cupynumeric.ndarray.stencil_hint#
- ndarray.stencil_hint( ) None#
Inform cuPyNumeric that this array will be used in a stencil computation in the following code.
This allows cuPyNumeric to allocate space for the “ghost” elements ahead of time, rather than discovering the full extent of accesses incrementally, and thus avoid intermediate copies.
For example, let’s say we have a 1-D array A of size 10 and we want to partition A across two GPUs. By default, A would be partitioned equally and each GPU gets an instance of size 5 (GPU0 gets elements 0-4, and GPU1 gets 5-9 inclusive). Suppose we use A in the stencil computation B = A[:9] + A[1:]. The runtime would now need to adjust the partitioning such that GPU0 has elements 0-5 and GPU1 has elements 4-9 inclusive. Since the original instance on GPU0 does not cover index 5, cuPyNumeric needs to allocate a full new instance that covers 0-5, leading to an extra copy. In this case, if the code calls A.stencil_hint([1], [1]) to pre-allocate instances that contain the extra elements before it uses A, the extra copies can be avoided.
- Parameters:
Notes
This function currently does not behave as expected in the case where multiple CPU/OpenMP processors use the same system memory.
- Availability:
Multiple CPUs, Multiple GPUs