netcl wiki
knowledge

NetCL Curriculum & Learning Tracks

NetCL Curriculum & Learning Tracks

Select a track based on your current engineering focus. Each track is an end-to-end progression from mathematical mechanics to verified OpenCL GPU execution.

Curriculum Roadmap

flowchart TD A["1. Foundations & Loss Surfaces"] --> B["2. Calculus, Gradients & Tape"] B --> C["3. Neural Building Blocks"] C --> D["4. Optimization & AdamW"] D --> E["Track A: Vision & Convolutions"] D --> F["Track B: Large Language Models"] D --> G["Track C: Geometry & Clustering"] D --> H["Track D: Evolutionary Search"] D --> I["Track E: Systems & Runtime"] E --> E1["Convolutions & im2col"] E1 --> E2["Winograd & ResNet"] F --> F1["Tokens & Multi-Head Attention"] F1 --> F2["FlashAttention & Local Tiling"] G --> G1["KMeans & Geometry"] G1 --> G2["Spectral Graph Theory & SSL"] H --> H1["NES & Black-Box Search"] H1 --> H2["SepCMAES & GPU Populations"] I --> I1["Memory Pool & JIT Caching"] I1 --> I2["Multi-GPU Ring Reduction"]

Tracks Directory

TRACK 01

Foundations to First Model

5 modules · beginner

Core intuition from parameter curve fitting to reverse-mode autograd and training an MNIST classifier on GPU buffers.

Foundations → Calculus & Tape → Activations & Projections → AdamW → MNIST Tutorial
TRACK 02

PyTorch to OpenCL Migration

5 modules · intermediate

Transition guide for PyTorch practitioners: explicit command queues, tape autograd on raw memory, and multi-GPU shared memory scaling.

Quickstart → Tape vs Autograd → Fused Kernels → Data-Parallel Multi-GPU → Architecture Spec
TRACK 03

LLMs, Attention & FlashAttention

4 modules · advanced

From tokenization and query-key-value routing to GPU local memory tiling and FlashAttention kernel execution.

Tokens & Attention → RoPE & SwiGLU → Hardware & FlashAttention → Text & Tokenizer API
TRACK 04

Computer Vision & Convolutional Backbones

5 modules · intermediate

Spatial locality, 2D receptive fields, im2col GEMM acceleration, Winograd arithmetic, and ResNet skip connections.

Convolutions & Vision → im2col GEMM → Winograd Filtering → BatchNorm & Normalization → ResNet Architecture
TRACK 05

Unsupervised Geometry & Self-Supervised Learning

5 modules · intermediate

Voronoi quantization, graph Laplacians with GPU LOBPCG eigensolvers, and non-contrastive representation learning.

KMeans & Geometry → Spectral Graph Theory → SimCLR & BYOL → Metric Losses → Cluster & SSL APIs
TRACK 06

Evolutionary Search & Neuroevolution

3 modules · intermediate

Black-box optimization when gradients are unavailable: Natural Evolution Strategies (NES), SepCMAES, and NAS with weight inheritance.

Evolutionary Optimization → Neuroevolution & NAS → Evolutionary API
TRACK 07

Systems & Bare-Metal Kernel Engineering

6 modules · systems

OpenCL runtime internals, driver allocation pools, dynamic JIT caching, custom C kernels, and multi-GPU host IPC.

Stack Layers → Tensor Backend → Memory Pool → JIT Compiler → Custom OpenCL C → Multi-GPU IPC

Quick Navigation: Tutorials vs Reference Index