#version 440 core
layout(std430) buffer;

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

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

layout(set=0, binding=2) uniform constBuffer {
    ivec4 size;
    vec4 scale;
} uConstant;

layout (local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
void main()
{
	int pos = int(gl_GlobalInvocationID.x);
	// input output grid layout is NC4HW4

	ivec4 size = uConstant.size;
    if (pos < size.x) {
        uOutput.data[pos] = int(uInput.data[pos] * uConstant.scale.x + uConstant.scale.y);
    }
}