Use precise grid-space ray-tracing about mcx HOT 3 CLOSED

Implementation is progress. Initial result is very positive. MCX and MMC now match perfectly for both homogeneous and heterogeneous cases with scattering contrasts.

The boundary reflection code was also significantly simplified due to this new change.

Computational speed did get a hit - about 50% slow-down for Fermi (GTX 590) and 30% for Maxwell (GTX980Ti). NVVP profiler shows hot spots in the new ray-grid ray-tracing code. Need to further optimize to improve the speed.

Stay tuned.

from mcx.

Changes are completed

c7f9625

the precise ray-tracing algorithm does make a difference when an inclusion present scattering contrast (either in mus or in g). New algorithm returns smooth fluence profiles while the old algorithm generate a small discontinuity.

The computational speed was improved guided by NVVP profiling results. A bunch of register variables, which were actually stored in the local/global memory, were optimized by using shared memory. see my forum post at

https://devtalk.nvidia.com/default/topic/868118/?comment=4646254

These changes improved the performance on Maxwell by 50%; however, there is no noticeable speed improvement on Fermi or Kepler.

The old way of photon fluence accumulation, i.e. adding the photon weight to enclosing voxel at a fixed length (1mm) can still be simulated by setting "-F 1" or "--faststep 1". By default, faststep is disabled.

from mcx.

issue now closed

from mcx.