C Utilities (cutils)

Low-level C++/CUDA utility functions used internally by all pyc kernels. These are not part of the public API but are documented here for completeness.

CPU/C++ Utilities (utils.h)

Function	Description
clip(tensor, dim)	Extract column dim from a 2-D tensor as a 1-D view
format(vector<Tensor>*)	Stack a list of tensors as columns (each reshaped to Nx1) into an NxK tensor
format(vector<Tensor*>)	Pointer-vector overload of format
MakeOp(tensor*)	Create a TensorOptions matching the device and dtype of the input tensor
changedev(tensor*)	No-op in CPU build; moves tensor to the appropriate device in CUDA build
changedev(dev, tensor*)	Move tensor to named device string (e.g. "cuda:0"); no-op in CPU build

Low-level CUDA/C++ tensor utility functions.

(pyc)

Provides clip, format, tensor-option construction, and device transfer helpers used internally by the pyc modules.

Functions

torch::Tensor clip(torch::Tensor *inpt, int dim)

Clip tensor inpt along dimension dim.

Parameters:

• inpt – Input tensor.

• dim – Dimension to clip.

Returns:

Clipped tensor.

torch::Tensor format(std::vector<torch::Tensor> *inpt)

Stack a vector of tensors into a single tensor.

Parameters:

inpt – Vector of tensors.

Returns:

Stacked tensor.

torch::Tensor format(std::vector<torch::Tensor*> inpt)

Stack a vector of tensor pointers into a single tensor.

Parameters:

inpt – Vector of tensor pointers.

Returns:

Stacked tensor.

torch::TensorOptions MakeOp(torch::Tensor *x)

Create torch::TensorOptions matching the device and dtype of x.

Parameters:

x – Reference tensor.

Returns:

Matching TensorOptions.

void changedev(torch::Tensor *inpt)

Move tensor inpt to CPU in place.

Parameters:

inpt – Tensor to move.

torch::Tensor changedev(std::string dev, torch::Tensor *inx)

Copy tensor inx to the device specified by dev.

Parameters:

• dev – Device string (e.g. "cpu", "cuda:0").

• inx – Source tensor.

Returns:

Tensor on the target device.

CUDA Atomic Device Helpers (atomic.cuh)

These __device__ template functions are inlined into every CUDA kernel that needs them. They provide numerically stable variants of common operations.

Function	Description
_cofactor<T>(M, idy, idz)	Compute the (idy, idz) cofactor of a 3×3 matrix M
_div(p)	Safe reciprocal: returns 1/p or 0 if p == 0
_p2(p)	Squared value: (*p) × (*p)
_clp(p)	Round-trip clamp: rounds to 10 decimal places
_sqrt(p)	Sign-preserving square root: returns -√|p| for negative inputs
_cmp(xx, yy, xy)	Computes xy / √(xx × yy) (used for cosine-like ratios)
_arccos(sm, pz)	acos(pz / √sm) with safe division

CUDA Thread/Block Geometry (utils.cuh)

The blk_ inline computes the 2-D CUDA launch grid:

dim3 blk = blk_(num_elements, threads_per_block);

This is used in every CUDA kernel launch to ensure full coverage of the batch dimension.