Transient buffer lifetime reuse

At a glance

Question	Answer
Problem	Reuse compatible physical storage for large logical transients with disjoint lifetimes.
Reads / writes	The planner consumes declared lifetimes and emits allocation assignments.
Ownership	Logical resources remain graph-owned; compiled allocations follow the plan.
Output contract	Deterministic logical-to-physical assignments with peak-byte accounting.
Expected work	CPU-side interval planning over declared transient requests.
Chunks	Chunk identity remains logical and is not implicitly repacked.
Conditions / budgets	Planning occurs at graph compilation and never changes an encoding topology.
Neighborhood	graph resource lifetimes → planner → compiled transient allocations.

CostPlanner bookkeeping versus reduced peak GPU memory for heavyweight scratch.

Common mistakeDo not recycle tiny arena slots when stable offsets are more valuable than negligible savings.

Overview

Large GPU algorithms frequently need scratch buffers that are never live at the same time. A sort workspace may be dead before an FFT temporary is needed; hierarchical reduction partials disappear before a later solver phase.

node →       0 1 2 3 4 5 6 7 8 9
sort scratch █████
FFT temp             ██████
reduction                    ███

Allocating all three independently makes peak physical memory equal to their sum even though their lifetimes do not overlap. A graph compiler can instead map compatible logical buffers onto fewer physical allocations.

What is reused

Lifetime reuse targets heavyweight transient buffers: reduction partials, sort workspace, FFT temporaries, sparse workspace and similar scratch resources.

It deliberately does not reuse GPUValueArena scalar slots. Hundreds of 32-bit values cost only hundreds or thousands of bytes, while stable scalar offsets make WGSL, traces and debugging substantially easier to understand.

Safety

Two logical buffers can share storage only when:

1. their graph-node lifetimes do not overlap;
2. the physical allocation satisfies the required usage flags;
3. the physical allocation is at least as large as the logical buffer.

The first planner uses greedy best-fit interval reuse. It is deterministic and intentionally conservative.

Example

A: 8 MiB, nodes 0–3 ─────┐
                         ├─ physical allocation 0: 8 MiB
B: 6 MiB, nodes 5–8 ─────┘

C: 4 MiB, nodes 2–7 ─────── physical allocation 1: 4 MiB

Logical storage is 18 MiB; peak physical storage is 12 MiB.

Compiler integration

This PR introduces the planning policy separately from physical graph allocation. The next integration step can derive lifetimes from resource accesses in the dependency-ordered graph and feed assignments into transient allocation. Keeping policy separate makes it straightforward to test aliasing safety before changing resource ownership.