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 "OpenCL/simd.c"
23 __constant u32 padding[8] =
35 static void md5_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4])
59 MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
60 MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
61 MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02);
62 MD5_STEP (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03);
63 MD5_STEP (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00);
64 MD5_STEP (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01);
65 MD5_STEP (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02);
66 MD5_STEP (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03);
67 MD5_STEP (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00);
68 MD5_STEP (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01);
69 MD5_STEP (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02);
70 MD5_STEP (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03);
71 MD5_STEP (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00);
72 MD5_STEP (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01);
73 MD5_STEP (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02);
74 MD5_STEP (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03);
76 MD5_STEP (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10);
77 MD5_STEP (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11);
78 MD5_STEP (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12);
79 MD5_STEP (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13);
80 MD5_STEP (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10);
81 MD5_STEP (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11);
82 MD5_STEP (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12);
83 MD5_STEP (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13);
84 MD5_STEP (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10);
85 MD5_STEP (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11);
86 MD5_STEP (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12);
87 MD5_STEP (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13);
88 MD5_STEP (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10);
89 MD5_STEP (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11);
90 MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12);
91 MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13);
93 MD5_STEP (MD5_H , a, b, c, d, w5_t, MD5C20, MD5S20);
94 MD5_STEP (MD5_H , d, a, b, c, w8_t, MD5C21, MD5S21);
95 MD5_STEP (MD5_H , c, d, a, b, wb_t, MD5C22, MD5S22);
96 MD5_STEP (MD5_H , b, c, d, a, we_t, MD5C23, MD5S23);
97 MD5_STEP (MD5_H , a, b, c, d, w1_t, MD5C24, MD5S20);
98 MD5_STEP (MD5_H , d, a, b, c, w4_t, MD5C25, MD5S21);
99 MD5_STEP (MD5_H , c, d, a, b, w7_t, MD5C26, MD5S22);
100 MD5_STEP (MD5_H , b, c, d, a, wa_t, MD5C27, MD5S23);
101 MD5_STEP (MD5_H , a, b, c, d, wd_t, MD5C28, MD5S20);
102 MD5_STEP (MD5_H , d, a, b, c, w0_t, MD5C29, MD5S21);
103 MD5_STEP (MD5_H , c, d, a, b, w3_t, MD5C2a, MD5S22);
104 MD5_STEP (MD5_H , b, c, d, a, w6_t, MD5C2b, MD5S23);
105 MD5_STEP (MD5_H , a, b, c, d, w9_t, MD5C2c, MD5S20);
106 MD5_STEP (MD5_H , d, a, b, c, wc_t, MD5C2d, MD5S21);
107 MD5_STEP (MD5_H , c, d, a, b, wf_t, MD5C2e, MD5S22);
108 MD5_STEP (MD5_H , b, c, d, a, w2_t, MD5C2f, MD5S23);
110 MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30);
111 MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31);
112 MD5_STEP (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32);
113 MD5_STEP (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33);
114 MD5_STEP (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30);
115 MD5_STEP (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31);
116 MD5_STEP (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32);
117 MD5_STEP (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33);
118 MD5_STEP (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30);
119 MD5_STEP (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31);
120 MD5_STEP (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32);
121 MD5_STEP (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33);
122 MD5_STEP (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30);
123 MD5_STEP (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31);
124 MD5_STEP (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32);
125 MD5_STEP (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33);
133 __kernel void m10420_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 pdf_t *pdf_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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
139 const u32 lid = get_local_id (0);
145 const u32 gid = get_global_id (0);
147 if (gid >= gid_max) return;
152 pws0[0] = pws[gid].i[0];
153 pws0[1] = pws[gid].i[1];
154 pws0[2] = pws[gid].i[2];
155 pws0[3] = pws[gid].i[3];
156 pws1[0] = pws[gid].i[4];
157 pws1[1] = pws[gid].i[5];
158 pws1[2] = pws[gid].i[6];
159 pws1[3] = pws[gid].i[7];
161 const u32 pw_l_len = pws[gid].pw_len;
169 o_buf[0] = pdf_bufs[salt_pos].o_buf[0];
170 o_buf[1] = pdf_bufs[salt_pos].o_buf[1];
171 o_buf[2] = pdf_bufs[salt_pos].o_buf[2];
172 o_buf[3] = pdf_bufs[salt_pos].o_buf[3];
173 o_buf[4] = pdf_bufs[salt_pos].o_buf[4];
174 o_buf[5] = pdf_bufs[salt_pos].o_buf[5];
175 o_buf[6] = pdf_bufs[salt_pos].o_buf[6];
176 o_buf[7] = pdf_bufs[salt_pos].o_buf[7];
178 u32 P = pdf_bufs[salt_pos].P;
182 id_buf[0] = pdf_bufs[salt_pos].id_buf[0];
183 id_buf[1] = pdf_bufs[salt_pos].id_buf[1];
184 id_buf[2] = pdf_bufs[salt_pos].id_buf[2];
185 id_buf[3] = pdf_bufs[salt_pos].id_buf[3];
191 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos += VECT_SIZE)
193 const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos);
195 const u32x pw_len = pw_l_len + pw_r_len;
197 u32x wordr0[4] = { 0 };
198 u32x wordr1[4] = { 0 };
199 u32x wordr2[4] = { 0 };
200 u32x wordr3[4] = { 0 };
202 wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
203 wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
204 wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
205 wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
206 wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
207 wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
208 wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
209 wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);
211 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
213 switch_buffer_by_offset_le (wordr0, wordr1, wordr2, wordr3, pw_l_len);
218 w0[0] = wordl0[0] | wordr0[0];
219 w0[1] = wordl0[1] | wordr0[1];
220 w0[2] = wordl0[2] | wordr0[2];
221 w0[3] = wordl0[3] | wordr0[3];
225 w1[0] = wordl1[0] | wordr1[0];
226 w1[1] = wordl1[1] | wordr1[1];
227 w1[2] = wordl1[2] | wordr1[2];
228 w1[3] = wordl1[3] | wordr1[3];
232 w2[0] = wordl2[0] | wordr2[0];
233 w2[1] = wordl2[1] | wordr2[1];
234 w2[2] = wordl2[2] | wordr2[2];
235 w2[3] = wordl2[3] | wordr2[3];
239 w3[0] = wordl3[0] | wordr3[0];
240 w3[1] = wordl3[1] | wordr3[1];
241 w3[2] = wordl3[2] | wordr3[2];
242 w3[3] = wordl3[3] | wordr3[3];
249 // max length supported by pdf11 is 32
251 w0_t[0] = padding[0];
252 w0_t[1] = padding[1];
253 w0_t[2] = padding[2];
254 w0_t[3] = padding[3];
255 w1_t[0] = padding[4];
256 w1_t[1] = padding[5];
257 w1_t[2] = padding[6];
258 w1_t[3] = padding[7];
268 switch_buffer_by_offset_le (w0_t, w1_t, w2_t, w3_t, pw_len);
271 // truncate at 32 is wanted, not a bug!
298 md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
317 md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
320 u32x b = digest[1] & 0xff;
331 __kernel void m10420_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 pdf_t *pdf_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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
335 __kernel void m10420_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 pdf_t *pdf_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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
339 __kernel void m10420_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 pdf_t *pdf_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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
345 const u32 lid = get_local_id (0);
351 const u32 gid = get_global_id (0);
353 if (gid >= gid_max) return;
358 pws0[0] = pws[gid].i[0];
359 pws0[1] = pws[gid].i[1];
360 pws0[2] = pws[gid].i[2];
361 pws0[3] = pws[gid].i[3];
362 pws1[0] = pws[gid].i[4];
363 pws1[1] = pws[gid].i[5];
364 pws1[2] = pws[gid].i[6];
365 pws1[3] = pws[gid].i[7];
367 const u32 pw_l_len = pws[gid].pw_len;
373 const u32 search[4] =
375 digests_buf[digests_offset].digest_buf[DGST_R0],
376 digests_buf[digests_offset].digest_buf[DGST_R1],
377 digests_buf[digests_offset].digest_buf[DGST_R2],
378 digests_buf[digests_offset].digest_buf[DGST_R3]
387 o_buf[0] = pdf_bufs[salt_pos].o_buf[0];
388 o_buf[1] = pdf_bufs[salt_pos].o_buf[1];
389 o_buf[2] = pdf_bufs[salt_pos].o_buf[2];
390 o_buf[3] = pdf_bufs[salt_pos].o_buf[3];
391 o_buf[4] = pdf_bufs[salt_pos].o_buf[4];
392 o_buf[5] = pdf_bufs[salt_pos].o_buf[5];
393 o_buf[6] = pdf_bufs[salt_pos].o_buf[6];
394 o_buf[7] = pdf_bufs[salt_pos].o_buf[7];
396 u32 P = pdf_bufs[salt_pos].P;
400 id_buf[0] = pdf_bufs[salt_pos].id_buf[0];
401 id_buf[1] = pdf_bufs[salt_pos].id_buf[1];
402 id_buf[2] = pdf_bufs[salt_pos].id_buf[2];
403 id_buf[3] = pdf_bufs[salt_pos].id_buf[3];
409 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos += VECT_SIZE)
411 const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos);
413 const u32x pw_len = pw_l_len + pw_r_len;
415 u32x wordr0[4] = { 0 };
416 u32x wordr1[4] = { 0 };
417 u32x wordr2[4] = { 0 };
418 u32x wordr3[4] = { 0 };
420 wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
421 wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
422 wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
423 wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
424 wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
425 wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
426 wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
427 wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);
429 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
431 switch_buffer_by_offset_le (wordr0, wordr1, wordr2, wordr3, pw_l_len);
436 w0[0] = wordl0[0] | wordr0[0];
437 w0[1] = wordl0[1] | wordr0[1];
438 w0[2] = wordl0[2] | wordr0[2];
439 w0[3] = wordl0[3] | wordr0[3];
443 w1[0] = wordl1[0] | wordr1[0];
444 w1[1] = wordl1[1] | wordr1[1];
445 w1[2] = wordl1[2] | wordr1[2];
446 w1[3] = wordl1[3] | wordr1[3];
450 w2[0] = wordl2[0] | wordr2[0];
451 w2[1] = wordl2[1] | wordr2[1];
452 w2[2] = wordl2[2] | wordr2[2];
453 w2[3] = wordl2[3] | wordr2[3];
457 w3[0] = wordl3[0] | wordr3[0];
458 w3[1] = wordl3[1] | wordr3[1];
459 w3[2] = wordl3[2] | wordr3[2];
460 w3[3] = wordl3[3] | wordr3[3];
467 // max length supported by pdf11 is 32
469 w0_t[0] = padding[0];
470 w0_t[1] = padding[1];
471 w0_t[2] = padding[2];
472 w0_t[3] = padding[3];
473 w1_t[0] = padding[4];
474 w1_t[1] = padding[5];
475 w1_t[2] = padding[6];
476 w1_t[3] = padding[7];
486 switch_buffer_by_offset_le (w0_t, w1_t, w2_t, w3_t, pw_len);
489 // truncate at 32 is wanted, not a bug!
516 md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
535 md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
538 u32x b = digest[1] & 0xff;
549 __kernel void m10420_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 pdf_t *pdf_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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
553 __kernel void m10420_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 pdf_t *pdf_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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)