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 __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)
14 const u32 gid = get_global_id (0);
16 if (gid >= gid_max) return;
18 const u32 pw_len = pws[gid].pw_len;
25 w0[0] = pws[gid].i[ 0];
26 w0[1] = pws[gid].i[ 1];
27 w0[2] = pws[gid].i[ 2];
28 w0[3] = pws[gid].i[ 3];
29 w1[0] = pws[gid].i[ 4];
30 w1[1] = pws[gid].i[ 5];
31 w1[2] = pws[gid].i[ 6];
32 w1[3] = pws[gid].i[ 7];
33 w2[0] = pws[gid].i[ 8];
34 w2[1] = pws[gid].i[ 9];
35 w2[2] = pws[gid].i[10];
36 w2[3] = pws[gid].i[11];
37 w3[0] = pws[gid].i[12];
38 w3[1] = pws[gid].i[13];
39 w3[2] = pws[gid].i[14];
40 w3[3] = pws[gid].i[15];
42 const u32 w0r = bfs_buf[0].i;
44 pws_amp[gid].i[ 0] = w0[0] | w0r;
45 pws_amp[gid].i[ 1] = w0[1];
46 pws_amp[gid].i[ 2] = w0[2];
47 pws_amp[gid].i[ 3] = w0[3];
48 pws_amp[gid].i[ 4] = w1[0];
49 pws_amp[gid].i[ 5] = w1[1];
50 pws_amp[gid].i[ 6] = w1[2];
51 pws_amp[gid].i[ 7] = w1[3];
52 pws_amp[gid].i[ 8] = w2[0];
53 pws_amp[gid].i[ 9] = w2[1];
54 pws_amp[gid].i[10] = w2[2];
55 pws_amp[gid].i[11] = w2[3];
56 pws_amp[gid].i[12] = w3[0];
57 pws_amp[gid].i[13] = w3[1];
58 pws_amp[gid].i[14] = w3[2];
59 pws_amp[gid].i[15] = w3[3];
61 pws_amp[gid].pw_len = pw_len;