layout(set=0, binding=0) writeonly buffer destBuffer{
    FLOAT data[];
} uOutput;

layout(set=0, binding=1) readonly buffer sourceBuffer0{
    FLOAT data[];
} uInput;

layout(set=0, binding=2) readonly uniform constBuffer {
    ivec4 size; // kw, kh, height, kw*kh*c/4
} uConstant;

layout (local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
#define UP_DIV(x, y) (((x)+(y)-1)/(y))

void main()
{
    ivec3 posTmp = ivec3(gl_GlobalInvocationID);
    if (posTmp.x < uConstant.size.w)
    {
        ivec2 pos;
        int kernelSize = uConstant.size.x * uConstant.size.y;
        pos.x = posTmp.x % kernelSize;
        pos.y = posTmp.x / kernelSize;
        int outputPos = pos.x * 4 + (pos.y / 4) * kernelSize * 4 + (pos.y % 4);
        uOutput.data[outputPos] = uInput.data[posTmp.x];
    }
}