Module tensor 

Re-exports

pub use identity::*;

pub use view::*;

Modules

identity

index_offset_contiguous

index_offset_contiguous_fastdivmod

index_offset_with_layout

layout

view

virtual

Structs

ErasedTensor

A tensor that has forgotten its line width. See the module docs.

ErasedTensorExpand

Expand type for ErasedTensor.

ErasedView

Any view, at a width named once here and forgotten after.

ErasedViewMut

Any mutable view. As ErasedView, with the write half as well.

TensorHandle

Tensor representation containing a server handle as well as basic tensor metadata.,

WriteOnly

Visibility marker for a destination that is written and never read.

Enums

MatrixBatchLayout

Layout for matrix batch tensors, i.e. tensors whose interpretation is a bunch of batched matrices of 2 dimensions

Constants

NUM_SM_APPROX

Traits

ErasedBacking

A backing that can be wrapped at visibility IO.

ErasedIo

The visibility markers an ErasedTensor accepts.

ErasedIoRead

A visibility that permits reads.

ErasedIoWrite

A visibility that permits writes.

ErasedTensorOperationsExpand

What a backing has to answer to be an ErasedTensor.

ReadsLines

A backing whose __expand_read_line_method is real rather than the default.

WritesLines

A backing whose __expand_write_line_method is real rather than the default.

Functions

compact_strides

copy_gpu_ref

Make a jit tensor contiguous.

copy_into

Copies the input tensor into the output tensor following the strides.

index_offset_contiguous

Returns the offset of the tensor corresponding to a contiguous layout.

index_offset_contiguous_fastdivmod

Returns the offset of the tensor corresponding to a contiguous layout.

index_offset_with_layout

Returns the offset of the tensor corresponding to the layout tensor.

into_contiguous

Make a jit tensor contiguous.

into_contiguous_packed

Make a jit tensor contiguous, using the pitched allocator if available. See create_tensor. Handles unpacking and repacking packed tensors (i.e. quantized values). shape refers to the actual (unpacked) shape of the tensor, while packing specifies the number of elements in each storage element.

into_contiguous_packed_ref

Make a jit tensor contiguous.

into_contiguous_pitched

Make a jit tensor contiguous, using the pitched allocator if available. See create_tensor.

is_contiguous

Checks if the tensor associated with the given shape and strides is contiguous.

is_contiguous_pitched

Checks if a tensor is only strided on the last dimension, and could be safely reinterpreted as a 2D tensor with unit stride on the last dimension. This will always hold for non-permuted tensors allocated on a runtime.

launch_copy_perpendicular_ref

Launches the perpendicular contiguous kernel.

launch_into_contiguous_perpendicular

Launches the perpendicular contiguous kernel.

matrix_batch_layout

Return the layout of a matrix batch given the strides.