torchlet

A lightweight torch like tensor computation library implemented from scratch in C and CUDA.

I created this to better understand the inner workings of tensor libraries like PyTorch, making it an excellent resource for learning how these libraries work under the hood.

Demo

What it feels like to train a neural network on GPU as compared to CPU.

m.mp4

Features

• Basic tensor operations implemented in C and CUDA
• Simple and readable implementation focused on learning
• GPU acceleration support through CUDA
• Zero external dependencies
• Clear code structure for educational purposes

Prerequisites

• C compiler (gcc/clang)
• CUDA Toolkit
• Make

Getting Started

Basic Usage

Here's a simple example demonstrating matrix multiplication with torchlet:

// main.cu
#include "tensor.h"

int main() {
    // Create two random tensors
    Tensor *a = tensor_randn((size_t[]){4, 5}, 2);  // 4x5 matrix
    Tensor *b = tensor_randn((size_t[]){5, 3}, 2);  // 5x3 matrix
    
    // Perform matrix multiplication
    Tensor *c = tensor_matmul(a, b);  // Results in a 4x3 matrix
    
    // Display results
    tensor_print(c);   // Print tensor values
    tensor_debug(c);   // Print tensor metadata
    
    // Clean up
    tensor_free(a);
    tensor_free(b);
    tensor_free(c);
    
    return 0;
}

Check out ./mnist_cpu.c and ./mnist_gpu.cu to understand how to train a neural network using this library.

Compiling Your Code

nvcc -O3 -Xcompiler -Wall -Xcompiler -Werror -arch=sm_86 -lm tensor.c tensor.cu main.cu -Xcompiler -fopenmp -o main.out

Running the Example

./main

Contributing

Contributions are welcome! If you find a bug, have an idea for an enhancement, or want to contribute in any way, feel free to open an issue or submit a pull request.

License

This project is licensed under the AGPL3 License. For details, see the LICENSE file.