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 #define SIPROUND(v0,v1,v2,v3) \
45 (v1) = rotl64 ((v1), 13); \
47 (v0) = rotl64 ((v0), 32); \
49 (v3) = rotl64 ((v3), 16); \
52 (v3) = rotl64 ((v3), 21); \
55 (v1) = rotl64 ((v1), 17); \
57 (v2) = rotl64 ((v2), 32);
59 __device__ __constant__ comb_t c_combs[1024];
61 extern "C" __global__ void __launch_bounds__ (256, 1) m10100_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 void *esalt_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)
67 const u32 lid = threadIdx.x;
73 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
75 if (gid >= gid_max) return;
79 wordl0[0] = pws[gid].i[ 0];
80 wordl0[1] = pws[gid].i[ 1];
81 wordl0[2] = pws[gid].i[ 2];
82 wordl0[3] = pws[gid].i[ 3];
86 wordl1[0] = pws[gid].i[ 4];
87 wordl1[1] = pws[gid].i[ 5];
88 wordl1[2] = pws[gid].i[ 6];
89 wordl1[3] = pws[gid].i[ 7];
105 const u32 pw_l_len = pws[gid].pw_len;
107 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
109 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, c_combs[0].pw_len);
116 u64 v0p = SIPHASHM_0;
117 u64 v1p = SIPHASHM_1;
118 u64 v2p = SIPHASHM_2;
119 u64 v3p = SIPHASHM_3;
121 v0p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[1], salt_bufs[salt_pos].salt_buf[0]);
122 v1p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[3], salt_bufs[salt_pos].salt_buf[2]);
123 v2p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[1], salt_bufs[salt_pos].salt_buf[0]);
124 v3p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[3], salt_bufs[salt_pos].salt_buf[2]);
130 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
132 const u32 pw_r_len = c_combs[il_pos].pw_len;
134 const u32 pw_len = pw_l_len + pw_r_len;
138 wordr0[0] = c_combs[il_pos].i[0];
139 wordr0[1] = c_combs[il_pos].i[1];
140 wordr0[2] = c_combs[il_pos].i[2];
141 wordr0[3] = c_combs[il_pos].i[3];
145 wordr1[0] = c_combs[il_pos].i[4];
146 wordr1[1] = c_combs[il_pos].i[5];
147 wordr1[2] = c_combs[il_pos].i[6];
148 wordr1[3] = c_combs[il_pos].i[7];
164 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
166 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
171 w[ 0] = wordl0[0] | wordr0[0];
172 w[ 1] = wordl0[1] | wordr0[1];
173 w[ 2] = wordl0[2] | wordr0[2];
174 w[ 3] = wordl0[3] | wordr0[3];
175 w[ 4] = wordl1[0] | wordr1[0];
176 w[ 5] = wordl1[1] | wordr1[1];
177 w[ 6] = wordl1[2] | wordr1[2];
178 w[ 7] = wordl1[3] | wordr1[3];
179 w[ 8] = wordl2[0] | wordr2[0];
180 w[ 9] = wordl2[1] | wordr2[1];
181 w[10] = wordl2[2] | wordr2[2];
182 w[11] = wordl2[3] | wordr2[3];
183 w[12] = wordl3[0] | wordr3[0];
184 w[13] = wordl3[1] | wordr3[1];
185 w[14] = wordl3[2] | wordr3[2];
186 w[15] = wordl3[3] | wordr3[3];
188 u64 *w_ptr = (u64 *) w;
190 w_ptr[pw_len / 8] |= (u64) pw_len << 56;
200 for (i = 0, j = 0; i <= pw_len; i += 8, j += 2)
202 u64x m = hl32_to_64 (w[j + 1], w[j + 0]);
206 SIPROUND (v0, v1, v2, v3);
207 SIPROUND (v0, v1, v2, v3);
214 SIPROUND (v0, v1, v2, v3);
215 SIPROUND (v0, v1, v2, v3);
216 SIPROUND (v0, v1, v2, v3);
217 SIPROUND (v0, v1, v2, v3);
219 const u64x v = v0 ^ v1 ^ v2 ^ v3;
221 const u32x a = l32_from_64 (v);
222 const u32x b = h32_from_64 (v);
229 #include VECT_COMPARE_M
233 extern "C" __global__ void __launch_bounds__ (256, 1) m10100_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 void *esalt_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)
237 extern "C" __global__ void __launch_bounds__ (256, 1) m10100_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 void *esalt_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)
241 extern "C" __global__ void __launch_bounds__ (256, 1) m10100_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 void *esalt_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)
247 const u32 lid = threadIdx.x;
253 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
255 if (gid >= gid_max) return;
259 wordl0[0] = pws[gid].i[ 0];
260 wordl0[1] = pws[gid].i[ 1];
261 wordl0[2] = pws[gid].i[ 2];
262 wordl0[3] = pws[gid].i[ 3];
266 wordl1[0] = pws[gid].i[ 4];
267 wordl1[1] = pws[gid].i[ 5];
268 wordl1[2] = pws[gid].i[ 6];
269 wordl1[3] = pws[gid].i[ 7];
285 const u32 pw_l_len = pws[gid].pw_len;
287 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
289 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, c_combs[0].pw_len);
296 u64 v0p = SIPHASHM_0;
297 u64 v1p = SIPHASHM_1;
298 u64 v2p = SIPHASHM_2;
299 u64 v3p = SIPHASHM_3;
301 v0p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[1], salt_bufs[salt_pos].salt_buf[0]);
302 v1p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[3], salt_bufs[salt_pos].salt_buf[2]);
303 v2p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[1], salt_bufs[salt_pos].salt_buf[0]);
304 v3p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[3], salt_bufs[salt_pos].salt_buf[2]);
310 const u32 search[4] =
312 digests_buf[digests_offset].digest_buf[DGST_R0],
313 digests_buf[digests_offset].digest_buf[DGST_R1],
314 digests_buf[digests_offset].digest_buf[DGST_R2],
315 digests_buf[digests_offset].digest_buf[DGST_R3]
322 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
324 const u32 pw_r_len = c_combs[il_pos].pw_len;
326 const u32 pw_len = pw_l_len + pw_r_len;
330 wordr0[0] = c_combs[il_pos].i[0];
331 wordr0[1] = c_combs[il_pos].i[1];
332 wordr0[2] = c_combs[il_pos].i[2];
333 wordr0[3] = c_combs[il_pos].i[3];
337 wordr1[0] = c_combs[il_pos].i[4];
338 wordr1[1] = c_combs[il_pos].i[5];
339 wordr1[2] = c_combs[il_pos].i[6];
340 wordr1[3] = c_combs[il_pos].i[7];
356 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
358 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
363 w[ 0] = wordl0[0] | wordr0[0];
364 w[ 1] = wordl0[1] | wordr0[1];
365 w[ 2] = wordl0[2] | wordr0[2];
366 w[ 3] = wordl0[3] | wordr0[3];
367 w[ 4] = wordl1[0] | wordr1[0];
368 w[ 5] = wordl1[1] | wordr1[1];
369 w[ 6] = wordl1[2] | wordr1[2];
370 w[ 7] = wordl1[3] | wordr1[3];
371 w[ 8] = wordl2[0] | wordr2[0];
372 w[ 9] = wordl2[1] | wordr2[1];
373 w[10] = wordl2[2] | wordr2[2];
374 w[11] = wordl2[3] | wordr2[3];
375 w[12] = wordl3[0] | wordr3[0];
376 w[13] = wordl3[1] | wordr3[1];
377 w[14] = wordl3[2] | wordr3[2];
378 w[15] = wordl3[3] | wordr3[3];
380 u64 *w_ptr = (u64 *) w;
382 w_ptr[pw_len / 8] |= (u64) pw_len << 56;
392 for (i = 0, j = 0; i <= pw_len; i += 8, j += 2)
394 u64x m = hl32_to_64 (w[j + 1], w[j + 0]);
398 SIPROUND (v0, v1, v2, v3);
399 SIPROUND (v0, v1, v2, v3);
406 SIPROUND (v0, v1, v2, v3);
407 SIPROUND (v0, v1, v2, v3);
408 SIPROUND (v0, v1, v2, v3);
409 SIPROUND (v0, v1, v2, v3);
411 const u64x v = v0 ^ v1 ^ v2 ^ v3;
413 const u32x a = l32_from_64 (v);
414 const u32x b = h32_from_64 (v);
421 #include VECT_COMPARE_S
425 extern "C" __global__ void __launch_bounds__ (256, 1) m10100_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 void *esalt_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)
429 extern "C" __global__ void __launch_bounds__ (256, 1) m10100_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 void *esalt_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)