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 __device__ __constant__ gpu_rule_t c_rules[1024];
47 extern "C" __global__ void __launch_bounds__ (256, 1) m00000_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)
53 const u32 lid = threadIdx.x;
59 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
61 if (gid >= gid_max) return;
65 pw_buf0[0] = pws[gid].i[ 0];
66 pw_buf0[1] = pws[gid].i[ 1];
67 pw_buf0[2] = pws[gid].i[ 2];
68 pw_buf0[3] = pws[gid].i[ 3];
72 pw_buf1[0] = pws[gid].i[ 4];
73 pw_buf1[1] = pws[gid].i[ 5];
74 pw_buf1[2] = pws[gid].i[ 6];
75 pw_buf1[3] = pws[gid].i[ 7];
77 const u32 pw_len = pws[gid].pw_len;
83 for (u32 il_pos = 0; il_pos < rules_cnt; il_pos++)
113 const u32 out_len = apply_rules (c_rules[il_pos].cmds, w0, w1, pw_len);
115 append_0x80_2 (w0, w1, out_len);
126 MD5_STEP (MD5_Fo, a, b, c, d, w0[0], MD5C00, MD5S00);
127 MD5_STEP (MD5_Fo, d, a, b, c, w0[1], MD5C01, MD5S01);
128 MD5_STEP (MD5_Fo, c, d, a, b, w0[2], MD5C02, MD5S02);
129 MD5_STEP (MD5_Fo, b, c, d, a, w0[3], MD5C03, MD5S03);
130 MD5_STEP (MD5_Fo, a, b, c, d, w1[0], MD5C04, MD5S00);
131 MD5_STEP (MD5_Fo, d, a, b, c, w1[1], MD5C05, MD5S01);
132 MD5_STEP (MD5_Fo, c, d, a, b, w1[2], MD5C06, MD5S02);
133 MD5_STEP (MD5_Fo, b, c, d, a, w1[3], MD5C07, MD5S03);
134 MD5_STEP (MD5_Fo, a, b, c, d, w2[0], MD5C08, MD5S00);
135 MD5_STEP (MD5_Fo, d, a, b, c, w2[1], MD5C09, MD5S01);
136 MD5_STEP (MD5_Fo, c, d, a, b, w2[2], MD5C0a, MD5S02);
137 MD5_STEP (MD5_Fo, b, c, d, a, w2[3], MD5C0b, MD5S03);
138 MD5_STEP (MD5_Fo, a, b, c, d, w3[0], MD5C0c, MD5S00);
139 MD5_STEP (MD5_Fo, d, a, b, c, w3[1], MD5C0d, MD5S01);
140 MD5_STEP (MD5_Fo, c, d, a, b, w3[2], MD5C0e, MD5S02);
141 MD5_STEP (MD5_Fo, b, c, d, a, w3[3], MD5C0f, MD5S03);
143 MD5_STEP (MD5_Go, a, b, c, d, w0[1], MD5C10, MD5S10);
144 MD5_STEP (MD5_Go, d, a, b, c, w1[2], MD5C11, MD5S11);
145 MD5_STEP (MD5_Go, c, d, a, b, w2[3], MD5C12, MD5S12);
146 MD5_STEP (MD5_Go, b, c, d, a, w0[0], MD5C13, MD5S13);
147 MD5_STEP (MD5_Go, a, b, c, d, w1[1], MD5C14, MD5S10);
148 MD5_STEP (MD5_Go, d, a, b, c, w2[2], MD5C15, MD5S11);
149 MD5_STEP (MD5_Go, c, d, a, b, w3[3], MD5C16, MD5S12);
150 MD5_STEP (MD5_Go, b, c, d, a, w1[0], MD5C17, MD5S13);
151 MD5_STEP (MD5_Go, a, b, c, d, w2[1], MD5C18, MD5S10);
152 MD5_STEP (MD5_Go, d, a, b, c, w3[2], MD5C19, MD5S11);
153 MD5_STEP (MD5_Go, c, d, a, b, w0[3], MD5C1a, MD5S12);
154 MD5_STEP (MD5_Go, b, c, d, a, w2[0], MD5C1b, MD5S13);
155 MD5_STEP (MD5_Go, a, b, c, d, w3[1], MD5C1c, MD5S10);
156 MD5_STEP (MD5_Go, d, a, b, c, w0[2], MD5C1d, MD5S11);
157 MD5_STEP (MD5_Go, c, d, a, b, w1[3], MD5C1e, MD5S12);
158 MD5_STEP (MD5_Go, b, c, d, a, w3[0], MD5C1f, MD5S13);
160 MD5_STEP (MD5_H1, a, b, c, d, w1[1], MD5C20, MD5S20);
161 MD5_STEP (MD5_H2, d, a, b, c, w2[0], MD5C21, MD5S21);
162 MD5_STEP (MD5_H1, c, d, a, b, w2[3], MD5C22, MD5S22);
163 MD5_STEP (MD5_H2, b, c, d, a, w3[2], MD5C23, MD5S23);
164 MD5_STEP (MD5_H1, a, b, c, d, w0[1], MD5C24, MD5S20);
165 MD5_STEP (MD5_H2, d, a, b, c, w1[0], MD5C25, MD5S21);
166 MD5_STEP (MD5_H1, c, d, a, b, w1[3], MD5C26, MD5S22);
167 MD5_STEP (MD5_H2, b, c, d, a, w2[2], MD5C27, MD5S23);
168 MD5_STEP (MD5_H1, a, b, c, d, w3[1], MD5C28, MD5S20);
169 MD5_STEP (MD5_H2, d, a, b, c, w0[0], MD5C29, MD5S21);
170 MD5_STEP (MD5_H1, c, d, a, b, w0[3], MD5C2a, MD5S22);
171 MD5_STEP (MD5_H2, b, c, d, a, w1[2], MD5C2b, MD5S23);
172 MD5_STEP (MD5_H1, a, b, c, d, w2[1], MD5C2c, MD5S20);
173 MD5_STEP (MD5_H2, d, a, b, c, w3[0], MD5C2d, MD5S21);
174 MD5_STEP (MD5_H1, c, d, a, b, w3[3], MD5C2e, MD5S22);
175 MD5_STEP (MD5_H2, b, c, d, a, w0[2], MD5C2f, MD5S23);
177 MD5_STEP (MD5_I , a, b, c, d, w0[0], MD5C30, MD5S30);
178 MD5_STEP (MD5_I , d, a, b, c, w1[3], MD5C31, MD5S31);
179 MD5_STEP (MD5_I , c, d, a, b, w3[2], MD5C32, MD5S32);
180 MD5_STEP (MD5_I , b, c, d, a, w1[1], MD5C33, MD5S33);
181 MD5_STEP (MD5_I , a, b, c, d, w3[0], MD5C34, MD5S30);
182 MD5_STEP (MD5_I , d, a, b, c, w0[3], MD5C35, MD5S31);
183 MD5_STEP (MD5_I , c, d, a, b, w2[2], MD5C36, MD5S32);
184 MD5_STEP (MD5_I , b, c, d, a, w0[1], MD5C37, MD5S33);
185 MD5_STEP (MD5_I , a, b, c, d, w2[0], MD5C38, MD5S30);
186 MD5_STEP (MD5_I , d, a, b, c, w3[3], MD5C39, MD5S31);
187 MD5_STEP (MD5_I , c, d, a, b, w1[2], MD5C3a, MD5S32);
188 MD5_STEP (MD5_I , b, c, d, a, w3[1], MD5C3b, MD5S33);
189 MD5_STEP (MD5_I , a, b, c, d, w1[0], MD5C3c, MD5S30);
190 MD5_STEP (MD5_I , d, a, b, c, w2[3], MD5C3d, MD5S31);
191 MD5_STEP (MD5_I , c, d, a, b, w0[2], MD5C3e, MD5S32);
192 MD5_STEP (MD5_I , b, c, d, a, w2[1], MD5C3f, MD5S33);
200 #include VECT_COMPARE_M
204 extern "C" __global__ void __launch_bounds__ (256, 1) m00000_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)
208 extern "C" __global__ void __launch_bounds__ (256, 1) m00000_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)
212 extern "C" __global__ void __launch_bounds__ (256, 1) m00000_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)
218 const u32 lid = threadIdx.x;
224 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
226 if (gid >= gid_max) return;
230 pw_buf0[0] = pws[gid].i[ 0];
231 pw_buf0[1] = pws[gid].i[ 1];
232 pw_buf0[2] = pws[gid].i[ 2];
233 pw_buf0[3] = pws[gid].i[ 3];
237 pw_buf1[0] = pws[gid].i[ 4];
238 pw_buf1[1] = pws[gid].i[ 5];
239 pw_buf1[2] = pws[gid].i[ 6];
240 pw_buf1[3] = pws[gid].i[ 7];
242 const u32 pw_len = pws[gid].pw_len;
248 const u32 search[4] =
250 digests_buf[digests_offset].digest_buf[DGST_R0],
251 digests_buf[digests_offset].digest_buf[DGST_R1],
252 digests_buf[digests_offset].digest_buf[DGST_R2],
253 digests_buf[digests_offset].digest_buf[DGST_R3]
260 for (u32 il_pos = 0; il_pos < rules_cnt; il_pos++)
290 const u32 out_len = apply_rules (c_rules[il_pos].cmds, w0, w1, pw_len);
292 append_0x80_2 (w0, w1, out_len);
303 MD5_STEP (MD5_Fo, a, b, c, d, w0[0], MD5C00, MD5S00);
304 MD5_STEP (MD5_Fo, d, a, b, c, w0[1], MD5C01, MD5S01);
305 MD5_STEP (MD5_Fo, c, d, a, b, w0[2], MD5C02, MD5S02);
306 MD5_STEP (MD5_Fo, b, c, d, a, w0[3], MD5C03, MD5S03);
307 MD5_STEP (MD5_Fo, a, b, c, d, w1[0], MD5C04, MD5S00);
308 MD5_STEP (MD5_Fo, d, a, b, c, w1[1], MD5C05, MD5S01);
309 MD5_STEP (MD5_Fo, c, d, a, b, w1[2], MD5C06, MD5S02);
310 MD5_STEP (MD5_Fo, b, c, d, a, w1[3], MD5C07, MD5S03);
311 MD5_STEP (MD5_Fo, a, b, c, d, w2[0], MD5C08, MD5S00);
312 MD5_STEP (MD5_Fo, d, a, b, c, w2[1], MD5C09, MD5S01);
313 MD5_STEP (MD5_Fo, c, d, a, b, w2[2], MD5C0a, MD5S02);
314 MD5_STEP (MD5_Fo, b, c, d, a, w2[3], MD5C0b, MD5S03);
315 MD5_STEP (MD5_Fo, a, b, c, d, w3[0], MD5C0c, MD5S00);
316 MD5_STEP (MD5_Fo, d, a, b, c, w3[1], MD5C0d, MD5S01);
317 MD5_STEP (MD5_Fo, c, d, a, b, w3[2], MD5C0e, MD5S02);
318 MD5_STEP (MD5_Fo, b, c, d, a, w3[3], MD5C0f, MD5S03);
320 MD5_STEP (MD5_Go, a, b, c, d, w0[1], MD5C10, MD5S10);
321 MD5_STEP (MD5_Go, d, a, b, c, w1[2], MD5C11, MD5S11);
322 MD5_STEP (MD5_Go, c, d, a, b, w2[3], MD5C12, MD5S12);
323 MD5_STEP (MD5_Go, b, c, d, a, w0[0], MD5C13, MD5S13);
324 MD5_STEP (MD5_Go, a, b, c, d, w1[1], MD5C14, MD5S10);
325 MD5_STEP (MD5_Go, d, a, b, c, w2[2], MD5C15, MD5S11);
326 MD5_STEP (MD5_Go, c, d, a, b, w3[3], MD5C16, MD5S12);
327 MD5_STEP (MD5_Go, b, c, d, a, w1[0], MD5C17, MD5S13);
328 MD5_STEP (MD5_Go, a, b, c, d, w2[1], MD5C18, MD5S10);
329 MD5_STEP (MD5_Go, d, a, b, c, w3[2], MD5C19, MD5S11);
330 MD5_STEP (MD5_Go, c, d, a, b, w0[3], MD5C1a, MD5S12);
331 MD5_STEP (MD5_Go, b, c, d, a, w2[0], MD5C1b, MD5S13);
332 MD5_STEP (MD5_Go, a, b, c, d, w3[1], MD5C1c, MD5S10);
333 MD5_STEP (MD5_Go, d, a, b, c, w0[2], MD5C1d, MD5S11);
334 MD5_STEP (MD5_Go, c, d, a, b, w1[3], MD5C1e, MD5S12);
335 MD5_STEP (MD5_Go, b, c, d, a, w3[0], MD5C1f, MD5S13);
337 MD5_STEP (MD5_H1, a, b, c, d, w1[1], MD5C20, MD5S20);
338 MD5_STEP (MD5_H2, d, a, b, c, w2[0], MD5C21, MD5S21);
339 MD5_STEP (MD5_H1, c, d, a, b, w2[3], MD5C22, MD5S22);
340 MD5_STEP (MD5_H2, b, c, d, a, w3[2], MD5C23, MD5S23);
341 MD5_STEP (MD5_H1, a, b, c, d, w0[1], MD5C24, MD5S20);
342 MD5_STEP (MD5_H2, d, a, b, c, w1[0], MD5C25, MD5S21);
343 MD5_STEP (MD5_H1, c, d, a, b, w1[3], MD5C26, MD5S22);
344 MD5_STEP (MD5_H2, b, c, d, a, w2[2], MD5C27, MD5S23);
345 MD5_STEP (MD5_H1, a, b, c, d, w3[1], MD5C28, MD5S20);
346 MD5_STEP (MD5_H2, d, a, b, c, w0[0], MD5C29, MD5S21);
347 MD5_STEP (MD5_H1, c, d, a, b, w0[3], MD5C2a, MD5S22);
348 MD5_STEP (MD5_H2, b, c, d, a, w1[2], MD5C2b, MD5S23);
349 MD5_STEP (MD5_H1, a, b, c, d, w2[1], MD5C2c, MD5S20);
350 MD5_STEP (MD5_H2, d, a, b, c, w3[0], MD5C2d, MD5S21);
351 MD5_STEP (MD5_H1, c, d, a, b, w3[3], MD5C2e, MD5S22);
352 MD5_STEP (MD5_H2, b, c, d, a, w0[2], MD5C2f, MD5S23);
354 MD5_STEP (MD5_I , a, b, c, d, w0[0], MD5C30, MD5S30);
355 MD5_STEP (MD5_I , d, a, b, c, w1[3], MD5C31, MD5S31);
356 MD5_STEP (MD5_I , c, d, a, b, w3[2], MD5C32, MD5S32);
357 MD5_STEP (MD5_I , b, c, d, a, w1[1], MD5C33, MD5S33);
358 MD5_STEP (MD5_I , a, b, c, d, w3[0], MD5C34, MD5S30);
359 MD5_STEP (MD5_I , d, a, b, c, w0[3], MD5C35, MD5S31);
360 MD5_STEP (MD5_I , c, d, a, b, w2[2], MD5C36, MD5S32);
361 MD5_STEP (MD5_I , b, c, d, a, w0[1], MD5C37, MD5S33);
362 MD5_STEP (MD5_I , a, b, c, d, w2[0], MD5C38, MD5S30);
363 MD5_STEP (MD5_I , d, a, b, c, w3[3], MD5C39, MD5S31);
364 MD5_STEP (MD5_I , c, d, a, b, w1[2], MD5C3a, MD5S32);
365 MD5_STEP (MD5_I , b, c, d, a, w3[1], MD5C3b, MD5S33);
366 MD5_STEP (MD5_I , a, b, c, d, w1[0], MD5C3c, MD5S30);
368 bool q_cond = (search[0] != a);
370 if (q_cond) continue;
372 MD5_STEP (MD5_I , d, a, b, c, w2[3], MD5C3d, MD5S31);
373 MD5_STEP (MD5_I , c, d, a, b, w0[2], MD5C3e, MD5S32);
374 MD5_STEP (MD5_I , b, c, d, a, w2[1], MD5C3f, MD5S33);
382 #include VECT_COMPARE_S
386 extern "C" __global__ void __launch_bounds__ (256, 1) m00000_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)
390 extern "C" __global__ void __launch_bounds__ (256, 1) m00000_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)