3 #extension GL_KHR_shader_subgroup_quad: enable
\r
5 layout (local_size_x = 8) in;
\r
7 layout(binding = 0) buffer Buffers
\r
17 uint invocation = (gl_SubgroupInvocationID + gl_SubgroupSize) % 4;
\r
19 data[invocation].f4.x = subgroupQuadBroadcast(data[0].f4.x, 1);
\r
20 data[invocation].f4.xy = subgroupQuadBroadcast(data[1].f4.xy, 1);
\r
21 data[invocation].f4.xyz = subgroupQuadBroadcast(data[2].f4.xyz, 1);
\r
22 data[invocation].f4 = subgroupQuadBroadcast(data[3].f4, 1);
\r
24 data[invocation].i4.x = subgroupQuadBroadcast(data[0].i4.x, 1);
\r
25 data[invocation].i4.xy = subgroupQuadBroadcast(data[1].i4.xy, 1);
\r
26 data[invocation].i4.xyz = subgroupQuadBroadcast(data[2].i4.xyz, 1);
\r
27 data[invocation].i4 = subgroupQuadBroadcast(data[3].i4, 1);
\r
29 data[invocation].u4.x = subgroupQuadBroadcast(data[0].u4.x, 1);
\r
30 data[invocation].u4.xy = subgroupQuadBroadcast(data[1].u4.xy, 1);
\r
31 data[invocation].u4.xyz = subgroupQuadBroadcast(data[2].u4.xyz, 1);
\r
32 data[invocation].u4 = subgroupQuadBroadcast(data[3].u4, 1);
\r
34 data[invocation].d4.x = subgroupQuadBroadcast(data[0].d4.x, 1);
\r
35 data[invocation].d4.xy = subgroupQuadBroadcast(data[1].d4.xy, 1);
\r
36 data[invocation].d4.xyz = subgroupQuadBroadcast(data[2].d4.xyz, 1);
\r
37 data[invocation].d4 = subgroupQuadBroadcast(data[3].d4, 1);
\r
39 data[invocation].i4.x = int(subgroupQuadBroadcast(data[0].i4.x < 0, 1));
\r
40 data[invocation].i4.xy = ivec2(subgroupQuadBroadcast(lessThan(data[1].i4.xy, ivec2(0)), 1));
\r
41 data[invocation].i4.xyz = ivec3(subgroupQuadBroadcast(lessThan(data[1].i4.xyz, ivec3(0)), 1));
\r
42 data[invocation].i4 = ivec4(subgroupQuadBroadcast(lessThan(data[1].i4, ivec4(0)), 1));
\r
44 data[invocation].f4.x = subgroupQuadSwapHorizontal(data[0].f4.x);
\r
45 data[invocation].f4.xy = subgroupQuadSwapHorizontal(data[1].f4.xy);
\r
46 data[invocation].f4.xyz = subgroupQuadSwapHorizontal(data[2].f4.xyz);
\r
47 data[invocation].f4 = subgroupQuadSwapHorizontal(data[3].f4);
\r
49 data[invocation].i4.x = subgroupQuadSwapHorizontal(data[0].i4.x);
\r
50 data[invocation].i4.xy = subgroupQuadSwapHorizontal(data[1].i4.xy);
\r
51 data[invocation].i4.xyz = subgroupQuadSwapHorizontal(data[2].i4.xyz);
\r
52 data[invocation].i4 = subgroupQuadSwapHorizontal(data[3].i4);
\r
54 data[invocation].u4.x = subgroupQuadSwapHorizontal(data[0].u4.x);
\r
55 data[invocation].u4.xy = subgroupQuadSwapHorizontal(data[1].u4.xy);
\r
56 data[invocation].u4.xyz = subgroupQuadSwapHorizontal(data[2].u4.xyz);
\r
57 data[invocation].u4 = subgroupQuadSwapHorizontal(data[3].u4);
\r
59 data[invocation].d4.x = subgroupQuadSwapHorizontal(data[0].d4.x);
\r
60 data[invocation].d4.xy = subgroupQuadSwapHorizontal(data[1].d4.xy);
\r
61 data[invocation].d4.xyz = subgroupQuadSwapHorizontal(data[2].d4.xyz);
\r
62 data[invocation].d4 = subgroupQuadSwapHorizontal(data[3].d4);
\r
64 data[invocation].i4.x = int(subgroupQuadSwapHorizontal(data[0].i4.x < 0));
\r
65 data[invocation].i4.xy = ivec2(subgroupQuadSwapHorizontal(lessThan(data[1].i4.xy, ivec2(0))));
\r
66 data[invocation].i4.xyz = ivec3(subgroupQuadSwapHorizontal(lessThan(data[1].i4.xyz, ivec3(0))));
\r
67 data[invocation].i4 = ivec4(subgroupQuadSwapHorizontal(lessThan(data[1].i4, ivec4(0))));
\r
69 data[invocation].f4.x = subgroupQuadSwapVertical(data[0].f4.x);
\r
70 data[invocation].f4.xy = subgroupQuadSwapVertical(data[1].f4.xy);
\r
71 data[invocation].f4.xyz = subgroupQuadSwapVertical(data[2].f4.xyz);
\r
72 data[invocation].f4 = subgroupQuadSwapVertical(data[3].f4);
\r
74 data[invocation].i4.x = subgroupQuadSwapVertical(data[0].i4.x);
\r
75 data[invocation].i4.xy = subgroupQuadSwapVertical(data[1].i4.xy);
\r
76 data[invocation].i4.xyz = subgroupQuadSwapVertical(data[2].i4.xyz);
\r
77 data[invocation].i4 = subgroupQuadSwapVertical(data[3].i4);
\r
79 data[invocation].u4.x = subgroupQuadSwapVertical(data[0].u4.x);
\r
80 data[invocation].u4.xy = subgroupQuadSwapVertical(data[1].u4.xy);
\r
81 data[invocation].u4.xyz = subgroupQuadSwapVertical(data[2].u4.xyz);
\r
82 data[invocation].u4 = subgroupQuadSwapVertical(data[3].u4);
\r
84 data[invocation].d4.x = subgroupQuadSwapVertical(data[0].d4.x);
\r
85 data[invocation].d4.xy = subgroupQuadSwapVertical(data[1].d4.xy);
\r
86 data[invocation].d4.xyz = subgroupQuadSwapVertical(data[2].d4.xyz);
\r
87 data[invocation].d4 = subgroupQuadSwapVertical(data[3].d4);
\r
89 data[invocation].i4.x = int(subgroupQuadSwapVertical(data[0].i4.x < 0));
\r
90 data[invocation].i4.xy = ivec2(subgroupQuadSwapVertical(lessThan(data[1].i4.xy, ivec2(0))));
\r
91 data[invocation].i4.xyz = ivec3(subgroupQuadSwapVertical(lessThan(data[1].i4.xyz, ivec3(0))));
\r
92 data[invocation].i4 = ivec4(subgroupQuadSwapVertical(lessThan(data[1].i4, ivec4(0))));
\r
94 data[invocation].f4.x = subgroupQuadSwapDiagonal(data[0].f4.x);
\r
95 data[invocation].f4.xy = subgroupQuadSwapDiagonal(data[1].f4.xy);
\r
96 data[invocation].f4.xyz = subgroupQuadSwapDiagonal(data[2].f4.xyz);
\r
97 data[invocation].f4 = subgroupQuadSwapDiagonal(data[3].f4);
\r
99 data[invocation].i4.x = subgroupQuadSwapDiagonal(data[0].i4.x);
\r
100 data[invocation].i4.xy = subgroupQuadSwapDiagonal(data[1].i4.xy);
\r
101 data[invocation].i4.xyz = subgroupQuadSwapDiagonal(data[2].i4.xyz);
\r
102 data[invocation].i4 = subgroupQuadSwapDiagonal(data[3].i4);
\r
104 data[invocation].u4.x = subgroupQuadSwapDiagonal(data[0].u4.x);
\r
105 data[invocation].u4.xy = subgroupQuadSwapDiagonal(data[1].u4.xy);
\r
106 data[invocation].u4.xyz = subgroupQuadSwapDiagonal(data[2].u4.xyz);
\r
107 data[invocation].u4 = subgroupQuadSwapDiagonal(data[3].u4);
\r
109 data[invocation].d4.x = subgroupQuadSwapDiagonal(data[0].d4.x);
\r
110 data[invocation].d4.xy = subgroupQuadSwapDiagonal(data[1].d4.xy);
\r
111 data[invocation].d4.xyz = subgroupQuadSwapDiagonal(data[2].d4.xyz);
\r
112 data[invocation].d4 = subgroupQuadSwapDiagonal(data[3].d4);
\r
114 data[invocation].i4.x = int(subgroupQuadSwapDiagonal(data[0].i4.x < 0));
\r
115 data[invocation].i4.xy = ivec2(subgroupQuadSwapDiagonal(lessThan(data[1].i4.xy, ivec2(0))));
\r
116 data[invocation].i4.xyz = ivec3(subgroupQuadSwapDiagonal(lessThan(data[1].i4.xyz, ivec3(0))));
\r
117 data[invocation].i4 = ivec4(subgroupQuadSwapDiagonal(lessThan(data[1].i4, ivec4(0))));
\r