Test/spv.subgroupShuffle.comp

   1 #version 450\r
   2 \r
   3 #extension GL_KHR_shader_subgroup_shuffle: 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     data[invocation].f4.x   = subgroupShuffle(data[0].f4.x,    invocation);\r
  20     data[invocation].f4.xy  = subgroupShuffle(data[1].f4.xy,   invocation);\r
  21     data[invocation].f4.xyz = subgroupShuffle(data[2].f4.xyz,  invocation);\r
  22     data[invocation].f4     = subgroupShuffle(data[3].f4,      invocation);\r
  23 \r
  24     data[invocation].i4.x   = subgroupShuffle(data[0].i4.x,    invocation);\r
  25     data[invocation].i4.xy  = subgroupShuffle(data[1].i4.xy,   invocation);\r
  26     data[invocation].i4.xyz = subgroupShuffle(data[2].i4.xyz,  invocation);\r
  27     data[invocation].i4     = subgroupShuffle(data[3].i4,      invocation);\r
  28 \r
  29     data[invocation].u4.x   = subgroupShuffle(data[0].u4.x,    invocation);\r
  30     data[invocation].u4.xy  = subgroupShuffle(data[1].u4.xy,   invocation);\r
  31     data[invocation].u4.xyz = subgroupShuffle(data[2].u4.xyz,  invocation);\r
  32     data[invocation].u4     = subgroupShuffle(data[3].u4,      invocation);\r
  33 \r
  34     data[invocation].d4.x   = subgroupShuffle(data[0].d4.x,    invocation);\r
  35     data[invocation].d4.xy  = subgroupShuffle(data[1].d4.xy,   invocation);\r
  36     data[invocation].d4.xyz = subgroupShuffle(data[2].d4.xyz,  invocation);\r
  37     data[invocation].d4     = subgroupShuffle(data[3].d4,      invocation);\r
  38 \r
  39     data[invocation].i4.x   =   int(subgroupShuffle(data[0].i4.x < 0,                   invocation));\r
  40     data[invocation].i4.xy  = ivec2(subgroupShuffle(lessThan(data[1].i4.xy, ivec2(0)),  invocation));\r
  41     data[invocation].i4.xyz = ivec3(subgroupShuffle(lessThan(data[1].i4.xyz, ivec3(0)), invocation));\r
  42     data[invocation].i4     = ivec4(subgroupShuffle(lessThan(data[1].i4, ivec4(0)),     invocation));\r
  43 \r
  44     data[invocation].f4.x   = subgroupShuffleXor(data[0].f4.x,    invocation);\r
  45     data[invocation].f4.xy  = subgroupShuffleXor(data[1].f4.xy,   invocation);\r
  46     data[invocation].f4.xyz = subgroupShuffleXor(data[2].f4.xyz,  invocation);\r
  47     data[invocation].f4     = subgroupShuffleXor(data[3].f4,      invocation);\r
  48 \r
  49     data[invocation].i4.x   = subgroupShuffleXor(data[0].i4.x,    invocation);\r
  50     data[invocation].i4.xy  = subgroupShuffleXor(data[1].i4.xy,   invocation);\r
  51     data[invocation].i4.xyz = subgroupShuffleXor(data[2].i4.xyz,  invocation);\r
  52     data[invocation].i4     = subgroupShuffleXor(data[3].i4,      invocation);\r
  53 \r
  54     data[invocation].u4.x   = subgroupShuffleXor(data[0].u4.x,    invocation);\r
  55     data[invocation].u4.xy  = subgroupShuffleXor(data[1].u4.xy,   invocation);\r
  56     data[invocation].u4.xyz = subgroupShuffleXor(data[2].u4.xyz,  invocation);\r
  57     data[invocation].u4     = subgroupShuffleXor(data[3].u4,      invocation);\r
  58 \r
  59     data[invocation].d4.x   = subgroupShuffleXor(data[0].d4.x,    invocation);\r
  60     data[invocation].d4.xy  = subgroupShuffleXor(data[1].d4.xy,   invocation);\r
  61     data[invocation].d4.xyz = subgroupShuffleXor(data[2].d4.xyz,  invocation);\r
  62     data[invocation].d4     = subgroupShuffleXor(data[3].d4,      invocation);\r
  63 \r
  64     data[invocation].i4.x   =   int(subgroupShuffleXor(data[0].i4.x < 0,                   invocation));\r
  65     data[invocation].i4.xy  = ivec2(subgroupShuffleXor(lessThan(data[1].i4.xy, ivec2(0)),  invocation));\r
  66     data[invocation].i4.xyz = ivec3(subgroupShuffleXor(lessThan(data[1].i4.xyz, ivec3(0)), invocation));\r
  67     data[invocation].i4     = ivec4(subgroupShuffleXor(lessThan(data[1].i4, ivec4(0)),     invocation));\r
  68 }\r