#version 450 core
layout(std430) buffer;

layout(set=0, binding=0) writeonly uniform image2D uOutput;
layout(set=0, binding=1) uniform sampler2D uInput;
layout(set=0, binding=2) uniform sampler2D uSlope;

layout(set = 0, binding = 3) uniform reluBuffer{
    ivec4 imgSize;
    float slope;
}uPreluParam;

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

void main()
{
	ivec3 pos = ivec3(gl_GlobalInvocationID);
	ivec3 imgSize = uPreluParam.imgSize.xyz;
	int oz = pos.z % imgSize.z;
    int ob = pos.z / imgSize.z;
    if(pos.x < imgSize.x && pos.y < imgSize.y)
	{
        ivec2 imgPos = ivec2(pos.x + oz * uPreluParam.imgSize.x, pos.y + ob * uPreluParam.imgSize.y);
		vec4 slope = texelFetch(uSlope, ivec2(oz, 0), 0);
		vec4 dataIn = texelFetch(uInput, imgPos, 0);
		vec4 dataTemp = dataIn * slope;
		bvec4 lessZero = bvec4(lessThan(dataIn, vec4(0.0)));
		imageStore(uOutput, imgPos, mix(dataIn, dataTemp, lessZero));
	}

}