Compute 0.1.0

Compute 0.1.0 provides a TTX SDK for Tetrodotoxin's compute infrastructure, with separate CPU and CUDA engines.

Main update:

This refresh replaces the earlier 0.1.0 artifacts with the command-buffer model. Applications record work through shared Compute contracts and submit it in authored order. Providers manage execution and retain the resources needed by their results.

Features:

• Package the shared Compute SDK independently of the CPU and CUDA implementations.
• Record image operations into command buffers, with explicit submission and completion. Result access waits for the work needed to produce that result.
• Compose CPU and CUDA providers through the same TTX contracts. Policy requirements apply to every candidate during provider selection.
• Keep compatible CUDA results on the device through ordered stream dependencies.
• Retain inputs through their supplied interfaces or capture them while accessible. Result acquisitions remain valid independently of the originating engine acquisition.
• Report final resource cleanup through TTX Retained, including pending work and retained inputs.
• Preserve CUDA-specific compilation and device APIs for consumers that require them.
• Supply static and shared Compute and CPU packages for Linux, Windows and Wasm, plus CUDA packages for Linux and Windows.

This refreshed release replaces the earlier eager image API. Invert and Pixels have new contract UUIDs, and consumers must update their archive checksums and migrate to command buffers. The v0.1.0 tag and source snapshot identify the replacement revision.

SDK imports use project = "Compute" and select library = "compute", "cpu" or "cuda". The bundle uses TTX 0.4.1, Perimortem 0.4.0 and Toolchain 0.4.2. NVIDIA runtime components retain their supplied license terms; the installed NVIDIA driver remains a host requirement.

Native Windows execution remains unvalidated.