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"
20 #define COMPARE_S "check_single_comp4.c"
21 #define COMPARE_M "check_multi_comp4.c"
23 #define PERM_OP(a,b,tt,n,m) \
33 #define HPERM_OP(a,tt,n,m) \
39 tt = tt >> (16 + n); \
43 __constant u32 c_SPtrans[8][64] =
46 0x02080800, 0x00080000, 0x02000002, 0x02080802,
47 0x02000000, 0x00080802, 0x00080002, 0x02000002,
48 0x00080802, 0x02080800, 0x02080000, 0x00000802,
49 0x02000802, 0x02000000, 0x00000000, 0x00080002,
50 0x00080000, 0x00000002, 0x02000800, 0x00080800,
51 0x02080802, 0x02080000, 0x00000802, 0x02000800,
52 0x00000002, 0x00000800, 0x00080800, 0x02080002,
53 0x00000800, 0x02000802, 0x02080002, 0x00000000,
54 0x00000000, 0x02080802, 0x02000800, 0x00080002,
55 0x02080800, 0x00080000, 0x00000802, 0x02000800,
56 0x02080002, 0x00000800, 0x00080800, 0x02000002,
57 0x00080802, 0x00000002, 0x02000002, 0x02080000,
58 0x02080802, 0x00080800, 0x02080000, 0x02000802,
59 0x02000000, 0x00000802, 0x00080002, 0x00000000,
60 0x00080000, 0x02000000, 0x02000802, 0x02080800,
61 0x00000002, 0x02080002, 0x00000800, 0x00080802,
63 0x40108010, 0x00000000, 0x00108000, 0x40100000,
64 0x40000010, 0x00008010, 0x40008000, 0x00108000,
65 0x00008000, 0x40100010, 0x00000010, 0x40008000,
66 0x00100010, 0x40108000, 0x40100000, 0x00000010,
67 0x00100000, 0x40008010, 0x40100010, 0x00008000,
68 0x00108010, 0x40000000, 0x00000000, 0x00100010,
69 0x40008010, 0x00108010, 0x40108000, 0x40000010,
70 0x40000000, 0x00100000, 0x00008010, 0x40108010,
71 0x00100010, 0x40108000, 0x40008000, 0x00108010,
72 0x40108010, 0x00100010, 0x40000010, 0x00000000,
73 0x40000000, 0x00008010, 0x00100000, 0x40100010,
74 0x00008000, 0x40000000, 0x00108010, 0x40008010,
75 0x40108000, 0x00008000, 0x00000000, 0x40000010,
76 0x00000010, 0x40108010, 0x00108000, 0x40100000,
77 0x40100010, 0x00100000, 0x00008010, 0x40008000,
78 0x40008010, 0x00000010, 0x40100000, 0x00108000,
80 0x04000001, 0x04040100, 0x00000100, 0x04000101,
81 0x00040001, 0x04000000, 0x04000101, 0x00040100,
82 0x04000100, 0x00040000, 0x04040000, 0x00000001,
83 0x04040101, 0x00000101, 0x00000001, 0x04040001,
84 0x00000000, 0x00040001, 0x04040100, 0x00000100,
85 0x00000101, 0x04040101, 0x00040000, 0x04000001,
86 0x04040001, 0x04000100, 0x00040101, 0x04040000,
87 0x00040100, 0x00000000, 0x04000000, 0x00040101,
88 0x04040100, 0x00000100, 0x00000001, 0x00040000,
89 0x00000101, 0x00040001, 0x04040000, 0x04000101,
90 0x00000000, 0x04040100, 0x00040100, 0x04040001,
91 0x00040001, 0x04000000, 0x04040101, 0x00000001,
92 0x00040101, 0x04000001, 0x04000000, 0x04040101,
93 0x00040000, 0x04000100, 0x04000101, 0x00040100,
94 0x04000100, 0x00000000, 0x04040001, 0x00000101,
95 0x04000001, 0x00040101, 0x00000100, 0x04040000,
97 0x00401008, 0x10001000, 0x00000008, 0x10401008,
98 0x00000000, 0x10400000, 0x10001008, 0x00400008,
99 0x10401000, 0x10000008, 0x10000000, 0x00001008,
100 0x10000008, 0x00401008, 0x00400000, 0x10000000,
101 0x10400008, 0x00401000, 0x00001000, 0x00000008,
102 0x00401000, 0x10001008, 0x10400000, 0x00001000,
103 0x00001008, 0x00000000, 0x00400008, 0x10401000,
104 0x10001000, 0x10400008, 0x10401008, 0x00400000,
105 0x10400008, 0x00001008, 0x00400000, 0x10000008,
106 0x00401000, 0x10001000, 0x00000008, 0x10400000,
107 0x10001008, 0x00000000, 0x00001000, 0x00400008,
108 0x00000000, 0x10400008, 0x10401000, 0x00001000,
109 0x10000000, 0x10401008, 0x00401008, 0x00400000,
110 0x10401008, 0x00000008, 0x10001000, 0x00401008,
111 0x00400008, 0x00401000, 0x10400000, 0x10001008,
112 0x00001008, 0x10000000, 0x10000008, 0x10401000,
114 0x08000000, 0x00010000, 0x00000400, 0x08010420,
115 0x08010020, 0x08000400, 0x00010420, 0x08010000,
116 0x00010000, 0x00000020, 0x08000020, 0x00010400,
117 0x08000420, 0x08010020, 0x08010400, 0x00000000,
118 0x00010400, 0x08000000, 0x00010020, 0x00000420,
119 0x08000400, 0x00010420, 0x00000000, 0x08000020,
120 0x00000020, 0x08000420, 0x08010420, 0x00010020,
121 0x08010000, 0x00000400, 0x00000420, 0x08010400,
122 0x08010400, 0x08000420, 0x00010020, 0x08010000,
123 0x00010000, 0x00000020, 0x08000020, 0x08000400,
124 0x08000000, 0x00010400, 0x08010420, 0x00000000,
125 0x00010420, 0x08000000, 0x00000400, 0x00010020,
126 0x08000420, 0x00000400, 0x00000000, 0x08010420,
127 0x08010020, 0x08010400, 0x00000420, 0x00010000,
128 0x00010400, 0x08010020, 0x08000400, 0x00000420,
129 0x00000020, 0x00010420, 0x08010000, 0x08000020,
131 0x80000040, 0x00200040, 0x00000000, 0x80202000,
132 0x00200040, 0x00002000, 0x80002040, 0x00200000,
133 0x00002040, 0x80202040, 0x00202000, 0x80000000,
134 0x80002000, 0x80000040, 0x80200000, 0x00202040,
135 0x00200000, 0x80002040, 0x80200040, 0x00000000,
136 0x00002000, 0x00000040, 0x80202000, 0x80200040,
137 0x80202040, 0x80200000, 0x80000000, 0x00002040,
138 0x00000040, 0x00202000, 0x00202040, 0x80002000,
139 0x00002040, 0x80000000, 0x80002000, 0x00202040,
140 0x80202000, 0x00200040, 0x00000000, 0x80002000,
141 0x80000000, 0x00002000, 0x80200040, 0x00200000,
142 0x00200040, 0x80202040, 0x00202000, 0x00000040,
143 0x80202040, 0x00202000, 0x00200000, 0x80002040,
144 0x80000040, 0x80200000, 0x00202040, 0x00000000,
145 0x00002000, 0x80000040, 0x80002040, 0x80202000,
146 0x80200000, 0x00002040, 0x00000040, 0x80200040,
148 0x00004000, 0x00000200, 0x01000200, 0x01000004,
149 0x01004204, 0x00004004, 0x00004200, 0x00000000,
150 0x01000000, 0x01000204, 0x00000204, 0x01004000,
151 0x00000004, 0x01004200, 0x01004000, 0x00000204,
152 0x01000204, 0x00004000, 0x00004004, 0x01004204,
153 0x00000000, 0x01000200, 0x01000004, 0x00004200,
154 0x01004004, 0x00004204, 0x01004200, 0x00000004,
155 0x00004204, 0x01004004, 0x00000200, 0x01000000,
156 0x00004204, 0x01004000, 0x01004004, 0x00000204,
157 0x00004000, 0x00000200, 0x01000000, 0x01004004,
158 0x01000204, 0x00004204, 0x00004200, 0x00000000,
159 0x00000200, 0x01000004, 0x00000004, 0x01000200,
160 0x00000000, 0x01000204, 0x01000200, 0x00004200,
161 0x00000204, 0x00004000, 0x01004204, 0x01000000,
162 0x01004200, 0x00000004, 0x00004004, 0x01004204,
163 0x01000004, 0x01004200, 0x01004000, 0x00004004,
165 0x20800080, 0x20820000, 0x00020080, 0x00000000,
166 0x20020000, 0x00800080, 0x20800000, 0x20820080,
167 0x00000080, 0x20000000, 0x00820000, 0x00020080,
168 0x00820080, 0x20020080, 0x20000080, 0x20800000,
169 0x00020000, 0x00820080, 0x00800080, 0x20020000,
170 0x20820080, 0x20000080, 0x00000000, 0x00820000,
171 0x20000000, 0x00800000, 0x20020080, 0x20800080,
172 0x00800000, 0x00020000, 0x20820000, 0x00000080,
173 0x00800000, 0x00020000, 0x20000080, 0x20820080,
174 0x00020080, 0x20000000, 0x00000000, 0x00820000,
175 0x20800080, 0x20020080, 0x20020000, 0x00800080,
176 0x20820000, 0x00000080, 0x00800080, 0x20020000,
177 0x20820080, 0x00800000, 0x20800000, 0x20000080,
178 0x00820000, 0x00020080, 0x20020080, 0x20800000,
179 0x00000080, 0x20820000, 0x00820080, 0x00000000,
180 0x20000000, 0x20800080, 0x00020000, 0x00820080,
183 __constant u32 c_skb[8][64] =
185 /* for C bits (numbered as per FIPS 46) 1 2 3 4 5 6 */
186 0x00000000, 0x00000010, 0x20000000, 0x20000010,
187 0x00010000, 0x00010010, 0x20010000, 0x20010010,
188 0x00000800, 0x00000810, 0x20000800, 0x20000810,
189 0x00010800, 0x00010810, 0x20010800, 0x20010810,
190 0x00000020, 0x00000030, 0x20000020, 0x20000030,
191 0x00010020, 0x00010030, 0x20010020, 0x20010030,
192 0x00000820, 0x00000830, 0x20000820, 0x20000830,
193 0x00010820, 0x00010830, 0x20010820, 0x20010830,
194 0x00080000, 0x00080010, 0x20080000, 0x20080010,
195 0x00090000, 0x00090010, 0x20090000, 0x20090010,
196 0x00080800, 0x00080810, 0x20080800, 0x20080810,
197 0x00090800, 0x00090810, 0x20090800, 0x20090810,
198 0x00080020, 0x00080030, 0x20080020, 0x20080030,
199 0x00090020, 0x00090030, 0x20090020, 0x20090030,
200 0x00080820, 0x00080830, 0x20080820, 0x20080830,
201 0x00090820, 0x00090830, 0x20090820, 0x20090830,
202 /* for C bits (numbered as per FIPS 46) 7 8 10 11 12 13 */
203 0x00000000, 0x02000000, 0x00002000, 0x02002000,
204 0x00200000, 0x02200000, 0x00202000, 0x02202000,
205 0x00000004, 0x02000004, 0x00002004, 0x02002004,
206 0x00200004, 0x02200004, 0x00202004, 0x02202004,
207 0x00000400, 0x02000400, 0x00002400, 0x02002400,
208 0x00200400, 0x02200400, 0x00202400, 0x02202400,
209 0x00000404, 0x02000404, 0x00002404, 0x02002404,
210 0x00200404, 0x02200404, 0x00202404, 0x02202404,
211 0x10000000, 0x12000000, 0x10002000, 0x12002000,
212 0x10200000, 0x12200000, 0x10202000, 0x12202000,
213 0x10000004, 0x12000004, 0x10002004, 0x12002004,
214 0x10200004, 0x12200004, 0x10202004, 0x12202004,
215 0x10000400, 0x12000400, 0x10002400, 0x12002400,
216 0x10200400, 0x12200400, 0x10202400, 0x12202400,
217 0x10000404, 0x12000404, 0x10002404, 0x12002404,
218 0x10200404, 0x12200404, 0x10202404, 0x12202404,
219 /* for C bits (numbered as per FIPS 46) 14 15 16 17 19 20 */
220 0x00000000, 0x00000001, 0x00040000, 0x00040001,
221 0x01000000, 0x01000001, 0x01040000, 0x01040001,
222 0x00000002, 0x00000003, 0x00040002, 0x00040003,
223 0x01000002, 0x01000003, 0x01040002, 0x01040003,
224 0x00000200, 0x00000201, 0x00040200, 0x00040201,
225 0x01000200, 0x01000201, 0x01040200, 0x01040201,
226 0x00000202, 0x00000203, 0x00040202, 0x00040203,
227 0x01000202, 0x01000203, 0x01040202, 0x01040203,
228 0x08000000, 0x08000001, 0x08040000, 0x08040001,
229 0x09000000, 0x09000001, 0x09040000, 0x09040001,
230 0x08000002, 0x08000003, 0x08040002, 0x08040003,
231 0x09000002, 0x09000003, 0x09040002, 0x09040003,
232 0x08000200, 0x08000201, 0x08040200, 0x08040201,
233 0x09000200, 0x09000201, 0x09040200, 0x09040201,
234 0x08000202, 0x08000203, 0x08040202, 0x08040203,
235 0x09000202, 0x09000203, 0x09040202, 0x09040203,
236 /* for C bits (numbered as per FIPS 46) 21 23 24 26 27 28 */
237 0x00000000, 0x00100000, 0x00000100, 0x00100100,
238 0x00000008, 0x00100008, 0x00000108, 0x00100108,
239 0x00001000, 0x00101000, 0x00001100, 0x00101100,
240 0x00001008, 0x00101008, 0x00001108, 0x00101108,
241 0x04000000, 0x04100000, 0x04000100, 0x04100100,
242 0x04000008, 0x04100008, 0x04000108, 0x04100108,
243 0x04001000, 0x04101000, 0x04001100, 0x04101100,
244 0x04001008, 0x04101008, 0x04001108, 0x04101108,
245 0x00020000, 0x00120000, 0x00020100, 0x00120100,
246 0x00020008, 0x00120008, 0x00020108, 0x00120108,
247 0x00021000, 0x00121000, 0x00021100, 0x00121100,
248 0x00021008, 0x00121008, 0x00021108, 0x00121108,
249 0x04020000, 0x04120000, 0x04020100, 0x04120100,
250 0x04020008, 0x04120008, 0x04020108, 0x04120108,
251 0x04021000, 0x04121000, 0x04021100, 0x04121100,
252 0x04021008, 0x04121008, 0x04021108, 0x04121108,
253 /* for D bits (numbered as per FIPS 46) 1 2 3 4 5 6 */
254 0x00000000, 0x10000000, 0x00010000, 0x10010000,
255 0x00000004, 0x10000004, 0x00010004, 0x10010004,
256 0x20000000, 0x30000000, 0x20010000, 0x30010000,
257 0x20000004, 0x30000004, 0x20010004, 0x30010004,
258 0x00100000, 0x10100000, 0x00110000, 0x10110000,
259 0x00100004, 0x10100004, 0x00110004, 0x10110004,
260 0x20100000, 0x30100000, 0x20110000, 0x30110000,
261 0x20100004, 0x30100004, 0x20110004, 0x30110004,
262 0x00001000, 0x10001000, 0x00011000, 0x10011000,
263 0x00001004, 0x10001004, 0x00011004, 0x10011004,
264 0x20001000, 0x30001000, 0x20011000, 0x30011000,
265 0x20001004, 0x30001004, 0x20011004, 0x30011004,
266 0x00101000, 0x10101000, 0x00111000, 0x10111000,
267 0x00101004, 0x10101004, 0x00111004, 0x10111004,
268 0x20101000, 0x30101000, 0x20111000, 0x30111000,
269 0x20101004, 0x30101004, 0x20111004, 0x30111004,
270 /* for D bits (numbered as per FIPS 46) 8 9 11 12 13 14 */
271 0x00000000, 0x08000000, 0x00000008, 0x08000008,
272 0x00000400, 0x08000400, 0x00000408, 0x08000408,
273 0x00020000, 0x08020000, 0x00020008, 0x08020008,
274 0x00020400, 0x08020400, 0x00020408, 0x08020408,
275 0x00000001, 0x08000001, 0x00000009, 0x08000009,
276 0x00000401, 0x08000401, 0x00000409, 0x08000409,
277 0x00020001, 0x08020001, 0x00020009, 0x08020009,
278 0x00020401, 0x08020401, 0x00020409, 0x08020409,
279 0x02000000, 0x0A000000, 0x02000008, 0x0A000008,
280 0x02000400, 0x0A000400, 0x02000408, 0x0A000408,
281 0x02020000, 0x0A020000, 0x02020008, 0x0A020008,
282 0x02020400, 0x0A020400, 0x02020408, 0x0A020408,
283 0x02000001, 0x0A000001, 0x02000009, 0x0A000009,
284 0x02000401, 0x0A000401, 0x02000409, 0x0A000409,
285 0x02020001, 0x0A020001, 0x02020009, 0x0A020009,
286 0x02020401, 0x0A020401, 0x02020409, 0x0A020409,
287 /* for D bits (numbered as per FIPS 46) 16 17 18 19 20 21 */
288 0x00000000, 0x00000100, 0x00080000, 0x00080100,
289 0x01000000, 0x01000100, 0x01080000, 0x01080100,
290 0x00000010, 0x00000110, 0x00080010, 0x00080110,
291 0x01000010, 0x01000110, 0x01080010, 0x01080110,
292 0x00200000, 0x00200100, 0x00280000, 0x00280100,
293 0x01200000, 0x01200100, 0x01280000, 0x01280100,
294 0x00200010, 0x00200110, 0x00280010, 0x00280110,
295 0x01200010, 0x01200110, 0x01280010, 0x01280110,
296 0x00000200, 0x00000300, 0x00080200, 0x00080300,
297 0x01000200, 0x01000300, 0x01080200, 0x01080300,
298 0x00000210, 0x00000310, 0x00080210, 0x00080310,
299 0x01000210, 0x01000310, 0x01080210, 0x01080310,
300 0x00200200, 0x00200300, 0x00280200, 0x00280300,
301 0x01200200, 0x01200300, 0x01280200, 0x01280300,
302 0x00200210, 0x00200310, 0x00280210, 0x00280310,
303 0x01200210, 0x01200310, 0x01280210, 0x01280310,
304 /* for D bits (numbered as per FIPS 46) 22 23 24 25 27 28 */
305 0x00000000, 0x04000000, 0x00040000, 0x04040000,
306 0x00000002, 0x04000002, 0x00040002, 0x04040002,
307 0x00002000, 0x04002000, 0x00042000, 0x04042000,
308 0x00002002, 0x04002002, 0x00042002, 0x04042002,
309 0x00000020, 0x04000020, 0x00040020, 0x04040020,
310 0x00000022, 0x04000022, 0x00040022, 0x04040022,
311 0x00002020, 0x04002020, 0x00042020, 0x04042020,
312 0x00002022, 0x04002022, 0x00042022, 0x04042022,
313 0x00000800, 0x04000800, 0x00040800, 0x04040800,
314 0x00000802, 0x04000802, 0x00040802, 0x04040802,
315 0x00002800, 0x04002800, 0x00042800, 0x04042800,
316 0x00002802, 0x04002802, 0x00042802, 0x04042802,
317 0x00000820, 0x04000820, 0x00040820, 0x04040820,
318 0x00000822, 0x04000822, 0x00040822, 0x04040822,
319 0x00002820, 0x04002820, 0x00042820, 0x04042820,
320 0x00002822, 0x04002822, 0x00042822, 0x04042822
323 __constant u32 shifts3s0[16] = { 1, 1, 2, 2, 2, 2, 2, 2, 1, 2, 2, 2, 2, 2, 2, 1 };
324 __constant u32 shifts3s1[16] = { 27, 27, 26, 26, 26, 26, 26, 26, 27, 26, 26, 26, 26, 26, 26, 27 };
327 #define BOX(i,n,S) (u32x) ((S)[(n)][(i)])
331 #define BOX(i,n,S) (u32x) ((S)[(n)][(i).s0], (S)[(n)][(i).s1])
335 #define BOX(i,n,S) (u32x) ((S)[(n)][(i).s0], (S)[(n)][(i).s1], (S)[(n)][(i).s2], (S)[(n)][(i).s3])
338 static void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local u32 s_SPtrans[8][64])
344 for (u32 i = 0; i < 16; i += 2)
349 u = Kc[i + 0] ^ rotl32 (r, 30u);
350 t = Kd[i + 0] ^ rotl32 (r, 26u);
353 ^ BOX (amd_bfe (u, 0, 6), 0, s_SPtrans)
354 ^ BOX (amd_bfe (u, 8, 6), 2, s_SPtrans)
355 ^ BOX (amd_bfe (u, 16, 6), 4, s_SPtrans)
356 ^ BOX (amd_bfe (u, 24, 6), 6, s_SPtrans)
357 ^ BOX (amd_bfe (t, 0, 6), 1, s_SPtrans)
358 ^ BOX (amd_bfe (t, 8, 6), 3, s_SPtrans)
359 ^ BOX (amd_bfe (t, 16, 6), 5, s_SPtrans)
360 ^ BOX (amd_bfe (t, 24, 6), 7, s_SPtrans);
362 u = Kc[i + 1] ^ rotl32 (l, 30u);
363 t = Kd[i + 1] ^ rotl32 (l, 26u);
366 ^ BOX (amd_bfe (u, 0, 6), 0, s_SPtrans)
367 ^ BOX (amd_bfe (u, 8, 6), 2, s_SPtrans)
368 ^ BOX (amd_bfe (u, 16, 6), 4, s_SPtrans)
369 ^ BOX (amd_bfe (u, 24, 6), 6, s_SPtrans)
370 ^ BOX (amd_bfe (t, 0, 6), 1, s_SPtrans)
371 ^ BOX (amd_bfe (t, 8, 6), 3, s_SPtrans)
372 ^ BOX (amd_bfe (t, 16, 6), 5, s_SPtrans)
373 ^ BOX (amd_bfe (t, 24, 6), 7, s_SPtrans);
380 static void _des_crypt_keysetup (u32 c, u32 d, u32 Kc[16], u32 Kd[16], __local u32 s_skb[8][64])
384 PERM_OP (d, c, tt, 4, 0x0f0f0f0f);
385 HPERM_OP (c, tt, 2, 0xcccc0000);
386 HPERM_OP (d, tt, 2, 0xcccc0000);
387 PERM_OP (d, c, tt, 1, 0x55555555);
388 PERM_OP (c, d, tt, 8, 0x00ff00ff);
389 PERM_OP (d, c, tt, 1, 0x55555555);
391 d = ((d & 0x000000ff) << 16)
392 | ((d & 0x0000ff00) << 0)
393 | ((d & 0x00ff0000) >> 16)
394 | ((c & 0xf0000000) >> 4);
399 for (u32 i = 0; i < 16; i++)
401 c = c >> shifts3s0[i] | c << shifts3s1[i];
402 d = d >> shifts3s0[i] | d << shifts3s1[i];
407 const u32 c00 = (c >> 0) & 0x0000003f;
408 const u32 c06 = (c >> 6) & 0x00383003;
409 const u32 c07 = (c >> 7) & 0x0000003c;
410 const u32 c13 = (c >> 13) & 0x0000060f;
411 const u32 c20 = (c >> 20) & 0x00000001;
413 u32 s = BOX (((c00 >> 0) & 0xff), 0, s_skb)
414 | BOX (((c06 >> 0) & 0xff)
415 |((c07 >> 0) & 0xff), 1, s_skb)
416 | BOX (((c13 >> 0) & 0xff)
417 |((c06 >> 8) & 0xff), 2, s_skb)
418 | BOX (((c20 >> 0) & 0xff)
420 |((c06 >> 16) & 0xff), 3, s_skb);
422 const u32 d00 = (d >> 0) & 0x00003c3f;
423 const u32 d07 = (d >> 7) & 0x00003f03;
424 const u32 d21 = (d >> 21) & 0x0000000f;
425 const u32 d22 = (d >> 22) & 0x00000030;
427 u32 t = BOX (((d00 >> 0) & 0xff), 4, s_skb)
428 | BOX (((d07 >> 0) & 0xff)
429 |((d00 >> 8) & 0xff), 5, s_skb)
430 | BOX (((d07 >> 8) & 0xff), 6, s_skb)
431 | BOX (((d21 >> 0) & 0xff)
432 |((d22 >> 0) & 0xff), 7, s_skb);
434 Kc[i] = ((t << 16) | (s & 0x0000ffff));
435 Kd[i] = ((s >> 16) | (t & 0xffff0000));
439 static void transform_netntlmv1_key (const u32 w0, const u32 w1, u32 out[2])
442 const uchar4 t0 = as_uchar4 (w0);
443 const uchar4 t1 = as_uchar4 (w1);
448 k0.s0 = (t0.s0 >> 0);
449 k0.s1 = (t0.s0 << 7) | (t0.s1 >> 1);
450 k0.s2 = (t0.s1 << 6) | (t0.s2 >> 2);
451 k0.s3 = (t0.s2 << 5) | (t0.s3 >> 3);
452 k1.s0 = (t0.s3 << 4) | (t1.s0 >> 4);
453 k1.s1 = (t1.s0 << 3) | (t1.s1 >> 5);
454 k1.s2 = (t1.s1 << 2) | (t1.s2 >> 6);
455 k1.s3 = (t1.s2 << 1);
457 out[0] = as_uint (k0);
458 out[1] = as_uint (k1);
462 const uchar8 t0 = as_uchar8 (w0);
463 const uchar8 t1 = as_uchar8 (w1);
468 k0.s0 = (t0.s0 >> 0);
469 k0.s1 = (t0.s0 << 7) | (t0.s1 >> 1);
470 k0.s2 = (t0.s1 << 6) | (t0.s2 >> 2);
471 k0.s3 = (t0.s2 << 5) | (t0.s3 >> 3);
472 k1.s0 = (t0.s3 << 4) | (t1.s0 >> 4);
473 k1.s1 = (t1.s0 << 3) | (t1.s1 >> 5);
474 k1.s2 = (t1.s1 << 2) | (t1.s2 >> 6);
475 k1.s3 = (t1.s2 << 1);
477 k0.s4 = (t0.s4 >> 0);
478 k0.s5 = (t0.s4 << 7) | (t0.s5 >> 1);
479 k0.s6 = (t0.s5 << 6) | (t0.s6 >> 2);
480 k0.s7 = (t0.s6 << 5) | (t0.s7 >> 3);
481 k1.s4 = (t0.s7 << 4) | (t1.s4 >> 4);
482 k1.s5 = (t1.s4 << 3) | (t1.s5 >> 5);
483 k1.s6 = (t1.s5 << 2) | (t1.s6 >> 6);
484 k1.s7 = (t1.s6 << 1);
486 out[0] = as_uint2 (k0);
487 out[1] = as_uint2 (k1);
491 const uchar16 t0 = as_uchar16 (w0);
492 const uchar16 t1 = as_uchar16 (w1);
497 k0.s0 = (t0.s0 >> 0);
498 k0.s1 = (t0.s0 << 7) | (t0.s1 >> 1);
499 k0.s2 = (t0.s1 << 6) | (t0.s2 >> 2);
500 k0.s3 = (t0.s2 << 5) | (t0.s3 >> 3);
501 k1.s0 = (t0.s3 << 4) | (t1.s0 >> 4);
502 k1.s1 = (t1.s0 << 3) | (t1.s1 >> 5);
503 k1.s2 = (t1.s1 << 2) | (t1.s2 >> 6);
504 k1.s3 = (t1.s2 << 1);
506 k0.s4 = (t0.s4 >> 0);
507 k0.s5 = (t0.s4 << 7) | (t0.s5 >> 1);
508 k0.s6 = (t0.s5 << 6) | (t0.s6 >> 2);
509 k0.s7 = (t0.s6 << 5) | (t0.s7 >> 3);
510 k1.s4 = (t0.s7 << 4) | (t1.s4 >> 4);
511 k1.s5 = (t1.s4 << 3) | (t1.s5 >> 5);
512 k1.s6 = (t1.s5 << 2) | (t1.s6 >> 6);
513 k1.s7 = (t1.s6 << 1);
515 k0.s8 = (t0.s8 >> 0);
516 k0.s9 = (t0.s8 << 7) | (t0.s9 >> 1);
517 k0.sa = (t0.s9 << 6) | (t0.sa >> 2);
518 k0.sb = (t0.sa << 5) | (t0.sb >> 3);
519 k1.s8 = (t0.sb << 4) | (t1.s8 >> 4);
520 k1.s9 = (t1.s8 << 3) | (t1.s9 >> 5);
521 k1.sa = (t1.s9 << 2) | (t1.sa >> 6);
522 k1.sb = (t1.sa << 1);
524 k0.sc = (t0.sc >> 0);
525 k0.sd = (t0.sc << 7) | (t0.sd >> 1);
526 k0.se = (t0.sd << 6) | (t0.se >> 2);
527 k0.sf = (t0.se << 5) | (t0.sf >> 3);
528 k1.sc = (t0.sf << 4) | (t1.sc >> 4);
529 k1.sd = (t1.sc << 3) | (t1.sd >> 5);
530 k1.se = (t1.sd << 2) | (t1.se >> 6);
531 k1.sf = (t1.se << 1);
533 out[0] = as_uint4 (k0);
534 out[1] = as_uint4 (k1);
538 static void m05500m (__local u32 s_SPtrans[8][64], __local u32 s_skb[8][64], u32 w[16], const u32 pw_len, __global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global u32 * 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)
544 const u32 gid = get_global_id (0);
545 const u32 lid = get_local_id (0);
551 const u32 s0 = salt_bufs[salt_pos].salt_buf[0];
552 const u32 s1 = salt_bufs[salt_pos].salt_buf[1];
553 const u32 s2 = salt_bufs[salt_pos].salt_buf[2];
566 for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos++)
568 const u32 w0r = words_buf_r[il_pos];
570 const u32 w0 = w0l | w0r;
594 MD4_STEP (MD4_Fo, a, b, c, d, w0_t, MD4C00, MD4S00);
595 MD4_STEP (MD4_Fo, d, a, b, c, w1_t, MD4C00, MD4S01);
596 MD4_STEP (MD4_Fo, c, d, a, b, w2_t, MD4C00, MD4S02);
597 MD4_STEP (MD4_Fo, b, c, d, a, w3_t, MD4C00, MD4S03);
598 MD4_STEP (MD4_Fo, a, b, c, d, w4_t, MD4C00, MD4S00);
599 MD4_STEP (MD4_Fo, d, a, b, c, w5_t, MD4C00, MD4S01);
600 MD4_STEP (MD4_Fo, c, d, a, b, w6_t, MD4C00, MD4S02);
601 MD4_STEP (MD4_Fo, b, c, d, a, w7_t, MD4C00, MD4S03);
602 MD4_STEP (MD4_Fo, a, b, c, d, w8_t, MD4C00, MD4S00);
603 MD4_STEP (MD4_Fo, d, a, b, c, w9_t, MD4C00, MD4S01);
604 MD4_STEP (MD4_Fo, c, d, a, b, wa_t, MD4C00, MD4S02);
605 MD4_STEP (MD4_Fo, b, c, d, a, wb_t, MD4C00, MD4S03);
606 MD4_STEP (MD4_Fo, a, b, c, d, wc_t, MD4C00, MD4S00);
607 MD4_STEP (MD4_Fo, d, a, b, c, wd_t, MD4C00, MD4S01);
608 MD4_STEP (MD4_Fo, c, d, a, b, we_t, MD4C00, MD4S02);
609 MD4_STEP (MD4_Fo, b, c, d, a, wf_t, MD4C00, MD4S03);
611 MD4_STEP (MD4_Go, a, b, c, d, w0_t, MD4C01, MD4S10);
612 MD4_STEP (MD4_Go, d, a, b, c, w4_t, MD4C01, MD4S11);
613 MD4_STEP (MD4_Go, c, d, a, b, w8_t, MD4C01, MD4S12);
614 MD4_STEP (MD4_Go, b, c, d, a, wc_t, MD4C01, MD4S13);
615 MD4_STEP (MD4_Go, a, b, c, d, w1_t, MD4C01, MD4S10);
616 MD4_STEP (MD4_Go, d, a, b, c, w5_t, MD4C01, MD4S11);
617 MD4_STEP (MD4_Go, c, d, a, b, w9_t, MD4C01, MD4S12);
618 MD4_STEP (MD4_Go, b, c, d, a, wd_t, MD4C01, MD4S13);
619 MD4_STEP (MD4_Go, a, b, c, d, w2_t, MD4C01, MD4S10);
620 MD4_STEP (MD4_Go, d, a, b, c, w6_t, MD4C01, MD4S11);
621 MD4_STEP (MD4_Go, c, d, a, b, wa_t, MD4C01, MD4S12);
622 MD4_STEP (MD4_Go, b, c, d, a, we_t, MD4C01, MD4S13);
623 MD4_STEP (MD4_Go, a, b, c, d, w3_t, MD4C01, MD4S10);
624 MD4_STEP (MD4_Go, d, a, b, c, w7_t, MD4C01, MD4S11);
625 MD4_STEP (MD4_Go, c, d, a, b, wb_t, MD4C01, MD4S12);
626 MD4_STEP (MD4_Go, b, c, d, a, wf_t, MD4C01, MD4S13);
628 MD4_STEP (MD4_H , a, b, c, d, w0_t, MD4C02, MD4S20);
629 MD4_STEP (MD4_H , d, a, b, c, w8_t, MD4C02, MD4S21);
630 MD4_STEP (MD4_H , c, d, a, b, w4_t, MD4C02, MD4S22);
631 MD4_STEP (MD4_H , b, c, d, a, wc_t, MD4C02, MD4S23);
632 MD4_STEP (MD4_H , a, b, c, d, w2_t, MD4C02, MD4S20);
633 MD4_STEP (MD4_H , d, a, b, c, wa_t, MD4C02, MD4S21);
634 MD4_STEP (MD4_H , c, d, a, b, w6_t, MD4C02, MD4S22);
635 MD4_STEP (MD4_H , b, c, d, a, we_t, MD4C02, MD4S23);
636 MD4_STEP (MD4_H , a, b, c, d, w1_t, MD4C02, MD4S20);
637 MD4_STEP (MD4_H , d, a, b, c, w9_t, MD4C02, MD4S21);
638 MD4_STEP (MD4_H , c, d, a, b, w5_t, MD4C02, MD4S22);
639 MD4_STEP (MD4_H , b, c, d, a, wd_t, MD4C02, MD4S23);
640 MD4_STEP (MD4_H , a, b, c, d, w3_t, MD4C02, MD4S20);
641 MD4_STEP (MD4_H , d, a, b, c, wb_t, MD4C02, MD4S21);
643 if (allx (s2 != ((d + MD4M_D) >> 16))) continue;
645 MD4_STEP (MD4_H , c, d, a, b, w7_t, MD4C02, MD4S22);
646 MD4_STEP (MD4_H , b, c, d, a, wf_t, MD4C02, MD4S23);
659 transform_netntlmv1_key (a, b, key);
664 _des_crypt_keysetup (key[0], key[1], Kc, Kd, s_skb);
668 _des_crypt_encrypt (iv1, data, Kc, Kd, s_SPtrans);
674 transform_netntlmv1_key (((b >> 24) | (c << 8)), ((c >> 24) | (d << 8)), key);
676 _des_crypt_keysetup (key[0], key[1], Kc, Kd, s_skb);
680 _des_crypt_encrypt (iv2, data, Kc, Kd, s_SPtrans);
686 const u32 r0 = iv1[0];
687 const u32 r1 = iv1[1];
688 const u32 r2 = iv2[0];
689 const u32 r3 = iv2[1];
695 static void m05500s (__local u32 s_SPtrans[8][64], __local u32 s_skb[8][64], u32 w[16], const u32 pw_len, __global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global u32 * 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)
701 const u32 gid = get_global_id (0);
702 const u32 lid = get_local_id (0);
708 const u32 s0 = salt_bufs[salt_pos].salt_buf[0];
709 const u32 s1 = salt_bufs[salt_pos].salt_buf[1];
710 const u32 s2 = salt_bufs[salt_pos].salt_buf[2];
721 const u32 search[4] =
723 digests_buf[digests_offset].digest_buf[DGST_R0],
724 digests_buf[digests_offset].digest_buf[DGST_R1],
725 digests_buf[digests_offset].digest_buf[DGST_R2],
726 digests_buf[digests_offset].digest_buf[DGST_R3]
735 for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos++)
737 const u32 w0r = words_buf_r[il_pos];
739 const u32 w0 = w0l | w0r;
763 MD4_STEP (MD4_Fo, a, b, c, d, w0_t, MD4C00, MD4S00);
764 MD4_STEP (MD4_Fo, d, a, b, c, w1_t, MD4C00, MD4S01);
765 MD4_STEP (MD4_Fo, c, d, a, b, w2_t, MD4C00, MD4S02);
766 MD4_STEP (MD4_Fo, b, c, d, a, w3_t, MD4C00, MD4S03);
767 MD4_STEP (MD4_Fo, a, b, c, d, w4_t, MD4C00, MD4S00);
768 MD4_STEP (MD4_Fo, d, a, b, c, w5_t, MD4C00, MD4S01);
769 MD4_STEP (MD4_Fo, c, d, a, b, w6_t, MD4C00, MD4S02);
770 MD4_STEP (MD4_Fo, b, c, d, a, w7_t, MD4C00, MD4S03);
771 MD4_STEP (MD4_Fo, a, b, c, d, w8_t, MD4C00, MD4S00);
772 MD4_STEP (MD4_Fo, d, a, b, c, w9_t, MD4C00, MD4S01);
773 MD4_STEP (MD4_Fo, c, d, a, b, wa_t, MD4C00, MD4S02);
774 MD4_STEP (MD4_Fo, b, c, d, a, wb_t, MD4C00, MD4S03);
775 MD4_STEP (MD4_Fo, a, b, c, d, wc_t, MD4C00, MD4S00);
776 MD4_STEP (MD4_Fo, d, a, b, c, wd_t, MD4C00, MD4S01);
777 MD4_STEP (MD4_Fo, c, d, a, b, we_t, MD4C00, MD4S02);
778 MD4_STEP (MD4_Fo, b, c, d, a, wf_t, MD4C00, MD4S03);
780 MD4_STEP (MD4_Go, a, b, c, d, w0_t, MD4C01, MD4S10);
781 MD4_STEP (MD4_Go, d, a, b, c, w4_t, MD4C01, MD4S11);
782 MD4_STEP (MD4_Go, c, d, a, b, w8_t, MD4C01, MD4S12);
783 MD4_STEP (MD4_Go, b, c, d, a, wc_t, MD4C01, MD4S13);
784 MD4_STEP (MD4_Go, a, b, c, d, w1_t, MD4C01, MD4S10);
785 MD4_STEP (MD4_Go, d, a, b, c, w5_t, MD4C01, MD4S11);
786 MD4_STEP (MD4_Go, c, d, a, b, w9_t, MD4C01, MD4S12);
787 MD4_STEP (MD4_Go, b, c, d, a, wd_t, MD4C01, MD4S13);
788 MD4_STEP (MD4_Go, a, b, c, d, w2_t, MD4C01, MD4S10);
789 MD4_STEP (MD4_Go, d, a, b, c, w6_t, MD4C01, MD4S11);
790 MD4_STEP (MD4_Go, c, d, a, b, wa_t, MD4C01, MD4S12);
791 MD4_STEP (MD4_Go, b, c, d, a, we_t, MD4C01, MD4S13);
792 MD4_STEP (MD4_Go, a, b, c, d, w3_t, MD4C01, MD4S10);
793 MD4_STEP (MD4_Go, d, a, b, c, w7_t, MD4C01, MD4S11);
794 MD4_STEP (MD4_Go, c, d, a, b, wb_t, MD4C01, MD4S12);
795 MD4_STEP (MD4_Go, b, c, d, a, wf_t, MD4C01, MD4S13);
797 MD4_STEP (MD4_H , a, b, c, d, w0_t, MD4C02, MD4S20);
798 MD4_STEP (MD4_H , d, a, b, c, w8_t, MD4C02, MD4S21);
799 MD4_STEP (MD4_H , c, d, a, b, w4_t, MD4C02, MD4S22);
800 MD4_STEP (MD4_H , b, c, d, a, wc_t, MD4C02, MD4S23);
801 MD4_STEP (MD4_H , a, b, c, d, w2_t, MD4C02, MD4S20);
802 MD4_STEP (MD4_H , d, a, b, c, wa_t, MD4C02, MD4S21);
803 MD4_STEP (MD4_H , c, d, a, b, w6_t, MD4C02, MD4S22);
804 MD4_STEP (MD4_H , b, c, d, a, we_t, MD4C02, MD4S23);
805 MD4_STEP (MD4_H , a, b, c, d, w1_t, MD4C02, MD4S20);
806 MD4_STEP (MD4_H , d, a, b, c, w9_t, MD4C02, MD4S21);
807 MD4_STEP (MD4_H , c, d, a, b, w5_t, MD4C02, MD4S22);
808 MD4_STEP (MD4_H , b, c, d, a, wd_t, MD4C02, MD4S23);
809 MD4_STEP (MD4_H , a, b, c, d, w3_t, MD4C02, MD4S20);
810 MD4_STEP (MD4_H , d, a, b, c, wb_t, MD4C02, MD4S21);
812 if (allx (s2 != ((d + MD4M_D) >> 16))) continue;
814 MD4_STEP (MD4_H , c, d, a, b, w7_t, MD4C02, MD4S22);
815 MD4_STEP (MD4_H , b, c, d, a, wf_t, MD4C02, MD4S23);
828 transform_netntlmv1_key (a, b, key);
833 _des_crypt_keysetup (key[0], key[1], Kc, Kd, s_skb);
837 _des_crypt_encrypt (iv1, data, Kc, Kd, s_SPtrans);
844 transform_netntlmv1_key (((b >> 24) | (c << 8)), ((c >> 24) | (d << 8)), key);
846 _des_crypt_keysetup (key[0], key[1], Kc, Kd, s_skb);
850 _des_crypt_encrypt (iv2, data, Kc, Kd, s_SPtrans);
862 const u32 r0 = iv1[0];
863 const u32 r1 = iv1[1];
864 const u32 r2 = iv2[0];
865 const u32 r3 = iv2[1];
871 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05500_m04 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global u32 * 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)
877 const u32 gid = get_global_id (0);
878 const u32 lid = get_local_id (0);
882 w[ 0] = pws[gid].i[ 0];
883 w[ 1] = pws[gid].i[ 1];
884 w[ 2] = pws[gid].i[ 2];
885 w[ 3] = pws[gid].i[ 3];
896 w[14] = pws[gid].i[14];
899 const u32 pw_len = pws[gid].pw_len;
905 __local u32 s_SPtrans[8][64];
906 __local u32 s_skb[8][64];
908 s_SPtrans[0][lid] = c_SPtrans[0][lid];
909 s_SPtrans[1][lid] = c_SPtrans[1][lid];
910 s_SPtrans[2][lid] = c_SPtrans[2][lid];
911 s_SPtrans[3][lid] = c_SPtrans[3][lid];
912 s_SPtrans[4][lid] = c_SPtrans[4][lid];
913 s_SPtrans[5][lid] = c_SPtrans[5][lid];
914 s_SPtrans[6][lid] = c_SPtrans[6][lid];
915 s_SPtrans[7][lid] = c_SPtrans[7][lid];
917 s_skb[0][lid] = c_skb[0][lid];
918 s_skb[1][lid] = c_skb[1][lid];
919 s_skb[2][lid] = c_skb[2][lid];
920 s_skb[3][lid] = c_skb[3][lid];
921 s_skb[4][lid] = c_skb[4][lid];
922 s_skb[5][lid] = c_skb[5][lid];
923 s_skb[6][lid] = c_skb[6][lid];
924 s_skb[7][lid] = c_skb[7][lid];
926 barrier (CLK_LOCAL_MEM_FENCE);
928 if (gid >= gid_max) return;
934 m05500m (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);
937 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05500_m08 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global u32 * 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)
943 const u32 gid = get_global_id (0);
944 const u32 lid = get_local_id (0);
948 w[ 0] = pws[gid].i[ 0];
949 w[ 1] = pws[gid].i[ 1];
950 w[ 2] = pws[gid].i[ 2];
951 w[ 3] = pws[gid].i[ 3];
952 w[ 4] = pws[gid].i[ 4];
953 w[ 5] = pws[gid].i[ 5];
954 w[ 6] = pws[gid].i[ 6];
955 w[ 7] = pws[gid].i[ 7];
962 w[14] = pws[gid].i[14];
965 const u32 pw_len = pws[gid].pw_len;
971 __local u32 s_SPtrans[8][64];
972 __local u32 s_skb[8][64];
974 s_SPtrans[0][lid] = c_SPtrans[0][lid];
975 s_SPtrans[1][lid] = c_SPtrans[1][lid];
976 s_SPtrans[2][lid] = c_SPtrans[2][lid];
977 s_SPtrans[3][lid] = c_SPtrans[3][lid];
978 s_SPtrans[4][lid] = c_SPtrans[4][lid];
979 s_SPtrans[5][lid] = c_SPtrans[5][lid];
980 s_SPtrans[6][lid] = c_SPtrans[6][lid];
981 s_SPtrans[7][lid] = c_SPtrans[7][lid];
983 s_skb[0][lid] = c_skb[0][lid];
984 s_skb[1][lid] = c_skb[1][lid];
985 s_skb[2][lid] = c_skb[2][lid];
986 s_skb[3][lid] = c_skb[3][lid];
987 s_skb[4][lid] = c_skb[4][lid];
988 s_skb[5][lid] = c_skb[5][lid];
989 s_skb[6][lid] = c_skb[6][lid];
990 s_skb[7][lid] = c_skb[7][lid];
992 barrier (CLK_LOCAL_MEM_FENCE);
994 if (gid >= gid_max) return;
1000 m05500m (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);
1003 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05500_m16 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global u32 * 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)
1007 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05500_s04 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global u32 * 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)
1013 const u32 gid = get_global_id (0);
1014 const u32 lid = get_local_id (0);
1018 w[ 0] = pws[gid].i[ 0];
1019 w[ 1] = pws[gid].i[ 1];
1020 w[ 2] = pws[gid].i[ 2];
1021 w[ 3] = pws[gid].i[ 3];
1032 w[14] = pws[gid].i[14];
1035 const u32 pw_len = pws[gid].pw_len;
1041 __local u32 s_SPtrans[8][64];
1042 __local u32 s_skb[8][64];
1044 s_SPtrans[0][lid] = c_SPtrans[0][lid];
1045 s_SPtrans[1][lid] = c_SPtrans[1][lid];
1046 s_SPtrans[2][lid] = c_SPtrans[2][lid];
1047 s_SPtrans[3][lid] = c_SPtrans[3][lid];
1048 s_SPtrans[4][lid] = c_SPtrans[4][lid];
1049 s_SPtrans[5][lid] = c_SPtrans[5][lid];
1050 s_SPtrans[6][lid] = c_SPtrans[6][lid];
1051 s_SPtrans[7][lid] = c_SPtrans[7][lid];
1053 s_skb[0][lid] = c_skb[0][lid];
1054 s_skb[1][lid] = c_skb[1][lid];
1055 s_skb[2][lid] = c_skb[2][lid];
1056 s_skb[3][lid] = c_skb[3][lid];
1057 s_skb[4][lid] = c_skb[4][lid];
1058 s_skb[5][lid] = c_skb[5][lid];
1059 s_skb[6][lid] = c_skb[6][lid];
1060 s_skb[7][lid] = c_skb[7][lid];
1062 barrier (CLK_LOCAL_MEM_FENCE);
1064 if (gid >= gid_max) return;
1070 m05500s (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);
1073 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05500_s08 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global u32 * 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)
1079 const u32 gid = get_global_id (0);
1080 const u32 lid = get_local_id (0);
1084 w[ 0] = pws[gid].i[ 0];
1085 w[ 1] = pws[gid].i[ 1];
1086 w[ 2] = pws[gid].i[ 2];
1087 w[ 3] = pws[gid].i[ 3];
1088 w[ 4] = pws[gid].i[ 4];
1089 w[ 5] = pws[gid].i[ 5];
1090 w[ 6] = pws[gid].i[ 6];
1091 w[ 7] = pws[gid].i[ 7];
1098 w[14] = pws[gid].i[14];
1101 const u32 pw_len = pws[gid].pw_len;
1107 __local u32 s_SPtrans[8][64];
1108 __local u32 s_skb[8][64];
1110 s_SPtrans[0][lid] = c_SPtrans[0][lid];
1111 s_SPtrans[1][lid] = c_SPtrans[1][lid];
1112 s_SPtrans[2][lid] = c_SPtrans[2][lid];
1113 s_SPtrans[3][lid] = c_SPtrans[3][lid];
1114 s_SPtrans[4][lid] = c_SPtrans[4][lid];
1115 s_SPtrans[5][lid] = c_SPtrans[5][lid];
1116 s_SPtrans[6][lid] = c_SPtrans[6][lid];
1117 s_SPtrans[7][lid] = c_SPtrans[7][lid];
1119 s_skb[0][lid] = c_skb[0][lid];
1120 s_skb[1][lid] = c_skb[1][lid];
1121 s_skb[2][lid] = c_skb[2][lid];
1122 s_skb[3][lid] = c_skb[3][lid];
1123 s_skb[4][lid] = c_skb[4][lid];
1124 s_skb[5][lid] = c_skb[5][lid];
1125 s_skb[6][lid] = c_skb[6][lid];
1126 s_skb[7][lid] = c_skb[7][lid];
1128 barrier (CLK_LOCAL_MEM_FENCE);
1130 if (gid >= gid_max) return;
1136 m05500s (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);
1139 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05500_s16 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global u32 * 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)