2 * Author......: Jens Steube <jens.steube@gmail.com>
8 #include "include/constants.h"
9 #include "include/kernel_vendor.h"
16 #include "include/kernel_functions.c"
17 #include "types_ocl.c"
19 #include "include/rp_gpu.h"
22 #define COMPARE_S "check_single_comp4.c"
23 #define COMPARE_M "check_multi_comp4.c"
25 #define PERM_OP(a,b,tt,n,m) \
35 #define HPERM_OP(a,tt,n,m) \
41 tt = tt >> (16 + n); \
45 __constant u32 c_SPtrans[8][64] =
48 0x02080800, 0x00080000, 0x02000002, 0x02080802,
49 0x02000000, 0x00080802, 0x00080002, 0x02000002,
50 0x00080802, 0x02080800, 0x02080000, 0x00000802,
51 0x02000802, 0x02000000, 0x00000000, 0x00080002,
52 0x00080000, 0x00000002, 0x02000800, 0x00080800,
53 0x02080802, 0x02080000, 0x00000802, 0x02000800,
54 0x00000002, 0x00000800, 0x00080800, 0x02080002,
55 0x00000800, 0x02000802, 0x02080002, 0x00000000,
56 0x00000000, 0x02080802, 0x02000800, 0x00080002,
57 0x02080800, 0x00080000, 0x00000802, 0x02000800,
58 0x02080002, 0x00000800, 0x00080800, 0x02000002,
59 0x00080802, 0x00000002, 0x02000002, 0x02080000,
60 0x02080802, 0x00080800, 0x02080000, 0x02000802,
61 0x02000000, 0x00000802, 0x00080002, 0x00000000,
62 0x00080000, 0x02000000, 0x02000802, 0x02080800,
63 0x00000002, 0x02080002, 0x00000800, 0x00080802,
65 0x40108010, 0x00000000, 0x00108000, 0x40100000,
66 0x40000010, 0x00008010, 0x40008000, 0x00108000,
67 0x00008000, 0x40100010, 0x00000010, 0x40008000,
68 0x00100010, 0x40108000, 0x40100000, 0x00000010,
69 0x00100000, 0x40008010, 0x40100010, 0x00008000,
70 0x00108010, 0x40000000, 0x00000000, 0x00100010,
71 0x40008010, 0x00108010, 0x40108000, 0x40000010,
72 0x40000000, 0x00100000, 0x00008010, 0x40108010,
73 0x00100010, 0x40108000, 0x40008000, 0x00108010,
74 0x40108010, 0x00100010, 0x40000010, 0x00000000,
75 0x40000000, 0x00008010, 0x00100000, 0x40100010,
76 0x00008000, 0x40000000, 0x00108010, 0x40008010,
77 0x40108000, 0x00008000, 0x00000000, 0x40000010,
78 0x00000010, 0x40108010, 0x00108000, 0x40100000,
79 0x40100010, 0x00100000, 0x00008010, 0x40008000,
80 0x40008010, 0x00000010, 0x40100000, 0x00108000,
82 0x04000001, 0x04040100, 0x00000100, 0x04000101,
83 0x00040001, 0x04000000, 0x04000101, 0x00040100,
84 0x04000100, 0x00040000, 0x04040000, 0x00000001,
85 0x04040101, 0x00000101, 0x00000001, 0x04040001,
86 0x00000000, 0x00040001, 0x04040100, 0x00000100,
87 0x00000101, 0x04040101, 0x00040000, 0x04000001,
88 0x04040001, 0x04000100, 0x00040101, 0x04040000,
89 0x00040100, 0x00000000, 0x04000000, 0x00040101,
90 0x04040100, 0x00000100, 0x00000001, 0x00040000,
91 0x00000101, 0x00040001, 0x04040000, 0x04000101,
92 0x00000000, 0x04040100, 0x00040100, 0x04040001,
93 0x00040001, 0x04000000, 0x04040101, 0x00000001,
94 0x00040101, 0x04000001, 0x04000000, 0x04040101,
95 0x00040000, 0x04000100, 0x04000101, 0x00040100,
96 0x04000100, 0x00000000, 0x04040001, 0x00000101,
97 0x04000001, 0x00040101, 0x00000100, 0x04040000,
99 0x00401008, 0x10001000, 0x00000008, 0x10401008,
100 0x00000000, 0x10400000, 0x10001008, 0x00400008,
101 0x10401000, 0x10000008, 0x10000000, 0x00001008,
102 0x10000008, 0x00401008, 0x00400000, 0x10000000,
103 0x10400008, 0x00401000, 0x00001000, 0x00000008,
104 0x00401000, 0x10001008, 0x10400000, 0x00001000,
105 0x00001008, 0x00000000, 0x00400008, 0x10401000,
106 0x10001000, 0x10400008, 0x10401008, 0x00400000,
107 0x10400008, 0x00001008, 0x00400000, 0x10000008,
108 0x00401000, 0x10001000, 0x00000008, 0x10400000,
109 0x10001008, 0x00000000, 0x00001000, 0x00400008,
110 0x00000000, 0x10400008, 0x10401000, 0x00001000,
111 0x10000000, 0x10401008, 0x00401008, 0x00400000,
112 0x10401008, 0x00000008, 0x10001000, 0x00401008,
113 0x00400008, 0x00401000, 0x10400000, 0x10001008,
114 0x00001008, 0x10000000, 0x10000008, 0x10401000,
116 0x08000000, 0x00010000, 0x00000400, 0x08010420,
117 0x08010020, 0x08000400, 0x00010420, 0x08010000,
118 0x00010000, 0x00000020, 0x08000020, 0x00010400,
119 0x08000420, 0x08010020, 0x08010400, 0x00000000,
120 0x00010400, 0x08000000, 0x00010020, 0x00000420,
121 0x08000400, 0x00010420, 0x00000000, 0x08000020,
122 0x00000020, 0x08000420, 0x08010420, 0x00010020,
123 0x08010000, 0x00000400, 0x00000420, 0x08010400,
124 0x08010400, 0x08000420, 0x00010020, 0x08010000,
125 0x00010000, 0x00000020, 0x08000020, 0x08000400,
126 0x08000000, 0x00010400, 0x08010420, 0x00000000,
127 0x00010420, 0x08000000, 0x00000400, 0x00010020,
128 0x08000420, 0x00000400, 0x00000000, 0x08010420,
129 0x08010020, 0x08010400, 0x00000420, 0x00010000,
130 0x00010400, 0x08010020, 0x08000400, 0x00000420,
131 0x00000020, 0x00010420, 0x08010000, 0x08000020,
133 0x80000040, 0x00200040, 0x00000000, 0x80202000,
134 0x00200040, 0x00002000, 0x80002040, 0x00200000,
135 0x00002040, 0x80202040, 0x00202000, 0x80000000,
136 0x80002000, 0x80000040, 0x80200000, 0x00202040,
137 0x00200000, 0x80002040, 0x80200040, 0x00000000,
138 0x00002000, 0x00000040, 0x80202000, 0x80200040,
139 0x80202040, 0x80200000, 0x80000000, 0x00002040,
140 0x00000040, 0x00202000, 0x00202040, 0x80002000,
141 0x00002040, 0x80000000, 0x80002000, 0x00202040,
142 0x80202000, 0x00200040, 0x00000000, 0x80002000,
143 0x80000000, 0x00002000, 0x80200040, 0x00200000,
144 0x00200040, 0x80202040, 0x00202000, 0x00000040,
145 0x80202040, 0x00202000, 0x00200000, 0x80002040,
146 0x80000040, 0x80200000, 0x00202040, 0x00000000,
147 0x00002000, 0x80000040, 0x80002040, 0x80202000,
148 0x80200000, 0x00002040, 0x00000040, 0x80200040,
150 0x00004000, 0x00000200, 0x01000200, 0x01000004,
151 0x01004204, 0x00004004, 0x00004200, 0x00000000,
152 0x01000000, 0x01000204, 0x00000204, 0x01004000,
153 0x00000004, 0x01004200, 0x01004000, 0x00000204,
154 0x01000204, 0x00004000, 0x00004004, 0x01004204,
155 0x00000000, 0x01000200, 0x01000004, 0x00004200,
156 0x01004004, 0x00004204, 0x01004200, 0x00000004,
157 0x00004204, 0x01004004, 0x00000200, 0x01000000,
158 0x00004204, 0x01004000, 0x01004004, 0x00000204,
159 0x00004000, 0x00000200, 0x01000000, 0x01004004,
160 0x01000204, 0x00004204, 0x00004200, 0x00000000,
161 0x00000200, 0x01000004, 0x00000004, 0x01000200,
162 0x00000000, 0x01000204, 0x01000200, 0x00004200,
163 0x00000204, 0x00004000, 0x01004204, 0x01000000,
164 0x01004200, 0x00000004, 0x00004004, 0x01004204,
165 0x01000004, 0x01004200, 0x01004000, 0x00004004,
167 0x20800080, 0x20820000, 0x00020080, 0x00000000,
168 0x20020000, 0x00800080, 0x20800000, 0x20820080,
169 0x00000080, 0x20000000, 0x00820000, 0x00020080,
170 0x00820080, 0x20020080, 0x20000080, 0x20800000,
171 0x00020000, 0x00820080, 0x00800080, 0x20020000,
172 0x20820080, 0x20000080, 0x00000000, 0x00820000,
173 0x20000000, 0x00800000, 0x20020080, 0x20800080,
174 0x00800000, 0x00020000, 0x20820000, 0x00000080,
175 0x00800000, 0x00020000, 0x20000080, 0x20820080,
176 0x00020080, 0x20000000, 0x00000000, 0x00820000,
177 0x20800080, 0x20020080, 0x20020000, 0x00800080,
178 0x20820000, 0x00000080, 0x00800080, 0x20020000,
179 0x20820080, 0x00800000, 0x20800000, 0x20000080,
180 0x00820000, 0x00020080, 0x20020080, 0x20800000,
181 0x00000080, 0x20820000, 0x00820080, 0x00000000,
182 0x20000000, 0x20800080, 0x00020000, 0x00820080,
185 __constant u32 c_skb[8][64] =
187 /* for C bits (numbered as per FIPS 46) 1 2 3 4 5 6 */
188 0x00000000, 0x00000010, 0x20000000, 0x20000010,
189 0x00010000, 0x00010010, 0x20010000, 0x20010010,
190 0x00000800, 0x00000810, 0x20000800, 0x20000810,
191 0x00010800, 0x00010810, 0x20010800, 0x20010810,
192 0x00000020, 0x00000030, 0x20000020, 0x20000030,
193 0x00010020, 0x00010030, 0x20010020, 0x20010030,
194 0x00000820, 0x00000830, 0x20000820, 0x20000830,
195 0x00010820, 0x00010830, 0x20010820, 0x20010830,
196 0x00080000, 0x00080010, 0x20080000, 0x20080010,
197 0x00090000, 0x00090010, 0x20090000, 0x20090010,
198 0x00080800, 0x00080810, 0x20080800, 0x20080810,
199 0x00090800, 0x00090810, 0x20090800, 0x20090810,
200 0x00080020, 0x00080030, 0x20080020, 0x20080030,
201 0x00090020, 0x00090030, 0x20090020, 0x20090030,
202 0x00080820, 0x00080830, 0x20080820, 0x20080830,
203 0x00090820, 0x00090830, 0x20090820, 0x20090830,
204 /* for C bits (numbered as per FIPS 46) 7 8 10 11 12 13 */
205 0x00000000, 0x02000000, 0x00002000, 0x02002000,
206 0x00200000, 0x02200000, 0x00202000, 0x02202000,
207 0x00000004, 0x02000004, 0x00002004, 0x02002004,
208 0x00200004, 0x02200004, 0x00202004, 0x02202004,
209 0x00000400, 0x02000400, 0x00002400, 0x02002400,
210 0x00200400, 0x02200400, 0x00202400, 0x02202400,
211 0x00000404, 0x02000404, 0x00002404, 0x02002404,
212 0x00200404, 0x02200404, 0x00202404, 0x02202404,
213 0x10000000, 0x12000000, 0x10002000, 0x12002000,
214 0x10200000, 0x12200000, 0x10202000, 0x12202000,
215 0x10000004, 0x12000004, 0x10002004, 0x12002004,
216 0x10200004, 0x12200004, 0x10202004, 0x12202004,
217 0x10000400, 0x12000400, 0x10002400, 0x12002400,
218 0x10200400, 0x12200400, 0x10202400, 0x12202400,
219 0x10000404, 0x12000404, 0x10002404, 0x12002404,
220 0x10200404, 0x12200404, 0x10202404, 0x12202404,
221 /* for C bits (numbered as per FIPS 46) 14 15 16 17 19 20 */
222 0x00000000, 0x00000001, 0x00040000, 0x00040001,
223 0x01000000, 0x01000001, 0x01040000, 0x01040001,
224 0x00000002, 0x00000003, 0x00040002, 0x00040003,
225 0x01000002, 0x01000003, 0x01040002, 0x01040003,
226 0x00000200, 0x00000201, 0x00040200, 0x00040201,
227 0x01000200, 0x01000201, 0x01040200, 0x01040201,
228 0x00000202, 0x00000203, 0x00040202, 0x00040203,
229 0x01000202, 0x01000203, 0x01040202, 0x01040203,
230 0x08000000, 0x08000001, 0x08040000, 0x08040001,
231 0x09000000, 0x09000001, 0x09040000, 0x09040001,
232 0x08000002, 0x08000003, 0x08040002, 0x08040003,
233 0x09000002, 0x09000003, 0x09040002, 0x09040003,
234 0x08000200, 0x08000201, 0x08040200, 0x08040201,
235 0x09000200, 0x09000201, 0x09040200, 0x09040201,
236 0x08000202, 0x08000203, 0x08040202, 0x08040203,
237 0x09000202, 0x09000203, 0x09040202, 0x09040203,
238 /* for C bits (numbered as per FIPS 46) 21 23 24 26 27 28 */
239 0x00000000, 0x00100000, 0x00000100, 0x00100100,
240 0x00000008, 0x00100008, 0x00000108, 0x00100108,
241 0x00001000, 0x00101000, 0x00001100, 0x00101100,
242 0x00001008, 0x00101008, 0x00001108, 0x00101108,
243 0x04000000, 0x04100000, 0x04000100, 0x04100100,
244 0x04000008, 0x04100008, 0x04000108, 0x04100108,
245 0x04001000, 0x04101000, 0x04001100, 0x04101100,
246 0x04001008, 0x04101008, 0x04001108, 0x04101108,
247 0x00020000, 0x00120000, 0x00020100, 0x00120100,
248 0x00020008, 0x00120008, 0x00020108, 0x00120108,
249 0x00021000, 0x00121000, 0x00021100, 0x00121100,
250 0x00021008, 0x00121008, 0x00021108, 0x00121108,
251 0x04020000, 0x04120000, 0x04020100, 0x04120100,
252 0x04020008, 0x04120008, 0x04020108, 0x04120108,
253 0x04021000, 0x04121000, 0x04021100, 0x04121100,
254 0x04021008, 0x04121008, 0x04021108, 0x04121108,
255 /* for D bits (numbered as per FIPS 46) 1 2 3 4 5 6 */
256 0x00000000, 0x10000000, 0x00010000, 0x10010000,
257 0x00000004, 0x10000004, 0x00010004, 0x10010004,
258 0x20000000, 0x30000000, 0x20010000, 0x30010000,
259 0x20000004, 0x30000004, 0x20010004, 0x30010004,
260 0x00100000, 0x10100000, 0x00110000, 0x10110000,
261 0x00100004, 0x10100004, 0x00110004, 0x10110004,
262 0x20100000, 0x30100000, 0x20110000, 0x30110000,
263 0x20100004, 0x30100004, 0x20110004, 0x30110004,
264 0x00001000, 0x10001000, 0x00011000, 0x10011000,
265 0x00001004, 0x10001004, 0x00011004, 0x10011004,
266 0x20001000, 0x30001000, 0x20011000, 0x30011000,
267 0x20001004, 0x30001004, 0x20011004, 0x30011004,
268 0x00101000, 0x10101000, 0x00111000, 0x10111000,
269 0x00101004, 0x10101004, 0x00111004, 0x10111004,
270 0x20101000, 0x30101000, 0x20111000, 0x30111000,
271 0x20101004, 0x30101004, 0x20111004, 0x30111004,
272 /* for D bits (numbered as per FIPS 46) 8 9 11 12 13 14 */
273 0x00000000, 0x08000000, 0x00000008, 0x08000008,
274 0x00000400, 0x08000400, 0x00000408, 0x08000408,
275 0x00020000, 0x08020000, 0x00020008, 0x08020008,
276 0x00020400, 0x08020400, 0x00020408, 0x08020408,
277 0x00000001, 0x08000001, 0x00000009, 0x08000009,
278 0x00000401, 0x08000401, 0x00000409, 0x08000409,
279 0x00020001, 0x08020001, 0x00020009, 0x08020009,
280 0x00020401, 0x08020401, 0x00020409, 0x08020409,
281 0x02000000, 0x0A000000, 0x02000008, 0x0A000008,
282 0x02000400, 0x0A000400, 0x02000408, 0x0A000408,
283 0x02020000, 0x0A020000, 0x02020008, 0x0A020008,
284 0x02020400, 0x0A020400, 0x02020408, 0x0A020408,
285 0x02000001, 0x0A000001, 0x02000009, 0x0A000009,
286 0x02000401, 0x0A000401, 0x02000409, 0x0A000409,
287 0x02020001, 0x0A020001, 0x02020009, 0x0A020009,
288 0x02020401, 0x0A020401, 0x02020409, 0x0A020409,
289 /* for D bits (numbered as per FIPS 46) 16 17 18 19 20 21 */
290 0x00000000, 0x00000100, 0x00080000, 0x00080100,
291 0x01000000, 0x01000100, 0x01080000, 0x01080100,
292 0x00000010, 0x00000110, 0x00080010, 0x00080110,
293 0x01000010, 0x01000110, 0x01080010, 0x01080110,
294 0x00200000, 0x00200100, 0x00280000, 0x00280100,
295 0x01200000, 0x01200100, 0x01280000, 0x01280100,
296 0x00200010, 0x00200110, 0x00280010, 0x00280110,
297 0x01200010, 0x01200110, 0x01280010, 0x01280110,
298 0x00000200, 0x00000300, 0x00080200, 0x00080300,
299 0x01000200, 0x01000300, 0x01080200, 0x01080300,
300 0x00000210, 0x00000310, 0x00080210, 0x00080310,
301 0x01000210, 0x01000310, 0x01080210, 0x01080310,
302 0x00200200, 0x00200300, 0x00280200, 0x00280300,
303 0x01200200, 0x01200300, 0x01280200, 0x01280300,
304 0x00200210, 0x00200310, 0x00280210, 0x00280310,
305 0x01200210, 0x01200310, 0x01280210, 0x01280310,
306 /* for D bits (numbered as per FIPS 46) 22 23 24 25 27 28 */
307 0x00000000, 0x04000000, 0x00040000, 0x04040000,
308 0x00000002, 0x04000002, 0x00040002, 0x04040002,
309 0x00002000, 0x04002000, 0x00042000, 0x04042000,
310 0x00002002, 0x04002002, 0x00042002, 0x04042002,
311 0x00000020, 0x04000020, 0x00040020, 0x04040020,
312 0x00000022, 0x04000022, 0x00040022, 0x04040022,
313 0x00002020, 0x04002020, 0x00042020, 0x04042020,
314 0x00002022, 0x04002022, 0x00042022, 0x04042022,
315 0x00000800, 0x04000800, 0x00040800, 0x04040800,
316 0x00000802, 0x04000802, 0x00040802, 0x04040802,
317 0x00002800, 0x04002800, 0x00042800, 0x04042800,
318 0x00002802, 0x04002802, 0x00042802, 0x04042802,
319 0x00000820, 0x04000820, 0x00040820, 0x04040820,
320 0x00000822, 0x04000822, 0x00040822, 0x04040822,
321 0x00002820, 0x04002820, 0x00042820, 0x04042820,
322 0x00002822, 0x04002822, 0x00042822, 0x04042822
325 __constant u32 shifts3s0[16] = { 1, 1, 2, 2, 2, 2, 2, 2, 1, 2, 2, 2, 2, 2, 2, 1 };
326 __constant u32 shifts3s1[16] = { 27, 27, 26, 26, 26, 26, 26, 26, 27, 26, 26, 26, 26, 26, 26, 27 };
328 #define LM_IV_0_IP_RR3 0x2400b807
329 #define LM_IV_1_IP_RR3 0xaa190747
332 #define BOX(i,n,S) u32 ((S)[(n)][(i)])
336 #define BOX(i,n,S) u32 ((S)[(n)][(i).s0], (S)[(n)][(i).s1])
340 #define BOX(i,n,S) u32 ((S)[(n)][(i).s0], (S)[(n)][(i).s1], (S)[(n)][(i).s2], (S)[(n)][(i).s3])
343 static void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local u32 s_SPtrans[8][64])
349 for (u32 i = 0; i < 16; i += 2)
354 u = Kc[i + 0] ^ rotl32 (r, 30u);
355 t = Kd[i + 0] ^ rotl32 (r, 26u);
358 ^ BOX (amd_bfe (u, 0, 6), 0, s_SPtrans)
359 ^ BOX (amd_bfe (u, 8, 6), 2, s_SPtrans)
360 ^ BOX (amd_bfe (u, 16, 6), 4, s_SPtrans)
361 ^ BOX (amd_bfe (u, 24, 6), 6, s_SPtrans)
362 ^ BOX (amd_bfe (t, 0, 6), 1, s_SPtrans)
363 ^ BOX (amd_bfe (t, 8, 6), 3, s_SPtrans)
364 ^ BOX (amd_bfe (t, 16, 6), 5, s_SPtrans)
365 ^ BOX (amd_bfe (t, 24, 6), 7, s_SPtrans);
367 u = Kc[i + 1] ^ rotl32 (l, 30u);
368 t = Kd[i + 1] ^ rotl32 (l, 26u);
371 ^ BOX (amd_bfe (u, 0, 6), 0, s_SPtrans)
372 ^ BOX (amd_bfe (u, 8, 6), 2, s_SPtrans)
373 ^ BOX (amd_bfe (u, 16, 6), 4, s_SPtrans)
374 ^ BOX (amd_bfe (u, 24, 6), 6, s_SPtrans)
375 ^ BOX (amd_bfe (t, 0, 6), 1, s_SPtrans)
376 ^ BOX (amd_bfe (t, 8, 6), 3, s_SPtrans)
377 ^ BOX (amd_bfe (t, 16, 6), 5, s_SPtrans)
378 ^ BOX (amd_bfe (t, 24, 6), 7, s_SPtrans);
381 iv[0] = rotl32 (l, 29);
382 iv[1] = rotl32 (r, 29);
385 static void _des_crypt_keysetup (u32 c, u32 d, u32 Kc[16], u32 Kd[16], __local u32 s_skb[8][64])
389 PERM_OP (d, c, tt, 4, 0x0f0f0f0f);
390 HPERM_OP (c, tt, 2, 0xcccc0000);
391 HPERM_OP (d, tt, 2, 0xcccc0000);
392 PERM_OP (d, c, tt, 1, 0x55555555);
393 PERM_OP (c, d, tt, 8, 0x00ff00ff);
394 PERM_OP (d, c, tt, 1, 0x55555555);
396 d = ((d & 0x000000ff) << 16)
397 | ((d & 0x0000ff00) << 0)
398 | ((d & 0x00ff0000) >> 16)
399 | ((c & 0xf0000000) >> 4);
404 for (u32 i = 0; i < 16; i++)
406 c = c >> shifts3s0[i] | c << shifts3s1[i];
407 d = d >> shifts3s0[i] | d << shifts3s1[i];
412 const u32 c00 = (c >> 0) & 0x0000003f;
413 const u32 c06 = (c >> 6) & 0x00383003;
414 const u32 c07 = (c >> 7) & 0x0000003c;
415 const u32 c13 = (c >> 13) & 0x0000060f;
416 const u32 c20 = (c >> 20) & 0x00000001;
418 u32 s = BOX (((c00 >> 0) & 0xff), 0, s_skb)
419 | BOX (((c06 >> 0) & 0xff)
420 |((c07 >> 0) & 0xff), 1, s_skb)
421 | BOX (((c13 >> 0) & 0xff)
422 |((c06 >> 8) & 0xff), 2, s_skb)
423 | BOX (((c20 >> 0) & 0xff)
425 |((c06 >> 16) & 0xff), 3, s_skb);
427 const u32 d00 = (d >> 0) & 0x00003c3f;
428 const u32 d07 = (d >> 7) & 0x00003f03;
429 const u32 d21 = (d >> 21) & 0x0000000f;
430 const u32 d22 = (d >> 22) & 0x00000030;
432 u32 t = BOX (((d00 >> 0) & 0xff), 4, s_skb)
433 | BOX (((d07 >> 0) & 0xff)
434 |((d00 >> 8) & 0xff), 5, s_skb)
435 | BOX (((d07 >> 8) & 0xff), 6, s_skb)
436 | BOX (((d21 >> 0) & 0xff)
437 |((d22 >> 0) & 0xff), 7, s_skb);
439 Kc[i] = ((t << 16) | (s & 0x0000ffff));
440 Kd[i] = ((s >> 16) | (t & 0xffff0000));
444 static void transform_netntlmv1_key (const u32 w0, const u32 w1, u32 out[2])
447 const uchar4 t0 = as_uchar4 (w0);
448 const uchar4 t1 = as_uchar4 (w1);
453 k0.s0 = (t0.s0 >> 0);
454 k0.s1 = (t0.s0 << 7) | (t0.s1 >> 1);
455 k0.s2 = (t0.s1 << 6) | (t0.s2 >> 2);
456 k0.s3 = (t0.s2 << 5) | (t0.s3 >> 3);
457 k1.s0 = (t0.s3 << 4) | (t1.s0 >> 4);
458 k1.s1 = (t1.s0 << 3) | (t1.s1 >> 5);
459 k1.s2 = (t1.s1 << 2) | (t1.s2 >> 6);
460 k1.s3 = (t1.s2 << 1);
462 out[0] = as_uint (k0);
463 out[1] = as_uint (k1);
467 const uchar8 t0 = as_uchar8 (w0);
468 const uchar8 t1 = as_uchar8 (w1);
473 k0.s0 = (t0.s0 >> 0);
474 k0.s1 = (t0.s0 << 7) | (t0.s1 >> 1);
475 k0.s2 = (t0.s1 << 6) | (t0.s2 >> 2);
476 k0.s3 = (t0.s2 << 5) | (t0.s3 >> 3);
477 k1.s0 = (t0.s3 << 4) | (t1.s0 >> 4);
478 k1.s1 = (t1.s0 << 3) | (t1.s1 >> 5);
479 k1.s2 = (t1.s1 << 2) | (t1.s2 >> 6);
480 k1.s3 = (t1.s2 << 1);
482 k0.s4 = (t0.s4 >> 0);
483 k0.s5 = (t0.s4 << 7) | (t0.s5 >> 1);
484 k0.s6 = (t0.s5 << 6) | (t0.s6 >> 2);
485 k0.s7 = (t0.s6 << 5) | (t0.s7 >> 3);
486 k1.s4 = (t0.s7 << 4) | (t1.s4 >> 4);
487 k1.s5 = (t1.s4 << 3) | (t1.s5 >> 5);
488 k1.s6 = (t1.s5 << 2) | (t1.s6 >> 6);
489 k1.s7 = (t1.s6 << 1);
491 out[0] = as_uint2 (k0);
492 out[1] = as_uint2 (k1);
496 const uchar16 t0 = as_uchar16 (w0);
497 const uchar16 t1 = as_uchar16 (w1);
502 k0.s0 = (t0.s0 >> 0);
503 k0.s1 = (t0.s0 << 7) | (t0.s1 >> 1);
504 k0.s2 = (t0.s1 << 6) | (t0.s2 >> 2);
505 k0.s3 = (t0.s2 << 5) | (t0.s3 >> 3);
506 k1.s0 = (t0.s3 << 4) | (t1.s0 >> 4);
507 k1.s1 = (t1.s0 << 3) | (t1.s1 >> 5);
508 k1.s2 = (t1.s1 << 2) | (t1.s2 >> 6);
509 k1.s3 = (t1.s2 << 1);
511 k0.s4 = (t0.s4 >> 0);
512 k0.s5 = (t0.s4 << 7) | (t0.s5 >> 1);
513 k0.s6 = (t0.s5 << 6) | (t0.s6 >> 2);
514 k0.s7 = (t0.s6 << 5) | (t0.s7 >> 3);
515 k1.s4 = (t0.s7 << 4) | (t1.s4 >> 4);
516 k1.s5 = (t1.s4 << 3) | (t1.s5 >> 5);
517 k1.s6 = (t1.s5 << 2) | (t1.s6 >> 6);
518 k1.s7 = (t1.s6 << 1);
520 k0.s8 = (t0.s8 >> 0);
521 k0.s9 = (t0.s8 << 7) | (t0.s9 >> 1);
522 k0.sa = (t0.s9 << 6) | (t0.sa >> 2);
523 k0.sb = (t0.sa << 5) | (t0.sb >> 3);
524 k1.s8 = (t0.sb << 4) | (t1.s8 >> 4);
525 k1.s9 = (t1.s8 << 3) | (t1.s9 >> 5);
526 k1.sa = (t1.s9 << 2) | (t1.sa >> 6);
527 k1.sb = (t1.sa << 1);
529 k0.sc = (t0.sc >> 0);
530 k0.sd = (t0.sc << 7) | (t0.sd >> 1);
531 k0.se = (t0.sd << 6) | (t0.se >> 2);
532 k0.sf = (t0.se << 5) | (t0.sf >> 3);
533 k1.sc = (t0.sf << 4) | (t1.sc >> 4);
534 k1.sd = (t1.sc << 3) | (t1.sd >> 5);
535 k1.se = (t1.sd << 2) | (t1.se >> 6);
536 k1.sf = (t1.se << 1);
538 out[0] = as_uint4 (k0);
539 out[1] = as_uint4 (k1);
543 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m03000_m04 (__global pw_t *pws, __global gpu_rule_t * rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
549 const u32 lid = get_local_id (0);
555 const u32 gid = get_global_id (0);
559 pw_buf[0] = pws[gid].i[ 0];
560 pw_buf[1] = pws[gid].i[ 1];
564 const u32 pw_len = pws[gid].pw_len;
570 __local u32 s_SPtrans[8][64];
571 __local u32 s_skb[8][64];
573 s_SPtrans[0][lid] = c_SPtrans[0][lid];
574 s_SPtrans[1][lid] = c_SPtrans[1][lid];
575 s_SPtrans[2][lid] = c_SPtrans[2][lid];
576 s_SPtrans[3][lid] = c_SPtrans[3][lid];
577 s_SPtrans[4][lid] = c_SPtrans[4][lid];
578 s_SPtrans[5][lid] = c_SPtrans[5][lid];
579 s_SPtrans[6][lid] = c_SPtrans[6][lid];
580 s_SPtrans[7][lid] = c_SPtrans[7][lid];
582 s_skb[0][lid] = c_skb[0][lid];
583 s_skb[1][lid] = c_skb[1][lid];
584 s_skb[2][lid] = c_skb[2][lid];
585 s_skb[3][lid] = c_skb[3][lid];
586 s_skb[4][lid] = c_skb[4][lid];
587 s_skb[5][lid] = c_skb[5][lid];
588 s_skb[6][lid] = c_skb[6][lid];
589 s_skb[7][lid] = c_skb[7][lid];
591 barrier (CLK_LOCAL_MEM_FENCE);
593 if (gid >= gid_max) return;
599 for (u32 il_pos = 0; il_pos < rules_cnt; il_pos++)
629 u32 out_len = apply_rules (rules_buf[il_pos].cmds, w0, w1, pw_len);
631 out_len = (out_len >= 7) ? 7 : out_len;
635 transform_netntlmv1_key (w0[0], w0[1], key);
637 const u32 c = key[0];
638 const u32 d = key[1];
643 _des_crypt_keysetup (c, d, Kc, Kd, s_skb);
647 data[0] = LM_IV_0_IP_RR3;
648 data[1] = LM_IV_1_IP_RR3;
652 _des_crypt_encrypt (iv, data, Kc, Kd, s_SPtrans);
654 const u32 r0 = iv[0];
655 const u32 r1 = iv[1];
663 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m03000_m08 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
667 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m03000_m16 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
671 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m03000_s04 (__global pw_t *pws, __global gpu_rule_t * rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
677 const u32 lid = get_local_id (0);
683 const u32 gid = get_global_id (0);
687 pw_buf[0] = pws[gid].i[ 0];
688 pw_buf[1] = pws[gid].i[ 1];
692 const u32 pw_len = pws[gid].pw_len;
698 __local u32 s_SPtrans[8][64];
699 __local u32 s_skb[8][64];
701 s_SPtrans[0][lid] = c_SPtrans[0][lid];
702 s_SPtrans[1][lid] = c_SPtrans[1][lid];
703 s_SPtrans[2][lid] = c_SPtrans[2][lid];
704 s_SPtrans[3][lid] = c_SPtrans[3][lid];
705 s_SPtrans[4][lid] = c_SPtrans[4][lid];
706 s_SPtrans[5][lid] = c_SPtrans[5][lid];
707 s_SPtrans[6][lid] = c_SPtrans[6][lid];
708 s_SPtrans[7][lid] = c_SPtrans[7][lid];
710 s_skb[0][lid] = c_skb[0][lid];
711 s_skb[1][lid] = c_skb[1][lid];
712 s_skb[2][lid] = c_skb[2][lid];
713 s_skb[3][lid] = c_skb[3][lid];
714 s_skb[4][lid] = c_skb[4][lid];
715 s_skb[5][lid] = c_skb[5][lid];
716 s_skb[6][lid] = c_skb[6][lid];
717 s_skb[7][lid] = c_skb[7][lid];
719 barrier (CLK_LOCAL_MEM_FENCE);
721 if (gid >= gid_max) return;
727 const u32 search[4] =
729 digests_buf[digests_offset].digest_buf[DGST_R0],
730 digests_buf[digests_offset].digest_buf[DGST_R1],
731 digests_buf[digests_offset].digest_buf[DGST_R2],
732 digests_buf[digests_offset].digest_buf[DGST_R3]
739 for (u32 il_pos = 0; il_pos < rules_cnt; il_pos++)
769 u32 out_len = apply_rules (rules_buf[il_pos].cmds, w0, w1, pw_len);
771 out_len = (out_len >= 7) ? 7 : out_len;
775 transform_netntlmv1_key (w0[0], w0[1], key);
777 const u32 c = key[0];
778 const u32 d = key[1];
783 _des_crypt_keysetup (c, d, Kc, Kd, s_skb);
787 data[0] = LM_IV_0_IP_RR3;
788 data[1] = LM_IV_1_IP_RR3;
792 _des_crypt_encrypt (iv, data, Kc, Kd, s_SPtrans);
794 const u32 r0 = iv[0];
795 const u32 r1 = iv[1];
803 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m03000_s08 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
807 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m03000_s16 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)