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__ comb_t c_combs[1024];
154 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 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 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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
160 const u32 lid = threadIdx.x;
166 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
168 if (gid >= gid_max) return;
172 wordl0[0] = pws[gid].i[ 0];
173 wordl0[1] = pws[gid].i[ 1];
174 wordl0[2] = pws[gid].i[ 2];
175 wordl0[3] = pws[gid].i[ 3];
179 wordl1[0] = pws[gid].i[ 4];
180 wordl1[1] = pws[gid].i[ 5];
181 wordl1[2] = pws[gid].i[ 6];
182 wordl1[3] = pws[gid].i[ 7];
198 const u32 pw_l_len = pws[gid].pw_len;
200 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
202 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, c_combs[0].pw_len);
209 __shared__ RC4_KEY rc4_keys[64];
210 RC4_KEY *rc4_key = &rc4_keys[lid];
216 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
220 wordr0[0] = c_combs[il_pos].i[0];
221 wordr0[1] = c_combs[il_pos].i[1];
222 wordr0[2] = c_combs[il_pos].i[2];
223 wordr0[3] = c_combs[il_pos].i[3];
227 wordr1[0] = c_combs[il_pos].i[4];
228 wordr1[1] = c_combs[il_pos].i[5];
229 wordr1[2] = c_combs[il_pos].i[6];
230 wordr1[3] = c_combs[il_pos].i[7];
246 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
248 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
253 w0[0] = wordl0[0] | wordr0[0];
254 w0[1] = wordl0[1] | wordr0[1];
265 rc4_init_16 (rc4_key, key);
269 rc4_next_16 (rc4_key, 0, 0, padding, out);
271 const u32x r0 = out[0];
272 const u32x r1 = out[1];
273 const u32x r2 = out[2];
274 const u32x r3 = out[3];
276 #include VECT_COMPARE_M
280 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 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 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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
284 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 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 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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
288 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 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 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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
294 const u32 lid = threadIdx.x;
300 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
302 if (gid >= gid_max) return;
306 wordl0[0] = pws[gid].i[ 0];
307 wordl0[1] = pws[gid].i[ 1];
308 wordl0[2] = pws[gid].i[ 2];
309 wordl0[3] = pws[gid].i[ 3];
313 wordl1[0] = pws[gid].i[ 4];
314 wordl1[1] = pws[gid].i[ 5];
315 wordl1[2] = pws[gid].i[ 6];
316 wordl1[3] = pws[gid].i[ 7];
332 const u32 pw_l_len = pws[gid].pw_len;
334 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
336 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, c_combs[0].pw_len);
343 const u32 search[4] =
345 digests_buf[digests_offset].digest_buf[DGST_R0],
346 digests_buf[digests_offset].digest_buf[DGST_R1],
347 digests_buf[digests_offset].digest_buf[DGST_R2],
348 digests_buf[digests_offset].digest_buf[DGST_R3]
355 __shared__ RC4_KEY rc4_keys[64];
356 RC4_KEY *rc4_key = &rc4_keys[lid];
362 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
366 wordr0[0] = c_combs[il_pos].i[0];
367 wordr0[1] = c_combs[il_pos].i[1];
368 wordr0[2] = c_combs[il_pos].i[2];
369 wordr0[3] = c_combs[il_pos].i[3];
373 wordr1[0] = c_combs[il_pos].i[4];
374 wordr1[1] = c_combs[il_pos].i[5];
375 wordr1[2] = c_combs[il_pos].i[6];
376 wordr1[3] = c_combs[il_pos].i[7];
392 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
394 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
399 w0[0] = wordl0[0] | wordr0[0];
400 w0[1] = wordl0[1] | wordr0[1];
411 rc4_init_16 (rc4_key, key);
415 rc4_next_16 (rc4_key, 0, 0, padding, out);
417 const u32x r0 = out[0];
418 const u32x r1 = out[1];
419 const u32x r2 = out[2];
420 const u32x r3 = out[3];
422 #include VECT_COMPARE_S
426 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 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 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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
430 extern "C" __global__ void __launch_bounds__ (64, 1) m10410_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 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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)