2 * Author......: Jens Steube <jens.steube@gmail.com>
6 #include "inc_hash_constants.h"
7 #include "inc_vendor.cl"
8 #include "inc_types.cl"
10 inline void switch_buffer_by_offset_le (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 offset)
12 #if defined IS_AMD || defined IS_GENERIC
13 const int offset_mod_4 = offset & 3;
15 const int offset_minus_4 = 4 - offset;
20 w3[2] = amd_bytealign ( 0, w3[1], offset_minus_4);
21 w3[1] = amd_bytealign (w3[1], w3[0], offset_minus_4);
22 w3[0] = amd_bytealign (w3[0], w2[3], offset_minus_4);
23 w2[3] = amd_bytealign (w2[3], w2[2], offset_minus_4);
24 w2[2] = amd_bytealign (w2[2], w2[1], offset_minus_4);
25 w2[1] = amd_bytealign (w2[1], w2[0], offset_minus_4);
26 w2[0] = amd_bytealign (w2[0], w1[3], offset_minus_4);
27 w1[3] = amd_bytealign (w1[3], w1[2], offset_minus_4);
28 w1[2] = amd_bytealign (w1[2], w1[1], offset_minus_4);
29 w1[1] = amd_bytealign (w1[1], w1[0], offset_minus_4);
30 w1[0] = amd_bytealign (w1[0], w0[3], offset_minus_4);
31 w0[3] = amd_bytealign (w0[3], w0[2], offset_minus_4);
32 w0[2] = amd_bytealign (w0[2], w0[1], offset_minus_4);
33 w0[1] = amd_bytealign (w0[1], w0[0], offset_minus_4);
34 w0[0] = amd_bytealign (w0[0], 0, offset_minus_4);
36 if (offset_mod_4 == 0)
58 w3[2] = amd_bytealign ( 0, w3[0], offset_minus_4);
59 w3[1] = amd_bytealign (w3[0], w2[3], offset_minus_4);
60 w3[0] = amd_bytealign (w2[3], w2[2], offset_minus_4);
61 w2[3] = amd_bytealign (w2[2], w2[1], offset_minus_4);
62 w2[2] = amd_bytealign (w2[1], w2[0], offset_minus_4);
63 w2[1] = amd_bytealign (w2[0], w1[3], offset_minus_4);
64 w2[0] = amd_bytealign (w1[3], w1[2], offset_minus_4);
65 w1[3] = amd_bytealign (w1[2], w1[1], offset_minus_4);
66 w1[2] = amd_bytealign (w1[1], w1[0], offset_minus_4);
67 w1[1] = amd_bytealign (w1[0], w0[3], offset_minus_4);
68 w1[0] = amd_bytealign (w0[3], w0[2], offset_minus_4);
69 w0[3] = amd_bytealign (w0[2], w0[1], offset_minus_4);
70 w0[2] = amd_bytealign (w0[1], w0[0], offset_minus_4);
71 w0[1] = amd_bytealign (w0[0], 0, offset_minus_4);
74 if (offset_mod_4 == 0)
95 w3[2] = amd_bytealign ( 0, w2[3], offset_minus_4);
96 w3[1] = amd_bytealign (w2[3], w2[2], offset_minus_4);
97 w3[0] = amd_bytealign (w2[2], w2[1], offset_minus_4);
98 w2[3] = amd_bytealign (w2[1], w2[0], offset_minus_4);
99 w2[2] = amd_bytealign (w2[0], w1[3], offset_minus_4);
100 w2[1] = amd_bytealign (w1[3], w1[2], offset_minus_4);
101 w2[0] = amd_bytealign (w1[2], w1[1], offset_minus_4);
102 w1[3] = amd_bytealign (w1[1], w1[0], offset_minus_4);
103 w1[2] = amd_bytealign (w1[0], w0[3], offset_minus_4);
104 w1[1] = amd_bytealign (w0[3], w0[2], offset_minus_4);
105 w1[0] = amd_bytealign (w0[2], w0[1], offset_minus_4);
106 w0[3] = amd_bytealign (w0[1], w0[0], offset_minus_4);
107 w0[2] = amd_bytealign (w0[0], 0, offset_minus_4);
111 if (offset_mod_4 == 0)
131 w3[2] = amd_bytealign ( 0, w2[2], offset_minus_4);
132 w3[1] = amd_bytealign (w2[2], w2[1], offset_minus_4);
133 w3[0] = amd_bytealign (w2[1], w2[0], offset_minus_4);
134 w2[3] = amd_bytealign (w2[0], w1[3], offset_minus_4);
135 w2[2] = amd_bytealign (w1[3], w1[2], offset_minus_4);
136 w2[1] = amd_bytealign (w1[2], w1[1], offset_minus_4);
137 w2[0] = amd_bytealign (w1[1], w1[0], offset_minus_4);
138 w1[3] = amd_bytealign (w1[0], w0[3], offset_minus_4);
139 w1[2] = amd_bytealign (w0[3], w0[2], offset_minus_4);
140 w1[1] = amd_bytealign (w0[2], w0[1], offset_minus_4);
141 w1[0] = amd_bytealign (w0[1], w0[0], offset_minus_4);
142 w0[3] = amd_bytealign (w0[0], 0, offset_minus_4);
147 if (offset_mod_4 == 0)
166 w3[2] = amd_bytealign ( 0, w2[1], offset_minus_4);
167 w3[1] = amd_bytealign (w2[1], w2[0], offset_minus_4);
168 w3[0] = amd_bytealign (w2[0], w1[3], offset_minus_4);
169 w2[3] = amd_bytealign (w1[3], w1[2], offset_minus_4);
170 w2[2] = amd_bytealign (w1[2], w1[1], offset_minus_4);
171 w2[1] = amd_bytealign (w1[1], w1[0], offset_minus_4);
172 w2[0] = amd_bytealign (w1[0], w0[3], offset_minus_4);
173 w1[3] = amd_bytealign (w0[3], w0[2], offset_minus_4);
174 w1[2] = amd_bytealign (w0[2], w0[1], offset_minus_4);
175 w1[1] = amd_bytealign (w0[1], w0[0], offset_minus_4);
176 w1[0] = amd_bytealign (w0[0], 0, offset_minus_4);
182 if (offset_mod_4 == 0)
200 w3[2] = amd_bytealign ( 0, w2[0], offset_minus_4);
201 w3[1] = amd_bytealign (w2[0], w1[3], offset_minus_4);
202 w3[0] = amd_bytealign (w1[3], w1[2], offset_minus_4);
203 w2[3] = amd_bytealign (w1[2], w1[1], offset_minus_4);
204 w2[2] = amd_bytealign (w1[1], w1[0], offset_minus_4);
205 w2[1] = amd_bytealign (w1[0], w0[3], offset_minus_4);
206 w2[0] = amd_bytealign (w0[3], w0[2], offset_minus_4);
207 w1[3] = amd_bytealign (w0[2], w0[1], offset_minus_4);
208 w1[2] = amd_bytealign (w0[1], w0[0], offset_minus_4);
209 w1[1] = amd_bytealign (w0[0], 0, offset_minus_4);
216 if (offset_mod_4 == 0)
233 w3[2] = amd_bytealign ( 0, w1[3], offset_minus_4);
234 w3[1] = amd_bytealign (w1[3], w1[2], offset_minus_4);
235 w3[0] = amd_bytealign (w1[2], w1[1], offset_minus_4);
236 w2[3] = amd_bytealign (w1[1], w1[0], offset_minus_4);
237 w2[2] = amd_bytealign (w1[0], w0[3], offset_minus_4);
238 w2[1] = amd_bytealign (w0[3], w0[2], offset_minus_4);
239 w2[0] = amd_bytealign (w0[2], w0[1], offset_minus_4);
240 w1[3] = amd_bytealign (w0[1], w0[0], offset_minus_4);
241 w1[2] = amd_bytealign (w0[0], 0, offset_minus_4);
249 if (offset_mod_4 == 0)
265 w3[2] = amd_bytealign ( 0, w1[2], offset_minus_4);
266 w3[1] = amd_bytealign (w1[2], w1[1], offset_minus_4);
267 w3[0] = amd_bytealign (w1[1], w1[0], offset_minus_4);
268 w2[3] = amd_bytealign (w1[0], w0[3], offset_minus_4);
269 w2[2] = amd_bytealign (w0[3], w0[2], offset_minus_4);
270 w2[1] = amd_bytealign (w0[2], w0[1], offset_minus_4);
271 w2[0] = amd_bytealign (w0[1], w0[0], offset_minus_4);
272 w1[3] = amd_bytealign (w0[0], 0, offset_minus_4);
281 if (offset_mod_4 == 0)
296 w3[2] = amd_bytealign ( 0, w1[1], offset_minus_4);
297 w3[1] = amd_bytealign (w1[1], w1[0], offset_minus_4);
298 w3[0] = amd_bytealign (w1[0], w0[3], offset_minus_4);
299 w2[3] = amd_bytealign (w0[3], w0[2], offset_minus_4);
300 w2[2] = amd_bytealign (w0[2], w0[1], offset_minus_4);
301 w2[1] = amd_bytealign (w0[1], w0[0], offset_minus_4);
302 w2[0] = amd_bytealign (w0[0], 0, offset_minus_4);
312 if (offset_mod_4 == 0)
326 w3[2] = amd_bytealign ( 0, w1[0], offset_minus_4);
327 w3[1] = amd_bytealign (w1[0], w0[3], offset_minus_4);
328 w3[0] = amd_bytealign (w0[3], w0[2], offset_minus_4);
329 w2[3] = amd_bytealign (w0[2], w0[1], offset_minus_4);
330 w2[2] = amd_bytealign (w0[1], w0[0], offset_minus_4);
331 w2[1] = amd_bytealign (w0[0], 0, offset_minus_4);
342 if (offset_mod_4 == 0)
355 w3[2] = amd_bytealign ( 0, w0[3], offset_minus_4);
356 w3[1] = amd_bytealign (w0[3], w0[2], offset_minus_4);
357 w3[0] = amd_bytealign (w0[2], w0[1], offset_minus_4);
358 w2[3] = amd_bytealign (w0[1], w0[0], offset_minus_4);
359 w2[2] = amd_bytealign (w0[0], 0, offset_minus_4);
371 if (offset_mod_4 == 0)
383 w3[2] = amd_bytealign ( 0, w0[2], offset_minus_4);
384 w3[1] = amd_bytealign (w0[2], w0[1], offset_minus_4);
385 w3[0] = amd_bytealign (w0[1], w0[0], offset_minus_4);
386 w2[3] = amd_bytealign (w0[0], 0, offset_minus_4);
399 if (offset_mod_4 == 0)
410 w3[2] = amd_bytealign ( 0, w0[1], offset_minus_4);
411 w3[1] = amd_bytealign (w0[1], w0[0], offset_minus_4);
412 w3[0] = amd_bytealign (w0[0], 0, offset_minus_4);
426 if (offset_mod_4 == 0)
436 w3[2] = amd_bytealign ( 0, w0[0], offset_minus_4);
437 w3[1] = amd_bytealign (w0[0], 0, offset_minus_4);
452 if (offset_mod_4 == 0)
463 const int offset_minus_4 = 4 - (offset % 4);
465 const int selector = (0x76543210 >> (offset_minus_4 * 4)) & 0xffff;
470 w3[1] = __byte_perm (w3[0], w3[1], selector);
471 w3[0] = __byte_perm (w2[3], w3[0], selector);
472 w2[3] = __byte_perm (w2[2], w2[3], selector);
473 w2[2] = __byte_perm (w2[1], w2[2], selector);
474 w2[1] = __byte_perm (w2[0], w2[1], selector);
475 w2[0] = __byte_perm (w1[3], w2[0], selector);
476 w1[3] = __byte_perm (w1[2], w1[3], selector);
477 w1[2] = __byte_perm (w1[1], w1[2], selector);
478 w1[1] = __byte_perm (w1[0], w1[1], selector);
479 w1[0] = __byte_perm (w0[3], w1[0], selector);
480 w0[3] = __byte_perm (w0[2], w0[3], selector);
481 w0[2] = __byte_perm (w0[1], w0[2], selector);
482 w0[1] = __byte_perm (w0[0], w0[1], selector);
483 w0[0] = __byte_perm ( 0, w0[0], selector);
488 w3[1] = __byte_perm (w2[3], w3[0], selector);
489 w3[0] = __byte_perm (w2[2], w2[3], selector);
490 w2[3] = __byte_perm (w2[1], w2[2], selector);
491 w2[2] = __byte_perm (w2[0], w2[1], selector);
492 w2[1] = __byte_perm (w1[3], w2[0], selector);
493 w2[0] = __byte_perm (w1[2], w1[3], selector);
494 w1[3] = __byte_perm (w1[1], w1[2], selector);
495 w1[2] = __byte_perm (w1[0], w1[1], selector);
496 w1[1] = __byte_perm (w0[3], w1[0], selector);
497 w1[0] = __byte_perm (w0[2], w0[3], selector);
498 w0[3] = __byte_perm (w0[1], w0[2], selector);
499 w0[2] = __byte_perm (w0[0], w0[1], selector);
500 w0[1] = __byte_perm ( 0, w0[0], selector);
506 w3[1] = __byte_perm (w2[2], w2[3], selector);
507 w3[0] = __byte_perm (w2[1], w2[2], selector);
508 w2[3] = __byte_perm (w2[0], w2[1], selector);
509 w2[2] = __byte_perm (w1[3], w2[0], selector);
510 w2[1] = __byte_perm (w1[2], w1[3], selector);
511 w2[0] = __byte_perm (w1[1], w1[2], selector);
512 w1[3] = __byte_perm (w1[0], w1[1], selector);
513 w1[2] = __byte_perm (w0[3], w1[0], selector);
514 w1[1] = __byte_perm (w0[2], w0[3], selector);
515 w1[0] = __byte_perm (w0[1], w0[2], selector);
516 w0[3] = __byte_perm (w0[0], w0[1], selector);
517 w0[2] = __byte_perm ( 0, w0[0], selector);
524 w3[1] = __byte_perm (w2[1], w2[2], selector);
525 w3[0] = __byte_perm (w2[0], w2[1], selector);
526 w2[3] = __byte_perm (w1[3], w2[0], selector);
527 w2[2] = __byte_perm (w1[2], w1[3], selector);
528 w2[1] = __byte_perm (w1[1], w1[2], selector);
529 w2[0] = __byte_perm (w1[0], w1[1], selector);
530 w1[3] = __byte_perm (w0[3], w1[0], selector);
531 w1[2] = __byte_perm (w0[2], w0[3], selector);
532 w1[1] = __byte_perm (w0[1], w0[2], selector);
533 w1[0] = __byte_perm (w0[0], w0[1], selector);
534 w0[3] = __byte_perm ( 0, w0[0], selector);
542 w3[1] = __byte_perm (w2[0], w2[1], selector);
543 w3[0] = __byte_perm (w1[3], w2[0], selector);
544 w2[3] = __byte_perm (w1[2], w1[3], selector);
545 w2[2] = __byte_perm (w1[1], w1[2], selector);
546 w2[1] = __byte_perm (w1[0], w1[1], selector);
547 w2[0] = __byte_perm (w0[3], w1[0], selector);
548 w1[3] = __byte_perm (w0[2], w0[3], selector);
549 w1[2] = __byte_perm (w0[1], w0[2], selector);
550 w1[1] = __byte_perm (w0[0], w0[1], selector);
551 w1[0] = __byte_perm ( 0, w0[0], selector);
560 w3[1] = __byte_perm (w1[3], w2[0], selector);
561 w3[0] = __byte_perm (w1[2], w1[3], selector);
562 w2[3] = __byte_perm (w1[1], w1[2], selector);
563 w2[2] = __byte_perm (w1[0], w1[1], selector);
564 w2[1] = __byte_perm (w0[3], w1[0], selector);
565 w2[0] = __byte_perm (w0[2], w0[3], selector);
566 w1[3] = __byte_perm (w0[1], w0[2], selector);
567 w1[2] = __byte_perm (w0[0], w0[1], selector);
568 w1[1] = __byte_perm ( 0, w0[0], selector);
578 w3[1] = __byte_perm (w1[2], w1[3], selector);
579 w3[0] = __byte_perm (w1[1], w1[2], selector);
580 w2[3] = __byte_perm (w1[0], w1[1], selector);
581 w2[2] = __byte_perm (w0[3], w1[0], selector);
582 w2[1] = __byte_perm (w0[2], w0[3], selector);
583 w2[0] = __byte_perm (w0[1], w0[2], selector);
584 w1[3] = __byte_perm (w0[0], w0[1], selector);
585 w1[2] = __byte_perm ( 0, w0[0], selector);
596 w3[1] = __byte_perm (w1[1], w1[2], selector);
597 w3[0] = __byte_perm (w1[0], w1[1], selector);
598 w2[3] = __byte_perm (w0[3], w1[0], selector);
599 w2[2] = __byte_perm (w0[2], w0[3], selector);
600 w2[1] = __byte_perm (w0[1], w0[2], selector);
601 w2[0] = __byte_perm (w0[0], w0[1], selector);
602 w1[3] = __byte_perm ( 0, w0[0], selector);
614 w3[1] = __byte_perm (w1[0], w1[1], selector);
615 w3[0] = __byte_perm (w0[3], w1[0], selector);
616 w2[3] = __byte_perm (w0[2], w0[3], selector);
617 w2[2] = __byte_perm (w0[1], w0[2], selector);
618 w2[1] = __byte_perm (w0[0], w0[1], selector);
619 w2[0] = __byte_perm ( 0, w0[0], selector);
632 w3[1] = __byte_perm (w0[3], w1[0], selector);
633 w3[0] = __byte_perm (w0[2], w0[3], selector);
634 w2[3] = __byte_perm (w0[1], w0[2], selector);
635 w2[2] = __byte_perm (w0[0], w0[1], selector);
636 w2[1] = __byte_perm ( 0, w0[0], selector);
650 w3[1] = __byte_perm (w0[2], w0[3], selector);
651 w3[0] = __byte_perm (w0[1], w0[2], selector);
652 w2[3] = __byte_perm (w0[0], w0[1], selector);
653 w2[2] = __byte_perm ( 0, w0[0], selector);
668 w3[1] = __byte_perm (w0[1], w0[2], selector);
669 w3[0] = __byte_perm (w0[0], w0[1], selector);
670 w2[3] = __byte_perm ( 0, w0[0], selector);
686 w3[1] = __byte_perm (w0[0], w0[1], selector);
687 w3[0] = __byte_perm ( 0, w0[0], selector);
704 w3[1] = __byte_perm ( 0, w0[0], selector);
724 __kernel void amp (__global pw_t *pws, __global pw_t *pws_amp, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, const u32 combs_mode, const u32 gid_max)
726 const u32 gid = get_global_id (0);
728 if (gid >= gid_max) return;
730 const u32 pw_l_len = pws[gid].pw_len;
734 wordl0[0] = pws[gid].i[ 0];
735 wordl0[1] = pws[gid].i[ 1];
736 wordl0[2] = pws[gid].i[ 2];
737 wordl0[3] = pws[gid].i[ 3];
741 wordl1[0] = pws[gid].i[ 4];
742 wordl1[1] = pws[gid].i[ 5];
743 wordl1[2] = pws[gid].i[ 6];
744 wordl1[3] = pws[gid].i[ 7];
760 const u32 pw_r_len = combs_buf[0].pw_len;
764 wordr0[0] = combs_buf[0].i[0];
765 wordr0[1] = combs_buf[0].i[1];
766 wordr0[2] = combs_buf[0].i[2];
767 wordr0[3] = combs_buf[0].i[3];
771 wordr1[0] = combs_buf[0].i[4];
772 wordr1[1] = combs_buf[0].i[5];
773 wordr1[2] = combs_buf[0].i[6];
774 wordr1[3] = combs_buf[0].i[7];
790 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
792 switch_buffer_by_offset_le (wordr0, wordr1, wordr2, wordr3, pw_l_len);
795 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
797 switch_buffer_by_offset_le (wordl0, wordl1, wordl2, wordl3, pw_r_len);
802 w0[0] = wordl0[0] | wordr0[0];
803 w0[1] = wordl0[1] | wordr0[1];
804 w0[2] = wordl0[2] | wordr0[2];
805 w0[3] = wordl0[3] | wordr0[3];
809 w1[0] = wordl1[0] | wordr1[0];
810 w1[1] = wordl1[1] | wordr1[1];
811 w1[2] = wordl1[2] | wordr1[2];
812 w1[3] = wordl1[3] | wordr1[3];
816 w2[0] = wordl2[0] | wordr2[0];
817 w2[1] = wordl2[1] | wordr2[1];
818 w2[2] = wordl2[2] | wordr2[2];
819 w2[3] = wordl2[3] | wordr2[3];
823 w3[0] = wordl3[0] | wordr3[0];
824 w3[1] = wordl3[1] | wordr3[1];
825 w3[2] = wordl3[2] | wordr3[2];
826 w3[3] = wordl3[3] | wordr3[3];
828 const u32 pw_len = pw_l_len + pw_r_len;
830 pws_amp[gid].i[ 0] = w0[0];
831 pws_amp[gid].i[ 1] = w0[1];
832 pws_amp[gid].i[ 2] = w0[2];
833 pws_amp[gid].i[ 3] = w0[3];
834 pws_amp[gid].i[ 4] = w1[0];
835 pws_amp[gid].i[ 5] = w1[1];
836 pws_amp[gid].i[ 6] = w1[2];
837 pws_amp[gid].i[ 7] = w1[3];
838 pws_amp[gid].i[ 8] = w2[0];
839 pws_amp[gid].i[ 9] = w2[1];
840 pws_amp[gid].i[10] = w2[2];
841 pws_amp[gid].i[11] = w2[3];
842 pws_amp[gid].i[12] = w3[0];
843 pws_amp[gid].i[13] = w3[1];
844 pws_amp[gid].i[14] = w3[2];
845 pws_amp[gid].i[15] = w3[3];
847 pws_amp[gid].pw_len = pw_len;