Computes the quotient and remainder in a single method.

Helps with visibility of barrier init operations across warps / cta / cluster Available as a separate function so as to batch inits across barriers an

* Find divisor, using find_log2 */

Stride between real and imaginary parts

Initializes members. This variant of the method should only be used when problem_shape and tile_shape contain modes of only rank 1.

Compute element-wise natural logarithm of the input tensor. :param a: Input tensor :type a: Union[TensorSSA, Numeric] :param fastmath: En

Compute element-wise base-2 logarithm of the input tensor. :param a: Input tensor :type a: Union[TensorSSA, Numeric] :param fastmath: Ena

Partition a buffer into a given layout and tiler.

Executes one test

Returns the sum of all elements

Helper to rematerialize block Idx. Reduces register liveness.

Create a test-running function with the given specification

* Performs a non-blocking allocation of TMEM. * @param num_columns Number of columns being freed. Must be 32 <= num_columns <= 512 and power of 2.

Broadcast self.layout to shape

Compute the number of contiguous elements needed to store a tensor with the given size

Create GPU tensor from either a new or existing CPU tensor. :param torch_tensor_cpu: Optional existing CPU tensor to reuse. If None, creates a ne

Returns the depth (nesting level) of a tuple, layout, or tensor. The depth of a tuple is the maximum depth of its elements plus 1. For an emp

Returns the cached result

Flattens a potentially nested tuple structure into a flat tuple. This function recursively traverses the input structure and flattens it into

Ctors.

Loads a fragment from memory with additional logical offset

Element-wise maximum of 2 CuTe tensors

Synchronization call. Blocks until barriers are initialized in shared memory.

Stores a fragment to memory with additional pointer offset

Get or create a struct type. Parameters ---------- name : Optional[str] The name of the struct type. If not provi

Apply a function to the leaf nodes of nested tuple structures. This function traverses nested tuple structures in parallel and applies the f

Computes laneId within a warp

Helper to rematerialize block Idx. Reduces register liveness.

Verifies the following properties: If ref_dtype is not void: 1) Either ``tensor`` or ``ref_tensor`` must be set (

Generates dynamic expression for constant values.

Unpack work tile info from an rmem tensor of shape (4,).

Computes the grid size given a chosen threadblock shape

Loads a fragment from memory with additional logical offset

Create a fake tensor with the specified shape, element type, and a compact memory layout. :param dtype: Data type of the tensor elements.

Least positive value

Get the nodes in the unique lexicographical topological order It generates a unique ordering of nodes by first sorting topologically

Constructs a matrix from a uniform element 1

Returns the pointer to the tensor data. This is either the device pointer or the data pointer if the data is not in a device.

Return the LLVM struct type (resolving and caching if needed).

Computes padding in shared memory to perform efficient transpose without bank conflicts.

Number of SK blocks per region (splitting factor)

Number of SK regions

Returns the stride of the layout

Returns the stride of the layout

Helper function to return true when called by thread 0 of threadblock 0.

Returns the GEMM volume in thread block tiles

Transform the mma operands to the required types