cute::Layout - CUTLASS

Overview

cute::Layout is the fundamental abstraction in CuTe for describing the mapping between logical multi-dimensional coordinates and linear memory offsets. A Layout is defined by a Shape (extents in each dimension) and a Stride (steps between adjacent elements in each dimension).

Class Template

template <class Shape, class Stride = LayoutLeft::Apply<Shape>>
struct Layout : private cute::tuple<Shape, Stride>
{
  static constexpr int rank = rank_v<Shape>;
  
  CUTE_HOST_DEVICE constexpr
  Layout(Shape const& shape = {}, Stride const& stride = {});
};

Template Parameters

Shape

IntTuple

The shape of the layout, specifying the extent in each mode. Can be a single integer or a hierarchical tuple of integers.

Stride

IntTuple

default:"LayoutLeft::Apply<Shape>"

The stride of the layout, specifying the step size in each mode. Defaults to column-major (LayoutLeft) ordering.

Source Location

include/cute/layout.hpp:98-311

Member Functions

Accessors

shape()

template <int... I>
CUTE_HOST_DEVICE constexpr decltype(auto)
shape() const;

Returns the shape of the layout or a specific mode. Example:

auto layout = make_layout(make_shape(16, 32), make_stride(1, 16));
auto s = layout.shape();      // Returns (16, 32)
auto s0 = layout.shape<0>();  // Returns 16

stride()

template <int... I>
CUTE_HOST_DEVICE constexpr decltype(auto)
stride() const;

Returns the stride of the layout or a specific mode. Example:

auto layout = make_layout(make_shape(16, 32), make_stride(1, 16));
auto d = layout.stride();      // Returns (1, 16)
auto d0 = layout.stride<0>();  // Returns 1

Mapping Operations

operator()

template <class Coord>
CUTE_HOST_DEVICE constexpr auto
operator()(Coord const& coord) const;

Maps a logical coordinate to a linear index, or slices the layout if the coordinate contains _ (underscore). Parameters:

coord: Logical coordinate (single integer or tuple)

Returns: Linear index (integer) or sliced layout Example:

auto layout = make_layout(make_shape(4, 8), make_stride(1, 4));
int idx = layout(make_coord(2, 3));  // Returns 2 + 3*4 = 14
auto sub = layout(make_coord(_, 3)); // Returns sliced layout

Composition

compose()

template <class OtherLayout>
CUTE_HOST_DEVICE constexpr auto
compose(OtherLayout const& other) const;

Composes this layout with another layout: result(c) = this(other(c)). Example:

auto outer = make_layout(make_shape(64));
auto inner = make_layout(make_shape(8, 8));
auto composed = outer.compose(inner);

with_shape()

template <class OtherShape>
CUTE_HOST_DEVICE constexpr auto
with_shape(OtherShape const& shape) const;

Composes this layout with a layout having the given shape.

Tiling

tile()

template <class OtherLayout>
CUTE_HOST_DEVICE constexpr auto
tile(OtherLayout const& other) const;

Tiles this layout by dividing it into blocks defined by the other layout. Example:

auto layout = make_layout(make_shape(64, 64));
auto tiled = layout.tile(make_shape(16, 16));
// Result has shape ((16,16), (4,4))

Coordinate Conversion

get_hier_coord()

template <class IInt>
CUTE_HOST_DEVICE constexpr auto
get_hier_coord(IInt const& idx) const;

Converts a linear index to hierarchical logical coordinates. Returns: Coordinate with the same hierarchical structure as the shape

get_flat_coord()

template <class IInt>
CUTE_HOST_DEVICE constexpr auto
get_flat_coord(IInt const& idx) const;

Converts a linear index to flat logical coordinates. Returns: Flat rank-N coordinate tuple

get_1d_coord()

template <class IInt>
CUTE_HOST_DEVICE constexpr auto
get_1d_coord(IInt const& idx) const;

Converts a linear index to a generalized column-major 1D coordinate. Returns: Single integer coordinate

Factory Functions

make_layout()

template <class Shape, class Stride>
CUTE_HOST_DEVICE constexpr auto
make_layout(Shape const& shape, Stride const& stride);

template <class Shape>
CUTE_HOST_DEVICE constexpr auto
make_layout(Shape const& shape);

Creates a Layout with the specified shape and stride. If stride is omitted, defaults to LayoutLeft (column-major). Example:

// Explicit shape and stride
auto layout1 = make_layout(make_shape(16, 32), 
                           make_stride(1, 16));

// Column-major default
auto layout2 = make_layout(make_shape(16, 32));
// Equivalent to: make_stride(1, 16)

// Row-major
auto layout3 = make_layout(make_shape(16, 32), LayoutRight{});

make_identity_layout()

template <class Shape>
CUTE_HOST_DEVICE constexpr auto
make_identity_layout(Shape const& shape);

Creates an identity layout where layout(i) = i.

make_layout_like()

template <class Shape, class Stride>
CUTE_HOST_DEVICE constexpr auto
make_layout_like(Layout<Shape,Stride> const& layout);

Creates a compact layout with the same shape and stride ordering as the input layout.

Utility Functions

size()

template <int... Is, class Shape, class Stride>
CUTE_HOST_DEVICE constexpr auto
size(Layout<Shape,Stride> const& layout);

Returns the number of elements in the layout or a specific mode.

rank()

template <class Shape, class Stride>
CUTE_HOST_DEVICE constexpr auto
rank(Layout<Shape,Stride> const& layout);

Returns the number of modes (dimensions) in the layout.

depth()

template <class Shape, class Stride>
CUTE_HOST_DEVICE constexpr auto
depth(Layout<Shape,Stride> const& layout);

Returns the hierarchical depth of the layout.

cosize()

template <int... Is, class Shape, class Stride>
CUTE_HOST_DEVICE constexpr auto
cosize(Layout<Shape,Stride> const& layout);

Returns the codomain size (maximum value + 1 that the layout can produce).

coalesce()

template <class Shape, class Stride>
CUTE_HOST_DEVICE constexpr auto
coalesce(Layout<Shape,Stride> const& layout);

Simplifies the layout by combining adjacent modes that can be merged. Example:

auto layout = make_layout(make_shape(make_shape(4, 8), 2),
                          make_stride(make_stride(1, 4), 32));
auto simplified = coalesce(layout);
// Result: shape (32, 2), stride (1, 32)

filter()

template <class Shape, class Stride>
CUTE_HOST_DEVICE constexpr auto
filter(Layout<Shape,Stride> const& layout);

Removes all stride-0 modes and size-1 modes from the layout.

complement()

template <class Shape, class Stride>
CUTE_HOST_DEVICE constexpr auto
complement(Layout<Shape,Stride> const& layout);

Builds the complement layout that points to elements not covered by the input layout.

right_inverse()

template <class Shape, class Stride>
CUTE_HOST_DEVICE constexpr auto
right_inverse(Layout<Shape,Stride> const& layout);

Builds the right-inverse layout such that layout(result(i)) == i.

Common Layout Types

LayoutLeft

Column-major (Fortran-style) layout where the first mode has stride-1.

LayoutRight

Row-major (C-style) layout where the last mode has stride-1.

Usage Examples

Basic Layout Creation

#include <cute/layout.hpp>

using namespace cute;

// Column-major 16x32 layout
auto layout = make_layout(make_shape(16, 32));

// Access elements
int idx = layout(5, 10);  // Returns 5 + 10*16 = 165

// Query properties
int total_size = size(layout);        // 512
int rows = size<0>(layout);           // 16
int cols = size<1>(layout);           // 32

Hierarchical Layouts

// Create a 2-level hierarchical layout
auto shape = make_shape(make_shape(4, 8), make_shape(2, 4));
auto layout = make_layout(shape);

// Layout has shape ((4,8),(2,4)) representing a 32x8 matrix
// organized as 2x4 tiles of 4x8 blocks

Slicing and Tiling

// Create a layout
auto layout = make_layout(make_shape(64, 64));

// Slice to get a sub-layout
auto col_slice = layout(_, 10);  // Get column 10

// Tile into 8x8 blocks
auto tiled = layout.tile(make_shape(8, 8));
// Shape: ((8,8), (8,8))

Working with Strides

// Row-major layout
auto row_major = make_layout(make_shape(16, 32), 
                             make_stride(32, 1));

// Strided layout (every other element)
auto strided = make_layout(make_shape(8, 16),
                           make_stride(2, 32));

// Swizzled layout for bank conflict avoidance
auto swizzled = make_layout(make_shape(32, 32),
                            make_stride(1, 33));

C++ Templates

CuTe Library

Python API

CuTe DSL

​Overview

​Class Template

​Template Parameters

​Source Location

​Member Functions

​Accessors

​shape()

​stride()

​Mapping Operations

​operator()

​Composition

​compose()

​with_shape()

​Tiling

​tile()

​Coordinate Conversion

​get_hier_coord()

​get_flat_coord()

​get_1d_coord()

​Factory Functions

​make_layout()

​make_identity_layout()

​make_layout_like()

​Utility Functions

​size()

​rank()

​depth()

​cosize()

​coalesce()

​filter()

​complement()

​right_inverse()

​Common Layout Types

​LayoutLeft

​LayoutRight

​Usage Examples

​Basic Layout Creation

​Hierarchical Layouts

​Slicing and Tiling

​Working with Strides

​See Also

Build docs developers (and LLMs) love