NetCL API Reference & Symbol Index
NetCL API Reference & Symbol Index
Direct index for all public namespaces, class signatures, and GPU kernel wrappers in NetCL. For step-by-step conceptual walkthroughs, see Learning Tracks.
1. Core Runtime, Memory & Precision
The foundation of NetCL manages OpenCL platforms, command queues, raw GPU memory buffers, and mixed-precision execution:
- Core & Device Manager:
DeviceManager,DeviceHandle,OpenCLBackend, and hardware discovery. - Tensor Backend:
Tensor,BufferPool, memory allocations, strides, and slicing. - Runtime & Kernel Compilation: Dynamic OpenCL C JIT compilation, caching, and workgroup dispatch.
- Automatic Mixed Precision (AMP):
autocast,GradScaler, and FP16 kernel execution.
2. Autograd & Computational Graph
- Autograd API:
Tape,Node,ag.tensor(), reverse-mode backpropagation, and graph retention. - Ops Reference: Low-level mathematical operations, GEMM matrix multiplications, and fused kernel primitives.
3. Neural Network Layers & Architectures
- Neural Network Modules (nn):
Linear,Conv2d,BatchNorm2d,LayerNorm,RMSNorm,Dropout,Embedding,Sequential, and initialization utilities (kaiming_uniform,xavier_normal). - Trainer API: High-level
Trainer, training loop abstractions, evaluation callbacks, and checkpointing.
4. Optimizers & Learning Rate Schedulers
- Optimizers (optim):
SGD,Adam,AdamW(with fused single-pass OpenCL kernels),RMSprop,WarmupCosine,StepLR, and gradient clipping (clip_grad_norm).
5. Data Pipelines, Storage & Profiling
- Data Pipeline:
Dataset,DataLoader,TensorDataset, batch collators, and host worker prefetching. - Serialization & I/O: Safe checkpoint saving, NumPy export, and weights serialization.
- Profiling & Diagnostics: OpenCL event timers, kernel execution latency, and memory bandwidth profilers.
6. Specialized Domains
- Text & Tokenization:
BPETokenizer,RegexTokenizer, vocab building, encoding, decoding, and chat template formatting. - Clustering & Vector Search:
KMeans,MiniBatchKMeans,SpectralClustering,GaussianMixture, and device-side argmin kernels. - Self-Supervised Learning (SSL):
SimCLR,BYOL,SwAV, InfoNCE loss, and linear probe evaluation monitors. - Evolutionary Optimization:
GA,OpenAIES,SepCMAES,ArchSpace, and GPU-resident population mutation. - Distributed Multi-Device:
all_reduce,broadcast,scatter,gather, and workstation shared-memory IPC.