Expand description
The Kernel trait and kernel-match / cost types (§4.2).
Structs§
- Cost
- A cost estimate for running a kernel, consumed by the placement cost model
(
docs/ORT2.md§6). All time fields are in microseconds; a fuller model (roofline, calibration) lands inonnx-runtime-cost-model(Phase 2). - Kernel
Variant Selection - The concrete implementation selected by a kernel’s internal dispatcher.
- View
Output - A zero-copy view output: a kernel’s declaration that one of its outputs
is a strided view aliasing one of its inputs’ buffers, rather than freshly
computed bytes (
docs/ORT2.md§5.4, lazy PyTorch-style views).
Enums§
- Capture
Support - Whether a compiled kernel can participate in device-graph capture.
- Kernel
Input - An executor-delivered kernel input. Existing EPs receive
Tensorvariants; an EP advertising thenxrtcapability may receive a lazyWeightat thepkg.nxrt::BlockQuantizedMoEboundary. - Kernel
Match - Result of
crate::ExecutionProvider::supports_op.
Constants§
- ARG_
KERNEL_ VARIANT - Trace-argument key carrying the selected kernel implementation.
- ARG_
KERNEL_ VARIANT_ REASON - Trace-argument key carrying why the kernel variant was selected.
Traits§
- Kernel
- A kernel ready to execute a specific op with specific shapes (§4.2).
Functions§
- kernel_
variant_ tracing_ enabled - Whether kernel-variant trace annotations would currently be recorded.
- record_
kernel_ variant_ selection - Record a selected kernel variant on the active runtime trace span.
- record_
kernel_ variant_ stage_ selection - Record a selected kernel variant for a named sub-decision of the current op.