Tensor.Types.ixx File Reference

Core shape, stride, and index types for the Mila tensor API. More...

#include <array>
#include <string>
#include <cstdint>
#include <limits>
#include <stdexcept>
#include <algorithm>
#include <initializer_list>

Classes
struct	Mila::Dnn::TensorShape
	Fixed-capacity inline shape descriptor for N-dimensional tensors. More...

Namespaces
namespace	Mila
	Mila main API namespace.

Typedefs
using	Mila::Dnn::dim_t = int64_t
	Integer type used for tensor dimensions and indices.
using	Mila::Dnn::index_t = TensorShape
	Index descriptor for multi-dimensional element access.
using	Mila::Dnn::shape_t = TensorShape
	Row-major shape descriptor for tensor dimensional sizes.
using	Mila::Dnn::stride_t = TensorShape
	Stride descriptor (in elements) for each tensor dimension, row-major layout.

Functions
constexpr dim_t	Mila::Dnn::elementCount (const shape_t &shape) noexcept
	Product of a shape's extents – the number of logical elements it describes.
std::string	Mila::Dnn::indexToString (const index_t &index)
int	Mila::Dnn::narrowToKernelIndex (dim_t value)
	Narrow a dim_t to the 32-bit index type used inside device kernels.
std::string	Mila::Dnn::shapeToString (const shape_t &shape)
std::string	Mila::Dnn::strideToString (const stride_t &stride)

Detailed Description

Core shape, stride, and index types for the Mila tensor API.

Defines TensorShape – a fixed-capacity inline shape descriptor – and the shape_t, stride_t, and index_t aliases used throughout the framework.

Typedef Documentation

index_t

using Mila::Dnn::index_t = TensorShape

export

Index descriptor for multi-dimensional element access.

One index per tensor dimension. Valid indices satisfy: 0 <= index[i] < shape[i].

shape_t

using Mila::Dnn::shape_t = TensorShape

export

Row-major shape descriptor for tensor dimensional sizes.

• {} : scalar (rank 0)
• {n} : 1D tensor of length n
• {m, n} : 2D tensor, m rows and n columns
• {B, T, C} : 3D activation (batch, sequence, channels)
• {B, H, T, D} : 4D attention tensor

A zero in any position indicates an empty tensor.

stride_t

using Mila::Dnn::stride_t = TensorShape

export

Stride descriptor (in elements) for each tensor dimension, row-major layout.

stride_t[i] is the element count to advance one step along dimension i. Length equals shape.size(); empty for scalars.

Function Documentation

elementCount()

dim_t Mila::Dnn::elementCount ( const shape_t & shape )

constexprexportnoexcept

Product of a shape's extents – the number of logical elements it describes.

Distinct from TensorShape::size(), which is the rank. A rank-0 shape holds one element, so the empty product of 1 is the correct answer rather than a special case.

narrowToKernelIndex()

int Mila::Dnn::narrowToKernelIndex ( dim_t value )

inlineexport

Narrow a dim_t to the 32-bit index type used inside device kernels.

dim_t is the type of every axis extent, position, and element count at the API, component, and operation-interface layers. Kernels index in 32-bit for register pressure and address-arithmetic cost, so exactly one narrowing point exists per call path: the operation that launches the kernel. Routing that narrowing through this function keeps the membrane greppable and gives the out-of-range case a diagnosable failure instead of a silent wrap.

The check is host-side and runs once per launch, not per element – negligible against launch overhead. The margin is not theoretical: a Gemma 4 12B embedding table is 262144 x 3840 ~ 1.0e9 elements, roughly half of INT_MAX.

Exceptions

std::out_of_range

if the value does not fit in an int.