Description of the issue:

Issues

Symptom:

clBuildProgram kernel compilation failure when running super-resolution model (pnat-v1-fp16-576x672-2x-ox.onnx) on GPU. The failing node is depthtospace:/gnet/DepthToSpace with depth_to_space_ref implementation.

compile_graph: depthtospace:/gnet/DepthToSpace
  Type: depth_to_space
  Input layout 0: f16:b_fs_yx_fsv16:8x12x576x672 (4D)
  Input layout 1: f16:bfyx:8x3x1152x1344:nopad (4D)
  Output layout 0: f16:bfzyx:8x3x1x1152x1344 (5D)   ← problem here
  Selected impl: depth_to_space_ref

error: too many arguments provided to function-like macro invocation
  INPUT0_GET_INDEX(batch, input_feature, input_z, input_y, input_x)  ← 5 args
note: macro 'INPUT0_GET_INDEX' defined here
  #define INPUT0_GET_INDEX(b, f, y, x)  ← only accepts 4 args

Root causes

depth_to_space is rank-preserving by op semantics — input and output must have the same number of dimensions. However, can_fuse_reorder() and can_fuse_reorder_to_prev() in layout_optimizer.cpp unconditionally returned true for depth_to_space, allowing a rank-changing reorder (bfyx → bfzyx) to be fused in.

The fusion chain:

1. reorder_inputs pass inserts reorder_42 (bfyx → bfzyx) after depth_to_space for downstream Reshape_2
2. remove_redundant_reorders pass calls can_fuse_reorder_to_prev() which returns true unconditionally
3. reorder_42 is fused into depth_to_space, changing its output from 4D (bfyx) to 5D (bfzyx)
4. Kernel compilation: INPUT0 is 4D (b_fs_yx_fsv16) but OUTPUT is now 5D (bfzyx) → INPUT0_GET_INDEX macro argument count mismatch → clBuildProgram crash

How to fix it

Add a rank equality guard in two functions of layout_optimizer.cpp — separate depth_to_space from the blanket "return true" group and allow reorder fusion only when source and target ranks match:

Same-rank reorder fusion (e.g., bfyx → b_fs_yx_fsv16) remains allowed. The rank-changing reorder stays as a separate optimized node with memory sharing (zero-copy), so there is no performance penalty.

After fix rank-changing reorder fusion

Reproduced steps

benchmark_app -d GPU -m pnat-v1-fp16-576x672-2x-ox.onnx -nireq 1 -nstreams 1 -hint none

Tickets:

• CVS-179965

AI Assistance:

• AI assistance used: yes
• AI was used for debug log analysis (identifying the reorder fusion chain through 130K-line GPU debug logs), OpenVINO op spec audit (confirming depth_to_space is rank-preserving), and drafting test cases. All code changes were reviewed, the fix was validated by running the model end-to-end (121 iterations, 2.00 FPS), and unit tests were manually verified.