Crate baracuda_kernels_types 

baracuda-kernels-types

Shared type vocabulary for the baracuda ML kernel facade.

This crate has no behavior of its own — it ships pure-data types that are common to every member of the kernel facade ecosystem:

• The KernelDtype umbrella marker trait + the Element / IntElement / FpElement / BinElement / BiasElement trait hierarchy plus the ScalarType alpha/beta projection.
• Wrapper types (S8, U8, S4, U4, Bin, F32Strict, Fp8E4M3, Fp8E5M2) that drive kernel selection at the Rust type level.
• Tag enums (ElementKind, MathPrecision, BiasElementKind, LayoutSku, ArchSku, EpilogueKind, ActivationKind).
• Borrowed device-buffer views (MatrixRef, MatrixMut, VectorRef).
• Plan-layer descriptors (PlanPreference, PrecisionGuarantee, Workspace).

1.0-freeze stability (Phase 28)

Most op-family discriminant enums (BinaryKind, UnaryKind, ReduceKind, ScanKind, SoftmaxKind, NormalizationKind, LossKind, LinalgKind, ConvKind, PoolKind, AttentionKind, IndexingKind, SegmentKind, EmbeddingKind, QuantizeKind, GgufBlockFormat, MoeKind, SortKind, ImageKind, FftKind, RandomKind, ShapeLayoutKind, BinaryCmpKind, TernaryKind, GatedActivationKind, ArgReduceKind, PadMode) plus the category / backend tag enums (OpCategory, BackendKind) and the auxiliary index-dtype tag enums (IndexElementKind, IndexOutputKind) are marked #[non_exhaustive] as of Phase 28. Downstream match arms must include a _ => catch-all to remain forward-compatible with new variants.

The kernel-dispatch-keying enums ElementKind, BiasElementKind, LayoutSku, ArchSku, EpilogueKind, ActivationKind, and Workspace<'a> are intentionally left exhaustive — they’re the keys per-arch / per-layout / per-epilogue / per-bias-dtype kernel dispatchers exhaustively match on, so adding a new variant is a deliberate workspace-wide event that ought to surface as a build break across every match site.

The types here were previously defined in baracuda-cutlass::types; they were lifted out so that baracuda-kernels (the unified ML op facade) and any sibling wrapper crate (baracuda-cublas, baracuda-cudnn, …) can share one vocabulary instead of each re-declaring its own.

The trait Element was previously named CutlassElement; baracuda-cutlass keeps the old name available as a re-export for back-compat. The semantics are unchanged.

Re-exports

pub use element::BiasElement;

pub use element::BiasElementKind;

pub use element::Bin;

pub use element::BinElement;

pub use element::Bool;

pub use element::Complex32;

pub use element::Complex64;

pub use element::Element;

pub use element::ElementKind;

pub use element::F32Strict;

pub use element::Fp8E4M3;

pub use element::Fp8E5M2;

pub use element::FpElement;

pub use element::IndexElement;

pub use element::IndexElementKind;

pub use element::IndexOutputElement;

pub use element::IndexOutputKind;

pub use element::IntElement;

pub use element::KernelDtype;

pub use element::MathPrecision;

pub use element::S4;

pub use element::S8;

pub use element::ScalarType;

pub use element::U4;

pub use element::U8;

pub use layout::ActivationKind;

pub use layout::ArchSku;

pub use layout::EpilogueKind;

pub use layout::LayoutSku;

pub use matrix::MatrixMut;

pub use matrix::MatrixRef;

pub use matrix::VectorRef;

pub use ops::ArgReduceKind;

pub use ops::AttentionKind;

pub use ops::BinaryCmpKind;

pub use ops::BinaryKind;

pub use ops::ConvKind;

pub use ops::CrossEntropyTargetKind;

pub use ops::EmbeddingKind;

pub use ops::FftKind;

pub use ops::FillMode;

pub use ops::GatedActivationKind;

pub use ops::GgufBlockFormat;

pub use ops::ImageKind;

pub use ops::IndexingKind;

pub use ops::LinalgKind;

pub use ops::LossKind;

pub use ops::LossReduction;

pub use ops::MoeKind;

pub use ops::NormalizationKind;

pub use ops::PadMode;

pub use ops::PoolKind;

pub use ops::QuantizeKind;

pub use ops::RandomKind;

pub use ops::ReduceKind;

pub use ops::ReduceToOp;

pub use ops::ScanKind;

pub use ops::SegmentKind;

pub use ops::ShapeLayoutKind;

pub use ops::SoftmaxKind;

pub use ops::SortKind;

pub use ops::TernaryKind;

pub use ops::UnaryKind;

pub use plan::PlanPreference;

pub use plan::PrecisionGuarantee;

pub use plan::Workspace;

pub use sku::BackendKind;

pub use sku::KernelSku;

pub use sku::OpCategory;

pub use tensor::contiguous_stride;

pub use tensor::strides_equal;

pub use tensor::TensorMut;

pub use tensor::TensorRef;

Modules

element

Element types and trait hierarchy shared across baracuda kernel wrappers.

layout

Layout / arch / epilogue / activation tags shared across kernel families.

matrix

Borrowed views of device-resident matrices and vectors.

ops

Per-category op discriminant enums.

plan

Plan-layer descriptors shared across kernel families: caller preferences, workspace handles, and the numerical-guarantee record every plan exposes.

sku

Generalized kernel SKU descriptor — covers every op category.

tensor

Borrowed views of device-resident N-dimensional tensors.