Expand description
Coordinated GPU/NPU execution with inferred memory dependencies. Coordinated execution for GPU and NPU kernels.
Device placement is explicit. Contracts describe memory access; prepared
graphs infer dependencies and retain all allocations through completion.
The low-level GPU API remains available in crate::gpu.
Structs§
- Binding
Contract - A fixed-specialization buffer argument contract.
- Buffer
- A cloneable, tracked allocation. Every alias shares lifetime and access state.
- Buffer
View - An owned checked view. Keeping a view keeps its root allocation alive.
- Completion
- An owned completion observer. Dropping it does not cancel submitted work.
- Completion
Profile - Host-observed submission latency, including dependency and driver overhead. This is not a measurement of GPU kernel execution time.
- Executable
Graph - A prepared graph with fixed bindings and reusable bounded submission slots.
- Execution
Trace - Bounded trace of host-observed synchronized region latencies, including driver and scheduling overhead. These are never GPU timestamps.
- GpuAccess
- A buffer region and the access performed during an external GPU handoff.
- GpuDevice
- A selected GPU and its verified compilation target.
- GpuKernel
- Trusted, fixed GPU launch dimensions and argument contract.
- Graph
- A pipeline builder. Each operation declares placement and actual buffer use.
- Kernel
Contract - The trusted ABI of a fixed kernel specialization.
- Native
Session - Owned stream-bound model state scheduled alongside prepared GPU graphs.
- Node
- An operation index, useful for diagnostics. Ordering is inferred from access.
- NpuDevice
- An explicitly selected NPU ordinal.
- Read
Guard - A host read lease. The slice cannot outlive the guard.
- Runtime
- Shared scheduler and allocation domain. Clones use the same dependency state.
- Runtime
Options - Runtime limits fixed before workers or prepared run slots are created.
- Statistics
- Cumulative counters for one coordinated runtime. Native low-level calls are outside this accounting domain. Reads are snapshots, not global barriers.
- Trace
Event - A synchronized native region as observed by its execution worker.
- Write
Guard - An exclusive host write lease. Releasing it marks host writes authoritative.
Enums§
- Access
- Device access used for dependency inference and kernel contract validation.
- GpuLane
- Independent GPU scheduling lanes. Memory hazards apply across every lane.
- Memory
Placement - Explicit storage placement. Shared storage requires a resident NPU program to create its device mapping once, before execution.