2 * Author......: Jens Steube <jens.steube@gmail.com>
8 #include "include/constants.h"
9 #include "include/kernel_vendor.h"
24 #include "include/kernel_functions.c"
26 #include "common_nv.c"
29 #define VECT_COMPARE_S "check_single_vect1_comp4.c"
30 #define VECT_COMPARE_M "check_multi_vect1_comp4.c"
34 #define VECT_COMPARE_S "check_single_vect2_comp4.c"
35 #define VECT_COMPARE_M "check_multi_vect2_comp4.c"
39 #define VECT_COMPARE_S "check_single_vect4_comp4.c"
40 #define VECT_COMPARE_M "check_multi_vect4_comp4.c"
43 __device__ static void md5_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4])
67 MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
68 MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
69 MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02);
70 MD5_STEP (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03);
71 MD5_STEP (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00);
72 MD5_STEP (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01);
73 MD5_STEP (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02);
74 MD5_STEP (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03);
75 MD5_STEP (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00);
76 MD5_STEP (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01);
77 MD5_STEP (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02);
78 MD5_STEP (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03);
79 MD5_STEP (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00);
80 MD5_STEP (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01);
81 MD5_STEP (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02);
82 MD5_STEP (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03);
84 MD5_STEP (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10);
85 MD5_STEP (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11);
86 MD5_STEP (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12);
87 MD5_STEP (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13);
88 MD5_STEP (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10);
89 MD5_STEP (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11);
90 MD5_STEP (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12);
91 MD5_STEP (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13);
92 MD5_STEP (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10);
93 MD5_STEP (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11);
94 MD5_STEP (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12);
95 MD5_STEP (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13);
96 MD5_STEP (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10);
97 MD5_STEP (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11);
98 MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12);
99 MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13);
101 MD5_STEP (MD5_H , a, b, c, d, w5_t, MD5C20, MD5S20);
102 MD5_STEP (MD5_H , d, a, b, c, w8_t, MD5C21, MD5S21);
103 MD5_STEP (MD5_H , c, d, a, b, wb_t, MD5C22, MD5S22);
104 MD5_STEP (MD5_H , b, c, d, a, we_t, MD5C23, MD5S23);
105 MD5_STEP (MD5_H , a, b, c, d, w1_t, MD5C24, MD5S20);
106 MD5_STEP (MD5_H , d, a, b, c, w4_t, MD5C25, MD5S21);
107 MD5_STEP (MD5_H , c, d, a, b, w7_t, MD5C26, MD5S22);
108 MD5_STEP (MD5_H , b, c, d, a, wa_t, MD5C27, MD5S23);
109 MD5_STEP (MD5_H , a, b, c, d, wd_t, MD5C28, MD5S20);
110 MD5_STEP (MD5_H , d, a, b, c, w0_t, MD5C29, MD5S21);
111 MD5_STEP (MD5_H , c, d, a, b, w3_t, MD5C2a, MD5S22);
112 MD5_STEP (MD5_H , b, c, d, a, w6_t, MD5C2b, MD5S23);
113 MD5_STEP (MD5_H , a, b, c, d, w9_t, MD5C2c, MD5S20);
114 MD5_STEP (MD5_H , d, a, b, c, wc_t, MD5C2d, MD5S21);
115 MD5_STEP (MD5_H , c, d, a, b, wf_t, MD5C2e, MD5S22);
116 MD5_STEP (MD5_H , b, c, d, a, w2_t, MD5C2f, MD5S23);
118 MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30);
119 MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31);
120 MD5_STEP (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32);
121 MD5_STEP (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33);
122 MD5_STEP (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30);
123 MD5_STEP (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31);
124 MD5_STEP (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32);
125 MD5_STEP (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33);
126 MD5_STEP (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30);
127 MD5_STEP (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31);
128 MD5_STEP (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32);
129 MD5_STEP (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33);
130 MD5_STEP (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30);
131 MD5_STEP (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31);
132 MD5_STEP (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32);
133 MD5_STEP (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33);
141 __device__ static void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
143 w0[0] = w0[0] ^ 0x36363636;
144 w0[1] = w0[1] ^ 0x36363636;
145 w0[2] = w0[2] ^ 0x36363636;
146 w0[3] = w0[3] ^ 0x36363636;
147 w1[0] = w1[0] ^ 0x36363636;
148 w1[1] = w1[1] ^ 0x36363636;
149 w1[2] = w1[2] ^ 0x36363636;
150 w1[3] = w1[3] ^ 0x36363636;
151 w2[0] = w2[0] ^ 0x36363636;
152 w2[1] = w2[1] ^ 0x36363636;
153 w2[2] = w2[2] ^ 0x36363636;
154 w2[3] = w2[3] ^ 0x36363636;
155 w3[0] = w3[0] ^ 0x36363636;
156 w3[1] = w3[1] ^ 0x36363636;
157 w3[2] = w3[2] ^ 0x36363636;
158 w3[3] = w3[3] ^ 0x36363636;
165 md5_transform (w0, w1, w2, w3, ipad);
167 w0[0] = w0[0] ^ 0x6a6a6a6a;
168 w0[1] = w0[1] ^ 0x6a6a6a6a;
169 w0[2] = w0[2] ^ 0x6a6a6a6a;
170 w0[3] = w0[3] ^ 0x6a6a6a6a;
171 w1[0] = w1[0] ^ 0x6a6a6a6a;
172 w1[1] = w1[1] ^ 0x6a6a6a6a;
173 w1[2] = w1[2] ^ 0x6a6a6a6a;
174 w1[3] = w1[3] ^ 0x6a6a6a6a;
175 w2[0] = w2[0] ^ 0x6a6a6a6a;
176 w2[1] = w2[1] ^ 0x6a6a6a6a;
177 w2[2] = w2[2] ^ 0x6a6a6a6a;
178 w2[3] = w2[3] ^ 0x6a6a6a6a;
179 w3[0] = w3[0] ^ 0x6a6a6a6a;
180 w3[1] = w3[1] ^ 0x6a6a6a6a;
181 w3[2] = w3[2] ^ 0x6a6a6a6a;
182 w3[3] = w3[3] ^ 0x6a6a6a6a;
189 md5_transform (w0, w1, w2, w3, opad);
192 __device__ static void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
199 md5_transform (w0, w1, w2, w3, digest);
215 w3[2] = (64 + 16) * 8;
223 md5_transform (w0, w1, w2, w3, digest);
226 __device__ __constant__ bf_t c_bfs[1024];
228 __device__ static void m05300m (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const u32 pw_len, const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const ikepsk_t *ikepsk_bufs, u32 *d_return_buf, 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, u32 s_msg_buf[128])
234 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
235 const u32 lid = threadIdx.x;
241 const u32 nr_len = ikepsk_bufs[salt_pos].nr_len;
242 const u32 msg_len = ikepsk_bufs[salt_pos].msg_len;
246 salt_buf0[0] = ikepsk_bufs[salt_pos].nr_buf[ 0];
247 salt_buf0[1] = ikepsk_bufs[salt_pos].nr_buf[ 1];
248 salt_buf0[2] = ikepsk_bufs[salt_pos].nr_buf[ 2];
249 salt_buf0[3] = ikepsk_bufs[salt_pos].nr_buf[ 3];
253 salt_buf1[0] = ikepsk_bufs[salt_pos].nr_buf[ 4];
254 salt_buf1[1] = ikepsk_bufs[salt_pos].nr_buf[ 5];
255 salt_buf1[2] = ikepsk_bufs[salt_pos].nr_buf[ 6];
256 salt_buf1[3] = ikepsk_bufs[salt_pos].nr_buf[ 7];
260 salt_buf2[0] = ikepsk_bufs[salt_pos].nr_buf[ 8];
261 salt_buf2[1] = ikepsk_bufs[salt_pos].nr_buf[ 9];
262 salt_buf2[2] = ikepsk_bufs[salt_pos].nr_buf[10];
263 salt_buf2[3] = ikepsk_bufs[salt_pos].nr_buf[11];
267 salt_buf3[0] = ikepsk_bufs[salt_pos].nr_buf[12];
268 salt_buf3[1] = ikepsk_bufs[salt_pos].nr_buf[13];
278 for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos++)
280 const u32 w0r = c_bfs[il_pos].i;
319 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
321 w0_t[0] = salt_buf0[0];
322 w0_t[1] = salt_buf0[1];
323 w0_t[2] = salt_buf0[2];
324 w0_t[3] = salt_buf0[3];
325 w1_t[0] = salt_buf1[0];
326 w1_t[1] = salt_buf1[1];
327 w1_t[2] = salt_buf1[2];
328 w1_t[3] = salt_buf1[3];
329 w2_t[0] = salt_buf2[0];
330 w2_t[1] = salt_buf2[1];
331 w2_t[2] = salt_buf2[2];
332 w2_t[3] = salt_buf2[3];
333 w3_t[0] = salt_buf3[0];
334 w3_t[1] = salt_buf3[1];
335 w3_t[2] = (64 + nr_len) * 8;
340 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
359 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
364 for (left = ikepsk_bufs[salt_pos].msg_len, off = 0; left >= 56; left -= 64, off += 16)
366 w0_t[0] = s_msg_buf[off + 0];
367 w0_t[1] = s_msg_buf[off + 1];
368 w0_t[2] = s_msg_buf[off + 2];
369 w0_t[3] = s_msg_buf[off + 3];
370 w1_t[0] = s_msg_buf[off + 4];
371 w1_t[1] = s_msg_buf[off + 5];
372 w1_t[2] = s_msg_buf[off + 6];
373 w1_t[3] = s_msg_buf[off + 7];
374 w2_t[0] = s_msg_buf[off + 8];
375 w2_t[1] = s_msg_buf[off + 9];
376 w2_t[2] = s_msg_buf[off + 10];
377 w2_t[3] = s_msg_buf[off + 11];
378 w3_t[0] = s_msg_buf[off + 12];
379 w3_t[1] = s_msg_buf[off + 13];
380 w3_t[2] = s_msg_buf[off + 14];
381 w3_t[3] = s_msg_buf[off + 15];
383 md5_transform (w0_t, w1_t, w2_t, w3_t, ipad);
386 w0_t[0] = s_msg_buf[off + 0];
387 w0_t[1] = s_msg_buf[off + 1];
388 w0_t[2] = s_msg_buf[off + 2];
389 w0_t[3] = s_msg_buf[off + 3];
390 w1_t[0] = s_msg_buf[off + 4];
391 w1_t[1] = s_msg_buf[off + 5];
392 w1_t[2] = s_msg_buf[off + 6];
393 w1_t[3] = s_msg_buf[off + 7];
394 w2_t[0] = s_msg_buf[off + 8];
395 w2_t[1] = s_msg_buf[off + 9];
396 w2_t[2] = s_msg_buf[off + 10];
397 w2_t[3] = s_msg_buf[off + 11];
398 w3_t[0] = s_msg_buf[off + 12];
399 w3_t[1] = s_msg_buf[off + 13];
400 w3_t[2] = (64 + msg_len) * 8;
403 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
405 const u32x r0 = digest[0];
406 const u32x r1 = digest[3];
407 const u32x r2 = digest[2];
408 const u32x r3 = digest[1];
410 #include VECT_COMPARE_M
414 __device__ static void m05300s (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const u32 pw_len, const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const ikepsk_t *ikepsk_bufs, u32 *d_return_buf, 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, u32 s_msg_buf[128])
420 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
421 const u32 lid = threadIdx.x;
427 const u32 nr_len = ikepsk_bufs[salt_pos].nr_len;
428 const u32 msg_len = ikepsk_bufs[salt_pos].msg_len;
432 salt_buf0[0] = ikepsk_bufs[salt_pos].nr_buf[ 0];
433 salt_buf0[1] = ikepsk_bufs[salt_pos].nr_buf[ 1];
434 salt_buf0[2] = ikepsk_bufs[salt_pos].nr_buf[ 2];
435 salt_buf0[3] = ikepsk_bufs[salt_pos].nr_buf[ 3];
439 salt_buf1[0] = ikepsk_bufs[salt_pos].nr_buf[ 4];
440 salt_buf1[1] = ikepsk_bufs[salt_pos].nr_buf[ 5];
441 salt_buf1[2] = ikepsk_bufs[salt_pos].nr_buf[ 6];
442 salt_buf1[3] = ikepsk_bufs[salt_pos].nr_buf[ 7];
446 salt_buf2[0] = ikepsk_bufs[salt_pos].nr_buf[ 8];
447 salt_buf2[1] = ikepsk_bufs[salt_pos].nr_buf[ 9];
448 salt_buf2[2] = ikepsk_bufs[salt_pos].nr_buf[10];
449 salt_buf2[3] = ikepsk_bufs[salt_pos].nr_buf[11];
453 salt_buf3[0] = ikepsk_bufs[salt_pos].nr_buf[12];
454 salt_buf3[1] = ikepsk_bufs[salt_pos].nr_buf[13];
462 const u32 search[4] =
464 digests_buf[digests_offset].digest_buf[DGST_R0],
465 digests_buf[digests_offset].digest_buf[DGST_R1],
466 digests_buf[digests_offset].digest_buf[DGST_R2],
467 digests_buf[digests_offset].digest_buf[DGST_R3]
476 for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos++)
478 const u32 w0r = c_bfs[il_pos].i;
517 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
519 w0_t[0] = salt_buf0[0];
520 w0_t[1] = salt_buf0[1];
521 w0_t[2] = salt_buf0[2];
522 w0_t[3] = salt_buf0[3];
523 w1_t[0] = salt_buf1[0];
524 w1_t[1] = salt_buf1[1];
525 w1_t[2] = salt_buf1[2];
526 w1_t[3] = salt_buf1[3];
527 w2_t[0] = salt_buf2[0];
528 w2_t[1] = salt_buf2[1];
529 w2_t[2] = salt_buf2[2];
530 w2_t[3] = salt_buf2[3];
531 w3_t[0] = salt_buf3[0];
532 w3_t[1] = salt_buf3[1];
533 w3_t[2] = (64 + nr_len) * 8;
538 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
557 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
562 for (left = ikepsk_bufs[salt_pos].msg_len, off = 0; left >= 56; left -= 64, off += 16)
564 w0_t[0] = s_msg_buf[off + 0];
565 w0_t[1] = s_msg_buf[off + 1];
566 w0_t[2] = s_msg_buf[off + 2];
567 w0_t[3] = s_msg_buf[off + 3];
568 w1_t[0] = s_msg_buf[off + 4];
569 w1_t[1] = s_msg_buf[off + 5];
570 w1_t[2] = s_msg_buf[off + 6];
571 w1_t[3] = s_msg_buf[off + 7];
572 w2_t[0] = s_msg_buf[off + 8];
573 w2_t[1] = s_msg_buf[off + 9];
574 w2_t[2] = s_msg_buf[off + 10];
575 w2_t[3] = s_msg_buf[off + 11];
576 w3_t[0] = s_msg_buf[off + 12];
577 w3_t[1] = s_msg_buf[off + 13];
578 w3_t[2] = s_msg_buf[off + 14];
579 w3_t[3] = s_msg_buf[off + 15];
581 md5_transform (w0_t, w1_t, w2_t, w3_t, ipad);
584 w0_t[0] = s_msg_buf[off + 0];
585 w0_t[1] = s_msg_buf[off + 1];
586 w0_t[2] = s_msg_buf[off + 2];
587 w0_t[3] = s_msg_buf[off + 3];
588 w1_t[0] = s_msg_buf[off + 4];
589 w1_t[1] = s_msg_buf[off + 5];
590 w1_t[2] = s_msg_buf[off + 6];
591 w1_t[3] = s_msg_buf[off + 7];
592 w2_t[0] = s_msg_buf[off + 8];
593 w2_t[1] = s_msg_buf[off + 9];
594 w2_t[2] = s_msg_buf[off + 10];
595 w2_t[3] = s_msg_buf[off + 11];
596 w3_t[0] = s_msg_buf[off + 12];
597 w3_t[1] = s_msg_buf[off + 13];
598 w3_t[2] = (64 + msg_len) * 8;
601 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
603 const u32x r0 = digest[0];
604 const u32x r1 = digest[3];
605 const u32x r2 = digest[2];
606 const u32x r3 = digest[1];
608 #include VECT_COMPARE_S
612 extern "C" __global__ void __launch_bounds__ (256, 1) m05300_m04 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const ikepsk_t *ikepsk_bufs, u32 *d_return_buf, 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)
618 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
619 const u32 lid = threadIdx.x;
623 w0[0] = pws[gid].i[ 0];
624 w0[1] = pws[gid].i[ 1];
625 w0[2] = pws[gid].i[ 2];
626 w0[3] = pws[gid].i[ 3];
649 const u32 pw_len = pws[gid].pw_len;
655 __shared__ u32 s_msg_buf[128];
659 s_msg_buf[lid] = ikepsk_bufs[salt_pos].msg_buf[lid];
664 if (gid >= gid_max) return;
670 m05300m (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, 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, ikepsk_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, s_msg_buf);
673 extern "C" __global__ void __launch_bounds__ (256, 1) m05300_m08 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const ikepsk_t *ikepsk_bufs, u32 *d_return_buf, 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 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
680 const u32 lid = threadIdx.x;
684 w0[0] = pws[gid].i[ 0];
685 w0[1] = pws[gid].i[ 1];
686 w0[2] = pws[gid].i[ 2];
687 w0[3] = pws[gid].i[ 3];
691 w1[0] = pws[gid].i[ 4];
692 w1[1] = pws[gid].i[ 5];
693 w1[2] = pws[gid].i[ 6];
694 w1[3] = pws[gid].i[ 7];
710 const u32 pw_len = pws[gid].pw_len;
716 __shared__ u32 s_msg_buf[128];
720 s_msg_buf[lid] = ikepsk_bufs[salt_pos].msg_buf[lid];
725 if (gid >= gid_max) return;
731 m05300m (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, 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, ikepsk_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, s_msg_buf);
734 extern "C" __global__ void __launch_bounds__ (256, 1) m05300_m16 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const ikepsk_t *ikepsk_bufs, u32 *d_return_buf, 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)
740 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
741 const u32 lid = threadIdx.x;
745 w0[0] = pws[gid].i[ 0];
746 w0[1] = pws[gid].i[ 1];
747 w0[2] = pws[gid].i[ 2];
748 w0[3] = pws[gid].i[ 3];
752 w1[0] = pws[gid].i[ 4];
753 w1[1] = pws[gid].i[ 5];
754 w1[2] = pws[gid].i[ 6];
755 w1[3] = pws[gid].i[ 7];
759 w2[0] = pws[gid].i[ 8];
760 w2[1] = pws[gid].i[ 9];
761 w2[2] = pws[gid].i[10];
762 w2[3] = pws[gid].i[11];
766 w3[0] = pws[gid].i[12];
767 w3[1] = pws[gid].i[13];
771 const u32 pw_len = pws[gid].pw_len;
777 __shared__ u32 s_msg_buf[128];
781 s_msg_buf[lid] = ikepsk_bufs[salt_pos].msg_buf[lid];
786 if (gid >= gid_max) return;
792 m05300m (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, 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, ikepsk_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, s_msg_buf);
795 extern "C" __global__ void __launch_bounds__ (256, 1) m05300_s04 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const ikepsk_t *ikepsk_bufs, u32 *d_return_buf, 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)
801 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
802 const u32 lid = threadIdx.x;
806 w0[0] = pws[gid].i[ 0];
807 w0[1] = pws[gid].i[ 1];
808 w0[2] = pws[gid].i[ 2];
809 w0[3] = pws[gid].i[ 3];
832 const u32 pw_len = pws[gid].pw_len;
838 __shared__ u32 s_msg_buf[128];
842 s_msg_buf[lid] = ikepsk_bufs[salt_pos].msg_buf[lid];
847 if (gid >= gid_max) return;
853 m05300s (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, 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, ikepsk_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, s_msg_buf);
856 extern "C" __global__ void __launch_bounds__ (256, 1) m05300_s08 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const ikepsk_t *ikepsk_bufs, u32 *d_return_buf, 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)
862 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
863 const u32 lid = threadIdx.x;
867 w0[0] = pws[gid].i[ 0];
868 w0[1] = pws[gid].i[ 1];
869 w0[2] = pws[gid].i[ 2];
870 w0[3] = pws[gid].i[ 3];
874 w1[0] = pws[gid].i[ 4];
875 w1[1] = pws[gid].i[ 5];
876 w1[2] = pws[gid].i[ 6];
877 w1[3] = pws[gid].i[ 7];
893 const u32 pw_len = pws[gid].pw_len;
899 __shared__ u32 s_msg_buf[128];
903 s_msg_buf[lid] = ikepsk_bufs[salt_pos].msg_buf[lid];
908 if (gid >= gid_max) return;
914 m05300s (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, 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, ikepsk_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, s_msg_buf);
917 extern "C" __global__ void __launch_bounds__ (256, 1) m05300_s16 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const ikepsk_t *ikepsk_bufs, u32 *d_return_buf, 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)
923 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
924 const u32 lid = threadIdx.x;
928 w0[0] = pws[gid].i[ 0];
929 w0[1] = pws[gid].i[ 1];
930 w0[2] = pws[gid].i[ 2];
931 w0[3] = pws[gid].i[ 3];
935 w1[0] = pws[gid].i[ 4];
936 w1[1] = pws[gid].i[ 5];
937 w1[2] = pws[gid].i[ 6];
938 w1[3] = pws[gid].i[ 7];
942 w2[0] = pws[gid].i[ 8];
943 w2[1] = pws[gid].i[ 9];
944 w2[2] = pws[gid].i[10];
945 w2[3] = pws[gid].i[11];
949 w3[0] = pws[gid].i[12];
950 w3[1] = pws[gid].i[13];
954 const u32 pw_len = pws[gid].pw_len;
960 __shared__ u32 s_msg_buf[128];
964 s_msg_buf[lid] = ikepsk_bufs[salt_pos].msg_buf[lid];
969 if (gid >= gid_max) return;
975 m05300s (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, 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, ikepsk_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, s_msg_buf);