3 #extension GL_KHR_shader_subgroup_ballot: enable
\r
5 layout (local_size_x = 8, local_size_y = 8, local_size_z = 1) in;
\r
7 layout(binding = 0) buffer Buffers
\r
16 uint invocation = (gl_SubgroupInvocationID + gl_SubgroupSize) % 4u;
\r
18 uvec4 relMask = gl_SubgroupEqMask +
\r
24 uvec4 result = subgroupBallot(true);
\r
26 data[0].u4.x = subgroupBallotBitCount(result);
\r
27 data[0].u4.y = subgroupBallotBitExtract(result, 0u) ? 1u : 0u;
\r
28 data[0].u4.z = subgroupBallotInclusiveBitCount(result) + subgroupBallotExclusiveBitCount(result);
\r
29 data[0].u4.w = subgroupBallotFindLSB(result) + subgroupBallotFindMSB(result);
\r
31 if ((relMask == result) && subgroupInverseBallot(data[0].u4))
\r
33 data[1].f4.x = subgroupBroadcast(data[0].f4.x, 3u);
\r
34 data[1].f4.xy = subgroupBroadcast(data[1].f4.xy, 3u);
\r
35 data[1].f4.xyz = subgroupBroadcast(data[2].f4.xyz, 3u);
\r
36 data[1].f4 = subgroupBroadcast(data[3].f4, 3u);
\r
38 data[2].i4.x = subgroupBroadcast(data[0].i4.x, 2u);
\r
39 data[2].i4.xy = subgroupBroadcast(data[1].i4.xy, 2u);
\r
40 data[2].i4.xyz = subgroupBroadcast(data[2].i4.xyz, 2u);
\r
41 data[2].i4 = subgroupBroadcast(data[3].i4, 2u);
\r
43 data[3].u4.x = subgroupBroadcast(data[0].u4.x, 1u);
\r
44 data[3].u4.xy = subgroupBroadcast(data[1].u4.xy, 1u);
\r
45 data[3].u4.xyz = subgroupBroadcast(data[2].u4.xyz, 1u);
\r
46 data[3].u4 = subgroupBroadcast(data[3].u4, 1u);
\r
48 data[0].i4.x = int(subgroupBroadcast(data[0].i4.x < 0, 1u));
\r
49 data[0].i4.xy = ivec2(subgroupBroadcast(lessThan(data[1].i4.xy, ivec2(0)), 1u));
\r
50 data[0].i4.xyz = ivec3(subgroupBroadcast(lessThan(data[1].i4.xyz, ivec3(0)), 1u));
\r
51 data[0].i4 = ivec4(subgroupBroadcast(lessThan(data[1].i4, ivec4(0)), 1u));
\r
55 data[1].f4.x = subgroupBroadcastFirst(data[0].f4.x);
\r
56 data[1].f4.xy = subgroupBroadcastFirst(data[1].f4.xy);
\r
57 data[1].f4.xyz = subgroupBroadcastFirst(data[2].f4.xyz);
\r
58 data[1].f4 = subgroupBroadcastFirst(data[3].f4);
\r
60 data[2].i4.x = subgroupBroadcastFirst(data[0].i4.x);
\r
61 data[2].i4.xy = subgroupBroadcastFirst(data[1].i4.xy);
\r
62 data[2].i4.xyz = subgroupBroadcastFirst(data[2].i4.xyz);
\r
63 data[2].i4 = subgroupBroadcastFirst(data[3].i4);
\r
65 data[3].u4.x = subgroupBroadcastFirst(data[0].u4.x);
\r
66 data[3].u4.xy = subgroupBroadcastFirst(data[1].u4.xy);
\r
67 data[3].u4.xyz = subgroupBroadcastFirst(data[2].u4.xyz);
\r
68 data[3].u4 = subgroupBroadcastFirst(data[3].u4);
\r
70 data[0].i4.x = int(subgroupBroadcastFirst(data[0].i4.x < 0));
\r
71 data[0].i4.xy = ivec2(subgroupBroadcastFirst(lessThan(data[1].i4.xy, ivec2(0))));
\r
72 data[0].i4.xyz = ivec3(subgroupBroadcastFirst(lessThan(data[1].i4.xyz, ivec3(0))));
\r
73 data[0].i4 = ivec4(subgroupBroadcastFirst(lessThan(data[1].i4, ivec4(0))));
\r