1#version 450
 2
 3#include "generic_head.glsl"
 4#include "types.glsl"
 5
 6#extension GL_EXT_control_flow_attributes : enable
 7
 8#define FLT_MAX 3.402823466e+38F
 9
10layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
11
12layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
13layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
14
15layout (constant_id = 0) const uint BLOCK_SIZE = 32;
16
17shared FLOAT_TYPE tmpmax[BLOCK_SIZE];
18shared uint tmp[BLOCK_SIZE];
19
20void main() {
21    const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
22    const uint col = gl_LocalInvocationID.x;
23
24    if (row >= p.KY) {
25        return;
26    }
27
28    A_TYPE amax = -FLT_MAX;
29    uint acol = col;
30
31    if (col < p.KX) {
32        amax = data_a[row*p.KX + col];
33    }
34
35    for (uint i = col + BLOCK_SIZE; i < p.KX; i += BLOCK_SIZE) {
36        A_TYPE val = data_a[row*p.KX + i];
37        if (val > amax) {
38            amax = val;
39            acol = i;
40        }
41    }
42
43    tmp[col] = acol;
44    tmpmax[col] = amax;
45
46    barrier();
47    [[unroll]] for (int s = int(BLOCK_SIZE) / 2; s > 0; s >>= 1) {
48        if (col < s && col + s < p.KX) {
49            if (tmpmax[col] < tmpmax[col + s]) {
50                tmpmax[col] = tmpmax[col + s];
51                tmp[col] = tmp[col + s];
52            }
53        }
54        barrier();
55    }
56
57    if (col == 0) {
58        data_d[row] = D_TYPE(tmp[0]);
59    }
60}