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"
38 __device__ __constant__ u32 padding[8] =
58 __device__ static void swap (RC4_KEY *rc4_key, const u8 i, const u8 j)
63 rc4_key->S[i] = rc4_key->S[j];
67 __device__ static void rc4_init_16 (RC4_KEY *rc4_key, const u32 data[4])
72 u32 *ptr = (u32 *) rc4_key->S;
75 for (u32 i = 0; i < 64; i++)
80 const u32 d0 = data[0] >> 0;
81 const u32 d1 = data[0] >> 8;
82 const u32 d2 = data[0] >> 16;
83 const u32 d3 = data[0] >> 24;
84 const u32 d4 = data[1] >> 0;
90 for (i = 0; i < 255; i += 5)
92 j += rc4_key->S[i + 0]; j += d0; swap (rc4_key, i + 0, j);
93 j += rc4_key->S[i + 1]; j += d1; swap (rc4_key, i + 1, j);
94 j += rc4_key->S[i + 2]; j += d2; swap (rc4_key, i + 2, j);
95 j += rc4_key->S[i + 3]; j += d3; swap (rc4_key, i + 3, j);
96 j += rc4_key->S[i + 4]; j += d4; swap (rc4_key, i + 4, j);
99 j += rc4_key->S[i + 0]; j += d0; swap (rc4_key, i + 0, j);
102 __device__ static u8 rc4_next_16 (RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4])
104 for (u32 k = 0; k < 4; k++)
113 swap (rc4_key, i, j);
115 idx = rc4_key->S[i] + rc4_key->S[j];
117 xor4 |= rc4_key->S[idx] << 0;
122 swap (rc4_key, i, j);
124 idx = rc4_key->S[i] + rc4_key->S[j];
126 xor4 |= rc4_key->S[idx] << 8;
131 swap (rc4_key, i, j);
133 idx = rc4_key->S[i] + rc4_key->S[j];
135 xor4 |= rc4_key->S[idx] << 16;
140 swap (rc4_key, i, j);
142 idx = rc4_key->S[i] + rc4_key->S[j];
144 xor4 |= rc4_key->S[idx] << 24;
146 out[k] = in[k] ^ xor4;
152 __device__ __constant__ bf_t c_bfs[1024];
154 __device__ static void m10410m (RC4_KEY rc4_keys[64], 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 u32x * words_buf_r, 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 pdf_t *pdf_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)
160 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
161 const u32 lid = threadIdx.x;
163 RC4_KEY *rc4_key = &rc4_keys[lid];
171 for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos++)
173 const u32 w0r = c_bfs[il_pos].i;
186 rc4_init_16 (rc4_key, key);
190 rc4_next_16 (rc4_key, 0, 0, padding, out);
192 const u32x r0 = out[0];
193 const u32x r1 = out[1];
194 const u32x r2 = out[2];
195 const u32x r3 = out[3];
197 #include VECT_COMPARE_M
201 __device__ static void m10410s (RC4_KEY rc4_keys[64], 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 u32x * words_buf_r, 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 pdf_t *pdf_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)
207 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
208 const u32 lid = threadIdx.x;
210 RC4_KEY *rc4_key = &rc4_keys[lid];
216 const u32 search[4] =
218 digests_buf[digests_offset].digest_buf[DGST_R0],
219 digests_buf[digests_offset].digest_buf[DGST_R1],
220 digests_buf[digests_offset].digest_buf[DGST_R2],
221 digests_buf[digests_offset].digest_buf[DGST_R3]
230 for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos++)
232 const u32 w0r = c_bfs[il_pos].i;
245 rc4_init_16 (rc4_key, key);
249 rc4_next_16 (rc4_key, 0, 0, padding, out);
251 const u32x r0 = out[0];
252 const u32x r1 = out[1];
253 const u32x r2 = out[2];
254 const u32x r3 = out[3];
256 #include VECT_COMPARE_S
260 extern "C" __global__ void __launch_bounds__ (64, 1) m10410_m04 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const u32x * words_buf_r, 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 pdf_t *pdf_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)
266 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
268 if (gid >= gid_max) return;
272 w0[0] = pws[gid].i[ 0];
273 w0[1] = pws[gid].i[ 1];
274 w0[2] = pws[gid].i[ 2];
275 w0[3] = pws[gid].i[ 3];
298 const u32 pw_len = pws[gid].pw_len;
304 __shared__ RC4_KEY rc4_keys[64];
306 m10410m (rc4_keys, w0, w1, w2, w3, 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, pdf_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);
309 extern "C" __global__ void __launch_bounds__ (64, 1) m10410_m08 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const u32x * words_buf_r, 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 pdf_t *pdf_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)
315 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
317 if (gid >= gid_max) return;
321 w0[0] = pws[gid].i[ 0];
322 w0[1] = pws[gid].i[ 1];
323 w0[2] = pws[gid].i[ 2];
324 w0[3] = pws[gid].i[ 3];
328 w1[0] = pws[gid].i[ 4];
329 w1[1] = pws[gid].i[ 5];
330 w1[2] = pws[gid].i[ 6];
331 w1[3] = pws[gid].i[ 7];
347 const u32 pw_len = pws[gid].pw_len;
353 __shared__ RC4_KEY rc4_keys[64];
355 m10410m (rc4_keys, w0, w1, w2, w3, 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, pdf_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);
358 extern "C" __global__ void __launch_bounds__ (64, 1) m10410_m16 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const u32x * words_buf_r, 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 pdf_t *pdf_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)
364 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
366 if (gid >= gid_max) return;
370 w0[0] = pws[gid].i[ 0];
371 w0[1] = pws[gid].i[ 1];
372 w0[2] = pws[gid].i[ 2];
373 w0[3] = pws[gid].i[ 3];
377 w1[0] = pws[gid].i[ 4];
378 w1[1] = pws[gid].i[ 5];
379 w1[2] = pws[gid].i[ 6];
380 w1[3] = pws[gid].i[ 7];
384 w2[0] = pws[gid].i[ 8];
385 w2[1] = pws[gid].i[ 9];
386 w2[2] = pws[gid].i[10];
387 w2[3] = pws[gid].i[11];
391 w3[0] = pws[gid].i[12];
392 w3[1] = pws[gid].i[13];
396 const u32 pw_len = pws[gid].pw_len;
402 __shared__ RC4_KEY rc4_keys[64];
404 m10410m (rc4_keys, w0, w1, w2, w3, 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, pdf_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);
407 extern "C" __global__ void __launch_bounds__ (64, 1) m10410_s04 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const u32x * words_buf_r, 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 pdf_t *pdf_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)
413 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
415 if (gid >= gid_max) return;
419 w0[0] = pws[gid].i[ 0];
420 w0[1] = pws[gid].i[ 1];
421 w0[2] = pws[gid].i[ 2];
422 w0[3] = pws[gid].i[ 3];
445 const u32 pw_len = pws[gid].pw_len;
451 __shared__ RC4_KEY rc4_keys[64];
453 m10410s (rc4_keys, w0, w1, w2, w3, 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, pdf_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);
456 extern "C" __global__ void __launch_bounds__ (64, 1) m10410_s08 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const u32x * words_buf_r, 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 pdf_t *pdf_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)
462 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
464 if (gid >= gid_max) return;
468 w0[0] = pws[gid].i[ 0];
469 w0[1] = pws[gid].i[ 1];
470 w0[2] = pws[gid].i[ 2];
471 w0[3] = pws[gid].i[ 3];
475 w1[0] = pws[gid].i[ 4];
476 w1[1] = pws[gid].i[ 5];
477 w1[2] = pws[gid].i[ 6];
478 w1[3] = pws[gid].i[ 7];
494 const u32 pw_len = pws[gid].pw_len;
500 __shared__ RC4_KEY rc4_keys[64];
502 m10410s (rc4_keys, w0, w1, w2, w3, 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, pdf_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);
505 extern "C" __global__ void __launch_bounds__ (256, 1) m10410_s16 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const u32x * words_buf_r, 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 pdf_t *pdf_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)
511 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
513 if (gid >= gid_max) return;
517 w0[0] = pws[gid].i[ 0];
518 w0[1] = pws[gid].i[ 1];
519 w0[2] = pws[gid].i[ 2];
520 w0[3] = pws[gid].i[ 3];
524 w1[0] = pws[gid].i[ 4];
525 w1[1] = pws[gid].i[ 5];
526 w1[2] = pws[gid].i[ 6];
527 w1[3] = pws[gid].i[ 7];
531 w2[0] = pws[gid].i[ 8];
532 w2[1] = pws[gid].i[ 9];
533 w2[2] = pws[gid].i[10];
534 w2[3] = pws[gid].i[11];
538 w3[0] = pws[gid].i[12];
539 w3[1] = pws[gid].i[13];
543 const u32 pw_len = pws[gid].pw_len;
549 __shared__ RC4_KEY rc4_keys[64];
551 m10410s (rc4_keys, w0, w1, w2, w3, 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, pdf_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);