Quantified API

《▒ Welcome to the New Generation ▒》

With the release of V2, Quantified API has been upgraded by a LOT.

The scheduler, CPU runtime, caching system, task graphs, GPU routing, compatibility layer, supported platforms, build system, and general API design have all been heavily improved or completely rebuilt.

Quantified API now also has full official documentation, covering everything from basic installation to advanced compute workloads, GPU routing, caching, migration, troubleshooting, and examples :DDD

《▒ What It Actually Does ▒》

• Proper async scheduling with worker lanes, workload routing, telemetry, and sane fallback behaviour
• Parallel execution for workloads that can be safely divided across multiple CPU threads
• Dependency-aware task graphs for workflows where jobs need to run in a specific order
• Backend-aware compute routing across CPU, Vulkan, and OpenCL
• Duplicate task suppression so repeated identical work can be merged instead of executed hundreds of times
• Advanced caching with memory, disk, persistence, expiration, compression, and refresh support
• Automatic mod registration through Fabric, Forge, and NeoForge metadata
• Live diagnostics through runtime telemetry and the Quantified webpanel

《▒ Vulkan and OpenCL Acceleration ▒》

Quantified API provides Vulkan compute acceleration that simply works without requiring users to manually configure a bunch of random settings first.

Vulkan is the preferred GPU backend, while OpenCL remains available as a compatibility and fallback option.

GPU acceleration isn't blindly used for every tiny task either. Quantified API considers the workload, runtime state, available hardware, and preferred backend before deciding where the work should run.

• Isolated Vulkan probing so broken drivers or loader issues don't instantly brick the game
• Deferred initialization so GPU startup doesn't fight with Minecraft while the game is still loading
• Cleaner runtime states so failed or half-initialized backends aren't treated as ready
• Automatic fallback routing from Vulkan to OpenCL and finally to CPU when needed
• Device selection and diagnostics through the Quantified webpanel

《▒ Faster Runtime ▒》

The runtime underneath Quantified API was heavily redesigned to reduce the amount of overhead added around the work mods actually want to execute.

• Lower overhead for tiny and unique tasks
• Much stronger duplicate burst suppression
• Cheaper parallel workload orchestration
• Massively faster execution for smaller task graphs
• Better handling of thread-safe and non-thread-safe workloads
• Improved persistent and asynchronous cache behaviour

Basically, it now spends a lot less time managing work and a lot more time actually doing it. Crazy idea, ik xd

《▒ Performance Improvements ▒》

Internal runtime benchmarks against the older v1.4.4 release showed major improvements across all tested workload types.

• Tiny unique task bursts: around 3.95x faster
• Duplicate task bursts: around 45.4x faster
• Small parallel workloads: around 5.54x faster
• Medium parallel workloads: around 9.16x faster
• Small DAG workloads: up to around 2952x faster in the tested setup

Duplicate suppression also improved from 2048 real executions to only 1 real execution in the tested duplicate burst.

So yh, V2 wasn't exactly a small update :DDD

《▒ Quantified Webpanel ▒》

Quantified API includes a live webpanel so developers and users can see what the runtime is actually doing instead of blindly trusting the mysterious background threads.