#version 440 core
layout(std430) buffer;

layout(set=0, binding=0) writeonly buffer pointoffsetSum{
    highp uvec4 data[];
} uPointoffsetSum;


layout(set=0, binding=1) readonly buffer pointoffset{
    highp uvec4 data[];
} uPointoffset;

layout(set=0, binding=2) uniform constBuffer {
    ivec4 point; // point size
} uConstant;

layout (constant_id = 3) const int UNIT = 128;
layout (constant_id = 4) const int LOCAL_SIZE = 1;
layout (local_size_x = 1, local_size_y = 1, local_size_z = 1) in;

void main()
{
    int tId = int(gl_LocalInvocationID.x);
    int size = (uConstant.point.x + 3) / 4;
    int curOffset = 0;
    uvec4 threadBuffer[UNIT];
    uint sum = 0;
    for (int i=0; i<size; ++i) {
        uvec4 p0 = uPointoffset.data[i];
        p0.y = p0.y + p0.x;
        p0.z = p0.z + p0.y;
        p0.w = p0.w + p0.z;
        p0 = p0 + uvec4(sum);
        uPointoffsetSum.data[i] = p0;
        sum = p0.w;
    }
    return;
}