[GPU] Fix fused eltwise rank reduction

[wilson-seok]

Details

This fixes two Intel GPU accuracy issues in MotionDeblur/df1:

• Keep per-channel eltwise constants on the correct semantic axis when moving them through rank-changing Squeeze/Reshape operations. The GPU-local transformation accepts NumPy broadcasting only and rejects ambiguous, dynamic, overflowing, and PDPD cases before mutating the graph.
• Prevent higher-rank fused eltwise peers from being indexed with an incompatible reduced-rank host layout. A shared overflow-safe fold proof validates peers against the actual reduced layout. Rank-reducing reorder fusion is declined whenever any external peer cannot be represented exactly, preserving the producer rank instead.

The common transformation implementation remains unchanged, so CPU, NPU, and external common-pass users are unaffected.

Validation

Tested on Intel Arc B580 (GPU.1):

• Focused rank/fold matrix: 41/41, repeated 3 times
  • 7-case 5D broadcast matrix
  • all 16 spatial broadcast masks for a 6D host, including XW/XWY
  • 15 canonicalization guard tests
• Full Permute suite: 844/844
• Relevant neighboring fusion suite: 1957 passed, 21 hardware skips, 0 failures
• GPU-local transformation tests: 10/10
• Unchanged common transformation tests: 18/18

MotionDeblur/df1 GPU model MSE versus captured reference:

Configuration	Before	After
f16 default	2.383282e-02	9.524043e-06
f32 inference hint	2.376328e-02	3.140034e-11

After-fix model MSEs were also reproduced on Intel PTL-H (366H): 9.5240e-06 for default inference and 3.1400e-11 for the f32 hint.

All four f16 outputs are finite and stable; outputs 0/2/3 improve and output 1 is unchanged. The full runnable Blackmagic GPU.1 inventory shows no non-df1 MSE regression.

Output	BEFORE (none)	Fix 1/2 only	Fix 3 only	AFTER (all)	Fixed by
output_0	1.04e‑2	1.18e‑4	1.04e‑2	1.99e‑6	Fix 1/2
output_1	2.24e‑4	2.24e‑4	2.24e‑4	2.24e‑4	— (off both paths)
output_2	2.52e‑1	2.52e‑1	1.87e‑4	1.87e‑4	Fix 3
output_3	1.11	1.11	3.41e‑4	3.41e‑4	Fix 3

The two defects are orthogonal — each output is recovered by exactly one fix:

• Fix 1/2 (runtime fold) → recovers output_0 only. Fix 3 alone leaves it broken.
• Fix 3 (transformation) → recovers output_2 & output_3 only. Fix 1/2 alone leaves them broken.
• Fix 3 additionally sharpens output_0 in the full build (1.18e‑4 → 1.99e‑6, its per-channel input gets cleaner).

This corrects what I told you two turns ago. I had assumed Fix 3 → Defect A and Fix 1/2 → Defect B by reasoning; the measurement confirms the pairing but pins it to specific outputs, and shows neither fix alone is sufficient.

[Copilot]

Pull request overview

This PR addresses Intel GPU accuracy regressions caused by fused eltwise peers becoming rank-incompatible with a rank-reduced “host” layout (e.g., 5D peer being indexed as 4D after layout optimization), and by incorrectly moving per-channel constants across rank-changing data-movement ops.

Changes:

• Add an overflow-safe folding proof (fold_higher_rank_fused_peer) and use it in OCL fused-op shape canonicalization to safely represent higher-rank peers at the host rank when possible.
• Prevent rank-reducing reorder fusion when any fused higher-rank eltwise peer cannot be represented exactly at the reduced layout (keeps producer rank to preserve correct indexing).
• Introduce a GPU-local “move per-channel eltwise up through reshape/squeeze/unsqueeze” transformation and add targeted unit/regression tests (including df1-like reproducers and broadcast matrices).

Reviewed changes

Copilot reviewed 9 out of 9 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
src/plugins/intel_gpu/src/graph/impls/ocl/kernel_selector_helper.h	Canonicalize fused eltwise peer shapes by folding higher-rank peers to the host rank when provable; avoid assertions for unsupported cases.
src/plugins/intel_gpu/src/graph/include/fused_shape_utils.hpp	New shared, overflow-safe helper implementing the fold proof for higher-rank fused peers.
src/plugins/intel_gpu/src/graph/layout_optimizer.cpp	Gate rank-reducing reorder fusion on whether all higher-rank fused eltwise peers can fold to the reduced layout.
src/plugins/intel_gpu/src/plugin/transformations/move_eltwise_up_data_movement_per_channel.hpp	New GPU-local MatcherPass declaration for per-channel eltwise motion through rank-changing ops.
src/plugins/intel_gpu/src/plugin/transformations/move_eltwise_up_data_movement_per_channel.cpp	GPU-local transformation implementation with stricter guards (NumPy-only, static/overflow-safe/unique-axis checks).
src/plugins/intel_gpu/src/plugin/transformations_pipeline.cpp	Register the GPU-local per-channel eltwise move pass in the GPU transformation pipeline.
src/plugins/intel_gpu/tests/unit/transformations/move_eltwise_up_data_movement_per_channel_test.cpp	Unit tests validating legal moves and explicit rejection/guard cases for the new GPU-local pass.
src/plugins/intel_gpu/tests/unit/module_tests/canonicalize_fused_shapes_test.cpp	Unit coverage for the fold proof across equal-total, broadcast, padding/format/dynamic/overflow rejection scenarios.
src/plugins/intel_gpu/tests/unit/fusions/permute_fusion_test.cpp	Large end-to-end regression coverage reproducing df1-like rank mismatch and broadcast matrices, asserting both fusion state and numerical correctness.