oa::Engine
The sole local owner of device, memory, queues, kernels, scheduling, and profiling. Optional systems borrow it through composition.
High Performance Compute
The compute layer.
One capability-gated runtime.
A low-level Rust library with native Python bindings for compute, Ml, Vision, Audio, Media, Data, Crypto, Render, Ui, and Plot. The current complete acceptance pack is Intel Iris Xe; every additional device family is qualified separately.
Architecture
Four pillars of the compute substrate.
Memory
oa::Engine-owned allocation, bindless resources, bounded upload/readback rings, and graph lifetime analysis behind semantic values.
Compute Kernels
One embedded manifest owns compilation, stable identities, capability requirements, routing, hashes, and validation fixtures.
Cryptography
Strict host primitives plus capability-gated Vulkan batch hashing and public-data acceleration, with no certification claim.
Runtime Integration
Core, Ml, Vision, Audio, Media, Render, Ui, Plot, and Crypto borrow one oa::Engine without exposing Vulkan machinery publicly.
Capabilities · Compatibility
Everything between your code and the hardware.
One API, capability-gated routes. oa::Engine owns local execution while domain modules such as oa::matrix preserve the semantic contract across Rust and Python. Hardware-specific decisions stay private; unsupported and unqualified routes stay explicit.
oa::matrix and domain modules
Stateless, schema-owned transformations share validation, effects, differentiation, lowering metadata, and paired language surfaces.
oa::Event
Every submission returns its exact completion value. Waiting, polling, dependency chaining, and failure remain deliberate boundaries.
oa::ExecutionPlan
Immutable captured work can be submitted repeatedly without exposing the private executable Vulkan graph or rebuilding stable regions.
Vendor-neutral contracts
Applications express semantic work once. Kernel, precision, subgroup, media, and memory routes remain internal capability decisions.
Device-resident semantics
Matrices, images, audio, video frames, textures, and meshes may share storage without surrendering their public meaning.
Embedded kernels
Kernels compile and validate at build time, embed behind stable identities, and fail closed when a required device capability is absent.
Qualified hardware packs
The current Core reference is Intel Iris Xe FP32. NVIDIA, AMD, mobile, and datacenter routes remain explicitly unverified until requalified.
Foundation
The oa::Library
Black magic mixed with alien technology. A hand-crafted Rust library for matrix operations, graph execution, machine learning, Vision, Audio, Media, Render, Ui, Plot, cryptography, and the runtime beneath them all.
Values carry semantics, stateless operations in oa::matrix and sibling modules describe transformations, and sessions own stateful activity. The semantic operation graph stays separate from the private Vulkan execution graph.
Every abstraction earned. One oa::Engine owner, one schema authority, one embedded kernel manifest, and no hidden path from eager thought to captured replay.
Compute Engine
oa::Engine
The all-seeing eye. One pinned object holds the local compute world—device, memory, queues, kernels, scheduling, and profiling. Every matrix, image, waveform, frame, and plot moves through a single point of consciousness; the modules borrow its power without seeing the machinery beneath.
Black magic mixed with alien technology, engineered as explicit Rust and Vulkan. Operations describe intent, submission returns the exact oa::Event, and recurring work condenses into immutable oa::ExecutionPlan values while the semantic graph remains separate from the executable Vulkan graph.
oa::Engine
Ownership
oa::Event
Completion
oa::ExecutionPlan
Replay
oa::matrix
Operations
Captured Execution
oa::ExecutionPlan
Capture a reusable semantic region once, receive an immutableoa::ExecutionPlan, then submit that plan through the same oa::Engine and exact-event completion contract as eager work.
The private executable graph tracks resource reads, writes, aliases, lifetime edges, barriers, and lowering provenance. The public plan owns replay, not raw Vulkan graph authoring or inspection.
Automatic OaDnn capture and training-program replay use this same boundary. Performance claims remain workload-, device-, driver-, and measurement-specific.
