Plast: A High-Performance Deep Learning Engine from Scratch

Plast is a professional-grade, zero-dependency deep learning framework built entirely from the ground up in C and CUDA. It bridges the gap between high-level AI research and low-level systems engineering, delivering a high-performance training engine that rivals the core implementations of industry standards.

Engineering Excellence & Technical Depth

1. Production-Ready Autograd Engine

Plast implements a sophisticated Automatic Differentiation engine based on Dynamic Computational Graphs.

• Topological Optimization: Intelligently schedules operations via DAG-based topological sorting to minimize memory footprint.
• Robust Gradient Propagation: Advanced accumulation logic that correctly handles complex edge cases like node re-use and non-contiguous gradient flow.
• Extensible Operator API: A clean, decoupled architecture allows for seamless integration of new mathematical operations without graph-level changes.

2. High-Performance numerical Kernels

Engineered for raw throughput across heterogeneous hardware:

• Massively Parallel CUDA Kernels: Tiled matrix operations with shared memory optimizations, designed to exploit maximum warp occupancy and memory coalescing on NVIDIA GPUs.
• SIMD Optimized CPU Backends: Leverages AVX/NEON intrinsics and OpenMP multi-threading to achieve near-theoretical peak performance on modern CPUs.
• Stride-Aware Logic: Efficiently handles arbitrary tensor layouts (slices, transposes, views) through zero-copy "virtual" tensors and localized packing utilities.

3. Industrial-Grade Memory Management

Plast avoids the overhead of standard garbage collection through a custom Arena Allocation system:

• Zero-Latency Training: Pre-allocates memory for the entire computation graph, eliminating the need for malloc/free during the critical training path.
• Deterministic Memory Profile: Perfect for resource-constrained environments where predictable memory behavior is non-negotiable.
• Cross-Device Unified Interface: A consistent memory management layer across CPU and GPU memory spaces.

System Architecture

graph TD
    A[User API] --> B[Dynamic Computation Graph]
    B --> C[Optimized Task Scheduler]
    C --> D[Hardware-Specific Backends]
    D --> E[AVX/SIMD CPU Kernels]
    D --> F[NVIDIA CUDA Kernels]
    G[Arena Memory Manager] --> B
    G --> D

Loading

Roadmap & Future Vision

• Python Ecosystem Integration: CFFI/Pybind11 bindings for seamless interoperability with the broader AI ecosystem.
• Mixed Precision Training: Support for FP16 and BF16 to accelerate training on modern Tensor Cores.
• Distributed High-Performance Computing: Cluster-scale training support via MPI and NCCL.
• Layer-Level Abstractions: High-level modules for Conv2D, Transformers, and BatchNorm.