[dganellari]

Enable guarded producer/consumer fusion in the DaCe backend while retaining the original model field operators. Both extensions are opt-in and reuse existing GT4Py transformations.

• Extend VerticalSplitMapRange with allow_shared_data. Require disjoint pointwise writers and one covering writer for each reader. Preserve external stores and later consumers; reject unknown or unsafe partitions. Applications select candidates through the existing auto-optimizer callback.
• Add bounded scan-input fusion using fuse_as_fieldop once per original scan, with optional per-input selection, alias guards and shifted-input handling. The optional field_operator preparation combines local coefficient expressions within their original function boundary. Calls to upstream field operators remain boundaries. Run the standard field-view transforms before fusion and again after a successful rewrite.
• Convert array scalars before formatting the non-contiguous-domain warning, so a CuPy warning cannot abort compilation of an otherwise valid program.

The design ADR describes options, examples and safety limits. This branch is based on amd_chiplet_setting (eb763b97).

Reproduction with ICON4Py

Companion application integration and controlled runner: C2SM/icon4py#1484.

The reproduction instructions pin this compiler and the measured DaCe base. The runner selects the passes through normal ICON4Py model options, with separate or combined comparisons on global/120 and regional/120. It restores arrays and scalar state, checks 148 fields, uses balanced A/B quartets with interleaved A/A controls, and saves device/wall samples, SDFGs, generated code and source provenance. No compiler overlay or model rewrite is required.

Published controlled compiler-only measurements, original versus both passes:

GPU	Regional device-time reduction	Global device-time reduction
MI300A	5.99%	4.81%
GH200	2.06%	4.45%

These are the archived GPU experiments. The newly packaged public PR pair still requires its end-to-end GPU replay. The figures are combined-treatment results, not a universal profitability guarantee or isolated attribution to each pass. The ICON4Py PR documents wall time, timing controls, generated-code evidence and the original grid-dependent vendor gap.

Validation

• Focused fusion/backend suites: 74 tests passed in branch validation.
• Domain utility suite, including the array-scalar warning regression: 37 passed.
• gt4py.next doctests: 111 passed, 2 skipped.
• Staged pre-commit checks, including mypy and dependency boundaries, passed.

Requirements

• Fixes and features have corresponding tests.
• Design decisions are documented in the ADR.

[dganellari]

Split this into three PRs so the changes can be reviewed separately:

• perf[next-dace]: allow guarded fusion through shared outputs #2922: shared-output fusion in the DaCe transformations
• perf[next-dace]: add bounded scan-input fusion #2923: scan-input fusion at the GTIR level
• fix[next]: format domain warnings for array scalars #2924: domain-warning fix for array scalars

Closing this PR in favor of those three. Keeping this branch available because our earlier GPU results refer to it.