triton.experimental.gluon.language.amd.cdna5.PartitionedSharedLayout

class triton.experimental.gluon.language.amd.cdna5.PartitionedSharedLayout(self, num_partitions: int, num_groups: int, partition_dim: int, partition_layout: SharedLayout)

Represents a partitioned shared memory layout that splits a tensor across multiple physical shared memory partitions.

This reduces shared memory partition conflicts by placing different pieces of a tensor in separate physical memory slots.

Parameters:

• num_partitions (int) – Number of physical memory partitions.

• num_groups (int) – Number of groups (each group has num_partitions pieces).

• partition_dim (int) – Dimension along which to partition.

• partition_layout (SharedLayout) – Inner layout for each piece (e.g., SwizzledSharedLayout or PaddedSharedLayout).

__init__(self, num_partitions: int, num_groups: int, partition_dim: int, partition_layout: SharedLayout) → None

Methods

__init__(self, num_partitions, num_groups, ...)
format_hardware_view(self, shape)
format_tensor_view(self, shape)
mangle(self)

Attributes

type
num_partitions
num_groups
partition_dim
partition_layout