2 * Author......: Jens Steube <jens.steube@gmail.com>
8 #include "include/constants.h"
9 #include "include/kernel_vendor.h"
16 #include "include/kernel_functions.c"
17 #include "OpenCL/types_ocl.c"
18 #include "OpenCL/common.c"
20 #define COMPARE_S "OpenCL/check_single_comp4.c"
21 #define COMPARE_M "OpenCL/check_multi_comp4.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 __attribute__((reqd_work_group_size (64, 1, 1))) 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;
151 wordl0[0] = pws[gid].i[ 0];
152 wordl0[1] = pws[gid].i[ 1];
153 wordl0[2] = pws[gid].i[ 2];
154 wordl0[3] = pws[gid].i[ 3];
158 wordl1[0] = pws[gid].i[ 4];
159 wordl1[1] = pws[gid].i[ 5];
160 wordl1[2] = pws[gid].i[ 6];
161 wordl1[3] = pws[gid].i[ 7];
177 const u32 pw_l_len = pws[gid].pw_len;
179 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
181 append_0x80_2x4 (wordl0, wordl1, pw_l_len);
183 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
192 o_buf[0] = pdf_bufs[salt_pos].o_buf[0];
193 o_buf[1] = pdf_bufs[salt_pos].o_buf[1];
194 o_buf[2] = pdf_bufs[salt_pos].o_buf[2];
195 o_buf[3] = pdf_bufs[salt_pos].o_buf[3];
196 o_buf[4] = pdf_bufs[salt_pos].o_buf[4];
197 o_buf[5] = pdf_bufs[salt_pos].o_buf[5];
198 o_buf[6] = pdf_bufs[salt_pos].o_buf[6];
199 o_buf[7] = pdf_bufs[salt_pos].o_buf[7];
201 u32 P = pdf_bufs[salt_pos].P;
205 id_buf[0] = pdf_bufs[salt_pos].id_buf[0];
206 id_buf[1] = pdf_bufs[salt_pos].id_buf[1];
207 id_buf[2] = pdf_bufs[salt_pos].id_buf[2];
208 id_buf[3] = pdf_bufs[salt_pos].id_buf[3];
214 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
216 const u32 pw_r_len = combs_buf[il_pos].pw_len;
218 const u32 pw_len = pw_l_len + pw_r_len;
222 wordr0[0] = combs_buf[il_pos].i[0];
223 wordr0[1] = combs_buf[il_pos].i[1];
224 wordr0[2] = combs_buf[il_pos].i[2];
225 wordr0[3] = combs_buf[il_pos].i[3];
229 wordr1[0] = combs_buf[il_pos].i[4];
230 wordr1[1] = combs_buf[il_pos].i[5];
231 wordr1[2] = combs_buf[il_pos].i[6];
232 wordr1[3] = combs_buf[il_pos].i[7];
248 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
250 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
255 w0[0] = wordl0[0] | wordr0[0];
256 w0[1] = wordl0[1] | wordr0[1];
257 w0[2] = wordl0[2] | wordr0[2];
258 w0[3] = wordl0[3] | wordr0[3];
262 w1[0] = wordl1[0] | wordr1[0];
263 w1[1] = wordl1[1] | wordr1[1];
264 w1[2] = wordl1[2] | wordr1[2];
265 w1[3] = wordl1[3] | wordr1[3];
269 w2[0] = wordl2[0] | wordr2[0];
270 w2[1] = wordl2[1] | wordr2[1];
271 w2[2] = wordl2[2] | wordr2[2];
272 w2[3] = wordl2[3] | wordr2[3];
276 w3[0] = wordl3[0] | wordr3[0];
277 w3[1] = wordl3[1] | wordr3[1];
278 w3[2] = wordl3[2] | wordr3[2];
279 w3[3] = wordl3[3] | wordr3[3];
286 // max length supported by pdf11 is 32
288 w0_t[0] = padding[0];
289 w0_t[1] = padding[1];
290 w0_t[2] = padding[2];
291 w0_t[3] = padding[3];
292 w1_t[0] = padding[4];
293 w1_t[1] = padding[5];
294 w1_t[2] = padding[6];
295 w1_t[3] = padding[7];
305 switch_buffer_by_offset (w0_t, w1_t, w2_t, w3_t, pw_len);
308 // truncate at 32 is wanted, not a bug!
335 md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
354 md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
357 u32 b = digest[1] & 0xff;
368 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) 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)
372 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) 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)
376 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) 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)
382 const u32 lid = get_local_id (0);
388 const u32 gid = get_global_id (0);
390 if (gid >= gid_max) return;
394 wordl0[0] = pws[gid].i[ 0];
395 wordl0[1] = pws[gid].i[ 1];
396 wordl0[2] = pws[gid].i[ 2];
397 wordl0[3] = pws[gid].i[ 3];
401 wordl1[0] = pws[gid].i[ 4];
402 wordl1[1] = pws[gid].i[ 5];
403 wordl1[2] = pws[gid].i[ 6];
404 wordl1[3] = pws[gid].i[ 7];
420 const u32 pw_l_len = pws[gid].pw_len;
422 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
424 append_0x80_2x4 (wordl0, wordl1, pw_l_len);
426 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
433 const u32 search[4] =
435 digests_buf[digests_offset].digest_buf[DGST_R0],
436 digests_buf[digests_offset].digest_buf[DGST_R1],
437 digests_buf[digests_offset].digest_buf[DGST_R2],
438 digests_buf[digests_offset].digest_buf[DGST_R3]
447 o_buf[0] = pdf_bufs[salt_pos].o_buf[0];
448 o_buf[1] = pdf_bufs[salt_pos].o_buf[1];
449 o_buf[2] = pdf_bufs[salt_pos].o_buf[2];
450 o_buf[3] = pdf_bufs[salt_pos].o_buf[3];
451 o_buf[4] = pdf_bufs[salt_pos].o_buf[4];
452 o_buf[5] = pdf_bufs[salt_pos].o_buf[5];
453 o_buf[6] = pdf_bufs[salt_pos].o_buf[6];
454 o_buf[7] = pdf_bufs[salt_pos].o_buf[7];
456 u32 P = pdf_bufs[salt_pos].P;
460 id_buf[0] = pdf_bufs[salt_pos].id_buf[0];
461 id_buf[1] = pdf_bufs[salt_pos].id_buf[1];
462 id_buf[2] = pdf_bufs[salt_pos].id_buf[2];
463 id_buf[3] = pdf_bufs[salt_pos].id_buf[3];
469 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
471 const u32 pw_r_len = combs_buf[il_pos].pw_len;
473 const u32 pw_len = pw_l_len + pw_r_len;
477 wordr0[0] = combs_buf[il_pos].i[0];
478 wordr0[1] = combs_buf[il_pos].i[1];
479 wordr0[2] = combs_buf[il_pos].i[2];
480 wordr0[3] = combs_buf[il_pos].i[3];
484 wordr1[0] = combs_buf[il_pos].i[4];
485 wordr1[1] = combs_buf[il_pos].i[5];
486 wordr1[2] = combs_buf[il_pos].i[6];
487 wordr1[3] = combs_buf[il_pos].i[7];
503 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
505 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
510 w0[0] = wordl0[0] | wordr0[0];
511 w0[1] = wordl0[1] | wordr0[1];
512 w0[2] = wordl0[2] | wordr0[2];
513 w0[3] = wordl0[3] | wordr0[3];
517 w1[0] = wordl1[0] | wordr1[0];
518 w1[1] = wordl1[1] | wordr1[1];
519 w1[2] = wordl1[2] | wordr1[2];
520 w1[3] = wordl1[3] | wordr1[3];
524 w2[0] = wordl2[0] | wordr2[0];
525 w2[1] = wordl2[1] | wordr2[1];
526 w2[2] = wordl2[2] | wordr2[2];
527 w2[3] = wordl2[3] | wordr2[3];
531 w3[0] = wordl3[0] | wordr3[0];
532 w3[1] = wordl3[1] | wordr3[1];
533 w3[2] = wordl3[2] | wordr3[2];
534 w3[3] = wordl3[3] | wordr3[3];
541 // max length supported by pdf11 is 32
543 w0_t[0] = padding[0];
544 w0_t[1] = padding[1];
545 w0_t[2] = padding[2];
546 w0_t[3] = padding[3];
547 w1_t[0] = padding[4];
548 w1_t[1] = padding[5];
549 w1_t[2] = padding[6];
550 w1_t[3] = padding[7];
560 switch_buffer_by_offset (w0_t, w1_t, w2_t, w3_t, pw_len);
563 // truncate at 32 is wanted, not a bug!
590 md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
609 md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
612 u32 b = digest[1] & 0xff;
623 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) 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)
627 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) 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)