#version 440 core
layout(std140) buffer;
layout(set=0, binding=0) writeonly uniform image2D uOutput;
layout(set=0, binding=1) uniform sampler2D uInput;

layout(set = 0, binding = 2) readonly buffer scaleBuffer{
    vec4 data[];
}uScale;

layout(set = 0, binding = 3) readonly buffer biasBuffer{
    vec4 data[];
}uBias;

layout(set = 0, binding = 4) uniform constBuffer{
    ivec4 imgSize;
}uConst;

layout(local_size_x = 16, local_size_y = 16, local_size_z = 1) in;

void main()
{
    ivec3 pos = ivec3(gl_GlobalInvocationID);
    ivec3 imgSize = uConst.imgSize.xyz;

    if(all(lessThan(pos.xy, imgSize.xy)))
    {
        int channelIndex = pos.z % uConst.imgSize.z;
        int batchIndex = pos.z / uConst.imgSize.z;
        vec4 scale = uScale.data[channelIndex];
        vec4 bias = uBias.data[channelIndex];
        ivec2 imgPos = ivec2(pos.x+channelIndex*uConst.imgSize.x, pos.y+batchIndex*uConst.imgSize.y);

        vec4 color = texelFetch(uInput, imgPos, 0);
        vec4 res = color * scale + bias;
        imageStore(uOutput, imgPos, res);
    }

}