Expand description
Physical operators (docs/SPEC.md §2.3): the Gpu*Exec DataFusion
ExecutionPlans, the backend-agnostic GpuOperator that runs each batch
on the local backend with CPU fallback, and the conformance suite (in
tests/) that pins every operator to the stock DataFusion semantics.
GpuFilterExec— fused filter + projectionGpuHashJoinExec— inner hash join on oneInt64keyGpuAggregateExec—SUM/COUNT/MIN/MAXgrouped by oneInt64keyGpuVectorDistanceExec— L2 / cosine distance column
Every exec records the planned backend (its EXPLAIN tag) but selects the
real backend at execute() time from the local hardware detector, so a
plan shipped to a machine without a GPU still runs correctly.
Dependency direction: depends on oxidelake-core, oxidelake-memory, oxidelake-device.
Re-exports§
pub use backend::init_local_backend;pub use backend::local_backend;pub use exec::GpuAggregateExec;pub use exec::GpuFilterExec;pub use exec::GpuHashJoinExec;pub use exec::GpuVectorDistanceExec;pub use exec::aggregate_output_schema;pub use operator::GpuOperator;pub use operator::JoinBuild;pub use operator::JoinProbe;pub use predict::ACTIVATION_KEY;pub use predict::Activation;pub use predict::Model;pub use predict::ModelSpec;pub use predict::PREDICT;pub use predict::predict_udf;pub use udf::COSINE_DISTANCE;pub use udf::L2_DISTANCE;pub use udf::cosine_distance_udf;pub use udf::distance_metric_for;pub use udf::distance_udf_name;pub use udf::l2_distance_udf;pub use udf::literal_query;pub use udf::oxide_udfs;pub use udf::query_literal;
Modules§
- backend
- Process-wide backend selection for operators.
- display
EXPLAINrendering helpers shared by the execs.- exec
- The four
Gpu*ExecDataFusionExecutionPlans and their shared plumbing. - operator
GpuOperator: runs one operator’s batches on a backend with per-batch CPU fallback and telemetry.- params
- Operator parameter types and the bounded v1 GPU type coverage.
- predict
- The
predictscalar UDF: run a trained model over a column, in SQL. - udf
l2_distance/cosine_distancescalar UDFs (docs/SPEC.md §2.8).
Structs§
- Aggregate
Spec - A grouped aggregation over one
Int64key.
Enums§
- Aggregate
Function - Aggregate functions supported by the grouped-aggregation kernel.
- Comparison
- Comparison operators supported by the fused filter kernel.
- Distance
Metric - Distance metric for the vector kernel.
- Literal
- A literal a column is compared against.
- Predicate
- A filter predicate in the bounded v1 grammar: column-vs-literal comparisons
combined with
AND.
Constants§
- FEATURE_
MASK - Which optional capabilities of this build change what a plan can say, as a bitmask.
Functions§
- gpu_
eligible_ scalar truefor the scalar types the v1 GPU kernels accept as filter, join-key and aggregate inputs.- vector_
dimension - For a
FixedSizeList<Float32, n>column returnsSome(n); otherwiseNone.