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"
27 #include "include/rp_gpu.h"
31 #define VECT_COMPARE_S "check_single_vect1_comp4.c"
32 #define VECT_COMPARE_M "check_multi_vect1_comp4.c"
36 #define VECT_COMPARE_S "check_single_vect2_comp4.c"
37 #define VECT_COMPARE_M "check_multi_vect2_comp4.c"
41 #define VECT_COMPARE_S "check_single_vect4_comp4.c"
42 #define VECT_COMPARE_M "check_multi_vect4_comp4.c"
45 #define SIPROUND(v0,v1,v2,v3) \
47 (v1) = rotl64 ((v1), 13); \
49 (v0) = rotl64 ((v0), 32); \
51 (v3) = rotl64 ((v3), 16); \
54 (v3) = rotl64 ((v3), 21); \
57 (v1) = rotl64 ((v1), 17); \
59 (v2) = rotl64 ((v2), 32);
61 __device__ __constant__ gpu_rule_t c_rules[1024];
63 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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
69 const u32 lid = threadIdx.x;
75 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
77 if (gid >= gid_max) return;
81 pw_buf0[0] = pws[gid].i[ 0];
82 pw_buf0[1] = pws[gid].i[ 1];
83 pw_buf0[2] = pws[gid].i[ 2];
84 pw_buf0[3] = pws[gid].i[ 3];
88 pw_buf1[0] = pws[gid].i[ 4];
89 pw_buf1[1] = pws[gid].i[ 5];
90 pw_buf1[2] = pws[gid].i[ 6];
91 pw_buf1[3] = pws[gid].i[ 7];
93 const u32 pw_len = pws[gid].pw_len;
100 u64 v1p = SIPHASHM_1;
101 u64 v2p = SIPHASHM_2;
102 u64 v3p = SIPHASHM_3;
104 v0p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[1], salt_bufs[salt_pos].salt_buf[0]);
105 v1p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[3], salt_bufs[salt_pos].salt_buf[2]);
106 v2p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[1], salt_bufs[salt_pos].salt_buf[0]);
107 v3p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[3], salt_bufs[salt_pos].salt_buf[2]);
113 for (u32 il_pos = 0; il_pos < rules_cnt; il_pos++)
134 const u32x out_len = apply_rules (c_rules[il_pos].cmds, &w[0], &w[4], pw_len);
136 u64 *w_ptr = (u64 *) w;
138 w_ptr[out_len / 8] |= (u64) out_len << 56;
148 for (i = 0, j = 0; i <= pw_len; i += 8, j += 2)
150 u64x m = hl32_to_64 (w[j + 1], w[j + 0]);
154 SIPROUND (v0, v1, v2, v3);
155 SIPROUND (v0, v1, v2, v3);
162 SIPROUND (v0, v1, v2, v3);
163 SIPROUND (v0, v1, v2, v3);
164 SIPROUND (v0, v1, v2, v3);
165 SIPROUND (v0, v1, v2, v3);
167 const u64x v = v0 ^ v1 ^ v2 ^ v3;
169 const u32x a = l32_from_64 (v);
170 const u32x b = h32_from_64 (v);
177 #include VECT_COMPARE_M
181 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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
185 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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
189 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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
195 const u32 lid = threadIdx.x;
201 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
203 if (gid >= gid_max) return;
207 pw_buf0[0] = pws[gid].i[ 0];
208 pw_buf0[1] = pws[gid].i[ 1];
209 pw_buf0[2] = pws[gid].i[ 2];
210 pw_buf0[3] = pws[gid].i[ 3];
214 pw_buf1[0] = pws[gid].i[ 4];
215 pw_buf1[1] = pws[gid].i[ 5];
216 pw_buf1[2] = pws[gid].i[ 6];
217 pw_buf1[3] = pws[gid].i[ 7];
219 const u32 pw_len = pws[gid].pw_len;
225 u64 v0p = SIPHASHM_0;
226 u64 v1p = SIPHASHM_1;
227 u64 v2p = SIPHASHM_2;
228 u64 v3p = SIPHASHM_3;
230 v0p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[1], salt_bufs[salt_pos].salt_buf[0]);
231 v1p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[3], salt_bufs[salt_pos].salt_buf[2]);
232 v2p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[1], salt_bufs[salt_pos].salt_buf[0]);
233 v3p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[3], salt_bufs[salt_pos].salt_buf[2]);
239 const u32 search[4] =
241 digests_buf[digests_offset].digest_buf[DGST_R0],
242 digests_buf[digests_offset].digest_buf[DGST_R1],
243 digests_buf[digests_offset].digest_buf[DGST_R2],
244 digests_buf[digests_offset].digest_buf[DGST_R3]
251 for (u32 il_pos = 0; il_pos < rules_cnt; il_pos++)
272 const u32x out_len = apply_rules (c_rules[il_pos].cmds, &w[0], &w[4], pw_len);
274 u64 *w_ptr = (u64 *) w;
276 w_ptr[out_len / 8] |= (u64) out_len << 56;
286 for (i = 0, j = 0; i <= pw_len; i += 8, j += 2)
288 u64x m = hl32_to_64 (w[j + 1], w[j + 0]);
292 SIPROUND (v0, v1, v2, v3);
293 SIPROUND (v0, v1, v2, v3);
300 SIPROUND (v0, v1, v2, v3);
301 SIPROUND (v0, v1, v2, v3);
302 SIPROUND (v0, v1, v2, v3);
303 SIPROUND (v0, v1, v2, v3);
305 const u64x v = v0 ^ v1 ^ v2 ^ v3;
307 const u32x a = l32_from_64 (v);
308 const u32x b = h32_from_64 (v);
315 #include VECT_COMPARE_S
319 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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
323 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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)