The first area of focus for reduced-precision work will be spatial searches and optimization methods.
Spatial search acceleration data structures are one obvious target for use of reduced precision. For
example, on meshed geometries, spatial searches using reduced-precision bounding boxes can be
used to produce the same intersection tests on triangle primitives. We will explore similar
techniques for utilized reduced-precision operations in other search data structures while preserving
the accuracy and robustness of the results. In addition, variance reduction methods also represent
an opportunity to leverage reduced precision in order to improve performance while, at the same
time, reducing memory use. For example, reduced-precision weight window parameters would
be straightforward way to reduce memory usage, but a direct conversion may result in changes
to the total weight of all particles in the simulation.
The first area of focus for reduced-precision work will be spatial searches and optimization methods.
Spatial search acceleration data structures are one obvious target for use of reduced precision. For
example, on meshed geometries, spatial searches using reduced-precision bounding boxes can be
used to produce the same intersection tests on triangle primitives. We will explore similar
techniques for utilized reduced-precision operations in other search data structures while preserving
the accuracy and robustness of the results. In addition, variance reduction methods also represent
an opportunity to leverage reduced precision in order to improve performance while, at the same
time, reducing memory use. For example, reduced-precision weight window parameters would
be straightforward way to reduce memory usage, but a direct conversion may result in changes
to the total weight of all particles in the simulation.