Module optimizer 

Optimisation passes for the OxiCUDA computation graph.

Three complementary passes transform a validated ComputeGraph into a form ready for execution:

• fusion — Identifies chains of element-wise kernels that can be merged into a single PTX kernel, reducing kernel-launch overhead and improving cache behaviour.

• fusion_reduction — Identifies reduction-chained element-wise regions (LayerNorm, softmax, …) — single-entry/single-exit broadcast diamonds that the linear chain fuser cannot capture — and collapses each into one fused kernel.

• memory — Assigns device memory slots to logical buffers using live-interval graph colouring, maximising buffer reuse and minimising peak device memory usage.

• stream — Partitions independent nodes onto separate CUDA streams for concurrent GPU execution, and identifies the cross-stream event synchronisation points needed to enforce dependencies.

Re-exports

pub use fusion::FusionGroup;

pub use fusion::FusionPlan;

pub use fusion::analyse as fusion_analyse;

pub use fusion_reduction::ReductionFusionGroup;

pub use fusion_reduction::ReductionFusionPlan;

pub use fusion_reduction::ReductionPattern;

pub use fusion_reduction::analyse as reduction_fusion_analyse;

pub use fusion_reduction::rewrite as reduction_fusion_rewrite;

pub use memory::MemoryPlan;

pub use memory::SlotAssignment;

pub use memory::analyse as memory_analyse;

pub use stream::StreamPlan;

pub use stream::SyncPoint;

pub use stream::analyse as stream_analyse;

Modules

fusion

Operator fusion pass — merges chains of fusible element-wise kernels.

fusion_reduction

Reduction-pattern fusion pass — fuses reduction-chained element-wise regions (LayerNorm, softmax, …) that the linear element-wise fuser in crate::optimizer::fusion cannot capture.

memory

Memory planning — buffer allocation via live-interval graph colouring.

stream

Stream partitioning — assigns nodes to independent CUDA streams.