Tensr Documentation

Welcome to Tensr - a powerful, superfast multidimensional tensor library for C/C++!

Overview

Tensr is a high-performance tensor computation library designed for scientific computing, machine learning, and numerical analysis. It provides a clean, intuitive API for both C and C++ with full GPU acceleration support.

Key Features

⚡ High Performance: Optimized for speed with SIMD and multi-threading
🚀 GPU Acceleration: Full CUDA support for parallel computing
🎯 Simple API: Clean, intuitive interfaces for C and C++
📦 Comprehensive: All essential tensor operations
✅ Production Ready: Thoroughly tested and documented
🌍 Cross-Platform: Windows, Linux, and macOS support
🔧 Zero Dependencies: No external dependencies for core library

Quick Example

C API (NumPy-Style)

#include <tensr/tensr.h>
#include <tensr/tensr_array.h>

int main() {
    /* Create arrays from data (like np.array([1, 2, 3])) */
    float data_a[] = {1, 2, 3};
    float data_b[] = {4, 5, 6};
    size_t shape[] = {3};

    Tensor* a = tensr_from_array(shape, 1, TENSR_FLOAT32, TENSR_CPU, data_a);
    Tensor* b = tensr_from_array(shape, 1, TENSR_FLOAT32, TENSR_CPU, data_b);

    /* Element-wise operations (like a + b, a * b, a ** 2) */
    Tensor* sum = tensr_add(a, b);
    Tensor* product = tensr_mul(a, b);
    Tensor* squared = tensr_pow(a, 2.0);

    /* Print results */
    tensr_print(sum);      /* [5, 7, 9] */
    tensr_print(product);  /* [4, 10, 18] */
    tensr_print(squared);  /* [1, 4, 9] */

    /* Cleanup */
    tensr_free(a);
    tensr_free(b);
    tensr_free(sum);
    tensr_free(product);
    tensr_free(squared);

    return 0;
}

C++ API

#include <tensr/tensr.hpp>

int main() {
    /* Create tensors */
    auto eye = tensr::Tensor::eye(3);
    auto rand_mat = tensr::Tensor::rand({3, 3});

    /* Matrix multiplication */
    auto result = eye.matmul(rand_mat);

    /* Print result */
    result.print();

    return 0;
}

Installation

Download prebuilt binaries from the GitHub Releases page. We publish platform-specific artifacts (headers + static/shared libraries) for each tagged release.

Windows (PowerShell)

# 1. Go to: https://github.com/muhammad-fiaz/tensr/releases
# 2. Download the desired `tensr-<version>-windows-x64.zip` asset
Expand-Archive -Path .\tensr-<version>-windows-x64.zip -DestinationPath .\tensr
cd .\tensr
# Headers are in include\, libraries in lib\, and binaries (if any) in bin\

Linux / macOS

# 1. Go to: https://github.com/muhammad-fiaz/tensr/releases
# 2. Download the desired `tensr-<version>-linux-x64.tar.gz` (or macOS tarball)
tar -xzf tensr-<version>-linux-x64.tar.gz
cd tensr-<version>
# Headers in include/, libs in lib/, binaries in bin/

If you prefer to build from source, follow the instructions below.

From Source

git clone https://github.com/muhammad-fiaz/tensr.git
cd tensr
xmake build
xmake install

Core Operations

Tensor Creation

Create tensors with various initialization methods:

zeros() - All zeros
ones() - All ones
full() - Fill with value
arange() - Evenly spaced values
linspace() - Linearly spaced values
eye() - Identity matrix
rand() - Random uniform
randn() - Random normal

Arithmetic Operations

Perform element-wise operations:

Basic: add(), sub(), mul(), div()
Mathematical: pow(), sqrt(), exp(), log()
Trigonometric: sin(), cos(), tan()

Linear Algebra

Advanced matrix operations:

dot() - Dot product
matmul() - Matrix multiplication
inv() - Matrix inverse
det() - Determinant
svd() - Singular value decomposition
eig() - Eigenvalues and eigenvectors

Reduction Operations

Aggregate tensor values:

sum() - Sum of elements
mean() - Mean value
max(), min() - Maximum and minimum
argmax(), argmin() - Indices of extrema

GPU Support

Tensr supports multiple accelerator backends:

/* CUDA GPU */
Tensor* t = tensr_zeros(shape, 2, TENSR_FLOAT32, TENSR_CUDA);

/* Transfer between devices */
tensr_to_device(t, TENSR_CUDA, 0);

Supported devices: - CPU: Standard CPU execution - CUDA: NVIDIA GPU acceleration - XPU: Intel XPU support - NPU: Neural Processing Unit - TPU: Tensor Processing Unit

Performance Tips

Optimization Guidelines

Use GPU for large tensors (>1000 elements)
Batch operations when possible
Reuse tensors to minimize allocations
Use appropriate data types (float32 vs float64)
Profile your code to identify bottlenecks

Best Practices

Do's ✅

Free tensors when done to avoid memory leaks
Check return values for NULL
Use GPU for large-scale computations
Profile your code for bottlenecks
Use appropriate data types

Don'ts ❌

Don't mix tensors from different devices without transfer
Don't modify tensor data directly
Don't forget to synchronize after GPU operations
Don't use debug builds in production
Don't ignore compiler warnings

Getting Help

Contributing

We welcome contributions! See our Contributing Guide for details.

License

Tensr is licensed under the Apache License 2.0. See LICENSE for details.

Author: Muhammad Fiaz
Email: contact@muhammadfiaz.com
GitHub: @muhammad-fiaz