Test/spv.subgroupBallot.comp

   1 #version 450\r
   2 \r
   3 #extension GL_KHR_shader_subgroup_ballot: enable\r
   4 \r
   5 layout (local_size_x = 8, local_size_y = 8, local_size_z = 1) in;\r
   6 \r
   7 layout(binding = 0) buffer Buffers\r
   8 {\r
   9     vec4  f4;\r
  10     ivec4 i4;\r
  11     uvec4 u4;\r
  12     dvec4 d4;\r
  13 } data[4];\r
  14 \r
  15 void main()\r
  16 {\r
  17     uint invocation = (gl_SubgroupInvocationID + gl_SubgroupSize) % 4;\r
  18 \r
  19     uvec4 relMask = gl_SubgroupEqMask +\r
  20                        gl_SubgroupGeMask +\r
  21                        gl_SubgroupGtMask +\r
  22                        gl_SubgroupLeMask +\r
  23                        gl_SubgroupLtMask;\r
  24 \r
  25     uvec4 result = subgroupBallot(true);\r
  26 \r
  27     data[invocation].u4.x = subgroupBallotBitCount(result);\r
  28     data[invocation].u4.y = subgroupBallotBitExtract(result, 0) ? 1 : 0;\r
  29     data[invocation].u4.z = subgroupBallotInclusiveBitCount(result) + subgroupBallotExclusiveBitCount(result);\r
  30     data[invocation].u4.w = subgroupBallotFindLSB(result) + subgroupBallotFindMSB(result);\r
  31 \r
  32     if ((relMask == result) && subgroupInverseBallot(data[0].u4))\r
  33     {\r
  34         data[invocation].f4.x   = subgroupBroadcast(data[0].f4.x,    3);\r
  35         data[invocation].f4.xy  = subgroupBroadcast(data[1].f4.xy,   3);\r
  36         data[invocation].f4.xyz = subgroupBroadcast(data[2].f4.xyz,  3);\r
  37         data[invocation].f4     = subgroupBroadcast(data[3].f4,      3);\r
  38 \r
  39         data[invocation].i4.x   = subgroupBroadcast(data[0].i4.x,    2);\r
  40         data[invocation].i4.xy  = subgroupBroadcast(data[1].i4.xy,   2);\r
  41         data[invocation].i4.xyz = subgroupBroadcast(data[2].i4.xyz,  2);\r
  42         data[invocation].i4     = subgroupBroadcast(data[3].i4,      2);\r
  43 \r
  44         data[invocation].u4.x   = subgroupBroadcast(data[0].u4.x,    1);\r
  45         data[invocation].u4.xy  = subgroupBroadcast(data[1].u4.xy,   1);\r
  46         data[invocation].u4.xyz = subgroupBroadcast(data[2].u4.xyz,  1);\r
  47         data[invocation].u4     = subgroupBroadcast(data[3].u4,      1);\r
  48 \r
  49         data[invocation].d4.x   = subgroupBroadcast(data[0].d4.x,    0);\r
  50         data[invocation].d4.xy  = subgroupBroadcast(data[1].d4.xy,   0);\r
  51         data[invocation].d4.xyz = subgroupBroadcast(data[2].d4.xyz,  0);\r
  52         data[invocation].d4     = subgroupBroadcast(data[3].d4,      0);\r
  53 \r
  54         data[invocation].i4.x   = int(subgroupBroadcast(data[0].i4.x < 0,            1));\r
  55         data[invocation].i4.xy  = ivec2(subgroupBroadcast(lessThan(data[1].i4.xy, ivec2(0)), 1));\r
  56         data[invocation].i4.xyz = ivec3(subgroupBroadcast(lessThan(data[1].i4.xyz, ivec3(0)), 1));\r
  57         data[invocation].i4     = ivec4(subgroupBroadcast(lessThan(data[1].i4, ivec4(0)), 1));\r
  58     }\r
  59     else\r
  60     {\r
  61         data[invocation].f4.x   = subgroupBroadcastFirst(data[0].f4.x);\r
  62         data[invocation].f4.xy  = subgroupBroadcastFirst(data[1].f4.xy);\r
  63         data[invocation].f4.xyz = subgroupBroadcastFirst(data[2].f4.xyz);\r
  64         data[invocation].f4     = subgroupBroadcastFirst(data[3].f4);\r
  65 \r
  66         data[invocation].i4.x   = subgroupBroadcastFirst(data[0].i4.x);\r
  67         data[invocation].i4.xy  = subgroupBroadcastFirst(data[1].i4.xy);\r
  68         data[invocation].i4.xyz = subgroupBroadcastFirst(data[2].i4.xyz);\r
  69         data[invocation].i4     = subgroupBroadcastFirst(data[3].i4);\r
  70 \r
  71         data[invocation].u4.x   = subgroupBroadcastFirst(data[0].u4.x);\r
  72         data[invocation].u4.xy  = subgroupBroadcastFirst(data[1].u4.xy);\r
  73         data[invocation].u4.xyz = subgroupBroadcastFirst(data[2].u4.xyz);\r
  74         data[invocation].u4     = subgroupBroadcastFirst(data[3].u4);\r
  75 \r
  76         data[invocation].d4.x   = subgroupBroadcastFirst(data[0].d4.x);\r
  77         data[invocation].d4.xy  = subgroupBroadcastFirst(data[1].d4.xy);\r
  78         data[invocation].d4.xyz = subgroupBroadcastFirst(data[2].d4.xyz);\r
  79         data[invocation].d4     = subgroupBroadcastFirst(data[3].d4);\r
  80 \r
  81         data[invocation].i4.x   = int(subgroupBroadcastFirst(data[0].i4.x < 0));\r
  82         data[invocation].i4.xy  = ivec2(subgroupBroadcastFirst(lessThan(data[1].i4.xy, ivec2(0))));\r
  83         data[invocation].i4.xyz = ivec3(subgroupBroadcastFirst(lessThan(data[1].i4.xyz, ivec3(0))));\r
  84         data[invocation].i4     = ivec4(subgroupBroadcastFirst(lessThan(data[1].i4, ivec4(0))));\r
  85     }\r
  86 }\r