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 u32x swap_workaround (const u32x v)
14 return (v << 24) + ((v & 0x0000FF00) << 8) + ((v & 0x00FF0000) >> 8) + (v >> 24);
17 #include "include/rp_gpu.h"
20 __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)
22 const u32 gid = get_global_id (0);
24 if (gid >= gid_max) return;
26 const u32 pw_len = pws[gid].pw_len;
31 w0[0] = pws[gid].i[ 0];
32 w0[1] = pws[gid].i[ 1];
33 w0[2] = pws[gid].i[ 2];
34 w0[3] = pws[gid].i[ 3];
35 w1[0] = pws[gid].i[ 4];
36 w1[1] = pws[gid].i[ 5];
37 w1[2] = pws[gid].i[ 6];
38 w1[3] = pws[gid].i[ 7];
40 const u32 out_len = apply_rules (rules_buf[0].cmds, w0, w1, pw_len);
42 pws_amp[gid].i[0] = w0[0];
43 pws_amp[gid].i[1] = w0[1];
44 pws_amp[gid].i[2] = w0[2];
45 pws_amp[gid].i[3] = w0[3];
46 pws_amp[gid].i[4] = w1[0];
47 pws_amp[gid].i[5] = w1[1];
48 pws_amp[gid].i[6] = w1[2];
49 pws_amp[gid].i[7] = w1[3];
51 pws_amp[gid].pw_len = out_len;