2 * Author......: Jens Steube <jens.steube@gmail.com>
10 #include "include/constants.h"
11 #include "include/kernel_vendor.h"
18 #include "include/kernel_functions.c"
19 #include "OpenCL/types_ocl.c"
20 #include "OpenCL/common.c"
21 #include "OpenCL/simd.c"
23 #define PERM_OP(a,b,tt,n,m) \
33 #define HPERM_OP(a,tt,n,m) \
39 tt = tt >> (16 + n); \
45 PERM_OP (r, l, tt, 4, 0x0f0f0f0f); \
46 PERM_OP (l, r, tt, 16, 0x0000ffff); \
47 PERM_OP (r, l, tt, 2, 0x33333333); \
48 PERM_OP (l, r, tt, 8, 0x00ff00ff); \
49 PERM_OP (r, l, tt, 1, 0x55555555); \
54 PERM_OP (l, r, tt, 1, 0x55555555); \
55 PERM_OP (r, l, tt, 8, 0x00ff00ff); \
56 PERM_OP (l, r, tt, 2, 0x33333333); \
57 PERM_OP (r, l, tt, 16, 0x0000ffff); \
58 PERM_OP (l, r, tt, 4, 0x0f0f0f0f); \
61 __constant u8 ascii_to_ebcdic_pc[256] =
63 // little hack, can't crack 0-bytes in password, but who cares
64 // 0xab, 0xa8, 0xae, 0xad, 0xc4, 0xf1, 0xf7, 0xf4, 0x86, 0xa1, 0xe0, 0xbc, 0xb3, 0xb0, 0xb6, 0xb5,
65 0x2a, 0xa8, 0xae, 0xad, 0xc4, 0xf1, 0xf7, 0xf4, 0x86, 0xa1, 0xe0, 0xbc, 0xb3, 0xb0, 0xb6, 0xb5,
66 0x8a, 0x89, 0x8f, 0x8c, 0xd3, 0xd0, 0xce, 0xe6, 0x9b, 0x98, 0xd5, 0xe5, 0x92, 0x91, 0x97, 0x94,
67 0x2a, 0x34, 0x54, 0x5d, 0x1c, 0x73, 0x0b, 0x51, 0x31, 0x10, 0x13, 0x37, 0x7c, 0x6b, 0x3d, 0x68,
68 0x4a, 0x49, 0x4f, 0x4c, 0x43, 0x40, 0x46, 0x45, 0x5b, 0x58, 0x5e, 0x16, 0x32, 0x57, 0x76, 0x75,
69 0x52, 0x29, 0x2f, 0x2c, 0x23, 0x20, 0x26, 0x25, 0x3b, 0x38, 0x08, 0x0e, 0x0d, 0x02, 0x01, 0x07,
70 0x04, 0x1a, 0x19, 0x6e, 0x6d, 0x62, 0x61, 0x67, 0x64, 0x7a, 0x79, 0x3e, 0x6b, 0x1f, 0x15, 0x70,
71 0x58, 0xa8, 0xae, 0xad, 0xa2, 0xa1, 0xa7, 0xa4, 0xba, 0xb9, 0x89, 0x8f, 0x8c, 0x83, 0x80, 0x86,
72 0x85, 0x9b, 0x98, 0xef, 0xec, 0xe3, 0xe0, 0xe6, 0xe5, 0xfb, 0xf8, 0x2a, 0x7f, 0x0b, 0xe9, 0xa4,
73 0xea, 0xe9, 0xef, 0xec, 0xe3, 0x80, 0xa7, 0x85, 0xfb, 0xf8, 0xfe, 0xfd, 0xf2, 0xb9, 0xbf, 0x9d,
74 0xcb, 0xc8, 0x9e, 0xcd, 0xc2, 0xc1, 0xc7, 0xba, 0xda, 0xd9, 0xdf, 0xdc, 0xa2, 0x83, 0xd6, 0x68,
75 0x29, 0x2f, 0x2c, 0x23, 0x20, 0x26, 0x25, 0x3b, 0x38, 0x08, 0x0e, 0x0d, 0x02, 0x01, 0x07, 0x04,
76 0x1a, 0x19, 0x6e, 0x6d, 0x62, 0x61, 0x67, 0x64, 0x7a, 0x79, 0x4a, 0x49, 0x4f, 0x4c, 0x43, 0x40,
77 0x46, 0x45, 0x5b, 0xab, 0xbf, 0xbc, 0xb3, 0xb0, 0xb6, 0xb5, 0x8a, 0x9e, 0x9d, 0x92, 0x91, 0x97,
78 0x94, 0xea, 0xfe, 0xfd, 0xf2, 0xf1, 0xf7, 0xf4, 0xcb, 0xc8, 0xce, 0xcd, 0xc2, 0xc1, 0xc7, 0xc4,
79 0xda, 0xd9, 0xdf, 0xdc, 0xd3, 0xd0, 0xd6, 0xd5, 0x3e, 0x3d, 0x32, 0x31, 0x37, 0x34, 0x1f, 0x1c,
80 0x13, 0x10, 0x16, 0x15, 0x7f, 0x7c, 0x73, 0x70, 0x76, 0x75, 0x5e, 0x5d, 0x52, 0x51, 0x57, 0x54,
83 __constant u32 c_SPtrans[8][64] =
86 0x02080800, 0x00080000, 0x02000002, 0x02080802,
87 0x02000000, 0x00080802, 0x00080002, 0x02000002,
88 0x00080802, 0x02080800, 0x02080000, 0x00000802,
89 0x02000802, 0x02000000, 0x00000000, 0x00080002,
90 0x00080000, 0x00000002, 0x02000800, 0x00080800,
91 0x02080802, 0x02080000, 0x00000802, 0x02000800,
92 0x00000002, 0x00000800, 0x00080800, 0x02080002,
93 0x00000800, 0x02000802, 0x02080002, 0x00000000,
94 0x00000000, 0x02080802, 0x02000800, 0x00080002,
95 0x02080800, 0x00080000, 0x00000802, 0x02000800,
96 0x02080002, 0x00000800, 0x00080800, 0x02000002,
97 0x00080802, 0x00000002, 0x02000002, 0x02080000,
98 0x02080802, 0x00080800, 0x02080000, 0x02000802,
99 0x02000000, 0x00000802, 0x00080002, 0x00000000,
100 0x00080000, 0x02000000, 0x02000802, 0x02080800,
101 0x00000002, 0x02080002, 0x00000800, 0x00080802,
104 0x40108010, 0x00000000, 0x00108000, 0x40100000,
105 0x40000010, 0x00008010, 0x40008000, 0x00108000,
106 0x00008000, 0x40100010, 0x00000010, 0x40008000,
107 0x00100010, 0x40108000, 0x40100000, 0x00000010,
108 0x00100000, 0x40008010, 0x40100010, 0x00008000,
109 0x00108010, 0x40000000, 0x00000000, 0x00100010,
110 0x40008010, 0x00108010, 0x40108000, 0x40000010,
111 0x40000000, 0x00100000, 0x00008010, 0x40108010,
112 0x00100010, 0x40108000, 0x40008000, 0x00108010,
113 0x40108010, 0x00100010, 0x40000010, 0x00000000,
114 0x40000000, 0x00008010, 0x00100000, 0x40100010,
115 0x00008000, 0x40000000, 0x00108010, 0x40008010,
116 0x40108000, 0x00008000, 0x00000000, 0x40000010,
117 0x00000010, 0x40108010, 0x00108000, 0x40100000,
118 0x40100010, 0x00100000, 0x00008010, 0x40008000,
119 0x40008010, 0x00000010, 0x40100000, 0x00108000,
122 0x04000001, 0x04040100, 0x00000100, 0x04000101,
123 0x00040001, 0x04000000, 0x04000101, 0x00040100,
124 0x04000100, 0x00040000, 0x04040000, 0x00000001,
125 0x04040101, 0x00000101, 0x00000001, 0x04040001,
126 0x00000000, 0x00040001, 0x04040100, 0x00000100,
127 0x00000101, 0x04040101, 0x00040000, 0x04000001,
128 0x04040001, 0x04000100, 0x00040101, 0x04040000,
129 0x00040100, 0x00000000, 0x04000000, 0x00040101,
130 0x04040100, 0x00000100, 0x00000001, 0x00040000,
131 0x00000101, 0x00040001, 0x04040000, 0x04000101,
132 0x00000000, 0x04040100, 0x00040100, 0x04040001,
133 0x00040001, 0x04000000, 0x04040101, 0x00000001,
134 0x00040101, 0x04000001, 0x04000000, 0x04040101,
135 0x00040000, 0x04000100, 0x04000101, 0x00040100,
136 0x04000100, 0x00000000, 0x04040001, 0x00000101,
137 0x04000001, 0x00040101, 0x00000100, 0x04040000,
140 0x00401008, 0x10001000, 0x00000008, 0x10401008,
141 0x00000000, 0x10400000, 0x10001008, 0x00400008,
142 0x10401000, 0x10000008, 0x10000000, 0x00001008,
143 0x10000008, 0x00401008, 0x00400000, 0x10000000,
144 0x10400008, 0x00401000, 0x00001000, 0x00000008,
145 0x00401000, 0x10001008, 0x10400000, 0x00001000,
146 0x00001008, 0x00000000, 0x00400008, 0x10401000,
147 0x10001000, 0x10400008, 0x10401008, 0x00400000,
148 0x10400008, 0x00001008, 0x00400000, 0x10000008,
149 0x00401000, 0x10001000, 0x00000008, 0x10400000,
150 0x10001008, 0x00000000, 0x00001000, 0x00400008,
151 0x00000000, 0x10400008, 0x10401000, 0x00001000,
152 0x10000000, 0x10401008, 0x00401008, 0x00400000,
153 0x10401008, 0x00000008, 0x10001000, 0x00401008,
154 0x00400008, 0x00401000, 0x10400000, 0x10001008,
155 0x00001008, 0x10000000, 0x10000008, 0x10401000,
158 0x08000000, 0x00010000, 0x00000400, 0x08010420,
159 0x08010020, 0x08000400, 0x00010420, 0x08010000,
160 0x00010000, 0x00000020, 0x08000020, 0x00010400,
161 0x08000420, 0x08010020, 0x08010400, 0x00000000,
162 0x00010400, 0x08000000, 0x00010020, 0x00000420,
163 0x08000400, 0x00010420, 0x00000000, 0x08000020,
164 0x00000020, 0x08000420, 0x08010420, 0x00010020,
165 0x08010000, 0x00000400, 0x00000420, 0x08010400,
166 0x08010400, 0x08000420, 0x00010020, 0x08010000,
167 0x00010000, 0x00000020, 0x08000020, 0x08000400,
168 0x08000000, 0x00010400, 0x08010420, 0x00000000,
169 0x00010420, 0x08000000, 0x00000400, 0x00010020,
170 0x08000420, 0x00000400, 0x00000000, 0x08010420,
171 0x08010020, 0x08010400, 0x00000420, 0x00010000,
172 0x00010400, 0x08010020, 0x08000400, 0x00000420,
173 0x00000020, 0x00010420, 0x08010000, 0x08000020,
176 0x80000040, 0x00200040, 0x00000000, 0x80202000,
177 0x00200040, 0x00002000, 0x80002040, 0x00200000,
178 0x00002040, 0x80202040, 0x00202000, 0x80000000,
179 0x80002000, 0x80000040, 0x80200000, 0x00202040,
180 0x00200000, 0x80002040, 0x80200040, 0x00000000,
181 0x00002000, 0x00000040, 0x80202000, 0x80200040,
182 0x80202040, 0x80200000, 0x80000000, 0x00002040,
183 0x00000040, 0x00202000, 0x00202040, 0x80002000,
184 0x00002040, 0x80000000, 0x80002000, 0x00202040,
185 0x80202000, 0x00200040, 0x00000000, 0x80002000,
186 0x80000000, 0x00002000, 0x80200040, 0x00200000,
187 0x00200040, 0x80202040, 0x00202000, 0x00000040,
188 0x80202040, 0x00202000, 0x00200000, 0x80002040,
189 0x80000040, 0x80200000, 0x00202040, 0x00000000,
190 0x00002000, 0x80000040, 0x80002040, 0x80202000,
191 0x80200000, 0x00002040, 0x00000040, 0x80200040,
194 0x00004000, 0x00000200, 0x01000200, 0x01000004,
195 0x01004204, 0x00004004, 0x00004200, 0x00000000,
196 0x01000000, 0x01000204, 0x00000204, 0x01004000,
197 0x00000004, 0x01004200, 0x01004000, 0x00000204,
198 0x01000204, 0x00004000, 0x00004004, 0x01004204,
199 0x00000000, 0x01000200, 0x01000004, 0x00004200,
200 0x01004004, 0x00004204, 0x01004200, 0x00000004,
201 0x00004204, 0x01004004, 0x00000200, 0x01000000,
202 0x00004204, 0x01004000, 0x01004004, 0x00000204,
203 0x00004000, 0x00000200, 0x01000000, 0x01004004,
204 0x01000204, 0x00004204, 0x00004200, 0x00000000,
205 0x00000200, 0x01000004, 0x00000004, 0x01000200,
206 0x00000000, 0x01000204, 0x01000200, 0x00004200,
207 0x00000204, 0x00004000, 0x01004204, 0x01000000,
208 0x01004200, 0x00000004, 0x00004004, 0x01004204,
209 0x01000004, 0x01004200, 0x01004000, 0x00004004,
212 0x20800080, 0x20820000, 0x00020080, 0x00000000,
213 0x20020000, 0x00800080, 0x20800000, 0x20820080,
214 0x00000080, 0x20000000, 0x00820000, 0x00020080,
215 0x00820080, 0x20020080, 0x20000080, 0x20800000,
216 0x00020000, 0x00820080, 0x00800080, 0x20020000,
217 0x20820080, 0x20000080, 0x00000000, 0x00820000,
218 0x20000000, 0x00800000, 0x20020080, 0x20800080,
219 0x00800000, 0x00020000, 0x20820000, 0x00000080,
220 0x00800000, 0x00020000, 0x20000080, 0x20820080,
221 0x00020080, 0x20000000, 0x00000000, 0x00820000,
222 0x20800080, 0x20020080, 0x20020000, 0x00800080,
223 0x20820000, 0x00000080, 0x00800080, 0x20020000,
224 0x20820080, 0x00800000, 0x20800000, 0x20000080,
225 0x00820000, 0x00020080, 0x20020080, 0x20800000,
226 0x00000080, 0x20820000, 0x00820080, 0x00000000,
227 0x20000000, 0x20800080, 0x00020000, 0x00820080,
231 __constant u32 c_skb[8][64] =
234 0x00000000, 0x00000010, 0x20000000, 0x20000010,
235 0x00010000, 0x00010010, 0x20010000, 0x20010010,
236 0x00000800, 0x00000810, 0x20000800, 0x20000810,
237 0x00010800, 0x00010810, 0x20010800, 0x20010810,
238 0x00000020, 0x00000030, 0x20000020, 0x20000030,
239 0x00010020, 0x00010030, 0x20010020, 0x20010030,
240 0x00000820, 0x00000830, 0x20000820, 0x20000830,
241 0x00010820, 0x00010830, 0x20010820, 0x20010830,
242 0x00080000, 0x00080010, 0x20080000, 0x20080010,
243 0x00090000, 0x00090010, 0x20090000, 0x20090010,
244 0x00080800, 0x00080810, 0x20080800, 0x20080810,
245 0x00090800, 0x00090810, 0x20090800, 0x20090810,
246 0x00080020, 0x00080030, 0x20080020, 0x20080030,
247 0x00090020, 0x00090030, 0x20090020, 0x20090030,
248 0x00080820, 0x00080830, 0x20080820, 0x20080830,
249 0x00090820, 0x00090830, 0x20090820, 0x20090830,
252 0x00000000, 0x02000000, 0x00002000, 0x02002000,
253 0x00200000, 0x02200000, 0x00202000, 0x02202000,
254 0x00000004, 0x02000004, 0x00002004, 0x02002004,
255 0x00200004, 0x02200004, 0x00202004, 0x02202004,
256 0x00000400, 0x02000400, 0x00002400, 0x02002400,
257 0x00200400, 0x02200400, 0x00202400, 0x02202400,
258 0x00000404, 0x02000404, 0x00002404, 0x02002404,
259 0x00200404, 0x02200404, 0x00202404, 0x02202404,
260 0x10000000, 0x12000000, 0x10002000, 0x12002000,
261 0x10200000, 0x12200000, 0x10202000, 0x12202000,
262 0x10000004, 0x12000004, 0x10002004, 0x12002004,
263 0x10200004, 0x12200004, 0x10202004, 0x12202004,
264 0x10000400, 0x12000400, 0x10002400, 0x12002400,
265 0x10200400, 0x12200400, 0x10202400, 0x12202400,
266 0x10000404, 0x12000404, 0x10002404, 0x12002404,
267 0x10200404, 0x12200404, 0x10202404, 0x12202404,
270 0x00000000, 0x00000001, 0x00040000, 0x00040001,
271 0x01000000, 0x01000001, 0x01040000, 0x01040001,
272 0x00000002, 0x00000003, 0x00040002, 0x00040003,
273 0x01000002, 0x01000003, 0x01040002, 0x01040003,
274 0x00000200, 0x00000201, 0x00040200, 0x00040201,
275 0x01000200, 0x01000201, 0x01040200, 0x01040201,
276 0x00000202, 0x00000203, 0x00040202, 0x00040203,
277 0x01000202, 0x01000203, 0x01040202, 0x01040203,
278 0x08000000, 0x08000001, 0x08040000, 0x08040001,
279 0x09000000, 0x09000001, 0x09040000, 0x09040001,
280 0x08000002, 0x08000003, 0x08040002, 0x08040003,
281 0x09000002, 0x09000003, 0x09040002, 0x09040003,
282 0x08000200, 0x08000201, 0x08040200, 0x08040201,
283 0x09000200, 0x09000201, 0x09040200, 0x09040201,
284 0x08000202, 0x08000203, 0x08040202, 0x08040203,
285 0x09000202, 0x09000203, 0x09040202, 0x09040203,
288 0x00000000, 0x00100000, 0x00000100, 0x00100100,
289 0x00000008, 0x00100008, 0x00000108, 0x00100108,
290 0x00001000, 0x00101000, 0x00001100, 0x00101100,
291 0x00001008, 0x00101008, 0x00001108, 0x00101108,
292 0x04000000, 0x04100000, 0x04000100, 0x04100100,
293 0x04000008, 0x04100008, 0x04000108, 0x04100108,
294 0x04001000, 0x04101000, 0x04001100, 0x04101100,
295 0x04001008, 0x04101008, 0x04001108, 0x04101108,
296 0x00020000, 0x00120000, 0x00020100, 0x00120100,
297 0x00020008, 0x00120008, 0x00020108, 0x00120108,
298 0x00021000, 0x00121000, 0x00021100, 0x00121100,
299 0x00021008, 0x00121008, 0x00021108, 0x00121108,
300 0x04020000, 0x04120000, 0x04020100, 0x04120100,
301 0x04020008, 0x04120008, 0x04020108, 0x04120108,
302 0x04021000, 0x04121000, 0x04021100, 0x04121100,
303 0x04021008, 0x04121008, 0x04021108, 0x04121108,
306 0x00000000, 0x10000000, 0x00010000, 0x10010000,
307 0x00000004, 0x10000004, 0x00010004, 0x10010004,
308 0x20000000, 0x30000000, 0x20010000, 0x30010000,
309 0x20000004, 0x30000004, 0x20010004, 0x30010004,
310 0x00100000, 0x10100000, 0x00110000, 0x10110000,
311 0x00100004, 0x10100004, 0x00110004, 0x10110004,
312 0x20100000, 0x30100000, 0x20110000, 0x30110000,
313 0x20100004, 0x30100004, 0x20110004, 0x30110004,
314 0x00001000, 0x10001000, 0x00011000, 0x10011000,
315 0x00001004, 0x10001004, 0x00011004, 0x10011004,
316 0x20001000, 0x30001000, 0x20011000, 0x30011000,
317 0x20001004, 0x30001004, 0x20011004, 0x30011004,
318 0x00101000, 0x10101000, 0x00111000, 0x10111000,
319 0x00101004, 0x10101004, 0x00111004, 0x10111004,
320 0x20101000, 0x30101000, 0x20111000, 0x30111000,
321 0x20101004, 0x30101004, 0x20111004, 0x30111004,
324 0x00000000, 0x08000000, 0x00000008, 0x08000008,
325 0x00000400, 0x08000400, 0x00000408, 0x08000408,
326 0x00020000, 0x08020000, 0x00020008, 0x08020008,
327 0x00020400, 0x08020400, 0x00020408, 0x08020408,
328 0x00000001, 0x08000001, 0x00000009, 0x08000009,
329 0x00000401, 0x08000401, 0x00000409, 0x08000409,
330 0x00020001, 0x08020001, 0x00020009, 0x08020009,
331 0x00020401, 0x08020401, 0x00020409, 0x08020409,
332 0x02000000, 0x0A000000, 0x02000008, 0x0A000008,
333 0x02000400, 0x0A000400, 0x02000408, 0x0A000408,
334 0x02020000, 0x0A020000, 0x02020008, 0x0A020008,
335 0x02020400, 0x0A020400, 0x02020408, 0x0A020408,
336 0x02000001, 0x0A000001, 0x02000009, 0x0A000009,
337 0x02000401, 0x0A000401, 0x02000409, 0x0A000409,
338 0x02020001, 0x0A020001, 0x02020009, 0x0A020009,
339 0x02020401, 0x0A020401, 0x02020409, 0x0A020409,
342 0x00000000, 0x00000100, 0x00080000, 0x00080100,
343 0x01000000, 0x01000100, 0x01080000, 0x01080100,
344 0x00000010, 0x00000110, 0x00080010, 0x00080110,
345 0x01000010, 0x01000110, 0x01080010, 0x01080110,
346 0x00200000, 0x00200100, 0x00280000, 0x00280100,
347 0x01200000, 0x01200100, 0x01280000, 0x01280100,
348 0x00200010, 0x00200110, 0x00280010, 0x00280110,
349 0x01200010, 0x01200110, 0x01280010, 0x01280110,
350 0x00000200, 0x00000300, 0x00080200, 0x00080300,
351 0x01000200, 0x01000300, 0x01080200, 0x01080300,
352 0x00000210, 0x00000310, 0x00080210, 0x00080310,
353 0x01000210, 0x01000310, 0x01080210, 0x01080310,
354 0x00200200, 0x00200300, 0x00280200, 0x00280300,
355 0x01200200, 0x01200300, 0x01280200, 0x01280300,
356 0x00200210, 0x00200310, 0x00280210, 0x00280310,
357 0x01200210, 0x01200310, 0x01280210, 0x01280310,
360 0x00000000, 0x04000000, 0x00040000, 0x04040000,
361 0x00000002, 0x04000002, 0x00040002, 0x04040002,
362 0x00002000, 0x04002000, 0x00042000, 0x04042000,
363 0x00002002, 0x04002002, 0x00042002, 0x04042002,
364 0x00000020, 0x04000020, 0x00040020, 0x04040020,
365 0x00000022, 0x04000022, 0x00040022, 0x04040022,
366 0x00002020, 0x04002020, 0x00042020, 0x04042020,
367 0x00002022, 0x04002022, 0x00042022, 0x04042022,
368 0x00000800, 0x04000800, 0x00040800, 0x04040800,
369 0x00000802, 0x04000802, 0x00040802, 0x04040802,
370 0x00002800, 0x04002800, 0x00042800, 0x04042800,
371 0x00002802, 0x04002802, 0x00042802, 0x04042802,
372 0x00000820, 0x04000820, 0x00040820, 0x04040820,
373 0x00000822, 0x04000822, 0x00040822, 0x04040822,
374 0x00002820, 0x04002820, 0x00042820, 0x04042820,
375 0x00002822, 0x04002822, 0x00042822, 0x04042822
380 #define BOX(i,n,S) (S)[(n)][(i)]
382 #define BOX(i,n,S) (u32x) ((S)[(n)][(i).s0], (S)[(n)][(i).s1])
384 #define BOX(i,n,S) (u32x) ((S)[(n)][(i).s0], (S)[(n)][(i).s1], (S)[(n)][(i).s2], (S)[(n)][(i).s3])
386 #define BOX(i,n,S) (u32x) ((S)[(n)][(i).s0], (S)[(n)][(i).s1], (S)[(n)][(i).s2], (S)[(n)][(i).s3], (S)[(n)][(i).s4], (S)[(n)][(i).s5], (S)[(n)][(i).s6], (S)[(n)][(i).s7])
390 #define BOX1(i,S) (S)[(i)]
392 #define BOX1(i,S) (u32x) ((S)[(i).s0], (S)[(i).s1])
394 #define BOX1(i,S) (u32x) ((S)[(i).s0], (S)[(i).s1], (S)[(i).s2], (S)[(i).s3])
396 #define BOX1(i,S) (u32x) ((S)[(i).s0], (S)[(i).s1], (S)[(i).s2], (S)[(i).s3], (S)[(i).s4], (S)[(i).s5], (S)[(i).s6], (S)[(i).s7])
399 static void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 s_SPtrans[8][64])
405 for (u32 i = 0; i < 16; i += 2)
411 t = Kd[i + 0] ^ rotl32 (r, 28u);
413 l ^= BOX (((u >> 2) & 0x3f), 0, s_SPtrans)
414 | BOX (((u >> 10) & 0x3f), 2, s_SPtrans)
415 | BOX (((u >> 18) & 0x3f), 4, s_SPtrans)
416 | BOX (((u >> 26) & 0x3f), 6, s_SPtrans)
417 | BOX (((t >> 2) & 0x3f), 1, s_SPtrans)
418 | BOX (((t >> 10) & 0x3f), 3, s_SPtrans)
419 | BOX (((t >> 18) & 0x3f), 5, s_SPtrans)
420 | BOX (((t >> 26) & 0x3f), 7, s_SPtrans);
423 t = Kd[i + 1] ^ rotl32 (l, 28u);
425 r ^= BOX (((u >> 2) & 0x3f), 0, s_SPtrans)
426 | BOX (((u >> 10) & 0x3f), 2, s_SPtrans)
427 | BOX (((u >> 18) & 0x3f), 4, s_SPtrans)
428 | BOX (((u >> 26) & 0x3f), 6, s_SPtrans)
429 | BOX (((t >> 2) & 0x3f), 1, s_SPtrans)
430 | BOX (((t >> 10) & 0x3f), 3, s_SPtrans)
431 | BOX (((t >> 18) & 0x3f), 5, s_SPtrans)
432 | BOX (((t >> 26) & 0x3f), 7, s_SPtrans);
439 static void _des_crypt_keysetup (u32x c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 s_skb[8][64])
443 PERM_OP (d, c, tt, 4, 0x0f0f0f0f);
444 HPERM_OP (c, tt, 2, 0xcccc0000);
445 HPERM_OP (d, tt, 2, 0xcccc0000);
446 PERM_OP (d, c, tt, 1, 0x55555555);
447 PERM_OP (c, d, tt, 8, 0x00ff00ff);
448 PERM_OP (d, c, tt, 1, 0x55555555);
450 d = ((d & 0x000000ff) << 16)
451 | ((d & 0x0000ff00) << 0)
452 | ((d & 0x00ff0000) >> 16)
453 | ((c & 0xf0000000) >> 4);
458 for (u32 i = 0; i < 16; i++)
460 if ((i < 2) || (i == 8) || (i == 15))
462 c = ((c >> 1) | (c << 27));
463 d = ((d >> 1) | (d << 27));
467 c = ((c >> 2) | (c << 26));
468 d = ((d >> 2) | (d << 26));
474 const u32x c00 = (c >> 0) & 0x0000003f;
475 const u32x c06 = (c >> 6) & 0x00383003;
476 const u32x c07 = (c >> 7) & 0x0000003c;
477 const u32x c13 = (c >> 13) & 0x0000060f;
478 const u32x c20 = (c >> 20) & 0x00000001;
480 u32x s = BOX (((c00 >> 0) & 0xff), 0, s_skb)
481 | BOX (((c06 >> 0) & 0xff)
482 |((c07 >> 0) & 0xff), 1, s_skb)
483 | BOX (((c13 >> 0) & 0xff)
484 |((c06 >> 8) & 0xff), 2, s_skb)
485 | BOX (((c20 >> 0) & 0xff)
487 |((c06 >> 16) & 0xff), 3, s_skb);
489 const u32x d00 = (d >> 0) & 0x00003c3f;
490 const u32x d07 = (d >> 7) & 0x00003f03;
491 const u32x d21 = (d >> 21) & 0x0000000f;
492 const u32x d22 = (d >> 22) & 0x00000030;
494 u32x t = BOX (((d00 >> 0) & 0xff), 4, s_skb)
495 | BOX (((d07 >> 0) & 0xff)
496 |((d00 >> 8) & 0xff), 5, s_skb)
497 | BOX (((d07 >> 8) & 0xff), 6, s_skb)
498 | BOX (((d21 >> 0) & 0xff)
499 |((d22 >> 0) & 0xff), 7, s_skb);
501 Kc[i] = ((t << 16) | (s & 0x0000ffff));
502 Kd[i] = ((s >> 16) | (t & 0xffff0000));
504 Kc[i] = rotl32 (Kc[i], 2u);
505 Kd[i] = rotl32 (Kd[i], 2u);
509 static void transform_racf_key (const u32x w0, const u32x w1, u32x key[2])
511 key[0] = BOX1 (((w0 >> 0) & 0xff), ascii_to_ebcdic_pc) << 0
512 | BOX1 (((w0 >> 8) & 0xff), ascii_to_ebcdic_pc) << 8
513 | BOX1 (((w0 >> 16) & 0xff), ascii_to_ebcdic_pc) << 16
514 | BOX1 (((w0 >> 24) & 0xff), ascii_to_ebcdic_pc) << 24;
516 key[1] = BOX1 (((w1 >> 0) & 0xff), ascii_to_ebcdic_pc) << 0
517 | BOX1 (((w1 >> 8) & 0xff), ascii_to_ebcdic_pc) << 8
518 | BOX1 (((w1 >> 16) & 0xff), ascii_to_ebcdic_pc) << 16
519 | BOX1 (((w1 >> 24) & 0xff), ascii_to_ebcdic_pc) << 24;
522 static void m08500m (__local u32 s_SPtrans[8][64], __local u32 s_skb[8][64], u32 w[16], const u32 pw_len, __global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant u32x * words_buf_r, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 bfs_cnt, const u32 digests_cnt, const u32 digests_offset)
528 const u32 gid = get_global_id (0);
529 const u32 lid = get_local_id (0);
537 salt_buf0[0] = salt_bufs[salt_pos].salt_buf_pc[0];
538 salt_buf0[1] = salt_bufs[salt_pos].salt_buf_pc[1];
548 for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos += VECT_SIZE)
550 const u32x w0r = words_buf_r[il_pos / VECT_SIZE];
552 const u32x w0 = w0l | w0r;
556 transform_racf_key (w0, w1, key);
558 const u32x c = key[0];
559 const u32x d = key[1];
564 _des_crypt_keysetup (c, d, Kc, Kd, s_skb);
568 data[0] = salt_buf0[0];
569 data[1] = salt_buf0[1];
573 _des_crypt_encrypt (iv, data, Kc, Kd, s_SPtrans);
578 COMPARE_M_SIMD (iv[0], iv[1], iv2, iv3);
582 static void m08500s (__local u32 s_SPtrans[8][64], __local u32 s_skb[8][64], u32 w[16], const u32 pw_len, __global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant u32x * words_buf_r, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 bfs_cnt, const u32 digests_cnt, const u32 digests_offset)
588 const u32 gid = get_global_id (0);
589 const u32 lid = get_local_id (0);
597 salt_buf0[0] = salt_bufs[salt_pos].salt_buf_pc[0];
598 salt_buf0[1] = salt_bufs[salt_pos].salt_buf_pc[1];
604 const u32 search[4] =
606 digests_buf[digests_offset].digest_buf[DGST_R0],
607 digests_buf[digests_offset].digest_buf[DGST_R1],
608 digests_buf[digests_offset].digest_buf[DGST_R2],
609 digests_buf[digests_offset].digest_buf[DGST_R3]
620 for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos += VECT_SIZE)
622 const u32x w0r = words_buf_r[il_pos / VECT_SIZE];
624 const u32x w0 = w0l | w0r;
628 transform_racf_key (w0, w1, key);
630 const u32x c = key[0];
631 const u32x d = key[1];
636 _des_crypt_keysetup (c, d, Kc, Kd, s_skb);
640 data[0] = salt_buf0[0];
641 data[1] = salt_buf0[1];
645 _des_crypt_encrypt (iv, data, Kc, Kd, s_SPtrans);
650 COMPARE_S_SIMD (iv[0], iv[1], iv2, iv3);
654 __kernel void m08500_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant u32x * words_buf_r, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
660 const u32 gid = get_global_id (0);
661 const u32 lid = get_local_id (0);
662 const u32 lsz = get_local_size (0);
668 __local u32 s_SPtrans[8][64];
669 __local u32 s_skb[8][64];
671 for (u32 i = lid; i < 64; i += lsz)
673 s_SPtrans[0][i] = c_SPtrans[0][i];
674 s_SPtrans[1][i] = c_SPtrans[1][i];
675 s_SPtrans[2][i] = c_SPtrans[2][i];
676 s_SPtrans[3][i] = c_SPtrans[3][i];
677 s_SPtrans[4][i] = c_SPtrans[4][i];
678 s_SPtrans[5][i] = c_SPtrans[5][i];
679 s_SPtrans[6][i] = c_SPtrans[6][i];
680 s_SPtrans[7][i] = c_SPtrans[7][i];
682 s_skb[0][i] = c_skb[0][i];
683 s_skb[1][i] = c_skb[1][i];
684 s_skb[2][i] = c_skb[2][i];
685 s_skb[3][i] = c_skb[3][i];
686 s_skb[4][i] = c_skb[4][i];
687 s_skb[5][i] = c_skb[5][i];
688 s_skb[6][i] = c_skb[6][i];
689 s_skb[7][i] = c_skb[7][i];
692 barrier (CLK_LOCAL_MEM_FENCE);
694 if (gid >= gid_max) return;
702 w[ 0] = pws[gid].i[ 0];
703 w[ 1] = pws[gid].i[ 1];
719 const u32 pw_len = pws[gid].pw_len;
725 m08500m (s_SPtrans, s_skb, w, pw_len, pws, rules_buf, combs_buf, words_buf_r, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset);
728 __kernel void m08500_m08 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant u32x * words_buf_r, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
732 __kernel void m08500_m16 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant u32x * words_buf_r, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
736 __kernel void m08500_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant u32x * words_buf_r, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
742 const u32 gid = get_global_id (0);
743 const u32 lid = get_local_id (0);
744 const u32 lsz = get_local_size (0);
750 __local u32 s_SPtrans[8][64];
751 __local u32 s_skb[8][64];
753 for (u32 i = lid; i < 64; i += lsz)
755 s_SPtrans[0][i] = c_SPtrans[0][i];
756 s_SPtrans[1][i] = c_SPtrans[1][i];
757 s_SPtrans[2][i] = c_SPtrans[2][i];
758 s_SPtrans[3][i] = c_SPtrans[3][i];
759 s_SPtrans[4][i] = c_SPtrans[4][i];
760 s_SPtrans[5][i] = c_SPtrans[5][i];
761 s_SPtrans[6][i] = c_SPtrans[6][i];
762 s_SPtrans[7][i] = c_SPtrans[7][i];
764 s_skb[0][i] = c_skb[0][i];
765 s_skb[1][i] = c_skb[1][i];
766 s_skb[2][i] = c_skb[2][i];
767 s_skb[3][i] = c_skb[3][i];
768 s_skb[4][i] = c_skb[4][i];
769 s_skb[5][i] = c_skb[5][i];
770 s_skb[6][i] = c_skb[6][i];
771 s_skb[7][i] = c_skb[7][i];
774 barrier (CLK_LOCAL_MEM_FENCE);
776 if (gid >= gid_max) return;
784 w[ 0] = pws[gid].i[ 0];
785 w[ 1] = pws[gid].i[ 1];
801 const u32 pw_len = pws[gid].pw_len;
807 m08500s (s_SPtrans, s_skb, w, pw_len, pws, rules_buf, combs_buf, words_buf_r, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset);
810 __kernel void m08500_s08 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant u32x * words_buf_r, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
814 __kernel void m08500_s16 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant u32x * words_buf_r, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)