Crate optirs_gpu 

OptiRS GPU - GPU Acceleration for ML Optimization

Version: 0.1.0 Status: Framework Ready (GPU Kernels Coming Soon)

optirs-gpu provides GPU acceleration for OptiRS optimizers, built on SciRS2’s GPU abstractions.

Dependencies

• scirs2-core 0.1.1 - Required foundation

Implementation Status (v0.1.0)

• ✅ GPU context management
• ✅ Multi-backend support framework (CUDA, Metal, OpenCL, WebGPU)
• ✅ Memory transfer utilities
• ✅ Configuration and initialization
• 🚧 GPU kernels (in development)
• 🚧 Tensor cores support (in development)
• 📝 Multi-GPU coordination (planned)

Status: Coming Soon

This crate is under active development. GPU acceleration will leverage:

• scirs2_core::gpu for GPU context and memory management
• scirs2_core::tensor_cores for mixed-precision training
• scirs2_core::array_protocol::GPUArray for zero-copy operations

Planned Features

Multi-Backend Support

• CUDA - NVIDIA GPUs with full tensor core support
• Metal - Apple Silicon M1/M2/M3 with unified memory
• OpenCL - Cross-platform GPU compute
• WebGPU - Browser and cross-platform support

Performance Optimizations

• Tensor Cores - FP16/BF16 mixed-precision training
• Memory Pools - Advanced GPU memory management
• Kernel Fusion - Optimized kernel execution
• Multi-GPU - Distributed optimization across GPUs

Expected Speedup

• 10-50x for large models (1M+ parameters)
• 100x+ for very large models (100M+ parameters)
• Near-linear scaling with multiple GPUs

Example Usage (Future)

use optirs_gpu::GpuOptimizer;
use optirs::prelude::*;
use scirs2_core::ndarray::Array1;

// Create GPU-accelerated optimizer
let optimizer = Adam::new(0.001);
let mut gpu_opt = GpuOptimizer::new(optimizer)?;

// Use like any optimizer - GPU acceleration is automatic
let params = Array1::from_elem(1_000_000, 1.0);
let grads = Array1::from_elem(1_000_000, 0.01);
let updated = gpu_opt.step(&params, &grads)?;

Architecture

Built exclusively on SciRS2:

• GPU Context: scirs2_core::gpu::GpuContext
• GPU Memory: scirs2_core::gpu::GpuBuffer
• GPU Kernels: scirs2_core::gpu::GpuKernel
• Tensor Cores: scirs2_core::tensor_cores
• Zero-Copy: scirs2_core::array_protocol::GPUArray

Contributing

GPU acceleration development follows SciRS2 integration guidelines. All GPU operations must use scirs2_core::gpu abstractions.

Re-exports

pub use backends::GpuBackend;

pub use memory::MemoryPool;

Modules

backends

GPU Backend Abstraction Layer

kernels

GPU Kernel Management and Compilation

memory

multi_gpu

tensor_cores

utils

Enums

GpuOptimError

Error type for GPU optimizer operations

Traits

GpuOptimizer

Trait for GPU-accelerated optimizers