! try to get it more cache friendly with multiple loops
    do iy=1,Ny
        do ix=1,Nx
            do i=1,Ntrl
                Aidx = condA(i) + 1
                Bidx = condB(i) + 1

                sumA(ix,iy,Aidx) = sumA(ix,iy,Aidx) + dat(i,ix,iy)
                sumB(ix,iy,Bidx) = sumB(ix,iy,Bidx) + dat(i,ix,iy)
                sumAB(ix,iy,Aidx,Bidx) = sumAB(ix,iy,Aidx,Bidx) + dat(i,ix,iy)
                sumTot(ix,iy) = sumTot(ix,iy) + dat(i,ix,iy)
            enddo
        enddo
    enddo
    ! trial numbers seperately
    do i=1,Ntrl
        Aidx = condA(i) + 1
        Bidx = condB(i) + 1
        NtrlA(Aidx) = NtrlA(Aidx) + 1
        NtrlB(Bidx) = NtrlB(Bidx) + 1
        NtrlAB(Aidx,Bidx) = NtrlAB(Aidx,Bidx) + 1
    enddo