[Gretl-devel] slowness with panel sample restrictions (?) or pxnobs() function (?)

hi, in relation to the feature request about cross-section dependence testing in panels I've encountered some performance bottleneck. Here's working code given some panel variable 'u': <hansl> series unit = $obsmajor N = max(unit) #### more elegant attempt (but slow for some reason...) rhos = zeros(N,N) Tcomms = zeros(N,N) loop i=1..N -q loop j=(i+1)..N -q smpl (unit == i || unit == j) --restrict --replace smpl pxnobs(u) == 2 --restrict # both units have values there matrix uij = { u } # hope it's stacked, so that upper half is unit i Tcomms[i,j] = rows(uij) / 2 rhos[i,j] = corr( uij[1: Tcomms[i,j]], uij[Tcomms[i,j] + 1 : ] ) endloop endloop # Pesaran's CD (unbalanced) scalar CD = sqrt( 2 / (N*(N-1)) ) * sum( sqrt(Tcomms) .* rhos ) </hansl> This takes quite long to run (several seconds for datasets like abdata.gdt). I've had a less elegant implementation before with more code and more copying to temporary matrices and so on (including the nested loop) which was much faster. I'm wondering whether perhaps the sample restriction with the pxnobs() function is slow. thanks, sven