2 * Author......: Jens Steube <jens.steube@gmail.com>
8 #include "include/constants.h"
9 #include "include/kernel_vendor.h"
28 #include "include/kernel_functions.c"
29 #include "types_amd.c"
30 #include "common_amd.c"
33 #define VECT_COMPARE_S "check_single_vect1_comp4.c"
34 #define VECT_COMPARE_M "check_multi_vect1_comp4.c"
38 #define VECT_COMPARE_S "check_single_vect2_comp4.c"
39 #define VECT_COMPARE_M "check_multi_vect2_comp4.c"
43 #define VECT_COMPARE_S "check_single_vect4_comp4.c"
44 #define VECT_COMPARE_M "check_multi_vect4_comp4.c"
47 #define PERM_OP(a,b,tt,n,m) \
57 #define HPERM_OP(a,tt,n,m) \
63 tt = tt >> (16 + n); \
67 __constant u32 c_SPtrans[8][64] =
70 0x00820200, 0x00020000, 0x80800000, 0x80820200,
71 0x00800000, 0x80020200, 0x80020000, 0x80800000,
72 0x80020200, 0x00820200, 0x00820000, 0x80000200,
73 0x80800200, 0x00800000, 0x00000000, 0x80020000,
74 0x00020000, 0x80000000, 0x00800200, 0x00020200,
75 0x80820200, 0x00820000, 0x80000200, 0x00800200,
76 0x80000000, 0x00000200, 0x00020200, 0x80820000,
77 0x00000200, 0x80800200, 0x80820000, 0x00000000,
78 0x00000000, 0x80820200, 0x00800200, 0x80020000,
79 0x00820200, 0x00020000, 0x80000200, 0x00800200,
80 0x80820000, 0x00000200, 0x00020200, 0x80800000,
81 0x80020200, 0x80000000, 0x80800000, 0x00820000,
82 0x80820200, 0x00020200, 0x00820000, 0x80800200,
83 0x00800000, 0x80000200, 0x80020000, 0x00000000,
84 0x00020000, 0x00800000, 0x80800200, 0x00820200,
85 0x80000000, 0x80820000, 0x00000200, 0x80020200,
87 0x10042004, 0x00000000, 0x00042000, 0x10040000,
88 0x10000004, 0x00002004, 0x10002000, 0x00042000,
89 0x00002000, 0x10040004, 0x00000004, 0x10002000,
90 0x00040004, 0x10042000, 0x10040000, 0x00000004,
91 0x00040000, 0x10002004, 0x10040004, 0x00002000,
92 0x00042004, 0x10000000, 0x00000000, 0x00040004,
93 0x10002004, 0x00042004, 0x10042000, 0x10000004,
94 0x10000000, 0x00040000, 0x00002004, 0x10042004,
95 0x00040004, 0x10042000, 0x10002000, 0x00042004,
96 0x10042004, 0x00040004, 0x10000004, 0x00000000,
97 0x10000000, 0x00002004, 0x00040000, 0x10040004,
98 0x00002000, 0x10000000, 0x00042004, 0x10002004,
99 0x10042000, 0x00002000, 0x00000000, 0x10000004,
100 0x00000004, 0x10042004, 0x00042000, 0x10040000,
101 0x10040004, 0x00040000, 0x00002004, 0x10002000,
102 0x10002004, 0x00000004, 0x10040000, 0x00042000,
104 0x41000000, 0x01010040, 0x00000040, 0x41000040,
105 0x40010000, 0x01000000, 0x41000040, 0x00010040,
106 0x01000040, 0x00010000, 0x01010000, 0x40000000,
107 0x41010040, 0x40000040, 0x40000000, 0x41010000,
108 0x00000000, 0x40010000, 0x01010040, 0x00000040,
109 0x40000040, 0x41010040, 0x00010000, 0x41000000,
110 0x41010000, 0x01000040, 0x40010040, 0x01010000,
111 0x00010040, 0x00000000, 0x01000000, 0x40010040,
112 0x01010040, 0x00000040, 0x40000000, 0x00010000,
113 0x40000040, 0x40010000, 0x01010000, 0x41000040,
114 0x00000000, 0x01010040, 0x00010040, 0x41010000,
115 0x40010000, 0x01000000, 0x41010040, 0x40000000,
116 0x40010040, 0x41000000, 0x01000000, 0x41010040,
117 0x00010000, 0x01000040, 0x41000040, 0x00010040,
118 0x01000040, 0x00000000, 0x41010000, 0x40000040,
119 0x41000000, 0x40010040, 0x00000040, 0x01010000,
121 0x00100402, 0x04000400, 0x00000002, 0x04100402,
122 0x00000000, 0x04100000, 0x04000402, 0x00100002,
123 0x04100400, 0x04000002, 0x04000000, 0x00000402,
124 0x04000002, 0x00100402, 0x00100000, 0x04000000,
125 0x04100002, 0x00100400, 0x00000400, 0x00000002,
126 0x00100400, 0x04000402, 0x04100000, 0x00000400,
127 0x00000402, 0x00000000, 0x00100002, 0x04100400,
128 0x04000400, 0x04100002, 0x04100402, 0x00100000,
129 0x04100002, 0x00000402, 0x00100000, 0x04000002,
130 0x00100400, 0x04000400, 0x00000002, 0x04100000,
131 0x04000402, 0x00000000, 0x00000400, 0x00100002,
132 0x00000000, 0x04100002, 0x04100400, 0x00000400,
133 0x04000000, 0x04100402, 0x00100402, 0x00100000,
134 0x04100402, 0x00000002, 0x04000400, 0x00100402,
135 0x00100002, 0x00100400, 0x04100000, 0x04000402,
136 0x00000402, 0x04000000, 0x04000002, 0x04100400,
138 0x02000000, 0x00004000, 0x00000100, 0x02004108,
139 0x02004008, 0x02000100, 0x00004108, 0x02004000,
140 0x00004000, 0x00000008, 0x02000008, 0x00004100,
141 0x02000108, 0x02004008, 0x02004100, 0x00000000,
142 0x00004100, 0x02000000, 0x00004008, 0x00000108,
143 0x02000100, 0x00004108, 0x00000000, 0x02000008,
144 0x00000008, 0x02000108, 0x02004108, 0x00004008,
145 0x02004000, 0x00000100, 0x00000108, 0x02004100,
146 0x02004100, 0x02000108, 0x00004008, 0x02004000,
147 0x00004000, 0x00000008, 0x02000008, 0x02000100,
148 0x02000000, 0x00004100, 0x02004108, 0x00000000,
149 0x00004108, 0x02000000, 0x00000100, 0x00004008,
150 0x02000108, 0x00000100, 0x00000000, 0x02004108,
151 0x02004008, 0x02004100, 0x00000108, 0x00004000,
152 0x00004100, 0x02004008, 0x02000100, 0x00000108,
153 0x00000008, 0x00004108, 0x02004000, 0x02000008,
155 0x20000010, 0x00080010, 0x00000000, 0x20080800,
156 0x00080010, 0x00000800, 0x20000810, 0x00080000,
157 0x00000810, 0x20080810, 0x00080800, 0x20000000,
158 0x20000800, 0x20000010, 0x20080000, 0x00080810,
159 0x00080000, 0x20000810, 0x20080010, 0x00000000,
160 0x00000800, 0x00000010, 0x20080800, 0x20080010,
161 0x20080810, 0x20080000, 0x20000000, 0x00000810,
162 0x00000010, 0x00080800, 0x00080810, 0x20000800,
163 0x00000810, 0x20000000, 0x20000800, 0x00080810,
164 0x20080800, 0x00080010, 0x00000000, 0x20000800,
165 0x20000000, 0x00000800, 0x20080010, 0x00080000,
166 0x00080010, 0x20080810, 0x00080800, 0x00000010,
167 0x20080810, 0x00080800, 0x00080000, 0x20000810,
168 0x20000010, 0x20080000, 0x00080810, 0x00000000,
169 0x00000800, 0x20000010, 0x20000810, 0x20080800,
170 0x20080000, 0x00000810, 0x00000010, 0x20080010,
172 0x00001000, 0x00000080, 0x00400080, 0x00400001,
173 0x00401081, 0x00001001, 0x00001080, 0x00000000,
174 0x00400000, 0x00400081, 0x00000081, 0x00401000,
175 0x00000001, 0x00401080, 0x00401000, 0x00000081,
176 0x00400081, 0x00001000, 0x00001001, 0x00401081,
177 0x00000000, 0x00400080, 0x00400001, 0x00001080,
178 0x00401001, 0x00001081, 0x00401080, 0x00000001,
179 0x00001081, 0x00401001, 0x00000080, 0x00400000,
180 0x00001081, 0x00401000, 0x00401001, 0x00000081,
181 0x00001000, 0x00000080, 0x00400000, 0x00401001,
182 0x00400081, 0x00001081, 0x00001080, 0x00000000,
183 0x00000080, 0x00400001, 0x00000001, 0x00400080,
184 0x00000000, 0x00400081, 0x00400080, 0x00001080,
185 0x00000081, 0x00001000, 0x00401081, 0x00400000,
186 0x00401080, 0x00000001, 0x00001001, 0x00401081,
187 0x00400001, 0x00401080, 0x00401000, 0x00001001,
189 0x08200020, 0x08208000, 0x00008020, 0x00000000,
190 0x08008000, 0x00200020, 0x08200000, 0x08208020,
191 0x00000020, 0x08000000, 0x00208000, 0x00008020,
192 0x00208020, 0x08008020, 0x08000020, 0x08200000,
193 0x00008000, 0x00208020, 0x00200020, 0x08008000,
194 0x08208020, 0x08000020, 0x00000000, 0x00208000,
195 0x08000000, 0x00200000, 0x08008020, 0x08200020,
196 0x00200000, 0x00008000, 0x08208000, 0x00000020,
197 0x00200000, 0x00008000, 0x08000020, 0x08208020,
198 0x00008020, 0x08000000, 0x00000000, 0x00208000,
199 0x08200020, 0x08008020, 0x08008000, 0x00200020,
200 0x08208000, 0x00000020, 0x00200020, 0x08008000,
201 0x08208020, 0x00200000, 0x08200000, 0x08000020,
202 0x00208000, 0x00008020, 0x08008020, 0x08200000,
203 0x00000020, 0x08208000, 0x00208020, 0x00000000,
204 0x08000000, 0x08200020, 0x00008000, 0x00208020
207 __constant u32 c_skb[8][64] =
209 /* for C bits (numbered as per FIPS 46) 1 2 3 4 5 6 */
210 0x00000000, 0x00000010, 0x20000000, 0x20000010,
211 0x00010000, 0x00010010, 0x20010000, 0x20010010,
212 0x00000800, 0x00000810, 0x20000800, 0x20000810,
213 0x00010800, 0x00010810, 0x20010800, 0x20010810,
214 0x00000020, 0x00000030, 0x20000020, 0x20000030,
215 0x00010020, 0x00010030, 0x20010020, 0x20010030,
216 0x00000820, 0x00000830, 0x20000820, 0x20000830,
217 0x00010820, 0x00010830, 0x20010820, 0x20010830,
218 0x00080000, 0x00080010, 0x20080000, 0x20080010,
219 0x00090000, 0x00090010, 0x20090000, 0x20090010,
220 0x00080800, 0x00080810, 0x20080800, 0x20080810,
221 0x00090800, 0x00090810, 0x20090800, 0x20090810,
222 0x00080020, 0x00080030, 0x20080020, 0x20080030,
223 0x00090020, 0x00090030, 0x20090020, 0x20090030,
224 0x00080820, 0x00080830, 0x20080820, 0x20080830,
225 0x00090820, 0x00090830, 0x20090820, 0x20090830,
226 /* for C bits (numbered as per FIPS 46) 7 8 10 11 12 13 */
227 0x00000000, 0x02000000, 0x00002000, 0x02002000,
228 0x00200000, 0x02200000, 0x00202000, 0x02202000,
229 0x00000004, 0x02000004, 0x00002004, 0x02002004,
230 0x00200004, 0x02200004, 0x00202004, 0x02202004,
231 0x00000400, 0x02000400, 0x00002400, 0x02002400,
232 0x00200400, 0x02200400, 0x00202400, 0x02202400,
233 0x00000404, 0x02000404, 0x00002404, 0x02002404,
234 0x00200404, 0x02200404, 0x00202404, 0x02202404,
235 0x10000000, 0x12000000, 0x10002000, 0x12002000,
236 0x10200000, 0x12200000, 0x10202000, 0x12202000,
237 0x10000004, 0x12000004, 0x10002004, 0x12002004,
238 0x10200004, 0x12200004, 0x10202004, 0x12202004,
239 0x10000400, 0x12000400, 0x10002400, 0x12002400,
240 0x10200400, 0x12200400, 0x10202400, 0x12202400,
241 0x10000404, 0x12000404, 0x10002404, 0x12002404,
242 0x10200404, 0x12200404, 0x10202404, 0x12202404,
243 /* for C bits (numbered as per FIPS 46) 14 15 16 17 19 20 */
244 0x00000000, 0x00000001, 0x00040000, 0x00040001,
245 0x01000000, 0x01000001, 0x01040000, 0x01040001,
246 0x00000002, 0x00000003, 0x00040002, 0x00040003,
247 0x01000002, 0x01000003, 0x01040002, 0x01040003,
248 0x00000200, 0x00000201, 0x00040200, 0x00040201,
249 0x01000200, 0x01000201, 0x01040200, 0x01040201,
250 0x00000202, 0x00000203, 0x00040202, 0x00040203,
251 0x01000202, 0x01000203, 0x01040202, 0x01040203,
252 0x08000000, 0x08000001, 0x08040000, 0x08040001,
253 0x09000000, 0x09000001, 0x09040000, 0x09040001,
254 0x08000002, 0x08000003, 0x08040002, 0x08040003,
255 0x09000002, 0x09000003, 0x09040002, 0x09040003,
256 0x08000200, 0x08000201, 0x08040200, 0x08040201,
257 0x09000200, 0x09000201, 0x09040200, 0x09040201,
258 0x08000202, 0x08000203, 0x08040202, 0x08040203,
259 0x09000202, 0x09000203, 0x09040202, 0x09040203,
260 /* for C bits (numbered as per FIPS 46) 21 23 24 26 27 28 */
261 0x00000000, 0x00100000, 0x00000100, 0x00100100,
262 0x00000008, 0x00100008, 0x00000108, 0x00100108,
263 0x00001000, 0x00101000, 0x00001100, 0x00101100,
264 0x00001008, 0x00101008, 0x00001108, 0x00101108,
265 0x04000000, 0x04100000, 0x04000100, 0x04100100,
266 0x04000008, 0x04100008, 0x04000108, 0x04100108,
267 0x04001000, 0x04101000, 0x04001100, 0x04101100,
268 0x04001008, 0x04101008, 0x04001108, 0x04101108,
269 0x00020000, 0x00120000, 0x00020100, 0x00120100,
270 0x00020008, 0x00120008, 0x00020108, 0x00120108,
271 0x00021000, 0x00121000, 0x00021100, 0x00121100,
272 0x00021008, 0x00121008, 0x00021108, 0x00121108,
273 0x04020000, 0x04120000, 0x04020100, 0x04120100,
274 0x04020008, 0x04120008, 0x04020108, 0x04120108,
275 0x04021000, 0x04121000, 0x04021100, 0x04121100,
276 0x04021008, 0x04121008, 0x04021108, 0x04121108,
277 /* for D bits (numbered as per FIPS 46) 1 2 3 4 5 6 */
278 0x00000000, 0x10000000, 0x00010000, 0x10010000,
279 0x00000004, 0x10000004, 0x00010004, 0x10010004,
280 0x20000000, 0x30000000, 0x20010000, 0x30010000,
281 0x20000004, 0x30000004, 0x20010004, 0x30010004,
282 0x00100000, 0x10100000, 0x00110000, 0x10110000,
283 0x00100004, 0x10100004, 0x00110004, 0x10110004,
284 0x20100000, 0x30100000, 0x20110000, 0x30110000,
285 0x20100004, 0x30100004, 0x20110004, 0x30110004,
286 0x00001000, 0x10001000, 0x00011000, 0x10011000,
287 0x00001004, 0x10001004, 0x00011004, 0x10011004,
288 0x20001000, 0x30001000, 0x20011000, 0x30011000,
289 0x20001004, 0x30001004, 0x20011004, 0x30011004,
290 0x00101000, 0x10101000, 0x00111000, 0x10111000,
291 0x00101004, 0x10101004, 0x00111004, 0x10111004,
292 0x20101000, 0x30101000, 0x20111000, 0x30111000,
293 0x20101004, 0x30101004, 0x20111004, 0x30111004,
294 /* for D bits (numbered as per FIPS 46) 8 9 11 12 13 14 */
295 0x00000000, 0x08000000, 0x00000008, 0x08000008,
296 0x00000400, 0x08000400, 0x00000408, 0x08000408,
297 0x00020000, 0x08020000, 0x00020008, 0x08020008,
298 0x00020400, 0x08020400, 0x00020408, 0x08020408,
299 0x00000001, 0x08000001, 0x00000009, 0x08000009,
300 0x00000401, 0x08000401, 0x00000409, 0x08000409,
301 0x00020001, 0x08020001, 0x00020009, 0x08020009,
302 0x00020401, 0x08020401, 0x00020409, 0x08020409,
303 0x02000000, 0x0A000000, 0x02000008, 0x0A000008,
304 0x02000400, 0x0A000400, 0x02000408, 0x0A000408,
305 0x02020000, 0x0A020000, 0x02020008, 0x0A020008,
306 0x02020400, 0x0A020400, 0x02020408, 0x0A020408,
307 0x02000001, 0x0A000001, 0x02000009, 0x0A000009,
308 0x02000401, 0x0A000401, 0x02000409, 0x0A000409,
309 0x02020001, 0x0A020001, 0x02020009, 0x0A020009,
310 0x02020401, 0x0A020401, 0x02020409, 0x0A020409,
311 /* for D bits (numbered as per FIPS 46) 16 17 18 19 20 21 */
312 0x00000000, 0x00000100, 0x00080000, 0x00080100,
313 0x01000000, 0x01000100, 0x01080000, 0x01080100,
314 0x00000010, 0x00000110, 0x00080010, 0x00080110,
315 0x01000010, 0x01000110, 0x01080010, 0x01080110,
316 0x00200000, 0x00200100, 0x00280000, 0x00280100,
317 0x01200000, 0x01200100, 0x01280000, 0x01280100,
318 0x00200010, 0x00200110, 0x00280010, 0x00280110,
319 0x01200010, 0x01200110, 0x01280010, 0x01280110,
320 0x00000200, 0x00000300, 0x00080200, 0x00080300,
321 0x01000200, 0x01000300, 0x01080200, 0x01080300,
322 0x00000210, 0x00000310, 0x00080210, 0x00080310,
323 0x01000210, 0x01000310, 0x01080210, 0x01080310,
324 0x00200200, 0x00200300, 0x00280200, 0x00280300,
325 0x01200200, 0x01200300, 0x01280200, 0x01280300,
326 0x00200210, 0x00200310, 0x00280210, 0x00280310,
327 0x01200210, 0x01200310, 0x01280210, 0x01280310,
328 /* for D bits (numbered as per FIPS 46) 22 23 24 25 27 28 */
329 0x00000000, 0x04000000, 0x00040000, 0x04040000,
330 0x00000002, 0x04000002, 0x00040002, 0x04040002,
331 0x00002000, 0x04002000, 0x00042000, 0x04042000,
332 0x00002002, 0x04002002, 0x00042002, 0x04042002,
333 0x00000020, 0x04000020, 0x00040020, 0x04040020,
334 0x00000022, 0x04000022, 0x00040022, 0x04040022,
335 0x00002020, 0x04002020, 0x00042020, 0x04042020,
336 0x00002022, 0x04002022, 0x00042022, 0x04042022,
337 0x00000800, 0x04000800, 0x00040800, 0x04040800,
338 0x00000802, 0x04000802, 0x00040802, 0x04040802,
339 0x00002800, 0x04002800, 0x00042800, 0x04042800,
340 0x00002802, 0x04002802, 0x00042802, 0x04042802,
341 0x00000820, 0x04000820, 0x00040820, 0x04040820,
342 0x00000822, 0x04000822, 0x00040822, 0x04040822,
343 0x00002820, 0x04002820, 0x00042820, 0x04042820,
344 0x00002822, 0x04002822, 0x00042822, 0x04042822
347 __constant u32 shifts3s0[16] = { 1, 1, 2, 2, 2, 2, 2, 2, 1, 2, 2, 2, 2, 2, 2, 1 };
348 __constant u32 shifts3s1[16] = { 27, 27, 26, 26, 26, 26, 26, 26, 27, 26, 26, 26, 26, 26, 26, 27 };
351 #define BOX(i,n,S) (u32x) ((S)[(n)][(i)])
355 #define BOX(i,n,S) (u32x) ((S)[(n)][(i).s0], (S)[(n)][(i).s1])
359 #define BOX(i,n,S) (u32x) ((S)[(n)][(i).s0], (S)[(n)][(i).s1], (S)[(n)][(i).s2], (S)[(n)][(i).s3])
362 static void _des_crypt_keysetup (u32x c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 s_skb[8][64])
366 PERM_OP (d, c, tt, 4, 0x0f0f0f0f);
367 HPERM_OP (c, tt, 2, 0xcccc0000);
368 HPERM_OP (d, tt, 2, 0xcccc0000);
369 PERM_OP (d, c, tt, 1, 0x55555555);
370 PERM_OP (c, d, tt, 8, 0x00ff00ff);
371 PERM_OP (d, c, tt, 1, 0x55555555);
373 d = ((d & 0x000000ff) << 16)
374 | ((d & 0x0000ff00) << 0)
375 | ((d & 0x00ff0000) >> 16)
376 | ((c & 0xf0000000) >> 4);
381 for (u32 i = 0; i < 16; i++)
383 c = c >> shifts3s0[i] | c << shifts3s1[i];
384 d = d >> shifts3s0[i] | d << shifts3s1[i];
389 const u32x c00 = (c >> 0) & 0x0000003f;
390 const u32x c06 = (c >> 6) & 0x00383003;
391 const u32x c07 = (c >> 7) & 0x0000003c;
392 const u32x c13 = (c >> 13) & 0x0000060f;
393 const u32x c20 = (c >> 20) & 0x00000001;
395 u32x s = BOX (((c00 >> 0) & 0xff), 0, s_skb)
396 | BOX (((c06 >> 0) & 0xff)
397 |((c07 >> 0) & 0xff), 1, s_skb)
398 | BOX (((c13 >> 0) & 0xff)
399 |((c06 >> 8) & 0xff), 2, s_skb)
400 | BOX (((c20 >> 0) & 0xff)
402 |((c06 >> 16) & 0xff), 3, s_skb);
404 const u32x d00 = (d >> 0) & 0x00003c3f;
405 const u32x d07 = (d >> 7) & 0x00003f03;
406 const u32x d21 = (d >> 21) & 0x0000000f;
407 const u32x d22 = (d >> 22) & 0x00000030;
409 u32x t = BOX (((d00 >> 0) & 0xff), 4, s_skb)
410 | BOX (((d07 >> 0) & 0xff)
411 |((d00 >> 8) & 0xff), 5, s_skb)
412 | BOX (((d07 >> 8) & 0xff), 6, s_skb)
413 | BOX (((d21 >> 0) & 0xff)
414 |((d22 >> 0) & 0xff), 7, s_skb);
416 Kc[i] = ((t << 16) | (s & 0x0000ffff));
417 Kd[i] = ((s >> 16) | (t & 0xffff0000));
421 static void _des_crypt_encrypt (u32x iv[2], u32 mask, u32x Kc[16], u32x Kd[16], __local u32 s_SPtrans[8][64])
423 const u32 E1 = (mask >> 2) & 0x3f0;
424 const u32 E0 = mask & 0x3f;
429 for (u32 i = 0; i < 25; i++)
431 for (u32 j = 0; j < 16; j += 2)
448 ^ BOX (amd_bfe (u, 0, 6), 0, s_SPtrans)
449 ^ BOX (amd_bfe (u, 8, 6), 2, s_SPtrans)
450 ^ BOX (amd_bfe (u, 16, 6), 4, s_SPtrans)
451 ^ BOX (amd_bfe (u, 24, 6), 6, s_SPtrans)
452 ^ BOX (amd_bfe (t, 0, 6), 1, s_SPtrans)
453 ^ BOX (amd_bfe (t, 8, 6), 3, s_SPtrans)
454 ^ BOX (amd_bfe (t, 16, 6), 5, s_SPtrans)
455 ^ BOX (amd_bfe (t, 24, 6), 7, s_SPtrans);
469 ^ BOX (amd_bfe (u, 0, 6), 0, s_SPtrans)
470 ^ BOX (amd_bfe (u, 8, 6), 2, s_SPtrans)
471 ^ BOX (amd_bfe (u, 16, 6), 4, s_SPtrans)
472 ^ BOX (amd_bfe (u, 24, 6), 6, s_SPtrans)
473 ^ BOX (amd_bfe (t, 0, 6), 1, s_SPtrans)
474 ^ BOX (amd_bfe (t, 8, 6), 3, s_SPtrans)
475 ^ BOX (amd_bfe (t, 16, 6), 5, s_SPtrans)
476 ^ BOX (amd_bfe (t, 24, 6), 7, s_SPtrans);
486 iv[0] = rotl32 (r, 31);
487 iv[1] = rotl32 (l, 31);
490 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01500_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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
496 const u32 lid = get_local_id (0);
502 const u32 gid = get_global_id (0);
506 wordl0[0] = pws[gid].i[ 0];
507 wordl0[1] = pws[gid].i[ 1];
532 const u32 pw_l_len = pws[gid].pw_len;
534 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
536 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
543 __local u32 s_SPtrans[8][64];
545 __local u32 s_skb[8][64];
547 s_SPtrans[0][lid] = c_SPtrans[0][lid];
548 s_SPtrans[1][lid] = c_SPtrans[1][lid];
549 s_SPtrans[2][lid] = c_SPtrans[2][lid];
550 s_SPtrans[3][lid] = c_SPtrans[3][lid];
551 s_SPtrans[4][lid] = c_SPtrans[4][lid];
552 s_SPtrans[5][lid] = c_SPtrans[5][lid];
553 s_SPtrans[6][lid] = c_SPtrans[6][lid];
554 s_SPtrans[7][lid] = c_SPtrans[7][lid];
556 s_skb[0][lid] = c_skb[0][lid];
557 s_skb[1][lid] = c_skb[1][lid];
558 s_skb[2][lid] = c_skb[2][lid];
559 s_skb[3][lid] = c_skb[3][lid];
560 s_skb[4][lid] = c_skb[4][lid];
561 s_skb[5][lid] = c_skb[5][lid];
562 s_skb[6][lid] = c_skb[6][lid];
563 s_skb[7][lid] = c_skb[7][lid];
565 barrier (CLK_LOCAL_MEM_FENCE);
567 if (gid >= gid_max) return;
573 const u32 mask = salt_bufs[salt_pos].salt_buf[0];
579 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
581 const u32 pw_r_len = combs_buf[il_pos].pw_len;
583 u32 pw_len = pw_l_len + pw_r_len;
585 pw_len = (pw_len >= 8) ? 8 : pw_len;
589 wordr0[0] = combs_buf[il_pos].i[0];
590 wordr0[1] = combs_buf[il_pos].i[1];
615 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
617 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
622 w0[0] = wordl0[0] | wordr0[0];
623 w0[1] = wordl0[1] | wordr0[1];
650 data[0] = (w0[0] << 1) & 0xfefefefe;
651 data[1] = (w0[1] << 1) & 0xfefefefe;
656 _des_crypt_keysetup (data[0], data[1], Kc, Kd, s_skb);
660 _des_crypt_encrypt (iv, mask, Kc, Kd, s_SPtrans);
662 const u32x r0 = iv[0];
663 const u32x r1 = iv[1];
667 #include VECT_COMPARE_M
671 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01500_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
675 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01500_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
679 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01500_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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
685 const u32 lid = get_local_id (0);
691 const u32 gid = get_global_id (0);
695 wordl0[0] = pws[gid].i[ 0];
696 wordl0[1] = pws[gid].i[ 1];
721 const u32 pw_l_len = pws[gid].pw_len;
723 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
725 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
732 __local u32 s_SPtrans[8][64];
734 __local u32 s_skb[8][64];
736 s_SPtrans[0][lid] = c_SPtrans[0][lid];
737 s_SPtrans[1][lid] = c_SPtrans[1][lid];
738 s_SPtrans[2][lid] = c_SPtrans[2][lid];
739 s_SPtrans[3][lid] = c_SPtrans[3][lid];
740 s_SPtrans[4][lid] = c_SPtrans[4][lid];
741 s_SPtrans[5][lid] = c_SPtrans[5][lid];
742 s_SPtrans[6][lid] = c_SPtrans[6][lid];
743 s_SPtrans[7][lid] = c_SPtrans[7][lid];
745 s_skb[0][lid] = c_skb[0][lid];
746 s_skb[1][lid] = c_skb[1][lid];
747 s_skb[2][lid] = c_skb[2][lid];
748 s_skb[3][lid] = c_skb[3][lid];
749 s_skb[4][lid] = c_skb[4][lid];
750 s_skb[5][lid] = c_skb[5][lid];
751 s_skb[6][lid] = c_skb[6][lid];
752 s_skb[7][lid] = c_skb[7][lid];
754 barrier (CLK_LOCAL_MEM_FENCE);
756 if (gid >= gid_max) return;
762 const u32 mask = salt_bufs[salt_pos].salt_buf[0];
768 const u32 search[4] =
770 digests_buf[digests_offset].digest_buf[DGST_R0],
771 digests_buf[digests_offset].digest_buf[DGST_R1],
772 digests_buf[digests_offset].digest_buf[DGST_R2],
773 digests_buf[digests_offset].digest_buf[DGST_R3]
780 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
782 const u32 pw_r_len = combs_buf[il_pos].pw_len;
784 u32 pw_len = pw_l_len + pw_r_len;
786 pw_len = (pw_len >= 8) ? 8 : pw_len;
790 wordr0[0] = combs_buf[il_pos].i[0];
791 wordr0[1] = combs_buf[il_pos].i[1];
816 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
818 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
823 w0[0] = wordl0[0] | wordr0[0];
824 w0[1] = wordl0[1] | wordr0[1];
851 data[0] = (w0[0] << 1) & 0xfefefefe;
852 data[1] = (w0[1] << 1) & 0xfefefefe;
857 _des_crypt_keysetup (data[0], data[1], Kc, Kd, s_skb);
861 _des_crypt_encrypt (iv, mask, Kc, Kd, s_SPtrans);
863 const u32x r0 = iv[0];
864 const u32x r1 = iv[1];
868 #include VECT_COMPARE_S
872 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01500_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
876 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01500_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)