2 * Author......: Jens Steube <jens.steube@gmail.com>
8 #include "include/constants.h"
9 #include "include/kernel_vendor.h"
28 #include "include/kernel_functions.c"
29 #include "types_amd.c"
30 #include "common_amd.c"
33 #define VECT_COMPARE_S "check_single_vect1_comp4.c"
34 #define VECT_COMPARE_M "check_multi_vect1_comp4.c"
38 #define VECT_COMPARE_S "check_single_vect2_comp4.c"
39 #define VECT_COMPARE_M "check_multi_vect2_comp4.c"
43 #define VECT_COMPARE_S "check_single_vect4_comp4.c"
44 #define VECT_COMPARE_M "check_multi_vect4_comp4.c"
47 __constant u32 padding[8] =
59 static void md5_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4])
83 MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
84 MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
85 MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02);
86 MD5_STEP (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03);
87 MD5_STEP (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00);
88 MD5_STEP (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01);
89 MD5_STEP (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02);
90 MD5_STEP (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03);
91 MD5_STEP (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00);
92 MD5_STEP (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01);
93 MD5_STEP (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02);
94 MD5_STEP (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03);
95 MD5_STEP (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00);
96 MD5_STEP (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01);
97 MD5_STEP (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02);
98 MD5_STEP (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03);
100 MD5_STEP (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10);
101 MD5_STEP (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11);
102 MD5_STEP (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12);
103 MD5_STEP (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13);
104 MD5_STEP (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10);
105 MD5_STEP (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11);
106 MD5_STEP (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12);
107 MD5_STEP (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13);
108 MD5_STEP (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10);
109 MD5_STEP (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11);
110 MD5_STEP (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12);
111 MD5_STEP (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13);
112 MD5_STEP (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10);
113 MD5_STEP (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11);
114 MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12);
115 MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13);
117 MD5_STEP (MD5_H , a, b, c, d, w5_t, MD5C20, MD5S20);
118 MD5_STEP (MD5_H , d, a, b, c, w8_t, MD5C21, MD5S21);
119 MD5_STEP (MD5_H , c, d, a, b, wb_t, MD5C22, MD5S22);
120 MD5_STEP (MD5_H , b, c, d, a, we_t, MD5C23, MD5S23);
121 MD5_STEP (MD5_H , a, b, c, d, w1_t, MD5C24, MD5S20);
122 MD5_STEP (MD5_H , d, a, b, c, w4_t, MD5C25, MD5S21);
123 MD5_STEP (MD5_H , c, d, a, b, w7_t, MD5C26, MD5S22);
124 MD5_STEP (MD5_H , b, c, d, a, wa_t, MD5C27, MD5S23);
125 MD5_STEP (MD5_H , a, b, c, d, wd_t, MD5C28, MD5S20);
126 MD5_STEP (MD5_H , d, a, b, c, w0_t, MD5C29, MD5S21);
127 MD5_STEP (MD5_H , c, d, a, b, w3_t, MD5C2a, MD5S22);
128 MD5_STEP (MD5_H , b, c, d, a, w6_t, MD5C2b, MD5S23);
129 MD5_STEP (MD5_H , a, b, c, d, w9_t, MD5C2c, MD5S20);
130 MD5_STEP (MD5_H , d, a, b, c, wc_t, MD5C2d, MD5S21);
131 MD5_STEP (MD5_H , c, d, a, b, wf_t, MD5C2e, MD5S22);
132 MD5_STEP (MD5_H , b, c, d, a, w2_t, MD5C2f, MD5S23);
134 MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30);
135 MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31);
136 MD5_STEP (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32);
137 MD5_STEP (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33);
138 MD5_STEP (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30);
139 MD5_STEP (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31);
140 MD5_STEP (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32);
141 MD5_STEP (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33);
142 MD5_STEP (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30);
143 MD5_STEP (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31);
144 MD5_STEP (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32);
145 MD5_STEP (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33);
146 MD5_STEP (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30);
147 MD5_STEP (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31);
148 MD5_STEP (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32);
149 MD5_STEP (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33);
157 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m10420_m04 (__global pw_t *pws, __global gpu_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)
163 const u32 lid = get_local_id (0);
169 const u32 gid = get_global_id (0);
171 if (gid >= gid_max) return;
175 wordl0[0] = pws[gid].i[ 0];
176 wordl0[1] = pws[gid].i[ 1];
177 wordl0[2] = pws[gid].i[ 2];
178 wordl0[3] = pws[gid].i[ 3];
182 wordl1[0] = pws[gid].i[ 4];
183 wordl1[1] = pws[gid].i[ 5];
184 wordl1[2] = pws[gid].i[ 6];
185 wordl1[3] = pws[gid].i[ 7];
201 const u32 pw_l_len = pws[gid].pw_len;
203 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
205 append_0x80_2 (wordl0, wordl1, pw_l_len);
207 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
216 o_buf[0] = pdf_bufs[salt_pos].o_buf[0];
217 o_buf[1] = pdf_bufs[salt_pos].o_buf[1];
218 o_buf[2] = pdf_bufs[salt_pos].o_buf[2];
219 o_buf[3] = pdf_bufs[salt_pos].o_buf[3];
220 o_buf[4] = pdf_bufs[salt_pos].o_buf[4];
221 o_buf[5] = pdf_bufs[salt_pos].o_buf[5];
222 o_buf[6] = pdf_bufs[salt_pos].o_buf[6];
223 o_buf[7] = pdf_bufs[salt_pos].o_buf[7];
225 u32 P = pdf_bufs[salt_pos].P;
229 id_buf[0] = pdf_bufs[salt_pos].id_buf[0];
230 id_buf[1] = pdf_bufs[salt_pos].id_buf[1];
231 id_buf[2] = pdf_bufs[salt_pos].id_buf[2];
232 id_buf[3] = pdf_bufs[salt_pos].id_buf[3];
238 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
240 const u32 pw_r_len = combs_buf[il_pos].pw_len;
242 const u32 pw_len = pw_l_len + pw_r_len;
246 wordr0[0] = combs_buf[il_pos].i[0];
247 wordr0[1] = combs_buf[il_pos].i[1];
248 wordr0[2] = combs_buf[il_pos].i[2];
249 wordr0[3] = combs_buf[il_pos].i[3];
253 wordr1[0] = combs_buf[il_pos].i[4];
254 wordr1[1] = combs_buf[il_pos].i[5];
255 wordr1[2] = combs_buf[il_pos].i[6];
256 wordr1[3] = combs_buf[il_pos].i[7];
272 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
274 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
279 w0[0] = wordl0[0] | wordr0[0];
280 w0[1] = wordl0[1] | wordr0[1];
281 w0[2] = wordl0[2] | wordr0[2];
282 w0[3] = wordl0[3] | wordr0[3];
286 w1[0] = wordl1[0] | wordr1[0];
287 w1[1] = wordl1[1] | wordr1[1];
288 w1[2] = wordl1[2] | wordr1[2];
289 w1[3] = wordl1[3] | wordr1[3];
293 w2[0] = wordl2[0] | wordr2[0];
294 w2[1] = wordl2[1] | wordr2[1];
295 w2[2] = wordl2[2] | wordr2[2];
296 w2[3] = wordl2[3] | wordr2[3];
300 w3[0] = wordl3[0] | wordr3[0];
301 w3[1] = wordl3[1] | wordr3[1];
302 w3[2] = wordl3[2] | wordr3[2];
303 w3[3] = wordl3[3] | wordr3[3];
310 // max length supported by pdf11 is 32
312 w0_t[0] = padding[0];
313 w0_t[1] = padding[1];
314 w0_t[2] = padding[2];
315 w0_t[3] = padding[3];
316 w1_t[0] = padding[4];
317 w1_t[1] = padding[5];
318 w1_t[2] = padding[6];
319 w1_t[3] = padding[7];
329 switch_buffer_by_offset (w0_t, w1_t, w2_t, w3_t, pw_len);
332 // truncate at 32 is wanted, not a bug!
359 md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
378 md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
381 u32x b = digest[1] & 0xff;
388 #include VECT_COMPARE_M
392 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m10420_m08 (__global pw_t *pws, __global gpu_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)
396 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m10420_m16 (__global pw_t *pws, __global gpu_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)
400 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m10420_s04 (__global pw_t *pws, __global gpu_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)
406 const u32 lid = get_local_id (0);
412 const u32 gid = get_global_id (0);
414 if (gid >= gid_max) return;
418 wordl0[0] = pws[gid].i[ 0];
419 wordl0[1] = pws[gid].i[ 1];
420 wordl0[2] = pws[gid].i[ 2];
421 wordl0[3] = pws[gid].i[ 3];
425 wordl1[0] = pws[gid].i[ 4];
426 wordl1[1] = pws[gid].i[ 5];
427 wordl1[2] = pws[gid].i[ 6];
428 wordl1[3] = pws[gid].i[ 7];
444 const u32 pw_l_len = pws[gid].pw_len;
446 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
448 append_0x80_2 (wordl0, wordl1, pw_l_len);
450 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
457 const u32 search[4] =
459 digests_buf[digests_offset].digest_buf[DGST_R0],
460 digests_buf[digests_offset].digest_buf[DGST_R1],
461 digests_buf[digests_offset].digest_buf[DGST_R2],
462 digests_buf[digests_offset].digest_buf[DGST_R3]
471 o_buf[0] = pdf_bufs[salt_pos].o_buf[0];
472 o_buf[1] = pdf_bufs[salt_pos].o_buf[1];
473 o_buf[2] = pdf_bufs[salt_pos].o_buf[2];
474 o_buf[3] = pdf_bufs[salt_pos].o_buf[3];
475 o_buf[4] = pdf_bufs[salt_pos].o_buf[4];
476 o_buf[5] = pdf_bufs[salt_pos].o_buf[5];
477 o_buf[6] = pdf_bufs[salt_pos].o_buf[6];
478 o_buf[7] = pdf_bufs[salt_pos].o_buf[7];
480 u32 P = pdf_bufs[salt_pos].P;
484 id_buf[0] = pdf_bufs[salt_pos].id_buf[0];
485 id_buf[1] = pdf_bufs[salt_pos].id_buf[1];
486 id_buf[2] = pdf_bufs[salt_pos].id_buf[2];
487 id_buf[3] = pdf_bufs[salt_pos].id_buf[3];
493 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
495 const u32 pw_r_len = combs_buf[il_pos].pw_len;
497 const u32 pw_len = pw_l_len + pw_r_len;
501 wordr0[0] = combs_buf[il_pos].i[0];
502 wordr0[1] = combs_buf[il_pos].i[1];
503 wordr0[2] = combs_buf[il_pos].i[2];
504 wordr0[3] = combs_buf[il_pos].i[3];
508 wordr1[0] = combs_buf[il_pos].i[4];
509 wordr1[1] = combs_buf[il_pos].i[5];
510 wordr1[2] = combs_buf[il_pos].i[6];
511 wordr1[3] = combs_buf[il_pos].i[7];
527 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
529 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
534 w0[0] = wordl0[0] | wordr0[0];
535 w0[1] = wordl0[1] | wordr0[1];
536 w0[2] = wordl0[2] | wordr0[2];
537 w0[3] = wordl0[3] | wordr0[3];
541 w1[0] = wordl1[0] | wordr1[0];
542 w1[1] = wordl1[1] | wordr1[1];
543 w1[2] = wordl1[2] | wordr1[2];
544 w1[3] = wordl1[3] | wordr1[3];
548 w2[0] = wordl2[0] | wordr2[0];
549 w2[1] = wordl2[1] | wordr2[1];
550 w2[2] = wordl2[2] | wordr2[2];
551 w2[3] = wordl2[3] | wordr2[3];
555 w3[0] = wordl3[0] | wordr3[0];
556 w3[1] = wordl3[1] | wordr3[1];
557 w3[2] = wordl3[2] | wordr3[2];
558 w3[3] = wordl3[3] | wordr3[3];
565 // max length supported by pdf11 is 32
567 w0_t[0] = padding[0];
568 w0_t[1] = padding[1];
569 w0_t[2] = padding[2];
570 w0_t[3] = padding[3];
571 w1_t[0] = padding[4];
572 w1_t[1] = padding[5];
573 w1_t[2] = padding[6];
574 w1_t[3] = padding[7];
584 switch_buffer_by_offset (w0_t, w1_t, w2_t, w3_t, pw_len);
587 // truncate at 32 is wanted, not a bug!
614 md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
633 md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
636 u32x b = digest[1] & 0xff;
643 #include VECT_COMPARE_S
647 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m10420_s08 (__global pw_t *pws, __global gpu_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)
651 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m10420_s16 (__global pw_t *pws, __global gpu_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)