Crate runmat_accelerate 

RunMat Accelerate: GPU Acceleration Abstraction Layer

Goals:

• Provide a backend-agnostic API surface that maps RunMat operations to GPU kernels.
• Support multiple backends via features (CUDA, ROCm, Metal, Vulkan, OpenCL, wgpu).
• Allow zero-copy interop with runmat-builtins::Matrix where possible.
• Defer actual kernel authoring to backend crates/modules; this crate defines traits and wiring.

Re-exports

pub use native_auto::apply_auto_offload_calibration_from_file;

pub use native_auto::auto_offload_report;

pub use native_auto::is_sink;

pub use native_auto::prepare_builtin_args;

pub use native_auto::promote_binary;

pub use native_auto::promote_reduction_args;

pub use native_auto::promote_unary;

pub use native_auto::reset_auto_offload_log;

pub use native_auto::AutoOffloadCalibrationOutcome;

pub use native_auto::AutoOffloadCalibrationSummary;

pub use native_auto::AutoOffloadDecisionEntry;

pub use native_auto::AutoOffloadDisposition;

pub use native_auto::AutoOffloadReport;

pub use native_auto::BinaryOp;

pub use native_auto::CachedProviderInfo;

pub use native_auto::DecisionReason;

pub use native_auto::ReductionOp;

pub use native_auto::ThresholdBase;

pub use native_auto::ThresholdDelta;

pub use native_auto::ThresholdDeltaEntry;

pub use native_auto::ThresholdSnapshot;

pub use native_auto::UnaryOp;

pub use fusion::*;

pub use graph::*;

Modules

backend

fusion

fusion_exec

fusion_residency

graph

native_auto

precision

simple_provider

telemetry

Structs

AccelerateInitOptions

Initialization options for selecting and configuring the acceleration provider.

Accelerator

High-level façade for accelerated operations, falling back to runmat-runtime.

AutoOffloadOptions

Configuration passed to the native auto-offload planner.

DeviceInfo

Device descriptor used for selection and capabilities query.

Planner

Planner determines whether to execute on CPU or a selected backend. This will eventually consult sizes, heuristics, and device availability.

Enums

AccelPowerPreference

Power preference used when initializing a WGPU backend

AccelerateProviderPreference

Preferred acceleration provider selection

AutoOffloadLogLevel

Logging verbosity for auto-offload promotion decisions.

DeviceKind

High-level device kind. Concrete selection is provided by backend.

ExecutionTarget

ReductionAxes

Traits

AccelerateBackend

Core backend interface that concrete backends must implement.

BufferHandle

Abstract buffer that may reside on device or be host-pinned.

DeviceMatrix

Abstract matrix allocated on a device backend.

Functions

configure_auto_offload

initialize_acceleration_provider

Initialize the acceleration provider using default options.

initialize_acceleration_provider_with

Initialize the global acceleration provider using the supplied options.

value_is_all_keyword