gorgonia/gorgonia
Gorgonia
Gorgonia is a Go library that provides an automatic differentiation engine and a computation graph framework for building and training neural networks. It functions as a CUDA-accelerated tensor library and a SIMD-optimized math library, enabling machine learning workflows entirely within the Go ecosystem.
The library distinguishes itself through a dual-backend architecture that dispatches neural network operations to either a GPU or CPU depending on CUDA availability at runtime. It constructs differentiable directed acyclic graphs of tensor operations, supports reverse-mode automatic gradient computation for backpropagation, and executes these graphs using a dedicated virtual machine. Performance is further enhanced through compile-time SIMD dispatch for SSE or AVX instruction sets and hand-optimized x86-64 assembly for vector math.
Gorgonia includes a CUDA code generation pipeline that produces hardware-specific kernel bindings and manages GPU memory and execution through the low-level CUDA Driver API. It supports GPU-accelerated neural network operators such as convolution and pooling, with automatic CPU fallback when CUDA is unavailable. The library also covers multivariate linear regression and convolutional neural network training for tasks like image classification.
Features
- Go Machine Learning Libraries - Provides a Go library for building and training neural networks with automatic differentiation and computation graph execution.
- Automatic Differentiation Engines - Provides an automatic differentiation engine that computes gradients via reverse-mode differentiation for backpropagation.
- Computational Graph Frameworks - Ships a directed acyclic graph framework for defining and executing mathematical operations in neural network models.
- CUDA Accelerated Neural Networks - Executes neural network operations on CUDA-compatible GPUs with automatic CPU fallback.
- Tensor Operation Accelerators - Offloads tensor operations to CUDA-enabled GPUs to speed up neural network training and inference.
- CUDA-Accelerated Libraries - Provides a tensor computation library that offloads operations to NVIDIA GPUs using CUDA.
- Tensor-Matrix Accelerators - Offloads tensor and matrix operations to NVIDIA GPUs using CUDA to speed up training and inference.
- Neural Network Training Frameworks - Builds and trains neural networks using automatic differentiation and graph computation for machine learning tasks.
- Reverse-Mode Differentiation - Computes gradients of any tensor expression in a graph using reverse-mode automatic differentiation.
- SIMD-Accelerated Arithmetic - Uses platform-specific SIMD instructions to accelerate mathematical operations for neural network training.
- Float32 Operations - Provides SIMD-accelerated float32 vector math as a core performance feature for neural network computations.
- Float64 Operations - Provides SIMD-accelerated float64 vector math as a core performance feature for neural network computations.
- x86-64 Assembly Accelerators - Uses hand-optimized x86-64 assembly instructions to accelerate float32 and float64 vector operations.
- SIMD Instruction Set Toggles - Provides build-time flags to enable SSE or AVX instruction sets for accelerated vector math.
- CUDA Driver API Integrations - Manages GPU memory and kernel execution directly through the low-level CUDA Driver API from Go code.
- GPU Kernel Code Generators - Produces CUDA source files needed to execute operations on a GPU using a code generation tool.
- GPU-CPU Backend Dispatchers - Routes neural network operations to either GPU or CPU backends at runtime based on CUDA availability.
- Graph-Based Computational Execution - Constructs and runs directed acyclic graphs of mathematical operations for model training and inference.
- Graph Construction Engines - Constructs a directed acyclic graph of mathematical operations that defines the structure of a neural network.
- GPU-Accelerated Computation - Offloads tensor and matrix operations to NVIDIA GPUs via CUDA for faster training and inference in Go.
- General-Purpose Tensor Operations - Offloads general tensor operations to NVIDIA GPUs using CUDA for accelerated neural network training and inference.
- CUDA Kernel Generators - Generates hardware-specific CUDA kernel source files and Go bindings for GPU acceleration.
- Computation Graph Runtimes - Evaluates computational graphs by traversing nodes and computing results using a dedicated virtual machine runtime.
- Convolutional Neural Networks - Constructs and executes convolutional neural network models for image classification tasks like digit recognition.
- CUDA Binding Generators - Creates CUDA kernel bindings tailored to the user's specific GPU hardware using a dedicated code generation tool.
- Linear Regression - Fits a linear regression model to datasets with multiple input features using gradient-based optimization.
- Context Initializers - Creates and manages a CUDA driver context for executing GPU operations from application code.
- SIMD Optimization - Uses platform-specific SIMD instructions on amd64 CPUs to accelerate vector and matrix operations.
- Build-Time SIMD Toggles - Enables SSE or AVX instruction sets at build time to accelerate vector math without changing application code.
- Assembly Kernels - Uses hand-optimized CPU assembly instructions to speed up mathematical operations on compatible architectures.
- CUDA Binding Generators - Creates hardware-specific CUDA bindings using a code generation tool to enable GPU-accelerated computation.
- CUDA Memory Transfers - Transfers data to GPU memory using the CUDA Driver API for accelerated computation.
- Virtual Machine Execution - Runs a pre-built computational graph on a dedicated virtual machine, computing every operation in the graph.
- AI and Machine Learning - Library for machine learning in Go.
- Data Processing and Machine Learning - Library for deep learning and tensor computation.
- General Machine Learning - Deep learning and graph-based computation library.
- Machine Learning - Graph-based computation library for machine learning.
- Machine Learning and AI - Graph-based computation library for building neural networks.
