Closed gmarkall closed 12 years ago
Great! Does it show any performance improvements?
It shaves about 10-15% off the simulation time for the advection, diffusion and advection-diffusion test cases. However, they're pretty small tests so I'd expect more speedup with bigger simulations.
Presently we transfer the fields from the GPU back to the host at the end of every timestep. This is inefficent. Instead, these changes only transfer the fields back to the host when a call to output_state occurs.
A summary of the changes are:
These changes have passed the tests on the fluidity-dev queue on the buildbot.