• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1#version 450
2
3#extension GL_KHR_shader_subgroup_clustered: enable
4#extension GL_EXT_shader_explicit_arithmetic_types_int8: enable
5#extension GL_EXT_shader_explicit_arithmetic_types_int16: enable
6#extension GL_EXT_shader_explicit_arithmetic_types_int64: enable
7#extension GL_EXT_shader_explicit_arithmetic_types_float16: enable
8
9layout (local_size_x = 8) in;
10
11layout(binding = 0) buffer Buffers
12{
13    i8vec4 i8;
14    u8vec4 u8;
15    i16vec4 i16;
16    u16vec4 u16;
17    i64vec4 i64;
18    u64vec4 u64;
19    f16vec4 f16;
20} data[4];
21
22void main()
23{
24    uint invocation = (gl_SubgroupInvocationID + gl_SubgroupSize) % 4;
25
26    data[invocation].i8.x   = subgroupClusteredAdd(data[0].i8.x, 1);
27    data[invocation].i8.xy  = subgroupClusteredAdd(data[1].i8.xy, 1);
28    data[invocation].i8.xyz = subgroupClusteredAdd(data[2].i8.xyz, 1);
29    data[invocation].i8     = subgroupClusteredAdd(data[3].i8, 1);
30
31    data[invocation].i8.x   = subgroupClusteredMul(data[0].i8.x, 1);
32    data[invocation].i8.xy  = subgroupClusteredMul(data[1].i8.xy, 1);
33    data[invocation].i8.xyz = subgroupClusteredMul(data[2].i8.xyz, 1);
34    data[invocation].i8     = subgroupClusteredMul(data[3].i8, 1);
35
36    data[invocation].i8.x   = subgroupClusteredMin(data[0].i8.x, 1);
37    data[invocation].i8.xy  = subgroupClusteredMin(data[1].i8.xy, 1);
38    data[invocation].i8.xyz = subgroupClusteredMin(data[2].i8.xyz, 1);
39    data[invocation].i8     = subgroupClusteredMin(data[3].i8, 1);
40
41    data[invocation].i8.x   = subgroupClusteredMax(data[0].i8.x, 1);
42    data[invocation].i8.xy  = subgroupClusteredMax(data[1].i8.xy, 1);
43    data[invocation].i8.xyz = subgroupClusteredMax(data[2].i8.xyz, 1);
44    data[invocation].i8     = subgroupClusteredMax(data[3].i8, 1);
45
46    data[invocation].i8.x   = subgroupClusteredAnd(data[0].i8.x, 1);
47    data[invocation].i8.xy  = subgroupClusteredAnd(data[1].i8.xy, 1);
48    data[invocation].i8.xyz = subgroupClusteredAnd(data[2].i8.xyz, 1);
49    data[invocation].i8     = subgroupClusteredAnd(data[3].i8, 1);
50
51    data[invocation].i8.x   = subgroupClusteredOr(data[0].i8.x, 1);
52    data[invocation].i8.xy  = subgroupClusteredOr(data[1].i8.xy, 1);
53    data[invocation].i8.xyz = subgroupClusteredOr(data[2].i8.xyz, 1);
54    data[invocation].i8     = subgroupClusteredOr(data[3].i8, 1);
55
56    data[invocation].i8.x   = subgroupClusteredXor(data[0].i8.x, 1);
57    data[invocation].i8.xy  = subgroupClusteredXor(data[1].i8.xy, 1);
58    data[invocation].i8.xyz = subgroupClusteredXor(data[2].i8.xyz, 1);
59    data[invocation].i8     = subgroupClusteredXor(data[3].i8, 1);
60
61    data[invocation].u8.x   = subgroupClusteredAdd(data[0].u8.x, 1);
62    data[invocation].u8.xy  = subgroupClusteredAdd(data[1].u8.xy, 1);
63    data[invocation].u8.xyz = subgroupClusteredAdd(data[2].u8.xyz, 1);
64    data[invocation].u8     = subgroupClusteredAdd(data[3].u8, 1);
65
66    data[invocation].u8.x   = subgroupClusteredMul(data[0].u8.x, 1);
67    data[invocation].u8.xy  = subgroupClusteredMul(data[1].u8.xy, 1);
68    data[invocation].u8.xyz = subgroupClusteredMul(data[2].u8.xyz, 1);
69    data[invocation].u8     = subgroupClusteredMul(data[3].u8, 1);
70
71    data[invocation].u8.x   = subgroupClusteredMin(data[0].u8.x, 1);
72    data[invocation].u8.xy  = subgroupClusteredMin(data[1].u8.xy, 1);
73    data[invocation].u8.xyz = subgroupClusteredMin(data[2].u8.xyz, 1);
74    data[invocation].u8     = subgroupClusteredMin(data[3].u8, 1);
75
76    data[invocation].u8.x   = subgroupClusteredMax(data[0].u8.x, 1);
77    data[invocation].u8.xy  = subgroupClusteredMax(data[1].u8.xy, 1);
78    data[invocation].u8.xyz = subgroupClusteredMax(data[2].u8.xyz, 1);
79    data[invocation].u8     = subgroupClusteredMax(data[3].u8, 1);
80
81    data[invocation].u8.x   = subgroupClusteredAnd(data[0].u8.x, 1);
82    data[invocation].u8.xy  = subgroupClusteredAnd(data[1].u8.xy, 1);
83    data[invocation].u8.xyz = subgroupClusteredAnd(data[2].u8.xyz, 1);
84    data[invocation].u8     = subgroupClusteredAnd(data[3].u8, 1);
85
86    data[invocation].u8.x   = subgroupClusteredOr(data[0].u8.x, 1);
87    data[invocation].u8.xy  = subgroupClusteredOr(data[1].u8.xy, 1);
88    data[invocation].u8.xyz = subgroupClusteredOr(data[2].u8.xyz, 1);
89    data[invocation].u8     = subgroupClusteredOr(data[3].u8, 1);
90
91    data[invocation].u8.x   = subgroupClusteredXor(data[0].u8.x, 1);
92    data[invocation].u8.xy  = subgroupClusteredXor(data[1].u8.xy, 1);
93    data[invocation].u8.xyz = subgroupClusteredXor(data[2].u8.xyz, 1);
94    data[invocation].u8     = subgroupClusteredXor(data[3].u8, 1);
95
96    data[invocation].i16.x   = subgroupClusteredAdd(data[0].i16.x, 1);
97    data[invocation].i16.xy  = subgroupClusteredAdd(data[1].i16.xy, 1);
98    data[invocation].i16.xyz = subgroupClusteredAdd(data[2].i16.xyz, 1);
99    data[invocation].i16     = subgroupClusteredAdd(data[3].i16, 1);
100
101    data[invocation].i16.x   = subgroupClusteredMul(data[0].i16.x, 1);
102    data[invocation].i16.xy  = subgroupClusteredMul(data[1].i16.xy, 1);
103    data[invocation].i16.xyz = subgroupClusteredMul(data[2].i16.xyz, 1);
104    data[invocation].i16     = subgroupClusteredMul(data[3].i16, 1);
105
106    data[invocation].i16.x   = subgroupClusteredMin(data[0].i16.x, 1);
107    data[invocation].i16.xy  = subgroupClusteredMin(data[1].i16.xy, 1);
108    data[invocation].i16.xyz = subgroupClusteredMin(data[2].i16.xyz, 1);
109    data[invocation].i16     = subgroupClusteredMin(data[3].i16, 1);
110
111    data[invocation].i16.x   = subgroupClusteredMax(data[0].i16.x, 1);
112    data[invocation].i16.xy  = subgroupClusteredMax(data[1].i16.xy, 1);
113    data[invocation].i16.xyz = subgroupClusteredMax(data[2].i16.xyz, 1);
114    data[invocation].i16     = subgroupClusteredMax(data[3].i16, 1);
115
116    data[invocation].i16.x   = subgroupClusteredAnd(data[0].i16.x, 1);
117    data[invocation].i16.xy  = subgroupClusteredAnd(data[1].i16.xy, 1);
118    data[invocation].i16.xyz = subgroupClusteredAnd(data[2].i16.xyz, 1);
119    data[invocation].i16     = subgroupClusteredAnd(data[3].i16, 1);
120
121    data[invocation].i16.x   = subgroupClusteredOr(data[0].i16.x, 1);
122    data[invocation].i16.xy  = subgroupClusteredOr(data[1].i16.xy, 1);
123    data[invocation].i16.xyz = subgroupClusteredOr(data[2].i16.xyz, 1);
124    data[invocation].i16     = subgroupClusteredOr(data[3].i16, 1);
125
126    data[invocation].i16.x   = subgroupClusteredXor(data[0].i16.x, 1);
127    data[invocation].i16.xy  = subgroupClusteredXor(data[1].i16.xy, 1);
128    data[invocation].i16.xyz = subgroupClusteredXor(data[2].i16.xyz, 1);
129    data[invocation].i16     = subgroupClusteredXor(data[3].i16, 1);
130
131    data[invocation].u16.x   = subgroupClusteredAdd(data[0].u16.x, 1);
132    data[invocation].u16.xy  = subgroupClusteredAdd(data[1].u16.xy, 1);
133    data[invocation].u16.xyz = subgroupClusteredAdd(data[2].u16.xyz, 1);
134    data[invocation].u16     = subgroupClusteredAdd(data[3].u16, 1);
135
136    data[invocation].u16.x   = subgroupClusteredMul(data[0].u16.x, 1);
137    data[invocation].u16.xy  = subgroupClusteredMul(data[1].u16.xy, 1);
138    data[invocation].u16.xyz = subgroupClusteredMul(data[2].u16.xyz, 1);
139    data[invocation].u16     = subgroupClusteredMul(data[3].u16, 1);
140
141    data[invocation].u16.x   = subgroupClusteredMin(data[0].u16.x, 1);
142    data[invocation].u16.xy  = subgroupClusteredMin(data[1].u16.xy, 1);
143    data[invocation].u16.xyz = subgroupClusteredMin(data[2].u16.xyz, 1);
144    data[invocation].u16     = subgroupClusteredMin(data[3].u16, 1);
145
146    data[invocation].u16.x   = subgroupClusteredMax(data[0].u16.x, 1);
147    data[invocation].u16.xy  = subgroupClusteredMax(data[1].u16.xy, 1);
148    data[invocation].u16.xyz = subgroupClusteredMax(data[2].u16.xyz, 1);
149    data[invocation].u16     = subgroupClusteredMax(data[3].u16, 1);
150
151    data[invocation].u16.x   = subgroupClusteredAnd(data[0].u16.x, 1);
152    data[invocation].u16.xy  = subgroupClusteredAnd(data[1].u16.xy, 1);
153    data[invocation].u16.xyz = subgroupClusteredAnd(data[2].u16.xyz, 1);
154    data[invocation].u16     = subgroupClusteredAnd(data[3].u16, 1);
155
156    data[invocation].u16.x   = subgroupClusteredOr(data[0].u16.x, 1);
157    data[invocation].u16.xy  = subgroupClusteredOr(data[1].u16.xy, 1);
158    data[invocation].u16.xyz = subgroupClusteredOr(data[2].u16.xyz, 1);
159    data[invocation].u16     = subgroupClusteredOr(data[3].u16, 1);
160
161    data[invocation].u16.x   = subgroupClusteredXor(data[0].u16.x, 1);
162    data[invocation].u16.xy  = subgroupClusteredXor(data[1].u16.xy, 1);
163    data[invocation].u16.xyz = subgroupClusteredXor(data[2].u16.xyz, 1);
164    data[invocation].u16     = subgroupClusteredXor(data[3].u16, 1);
165
166    data[invocation].i64.x   = subgroupClusteredAdd(data[0].i64.x, 1);
167    data[invocation].i64.xy  = subgroupClusteredAdd(data[1].i64.xy, 1);
168    data[invocation].i64.xyz = subgroupClusteredAdd(data[2].i64.xyz, 1);
169    data[invocation].i64     = subgroupClusteredAdd(data[3].i64, 1);
170
171    data[invocation].i64.x   = subgroupClusteredMul(data[0].i64.x, 1);
172    data[invocation].i64.xy  = subgroupClusteredMul(data[1].i64.xy, 1);
173    data[invocation].i64.xyz = subgroupClusteredMul(data[2].i64.xyz, 1);
174    data[invocation].i64     = subgroupClusteredMul(data[3].i64, 1);
175
176    data[invocation].i64.x   = subgroupClusteredMin(data[0].i64.x, 1);
177    data[invocation].i64.xy  = subgroupClusteredMin(data[1].i64.xy, 1);
178    data[invocation].i64.xyz = subgroupClusteredMin(data[2].i64.xyz, 1);
179    data[invocation].i64     = subgroupClusteredMin(data[3].i64, 1);
180
181    data[invocation].i64.x   = subgroupClusteredMax(data[0].i64.x, 1);
182    data[invocation].i64.xy  = subgroupClusteredMax(data[1].i64.xy, 1);
183    data[invocation].i64.xyz = subgroupClusteredMax(data[2].i64.xyz, 1);
184    data[invocation].i64     = subgroupClusteredMax(data[3].i64, 1);
185
186    data[invocation].i64.x   = subgroupClusteredAnd(data[0].i64.x, 1);
187    data[invocation].i64.xy  = subgroupClusteredAnd(data[1].i64.xy, 1);
188    data[invocation].i64.xyz = subgroupClusteredAnd(data[2].i64.xyz, 1);
189    data[invocation].i64     = subgroupClusteredAnd(data[3].i64, 1);
190
191    data[invocation].i64.x   = subgroupClusteredOr(data[0].i64.x, 1);
192    data[invocation].i64.xy  = subgroupClusteredOr(data[1].i64.xy, 1);
193    data[invocation].i64.xyz = subgroupClusteredOr(data[2].i64.xyz, 1);
194    data[invocation].i64     = subgroupClusteredOr(data[3].i64, 1);
195
196    data[invocation].i64.x   = subgroupClusteredXor(data[0].i64.x, 1);
197    data[invocation].i64.xy  = subgroupClusteredXor(data[1].i64.xy, 1);
198    data[invocation].i64.xyz = subgroupClusteredXor(data[2].i64.xyz, 1);
199    data[invocation].i64     = subgroupClusteredXor(data[3].i64, 1);
200
201    data[invocation].u64.x   = subgroupClusteredAdd(data[0].u64.x, 1);
202    data[invocation].u64.xy  = subgroupClusteredAdd(data[1].u64.xy, 1);
203    data[invocation].u64.xyz = subgroupClusteredAdd(data[2].u64.xyz, 1);
204    data[invocation].u64     = subgroupClusteredAdd(data[3].u64, 1);
205
206    data[invocation].u64.x   = subgroupClusteredMul(data[0].u64.x, 1);
207    data[invocation].u64.xy  = subgroupClusteredMul(data[1].u64.xy, 1);
208    data[invocation].u64.xyz = subgroupClusteredMul(data[2].u64.xyz, 1);
209    data[invocation].u64     = subgroupClusteredMul(data[3].u64, 1);
210
211    data[invocation].u64.x   = subgroupClusteredMin(data[0].u64.x, 1);
212    data[invocation].u64.xy  = subgroupClusteredMin(data[1].u64.xy, 1);
213    data[invocation].u64.xyz = subgroupClusteredMin(data[2].u64.xyz, 1);
214    data[invocation].u64     = subgroupClusteredMin(data[3].u64, 1);
215
216    data[invocation].u64.x   = subgroupClusteredMax(data[0].u64.x, 1);
217    data[invocation].u64.xy  = subgroupClusteredMax(data[1].u64.xy, 1);
218    data[invocation].u64.xyz = subgroupClusteredMax(data[2].u64.xyz, 1);
219    data[invocation].u64     = subgroupClusteredMax(data[3].u64, 1);
220
221    data[invocation].u64.x   = subgroupClusteredAnd(data[0].u64.x, 1);
222    data[invocation].u64.xy  = subgroupClusteredAnd(data[1].u64.xy, 1);
223    data[invocation].u64.xyz = subgroupClusteredAnd(data[2].u64.xyz, 1);
224    data[invocation].u64     = subgroupClusteredAnd(data[3].u64, 1);
225
226    data[invocation].u64.x   = subgroupClusteredOr(data[0].u64.x, 1);
227    data[invocation].u64.xy  = subgroupClusteredOr(data[1].u64.xy, 1);
228    data[invocation].u64.xyz = subgroupClusteredOr(data[2].u64.xyz, 1);
229    data[invocation].u64     = subgroupClusteredOr(data[3].u64, 1);
230
231    data[invocation].u64.x   = subgroupClusteredXor(data[0].u64.x, 1);
232    data[invocation].u64.xy  = subgroupClusteredXor(data[1].u64.xy, 1);
233    data[invocation].u64.xyz = subgroupClusteredXor(data[2].u64.xyz, 1);
234    data[invocation].u64     = subgroupClusteredXor(data[3].u64, 1);
235
236    data[invocation].f16.x   = subgroupClusteredAdd(data[0].f16.x, 1);
237    data[invocation].f16.xy  = subgroupClusteredAdd(data[1].f16.xy, 1);
238    data[invocation].f16.xyz = subgroupClusteredAdd(data[2].f16.xyz, 1);
239    data[invocation].f16     = subgroupClusteredAdd(data[3].f16, 1);
240
241    data[invocation].f16.x   = subgroupClusteredMul(data[0].f16.x, 1);
242    data[invocation].f16.xy  = subgroupClusteredMul(data[1].f16.xy, 1);
243    data[invocation].f16.xyz = subgroupClusteredMul(data[2].f16.xyz, 1);
244    data[invocation].f16     = subgroupClusteredMul(data[3].f16, 1);
245
246    data[invocation].f16.x   = subgroupClusteredMin(data[0].f16.x, 1);
247    data[invocation].f16.xy  = subgroupClusteredMin(data[1].f16.xy, 1);
248    data[invocation].f16.xyz = subgroupClusteredMin(data[2].f16.xyz, 1);
249    data[invocation].f16     = subgroupClusteredMin(data[3].f16, 1);
250
251    data[invocation].f16.x   = subgroupClusteredMax(data[0].f16.x, 1);
252    data[invocation].f16.xy  = subgroupClusteredMax(data[1].f16.xy, 1);
253    data[invocation].f16.xyz = subgroupClusteredMax(data[2].f16.xyz, 1);
254    data[invocation].f16     = subgroupClusteredMax(data[3].f16, 1);
255}
256