[GPU] Canonicalize 3d shape for onednn conv/deconv post operations #32391

sungeunk · 2025-10-14T05:56:20Z

Description of the issue(symptom, root-cause, how it was resolved)

onednn 3d conv post-op mem_desc needs to be canonicalized to 4d when conv output is blocked

The code and line that caused this issue (if it is not changed directly)

src/plugins/intel_gpu/src/graph/program_node.cpp

Reproduction step and snapshot (if applicable. Do not attach for customer model)

reproduction step and model are attached in the ticket.

// need to convert IR: embedding_model.onnx -> FP32 -> INT8
$ ovc embedding_model.onnx --output_model model_FP32/embedding_model.xml --input "input[?,50,29]" --compress_to_fp16 False
$ python int8_quantization.py

// Run test
$ python openvino_script.py --device GPU.1 --model ov_onnx_model/int8/model_INT8.xml --batch 1

Problematic graph

It doesn't rely on graph patterns.

Checklist

Is it a proper fix? (not a workaround)
Did you include test case for this fix, if necessary?
Did you review existing test that can be extended to cover this scenario? Which test did you review?
-- No test for this issue.

Tickets:

174583

…ormat

sungeunk · 2025-10-14T12:18:03Z

[Fixed] ov_gpu_unit_tests issue can be reproduced on local machine(A770).

jade-cho

LGTM

sungeunk · 2025-10-15T05:38:01Z

Pass LLM daily test on BMG/A770/LNL

yeonbok · 2025-10-16T04:24:42Z

src/plugins/intel_gpu/src/graph/program_node.cpp

+
+                            dnnl::memory::desc in_scale_desc;
+                            if (is_type<gemm>() || is_type<fully_connected>()) {
+                                in_scale_desc = onednn::layout_to_memory_desc(in_scale, onednn::get_default_data_format(in_scale));


Don't gemm and fc need need_blocked when blocked?

It seems gemm/fc don't support the need_blocked flag. fc test-cases in unit-test are failed with the need_blocked flag.

Sorry but it is difficult to understand the logic. Let's discuss offline together with @jade-cho .

@jade-cho is "need_blocked" proper name? Maybe "allow_blocked" is the right name? actually the behavior seems similar to !flatten :(

why should we treat gemm and fc differently? If test-case fails, we may change the test case.

Maybe should we implement the logic withint the function itself?

discussed offline. We will clean this up after initial PR is merged.

isanghao · 2025-10-17T02:16:04Z

src/plugins/intel_gpu/src/graph/program_node.cpp

+                                    auto mem_flag = cldnn::format::is_blocked(get_output_layout().format) ?
+                                        onednn::mem_flags::need_blocked : onednn::mem_flags::None;
+                                    out_scale_desc = onednn::layout_to_memory_desc(out_scale, dnnl::memory::format_tag::undef, mem_flag);
+                                }


By Sungeun)
step1) Introduce lambda function

To be done by Jade)
step2) Introduce new function: fused_op_layout_to_memory_desc(fused_op_layout, layer_output_layout ...)
step2) change name: need_blocked --> respect_ov_layout
step2) remove conditional handling for gemm and FC

isanghao · 2025-10-17T02:16:34Z

src/plugins/intel_gpu/src/graph/program_node.cpp

+
+                            dnnl::memory::desc in_scale_desc;
+                            if (is_type<gemm>() || is_type<fully_connected>()) {
+                                in_scale_desc = onednn::layout_to_memory_desc(in_scale, onednn::get_default_data_format(in_scale));


discussed offline. We will clean this up after initial PR is merged.

isanghao · 2025-10-17T02:16:55Z

src/plugins/intel_gpu/src/graph/program_node.cpp

+                                    auto mem_flag = cldnn::format::is_blocked(get_output_layout().format) ?
+                                        onednn::mem_flags::need_blocked : onednn::mem_flags::None;
+                                    out_scale_desc = onednn::layout_to_memory_desc(out_scale, dnnl::memory::format_tag::undef, mem_flag);
+                                }


random spot) could you add a test where 1d conv is fused with quantize and mismatch happens between layout and rank?

isanghao · 2025-10-17T05:00:57Z

no perf issue from dGPU daily test

isanghao

LGTM

[GPU] set need_blocked to onednn::layout_to_memory_desc for blocked f…

95a76d6

…ormat

sungeunk added the category: GPU OpenVINO GPU plugin label Oct 14, 2025

sungeunk requested review from a team as code owners October 14, 2025 05:56

p-durandin added this to the 2025.4 milestone Oct 14, 2025

fix cpplint issues

456e9eb

not set need_blocked for gemm/fc due to unit-test failures.

bac797b

jade-cho approved these changes Oct 15, 2025

View reviewed changes

yeonbok reviewed Oct 16, 2025

View reviewed changes

isanghao reviewed Oct 17, 2025

View reviewed changes

sungeunk added 2 commits October 17, 2025 18:24

add unit-test

ae921d6

consolidate duplicate code with a lambda

6fe8e47

sungeunk changed the title ~~[GPU] set need_blocked to onednn::layout_to_memory_desc for blocked format~~ [GPU] Canonicalize 3d shape to handle blocked format for onednn conv/deconv Oct 17, 2025

sungeunk requested review from isanghao and jade-cho October 17, 2025 10:59

isanghao changed the title ~~[GPU] Canonicalize 3d shape to handle blocked format for onednn conv/deconv~~ [GPU] Canonicalize 3d shape for onednn conv/deconv post operations Oct 20, 2025

isanghao approved these changes Oct 20, 2025

View reviewed changes

isanghao added this pull request to the merge queue Oct 20, 2025

geunhwan added the Code Freeze label Oct 20, 2025

Merged via the queue into openvinotoolkit:master with commit e3a81e1 Oct 20, 2025
187 checks passed

[GPU] Canonicalize 3d shape for onednn conv/deconv post operations #32391

[GPU] Canonicalize 3d shape for onednn conv/deconv post operations #32391

Uh oh!

Conversation

sungeunk commented Oct 14, 2025 • edited by isanghao Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description of the issue(symptom, root-cause, how it was resolved)

The code and line that caused this issue (if it is not changed directly)

Reproduction step and snapshot (if applicable. Do not attach for customer model)

Problematic graph

Checklist

Tickets:

Uh oh!

sungeunk commented Oct 14, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

jade-cho left a comment

Choose a reason for hiding this comment

Uh oh!

sungeunk commented Oct 15, 2025

Uh oh!

yeonbok Oct 16, 2025

Choose a reason for hiding this comment

Uh oh!

sungeunk Oct 16, 2025

Choose a reason for hiding this comment

Uh oh!

isanghao Oct 16, 2025

Choose a reason for hiding this comment

Uh oh!

isanghao Oct 17, 2025

Choose a reason for hiding this comment

Uh oh!

isanghao Oct 17, 2025

Choose a reason for hiding this comment

Uh oh!

isanghao Oct 17, 2025

Choose a reason for hiding this comment

Uh oh!

isanghao Oct 17, 2025

Choose a reason for hiding this comment

Uh oh!

isanghao commented Oct 17, 2025

Uh oh!

isanghao left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

sungeunk commented Oct 14, 2025 •

edited by isanghao

Loading

sungeunk commented Oct 14, 2025 •

edited

Loading