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

layout(set=0, binding=1) readonly buffer s0{
    FLOAT data[];
} uInput;

layout(set=0, binding=2) readonly buffer s1{
    int data[];
} uSrcOffset;

layout(set=0, binding=3) readonly buffer s2{
    int data[];
} uDstOffset;

layout(set=0, binding=4) uniform constBuffer{
    ivec4 stride;
    ivec4 size;
    ivec4 extent;
    ivec4 step;
    ivec4 iter;
} uConstant;

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.step.w)
    {
        ivec3 pos;
        int regionInsideIndex = posTmp.x % uConstant.size.w;
        int regionOutsideIndex = posTmp.x / uConstant.size.w;
        pos.x = regionInsideIndex / (uConstant.size.y * uConstant.size.z);
        int subIndex = regionInsideIndex % (uConstant.size.y * uConstant.size.z);
        pos.z = subIndex % uConstant.size.z;
        pos.y = subIndex / uConstant.size.z;
        int srcBasicOffset, dstBasicOffset;
        if (uConstant.iter.y > 0) {
            srcBasicOffset = uConstant.step.y * int(uSrcOffset.data[regionOutsideIndex]);
        } else {
            srcBasicOffset = uConstant.step.y * regionOutsideIndex;;
        }
        if (uConstant.iter.x > 0) {
            dstBasicOffset = uConstant.step.x * int(uDstOffset.data[regionOutsideIndex]);
        } else {
            dstBasicOffset = uConstant.step.x * regionOutsideIndex;;
        }
        int srcOffset = srcBasicOffset
            + uConstant.stride.w
            + uConstant.stride.z * pos.z
            + uConstant.stride.y * pos.y
            + uConstant.stride.x * pos.x
            ;
        
        int dstOffset = dstBasicOffset
            + uConstant.extent.w
            + pos.x * uConstant.extent.x
            + pos.y * uConstant.extent.y
            + pos.z * uConstant.extent.z
            ;
        uOutput.data[dstOffset] = uInput.data[srcOffset];
    }
}