glsl.450.subgroupQuad.comp 6.3 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119
  1. #version 450
  2. #extension GL_KHR_shader_subgroup_quad: enable
  3. layout (local_size_x = 8) in;
  4. layout(binding = 0) buffer Buffers
  5. {
  6. vec4 f4;
  7. ivec4 i4;
  8. uvec4 u4;
  9. dvec4 d4;
  10. } data[4];
  11. void main()
  12. {
  13. uint invocation = (gl_SubgroupInvocationID + gl_SubgroupSize) % 4;
  14. data[invocation].f4.x = subgroupQuadBroadcast(data[0].f4.x, 1);
  15. data[invocation].f4.xy = subgroupQuadBroadcast(data[1].f4.xy, 1);
  16. data[invocation].f4.xyz = subgroupQuadBroadcast(data[2].f4.xyz, 1);
  17. data[invocation].f4 = subgroupQuadBroadcast(data[3].f4, 1);
  18. data[invocation].i4.x = subgroupQuadBroadcast(data[0].i4.x, 1);
  19. data[invocation].i4.xy = subgroupQuadBroadcast(data[1].i4.xy, 1);
  20. data[invocation].i4.xyz = subgroupQuadBroadcast(data[2].i4.xyz, 1);
  21. data[invocation].i4 = subgroupQuadBroadcast(data[3].i4, 1);
  22. data[invocation].u4.x = subgroupQuadBroadcast(data[0].u4.x, 1);
  23. data[invocation].u4.xy = subgroupQuadBroadcast(data[1].u4.xy, 1);
  24. data[invocation].u4.xyz = subgroupQuadBroadcast(data[2].u4.xyz, 1);
  25. data[invocation].u4 = subgroupQuadBroadcast(data[3].u4, 1);
  26. data[invocation].d4.x = subgroupQuadBroadcast(data[0].d4.x, 1);
  27. data[invocation].d4.xy = subgroupQuadBroadcast(data[1].d4.xy, 1);
  28. data[invocation].d4.xyz = subgroupQuadBroadcast(data[2].d4.xyz, 1);
  29. data[invocation].d4 = subgroupQuadBroadcast(data[3].d4, 1);
  30. data[invocation].i4.x = int(subgroupQuadBroadcast(data[0].i4.x < 0, 1));
  31. data[invocation].i4.xy = ivec2(subgroupQuadBroadcast(lessThan(data[1].i4.xy, ivec2(0)), 1));
  32. data[invocation].i4.xyz = ivec3(subgroupQuadBroadcast(lessThan(data[1].i4.xyz, ivec3(0)), 1));
  33. data[invocation].i4 = ivec4(subgroupQuadBroadcast(lessThan(data[1].i4, ivec4(0)), 1));
  34. data[invocation].f4.x = subgroupQuadSwapHorizontal(data[0].f4.x);
  35. data[invocation].f4.xy = subgroupQuadSwapHorizontal(data[1].f4.xy);
  36. data[invocation].f4.xyz = subgroupQuadSwapHorizontal(data[2].f4.xyz);
  37. data[invocation].f4 = subgroupQuadSwapHorizontal(data[3].f4);
  38. data[invocation].i4.x = subgroupQuadSwapHorizontal(data[0].i4.x);
  39. data[invocation].i4.xy = subgroupQuadSwapHorizontal(data[1].i4.xy);
  40. data[invocation].i4.xyz = subgroupQuadSwapHorizontal(data[2].i4.xyz);
  41. data[invocation].i4 = subgroupQuadSwapHorizontal(data[3].i4);
  42. data[invocation].u4.x = subgroupQuadSwapHorizontal(data[0].u4.x);
  43. data[invocation].u4.xy = subgroupQuadSwapHorizontal(data[1].u4.xy);
  44. data[invocation].u4.xyz = subgroupQuadSwapHorizontal(data[2].u4.xyz);
  45. data[invocation].u4 = subgroupQuadSwapHorizontal(data[3].u4);
  46. data[invocation].d4.x = subgroupQuadSwapHorizontal(data[0].d4.x);
  47. data[invocation].d4.xy = subgroupQuadSwapHorizontal(data[1].d4.xy);
  48. data[invocation].d4.xyz = subgroupQuadSwapHorizontal(data[2].d4.xyz);
  49. data[invocation].d4 = subgroupQuadSwapHorizontal(data[3].d4);
  50. data[invocation].i4.x = int(subgroupQuadSwapHorizontal(data[0].i4.x < 0));
  51. data[invocation].i4.xy = ivec2(subgroupQuadSwapHorizontal(lessThan(data[1].i4.xy, ivec2(0))));
  52. data[invocation].i4.xyz = ivec3(subgroupQuadSwapHorizontal(lessThan(data[1].i4.xyz, ivec3(0))));
  53. data[invocation].i4 = ivec4(subgroupQuadSwapHorizontal(lessThan(data[1].i4, ivec4(0))));
  54. data[invocation].f4.x = subgroupQuadSwapVertical(data[0].f4.x);
  55. data[invocation].f4.xy = subgroupQuadSwapVertical(data[1].f4.xy);
  56. data[invocation].f4.xyz = subgroupQuadSwapVertical(data[2].f4.xyz);
  57. data[invocation].f4 = subgroupQuadSwapVertical(data[3].f4);
  58. data[invocation].i4.x = subgroupQuadSwapVertical(data[0].i4.x);
  59. data[invocation].i4.xy = subgroupQuadSwapVertical(data[1].i4.xy);
  60. data[invocation].i4.xyz = subgroupQuadSwapVertical(data[2].i4.xyz);
  61. data[invocation].i4 = subgroupQuadSwapVertical(data[3].i4);
  62. data[invocation].u4.x = subgroupQuadSwapVertical(data[0].u4.x);
  63. data[invocation].u4.xy = subgroupQuadSwapVertical(data[1].u4.xy);
  64. data[invocation].u4.xyz = subgroupQuadSwapVertical(data[2].u4.xyz);
  65. data[invocation].u4 = subgroupQuadSwapVertical(data[3].u4);
  66. data[invocation].d4.x = subgroupQuadSwapVertical(data[0].d4.x);
  67. data[invocation].d4.xy = subgroupQuadSwapVertical(data[1].d4.xy);
  68. data[invocation].d4.xyz = subgroupQuadSwapVertical(data[2].d4.xyz);
  69. data[invocation].d4 = subgroupQuadSwapVertical(data[3].d4);
  70. data[invocation].i4.x = int(subgroupQuadSwapVertical(data[0].i4.x < 0));
  71. data[invocation].i4.xy = ivec2(subgroupQuadSwapVertical(lessThan(data[1].i4.xy, ivec2(0))));
  72. data[invocation].i4.xyz = ivec3(subgroupQuadSwapVertical(lessThan(data[1].i4.xyz, ivec3(0))));
  73. data[invocation].i4 = ivec4(subgroupQuadSwapVertical(lessThan(data[1].i4, ivec4(0))));
  74. data[invocation].f4.x = subgroupQuadSwapDiagonal(data[0].f4.x);
  75. data[invocation].f4.xy = subgroupQuadSwapDiagonal(data[1].f4.xy);
  76. data[invocation].f4.xyz = subgroupQuadSwapDiagonal(data[2].f4.xyz);
  77. data[invocation].f4 = subgroupQuadSwapDiagonal(data[3].f4);
  78. data[invocation].i4.x = subgroupQuadSwapDiagonal(data[0].i4.x);
  79. data[invocation].i4.xy = subgroupQuadSwapDiagonal(data[1].i4.xy);
  80. data[invocation].i4.xyz = subgroupQuadSwapDiagonal(data[2].i4.xyz);
  81. data[invocation].i4 = subgroupQuadSwapDiagonal(data[3].i4);
  82. data[invocation].u4.x = subgroupQuadSwapDiagonal(data[0].u4.x);
  83. data[invocation].u4.xy = subgroupQuadSwapDiagonal(data[1].u4.xy);
  84. data[invocation].u4.xyz = subgroupQuadSwapDiagonal(data[2].u4.xyz);
  85. data[invocation].u4 = subgroupQuadSwapDiagonal(data[3].u4);
  86. data[invocation].d4.x = subgroupQuadSwapDiagonal(data[0].d4.x);
  87. data[invocation].d4.xy = subgroupQuadSwapDiagonal(data[1].d4.xy);
  88. data[invocation].d4.xyz = subgroupQuadSwapDiagonal(data[2].d4.xyz);
  89. data[invocation].d4 = subgroupQuadSwapDiagonal(data[3].d4);
  90. data[invocation].i4.x = int(subgroupQuadSwapDiagonal(data[0].i4.x < 0));
  91. data[invocation].i4.xy = ivec2(subgroupQuadSwapDiagonal(lessThan(data[1].i4.xy, ivec2(0))));
  92. data[invocation].i4.xyz = ivec3(subgroupQuadSwapDiagonal(lessThan(data[1].i4.xyz, ivec3(0))));
  93. data[invocation].i4 = ivec4(subgroupQuadSwapDiagonal(lessThan(data[1].i4, ivec4(0))));
  94. }