2 * Author......: Jens Steube <jens.steube@gmail.com>
10 #include "inc_vendor.cl"
11 #include "inc_hash_constants.h"
12 #include "inc_hash_functions.cl"
13 #include "inc_types.cl"
14 #include "inc_common.cl"
17 #include "inc_simd.cl"
19 void md5_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4])
43 MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
44 MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
45 MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02);
46 MD5_STEP (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03);
47 MD5_STEP (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00);
48 MD5_STEP (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01);
49 MD5_STEP (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02);
50 MD5_STEP (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03);
51 MD5_STEP (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00);
52 MD5_STEP (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01);
53 MD5_STEP (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02);
54 MD5_STEP (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03);
55 MD5_STEP (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00);
56 MD5_STEP (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01);
57 MD5_STEP (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02);
58 MD5_STEP (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03);
60 MD5_STEP (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10);
61 MD5_STEP (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11);
62 MD5_STEP (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12);
63 MD5_STEP (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13);
64 MD5_STEP (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10);
65 MD5_STEP (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11);
66 MD5_STEP (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12);
67 MD5_STEP (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13);
68 MD5_STEP (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10);
69 MD5_STEP (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11);
70 MD5_STEP (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12);
71 MD5_STEP (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13);
72 MD5_STEP (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10);
73 MD5_STEP (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11);
74 MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12);
75 MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13);
77 MD5_STEP (MD5_H , a, b, c, d, w5_t, MD5C20, MD5S20);
78 MD5_STEP (MD5_H , d, a, b, c, w8_t, MD5C21, MD5S21);
79 MD5_STEP (MD5_H , c, d, a, b, wb_t, MD5C22, MD5S22);
80 MD5_STEP (MD5_H , b, c, d, a, we_t, MD5C23, MD5S23);
81 MD5_STEP (MD5_H , a, b, c, d, w1_t, MD5C24, MD5S20);
82 MD5_STEP (MD5_H , d, a, b, c, w4_t, MD5C25, MD5S21);
83 MD5_STEP (MD5_H , c, d, a, b, w7_t, MD5C26, MD5S22);
84 MD5_STEP (MD5_H , b, c, d, a, wa_t, MD5C27, MD5S23);
85 MD5_STEP (MD5_H , a, b, c, d, wd_t, MD5C28, MD5S20);
86 MD5_STEP (MD5_H , d, a, b, c, w0_t, MD5C29, MD5S21);
87 MD5_STEP (MD5_H , c, d, a, b, w3_t, MD5C2a, MD5S22);
88 MD5_STEP (MD5_H , b, c, d, a, w6_t, MD5C2b, MD5S23);
89 MD5_STEP (MD5_H , a, b, c, d, w9_t, MD5C2c, MD5S20);
90 MD5_STEP (MD5_H , d, a, b, c, wc_t, MD5C2d, MD5S21);
91 MD5_STEP (MD5_H , c, d, a, b, wf_t, MD5C2e, MD5S22);
92 MD5_STEP (MD5_H , b, c, d, a, w2_t, MD5C2f, MD5S23);
94 MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30);
95 MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31);
96 MD5_STEP (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32);
97 MD5_STEP (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33);
98 MD5_STEP (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30);
99 MD5_STEP (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31);
100 MD5_STEP (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32);
101 MD5_STEP (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33);
102 MD5_STEP (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30);
103 MD5_STEP (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31);
104 MD5_STEP (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32);
105 MD5_STEP (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33);
106 MD5_STEP (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30);
107 MD5_STEP (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31);
108 MD5_STEP (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32);
109 MD5_STEP (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33);
117 void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
119 w0[0] = w0[0] ^ 0x36363636;
120 w0[1] = w0[1] ^ 0x36363636;
121 w0[2] = w0[2] ^ 0x36363636;
122 w0[3] = w0[3] ^ 0x36363636;
123 w1[0] = w1[0] ^ 0x36363636;
124 w1[1] = w1[1] ^ 0x36363636;
125 w1[2] = w1[2] ^ 0x36363636;
126 w1[3] = w1[3] ^ 0x36363636;
127 w2[0] = w2[0] ^ 0x36363636;
128 w2[1] = w2[1] ^ 0x36363636;
129 w2[2] = w2[2] ^ 0x36363636;
130 w2[3] = w2[3] ^ 0x36363636;
131 w3[0] = w3[0] ^ 0x36363636;
132 w3[1] = w3[1] ^ 0x36363636;
133 w3[2] = w3[2] ^ 0x36363636;
134 w3[3] = w3[3] ^ 0x36363636;
141 md5_transform (w0, w1, w2, w3, ipad);
143 w0[0] = w0[0] ^ 0x6a6a6a6a;
144 w0[1] = w0[1] ^ 0x6a6a6a6a;
145 w0[2] = w0[2] ^ 0x6a6a6a6a;
146 w0[3] = w0[3] ^ 0x6a6a6a6a;
147 w1[0] = w1[0] ^ 0x6a6a6a6a;
148 w1[1] = w1[1] ^ 0x6a6a6a6a;
149 w1[2] = w1[2] ^ 0x6a6a6a6a;
150 w1[3] = w1[3] ^ 0x6a6a6a6a;
151 w2[0] = w2[0] ^ 0x6a6a6a6a;
152 w2[1] = w2[1] ^ 0x6a6a6a6a;
153 w2[2] = w2[2] ^ 0x6a6a6a6a;
154 w2[3] = w2[3] ^ 0x6a6a6a6a;
155 w3[0] = w3[0] ^ 0x6a6a6a6a;
156 w3[1] = w3[1] ^ 0x6a6a6a6a;
157 w3[2] = w3[2] ^ 0x6a6a6a6a;
158 w3[3] = w3[3] ^ 0x6a6a6a6a;
165 md5_transform (w0, w1, w2, w3, opad);
168 void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
175 md5_transform (w0, w1, w2, w3, digest);
191 w3[2] = (64 + 16) * 8;
199 md5_transform (w0, w1, w2, w3, digest);
202 __kernel void m00050_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_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_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)
208 const u32 lid = get_local_id (0);
214 const u32 gid = get_global_id (0);
216 if (gid >= gid_max) return;
221 pw_buf0[0] = pws[gid].i[0];
222 pw_buf0[1] = pws[gid].i[1];
223 pw_buf0[2] = pws[gid].i[2];
224 pw_buf0[3] = pws[gid].i[3];
225 pw_buf1[0] = pws[gid].i[4];
226 pw_buf1[1] = pws[gid].i[5];
227 pw_buf1[2] = pws[gid].i[6];
228 pw_buf1[3] = pws[gid].i[7];
230 const u32 pw_len = pws[gid].pw_len;
241 salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0];
242 salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1];
243 salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2];
244 salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3];
245 salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
246 salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
247 salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
248 salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
249 salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8];
250 salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9];
251 salt_buf2[2] = salt_bufs[salt_pos].salt_buf[10];
252 salt_buf2[3] = salt_bufs[salt_pos].salt_buf[11];
253 salt_buf3[0] = salt_bufs[salt_pos].salt_buf[12];
254 salt_buf3[1] = salt_bufs[salt_pos].salt_buf[13];
255 salt_buf3[2] = salt_bufs[salt_pos].salt_buf[14];
256 salt_buf3[3] = salt_bufs[salt_pos].salt_buf[15];
258 const u32 salt_len = salt_bufs[salt_pos].salt_len;
264 for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
271 apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
280 hmac_md5_pad (w0, w1, w2, w3, ipad, opad);
282 w0[0] = salt_buf0[0];
283 w0[1] = salt_buf0[1];
284 w0[2] = salt_buf0[2];
285 w0[3] = salt_buf0[3];
286 w1[0] = salt_buf1[0];
287 w1[1] = salt_buf1[1];
288 w1[2] = salt_buf1[2];
289 w1[3] = salt_buf1[3];
290 w2[0] = salt_buf2[0];
291 w2[1] = salt_buf2[1];
292 w2[2] = salt_buf2[2];
293 w2[3] = salt_buf2[3];
294 w3[0] = salt_buf3[0];
295 w3[1] = salt_buf3[1];
296 w3[2] = (64 + salt_len) * 8;
301 hmac_md5_run (w0, w1, w2, w3, ipad, opad, digest);
303 COMPARE_M_SIMD (digest[0], digest[3], digest[2], digest[1]);
307 __kernel void m00050_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_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_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)
311 __kernel void m00050_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_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_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)
315 __kernel void m00050_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_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_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)
321 const u32 lid = get_local_id (0);
327 const u32 gid = get_global_id (0);
329 if (gid >= gid_max) return;
334 pw_buf0[0] = pws[gid].i[0];
335 pw_buf0[1] = pws[gid].i[1];
336 pw_buf0[2] = pws[gid].i[2];
337 pw_buf0[3] = pws[gid].i[3];
338 pw_buf1[0] = pws[gid].i[4];
339 pw_buf1[1] = pws[gid].i[5];
340 pw_buf1[2] = pws[gid].i[6];
341 pw_buf1[3] = pws[gid].i[7];
343 const u32 pw_len = pws[gid].pw_len;
354 salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0];
355 salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1];
356 salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2];
357 salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3];
358 salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
359 salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
360 salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
361 salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
362 salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8];
363 salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9];
364 salt_buf2[2] = salt_bufs[salt_pos].salt_buf[10];
365 salt_buf2[3] = salt_bufs[salt_pos].salt_buf[11];
366 salt_buf3[0] = salt_bufs[salt_pos].salt_buf[12];
367 salt_buf3[1] = salt_bufs[salt_pos].salt_buf[13];
368 salt_buf3[2] = salt_bufs[salt_pos].salt_buf[14];
369 salt_buf3[3] = salt_bufs[salt_pos].salt_buf[15];
371 const u32 salt_len = salt_bufs[salt_pos].salt_len;
377 const u32 search[4] =
379 digests_buf[digests_offset].digest_buf[DGST_R0],
380 digests_buf[digests_offset].digest_buf[DGST_R1],
381 digests_buf[digests_offset].digest_buf[DGST_R2],
382 digests_buf[digests_offset].digest_buf[DGST_R3]
389 for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
396 apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
405 hmac_md5_pad (w0, w1, w2, w3, ipad, opad);
407 w0[0] = salt_buf0[0];
408 w0[1] = salt_buf0[1];
409 w0[2] = salt_buf0[2];
410 w0[3] = salt_buf0[3];
411 w1[0] = salt_buf1[0];
412 w1[1] = salt_buf1[1];
413 w1[2] = salt_buf1[2];
414 w1[3] = salt_buf1[3];
415 w2[0] = salt_buf2[0];
416 w2[1] = salt_buf2[1];
417 w2[2] = salt_buf2[2];
418 w2[3] = salt_buf2[3];
419 w3[0] = salt_buf3[0];
420 w3[1] = salt_buf3[1];
421 w3[2] = (64 + salt_len) * 8;
426 hmac_md5_run (w0, w1, w2, w3, ipad, opad, digest);
428 COMPARE_S_SIMD (digest[0], digest[3], digest[2], digest[1]);
432 __kernel void m00050_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_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_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)
436 __kernel void m00050_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_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_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)