2 * Author......: Jens Steube <jens.steube@gmail.com>
8 #include "include/constants.h"
9 #include "include/kernel_vendor.h"
10 #include "types_amd.c"
12 static void switch_buffer_by_offset (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const u32 offset)
14 const int offset_mod_4 = offset & 3;
16 const int offset_minus_4 = 4 - offset;
21 w3[2] = amd_bytealign ( 0, w3[1], offset_minus_4);
22 w3[1] = amd_bytealign (w3[1], w3[0], offset_minus_4);
23 w3[0] = amd_bytealign (w3[0], w2[3], offset_minus_4);
24 w2[3] = amd_bytealign (w2[3], w2[2], offset_minus_4);
25 w2[2] = amd_bytealign (w2[2], w2[1], offset_minus_4);
26 w2[1] = amd_bytealign (w2[1], w2[0], offset_minus_4);
27 w2[0] = amd_bytealign (w2[0], w1[3], offset_minus_4);
28 w1[3] = amd_bytealign (w1[3], w1[2], offset_minus_4);
29 w1[2] = amd_bytealign (w1[2], w1[1], offset_minus_4);
30 w1[1] = amd_bytealign (w1[1], w1[0], offset_minus_4);
31 w1[0] = amd_bytealign (w1[0], w0[3], offset_minus_4);
32 w0[3] = amd_bytealign (w0[3], w0[2], offset_minus_4);
33 w0[2] = amd_bytealign (w0[2], w0[1], offset_minus_4);
34 w0[1] = amd_bytealign (w0[1], w0[0], offset_minus_4);
35 w0[0] = amd_bytealign (w0[0], 0, offset_minus_4);
37 if (offset_mod_4 == 0)
59 w3[2] = amd_bytealign ( 0, w3[0], offset_minus_4);
60 w3[1] = amd_bytealign (w3[0], w2[3], offset_minus_4);
61 w3[0] = amd_bytealign (w2[3], w2[2], offset_minus_4);
62 w2[3] = amd_bytealign (w2[2], w2[1], offset_minus_4);
63 w2[2] = amd_bytealign (w2[1], w2[0], offset_minus_4);
64 w2[1] = amd_bytealign (w2[0], w1[3], offset_minus_4);
65 w2[0] = amd_bytealign (w1[3], w1[2], offset_minus_4);
66 w1[3] = amd_bytealign (w1[2], w1[1], offset_minus_4);
67 w1[2] = amd_bytealign (w1[1], w1[0], offset_minus_4);
68 w1[1] = amd_bytealign (w1[0], w0[3], offset_minus_4);
69 w1[0] = amd_bytealign (w0[3], w0[2], offset_minus_4);
70 w0[3] = amd_bytealign (w0[2], w0[1], offset_minus_4);
71 w0[2] = amd_bytealign (w0[1], w0[0], offset_minus_4);
72 w0[1] = amd_bytealign (w0[0], 0, offset_minus_4);
75 if (offset_mod_4 == 0)
96 w3[2] = amd_bytealign ( 0, w2[3], offset_minus_4);
97 w3[1] = amd_bytealign (w2[3], w2[2], offset_minus_4);
98 w3[0] = amd_bytealign (w2[2], w2[1], offset_minus_4);
99 w2[3] = amd_bytealign (w2[1], w2[0], offset_minus_4);
100 w2[2] = amd_bytealign (w2[0], w1[3], offset_minus_4);
101 w2[1] = amd_bytealign (w1[3], w1[2], offset_minus_4);
102 w2[0] = amd_bytealign (w1[2], w1[1], offset_minus_4);
103 w1[3] = amd_bytealign (w1[1], w1[0], offset_minus_4);
104 w1[2] = amd_bytealign (w1[0], w0[3], offset_minus_4);
105 w1[1] = amd_bytealign (w0[3], w0[2], offset_minus_4);
106 w1[0] = amd_bytealign (w0[2], w0[1], offset_minus_4);
107 w0[3] = amd_bytealign (w0[1], w0[0], offset_minus_4);
108 w0[2] = amd_bytealign (w0[0], 0, offset_minus_4);
112 if (offset_mod_4 == 0)
132 w3[2] = amd_bytealign ( 0, w2[2], offset_minus_4);
133 w3[1] = amd_bytealign (w2[2], w2[1], offset_minus_4);
134 w3[0] = amd_bytealign (w2[1], w2[0], offset_minus_4);
135 w2[3] = amd_bytealign (w2[0], w1[3], offset_minus_4);
136 w2[2] = amd_bytealign (w1[3], w1[2], offset_minus_4);
137 w2[1] = amd_bytealign (w1[2], w1[1], offset_minus_4);
138 w2[0] = amd_bytealign (w1[1], w1[0], offset_minus_4);
139 w1[3] = amd_bytealign (w1[0], w0[3], offset_minus_4);
140 w1[2] = amd_bytealign (w0[3], w0[2], offset_minus_4);
141 w1[1] = amd_bytealign (w0[2], w0[1], offset_minus_4);
142 w1[0] = amd_bytealign (w0[1], w0[0], offset_minus_4);
143 w0[3] = amd_bytealign (w0[0], 0, offset_minus_4);
148 if (offset_mod_4 == 0)
167 w3[2] = amd_bytealign ( 0, w2[1], offset_minus_4);
168 w3[1] = amd_bytealign (w2[1], w2[0], offset_minus_4);
169 w3[0] = amd_bytealign (w2[0], w1[3], offset_minus_4);
170 w2[3] = amd_bytealign (w1[3], w1[2], offset_minus_4);
171 w2[2] = amd_bytealign (w1[2], w1[1], offset_minus_4);
172 w2[1] = amd_bytealign (w1[1], w1[0], offset_minus_4);
173 w2[0] = amd_bytealign (w1[0], w0[3], offset_minus_4);
174 w1[3] = amd_bytealign (w0[3], w0[2], offset_minus_4);
175 w1[2] = amd_bytealign (w0[2], w0[1], offset_minus_4);
176 w1[1] = amd_bytealign (w0[1], w0[0], offset_minus_4);
177 w1[0] = amd_bytealign (w0[0], 0, offset_minus_4);
183 if (offset_mod_4 == 0)
201 w3[2] = amd_bytealign ( 0, w2[0], offset_minus_4);
202 w3[1] = amd_bytealign (w2[0], w1[3], offset_minus_4);
203 w3[0] = amd_bytealign (w1[3], w1[2], offset_minus_4);
204 w2[3] = amd_bytealign (w1[2], w1[1], offset_minus_4);
205 w2[2] = amd_bytealign (w1[1], w1[0], offset_minus_4);
206 w2[1] = amd_bytealign (w1[0], w0[3], offset_minus_4);
207 w2[0] = amd_bytealign (w0[3], w0[2], offset_minus_4);
208 w1[3] = amd_bytealign (w0[2], w0[1], offset_minus_4);
209 w1[2] = amd_bytealign (w0[1], w0[0], offset_minus_4);
210 w1[1] = amd_bytealign (w0[0], 0, offset_minus_4);
217 if (offset_mod_4 == 0)
234 w3[2] = amd_bytealign ( 0, w1[3], offset_minus_4);
235 w3[1] = amd_bytealign (w1[3], w1[2], offset_minus_4);
236 w3[0] = amd_bytealign (w1[2], w1[1], offset_minus_4);
237 w2[3] = amd_bytealign (w1[1], w1[0], offset_minus_4);
238 w2[2] = amd_bytealign (w1[0], w0[3], offset_minus_4);
239 w2[1] = amd_bytealign (w0[3], w0[2], offset_minus_4);
240 w2[0] = amd_bytealign (w0[2], w0[1], offset_minus_4);
241 w1[3] = amd_bytealign (w0[1], w0[0], offset_minus_4);
242 w1[2] = amd_bytealign (w0[0], 0, offset_minus_4);
250 if (offset_mod_4 == 0)
266 w3[2] = amd_bytealign ( 0, w1[2], offset_minus_4);
267 w3[1] = amd_bytealign (w1[2], w1[1], offset_minus_4);
268 w3[0] = amd_bytealign (w1[1], w1[0], offset_minus_4);
269 w2[3] = amd_bytealign (w1[0], w0[3], offset_minus_4);
270 w2[2] = amd_bytealign (w0[3], w0[2], offset_minus_4);
271 w2[1] = amd_bytealign (w0[2], w0[1], offset_minus_4);
272 w2[0] = amd_bytealign (w0[1], w0[0], offset_minus_4);
273 w1[3] = amd_bytealign (w0[0], 0, offset_minus_4);
282 if (offset_mod_4 == 0)
297 w3[2] = amd_bytealign ( 0, w1[1], offset_minus_4);
298 w3[1] = amd_bytealign (w1[1], w1[0], offset_minus_4);
299 w3[0] = amd_bytealign (w1[0], w0[3], offset_minus_4);
300 w2[3] = amd_bytealign (w0[3], w0[2], offset_minus_4);
301 w2[2] = amd_bytealign (w0[2], w0[1], offset_minus_4);
302 w2[1] = amd_bytealign (w0[1], w0[0], offset_minus_4);
303 w2[0] = amd_bytealign (w0[0], 0, offset_minus_4);
313 if (offset_mod_4 == 0)
327 w3[2] = amd_bytealign ( 0, w1[0], offset_minus_4);
328 w3[1] = amd_bytealign (w1[0], w0[3], offset_minus_4);
329 w3[0] = amd_bytealign (w0[3], w0[2], offset_minus_4);
330 w2[3] = amd_bytealign (w0[2], w0[1], offset_minus_4);
331 w2[2] = amd_bytealign (w0[1], w0[0], offset_minus_4);
332 w2[1] = amd_bytealign (w0[0], 0, offset_minus_4);
343 if (offset_mod_4 == 0)
356 w3[2] = amd_bytealign ( 0, w0[3], offset_minus_4);
357 w3[1] = amd_bytealign (w0[3], w0[2], offset_minus_4);
358 w3[0] = amd_bytealign (w0[2], w0[1], offset_minus_4);
359 w2[3] = amd_bytealign (w0[1], w0[0], offset_minus_4);
360 w2[2] = amd_bytealign (w0[0], 0, offset_minus_4);
372 if (offset_mod_4 == 0)
384 w3[2] = amd_bytealign ( 0, w0[2], offset_minus_4);
385 w3[1] = amd_bytealign (w0[2], w0[1], offset_minus_4);
386 w3[0] = amd_bytealign (w0[1], w0[0], offset_minus_4);
387 w2[3] = amd_bytealign (w0[0], 0, offset_minus_4);
400 if (offset_mod_4 == 0)
411 w3[2] = amd_bytealign ( 0, w0[1], offset_minus_4);
412 w3[1] = amd_bytealign (w0[1], w0[0], offset_minus_4);
413 w3[0] = amd_bytealign (w0[0], 0, offset_minus_4);
427 if (offset_mod_4 == 0)
437 w3[2] = amd_bytealign ( 0, w0[0], offset_minus_4);
438 w3[1] = amd_bytealign (w0[0], 0, offset_minus_4);
453 if (offset_mod_4 == 0)
463 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) amp (__global pw_t *pws, __global pw_t *pws_amp, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, const u32 combs_mode, const u32 gid_max)
465 const u32 gid = get_global_id (0);
467 if (gid >= gid_max) return;
469 const u32 pw_l_len = pws[gid].pw_len;
473 wordl0[0] = pws[gid].i[ 0];
474 wordl0[1] = pws[gid].i[ 1];
475 wordl0[2] = pws[gid].i[ 2];
476 wordl0[3] = pws[gid].i[ 3];
480 wordl1[0] = pws[gid].i[ 4];
481 wordl1[1] = pws[gid].i[ 5];
482 wordl1[2] = pws[gid].i[ 6];
483 wordl1[3] = pws[gid].i[ 7];
499 const u32 pw_r_len = combs_buf[0].pw_len;
503 wordr0[0] = combs_buf[0].i[0];
504 wordr0[1] = combs_buf[0].i[1];
505 wordr0[2] = combs_buf[0].i[2];
506 wordr0[3] = combs_buf[0].i[3];
510 wordr1[0] = combs_buf[0].i[4];
511 wordr1[1] = combs_buf[0].i[5];
512 wordr1[2] = combs_buf[0].i[6];
513 wordr1[3] = combs_buf[0].i[7];
529 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
531 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
534 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
536 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, pw_r_len);
541 w0[0] = wordl0[0] | wordr0[0];
542 w0[1] = wordl0[1] | wordr0[1];
543 w0[2] = wordl0[2] | wordr0[2];
544 w0[3] = wordl0[3] | wordr0[3];
548 w1[0] = wordl1[0] | wordr1[0];
549 w1[1] = wordl1[1] | wordr1[1];
550 w1[2] = wordl1[2] | wordr1[2];
551 w1[3] = wordl1[3] | wordr1[3];
555 w2[0] = wordl2[0] | wordr2[0];
556 w2[1] = wordl2[1] | wordr2[1];
557 w2[2] = wordl2[2] | wordr2[2];
558 w2[3] = wordl2[3] | wordr2[3];
562 w3[0] = wordl3[0] | wordr3[0];
563 w3[1] = wordl3[1] | wordr3[1];
564 w3[2] = wordl3[2] | wordr3[2];
565 w3[3] = wordl3[3] | wordr3[3];
567 const u32 pw_len = pw_l_len + pw_r_len;
569 pws_amp[gid].i[ 0] = w0[0];
570 pws_amp[gid].i[ 1] = w0[1];
571 pws_amp[gid].i[ 2] = w0[2];
572 pws_amp[gid].i[ 3] = w0[3];
573 pws_amp[gid].i[ 4] = w1[0];
574 pws_amp[gid].i[ 5] = w1[1];
575 pws_amp[gid].i[ 6] = w1[2];
576 pws_amp[gid].i[ 7] = w1[3];
577 pws_amp[gid].i[ 8] = w2[0];
578 pws_amp[gid].i[ 9] = w2[1];
579 pws_amp[gid].i[10] = w2[2];
580 pws_amp[gid].i[11] = w2[3];
581 pws_amp[gid].i[12] = w3[0];
582 pws_amp[gid].i[13] = w3[1];
583 pws_amp[gid].i[14] = w3[2];
584 pws_amp[gid].i[15] = w3[3];
586 pws_amp[gid].pw_len = pw_len;