OptiRS GPU

GPU acceleration and multi-GPU optimization for the OptiRS machine learning optimization library.

Overview

OptiRS-GPU provides hardware acceleration for machine learning optimization workloads across multiple GPU backends. This crate enables high-performance, parallel optimization on CUDA, Metal, OpenCL, and WebGPU platforms, with automatic device selection and memory management.

Features

• Multi-Backend Support: CUDA, Metal, OpenCL, and WebGPU compatibility
• Automatic Device Detection: Intelligent GPU selection and fallback mechanisms
• Multi-GPU Training: Distributed optimization across multiple GPUs
• Memory Management: Efficient GPU memory allocation and transfer
• Async Operations: Non-blocking GPU kernel execution
• Cross-Platform: Unified API across different GPU ecosystems
• Performance Monitoring: GPU utilization and memory usage tracking

Supported Backends

CUDA (NVIDIA)

• Full CUDA runtime integration via cudarc
• Support for CUDA 12.0+ features
• Optimized kernels for tensor operations
• Multi-GPU scaling with NCCL integration

Metal (Apple Silicon)

• Native Metal Performance Shaders integration
• Optimized for Apple M1/M2/M3 architecture
• Unified memory architecture support
• Metal Performance Shaders Graph integration

OpenCL (Cross-platform)

• OpenCL 3.0 support for maximum compatibility
• Vendor-agnostic GPU acceleration
• Support for Intel, AMD, and other OpenCL devices
• Custom kernel compilation and caching

WebGPU (Web/Native)

• Cross-platform GPU acceleration via wgpu
• WebAssembly compatibility for web deployment
• Vulkan, DirectX, and Metal backend support
• Portable compute shaders

Installation

Add this to your Cargo.toml:

[dependencies]
optirs-gpu = "0.1.0-beta.3"
scirs2-core = "0.1.0-rc.2"  # Required foundation

Feature Selection

Enable specific GPU backends:

[dependencies]
optirs-gpu = { version = "0.1.0-beta.3", features = ["cuda", "metal"] }

Available features:

• cuda: NVIDIA CUDA support
• metal: Apple Metal support
• opencl: OpenCL support
• wgpu: WebGPU support (enabled by default)

Usage

Basic GPU Optimization

use optirs_gpu::{GpuOptimizer, DeviceManager};
use optirs_core::optimizers::Adam;

// Initialize GPU device manager
let device_manager = DeviceManager::new().await?;
let device = device_manager.select_best_device()?;

// Create GPU-accelerated optimizer
let mut optimizer = GpuOptimizer::new(device)
    .with_optimizer(Adam::new(0.001))
    .build()?;

// Your model parameters (automatically transferred to GPU)
let mut params = optimizer.create_tensor(&[1024, 512])?;
let grads = optimizer.create_tensor_from_slice(&gradient_data)?;

// Perform optimization step on GPU
optimizer.step(&mut params, &grads).await?;

Multi-GPU Training

use optirs_gpu::{MultiGpuOptimizer, DataParallelStrategy};

// Setup multi-GPU training
let mut multi_gpu = MultiGpuOptimizer::new()
    .with_strategy(DataParallelStrategy::AllReduce)
    .with_devices(&device_manager.available_devices())
    .build()?;

// Distribute model across GPUs
multi_gpu.distribute_model(&model_parameters).await?;

// Synchronized optimization across all GPUs
multi_gpu.step_synchronized(&gradients).await?;

Backend-Specific Features

CUDA-Specific Operations

use optirs_gpu::cuda::{CudaContext, CudaStream};

let cuda_ctx = CudaContext::new(device_id)?;
let stream = CudaStream::new(&cuda_ctx)?;

// Custom CUDA kernels
let result = stream.launch_kernel("custom_optimizer", &params, &config).await?;

Metal-Specific Operations

use optirs_gpu::metal::{MetalDevice, MetalCommandQueue};

let metal_device = MetalDevice::system_default()?;
let command_queue = metal_device.new_command_queue();

// Metal Performance Shaders integration
let mps_optimizer = command_queue.create_mps_optimizer(&config)?;

Architecture

Device Management

• Automatic GPU detection and selection
• Device capability querying
• Memory and compute capability assessment
• Fallback mechanism for unsupported operations

Memory Management

• Efficient GPU memory allocation
• Automatic CPU-GPU data transfer
• Memory pool management
• Garbage collection for GPU resources

Kernel Management

• Optimized compute kernels for common operations
• JIT compilation and caching
• Kernel fusion for performance optimization
• Platform-specific optimizations

Performance Considerations

Memory Transfer Optimization

• Asynchronous CPU-GPU transfers
• Memory pinning for faster transfers
• Batch operations to reduce transfer overhead
• Smart caching of frequently accessed data

Compute Optimization

• Kernel fusion for reduced memory bandwidth
• Optimal thread block sizing
• Memory coalescing patterns
• Shared memory utilization

Multi-GPU Scaling

• Efficient gradient synchronization
• Load balancing across devices
• Topology-aware communication
• Overlap communication and computation

Error Handling

OptiRS-GPU provides comprehensive error handling for GPU operations:

use optirs_gpu::error::{GpuError, GpuResult};

match optimizer.step(&mut params, &grads).await {
    Ok(()) => println!("Optimization successful"),
    Err(GpuError::OutOfMemory) => {
        // Handle GPU memory exhaustion
        optimizer.clear_cache()?;
    }
    Err(GpuError::DeviceNotAvailable) => {
        // Fallback to CPU optimization
        fallback_to_cpu_optimizer()?;
    }
    Err(e) => return Err(e.into()),
}

Benchmarking

OptiRS-GPU includes built-in benchmarking tools:

use optirs_gpu::benchmarks::{GpuBenchmark, BenchmarkConfig};

let benchmark = GpuBenchmark::new()
    .with_config(BenchmarkConfig::default())
    .with_optimizer(optimizer)
    .build()?;

let results = benchmark.run_performance_suite().await?;
println!("GPU throughput: {:.2} GFLOPS", results.throughput);

Platform Support

Contributing

OptiRS follows the Cool Japan organization's development standards. See the main OptiRS repository for contribution guidelines.

License

This project is licensed under either of:

• Apache License, Version 2.0
• MIT License

at your option.