OpenCL Kernels: Use a special kernel to initialize the password buffer used during autotune measurements to reduce startup time

2018-01-23 20:33:26 +01:00
parent 7a8239b4c1
commit 7062425d2b
6 changed files with 148 additions and 8 deletions
@@ -61594,3 +61594,84 @@ __kernel void gpu_memset (__global uint4 *buf, const u32 value, const u64 gid_ma

  buf[gid] = (uint4) (value);
 }
+
+__kernel void gpu_atinit (__global pw_t *buf, const u64 gid_max)
+{
+  const u64 gid = get_global_id (0);
+
+  if (gid >= gid_max) return;
+
+  const u32 l32 = l32_from_64_S (gid);
+  const u32 h32 = h32_from_64_S (gid);
+
+  pw_t pw;
+
+  pw.i[ 0] = 0x5c5c5c5c ^ l32;
+  pw.i[ 1] = 0x36363636 ^ h32;
+  pw.i[ 2] = 0;
+  pw.i[ 3] = 0;
+  pw.i[ 4] = 0;
+  pw.i[ 5] = 0;
+  pw.i[ 6] = 0;
+  pw.i[ 7] = 0;
+  pw.i[ 8] = 0;
+  pw.i[ 9] = 0;
+  pw.i[10] = 0;
+  pw.i[11] = 0;
+  pw.i[12] = 0;
+  pw.i[13] = 0;
+  pw.i[14] = 0;
+  pw.i[15] = 0;
+  pw.i[16] = 0;
+  pw.i[17] = 0;
+  pw.i[18] = 0;
+  pw.i[19] = 0;
+  pw.i[20] = 0;
+  pw.i[21] = 0;
+  pw.i[22] = 0;
+  pw.i[23] = 0;
+  pw.i[24] = 0;
+  pw.i[25] = 0;
+  pw.i[26] = 0;
+  pw.i[27] = 0;
+  pw.i[28] = 0;
+  pw.i[29] = 0;
+  pw.i[30] = 0;
+  pw.i[31] = 0;
+  pw.i[32] = 0;
+  pw.i[33] = 0;
+  pw.i[34] = 0;
+  pw.i[35] = 0;
+  pw.i[36] = 0;
+  pw.i[37] = 0;
+  pw.i[38] = 0;
+  pw.i[39] = 0;
+  pw.i[40] = 0;
+  pw.i[41] = 0;
+  pw.i[42] = 0;
+  pw.i[43] = 0;
+  pw.i[44] = 0;
+  pw.i[45] = 0;
+  pw.i[46] = 0;
+  pw.i[47] = 0;
+  pw.i[48] = 0;
+  pw.i[49] = 0;
+  pw.i[50] = 0;
+  pw.i[51] = 0;
+  pw.i[52] = 0;
+  pw.i[53] = 0;
+  pw.i[54] = 0;
+  pw.i[55] = 0;
+  pw.i[56] = 0;
+  pw.i[57] = 0;
+  pw.i[58] = 0;
+  pw.i[59] = 0;
+  pw.i[60] = 0;
+  pw.i[61] = 0;
+  pw.i[62] = 0;
+  pw.i[63] = 0; // yep that's faster
+
+  pw.pw_len = 1 + (l32 & 15);
+
+  buf[gid] = pw;
+}