2 * Author......: Jens Steube <jens.steube@gmail.com>
10 #include "include/constants.h"
11 #include "include/kernel_vendor.h"
18 #include "include/kernel_functions.c"
19 #include "OpenCL/types_ocl.c"
20 #include "OpenCL/common.c"
21 #include "include/rp_kernel.h"
22 #include "OpenCL/rp.c"
23 #include "OpenCL/simd.c"
25 static void md5_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4])
49 MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
50 MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
51 MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02);
52 MD5_STEP (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03);
53 MD5_STEP (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00);
54 MD5_STEP (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01);
55 MD5_STEP (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02);
56 MD5_STEP (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03);
57 MD5_STEP (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00);
58 MD5_STEP (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01);
59 MD5_STEP (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02);
60 MD5_STEP (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03);
61 MD5_STEP (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00);
62 MD5_STEP (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01);
63 MD5_STEP (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02);
64 MD5_STEP (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03);
66 MD5_STEP (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10);
67 MD5_STEP (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11);
68 MD5_STEP (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12);
69 MD5_STEP (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13);
70 MD5_STEP (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10);
71 MD5_STEP (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11);
72 MD5_STEP (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12);
73 MD5_STEP (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13);
74 MD5_STEP (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10);
75 MD5_STEP (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11);
76 MD5_STEP (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12);
77 MD5_STEP (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13);
78 MD5_STEP (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10);
79 MD5_STEP (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11);
80 MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12);
81 MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13);
83 MD5_STEP (MD5_H , a, b, c, d, w5_t, MD5C20, MD5S20);
84 MD5_STEP (MD5_H , d, a, b, c, w8_t, MD5C21, MD5S21);
85 MD5_STEP (MD5_H , c, d, a, b, wb_t, MD5C22, MD5S22);
86 MD5_STEP (MD5_H , b, c, d, a, we_t, MD5C23, MD5S23);
87 MD5_STEP (MD5_H , a, b, c, d, w1_t, MD5C24, MD5S20);
88 MD5_STEP (MD5_H , d, a, b, c, w4_t, MD5C25, MD5S21);
89 MD5_STEP (MD5_H , c, d, a, b, w7_t, MD5C26, MD5S22);
90 MD5_STEP (MD5_H , b, c, d, a, wa_t, MD5C27, MD5S23);
91 MD5_STEP (MD5_H , a, b, c, d, wd_t, MD5C28, MD5S20);
92 MD5_STEP (MD5_H , d, a, b, c, w0_t, MD5C29, MD5S21);
93 MD5_STEP (MD5_H , c, d, a, b, w3_t, MD5C2a, MD5S22);
94 MD5_STEP (MD5_H , b, c, d, a, w6_t, MD5C2b, MD5S23);
95 MD5_STEP (MD5_H , a, b, c, d, w9_t, MD5C2c, MD5S20);
96 MD5_STEP (MD5_H , d, a, b, c, wc_t, MD5C2d, MD5S21);
97 MD5_STEP (MD5_H , c, d, a, b, wf_t, MD5C2e, MD5S22);
98 MD5_STEP (MD5_H , b, c, d, a, w2_t, MD5C2f, MD5S23);
100 MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30);
101 MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31);
102 MD5_STEP (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32);
103 MD5_STEP (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33);
104 MD5_STEP (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30);
105 MD5_STEP (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31);
106 MD5_STEP (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32);
107 MD5_STEP (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33);
108 MD5_STEP (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30);
109 MD5_STEP (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31);
110 MD5_STEP (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32);
111 MD5_STEP (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33);
112 MD5_STEP (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30);
113 MD5_STEP (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31);
114 MD5_STEP (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32);
115 MD5_STEP (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33);
123 static void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
125 w0[0] = w0[0] ^ 0x36363636;
126 w0[1] = w0[1] ^ 0x36363636;
127 w0[2] = w0[2] ^ 0x36363636;
128 w0[3] = w0[3] ^ 0x36363636;
129 w1[0] = w1[0] ^ 0x36363636;
130 w1[1] = w1[1] ^ 0x36363636;
131 w1[2] = w1[2] ^ 0x36363636;
132 w1[3] = w1[3] ^ 0x36363636;
133 w2[0] = w2[0] ^ 0x36363636;
134 w2[1] = w2[1] ^ 0x36363636;
135 w2[2] = w2[2] ^ 0x36363636;
136 w2[3] = w2[3] ^ 0x36363636;
137 w3[0] = w3[0] ^ 0x36363636;
138 w3[1] = w3[1] ^ 0x36363636;
139 w3[2] = w3[2] ^ 0x36363636;
140 w3[3] = w3[3] ^ 0x36363636;
147 md5_transform (w0, w1, w2, w3, ipad);
149 w0[0] = w0[0] ^ 0x6a6a6a6a;
150 w0[1] = w0[1] ^ 0x6a6a6a6a;
151 w0[2] = w0[2] ^ 0x6a6a6a6a;
152 w0[3] = w0[3] ^ 0x6a6a6a6a;
153 w1[0] = w1[0] ^ 0x6a6a6a6a;
154 w1[1] = w1[1] ^ 0x6a6a6a6a;
155 w1[2] = w1[2] ^ 0x6a6a6a6a;
156 w1[3] = w1[3] ^ 0x6a6a6a6a;
157 w2[0] = w2[0] ^ 0x6a6a6a6a;
158 w2[1] = w2[1] ^ 0x6a6a6a6a;
159 w2[2] = w2[2] ^ 0x6a6a6a6a;
160 w2[3] = w2[3] ^ 0x6a6a6a6a;
161 w3[0] = w3[0] ^ 0x6a6a6a6a;
162 w3[1] = w3[1] ^ 0x6a6a6a6a;
163 w3[2] = w3[2] ^ 0x6a6a6a6a;
164 w3[3] = w3[3] ^ 0x6a6a6a6a;
171 md5_transform (w0, w1, w2, w3, opad);
174 static 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])
181 md5_transform (w0, w1, w2, w3, digest);
197 w3[2] = (64 + 16) * 8;
205 md5_transform (w0, w1, w2, w3, digest);
208 __kernel void m05300_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 ikepsk_t *ikepsk_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)
214 const u32 gid = get_global_id (0);
215 const u32 lid = get_local_id (0);
216 const u32 lsz = get_local_size (0);
224 for (u32 i = lid; i < 16; i += lsz)
226 w_s[i] = ikepsk_bufs[salt_pos].nr_buf[i];
229 barrier (CLK_LOCAL_MEM_FENCE);
231 __local u32 s_msg_buf[128];
233 for (u32 i = lid; i < 128; i += lsz)
235 s_msg_buf[i] = ikepsk_bufs[salt_pos].msg_buf[i];
238 barrier (CLK_LOCAL_MEM_FENCE);
240 if (gid >= gid_max) return;
246 const u32 nr_len = ikepsk_bufs[salt_pos].nr_len;
247 const u32 msg_len = ikepsk_bufs[salt_pos].msg_len;
255 pw_buf0[0] = pws[gid].i[ 0];
256 pw_buf0[1] = pws[gid].i[ 1];
257 pw_buf0[2] = pws[gid].i[ 2];
258 pw_buf0[3] = pws[gid].i[ 3];
262 pw_buf1[0] = pws[gid].i[ 4];
263 pw_buf1[1] = pws[gid].i[ 5];
264 pw_buf1[2] = pws[gid].i[ 6];
265 pw_buf1[3] = pws[gid].i[ 7];
267 const u32 pw_len = pws[gid].pw_len;
273 for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
280 apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
317 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
333 w3_t[2] = (64 + nr_len) * 8;
338 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
357 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
362 for (left = ikepsk_bufs[salt_pos].msg_len, off = 0; left >= 56; left -= 64, off += 16)
364 w0_t[0] = s_msg_buf[off + 0];
365 w0_t[1] = s_msg_buf[off + 1];
366 w0_t[2] = s_msg_buf[off + 2];
367 w0_t[3] = s_msg_buf[off + 3];
368 w1_t[0] = s_msg_buf[off + 4];
369 w1_t[1] = s_msg_buf[off + 5];
370 w1_t[2] = s_msg_buf[off + 6];
371 w1_t[3] = s_msg_buf[off + 7];
372 w2_t[0] = s_msg_buf[off + 8];
373 w2_t[1] = s_msg_buf[off + 9];
374 w2_t[2] = s_msg_buf[off + 10];
375 w2_t[3] = s_msg_buf[off + 11];
376 w3_t[0] = s_msg_buf[off + 12];
377 w3_t[1] = s_msg_buf[off + 13];
378 w3_t[2] = s_msg_buf[off + 14];
379 w3_t[3] = s_msg_buf[off + 15];
381 md5_transform (w0_t, w1_t, w2_t, w3_t, ipad);
384 w0_t[0] = s_msg_buf[off + 0];
385 w0_t[1] = s_msg_buf[off + 1];
386 w0_t[2] = s_msg_buf[off + 2];
387 w0_t[3] = s_msg_buf[off + 3];
388 w1_t[0] = s_msg_buf[off + 4];
389 w1_t[1] = s_msg_buf[off + 5];
390 w1_t[2] = s_msg_buf[off + 6];
391 w1_t[3] = s_msg_buf[off + 7];
392 w2_t[0] = s_msg_buf[off + 8];
393 w2_t[1] = s_msg_buf[off + 9];
394 w2_t[2] = s_msg_buf[off + 10];
395 w2_t[3] = s_msg_buf[off + 11];
396 w3_t[0] = s_msg_buf[off + 12];
397 w3_t[1] = s_msg_buf[off + 13];
398 w3_t[2] = (64 + msg_len) * 8;
401 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
403 COMPARE_M_SIMD (digest[0], digest[3], digest[2], digest[1]);
407 __kernel void m05300_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 ikepsk_t *ikepsk_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)
411 __kernel void m05300_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 ikepsk_t *ikepsk_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)
415 __kernel void m05300_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 ikepsk_t *ikepsk_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)
421 const u32 gid = get_global_id (0);
422 const u32 lid = get_local_id (0);
423 const u32 lsz = get_local_size (0);
431 for (u32 i = lid; i < 16; i += lsz)
433 w_s[i] = ikepsk_bufs[salt_pos].nr_buf[i];
436 barrier (CLK_LOCAL_MEM_FENCE);
438 __local u32 s_msg_buf[128];
440 for (u32 i = lid; i < 128; i += lsz)
442 s_msg_buf[i] = ikepsk_bufs[salt_pos].msg_buf[i];
445 barrier (CLK_LOCAL_MEM_FENCE);
447 if (gid >= gid_max) return;
453 const u32 nr_len = ikepsk_bufs[salt_pos].nr_len;
454 const u32 msg_len = ikepsk_bufs[salt_pos].msg_len;
462 pw_buf0[0] = pws[gid].i[ 0];
463 pw_buf0[1] = pws[gid].i[ 1];
464 pw_buf0[2] = pws[gid].i[ 2];
465 pw_buf0[3] = pws[gid].i[ 3];
469 pw_buf1[0] = pws[gid].i[ 4];
470 pw_buf1[1] = pws[gid].i[ 5];
471 pw_buf1[2] = pws[gid].i[ 6];
472 pw_buf1[3] = pws[gid].i[ 7];
474 const u32 pw_len = pws[gid].pw_len;
480 const u32 search[4] =
482 digests_buf[digests_offset].digest_buf[DGST_R0],
483 digests_buf[digests_offset].digest_buf[DGST_R1],
484 digests_buf[digests_offset].digest_buf[DGST_R2],
485 digests_buf[digests_offset].digest_buf[DGST_R3]
492 for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
499 apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
536 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
552 w3_t[2] = (64 + nr_len) * 8;
557 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
576 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
581 for (left = ikepsk_bufs[salt_pos].msg_len, off = 0; left >= 56; left -= 64, off += 16)
583 w0_t[0] = s_msg_buf[off + 0];
584 w0_t[1] = s_msg_buf[off + 1];
585 w0_t[2] = s_msg_buf[off + 2];
586 w0_t[3] = s_msg_buf[off + 3];
587 w1_t[0] = s_msg_buf[off + 4];
588 w1_t[1] = s_msg_buf[off + 5];
589 w1_t[2] = s_msg_buf[off + 6];
590 w1_t[3] = s_msg_buf[off + 7];
591 w2_t[0] = s_msg_buf[off + 8];
592 w2_t[1] = s_msg_buf[off + 9];
593 w2_t[2] = s_msg_buf[off + 10];
594 w2_t[3] = s_msg_buf[off + 11];
595 w3_t[0] = s_msg_buf[off + 12];
596 w3_t[1] = s_msg_buf[off + 13];
597 w3_t[2] = s_msg_buf[off + 14];
598 w3_t[3] = s_msg_buf[off + 15];
600 md5_transform (w0_t, w1_t, w2_t, w3_t, ipad);
603 w0_t[0] = s_msg_buf[off + 0];
604 w0_t[1] = s_msg_buf[off + 1];
605 w0_t[2] = s_msg_buf[off + 2];
606 w0_t[3] = s_msg_buf[off + 3];
607 w1_t[0] = s_msg_buf[off + 4];
608 w1_t[1] = s_msg_buf[off + 5];
609 w1_t[2] = s_msg_buf[off + 6];
610 w1_t[3] = s_msg_buf[off + 7];
611 w2_t[0] = s_msg_buf[off + 8];
612 w2_t[1] = s_msg_buf[off + 9];
613 w2_t[2] = s_msg_buf[off + 10];
614 w2_t[3] = s_msg_buf[off + 11];
615 w3_t[0] = s_msg_buf[off + 12];
616 w3_t[1] = s_msg_buf[off + 13];
617 w3_t[2] = (64 + msg_len) * 8;
620 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
622 COMPARE_S_SIMD (digest[0], digest[3], digest[2], digest[1]);
626 __kernel void m05300_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 ikepsk_t *ikepsk_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)
630 __kernel void m05300_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 ikepsk_t *ikepsk_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)