This repository is a clean public implementation. It does not contain the proprietary AGX policy engine, optimization algorithms, internal scheduler, cache system, mesh enhancement system, interception layer, or commercial rendering backend.

Current features

Device and backend abstraction
Safe buffer allocation, upload, and download
CPU reference kernels
CUDA-like submit-and-dispatch learning model
Small examples that compile independently

Build

Requirements: Free Pascal Compiler 3.2 or newer and make.

make
make test

Run

./build/hello_device
./build/compute_demo

Architecture

Applications use the public runtime API. The included CPU backend is intentionally simple and readable. Future community backends may target Metal, Vulkan Compute, or DirectCompute without changing application code.

Project status

Experimental. The API may change before version 1.0.

Trademark

The license covers source code, not the AGX or Agila names, logos, or branding. See TRADEMARKS.md.