Module frontend 

Cube Frontend Types.

Modules

ABSOLUTE_POS

The position of the working unit in the whole cube kernel, without regards to cubes and axis.

ABSOLUTE_POS_X

The index of the working unit in the whole cube kernel along the X axis, without regards to cubes.

ABSOLUTE_POS_Y

The index of the working unit in the whole cube kernel along the Y axis, without regards to cubes.

ABSOLUTE_POS_Z

The index of the working unit in the whole cube kernel along the Z axis, without regards to cubes.

CUBE_CLUSTER_DIM

The total amount of cubes in a cluster.

CUBE_CLUSTER_DIM_X

The dimension of the cluster along the X axis.

CUBE_CLUSTER_DIM_Y

The dimension of the cluster along the Y axis.

CUBE_CLUSTER_DIM_Z

The dimension of the cluster along the Z axis.

CUBE_COUNT

The number of cubes launched.

CUBE_COUNT_X

The number of cubes launched along the X axis.

CUBE_COUNT_Y

The number of cubes launched along the Y axis.

CUBE_COUNT_Z

The number of cubes launched along the Z axis.

CUBE_DIM

The total amount of working units in a cube.

CUBE_DIM_X

The dimension of the cube along the X axis.

CUBE_DIM_Y

The dimension of the cube along the Y axis.

CUBE_DIM_Z

The dimension of the cube along the Z axis.

CUBE_POS

The cube position, without regards to axis.

CUBE_POS_CLUSTER

The cube position within the cluster.

CUBE_POS_CLUSTER_X

The cube position in the cluster along the X axis.

CUBE_POS_CLUSTER_Y

The cube position in the cluster along the Y axis.

CUBE_POS_CLUSTER_Z

The cube position in the cluster along the Z axis.

CUBE_POS_X

The cube position along the X axis.

CUBE_POS_Y

The cube position along the Y axis.

CUBE_POS_Z

The cube position along the Z axis.

PLANE_DIM

The total amount of working units in a plane.

PLANE_POS

The position of the plane within the cube (plane/warp/subgroup index).

UNIT_POS

The position of the working unit inside the cube, without regards to axis.

UNIT_POS_PLANE

The relative position of the working unit inside the plane, without regards to cube dimensions.

UNIT_POS_X

The position of the working unit inside the cube along the X axis.

UNIT_POS_Y

The position of the working unit inside the cube along the Y axis.

UNIT_POS_Z

The position of the working unit inside the cube along the Z axis.

assign

barrier

This module exposes barrier for asynchronous data transfer

branch

clamp

clamp_max

clamp_min

cmma

This module exposes cooperative matrix-multiply and accumulate operations.

comptime

Module containing compile-time information about the current runtime.

comptime_error

copy

copy_bulk

cube_comment

div_ceil

erf

fma

Expand method of fma().

hypot

index

index_mut

max

min

not

plane_all

Module containing the expand function for plane_all().

plane_any

Module containing the expand function for plane_any().

plane_ballot

Module containing the expand function for plane_ballot().

plane_broadcast

Module containing the expand function for plane_broadcast().

plane_elect

Module containing the expand function for plane_elect().

plane_exclusive_prod

Module containing the expand function for plane_exclusive_prod().

plane_exclusive_sum

Module containing the expand function for plane_exclusive_sum().

plane_inclusive_prod

Module containing the expand function for plane_inclusive_prod().

plane_inclusive_sum

Module containing the expand function for plane_inclusive_sum().

plane_max

Module containing the expand function for plane_max().

plane_min

Module containing the expand function for plane_min().

plane_prod

Module containing the expand function for plane_prod().

plane_shuffle

Module containing the expand function for plane_shuffle().

plane_shuffle_down

Module containing the expand function for plane_shuffle_down().

plane_shuffle_up

Module containing the expand function for plane_shuffle_up().

plane_shuffle_xor

Module containing the expand function for plane_shuffle_xor().

plane_sum

Module containing the expand function for plane_sum().

push_validation_error

range

range_stepped

rhypot

select

select_many

set_polyfill

Expand module of set_polyfill().

slice

storage_type_of

sub

synchronization

tma_group_commit

tma_group_wait

tma_group_wait_read

tma_store_1d

tma_store_2d

tma_store_3d

tma_store_4d

tma_store_5d

type_of

Macros

__expand_debug_print

Print a formatted message using the target’s debug print facilities. The format string is target specific, but Vulkan and CUDA both use the C++ conventions. WGSL isn’t currently supported.

debug_print

Print a formatted message using the target’s debug print facilities. The format string is target specific, but Vulkan and CUDA both use the C++ conventions. WGSL isn’t currently supported.

Structs

Array

A contiguous array of elements.

Atomic

An atomic numerical type wrapping a normal numeric primitive. Enables the use of atomic operations, while disabling normal operations. In WGSL, this is a separate type - on CUDA/SPIR-V it can theoretically be bitcast to a normal number, but this isn’t recommended.

BufferBinding

Buffer representation with a reference to the server handle.

BufferCompilationArg

ComptimeCell

A cell that can store and mutate a cube type during comptime.

ComptimeCellExpand

Expand type of ComptimeCell.

Const

DynamicScalar

A fake element type that can be configured to map to any other element type.

DynamicSize

A fake constant type that can be configured to map to any comptime value.

Im2col

Im2col indexing. Loads a “column” (not the same column as im2col) of pixels into shared memory, with a certain offset (kernel position). The corners are the bounds to load pixels from at offset 0, so the top left corner of the kernel. The offset is added to the corner offsets, so a (-1, -1) corner will stop the bounding box at (1, 1) for kernel offset (2, 2).

Im2colArgs

Args for im2col tensor maps

Im2colCompilationArg

Im2colExpand

Im2colLaunch

Im2colWide

1D im2col, not properly supported yet

Im2colWideArgs

Args for im2col wide tensor maps

Im2colWideCompilationArg

Im2colWideExpand

Im2colWideLaunch

InputScalar

A way to define an input scalar without a generic attached to it.

InputScalarCompilationArg

InputScalarExpand

NativeExpand

Expand type of a native GPU type, i.e. scalar primitives, arrays, shared memory.

OptionExpand

OrderingExpand

OwnedTensor

OwnedTensorExpand

RangeExpand

RangeFromExpand

RangeFullExpand

RangeInclusiveExpand

RangeToExpand

RangeToInclusiveExpand

ReadOnly

ReadWrite

Registry

It is similar to a map, but where the keys are stored at comptime, but the values can be runtime variables.

RuntimeCell

RuntimeCellExpand

Sequence

A sequence of cube types that is inlined during compilation.

SequenceArg

SequenceCompilationArg

SequenceExpand

Expand type of Sequence.

Shared

SteppedRangeExpand

SwitchExpand

SwitchExpandExpr

Tensor

The tensor type is a wrapper around [T] that comes with more metadata such as stride and shape.

TensorBinding

Tensor representation with a reference to the server handle, the strides and the shape.

TensorCompilationArg

Compilation argument for a tensor.

TensorExpand

TensorLayout

TensorMap

A CUDA CUtensorMap object. Represents a tensor encoded with a lot of metadata, and is an opaque packed object at runtime. Does not support retrieving any shapes or strides, nor does it give access to the pointer. So these need to be passed separately in an aliased Tensor if needed.

TensorMapArg

Grid constant tensor map, currently only maps to CUDA tensormap. May be interleaved or swizzled, but last dimension must be contiguous (since strides don’t include the last dimension).

TensorMeta

TensorMetaCompilationArg

TensorMetaExpand

TensorMetaLaunch

TensorReinterpret

TensorView

TensorViewBuilder

TensorViewBuilderCompilationArg

TensorViewBuilderExpand

TensorViewBuilderLaunch

TensorViewExpand

Tiled

Regular tiled tensor map

TiledArgs

Args for tiled tensor maps

TiledCompilationArg

TiledExpand

TiledLaunch

Vector

A contiguous list of elements that supports auto-vectorized operations.

Enums

BufferArg

ComptimeOption

ComptimeOptionArgs

ComptimeOptionCompilationArg

ComptimeOptionExpand

IfElseExpand

IfElseExprExpand

MatchExpand

MatchExpandExpr

OobFill

What value to use when filling out of bounds values

OptionArgs

OptionCompilationArg

TensorArg

Argument to be used for tensors passed as arguments to kernels.

TensorClampMode

TensorMapFormat

Format of TensorMap

TensorMapInterleave

Interleave setting for TensorMap

TensorMapPrefetch

Additional prefetching to perform during load Specifies L2 fetch size which indicates the byte granularity at which L2 requests are filled from DRAM

TensorMapSwizzle

Data are organized in a specific order in global memory; however, this may not match the order in which the application accesses data in shared memory. This difference in data organization may cause bank conflicts when shared memory is accessed. In order to avoid this problem, data can be loaded to shared memory with shuffling across shared memory banks. When interleave is TensorMapInterleave::B32, swizzle must be TensorMapSwizzle::B32. Other interleave modes can have any swizzling pattern.

Constants

ABSOLUTE_POS

The position of the working unit in the whole cube kernel, without regards to cubes and axis.

ABSOLUTE_POS_X

The index of the working unit in the whole cube kernel along the X axis, without regards to cubes.

ABSOLUTE_POS_Y

The index of the working unit in the whole cube kernel along the Y axis, without regards to cubes.

ABSOLUTE_POS_Z

The index of the working unit in the whole cube kernel along the Z axis, without regards to cubes.

CUBE_CLUSTER_DIM

The total amount of cubes in a cluster.

CUBE_CLUSTER_DIM_X

The dimension of the cluster along the X axis.

CUBE_CLUSTER_DIM_Y

The dimension of the cluster along the Y axis.

CUBE_CLUSTER_DIM_Z

The dimension of the cluster along the Z axis.

CUBE_COUNT

The number of cubes launched.

CUBE_COUNT_X

The number of cubes launched along the X axis.

CUBE_COUNT_Y

The number of cubes launched along the Y axis.

CUBE_COUNT_Z

The number of cubes launched along the Z axis.

CUBE_DIM

The total amount of working units in a cube.

CUBE_DIM_X

The dimension of the cube along the X axis.

CUBE_DIM_Y

The dimension of the cube along the Y axis.

CUBE_DIM_Z

The dimension of the cube along the Z axis.

CUBE_POS

The cube position, without regards to axis.

CUBE_POS_CLUSTER

The cube position within the cluster.

CUBE_POS_CLUSTER_X

The cube position in the cluster along the X axis.

CUBE_POS_CLUSTER_Y

The cube position in the cluster along the Y axis.

CUBE_POS_CLUSTER_Z

The cube position in the cluster along the Z axis.

CUBE_POS_X

The cube position along the X axis.

CUBE_POS_Y

The cube position along the Y axis.

CUBE_POS_Z

The cube position along the Z axis.

PLANE_DIM

The total amount of working units in a plane.

PLANE_POS

The position of the plane within the cube (plane/warp/subgroup index).

UNIT_POS

The position of the working unit inside the cube, without regards to axis.

UNIT_POS_PLANE

The relative position of the working unit inside the plane, without regards to cube dimensions.

UNIT_POS_X

The position of the working unit inside the cube along the X axis.

UNIT_POS_Y

The position of the working unit inside the cube along the Y axis.

UNIT_POS_Z

The position of the working unit inside the cube along the Z axis.

Traits

Abs

AbsExpand

AddAssignExpand

AddExpand

AndExpand

ArcCos

ArcCosExpand

ArcCosh

ArcCoshExpand

ArcSin

ArcSinExpand

ArcSinh

ArcSinhExpand

ArcTan

ArcTan2

ArcTan2Expand

ArcTanExpand

ArcTanh

ArcTanhExpand

AsDerefExpand

AsDerefMutExpand

AsMutExpand

Expand version of AsMut. The Self version must be implemented by all ExpandTypes, since CubeCL also uses it to implement &mut x.

AsRefExpand

Expand version of AsRef. Like AsRef<Self> it’s implemented for all ExpandTypes. This is called when the Rust code uses &x.

Assign

BitAndAssignExpand

BitAndExpand

BitOrAssignExpand

BitOrExpand

BitXorAssignExpand

BitXorExpand

BoolOps

Extension trait for bool.

Cast

Enable elegant casting from any to any CubeElem

Ceil

CeilExpand

CloneExpand

CompilationArg

Argument used during the compilation of kernels.

Cos

CosExpand

Cosh

CoshExpand

CountOnes

CountOnesExpand

CubeAdd

CubeAddAssign

CubeAnd

CubeBitAnd

CubeBitAndAssign

CubeBitOr

CubeBitOrAssign

CubeBitXor

CubeBitXorAssign

CubeComptime

A type that can be used as a kernel comptime argument. Note that a type doesn’t need to implement CubeComptime to be used as a comptime argument. However, this facilitate the declaration of generic cube types.

CubeDebug

CubeDiv

CubeDivAssign

CubeEnum

CubeIndex

Trait bound that can be used to guarantee the expand also implements IndexExpand

CubeIndexMut

CubeMul

CubeMulAssign

CubeNeg

CubeNot

CubeOption

Extensions for Option

CubeOptionDefault

Extensions for Option that require default

CubeOr

CubeOrd

CubeOrdering

CubePartialEq

CubePartialOrd

CubePrimitive

Form of CubeType that encapsulates all primitive types: Numeric, UInt, Bool

CubePrimitiveExpand

CubeRem

CubeRemAssign

CubeShl

CubeShlAssign

CubeShr

CubeShrAssign

CubeSub

CubeSubAssign

CubeType

Types used in a cube function must implement this trait

DefaultExpand

Degrees

DegreesExpand

DerefExpand

CubeCL version of Deref. Unlike those traits, this trait produces owned values directly. Maps to *x.

DivAssignExpand

DivCeil

DivCeilExpand

DivExpand

Dot

DotExpand

Erf

ErfExpand

Exp

ExpExpand

ExpandTypeClone

Expm1

Expm1Expand

FindFirstSet

FindFirstSetExpand

Float

Floating point numbers. Used as input in float kernels

FloatBits

FloatBitsExpand

FloatOps

FloatOpsExpand

Floor

FloorExpand

Hypot

HypotExpand

IndexExpand

IndexMutExpand

Int

Signed or unsigned integer. Used as input in int kernels

IntoComptime

Trait for marking a function return value as comptime when the compiler can’t infer it.

IntoExpand

IntoMut

Convert an expand type to a version with mutable registers when necessary.

IntoRuntime

Trait useful to convert a comptime value into runtime value.

InverseSqrt

InverseSqrtExpand

IsInf

IsInfExpand

IsNan

IsNanExpand

Iterable

Something that can be iterated on by a for loop. Currently only includes Range, StepBy and Sequence.

LaunchArg

Defines how a launch argument can be expanded.

LeadingZeros

LeadingZerosExpand

List

Type from which we can read/to which we can write values in cube functions.

ListExpand

Type from which we can read/to which we can write values in cube functions.

Log

Log1p

Log1pExpand

LogExpand

Magnitude

MagnitudeExpand

ModFloor

ModFloorExpand

MulAssignExpand

MulExpand

MulHi

MulHiExpand

NativeAssign

Trait for native types that can be assigned. For non-native composites, use the normal Assign.

NativeCubeType

NegExpand

Normalize

NormalizeExpand

NotExpand

Numeric

Type that encompasses both (unsigned or signed) integers and floats Used in kernels that should work for both.

OneExpand

OptionExt

OrExpand

OrdExpand

PartialEqExpand

PartialOrdExpand

Powf

PowfExpand

Powi

PowiExpand

Radians

RadiansExpand

Recip

RecipExpand

RegistryQuery

To find an item from the registry, the query must be able to be translated to the actual key type.

Reinterpret

Enables reinterpetring the bits from any value to any other type of the same size.

RemAssignExpand

RemExpand

ReverseBits

ReverseBitsExpand

Rhypot

RhypotExpand

Round

RoundExpand

RuntimeAssign

SaturatingAdd

SaturatingAddExpand

SaturatingSub

SaturatingSubExpand

Scalar

Marker trait for scalar primitives. Should be implemented for all scalar CubePrimitives, but not for Vector or non-standard primitives like Barrier. Alternatively, treat these as types that can be stored in a [Vector]

ScalarArgSettings

Similar to [ArgSettings], however only for scalar types that don’t depend on the Runtime trait.

ShlAssignExpand

ShlExpand

ShrAssignExpand

ShrExpand

Sin

SinExpand

Sinh

SinhExpand

Size

SizedContainer

SizedContainerExpand

SliceExt

SliceOperator

SliceOperatorExpand

SliceVectorExt

SliceVisibility

Sqrt

SqrtExpand

SubAssignExpand

SubExpand

Tan

TanExpand

Tanh

TanhExpand

TensorMapKind

TrailingZeros

TrailingZerosExpand

Trunc

TruncExpand

VectorSum

VectorSumExpand

Vectorized

VectorizedExpand

ZeroExpand

Functions

__expand_assign

__expand_deref

assign_binary_op_expand

break_expand

copy

Copy one element between two array-likes without intermediates.

copy_bulk

Bulk copy length elements between two array-likes without intermediates.

debug_call_expand

Calls a function and inserts debug symbols if debug is enabled.

debug_source_expand

Adds source instruction if debug is enabled

debug_var_expand

Registers name for an expand if possible

div_ceil

erf

expand_erf

expand_himul_64

expand_himul_sim

expand_hypot

expand_rhypot

fast_math_expand

fma

Fused multiply-add A*B+C.

for_expand

from_raw_parts

hypot

Computes the hypotenuse of a right triangle given the lengths of the other two sides.

if_else_expand

if_else_expr_expand

if_expand

init_expand

into_mut_assign

loop_expand

match_expand

match_expand_expr

max

The maximum of two values, not requiring Ord. Provided for clarity in certain cases, though clamp_min may sometimes be more clear.

min

The minimum of two values, not requiring Ord. Provided for clarity in certain cases, though clamp_max may sometimes be more clear.

plane_all

Perform a reduce all operation across all units in a plane.

plane_any

Perform a reduce any operation across all units in a plane.

plane_ballot

Perform a ballot operation across all units in a plane. Returns a set of 32-bit bitfields as a Vector, with each element containing the value from 32 invocations. Note that vector size will always be set to 4 even for PLANE_DIM <= 64, because we can’t retrieve the actual plane size at expand time. Use the runtimePLANE_DIM to index appropriately.

plane_broadcast

Broadcasts the value from the specified plane unit at the given index to all active units within that plane. Requires a constant index. For non-constant indices, use plane_shuffle().

plane_elect

Returns true if the cube unit has the lowest plane_unit_id among active unit in the plane

plane_exclusive_prod

Perform an exclusive product operation across all units in a plane. This multiplies all values to the “left” of the unit, excluding this unit’s value. The 0th unit will be set to E::one(). Also known as “exclusive prefix product” or “exclusive scan”.

plane_exclusive_sum

Perform an exclusive sum operation across all units in a plane. This sums all values to the “left” of the unit, excluding this unit’s value. The 0th unit will be set to E::zero(). Also known as “exclusive prefix sum” or “exclusive scan”.

plane_inclusive_prod

Perform an inclusive product operation across all units in a plane. This multiplies all values to the “left” of the unit, including this unit’s value. Also known as “prefix product” or “inclusive scan”.

plane_inclusive_sum

Perform an inclusive sum operation across all units in a plane. This sums all values to the “left” of the unit, including this unit’s value. Also known as “prefix sum” or “inclusive scan”.

plane_max

Perform a reduce max operation across all units in a plane.

plane_min

Perform a reduce min operation across all units in a plane.

plane_prod

Perform a reduce prod operation across all units in a plane.

plane_shuffle

Perform an arbitrary lane shuffle operation across the plane. Each unit reads the value from the specified source lane.

plane_shuffle_down

Perform a shuffle down operation across the plane. Each unit reads the value from a unit with a higher lane ID (current_id + delta). Units at the end will read from themselves if (lane_id + delta >= plane_dim).

plane_shuffle_up

Perform a shuffle up operation across the plane. Each unit reads the value from a unit with a lower lane ID (current_id - delta). Units with lane_id < delta will read from themselves (no change).

plane_shuffle_xor

Perform a shuffle XOR operation across the plane. Each unit exchanges its value with another unit at an index determined by XOR with the mask. This is useful for butterfly reduction patterns.

plane_sum

Perform a reduce sum operation across all units in a plane.

printf_expand

Prints a formatted message using the print debug layer in Vulkan, or printf in CUDA.

push_validation_error

Push a validation error that will make the kernel compilation to fail.

range

integer range. Equivalent to:

range_stepped

Stepped range. Equivalent to:

return_expand

rhypot

Computes the reciprocal of the hypotenuse of a right triangle given the lengths of the other two sides.

select

Executes both branches, then selects a value based on the condition. This should be branchless, but might depend on the compiler.

select_many

Same as select() but with vectors instead.

set_polyfill

Change the meaning of the given cube primitive type during compilation.

storage_type_of

switch_expand

switch_expand_expr

sync_async_proxy_shared

sync_async_proxy_shared is a synchronization fence for the experimental SM 9.0+ copy functions, applying bidirectionally between the async proxy (i.e. TMA) and shared memory. Should be used after initializing the barriers, and before the copy operation. PTX: fence.proxy.async.shared::cta Experimental and subject to change.

sync_cube

Coordinates the following among all invocations in the current cube:

sync_plane

Synchronizes units within their plane (e.g., warp or SIMD group).

sync_storage

Sync_storage is the same but change “cube address space(shared memory)” to “storage address space(input args)”. But the set of invocations that are collaborating is still only the invocations in the same cube.There is no guarantee about using barriers alone to make the writes to storage buffer in one cube become visible to invocations in a different cube.

tma_group_commit

Commit an async tensor operation. Not sure how this works, poor docs. But you need to call it after a write, but not after reads.

tma_group_wait

Wait until at most max_pending TMA copy operations are in flight.

tma_group_wait_read

Wait TMA copy operations have finished reading from shared memory, with at most max_pending operations being unfinished.

tma_store_1d

Copy a tile from a shared memory src to a global memory dst, with the provided offsets. Should be combined with memcpy_async_tensor_commit and memcpy_async_tensor_wait_read.

tma_store_2d

Copy a tile from a shared memory src to a global memory dst, with the provided offsets. Should be combined with memcpy_async_tensor_commit and memcpy_async_tensor_wait_read.

tma_store_3d

Copy a tile from a shared memory src to a global memory dst, with the provided offsets. Should be combined with memcpy_async_tensor_commit and memcpy_async_tensor_wait_read.

tma_store_4d

Copy a tile from a shared memory src to a global memory dst, with the provided offsets. Should be combined with memcpy_async_tensor_commit and memcpy_async_tensor_wait_read.

tma_store_5d

Copy a tile from a shared memory src to a global memory dst, with the provided offsets. Should be combined with memcpy_async_tensor_commit and memcpy_async_tensor_wait_read.

type_of

unary_expand

unary_expand_fixed_output

workgroup_uniform_load

Barrier, then load reference with the result marked workgroup-uniform — mirrors WGSL’s workgroupUniformLoad. Lets a workgroup-shared value gate control flow that contains barriers. Non-WGSL backends lower it to [sync_cube] plus a plain load.

workgroup_uniform_load_atomic

Atomic counterpart of [workgroup_uniform_load]: barrier + atomic load, returning the underlying numeric (WGSL’s atomic workgroupUniformLoad overload).

Type Aliases

SharedExpand

SliceExpand