| blob | 61764e56209c85b60e2d2eef0ca4b8cbab369bbb |
1 #version 320 es\r
2 \r
3 #extension GL_KHR_shader_subgroup_clustered: enable\r
4 \r
5 layout (local_size_x = 8) 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 } data[4];\r
13 \r
14 void main()\r
15 {\r
16 uint invocation = (gl_SubgroupInvocationID + gl_SubgroupSize) % 4u;\r
17 \r
18 data[0].f4.x = subgroupClusteredAdd(data[0].f4.x, 1u);\r
19 data[0].f4.xy = subgroupClusteredAdd(data[1].f4.xy, 1u);\r
20 data[0].f4.xyz = subgroupClusteredAdd(data[2].f4.xyz, 1u);\r
21 data[0].f4 = subgroupClusteredAdd(data[3].f4, 1u);\r
22 \r
23 data[1].i4.x = subgroupClusteredAdd(data[0].i4.x, 1u);\r
24 data[1].i4.xy = subgroupClusteredAdd(data[1].i4.xy, 1u);\r
25 data[1].i4.xyz = subgroupClusteredAdd(data[2].i4.xyz, 1u);\r
26 data[1].i4 = subgroupClusteredAdd(data[3].i4, 1u);\r
27 \r
28 data[2].u4.x = subgroupClusteredAdd(data[0].u4.x, 1u);\r
29 data[2].u4.xy = subgroupClusteredAdd(data[1].u4.xy, 1u);\r
30 data[2].u4.xyz = subgroupClusteredAdd(data[2].u4.xyz, 1u);\r
31 data[2].u4 = subgroupClusteredAdd(data[3].u4, 1u);\r
32 \r
33 data[3].f4.x = subgroupClusteredMul(data[0].f4.x, 1u);\r
34 data[3].f4.xy = subgroupClusteredMul(data[1].f4.xy, 1u);\r
35 data[3].f4.xyz = subgroupClusteredMul(data[2].f4.xyz, 1u);\r
36 data[3].f4 = subgroupClusteredMul(data[3].f4, 1u);\r
37 \r
38 data[0].i4.x = subgroupClusteredMul(data[0].i4.x, 1u);\r
39 data[0].i4.xy = subgroupClusteredMul(data[1].i4.xy, 1u);\r
40 data[0].i4.xyz = subgroupClusteredMul(data[2].i4.xyz, 1u);\r
41 data[0].i4 = subgroupClusteredMul(data[3].i4, 1u);\r
42 \r
43 data[1].u4.x = subgroupClusteredMul(data[0].u4.x, 1u);\r
44 data[1].u4.xy = subgroupClusteredMul(data[1].u4.xy, 1u);\r
45 data[1].u4.xyz = subgroupClusteredMul(data[2].u4.xyz, 1u);\r
46 data[1].u4 = subgroupClusteredMul(data[3].u4, 1u);\r
47 \r
48 data[2].f4.x = subgroupClusteredMin(data[0].f4.x, 1u);\r
49 data[2].f4.xy = subgroupClusteredMin(data[1].f4.xy, 1u);\r
50 data[2].f4.xyz = subgroupClusteredMin(data[2].f4.xyz, 1u);\r
51 data[2].f4 = subgroupClusteredMin(data[3].f4, 1u);\r
52 \r
53 data[3].i4.x = subgroupClusteredMin(data[0].i4.x, 1u);\r
54 data[3].i4.xy = subgroupClusteredMin(data[1].i4.xy, 1u);\r
55 data[3].i4.xyz = subgroupClusteredMin(data[2].i4.xyz, 1u);\r
56 data[3].i4 = subgroupClusteredMin(data[3].i4, 1u);\r
57 \r
58 data[0].u4.x = subgroupClusteredMin(data[0].u4.x, 1u);\r
59 data[0].u4.xy = subgroupClusteredMin(data[1].u4.xy, 1u);\r
60 data[0].u4.xyz = subgroupClusteredMin(data[2].u4.xyz, 1u);\r
61 data[0].u4 = subgroupClusteredMin(data[3].u4, 1u);\r
62 \r
63 data[1].f4.x = subgroupClusteredMax(data[0].f4.x, 1u);\r
64 data[1].f4.xy = subgroupClusteredMax(data[1].f4.xy, 1u);\r
65 data[1].f4.xyz = subgroupClusteredMax(data[2].f4.xyz, 1u);\r
66 data[1].f4 = subgroupClusteredMax(data[3].f4, 1u);\r
67 \r
68 data[2].i4.x = subgroupClusteredMax(data[0].i4.x, 1u);\r
69 data[2].i4.xy = subgroupClusteredMax(data[1].i4.xy, 1u);\r
70 data[2].i4.xyz = subgroupClusteredMax(data[2].i4.xyz, 1u);\r
71 data[2].i4 = subgroupClusteredMax(data[3].i4, 1u);\r
72 \r
73 data[3].u4.x = subgroupClusteredMax(data[0].u4.x, 1u);\r
74 data[3].u4.xy = subgroupClusteredMax(data[1].u4.xy, 1u);\r
75 data[3].u4.xyz = subgroupClusteredMax(data[2].u4.xyz, 1u);\r
76 data[3].u4 = subgroupClusteredMax(data[3].u4, 1u);\r
77 \r
78 data[0].i4.x = subgroupClusteredAnd(data[0].i4.x, 1u);\r
79 data[0].i4.xy = subgroupClusteredAnd(data[1].i4.xy, 1u);\r
80 data[0].i4.xyz = subgroupClusteredAnd(data[2].i4.xyz, 1u);\r
81 data[0].i4 = subgroupClusteredAnd(data[3].i4, 1u);\r
82 \r
83 data[1].u4.x = subgroupClusteredAnd(data[0].u4.x, 1u);\r
84 data[1].u4.xy = subgroupClusteredAnd(data[1].u4.xy, 1u);\r
85 data[1].u4.xyz = subgroupClusteredAnd(data[2].u4.xyz, 1u);\r
86 data[1].u4 = subgroupClusteredAnd(data[3].u4, 1u);\r
87 \r
88 data[2].i4.x = int(subgroupClusteredAnd(data[0].i4.x < 0, 1u));\r
89 data[2].i4.xy = ivec2(subgroupClusteredAnd(lessThan(data[1].i4.xy, ivec2(0)), 1u));\r
90 data[2].i4.xyz = ivec3(subgroupClusteredAnd(lessThan(data[1].i4.xyz, ivec3(0)), 1u));\r
91 data[2].i4 = ivec4(subgroupClusteredAnd(lessThan(data[1].i4, ivec4(0)), 1u));\r
92 \r
93 data[3].i4.x = subgroupClusteredOr(data[0].i4.x, 1u);\r
94 data[3].i4.xy = subgroupClusteredOr(data[1].i4.xy, 1u);\r
95 data[3].i4.xyz = subgroupClusteredOr(data[2].i4.xyz, 1u);\r
96 data[3].i4 = subgroupClusteredOr(data[3].i4, 1u);\r
97 \r
98 data[0].u4.x = subgroupClusteredOr(data[0].u4.x, 1u);\r
99 data[0].u4.xy = subgroupClusteredOr(data[1].u4.xy, 1u);\r
100 data[0].u4.xyz = subgroupClusteredOr(data[2].u4.xyz, 1u);\r
101 data[0].u4 = subgroupClusteredOr(data[3].u4, 1u);\r
102 \r
103 data[1].i4.x = int(subgroupClusteredOr(data[0].i4.x < 0, 1u));\r
104 data[1].i4.xy = ivec2(subgroupClusteredOr(lessThan(data[1].i4.xy, ivec2(0)), 1u));\r
105 data[1].i4.xyz = ivec3(subgroupClusteredOr(lessThan(data[1].i4.xyz, ivec3(0)), 1u));\r
106 data[1].i4 = ivec4(subgroupClusteredOr(lessThan(data[1].i4, ivec4(0)), 1u));\r
107 \r
108 data[2].i4.x = subgroupClusteredXor(data[0].i4.x, 1u);\r
109 data[2].i4.xy = subgroupClusteredXor(data[1].i4.xy, 1u);\r
110 data[2].i4.xyz = subgroupClusteredXor(data[2].i4.xyz, 1u);\r
111 data[2].i4 = subgroupClusteredXor(data[3].i4, 1u);\r
112 \r
113 data[3].u4.x = subgroupClusteredXor(data[0].u4.x, 1u);\r
114 data[3].u4.xy = subgroupClusteredXor(data[1].u4.xy, 1u);\r
115 data[3].u4.xyz = subgroupClusteredXor(data[2].u4.xyz, 1u);\r
116 data[3].u4 = subgroupClusteredXor(data[3].u4, 1u);\r
117 \r
118 data[0].i4.x = int(subgroupClusteredXor(data[0].i4.x < 0, 1u));\r
119 data[0].i4.xy = ivec2(subgroupClusteredXor(lessThan(data[1].i4.xy, ivec2(0)), 1u));\r
120 data[0].i4.xyz = ivec3(subgroupClusteredXor(lessThan(data[1].i4.xyz, ivec3(0)), 1u));\r
121 data[0].i4 = ivec4(subgroupClusteredXor(lessThan(data[1].i4, ivec4(0)), 1u));\r
122 }\r