Onumpy - GPU-Accelerated NumPy for AMD GPUs

Onumpy is a drop-in replacement for NumPy that automatically accelerates large array operations on AMD GPUs using OpenCL and rocBLAS.

🚀 Key Features

• Automatic GPU Acceleration: Seamlessly uses GPU for large arrays (>100K elements)
• 100% NumPy Compatible: Drop-in replacement - no code changes needed
• AMD RDNA1 Optimized: Tuned for RX 5700 XT (wavefront 64, 20 CUs, 8GB VRAM)
• Smart Dispatch: Automatically selects GPU or CPU based on array size
• Memory Efficient: Offloads large arrays to GPU VRAM
• Native Math Functions: 2-5x speedup with native_exp, native_log, etc.
• Vectorized Operations: 1.5-3x speedup with float4 vectorization

📊 Performance

Matrix Operations (rocBLAS)

• 10-100x speedup for matrices >1000×1000
• Reduced CPU memory usage
• Better GPU utilization

Element-wise Operations (OpenCL)

• 2-5x speedup with native math functions (native_exp, native_log)
• 1.5-3x speedup with vectorized operations (float4)
• Parallel execution on 20 compute units

Reductions

• 10-50x speedup for arrays >1M elements
• Efficient GPU parallel reduction

🎯 Quick Start

Installation

# Clone repository
# NOTE: Replace YOUR_GITHUB_USERNAME with your actual GitHub username
git clone https://github.com/YOUR_GITHUB_USERNAME/Onumpy.git
cd Onumpy

# Install dependencies
pip install numpy pyopencl

# Build and install
cd numpy_gpu
python setup.py build_ext --inplace
python setup.py install

See Installation Guide for detailed instructions.

Basic Usage

from numpy_bridge import np

# Works exactly like NumPy
arr = np.array([1, 2, 3, 4, 5])

# Automatically uses GPU for large operations
large_array = np.random.rand(1000000)
result = np.exp(large_array)  # Uses GPU automatically

# Explicit GPU methods
result = np.gpu_exp(large_array)  # Force GPU usage

# Check GPU availability
if np.GPU_AVAILABLE:
    print("GPU acceleration enabled!")

Advanced Usage

from numpy_bridge import np

# Matrix operations automatically use GPU for large matrices
a = np.random.rand(2000, 2000).astype(np.float32)
b = np.random.rand(2000, 2000).astype(np.float32)
result = np.matmul(a, b)  # Automatically uses GPU

# Element-wise operations
large_array = np.random.rand(1000000).astype(np.float32)
result = np.exp(large_array)  # GPU-accelerated

# Explicit GPU control
result = np.gpu_matmul(a, b)  # Force GPU usage
result = np.gpu_exp(large_array)  # Force GPU usage

📖 Documentation

• Installation Guide - Detailed setup instructions
• Performance Benchmarks - Actual performance data
• API Reference - Complete API documentation
• Examples - Practical usage examples
• Consolidation Guide - Migration from standard NumPy

🔧 Requirements

• Python: 3.8+
• NumPy: 2.0+
• OpenCL: ROCm 5.7+ for AMD GPUs
• rocBLAS: Optional, for matrix operations (recommended)
• GPU: AMD GPU with ROCm support (RX 5700 XT tested)

🆚 Why Onumpy?

Feature	NumPy	CuPy	Onumpy
AMD GPU Support	❌	⚠️ Limited	✅ Native
Drop-in Replacement	-	✅	✅
Automatic Dispatch	❌	❌	✅
RDNA1 Optimized	❌	❌	✅
OpenCL Support	❌	❌	✅
rocBLAS Integration	❌	⚠️ Partial	✅ Full

Onumpy is specifically designed for AMD GPUs and provides native OpenCL/ROCm support, making it the ideal choice for AMD GPU users who want GPU acceleration without switching to NVIDIA hardware.

🏗️ Architecture

Onumpy/
├── numpy_bridge.py          # Unified NumPy bridge (auto GPU dispatch)
├── numpy_gpu/               # GPU-accelerated extension
│   ├── api/                 # Python API
│   ├── src/                 # C/C++ implementation
│   ├── kernels/             # OpenCL kernels (RDNA1-optimized)
│   └── setup.py             # Build configuration
├── benchmarks/              # Performance benchmarks
└── docs/                    # Documentation

🎓 Use Cases

Machine Learning Training

• Volkner: Mamba SSM and Transformer operations
• Large batch training data processing
• Model weight operations

Scientific Computing

• Large-scale numerical simulations
• Matrix operations on big datasets
• Element-wise operations on large arrays

Data Processing

• Vector operations on large datasets
• Mathematical transformations
• Statistical computations

🧪 Testing

# Run basic tests
cd numpy_gpu
python tests/test_basic.py

# Run benchmarks
python benchmarks/benchmark_matmul.py

# Run comprehensive benchmarks
cd ../benchmarks
python run_all_benchmarks.py

🤝 Contributing

Contributions welcome! Please see CONTRIBUTING.md for guidelines.

📄 License

https://github.com/CharlesMcGowen/ONumpy/blob/main/LICENSE

🙏 Acknowledgments

• NumPy team for the excellent base library
• AMD for ROCm and OpenCL support
• OpenCL community for GPU computing standards

📝 Notes

• This is a custom extension, not a fork of NumPy
• It extends NumPy with GPU capabilities while maintaining 100% compatibility
• Optimized specifically for AMD RX 5700 XT (RDNA1 architecture)
• Designed for Volkner's training workloads but works for any large array operations

---

Status: Alpha - Actively developed and tested on AMD RX 5700 XT