triton.experimental.gluon.language.nvidia.blackwell.tensor_memory_descriptor

class triton.experimental.gluon.language.nvidia.blackwell.tensor_memory_descriptor(self, handle, element_ty, shape, layout, alloc_shape)

Represents a tensor memory descriptor handle for Tensor Core Gen5 operations.

__init__(self, handle, element_ty, shape, layout, alloc_shape)

Methods

__init__(self, handle, element_ty, shape, ...)

get_reg_layout(self[, num_warps, ...])

Return the register layout used to access this tensor memory descriptor.

index(self, index[, _semantic])

Create a subview of tensor memory by indexing the first dimension.

load(self[, layout, _semantic, _generator])

Load a tensor from tensor memory.

load_max(self[, layout, abs, propagate_nan, ...])

Load a tensor from tensor memory with MAX reduction along the N-dimension.

load_min(self[, layout, abs, propagate_nan, ...])

Load a tensor from tensor memory with MIN reduction along the N-dimension.

slice(self, start, length[, _semantic])

Create a slice of the tensor memory descriptor along the last dimension.

store(self, value[, pred, _semantic])

Store a tensor into tensor memory.

Attributes

dtype

layout

rank

shape

type