-
Notifications
You must be signed in to change notification settings - Fork 3
halide_notes
-
the organization of computations and data for a given algorithm is constrained by a fundamental tension between parallelism, locality, and redundant computation of shared values
- This is addressed by a systematic model of schedules.
-
Stencil computation - repeated application of kernel across a dataset.
-
I believe the right way to program image processing pipelines is to separate the intrinsic algorithm—what is computed—from the concerns of e�ciently organizing it for machine execution—decisions about storage and the ordering of computation.
-
this is also why libraries of optimized code cannot deliver efficient performance when building real image processing pipelines: individually optimized subroutines do not compose into an optimized whole, since they cannot reorganize computation for locality or parallel execution across function boundaries.