1#version 450
 2
 3#include "generic_head.glsl"
 4#include "types.glsl"
 5
 6#extension GL_EXT_control_flow_attributes : enable
 7#define BLOCK_SIZE 512
 8
 9layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
10
11layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
12layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
13
14shared float tmp[BLOCK_SIZE];
15
16void main() {
17    const uint group_size = p.KX;
18    const float eps = p.param1;
19
20    const uint tid = gl_LocalInvocationID.x;
21    const uint start = gl_WorkGroupID.x * group_size + tid;
22    const uint end = (gl_WorkGroupID.x + 1) * group_size;
23
24    tmp[tid] = 0.0f;
25
26    // Calculate mean
27    [[unroll]] for (uint col = start; col < end; col += BLOCK_SIZE) {
28        tmp[tid] += float(data_a[col]);
29    }
30
31    // tmp up partial tmps and write back result
32    barrier();
33    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
34        if (tid < s) {
35            tmp[tid] += tmp[tid + s];
36        }
37        barrier();
38    }
39
40    const float mean = tmp[0] / group_size;
41    barrier();
42    tmp[tid] = 0.0f;
43
44    // Calculate variance
45    [[unroll]] for (uint col = start; col < end; col += BLOCK_SIZE) {
46        const float xi = float(data_a[col]) - mean;
47        data_d[col] = D_TYPE(xi);
48        tmp[tid] += xi * xi;
49    }
50
51    // sum up partial sums and write back result
52    barrier();
53    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
54        if (tid < s) {
55            tmp[tid] += tmp[tid + s];
56        }
57        barrier();
58    }
59
60    const float variance = tmp[0] / group_size;
61    const float scale = inversesqrt(variance + eps);
62
63    [[unroll]] for (uint col = start; col < end; col += BLOCK_SIZE) {
64        data_d[col] *= D_TYPE(scale);
65    }
66}