9 RWStructuredBuffer<Types> data;
\r
11 [numthreads(32, 16, 1)]
\r
12 void CSMain(uint3 dti : SV_DispatchThreadID)
\r
14 data[dti.x].u = WaveActiveSum(data[dti.x].u);
\r
15 data[dti.x].u.x = WaveActiveSum(data[dti.x].u.x);
\r
16 data[dti.x].u.xy = WaveActiveSum(data[dti.x].u.xy);
\r
17 data[dti.x].u.xyz = WaveActiveSum(data[dti.x].u.xyz);
\r
19 data[dti.x].i = WaveActiveSum(data[dti.x].i);
\r
20 data[dti.x].i.x = WaveActiveSum(data[dti.x].i.x);
\r
21 data[dti.x].i.xy = WaveActiveSum(data[dti.x].i.xy);
\r
22 data[dti.x].i.xyz = WaveActiveSum(data[dti.x].i.xyz);
\r
24 data[dti.x].f = WaveActiveSum(data[dti.x].f);
\r
25 data[dti.x].f.x = WaveActiveSum(data[dti.x].f.x);
\r
26 data[dti.x].f.xy = WaveActiveSum(data[dti.x].f.xy);
\r
27 data[dti.x].f.xyz = WaveActiveSum(data[dti.x].f.xyz);
\r
29 data[dti.x].d = WaveActiveSum(data[dti.x].d);
\r
30 data[dti.x].d.x = WaveActiveSum(data[dti.x].d.x);
\r
31 data[dti.x].d.xy = WaveActiveSum(data[dti.x].d.xy);
\r
32 data[dti.x].d.xyz = WaveActiveSum(data[dti.x].d.xyz);
\r
34 data[dti.x].u = WaveActiveProduct(data[dti.x].u);
\r
35 data[dti.x].u.x = WaveActiveProduct(data[dti.x].u.x);
\r
36 data[dti.x].u.xy = WaveActiveProduct(data[dti.x].u.xy);
\r
37 data[dti.x].u.xyz = WaveActiveProduct(data[dti.x].u.xyz);
\r
39 data[dti.x].i = WaveActiveProduct(data[dti.x].i);
\r
40 data[dti.x].i.x = WaveActiveProduct(data[dti.x].i.x);
\r
41 data[dti.x].i.xy = WaveActiveProduct(data[dti.x].i.xy);
\r
42 data[dti.x].i.xyz = WaveActiveProduct(data[dti.x].i.xyz);
\r
44 data[dti.x].f = WaveActiveProduct(data[dti.x].f);
\r
45 data[dti.x].f.x = WaveActiveProduct(data[dti.x].f.x);
\r
46 data[dti.x].f.xy = WaveActiveProduct(data[dti.x].f.xy);
\r
47 data[dti.x].f.xyz = WaveActiveProduct(data[dti.x].f.xyz);
\r
49 data[dti.x].d = WaveActiveProduct(data[dti.x].d);
\r
50 data[dti.x].d.x = WaveActiveProduct(data[dti.x].d.x);
\r
51 data[dti.x].d.xy = WaveActiveProduct(data[dti.x].d.xy);
\r
52 data[dti.x].d.xyz = WaveActiveProduct(data[dti.x].d.xyz);
\r
54 data[dti.x].u = WaveActiveMin(data[dti.x].u);
\r
55 data[dti.x].u.x = WaveActiveMin(data[dti.x].u.x);
\r
56 data[dti.x].u.xy = WaveActiveMin(data[dti.x].u.xy);
\r
57 data[dti.x].u.xyz = WaveActiveMin(data[dti.x].u.xyz);
\r
59 data[dti.x].i = WaveActiveMin(data[dti.x].i);
\r
60 data[dti.x].i.x = WaveActiveMin(data[dti.x].i.x);
\r
61 data[dti.x].i.xy = WaveActiveMin(data[dti.x].i.xy);
\r
62 data[dti.x].i.xyz = WaveActiveMin(data[dti.x].i.xyz);
\r
64 data[dti.x].f = WaveActiveMin(data[dti.x].f);
\r
65 data[dti.x].f.x = WaveActiveMin(data[dti.x].f.x);
\r
66 data[dti.x].f.xy = WaveActiveMin(data[dti.x].f.xy);
\r
67 data[dti.x].f.xyz = WaveActiveMin(data[dti.x].f.xyz);
\r
69 data[dti.x].d = WaveActiveMin(data[dti.x].d);
\r
70 data[dti.x].d.x = WaveActiveMin(data[dti.x].d.x);
\r
71 data[dti.x].d.xy = WaveActiveMin(data[dti.x].d.xy);
\r
72 data[dti.x].d.xyz = WaveActiveMin(data[dti.x].d.xyz);
\r
74 data[dti.x].u = WaveActiveMax(data[dti.x].u);
\r
75 data[dti.x].u.x = WaveActiveMax(data[dti.x].u.x);
\r
76 data[dti.x].u.xy = WaveActiveMax(data[dti.x].u.xy);
\r
77 data[dti.x].u.xyz = WaveActiveMax(data[dti.x].u.xyz);
\r
79 data[dti.x].i = WaveActiveMax(data[dti.x].i);
\r
80 data[dti.x].i.x = WaveActiveMax(data[dti.x].i.x);
\r
81 data[dti.x].i.xy = WaveActiveMax(data[dti.x].i.xy);
\r
82 data[dti.x].i.xyz = WaveActiveMax(data[dti.x].i.xyz);
\r
84 data[dti.x].f = WaveActiveMax(data[dti.x].f);
\r
85 data[dti.x].f.x = WaveActiveMax(data[dti.x].f.x);
\r
86 data[dti.x].f.xy = WaveActiveMax(data[dti.x].f.xy);
\r
87 data[dti.x].f.xyz = WaveActiveMax(data[dti.x].f.xyz);
\r
89 data[dti.x].d = WaveActiveMax(data[dti.x].d);
\r
90 data[dti.x].d.x = WaveActiveMax(data[dti.x].d.x);
\r
91 data[dti.x].d.xy = WaveActiveMax(data[dti.x].d.xy);
\r
92 data[dti.x].d.xyz = WaveActiveMax(data[dti.x].d.xyz);
\r
94 data[dti.x].u = WaveActiveBitAnd(data[dti.x].u);
\r
95 data[dti.x].u.x = WaveActiveBitAnd(data[dti.x].u.x);
\r
96 data[dti.x].u.xy = WaveActiveBitAnd(data[dti.x].u.xy);
\r
97 data[dti.x].u.xyz = WaveActiveBitAnd(data[dti.x].u.xyz);
\r
99 data[dti.x].i = WaveActiveBitAnd(data[dti.x].i);
\r
100 data[dti.x].i.x = WaveActiveBitAnd(data[dti.x].i.x);
\r
101 data[dti.x].i.xy = WaveActiveBitAnd(data[dti.x].i.xy);
\r
102 data[dti.x].i.xyz = WaveActiveBitAnd(data[dti.x].i.xyz);
\r
104 data[dti.x].u = WaveActiveBitOr(data[dti.x].u);
\r
105 data[dti.x].u.x = WaveActiveBitOr(data[dti.x].u.x);
\r
106 data[dti.x].u.xy = WaveActiveBitOr(data[dti.x].u.xy);
\r
107 data[dti.x].u.xyz = WaveActiveBitOr(data[dti.x].u.xyz);
\r
109 data[dti.x].i = WaveActiveBitOr(data[dti.x].i);
\r
110 data[dti.x].i.x = WaveActiveBitOr(data[dti.x].i.x);
\r
111 data[dti.x].i.xy = WaveActiveBitOr(data[dti.x].i.xy);
\r
112 data[dti.x].i.xyz = WaveActiveBitOr(data[dti.x].i.xyz);
\r
114 data[dti.x].u = WaveActiveBitXor(data[dti.x].u);
\r
115 data[dti.x].u.x = WaveActiveBitXor(data[dti.x].u.x);
\r
116 data[dti.x].u.xy = WaveActiveBitXor(data[dti.x].u.xy);
\r
117 data[dti.x].u.xyz = WaveActiveBitXor(data[dti.x].u.xyz);
\r
119 data[dti.x].i = WaveActiveBitXor(data[dti.x].i);
\r
120 data[dti.x].i.x = WaveActiveBitXor(data[dti.x].i.x);
\r
121 data[dti.x].i.xy = WaveActiveBitXor(data[dti.x].i.xy);
\r
122 data[dti.x].i.xyz = WaveActiveBitXor(data[dti.x].i.xyz);
\r
124 data[dti.x].u.x = WaveActiveCountBits(data[dti.x].u.x == 0);
\r