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"
18 #include "OpenCL/types_ocl.c"
19 #include "OpenCL/common.c"
21 #define COMPARE_S "OpenCL/check_single_comp4.c"
22 #define COMPARE_M "OpenCL/check_multi_comp4.c"
24 void md5_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4])
48 MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
49 MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
50 MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02);
51 MD5_STEP (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03);
52 MD5_STEP (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00);
53 MD5_STEP (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01);
54 MD5_STEP (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02);
55 MD5_STEP (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03);
56 MD5_STEP (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00);
57 MD5_STEP (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01);
58 MD5_STEP (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02);
59 MD5_STEP (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03);
60 MD5_STEP (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00);
61 MD5_STEP (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01);
62 MD5_STEP (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02);
63 MD5_STEP (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03);
65 MD5_STEP (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10);
66 MD5_STEP (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11);
67 MD5_STEP (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12);
68 MD5_STEP (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13);
69 MD5_STEP (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10);
70 MD5_STEP (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11);
71 MD5_STEP (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12);
72 MD5_STEP (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13);
73 MD5_STEP (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10);
74 MD5_STEP (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11);
75 MD5_STEP (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12);
76 MD5_STEP (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13);
77 MD5_STEP (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10);
78 MD5_STEP (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11);
79 MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12);
80 MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13);
82 MD5_STEP (MD5_H , a, b, c, d, w5_t, MD5C20, MD5S20);
83 MD5_STEP (MD5_H , d, a, b, c, w8_t, MD5C21, MD5S21);
84 MD5_STEP (MD5_H , c, d, a, b, wb_t, MD5C22, MD5S22);
85 MD5_STEP (MD5_H , b, c, d, a, we_t, MD5C23, MD5S23);
86 MD5_STEP (MD5_H , a, b, c, d, w1_t, MD5C24, MD5S20);
87 MD5_STEP (MD5_H , d, a, b, c, w4_t, MD5C25, MD5S21);
88 MD5_STEP (MD5_H , c, d, a, b, w7_t, MD5C26, MD5S22);
89 MD5_STEP (MD5_H , b, c, d, a, wa_t, MD5C27, MD5S23);
90 MD5_STEP (MD5_H , a, b, c, d, wd_t, MD5C28, MD5S20);
91 MD5_STEP (MD5_H , d, a, b, c, w0_t, MD5C29, MD5S21);
92 MD5_STEP (MD5_H , c, d, a, b, w3_t, MD5C2a, MD5S22);
93 MD5_STEP (MD5_H , b, c, d, a, w6_t, MD5C2b, MD5S23);
94 MD5_STEP (MD5_H , a, b, c, d, w9_t, MD5C2c, MD5S20);
95 MD5_STEP (MD5_H , d, a, b, c, wc_t, MD5C2d, MD5S21);
96 MD5_STEP (MD5_H , c, d, a, b, wf_t, MD5C2e, MD5S22);
97 MD5_STEP (MD5_H , b, c, d, a, w2_t, MD5C2f, MD5S23);
99 MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30);
100 MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31);
101 MD5_STEP (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32);
102 MD5_STEP (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33);
103 MD5_STEP (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30);
104 MD5_STEP (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31);
105 MD5_STEP (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32);
106 MD5_STEP (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33);
107 MD5_STEP (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30);
108 MD5_STEP (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31);
109 MD5_STEP (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32);
110 MD5_STEP (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33);
111 MD5_STEP (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30);
112 MD5_STEP (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31);
113 MD5_STEP (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32);
114 MD5_STEP (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33);
122 void hmac_md5_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4])
124 w0[0] = w0[0] ^ 0x36363636;
125 w0[1] = w0[1] ^ 0x36363636;
126 w0[2] = w0[2] ^ 0x36363636;
127 w0[3] = w0[3] ^ 0x36363636;
128 w1[0] = w1[0] ^ 0x36363636;
129 w1[1] = w1[1] ^ 0x36363636;
130 w1[2] = w1[2] ^ 0x36363636;
131 w1[3] = w1[3] ^ 0x36363636;
132 w2[0] = w2[0] ^ 0x36363636;
133 w2[1] = w2[1] ^ 0x36363636;
134 w2[2] = w2[2] ^ 0x36363636;
135 w2[3] = w2[3] ^ 0x36363636;
136 w3[0] = w3[0] ^ 0x36363636;
137 w3[1] = w3[1] ^ 0x36363636;
138 w3[2] = w3[2] ^ 0x36363636;
139 w3[3] = w3[3] ^ 0x36363636;
146 md5_transform (w0, w1, w2, w3, ipad);
148 w0[0] = w0[0] ^ 0x6a6a6a6a;
149 w0[1] = w0[1] ^ 0x6a6a6a6a;
150 w0[2] = w0[2] ^ 0x6a6a6a6a;
151 w0[3] = w0[3] ^ 0x6a6a6a6a;
152 w1[0] = w1[0] ^ 0x6a6a6a6a;
153 w1[1] = w1[1] ^ 0x6a6a6a6a;
154 w1[2] = w1[2] ^ 0x6a6a6a6a;
155 w1[3] = w1[3] ^ 0x6a6a6a6a;
156 w2[0] = w2[0] ^ 0x6a6a6a6a;
157 w2[1] = w2[1] ^ 0x6a6a6a6a;
158 w2[2] = w2[2] ^ 0x6a6a6a6a;
159 w2[3] = w2[3] ^ 0x6a6a6a6a;
160 w3[0] = w3[0] ^ 0x6a6a6a6a;
161 w3[1] = w3[1] ^ 0x6a6a6a6a;
162 w3[2] = w3[2] ^ 0x6a6a6a6a;
163 w3[3] = w3[3] ^ 0x6a6a6a6a;
170 md5_transform (w0, w1, w2, w3, opad);
173 void hmac_md5_run (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4], u32 digest[4])
180 md5_transform (w0, w1, w2, w3, digest);
196 w3[2] = (64 + 16) * 8;
204 md5_transform (w0, w1, w2, w3, digest);
207 __kernel void m11900_init (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global pbkdf2_md5_tmp_t *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 pbkdf2_md5_t *esalt_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 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
213 const u32 gid = get_global_id (0);
215 if (gid >= gid_max) return;
219 w0[0] = pws[gid].i[ 0];
220 w0[1] = pws[gid].i[ 1];
221 w0[2] = pws[gid].i[ 2];
222 w0[3] = pws[gid].i[ 3];
226 w1[0] = pws[gid].i[ 4];
227 w1[1] = pws[gid].i[ 5];
228 w1[2] = pws[gid].i[ 6];
229 w1[3] = pws[gid].i[ 7];
233 w2[0] = pws[gid].i[ 8];
234 w2[1] = pws[gid].i[ 9];
235 w2[2] = pws[gid].i[10];
236 w2[3] = pws[gid].i[11];
240 w3[0] = pws[gid].i[12];
241 w3[1] = pws[gid].i[13];
242 w3[2] = pws[gid].i[14];
243 w3[3] = pws[gid].i[15];
249 const u32 salt_len = salt_bufs[salt_pos].salt_len;
256 esalt_buf0[0] = esalt_bufs[salt_pos].salt_buf[ 0];
257 esalt_buf0[1] = esalt_bufs[salt_pos].salt_buf[ 1];
258 esalt_buf0[2] = esalt_bufs[salt_pos].salt_buf[ 2];
259 esalt_buf0[3] = esalt_bufs[salt_pos].salt_buf[ 3];
260 esalt_buf1[0] = esalt_bufs[salt_pos].salt_buf[ 4];
261 esalt_buf1[1] = esalt_bufs[salt_pos].salt_buf[ 5];
262 esalt_buf1[2] = esalt_bufs[salt_pos].salt_buf[ 6];
263 esalt_buf1[3] = esalt_bufs[salt_pos].salt_buf[ 7];
264 esalt_buf2[0] = esalt_bufs[salt_pos].salt_buf[ 8];
265 esalt_buf2[1] = esalt_bufs[salt_pos].salt_buf[ 9];
266 esalt_buf2[2] = esalt_bufs[salt_pos].salt_buf[10];
267 esalt_buf2[3] = esalt_bufs[salt_pos].salt_buf[11];
268 esalt_buf3[0] = esalt_bufs[salt_pos].salt_buf[12];
269 esalt_buf3[1] = esalt_bufs[salt_pos].salt_buf[13];
270 esalt_buf3[2] = (64 + salt_len + 4) * 8;
276 hmac_md5_pad (w0, w1, w2, w3, ipad, opad);
278 tmps[gid].ipad[0] = ipad[0];
279 tmps[gid].ipad[1] = ipad[1];
280 tmps[gid].ipad[2] = ipad[2];
281 tmps[gid].ipad[3] = ipad[3];
283 tmps[gid].opad[0] = opad[0];
284 tmps[gid].opad[1] = opad[1];
285 tmps[gid].opad[2] = opad[2];
286 tmps[gid].opad[3] = opad[3];
288 for (u32 i = 0, j = 1; i < 4; i += 4, j += 1)
292 hmac_md5_run (esalt_buf0, esalt_buf1, esalt_buf2, esalt_buf3, ipad, opad, dgst);
294 tmps[gid].dgst[i + 0] = dgst[0];
295 tmps[gid].dgst[i + 1] = dgst[1];
296 tmps[gid].dgst[i + 2] = dgst[2];
297 tmps[gid].dgst[i + 3] = dgst[3];
299 tmps[gid].out[i + 0] = dgst[0];
300 tmps[gid].out[i + 1] = dgst[1];
301 tmps[gid].out[i + 2] = dgst[2];
302 tmps[gid].out[i + 3] = dgst[3];
306 __kernel void m11900_loop (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global pbkdf2_md5_tmp_t *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 pbkdf2_md5_t *esalt_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 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
308 const u32 gid = get_global_id (0);
310 if (gid >= gid_max) return;
314 ipad[0] = tmps[gid].ipad[0];
315 ipad[1] = tmps[gid].ipad[1];
316 ipad[2] = tmps[gid].ipad[2];
317 ipad[3] = tmps[gid].ipad[3];
321 opad[0] = tmps[gid].opad[0];
322 opad[1] = tmps[gid].opad[1];
323 opad[2] = tmps[gid].opad[2];
324 opad[3] = tmps[gid].opad[3];
326 for (u32 i = 0; i < 4; i += 4)
330 dgst[0] = tmps[gid].dgst[i + 0];
331 dgst[1] = tmps[gid].dgst[i + 1];
332 dgst[2] = tmps[gid].dgst[i + 2];
333 dgst[3] = tmps[gid].dgst[i + 3];
337 out[0] = tmps[gid].out[i + 0];
338 out[1] = tmps[gid].out[i + 1];
339 out[2] = tmps[gid].out[i + 2];
340 out[3] = tmps[gid].out[i + 3];
342 for (u32 j = 0; j < loop_cnt; j++)
363 w3[2] = (64 + 16) * 8;
366 hmac_md5_run (w0, w1, w2, w3, ipad, opad, dgst);
374 tmps[gid].dgst[i + 0] = dgst[0];
375 tmps[gid].dgst[i + 1] = dgst[1];
376 tmps[gid].dgst[i + 2] = dgst[2];
377 tmps[gid].dgst[i + 3] = dgst[3];
379 tmps[gid].out[i + 0] = out[0];
380 tmps[gid].out[i + 1] = out[1];
381 tmps[gid].out[i + 2] = out[2];
382 tmps[gid].out[i + 3] = out[3];
386 __kernel void m11900_comp (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global pbkdf2_md5_tmp_t *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 pbkdf2_md5_t *esalt_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 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
392 const u32 gid = get_global_id (0);
394 if (gid >= gid_max) return;
396 const u32 lid = get_local_id (0);
398 const u32 r0 = tmps[gid].out[DGST_R0];
399 const u32 r1 = tmps[gid].out[DGST_R1];
400 const u32 r2 = tmps[gid].out[DGST_R2];
401 const u32 r3 = tmps[gid].out[DGST_R3];