[GPU] fix to embedding_bag_ref kernel

[Roszczyk]

Details:

• INPUT1_OFFSET was added twice. First, to start_indices and to end_indices, which are the range of the for loop and later to i in the same for loop. So finally indices_index was equal 2*INPUT1_OFFSET+offsets[offsets_ind]
• in line 122 INPUT3_OFFSET was used to calculate weight_index while INPUT4_OFFSET represents the offset for weights. Setting INPUT3 is not required, and the check is set only to checking if INPUT4 is set, so it even might result in some crashes (INPUT3_OFFSET not declared).

Tickets:

• CVS-192572

AI

• AI was used to speed up writing unit tests.

[Copilot]

Pull request overview

Fixes indexing/offset handling in the Intel GPU OpenCL reference kernel for embedding_bag, addressing incorrect double-application of INPUT1_OFFSET in offsets-based mode and incorrect offset selection for per-sample weights in segments-based mode.

Changes:

• Fix OFFSETS_SUM start/end range computation to avoid adding INPUT1_OFFSET twice.
• Fix SEGMENTS_SUM per-sample weights indexing to use INPUT4_OFFSET (weights input) instead of INPUT3_OFFSET.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated no new comments.

Suppressed comments (2)

src/plugins/intel_gpu/src/kernel_selector/cl_kernels/embedding_bag_ref.cl:122

• [HIGH] Add a weighted segments_sum regression test where the weights tensor has a different physical offset from input 3. The current weighted test uses dense tensors (embedding_bag_gpu_test.cpp:776-777), making both offsets zero, so it cannot distinguish INPUT3_OFFSET from INPUT4_OFFSET and would not catch the original bug.

                uint weight_index = INPUT4_OFFSET + i;

src/plugins/intel_gpu/src/kernel_selector/cl_kernels/embedding_bag_ref.cl:56

• [HIGH] Add a regression test that runs offsets_sum with a non-zero physical offset on the indices tensor. The existing GPU tests allocate dense, unpadded indices (for example, embedding_bag_gpu_test.cpp:376), so INPUT1_OFFSET is zero and they pass with both the old double-offset calculation and this fix. A padded/cropped indices input is needed to prove the corrected addressing and prevent this bug from returning.

    uint start_indices = offsets[offsets_ind];