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"
20 void md5_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4])
44 MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
45 MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
46 MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02);
47 MD5_STEP (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03);
48 MD5_STEP (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00);
49 MD5_STEP (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01);
50 MD5_STEP (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02);
51 MD5_STEP (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03);
52 MD5_STEP (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00);
53 MD5_STEP (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01);
54 MD5_STEP (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02);
55 MD5_STEP (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03);
56 MD5_STEP (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00);
57 MD5_STEP (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01);
58 MD5_STEP (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02);
59 MD5_STEP (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03);
61 MD5_STEP (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10);
62 MD5_STEP (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11);
63 MD5_STEP (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12);
64 MD5_STEP (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13);
65 MD5_STEP (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10);
66 MD5_STEP (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11);
67 MD5_STEP (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12);
68 MD5_STEP (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13);
69 MD5_STEP (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10);
70 MD5_STEP (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11);
71 MD5_STEP (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12);
72 MD5_STEP (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13);
73 MD5_STEP (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10);
74 MD5_STEP (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11);
75 MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12);
76 MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13);
78 MD5_STEP (MD5_H , a, b, c, d, w5_t, MD5C20, MD5S20);
79 MD5_STEP (MD5_H , d, a, b, c, w8_t, MD5C21, MD5S21);
80 MD5_STEP (MD5_H , c, d, a, b, wb_t, MD5C22, MD5S22);
81 MD5_STEP (MD5_H , b, c, d, a, we_t, MD5C23, MD5S23);
82 MD5_STEP (MD5_H , a, b, c, d, w1_t, MD5C24, MD5S20);
83 MD5_STEP (MD5_H , d, a, b, c, w4_t, MD5C25, MD5S21);
84 MD5_STEP (MD5_H , c, d, a, b, w7_t, MD5C26, MD5S22);
85 MD5_STEP (MD5_H , b, c, d, a, wa_t, MD5C27, MD5S23);
86 MD5_STEP (MD5_H , a, b, c, d, wd_t, MD5C28, MD5S20);
87 MD5_STEP (MD5_H , d, a, b, c, w0_t, MD5C29, MD5S21);
88 MD5_STEP (MD5_H , c, d, a, b, w3_t, MD5C2a, MD5S22);
89 MD5_STEP (MD5_H , b, c, d, a, w6_t, MD5C2b, MD5S23);
90 MD5_STEP (MD5_H , a, b, c, d, w9_t, MD5C2c, MD5S20);
91 MD5_STEP (MD5_H , d, a, b, c, wc_t, MD5C2d, MD5S21);
92 MD5_STEP (MD5_H , c, d, a, b, wf_t, MD5C2e, MD5S22);
93 MD5_STEP (MD5_H , b, c, d, a, w2_t, MD5C2f, MD5S23);
95 MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30);
96 MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31);
97 MD5_STEP (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32);
98 MD5_STEP (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33);
99 MD5_STEP (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30);
100 MD5_STEP (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31);
101 MD5_STEP (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32);
102 MD5_STEP (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33);
103 MD5_STEP (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30);
104 MD5_STEP (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31);
105 MD5_STEP (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32);
106 MD5_STEP (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33);
107 MD5_STEP (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30);
108 MD5_STEP (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31);
109 MD5_STEP (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32);
110 MD5_STEP (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33);
118 void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
120 w0[0] = w0[0] ^ 0x36363636;
121 w0[1] = w0[1] ^ 0x36363636;
122 w0[2] = w0[2] ^ 0x36363636;
123 w0[3] = w0[3] ^ 0x36363636;
124 w1[0] = w1[0] ^ 0x36363636;
125 w1[1] = w1[1] ^ 0x36363636;
126 w1[2] = w1[2] ^ 0x36363636;
127 w1[3] = w1[3] ^ 0x36363636;
128 w2[0] = w2[0] ^ 0x36363636;
129 w2[1] = w2[1] ^ 0x36363636;
130 w2[2] = w2[2] ^ 0x36363636;
131 w2[3] = w2[3] ^ 0x36363636;
132 w3[0] = w3[0] ^ 0x36363636;
133 w3[1] = w3[1] ^ 0x36363636;
134 w3[2] = w3[2] ^ 0x36363636;
135 w3[3] = w3[3] ^ 0x36363636;
142 md5_transform (w0, w1, w2, w3, ipad);
144 w0[0] = w0[0] ^ 0x6a6a6a6a;
145 w0[1] = w0[1] ^ 0x6a6a6a6a;
146 w0[2] = w0[2] ^ 0x6a6a6a6a;
147 w0[3] = w0[3] ^ 0x6a6a6a6a;
148 w1[0] = w1[0] ^ 0x6a6a6a6a;
149 w1[1] = w1[1] ^ 0x6a6a6a6a;
150 w1[2] = w1[2] ^ 0x6a6a6a6a;
151 w1[3] = w1[3] ^ 0x6a6a6a6a;
152 w2[0] = w2[0] ^ 0x6a6a6a6a;
153 w2[1] = w2[1] ^ 0x6a6a6a6a;
154 w2[2] = w2[2] ^ 0x6a6a6a6a;
155 w2[3] = w2[3] ^ 0x6a6a6a6a;
156 w3[0] = w3[0] ^ 0x6a6a6a6a;
157 w3[1] = w3[1] ^ 0x6a6a6a6a;
158 w3[2] = w3[2] ^ 0x6a6a6a6a;
159 w3[3] = w3[3] ^ 0x6a6a6a6a;
166 md5_transform (w0, w1, w2, w3, opad);
169 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])
176 md5_transform (w0, w1, w2, w3, digest);
192 w3[2] = (64 + 16) * 8;
200 md5_transform (w0, w1, w2, w3, digest);
203 __kernel void m00060_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)
209 const u32 lid = get_local_id (0);
215 const u32 gid = get_global_id (0);
217 if (gid >= gid_max) return;
222 pw_buf0[0] = pws[gid].i[0];
223 pw_buf0[1] = pws[gid].i[1];
224 pw_buf0[2] = pws[gid].i[2];
225 pw_buf0[3] = pws[gid].i[3];
226 pw_buf1[0] = pws[gid].i[4];
227 pw_buf1[1] = pws[gid].i[5];
228 pw_buf1[2] = pws[gid].i[6];
229 pw_buf1[3] = pws[gid].i[7];
231 const u32 pw_len = pws[gid].pw_len;
242 salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0];
243 salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1];
244 salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2];
245 salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3];
246 salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
247 salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
248 salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
249 salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
250 salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8];
251 salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9];
252 salt_buf2[2] = salt_bufs[salt_pos].salt_buf[10];
253 salt_buf2[3] = salt_bufs[salt_pos].salt_buf[11];
254 salt_buf3[0] = salt_bufs[salt_pos].salt_buf[12];
255 salt_buf3[1] = salt_bufs[salt_pos].salt_buf[13];
256 salt_buf3[2] = salt_bufs[salt_pos].salt_buf[14];
257 salt_buf3[3] = salt_bufs[salt_pos].salt_buf[15];
259 const u32 salt_len = salt_bufs[salt_pos].salt_len;
270 w0_t[0] = salt_buf0[0];
271 w0_t[1] = salt_buf0[1];
272 w0_t[2] = salt_buf0[2];
273 w0_t[3] = salt_buf0[3];
274 w1_t[0] = salt_buf1[0];
275 w1_t[1] = salt_buf1[1];
276 w1_t[2] = salt_buf1[2];
277 w1_t[3] = salt_buf1[3];
278 w2_t[0] = salt_buf2[0];
279 w2_t[1] = salt_buf2[1];
280 w2_t[2] = salt_buf2[2];
281 w2_t[3] = salt_buf2[3];
282 w3_t[0] = salt_buf3[0];
283 w3_t[1] = salt_buf3[1];
284 w3_t[2] = salt_buf3[2];
285 w3_t[3] = salt_buf3[3];
290 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
296 for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
303 u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
305 out_len = mangle(w0, w1, out_len);
307 append_0x80_2x4_VV (w0, w1, out_len);
323 w3_t[2] = (64 + out_len) * 8;
328 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
330 COMPARE_M_SIMD (digest[0], digest[3], digest[2], digest[1]);
334 __kernel void m00060_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)
338 __kernel void m00060_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)
342 __kernel void m00060_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)
348 const u32 lid = get_local_id (0);
354 const u32 gid = get_global_id (0);
356 if (gid >= gid_max) return;
361 pw_buf0[0] = pws[gid].i[0];
362 pw_buf0[1] = pws[gid].i[1];
363 pw_buf0[2] = pws[gid].i[2];
364 pw_buf0[3] = pws[gid].i[3];
365 pw_buf1[0] = pws[gid].i[4];
366 pw_buf1[1] = pws[gid].i[5];
367 pw_buf1[2] = pws[gid].i[6];
368 pw_buf1[3] = pws[gid].i[7];
370 const u32 pw_len = pws[gid].pw_len;
381 salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0];
382 salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1];
383 salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2];
384 salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3];
385 salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
386 salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
387 salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
388 salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
389 salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8];
390 salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9];
391 salt_buf2[2] = salt_bufs[salt_pos].salt_buf[10];
392 salt_buf2[3] = salt_bufs[salt_pos].salt_buf[11];
393 salt_buf3[0] = salt_bufs[salt_pos].salt_buf[12];
394 salt_buf3[1] = salt_bufs[salt_pos].salt_buf[13];
395 salt_buf3[2] = salt_bufs[salt_pos].salt_buf[14];
396 salt_buf3[3] = salt_bufs[salt_pos].salt_buf[15];
398 const u32 salt_len = salt_bufs[salt_pos].salt_len;
409 w0_t[0] = salt_buf0[0];
410 w0_t[1] = salt_buf0[1];
411 w0_t[2] = salt_buf0[2];
412 w0_t[3] = salt_buf0[3];
413 w1_t[0] = salt_buf1[0];
414 w1_t[1] = salt_buf1[1];
415 w1_t[2] = salt_buf1[2];
416 w1_t[3] = salt_buf1[3];
417 w2_t[0] = salt_buf2[0];
418 w2_t[1] = salt_buf2[1];
419 w2_t[2] = salt_buf2[2];
420 w2_t[3] = salt_buf2[3];
421 w3_t[0] = salt_buf3[0];
422 w3_t[1] = salt_buf3[1];
423 w3_t[2] = salt_buf3[2];
424 w3_t[3] = salt_buf3[3];
429 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
435 const u32 search[4] =
437 digests_buf[digests_offset].digest_buf[DGST_R0],
438 digests_buf[digests_offset].digest_buf[DGST_R1],
439 digests_buf[digests_offset].digest_buf[DGST_R2],
440 digests_buf[digests_offset].digest_buf[DGST_R3]
447 for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
454 u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
456 out_len = mangle(w0, w1, out_len);
458 append_0x80_2x4_VV (w0, w1, out_len);
474 w3_t[2] = (64 + out_len) * 8;
479 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
481 COMPARE_S_SIMD (digest[0], digest[3], digest[2], digest[1]);
485 __kernel void m00060_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)
489 __kernel void m00060_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)