2 * Author......: Jens Steube <jens.steube@gmail.com>
10 #include "include/constants.h"
11 #include "include/kernel_vendor.h"
18 #include "include/kernel_functions.c"
19 #include "OpenCL/types_ocl.c"
20 #include "OpenCL/common.c"
21 #include "include/rp_kernel.h"
22 #include "OpenCL/rp.c"
23 #include "OpenCL/simd.c"
25 __kernel void m01000_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global 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 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
31 const u32 lid = get_local_id (0);
37 const u32 gid = get_global_id (0);
39 if (gid >= gid_max) return;
44 pw_buf0[0] = pws[gid].i[0];
45 pw_buf0[1] = pws[gid].i[1];
46 pw_buf0[2] = pws[gid].i[2];
47 pw_buf0[3] = pws[gid].i[3];
48 pw_buf1[0] = pws[gid].i[4];
49 pw_buf1[1] = pws[gid].i[5];
50 pw_buf1[2] = pws[gid].i[6];
51 pw_buf1[3] = pws[gid].i[7];
53 const u32 pw_len = pws[gid].pw_len;
59 for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
66 const u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
68 append_0x80_2x4_VV (w0, w1, out_len);
70 make_unicode (w1, w2, w3);
71 make_unicode (w0, w0, w1);
73 w3[2] = out_len * 8 * 2;
81 MD4_STEP (MD4_Fo, a, b, c, d, w0[0], MD4C00, MD4S00);
82 MD4_STEP (MD4_Fo, d, a, b, c, w0[1], MD4C00, MD4S01);
83 MD4_STEP (MD4_Fo, c, d, a, b, w0[2], MD4C00, MD4S02);
84 MD4_STEP (MD4_Fo, b, c, d, a, w0[3], MD4C00, MD4S03);
85 MD4_STEP (MD4_Fo, a, b, c, d, w1[0], MD4C00, MD4S00);
86 MD4_STEP (MD4_Fo, d, a, b, c, w1[1], MD4C00, MD4S01);
87 MD4_STEP (MD4_Fo, c, d, a, b, w1[2], MD4C00, MD4S02);
88 MD4_STEP (MD4_Fo, b, c, d, a, w1[3], MD4C00, MD4S03);
89 MD4_STEP (MD4_Fo, a, b, c, d, w2[0], MD4C00, MD4S00);
90 MD4_STEP (MD4_Fo, d, a, b, c, w2[1], MD4C00, MD4S01);
91 MD4_STEP (MD4_Fo, c, d, a, b, w2[2], MD4C00, MD4S02);
92 MD4_STEP (MD4_Fo, b, c, d, a, w2[3], MD4C00, MD4S03);
93 MD4_STEP (MD4_Fo, a, b, c, d, w3[0], MD4C00, MD4S00);
94 MD4_STEP (MD4_Fo, d, a, b, c, w3[1], MD4C00, MD4S01);
95 MD4_STEP (MD4_Fo, c, d, a, b, w3[2], MD4C00, MD4S02);
96 MD4_STEP (MD4_Fo, b, c, d, a, w3[3], MD4C00, MD4S03);
98 MD4_STEP (MD4_Go, a, b, c, d, w0[0], MD4C01, MD4S10);
99 MD4_STEP (MD4_Go, d, a, b, c, w1[0], MD4C01, MD4S11);
100 MD4_STEP (MD4_Go, c, d, a, b, w2[0], MD4C01, MD4S12);
101 MD4_STEP (MD4_Go, b, c, d, a, w3[0], MD4C01, MD4S13);
102 MD4_STEP (MD4_Go, a, b, c, d, w0[1], MD4C01, MD4S10);
103 MD4_STEP (MD4_Go, d, a, b, c, w1[1], MD4C01, MD4S11);
104 MD4_STEP (MD4_Go, c, d, a, b, w2[1], MD4C01, MD4S12);
105 MD4_STEP (MD4_Go, b, c, d, a, w3[1], MD4C01, MD4S13);
106 MD4_STEP (MD4_Go, a, b, c, d, w0[2], MD4C01, MD4S10);
107 MD4_STEP (MD4_Go, d, a, b, c, w1[2], MD4C01, MD4S11);
108 MD4_STEP (MD4_Go, c, d, a, b, w2[2], MD4C01, MD4S12);
109 MD4_STEP (MD4_Go, b, c, d, a, w3[2], MD4C01, MD4S13);
110 MD4_STEP (MD4_Go, a, b, c, d, w0[3], MD4C01, MD4S10);
111 MD4_STEP (MD4_Go, d, a, b, c, w1[3], MD4C01, MD4S11);
112 MD4_STEP (MD4_Go, c, d, a, b, w2[3], MD4C01, MD4S12);
113 MD4_STEP (MD4_Go, b, c, d, a, w3[3], MD4C01, MD4S13);
115 MD4_STEP (MD4_H , a, b, c, d, w0[0], MD4C02, MD4S20);
116 MD4_STEP (MD4_H , d, a, b, c, w2[0], MD4C02, MD4S21);
117 MD4_STEP (MD4_H , c, d, a, b, w1[0], MD4C02, MD4S22);
118 MD4_STEP (MD4_H , b, c, d, a, w3[0], MD4C02, MD4S23);
119 MD4_STEP (MD4_H , a, b, c, d, w0[2], MD4C02, MD4S20);
120 MD4_STEP (MD4_H , d, a, b, c, w2[2], MD4C02, MD4S21);
121 MD4_STEP (MD4_H , c, d, a, b, w1[2], MD4C02, MD4S22);
122 MD4_STEP (MD4_H , b, c, d, a, w3[2], MD4C02, MD4S23);
123 MD4_STEP (MD4_H , a, b, c, d, w0[1], MD4C02, MD4S20);
124 MD4_STEP (MD4_H , d, a, b, c, w2[1], MD4C02, MD4S21);
125 MD4_STEP (MD4_H , c, d, a, b, w1[1], MD4C02, MD4S22);
126 MD4_STEP (MD4_H , b, c, d, a, w3[1], MD4C02, MD4S23);
127 MD4_STEP (MD4_H , a, b, c, d, w0[3], MD4C02, MD4S20);
128 MD4_STEP (MD4_H , d, a, b, c, w2[3], MD4C02, MD4S21);
129 MD4_STEP (MD4_H , c, d, a, b, w1[3], MD4C02, MD4S22);
130 MD4_STEP (MD4_H , b, c, d, a, w3[3], MD4C02, MD4S23);
132 COMPARE_M_SIMD (a, d, c, b);
136 __kernel void m01000_m08 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global 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 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
140 __kernel void m01000_m16 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global 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 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
144 __kernel void m01000_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global 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 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
150 const u32 lid = get_local_id (0);
156 const u32 gid = get_global_id (0);
158 if (gid >= gid_max) return;
163 pw_buf0[0] = pws[gid].i[0];
164 pw_buf0[1] = pws[gid].i[1];
165 pw_buf0[2] = pws[gid].i[2];
166 pw_buf0[3] = pws[gid].i[3];
167 pw_buf1[0] = pws[gid].i[4];
168 pw_buf1[1] = pws[gid].i[5];
169 pw_buf1[2] = pws[gid].i[6];
170 pw_buf1[3] = pws[gid].i[7];
172 const u32 pw_len = pws[gid].pw_len;
178 const u32 search[4] =
180 digests_buf[digests_offset].digest_buf[DGST_R0],
181 digests_buf[digests_offset].digest_buf[DGST_R1],
182 digests_buf[digests_offset].digest_buf[DGST_R2],
183 digests_buf[digests_offset].digest_buf[DGST_R3]
190 for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
197 const u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
199 append_0x80_2x4_VV (w0, w1, out_len);
201 make_unicode (w1, w2, w3);
202 make_unicode (w0, w0, w1);
204 w3[2] = out_len * 8 * 2;
212 MD4_STEP (MD4_Fo, a, b, c, d, w0[0], MD4C00, MD4S00);
213 MD4_STEP (MD4_Fo, d, a, b, c, w0[1], MD4C00, MD4S01);
214 MD4_STEP (MD4_Fo, c, d, a, b, w0[2], MD4C00, MD4S02);
215 MD4_STEP (MD4_Fo, b, c, d, a, w0[3], MD4C00, MD4S03);
216 MD4_STEP (MD4_Fo, a, b, c, d, w1[0], MD4C00, MD4S00);
217 MD4_STEP (MD4_Fo, d, a, b, c, w1[1], MD4C00, MD4S01);
218 MD4_STEP (MD4_Fo, c, d, a, b, w1[2], MD4C00, MD4S02);
219 MD4_STEP (MD4_Fo, b, c, d, a, w1[3], MD4C00, MD4S03);
220 MD4_STEP (MD4_Fo, a, b, c, d, w2[0], MD4C00, MD4S00);
221 MD4_STEP (MD4_Fo, d, a, b, c, w2[1], MD4C00, MD4S01);
222 MD4_STEP (MD4_Fo, c, d, a, b, w2[2], MD4C00, MD4S02);
223 MD4_STEP (MD4_Fo, b, c, d, a, w2[3], MD4C00, MD4S03);
224 MD4_STEP (MD4_Fo, a, b, c, d, w3[0], MD4C00, MD4S00);
225 MD4_STEP (MD4_Fo, d, a, b, c, w3[1], MD4C00, MD4S01);
226 MD4_STEP (MD4_Fo, c, d, a, b, w3[2], MD4C00, MD4S02);
227 MD4_STEP (MD4_Fo, b, c, d, a, w3[3], MD4C00, MD4S03);
229 MD4_STEP (MD4_Go, a, b, c, d, w0[0], MD4C01, MD4S10);
230 MD4_STEP (MD4_Go, d, a, b, c, w1[0], MD4C01, MD4S11);
231 MD4_STEP (MD4_Go, c, d, a, b, w2[0], MD4C01, MD4S12);
232 MD4_STEP (MD4_Go, b, c, d, a, w3[0], MD4C01, MD4S13);
233 MD4_STEP (MD4_Go, a, b, c, d, w0[1], MD4C01, MD4S10);
234 MD4_STEP (MD4_Go, d, a, b, c, w1[1], MD4C01, MD4S11);
235 MD4_STEP (MD4_Go, c, d, a, b, w2[1], MD4C01, MD4S12);
236 MD4_STEP (MD4_Go, b, c, d, a, w3[1], MD4C01, MD4S13);
237 MD4_STEP (MD4_Go, a, b, c, d, w0[2], MD4C01, MD4S10);
238 MD4_STEP (MD4_Go, d, a, b, c, w1[2], MD4C01, MD4S11);
239 MD4_STEP (MD4_Go, c, d, a, b, w2[2], MD4C01, MD4S12);
240 MD4_STEP (MD4_Go, b, c, d, a, w3[2], MD4C01, MD4S13);
241 MD4_STEP (MD4_Go, a, b, c, d, w0[3], MD4C01, MD4S10);
242 MD4_STEP (MD4_Go, d, a, b, c, w1[3], MD4C01, MD4S11);
243 MD4_STEP (MD4_Go, c, d, a, b, w2[3], MD4C01, MD4S12);
244 MD4_STEP (MD4_Go, b, c, d, a, w3[3], MD4C01, MD4S13);
246 MD4_STEP (MD4_H , a, b, c, d, w0[0], MD4C02, MD4S20);
247 MD4_STEP (MD4_H , d, a, b, c, w2[0], MD4C02, MD4S21);
248 MD4_STEP (MD4_H , c, d, a, b, w1[0], MD4C02, MD4S22);
249 MD4_STEP (MD4_H , b, c, d, a, w3[0], MD4C02, MD4S23);
250 MD4_STEP (MD4_H , a, b, c, d, w0[2], MD4C02, MD4S20);
251 MD4_STEP (MD4_H , d, a, b, c, w2[2], MD4C02, MD4S21);
252 MD4_STEP (MD4_H , c, d, a, b, w1[2], MD4C02, MD4S22);
253 MD4_STEP (MD4_H , b, c, d, a, w3[2], MD4C02, MD4S23);
254 MD4_STEP (MD4_H , a, b, c, d, w0[1], MD4C02, MD4S20);
255 MD4_STEP (MD4_H , d, a, b, c, w2[1], MD4C02, MD4S21);
256 MD4_STEP (MD4_H , c, d, a, b, w1[1], MD4C02, MD4S22);
257 MD4_STEP (MD4_H , b, c, d, a, w3[1], MD4C02, MD4S23);
258 MD4_STEP (MD4_H , a, b, c, d, w0[3], MD4C02, MD4S20);
260 if (MATCHES_NONE_VS (a, search[0])) continue;
262 MD4_STEP (MD4_H , d, a, b, c, w2[3], MD4C02, MD4S21);
263 MD4_STEP (MD4_H , c, d, a, b, w1[3], MD4C02, MD4S22);
264 MD4_STEP (MD4_H , b, c, d, a, w3[3], MD4C02, MD4S23);
266 COMPARE_S_SIMD (a, d, c, b);
270 __kernel void m01000_s08 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global 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 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
274 __kernel void m01000_s16 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global 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 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)