#undef FLOAT4
#define FLOAT4 ivec4
layout(set=0, binding=0) writeonly buffer destBuffer{
    FLOAT4 data[];
} uOutput;

layout(set=0, binding=1) readonly buffer sourceBuffer0{
    FLOAT4 data[];
} uInput;
layout(set=0, binding=2) uniform constBuffer{
    ivec4 size; // x: limit, y: channelC4*b, z:height, w:width
    vec4 slope;
} uConstant;

//for dynamic change threads counts from outside
// from vkCreateComputePipelines->VkComputePipelineCreateInfo->VkPipelineShaderStageCreateInfo->VkSpecializationInfo->VkSpecializationMapEntry
// layout(local_size_x_id = 1, local_size_y_id = 2, local_size_z_id = 3) in;

layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;

void main()
{
    ivec3 posTmp = ivec3(gl_GlobalInvocationID);
    if (posTmp.x < uConstant.size.x) {
        FLOAT4 value = uInput.data[posTmp.x];
#ifdef NEG
        value = -value;
#endif
#ifdef SIGN
        value = sign(value);
#endif
#ifdef ABS
        value = abs(value);
#endif
#ifdef SQUARE
        value = value * value;
#endif
        uOutput.data[posTmp.x] = value;
    }
}