[GPU] Add SliceScatter-15 op support by wilson-seok · Pull Request #34085 · openvinotoolkit/openvino

wilson-seok · 2026-02-12T08:43:14Z

Details:

Extend GPU Plugin with operation "SliceScatter-15"
AI generated code

Tickets:

181088

Copilot

Pull request overview

This PR adds support for the SliceScatter-15 operation to the Intel GPU Plugin, implementing the complete operation stack from plugin layer through kernel execution, with comprehensive test coverage.

Changes:

Implements SliceScatter-15 operation with reference and optimized OpenCL kernels
Adds comprehensive unit and functional tests covering static/dynamic shapes and edge cases
Includes performance benchmarks for large tensor operations

Reviewed changes

Copilot reviewed 19 out of 19 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
src/plugins/intel_gpu/tests/unit/test_cases/slice_scatter_gpu_test.cpp	Unit tests for SliceScatter covering static/dynamic shapes, edge cases, and performance benchmarks
src/plugins/intel_gpu/tests/functional/shared_tests_instances/single_layer_tests/slice_scatter.cpp	Functional test instantiations for SliceScatter layer
src/plugins/intel_gpu/src/plugin/ops/slice_scatter.cpp	Plugin operation registration and primitive creation
src/plugins/intel_gpu/src/kernel_selector/kernels/slice_scatter/*.{h,cpp}	Kernel selector and implementations (reference and optimized)
src/plugins/intel_gpu/src/kernel_selector/common_types.h	Added SLICE_SCATTER kernel type enum
src/plugins/intel_gpu/src/kernel_selector/cl_kernels/slice_scatter_*.cl	OpenCL kernel implementations (reference and optimized)
src/plugins/intel_gpu/src/graph/slice_scatter.cpp	Graph-level primitive implementation and shape inference
src/plugins/intel_gpu/src/graph/include/slice_scatter_inst.h	Primitive instance header with helper classes
src/plugins/intel_gpu/src/graph/impls/ocl/slice_scatter.cpp	OCL implementation with kernel parameter preparation
src/plugins/intel_gpu/src/graph/impls/ocl/register.{hpp,cpp}	Registration of SliceScatter OCL implementation
src/plugins/intel_gpu/src/graph/registry/registry.hpp	Registry entry for SliceScatter primitive
src/plugins/intel_gpu/include/intel_gpu/primitives/slice_scatter.hpp	Public primitive header

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

src/plugins/intel_gpu/src/graph/impls/ocl/slice_scatter.cpp

Copilot

Pull request overview

Copilot reviewed 19 out of 19 changed files in this pull request and generated 3 comments.

src/plugins/intel_gpu/src/graph/impls/ocl_v2/slice_scatter_ref.cl

src/plugins/intel_gpu/src/graph/impls/ocl_v2/slice_scatter_opt.cl

src/plugins/intel_gpu/src/kernel_selector/kernels/slice_scatter/slice_scatter_kernel_ref.cpp

Copilot

Pull request overview

Copilot reviewed 20 out of 20 changed files in this pull request and generated 3 comments.

src/plugins/intel_gpu/src/kernel_selector/cl_kernels/slice_scatter_opt.cl

src/plugins/intel_gpu/src/kernel_selector/cl_kernels/slice_scatter_ref.cl

src/plugins/intel_gpu/tests/unit/test_cases/slice_scatter_gpu_test.cpp

isanghao · 2026-03-10T05:33:57Z

src/plugins/intel_gpu/src/graph/impls/ocl/slice_scatter.cpp

+namespace {
+template <typename T, class = typename std::enable_if<std::is_integral<T>::value>::type>
+std::vector<std::int64_t> extractIntegerData(const data_node& node, const stream& stream) {
+    mem_lock<T> lock{node.get_attached_memory_ptr(), stream};


please use mem_lock<T, mem_lock_type::read> because write does not happen.

Updated! Thanks!

isanghao · 2026-03-10T05:37:45Z

src/plugins/intel_gpu/src/graph/impls/ocl/slice_scatter.cpp

@@ -0,0 +1,208 @@
+// Copyright (C) 2018-2026 Intel Corporation


random spot) could you add this to ocl_v2 instead of ocl?

Updated! Thanks!

src/plugins/intel_gpu/src/graph/registry/registry.hpp

isanghao · 2026-03-11T07:13:11Z

src/plugins/intel_gpu/src/graph/registry/registry.hpp

 REGISTER_DEFAULT_IMPLS(roll, OCL_S);
 REGISTER_DEFAULT_IMPLS(shuffle_channels, OCL_S);
 REGISTER_DEFAULT_IMPLS(slice, OCL_S, OCL_D);
+REGISTER_IMPLS(slice_scatter);


for education purpose, could you tell me the difference between register_impls and register_default_impls?
By the way, those two categories are placed separately. If if is register_impls it should be placed near other register_impls.

Updated to proper position. Thanks!

maybe REGISTER_DEFAULT_IMPLS(slice_Scatter, OCL_S); would work with less code?

REGISTER_IMPLS() is better for current implementation. Because custom validate_impl() is required and
Using shape_types::any makes simpler code as slice_scatter impl supports static and dynamic.

- Add primitive definition (slice_scatter.hpp) - Add graph instance (slice_scatter_inst.h, slice_scatter.cpp) - Add ref OpenCL kernel with per-element scatter - Add opt OpenCL kernel with vectorized vstore8 for step=1 cases - Add kernel selector (ref + opt with priority-based selection) - Add OCL implementation bridge - Add plugin ops translation from ov::op::v15::SliceScatter - Register in common_types, register.hpp/cpp, registry.hpp - Add unit tests (84 tests: 3 data types x 28 cases) - Covers: basic bfyx, axes, step>1, negative step, clamping, INT_MAX stop, negative axes, spec examples, 1D/2D/5D tensors, dynamic shapes, caching - Add functional tests (60 tests: static, dynamic, precision) - Verified on GPU.0 (UHD 770 iGPU) and GPU.1 (Arc A770 dGPU)

…ter_kernel_utils.h Deduplicate identical addJitConstantsForParam function and associated constants (MAX_SUPPORTED_DIM, JIT_AXES_BUFF_SIZE_NAME) from slice_scatter_kernel_ref.cpp and slice_scatter_kernel_opt.cpp into a shared header slice_scatter_kernel_utils.h.

…tion, disable perf tests, use read-only mem_lock - Fix ceil division in slice_scatter_opt.cl bounds check for 4D/5D cases - Add axis range validation in slice_scatter_ref.cl and slice_scatter_opt.cl - Disable benchmark tests (DISABLED_ prefix) to avoid CI overhead - Use mem_lock<T, mem_lock_type::read> for read-only memory access

…orrect group

…registry.hpp

wilson-seok added the do_not_review label Feb 12, 2026

wilson-seok requested review from a team as code owners February 12, 2026 08:43

wilson-seok added the do_not_merge label Feb 12, 2026

wilson-seok requested review from Copilot and removed request for a team February 12, 2026 08:43

Copilot AI reviewed Feb 12, 2026

View reviewed changes

src/plugins/intel_gpu/src/graph/impls/ocl/slice_scatter.cpp Outdated Show resolved Hide resolved

github-actions bot added the category: GPU OpenVINO GPU plugin label Feb 12, 2026

wilson-seok removed do_not_review do_not_merge labels Feb 19, 2026

wilson-seok changed the title ~~[Do Not Merge][GPU] Add SliceScatter-15 op support~~ [GPU] Add SliceScatter-15 op support Feb 19, 2026

wilson-seok requested review from a team February 19, 2026 07:27

wilson-seok requested a review from Copilot February 27, 2026 14:00

Copilot started reviewing on behalf of wilson-seok February 27, 2026 14:01 View session

Copilot AI reviewed Feb 27, 2026

View reviewed changes

wilson-seok force-pushed the op_enablement/slice_scatter-15 branch from c6e8d4f to 7a9c887 Compare March 4, 2026 09:23

wilson-seok requested a review from Copilot March 4, 2026 11:13

Copilot started reviewing on behalf of wilson-seok March 4, 2026 11:14 View session

Copilot AI reviewed Mar 4, 2026

View reviewed changes

isanghao reviewed Mar 10, 2026

View reviewed changes

isanghao reviewed Mar 11, 2026

View reviewed changes

wilson-seok force-pushed the op_enablement/slice_scatter-15 branch from 5c2bdb5 to a55e7ea Compare March 18, 2026 01:11

wilson-seok added 6 commits March 18, 2026 10:44

Fix variable naming: transformed_axe -> transformed_axis

2652d38

Move slice_scatter from legacy ocl to ocl_v2 framework

82c2fd8

Fix clang-format: merge constructor to single line

f07eb9d

wilson-seok added 3 commits March 18, 2026 10:44

Apply clang-format to all changed files

e06afea

Revert style-only changes and move REGISTER_IMPLS(slice_scatter) to c…

a55e7ea

…orrect group

Move REGISTER_IMPLS(segment_max) to correct alphabetical position in …

6aa383a

…registry.hpp

		@@ -0,0 +1,208 @@
		// Copyright (C) 2018-2026 Intel Corporation

Conversation

wilson-seok commented Feb 12, 2026

Details:

Tickets:

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants