Edge-Computing-Engine

An open-source scientific computing engine designed for matrix computation and machine learning tasks, with CUDA acceleration support.

Overview

Edge-Computing-Engine provides a flexible computational graph implementation that supports:

1. Forward propagation for neural network inference
2. Backward propagation for training via automatic differentiation
3. CUDA-accelerated operations for improved performance on NVIDIA GPUs

The computational graph is built using nodes representing operations such as matrix multiplication, element-wise operations, and activation functions.

1. GPU Features

1.1 Operation Types

The framework supports the following operations:

• OP_INPUT: Input nodes for data and parameters
• OP_ADD: Element-wise addition
• OP_SUB: Element-wise subtraction
• OP_MUL: Element-wise multiplication
• OP_DIV: Element-wise division
• OP_MATMUL: Matrix multiplication
• OP_RELU: ReLU activation function
• OP_SIGMOID: Sigmoid activation function
• OP_TANH: Tanh activation function (declared but not implemented)
• OP_SOFTMAX: Softmax function (declared but not implemented)

1.2 Core Classes

1. Matrix_CU: A custom matrix class with CUDA support
2. ComputeNode: Represents a node in the computational graph
3. ComputeGraph: Manages the graph, including node creation, forward and backward propagation

1.3 CUDA Kernels

The code includes several CUDA kernel functions:

• reluForwardKernel/reluBackwardKernel: Forward and backward propagation for ReLU
• sigmoidForwardKernel/sigmoidBackwardKernel: Forward and backward propagation for Sigmoid
• matmulForwardKernel/matmulBackwardAKernel/matmulBackwardBKernel: Forward and backward propagation for matrix multiplication
• mulBackwardKernel: Backward propagation for element-wise multiplication
• divBackwardKernel: Backward propagation for element-wise division
• negGradKernel: Helper kernel for computing negative gradients

1.4 CUDA Example

// Create a computational graph
ComputeGraph graph;

// Create input and weight nodes
ComputeNode* input = graph.addInput(input_rows, input_cols, batch_size);
ComputeNode* weight = graph.addInput(input_cols, output_cols);

// Initialize with random values
input->value.randomInit();
weight->value.randomInit();

// Build forward computation graph
ComputeNode* fc1 = graph.addMatMul(input, weight);  // Fully connected layer
ComputeNode* output_node = graph.addReLU(fc1);      // Apply ReLU activation

// Perform forward propagation
graph.forward();

// Compute gradients via backward propagation
graph.backward(output_node);

1.5 GPU_Usage

root@5353:/backup# nvcc -arch=sm_80 -I/usr/local/cuda/include -L/usr/local/cuda/lib64 cuda_mat/mat_grad.cu -o main && ./main
root@5353:/backup# ./main
====== Configuration ======
Input matrix: 10000 x 1000
Weight matrix: 1000 x 500
Batch size: 32
Activation: ReLU

====== CUDA Results ======
Input first 5x5:
0.840188 0.394383 0.783099 0.79844 0.911647
0.45724 0.640368 0.717092 0.460067 0.54114
0.0860517 0.180886 0.426423 0.0470658 0.306902
0.0587587 0.0138811 0.622212 0.0391351 0.221551
0.456151 0.24875 0.0699601 0.742097 0.216935
Weight first 5x5:
0.817311 0.0247108 0.0146763 0.939293 0.502699
0.604392 0.921644 0.675689 0.948712 0.58803
0.824073 0.747934 0.0448163 0.757354 0.858343
0.308744 0.659798 0.941692 0.255731 0.539655
0.89383 0.609149 0.799556 0.726306 0.640965
Output first 5x5:
256.076 253.23 258.393 259.965 255.971
250.843 246.827 252.131 253.493 244.151
247.064 244.543 251.723 247.424 250.466
249.944 250.359 256.148 252.21 249.615
246.675 238.295 252.572 242.94 243.512
Input gradient first 5x5:
244.628 251.341 255.388 250.4 249.891
244.628 251.341 255.388 250.4 249.891
244.628 251.341 255.388 250.4 249.891
244.628 251.341 255.388 250.4 249.891
244.628 251.341 255.388 250.4 249.891
Weight gradient first 5x5:
159791 159791 159791 159791 159791
160010 160010 160010 160010 160010
160266 160266 160266 160266 160266
159899 159899 159899 159899 159899
159964 159964 159964 159964 159964

====== Performance ======
CUDA computation time: 2201.45 ms

1.6 Requirements

• CUDA-capable NVIDIA GPU (for GPU features)
• CUDA Toolkit
• C++11 or above
• nvcc: NVIDIA (R) Cuda compiler driver Copyright (c) 2005-2024 NVIDIA Corporation Built on Thu_Mar_28_02:18:24_PDT_2024 Cuda compilation tools, release 12.4, V12.4.131 Build cuda_12.4.r12.4/compiler.34097967_0

2. CPU Features

• read_csv(string &file_path): Reads a formatted file (csv) and returns a matrix with automatically calculated dimensions
• Formatted file writing interface (similar to pandas.to_csv)
• Broadcasting mechanism for matrices with padding interface
• Fully connected layer forward and backward propagation interfaces with automatic differentiation support
• Matrix differentiation and automatic differentiation interfaces
• Various matrix operations:
  • create(row,cols): Creates a matrix with specified dimensions
  • add, subtract, mul: Basic matrix operations
  • get_T: Matrix transposition
  • flatten: Returns a flattened array
  • matrix_rs: Matrix structure compression
  • matrix_sum, matrix_mean: Statistical operations
  • iloc: Matrix slicing
  • And many more
• Activation functions (Relu)
• Loss functions (MSE)
• Neural network components (convolutional layers, pooling layers)

2.1 CPU_autodiff_Installation

git clone git@github.com:AllenZYJ/Edge-Computing-Engine.git
cd Edge-Computing-Engine/install_diff && make && make install

Note: If you're not a root user, you may need to add sudo to the make install command.

2.2 CPU_Usage

root@5353:/backup# .g++ main.cpp -o main -lautodiff
root@5353:/backup# ./main
0.000000+-0.000378*1.000000
-0.000378+0.000000*1.000000
....
-0.000378+0.000000*1.000000
0.000000+-0.000378*1.000000
-0.000378+0.000000*1.000000
-0.000378+0.000000*1.000000
weight_1_grad:
[
   -0.000378    -0.000378    -0.000378 
   -0.000378    -0.000378    -0.000378 
   -0.000378    -0.000378    -0.000378 
]
neraul end;

3. Contributing

Contributions are welcome. To contribute:

1. Fork the repository
2. Create a new branch: git checkout -b my-new-feature
3. Make your changes and commit them: git commit -am 'Add some feature'
4. Push your changes: git push origin my-new-feature
5. Submit a pull request

Please ensure your code adheres to the existing style and passes tests before submitting.

Stargazers over time

Contact

• Email: zk@likedge.top or edge@ibooker.org.cn
• Website: Likedge

License

Edge-Computing-Engine is released under the Apache 2.0 License. See the LICENSE file for details.