Skip to main content

MlTensor

Namespace: MLLIB  ·  Library: Machine Learning Library

#include <ml/ml_tensor.h>

class MLLIB::MlTensor

N-dimensional row-major float32 tensor with shared-buffer storage, Eigen Map accessors and a non-owning view mode.

The shape is an arbitrary std::vector<int64_t> (matching ONNX Runtime), the buffer is a reference-counted std::vector<float> (or external memory in view mode), and copy / reshape / slice all stay O(1) by sharing the underlying storage. Eigen Map accessors give in-place row-major views of 2-D tensors; toMatrixXf / toMatrixXd produce column-major Eigen copies for code that needs the native Eigen layout.

src/examples/ex_ml/main.cpp
MatrixXf features(2, 3);
features << 1.0f, 2.0f, 0.5f,
    0.0f, -1.0f, 1.5f;
MlTensor batch(features);                          // owning row-major copy, shape {2, 3}
MlTensor flat = batch.reshape({6});                // shares the buffer
MlTensor::RowMajorMatrixMap rows = batch.matrix(); // zero-copy Eigen view

Public Methods​

MlTensor()​

Default constructor – creates an empty 0-element tensor with shape {}.

MlTensor(data, shape)​

Construct from a moved data buffer and an arbitrary shape.

The buffer size must equal the product of shape elements.

Parameters:

  • data : std::vector< float > && Flat float32 buffer (row-major order). Moved in.

  • shape : std::vector< int64_t > Dimension sizes (e.g. {batch, seq, features}).

MlTensor(data, shape)​

Construct by copying from a raw pointer and a shape.

Parameters:

  • data : *const float ** Pointer to contiguous float32 data (row-major).

  • shape : std::vector< int64_t > Dimension sizes.

MlTensor(mat)​

Construct from an Eigen column-major MatrixXf.

The data is copied and transposed into row-major layout. Shape is set to {rows, cols}.

Parameters:

  • mat : const Eigen::MatrixXf & Source matrix (column-major).

MlTensor(mat)​

Construct from an Eigen column-major MatrixXd.

The data is cast to float32 and stored in row-major layout. Shape is set to {rows, cols}.

Parameters:

  • mat : const Eigen::MatrixXd & Source matrix (column-major, double precision).

ndim()​

Returns:

  • int — Number of dimensions.

size()​

Returns:

  • int64_t — Total number of elements (product of all dimensions).

shape()​

Returns:

  • const std::vector< int64_t > & — Const reference to the full shape vector.

shape(dim)​

Parameters:

  • dim : int Dimension index (negative indices count from the end).

Returns:

  • int64_t — Size of the requested dimension.

rows()​

2-D convenience — equivalent to shape(0).

Asserts ndim >= 1.

Returns:

  • int — Number of rows.

cols()​

2-D convenience — equivalent to shape(1).

Asserts ndim >= 2.

Returns:

  • int — Number of columns.

data()​

Returns:

  • *float ** — Mutable pointer to the contiguous float32 buffer.

data()​

Returns:

  • *const float ** — Const pointer to the contiguous float32 buffer.

matrix()​

Mutable row-major Eigen::Map for 2-D tensors.

Asserts ndim == 2.

Returns:

  • RowMajorMatrixMap — Row-major mutable Map.

matrix()​

Const row-major Eigen::Map for 2-D tensors.

Asserts ndim == 2.

Returns:

  • ConstRowMajorMatrixMap — Row-major const Map.

toMatrixXf()​

Returns:

  • Eigen::MatrixXf — Column-major float copy of the 2-D data. Asserts ndim == 2.

toMatrixXd()​

Returns:

  • Eigen::MatrixXd — Column-major double copy of the 2-D data. Asserts ndim == 2.

reshape(newShape)​

Return a tensor that shares the same storage but has a different shape.

The total element count must be unchanged. If this tensor owns its data the result shares ownership (reference-counted). If this tensor is a view, the result is also a view into the same external buffer.

Parameters:

  • newShape : std::vector< int64_t > New shape (product must equal size()).

Returns:

  • MlTensor — Reshaped tensor (zero-copy).

isView()​

Returns:

  • bool — True if this tensor does not own its data buffer.

empty()​

Returns:

  • bool — True if the tensor contains zero elements.

Static Methods​

view(data, shape)​

Create a non-owning view over external mutable memory.

The caller is responsible for keeping the buffer alive for the lifetime of this tensor (and any copies / reshapes derived from it).

Parameters:

  • data : *float ** Pointer to external float32 data (row-major).

  • shape : std::vector< int64_t > Dimension sizes.

Returns:

fromBuffer(data, rows, cols)​

Copy raw data into a new owning tensor (legacy 2-D helper).

Parameters:

  • data : *const float ** Pointer to row-major float data.

  • rows : int Number of rows.

  • cols : int Number of columns.

Returns:

  • MlTensor — Newly constructed tensor with shape {rows, cols}.

Example​

Used in examples​

ex_ml​

ML library: tensors, ONNX inference through the MlModel interface, and the Python trainer.

src/examples/ex_ml/main.cpp, line 75
MatrixXf features(2, 3);
features << 1.0f, 2.0f, 0.5f,
    0.0f, -1.0f, 1.5f;
MlTensor batch(features);                          // owning row-major copy, shape {2, 3}
MlTensor flat = batch.reshape({6});                // shares the buffer
MlTensor::RowMajorMatrixMap rows = batch.matrix(); // zero-copy Eigen view
ok &= expect(batch.shape() == std::vector<int64_t>({2, 3}) && flat.ndim() == 1 && flat.data()[4] == -1.0f && rows(1, 2) == 1.5f && batch.toMatrixXd().isApprox(features.cast<double>()),
             "MlTensor: {2, 3} from Eigen, reshape shares data, matrix() maps it back");

Authors of this file​