cuda::experimental::places::blocked_partition

using cuda::experimental::places::blocked_partition = blocked_partition_custom<>

Partitions a multidimensional box or shape into contiguous blocks along a selected dimension.

This partitioning strategy divides the data space into contiguous blocks, distributing them across execution places. By default, partitioning occurs along the last dimension, but a specific dimension can be selected using the template parameter. This approach ensures good spatial locality and is particularly effective for regular data access patterns.

When mapping element coordinates (get_executor), the selected dimension is clamped to the highest axis whose extent is greater than one: -1 always selects that axis, and a larger explicit dimension is clamped down to it (e.g. blocked_partition_custom<2> on extents {n, 1, 1, 1} partitions along axis 0).