t2 >> 16) & 0xff] ^ td2[(t1 >>  8) & 0xff] ^ td3[t0 & 0xff] ^ ks[43];
-  t0 = td0[s0 >> 24] ^ td1[(s3 >> 16) & 0xff] ^ td2[(s2 >>  8) & 0xff] ^ td3[s1 & 0xff] ^ ks[44];
-  t1 = td0[s1 >> 24] ^ td1[(s0 >> 16) & 0xff] ^ td2[(s3 >>  8) & 0xff] ^ td3[s2 & 0xff] ^ ks[45];
-  t2 = td0[s2 >> 24] ^ td1[(s1 >> 16) & 0xff] ^ td2[(s0 >>  8) & 0xff] ^ td3[s3 & 0xff] ^ ks[46];
-  t3 = td0[s3 >> 24] ^ td1[(s2 >> 16) & 0xff] ^ td2[(s1 >>  8) & 0xff] ^ td3[s0 & 0xff] ^ ks[47];
-  s0 = td0[t0 >> 24] ^ td1[(t3 >> 16) & 0xff] ^ td2[(t2 >>  8) & 0xff] ^ td3[t1 & 0xff] ^ ks[48];
-  s1 = td0[t1 >> 24] ^ td1[(t0 >> 16) & 0xff] ^ td2[(t3 >>  8) & 0xff] ^ td3[t2 & 0xff] ^ ks[49];
-  s2 = td0[t2 >> 24] ^ td1[(t1 >> 16) & 0xff] ^ td2[(t0 >>  8) & 0xff] ^ td3[t3 & 0xff] ^ ks[50];
-  s3 = td0[t3 >> 24] ^ td1[(t2 >> 16) & 0xff] ^ td2[(t1 >>  8) & 0xff] ^ td3[t0 & 0xff] ^ ks[51];
-  t0 = td0[s0 >> 24] ^ td1[(s3 >> 16) & 0xff] ^ td2[(s2 >>  8) & 0xff] ^ td3[s1 & 0xff] ^ ks[52];
-  t1 = td0[s1 >> 24] ^ td1[(s0 >> 16) & 0xff] ^ td2[(s3 >>  8) & 0xff] ^ td3[s2 & 0xff] ^ ks[53];
-  t2 = td0[s2 >> 24] ^ td1[(s1 >> 16) & 0xff] ^ td2[(s0 >>  8) & 0xff] ^ td3[s3 & 0xff] ^ ks[54];
-  t3 = td0[s3 >> 24] ^ td1[(s2 >> 16) & 0xff] ^ td2[(s1 >>  8) & 0xff] ^ td3[s0 & 0xff] ^ ks[55];
-
-  out[0] = (td4[(t0 >> 24) & 0xff] & 0xff000000)
-         ^ (td4[(t3 >> 16) & 0xff] & 0x00ff0000)
-         ^ (td4[(t2 >>  8) & 0xff] & 0x0000ff00)
-         ^ (td4[(t1 >>  0) & 0xff] & 0x000000ff)
-         ^ ks[56];
-
-  out[1] = (td4[(t1 >> 24) & 0xff] & 0xff000000)
-         ^ (td4[(t0 >> 16) & 0xff] & 0x00ff0000)
-         ^ (td4[(t3 >>  8) & 0xff] & 0x0000ff00)
-         ^ (td4[(t2 >>  0) & 0xff] & 0x000000ff)
-         ^ ks[57];
-
-  out[2] = (td4[(t2 >> 24) & 0xff] & 0xff000000)
-         ^ (td4[(t1 >> 16) & 0xff] & 0x00ff0000)
-         ^ (td4[(t0 >>  8) & 0xff] & 0x0000ff00)
-         ^ (td4[(t3 >>  0) & 0xff] & 0x000000ff)
-         ^ ks[58];
-
-  out[3] = (td4[(t3 >> 24) & 0xff] & 0xff000000)
-         ^ (td4[(t2 >> 16) & 0xff] & 0x00ff0000)
-         ^ (td4[(t1 >>  8) & 0xff] & 0x0000ff00)
-         ^ (td4[(t0 >>  0) & 0xff] & 0x000000ff)
-         ^ ks[59];
-
-  out[0] = swap32 (out[0]);
-  out[1] = swap32 (out[1]);
-  out[2] = swap32 (out[2]);
-  out[3] = swap32 (out[3]);
-}
-
-static void aes256_encrypt (const u32 *ks, const u32 *in, u32 *out)
-{
-  u32 in_s[4];
-
-  in_s[0] = swap32 (in[0]);
-  in_s[1] = swap32 (in[1]);
-  in_s[2] = swap32 (in[2]);
-  in_s[3] = swap32 (in[3]);
-
-  u32 s0 = in_s[0] ^ ks[0];
-  u32 s1 = in_s[1] ^ ks[1];
-  u32 s2 = in_s[2] ^ ks[2];
-  u32 s3 = in_s[3] ^ ks[3];
-
-  u32 t0;
-  u32 t1;
-  u32 t2;
-  u32 t3;
-
-  t0 = te0[s0 >> 24] ^ te1[(s1 >> 16) & 0xff] ^ te2[(s2 >>  8) & 0xff] ^ te3[s3 & 0xff] ^ ks[ 4];
-  t1 = te0[s1 >> 24] ^ te1[(s2 >> 16) & 0xff] ^ te2[(s3 >>  8) & 0xff] ^ te3[s0 & 0xff] ^ ks[ 5];
-  t2 = te0[s2 >> 24] ^ te1[(s3 >> 16) & 0xff] ^ te2[(s0 >>  8) & 0xff] ^ te3[s1 & 0xff] ^ ks[ 6];
-  t3 = te0[s3 >> 24] ^ te1[(s0 >> 16) & 0xff] ^ te2[(s1 >>  8) & 0xff] ^ te3[s2 & 0xff] ^ ks[ 7];
-  s0 = te0[t0 >> 24] ^ te1[(t1 >> 16) & 0xff] ^ te2[(t2 >>  8) & 0xff] ^ te3[t3 & 0xff] ^ ks[ 8];
-  s1 = te0[t1 >> 24] ^ te1[(t2 >> 16) & 0xff] ^ te2[(t3 >>  8) & 0xff] ^ te3[t0 & 0xff] ^ ks[ 9];
-  s2 = te0[t2 >> 24] ^ te1[(t3 >> 16) & 0xff] ^ te2[(t0 >>  8) & 0xff] ^ te3[t1 & 0xff] ^ ks[10];
-  s3 = te0[t3 >> 24] ^ te1[(t0 >> 16) & 0xff] ^ te2[(t1 >>  8) & 0xff] ^ te3[t2 & 0xff] ^ ks[11];
-  t0 = te0[s0 >> 24] ^ te1[(s1 >> 16) & 0xff] ^ te2[(s2 >>  8) & 0xff] ^ te3[s3 & 0xff] ^ ks[12];
-  t1 = te0[s1 >> 24] ^ te1[(s2 >> 16) & 0xff] ^ te2[(s3 >>  8) & 0xff] ^ te3[s0 & 0xff] ^ ks[13];
-  t2 = te0[s2 >> 24] ^ te1[(s3 >> 16) & 0xff] ^ te2[(s0 >>  8) & 0xff] ^ te3[s1 & 0xff] ^ ks[14];
-  t3 = te0[s3 >> 24] ^ te1[(s0 >> 16) & 0xff] ^ te2[(s1 >>  8) & 0xff] ^ te3[s2 & 0xff] ^ ks[15];
-  s0 = te0[t0 >> 24] ^ te1[(t1 >> 16) & 0xff] ^ te2[(t2 >>  8) & 0xff] ^ te3[t3 & 0xff] ^ ks[16];
-  s1 = te0[t1 >> 24] ^ te1[(t2 >> 16) & 0xff] ^ te2[(t3 >>  8) & 0xff] ^ te3[t0 & 0xff] ^ ks[17];
-  s2 = te0[t2 >> 24] ^ te1[(t3 >> 16) & 0xff] ^ te2[(t0 >>  8) & 0xff] ^ te3[t1 & 0xff] ^ ks[18];
-  s3 = te0[t3 >> 24] ^ te1[(t0 >> 16) & 0xff] ^ te2[(t1 >>  8) & 0xff] ^ te3[t2 & 0xff] ^ ks[19];
-  t0 = te0[s0 >> 24] ^ te1[(s1 >> 16) & 0xff] ^ te2[(s2 >>  8) & 0xff] ^ te3[s3 & 0xff] ^ ks[20];
-  t1 = te0[s1 >> 24] ^ te1[(s2 >> 16) & 0xff] ^ te2[(s3 >>  8) & 0xff] ^ te3[s0 & 0xff] ^ ks[21];
-  t2 = te0[s2 >> 24] ^ te1[(s3 >> 16) & 0xff] ^ te2[(s0 >>  8) & 0xff] ^ te3[s1 & 0xff] ^ ks[22];
-  t3 = te0[s3 >> 24] ^ te1[(s0 >> 16) & 0xff] ^ te2[(s1 >>  8) & 0xff] ^ te3[s2 & 0xff] ^ ks[23];
-  s0 = te0[t0 >> 24] ^ te1[(t1 >> 16) & 0xff] ^ te2[(t2 >>  8) & 0xff] ^ te3[t3 & 0xff] ^ ks[24];
-  s1 = te0[t1 >> 24] ^ te1[(t2 >> 16) & 0xff] ^ te2[(t3 >>  8) & 0xff] ^ te3[t0 & 0xff] ^ ks[25];
-  s2 = te0[t2 >> 24] ^ te1[(t3 >> 16) & 0xff] ^ te2[(t0 >>  8) & 0xff] ^ te3[t1 & 0xff] ^ ks[26];
-  s3 = te0[t3 >> 24] ^ te1[(t0 >> 16) & 0xff] ^ te2[(t1 >>  8) & 0xff] ^ te3[t2 & 0xff] ^ ks[27];
-  t0 = te0[s0 >> 24] ^ te1[(s1 >> 16) & 0xff] ^ te2[(s2 >>  8) & 0xff] ^ te3[s3 & 0xff] ^ ks[28];
-  t1 = te0[s1 >> 24] ^ te1[(s2 >> 16) & 0xff] ^ te2[(s3 >>  8) & 0xff] ^ te3[s0 & 0xff] ^ ks[29];
-  t2 = te0[s2 >> 24] ^ te1[(s3 >> 16) & 0xff] ^ te2[(s0 >>  8) & 0xff] ^ te3[s1 & 0xff] ^ ks[30];
-  t3 = te0[s3 >> 24] ^ te1[(s0 >> 16) & 0xff] ^ te2[(s1 >>  8) & 0xff] ^ te3[s2 & 0xff] ^ ks[31];
-  s0 = te0[t0 >> 24] ^ te1[(t1 >> 16) & 0xff] ^ te2[(t2 >>  8) & 0xff] ^ te3[t3 & 0xff] ^ ks[32];
-  s1 = te0[t1 >> 24] ^ te1[(t2 >> 16) & 0xff] ^ te2[(t3 >>  8) & 0xff] ^ te3[t0 & 0xff] ^ ks[33];
-  s2 = te0[t2 >> 24] ^ te1[(t3 >> 16) & 0xff] ^ te2[(t0 >>  8) & 0xff] ^ te3[t1 & 0xff] ^ ks[34];
-  s3 = te0[t3 >> 24] ^ te1[(t0 >> 16) & 0xff] ^ te2[(t1 >>  8) & 0xff] ^ te3[t2 & 0xff] ^ ks[35];
-  t0 = te0[s0 >> 24] ^ te1[(s1 >> 16) & 0xff] ^ te2[(s2 >>  8) & 0xff] ^ te3[s3 & 0xff] ^ ks[36];
-  t1 = te0[s1 >> 24] ^ te1[(s2 >> 16) & 0xff] ^ te2[(s3 >>  8) & 0xff] ^ te3[s0 & 0xff] ^ ks[37];
-  t2 = te0[s2 >> 24] ^ te1[(s3 >> 16) & 0xff] ^ te2[(s0 >>  8) & 0xff] ^ te3[s1 & 0xff] ^ ks[38];
-  t3 = te0[s3 >> 24] ^ te1[(s0 >> 16) & 0xff] ^ te2[(s1 >>  8) & 0xff] ^ te3[s2 & 0xff] ^ ks[39];
-  s0 = te0[t0 >> 24] ^ te1[(t1 >> 16) & 0xff] ^ te2[(t2 >>  8) & 0xff] ^ te3[t3 & 0xff] ^ ks[40];
-  s1 = te0[t1 >> 24] ^ te1[(t2 >> 16) & 0xff] ^ te2[(t3 >>  8) & 0xff] ^ te3[t0 & 0xff] ^ ks[41];
-  s2 = te0[t2 >> 24] ^ te1[(t3 >> 16) & 0xff] ^ te2[(t0 >>  8) & 0xff] ^ te3[t1 & 0xff] ^ ks[42];
-  s3 = te0[t3 >> 24] ^ te1[(t0 >> 16) & 0xff] ^ te2[(t1 >>  8) & 0xff] ^ te3[t2 & 0xff] ^ ks[43];
-  t0 = te0[s0 >> 24] ^ te1[(s1 >> 16) & 0xff] ^ te2[(s2 >>  8) & 0xff] ^ te3[s3 & 0xff] ^ ks[44];
-  t1 = te0[s1 >> 24] ^ te1[(s2 >> 16) & 0xff] ^ te2[(s3 >>  8) & 0xff] ^ te3[s0 & 0xff] ^ ks[45];
-  t2 = te0[s2 >> 24] ^ te1[(s3 >> 16) & 0xff] ^ te2[(s0 >>  8) & 0xff] ^ te3[s1 & 0xff] ^ ks[46];
-  t3 = te0[s3 >> 24] ^ te1[(s0 >> 16) & 0xff] ^ te2[(s1 >>  8) & 0xff] ^ te3[s2 & 0xff] ^ ks[47];
-  s0 = te0[t0 >> 24] ^ te1[(t1 >> 16) & 0xff] ^ te2[(t2 >>  8) & 0xff] ^ te3[t3 & 0xff] ^ ks[48];
-  s1 = te0[t1 >> 24] ^ te1[(t2 >> 16) & 0xff] ^ te2[(t3 >>  8) & 0xff] ^ te3[t0 & 0xff] ^ ks[49];
-  s2 = te0[t2 >> 24] ^ te1[(t3 >> 16) & 0xff] ^ te2[(t0 >>  8) & 0xff] ^ te3[t1 & 0xff] ^ ks[50];
-  s3 = te0[t3 >> 24] ^ te1[(t0 >> 16) & 0xff] ^ te2[(t1 >>  8) & 0xff] ^ te3[t2 & 0xff] ^ ks[51];
-  t0 = te0[s0 >> 24] ^ te1[(s1 >> 16) & 0xff] ^ te2[(s2 >>  8) & 0xff] ^ te3[s3 & 0xff] ^ ks[52];
-  t1 = te0[s1 >> 24] ^ te1[(s2 >> 16) & 0xff] ^ te2[(s3 >>  8) & 0xff] ^ te3[s0 & 0xff] ^ ks[53];
-  t2 = te0[s2 >> 24] ^ te1[(s3 >> 16) & 0xff] ^ te2[(s0 >>  8) & 0xff] ^ te3[s1 & 0xff] ^ ks[54];
-  t3 = te0[s3 >> 24] ^ te1[(s0 >> 16) & 0xff] ^ te2[(s1 >>  8) & 0xff] ^ te3[s2 & 0xff] ^ ks[55];
-
-  out[0] = (te4[(t0 >> 24) & 0xff] & 0xff000000)
-         ^ (te4[(t1 >> 16) & 0xff] & 0x00ff0000)
-         ^ (te4[(t2 >>  8) & 0xff] & 0x0000ff00)
-         ^ (te4[(t3 >>  0) & 0xff] & 0x000000ff)
-         ^ ks[56];
-
-  out[1] = (te4[(t1 >> 24) & 0xff] & 0xff000000)
-         ^ (te4[(t2 >> 16) & 0xff] & 0x00ff0000)
-         ^ (te4[(t3 >>  8) & 0xff] & 0x0000ff00)
-         ^ (te4[(t0 >>  0) & 0xff] & 0x000000ff)
-         ^ ks[57];
-
-  out[2] = (te4[(t2 >> 24) & 0xff] & 0xff000000)
-         ^ (te4[(t3 >> 16) & 0xff] & 0x00ff0000)
-         ^ (te4[(t0 >>  8) & 0xff] & 0x0000ff00)
-         ^ (te4[(t1 >>  0) & 0xff] & 0x000000ff)
-         ^ ks[58];
-
-  out[3] = (te4[(t3 >> 24) & 0xff] & 0xff000000)
-         ^ (te4[(t0 >> 16) & 0xff] & 0x00ff0000)
-         ^ (te4[(t1 >>  8) & 0xff] & 0x0000ff00)
-         ^ (te4[(t2 >>  0) & 0xff] & 0x000000ff)
-         ^ ks[59];
-
-  out[0] = swap32 (out[0]);
-  out[1] = swap32 (out[1]);
-  out[2] = swap32 (out[2]);
-  out[3] = swap32 (out[3]);
-}
-
-static void aes256_decrypt_xts (const u32 *ukey1, const u32 *ukey2, const u32 *in, u32 *out)
-{
-  u32 T[4] = { 0 };
-  u32 Z[4] = { 0 };
-
-  out[0] = in[0];
-  out[1] = in[1];
-  out[2] = in[2];
-  out[3] = in[3];
-
-  u32 ks[60];
-
-  aes256_set_encrypt_key (ks, ukey2);
-  aes256_encrypt (ks, Z, T);
-
-  out[0] ^= T[0];
-  out[1] ^= T[1];
-  out[2] ^= T[2];
-  out[3] ^= T[3];
-
-  aes256_set_decrypt_key (ks, ukey1);
-  aes256_decrypt (ks, out, out);
-
-  out[0] ^= T[0];
-  out[1] ^= T[1];
-  out[2] ^= T[2];
-  out[3] ^= T[3];
-}
diff --git a/OpenCL/kernel_serpent256.c b/OpenCL/kernel_serpent256.c
new file mode 100644
index 0000000..8288980
--- /dev/null
+++ b/OpenCL/kernel_serpent256.c
@@ -0,0 +1,583 @@
+/* This is an independent implementation of the encryption algorithm:   */
+/*                                                                      */
+/*         Serpent by Ross Anderson, Eli Biham and Lars Knudsen         */
+/*                                                                      */
+/* which is a candidate algorithm in the Advanced Encryption Standard   */
+/* programme of the US National Institute of Standards and Technology.  */
+/*                                                                      */
+/* Copyright in this implementation is held by Dr B R Gladman but I     */
+/* hereby give permission for its free direct or derivative use subject */
+/* to acknowledgment of its origin and compliance with any conditions   */
+/* that the originators of the algorithm place on its exploitation.     */
+/*                                                                      */
+/* Dr Brian Gladman (gladman@seven77.demon.co.uk) 14th January 1999     */
+/*                                                                      */
+/* -------------------------------------------------------------------- */
+/*                                                                      */
+/* Cleaned and optimized for GPU use with oclHashcat by Jens Steube     */
+
+/* 15 terms */
+
+#define sb0(a,b,c,d,e,f,g,h)    \
+    t1 = a ^ d;     \
+    t2 = a & d;     \
+    t3 = c ^ t1;    \
+    t6 = b & t1;    \
+    t4 = b ^ t3;    \
+    t10 = ~t3;      \
+    h = t2 ^ t4;    \
+    t7 = a ^ t6;    \
+    t14 = ~t7;      \
+    t8 = c | t7;    \
+    t11 = t3 ^ t7;  \
+    g = t4 ^ t8;    \
+    t12 = h & t11;  \
+    f = t10 ^ t12;  \
+    e = t12 ^ t14
+
+/* 15 terms */
+
+#define ib0(a,b,c,d,e,f,g,h)    \
+    t1 = ~a;        \
+    t2 = a ^ b;     \
+    t3 = t1 | t2;   \
+    t4 = d ^ t3;    \
+    t7 = d & t2;    \
+    t5 = c ^ t4;    \
+    t8 = t1 ^ t7;   \
+    g = t2 ^ t5;    \
+    t11 = a & t4;   \
+    t9 = g & t8;    \
+    t14 = t5 ^ t8;  \
+    f = t4 ^ t9;    \
+    t12 = t5 | f;   \
+    h = t11 ^ t12;  \
+    e = h ^ t14
+
+/* 14 terms!  */
+
+#define sb1(a,b,c,d,e,f,g,h)    \
+    t1 = ~a;        \
+    t2 = b ^ t1;    \
+    t3 = a | t2;    \
+    t4 = d | t2;    \
+    t5 = c ^ t3;    \
+    g = d ^ t5;     \
+    t7 = b ^ t4;    \
+    t8 = t2 ^ g;    \
+    t9 = t5 & t7;   \
+    h = t8 ^ t9;    \
+    t11 = t5 ^ t7;  \
+    f = h ^ t11;    \
+    t13 = t8 & t11; \
+    e = t5 ^ t13
+
+/* 17 terms */
+
+#define ib1(a,b,c,d,e,f,g,h)    \
+    t1 = a ^ d;     \
+    t2 = a & b;     \
+    t3 = b ^ c;     \
+    t4 = a ^ t3;    \
+    t5 = b | d;     \
+    t7 = c | t1;    \
+    h = t4 ^ t5;    \
+    t8 = b ^ t7;    \
+    t11 = ~t2;      \
+    t9 = t4 & t8;   \
+    f = t1 ^ t9;    \
+    t13 = t9 ^ t11; \
+    t12 = h & f;    \
+    g = t12 ^ t13;  \
+    t15 = a & d;    \
+    t16 = c ^ t13;  \
+    e = t15 ^ t16
+
+/* 16 terms */
+
+#define sb2(a,b,c,d,e,f,g,h)    \
+    t1 = ~a;        \
+    t2 = b ^ d;     \
+    t3 = c & t1;    \
+    t13 = d | t1;   \
+    e = t2 ^ t3;    \
+    t5 = c ^ t1;    \
+    t6 = c ^ e;     \
+    t7 = b & t6;    \
+    t10 = e | t5;   \
+    h = t5 ^ t7;    \
+    t9 = d | t7;    \
+    t11 = t9 & t10; \
+    t14 = t2 ^ h;   \
+    g = a ^ t11;    \
+    t15 = g ^ t13;  \
+    f = t14 ^ t15
+
+/* 16 terms */
+
+#define ib2(a,b,c,d,e,f,g,h)    \
+    t1 = b ^ d;     \
+    t2 = ~t1;       \
+    t3 = a ^ c;     \
+    t4 = c ^ t1;    \
+    t7 = a | t2;    \
+    t5 = b & t4;    \
+    t8 = d ^ t7;    \
+    t11 = ~t4;      \
+    e = t3 ^ t5;    \
+    t9 = t3 | t8;   \
+    t14 = d & t11;  \
+    h = t1 ^ t9;    \
+    t12 = e | h;    \
+    f = t11 ^ t12;  \
+    t15 = t3 ^ t12; \
+    g = t14 ^ t15
+
+/* 17 terms */
+
+#define sb3(a,b,c,d,e,f,g,h)    \
+    t1 = a ^ c;     \
+    t2 = d ^ t1;    \
+    t3 = a & t2;    \
+    t4 = d ^ t3;    \
+    t5 = b & t4;    \
+    g = t2 ^ t5;    \
+    t7 = a | g;     \
+    t8 = b | d;     \
+    t11 = a | d;    \
+    t9 = t4 & t7;   \
+    f = t8 ^ t9;    \
+    t12 = b ^ t11;  \
+    t13 = g ^ t9;   \
+    t15 = t3 ^ t8;  \
+    h = t12 ^ t13;  \
+    t16 = c & t15;  \
+    e = t12 ^ t16
+
+/* 16 term solution that performs less well than 17 term one
+   in my environment (PPro/PII)
+
+#define sb3(a,b,c,d,e,f,g,h)    \
+    t1 = a ^ b;     \
+    t2 = a & c;     \
+    t3 = a | d;     \
+    t4 = c ^ d;     \
+    t5 = t1 & t3;   \
+    t6 = t2 | t5;   \
+    g = t4 ^ t6;    \
+    t8 = b ^ t3;    \
+    t9 = t6 ^ t8;   \
+    t10 = t4 & t9;  \
+    e = t1 ^ t10;   \
+    t12 = g & e;    \
+    f = t9 ^ t12;   \
+    t14 = b | d;    \
+    t15 = t4 ^ t12; \
+    h = t14 ^ t15
+*/
+
+/* 17 terms */
+
+#define ib3(a,b,c,d,e,f,g,h)    \
+    t1 = b ^ c;     \
+    t2 = b | c;     \
+    t3 = a ^ c;     \
+    t7 = a ^ d;     \
+    t4 = t2 ^ t3;   \
+    t5 = d | t4;    \
+    t9 = t2 ^ t7;   \
+    e = t1 ^ t5;    \
+    t8 = t1 | t5;   \
+    t11 = a & t4;   \
+    g = t8 ^ t9;    \
+    t12 = e | t9;   \
+    f = t11 ^ t12;  \
+    t14 = a & g;    \
+    t15 = t2 ^ t14; \
+    t16 = e & t15;  \
+    h = t4 ^ t16
+
+/* 15 terms */
+
+#define sb4(a,b,c,d,e,f,g,h)    \
+    t1 = a ^ d;     \
+    t2 = d & t1;    \
+    t3 = c ^ t2;    \
+    t4 = b | t3;    \
+    h = t1 ^ t4;    \
+    t6 = ~b;        \
+    t7 = t1 | t6;   \
+    e = t3 ^ t7;    \
+    t9 = a & e;     \
+    t10 = t1 ^ t6;  \
+    t11 = t4 & t10; \
+    g = t9 ^ t11;   \
+    t13 = a ^ t3;   \
+    t14 = t10 & g;  \
+    f = t13 ^ t14
+
+/* 17 terms */
+
+#define ib4(a,b,c,d,e,f,g,h)    \
+    t1 = c ^ d;     \
+    t2 = c | d;     \
+    t3 = b ^ t2;    \
+    t4 = a & t3;    \
+    f = t1 ^ t4;    \
+    t6 = a ^ d;     \
+    t7 = b | d;     \
+    t8 = t6 & t7;   \
+    h = t3 ^ t8;    \
+    t10 = ~a;       \
+    t11 = c ^ h;    \
+    t12 = t10 | t11;\
+    e = t3 ^ t12;   \
+    t14 = c | t4;   \
+    t15 = t7 ^ t14; \
+    t16 = h | t10;  \
+    g = t15 ^ t16
+
+/* 16 terms */
+
+#define sb5(a,b,c,d,e,f,g,h)    \
+    t1 = ~a;        \
+    t2 = a ^ b;     \
+    t3 = a ^ d;     \
+    t4 = c ^ t1;    \
+    t5 = t2 | t3;   \
+    e = t4 ^ t5;    \
+    t7 = d & e;     \
+    t8 = t2 ^ e;    \
+    t10 = t1 | e;   \
+    f = t7 ^ t8;    \
+    t11 = t2 | t7;  \
+    t12 = t3 ^ t10; \
+    t14 = b ^ t7;   \
+    g = t11 ^ t12;  \
+    t15 = f & t12;  \
+    h = t14 ^ t15
+
+/* 16 terms */
+
+#define ib5(a,b,c,d,e,f,g,h)    \
+    t1 = ~c;        \
+    t2 = b & t1;    \
+    t3 = d ^ t2;    \
+    t4 = a & t3;    \
+    t5 = b ^ t1;    \
+    h = t4 ^ t5;    \
+    t7 = b | h;     \
+    t8 = a & t7;    \
+    f = t3 ^ t8;    \
+    t10 = a | d;    \
+    t11 = t1 ^ t7;  \
+    e = t10 ^ t11;  \
+    t13 = a ^ c;    \
+    t14 = b & t10;  \
+    t15 = t4 | t13; \
+    g = t14 ^ t15
+
+/* 15 terms */
+
+#define sb6(a,b,c,d,e,f,g,h)    \
+    t1 = ~a;        \
+    t2 = a ^ d;     \
+    t3 = b ^ t2;    \
+    t4 = t1 | t2;   \
+    t5 = c ^ t4;    \
+    f = b ^ t5;     \
+    t13 = ~t5;      \
+    t7 = t2 | f;    \
+    t8 = d ^ t7;    \
+    t9 = t5 & t8;   \
+    g = t3 ^ t9;    \
+    t11 = t5 ^ t8;  \
+    e = g ^ t11;    \
+    t14 = t3 & t11; \
+    h = t13 ^ t14
+
+/* 15 terms */
+
+#define ib6(a,b,c,d,e,f,g,h)    \
+    t1 = ~a;        \
+    t2 = a ^ b;     \
+    t3 = c ^ t2;    \
+    t4 = c | t1;    \
+    t5 = d ^ t4;    \
+    t13 = d & t1;   \
+    f = t3 ^ t5;    \
+    t7 = t3 & t5;   \
+    t8 = t2 ^ t7;   \
+    t9 = b | t8;    \
+    h = t5 ^ t9;    \
+    t11 = b | h;    \
+    e = t8 ^ t11;   \
+    t14 = t3 ^ t11; \
+    g = t13 ^ t14
+
+/* 17 terms */
+
+#define sb7(a,b,c,d,e,f,g,h)    \
+    t1 = ~c;        \
+    t2 = b ^ c;     \
+    t3 = b | t1;    \
+    t4 = d ^ t3;    \
+    t5 = a & t4;    \
+    t7 = a ^ d;     \
+    h = t2 ^ t5;    \
+    t8 = b ^ t5;    \
+    t9 = t2 | t8;   \
+    t11 = d & t3;   \
+    f = t7 ^ t9;    \
+    t12 = t5 ^ f;   \
+    t15 = t1 | t4;  \
+    t13 = h & t12;  \
+    g = t11 ^ t13;  \
+    t16 = t12 ^ g;  \
+    e = t15 ^ t16
+
+/* 17 terms */
+
+#define ib7(a,b,c,d,e,f,g,h)    \
+    t1 = a & b;     \
+    t2 = a | b;     \
+    t3 = c | t1;    \
+    t4 = d & t2;    \
+    h = t3 ^ t4;    \
+    t6 = ~d;        \
+    t7 = b ^ t4;    \
+    t8 = h ^ t6;    \
+    t11 = c ^ t7;   \
+    t9 = t7 | t8;   \
+    f = a ^ t9;     \
+    t12 = d | f;    \
+    e = t11 ^ t12;  \
+    t14 = a & h;    \
+    t15 = t3 ^ f;   \
+    t16 = e ^ t14;  \
+    g = t15 ^ t16
+
+#define k_xor(r,a,b,c,d)    \
+    a ^= ks[4 * r +  8]; \
+    b ^= ks[4 * r +  9]; \
+    c ^= ks[4 * r + 10]; \
+    d ^= ks[4 * r + 11]
+
+#define k_set(r,a,b,c,d)    \
+    a = ks[4 * r +  8];  \
+    b = ks[4 * r +  9];  \
+    c = ks[4 * r + 10];  \
+    d = ks[4 * r + 11]
+
+#define k_get(r,a,b,c,d)    \
+    ks[4 * r +  8] = a;  \
+    ks[4 * r +  9] = b;  \
+    ks[4 * r + 10] = c;  \
+    ks[4 * r + 11] = d
+
+/* the linear transformation and its inverse    */
+
+#define rot(a,b,c,d)    \
+    a = rotl32(a, 13);    \
+    c = rotl32(c, 3);     \
+    d ^= c ^ (a << 3);  \
+    b ^= a ^ c;         \
+    d = rotl32(d, 7);     \
+    b = rotl32(b, 1);     \
+    a ^= b ^ d;         \
+    c ^= d ^ (b << 7);  \
+    a = rotl32(a, 5);     \
+    c = rotl32(c, 22)
+
+#define irot(a,b,c,d)   \
+    c = rotr32(c, 22);    \
+    a = rotr32(a, 5);     \
+    c ^= d ^ (b << 7);  \
+    a ^= b ^ d;         \
+    d = rotr32(d, 7);     \
+    b = rotr32(b, 1);     \
+    d ^= c ^ (a << 3);  \
+    b ^= a ^ c;         \
+    c = rotr32(c, 3);     \
+    a = rotr32(a, 13)
+
+static void serpent256_set_key (u32 *ks, const u32 *ukey)
+{
+  #pragma unroll
+  for (int i = 0; i < 8; i++)
+  {
+    ks[i] = ukey[i];
+  }
+
+  #pragma unroll
+  for (int i = 0; i < 132; i++)
+  {
+    ks[i + 8] = rotl32 (ks[i + 7] ^ ks[i + 5] ^ ks[i + 3] ^ ks[i + 0] ^ 0x9e3779b9 ^ i, 11);
+  }
+
+  u32  a,b,c,d,e,f,g,h;
+  u32  t1,t2,t3,t4,t5,t6,t7,t8,t9,t10,t11,t12,t13,t14,t15,t16;
+
+  k_set( 0,a,b,c,d); sb3(a,b,c,d,e,f,g,h); k_get( 0,e,f,g,h);
+  k_set( 1,a,b,c,d); sb2(a,b,c,d,e,f,g,h); k_get( 1,e,f,g,h);
+  k_set( 2,a,b,c,d); sb1(a,b,c,d,e,f,g,h); k_get( 2,e,f,g,h);
+  k_set( 3,a,b,c,d); sb0(a,b,c,d,e,f,g,h); k_get( 3,e,f,g,h);
+  k_set( 4,a,b,c,d); sb7(a,b,c,d,e,f,g,h); k_get( 4,e,f,g,h);
+  k_set( 5,a,b,c,d); sb6(a,b,c,d,e,f,g,h); k_get( 5,e,f,g,h);
+  k_set( 6,a,b,c,d); sb5(a,b,c,d,e,f,g,h); k_get( 6,e,f,g,h);
+  k_set( 7,a,b,c,d); sb4(a,b,c,d,e,f,g,h); k_get( 7,e,f,g,h);
+  k_set( 8,a,b,c,d); sb3(a,b,c,d,e,f,g,h); k_get( 8,e,f,g,h);
+  k_set( 9,a,b,c,d); sb2(a,b,c,d,e,f,g,h); k_get( 9,e,f,g,h);
+  k_set(10,a,b,c,d); sb1(a,b,c,d,e,f,g,h); k_get(10,e,f,g,h);
+  k_set(11,a,b,c,d); sb0(a,b,c,d,e,f,g,h); k_get(11,e,f,g,h);
+  k_set(12,a,b,c,d); sb7(a,b,c,d,e,f,g,h); k_get(12,e,f,g,h);
+  k_set(13,a,b,c,d); sb6(a,b,c,d,e,f,g,h); k_get(13,e,f,g,h);
+  k_set(14,a,b,c,d); sb5(a,b,c,d,e,f,g,h); k_get(14,e,f,g,h);
+  k_set(15,a,b,c,d); sb4(a,b,c,d,e,f,g,h); k_get(15,e,f,g,h);
+  k_set(16,a,b,c,d); sb3(a,b,c,d,e,f,g,h); k_get(16,e,f,g,h);
+  k_set(17,a,b,c,d); sb2(a,b,c,d,e,f,g,h); k_get(17,e,f,g,h);
+  k_set(18,a,b,c,d); sb1(a,b,c,d,e,f,g,h); k_get(18,e,f,g,h);
+  k_set(19,a,b,c,d); sb0(a,b,c,d,e,f,g,h); k_get(19,e,f,g,h);
+  k_set(20,a,b,c,d); sb7(a,b,c,d,e,f,g,h); k_get(20,e,f,g,h);
+  k_set(21,a,b,c,d); sb6(a,b,c,d,e,f,g,h); k_get(21,e,f,g,h);
+  k_set(22,a,b,c,d); sb5(a,b,c,d,e,f,g,h); k_get(22,e,f,g,h);
+  k_set(23,a,b,c,d); sb4(a,b,c,d,e,f,g,h); k_get(23,e,f,g,h);
+  k_set(24,a,b,c,d); sb3(a,b,c,d,e,f,g,h); k_get(24,e,f,g,h);
+  k_set(25,a,b,c,d); sb2(a,b,c,d,e,f,g,h); k_get(25,e,f,g,h);
+  k_set(26,a,b,c,d); sb1(a,b,c,d,e,f,g,h); k_get(26,e,f,g,h);
+  k_set(27,a,b,c,d); sb0(a,b,c,d,e,f,g,h); k_get(27,e,f,g,h);
+  k_set(28,a,b,c,d); sb7(a,b,c,d,e,f,g,h); k_get(28,e,f,g,h);
+  k_set(29,a,b,c,d); sb6(a,b,c,d,e,f,g,h); k_get(29,e,f,g,h);
+  k_set(30,a,b,c,d); sb5(a,b,c,d,e,f,g,h); k_get(30,e,f,g,h);
+  k_set(31,a,b,c,d); sb4(a,b,c,d,e,f,g,h); k_get(31,e,f,g,h);
+  k_set(32,a,b,c,d); sb3(a,b,c,d,e,f,g,h); k_get(32,e,f,g,h);
+}
+
+static void serpent256_encrypt (const u32 *ks, const u32 *in, u32 *out)
+{
+  u32  a,b,c,d,e,f,g,h;
+  u32  t1,t2,t3,t4,t5,t6,t7,t8,t9,t10,t11,t12,t13,t14,t15,t16;
+
+  a = in[0];
+  b = in[1];
+  c = in[2];
+  d = in[3];
+
+  k_xor( 0,a,b,c,d); sb0(a,b,c,d,e,f,g,h); rot(e,f,g,h);
+  k_xor( 1,e,f,g,h); sb1(e,f,g,h,a,b,c,d); rot(a,b,c,d);
+  k_xor( 2,a,b,c,d); sb2(a,b,c,d,e,f,g,h); rot(e,f,g,h);
+  k_xor( 3,e,f,g,h); sb3(e,f,g,h,a,b,c,d); rot(a,b,c,d);
+  k_xor( 4,a,b,c,d); sb4(a,b,c,d,e,f,g,h); rot(e,f,g,h);
+  k_xor( 5,e,f,g,h); sb5(e,f,g,h,a,b,c,d); rot(a,b,c,d);
+  k_xor( 6,a,b,c,d); sb6(a,b,c,d,e,f,g,h); rot(e,f,g,h);
+  k_xor( 7,e,f,g,h); sb7(e,f,g,h,a,b,c,d); rot(a,b,c,d);
+  k_xor( 8,a,b,c,d); sb0(a,b,c,d,e,f,g,h); rot(e,f,g,h);
+  k_xor( 9,e,f,g,h); sb1(e,f,g,h,a,b,c,d); rot(a,b,c,d);
+  k_xor(10,a,b,c,d); sb2(a,b,c,d,e,f,g,h); rot(e,f,g,h);
+  k_xor(11,e,f,g,h); sb3(e,f,g,h,a,b,c,d); rot(a,b,c,d);
+  k_xor(12,a,b,c,d); sb4(a,b,c,d,e,f,g,h); rot(e,f,g,h);
+  k_xor(13,e,f,g,h); sb5(e,f,g,h,a,b,c,d); rot(a,b,c,d);
+  k_xor(14,a,b,c,d); sb6(a,b,c,d,e,f,g,h); rot(e,f,g,h);
+  k_xor(15,e,f,g,h); sb7(e,f,g,h,a,b,c,d); rot(a,b,c,d);
+  k_xor(16,a,b,c,d); sb0(a,b,c,d,e,f,g,h); rot(e,f,g,h);
+  k_xor(17,e,f,g,h); sb1(e,f,g,h,a,b,c,d); rot(a,b,c,d);
+  k_xor(18,a,b,c,d); sb2(a,b,c,d,e,f,g,h); rot(e,f,g,h);
+  k_xor(19,e,f,g,h); sb3(e,f,g,h,a,b,c,d); rot(a,b,c,d);
+  k_xor(20,a,b,c,d); sb4(a,b,c,d,e,f,g,h); rot(e,f,g,h);
+  k_xor(21,e,f,g,h); sb5(e,f,g,h,a,b,c,d); rot(a,b,c,d);
+  k_xor(22,a,b,c,d); sb6(a,b,c,d,e,f,g,h); rot(e,f,g,h);
+  k_xor(23,e,f,g,h); sb7(e,f,g,h,a,b,c,d); rot(a,b,c,d);
+  k_xor(24,a,b,c,d); sb0(a,b,c,d,e,f,g,h); rot(e,f,g,h);
+  k_xor(25,e,f,g,h); sb1(e,f,g,h,a,b,c,d); rot(a,b,c,d);
+  k_xor(26,a,b,c,d); sb2(a,b,c,d,e,f,g,h); rot(e,f,g,h);
+  k_xor(27,e,f,g,h); sb3(e,f,g,h,a,b,c,d); rot(a,b,c,d);
+  k_xor(28,a,b,c,d); sb4(a,b,c,d,e,f,g,h); rot(e,f,g,h);
+  k_xor(29,e,f,g,h); sb5(e,f,g,h,a,b,c,d); rot(a,b,c,d);
+  k_xor(30,a,b,c,d); sb6(a,b,c,d,e,f,g,h); rot(e,f,g,h);
+  k_xor(31,e,f,g,h); sb7(e,f,g,h,a,b,c,d);
+  k_xor(32,a,b,c,d);
+
+  out[0] = a;
+  out[1] = b;
+  out[2] = c;
+  out[3] = d;
+}
+
+static void serpent256_decrypt (const u32 *ks, const u32 *in, u32 *out)
+{
+  u32  a,b,c,d,e,f,g,h;
+  u32  t1,t2,t3,t4,t5,t6,t7,t8,t9,t10,t11,t12,t13,t14,t15,t16;
+
+  a = in[0];
+  b = in[1];
+  c = in[2];
+  d = in[3];
+
+                                       k_xor(32,a,b,c,d);
+                 ib7(a,b,c,d,e,f,g,h); k_xor(31,e,f,g,h);
+  irot(e,f,g,h); ib6(e,f,g,h,a,b,c,d); k_xor(30,a,b,c,d);
+  irot(a,b,c,d); ib5(a,b,c,d,e,f,g,h); k_xor(29,e,f,g,h);
+  irot(e,f,g,h); ib4(e,f,g,h,a,b,c,d); k_xor(28,a,b,c,d);
+  irot(a,b,c,d); ib3(a,b,c,d,e,f,g,h); k_xor(27,e,f,g,h);
+  irot(e,f,g,h); ib2(e,f,g,h,a,b,c,d); k_xor(26,a,b,c,d);
+  irot(a,b,c,d); ib1(a,b,c,d,e,f,g,h); k_xor(25,e,f,g,h);
+  irot(e,f,g,h); ib0(e,f,g,h,a,b,c,d); k_xor(24,a,b,c,d);
+  irot(a,b,c,d); ib7(a,b,c,d,e,f,g,h); k_xor(23,e,f,g,h);
+  irot(e,f,g,h); ib6(e,f,g,h,a,b,c,d); k_xor(22,a,b,c,d);
+  irot(a,b,c,d); ib5(a,b,c,d,e,f,g,h); k_xor(21,e,f,g,h);
+  irot(e,f,g,h); ib4(e,f,g,h,a,b,c,d); k_xor(20,a,b,c,d);
+  irot(a,b,c,d); ib3(a,b,c,d,e,f,g,h); k_xor(19,e,f,g,h);
+  irot(e,f,g,h); ib2(e,f,g,h,a,b,c,d); k_xor(18,a,b,c,d);
+  irot(a,b,c,d); ib1(a,b,c,d,e,f,g,h); k_xor(17,e,f,g,h);
+  irot(e,f,g,h); ib0(e,f,g,h,a,b,c,d); k_xor(16,a,b,c,d);
+  irot(a,b,c,d); ib7(a,b,c,d,e,f,g,h); k_xor(15,e,f,g,h);
+  irot(e,f,g,h); ib6(e,f,g,h,a,b,c,d); k_xor(14,a,b,c,d);
+  irot(a,b,c,d); ib5(a,b,c,d,e,f,g,h); k_xor(13,e,f,g,h);
+  irot(e,f,g,h); ib4(e,f,g,h,a,b,c,d); k_xor(12,a,b,c,d);
+  irot(a,b,c,d); ib3(a,b,c,d,e,f,g,h); k_xor(11,e,f,g,h);
+  irot(e,f,g,h); ib2(e,f,g,h,a,b,c,d); k_xor(10,a,b,c,d);
+  irot(a,b,c,d); ib1(a,b,c,d,e,f,g,h); k_xor( 9,e,f,g,h);
+  irot(e,f,g,h); ib0(e,f,g,h,a,b,c,d); k_xor( 8,a,b,c,d);
+  irot(a,b,c,d); ib7(a,b,c,d,e,f,g,h); k_xor( 7,e,f,g,h);
+  irot(e,f,g,h); ib6(e,f,g,h,a,b,c,d); k_xor( 6,a,b,c,d);
+  irot(a,b,c,d); ib5(a,b,c,d,e,f,g,h); k_xor( 5,e,f,g,h);
+  irot(e,f,g,h); ib4(e,f,g,h,a,b,c,d); k_xor( 4,a,b,c,d);
+  irot(a,b,c,d); ib3(a,b,c,d,e,f,g,h); k_xor( 3,e,f,g,h);
+  irot(e,f,g,h); ib2(e,f,g,h,a,b,c,d); k_xor( 2,a,b,c,d);
+  irot(a,b,c,d); ib1(a,b,c,d,e,f,g,h); k_xor( 1,e,f,g,h);
+  irot(e,f,g,h); ib0(e,f,g,h,a,b,c,d); k_xor( 0,a,b,c,d);
+
+  out[0] = a;
+  out[1] = b;
+  out[2] = c;
+  out[3] = d;
+}
+
+static void serpent256_decrypt_xts (const u32 *ukey1, const u32 *ukey2, const u32 *in, u32 *out)
+{
+  u32 T[4] = { 0 };
+  u32 Z[4] = { 0 };
+
+  out[0] = in[0];
+  out[1] = in[1];
+  out[2] = in[2];
+  out[3] = in[3];
+
+  u32 ks[140];
+
+  serpent256_set_key (ks, ukey2);
+  serpent256_encrypt (ks, Z, T);
+
+  out[0] ^= T[0];
+  out[1] ^= T[1];
+  out[2] ^= T[2];
+  out[3] ^= T[3];
+
+  serpent256_set_key (ks, ukey1);
+  serpent256_decrypt (ks, out, out);
+
+  out[0] ^= T[0];
+  out[1] ^= T[1];
+  out[2] ^= T[2];
+  out[3] ^= T[3];
+}
diff --git a/OpenCL/kernel_serpent256_amd.c b/OpenCL/kernel_serpent256_amd.c
deleted file mode 100644
index 8288980..0000000
--- a/OpenCL/kernel_serpent256_amd.c
+++ /dev/null
@@ -1,583 +0,0 @@
-/* This is an independent implementation of the encryption algorithm:   */
-/*                                                                      */
-/*         Serpent by Ross Anderson, Eli Biham and Lars Knudsen         */
-/*                                                                      */
-/* which is a candidate algorithm in the Advanced Encryption Standard   */
-/* programme of the US National Institute of Standards and Technology.  */
-/*                                                                      */
-/* Copyright in this implementation is held by Dr B R Gladman but I     */
-/* hereby give permission for its free direct or derivative use subject */
-/* to acknowledgment of its origin and compliance with any conditions   */
-/* that the originators of the algorithm place on its exploitation.     */
-/*                                                                      */
-/* Dr Brian Gladman (gladman@seven77.demon.co.uk) 14th January 1999     */
-/*                                                                      */
-/* -------------------------------------------------------------------- */
-/*                                                                      */
-/* Cleaned and optimized for GPU use with oclHashcat by Jens Steube     */
-
-/* 15 terms */
-
-#define sb0(a,b,c,d,e,f,g,h)    \
-    t1 = a ^ d;     \
-    t2 = a & d;     \
-    t3 = c ^ t1;    \
-    t6 = b & t1;    \
-    t4 = b ^ t3;    \
-    t10 = ~t3;      \
-    h = t2 ^ t4;    \
-    t7 = a ^ t6;    \
-    t14 = ~t7;      \
-    t8 = c | t7;    \
-    t11 = t3 ^ t7;  \
-    g = t4 ^ t8;    \
-    t12 = h & t11;  \
-    f = t10 ^ t12;  \
-    e = t12 ^ t14
-
-/* 15 terms */
-
-#define ib0(a,b,c,d,e,f,g,h)    \
-    t1 = ~a;        \
-    t2 = a ^ b;     \
-    t3 = t1 | t2;   \
-    t4 = d ^ t3;    \
-    t7 = d & t2;    \
-    t5 = c ^ t4;    \
-    t8 = t1 ^ t7;   \
-    g = t2 ^ t5;    \
-    t11 = a & t4;   \
-    t9 = g & t8;    \
-    t14 = t5 ^ t8;  \
-    f = t4 ^ t9;    \
-    t12 = t5 | f;   \
-    h = t11 ^ t12;  \
-    e = h ^ t14
-
-/* 14 terms!  */
-
-#define sb1(a,b,c,d,e,f,g,h)    \
-    t1 = ~a;        \
-    t2 = b ^ t1;    \
-    t3 = a | t2;    \
-    t4 = d | t2;    \
-    t5 = c ^ t3;    \
-    g = d ^ t5;     \
-    t7 = b ^ t4;    \
-    t8 = t2 ^ g;    \
-    t9 = t5 & t7;   \
-    h = t8 ^ t9;    \
-    t11 = t5 ^ t7;  \
-    f = h ^ t11;    \
-    t13 = t8 & t11; \
-    e = t5 ^ t13
-
-/* 17 terms */
-
-#define ib1(a,b,c,d,e,f,g,h)    \
-    t1 = a ^ d;     \
-    t2 = a & b;     \
-    t3 = b ^ c;     \
-    t4 = a ^ t3;    \
-    t5 = b | d;     \
-    t7 = c | t1;    \
-    h = t4 ^ t5;    \
-    t8 = b ^ t7;    \
-    t11 = ~t2;      \
-    t9 = t4 & t8;   \
-    f = t1 ^ t9;    \
-    t13 = t9 ^ t11; \
-    t12 = h & f;    \
-    g = t12 ^ t13;  \
-    t15 = a & d;    \
-    t16 = c ^ t13;  \
-    e = t15 ^ t16
-
-/* 16 terms */
-
-#define sb2(a,b,c,d,e,f,g,h)    \
-    t1 = ~a;        \
-    t2 = b ^ d;     \
-    t3 = c & t1;    \
-    t13 = d | t1;   \
-    e = t2 ^ t3;    \
-    t5 = c ^ t1;    \
-    t6 = c ^ e;     \
-    t7 = b & t6;    \
-    t10 = e | t5;   \
-    h = t5 ^ t7;    \
-    t9 = d | t7;    \
-    t11 = t9 & t10; \
-    t14 = t2 ^ h;   \
-    g = a ^ t11;    \
-    t15 = g ^ t13;  \
-    f = t14 ^ t15
-
-/* 16 terms */
-
-#define ib2(a,b,c,d,e,f,g,h)    \
-    t1 = b ^ d;     \
-    t2 = ~t1;       \
-    t3 = a ^ c;     \
-    t4 = c ^ t1;    \
-    t7 = a | t2;    \
-    t5 = b & t4;    \
-    t8 = d ^ t7;    \
-    t11 = ~t4;      \
-    e = t3 ^ t5;    \
-    t9 = t3 | t8;   \
-    t14 = d & t11;  \
-    h = t1 ^ t9;    \
-    t12 = e | h;    \
-    f = t11 ^ t12;  \
-    t15 = t3 ^ t12; \
-    g = t14 ^ t15
-
-/* 17 terms */
-
-#define sb3(a,b,c,d,e,f,g,h)    \
-    t1 = a ^ c;     \
-    t2 = d ^ t1;    \
-    t3 = a & t2;    \
-    t4 = d ^ t3;    \
-    t5 = b & t4;    \
-    g = t2 ^ t5;    \
-    t7 = a | g;     \
-    t8 = b | d;     \
-    t11 = a | d;    \
-    t9 = t4 & t7;   \
-    f = t8 ^ t9;    \
-    t12 = b ^ t11;  \
-    t13 = g ^ t9;   \
-    t15 = t3 ^ t8;  \
-    h = t12 ^ t13;  \
-    t16 = c & t15;  \
-    e = t12 ^ t16
-
-/* 16 term solution that performs less well than 17 term one
-   in my environment (PPro/PII)
-
-#define sb3(a,b,c,d,e,f,g,h)    \
-    t1 = a ^ b;     \
-    t2 = a & c;     \
-    t3 = a | d;     \
-    t4 = c ^ d;     \
-    t5 = t1 & t3;   \
-    t6 = t2 | t5;   \
-    g = t4 ^ t6;    \
-    t8 = b ^ t3;    \
-    t9 = t6 ^ t8;   \
-    t10 = t4 & t9;  \
-    e = t1 ^ t10;   \
-    t12 = g & e;    \
-    f = t9 ^ t12;   \
-    t14 = b | d;    \
-    t15 = t4 ^ t12; \
-    h = t14 ^ t15
-*/
-
-/* 17 terms */
-
-#define ib3(a,b,c,d,e,f,g,h)    \
-    t1 = b ^ c;     \
-    t2 = b | c;     \
-    t3 = a ^ c;     \
-    t7 = a ^ d;     \
-    t4 = t2 ^ t3;   \
-    t5 = d | t4;    \
-    t9 = t2 ^ t7;   \
-    e = t1 ^ t5;    \
-    t8 = t1 | t5;   \
-    t11 = a & t4;   \
-    g = t8 ^ t9;    \
-    t12 = e | t9;   \
-    f = t11 ^ t12;  \
-    t14 = a & g;    \
-    t15 = t2 ^ t14; \
-    t16 = e & t15;  \
-    h = t4 ^ t16
-
-/* 15 terms */
-
-#define sb4(a,b,c,d,e,f,g,h)    \
-    t1 = a ^ d;     \
-    t2 = d & t1;    \
-    t3 = c ^ t2;    \
-    t4 = b | t3;    \
-    h = t1 ^ t4;    \
-    t6 = ~b;        \
-    t7 = t1 | t6;   \
-    e = t3 ^ t7;    \
-    t9 = a & e;     \
-    t10 = t1 ^ t6;  \
-    t11 = t4 & t10; \
-    g = t9 ^ t11;   \
-    t13 = a ^ t3;   \
-    t14 = t10 & g;  \
-    f = t13 ^ t14
-
-/* 17 terms */
-
-#define ib4(a,b,c,d,e,f,g,h)    \
-    t1 = c ^ d;     \
-    t2 = c | d;     \
-    t3 = b ^ t2;    \
-    t4 = a & t3;    \
-    f = t1 ^ t4;    \
-    t6 = a ^ d;     \
-    t7 = b | d;     \
-    t8 = t6 & t7;   \
-    h = t3 ^ t8;    \
-    t10 = ~a;       \
-    t11 = c ^ h;    \
-    t12 = t10 | t11;\
-    e = t3 ^ t12;   \
-    t14 = c | t4;   \
-    t15 = t7 ^ t14; \
-    t16 = h | t10;  \
-    g = t15 ^ t16
-
-/* 16 terms */
-
-#define sb5(a,b,c,d,e,f,g,h)    \
-    t1 = ~a;        \
-    t2 = a ^ b;     \
-    t3 = a ^ d;     \
-    t4 = c ^ t1;    \
-    t5 = t2 | t3;   \
-    e = t4 ^ t5;    \
-    t7 = d & e;     \
-    t8 = t2 ^ e;    \
-    t10 = t1 | e;   \
-    f = t7 ^ t8;    \
-    t11 = t2 | t7;  \
-    t12 = t3 ^ t10; \
-    t14 = b ^ t7;   \
-    g = t11 ^ t12;  \
-    t15 = f & t12;  \
-    h = t14 ^ t15
-
-/* 16 terms */
-
-#define ib5(a,b,c,d,e,f,g,h)    \
-    t1 = ~c;        \
-    t2 = b & t1;    \
-    t3 = d ^ t2;    \
-    t4 = a & t3;    \
-    t5 = b ^ t1;    \
-    h = t4 ^ t5;    \
-    t7 = b | h;     \
-    t8 = a & t7;    \
-    f = t3 ^ t8;    \
-    t10 = a | d;    \
-    t11 = t1 ^ t7;  \
-    e = t10 ^ t11;  \
-    t13 = a ^ c;    \
-    t14 = b & t10;  \
-    t15 = t4 | t13; \
-    g = t14 ^ t15
-
-/* 15 terms */
-
-#define sb6(a,b,c,d,e,f,g,h)    \
-    t1 = ~a;        \
-    t2 = a ^ d;     \
-    t3 = b ^ t2;    \
-    t4 = t1 | t2;   \
-    t5 = c ^ t4;    \
-    f = b ^ t5;     \
-    t13 = ~t5;      \
-    t7 = t2 | f;    \
-    t8 = d ^ t7;    \
-    t9 = t5 & t8;   \
-    g = t3 ^ t9;    \
-    t11 = t5 ^ t8;  \
-    e = g ^ t11;    \
-    t14 = t3 & t11; \
-    h = t13 ^ t14
-
-/* 15 terms */
-
-#define ib6(a,b,c,d,e,f,g,h)    \
-    t1 = ~a;        \
-    t2 = a ^ b;     \
-    t3 = c ^ t2;    \
-    t4 = c | t1;    \
-    t5 = d ^ t4;    \
-    t13 = d & t1;   \
-    f = t3 ^ t5;    \
-    t7 = t3 & t5;   \
-    t8 = t2 ^ t7;   \
-    t9 = b | t8;    \
-    h = t5 ^ t9;    \
-    t11 = b | h;    \
-    e = t8 ^ t11;   \
-    t14 = t3 ^ t11; \
-    g = t13 ^ t14
-
-/* 17 terms */
-
-#define sb7(a,b,c,d,e,f,g,h)    \
-    t1 = ~c;        \
-    t2 = b ^ c;     \
-    t3 = b | t1;    \
-    t4 = d ^ t3;    \
-    t5 = a & t4;    \
-    t7 = a ^ d;     \
-    h = t2 ^ t5;    \
-    t8 = b ^ t5;    \
-    t9 = t2 | t8;   \
-    t11 = d & t3;   \
-    f = t7 ^ t9;    \
-    t12 = t5 ^ f;   \
-    t15 = t1 | t4;  \
-    t13 = h & t12;  \
-    g = t11 ^ t13;  \
-    t16 = t12 ^ g;  \
-    e = t15 ^ t16
-
-/* 17 terms */
-
-#define ib7(a,b,c,d,e,f,g,h)    \
-    t1 = a & b;     \
-    t2 = a | b;     \
-    t3 = c | t1;    \
-    t4 = d & t2;    \
-    h = t3 ^ t4;    \
-    t6 = ~d;        \
-    t7 = b ^ t4;    \
-    t8 = h ^ t6;    \
-    t11 = c ^ t7;   \
-    t9 = t7 | t8;   \
-    f = a ^ t9;     \
-    t12 = d | f;    \
-    e = t11 ^ t12;  \
-    t14 = a & h;    \
-    t15 = t3 ^ f;   \
-    t16 = e ^ t14;  \
-    g = t15 ^ t16
-
-#define k_xor(r,a,b,c,d)    \
-    a ^= ks[4 * r +  8]; \
-    b ^= ks[4 * r +  9]; \
-    c ^= ks[4 * r + 10]; \
-    d ^= ks[4 * r + 11]
-
-#define k_set(r,a,b,c,d)    \
-    a = ks[4 * r +  8];  \
-    b = ks[4 * r +  9];  \
-    c = ks[4 * r + 10];  \
-    d = ks[4 * r + 11]
-
-#define k_get(r,a,b,c,d)    \
-    ks[4 * r +  8] = a;  \
-    ks[4 * r +  9] = b;  \
-    ks[4 * r + 10] = c;  \
-    ks[4 * r + 11] = d
-
-/* the linear transformation and its inverse    */
-
-#define rot(a,b,c,d)    \
-    a = rotl32(a, 13);    \
-    c = rotl32(c, 3);     \
-    d ^= c ^ (a << 3);  \
-    b ^= a ^ c;         \
-    d = rotl32(d, 7);     \
-    b = rotl32(b, 1);     \
-    a ^= b ^ d;         \
-    c ^= d ^ (b << 7);  \
-    a = rotl32(a, 5);     \
-    c = rotl32(c, 22)
-
-#define irot(a,b,c,d)   \
-    c = rotr32(c, 22);    \
-    a = rotr32(a, 5);     \
-    c ^= d ^ (b << 7);  \
-    a ^= b ^ d;         \
-    d = rotr32(d, 7);     \
-    b = rotr32(b, 1);     \
-    d ^= c ^ (a << 3);  \
-    b ^= a ^ c;         \
-    c = rotr32(c, 3);     \
-    a = rotr32(a, 13)
-
-static void serpent256_set_key (u32 *ks, const u32 *ukey)
-{
-  #pragma unroll
-  for (int i = 0; i < 8; i++)
-  {
-    ks[i] = ukey[i];
-  }
-
-  #pragma unroll
-  for (int i = 0; i < 132; i++)
-  {
-    ks[i + 8] = rotl32 (ks[i + 7] ^ ks[i + 5] ^ ks[i + 3] ^ ks[i + 0] ^ 0x9e3779b9 ^ i, 11);
-  }
-
-  u32  a,b,c,d,e,f,g,h;
-  u32  t1,t2,t3,t4,t5,t6,t7,t8,t9,t10,t11,t12,t13,t14,t15,t16;
-
-  k_set( 0,a,b,c,d); sb3(a,b,c,d,e,f,g,h); k_get( 0,e,f,g,h);
-  k_set( 1,a,b,c,d); sb2(a,b,c,d,e,f,g,h); k_get( 1,e,f,g,h);
-  k_set( 2,a,b,c,d); sb1(a,b,c,d,e,f,g,h); k_get( 2,e,f,g,h);
-  k_set( 3,a,b,c,d); sb0(a,b,c,d,e,f,g,h); k_get( 3,e,f,g,h);
-  k_set( 4,a,b,c,d); sb7(a,b,c,d,e,f,g,h); k_get( 4,e,f,g,h);
-  k_set( 5,a,b,c,d); sb6(a,b,c,d,e,f,g,h); k_get( 5,e,f,g,h);
-  k_set( 6,a,b,c,d); sb5(a,b,c,d,e,f,g,h); k_get( 6,e,f,g,h);
-  k_set( 7,a,b,c,d); sb4(a,b,c,d,e,f,g,h); k_get( 7,e,f,g,h);
-  k_set( 8,a,b,c,d); sb3(a,b,c,d,e,f,g,h); k_get( 8,e,f,g,h);
-  k_set( 9,a,b,c,d); sb2(a,b,c,d,e,f,g,h); k_get( 9,e,f,g,h);
-  k_set(10,a,b,c,d); sb1(a,b,c,d,e,f,g,h); k_get(10,e,f,g,h);
-  k_set(11,a,b,c,d); sb0(a,b,c,d,e,f,g,h); k_get(11,e,f,g,h);
-  k_set(12,a,b,c,d); sb7(a,b,c,d,e,f,g,h); k_get(12,e,f,g,h);
-  k_set(13,a,b,c,d); sb6(a,b,c,d,e,f,g,h); k_get(13,e,f,g,h);
-  k_set(14,a,b,c,d); sb5(a,b,c,d,e,f,g,h); k_get(14,e,f,g,h);
-  k_set(15,a,b,c,d); sb4(a,b,c,d,e,f,g,h); k_get(15,e,f,g,h);
-  k_set(16,a,b,c,d); sb3(a,b,c,d,e,f,g,h); k_get(16,e,f,g,h);
-  k_set(17,a,b,c,d); sb2(a,b,c,d,e,f,g,h); k_get(17,e,f,g,h);
-  k_set(18,a,b,c,d); sb1(a,b,c,d,e,f,g,h); k_get(18,e,f,g,h);
-  k_set(19,a,b,c,d); sb0(a,b,c,d,e,f,g,h); k_get(19,e,f,g,h);
-  k_set(20,a,b,c,d); sb7(a,b,c,d,e,f,g,h); k_get(20,e,f,g,h);
-  k_set(21,a,b,c,d); sb6(a,b,c,d,e,f,g,h); k_get(21,e,f,g,h);
-  k_set(22,a,b,c,d); sb5(a,b,c,d,e,f,g,h); k_get(22,e,f,g,h);
-  k_set(23,a,b,c,d); sb4(a,b,c,d,e,f,g,h); k_get(23,e,f,g,h);
-  k_set(24,a,b,c,d); sb3(a,b,c,d,e,f,g,h); k_get(24,e,f,g,h);
-  k_set(25,a,b,c,d); sb2(a,b,c,d,e,f,g,h); k_get(25,e,f,g,h);
-  k_set(26,a,b,c,d); sb1(a,b,c,d,e,f,g,h); k_get(26,e,f,g,h);
-  k_set(27,a,b,c,d); sb0(a,b,c,d,e,f,g,h); k_get(27,e,f,g,h);
-  k_set(28,a,b,c,d); sb7(a,b,c,d,e,f,g,h); k_get(28,e,f,g,h);
-  k_set(29,a,b,c,d); sb6(a,b,c,d,e,f,g,h); k_get(29,e,f,g,h);
-  k_set(30,a,b,c,d); sb5(a,b,c,d,e,f,g,h); k_get(30,e,f,g,h);
-  k_set(31,a,b,c,d); sb4(a,b,c,d,e,f,g,h); k_get(31,e,f,g,h);
-  k_set(32,a,b,c,d); sb3(a,b,c,d,e,f,g,h); k_get(32,e,f,g,h);
-}
-
-static void serpent256_encrypt (const u32 *ks, const u32 *in, u32 *out)
-{
-  u32  a,b,c,d,e,f,g,h;
-  u32  t1,t2,t3,t4,t5,t6,t7,t8,t9,t10,t11,t12,t13,t14,t15,t16;
-
-  a = in[0];
-  b = in[1];
-  c = in[2];
-  d = in[3];
-
-  k_xor( 0,a,b,c,d); sb0(a,b,c,d,e,f,g,h); rot(e,f,g,h);
-  k_xor( 1,e,f,g,h); sb1(e,f,g,h,a,b,c,d); rot(a,b,c,d);
-  k_xor( 2,a,b,c,d); sb2(a,b,c,d,e,f,g,h); rot(e,f,g,h);
-  k_xor( 3,e,f,g,h); sb3(e,f,g,h,a,b,c,d); rot(a,b,c,d);
-  k_xor( 4,a,b,c,d); sb4(a,b,c,d,e,f,g,h); rot(e,f,g,h);
-  k_xor( 5,e,f,g,h); sb5(e,f,g,h,a,b,c,d); rot(a,b,c,d);
-  k_xor( 6,a,b,c,d); sb6(a,b,c,d,e,f,g,h); rot(e,f,g,h);
-  k_xor( 7,e,f,g,h); sb7(e,f,g,h,a,b,c,d); rot(a,b,c,d);
-  k_xor( 8,a,b,c,d); sb0(a,b,c,d,e,f,g,h); rot(e,f,g,h);
-  k_xor( 9,e,f,g,h); sb1(e,f,g,h,a,b,c,d); rot(a,b,c,d);
-  k_xor(10,a,b,c,d); sb2(a,b,c,d,e,f,g,h); rot(e,f,g,h);
-  k_xor(11,e,f,g,h); sb3(e,f,g,h,a,b,c,d); rot(a,b,c,d);
-  k_xor(12,a,b,c,d); sb4(a,b,c,d,e,f,g,h); rot(e,f,g,h);
-  k_xor(13,e,f,g,h); sb5(e,f,g,h,a,b,c,d); rot(a,b,c,d);
-  k_xor(14,a,b,c,d); sb6(a,b,c,d,e,f,g,h); rot(e,f,g,h);
-  k_xor(15,e,f,g,h); sb7(e,f,g,h,a,b,c,d); rot(a,b,c,d);
-  k_xor(16,a,b,c,d); sb0(a,b,c,d,e,f,g,h); rot(e,f,g,h);
-  k_xor(17,e,f,g,h); sb1(e,f,g,h,a,b,c,d); rot(a,b,c,d);
-  k_xor(18,a,b,c,d); sb2(a,b,c,d,e,f,g,h); rot(e,f,g,h);
-  k_xor(19,e,f,g,h); sb3(e,f,g,h,a,b,c,d); rot(a,b,c,d);
-  k_xor(20,a,b,c,d); sb4(a,b,c,d,e,f,g,h); rot(e,f,g,h);
-  k_xor(21,e,f,g,h); sb5(e,f,g,h,a,b,c,d); rot(a,b,c,d);
-  k_xor(22,a,b,c,d); sb6(a,b,c,d,e,f,g,h); rot(e,f,g,h);
-  k_xor(23,e,f,g,h); sb7(e,f,g,h,a,b,c,d); rot(a,b,c,d);
-  k_xor(24,a,b,c,d); sb0(a,b,c,d,e,f,g,h); rot(e,f,g,h);
-  k_xor(25,e,f,g,h); sb1(e,f,g,h,a,b,c,d); rot(a,b,c,d);
-  k_xor(26,a,b,c,d); sb2(a,b,c,d,e,f,g,h); rot(e,f,g,h);
-  k_xor(27,e,f,g,h); sb3(e,f,g,h,a,b,c,d); rot(a,b,c,d);
-  k_xor(28,a,b,c,d); sb4(a,b,c,d,e,f,g,h); rot(e,f,g,h);
-  k_xor(29,e,f,g,h); sb5(e,f,g,h,a,b,c,d); rot(a,b,c,d);
-  k_xor(30,a,b,c,d); sb6(a,b,c,d,e,f,g,h); rot(e,f,g,h);
-  k_xor(31,e,f,g,h); sb7(e,f,g,h,a,b,c,d);
-  k_xor(32,a,b,c,d);
-
-  out[0] = a;
-  out[1] = b;
-  out[2] = c;
-  out[3] = d;
-}
-
-static void serpent256_decrypt (const u32 *ks, const u32 *in, u32 *out)
-{
-  u32  a,b,c,d,e,f,g,h;
-  u32  t1,t2,t3,t4,t5,t6,t7,t8,t9,t10,t11,t12,t13,t14,t15,t16;
-
-  a = in[0];
-  b = in[1];
-  c = in[2];
-  d = in[3];
-
-                                       k_xor(32,a,b,c,d);
-                 ib7(a,b,c,d,e,f,g,h); k_xor(31,e,f,g,h);
-  irot(e,f,g,h); ib6(e,f,g,h,a,b,c,d); k_xor(30,a,b,c,d);
-  irot(a,b,c,d); ib5(a,b,c,d,e,f,g,h); k_xor(29,e,f,g,h);
-  irot(e,f,g,h); ib4(e,f,g,h,a,b,c,d); k_xor(28,a,b,c,d);
-  irot(a,b,c,d); ib3(a,b,c,d,e,f,g,h); k_xor(27,e,f,g,h);
-  irot(e,f,g,h); ib2(e,f,g,h,a,b,c,d); k_xor(26,a,b,c,d);
-  irot(a,b,c,d); ib1(a,b,c,d,e,f,g,h); k_xor(25,e,f,g,h);
-  irot(e,f,g,h); ib0(e,f,g,h,a,b,c,d); k_xor(24,a,b,c,d);
-  irot(a,b,c,d); ib7(a,b,c,d,e,f,g,h); k_xor(23,e,f,g,h);
-  irot(e,f,g,h); ib6(e,f,g,h,a,b,c,d); k_xor(22,a,b,c,d);
-  irot(a,b,c,d); ib5(a,b,c,d,e,f,g,h); k_xor(21,e,f,g,h);
-  irot(e,f,g,h); ib4(e,f,g,h,a,b,c,d); k_xor(20,a,b,c,d);
-  irot(a,b,c,d); ib3(a,b,c,d,e,f,g,h); k_xor(19,e,f,g,h);
-  irot(e,f,g,h); ib2(e,f,g,h,a,b,c,d); k_xor(18,a,b,c,d);
-  irot(a,b,c,d); ib1(a,b,c,d,e,f,g,h); k_xor(17,e,f,g,h);
-  irot(e,f,g,h); ib0(e,f,g,h,a,b,c,d); k_xor(16,a,b,c,d);
-  irot(a,b,c,d); ib7(a,b,c,d,e,f,g,h); k_xor(15,e,f,g,h);
-  irot(e,f,g,h); ib6(e,f,g,h,a,b,c,d); k_xor(14,a,b,c,d);
-  irot(a,b,c,d); ib5(a,b,c,d,e,f,g,h); k_xor(13,e,f,g,h);
-  irot(e,f,g,h); ib4(e,f,g,h,a,b,c,d); k_xor(12,a,b,c,d);
-  irot(a,b,c,d); ib3(a,b,c,d,e,f,g,h); k_xor(11,e,f,g,h);
-  irot(e,f,g,h); ib2(e,f,g,h,a,b,c,d); k_xor(10,a,b,c,d);
-  irot(a,b,c,d); ib1(a,b,c,d,e,f,g,h); k_xor( 9,e,f,g,h);
-  irot(e,f,g,h); ib0(e,f,g,h,a,b,c,d); k_xor( 8,a,b,c,d);
-  irot(a,b,c,d); ib7(a,b,c,d,e,f,g,h); k_xor( 7,e,f,g,h);
-  irot(e,f,g,h); ib6(e,f,g,h,a,b,c,d); k_xor( 6,a,b,c,d);
-  irot(a,b,c,d); ib5(a,b,c,d,e,f,g,h); k_xor( 5,e,f,g,h);
-  irot(e,f,g,h); ib4(e,f,g,h,a,b,c,d); k_xor( 4,a,b,c,d);
-  irot(a,b,c,d); ib3(a,b,c,d,e,f,g,h); k_xor( 3,e,f,g,h);
-  irot(e,f,g,h); ib2(e,f,g,h,a,b,c,d); k_xor( 2,a,b,c,d);
-  irot(a,b,c,d); ib1(a,b,c,d,e,f,g,h); k_xor( 1,e,f,g,h);
-  irot(e,f,g,h); ib0(e,f,g,h,a,b,c,d); k_xor( 0,a,b,c,d);
-
-  out[0] = a;
-  out[1] = b;
-  out[2] = c;
-  out[3] = d;
-}
-
-static void serpent256_decrypt_xts (const u32 *ukey1, const u32 *ukey2, const u32 *in, u32 *out)
-{
-  u32 T[4] = { 0 };
-  u32 Z[4] = { 0 };
-
-  out[0] = in[0];
-  out[1] = in[1];
-  out[2] = in[2];
-  out[3] = in[3];
-
-  u32 ks[140];
-
-  serpent256_set_key (ks, ukey2);
-  serpent256_encrypt (ks, Z, T);
-
-  out[0] ^= T[0];
-  out[1] ^= T[1];
-  out[2] ^= T[2];
-  out[3] ^= T[3];
-
-  serpent256_set_key (ks, ukey1);
-  serpent256_decrypt (ks, out, out);
-
-  out[0] ^= T[0];
-  out[1] ^= T[1];
-  out[2] ^= T[2];
-  out[3] ^= T[3];
-}
diff --git a/OpenCL/kernel_twofish256.c b/OpenCL/kernel_twofish256.c
new file mode 100644
index 0000000..ef8a943
--- /dev/null
+++ b/OpenCL/kernel_twofish256.c
@@ -0,0 +1,466 @@
+/* This is an independent implementation of the encryption algorithm:   */
+/*                                                                      */
+/*         Twofish by Bruce Schneier and colleagues                     */
+/*                                                                      */
+/* which is a candidate algorithm in the Advanced Encryption Standard   */
+/* programme of the US National Institute of Standards and Technology.  */
+/*                                                                      */
+/* Copyright in this implementation is held by Dr B R Gladman but I     */
+/* hereby give permission for its free direct or derivative use subject */
+/* to acknowledgment of its origin and compliance with any conditions   */
+/* that the originators of t he algorithm place on its exploitation.    */
+/*                                                                      */
+/* My thanks to Doug Whiting and Niels Ferguson for comments that led   */
+/* to improvements in this implementation.                              */
+/*                                                                      */
+/* Dr Brian Gladman (gladman@seven77.demon.co.uk) 14th January 1999     */
+/*                                                                      */
+/* -------------------------------------------------------------------- */
+/*                                                                      */
+/* Cleaned and optimized for GPU use with oclHashcat by Jens Steube     */
+
+#define extract_byte(x,n) (((x) >> (8 * (n))) & 0xff)
+
+__constant u32 q_tab[2][256] =
+{
+  {
+    0xA9, 0x67, 0xB3, 0xE8, 0x04, 0xFD, 0xA3, 0x76, 0x9A, 0x92, 0x80, 0x78,
+    0xE4, 0xDD, 0xD1, 0x38, 0x0D, 0xC6, 0x35, 0x98, 0x18, 0xF7, 0xEC, 0x6C,
+    0x43, 0x75, 0x37, 0x26, 0xFA, 0x13, 0x94, 0x48, 0xF2, 0xD0, 0x8B, 0x30,
+    0x84, 0x54, 0xDF, 0x23, 0x19, 0x5B, 0x3D, 0x59, 0xF3, 0xAE, 0xA2, 0x82,
+    0x63, 0x01, 0x83, 0x2E, 0xD9, 0x51, 0x9B, 0x7C, 0xA6, 0xEB, 0xA5, 0xBE,
+    0x16, 0x0C, 0xE3, 0x61, 0xC0, 0x8C, 0x3A, 0xF5, 0x73, 0x2C, 0x25, 0x0B,
+    0xBB, 0x4E, 0x89, 0x6B, 0x53, 0x6A, 0xB4, 0xF1, 0xE1, 0xE6, 0xBD, 0x45,
+    0xE2, 0xF4, 0xB6, 0x66, 0xCC, 0x95, 0x03, 0x56, 0xD4, 0x1C, 0x1E, 0xD7,
+    0xFB, 0xC3, 0x8E, 0xB5, 0xE9, 0xCF, 0xBF, 0xBA, 0xEA, 0x77, 0x39, 0xAF,
+    0x33, 0xC9, 0x62, 0x71, 0x81, 0x79, 0x09, 0xAD, 0x24, 0xCD, 0xF9, 0xD8,
+    0xE5, 0xC5, 0xB9, 0x4D, 0x44, 0x08, 0x86, 0xE7, 0xA1, 0x1D, 0xAA, 0xED,
+    0x06, 0x70, 0xB2, 0xD2, 0x41, 0x7B, 0xA0, 0x11, 0x31, 0xC2, 0x27, 0x90,
+    0x20, 0xF6, 0x60, 0xFF, 0x96, 0x5C, 0xB1, 0xAB, 0x9E, 0x9C, 0x52, 0x1B,
+    0x5F, 0x93, 0x0A, 0xEF, 0x91, 0x85, 0x49, 0xEE, 0x2D, 0x4F, 0x8F, 0x3B,
+    0x47, 0x87, 0x6D, 0x46, 0xD6, 0x3E, 0x69, 0x64, 0x2A, 0xCE, 0xCB, 0x2F,
+    0xFC, 0x97, 0x05, 0x7A, 0xAC, 0x7F, 0xD5, 0x1A, 0x4B, 0x0E, 0xA7, 0x5A,
+    0x28, 0x14, 0x3F, 0x29, 0x88, 0x3C, 0x4C, 0x02, 0xB8, 0xDA, 0xB0, 0x17,
+    0x55, 0x1F, 0x8A, 0x7D, 0x57, 0xC7, 0x8D, 0x74, 0xB7, 0xC4, 0x9F, 0x72,
+    0x7E, 0x15, 0x22, 0x12, 0x58, 0x07, 0x99, 0x34, 0x6E, 0x50, 0xDE, 0x68,
+    0x65, 0xBC, 0xDB, 0xF8, 0xC8, 0xA8, 0x2B, 0x40, 0xDC, 0xFE, 0x32, 0xA4,
+    0xCA, 0x10, 0x21, 0xF0, 0xD3, 0x5D, 0x0F, 0x00, 0x6F, 0x9D, 0x36, 0x42,
+    0x4A, 0x5E, 0xC1, 0xE0
+  },
+  {
+    0x75, 0xF3, 0xC6, 0xF4, 0xDB, 0x7B, 0xFB, 0xC8, 0x4A, 0xD3, 0xE6, 0x6B,
+    0x45, 0x7D, 0xE8, 0x4B, 0xD6, 0x32, 0xD8, 0xFD, 0x37, 0x71, 0xF1, 0xE1,
+    0x30, 0x0F, 0xF8, 0x1B, 0x87, 0xFA, 0x06, 0x3F, 0x5E, 0xBA, 0xAE, 0x5B,
+    0x8A, 0x00, 0xBC, 0x9D, 0x6D, 0xC1, 0xB1, 0x0E, 0x80, 0x5D, 0xD2, 0xD5,
+    0xA0, 0x84, 0x07, 0x14, 0xB5, 0x90, 0x2C, 0xA3, 0xB2, 0x73, 0x4C, 0x54,
+    0x92, 0x74, 0x36, 0x51, 0x38, 0xB0, 0xBD, 0x5A, 0xFC, 0x60, 0x62, 0x96,
+    0x6C, 0x42, 0xF7, 0x10, 0x7C, 0x28, 0x27, 0x8C, 0x13, 0x95, 0x9C, 0xC7,
+    0x24, 0x46, 0x3B, 0x70, 0xCA, 0xE3, 0x85, 0xCB, 0x11, 0xD0, 0x93, 0xB8,
+    0xA6, 0x83, 0x20, 0xFF, 0x9F, 0x77, 0xC3, 0xCC, 0x03, 0x6F, 0x08, 0xBF,
+    0x40, 0xE7, 0x2B, 0xE2, 0x79, 0x0C, 0xAA, 0x82, 0x41, 0x3A, 0xEA, 0xB9,
+    0xE4, 0x9A, 0xA4, 0x97, 0x7E, 0xDA, 0x7A, 0x17, 0x66, 0x94, 0xA1, 0x1D,
+    0x3D, 0xF0, 0xDE, 0xB3, 0x0B, 0x72, 0xA7, 0x1C, 0xEF, 0xD1, 0x53, 0x3E,
+    0x8F, 0x33, 0x26, 0x5F, 0xEC, 0x76, 0x2A, 0x49, 0x81, 0x88, 0xEE, 0x21,
+    0xC4, 0x1A, 0xEB, 0xD9, 0xC5, 0x39, 0x99, 0xCD, 0xAD, 0x31, 0x8B, 0x01,
+    0x18, 0x23, 0xDD, 0x1F, 0x4E, 0x2D, 0xF9, 0x48, 0x4F, 0xF2, 0x65, 0x8E,
+    0x78, 0x5C, 0x58, 0x19, 0x8D, 0xE5, 0x98, 0x57, 0x67, 0x7F, 0x05, 0x64,
+    0xAF, 0x63, 0xB6, 0xFE, 0xF5, 0xB7, 0x3C, 0xA5, 0xCE, 0xE9, 0x68, 0x44,
+    0xE0, 0x4D, 0x43, 0x69, 0x29, 0x2E, 0xAC, 0x15, 0x59, 0xA8, 0x0A, 0x9E,
+    0x6E, 0x47, 0xDF, 0x34, 0x35, 0x6A, 0xCF, 0xDC, 0x22, 0xC9, 0xC0, 0x9B,
+    0x89, 0xD4, 0xED, 0xAB, 0x12, 0xA2, 0x0D, 0x52, 0xBB, 0x02, 0x2F, 0xA9,
+    0xD7, 0x61, 0x1E, 0xB4, 0x50, 0x04, 0xF6, 0xC2, 0x16, 0x25, 0x86, 0x56,
+    0x55, 0x09, 0xBE, 0x91
+  }
+};
+
+#define q(n,x) q_tab[n][x]
+
+__constant u32  m_tab[4][256] =
+{
+  { 0xBCBC3275, 0xECEC21F3, 0x202043C6, 0xB3B3C9F4, 0xDADA03DB, 0x02028B7B,
+    0xE2E22BFB, 0x9E9EFAC8, 0xC9C9EC4A, 0xD4D409D3, 0x18186BE6, 0x1E1E9F6B,
+    0x98980E45, 0xB2B2387D, 0xA6A6D2E8, 0x2626B74B, 0x3C3C57D6, 0x93938A32,
+    0x8282EED8, 0x525298FD, 0x7B7BD437, 0xBBBB3771, 0x5B5B97F1, 0x474783E1,
+    0x24243C30, 0x5151E20F, 0xBABAC6F8, 0x4A4AF31B, 0xBFBF4887, 0x0D0D70FA,
+    0xB0B0B306, 0x7575DE3F, 0xD2D2FD5E, 0x7D7D20BA, 0x666631AE, 0x3A3AA35B,
+    0x59591C8A, 0x00000000, 0xCDCD93BC, 0x1A1AE09D, 0xAEAE2C6D, 0x7F7FABC1,
+    0x2B2BC7B1, 0xBEBEB90E, 0xE0E0A080, 0x8A8A105D, 0x3B3B52D2, 0x6464BAD5,
+    0xD8D888A0, 0xE7E7A584, 0x5F5FE807, 0x1B1B1114, 0x2C2CC2B5, 0xFCFCB490,
+    0x3131272C, 0x808065A3, 0x73732AB2, 0x0C0C8173, 0x79795F4C, 0x6B6B4154,
+    0x4B4B0292, 0x53536974, 0x94948F36, 0x83831F51, 0x2A2A3638, 0xC4C49CB0,
+    0x2222C8BD, 0xD5D5F85A, 0xBDBDC3FC, 0x48487860, 0xFFFFCE62, 0x4C4C0796,
+    0x4141776C, 0xC7C7E642, 0xEBEB24F7, 0x1C1C1410, 0x5D5D637C, 0x36362228,
+    0x6767C027, 0xE9E9AF8C, 0x4444F913, 0x1414EA95, 0xF5F5BB9C, 0xCFCF18C7,
+    0x3F3F2D24, 0xC0C0E346, 0x7272DB3B, 0x54546C70, 0x29294CCA, 0xF0F035E3,
+    0x0808FE85, 0xC6C617CB, 0xF3F34F11, 0x8C8CE4D0, 0xA4A45993, 0xCACA96B8,
+    0x68683BA6, 0xB8B84D83, 0x38382820, 0xE5E52EFF, 0xADAD569F, 0x0B0B8477,
+    0xC8C81DC3, 0x9999FFCC, 0x5858ED03, 0x19199A6F, 0x0E0E0A08, 0x95957EBF,
+    0x70705040, 0xF7F730E7, 0x6E6ECF2B, 0x1F1F6EE2, 0xB5B53D79, 0x09090F0C,
+    0x616134AA, 0x57571682, 0x9F9F0B41, 0x9D9D803A, 0x111164EA, 0x2525CDB9,
+    0xAFAFDDE4, 0x4545089A, 0xDFDF8DA4, 0xA3A35C97, 0xEAEAD57E, 0x353558DA,
+    0xEDEDD07A, 0x4343FC17, 0xF8F8CB66, 0xFBFBB194, 0x3737D3A1, 0xFAFA401D,
+    0xC2C2683D, 0xB4B4CCF0, 0x32325DDE, 0x9C9C71B3, 0x5656E70B, 0xE3E3DA72,
+    0x878760A7, 0x15151B1C, 0xF9F93AEF, 0x6363BFD1, 0x3434A953, 0x9A9A853E,
+    0xB1B1428F, 0x7C7CD133, 0x88889B26, 0x3D3DA65F, 0xA1A1D7EC, 0xE4E4DF76,
+    0x8181942A, 0x91910149, 0x0F0FFB81, 0xEEEEAA88, 0x161661EE, 0xD7D77321,
+    0x9797F5C4, 0xA5A5A81A, 0xFEFE3FEB, 0x6D6DB5D9, 0x7878AEC5, 0xC5C56D39,
+    0x1D1DE599, 0x7676A4CD, 0x3E3EDCAD, 0xCBCB6731, 0xB6B6478B, 0xEFEF5B01,
+    0x12121E18, 0x6060C523, 0x6A6AB0DD, 0x4D4DF61F, 0xCECEE94E, 0xDEDE7C2D,
+    0x55559DF9, 0x7E7E5A48, 0x2121B24F, 0x03037AF2, 0xA0A02665, 0x5E5E198E,
+    0x5A5A6678, 0x65654B5C, 0x62624E58, 0xFDFD4519, 0x0606F48D, 0x404086E5,
+    0xF2F2BE98, 0x3333AC57, 0x17179067, 0x05058E7F, 0xE8E85E05, 0x4F4F7D64,
+    0x89896AAF, 0x10109563, 0x74742FB6, 0x0A0A75FE, 0x5C5C92F5, 0x9B9B74B7,
+    0x2D2D333C, 0x3030D6A5, 0x2E2E49CE, 0x494989E9, 0x46467268, 0x77775544,
+    0xA8A8D8E0, 0x9696044D, 0x2828BD43, 0xA9A92969, 0xD9D97929, 0x8686912E,
+    0xD1D187AC, 0xF4F44A15, 0x8D8D1559, 0xD6D682A8, 0xB9B9BC0A, 0x42420D9E,
+    0xF6F6C16E, 0x2F2FB847, 0xDDDD06DF, 0x23233934, 0xCCCC6235, 0xF1F1C46A,
+    0xC1C112CF, 0x8585EBDC, 0x8F8F9E22, 0x7171A1C9, 0x9090F0C0, 0xAAAA539B,
+    0x0101F189, 0x8B8BE1D4, 0x4E4E8CED, 0x8E8E6FAB, 0xABABA212, 0x6F6F3EA2,
+    0xE6E6540D, 0xDBDBF252, 0x92927BBB, 0xB7B7B602, 0x6969CA2F, 0x3939D9A9,
+    0xD3D30CD7, 0xA7A72361, 0xA2A2AD1E, 0xC3C399B4, 0x6C6C4450, 0x07070504,
+    0x04047FF6, 0x272746C2, 0xACACA716, 0xD0D07625, 0x50501386, 0xDCDCF756,
+    0x84841A55, 0xE1E15109, 0x7A7A25BE, 0x1313EF91 },
+
+  { 0xA9D93939, 0x67901717, 0xB3719C9C, 0xE8D2A6A6, 0x04050707, 0xFD985252,
+    0xA3658080, 0x76DFE4E4, 0x9A084545, 0x92024B4B, 0x80A0E0E0, 0x78665A5A,
+    0xE4DDAFAF, 0xDDB06A6A, 0xD1BF6363, 0x38362A2A, 0x0D54E6E6, 0xC6432020,
+    0x3562CCCC, 0x98BEF2F2, 0x181E1212, 0xF724EBEB, 0xECD7A1A1, 0x6C774141,
+    0x43BD2828, 0x7532BCBC, 0x37D47B7B, 0x269B8888, 0xFA700D0D, 0x13F94444,
+    0x94B1FBFB, 0x485A7E7E, 0xF27A0303, 0xD0E48C8C, 0x8B47B6B6, 0x303C2424,
+    0x84A5E7E7, 0x54416B6B, 0xDF06DDDD, 0x23C56060, 0x1945FDFD, 0x5BA33A3A,
+    0x3D68C2C2, 0x59158D8D, 0xF321ECEC, 0xAE316666, 0xA23E6F6F, 0x82165757,
+    0x63951010, 0x015BEFEF, 0x834DB8B8, 0x2E918686, 0xD9B56D6D, 0x511F8383,
+    0x9B53AAAA, 0x7C635D5D, 0xA63B6868, 0xEB3FFEFE, 0xA5D63030, 0xBE257A7A,
+    0x16A7ACAC, 0x0C0F0909, 0xE335F0F0, 0x6123A7A7, 0xC0F09090, 0x8CAFE9E9,
+    0x3A809D9D, 0xF5925C5C, 0x73810C0C, 0x2C273131, 0x2576D0D0, 0x0BE75656,
+    0xBB7B9292, 0x4EE9CECE, 0x89F10101, 0x6B9F1E1E, 0x53A93434, 0x6AC4F1F1,
+    0xB499C3C3, 0xF1975B5B, 0xE1834747, 0xE66B1818, 0xBDC82222, 0x450E9898,
+    0xE26E1F1F, 0xF4C9B3B3, 0xB62F7474, 0x66CBF8F8, 0xCCFF9999, 0x95EA1414,
+    0x03ED5858, 0x56F7DCDC, 0xD4E18B8B, 0x1C1B1515, 0x1EADA2A2, 0xD70CD3D3,
+    0xFB2BE2E2, 0xC31DC8C8, 0x8E195E5E, 0xB5C22C2C, 0xE9894949, 0xCF12C1C1,
+    0xBF7E9595, 0xBA207D7D, 0xEA641111, 0x77840B0B, 0x396DC5C5, 0xAF6A8989,
+    0x33D17C7C, 0xC9A17171, 0x62CEFFFF, 0x7137BBBB, 0x81FB0F0F, 0x793DB5B5,
+    0x0951E1E1, 0xADDC3E3E, 0x242D3F3F, 0xCDA47676, 0xF99D5555, 0xD8EE8282,
+    0xE5864040, 0xC5AE7878, 0xB9CD2525, 0x4D049696, 0x44557777, 0x080A0E0E,
+    0x86135050, 0xE730F7F7, 0xA1D33737, 0x1D40FAFA, 0xAA346161, 0xED8C4E4E,
+    0x06B3B0B0, 0x706C5454, 0xB22A7373, 0xD2523B3B, 0x410B9F9F, 0x7B8B0202,
+    0xA088D8D8, 0x114FF3F3, 0x3167CBCB, 0xC2462727, 0x27C06767, 0x90B4FCFC,
+    0x20283838, 0xF67F0404, 0x60784848, 0xFF2EE5E5, 0x96074C4C, 0x5C4B6565,
+    0xB1C72B2B, 0xAB6F8E8E, 0x9E0D4242, 0x9CBBF5F5, 0x52F2DBDB, 0x1BF34A4A,
+    0x5FA63D3D, 0x9359A4A4, 0x0ABCB9B9, 0xEF3AF9F9, 0x91EF1313, 0x85FE0808,
+    0x49019191, 0xEE611616, 0x2D7CDEDE, 0x4FB22121, 0x8F42B1B1, 0x3BDB7272,
+    0x47B82F2F, 0x8748BFBF, 0x6D2CAEAE, 0x46E3C0C0, 0xD6573C3C, 0x3E859A9A,
+    0x6929A9A9, 0x647D4F4F, 0x2A948181, 0xCE492E2E, 0xCB17C6C6, 0x2FCA6969,
+    0xFCC3BDBD, 0x975CA3A3, 0x055EE8E8, 0x7AD0EDED, 0xAC87D1D1, 0x7F8E0505,
+    0xD5BA6464, 0x1AA8A5A5, 0x4BB72626, 0x0EB9BEBE, 0xA7608787, 0x5AF8D5D5,
+    0x28223636, 0x14111B1B, 0x3FDE7575, 0x2979D9D9, 0x88AAEEEE, 0x3C332D2D,
+    0x4C5F7979, 0x02B6B7B7, 0xB896CACA, 0xDA583535, 0xB09CC4C4, 0x17FC4343,
+    0x551A8484, 0x1FF64D4D, 0x8A1C5959, 0x7D38B2B2, 0x57AC3333, 0xC718CFCF,
+    0x8DF40606, 0x74695353, 0xB7749B9B, 0xC4F59797, 0x9F56ADAD, 0x72DAE3E3,
+    0x7ED5EAEA, 0x154AF4F4, 0x229E8F8F, 0x12A2ABAB, 0x584E6262, 0x07E85F5F,
+    0x99E51D1D, 0x34392323, 0x6EC1F6F6, 0x50446C6C, 0xDE5D3232, 0x68724646,
+    0x6526A0A0, 0xBC93CDCD, 0xDB03DADA, 0xF8C6BABA, 0xC8FA9E9E, 0xA882D6D6,
+    0x2BCF6E6E, 0x40507070, 0xDCEB8585, 0xFE750A0A, 0x328A9393, 0xA48DDFDF,
+    0xCA4C2929, 0x10141C1C, 0x2173D7D7, 0xF0CCB4B4, 0xD309D4D4, 0x5D108A8A,
+    0x0FE25151, 0x00000000, 0x6F9A1919, 0x9DE01A1A, 0x368F9494, 0x42E6C7C7,
+    0x4AECC9C9, 0x5EFDD2D2, 0xC1AB7F7F, 0xE0D8A8A8 },
+
+  { 0xBC75BC32, 0xECF3EC21, 0x20C62043, 0xB3F4B3C9, 0xDADBDA03, 0x027B028B,
+    0xE2FBE22B, 0x9EC89EFA, 0xC94AC9EC, 0xD4D3D409, 0x18E6186B, 0x1E6B1E9F,
+    0x9845980E, 0xB27DB238, 0xA6E8A6D2, 0x264B26B7, 0x3CD63C57, 0x9332938A,
+    0x82D882EE, 0x52FD5298, 0x7B377BD4, 0xBB71BB37, 0x5BF15B97, 0x47E14783,
+    0x2430243C, 0x510F51E2, 0xBAF8BAC6, 0x4A1B4AF3, 0xBF87BF48, 0x0DFA0D70,
+    0xB006B0B3, 0x753F75DE, 0xD25ED2FD, 0x7DBA7D20, 0x66AE6631, 0x3A5B3AA3,
+    0x598A591C, 0x00000000, 0xCDBCCD93, 0x1A9D1AE0, 0xAE6DAE2C, 0x7FC17FAB,
+    0x2BB12BC7, 0xBE0EBEB9, 0xE080E0A0, 0x8A5D8A10, 0x3BD23B52, 0x64D564BA,
+    0xD8A0D888, 0xE784E7A5, 0x5F075FE8, 0x1B141B11, 0x2CB52CC2, 0xFC90FCB4,
+    0x312C3127, 0x80A38065, 0x73B2732A, 0x0C730C81, 0x794C795F, 0x6B546B41,
+    0x4B924B02, 0x53745369, 0x9436948F, 0x8351831F, 0x2A382A36, 0xC4B0C49C,
+    0x22BD22C8, 0xD55AD5F8, 0xBDFCBDC3, 0x48604878, 0xFF62FFCE, 0x4C964C07,
+    0x416C4177, 0xC742C7E6, 0xEBF7EB24, 0x1C101C14, 0x5D7C5D63, 0x36283622,
+    0x672767C0, 0xE98CE9AF, 0x441344F9, 0x149514EA, 0xF59CF5BB, 0xCFC7CF18,
+    0x3F243F2D, 0xC046C0E3, 0x723B72DB, 0x5470546C, 0x29CA294C, 0xF0E3F035,
+    0x088508FE, 0xC6CBC617, 0xF311F34F, 0x8CD08CE4, 0xA493A459, 0xCAB8CA96,
+    0x68A6683B, 0xB883B84D, 0x38203828, 0xE5FFE52E, 0xAD9FAD56, 0x0B770B84,
+    0xC8C3C81D, 0x99CC99FF, 0x580358ED, 0x196F199A, 0x0E080E0A, 0x95BF957E,
+    0x70407050, 0xF7E7F730, 0x6E2B6ECF, 0x1FE21F6E, 0xB579B53D, 0x090C090F,
+    0x61AA6134, 0x57825716, 0x9F419F0B, 0x9D3A9D80, 0x11EA1164, 0x25B925CD,
+    0xAFE4AFDD, 0x459A4508, 0xDFA4DF8D, 0xA397A35C, 0xEA7EEAD5, 0x35DA3558,
+    0xED7AEDD0, 0x431743FC, 0xF866F8CB, 0xFB94FBB1, 0x37A137D3, 0xFA1DFA40,
+    0xC23DC268, 0xB4F0B4CC, 0x32DE325D, 0x9CB39C71, 0x560B56E7, 0xE372E3DA,
+    0x87A78760, 0x151C151B, 0xF9EFF93A, 0x63D163BF, 0x345334A9, 0x9A3E9A85,
+    0xB18FB142, 0x7C337CD1, 0x8826889B, 0x3D5F3DA6, 0xA1ECA1D7, 0xE476E4DF,
+    0x812A8194, 0x91499101, 0x0F810FFB, 0xEE88EEAA, 0x16EE1661, 0xD721D773,
+    0x97C497F5, 0xA51AA5A8, 0xFEEBFE3F, 0x6DD96DB5, 0x78C578AE, 0xC539C56D,
+    0x1D991DE5, 0x76CD76A4, 0x3EAD3EDC, 0xCB31CB67, 0xB68BB647, 0xEF01EF5B,
+    0x1218121E, 0x602360C5, 0x6ADD6AB0, 0x4D1F4DF6, 0xCE4ECEE9, 0xDE2DDE7C,
+    0x55F9559D, 0x7E487E5A, 0x214F21B2, 0x03F2037A, 0xA065A026, 0x5E8E5E19,
+    0x5A785A66, 0x655C654B, 0x6258624E, 0xFD19FD45, 0x068D06F4, 0x40E54086,
+    0xF298F2BE, 0x335733AC, 0x17671790, 0x057F058E, 0xE805E85E, 0x4F644F7D,
+    0x89AF896A, 0x10631095, 0x74B6742F, 0x0AFE0A75, 0x5CF55C92, 0x9BB79B74,
+    0x2D3C2D33, 0x30A530D6, 0x2ECE2E49, 0x49E94989, 0x46684672, 0x77447755,
+    0xA8E0A8D8, 0x964D9604, 0x284328BD, 0xA969A929, 0xD929D979, 0x862E8691,
+    0xD1ACD187, 0xF415F44A, 0x8D598D15, 0xD6A8D682, 0xB90AB9BC, 0x429E420D,
+    0xF66EF6C1, 0x2F472FB8, 0xDDDFDD06, 0x23342339, 0xCC35CC62, 0xF16AF1C4,
+    0xC1CFC112, 0x85DC85EB, 0x8F228F9E, 0x71C971A1, 0x90C090F0, 0xAA9BAA53,
+    0x018901F1, 0x8BD48BE1, 0x4EED4E8C, 0x8EAB8E6F, 0xAB12ABA2, 0x6FA26F3E,
+    0xE60DE654, 0xDB52DBF2, 0x92BB927B, 0xB702B7B6, 0x692F69CA, 0x39A939D9,
+    0xD3D7D30C, 0xA761A723, 0xA21EA2AD, 0xC3B4C399, 0x6C506C44, 0x07040705,
+    0x04F6047F, 0x27C22746, 0xAC16ACA7, 0xD025D076, 0x50865013, 0xDC56DCF7,
+    0x8455841A, 0xE109E151, 0x7ABE7A25, 0x139113EF },
+
+  { 0xD939A9D9, 0x90176790, 0x719CB371, 0xD2A6E8D2, 0x05070405, 0x9852FD98,
+    0x6580A365, 0xDFE476DF, 0x08459A08, 0x024B9202, 0xA0E080A0, 0x665A7866,
+    0xDDAFE4DD, 0xB06ADDB0, 0xBF63D1BF, 0x362A3836, 0x54E60D54, 0x4320C643,
+    0x62CC3562, 0xBEF298BE, 0x1E12181E, 0x24EBF724, 0xD7A1ECD7, 0x77416C77,
+    0xBD2843BD, 0x32BC7532, 0xD47B37D4, 0x9B88269B, 0x700DFA70, 0xF94413F9,
+    0xB1FB94B1, 0x5A7E485A, 0x7A03F27A, 0xE48CD0E4, 0x47B68B47, 0x3C24303C,
+    0xA5E784A5, 0x416B5441, 0x06DDDF06, 0xC56023C5, 0x45FD1945, 0xA33A5BA3,
+    0x68C23D68, 0x158D5915, 0x21ECF321, 0x3166AE31, 0x3E6FA23E, 0x16578216,
+    0x95106395, 0x5BEF015B, 0x4DB8834D, 0x91862E91, 0xB56DD9B5, 0x1F83511F,
+    0x53AA9B53, 0x635D7C63, 0x3B68A63B, 0x3FFEEB3F, 0xD630A5D6, 0x257ABE25,
+    0xA7AC16A7, 0x0F090C0F, 0x35F0E335, 0x23A76123, 0xF090C0F0, 0xAFE98CAF,
+    0x809D3A80, 0x925CF592, 0x810C7381, 0x27312C27, 0x76D02576, 0xE7560BE7,
+    0x7B92BB7B, 0xE9CE4EE9, 0xF10189F1, 0x9F1E6B9F, 0xA93453A9, 0xC4F16AC4,
+    0x99C3B499, 0x975BF197, 0x8347E183, 0x6B18E66B, 0xC822BDC8, 0x0E98450E,
+    0x6E1FE26E, 0xC9B3F4C9, 0x2F74B62F, 0xCBF866CB, 0xFF99CCFF, 0xEA1495EA,
+    0xED5803ED, 0xF7DC56F7, 0xE18BD4E1, 0x1B151C1B, 0xADA21EAD, 0x0CD3D70C,
+    0x2BE2FB2B, 0x1DC8C31D, 0x195E8E19, 0xC22CB5C2, 0x8949E989, 0x12C1CF12,
+    0x7E95BF7E, 0x207DBA20, 0x6411EA64, 0x840B7784, 0x6DC5396D, 0x6A89AF6A,
+    0xD17C33D1, 0xA171C9A1, 0xCEFF62CE, 0x37BB7137, 0xFB0F81FB, 0x3DB5793D,
+    0x51E10951, 0xDC3EADDC, 0x2D3F242D, 0xA476CDA4, 0x9D55F99D, 0xEE82D8EE,
+    0x8640E586, 0xAE78C5AE, 0xCD25B9CD, 0x04964D04, 0x55774455, 0x0A0E080A,
+    0x13508613, 0x30F7E730, 0xD337A1D3, 0x40FA1D40, 0x3461AA34, 0x8C4EED8C,
+    0xB3B006B3, 0x6C54706C, 0x2A73B22A, 0x523BD252, 0x0B9F410B, 0x8B027B8B,
+    0x88D8A088, 0x4FF3114F, 0x67CB3167, 0x4627C246, 0xC06727C0, 0xB4FC90B4,
+    0x28382028, 0x7F04F67F, 0x78486078, 0x2EE5FF2E, 0x074C9607, 0x4B655C4B,
+    0xC72BB1C7, 0x6F8EAB6F, 0x0D429E0D, 0xBBF59CBB, 0xF2DB52F2, 0xF34A1BF3,
+    0xA63D5FA6, 0x59A49359, 0xBCB90ABC, 0x3AF9EF3A, 0xEF1391EF, 0xFE0885FE,
+    0x01914901, 0x6116EE61, 0x7CDE2D7C, 0xB2214FB2, 0x42B18F42, 0xDB723BDB,
+    0xB82F47B8, 0x48BF8748, 0x2CAE6D2C, 0xE3C046E3, 0x573CD657, 0x859A3E85,
+    0x29A96929, 0x7D4F647D, 0x94812A94, 0x492ECE49, 0x17C6CB17, 0xCA692FCA,
+    0xC3BDFCC3, 0x5CA3975C, 0x5EE8055E, 0xD0ED7AD0, 0x87D1AC87, 0x8E057F8E,
+    0xBA64D5BA, 0xA8A51AA8, 0xB7264BB7, 0xB9BE0EB9, 0x6087A760, 0xF8D55AF8,
+    0x22362822, 0x111B1411, 0xDE753FDE, 0x79D92979, 0xAAEE88AA, 0x332D3C33,
+    0x5F794C5F, 0xB6B702B6, 0x96CAB896, 0x5835DA58, 0x9CC4B09C, 0xFC4317FC,
+    0x1A84551A, 0xF64D1FF6, 0x1C598A1C, 0x38B27D38, 0xAC3357AC, 0x18CFC718,
+    0xF4068DF4, 0x69537469, 0x749BB774, 0xF597C4F5, 0x56AD9F56, 0xDAE372DA,
+    0xD5EA7ED5, 0x4AF4154A, 0x9E8F229E, 0xA2AB12A2, 0x4E62584E, 0xE85F07E8,
+    0xE51D99E5, 0x39233439, 0xC1F66EC1, 0x446C5044, 0x5D32DE5D, 0x72466872,
+    0x26A06526, 0x93CDBC93, 0x03DADB03, 0xC6BAF8C6, 0xFA9EC8FA, 0x82D6A882,
+    0xCF6E2BCF, 0x50704050, 0xEB85DCEB, 0x750AFE75, 0x8A93328A, 0x8DDFA48D,
+    0x4C29CA4C, 0x141C1014, 0x73D72173, 0xCCB4F0CC, 0x09D4D309, 0x108A5D10,
+    0xE2510FE2, 0x00000000, 0x9A196F9A, 0xE01A9DE0, 0x8F94368F, 0xE6C742E6,
+    0xECC94AEC, 0xFDD25EFD, 0xAB7FC1AB, 0xD8A8E0D8 }
+};
+
+#define mds(n,x) m_tab[n][x]
+
+static u32 h_fun (u32 *sk, u32 *lk, const u32 x, const u32 *key)
+{
+  u32  b0, b1, b2, b3;
+
+  b0 = extract_byte (x, 0);
+  b1 = extract_byte (x, 1);
+  b2 = extract_byte (x, 2);
+  b3 = extract_byte (x, 3);
+
+  b0 = q (1, b0) ^ extract_byte (key[3], 0);
+  b1 = q (0, b1) ^ extract_byte (key[3], 1);
+  b2 = q (0, b2) ^ extract_byte (key[3], 2);
+  b3 = q (1, b3) ^ extract_byte (key[3], 3);
+
+  b0 = q (1, b0) ^ extract_byte (key[2], 0);
+  b1 = q (1, b1) ^ extract_byte (key[2], 1);
+  b2 = q (0, b2) ^ extract_byte (key[2], 2);
+  b3 = q (0, b3) ^ extract_byte (key[2], 3);
+
+  b0 = q (0, (q (0, b0) ^ extract_byte (key[1], 0))) ^ extract_byte (key[0], 0);
+  b1 = q (0, (q (1, b1) ^ extract_byte (key[1], 1))) ^ extract_byte (key[0], 1);
+  b2 = q (1, (q (0, b2) ^ extract_byte (key[1], 2))) ^ extract_byte (key[0], 2);
+  b3 = q (1, (q (1, b3) ^ extract_byte (key[1], 3))) ^ extract_byte (key[0], 3);
+
+  return mds (0, b0) ^ mds (1, b1) ^ mds (2, b2) ^ mds (3, b3);
+}
+
+#define q40(x,k) q (0, q (0, q (1, q (1, x) ^ extract_byte (k[3], 0)) ^ extract_byte (k[2], 0)) ^ extract_byte (k[1], 0)) ^ extract_byte (k[0], 0)
+#define q41(x,k) q (0, q (1, q (1, q (0, x) ^ extract_byte (k[3], 1)) ^ extract_byte (k[2], 1)) ^ extract_byte (k[1], 1)) ^ extract_byte (k[0], 1)
+#define q42(x,k) q (1, q (0, q (0, q (0, x) ^ extract_byte (k[3], 2)) ^ extract_byte (k[2], 2)) ^ extract_byte (k[1], 2)) ^ extract_byte (k[0], 2)
+#define q43(x,k) q (1, q (1, q (0, q (1, x) ^ extract_byte (k[3], 3)) ^ extract_byte (k[2], 3)) ^ extract_byte (k[1], 3)) ^ extract_byte (k[0], 3)
+
+#define g1_fun(x)                                   \
+  (mds (0, q40 (extract_byte (x, 3), sk)) ^ \
+   mds (1, q41 (extract_byte (x, 0), sk)) ^ \
+   mds (2, q42 (extract_byte (x, 1), sk)) ^ \
+   mds (3, q43 (extract_byte (x, 2), sk)))
+
+#define g0_fun(x)                                   \
+  (mds (0, q40 (extract_byte (x, 0), sk)) ^ \
+   mds (1, q41 (extract_byte (x, 1), sk)) ^ \
+   mds (2, q42 (extract_byte (x, 2), sk)) ^ \
+   mds (3, q43 (extract_byte (x, 3), sk)))
+
+static u32 mds_rem (u32 p0, u32 p1)
+{
+  #define G_MOD 0x14d
+
+  for (int i = 0; i < 8; i++)
+  {
+    u32 t = p1 >> 24;
+
+    p1 = (p1 << 8) | (p0 >> 24);
+
+    p0 <<= 8;
+
+    u32 u = (t << 1);
+
+    if (t & 0x80) u ^= G_MOD;
+
+    p1 ^= t ^ (u << 16);
+
+    u ^= (t >> 1);
+
+    if (t & 0x01) u ^= G_MOD >> 1;
+
+    p1 ^= (u << 24) | (u << 8);
+  }
+
+  return p1;
+}
+
+static void twofish256_set_key (u32 *sk, u32 *lk, const u32 *ukey)
+{
+  u32 me_key[4];
+
+  me_key[0] = ukey[0];
+  me_key[1] = ukey[2];
+  me_key[2] = ukey[4];
+  me_key[3] = ukey[6];
+
+  u32 mo_key[4];
+
+  mo_key[0] = ukey[1];
+  mo_key[1] = ukey[3];
+  mo_key[2] = ukey[5];
+  mo_key[3] = ukey[7];
+
+  sk[3] = mds_rem (me_key[0], mo_key[0]);
+  sk[2] = mds_rem (me_key[1], mo_key[1]);
+  sk[1] = mds_rem (me_key[2], mo_key[2]);
+  sk[0] = mds_rem (me_key[3], mo_key[3]);
+
+  for (int i = 0; i < 40; i += 2)
+  {
+    u32 a = 0x01010101 * i;
+    u32 b = 0x01010101 + a;
+
+    a = h_fun (sk, lk, a, me_key);
+    b = h_fun (sk, lk, b, mo_key);
+
+    b = rotl32 (b, 8);
+
+    lk[i + 0] = a + b;
+    lk[i + 1] = rotl32 (a + 2 * b, 9);
+  }
+}
+
+#define f_rnd(i)                                                    \
+{                                                                   \
+  u32 t0 = g0_fun (data[0]);                                       \
+  u32 t1 = g1_fun (data[1]);                                       \
+  data[2] = rotr32 (data[2] ^ (t0 + t1 + lk[4 * (i) + 8]), 1);      \
+  data[3] = rotl32 (data[3], 1) ^ (t0 + 2 * t1 + lk[4 * (i) + 9]);  \
+  u32 t2 = g0_fun (data[2]);                                       \
+  u32 t3 = g1_fun (data[3]);                                       \
+  data[0] = rotr32 (data[0] ^ (t2 + t3 + lk[4 * (i) + 10]), 1);     \
+  data[1] = rotl32 (data[1], 1) ^ (t2 + 2 * t3 + lk[4 * (i) + 11]); \
+}
+
+static void twofish256_encrypt (const u32 *sk, const u32 *lk, const u32 *in, u32 *out)
+{
+  u32 data[4];
+
+  data[0] = in[0] ^ lk[0];
+  data[1] = in[1] ^ lk[1];
+  data[2] = in[2] ^ lk[2];
+  data[3] = in[3] ^ lk[3];
+
+  f_rnd (0);
+  f_rnd (1);
+  f_rnd (2);
+  f_rnd (3);
+  f_rnd (4);
+  f_rnd (5);
+  f_rnd (6);
+  f_rnd (7);
+
+  out[0] = data[2] ^ lk[4];
+  out[1] = data[3] ^ lk[5];
+  out[2] = data[0] ^ lk[6];
+  out[3] = data[1] ^ lk[7];
+}
+
+#define i_rnd(i)                                                    \
+{                                                                   \
+  u32 t0 = g0_fun (data[0]);                                       \
+  u32 t1 = g1_fun (data[1]);                                       \
+  data[2] = rotl32 (data[2], 1) ^ (t0 + t1 + lk[4 * (i) + 10]);     \
+  data[3] = rotr32 (data[3] ^ (t0 + 2 * t1 + lk[4 * (i) + 11]), 1); \
+  u32 t2 = g0_fun (data[2]);                                       \
+  u32 t3 = g1_fun (data[3]);                                       \
+  data[0] = rotl32 (data[0], 1) ^ (t2 + t3 + lk[4 * (i) +  8]);     \
+  data[1] = rotr32 (data[1] ^ (t2 + 2 * t3 + lk[4 * (i) +  9]), 1); \
+}
+
+static void twofish256_decrypt (const u32 *sk, const u32 *lk, const u32 *in, u32 *out)
+{
+  u32 data[4];
+
+  data[0] = in[0] ^ lk[4];
+  data[1] = in[1] ^ lk[5];
+  data[2] = in[2] ^ lk[6];
+  data[3] = in[3] ^ lk[7];
+
+  i_rnd (7);
+  i_rnd (6);
+  i_rnd (5);
+  i_rnd (4);
+  i_rnd (3);
+  i_rnd (2);
+  i_rnd (1);
+  i_rnd (0);
+
+  out[0] = data[2] ^ lk[0];
+  out[1] = data[3] ^ lk[1];
+  out[2] = data[0] ^ lk[2];
+  out[3] = data[1] ^ lk[3];
+}
+
+static void twofish256_decrypt_xts (const u32 *ukey1, const u32 *ukey2, const u32 *in, u32 *out)
+{
+  u32 T[4] = { 0 };
+  u32 Z[4] = { 0 };
+
+  out[0] = in[0];
+  out[1] = in[1];
+  out[2] = in[2];
+  out[3] = in[3];
+
+  u32 sk[4]; u32 lk[40];
+
+  twofish256_set_key (sk, lk, ukey2);
+  twofish256_encrypt (sk, lk, Z, T);
+
+  out[0] ^= T[0];
+  out[1] ^= T[1];
+  out[2] ^= T[2];
+  out[3] ^= T[3];
+
+  twofish256_set_key (sk, lk, ukey1);
+  twofish256_decrypt (sk, lk, out, out);
+
+  out[0] ^= T[0];
+  out[1] ^= T[1];
+  out[2] ^= T[2];
+  out[3] ^= T[3];
+}
diff --git a/OpenCL/kernel_twofish256_amd.c b/OpenCL/kernel_twofish256_amd.c
deleted file mode 100644
index ef8a943..0000000
--- a/OpenCL/kernel_twofish256_amd.c
+++ /dev/null
@@ -1,466 +0,0 @@
-/* This is an independent implementation of the encryption algorithm:   */
-/*                                                                      */
-/*         Twofish by Bruce Schneier and colleagues                     */
-/*                                                                      */
-/* which is a candidate algorithm in the Advanced Encryption Standard   */
-/* programme of the US National Institute of Standards and Technology.  */
-/*                                                                      */
-/* Copyright in this implementation is held by Dr B R Gladman but I     */
-/* hereby give permission for its free direct or derivative use subject */
-/* to acknowledgment of its origin and compliance with any conditions   */
-/* that the originators of t he algorithm place on its exploitation.    */
-/*                                                                      */
-/* My thanks to Doug Whiting and Niels Ferguson for comments that led   */
-/* to improvements in this implementation.                              */
-/*                                                                      */
-/* Dr Brian Gladman (gladman@seven77.demon.co.uk) 14th January 1999     */
-/*                                                                      */
-/* -------------------------------------------------------------------- */
-/*                                                                      */
-/* Cleaned and optimized for GPU use with oclHashcat by Jens Steube     */
-
-#define extract_byte(x,n) (((x) >> (8 * (n))) & 0xff)
-
-__constant u32 q_tab[2][256] =
-{
-  {
-    0xA9, 0x67, 0xB3, 0xE8, 0x04, 0xFD, 0xA3, 0x76, 0x9A, 0x92, 0x80, 0x78,
-    0xE4, 0xDD, 0xD1, 0x38, 0x0D, 0xC6, 0x35, 0x98, 0x18, 0xF7, 0xEC, 0x6C,
-    0x43, 0x75, 0x37, 0x26, 0xFA, 0x13, 0x94, 0x48, 0xF2, 0xD0, 0x8B, 0x30,
-    0x84, 0x54, 0xDF, 0x23, 0x19, 0x5B, 0x3D, 0x59, 0xF3, 0xAE, 0xA2, 0x82,
-    0x63, 0x01, 0x83, 0x2E, 0xD9, 0x51, 0x9B, 0x7C, 0xA6, 0xEB, 0xA5, 0xBE,
-    0x16, 0x0C, 0xE3, 0x61, 0xC0, 0x8C, 0x3A, 0xF5, 0x73, 0x2C, 0x25, 0x0B,
-    0xBB, 0x4E, 0x89, 0x6B, 0x53, 0x6A, 0xB4, 0xF1, 0xE1, 0xE6, 0xBD, 0x45,
-    0xE2, 0xF4, 0xB6, 0x66, 0xCC, 0x95, 0x03, 0x56, 0xD4, 0x1C, 0x1E, 0xD7,
-    0xFB, 0xC3, 0x8E, 0xB5, 0xE9, 0xCF, 0xBF, 0xBA, 0xEA, 0x77, 0x39, 0xAF,
-    0x33, 0xC9, 0x62, 0x71, 0x81, 0x79, 0x09, 0xAD, 0x24, 0xCD, 0xF9, 0xD8,
-    0xE5, 0xC5, 0xB9, 0x4D, 0x44, 0x08, 0x86, 0xE7, 0xA1, 0x1D, 0xAA, 0xED,
-    0x06, 0x70, 0xB2, 0xD2, 0x41, 0x7B, 0xA0, 0x11, 0x31, 0xC2, 0x27, 0x90,
-    0x20, 0xF6, 0x60, 0xFF, 0x96, 0x5C, 0xB1, 0xAB, 0x9E, 0x9C, 0x52, 0x1B,
-    0x5F, 0x93, 0x0A, 0xEF, 0x91, 0x85, 0x49, 0xEE, 0x2D, 0x4F, 0x8F, 0x3B,
-    0x47, 0x87, 0x6D, 0x46, 0xD6, 0x3E, 0x69, 0x64, 0x2A, 0xCE, 0xCB, 0x2F,
-    0xFC, 0x97, 0x05, 0x7A, 0xAC, 0x7F, 0xD5, 0x1A, 0x4B, 0x0E, 0xA7, 0x5A,
-    0x28, 0x14, 0x3F, 0x29, 0x88, 0x3C, 0x4C, 0x02, 0xB8, 0xDA, 0xB0, 0x17,
-    0x55, 0x1F, 0x8A, 0x7D, 0x57, 0xC7, 0x8D, 0x74, 0xB7, 0xC4, 0x9F, 0x72,
-    0x7E, 0x15, 0x22, 0x12, 0x58, 0x07, 0x99, 0x34, 0x6E, 0x50, 0xDE, 0x68,
-    0x65, 0xBC, 0xDB, 0xF8, 0xC8, 0xA8, 0x2B, 0x40, 0xDC, 0xFE, 0x32, 0xA4,
-    0xCA, 0x10, 0x21, 0xF0, 0xD3, 0x5D, 0x0F, 0x00, 0x6F, 0x9D, 0x36, 0x42,
-    0x4A, 0x5E, 0xC1, 0xE0
-  },
-  {
-    0x75, 0xF3, 0xC6, 0xF4, 0xDB, 0x7B, 0xFB, 0xC8, 0x4A, 0xD3, 0xE6, 0x6B,
-    0x45, 0x7D, 0xE8, 0x4B, 0xD6, 0x32, 0xD8, 0xFD, 0x37, 0x71, 0xF1, 0xE1,
-    0x30, 0x0F, 0xF8, 0x1B, 0x87, 0xFA, 0x06, 0x3F, 0x5E, 0xBA, 0xAE, 0x5B,
-    0x8A, 0x00, 0xBC, 0x9D, 0x6D, 0xC1, 0xB1, 0x0E, 0x80, 0x5D, 0xD2, 0xD5,
-    0xA0, 0x84, 0x07, 0x14, 0xB5, 0x90, 0x2C, 0xA3, 0xB2, 0x73, 0x4C, 0x54,
-    0x92, 0x74, 0x36, 0x51, 0x38, 0xB0, 0xBD, 0x5A, 0xFC, 0x60, 0x62, 0x96,
-    0x6C, 0x42, 0xF7, 0x10, 0x7C, 0x28, 0x27, 0x8C, 0x13, 0x95, 0x9C, 0xC7,
-    0x24, 0x46, 0x3B, 0x70, 0xCA, 0xE3, 0x85, 0xCB, 0x11, 0xD0, 0x93, 0xB8,
-    0xA6, 0x83, 0x20, 0xFF, 0x9F, 0x77, 0xC3, 0xCC, 0x03, 0x6F, 0x08, 0xBF,
-    0x40, 0xE7, 0x2B, 0xE2, 0x79, 0x0C, 0xAA, 0x82, 0x41, 0x3A, 0xEA, 0xB9,
-    0xE4, 0x9A, 0xA4, 0x97, 0x7E, 0xDA, 0x7A, 0x17, 0x66, 0x94, 0xA1, 0x1D,
-    0x3D, 0xF0, 0xDE, 0xB3, 0x0B, 0x72, 0xA7, 0x1C, 0xEF, 0xD1, 0x53, 0x3E,
-    0x8F, 0x33, 0x26, 0x5F, 0xEC, 0x76, 0x2A, 0x49, 0x81, 0x88, 0xEE, 0x21,
-    0xC4, 0x1A, 0xEB, 0xD9, 0xC5, 0x39, 0x99, 0xCD, 0xAD, 0x31, 0x8B, 0x01,
-    0x18, 0x23, 0xDD, 0x1F, 0x4E, 0x2D, 0xF9, 0x48, 0x4F, 0xF2, 0x65, 0x8E,
-    0x78, 0x5C, 0x58, 0x19, 0x8D, 0xE5, 0x98, 0x57, 0x67, 0x7F, 0x05, 0x64,
-    0xAF, 0x63, 0xB6, 0xFE, 0xF5, 0xB7, 0x3C, 0xA5, 0xCE, 0xE9, 0x68, 0x44,
-    0xE0, 0x4D, 0x43, 0x69, 0x29, 0x2E, 0xAC, 0x15, 0x59, 0xA8, 0x0A, 0x9E,
-    0x6E, 0x47, 0xDF, 0x34, 0x35, 0x6A, 0xCF, 0xDC, 0x22, 0xC9, 0xC0, 0x9B,
-    0x89, 0xD4, 0xED, 0xAB, 0x12, 0xA2, 0x0D, 0x52, 0xBB, 0x02, 0x2F, 0xA9,
-    0xD7, 0x61, 0x1E, 0xB4, 0x50, 0x04, 0xF6, 0xC2, 0x16, 0x25, 0x86, 0x56,
-    0x55, 0x09, 0xBE, 0x91
-  }
-};
-
-#define q(n,x) q_tab[n][x]
-
-__constant u32  m_tab[4][256] =
-{
-  { 0xBCBC3275, 0xECEC21F3, 0x202043C6, 0xB3B3C9F4, 0xDADA03DB, 0x02028B7B,
-    0xE2E22BFB, 0x9E9EFAC8, 0xC9C9EC4A, 0xD4D409D3, 0x18186BE6, 0x1E1E9F6B,
-    0x98980E45, 0xB2B2387D, 0xA6A6D2E8, 0x2626B74B, 0x3C3C57D6, 0x93938A32,
-    0x8282EED8, 0x525298FD, 0x7B7BD437, 0xBBBB3771, 0x5B5B97F1, 0x474783E1,
-    0x24243C30, 0x5151E20F, 0xBABAC6F8, 0x4A4AF31B, 0xBFBF4887, 0x0D0D70FA,
-    0xB0B0B306, 0x7575DE3F, 0xD2D2FD5E, 0x7D7D20BA, 0x666631AE, 0x3A3AA35B,
-    0x59591C8A, 0x00000000, 0xCDCD93BC, 0x1A1AE09D, 0xAEAE2C6D, 0x7F7FABC1,
-    0x2B2BC7B1, 0xBEBEB90E, 0xE0E0A080, 0x8A8A105D, 0x3B3B52D2, 0x6464BAD5,
-    0xD8D888A0, 0xE7E7A584, 0x5F5FE807, 0x1B1B1114, 0x2C2CC2B5, 0xFCFCB490,
-    0x3131272C, 0x808065A3, 0x73732AB2, 0x0C0C8173, 0x79795F4C, 0x6B6B4154,
-    0x4B4B0292, 0x53536974, 0x94948F36, 0x83831F51, 0x2A2A3638, 0xC4C49CB0,
-    0x2222C8BD, 0xD5D5F85A, 0xBDBDC3FC, 0x48487860, 0xFFFFCE62, 0x4C4C0796,
-    0x4141776C, 0xC7C7E642, 0xEBEB24F7, 0x1C1C1410, 0x5D5D637C, 0x36362228,
-    0x6767C027, 0xE9E9AF8C, 0x4444F913, 0x1414EA95, 0xF5F5BB9C, 0xCFCF18C7,
-    0x3F3F2D24, 0xC0C0E346, 0x7272DB3B, 0x54546C70, 0x29294CCA, 0xF0F035E3,
-    0x0808FE85, 0xC6C617CB, 0xF3F34F11, 0x8C8CE4D0, 0xA4A45993, 0xCACA96B8,
-    0x68683BA6, 0xB8B84D83, 0x38382820, 0xE5E52EFF, 0xADAD569F, 0x0B0B8477,
-    0xC8C81DC3, 0x9999FFCC, 0x5858ED03, 0x19199A6F, 0x0E0E0A08, 0x95957EBF,
-    0x70705040, 0xF7F730E7, 0x6E6ECF2B, 0x1F1F6EE2, 0xB5B53D79, 0x09090F0C,
-    0x616134AA, 0x57571682, 0x9F9F0B41, 0x9D9D803A, 0x111164EA, 0x2525CDB9,
-    0xAFAFDDE4, 0x4545089A, 0xDFDF8DA4, 0xA3A35C97, 0xEAEAD57E, 0x353558DA,
-    0xEDEDD07A, 0x4343FC17, 0xF8F8CB66, 0xFBFBB194, 0x3737D3A1, 0xFAFA401D,
-    0xC2C2683D, 0xB4B4CCF0, 0x32325DDE, 0x9C9C71B3, 0x5656E70B, 0xE3E3DA72,
-    0x878760A7, 0x15151B1C, 0xF9F93AEF, 0x6363BFD1, 0x3434A953, 0x9A9A853E,
-    0xB1B1428F, 0x7C7CD133, 0x88889B26, 0x3D3DA65F, 0xA1A1D7EC, 0xE4E4DF76,
-    0x8181942A, 0x91910149, 0x0F0FFB81, 0xEEEEAA88, 0x161661EE, 0xD7D77321,
-    0x9797F5C4, 0xA5A5A81A, 0xFEFE3FEB, 0x6D6DB5D9, 0x7878AEC5, 0xC5C56D39,
-    0x1D1DE599, 0x7676A4CD, 0x3E3EDCAD, 0xCBCB6731, 0xB6B6478B, 0xEFEF5B01,
-    0x12121E18, 0x6060C523, 0x6A6AB0DD, 0x4D4DF61F, 0xCECEE94E, 0xDEDE7C2D,
-    0x55559DF9, 0x7E7E5A48, 0x2121B24F, 0x03037AF2, 0xA0A02665, 0x5E5E198E,
-    0x5A5A6678, 0x65654B5C, 0x62624E58, 0xFDFD4519, 0x0606F48D, 0x404086E5,
-    0xF2F2BE98, 0x3333AC57, 0x17179067, 0x05058E7F, 0xE8E85E05, 0x4F4F7D64,
-    0x89896AAF, 0x10109563, 0x74742FB6, 0x0A0A75FE, 0x5C5C92F5, 0x9B9B74B7,
-    0x2D2D333C, 0x3030D6A5, 0x2E2E49CE, 0x494989E9, 0x46467268, 0x77775544,
-    0xA8A8D8E0, 0x9696044D, 0x2828BD43, 0xA9A92969, 0xD9D97929, 0x8686912E,
-    0xD1D187AC, 0xF4F44A15, 0x8D8D1559, 0xD6D682A8, 0xB9B9BC0A, 0x42420D9E,
-    0xF6F6C16E, 0x2F2FB847, 0xDDDD06DF, 0x23233934, 0xCCCC6235, 0xF1F1C46A,
-    0xC1C112CF, 0x8585EBDC, 0x8F8F9E22, 0x7171A1C9, 0x9090F0C0, 0xAAAA539B,
-    0x0101F189, 0x8B8BE1D4, 0x4E4E8CED, 0x8E8E6FAB, 0xABABA212, 0x6F6F3EA2,
-    0xE6E6540D, 0xDBDBF252, 0x92927BBB, 0xB7B7B602, 0x6969CA2F, 0x3939D9A9,
-    0xD3D30CD7, 0xA7A72361, 0xA2A2AD1E, 0xC3C399B4, 0x6C6C4450, 0x07070504,
-    0x04047FF6, 0x272746C2, 0xACACA716, 0xD0D07625, 0x50501386, 0xDCDCF756,
-    0x84841A55, 0xE1E15109, 0x7A7A25BE, 0x1313EF91 },
-
-  { 0xA9D93939, 0x67901717, 0xB3719C9C, 0xE8D2A6A6, 0x04050707, 0xFD985252,
-    0xA3658080, 0x76DFE4E4, 0x9A084545, 0x92024B4B, 0x80A0E0E0, 0x78665A5A,
-    0xE4DDAFAF, 0xDDB06A6A, 0xD1BF6363, 0x38362A2A, 0x0D54E6E6, 0xC6432020,
-    0x3562CCCC, 0x98BEF2F2, 0x181E1212, 0xF724EBEB, 0xECD7A1A1, 0x6C774141,
-    0x43BD2828, 0x7532BCBC, 0x37D47B7B, 0x269B8888, 0xFA700D0D, 0x13F94444,
-    0x94B1FBFB, 0x485A7E7E, 0xF27A0303, 0xD0E48C8C, 0x8B47B6B6, 0x303C2424,
-    0x84A5E7E7, 0x54416B6B, 0xDF06DDDD, 0x23C56060, 0x1945FDFD, 0x5BA33A3A,
-    0x3D68C2C2, 0x59158D8D, 0xF321ECEC, 0xAE316666, 0xA23E6F6F, 0x82165757,
-    0x63951010, 0x015BEFEF, 0x834DB8B8, 0x2E918686, 0xD9B56D6D, 0x511F8383,
-    0x9B53AAAA, 0x7C635D5D, 0xA63B6868, 0xEB3FFEFE, 0xA5D63030, 0xBE257A7A,
-    0x16A7ACAC, 0x0C0F0909, 0xE335F0F0, 0x6123A7A7, 0xC0F09090, 0x8CAFE9E9,
-    0x3A809D9D, 0xF5925C5C, 0x73810C0C, 0x2C273131, 0x2576D0D0, 0x0BE75656,
-    0xBB7B9292, 0x4EE9CECE, 0x89F10101, 0x6B9F1E1E, 0x53A93434, 0x6AC4F1F1,
-    0xB499C3C3, 0xF1975B5B, 0xE1834747, 0xE66B1818, 0xBDC82222, 0x450E9898,
-    0xE26E1F1F, 0xF4C9B3B3, 0xB62F7474, 0x66CBF8F8, 0xCCFF9999, 0x95EA1414,
-    0x03ED5858, 0x56F7DCDC, 0xD4E18B8B, 0x1C1B1515, 0x1EADA2A2, 0xD70CD3D3,
-    0xFB2BE2E2, 0xC31DC8C8, 0x8E195E5E, 0xB5C22C2C, 0xE9894949, 0xCF12C1C1,
-    0xBF7E9595, 0xBA207D7D, 0xEA641111, 0x77840B0B, 0x396DC5C5, 0xAF6A8989,
-    0x33D17C7C, 0xC9A17171, 0x62CEFFFF, 0x7137BBBB, 0x81FB0F0F, 0x793DB5B5,
-    0x0951E1E1, 0xADDC3E3E, 0x242D3F3F, 0xCDA47676, 0xF99D5555, 0xD8EE8282,
-    0xE5864040, 0xC5AE7878, 0xB9CD2525, 0x4D049696, 0x44557777, 0x080A0E0E,
-    0x86135050, 0xE730F7F7, 0xA1D33737, 0x1D40FAFA, 0xAA346161, 0xED8C4E4E,
-    0x06B3B0B0, 0x706C5454, 0xB22A7373, 0xD2523B3B, 0x410B9F9F, 0x7B8B0202,
-    0xA088D8D8, 0x114FF3F3, 0x3167CBCB, 0xC2462727, 0x27C06767, 0x90B4FCFC,
-    0x20283838, 0xF67F0404, 0x60784848, 0xFF2EE5E5, 0x96074C4C, 0x5C4B6565,
-    0xB1C72B2B, 0xAB6F8E8E, 0x9E0D4242, 0x9CBBF5F5, 0x52F2DBDB, 0x1BF34A4A,
-    0x5FA63D3D, 0x9359A4A4, 0x0ABCB9B9, 0xEF3AF9F9, 0x91EF1313, 0x85FE0808,
-    0x49019191, 0xEE611616, 0x2D7CDEDE, 0x4FB22121, 0x8F42B1B1, 0x3BDB7272,
-    0x47B82F2F, 0x8748BFBF, 0x6D2CAEAE, 0x46E3C0C0, 0xD6573C3C, 0x3E859A9A,
-    0x6929A9A9, 0x647D4F4F, 0x2A948181, 0xCE492E2E, 0xCB17C6C6, 0x2FCA6969,
-    0xFCC3BDBD, 0x975CA3A3, 0x055EE8E8, 0x7AD0EDED, 0xAC87D1D1, 0x7F8E0505,
-    0xD5BA6464, 0x1AA8A5A5, 0x4BB72626, 0x0EB9BEBE, 0xA7608787, 0x5AF8D5D5,
-    0x28223636, 0x14111B1B, 0x3FDE7575, 0x2979D9D9, 0x88AAEEEE, 0x3C332D2D,
-    0x4C5F7979, 0x02B6B7B7, 0xB896CACA, 0xDA583535, 0xB09CC4C4, 0x17FC4343,
-    0x551A8484, 0x1FF64D4D, 0x8A1C5959, 0x7D38B2B2, 0x57AC3333, 0xC718CFCF,
-    0x8DF40606, 0x74695353, 0xB7749B9B, 0xC4F59797, 0x9F56ADAD, 0x72DAE3E3,
-    0x7ED5EAEA, 0x154AF4F4, 0x229E8F8F, 0x12A2ABAB, 0x584E6262, 0x07E85F5F,
-    0x99E51D1D, 0x34392323, 0x6EC1F6F6, 0x50446C6C, 0xDE5D3232, 0x68724646,
-    0x6526A0A0, 0xBC93CDCD, 0xDB03DADA, 0xF8C6BABA, 0xC8FA9E9E, 0xA882D6D6,
-    0x2BCF6E6E, 0x40507070, 0xDCEB8585, 0xFE750A0A, 0x328A9393, 0xA48DDFDF,
-    0xCA4C2929, 0x10141C1C, 0x2173D7D7, 0xF0CCB4B4, 0xD309D4D4, 0x5D108A8A,
-    0x0FE25151, 0x00000000, 0x6F9A1919, 0x9DE01A1A, 0x368F9494, 0x42E6C7C7,
-    0x4AECC9C9, 0x5EFDD2D2, 0xC1AB7F7F, 0xE0D8A8A8 },
-
-  { 0xBC75BC32, 0xECF3EC21, 0x20C62043, 0xB3F4B3C9, 0xDADBDA03, 0x027B028B,
-    0xE2FBE22B, 0x9EC89EFA, 0xC94AC9EC, 0xD4D3D409, 0x18E6186B, 0x1E6B1E9F,
-    0x9845980E, 0xB27DB238, 0xA6E8A6D2, 0x264B26B7, 0x3CD63C57, 0x9332938A,
-    0x82D882EE, 0x52FD5298, 0x7B377BD4, 0xBB71BB37, 0x5BF15B97, 0x47E14783,
-    0x2430243C, 0x510F51E2, 0xBAF8BAC6, 0x4A1B4AF3, 0xBF87BF48, 0x0DFA0D70,
-    0xB006B0B3, 0x753F75DE, 0xD25ED2FD, 0x7DBA7D20, 0x66AE6631, 0x3A5B3AA3,
-    0x598A591C, 0x00000000, 0xCDBCCD93, 0x1A9D1AE0, 0xAE6DAE2C, 0x7FC17FAB,
-    0x2BB12BC7, 0xBE0EBEB9, 0xE080E0A0, 0x8A5D8A10, 0x3BD23B52, 0x64D564BA,
-    0xD8A0D888, 0xE784E7A5, 0x5F075FE8, 0x1B141B11, 0x2CB52CC2, 0xFC90FCB4,
-    0x312C3127, 0x80A38065, 0x73B2732A, 0x0C730C81, 0x794C795F, 0x6B546B41,
-    0x4B924B02, 0x53745369, 0x9436948F, 0x8351831F, 0x2A382A36, 0xC4B0C49C,
-    0x22BD22C8, 0xD55AD5F8, 0xBDFCBDC3, 0x48604878, 0xFF62FFCE, 0x4C964C07,
-    0x416C4177, 0xC742C7E6, 0xEBF7EB24, 0x1C101C14, 0x5D7C5D63, 0x36283622,
-    0x672767C0, 0xE98CE9AF, 0x441344F9, 0x149514EA, 0xF59CF5BB, 0xCFC7CF18,
-    0x3F243F2D, 0xC046C0E3, 0x723B72DB, 0x5470546C, 0x29CA294C, 0xF0E3F035,
-    0x088508FE, 0xC6CBC617, 0xF311F34F, 0x8CD08CE4, 0xA493A459, 0xCAB8CA96,
-    0x68A6683B, 0xB883B84D, 0x38203828, 0xE5FFE52E, 0xAD9FAD56, 0x0B770B84,
-    0xC8C3C81D, 0x99CC99FF, 0x580358ED, 0x196F199A, 0x0E080E0A, 0x95BF957E,
-    0x70407050, 0xF7E7F730, 0x6E2B6ECF, 0x1FE21F6E, 0xB579B53D, 0x090C090F,
-    0x61AA6134, 0x57825716, 0x9F419F0B, 0x9D3A9D80, 0x11EA1164, 0x25B925CD,
-    0xAFE4AFDD, 0x459A4508, 0xDFA4DF8D, 0xA397A35C, 0xEA7EEAD5, 0x35DA3558,
-    0xED7AEDD0, 0x431743FC, 0xF866F8CB, 0xFB94FBB1, 0x37A137D3, 0xFA1DFA40,
-    0xC23DC268, 0xB4F0B4CC, 0x32DE325D, 0x9CB39C71, 0x560B56E7, 0xE372E3DA,
-    0x87A78760, 0x151C151B, 0xF9EFF93A, 0x63D163BF, 0x345334A9, 0x9A3E9A85,
-    0xB18FB142, 0x7C337CD1, 0x8826889B, 0x3D5F3DA6, 0xA1ECA1D7, 0xE476E4DF,
-    0x812A8194, 0x91499101, 0x0F810FFB, 0xEE88EEAA, 0x16EE1661, 0xD721D773,
-    0x97C497F5, 0xA51AA5A8, 0xFEEBFE3F, 0x6DD96DB5, 0x78C578AE, 0xC539C56D,
-    0x1D991DE5, 0x76CD76A4, 0x3EAD3EDC, 0xCB31CB67, 0xB68BB647, 0xEF01EF5B,
-    0x1218121E, 0x602360C5, 0x6ADD6AB0, 0x4D1F4DF6, 0xCE4ECEE9, 0xDE2DDE7C,
-    0x55F9559D, 0x7E487E5A, 0x214F21B2, 0x03F2037A, 0xA065A026, 0x5E8E5E19,
-    0x5A785A66, 0x655C654B, 0x6258624E, 0xFD19FD45, 0x068D06F4, 0x40E54086,
-    0xF298F2BE, 0x335733AC, 0x17671790, 0x057F058E, 0xE805E85E, 0x4F644F7D,
-    0x89AF896A, 0x10631095, 0x74B6742F, 0x0AFE0A75, 0x5CF55C92, 0x9BB79B74,
-    0x2D3C2D33, 0x30A530D6, 0x2ECE2E49, 0x49E94989, 0x46684672, 0x77447755,
-    0xA8E0A8D8, 0x964D9604, 0x284328BD, 0xA969A929, 0xD929D979, 0x862E8691,
-    0xD1ACD187, 0xF415F44A, 0x8D598D15, 0xD6A8D682, 0xB90AB9BC, 0x429E420D,
-    0xF66EF6C1, 0x2F472FB8, 0xDDDFDD06, 0x23342339, 0xCC35CC62, 0xF16AF1C4,
-    0xC1CFC112, 0x85DC85EB, 0x8F228F9E, 0x71C971A1, 0x90C090F0, 0xAA9BAA53,
-    0x018901F1, 0x8BD48BE1, 0x4EED4E8C, 0x8EAB8E6F, 0xAB12ABA2, 0x6FA26F3E,
-    0xE60DE654, 0xDB52DBF2, 0x92BB927B, 0xB702B7B6, 0x692F69CA, 0x39A939D9,
-    0xD3D7D30C, 0xA761A723, 0xA21EA2AD, 0xC3B4C399, 0x6C506C44, 0x07040705,
-    0x04F6047F, 0x27C22746, 0xAC16ACA7, 0xD025D076, 0x50865013, 0xDC56DCF7,
-    0x8455841A, 0xE109E151, 0x7ABE7A25, 0x139113EF },
-
-  { 0xD939A9D9, 0x90176790, 0x719CB371, 0xD2A6E8D2, 0x05070405, 0x9852FD98,
-    0x6580A365, 0xDFE476DF, 0x08459A08, 0x024B9202, 0xA0E080A0, 0x665A7866,
-    0xDDAFE4DD, 0xB06ADDB0, 0xBF63D1BF, 0x362A3836, 0x54E60D54, 0x4320C643,
-    0x62CC3562, 0xBEF298BE, 0x1E12181E, 0x24EBF724, 0xD7A1ECD7, 0x77416C77,
-    0xBD2843BD, 0x32BC7532, 0xD47B37D4, 0x9B88269B, 0x700DFA70, 0xF94413F9,
-    0xB1FB94B1, 0x5A7E485A, 0x7A03F27A, 0xE48CD0E4, 0x47B68B47, 0x3C24303C,
-    0xA5E784A5, 0x416B5441, 0x06DDDF06, 0xC56023C5, 0x45FD1945, 0xA33A5BA3,
-    0x68C23D68, 0x158D5915, 0x21ECF321, 0x3166AE31, 0x3E6FA23E, 0x16578216,
-    0x95106395, 0x5BEF015B, 0x4DB8834D, 0x91862E91, 0xB56DD9B5, 0x1F83511F,
-    0x53AA9B53, 0x635D7C63, 0x3B68A63B, 0x3FFEEB3F, 0xD630A5D6, 0x257ABE25,
-    0xA7AC16A7, 0x0F090C0F, 0x35F0E335, 0x23A76123, 0xF090C0F0, 0xAFE98CAF,
-    0x809D3A80, 0x925CF592, 0x810C7381, 0x27312C27, 0x76D02576, 0xE7560BE7,
-    0x7B92BB7B, 0xE9CE4EE9, 0xF10189F1, 0x9F1E6B9F, 0xA93453A9, 0xC4F16AC4,
-    0x99C3B499, 0x975BF197, 0x8347E183, 0x6B18E66B, 0xC822BDC8, 0x0E98450E,
-    0x6E1FE26E, 0xC9B3F4C9, 0x2F74B62F, 0xCBF866CB, 0xFF99CCFF, 0xEA1495EA,
-    0xED5803ED, 0xF7DC56F7, 0xE18BD4E1, 0x1B151C1B, 0xADA21EAD, 0x0CD3D70C,
-    0x2BE2FB2B, 0x1DC8C31D, 0x195E8E19, 0xC22CB5C2, 0x8949E989, 0x12C1CF12,
-    0x7E95BF7E, 0x207DBA20, 0x6411EA64, 0x840B7784, 0x6DC5396D, 0x6A89AF6A,
-    0xD17C33D1, 0xA171C9A1, 0xCEFF62CE, 0x37BB7137, 0xFB0F81FB, 0x3DB5793D,
-    0x51E10951, 0xDC3EADDC, 0x2D3F242D, 0xA476CDA4, 0x9D55F99D, 0xEE82D8EE,
-    0x8640E586, 0xAE78C5AE, 0xCD25B9CD, 0x04964D04, 0x55774455, 0x0A0E080A,
-    0x13508613, 0x30F7E730, 0xD337A1D3, 0x40FA1D40, 0x3461AA34, 0x8C4EED8C,
-    0xB3B006B3, 0x6C54706C, 0x2A73B22A, 0x523BD252, 0x0B9F410B, 0x8B027B8B,
-    0x88D8A088, 0x4FF3114F, 0x67CB3167, 0x4627C246, 0xC06727C0, 0xB4FC90B4,
-    0x28382028, 0x7F04F67F, 0x78486078, 0x2EE5FF2E, 0x074C9607, 0x4B655C4B,
-    0xC72BB1C7, 0x6F8EAB6F, 0x0D429E0D, 0xBBF59CBB, 0xF2DB52F2, 0xF34A1BF3,
-    0xA63D5FA6, 0x59A49359, 0xBCB90ABC, 0x3AF9EF3A, 0xEF1391EF, 0xFE0885FE,
-    0x01914901, 0x6116EE61, 0x7CDE2D7C, 0xB2214FB2, 0x42B18F42, 0xDB723BDB,
-    0xB82F47B8, 0x48BF8748, 0x2CAE6D2C, 0xE3C046E3, 0x573CD657, 0x859A3E85,
-    0x29A96929, 0x7D4F647D, 0x94812A94, 0x492ECE49, 0x17C6CB17, 0xCA692FCA,
-    0xC3BDFCC3, 0x5CA3975C, 0x5EE8055E, 0xD0ED7AD0, 0x87D1AC87, 0x8E057F8E,
-    0xBA64D5BA, 0xA8A51AA8, 0xB7264BB7, 0xB9BE0EB9, 0x6087A760, 0xF8D55AF8,
-    0x22362822, 0x111B1411, 0xDE753FDE, 0x79D92979, 0xAAEE88AA, 0x332D3C33,
-    0x5F794C5F, 0xB6B702B6, 0x96CAB896, 0x5835DA58, 0x9CC4B09C, 0xFC4317FC,
-    0x1A84551A, 0xF64D1FF6, 0x1C598A1C, 0x38B27D38, 0xAC3357AC, 0x18CFC718,
-    0xF4068DF4, 0x69537469, 0x749BB774, 0xF597C4F5, 0x56AD9F56, 0xDAE372DA,
-    0xD5EA7ED5, 0x4AF4154A, 0x9E8F229E, 0xA2AB12A2, 0x4E62584E, 0xE85F07E8,
-    0xE51D99E5, 0x39233439, 0xC1F66EC1, 0x446C5044, 0x5D32DE5D, 0x72466872,
-    0x26A06526, 0x93CDBC93, 0x03DADB03, 0xC6BAF8C6, 0xFA9EC8FA, 0x82D6A882,
-    0xCF6E2BCF, 0x50704050, 0xEB85DCEB, 0x750AFE75, 0x8A93328A, 0x8DDFA48D,
-    0x4C29CA4C, 0x141C1014, 0x73D72173, 0xCCB4F0CC, 0x09D4D309, 0x108A5D10,
-    0xE2510FE2, 0x00000000, 0x9A196F9A, 0xE01A9DE0, 0x8F94368F, 0xE6C742E6,
-    0xECC94AEC, 0xFDD25EFD, 0xAB7FC1AB, 0xD8A8E0D8 }
-};
-
-#define mds(n,x) m_tab[n][x]
-
-static u32 h_fun (u32 *sk, u32 *lk, const u32 x, const u32 *key)
-{
-  u32  b0, b1, b2, b3;
-
-  b0 = extract_byte (x, 0);
-  b1 = extract_byte (x, 1);
-  b2 = extract_byte (x, 2);
-  b3 = extract_byte (x, 3);
-
-  b0 = q (1, b0) ^ extract_byte (key[3], 0);
-  b1 = q (0, b1) ^ extract_byte (key[3], 1);
-  b2 = q (0, b2) ^ extract_byte (key[3], 2);
-  b3 = q (1, b3) ^ extract_byte (key[3], 3);
-
-  b0 = q (1, b0) ^ extract_byte (key[2], 0);
-  b1 = q (1, b1) ^ extract_byte (key[2], 1);
-  b2 = q (0, b2) ^ extract_byte (key[2], 2);
-  b3 = q (0, b3) ^ extract_byte (key[2], 3);
-
-  b0 = q (0, (q (0, b0) ^ extract_byte (key[1], 0))) ^ extract_byte (key[0], 0);
-  b1 = q (0, (q (1, b1) ^ extract_byte (key[1], 1))) ^ extract_byte (key[0], 1);
-  b2 = q (1, (q (0, b2) ^ extract_byte (key[1], 2))) ^ extract_byte (key[0], 2);
-  b3 = q (1, (q (1, b3) ^ extract_byte (key[1], 3))) ^ extract_byte (key[0], 3);
-
-  return mds (0, b0) ^ mds (1, b1) ^ mds (2, b2) ^ mds (3, b3);
-}
-
-#define q40(x,k) q (0, q (0, q (1, q (1, x) ^ extract_byte (k[3], 0)) ^ extract_byte (k[2], 0)) ^ extract_byte (k[1], 0)) ^ extract_byte (k[0], 0)
-#define q41(x,k) q (0, q (1, q (1, q (0, x) ^ extract_byte (k[3], 1)) ^ extract_byte (k[2], 1)) ^ extract_byte (k[1], 1)) ^ extract_byte (k[0], 1)
-#define q42(x,k) q (1, q (0, q (0, q (0, x) ^ extract_byte (k[3], 2)) ^ extract_byte (k[2], 2)) ^ extract_byte (k[1], 2)) ^ extract_byte (k[0], 2)
-#define q43(x,k) q (1, q (1, q (0, q (1, x) ^ extract_byte (k[3], 3)) ^ extract_byte (k[2], 3)) ^ extract_byte (k[1], 3)) ^ extract_byte (k[0], 3)
-
-#define g1_fun(x)                                   \
-  (mds (0, q40 (extract_byte (x, 3), sk)) ^ \
-   mds (1, q41 (extract_byte (x, 0), sk)) ^ \
-   mds (2, q42 (extract_byte (x, 1), sk)) ^ \
-   mds (3, q43 (extract_byte (x, 2), sk)))
-
-#define g0_fun(x)                                   \
-  (mds (0, q40 (extract_byte (x, 0), sk)) ^ \
-   mds (1, q41 (extract_byte (x, 1), sk)) ^ \
-   mds (2, q42 (extract_byte (x, 2), sk)) ^ \
-   mds (3, q43 (extract_byte (x, 3), sk)))
-
-static u32 mds_rem (u32 p0, u32 p1)
-{
-  #define G_MOD 0x14d
-
-  for (int i = 0; i < 8; i++)
-  {
-    u32 t = p1 >> 24;
-
-    p1 = (p1 << 8) | (p0 >> 24);
-
-    p0 <<= 8;
-
-    u32 u = (t << 1);
-
-    if (t & 0x80) u ^= G_MOD;
-
-    p1 ^= t ^ (u << 16);
-
-    u ^= (t >> 1);
-
-    if (t & 0x01) u ^= G_MOD >> 1;
-
-    p1 ^= (u << 24) | (u << 8);
-  }
-
-  return p1;
-}
-
-static void twofish256_set_key (u32 *sk, u32 *lk, const u32 *ukey)
-{
-  u32 me_key[4];
-
-  me_key[0] = ukey[0];
-  me_key[1] = ukey[2];
-  me_key[2] = ukey[4];
-  me_key[3] = ukey[6];
-
-  u32 mo_key[4];
-
-  mo_key[0] = ukey[1];
-  mo_key[1] = ukey[3];
-  mo_key[2] = ukey[5];
-  mo_key[3] = ukey[7];
-
-  sk[3] = mds_rem (me_key[0], mo_key[0]);
-  sk[2] = mds_rem (me_key[1], mo_key[1]);
-  sk[1] = mds_rem (me_key[2], mo_key[2]);
-  sk[0] = mds_rem (me_key[3], mo_key[3]);
-
-  for (int i = 0; i < 40; i += 2)
-  {
-    u32 a = 0x01010101 * i;
-    u32 b = 0x01010101 + a;
-
-    a = h_fun (sk, lk, a, me_key);
-    b = h_fun (sk, lk, b, mo_key);
-
-    b = rotl32 (b, 8);
-
-    lk[i + 0] = a + b;
-    lk[i + 1] = rotl32 (a + 2 * b, 9);
-  }
-}
-
-#define f_rnd(i)                                                    \
-{                                                                   \
-  u32 t0 = g0_fun (data[0]);                                       \
-  u32 t1 = g1_fun (data[1]);                                       \
-  data[2] = rotr32 (data[2] ^ (t0 + t1 + lk[4 * (i) + 8]), 1);      \
-  data[3] = rotl32 (data[3], 1) ^ (t0 + 2 * t1 + lk[4 * (i) + 9]);  \
-  u32 t2 = g0_fun (data[2]);                                       \
-  u32 t3 = g1_fun (data[3]);                                       \
-  data[0] = rotr32 (data[0] ^ (t2 + t3 + lk[4 * (i) + 10]), 1);     \
-  data[1] = rotl32 (data[1], 1) ^ (t2 + 2 * t3 + lk[4 * (i) + 11]); \
-}
-
-static void twofish256_encrypt (const u32 *sk, const u32 *lk, const u32 *in, u32 *out)
-{
-  u32 data[4];
-
-  data[0] = in[0] ^ lk[0];
-  data[1] = in[1] ^ lk[1];
-  data[2] = in[2] ^ lk[2];
-  data[3] = in[3] ^ lk[3];
-
-  f_rnd (0);
-  f_rnd (1);
-  f_rnd (2);
-  f_rnd (3);
-  f_rnd (4);
-  f_rnd (5);
-  f_rnd (6);
-  f_rnd (7);
-
-  out[0] = data[2] ^ lk[4];
-  out[1] = data[3] ^ lk[5];
-  out[2] = data[0] ^ lk[6];
-  out[3] = data[1] ^ lk[7];
-}
-
-#define i_rnd(i)                                                    \
-{                                                                   \
-  u32 t0 = g0_fun (data[0]);                                       \
-  u32 t1 = g1_fun (data[1]);                                       \
-  data[2] = rotl32 (data[2], 1) ^ (t0 + t1 + lk[4 * (i) + 10]);     \
-  data[3] = rotr32 (data[3] ^ (t0 + 2 * t1 + lk[4 * (i) + 11]), 1); \
-  u32 t2 = g0_fun (data[2]);                                       \
-  u32 t3 = g1_fun (data[3]);                                       \
-  data[0] = rotl32 (data[0], 1) ^ (t2 + t3 + lk[4 * (i) +  8]);     \
-  data[1] = rotr32 (data[1] ^ (t2 + 2 * t3 + lk[4 * (i) +  9]), 1); \
-}
-
-static void twofish256_decrypt (const u32 *sk, const u32 *lk, const u32 *in, u32 *out)
-{
-  u32 data[4];
-
-  data[0] = in[0] ^ lk[4];
-  data[1] = in[1] ^ lk[5];
-  data[2] = in[2] ^ lk[6];
-  data[3] = in[3] ^ lk[7];
-
-  i_rnd (7);
-  i_rnd (6);
-  i_rnd (5);
-  i_rnd (4);
-  i_rnd (3);
-  i_rnd (2);
-  i_rnd (1);
-  i_rnd (0);
-
-  out[0] = data[2] ^ lk[0];
-  out[1] = data[3] ^ lk[1];
-  out[2] = data[0] ^ lk[2];
-  out[3] = data[1] ^ lk[3];
-}
-
-static void twofish256_decrypt_xts (const u32 *ukey1, const u32 *ukey2, const u32 *in, u32 *out)
-{
-  u32 T[4] = { 0 };
-  u32 Z[4] = { 0 };
-
-  out[0] = in[0];
-  out[1] = in[1];
-  out[2] = in[2];
-  out[3] = in[3];
-
-  u32 sk[4]; u32 lk[40];
-
-  twofish256_set_key (sk, lk, ukey2);
-  twofish256_encrypt (sk, lk, Z, T);
-
-  out[0] ^= T[0];
-  out[1] ^= T[1];
-  out[2] ^= T[2];
-  out[3] ^= T[3];
-
-  twofish256_set_key (sk, lk, ukey1);
-  twofish256_decrypt (sk, lk, out, out);
-
-  out[0] ^= T[0];
-  out[1] ^= T[1];
-  out[2] ^= T[2];
-  out[3] ^= T[3];
-}
diff --git a/OpenCL/m06211.cl b/OpenCL/m06211.cl
index a1f9cd3..40dcf61 100644
--- a/OpenCL/m06211.cl
+++ b/OpenCL/m06211.cl
@@ -17,9 +17,9 @@
 #include "OpenCL/types_ocl.c"
 #include "OpenCL/common.c"
 
-#include "OpenCL/kernel_aes256_amd.c"
-#include "OpenCL/kernel_twofish256_amd.c"
-#include "OpenCL/kernel_serpent256_amd.c"
+#include "OpenCL/kernel_aes256.c"
+#include "OpenCL/kernel_twofish256.c"
+#include "OpenCL/kernel_serpent256.c"
 
 static void ripemd160_transform (const u32 w[16], u32 dgst[5])
 {
diff --git a/OpenCL/m06212.cl b/OpenCL/m06212.cl
index 83327f4..dfd958c 100644
--- a/OpenCL/m06212.cl
+++ b/OpenCL/m06212.cl
@@ -17,9 +17,9 @@
 #include "OpenCL/types_ocl.c"
 #include "OpenCL/common.c"
 
-#include "OpenCL/kernel_aes256_amd.c"
-#include "OpenCL/kernel_twofish256_amd.c"
-#include "OpenCL/kernel_serpent256_amd.c"
+#include "OpenCL/kernel_aes256.c"
+#include "OpenCL/kernel_twofish256.c"
+#include "OpenCL/kernel_serpent256.c"
 
 static void ripemd160_transform (const u32 w[16], u32 dgst[5])
 {
diff --git a/OpenCL/m06213.cl b/OpenCL/m06213.cl
index 4df4d9e..ce072bb 100644
--- a/OpenCL/m06213.cl
+++ b/OpenCL/m06213.cl
@@ -17,9 +17,9 @@
 #include "OpenCL/types_ocl.c"
 #include "OpenCL/common.c"
 
-#include "OpenCL/kernel_aes256_amd.c"
-#include "OpenCL/kernel_twofish256_amd.c"
-#include "OpenCL/kernel_serpent256_amd.c"
+#include "OpenCL/kernel_aes256.c"
+#include "OpenCL/kernel_twofish256.c"
+#include "OpenCL/kernel_serpent256.c"
 
 static void ripemd160_transform (const u32 w[16], u32 dgst[5])
 {
diff --git a/OpenCL/m06221.cl b/OpenCL/m06221.cl
index 3cd7408..d2d5d04 100644
--- a/OpenCL/m06221.cl
+++ b/OpenCL/m06221.cl
@@ -17,9 +17,9 @@
 #include "OpenCL/types_ocl.c"
 #include "OpenCL/common.c"
 
-#include "OpenCL/kernel_aes256_amd.c"
-#include "OpenCL/kernel_twofish256_amd.c"
-#include "OpenCL/kernel_serpent256_amd.c"
+#include "OpenCL/kernel_aes256.c"
+#include "OpenCL/kernel_twofish256.c"
+#include "OpenCL/kernel_serpent256.c"
 
 __constant u64 k_sha512[80] =
 {
diff --git a/OpenCL/m06222.cl b/OpenCL/m06222.cl
index 7246459..5b29270 100644
--- a/OpenCL/m06222.cl
+++ b/OpenCL/m06222.cl
@@ -17,9 +17,9 @@
 #include "OpenCL/types_ocl.c"
 #include "OpenCL/common.c"
 
-#include "OpenCL/kernel_aes256_amd.c"
-#include "OpenCL/kernel_twofish256_amd.c"
-#include "OpenCL/kernel_serpent256_amd.c"
+#include "OpenCL/kernel_aes256.c"
+#include "OpenCL/kernel_twofish256.c"
+#include "OpenCL/kernel_serpent256.c"
 
 __constant u64 k_sha512[80] =
 {
diff --git a/OpenCL/m06223.cl b/OpenCL/m06223.cl
index 4f4ccf6..5c8eda4 100644
--- a/OpenCL/m06223.cl
+++ b/OpenCL/m06223.cl
@@ -17,9 +17,9 @@
 #include "OpenCL/types_ocl.c"
 #include "OpenCL/common.c"
 
-#include "OpenCL/kernel_aes256_amd.c"
-#include "OpenCL/kernel_twofish256_amd.c"
-#include "OpenCL/kernel_serpent256_amd.c"
+#include "OpenCL/kernel_aes256.c"
+#include "OpenCL/kernel_twofish256.c"
+#include "OpenCL/kernel_serpent256.c"
 
 __constant u64 k_sha512[80] =
 {
diff --git a/OpenCL/m06231.cl b/OpenCL/m06231.cl
index f9ecaa1..ec85d19 100644
--- a/OpenCL/m06231.cl
+++ b/OpenCL/m06231.cl
@@ -19,9 +19,9 @@
 #include "OpenCL/types_ocl.c"
 #include "OpenCL/common.c"
 
-#include "OpenCL/kernel_aes256_amd.c"
-#include "OpenCL/kernel_twofish256_amd.c"
-#include "OpenCL/kernel_serpent256_amd.c"
+#include "OpenCL/kernel_aes256.c"
+#include "OpenCL/kernel_twofish256.c"
+#include "OpenCL/kernel_serpent256.c"
 
 #define R 10
 
diff --git a/OpenCL/m06232.cl b/OpenCL/m06232.cl
index 4315f7c..475c8dc 100644
--- a/OpenCL/m06232.cl
+++ b/OpenCL/m06232.cl
@@ -19,9 +19,9 @@
 #include "OpenCL/types_ocl.c"
 #include "OpenCL/common.c"
 
-#include "OpenCL/kernel_aes256_amd.c"
-#include "OpenCL/kernel_twofish256_amd.c"
-#include "OpenCL/kernel_serpent256_amd.c"
+#include "OpenCL/kernel_aes256.c"
+#include "OpenCL/kernel_twofish256.c"
+#include "OpenCL/kernel_serpent256.c"
 
 #define R 10
 
diff --git a/OpenCL/m06233.cl b/OpenCL/m06233.cl
index 16a7888..aa81f31 100644
--- a/OpenCL/m06233.cl
+++ b/OpenCL/m06233.cl
@@ -19,9 +19,9 @@
 #include "OpenCL/types_ocl.c"
 #include "OpenCL/common.c"
 
-#include "OpenCL/kernel_aes256_amd.c"
-#include "OpenCL/kernel_twofish256_amd.c"
-#include "OpenCL/kernel_serpent256_amd.c"
+#include "OpenCL/kernel_aes256.c"
+#include "OpenCL/kernel_twofish256.c"
+#include "OpenCL/kernel_serpent256.c"
 
 #define R 10
 
diff --git a/include/ext_dummy.h b/include/ext_dummy.h
deleted file mode 100644
index c5139ae..0000000
--- a/include/ext_dummy.h
+++ /dev/null
@@ -1,15 +0,0 @@
-/**
- * Author......: Jens Steube <jens.steube@gmail.com>
- * License.....: MIT
- */
-
-#ifndef EXT_SMI_H
-#define EXT_SMI_H
-
-#ifdef HAVE_HWMON
-
-#include <common.h>
-
-#endif // HAVE_HWMON
-
-#endif
diff --git a/include/ext_smi.h b/include/ext_smi.h
deleted file mode 100644
index 3e88bde..0000000
--- a/include/ext_smi.h
+++ /dev/null
@@ -1,20 +0,0 @@
-/**
- * Author......: Jens Steube <jens.steube@gmail.com>
- * License.....: MIT
- */
-
-#ifndef EXT_SMI_H
-#define EXT_SMI_H
-
-#ifdef HAVE_HWMON
-
-#include <common.h>
-
-#define SMI_OK    0
-#define SMI_NOBIN 1
-
-int hc_nvidia_smi (int dev, int *temperature, int *gpu);
-
-#endif // HAVE_HWMON
-
-#endif
diff --git a/include/shared.h b/include/shared.h
index 08d0cd8..c10c7d2 100644
--- a/include/shared.h
+++ b/include/shared.h
@@ -100,10 +100,7 @@
  * temperature management
  */
 
-#ifdef OSX
-#include <ext_smi.h>
-#include <ext_dummy.h>
-#elif _WIN
+#if _WIN
 #include <ext_ADL.h>
 #include <ext_nvapi.h>
 #else
diff --git a/src/ext_dummy.c b/src/ext_dummy.c
deleted file mode 100644
index 3b09b02..0000000
--- a/src/ext_dummy.c
+++ /dev/null
@@ -1 +0,0 @@
-#include <ext_dummy.h>
diff --git a/src/ext_smi.c b/src/ext_smi.c
deleted file mode 100644
index ef032bf..0000000
--- a/src/ext_smi.c
+++ /dev/null
@@ -1,68 +0,0 @@
-/**
- * Author......: Jens Steube <jens.steube@gmail.com>
- * License.....: MIT
- */
-
-#include <ext_smi.h>
-
-int hc_nvidia_smi (int dev, int *temperature, int *gpu)
-{
-  char cmd[256] = { 0 };
-
-  snprintf (cmd, sizeof (cmd) - 1, "nvidia-smi -q -g %d", dev);
-
-  FILE *fp = popen (cmd, "r");
-
-  if (fp == NULL)
-  {
-    log_info ("WARN: %s\n", "nvidia-smi is missing!");
-
-    return SMI_NOBIN;
-  }
-
-  int temp_found = 0;
-  int util_found = 0;
-
-  char token[32];
-
-  while (fscanf (fp, " %31s ", token) == 1)
-  {
-    if (strcmp (token, "Temperature") == 0)
-    {
-      if (fscanf (fp, " : %4s C", token) == 1) *temperature = atoi (token);
-
-      temp_found = 1;
-    }
-
-    if (temp_found == 1)
-    {
-      if ((strcmp (token, "GPU") == 0) || (strcmp (token, "Gpu") == 0))
-      {
-        if (fscanf (fp, " : %4s C", token) == 1) *temperature = atoi (token);
-
-        temp_found = 0;
-      }
-    }
-
-    if (strcmp (token, "Utilization") == 0)
-    {
-      util_found = 1;
-
-      temp_found = 0;
-    }
-
-    if (util_found == 1)
-    {
-      if ((strcmp (token, "GPU") == 0) || (strcmp (token, "Gpu") == 0))
-      {
-        if (fscanf (fp, " : %2s%%", token) == 1) *gpu = atoi (token);
-
-        util_found = 0;
-      }
-    }
-  }
-
-  pclose (fp);
-
-  return (SMI_OK);
-}
diff --git a/tools/package_bin.sh b/tools/package_bin.sh
index e0f8794..3dad855 100755
--- a/tools/package_bin.sh
+++ b/tools/package_bin.sh
@@ -12,6 +12,7 @@ rm -rf $OUT
 rm -rf $OUT.7z
 
 mkdir -p $OUT
+mkdir -p $OUT/include
 
 cp    $IN/oclHashcat??.exe                      $OUT/
 cp    $IN/oclHashcat??.bin                      $OUT/
@@ -26,7 +27,10 @@ cp    $IN/example.dict                          $OUT/
 cp    $IN/example[0123456789]*.hash             $OUT/
 cp    $IN/example[0123456789]*.cmd              $OUT/
 
-cp -r $IN/include                               $OUT/
+cp -r $IN/include/constants.h                   $OUT/include
+cp -r $IN/include/kernel_functions.c            $OUT/include
+cp -r $IN/include/kernel_vendor.h               $OUT/include
+cp -r $IN/include/rp_kernel.h                   $OUT/include
 cp -r $IN/OpenCL                                $OUT/
 
 # since for the binary distribution we still use .bin, we need to rewrite the commands