Through the use of the dimensional splitting ''cascade'' method of gri
d-to-grid interpolation, it is shown that consistently high-order-accu
rate semi-Lagrangian integration of a three-dimensional hydrostatic pr
imitive equations model can be carried out using forward (downstream)
trajectories instead of the backward (upstream) trajectory computation
s that are more commonly employed in semi-Lagrangian models. Apart fro
m the efficiency resulting directly from the adoption of the cascade m
ethod, improved computational performance is achieved partly by the se
lective implicit treatment of only the deepest vertical gravity modes
and partly by obviating the need to iterate the estimation of each tra
jectory's location. Perhaps the main distinction of our present semi-L
agrangian method is its inherent exact conservation of mass and passiv
e tracers. This is achieved by adopting a simple variant of the cascad
e interpolation that incorporates mass (and tracer) conservation direc
tly and at only a very modest additional cost. The conserving cascade,
which is described in detail, is a generic algorithm that can be appl
ied at arbitrary order of accuracy. Tests of the new mass-conserving s
cheme in a regional forecast model show small but consistent improveme
nts in accuracy at 48 h. It is suggested that the benefits to extended
global forecasting and simulation should be much greater.