DocsBackend

Backend API Reference

The cyxwiz-backend library provides the core computational functionality for CyxWiz, including tensor operations, neural network layers, optimizers, and GPU-accelerated algorithms.

Overview

The backend is a shared library (DLL/SO) used by:

CyxWiz Engine - For local training and inference
CyxWiz Server Node - For distributed job execution
pycyxwiz - Python bindings for scripting

Documentation Sections

Section	Description
Tensor Operations	Core tensor class and operations
Device Management	GPU/CPU device selection
Neural Network Layers	Layer implementations
Optimizers	Training optimizers
Loss Functions	Loss function implementations

Quick Start

C++ Usage

#include <cyxwiz/cyxwiz.h>

int main() {
    // Initialize backend
    cyxwiz::Initialize();

    // Create tensor
    cyxwiz::Tensor x({1.0f, 2.0f, 3.0f, 4.0f}, {2, 2});

    // Build model
    cyxwiz::Sequential model;
    model.Add(std::make_unique<cyxwiz::Linear>(2, 4));
    model.Add(std::make_unique<cyxwiz::ReLU>());
    model.Add(std::make_unique<cyxwiz::Linear>(4, 1));

    // Forward pass
    cyxwiz::Tensor output = model.Forward(x);

    // Cleanup
    cyxwiz::Shutdown();
    return 0;
}

Python Usage

import pycyxwiz as cyx

# Initialize
cyx.initialize()

# Create tensor
x = cyx.Tensor([1.0, 2.0, 3.0, 4.0], shape=[2, 2])

# Build model
model = cyx.Sequential()
model.add(cyx.Linear(2, 4))
model.add(cyx.ReLU())
model.add(cyx.Linear(4, 1))

# Forward pass
output = model.forward(x)
print(output.data())

Available Layers

Layer	Description	Parameters
Linear	Fully connected	in_features, out_features
Conv2d	2D convolution	in_ch, out_ch, kernel, stride, padding
BatchNorm2d	Batch normalization	num_features
Dropout	Dropout regularization	probability
ReLU	ReLU activation	-
Sigmoid	Sigmoid activation	-
Softmax	Softmax activation	dim

Available Optimizers

Optimizer	Description	Key Parameters
SGD	Stochastic gradient descent	lr, momentum, weight_decay
Adam	Adaptive moments	lr, betas, eps, weight_decay
AdamW	Adam with decoupled decay	lr, betas, eps, weight_decay
RMSprop	Root mean square prop	lr, alpha, eps

Available Loss Functions

Loss	Description	Use Case
MSELoss	Mean squared error	Regression
CrossEntropyLoss	Cross entropy	Classification
BCELoss	Binary cross entropy	Binary classification
L1Loss	Mean absolute error	Robust regression

GPU Acceleration

The backend uses ArrayFire for GPU acceleration:

// Check available backends
if (af::isBackendAvailable(AF_BACKEND_CUDA)) {
    cyxwiz::Device::SetDevice(cyxwiz::Device::Type::CUDA);
} else if (af::isBackendAvailable(AF_BACKEND_OPENCL)) {
    cyxwiz::Device::SetDevice(cyxwiz::Device::Type::OpenCL);
} else {
    cyxwiz::Device::SetDevice(cyxwiz::Device::Type::CPU);
}

Backend Priority

CUDA - NVIDIA GPUs (best performance)
OpenCL - AMD/Intel GPUs
CPU - Fallback (always available)

Platform Notes

Windows

Library: cyxwiz-backend.dll
Import lib: cyxwiz-backend.lib
Requires: MSVC 2022+

Linux

Library: libcyxwiz-backend.so
Requires: GCC 10+

macOS

Library: libcyxwiz-backend.dylib
Requires: Clang 12+