cuda::experimental::places::make_partition_placement_provider#

inline auto cuda::experimental::places::make_partition_placement_provider(
const cute_partition_descriptor &partition,
dim4 data_dims,
size_t total_size,
size_t elemsize
)#

Composite data place whose partitioner is a cute_partition object.

Like data_place_composite but ownership is defined by the partition object value (a bare partition_fn_t cannot carry its leaves), so a canonical descriptor is stored on the place. Because a padded partition is intrinsically specific to one tensor, such a place is per-tensor by nature. This place supports both shaped raw allocations (allocate_nd(data_dims, elemsize)) and STF logical data.

Owner provider for localized_array from a partition descriptor.

Tries the analytic block plan first (exact owners, byte-true placement statistics: the sample counters then hold byte counts); falls back to the sampled majority vote for layouts denser than the placement blocks.